diff --git "a/segmentation_1_4_0_fp32_combined/vaiml_par_0/0/tensor_dims.json" "b/segmentation_1_4_0_fp32_combined/vaiml_par_0/0/tensor_dims.json" new file mode 100644--- /dev/null +++ "b/segmentation_1_4_0_fp32_combined/vaiml_par_0/0/tensor_dims.json" @@ -0,0 +1,32071 @@ +{ + "__VERSION__": 1, + "Div_2": { + "MulAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 24, + "ifmsv_height": 20, + "ifmsv_size": 1920, + "ifmsv_width": 4, + "num_kernel_iters": 8, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 0.00390625, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 180, + 320, + 4 + ], + "dims": [ + 184, + 320, + 4 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 180, + 320, + 4 + ], + "dims": [ + 192, + 320, + 4 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Input.0", + "type": "external" + } + ], + "layer_order": 1 + }, + "Slice_7": { + "SliceHCWC8Adf": { + "aie_arch": "aie2p", + "axis_letter": "W", + "dim_c": 184, + "dim_h": 320, + "dim_w": 4, + "dtype": "bfloat16", + "end": 3, + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "num_ifm_shim_ch": 2, + "num_ofm_shim_ch": 2, + "output_l3_format": "HCWN_C8", + "start": 0, + "step": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 180, + 320, + 4 + ], + "dims": [ + 184, + 320, + 4 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 180, + 320, + 3 + ], + "dims": [ + 184, + 320, + 3 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Div_2", + "type": "internal" + } + ], + "layer_order": 2 + }, + "Generated-#0": { + "BufferPadAdf": { + "aie_arch": "aie2p", + "dim_0": 320, + "dim_0_padded": 320, + "dim_1": 23, + "dim_1_padded": 23, + "dim_2": 3, + "dim_2_padded": 8, + "dim_3": 8, + "dim_3_padded": 8, + "dtype": "bfloat16", + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "output_l3_format": "HCWN_C8" + }, + "dims_and_bounds": [ + { + "bounds": [ + 180, + 320, + 3 + ], + "dims": [ + 184, + 320, + 3 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 180, + 320, + 3 + ], + "dims": [ + 184, + 320, + 8 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Slice_7", + "type": "internal" + } + ], + "layer_order": 3 + }, + "Generated-#2": { + "Transpose4dAdf": { + "aie_arch": "aie2p", + "dim_0": 320, + "dim_1": 23, + "dim_2": 8, + "dim_3": 8, + "dtype": "bfloat16", + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "output_l3_format": "HCWN_C8", + "perm": 10 + }, + "dims_and_bounds": [ + { + "bounds": [ + 180, + 320, + 3 + ], + "dims": [ + 184, + 320, + 8 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 3, + 180, + 320 + ], + "dims": [ + 8, + 184, + 320 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Generated-#0", + "type": "internal" + } + ], + "layer_order": 4 + }, + "Generated-#4": { + "BufferUnpadAdf": { + "aie_arch": "aie2p", + "dim_0": 184, + "dim_0_unpadded": 180, + "dim_1": 1, + "dim_1_unpadded": 1, + "dim_2": 320, + "dim_2_unpadded": 320, + "dim_3": 8, + "dim_3_unpadded": 8, + "dtype": "bfloat16", + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "output_l3_format": "HCWN_C8" + }, + "dims_and_bounds": [ + { + "bounds": [ + 3, + 180, + 320 + ], + "dims": [ + 8, + 184, + 320 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 3, + 180, + 320 + ], + "dims": [ + 8, + 180, + 320 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Generated-#2", + "type": "internal" + } + ], + "layer_order": 5 + }, + "Sub_14": { + "AddBf16": { + "act": 0, + "act_type": "LINEAR", + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 8, + "ifmsv_height": 5, + "ifmsv_size": 640, + "ifmsv_width": 16, + "num_elems": 57600, + "num_kernel_iters": 180, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "constants_reader": { + "Const.0": { + "bytes_to_skip": 20764800, + "channel_stride": 8, + "depth": 1, + "dtype": "bfloat16", + "height": 180, + "nibbles": 4, + "num_reads": 86400, + "rank": 4, + "reads_per_line": 3, + "width": 320 + } + }, + "dims_and_bounds": [ + { + "bounds": [ + 3, + 180, + 320 + ], + "dims": [ + 8, + 180, + 320 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 3, + 180, + 320 + ], + "dims": [ + 8, + 180, + 320 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 3, + 180, + 320 + ], + "dims": [ + 8, + 180, + 320 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Generated-#4", + "type": "internal" + }, + { + "name": "Const.0", + "type": "constant" + } + ], + "layer_order": 6 + }, + "Div_16": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 8, + "ifmsv_height": 5, + "ifmsv_size": 640, + "ifmsv_width": 16, + "num_elems": 57600, + "num_kernel_iters": 180, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "constants_reader": { + "Const.1": { + "bytes_to_skip": 20764800, + "channel_stride": 8, + "depth": 1, + "dtype": "bfloat16", + "height": 180, + "nibbles": 4, + "num_reads": 86400, + "rank": 4, + "reads_per_line": 3, + "width": 320 + } + }, + "dims_and_bounds": [ + { + "bounds": [ + 3, + 180, + 320 + ], + "dims": [ + 8, + 180, + 320 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 3, + 180, + 320 + ], + "dims": [ + 8, + 180, + 320 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 3, + 180, + 320 + ], + "dims": [ + 8, + 180, + 320 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Sub_14", + "type": "internal" + }, + { + "name": "Const.1", + "type": "constant" + } + ], + "layer_order": 7 + }, + "Conv_17": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 64, + "data_io.ofm.dimensions.width": 32, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 1, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 8, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 1, + "ifm_height": 180, + "ifm_sign": 0, + "ifm_sv_depth": 8, + "ifm_sv_height": 10, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 66, + "ifm_width": 320, + "ifm_x_iter": 5, + "ifm_y_iter": 6, + "kernel_height": 3, + "kernel_width": 3, + "ksize.height": 3, + "ksize.width": 3, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 1, + "num_ifm_height_iters": 6, + "num_ifm_width_iters": 5, + "num_ofm_depth_iters": 1, + "ofm_depth": 16, + "ofm_depth_iter": 1, + "ofm_height": 90, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 8, + "ofm_sv_height": 4, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 32, + "ofm_width": 160, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 1, + "pad_right": 0, + "pad_top": 1, + "pad_value": 0, + "padding_sv_height": 10, + "padding_sv_width": 66, + "psum_buff_offset_scaled": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 2, + "stride_height": 2, + "stride_log2.stride_log2_height": 1, + "stride_log2.stride_log2_width": 1, + "stride_w": 2, + "stride_width": 2, + "zp_en": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 3, + 180, + 320 + ], + "dims": [ + 8, + 180, + 320 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Div_16", + "type": "internal" + } + ], + "layer_order": 8, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 144, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 8, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 3, + "kernel_width": 3, + "krnl_w_div": 1, + "num_bias_sub_groups": 1, + "ofm_depth_iter": 1, + "ofm_depth_padded": 32, + "output_stride": 8, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 0, + "wts_no_of_reads": 864 + } + }, + "Add_19": { + "AddAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 8, + "ifmsv_height": 23, + "ifmsv_size": 1472, + "ifmsv_width": 8, + "num_kernel_iters": 20, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 3, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 32, + 92, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_17", + "type": "internal" + } + ], + "layer_order": 9 + }, + "Clip_22": { + "ClipBf16": { + "aie_arch": "aie2p", + "clamp_max": 6, + "clamp_min": 0, + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 8, + "ifmsv_height": 12, + "ifmsv_size": 1920, + "ifmsv_width": 20, + "num_kernel_iters": 16, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Add_19", + "type": "internal" + } + ], + "layer_order": 10 + }, + "Div_24": { + "MulAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 8, + "ifmsv_height": 23, + "ifmsv_size": 1472, + "ifmsv_width": 8, + "num_kernel_iters": 20, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 0.166015625, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 32, + 92, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Clip_22", + "type": "internal" + } + ], + "layer_order": 11 + }, + "Mul_25": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 8, + "ifmsv_height": 8, + "ifmsv_size": 1024, + "ifmsv_width": 16, + "num_elems": 14400, + "num_kernel_iters": 30, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_17", + "type": "internal" + }, + { + "name": "Div_24", + "type": "internal" + } + ], + "layer_order": 12 + }, + "Conv_26": { + "Conv2DBf16": { + "act": 1, + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 0, + "batch_size": 1, + "bias_shift": 0, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 0, + "data_io.ifm.dimensions.batch.padding": 0, + "data_io.ifm.dimensions.batch.size": 1, + "data_io.ofm.dimensions.batch.padding": 0, + "data_io.ofm.dimensions.batch.size": 1, + "data_io.ofm.dimensions.width.padding": 0, + "func_type": 10, + "ifm_depth": 16, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 1, + "ifm_height": 90, + "ifm_sign": 0, + "ifm_sv_depth": 8, + "ifm_sv_depth.size": 8, + "ifm_sv_height": 8, + "ifm_sv_height.size": 8, + "ifm_sv_width": 18, + "ifm_sv_width.size": 18, + "ifm_width": 160, + "ifm_x_iter": 10, + "ifm_y_iter": 4, + "ifmsv_size": 1280, + "kernel_height": 3, + "kernel_width": 3, + "num_adf_cols": 4, + "num_adf_rows": 4, + "ofm_depth": 32, + "ofm_depth_iter": 1, + "ofm_height": 90, + "ofm_len": 40, + "ofm_sv_depth": 8, + "ofm_sv_height": 6, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 17, + "ofm_width": 159, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 1, + "pad_left": 1, + "pad_right": 1, + "pad_top": 1, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 1, + "stride": 1, + "stride_height": 1, + "stride_width": 1, + "zp_en": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 32, + 96, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Mul_25", + "type": "internal" + } + ], + "layer_order": 13, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 144, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 8, + "is_dwc": true, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 3, + "kernel_width": 4, + "krnl_w_div": 1, + "num_bias_sub_groups": 1, + "ofm_depth_iter": 1, + "ofm_depth_padded": 32, + "output_stride": 8, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 10944, + "wts_no_of_reads": 144 + } + }, + "Conv_28": { + "AddBf16": { + "act": 0, + "act_type": "LINEAR", + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 8, + "ifmsv_height": 2, + "ifmsv_size": 512, + "ifmsv_width": 32, + "num_elems": 14400, + "num_kernel_iters": 60, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 64, + "data_io.ofm.dimensions.width": 32, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 11, + "ifm_depth": 16, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 1, + "ifm_height": 90, + "ifm_sign": 0, + "ifm_sv_depth": 64, + "ifm_sv_height": 2, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 32, + "ifm_width": 160, + "ifm_x_iter": 5, + "ifm_y_iter": 12, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 1, + "num_ifm_height_iters": 12, + "num_ifm_width_iters": 5, + "num_ofm_depth_iters": 1, + "ofm_depth": 16, + "ofm_depth_iter": 1, + "ofm_height": 90, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 8, + "ofm_sv_height": 2, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 32, + "ofm_width": 160, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 2, + "padding_sv_width": 32, + "psum_buff_offset_scaled": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_26", + "type": "internal" + }, + { + "name": "Mul_25", + "type": "internal" + } + ], + "layer_order": 14, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 144, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 64, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 1, + "ofm_depth_iter": 1, + "ofm_depth_padded": 32, + "output_stride": 8, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 13248, + "wts_no_of_reads": 768 + } + }, + "Conv_30": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 1, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 12, + "data_io.ofm.dimensions.width": 8, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 16, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 1, + "ifm_height": 90, + "ifm_sign": 0, + "ifm_sv_depth": 64, + "ifm_sv_height": 9, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 8, + "ifm_width": 160, + "ifm_x_iter": 5, + "ifm_y_iter": 10, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 0, + "lrelu_alpha_kernel": 0, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 1, + "num_ifm_height_iters": 10, + "num_ifm_width_iters": 5, + "num_ofm_depth_iters": 1, + "ofm_depth": 64, + "ofm_depth_iter": 1, + "ofm_height": 90, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 16, + "ofm_sv_height": 9, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 8, + "ofm_width": 160, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 9, + "padding_sv_width": 8, + "psum_buff_offset_scaled": 1, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 64, + 90, + 160 + ], + "dims": [ + 64, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_28", + "type": "internal" + } + ], + "layer_order": 15, + "split_width": true, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 288, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 64, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 2, + "ofm_depth_iter": 1, + "ofm_depth_padded": 64, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 23040, + "wts_no_of_reads": 1536 + } + }, + "Conv_32": { + "Conv2DBf16": { + "act": 1, + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 0, + "batch_size": 1, + "bias_shift": 0, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 0, + "data_io.ifm.dimensions.batch.padding": 0, + "data_io.ifm.dimensions.batch.size": 1, + "data_io.ofm.dimensions.batch.padding": 0, + "data_io.ofm.dimensions.batch.size": 1, + "data_io.ofm.dimensions.width.padding": 0, + "func_type": 10, + "ifm_depth": 64, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 1, + "ifm_height": 160, + "ifm_sign": 0, + "ifm_sv_depth": 8, + "ifm_sv_depth.size": 8, + "ifm_sv_height": 6, + "ifm_sv_height.size": 6, + "ifm_sv_width": 26, + "ifm_sv_width.size": 26, + "ifm_width": 90, + "ifm_x_iter": 7, + "ifm_y_iter": 6, + "ifmsv_size": 1344, + "kernel_height": 3, + "kernel_width": 3, + "num_adf_cols": 4, + "num_adf_rows": 4, + "ofm_depth": 64, + "ofm_depth_iter": 2, + "ofm_height": 80, + "ofm_len": 84, + "ofm_sv_depth": 8, + "ofm_sv_height": 2, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 13, + "ofm_width": 44, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 1, + "pad_right": 0, + "pad_top": 1, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 1, + "stride": 2, + "stride_height": 2, + "stride_width": 2, + "zp_en": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 64, + 90, + 160 + ], + "dims": [ + 64, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 64, + 45, + 80 + ], + "dims": [ + 64, + 48, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_30", + "type": "internal" + } + ], + "layer_order": 16, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 144, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 8, + "is_dwc": true, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 3, + "kernel_width": 4, + "krnl_w_div": 1, + "num_bias_sub_groups": 1, + "ofm_depth_iter": 2, + "ofm_depth_padded": 64, + "output_stride": 8, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 42624, + "wts_no_of_reads": 144 + } + }, + "Generated-#60": { + "BufferUnpadAdf": { + "aie_arch": "aie2p", + "dim_0": 48, + "dim_0_unpadded": 45, + "dim_1": 8, + "dim_1_unpadded": 8, + "dim_2": 80, + "dim_2_unpadded": 80, + "dim_3": 8, + "dim_3_unpadded": 8, + "dtype": "bfloat16", + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "output_l3_format": "HCWN_C8" + }, + "dims_and_bounds": [ + { + "bounds": [ + 64, + 45, + 80 + ], + "dims": [ + 64, + 48, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 64, + 45, + 80 + ], + "dims": [ + 64, + 45, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_32", + "type": "internal" + } + ], + "layer_order": 17 + }, + "Conv_34": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 0, + "data_io.ofm.dimensions.width": 16, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 64, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 1, + "ifm_height": 45, + "ifm_sign": 0, + "ifm_sv_depth": 64, + "ifm_sv_height": 4, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 16, + "ifm_width": 80, + "ifm_x_iter": 5, + "ifm_y_iter": 3, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 1, + "num_ifm_height_iters": 3, + "num_ifm_width_iters": 5, + "num_ofm_depth_iters": 1, + "ofm_depth": 24, + "ofm_depth_iter": 1, + "ofm_height": 45, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 16, + "ofm_sv_height": 4, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 16, + "ofm_width": 80, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 4, + "padding_sv_width": 16, + "psum_buff_offset_scaled": 1, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 64, + 45, + 80 + ], + "dims": [ + 64, + 45, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 24, + 45, + 80 + ], + "dims": [ + 64, + 48, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Generated-#60", + "type": "internal" + } + ], + "layer_order": 18, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 288, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 64, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 2, + "ofm_depth_iter": 1, + "ofm_depth_padded": 64, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 47232, + "wts_no_of_reads": 1536 + } + }, + "Conv_35": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 1, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 64, + "data_io.ofm.dimensions.width": 32, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 24, + "ifm_depth_concat_extend": 40, + "ifm_depth_iter": 1, + "ifm_height": 45, + "ifm_sign": 0, + "ifm_sv_depth": 64, + "ifm_sv_height": 2, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 32, + "ifm_width": 80, + "ifm_x_iter": 3, + "ifm_y_iter": 6, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 0, + "lrelu_alpha_kernel": 0, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 1, + "num_ifm_height_iters": 6, + "num_ifm_width_iters": 3, + "num_ofm_depth_iters": 1, + "ofm_depth": 72, + "ofm_depth_iter": 1, + "ofm_height": 45, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 24, + "ofm_sv_height": 2, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 32, + "ofm_width": 80, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 2, + "padding_sv_width": 32, + "psum_buff_offset_scaled": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 24, + 45, + 80 + ], + "dims": [ + 64, + 48, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 72, + 45, + 80 + ], + "dims": [ + 96, + 48, + 96 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_34", + "type": "internal" + } + ], + "layer_order": 19, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 432, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 64, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 3, + "ofm_depth_iter": 1, + "ofm_depth_padded": 96, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 66816, + "wts_no_of_reads": 2304 + } + }, + "Conv_37": { + "Conv2DBf16": { + "act": 1, + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 0, + "batch_size": 1, + "bias_shift": 0, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 0, + "data_io.ifm.dimensions.batch.padding": 0, + "data_io.ifm.dimensions.batch.size": 1, + "data_io.ofm.dimensions.batch.padding": 0, + "data_io.ofm.dimensions.batch.size": 1, + "data_io.ofm.dimensions.width.padding": 0, + "func_type": 10, + "ifm_depth": 72, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 1, + "ifm_height": 80, + "ifm_sign": 0, + "ifm_sv_depth": 8, + "ifm_sv_depth.size": 8, + "ifm_sv_height": 8, + "ifm_sv_height.size": 8, + "ifm_sv_width": 18, + "ifm_sv_width.size": 18, + "ifm_width": 45, + "ifm_x_iter": 5, + "ifm_y_iter": 2, + "ifmsv_size": 1280, + "kernel_height": 3, + "kernel_width": 3, + "num_adf_cols": 4, + "num_adf_rows": 4, + "ofm_depth": 72, + "ofm_depth_iter": 3, + "ofm_height": 80, + "ofm_len": 30, + "ofm_sv_depth": 8, + "ofm_sv_height": 6, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 17, + "ofm_width": 44, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 1, + "pad_left": 1, + "pad_right": 1, + "pad_top": 1, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 1, + "stride": 1, + "stride_height": 1, + "stride_width": 1, + "zp_en": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 72, + 45, + 80 + ], + "dims": [ + 72, + 45, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 72, + 45, + 80 + ], + "dims": [ + 72, + 46, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_35", + "type": "internal" + } + ], + "layer_order": 20, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 144, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 8, + "is_dwc": true, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 3, + "kernel_width": 4, + "krnl_w_div": 1, + "num_bias_sub_groups": 1, + "ofm_depth_iter": 3, + "ofm_depth_padded": 96, + "output_stride": 8, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 96192, + "wts_no_of_reads": 144 + } + }, + "Generated-#62": { + "BufferUnpadAdf": { + "aie_arch": "aie2p", + "dim_0": 46, + "dim_0_unpadded": 45, + "dim_1": 9, + "dim_1_unpadded": 9, + "dim_2": 80, + "dim_2_unpadded": 80, + "dim_3": 8, + "dim_3_unpadded": 8, + "dtype": "bfloat16", + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "output_l3_format": "HCWN_C8" + }, + "dims_and_bounds": [ + { + "bounds": [ + 72, + 45, + 80 + ], + "dims": [ + 72, + 46, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 72, + 45, + 80 + ], + "dims": [ + 72, + 45, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_37", + "type": "internal" + } + ], + "layer_order": 21 + }, + "Conv_39": { + "AddBf16": { + "act": 0, + "act_type": "LINEAR", + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 4, + "ifmsv_size": 1024, + "ifmsv_width": 16, + "num_elems": 3600, + "num_kernel_iters": 15, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 0, + "data_io.ofm.dimensions.width": 16, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 11, + "ifm_depth": 72, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 1, + "ifm_height": 45, + "ifm_sign": 0, + "ifm_sv_depth": 72, + "ifm_sv_height": 4, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 16, + "ifm_width": 80, + "ifm_x_iter": 5, + "ifm_y_iter": 3, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 1, + "num_ifm_height_iters": 3, + "num_ifm_width_iters": 5, + "num_ofm_depth_iters": 1, + "ofm_depth": 24, + "ofm_depth_iter": 1, + "ofm_height": 45, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 16, + "ofm_sv_height": 4, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 16, + "ofm_width": 80, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 4, + "padding_sv_width": 16, + "psum_buff_offset_scaled": 1, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 72, + 45, + 80 + ], + "dims": [ + 72, + 45, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 24, + 45, + 80 + ], + "dims": [ + 64, + 48, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 24, + 45, + 80 + ], + "dims": [ + 64, + 48, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Generated-#62", + "type": "internal" + }, + { + "name": "Conv_34", + "type": "internal" + } + ], + "layer_order": 22, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 288, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 72, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 2, + "ofm_depth_iter": 1, + "ofm_depth_padded": 64, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 103104, + "wts_no_of_reads": 1728 + } + }, + "Conv_41": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 1, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 64, + "data_io.ofm.dimensions.width": 32, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 24, + "ifm_depth_concat_extend": 40, + "ifm_depth_iter": 1, + "ifm_height": 45, + "ifm_sign": 0, + "ifm_sv_depth": 64, + "ifm_sv_height": 2, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 32, + "ifm_width": 80, + "ifm_x_iter": 3, + "ifm_y_iter": 6, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 0, + "lrelu_alpha_kernel": 0, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 1, + "num_ifm_height_iters": 6, + "num_ifm_width_iters": 3, + "num_ofm_depth_iters": 1, + "ofm_depth": 72, + "ofm_depth_iter": 1, + "ofm_height": 45, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 24, + "ofm_sv_height": 2, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 32, + "ofm_width": 80, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 2, + "padding_sv_width": 32, + "psum_buff_offset_scaled": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 24, + 45, + 80 + ], + "dims": [ + 64, + 48, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 72, + 45, + 80 + ], + "dims": [ + 96, + 48, + 96 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_39", + "type": "internal" + } + ], + "layer_order": 23, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 432, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 64, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 3, + "ofm_depth_iter": 1, + "ofm_depth_padded": 96, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 124992, + "wts_no_of_reads": 2304 + } + }, + "Conv_43": { + "Conv2DBf16": { + "act": 1, + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 0, + "batch_size": 1, + "bias_shift": 0, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 0, + "data_io.ifm.dimensions.batch.padding": 0, + "data_io.ifm.dimensions.batch.size": 1, + "data_io.ofm.dimensions.batch.padding": 0, + "data_io.ofm.dimensions.batch.size": 1, + "data_io.ofm.dimensions.width.padding": 0, + "func_type": 10, + "ifm_depth": 96, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 1, + "ifm_height": 48, + "ifm_sign": 0, + "ifm_sv_depth": 8, + "ifm_sv_depth.size": 8, + "ifm_sv_height": 8, + "ifm_sv_height.size": 8, + "ifm_sv_width": 20, + "ifm_sv_width.size": 20, + "ifm_width": 96, + "ifm_x_iter": 5, + "ifm_y_iter": 3, + "ifmsv_size": 1280, + "kernel_height": 5, + "kernel_width": 5, + "num_adf_cols": 4, + "num_adf_rows": 4, + "ofm_depth": 96, + "ofm_depth_iter": 3, + "ofm_height": 24, + "ofm_len": 45, + "ofm_sv_depth": 8, + "ofm_sv_height": 2, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 7, + "ofm_width": 46, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 1, + "pad_left": 2, + "pad_right": 2, + "pad_top": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 1, + "stride": 2, + "stride_height": 2, + "stride_width": 2, + "zp_en": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 72, + 45, + 80 + ], + "dims": [ + 96, + 48, + 96 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 72, + 23, + 40 + ], + "dims": [ + 96, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_41", + "type": "internal" + } + ], + "layer_order": 24, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 144, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 8, + "is_dwc": true, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 5, + "kernel_width": 8, + "krnl_w_div": 1, + "num_bias_sub_groups": 1, + "ofm_depth_iter": 3, + "ofm_depth_padded": 96, + "output_stride": 8, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 154368, + "wts_no_of_reads": 480 + } + }, + "Generated-#6": { + "Transpose4dAdf": { + "aie_arch": "aie2p", + "dim_0": 23, + "dim_1": 9, + "dim_2": 40, + "dim_3": 8, + "dtype": "bfloat16", + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "output_l3_format": "HCWN_C8", + "perm": 6 + }, + "dims_and_bounds": [ + { + "bounds": [ + 72, + 23, + 40 + ], + "dims": [ + 72, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 72, + 1, + 920 + ], + "dims": [ + 72, + 1, + 920 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_43", + "type": "internal" + } + ], + "layer_order": 25 + }, + "Generated-#8": { + "ReduceMeanC8Bf16": { + "L3_depth": 72, + "L3_height": 1, + "L3_width": 920, + "aie_arch": "aie2p", + "compiler": "chess", + "depth_iter": 1, + "full_channel": 72, + "full_height": 1, + "full_width": 920, + "height_iter": 1, + "loop_num_channel": 3, + "loop_num_height": 1, + "loop_num_width": 15, + "num_iters": 15, + "pad_value": 0, + "reduce_axis": 2, + "reduce_dim": "W", + "scale": 0.0010833740234375, + "scale_shift": 0, + "sv_channel": 32, + "sv_height": 1, + "sv_width": 64, + "width_iter": 15 + }, + "dims_and_bounds": [ + { + "bounds": [ + 72, + 1, + 920 + ], + "dims": [ + 72, + 1, + 920 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 72, + 1, + 1 + ], + "dims": [ + 128, + 4, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Generated-#6", + "type": "internal" + } + ], + "layer_order": 26 + }, + "Conv_46": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 1, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 12, + "data_io.ofm.dimensions.width": 16, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 72, + "ifm_depth_concat_extend": 56, + "ifm_depth_iter": 1, + "ifm_height": 1, + "ifm_sign": 0, + "ifm_sv_depth": 72, + "ifm_sv_height": 1, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 16, + "ifm_width": 1, + "ifm_x_iter": 1, + "ifm_y_iter": 1, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 0, + "lrelu_alpha_kernel": 0, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 1, + "num_ifm_height_iters": 1, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 1, + "ofm_depth": 24, + "ofm_depth_iter": 1, + "ofm_height": 1, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 16, + "ofm_sv_height": 1, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 16, + "ofm_width": 1, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 1, + "padding_sv_width": 16, + "psum_buff_offset_scaled": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 72, + 1, + 1 + ], + "dims": [ + 128, + 4, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 24, + 1, + 1 + ], + "dims": [ + 64, + 1, + 64 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Generated-#8", + "type": "internal" + } + ], + "layer_order": 27, + "split_width": true, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 288, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 72, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 2, + "ofm_depth_iter": 1, + "ofm_depth_padded": 64, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 173376, + "wts_no_of_reads": 1728 + } + }, + "Conv_48": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 12, + "data_io.ofm.dimensions.width": 8, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 24, + "ifm_depth_concat_extend": 40, + "ifm_depth_iter": 1, + "ifm_height": 1, + "ifm_sign": 0, + "ifm_sv_depth": 64, + "ifm_sv_height": 1, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 8, + "ifm_width": 1, + "ifm_x_iter": 1, + "ifm_y_iter": 1, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 1, + "num_ifm_height_iters": 1, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 1, + "ofm_depth": 72, + "ofm_depth_iter": 1, + "ofm_height": 1, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 32, + "ofm_sv_height": 1, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 8, + "ofm_width": 1, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 1, + "padding_sv_width": 8, + "psum_buff_offset_scaled": 1, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 24, + 1, + 1 + ], + "dims": [ + 64, + 1, + 64 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 72, + 1, + 1 + ], + "dims": [ + 128, + 1, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_46", + "type": "internal" + } + ], + "layer_order": 28, + "split_width": true, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 576, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 64, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 4, + "ofm_depth_iter": 1, + "ofm_depth_padded": 128, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 195264, + "wts_no_of_reads": 3072 + } + }, + "Add_50": { + "AddAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 24, + "ifmsv_height": 16, + "ifmsv_size": 384, + "ifmsv_width": 1, + "num_kernel_iters": 1, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 3, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 72, + 1, + 1 + ], + "dims": [ + 128, + 1, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 72, + 1, + 1 + ], + "dims": [ + 96, + 64, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_48", + "type": "internal" + } + ], + "layer_order": 29 + }, + "Clip_53": { + "ClipBf16": { + "aie_arch": "aie2p", + "clamp_max": 6, + "clamp_min": 0, + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 32, + "ifmsv_height": 8, + "ifmsv_size": 256, + "ifmsv_width": 1, + "num_kernel_iters": 1, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 72, + 1, + 1 + ], + "dims": [ + 96, + 64, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 72, + 1, + 1 + ], + "dims": [ + 128, + 32, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Add_50", + "type": "internal" + } + ], + "layer_order": 30 + }, + "Div_55": { + "MulAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 32, + "ifmsv_height": 16, + "ifmsv_size": 512, + "ifmsv_width": 1, + "num_kernel_iters": 1, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 0.166015625, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 72, + 1, + 1 + ], + "dims": [ + 128, + 32, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 72, + 1, + 1 + ], + "dims": [ + 128, + 64, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Clip_53", + "type": "internal" + } + ], + "layer_order": 31 + }, + "Generated-#10": { + "TileAdf": { + "aie_arch": "aie2p", + "dtype": "bfloat16", + "i_dim_c": 72, + "i_dim_h": 1, + "i_dim_n": 1, + "i_dim_w": 1, + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "output_l3_format": "HCWN_C8", + "rep_dim_c": 1, + "rep_dim_h": 23, + "rep_dim_w": 40 + }, + "dims_and_bounds": [ + { + "bounds": [ + 72, + 1, + 1 + ], + "dims": [ + 72, + 1, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 72, + 23, + 40 + ], + "dims": [ + 72, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Div_55", + "type": "internal" + } + ], + "layer_order": 32 + }, + "Mul_56": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 8, + "ifmsv_height": 6, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_elems": 920, + "num_kernel_iters": 6, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 72, + 23, + 40 + ], + "dims": [ + 72, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 72, + 23, + 40 + ], + "dims": [ + 96, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 72, + 23, + 40 + ], + "dims": [ + 96, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Generated-#10", + "type": "internal" + }, + { + "name": "Conv_43", + "type": "internal" + } + ], + "layer_order": 33 + }, + "Conv_57": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 12, + "data_io.ofm.dimensions.width": 8, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 72, + "ifm_depth_concat_extend": 24, + "ifm_depth_iter": 1, + "ifm_height": 23, + "ifm_sign": 0, + "ifm_sv_depth": 72, + "ifm_sv_height": 6, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 8, + "ifm_width": 40, + "ifm_x_iter": 5, + "ifm_y_iter": 1, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 1, + "num_ifm_height_iters": 1, + "num_ifm_width_iters": 5, + "num_ofm_depth_iters": 1, + "ofm_depth": 40, + "ofm_depth_iter": 1, + "ofm_height": 23, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 16, + "ofm_sv_height": 6, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 8, + "ofm_width": 40, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 6, + "padding_sv_width": 8, + "psum_buff_offset_scaled": 1, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 72, + 23, + 40 + ], + "dims": [ + 96, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 40, + 23, + 40 + ], + "dims": [ + 64, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Mul_56", + "type": "internal" + } + ], + "layer_order": 34, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 288, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 72, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 2, + "ofm_depth_iter": 1, + "ofm_depth_padded": 64, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 234432, + "wts_no_of_reads": 1728 + } + }, + "Conv_58": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 1, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 64, + "data_io.ofm.dimensions.width": 32, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 40, + "ifm_depth_concat_extend": 24, + "ifm_depth_iter": 1, + "ifm_height": 23, + "ifm_sign": 0, + "ifm_sv_depth": 64, + "ifm_sv_height": 2, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 32, + "ifm_width": 40, + "ifm_x_iter": 2, + "ifm_y_iter": 3, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 0, + "lrelu_alpha_kernel": 0, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 1, + "num_ifm_height_iters": 3, + "num_ifm_width_iters": 2, + "num_ofm_depth_iters": 1, + "ofm_depth": 120, + "ofm_depth_iter": 1, + "ofm_height": 23, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 32, + "ofm_sv_height": 2, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 32, + "ofm_width": 40, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 2, + "padding_sv_width": 32, + "psum_buff_offset_scaled": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 40, + 23, + 40 + ], + "dims": [ + 64, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 120, + 23, + 40 + ], + "dims": [ + 128, + 24, + 64 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_57", + "type": "internal" + } + ], + "layer_order": 35, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 576, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 64, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 4, + "ofm_depth_iter": 1, + "ofm_depth_padded": 128, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 256320, + "wts_no_of_reads": 3072 + } + }, + "Conv_60": { + "Conv2DBf16": { + "act": 1, + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 0, + "batch_size": 1, + "bias_shift": 0, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 0, + "data_io.ifm.dimensions.batch.padding": 0, + "data_io.ifm.dimensions.batch.size": 1, + "data_io.ofm.dimensions.batch.padding": 0, + "data_io.ofm.dimensions.batch.size": 1, + "data_io.ofm.dimensions.width.padding": 0, + "func_type": 10, + "ifm_depth": 128, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 1, + "ifm_height": 24, + "ifm_sign": 0, + "ifm_sv_depth": 8, + "ifm_sv_depth.size": 8, + "ifm_sv_height": 10, + "ifm_sv_height.size": 10, + "ifm_sv_width": 12, + "ifm_sv_width.size": 12, + "ifm_width": 64, + "ifm_x_iter": 5, + "ifm_y_iter": 1, + "ifmsv_size": 960, + "kernel_height": 5, + "kernel_width": 5, + "num_adf_cols": 4, + "num_adf_rows": 4, + "ofm_depth": 128, + "ofm_depth_iter": 4, + "ofm_height": 24, + "ofm_len": 20, + "ofm_sv_depth": 8, + "ofm_sv_height": 6, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 5, + "ofm_width": 61, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 2, + "pad_left": 2, + "pad_right": 2, + "pad_top": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 1, + "stride": 1, + "stride_height": 1, + "stride_width": 1, + "zp_en": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 120, + 23, + 40 + ], + "dims": [ + 128, + 24, + 64 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 120, + 23, + 40 + ], + "dims": [ + 128, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_58", + "type": "internal" + } + ], + "layer_order": 36, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 144, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 8, + "is_dwc": true, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 5, + "kernel_width": 8, + "krnl_w_div": 1, + "num_bias_sub_groups": 1, + "ofm_depth_iter": 4, + "ofm_depth_padded": 128, + "output_stride": 8, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 295488, + "wts_no_of_reads": 480 + } + }, + "Generated-#12": { + "Transpose4dAdf": { + "aie_arch": "aie2p", + "dim_0": 23, + "dim_1": 15, + "dim_2": 40, + "dim_3": 8, + "dtype": "bfloat16", + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "output_l3_format": "HCWN_C8", + "perm": 6 + }, + "dims_and_bounds": [ + { + "bounds": [ + 120, + 23, + 40 + ], + "dims": [ + 120, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 120, + 1, + 920 + ], + "dims": [ + 120, + 1, + 920 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_60", + "type": "internal" + } + ], + "layer_order": 37 + }, + "Generated-#14": { + "ReduceMeanC8Bf16": { + "L3_depth": 120, + "L3_height": 1, + "L3_width": 920, + "aie_arch": "aie2p", + "compiler": "chess", + "depth_iter": 1, + "full_channel": 120, + "full_height": 1, + "full_width": 920, + "height_iter": 1, + "loop_num_channel": 4, + "loop_num_height": 1, + "loop_num_width": 15, + "num_iters": 15, + "pad_value": 0, + "reduce_axis": 2, + "reduce_dim": "W", + "scale": 0.0010833740234375, + "scale_shift": 0, + "sv_channel": 32, + "sv_height": 1, + "sv_width": 64, + "width_iter": 15 + }, + "dims_and_bounds": [ + { + "bounds": [ + 120, + 1, + 920 + ], + "dims": [ + 120, + 1, + 920 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 120, + 1, + 1 + ], + "dims": [ + 128, + 4, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Generated-#12", + "type": "internal" + } + ], + "layer_order": 38 + }, + "Conv_63": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 1, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 12, + "data_io.ofm.dimensions.width": 16, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 120, + "ifm_depth_concat_extend": 8, + "ifm_depth_iter": 1, + "ifm_height": 1, + "ifm_sign": 0, + "ifm_sv_depth": 120, + "ifm_sv_height": 1, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 16, + "ifm_width": 1, + "ifm_x_iter": 1, + "ifm_y_iter": 1, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 0, + "lrelu_alpha_kernel": 0, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 1, + "num_ifm_height_iters": 1, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 1, + "ofm_depth": 32, + "ofm_depth_iter": 1, + "ofm_height": 1, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 16, + "ofm_sv_height": 1, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 16, + "ofm_width": 1, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 1, + "padding_sv_width": 16, + "psum_buff_offset_scaled": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 120, + 1, + 1 + ], + "dims": [ + 128, + 4, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 32, + 1, + 1 + ], + "dims": [ + 64, + 1, + 64 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Generated-#14", + "type": "internal" + } + ], + "layer_order": 39, + "split_width": true, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 288, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 120, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 2, + "ofm_depth_iter": 1, + "ofm_depth_padded": 64, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 320832, + "wts_no_of_reads": 2880 + } + }, + "Conv_65": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 12, + "data_io.ofm.dimensions.width": 8, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 32, + "ifm_depth_concat_extend": 32, + "ifm_depth_iter": 1, + "ifm_height": 1, + "ifm_sign": 0, + "ifm_sv_depth": 64, + "ifm_sv_height": 1, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 8, + "ifm_width": 1, + "ifm_x_iter": 1, + "ifm_y_iter": 1, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 1, + "num_ifm_height_iters": 1, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 1, + "ofm_depth": 120, + "ofm_depth_iter": 1, + "ofm_height": 1, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 32, + "ofm_sv_height": 1, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 8, + "ofm_width": 1, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 1, + "padding_sv_width": 8, + "psum_buff_offset_scaled": 1, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 32, + 1, + 1 + ], + "dims": [ + 64, + 1, + 64 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 120, + 1, + 1 + ], + "dims": [ + 128, + 1, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_63", + "type": "internal" + } + ], + "layer_order": 40, + "split_width": true, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 576, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 64, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 4, + "ofm_depth_iter": 1, + "ofm_depth_padded": 128, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 356544, + "wts_no_of_reads": 3072 + } + }, + "Add_67": { + "AddAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 32, + "ifmsv_height": 12, + "ifmsv_size": 384, + "ifmsv_width": 1, + "num_kernel_iters": 1, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 3, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 120, + 1, + 1 + ], + "dims": [ + 128, + 1, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 120, + 1, + 1 + ], + "dims": [ + 128, + 48, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_65", + "type": "internal" + } + ], + "layer_order": 41 + }, + "Clip_70": { + "ClipBf16": { + "aie_arch": "aie2p", + "clamp_max": 6, + "clamp_min": 0, + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 32, + "ifmsv_height": 8, + "ifmsv_size": 256, + "ifmsv_width": 1, + "num_kernel_iters": 1, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 120, + 1, + 1 + ], + "dims": [ + 128, + 48, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 120, + 1, + 1 + ], + "dims": [ + 128, + 32, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Add_67", + "type": "internal" + } + ], + "layer_order": 42 + }, + "Div_72": { + "MulAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 32, + "ifmsv_height": 16, + "ifmsv_size": 512, + "ifmsv_width": 1, + "num_kernel_iters": 1, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 0.166015625, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 120, + 1, + 1 + ], + "dims": [ + 128, + 32, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 120, + 1, + 1 + ], + "dims": [ + 128, + 64, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Clip_70", + "type": "internal" + } + ], + "layer_order": 43 + }, + "Generated-#16": { + "TileAdf": { + "aie_arch": "aie2p", + "dtype": "bfloat16", + "i_dim_c": 120, + "i_dim_h": 1, + "i_dim_n": 1, + "i_dim_w": 1, + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "output_l3_format": "HCWN_C8", + "rep_dim_c": 1, + "rep_dim_h": 23, + "rep_dim_w": 40 + }, + "dims_and_bounds": [ + { + "bounds": [ + 120, + 1, + 1 + ], + "dims": [ + 120, + 1, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 120, + 23, + 40 + ], + "dims": [ + 120, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Div_72", + "type": "internal" + } + ], + "layer_order": 44 + }, + "Mul_73": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 6, + "ifmsv_size": 960, + "ifmsv_width": 10, + "num_elems": 920, + "num_kernel_iters": 8, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 120, + 23, + 40 + ], + "dims": [ + 120, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 120, + 23, + 40 + ], + "dims": [ + 128, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 120, + 23, + 40 + ], + "dims": [ + 128, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Generated-#16", + "type": "internal" + }, + { + "name": "Conv_60", + "type": "internal" + } + ], + "layer_order": 45 + }, + "Conv_74": { + "AddBf16": { + "act": 0, + "act_type": "LINEAR", + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 1, + "ifmsv_size": 640, + "ifmsv_width": 40, + "num_elems": 920, + "num_kernel_iters": 6, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 12, + "data_io.ofm.dimensions.width": 40, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 11, + "ifm_depth": 120, + "ifm_depth_concat_extend": 8, + "ifm_depth_iter": 2, + "ifm_height": 23, + "ifm_sign": 0, + "ifm_sv_depth": 64, + "ifm_sv_height": 1, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 40, + "ifm_width": 40, + "ifm_x_iter": 1, + "ifm_y_iter": 6, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 2, + "num_ifm_height_iters": 6, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 1, + "ofm_depth": 40, + "ofm_depth_iter": 1, + "ofm_height": 23, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 16, + "ofm_sv_height": 1, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 40, + "ofm_width": 40, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 1, + "padding_sv_width": 40, + "psum_buff_offset_scaled": 5, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 120, + 23, + 40 + ], + "dims": [ + 128, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 40, + 23, + 40 + ], + "dims": [ + 64, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 40, + 23, + 40 + ], + "dims": [ + 64, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Mul_73", + "type": "internal" + }, + { + "name": "Conv_57", + "type": "internal" + } + ], + "layer_order": 46, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 288, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 2, + "ifm_depth_padded": 128, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 2, + "ofm_depth_iter": 1, + "ofm_depth_padded": 64, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 395712, + "wts_no_of_reads": 1536 + } + }, + "Conv_76": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 1, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 64, + "data_io.ofm.dimensions.width": 32, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 40, + "ifm_depth_concat_extend": 24, + "ifm_depth_iter": 1, + "ifm_height": 23, + "ifm_sign": 0, + "ifm_sv_depth": 64, + "ifm_sv_height": 2, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 32, + "ifm_width": 40, + "ifm_x_iter": 2, + "ifm_y_iter": 3, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 0, + "lrelu_alpha_kernel": 0, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 1, + "num_ifm_height_iters": 3, + "num_ifm_width_iters": 2, + "num_ofm_depth_iters": 1, + "ofm_depth": 120, + "ofm_depth_iter": 1, + "ofm_height": 23, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 32, + "ofm_sv_height": 2, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 32, + "ofm_width": 40, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 2, + "padding_sv_width": 32, + "psum_buff_offset_scaled": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 40, + 23, + 40 + ], + "dims": [ + 64, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 120, + 23, + 40 + ], + "dims": [ + 128, + 24, + 64 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_74", + "type": "internal" + } + ], + "layer_order": 47, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 576, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 64, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 4, + "ofm_depth_iter": 1, + "ofm_depth_padded": 128, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 434880, + "wts_no_of_reads": 3072 + } + }, + "Conv_78": { + "Conv2DBf16": { + "act": 1, + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 0, + "batch_size": 1, + "bias_shift": 0, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 0, + "data_io.ifm.dimensions.batch.padding": 0, + "data_io.ifm.dimensions.batch.size": 1, + "data_io.ofm.dimensions.batch.padding": 0, + "data_io.ofm.dimensions.batch.size": 1, + "data_io.ofm.dimensions.width.padding": 0, + "func_type": 10, + "ifm_depth": 128, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 1, + "ifm_height": 24, + "ifm_sign": 0, + "ifm_sv_depth": 8, + "ifm_sv_depth.size": 8, + "ifm_sv_height": 10, + "ifm_sv_height.size": 10, + "ifm_sv_width": 12, + "ifm_sv_width.size": 12, + "ifm_width": 64, + "ifm_x_iter": 5, + "ifm_y_iter": 1, + "ifmsv_size": 960, + "kernel_height": 5, + "kernel_width": 5, + "num_adf_cols": 4, + "num_adf_rows": 4, + "ofm_depth": 128, + "ofm_depth_iter": 4, + "ofm_height": 24, + "ofm_len": 20, + "ofm_sv_depth": 8, + "ofm_sv_height": 6, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 5, + "ofm_width": 61, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 2, + "pad_left": 2, + "pad_right": 2, + "pad_top": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 1, + "stride": 1, + "stride_height": 1, + "stride_width": 1, + "zp_en": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 120, + 23, + 40 + ], + "dims": [ + 128, + 24, + 64 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 120, + 23, + 40 + ], + "dims": [ + 128, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_76", + "type": "internal" + } + ], + "layer_order": 48, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 144, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 8, + "is_dwc": true, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 5, + "kernel_width": 8, + "krnl_w_div": 1, + "num_bias_sub_groups": 1, + "ofm_depth_iter": 4, + "ofm_depth_padded": 128, + "output_stride": 8, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 474048, + "wts_no_of_reads": 480 + } + }, + "Generated-#18": { + "Transpose4dAdf": { + "aie_arch": "aie2p", + "dim_0": 23, + "dim_1": 15, + "dim_2": 40, + "dim_3": 8, + "dtype": "bfloat16", + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "output_l3_format": "HCWN_C8", + "perm": 6 + }, + "dims_and_bounds": [ + { + "bounds": [ + 120, + 23, + 40 + ], + "dims": [ + 120, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 120, + 1, + 920 + ], + "dims": [ + 120, + 1, + 920 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_78", + "type": "internal" + } + ], + "layer_order": 49 + }, + "Generated-#20": { + "ReduceMeanC8Bf16": { + "L3_depth": 120, + "L3_height": 1, + "L3_width": 920, + "aie_arch": "aie2p", + "compiler": "chess", + "depth_iter": 1, + "full_channel": 120, + "full_height": 1, + "full_width": 920, + "height_iter": 1, + "loop_num_channel": 4, + "loop_num_height": 1, + "loop_num_width": 15, + "num_iters": 15, + "pad_value": 0, + "reduce_axis": 2, + "reduce_dim": "W", + "scale": 0.0010833740234375, + "scale_shift": 0, + "sv_channel": 32, + "sv_height": 1, + "sv_width": 64, + "width_iter": 15 + }, + "dims_and_bounds": [ + { + "bounds": [ + 120, + 1, + 920 + ], + "dims": [ + 120, + 1, + 920 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 120, + 1, + 1 + ], + "dims": [ + 128, + 4, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Generated-#18", + "type": "internal" + } + ], + "layer_order": 50 + }, + "Conv_81": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 1, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 12, + "data_io.ofm.dimensions.width": 16, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 120, + "ifm_depth_concat_extend": 8, + "ifm_depth_iter": 1, + "ifm_height": 1, + "ifm_sign": 0, + "ifm_sv_depth": 120, + "ifm_sv_height": 1, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 16, + "ifm_width": 1, + "ifm_x_iter": 1, + "ifm_y_iter": 1, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 0, + "lrelu_alpha_kernel": 0, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 1, + "num_ifm_height_iters": 1, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 1, + "ofm_depth": 32, + "ofm_depth_iter": 1, + "ofm_height": 1, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 16, + "ofm_sv_height": 1, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 16, + "ofm_width": 1, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 1, + "padding_sv_width": 16, + "psum_buff_offset_scaled": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 120, + 1, + 1 + ], + "dims": [ + 128, + 4, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 32, + 1, + 1 + ], + "dims": [ + 64, + 1, + 64 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Generated-#20", + "type": "internal" + } + ], + "layer_order": 51, + "split_width": true, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 288, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 120, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 2, + "ofm_depth_iter": 1, + "ofm_depth_padded": 64, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 499392, + "wts_no_of_reads": 2880 + } + }, + "Conv_83": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 12, + "data_io.ofm.dimensions.width": 8, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 32, + "ifm_depth_concat_extend": 32, + "ifm_depth_iter": 1, + "ifm_height": 1, + "ifm_sign": 0, + "ifm_sv_depth": 64, + "ifm_sv_height": 1, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 8, + "ifm_width": 1, + "ifm_x_iter": 1, + "ifm_y_iter": 1, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 1, + "num_ifm_height_iters": 1, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 1, + "ofm_depth": 120, + "ofm_depth_iter": 1, + "ofm_height": 1, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 32, + "ofm_sv_height": 1, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 8, + "ofm_width": 1, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 1, + "padding_sv_width": 8, + "psum_buff_offset_scaled": 1, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 32, + 1, + 1 + ], + "dims": [ + 64, + 1, + 64 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 120, + 1, + 1 + ], + "dims": [ + 128, + 1, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_81", + "type": "internal" + } + ], + "layer_order": 52, + "split_width": true, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 576, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 64, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 4, + "ofm_depth_iter": 1, + "ofm_depth_padded": 128, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 535104, + "wts_no_of_reads": 3072 + } + }, + "Add_85": { + "AddAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 32, + "ifmsv_height": 12, + "ifmsv_size": 384, + "ifmsv_width": 1, + "num_kernel_iters": 1, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 3, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 120, + 1, + 1 + ], + "dims": [ + 128, + 1, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 120, + 1, + 1 + ], + "dims": [ + 128, + 48, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_83", + "type": "internal" + } + ], + "layer_order": 53 + }, + "Clip_88": { + "ClipBf16": { + "aie_arch": "aie2p", + "clamp_max": 6, + "clamp_min": 0, + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 32, + "ifmsv_height": 8, + "ifmsv_size": 256, + "ifmsv_width": 1, + "num_kernel_iters": 1, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 120, + 1, + 1 + ], + "dims": [ + 128, + 48, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 120, + 1, + 1 + ], + "dims": [ + 128, + 32, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Add_85", + "type": "internal" + } + ], + "layer_order": 54 + }, + "Div_90": { + "MulAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 32, + "ifmsv_height": 16, + "ifmsv_size": 512, + "ifmsv_width": 1, + "num_kernel_iters": 1, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 0.166015625, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 120, + 1, + 1 + ], + "dims": [ + 128, + 32, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 120, + 1, + 1 + ], + "dims": [ + 128, + 64, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Clip_88", + "type": "internal" + } + ], + "layer_order": 55 + }, + "Generated-#22": { + "TileAdf": { + "aie_arch": "aie2p", + "dtype": "bfloat16", + "i_dim_c": 120, + "i_dim_h": 1, + "i_dim_n": 1, + "i_dim_w": 1, + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "output_l3_format": "HCWN_C8", + "rep_dim_c": 1, + "rep_dim_h": 23, + "rep_dim_w": 40 + }, + "dims_and_bounds": [ + { + "bounds": [ + 120, + 1, + 1 + ], + "dims": [ + 120, + 1, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 120, + 23, + 40 + ], + "dims": [ + 120, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Div_90", + "type": "internal" + } + ], + "layer_order": 56 + }, + "Mul_91": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 6, + "ifmsv_size": 960, + "ifmsv_width": 10, + "num_elems": 920, + "num_kernel_iters": 8, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 120, + 23, + 40 + ], + "dims": [ + 120, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 120, + 23, + 40 + ], + "dims": [ + 128, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 120, + 23, + 40 + ], + "dims": [ + 128, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Generated-#22", + "type": "internal" + }, + { + "name": "Conv_78", + "type": "internal" + } + ], + "layer_order": 57 + }, + "Conv_92": { + "AddBf16": { + "act": 0, + "act_type": "LINEAR", + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 1, + "ifmsv_size": 640, + "ifmsv_width": 40, + "num_elems": 920, + "num_kernel_iters": 6, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 12, + "data_io.ofm.dimensions.width": 40, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 11, + "ifm_depth": 120, + "ifm_depth_concat_extend": 8, + "ifm_depth_iter": 2, + "ifm_height": 23, + "ifm_sign": 0, + "ifm_sv_depth": 64, + "ifm_sv_height": 1, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 40, + "ifm_width": 40, + "ifm_x_iter": 1, + "ifm_y_iter": 6, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 2, + "num_ifm_height_iters": 6, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 1, + "ofm_depth": 40, + "ofm_depth_iter": 1, + "ofm_height": 23, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 16, + "ofm_sv_height": 1, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 40, + "ofm_width": 40, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 1, + "padding_sv_width": 40, + "psum_buff_offset_scaled": 5, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 120, + 23, + 40 + ], + "dims": [ + 128, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 40, + 23, + 40 + ], + "dims": [ + 64, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 40, + 23, + 40 + ], + "dims": [ + 64, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Mul_91", + "type": "internal" + }, + { + "name": "Conv_74", + "type": "internal" + } + ], + "layer_order": 58, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 288, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 2, + "ifm_depth_padded": 128, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 2, + "ofm_depth_iter": 1, + "ofm_depth_padded": 64, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 574272, + "wts_no_of_reads": 1536 + } + }, + "Conv_94": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 64, + "data_io.ofm.dimensions.width": 32, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 40, + "ifm_depth_concat_extend": 24, + "ifm_depth_iter": 1, + "ifm_height": 23, + "ifm_sign": 0, + "ifm_sv_depth": 64, + "ifm_sv_height": 2, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 32, + "ifm_width": 40, + "ifm_x_iter": 2, + "ifm_y_iter": 3, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 1, + "num_ifm_height_iters": 3, + "num_ifm_width_iters": 2, + "num_ofm_depth_iters": 2, + "ofm_depth": 240, + "ofm_depth_iter": 2, + "ofm_height": 23, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 32, + "ofm_sv_height": 2, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 32, + "ofm_width": 40, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 2, + "padding_sv_width": 32, + "psum_buff_offset_scaled": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 40, + 23, + 40 + ], + "dims": [ + 64, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 240, + 23, + 40 + ], + "dims": [ + 256, + 24, + 64 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_92", + "type": "internal" + } + ], + "layer_order": 59, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 576, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 64, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 4, + "ofm_depth_iter": 2, + "ofm_depth_padded": 256, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 613440, + "wts_no_of_reads": 3072 + } + }, + "Add_96": { + "AddAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 6, + "ifmsv_size": 1920, + "ifmsv_width": 20, + "num_kernel_iters": 8, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 3, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 240, + 23, + 40 + ], + "dims": [ + 256, + 24, + 64 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 240, + 23, + 40 + ], + "dims": [ + 256, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_94", + "type": "internal" + } + ], + "layer_order": 60 + }, + "Clip_99": { + "ClipBf16": { + "aie_arch": "aie2p", + "clamp_max": 6, + "clamp_min": 0, + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 6, + "ifmsv_size": 1920, + "ifmsv_width": 20, + "num_kernel_iters": 8, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 240, + 23, + 40 + ], + "dims": [ + 256, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 240, + 23, + 40 + ], + "dims": [ + 256, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Add_96", + "type": "internal" + } + ], + "layer_order": 61 + }, + "Div_101": { + "MulAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 6, + "ifmsv_size": 1920, + "ifmsv_width": 20, + "num_kernel_iters": 8, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 0.166015625, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 240, + 23, + 40 + ], + "dims": [ + 256, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 240, + 23, + 40 + ], + "dims": [ + 256, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Clip_99", + "type": "internal" + } + ], + "layer_order": 62 + }, + "Mul_102": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_elems": 920, + "num_kernel_iters": 16, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 240, + 23, + 40 + ], + "dims": [ + 240, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 240, + 23, + 40 + ], + "dims": [ + 240, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 240, + 23, + 40 + ], + "dims": [ + 240, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_94", + "type": "internal" + }, + { + "name": "Div_101", + "type": "internal" + } + ], + "layer_order": 63 + }, + "Generated-#64": { + "BufferUnpadAdf": { + "aie_arch": "aie2p", + "dim_0": 24, + "dim_0_unpadded": 23, + "dim_1": 30, + "dim_1_unpadded": 30, + "dim_2": 40, + "dim_2_unpadded": 40, + "dim_3": 8, + "dim_3_unpadded": 8, + "dtype": "bfloat16", + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "output_l3_format": "HCWN_C8" + }, + "dims_and_bounds": [ + { + "bounds": [ + 240, + 23, + 40 + ], + "dims": [ + 240, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 240, + 23, + 40 + ], + "dims": [ + 240, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Mul_102", + "type": "internal" + } + ], + "layer_order": 64 + }, + "Conv_103": { + "Conv2DBf16": { + "act": 0, + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 0, + "batch_size": 1, + "bias_shift": 0, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 0, + "data_io.ifm.dimensions.batch.padding": 0, + "data_io.ifm.dimensions.batch.size": 1, + "data_io.ofm.dimensions.batch.padding": 0, + "data_io.ofm.dimensions.batch.size": 1, + "data_io.ofm.dimensions.width.padding": 0, + "func_type": 10, + "ifm_depth": 240, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 1, + "ifm_height": 23, + "ifm_sign": 0, + "ifm_sv_depth": 8, + "ifm_sv_depth.size": 8, + "ifm_sv_height": 10, + "ifm_sv_height.size": 10, + "ifm_sv_width": 10, + "ifm_sv_width.size": 10, + "ifm_width": 40, + "ifm_x_iter": 5, + "ifm_y_iter": 1, + "ifmsv_size": 960, + "kernel_height": 3, + "kernel_width": 3, + "num_adf_cols": 4, + "num_adf_rows": 4, + "ofm_depth": 256, + "ofm_depth_iter": 8, + "ofm_height": 12, + "ofm_len": 40, + "ofm_sv_depth": 8, + "ofm_sv_height": 4, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 5, + "ofm_width": 19, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 1, + "pad_right": 1, + "pad_top": 1, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 1, + "stride": 2, + "stride_height": 2, + "stride_width": 2, + "zp_en": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 240, + 23, + 40 + ], + "dims": [ + 240, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 240, + 12, + 20 + ], + "dims": [ + 256, + 16, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Generated-#64", + "type": "internal" + } + ], + "layer_order": 65, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 144, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 8, + "is_dwc": true, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 3, + "kernel_width": 4, + "krnl_w_div": 1, + "num_bias_sub_groups": 1, + "ofm_depth_iter": 8, + "ofm_depth_padded": 256, + "output_stride": 8, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 691776, + "wts_no_of_reads": 144 + } + }, + "Add_105": { + "AddAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 32, + "ifmsv_height": 3, + "ifmsv_size": 1920, + "ifmsv_width": 20, + "num_kernel_iters": 2, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 3, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 240, + 12, + 20 + ], + "dims": [ + 256, + 16, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 240, + 12, + 20 + ], + "dims": [ + 256, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_103", + "type": "internal" + } + ], + "layer_order": 66 + }, + "Clip_108": { + "ClipBf16": { + "aie_arch": "aie2p", + "clamp_max": 6, + "clamp_min": 0, + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 32, + "ifmsv_height": 3, + "ifmsv_size": 1920, + "ifmsv_width": 20, + "num_kernel_iters": 2, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 240, + 12, + 20 + ], + "dims": [ + 256, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 240, + 12, + 20 + ], + "dims": [ + 256, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Add_105", + "type": "internal" + } + ], + "layer_order": 67 + }, + "Div_110": { + "MulAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 32, + "ifmsv_height": 3, + "ifmsv_size": 1920, + "ifmsv_width": 20, + "num_kernel_iters": 2, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 0.166015625, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 240, + 12, + 20 + ], + "dims": [ + 256, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 240, + 12, + 20 + ], + "dims": [ + 256, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Clip_108", + "type": "internal" + } + ], + "layer_order": 68 + }, + "Mul_111": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_elems": 240, + "num_kernel_iters": 4, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 240, + 12, + 20 + ], + "dims": [ + 256, + 16, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 240, + 12, + 20 + ], + "dims": [ + 256, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 240, + 12, + 20 + ], + "dims": [ + 256, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_103", + "type": "internal" + }, + { + "name": "Div_110", + "type": "internal" + } + ], + "layer_order": 69 + }, + "Conv_112": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 64, + "data_io.ofm.dimensions.width": 32, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 240, + "ifm_depth_concat_extend": 16, + "ifm_depth_iter": 3, + "ifm_height": 12, + "ifm_sign": 0, + "ifm_sv_depth": 80, + "ifm_sv_height": 1, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 32, + "ifm_width": 20, + "ifm_x_iter": 1, + "ifm_y_iter": 3, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 3, + "num_ifm_height_iters": 3, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 1, + "ofm_depth": 80, + "ofm_depth_iter": 1, + "ofm_height": 12, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 24, + "ofm_sv_height": 1, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 32, + "ofm_width": 20, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 1, + "padding_sv_width": 32, + "psum_buff_offset_scaled": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 240, + 12, + 20 + ], + "dims": [ + 256, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 80, + 12, + 20 + ], + "dims": [ + 96, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Mul_111", + "type": "internal" + } + ], + "layer_order": 70, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 432, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 3, + "ifm_depth_padded": 240, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 3, + "ofm_depth_iter": 1, + "ofm_depth_padded": 96, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 710208, + "wts_no_of_reads": 2880 + } + }, + "Conv_113": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 64, + "data_io.ofm.dimensions.width": 32, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 80, + "ifm_depth_concat_extend": 16, + "ifm_depth_iter": 1, + "ifm_height": 12, + "ifm_sign": 0, + "ifm_sv_depth": 80, + "ifm_sv_height": 1, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 32, + "ifm_width": 20, + "ifm_x_iter": 1, + "ifm_y_iter": 3, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 1, + "num_ifm_height_iters": 3, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 2, + "ofm_depth": 200, + "ofm_depth_iter": 2, + "ofm_height": 12, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 32, + "ofm_sv_height": 1, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 32, + "ofm_width": 20, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 1, + "padding_sv_width": 32, + "psum_buff_offset_scaled": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 80, + 12, + 20 + ], + "dims": [ + 96, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 200, + 12, + 20 + ], + "dims": [ + 256, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_112", + "type": "internal" + } + ], + "layer_order": 71, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 576, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 80, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 4, + "ofm_depth_iter": 2, + "ofm_depth_padded": 256, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 819072, + "wts_no_of_reads": 3840 + } + }, + "Add_115": { + "AddAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 32, + "ifmsv_height": 3, + "ifmsv_size": 1920, + "ifmsv_width": 20, + "num_kernel_iters": 2, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 3, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 200, + 12, + 20 + ], + "dims": [ + 256, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 200, + 12, + 20 + ], + "dims": [ + 256, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_113", + "type": "internal" + } + ], + "layer_order": 72 + }, + "Clip_118": { + "ClipBf16": { + "aie_arch": "aie2p", + "clamp_max": 6, + "clamp_min": 0, + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 32, + "ifmsv_height": 3, + "ifmsv_size": 1920, + "ifmsv_width": 20, + "num_kernel_iters": 2, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 200, + 12, + 20 + ], + "dims": [ + 256, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 200, + 12, + 20 + ], + "dims": [ + 256, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Add_115", + "type": "internal" + } + ], + "layer_order": 73 + }, + "Div_120": { + "MulAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 32, + "ifmsv_height": 3, + "ifmsv_size": 1920, + "ifmsv_width": 20, + "num_kernel_iters": 2, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 0.166015625, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 200, + 12, + 20 + ], + "dims": [ + 256, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 200, + 12, + 20 + ], + "dims": [ + 256, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Clip_118", + "type": "internal" + } + ], + "layer_order": 74 + }, + "Mul_121": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_elems": 240, + "num_kernel_iters": 4, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 200, + 12, + 20 + ], + "dims": [ + 256, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 200, + 12, + 20 + ], + "dims": [ + 256, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 200, + 12, + 20 + ], + "dims": [ + 256, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_113", + "type": "internal" + }, + { + "name": "Div_120", + "type": "internal" + } + ], + "layer_order": 75 + }, + "Conv_122": { + "Conv2DBf16": { + "act": 0, + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 0, + "batch_size": 1, + "bias_shift": 0, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 0, + "data_io.ifm.dimensions.batch.padding": 0, + "data_io.ifm.dimensions.batch.size": 1, + "data_io.ofm.dimensions.batch.padding": 0, + "data_io.ofm.dimensions.batch.size": 1, + "data_io.ofm.dimensions.width.padding": 0, + "func_type": 10, + "ifm_depth": 256, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 1, + "ifm_height": 12, + "ifm_sign": 0, + "ifm_sv_depth": 8, + "ifm_sv_depth.size": 8, + "ifm_sv_height": 6, + "ifm_sv_height.size": 6, + "ifm_sv_width": 26, + "ifm_sv_width.size": 26, + "ifm_width": 20, + "ifm_x_iter": 1, + "ifm_y_iter": 1, + "ifmsv_size": 1344, + "kernel_height": 3, + "kernel_width": 3, + "num_adf_cols": 4, + "num_adf_rows": 4, + "ofm_depth": 224, + "ofm_depth_iter": 7, + "ofm_height": 12, + "ofm_len": 7, + "ofm_sv_depth": 8, + "ofm_sv_height": 4, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 25, + "ofm_width": 19, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 1, + "pad_left": 1, + "pad_right": 1, + "pad_top": 1, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 1, + "stride": 1, + "stride_height": 1, + "stride_width": 1, + "zp_en": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 200, + 12, + 20 + ], + "dims": [ + 256, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 200, + 12, + 20 + ], + "dims": [ + 224, + 16, + 24 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Mul_121", + "type": "internal" + } + ], + "layer_order": 76, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 144, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 8, + "is_dwc": true, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 3, + "kernel_width": 4, + "krnl_w_div": 1, + "num_bias_sub_groups": 1, + "ofm_depth_iter": 7, + "ofm_depth_padded": 224, + "output_stride": 8, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 915840, + "wts_no_of_reads": 144 + } + }, + "Add_124": { + "AddAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 32, + "ifmsv_height": 3, + "ifmsv_size": 1920, + "ifmsv_width": 20, + "num_kernel_iters": 2, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 3, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 200, + 12, + 20 + ], + "dims": [ + 224, + 16, + 24 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 200, + 12, + 20 + ], + "dims": [ + 256, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_122", + "type": "internal" + } + ], + "layer_order": 77 + }, + "Clip_127": { + "ClipBf16": { + "aie_arch": "aie2p", + "clamp_max": 6, + "clamp_min": 0, + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 32, + "ifmsv_height": 3, + "ifmsv_size": 1920, + "ifmsv_width": 20, + "num_kernel_iters": 2, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 200, + 12, + 20 + ], + "dims": [ + 256, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 200, + 12, + 20 + ], + "dims": [ + 256, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Add_124", + "type": "internal" + } + ], + "layer_order": 78 + }, + "Div_129": { + "MulAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 32, + "ifmsv_height": 3, + "ifmsv_size": 1920, + "ifmsv_width": 20, + "num_kernel_iters": 2, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 0.166015625, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 200, + 12, + 20 + ], + "dims": [ + 256, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 200, + 12, + 20 + ], + "dims": [ + 256, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Clip_127", + "type": "internal" + } + ], + "layer_order": 79 + }, + "Mul_130": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_elems": 240, + "num_kernel_iters": 4, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 200, + 12, + 20 + ], + "dims": [ + 224, + 16, + 24 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 200, + 12, + 20 + ], + "dims": [ + 256, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 200, + 12, + 20 + ], + "dims": [ + 256, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_122", + "type": "internal" + }, + { + "name": "Div_129", + "type": "internal" + } + ], + "layer_order": 80 + }, + "Conv_131": { + "AddBf16": { + "act": 0, + "act_type": "LINEAR", + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 24, + "ifmsv_height": 1, + "ifmsv_size": 768, + "ifmsv_width": 32, + "num_elems": 240, + "num_kernel_iters": 3, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 64, + "data_io.ofm.dimensions.width": 32, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 11, + "ifm_depth": 200, + "ifm_depth_concat_extend": 56, + "ifm_depth_iter": 2, + "ifm_height": 12, + "ifm_sign": 0, + "ifm_sv_depth": 104, + "ifm_sv_height": 1, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 32, + "ifm_width": 20, + "ifm_x_iter": 1, + "ifm_y_iter": 3, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 2, + "num_ifm_height_iters": 3, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 1, + "ofm_depth": 80, + "ofm_depth_iter": 1, + "ofm_height": 12, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 24, + "ofm_sv_height": 1, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 32, + "ofm_width": 20, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 1, + "padding_sv_width": 32, + "psum_buff_offset_scaled": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 200, + 12, + 20 + ], + "dims": [ + 256, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 80, + 12, + 20 + ], + "dims": [ + 96, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 80, + 12, + 20 + ], + "dims": [ + 96, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Mul_130", + "type": "internal" + }, + { + "name": "Conv_112", + "type": "internal" + } + ], + "layer_order": 81, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 432, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 2, + "ifm_depth_padded": 208, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 3, + "ofm_depth_iter": 1, + "ofm_depth_padded": 96, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 931968, + "wts_no_of_reads": 3744 + } + }, + "Conv_133": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 12, + "data_io.ofm.dimensions.width": 8, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 80, + "ifm_depth_concat_extend": 16, + "ifm_depth_iter": 1, + "ifm_height": 12, + "ifm_sign": 0, + "ifm_sv_depth": 80, + "ifm_sv_height": 3, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 8, + "ifm_width": 20, + "ifm_x_iter": 3, + "ifm_y_iter": 1, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 1, + "num_ifm_height_iters": 1, + "num_ifm_width_iters": 3, + "num_ofm_depth_iters": 1, + "ofm_depth": 184, + "ofm_depth_iter": 1, + "ofm_height": 12, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 48, + "ofm_sv_height": 3, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 8, + "ofm_width": 20, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 3, + "padding_sv_width": 8, + "psum_buff_offset_scaled": 1, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 80, + 12, + 20 + ], + "dims": [ + 96, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 12, + 24 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_131", + "type": "internal" + } + ], + "layer_order": 82, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 864, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 80, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 6, + "ofm_depth_iter": 1, + "ofm_depth_padded": 192, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 1025280, + "wts_no_of_reads": 5760 + } + }, + "Add_135": { + "AddAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_kernel_iters": 3, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 3, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 12, + 24 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_133", + "type": "internal" + } + ], + "layer_order": 83 + }, + "Clip_138": { + "ClipBf16": { + "aie_arch": "aie2p", + "clamp_max": 6, + "clamp_min": 0, + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_kernel_iters": 3, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Add_135", + "type": "internal" + } + ], + "layer_order": 84 + }, + "Div_140": { + "MulAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_kernel_iters": 3, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 0.166015625, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Clip_138", + "type": "internal" + } + ], + "layer_order": 85 + }, + "Mul_141": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_elems": 240, + "num_kernel_iters": 3, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 12, + 24 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_133", + "type": "internal" + }, + { + "name": "Div_140", + "type": "internal" + } + ], + "layer_order": 86 + }, + "Conv_142": { + "Conv2DBf16": { + "act": 0, + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 0, + "batch_size": 1, + "bias_shift": 0, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 0, + "data_io.ifm.dimensions.batch.padding": 0, + "data_io.ifm.dimensions.batch.size": 1, + "data_io.ofm.dimensions.batch.padding": 0, + "data_io.ofm.dimensions.batch.size": 1, + "data_io.ofm.dimensions.width.padding": 0, + "func_type": 10, + "ifm_depth": 192, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 1, + "ifm_height": 12, + "ifm_sign": 0, + "ifm_sv_depth": 8, + "ifm_sv_depth.size": 8, + "ifm_sv_height": 6, + "ifm_sv_height.size": 6, + "ifm_sv_width": 26, + "ifm_sv_width.size": 26, + "ifm_width": 20, + "ifm_x_iter": 1, + "ifm_y_iter": 1, + "ifmsv_size": 1344, + "kernel_height": 3, + "kernel_width": 3, + "num_adf_cols": 4, + "num_adf_rows": 4, + "ofm_depth": 192, + "ofm_depth_iter": 6, + "ofm_height": 12, + "ofm_len": 6, + "ofm_sv_depth": 8, + "ofm_sv_height": 4, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 25, + "ofm_width": 19, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 1, + "pad_left": 1, + "pad_right": 1, + "pad_top": 1, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 1, + "stride": 1, + "stride_height": 1, + "stride_width": 1, + "zp_en": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 16, + 24 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Mul_141", + "type": "internal" + } + ], + "layer_order": 87, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 144, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 8, + "is_dwc": true, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 3, + "kernel_width": 4, + "krnl_w_div": 1, + "num_bias_sub_groups": 1, + "ofm_depth_iter": 6, + "ofm_depth_padded": 192, + "output_stride": 8, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 1097856, + "wts_no_of_reads": 144 + } + }, + "Add_144": { + "AddAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_kernel_iters": 3, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 3, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 16, + 24 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_142", + "type": "internal" + } + ], + "layer_order": 88 + }, + "Clip_147": { + "ClipBf16": { + "aie_arch": "aie2p", + "clamp_max": 6, + "clamp_min": 0, + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_kernel_iters": 3, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Add_144", + "type": "internal" + } + ], + "layer_order": 89 + }, + "Div_149": { + "MulAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_kernel_iters": 3, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 0.166015625, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Clip_147", + "type": "internal" + } + ], + "layer_order": 90 + }, + "Mul_150": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_elems": 240, + "num_kernel_iters": 3, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 16, + 24 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_142", + "type": "internal" + }, + { + "name": "Div_149", + "type": "internal" + } + ], + "layer_order": 91 + }, + "Conv_151": { + "AddBf16": { + "act": 0, + "act_type": "LINEAR", + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 24, + "ifmsv_height": 1, + "ifmsv_size": 768, + "ifmsv_width": 32, + "num_elems": 240, + "num_kernel_iters": 3, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 64, + "data_io.ofm.dimensions.width": 32, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 11, + "ifm_depth": 184, + "ifm_depth_concat_extend": 8, + "ifm_depth_iter": 2, + "ifm_height": 12, + "ifm_sign": 0, + "ifm_sv_depth": 96, + "ifm_sv_height": 1, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 32, + "ifm_width": 20, + "ifm_x_iter": 1, + "ifm_y_iter": 3, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 2, + "num_ifm_height_iters": 3, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 1, + "ofm_depth": 80, + "ofm_depth_iter": 1, + "ofm_height": 12, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 24, + "ofm_sv_height": 1, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 32, + "ofm_width": 20, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 1, + "padding_sv_width": 32, + "psum_buff_offset_scaled": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 80, + 12, + 20 + ], + "dims": [ + 96, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 80, + 12, + 20 + ], + "dims": [ + 96, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Mul_150", + "type": "internal" + }, + { + "name": "Conv_131", + "type": "internal" + } + ], + "layer_order": 92, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 432, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 2, + "ifm_depth_padded": 192, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 3, + "ofm_depth_iter": 1, + "ofm_depth_padded": 96, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 1111680, + "wts_no_of_reads": 3456 + } + }, + "Conv_153": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 12, + "data_io.ofm.dimensions.width": 8, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 80, + "ifm_depth_concat_extend": 16, + "ifm_depth_iter": 1, + "ifm_height": 12, + "ifm_sign": 0, + "ifm_sv_depth": 80, + "ifm_sv_height": 3, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 8, + "ifm_width": 20, + "ifm_x_iter": 3, + "ifm_y_iter": 1, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 1, + "num_ifm_height_iters": 1, + "num_ifm_width_iters": 3, + "num_ofm_depth_iters": 1, + "ofm_depth": 184, + "ofm_depth_iter": 1, + "ofm_height": 12, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 48, + "ofm_sv_height": 3, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 8, + "ofm_width": 20, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 3, + "padding_sv_width": 8, + "psum_buff_offset_scaled": 1, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 80, + 12, + 20 + ], + "dims": [ + 96, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 12, + 24 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_151", + "type": "internal" + } + ], + "layer_order": 93, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 864, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 80, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 6, + "ofm_depth_iter": 1, + "ofm_depth_padded": 192, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 1198080, + "wts_no_of_reads": 5760 + } + }, + "Add_155": { + "AddAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_kernel_iters": 3, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 3, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 12, + 24 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_153", + "type": "internal" + } + ], + "layer_order": 94 + }, + "Clip_158": { + "ClipBf16": { + "aie_arch": "aie2p", + "clamp_max": 6, + "clamp_min": 0, + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_kernel_iters": 3, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Add_155", + "type": "internal" + } + ], + "layer_order": 95 + }, + "Div_160": { + "MulAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_kernel_iters": 3, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 0.166015625, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Clip_158", + "type": "internal" + } + ], + "layer_order": 96 + }, + "Mul_161": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_elems": 240, + "num_kernel_iters": 3, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 12, + 24 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_153", + "type": "internal" + }, + { + "name": "Div_160", + "type": "internal" + } + ], + "layer_order": 97 + }, + "Conv_162": { + "Conv2DBf16": { + "act": 0, + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 0, + "batch_size": 1, + "bias_shift": 0, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 0, + "data_io.ifm.dimensions.batch.padding": 0, + "data_io.ifm.dimensions.batch.size": 1, + "data_io.ofm.dimensions.batch.padding": 0, + "data_io.ofm.dimensions.batch.size": 1, + "data_io.ofm.dimensions.width.padding": 0, + "func_type": 10, + "ifm_depth": 192, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 1, + "ifm_height": 12, + "ifm_sign": 0, + "ifm_sv_depth": 8, + "ifm_sv_depth.size": 8, + "ifm_sv_height": 6, + "ifm_sv_height.size": 6, + "ifm_sv_width": 26, + "ifm_sv_width.size": 26, + "ifm_width": 20, + "ifm_x_iter": 1, + "ifm_y_iter": 1, + "ifmsv_size": 1344, + "kernel_height": 3, + "kernel_width": 3, + "num_adf_cols": 4, + "num_adf_rows": 4, + "ofm_depth": 192, + "ofm_depth_iter": 6, + "ofm_height": 12, + "ofm_len": 6, + "ofm_sv_depth": 8, + "ofm_sv_height": 4, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 25, + "ofm_width": 19, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 1, + "pad_left": 1, + "pad_right": 1, + "pad_top": 1, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 1, + "stride": 1, + "stride_height": 1, + "stride_width": 1, + "zp_en": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 16, + 24 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Mul_161", + "type": "internal" + } + ], + "layer_order": 98, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 144, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 8, + "is_dwc": true, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 3, + "kernel_width": 4, + "krnl_w_div": 1, + "num_bias_sub_groups": 1, + "ofm_depth_iter": 6, + "ofm_depth_padded": 192, + "output_stride": 8, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 1270656, + "wts_no_of_reads": 144 + } + }, + "Add_164": { + "AddAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_kernel_iters": 3, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 3, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 16, + 24 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_162", + "type": "internal" + } + ], + "layer_order": 99 + }, + "Clip_167": { + "ClipBf16": { + "aie_arch": "aie2p", + "clamp_max": 6, + "clamp_min": 0, + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_kernel_iters": 3, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Add_164", + "type": "internal" + } + ], + "layer_order": 100 + }, + "Div_169": { + "MulAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_kernel_iters": 3, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 0.166015625, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Clip_167", + "type": "internal" + } + ], + "layer_order": 101 + }, + "Mul_170": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_elems": 240, + "num_kernel_iters": 3, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 16, + 24 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_162", + "type": "internal" + }, + { + "name": "Div_169", + "type": "internal" + } + ], + "layer_order": 102 + }, + "Conv_171": { + "AddBf16": { + "act": 0, + "act_type": "LINEAR", + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 24, + "ifmsv_height": 1, + "ifmsv_size": 768, + "ifmsv_width": 32, + "num_elems": 240, + "num_kernel_iters": 3, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 64, + "data_io.ofm.dimensions.width": 32, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 11, + "ifm_depth": 184, + "ifm_depth_concat_extend": 8, + "ifm_depth_iter": 2, + "ifm_height": 12, + "ifm_sign": 0, + "ifm_sv_depth": 96, + "ifm_sv_height": 1, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 32, + "ifm_width": 20, + "ifm_x_iter": 1, + "ifm_y_iter": 3, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 2, + "num_ifm_height_iters": 3, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 1, + "ofm_depth": 80, + "ofm_depth_iter": 1, + "ofm_height": 12, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 24, + "ofm_sv_height": 1, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 32, + "ofm_width": 20, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 1, + "padding_sv_width": 32, + "psum_buff_offset_scaled": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 184, + 12, + 20 + ], + "dims": [ + 192, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 80, + 12, + 20 + ], + "dims": [ + 96, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 80, + 12, + 20 + ], + "dims": [ + 96, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Mul_170", + "type": "internal" + }, + { + "name": "Conv_151", + "type": "internal" + } + ], + "layer_order": 103, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 432, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 2, + "ifm_depth_padded": 192, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 3, + "ofm_depth_iter": 1, + "ofm_depth_padded": 96, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 1284480, + "wts_no_of_reads": 3456 + } + }, + "Conv_173": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 64, + "data_io.ofm.dimensions.width": 32, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 80, + "ifm_depth_concat_extend": 16, + "ifm_depth_iter": 1, + "ifm_height": 12, + "ifm_sign": 0, + "ifm_sv_depth": 80, + "ifm_sv_height": 1, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 32, + "ifm_width": 20, + "ifm_x_iter": 1, + "ifm_y_iter": 3, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 1, + "num_ifm_height_iters": 3, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 3, + "ofm_depth": 480, + "ofm_depth_iter": 3, + "ofm_height": 12, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 40, + "ofm_sv_height": 1, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 32, + "ofm_width": 20, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 1, + "padding_sv_width": 32, + "psum_buff_offset_scaled": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 80, + 12, + 20 + ], + "dims": [ + 96, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 480, + 12, + 20 + ], + "dims": [ + 480, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_171", + "type": "internal" + } + ], + "layer_order": 104, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 720, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 80, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 5, + "ofm_depth_iter": 3, + "ofm_depth_padded": 480, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 1370880, + "wts_no_of_reads": 4800 + } + }, + "Add_175": { + "AddAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 32, + "ifmsv_height": 3, + "ifmsv_size": 1920, + "ifmsv_width": 20, + "num_kernel_iters": 4, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 3, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 480, + 12, + 20 + ], + "dims": [ + 480, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 480, + 12, + 20 + ], + "dims": [ + 512, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_173", + "type": "internal" + } + ], + "layer_order": 105 + }, + "Clip_178": { + "ClipBf16": { + "aie_arch": "aie2p", + "clamp_max": 6, + "clamp_min": 0, + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 32, + "ifmsv_height": 3, + "ifmsv_size": 1920, + "ifmsv_width": 20, + "num_kernel_iters": 4, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 480, + 12, + 20 + ], + "dims": [ + 512, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 480, + 12, + 20 + ], + "dims": [ + 512, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Add_175", + "type": "internal" + } + ], + "layer_order": 106 + }, + "Div_180": { + "MulAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 32, + "ifmsv_height": 3, + "ifmsv_size": 1920, + "ifmsv_width": 20, + "num_kernel_iters": 4, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 0.166015625, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 480, + 12, + 20 + ], + "dims": [ + 512, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 480, + 12, + 20 + ], + "dims": [ + 512, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Clip_178", + "type": "internal" + } + ], + "layer_order": 107 + }, + "Mul_181": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_elems": 240, + "num_kernel_iters": 8, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 480, + 12, + 20 + ], + "dims": [ + 480, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 480, + 12, + 20 + ], + "dims": [ + 512, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 480, + 12, + 20 + ], + "dims": [ + 512, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_173", + "type": "internal" + }, + { + "name": "Div_180", + "type": "internal" + } + ], + "layer_order": 108 + }, + "Conv_182": { + "Conv2DBf16": { + "act": 0, + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 0, + "batch_size": 1, + "bias_shift": 0, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 0, + "data_io.ifm.dimensions.batch.padding": 0, + "data_io.ifm.dimensions.batch.size": 1, + "data_io.ofm.dimensions.batch.padding": 0, + "data_io.ofm.dimensions.batch.size": 1, + "data_io.ofm.dimensions.width.padding": 0, + "func_type": 10, + "ifm_depth": 512, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 1, + "ifm_height": 12, + "ifm_sign": 0, + "ifm_sv_depth": 8, + "ifm_sv_depth.size": 8, + "ifm_sv_height": 6, + "ifm_sv_height.size": 6, + "ifm_sv_width": 26, + "ifm_sv_width.size": 26, + "ifm_width": 20, + "ifm_x_iter": 1, + "ifm_y_iter": 1, + "ifmsv_size": 1344, + "kernel_height": 3, + "kernel_width": 3, + "num_adf_cols": 4, + "num_adf_rows": 4, + "ofm_depth": 480, + "ofm_depth_iter": 15, + "ofm_height": 12, + "ofm_len": 15, + "ofm_sv_depth": 8, + "ofm_sv_height": 4, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 25, + "ofm_width": 19, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 1, + "pad_left": 1, + "pad_right": 1, + "pad_top": 1, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 1, + "stride": 1, + "stride_height": 1, + "stride_width": 1, + "zp_en": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 480, + 12, + 20 + ], + "dims": [ + 512, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 480, + 12, + 20 + ], + "dims": [ + 480, + 16, + 24 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Mul_181", + "type": "internal" + } + ], + "layer_order": 109, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 144, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 8, + "is_dwc": true, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 3, + "kernel_width": 4, + "krnl_w_div": 1, + "num_bias_sub_groups": 1, + "ofm_depth_iter": 15, + "ofm_depth_padded": 480, + "output_stride": 8, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 1552320, + "wts_no_of_reads": 144 + } + }, + "Add_184": { + "AddAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 32, + "ifmsv_height": 3, + "ifmsv_size": 1920, + "ifmsv_width": 20, + "num_kernel_iters": 4, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 3, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 480, + 12, + 20 + ], + "dims": [ + 480, + 16, + 24 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 480, + 12, + 20 + ], + "dims": [ + 512, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_182", + "type": "internal" + } + ], + "layer_order": 110 + }, + "Clip_187": { + "ClipBf16": { + "aie_arch": "aie2p", + "clamp_max": 6, + "clamp_min": 0, + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 32, + "ifmsv_height": 3, + "ifmsv_size": 1920, + "ifmsv_width": 20, + "num_kernel_iters": 4, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 480, + 12, + 20 + ], + "dims": [ + 512, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 480, + 12, + 20 + ], + "dims": [ + 512, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Add_184", + "type": "internal" + } + ], + "layer_order": 111 + }, + "Div_189": { + "MulAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 32, + "ifmsv_height": 3, + "ifmsv_size": 1920, + "ifmsv_width": 20, + "num_kernel_iters": 4, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 0.166015625, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 480, + 12, + 20 + ], + "dims": [ + 512, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 480, + 12, + 20 + ], + "dims": [ + 512, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Clip_187", + "type": "internal" + } + ], + "layer_order": 112 + }, + "Mul_190": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_elems": 240, + "num_kernel_iters": 8, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 480, + 12, + 20 + ], + "dims": [ + 480, + 16, + 24 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 480, + 12, + 20 + ], + "dims": [ + 512, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 480, + 12, + 20 + ], + "dims": [ + 512, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_182", + "type": "internal" + }, + { + "name": "Div_189", + "type": "internal" + } + ], + "layer_order": 113 + }, + "Generated-#24": { + "Transpose4dAdf": { + "aie_arch": "aie2p", + "dim_0": 12, + "dim_1": 60, + "dim_2": 20, + "dim_3": 8, + "dtype": "bfloat16", + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "output_l3_format": "HCWN_C8", + "perm": 6 + }, + "dims_and_bounds": [ + { + "bounds": [ + 480, + 12, + 20 + ], + "dims": [ + 480, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 480, + 1, + 240 + ], + "dims": [ + 480, + 1, + 240 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Mul_190", + "type": "internal" + } + ], + "layer_order": 114 + }, + "Generated-#26": { + "ReduceMeanC8Bf16": { + "L3_depth": 480, + "L3_height": 1, + "L3_width": 240, + "aie_arch": "aie2p", + "compiler": "chess", + "depth_iter": 3, + "full_channel": 480, + "full_height": 1, + "full_width": 240, + "height_iter": 1, + "loop_num_channel": 12, + "loop_num_height": 1, + "loop_num_width": 5, + "num_iters": 15, + "pad_value": 0, + "reduce_axis": 2, + "reduce_dim": "W", + "scale": 0.004180908203125, + "scale_shift": 0, + "sv_channel": 40, + "sv_height": 1, + "sv_width": 48, + "width_iter": 5 + }, + "dims_and_bounds": [ + { + "bounds": [ + 480, + 1, + 240 + ], + "dims": [ + 480, + 1, + 240 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 480, + 1, + 1 + ], + "dims": [ + 480, + 4, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Generated-#24", + "type": "internal" + } + ], + "layer_order": 115 + }, + "Conv_192": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 1, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 12, + "data_io.ofm.dimensions.width": 8, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 480, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 4, + "ifm_height": 1, + "ifm_sign": 0, + "ifm_sv_depth": 120, + "ifm_sv_height": 1, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 8, + "ifm_width": 1, + "ifm_x_iter": 1, + "ifm_y_iter": 1, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 0, + "lrelu_alpha_kernel": 0, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 4, + "num_ifm_height_iters": 1, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 1, + "ofm_depth": 120, + "ofm_depth_iter": 1, + "ofm_height": 1, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 32, + "ofm_sv_height": 1, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 8, + "ofm_width": 1, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 1, + "padding_sv_width": 8, + "psum_buff_offset_scaled": 1, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 480, + 1, + 1 + ], + "dims": [ + 480, + 4, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 120, + 1, + 1 + ], + "dims": [ + 128, + 1, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Generated-#26", + "type": "internal" + } + ], + "layer_order": 116, + "split_width": true, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 576, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 4, + "ifm_depth_padded": 480, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 4, + "ofm_depth_iter": 1, + "ofm_depth_padded": 128, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 1586880, + "wts_no_of_reads": 5760 + } + }, + "Conv_194": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 12, + "data_io.ofm.dimensions.width": 8, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 120, + "ifm_depth_concat_extend": 8, + "ifm_depth_iter": 1, + "ifm_height": 1, + "ifm_sign": 0, + "ifm_sv_depth": 120, + "ifm_sv_height": 1, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 8, + "ifm_width": 1, + "ifm_x_iter": 1, + "ifm_y_iter": 1, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 1, + "num_ifm_height_iters": 1, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 4, + "ofm_depth": 480, + "ofm_depth_iter": 4, + "ofm_height": 1, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 32, + "ofm_sv_height": 1, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 8, + "ofm_width": 1, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 1, + "padding_sv_width": 8, + "psum_buff_offset_scaled": 1, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 120, + 1, + 1 + ], + "dims": [ + 128, + 1, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 480, + 1, + 1 + ], + "dims": [ + 512, + 1, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_192", + "type": "internal" + } + ], + "layer_order": 117, + "split_width": true, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 576, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 120, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 4, + "ofm_depth_iter": 4, + "ofm_depth_padded": 512, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 1872576, + "wts_no_of_reads": 5760 + } + }, + "Add_196": { + "AddAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 128, + "ifmsv_height": 3, + "ifmsv_size": 384, + "ifmsv_width": 1, + "num_kernel_iters": 1, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 3, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 480, + 1, + 1 + ], + "dims": [ + 512, + 1, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 480, + 1, + 1 + ], + "dims": [ + 512, + 12, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_194", + "type": "internal" + } + ], + "layer_order": 118 + }, + "Clip_199": { + "ClipBf16": { + "aie_arch": "aie2p", + "clamp_max": 6, + "clamp_min": 0, + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 128, + "ifmsv_height": 2, + "ifmsv_size": 256, + "ifmsv_width": 1, + "num_kernel_iters": 1, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 480, + 1, + 1 + ], + "dims": [ + 512, + 12, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 480, + 1, + 1 + ], + "dims": [ + 512, + 8, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Add_196", + "type": "internal" + } + ], + "layer_order": 119 + }, + "Div_201": { + "MulAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 128, + "ifmsv_height": 4, + "ifmsv_size": 512, + "ifmsv_width": 1, + "num_kernel_iters": 1, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 0.166015625, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 480, + 1, + 1 + ], + "dims": [ + 512, + 8, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 480, + 1, + 1 + ], + "dims": [ + 512, + 16, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Clip_199", + "type": "internal" + } + ], + "layer_order": 120 + }, + "Generated-#28": { + "TileAdf": { + "aie_arch": "aie2p", + "dtype": "bfloat16", + "i_dim_c": 480, + "i_dim_h": 1, + "i_dim_n": 1, + "i_dim_w": 1, + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "output_l3_format": "HCWN_C8", + "rep_dim_c": 1, + "rep_dim_h": 12, + "rep_dim_w": 20 + }, + "dims_and_bounds": [ + { + "bounds": [ + 480, + 1, + 1 + ], + "dims": [ + 480, + 1, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 480, + 12, + 20 + ], + "dims": [ + 480, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Div_201", + "type": "internal" + } + ], + "layer_order": 121 + }, + "Mul_202": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_elems": 240, + "num_kernel_iters": 8, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 480, + 12, + 20 + ], + "dims": [ + 480, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 480, + 12, + 20 + ], + "dims": [ + 512, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 480, + 12, + 20 + ], + "dims": [ + 512, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Generated-#28", + "type": "internal" + }, + { + "name": "Mul_190", + "type": "internal" + } + ], + "layer_order": 122 + }, + "Conv_203": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 0, + "data_io.ofm.dimensions.width": 32, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 480, + "ifm_depth_concat_extend": 32, + "ifm_depth_iter": 5, + "ifm_height": 12, + "ifm_sign": 0, + "ifm_sv_depth": 96, + "ifm_sv_height": 1, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 32, + "ifm_width": 20, + "ifm_x_iter": 1, + "ifm_y_iter": 3, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 5, + "num_ifm_height_iters": 3, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 1, + "ofm_depth": 112, + "ofm_depth_iter": 1, + "ofm_height": 12, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 32, + "ofm_sv_height": 1, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 32, + "ofm_width": 20, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 1, + "padding_sv_width": 32, + "psum_buff_offset_scaled": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 480, + 12, + 20 + ], + "dims": [ + 512, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 112, + 12, + 20 + ], + "dims": [ + 128, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Mul_202", + "type": "internal" + } + ], + "layer_order": 123, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 576, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 5, + "ifm_depth_padded": 480, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 4, + "ofm_depth_iter": 1, + "ofm_depth_padded": 128, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 2158272, + "wts_no_of_reads": 4608 + } + }, + "Conv_204": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 64, + "data_io.ofm.dimensions.width": 32, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 112, + "ifm_depth_concat_extend": 16, + "ifm_depth_iter": 2, + "ifm_height": 12, + "ifm_sign": 0, + "ifm_sv_depth": 64, + "ifm_sv_height": 1, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 32, + "ifm_width": 20, + "ifm_x_iter": 1, + "ifm_y_iter": 3, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 2, + "num_ifm_height_iters": 3, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 3, + "ofm_depth": 672, + "ofm_depth_iter": 3, + "ofm_height": 12, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 56, + "ofm_sv_height": 1, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 32, + "ofm_width": 20, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 1, + "padding_sv_width": 32, + "psum_buff_offset_scaled": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 112, + 12, + 20 + ], + "dims": [ + 128, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 672, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_203", + "type": "internal" + } + ], + "layer_order": 124, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 1008, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 2, + "ifm_depth_padded": 128, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 7, + "ofm_depth_iter": 3, + "ofm_depth_padded": 672, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 2446272, + "wts_no_of_reads": 5376 + } + }, + "Add_206": { + "AddAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_kernel_iters": 11, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 3, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 672, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 704, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_204", + "type": "internal" + } + ], + "layer_order": 125 + }, + "Clip_209": { + "ClipBf16": { + "aie_arch": "aie2p", + "clamp_max": 6, + "clamp_min": 0, + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_kernel_iters": 11, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 704, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 704, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Add_206", + "type": "internal" + } + ], + "layer_order": 126 + }, + "Div_211": { + "MulAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_kernel_iters": 11, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 0.166015625, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 704, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 704, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Clip_209", + "type": "internal" + } + ], + "layer_order": 127 + }, + "Mul_212": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_elems": 240, + "num_kernel_iters": 11, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 672, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 704, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 704, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_204", + "type": "internal" + }, + { + "name": "Div_211", + "type": "internal" + } + ], + "layer_order": 128 + }, + "Conv_213": { + "Conv2DBf16": { + "act": 0, + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 0, + "batch_size": 1, + "bias_shift": 0, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 0, + "data_io.ifm.dimensions.batch.padding": 0, + "data_io.ifm.dimensions.batch.size": 1, + "data_io.ofm.dimensions.batch.padding": 0, + "data_io.ofm.dimensions.batch.size": 1, + "data_io.ofm.dimensions.width.padding": 0, + "func_type": 10, + "ifm_depth": 704, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 1, + "ifm_height": 12, + "ifm_sign": 0, + "ifm_sv_depth": 8, + "ifm_sv_depth.size": 8, + "ifm_sv_height": 6, + "ifm_sv_height.size": 6, + "ifm_sv_width": 26, + "ifm_sv_width.size": 26, + "ifm_width": 20, + "ifm_x_iter": 1, + "ifm_y_iter": 1, + "ifmsv_size": 1344, + "kernel_height": 3, + "kernel_width": 3, + "num_adf_cols": 4, + "num_adf_rows": 4, + "ofm_depth": 672, + "ofm_depth_iter": 21, + "ofm_height": 12, + "ofm_len": 21, + "ofm_sv_depth": 8, + "ofm_sv_height": 4, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 25, + "ofm_width": 19, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 1, + "pad_left": 1, + "pad_right": 1, + "pad_top": 1, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 1, + "stride": 1, + "stride_height": 1, + "stride_width": 1, + "zp_en": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 704, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 672, + 16, + 24 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Mul_212", + "type": "internal" + } + ], + "layer_order": 129, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 144, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 8, + "is_dwc": true, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 3, + "kernel_width": 4, + "krnl_w_div": 1, + "num_bias_sub_groups": 1, + "ofm_depth_iter": 21, + "ofm_depth_padded": 672, + "output_stride": 8, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 2857536, + "wts_no_of_reads": 144 + } + }, + "Add_215": { + "AddAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_kernel_iters": 11, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 3, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 672, + 16, + 24 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 704, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_213", + "type": "internal" + } + ], + "layer_order": 130 + }, + "Clip_218": { + "ClipBf16": { + "aie_arch": "aie2p", + "clamp_max": 6, + "clamp_min": 0, + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_kernel_iters": 11, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 704, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 704, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Add_215", + "type": "internal" + } + ], + "layer_order": 131 + }, + "Div_220": { + "MulAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_kernel_iters": 11, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 0.166015625, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 704, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 704, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Clip_218", + "type": "internal" + } + ], + "layer_order": 132 + }, + "Mul_221": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_elems": 240, + "num_kernel_iters": 11, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 672, + 16, + 24 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 704, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 704, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_213", + "type": "internal" + }, + { + "name": "Div_220", + "type": "internal" + } + ], + "layer_order": 133 + }, + "Generated-#30": { + "Transpose4dAdf": { + "aie_arch": "aie2p", + "dim_0": 12, + "dim_1": 84, + "dim_2": 20, + "dim_3": 8, + "dtype": "bfloat16", + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "output_l3_format": "HCWN_C8", + "perm": 6 + }, + "dims_and_bounds": [ + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 672, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 1, + 240 + ], + "dims": [ + 672, + 1, + 240 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Mul_221", + "type": "internal" + } + ], + "layer_order": 134 + }, + "Generated-#32": { + "ReduceMeanC8Bf16": { + "L3_depth": 672, + "L3_height": 1, + "L3_width": 240, + "aie_arch": "aie2p", + "compiler": "chess", + "depth_iter": 3, + "full_channel": 672, + "full_height": 1, + "full_width": 240, + "height_iter": 1, + "loop_num_channel": 12, + "loop_num_height": 1, + "loop_num_width": 8, + "num_iters": 24, + "pad_value": 0, + "reduce_axis": 2, + "reduce_dim": "W", + "scale": 0.004180908203125, + "scale_shift": 0, + "sv_channel": 56, + "sv_height": 1, + "sv_width": 32, + "width_iter": 8 + }, + "dims_and_bounds": [ + { + "bounds": [ + 672, + 1, + 240 + ], + "dims": [ + 672, + 1, + 240 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 1, + 1 + ], + "dims": [ + 672, + 4, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Generated-#30", + "type": "internal" + } + ], + "layer_order": 135 + }, + "Conv_223": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 1, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 12, + "data_io.ofm.dimensions.width": 8, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 672, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 7, + "ifm_height": 1, + "ifm_sign": 0, + "ifm_sv_depth": 96, + "ifm_sv_height": 1, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 8, + "ifm_width": 1, + "ifm_x_iter": 1, + "ifm_y_iter": 1, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 0, + "lrelu_alpha_kernel": 0, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 7, + "num_ifm_height_iters": 1, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 1, + "ofm_depth": 168, + "ofm_depth_iter": 1, + "ofm_height": 1, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 48, + "ofm_sv_height": 1, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 8, + "ofm_width": 1, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 1, + "padding_sv_width": 8, + "psum_buff_offset_scaled": 1, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 672, + 1, + 1 + ], + "dims": [ + 672, + 4, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 168, + 1, + 1 + ], + "dims": [ + 192, + 1, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Generated-#32", + "type": "internal" + } + ], + "layer_order": 136, + "split_width": true, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 864, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 7, + "ifm_depth_padded": 672, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 6, + "ofm_depth_iter": 1, + "ofm_depth_padded": 192, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 2905920, + "wts_no_of_reads": 6912 + } + }, + "Conv_225": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 12, + "data_io.ofm.dimensions.width": 16, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 168, + "ifm_depth_concat_extend": 24, + "ifm_depth_iter": 1, + "ifm_height": 1, + "ifm_sign": 0, + "ifm_sv_depth": 168, + "ifm_sv_height": 1, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 16, + "ifm_width": 1, + "ifm_x_iter": 1, + "ifm_y_iter": 1, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 1, + "num_ifm_height_iters": 1, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 11, + "ofm_depth": 672, + "ofm_depth_iter": 11, + "ofm_height": 1, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 16, + "ofm_sv_height": 1, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 16, + "ofm_width": 1, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 1, + "padding_sv_width": 16, + "psum_buff_offset_scaled": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 168, + 1, + 1 + ], + "dims": [ + 192, + 1, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 1, + 1 + ], + "dims": [ + 704, + 1, + 64 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_223", + "type": "internal" + } + ], + "layer_order": 137, + "split_width": true, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 288, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 168, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 2, + "ofm_depth_iter": 11, + "ofm_depth_padded": 704, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 3510720, + "wts_no_of_reads": 4032 + } + }, + "Add_227": { + "AddAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 192, + "ifmsv_height": 2, + "ifmsv_size": 384, + "ifmsv_width": 1, + "num_kernel_iters": 1, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 3, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 672, + 1, + 1 + ], + "dims": [ + 704, + 1, + 64 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 1, + 1 + ], + "dims": [ + 768, + 8, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_225", + "type": "internal" + } + ], + "layer_order": 138 + }, + "Clip_230": { + "ClipBf16": { + "aie_arch": "aie2p", + "clamp_max": 6, + "clamp_min": 0, + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 256, + "ifmsv_height": 1, + "ifmsv_size": 256, + "ifmsv_width": 1, + "num_kernel_iters": 1, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 672, + 1, + 1 + ], + "dims": [ + 768, + 8, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 1, + 1 + ], + "dims": [ + 1024, + 4, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Add_227", + "type": "internal" + } + ], + "layer_order": 139 + }, + "Div_232": { + "MulAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 256, + "ifmsv_height": 2, + "ifmsv_size": 512, + "ifmsv_width": 1, + "num_kernel_iters": 1, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 0.166015625, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 672, + 1, + 1 + ], + "dims": [ + 1024, + 4, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 1, + 1 + ], + "dims": [ + 1024, + 8, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Clip_230", + "type": "internal" + } + ], + "layer_order": 140 + }, + "Generated-#34": { + "TileAdf": { + "aie_arch": "aie2p", + "dtype": "bfloat16", + "i_dim_c": 672, + "i_dim_h": 1, + "i_dim_n": 1, + "i_dim_w": 1, + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "output_l3_format": "HCWN_C8", + "rep_dim_c": 1, + "rep_dim_h": 12, + "rep_dim_w": 20 + }, + "dims_and_bounds": [ + { + "bounds": [ + 672, + 1, + 1 + ], + "dims": [ + 672, + 1, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 672, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Div_232", + "type": "internal" + } + ], + "layer_order": 141 + }, + "Mul_233": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_elems": 240, + "num_kernel_iters": 11, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 672, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 704, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 704, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Generated-#34", + "type": "internal" + }, + { + "name": "Mul_221", + "type": "internal" + } + ], + "layer_order": 142 + }, + "Conv_234": { + "AddBf16": { + "act": 0, + "act_type": "LINEAR", + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 32, + "ifmsv_height": 1, + "ifmsv_size": 1024, + "ifmsv_width": 32, + "num_elems": 240, + "num_kernel_iters": 3, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 0, + "data_io.ofm.dimensions.width": 32, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 11, + "ifm_depth": 672, + "ifm_depth_concat_extend": 32, + "ifm_depth_iter": 7, + "ifm_height": 12, + "ifm_sign": 0, + "ifm_sv_depth": 96, + "ifm_sv_height": 1, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 32, + "ifm_width": 20, + "ifm_x_iter": 1, + "ifm_y_iter": 3, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 7, + "num_ifm_height_iters": 3, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 1, + "ofm_depth": 112, + "ofm_depth_iter": 1, + "ofm_height": 12, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 32, + "ofm_sv_height": 1, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 32, + "ofm_width": 20, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 1, + "padding_sv_width": 32, + "psum_buff_offset_scaled": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 704, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 112, + 12, + 20 + ], + "dims": [ + 128, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 112, + 12, + 20 + ], + "dims": [ + 128, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Mul_233", + "type": "internal" + }, + { + "name": "Conv_203", + "type": "internal" + } + ], + "layer_order": 143, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 576, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 7, + "ifm_depth_padded": 672, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 4, + "ofm_depth_iter": 1, + "ofm_depth_padded": 128, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 4055616, + "wts_no_of_reads": 4608 + } + }, + "Conv_236": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 64, + "data_io.ofm.dimensions.width": 32, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 112, + "ifm_depth_concat_extend": 16, + "ifm_depth_iter": 2, + "ifm_height": 12, + "ifm_sign": 0, + "ifm_sv_depth": 64, + "ifm_sv_height": 1, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 32, + "ifm_width": 20, + "ifm_x_iter": 1, + "ifm_y_iter": 3, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 2, + "num_ifm_height_iters": 3, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 3, + "ofm_depth": 672, + "ofm_depth_iter": 3, + "ofm_height": 12, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 56, + "ofm_sv_height": 1, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 32, + "ofm_width": 20, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 1, + "padding_sv_width": 32, + "psum_buff_offset_scaled": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 112, + 12, + 20 + ], + "dims": [ + 128, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 672, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_234", + "type": "internal" + } + ], + "layer_order": 144, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 1008, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 2, + "ifm_depth_padded": 128, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 7, + "ofm_depth_iter": 3, + "ofm_depth_padded": 672, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 4458816, + "wts_no_of_reads": 5376 + } + }, + "Add_238": { + "AddAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_kernel_iters": 11, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 3, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 672, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 704, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_236", + "type": "internal" + } + ], + "layer_order": 145 + }, + "Clip_241": { + "ClipBf16": { + "aie_arch": "aie2p", + "clamp_max": 6, + "clamp_min": 0, + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_kernel_iters": 11, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 704, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 704, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Add_238", + "type": "internal" + } + ], + "layer_order": 146 + }, + "Div_243": { + "MulAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_kernel_iters": 11, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 0.166015625, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 704, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 704, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Clip_241", + "type": "internal" + } + ], + "layer_order": 147 + }, + "Mul_244": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_elems": 240, + "num_kernel_iters": 11, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 672, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 704, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 704, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_236", + "type": "internal" + }, + { + "name": "Div_243", + "type": "internal" + } + ], + "layer_order": 148 + }, + "Conv_245": { + "Conv2DBf16": { + "act": 0, + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 0, + "batch_size": 1, + "bias_shift": 0, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 0, + "data_io.ifm.dimensions.batch.padding": 0, + "data_io.ifm.dimensions.batch.size": 1, + "data_io.ofm.dimensions.batch.padding": 0, + "data_io.ofm.dimensions.batch.size": 1, + "data_io.ofm.dimensions.width.padding": 0, + "func_type": 10, + "ifm_depth": 704, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 1, + "ifm_height": 12, + "ifm_sign": 0, + "ifm_sv_depth": 8, + "ifm_sv_depth.size": 8, + "ifm_sv_height": 10, + "ifm_sv_height.size": 10, + "ifm_sv_width": 16, + "ifm_sv_width.size": 16, + "ifm_width": 20, + "ifm_x_iter": 3, + "ifm_y_iter": 2, + "ifmsv_size": 1280, + "kernel_height": 9, + "kernel_width": 9, + "num_adf_cols": 4, + "num_adf_rows": 4, + "ofm_depth": 672, + "ofm_depth_iter": 21, + "ofm_height": 12, + "ofm_len": 126, + "ofm_sv_depth": 8, + "ofm_sv_height": 2, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 5, + "ofm_width": 17, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 4, + "pad_left": 4, + "pad_right": 4, + "pad_top": 4, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 1, + "stride": 1, + "stride_height": 1, + "stride_width": 1, + "zp_en": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 704, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 672, + 16, + 24 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Mul_244", + "type": "internal" + } + ], + "layer_order": 149, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 144, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 8, + "is_dwc": true, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 9, + "kernel_width": 12, + "krnl_w_div": 1, + "num_bias_sub_groups": 1, + "ofm_depth_iter": 21, + "ofm_depth_padded": 672, + "output_stride": 8, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 4870080, + "wts_no_of_reads": 1296 + } + }, + "Add_247": { + "AddAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_kernel_iters": 11, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 3, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 672, + 16, + 24 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 704, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_245", + "type": "internal" + } + ], + "layer_order": 150 + }, + "Clip_250": { + "ClipBf16": { + "aie_arch": "aie2p", + "clamp_max": 6, + "clamp_min": 0, + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_kernel_iters": 11, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 704, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 704, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Add_247", + "type": "internal" + } + ], + "layer_order": 151 + }, + "Div_252": { + "MulAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_kernel_iters": 11, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 0.166015625, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 704, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 704, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Clip_250", + "type": "internal" + } + ], + "layer_order": 152 + }, + "Mul_253": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_elems": 240, + "num_kernel_iters": 11, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 672, + 16, + 24 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 704, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 704, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_245", + "type": "internal" + }, + { + "name": "Div_252", + "type": "internal" + } + ], + "layer_order": 153 + }, + "Generated-#36": { + "Transpose4dAdf": { + "aie_arch": "aie2p", + "dim_0": 12, + "dim_1": 84, + "dim_2": 20, + "dim_3": 8, + "dtype": "bfloat16", + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "output_l3_format": "HCWN_C8", + "perm": 6 + }, + "dims_and_bounds": [ + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 672, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 1, + 240 + ], + "dims": [ + 672, + 1, + 240 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Mul_253", + "type": "internal" + } + ], + "layer_order": 154 + }, + "Generated-#38": { + "ReduceMeanC8Bf16": { + "L3_depth": 672, + "L3_height": 1, + "L3_width": 240, + "aie_arch": "aie2p", + "compiler": "chess", + "depth_iter": 3, + "full_channel": 672, + "full_height": 1, + "full_width": 240, + "height_iter": 1, + "loop_num_channel": 12, + "loop_num_height": 1, + "loop_num_width": 8, + "num_iters": 24, + "pad_value": 0, + "reduce_axis": 2, + "reduce_dim": "W", + "scale": 0.004180908203125, + "scale_shift": 0, + "sv_channel": 56, + "sv_height": 1, + "sv_width": 32, + "width_iter": 8 + }, + "dims_and_bounds": [ + { + "bounds": [ + 672, + 1, + 240 + ], + "dims": [ + 672, + 1, + 240 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 1, + 1 + ], + "dims": [ + 672, + 4, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Generated-#36", + "type": "internal" + } + ], + "layer_order": 155 + }, + "Conv_255": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 1, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 12, + "data_io.ofm.dimensions.width": 8, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 672, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 7, + "ifm_height": 1, + "ifm_sign": 0, + "ifm_sv_depth": 96, + "ifm_sv_height": 1, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 8, + "ifm_width": 1, + "ifm_x_iter": 1, + "ifm_y_iter": 1, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 0, + "lrelu_alpha_kernel": 0, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 7, + "num_ifm_height_iters": 1, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 1, + "ofm_depth": 168, + "ofm_depth_iter": 1, + "ofm_height": 1, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 48, + "ofm_sv_height": 1, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 8, + "ofm_width": 1, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 1, + "padding_sv_width": 8, + "psum_buff_offset_scaled": 1, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 672, + 1, + 1 + ], + "dims": [ + 672, + 4, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 168, + 1, + 1 + ], + "dims": [ + 192, + 1, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Generated-#38", + "type": "internal" + } + ], + "layer_order": 156, + "split_width": true, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 864, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 7, + "ifm_depth_padded": 672, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 6, + "ofm_depth_iter": 1, + "ofm_depth_padded": 192, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 5208768, + "wts_no_of_reads": 6912 + } + }, + "Conv_257": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 12, + "data_io.ofm.dimensions.width": 16, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 168, + "ifm_depth_concat_extend": 24, + "ifm_depth_iter": 1, + "ifm_height": 1, + "ifm_sign": 0, + "ifm_sv_depth": 168, + "ifm_sv_height": 1, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 16, + "ifm_width": 1, + "ifm_x_iter": 1, + "ifm_y_iter": 1, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 1, + "num_ifm_height_iters": 1, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 11, + "ofm_depth": 672, + "ofm_depth_iter": 11, + "ofm_height": 1, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 16, + "ofm_sv_height": 1, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 16, + "ofm_width": 1, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 1, + "padding_sv_width": 16, + "psum_buff_offset_scaled": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 168, + 1, + 1 + ], + "dims": [ + 192, + 1, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 1, + 1 + ], + "dims": [ + 704, + 1, + 64 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_255", + "type": "internal" + } + ], + "layer_order": 157, + "split_width": true, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 288, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 168, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 2, + "ofm_depth_iter": 11, + "ofm_depth_padded": 704, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 5813568, + "wts_no_of_reads": 4032 + } + }, + "Add_259": { + "AddAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 192, + "ifmsv_height": 2, + "ifmsv_size": 384, + "ifmsv_width": 1, + "num_kernel_iters": 1, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 3, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 672, + 1, + 1 + ], + "dims": [ + 704, + 1, + 64 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 1, + 1 + ], + "dims": [ + 768, + 8, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_257", + "type": "internal" + } + ], + "layer_order": 158 + }, + "Clip_262": { + "ClipBf16": { + "aie_arch": "aie2p", + "clamp_max": 6, + "clamp_min": 0, + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 256, + "ifmsv_height": 1, + "ifmsv_size": 256, + "ifmsv_width": 1, + "num_kernel_iters": 1, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 672, + 1, + 1 + ], + "dims": [ + 768, + 8, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 1, + 1 + ], + "dims": [ + 1024, + 4, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Add_259", + "type": "internal" + } + ], + "layer_order": 159 + }, + "Div_264": { + "MulAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 256, + "ifmsv_height": 2, + "ifmsv_size": 512, + "ifmsv_width": 1, + "num_kernel_iters": 1, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 0.166015625, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 672, + 1, + 1 + ], + "dims": [ + 1024, + 4, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 1, + 1 + ], + "dims": [ + 1024, + 8, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Clip_262", + "type": "internal" + } + ], + "layer_order": 160 + }, + "Generated-#40": { + "TileAdf": { + "aie_arch": "aie2p", + "dtype": "bfloat16", + "i_dim_c": 672, + "i_dim_h": 1, + "i_dim_n": 1, + "i_dim_w": 1, + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "output_l3_format": "HCWN_C8", + "rep_dim_c": 1, + "rep_dim_h": 12, + "rep_dim_w": 20 + }, + "dims_and_bounds": [ + { + "bounds": [ + 672, + 1, + 1 + ], + "dims": [ + 672, + 1, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 672, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Div_264", + "type": "internal" + } + ], + "layer_order": 161 + }, + "Mul_265": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_elems": 240, + "num_kernel_iters": 11, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 672, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 704, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 704, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Generated-#40", + "type": "internal" + }, + { + "name": "Mul_253", + "type": "internal" + } + ], + "layer_order": 162 + }, + "Conv_266": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 64, + "data_io.ofm.dimensions.width": 32, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 672, + "ifm_depth_concat_extend": 32, + "ifm_depth_iter": 9, + "ifm_height": 12, + "ifm_sign": 0, + "ifm_sv_depth": 80, + "ifm_sv_height": 1, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 32, + "ifm_width": 20, + "ifm_x_iter": 1, + "ifm_y_iter": 3, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 9, + "num_ifm_height_iters": 3, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 1, + "ofm_depth": 160, + "ofm_depth_iter": 1, + "ofm_height": 12, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 40, + "ofm_sv_height": 1, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 32, + "ofm_width": 20, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 1, + "padding_sv_width": 32, + "psum_buff_offset_scaled": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 672, + 12, + 20 + ], + "dims": [ + 704, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 160, + 12, + 20 + ], + "dims": [ + 160, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Mul_265", + "type": "internal" + } + ], + "layer_order": 163, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 720, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 9, + "ifm_depth_padded": 720, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 5, + "ofm_depth_iter": 1, + "ofm_depth_padded": 160, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 6358464, + "wts_no_of_reads": 4800 + } + }, + "Conv_267": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 64, + "data_io.ofm.dimensions.width": 32, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 160, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 2, + "ifm_height": 12, + "ifm_sign": 0, + "ifm_sv_depth": 80, + "ifm_sv_height": 1, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 32, + "ifm_width": 20, + "ifm_x_iter": 1, + "ifm_y_iter": 3, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 2, + "num_ifm_height_iters": 3, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 6, + "ofm_depth": 960, + "ofm_depth_iter": 6, + "ofm_height": 12, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 40, + "ofm_sv_height": 1, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 32, + "ofm_width": 20, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 1, + "padding_sv_width": 32, + "psum_buff_offset_scaled": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 160, + 12, + 20 + ], + "dims": [ + 160, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_266", + "type": "internal" + } + ], + "layer_order": 164, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 720, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 2, + "ifm_depth_padded": 160, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 5, + "ofm_depth_iter": 6, + "ofm_depth_padded": 960, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 6902784, + "wts_no_of_reads": 4800 + } + }, + "Add_269": { + "AddAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 80, + "ifmsv_height": 1, + "ifmsv_size": 1600, + "ifmsv_width": 20, + "num_kernel_iters": 9, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 3, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_267", + "type": "internal" + } + ], + "layer_order": 165 + }, + "Clip_272": { + "ClipBf16": { + "aie_arch": "aie2p", + "clamp_max": 6, + "clamp_min": 0, + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 80, + "ifmsv_height": 1, + "ifmsv_size": 1600, + "ifmsv_width": 20, + "num_kernel_iters": 9, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Add_269", + "type": "internal" + } + ], + "layer_order": 166 + }, + "Div_274": { + "MulAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 80, + "ifmsv_height": 1, + "ifmsv_size": 1600, + "ifmsv_width": 20, + "num_kernel_iters": 9, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 0.166015625, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Clip_272", + "type": "internal" + } + ], + "layer_order": 167 + }, + "Mul_275": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 24, + "ifmsv_height": 2, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_elems": 240, + "num_kernel_iters": 20, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_267", + "type": "internal" + }, + { + "name": "Div_274", + "type": "internal" + } + ], + "layer_order": 168 + }, + "Conv_276": { + "Conv2DBf16": { + "act": 0, + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 0, + "batch_size": 1, + "bias_shift": 0, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 0, + "data_io.ifm.dimensions.batch.padding": 0, + "data_io.ifm.dimensions.batch.size": 1, + "data_io.ofm.dimensions.batch.padding": 0, + "data_io.ofm.dimensions.batch.size": 1, + "data_io.ofm.dimensions.width.padding": 0, + "func_type": 10, + "ifm_depth": 960, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 1, + "ifm_height": 12, + "ifm_sign": 0, + "ifm_sv_depth": 8, + "ifm_sv_depth.size": 8, + "ifm_sv_height": 10, + "ifm_sv_height.size": 10, + "ifm_sv_width": 16, + "ifm_sv_width.size": 16, + "ifm_width": 20, + "ifm_x_iter": 3, + "ifm_y_iter": 2, + "ifmsv_size": 1280, + "kernel_height": 9, + "kernel_width": 9, + "num_adf_cols": 4, + "num_adf_rows": 4, + "ofm_depth": 960, + "ofm_depth_iter": 30, + "ofm_height": 12, + "ofm_len": 180, + "ofm_sv_depth": 8, + "ofm_sv_height": 2, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 5, + "ofm_width": 17, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 4, + "pad_left": 4, + "pad_right": 4, + "pad_top": 4, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 1, + "stride": 1, + "stride_height": 1, + "stride_width": 1, + "zp_en": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 16, + 24 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Mul_275", + "type": "internal" + } + ], + "layer_order": 169, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 144, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 8, + "is_dwc": true, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 9, + "kernel_width": 12, + "krnl_w_div": 1, + "num_bias_sub_groups": 1, + "ofm_depth_iter": 30, + "ofm_depth_padded": 960, + "output_stride": 8, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 7628544, + "wts_no_of_reads": 1296 + } + }, + "Add_278": { + "AddAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 80, + "ifmsv_height": 1, + "ifmsv_size": 1600, + "ifmsv_width": 20, + "num_kernel_iters": 9, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 3, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 16, + 24 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_276", + "type": "internal" + } + ], + "layer_order": 170 + }, + "Clip_281": { + "ClipBf16": { + "aie_arch": "aie2p", + "clamp_max": 6, + "clamp_min": 0, + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 80, + "ifmsv_height": 1, + "ifmsv_size": 1600, + "ifmsv_width": 20, + "num_kernel_iters": 9, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Add_278", + "type": "internal" + } + ], + "layer_order": 171 + }, + "Div_283": { + "MulAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 80, + "ifmsv_height": 1, + "ifmsv_size": 1600, + "ifmsv_width": 20, + "num_kernel_iters": 9, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 0.166015625, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Clip_281", + "type": "internal" + } + ], + "layer_order": 172 + }, + "Mul_284": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 24, + "ifmsv_height": 2, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_elems": 240, + "num_kernel_iters": 20, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_276", + "type": "internal" + }, + { + "name": "Div_283", + "type": "internal" + } + ], + "layer_order": 173 + }, + "Generated-#42": { + "Transpose4dAdf": { + "aie_arch": "aie2p", + "dim_0": 12, + "dim_1": 120, + "dim_2": 20, + "dim_3": 8, + "dtype": "bfloat16", + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "output_l3_format": "HCWN_C8", + "perm": 6 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 1, + 240 + ], + "dims": [ + 960, + 1, + 240 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Mul_284", + "type": "internal" + } + ], + "layer_order": 174 + }, + "Generated-#44": { + "ReduceMeanC8Bf16": { + "L3_depth": 960, + "L3_height": 1, + "L3_width": 240, + "aie_arch": "aie2p", + "compiler": "chess", + "depth_iter": 6, + "full_channel": 960, + "full_height": 1, + "full_width": 240, + "height_iter": 1, + "loop_num_channel": 24, + "loop_num_height": 1, + "loop_num_width": 5, + "num_iters": 30, + "pad_value": 0, + "reduce_axis": 2, + "reduce_dim": "W", + "scale": 0.004180908203125, + "scale_shift": 0, + "sv_channel": 40, + "sv_height": 1, + "sv_width": 48, + "width_iter": 5 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 1, + 240 + ], + "dims": [ + 960, + 1, + 240 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 1, + 1 + ], + "dims": [ + 960, + 4, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Generated-#42", + "type": "internal" + } + ], + "layer_order": 175 + }, + "Conv_286": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 1, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 12, + "data_io.ofm.dimensions.width": 16, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 960, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 5, + "ifm_height": 1, + "ifm_sign": 0, + "ifm_sv_depth": 192, + "ifm_sv_height": 1, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 16, + "ifm_width": 1, + "ifm_x_iter": 1, + "ifm_y_iter": 1, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 0, + "lrelu_alpha_kernel": 0, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 5, + "num_ifm_height_iters": 1, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 4, + "ofm_depth": 240, + "ofm_depth_iter": 4, + "ofm_height": 1, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 16, + "ofm_sv_height": 1, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 16, + "ofm_width": 1, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 1, + "padding_sv_width": 16, + "psum_buff_offset_scaled": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 1, + 1 + ], + "dims": [ + 960, + 4, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 240, + 1, + 1 + ], + "dims": [ + 256, + 1, + 64 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Generated-#44", + "type": "internal" + } + ], + "layer_order": 176, + "split_width": true, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 288, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 5, + "ifm_depth_padded": 960, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 2, + "ofm_depth_iter": 4, + "ofm_depth_padded": 256, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 8112384, + "wts_no_of_reads": 4608 + } + }, + "Conv_288": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 12, + "data_io.ofm.dimensions.width": 8, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 240, + "ifm_depth_concat_extend": 16, + "ifm_depth_iter": 3, + "ifm_height": 1, + "ifm_sign": 0, + "ifm_sv_depth": 80, + "ifm_sv_height": 1, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 8, + "ifm_width": 1, + "ifm_x_iter": 1, + "ifm_y_iter": 1, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 3, + "num_ifm_height_iters": 1, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 5, + "ofm_depth": 960, + "ofm_depth_iter": 5, + "ofm_height": 1, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 48, + "ofm_sv_height": 1, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 8, + "ofm_width": 1, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 1, + "padding_sv_width": 8, + "psum_buff_offset_scaled": 1, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 240, + 1, + 1 + ], + "dims": [ + 256, + 1, + 64 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 1, + 1 + ], + "dims": [ + 960, + 1, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_286", + "type": "internal" + } + ], + "layer_order": 177, + "split_width": true, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 864, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 3, + "ifm_depth_padded": 240, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 6, + "ofm_depth_iter": 5, + "ofm_depth_padded": 960, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 9241344, + "wts_no_of_reads": 5760 + } + }, + "Add_290": { + "AddAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 384, + "ifmsv_height": 1, + "ifmsv_size": 384, + "ifmsv_width": 1, + "num_kernel_iters": 1, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 3, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 1, + 1 + ], + "dims": [ + 960, + 1, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 1, + 1 + ], + "dims": [ + 1536, + 4, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_288", + "type": "internal" + } + ], + "layer_order": 178 + }, + "Clip_293": { + "ClipBf16": { + "aie_arch": "aie2p", + "clamp_max": 6, + "clamp_min": 0, + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 256, + "ifmsv_height": 1, + "ifmsv_size": 256, + "ifmsv_width": 1, + "num_kernel_iters": 1, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 1, + 1 + ], + "dims": [ + 1536, + 4, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 1, + 1 + ], + "dims": [ + 1024, + 4, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Add_290", + "type": "internal" + } + ], + "layer_order": 179 + }, + "Div_295": { + "MulAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 256, + "ifmsv_height": 2, + "ifmsv_size": 512, + "ifmsv_width": 1, + "num_kernel_iters": 1, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 0.166015625, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 1, + 1 + ], + "dims": [ + 1024, + 4, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 1, + 1 + ], + "dims": [ + 1024, + 8, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Clip_293", + "type": "internal" + } + ], + "layer_order": 180 + }, + "Generated-#46": { + "TileAdf": { + "aie_arch": "aie2p", + "dtype": "bfloat16", + "i_dim_c": 960, + "i_dim_h": 1, + "i_dim_n": 1, + "i_dim_w": 1, + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "output_l3_format": "HCWN_C8", + "rep_dim_c": 1, + "rep_dim_h": 12, + "rep_dim_w": 20 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 1, + 1 + ], + "dims": [ + 960, + 1, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Div_295", + "type": "internal" + } + ], + "layer_order": 181 + }, + "Mul_296": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_elems": 240, + "num_kernel_iters": 15, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Generated-#46", + "type": "internal" + }, + { + "name": "Mul_284", + "type": "internal" + } + ], + "layer_order": 182 + }, + "Conv_297": { + "AddBf16": { + "act": 0, + "act_type": "LINEAR", + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 24, + "ifmsv_height": 1, + "ifmsv_size": 768, + "ifmsv_width": 32, + "num_elems": 240, + "num_kernel_iters": 6, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 64, + "data_io.ofm.dimensions.width": 32, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 11, + "ifm_depth": 960, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 9, + "ifm_height": 12, + "ifm_sign": 0, + "ifm_sv_depth": 112, + "ifm_sv_height": 1, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 32, + "ifm_width": 20, + "ifm_x_iter": 1, + "ifm_y_iter": 3, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 9, + "num_ifm_height_iters": 3, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 2, + "ofm_depth": 160, + "ofm_depth_iter": 2, + "ofm_height": 12, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 24, + "ofm_sv_height": 1, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 32, + "ofm_width": 20, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 1, + "padding_sv_width": 32, + "psum_buff_offset_scaled": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 160, + 12, + 20 + ], + "dims": [ + 192, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 160, + 12, + 20 + ], + "dims": [ + 192, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Mul_296", + "type": "internal" + }, + { + "name": "Conv_266", + "type": "internal" + } + ], + "layer_order": 183, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 432, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 9, + "ifm_depth_padded": 1008, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 3, + "ofm_depth_iter": 2, + "ofm_depth_padded": 192, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 10329984, + "wts_no_of_reads": 4032 + } + }, + "Conv_299": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 64, + "data_io.ofm.dimensions.width": 32, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 160, + "ifm_depth_concat_extend": 32, + "ifm_depth_iter": 2, + "ifm_height": 12, + "ifm_sign": 0, + "ifm_sv_depth": 80, + "ifm_sv_height": 1, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 32, + "ifm_width": 20, + "ifm_x_iter": 1, + "ifm_y_iter": 3, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 2, + "num_ifm_height_iters": 3, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 6, + "ofm_depth": 960, + "ofm_depth_iter": 6, + "ofm_height": 12, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 40, + "ofm_sv_height": 1, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 32, + "ofm_width": 20, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 1, + "padding_sv_width": 32, + "psum_buff_offset_scaled": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 160, + 12, + 20 + ], + "dims": [ + 192, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_297", + "type": "internal" + } + ], + "layer_order": 184, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 720, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 2, + "ifm_depth_padded": 160, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 5, + "ofm_depth_iter": 6, + "ofm_depth_padded": 960, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 11232000, + "wts_no_of_reads": 4800 + } + }, + "Add_301": { + "AddAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 80, + "ifmsv_height": 1, + "ifmsv_size": 1600, + "ifmsv_width": 20, + "num_kernel_iters": 9, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 3, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_299", + "type": "internal" + } + ], + "layer_order": 185 + }, + "Clip_304": { + "ClipBf16": { + "aie_arch": "aie2p", + "clamp_max": 6, + "clamp_min": 0, + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 80, + "ifmsv_height": 1, + "ifmsv_size": 1600, + "ifmsv_width": 20, + "num_kernel_iters": 9, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Add_301", + "type": "internal" + } + ], + "layer_order": 186 + }, + "Div_306": { + "MulAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 80, + "ifmsv_height": 1, + "ifmsv_size": 1600, + "ifmsv_width": 20, + "num_kernel_iters": 9, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 0.166015625, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Clip_304", + "type": "internal" + } + ], + "layer_order": 187 + }, + "Mul_307": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 24, + "ifmsv_height": 2, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_elems": 240, + "num_kernel_iters": 20, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_299", + "type": "internal" + }, + { + "name": "Div_306", + "type": "internal" + } + ], + "layer_order": 188 + }, + "Conv_308": { + "Conv2DBf16": { + "act": 0, + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 0, + "batch_size": 1, + "bias_shift": 0, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 0, + "data_io.ifm.dimensions.batch.padding": 0, + "data_io.ifm.dimensions.batch.size": 1, + "data_io.ofm.dimensions.batch.padding": 0, + "data_io.ofm.dimensions.batch.size": 1, + "data_io.ofm.dimensions.width.padding": 0, + "func_type": 10, + "ifm_depth": 960, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 1, + "ifm_height": 12, + "ifm_sign": 0, + "ifm_sv_depth": 8, + "ifm_sv_depth.size": 8, + "ifm_sv_height": 10, + "ifm_sv_height.size": 10, + "ifm_sv_width": 16, + "ifm_sv_width.size": 16, + "ifm_width": 20, + "ifm_x_iter": 3, + "ifm_y_iter": 2, + "ifmsv_size": 1280, + "kernel_height": 9, + "kernel_width": 9, + "num_adf_cols": 4, + "num_adf_rows": 4, + "ofm_depth": 960, + "ofm_depth_iter": 30, + "ofm_height": 12, + "ofm_len": 180, + "ofm_sv_depth": 8, + "ofm_sv_height": 2, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 5, + "ofm_width": 17, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 4, + "pad_left": 4, + "pad_right": 4, + "pad_top": 4, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 1, + "stride": 1, + "stride_height": 1, + "stride_width": 1, + "zp_en": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 16, + 24 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Mul_307", + "type": "internal" + } + ], + "layer_order": 189, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 144, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 8, + "is_dwc": true, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 9, + "kernel_width": 12, + "krnl_w_div": 1, + "num_bias_sub_groups": 1, + "ofm_depth_iter": 30, + "ofm_depth_padded": 960, + "output_stride": 8, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 11957760, + "wts_no_of_reads": 1296 + } + }, + "Add_310": { + "AddAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 80, + "ifmsv_height": 1, + "ifmsv_size": 1600, + "ifmsv_width": 20, + "num_kernel_iters": 9, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 3, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 16, + 24 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_308", + "type": "internal" + } + ], + "layer_order": 190 + }, + "Clip_313": { + "ClipBf16": { + "aie_arch": "aie2p", + "clamp_max": 6, + "clamp_min": 0, + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 80, + "ifmsv_height": 1, + "ifmsv_size": 1600, + "ifmsv_width": 20, + "num_kernel_iters": 9, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Add_310", + "type": "internal" + } + ], + "layer_order": 191 + }, + "Div_315": { + "MulAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 80, + "ifmsv_height": 1, + "ifmsv_size": 1600, + "ifmsv_width": 20, + "num_kernel_iters": 9, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 0.166015625, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Clip_313", + "type": "internal" + } + ], + "layer_order": 192 + }, + "Mul_316": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 24, + "ifmsv_height": 2, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_elems": 240, + "num_kernel_iters": 20, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_308", + "type": "internal" + }, + { + "name": "Div_315", + "type": "internal" + } + ], + "layer_order": 193 + }, + "Generated-#48": { + "Transpose4dAdf": { + "aie_arch": "aie2p", + "dim_0": 12, + "dim_1": 120, + "dim_2": 20, + "dim_3": 8, + "dtype": "bfloat16", + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "output_l3_format": "HCWN_C8", + "perm": 6 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 1, + 240 + ], + "dims": [ + 960, + 1, + 240 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Mul_316", + "type": "internal" + } + ], + "layer_order": 194 + }, + "Generated-#50": { + "ReduceMeanC8Bf16": { + "L3_depth": 960, + "L3_height": 1, + "L3_width": 240, + "aie_arch": "aie2p", + "compiler": "chess", + "depth_iter": 6, + "full_channel": 960, + "full_height": 1, + "full_width": 240, + "height_iter": 1, + "loop_num_channel": 24, + "loop_num_height": 1, + "loop_num_width": 5, + "num_iters": 30, + "pad_value": 0, + "reduce_axis": 2, + "reduce_dim": "W", + "scale": 0.004180908203125, + "scale_shift": 0, + "sv_channel": 40, + "sv_height": 1, + "sv_width": 48, + "width_iter": 5 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 1, + 240 + ], + "dims": [ + 960, + 1, + 240 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 1, + 1 + ], + "dims": [ + 960, + 4, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Generated-#48", + "type": "internal" + } + ], + "layer_order": 195 + }, + "Conv_318": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 1, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 12, + "data_io.ofm.dimensions.width": 16, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 960, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 5, + "ifm_height": 1, + "ifm_sign": 0, + "ifm_sv_depth": 192, + "ifm_sv_height": 1, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 16, + "ifm_width": 1, + "ifm_x_iter": 1, + "ifm_y_iter": 1, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 0, + "lrelu_alpha_kernel": 0, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 5, + "num_ifm_height_iters": 1, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 4, + "ofm_depth": 240, + "ofm_depth_iter": 4, + "ofm_height": 1, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 16, + "ofm_sv_height": 1, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 16, + "ofm_width": 1, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 1, + "padding_sv_width": 16, + "psum_buff_offset_scaled": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 1, + 1 + ], + "dims": [ + 960, + 4, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 240, + 1, + 1 + ], + "dims": [ + 256, + 1, + 64 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Generated-#50", + "type": "internal" + } + ], + "layer_order": 196, + "split_width": true, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 288, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 5, + "ifm_depth_padded": 960, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 2, + "ofm_depth_iter": 4, + "ofm_depth_padded": 256, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 12441600, + "wts_no_of_reads": 4608 + } + }, + "Conv_320": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 12, + "data_io.ofm.dimensions.width": 8, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 240, + "ifm_depth_concat_extend": 16, + "ifm_depth_iter": 3, + "ifm_height": 1, + "ifm_sign": 0, + "ifm_sv_depth": 80, + "ifm_sv_height": 1, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 8, + "ifm_width": 1, + "ifm_x_iter": 1, + "ifm_y_iter": 1, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 3, + "num_ifm_height_iters": 1, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 5, + "ofm_depth": 960, + "ofm_depth_iter": 5, + "ofm_height": 1, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 48, + "ofm_sv_height": 1, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 8, + "ofm_width": 1, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 1, + "padding_sv_width": 8, + "psum_buff_offset_scaled": 1, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 240, + 1, + 1 + ], + "dims": [ + 256, + 1, + 64 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 1, + 1 + ], + "dims": [ + 960, + 1, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_318", + "type": "internal" + } + ], + "layer_order": 197, + "split_width": true, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 864, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 3, + "ifm_depth_padded": 240, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 6, + "ofm_depth_iter": 5, + "ofm_depth_padded": 960, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 13570560, + "wts_no_of_reads": 5760 + } + }, + "Add_322": { + "AddAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 384, + "ifmsv_height": 1, + "ifmsv_size": 384, + "ifmsv_width": 1, + "num_kernel_iters": 1, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 3, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 1, + 1 + ], + "dims": [ + 960, + 1, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 1, + 1 + ], + "dims": [ + 1536, + 4, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_320", + "type": "internal" + } + ], + "layer_order": 198 + }, + "Clip_325": { + "ClipBf16": { + "aie_arch": "aie2p", + "clamp_max": 6, + "clamp_min": 0, + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 256, + "ifmsv_height": 1, + "ifmsv_size": 256, + "ifmsv_width": 1, + "num_kernel_iters": 1, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 1, + 1 + ], + "dims": [ + 1536, + 4, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 1, + 1 + ], + "dims": [ + 1024, + 4, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Add_322", + "type": "internal" + } + ], + "layer_order": 199 + }, + "Div_327": { + "MulAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 256, + "ifmsv_height": 2, + "ifmsv_size": 512, + "ifmsv_width": 1, + "num_kernel_iters": 1, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 0.166015625, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 1, + 1 + ], + "dims": [ + 1024, + 4, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 1, + 1 + ], + "dims": [ + 1024, + 8, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Clip_325", + "type": "internal" + } + ], + "layer_order": 200 + }, + "Generated-#52": { + "TileAdf": { + "aie_arch": "aie2p", + "dtype": "bfloat16", + "i_dim_c": 960, + "i_dim_h": 1, + "i_dim_n": 1, + "i_dim_w": 1, + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "output_l3_format": "HCWN_C8", + "rep_dim_c": 1, + "rep_dim_h": 12, + "rep_dim_w": 20 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 1, + 1 + ], + "dims": [ + 960, + 1, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Div_327", + "type": "internal" + } + ], + "layer_order": 201 + }, + "Mul_328": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_elems": 240, + "num_kernel_iters": 15, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Generated-#52", + "type": "internal" + }, + { + "name": "Mul_316", + "type": "internal" + } + ], + "layer_order": 202 + }, + "Conv_329": { + "AddBf16": { + "act": 0, + "act_type": "LINEAR", + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 24, + "ifmsv_height": 1, + "ifmsv_size": 768, + "ifmsv_width": 32, + "num_elems": 240, + "num_kernel_iters": 6, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 64, + "data_io.ofm.dimensions.width": 32, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 11, + "ifm_depth": 960, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 9, + "ifm_height": 12, + "ifm_sign": 0, + "ifm_sv_depth": 112, + "ifm_sv_height": 1, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 32, + "ifm_width": 20, + "ifm_x_iter": 1, + "ifm_y_iter": 3, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 9, + "num_ifm_height_iters": 3, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 2, + "ofm_depth": 160, + "ofm_depth_iter": 2, + "ofm_height": 12, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 24, + "ofm_sv_height": 1, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 32, + "ofm_width": 20, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 1, + "padding_sv_width": 32, + "psum_buff_offset_scaled": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 160, + 12, + 20 + ], + "dims": [ + 192, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 160, + 12, + 20 + ], + "dims": [ + 192, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Mul_328", + "type": "internal" + }, + { + "name": "Conv_297", + "type": "internal" + } + ], + "layer_order": 203, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 432, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 9, + "ifm_depth_padded": 1008, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 3, + "ofm_depth_iter": 2, + "ofm_depth_padded": 192, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 14659200, + "wts_no_of_reads": 4032 + } + }, + "Conv_331": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 64, + "data_io.ofm.dimensions.width": 32, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 160, + "ifm_depth_concat_extend": 32, + "ifm_depth_iter": 2, + "ifm_height": 12, + "ifm_sign": 0, + "ifm_sv_depth": 80, + "ifm_sv_height": 1, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 32, + "ifm_width": 20, + "ifm_x_iter": 1, + "ifm_y_iter": 3, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 2, + "num_ifm_height_iters": 3, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 6, + "ofm_depth": 960, + "ofm_depth_iter": 6, + "ofm_height": 12, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 40, + "ofm_sv_height": 1, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 32, + "ofm_width": 20, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 1, + "padding_sv_width": 32, + "psum_buff_offset_scaled": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 160, + 12, + 20 + ], + "dims": [ + 192, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_329", + "type": "internal" + } + ], + "layer_order": 204, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 720, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 2, + "ifm_depth_padded": 160, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 5, + "ofm_depth_iter": 6, + "ofm_depth_padded": 960, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 15561216, + "wts_no_of_reads": 4800 + } + }, + "Add_333": { + "AddAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 80, + "ifmsv_height": 1, + "ifmsv_size": 1600, + "ifmsv_width": 20, + "num_kernel_iters": 9, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 3, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_331", + "type": "internal" + } + ], + "layer_order": 205 + }, + "Clip_336": { + "ClipBf16": { + "aie_arch": "aie2p", + "clamp_max": 6, + "clamp_min": 0, + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 80, + "ifmsv_height": 1, + "ifmsv_size": 1600, + "ifmsv_width": 20, + "num_kernel_iters": 9, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Add_333", + "type": "internal" + } + ], + "layer_order": 206 + }, + "Div_338": { + "MulAttributeBroadcastingBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 80, + "ifmsv_height": 1, + "ifmsv_size": 1600, + "ifmsv_width": 20, + "num_kernel_iters": 9, + "ofm_shift": 0, + "ofm_shift_biased": 0, + "scalar": 0.166015625, + "scalar_position": 1, + "scalar_shift": 0, + "scalar_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Clip_336", + "type": "internal" + } + ], + "layer_order": 207 + }, + "Mul_339": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 24, + "ifmsv_height": 2, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_elems": 240, + "num_kernel_iters": 20, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_331", + "type": "internal" + }, + { + "name": "Div_338", + "type": "internal" + } + ], + "layer_order": 208 + }, + "Generated-#54": { + "Transpose4dAdf": { + "aie_arch": "aie2p", + "dim_0": 12, + "dim_1": 120, + "dim_2": 20, + "dim_3": 8, + "dtype": "bfloat16", + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "output_l3_format": "HCWN_C8", + "perm": 6 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 1, + 240 + ], + "dims": [ + 960, + 1, + 240 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Mul_339", + "type": "internal" + } + ], + "layer_order": 209 + }, + "Generated-#56": { + "ReduceMeanC8Bf16": { + "L3_depth": 960, + "L3_height": 1, + "L3_width": 240, + "aie_arch": "aie2p", + "compiler": "chess", + "depth_iter": 6, + "full_channel": 960, + "full_height": 1, + "full_width": 240, + "height_iter": 1, + "loop_num_channel": 24, + "loop_num_height": 1, + "loop_num_width": 5, + "num_iters": 30, + "pad_value": 0, + "reduce_axis": 2, + "reduce_dim": "W", + "scale": 0.004180908203125, + "scale_shift": 0, + "sv_channel": 40, + "sv_height": 1, + "sv_width": 48, + "width_iter": 5 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 1, + 240 + ], + "dims": [ + 960, + 1, + 240 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 960, + 1, + 1 + ], + "dims": [ + 960, + 4, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Generated-#54", + "type": "internal" + } + ], + "layer_order": 210 + }, + "Conv_343": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 12, + "data_io.ofm.dimensions.width": 16, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 960, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 5, + "ifm_height": 1, + "ifm_sign": 0, + "ifm_sv_depth": 192, + "ifm_sv_height": 1, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 16, + "ifm_width": 1, + "ifm_x_iter": 1, + "ifm_y_iter": 1, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 5, + "num_ifm_height_iters": 1, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 2, + "ofm_depth": 128, + "ofm_depth_iter": 2, + "ofm_height": 1, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 16, + "ofm_sv_height": 1, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 16, + "ofm_width": 1, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 1, + "padding_sv_width": 16, + "psum_buff_offset_scaled": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 1, + 1 + ], + "dims": [ + 960, + 4, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 128, + 1, + 1 + ], + "dims": [ + 128, + 1, + 64 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Generated-#56", + "type": "internal" + } + ], + "layer_order": 211, + "split_width": true, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 288, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 5, + "ifm_depth_padded": 960, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 2, + "ofm_depth_iter": 2, + "ofm_depth_padded": 128, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 16286976, + "wts_no_of_reads": 4608 + } + }, + "Sigmoid_344": { + "SigmoidTemplatedBf16": { + "ENABLE_FP16_AS_BF16": 0, + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 32, + "ifmsv_height": 16, + "ifmsv_size": 512, + "ifmsv_width": 1, + "num_kernel_iters": 1, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 128, + 1, + 1 + ], + "dims": [ + 128, + 1, + 64 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 128, + 1, + 1 + ], + "dims": [ + 128, + 64, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_343", + "type": "internal" + } + ], + "layer_order": 212 + }, + "Generated-#58": { + "TileAdf": { + "aie_arch": "aie2p", + "dtype": "bfloat16", + "i_dim_c": 128, + "i_dim_h": 1, + "i_dim_n": 1, + "i_dim_w": 1, + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "output_l3_format": "HCWN_C8", + "rep_dim_c": 1, + "rep_dim_h": 12, + "rep_dim_w": 20 + }, + "dims_and_bounds": [ + { + "bounds": [ + 128, + 1, + 1 + ], + "dims": [ + 128, + 1, + 1 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 128, + 12, + 20 + ], + "dims": [ + 128, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Sigmoid_344", + "type": "internal" + } + ], + "layer_order": 213 + }, + "Conv_340": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 1, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 0, + "data_io.ofm.dimensions.width": 32, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 16, + "ifm_depth": 960, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 10, + "ifm_height": 12, + "ifm_sign": 0, + "ifm_sv_depth": 96, + "ifm_sv_height": 1, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 32, + "ifm_width": 20, + "ifm_x_iter": 1, + "ifm_y_iter": 3, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 0, + "lrelu_alpha_kernel": 0, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 10, + "num_ifm_height_iters": 3, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 1, + "ofm_depth": 128, + "ofm_depth_iter": 1, + "ofm_height": 12, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 32, + "ofm_sv_height": 1, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 32, + "ofm_width": 20, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 1, + "padding_sv_width": 32, + "psum_buff_offset_scaled": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 32, + "ifmsv_height": 1, + "ifmsv_size": 1024, + "ifmsv_width": 32, + "num_elems": 240, + "num_kernel_iters": 3, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 960, + 12, + 20 + ], + "dims": [ + 960, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 128, + 12, + 20 + ], + "dims": [ + 128, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 128, + 12, + 20 + ], + "dims": [ + 128, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Mul_339", + "type": "internal" + }, + { + "name": "Generated-#58", + "type": "internal" + } + ], + "layer_order": 214, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 576, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 10, + "ifm_depth_padded": 960, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 4, + "ofm_depth_iter": 1, + "ofm_depth_padded": 128, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 16851456, + "wts_no_of_reads": 4608 + } + }, + "Split_349_Duplicated#0": { + "SliceHCWC8Adf": { + "aie_arch": "aie2p", + "axis_letter": "C", + "dim_c": 128, + "dim_h": 12, + "dim_w": 20, + "dtype": "bfloat16", + "end": 64, + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "num_ifm_shim_ch": 2, + "num_ofm_shim_ch": 2, + "output_l3_format": "HCWN_C8", + "start": 0, + "step": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 128, + 12, + 20 + ], + "dims": [ + 128, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 64, + 12, + 20 + ], + "dims": [ + 64, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_340", + "type": "internal" + } + ], + "layer_order": 215 + }, + "Split_349_Duplicated#1": { + "SliceHCWC8Adf": { + "aie_arch": "aie2p", + "axis_letter": "C", + "dim_c": 128, + "dim_h": 12, + "dim_w": 20, + "dtype": "bfloat16", + "end": 128, + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "num_ifm_shim_ch": 2, + "num_ofm_shim_ch": 2, + "output_l3_format": "HCWN_C8", + "start": 64, + "step": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 128, + 12, + 20 + ], + "dims": [ + 128, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 64, + 12, + 20 + ], + "dims": [ + 64, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_340", + "type": "internal" + } + ], + "layer_order": 216 + }, + "Concat_350": { + "Concat": { + "axis": 1, + "dims_and_bounds": [ + { + "bounds": [ + 64, + 12, + 20 + ], + "dims": [ + 64, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 64, + 12, + 20 + ], + "dims": [ + 64, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 128, + 12, + 20 + ], + "dims": [ + 128, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "dtype": "bfloat16" + }, + "input_layer_names": [ + { + "name": "Split_349_Duplicated#1", + "type": "internal" + }, + { + "name": "Input.4", + "type": "external" + } + ], + "layer_order": 217 + }, + "Conv_351": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 0, + "data_io.ofm.dimensions.width": 32, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 1, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 128, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 8, + "ifm_height": 12, + "ifm_sign": 0, + "ifm_sv_depth": 16, + "ifm_sv_height": 5, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 34, + "ifm_width": 20, + "ifm_x_iter": 1, + "ifm_y_iter": 1, + "kernel_height": 3, + "kernel_width": 3, + "ksize.height": 3, + "ksize.width": 3, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 8, + "num_ifm_height_iters": 1, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 2, + "ofm_depth": 128, + "ofm_depth_iter": 2, + "ofm_height": 12, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 16, + "ofm_sv_height": 3, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 32, + "ofm_width": 20, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 1, + "pad_left": 1, + "pad_right": 1, + "pad_top": 1, + "pad_value": 0, + "padding_sv_height": 5, + "padding_sv_width": 34, + "psum_buff_offset_scaled": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 128, + 12, + 20 + ], + "dims": [ + 128, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 128, + 12, + 20 + ], + "dims": [ + 128, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Concat_350", + "type": "internal" + } + ], + "layer_order": 218, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 288, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 8, + "ifm_depth_padded": 128, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 3, + "kernel_width": 3, + "krnl_w_div": 1, + "num_bias_sub_groups": 2, + "ofm_depth_iter": 2, + "ofm_depth_padded": 128, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 17427456, + "wts_no_of_reads": 3456 + } + }, + "Sigmoid_352": { + "SigmoidTemplatedBf16": { + "ENABLE_FP16_AS_BF16": 0, + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 32, + "ifmsv_height": 3, + "ifmsv_size": 1920, + "ifmsv_width": 20, + "num_kernel_iters": 1, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 128, + 12, + 20 + ], + "dims": [ + 128, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 128, + 12, + 20 + ], + "dims": [ + 128, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_351", + "type": "internal" + } + ], + "layer_order": 219 + }, + "Split_353_Duplicated#1": { + "SliceHCWC8Adf": { + "aie_arch": "aie2p", + "axis_letter": "C", + "dim_c": 128, + "dim_h": 12, + "dim_w": 20, + "dtype": "bfloat16", + "end": 128, + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "num_ifm_shim_ch": 2, + "num_ofm_shim_ch": 2, + "output_l3_format": "HCWN_C8", + "start": 64, + "step": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 128, + 12, + 20 + ], + "dims": [ + 128, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 64, + 12, + 20 + ], + "dims": [ + 64, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Sigmoid_352", + "type": "internal" + } + ], + "layer_order": 220 + }, + "Sub_359": { + "SubBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_elems": 240, + "num_kernel_iters": 1, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "constants_reader": { + "Const.2": { + "bytes_to_skip": 20764800, + "channel_stride": 8, + "depth": 1, + "dtype": "bfloat16", + "height": 12, + "nibbles": 4, + "num_reads": 360, + "rank": 4, + "reads_per_line": 3, + "width": 20 + } + }, + "dims_and_bounds": [ + { + "bounds": [ + 64, + 12, + 20 + ], + "dims": [ + 64, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 64, + 12, + 20 + ], + "dims": [ + 64, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 64, + 12, + 20 + ], + "dims": [ + 64, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Const.2", + "type": "constant" + }, + { + "name": "Split_353_Duplicated#1", + "type": "internal" + } + ], + "layer_order": 221 + }, + "Mul_360": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_elems": 240, + "num_kernel_iters": 1, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 64, + 12, + 20 + ], + "dims": [ + 64, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 64, + 12, + 20 + ], + "dims": [ + 64, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 64, + 12, + 20 + ], + "dims": [ + 64, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Sub_359", + "type": "internal" + }, + { + "name": "Input.4", + "type": "external" + } + ], + "layer_order": 222 + }, + "Split_353_Duplicated#0": { + "SliceHCWC8Adf": { + "aie_arch": "aie2p", + "axis_letter": "C", + "dim_c": 128, + "dim_h": 12, + "dim_w": 20, + "dtype": "bfloat16", + "end": 64, + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "num_ifm_shim_ch": 2, + "num_ofm_shim_ch": 2, + "output_l3_format": "HCWN_C8", + "start": 0, + "step": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 128, + 12, + 20 + ], + "dims": [ + 128, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 64, + 12, + 20 + ], + "dims": [ + 64, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Sigmoid_352", + "type": "internal" + } + ], + "layer_order": 223 + }, + "Mul_354": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_elems": 240, + "num_kernel_iters": 1, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 64, + 12, + 20 + ], + "dims": [ + 64, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 64, + 12, + 20 + ], + "dims": [ + 64, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 64, + 12, + 20 + ], + "dims": [ + 64, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Split_353_Duplicated#0", + "type": "internal" + }, + { + "name": "Input.4", + "type": "external" + } + ], + "layer_order": 224 + }, + "Concat_355": { + "Concat": { + "axis": 1, + "dims_and_bounds": [ + { + "bounds": [ + 64, + 12, + 20 + ], + "dims": [ + 64, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 64, + 12, + 20 + ], + "dims": [ + 64, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 128, + 12, + 20 + ], + "dims": [ + 128, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "dtype": "bfloat16" + }, + "input_layer_names": [ + { + "name": "Split_349_Duplicated#1", + "type": "internal" + }, + { + "name": "Mul_354", + "type": "internal" + } + ], + "layer_order": 225 + }, + "Conv_356": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 0, + "data_io.ofm.dimensions.width": 32, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 1, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 128, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 8, + "ifm_height": 12, + "ifm_sign": 0, + "ifm_sv_depth": 16, + "ifm_sv_height": 5, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 34, + "ifm_width": 20, + "ifm_x_iter": 1, + "ifm_y_iter": 1, + "kernel_height": 3, + "kernel_width": 3, + "ksize.height": 3, + "ksize.width": 3, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 8, + "num_ifm_height_iters": 1, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 1, + "ofm_depth": 64, + "ofm_depth_iter": 1, + "ofm_height": 12, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 16, + "ofm_sv_height": 3, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 32, + "ofm_width": 20, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 1, + "pad_left": 1, + "pad_right": 1, + "pad_top": 1, + "pad_value": 0, + "padding_sv_height": 5, + "padding_sv_width": 34, + "psum_buff_offset_scaled": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 128, + 12, + 20 + ], + "dims": [ + 128, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 64, + 12, + 20 + ], + "dims": [ + 64, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Concat_355", + "type": "internal" + } + ], + "layer_order": 226, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 288, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 8, + "ifm_depth_padded": 128, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 3, + "kernel_width": 3, + "krnl_w_div": 1, + "num_bias_sub_groups": 2, + "ofm_depth_iter": 1, + "ofm_depth_padded": 64, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 18109440, + "wts_no_of_reads": 3456 + } + }, + "Tanh_357": { + "TanhTemplatedBf16": { + "ENABLE_FP16_AS_BF16": 0, + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_kernel_iters": 1, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 64, + 12, + 20 + ], + "dims": [ + 64, + 12, + 32 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 64, + 12, + 20 + ], + "dims": [ + 64, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_356", + "type": "internal" + } + ], + "layer_order": 227 + }, + "Mul_361": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_elems": 240, + "num_kernel_iters": 1, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 64, + 12, + 20 + ], + "dims": [ + 64, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 64, + 12, + 20 + ], + "dims": [ + 64, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 64, + 12, + 20 + ], + "dims": [ + 64, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Split_353_Duplicated#1", + "type": "internal" + }, + { + "name": "Tanh_357", + "type": "internal" + } + ], + "layer_order": 228 + }, + "Add_362": { + "AddBf16": { + "act": 0, + "act_type": "LINEAR", + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 3, + "ifmsv_size": 960, + "ifmsv_width": 20, + "num_elems": 240, + "num_kernel_iters": 1, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 64, + 12, + 20 + ], + "dims": [ + 64, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 64, + 12, + 20 + ], + "dims": [ + 64, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 64, + 12, + 20 + ], + "dims": [ + 64, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "external_output_layer": 3, + "input_layer_names": [ + { + "name": "Mul_360", + "type": "internal" + }, + { + "name": "Mul_361", + "type": "internal" + } + ], + "layer_order": 229 + }, + "Concat_363": { + "Concat": { + "axis": 1, + "dims_and_bounds": [ + { + "bounds": [ + 64, + 12, + 20 + ], + "dims": [ + 64, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 64, + 12, + 20 + ], + "dims": [ + 64, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 128, + 12, + 20 + ], + "dims": [ + 128, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "dtype": "bfloat16" + }, + "input_layer_names": [ + { + "name": "Split_349_Duplicated#0", + "type": "internal" + }, + { + "name": "Add_362", + "type": "internal" + } + ], + "layer_order": 230 + }, + "Resize_365": { + "ResizeAdf": { + "co_trans_mode": 1, + "dim_0": 1, + "dim_1": 128, + "dim_2": 12, + "dim_3": 20, + "dtype": "bfloat16", + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "mode": 1, + "nearest_mode": 0, + "num_ifm_shim_ch": 2, + "num_ofm_shim_ch": 2, + "output_H": 24, + "output_W": 40, + "output_l3_format": "HCWN_C8" + }, + "dims_and_bounds": [ + { + "bounds": [ + 128, + 12, + 20 + ], + "dims": [ + 128, + 12, + 20 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 128, + 24, + 40 + ], + "dims": [ + 128, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Concat_363", + "type": "internal" + } + ], + "layer_order": 231 + }, + "Slice_371": { + "SliceHCWC8Adf": { + "aie_arch": "aie2p", + "axis_letter": "H", + "dim_c": 128, + "dim_h": 24, + "dim_w": 40, + "dtype": "bfloat16", + "end": 23, + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "num_ifm_shim_ch": 2, + "num_ofm_shim_ch": 2, + "output_l3_format": "HCWN_C8", + "start": 0, + "step": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 128, + 24, + 40 + ], + "dims": [ + 128, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 128, + 23, + 40 + ], + "dims": [ + 128, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Resize_365", + "type": "internal" + } + ], + "layer_order": 232 + }, + "AveragePool_346": { + "AvgPool2DBf16": { + "aie_arch": "aie2p", + "avgpool_skip_width": 0, + "batch_size": 1, + "compiler": "chess", + "data_io.ofm.dimensions.width": 40, + "data_io.ofm.dimensions.width.padding": 0, + "div_factor": 0.25, + "div_shift": 0, + "dtype": "bfloat16", + "ifm_sign": 0, + "ifm_sv_depth": 8, + "ifm_sv_height": 2, + "ifm_sv_height_eff": 2, + "ifm_sv_size": 1280, + "ifm_sv_width": 80, + "ifm_sv_width_eff": 80, + "kernel_height": 2, + "kernel_width": 2, + "ksize": 2, + "manual_ba_pad_val": 0, + "num_iter": 640, + "ofm_len": 96, + "ofm_sv_depth": 8, + "ofm_sv_height": 1, + "ofm_sv_width": 40, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "stride_height": 2, + "stride_log2": 1, + "stride_width": 2 + }, + "Conv2D": { + "act": 0, + "alpha": 0, + "alpha_scaled": 0, + "batch_size": 0, + "conv2d_ofm_pad": 0, + "conv_type": 0, + "func_type": 6, + "ifm_depth": 8, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 0, + "ifm_height": 0, + "ifm_sign": 0, + "ifm_sv_depth": 0, + "ifm_sv_height": 0, + "ifm_sv_width": 0, + "ifm_width": 0, + "ifm_x_iter": 0, + "ifm_y_iter": 0, + "kernel_height": 0, + "kernel_width": 0, + "num_adf_cols": 0, + "num_adf_rows": 0, + "ofm_depth": 8, + "ofm_depth_iter": 0, + "ofm_height": 0, + "ofm_sv_depth": 0, + "ofm_sv_height": 0, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 0, + "ofm_width": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "relu_int8_enable": 1, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_height": 0, + "stride_width": 0, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 3, + 180, + 320 + ], + "dims": [ + 8, + 180, + 320 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 3, + 90, + 160 + ], + "dims": [ + 8, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Generated-#4", + "type": "internal" + } + ], + "layer_order": 233, + "split_width": false + }, + "AveragePool_347": { + "AvgPool2DBf16": { + "aie_arch": "aie2p", + "avgpool_skip_width": 0, + "batch_size": 1, + "compiler": "chess", + "data_io.ofm.dimensions.width": 8, + "data_io.ofm.dimensions.width.padding": 0, + "div_factor": 0.25, + "div_shift": 0, + "dtype": "bfloat16", + "ifm_sign": 0, + "ifm_sv_depth": 8, + "ifm_sv_height": 12, + "ifm_sv_height_eff": 12, + "ifm_sv_size": 1536, + "ifm_sv_width": 16, + "ifm_sv_width_eff": 16, + "kernel_height": 2, + "kernel_width": 2, + "ksize": 2, + "manual_ba_pad_val": 0, + "num_iter": 768, + "ofm_len": 20, + "ofm_sv_depth": 8, + "ofm_sv_height": 6, + "ofm_sv_width": 8, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "stride_height": 2, + "stride_log2": 1, + "stride_width": 2 + }, + "Conv2D": { + "act": 0, + "alpha": 0, + "alpha_scaled": 0, + "batch_size": 0, + "conv2d_ofm_pad": 0, + "conv_type": 0, + "func_type": 6, + "ifm_depth": 8, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 0, + "ifm_height": 0, + "ifm_sign": 0, + "ifm_sv_depth": 0, + "ifm_sv_height": 0, + "ifm_sv_width": 0, + "ifm_width": 0, + "ifm_x_iter": 0, + "ifm_y_iter": 0, + "kernel_height": 0, + "kernel_width": 0, + "num_adf_cols": 0, + "num_adf_rows": 0, + "ofm_depth": 8, + "ofm_depth_iter": 0, + "ofm_height": 0, + "ofm_sv_depth": 0, + "ofm_sv_height": 0, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 0, + "ofm_width": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "relu_int8_enable": 1, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_height": 0, + "stride_width": 0, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 3, + 90, + 160 + ], + "dims": [ + 8, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 3, + 45, + 80 + ], + "dims": [ + 32, + 48, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "AveragePool_346", + "type": "internal" + } + ], + "layer_order": 234, + "split_width": false + }, + "AveragePool_348": { + "AvgPool2DBf16": { + "aie_arch": "aie2p", + "avgpool_skip_width": 0, + "batch_size": 1, + "compiler": "chess", + "data_io.ofm.dimensions.width": 8, + "data_io.ofm.dimensions.width.padding": 0, + "div_factor": 0.25, + "div_shift": 0, + "dtype": "bfloat16", + "ifm_sign": 0, + "ifm_sv_depth": 8, + "ifm_sv_height": 12, + "ifm_sv_height_eff": 12, + "ifm_sv_size": 1536, + "ifm_sv_width": 16, + "ifm_sv_width_eff": 16, + "kernel_height": 2, + "kernel_width": 2, + "ksize": 2, + "manual_ba_pad_val": 0, + "num_iter": 768, + "ofm_len": 5, + "ofm_sv_depth": 8, + "ofm_sv_height": 6, + "ofm_sv_width": 8, + "pad_bottom": 1, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "stride_height": 2, + "stride_log2": 1, + "stride_width": 2 + }, + "Conv2D": { + "act": 0, + "alpha": 0, + "alpha_scaled": 0, + "batch_size": 0, + "conv2d_ofm_pad": 0, + "conv_type": 0, + "func_type": 6, + "ifm_depth": 8, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 0, + "ifm_height": 0, + "ifm_sign": 0, + "ifm_sv_depth": 0, + "ifm_sv_height": 0, + "ifm_sv_width": 0, + "ifm_width": 0, + "ifm_x_iter": 0, + "ifm_y_iter": 0, + "kernel_height": 0, + "kernel_width": 0, + "num_adf_cols": 0, + "num_adf_rows": 0, + "ofm_depth": 8, + "ofm_depth_iter": 0, + "ofm_height": 0, + "ofm_sv_depth": 0, + "ofm_sv_height": 0, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 0, + "ofm_width": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "relu_int8_enable": 1, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_height": 0, + "stride_width": 0, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 3, + 45, + 80 + ], + "dims": [ + 32, + 48, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 3, + 23, + 40 + ], + "dims": [ + 32, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "AveragePool_347", + "type": "internal" + } + ], + "layer_order": 235, + "split_width": false + }, + "Concat_372": { + "Concat": { + "axis": 1, + "dims_and_bounds": [ + { + "bounds": [ + 128, + 23, + 40 + ], + "dims": [ + 128, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 40, + 23, + 40 + ], + "dims": [ + 40, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 3, + 23, + 40 + ], + "dims": [ + 8, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 171, + 23, + 40 + ], + "dims": [ + 176, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "dtype": "bfloat16" + }, + "input_layer_names": [ + { + "name": "Slice_371", + "type": "internal" + }, + { + "name": "Conv_92", + "type": "internal" + }, + { + "name": "AveragePool_348", + "type": "internal" + } + ], + "layer_order": 236 + }, + "Conv_373": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 1, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 0, + "data_io.ofm.dimensions.width": 48, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 1, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 176, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 11, + "ifm_height": 23, + "ifm_sign": 0, + "ifm_sv_depth": 16, + "ifm_sv_height": 5, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 50, + "ifm_width": 40, + "ifm_x_iter": 1, + "ifm_y_iter": 2, + "kernel_height": 3, + "kernel_width": 3, + "ksize.height": 3, + "ksize.width": 3, + "lrelu_alpha": 0, + "lrelu_alpha_kernel": 0, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 11, + "num_ifm_height_iters": 2, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 2, + "ofm_depth": 80, + "ofm_depth_iter": 2, + "ofm_height": 23, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 16, + "ofm_sv_height": 3, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 48, + "ofm_width": 40, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 1, + "pad_left": 1, + "pad_right": 1, + "pad_top": 1, + "pad_value": 0, + "padding_sv_height": 5, + "padding_sv_width": 50, + "psum_buff_offset_scaled": 3, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 171, + 23, + 40 + ], + "dims": [ + 176, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 80, + 23, + 40 + ], + "dims": [ + 128, + 24, + 48 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Concat_372", + "type": "internal" + } + ], + "layer_order": 237, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 288, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 11, + "ifm_depth_padded": 176, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 3, + "kernel_width": 3, + "krnl_w_div": 1, + "num_bias_sub_groups": 2, + "ofm_depth_iter": 2, + "ofm_depth_padded": 128, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 18450432, + "wts_no_of_reads": 3456 + } + }, + "Split_375_Duplicated#0": { + "SliceHCWC8Adf": { + "aie_arch": "aie2p", + "axis_letter": "C", + "dim_c": 80, + "dim_h": 23, + "dim_w": 40, + "dtype": "bfloat16", + "end": 40, + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "num_ifm_shim_ch": 2, + "num_ofm_shim_ch": 2, + "output_l3_format": "HCWN_C8", + "start": 0, + "step": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 80, + 23, + 40 + ], + "dims": [ + 80, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 40, + 23, + 40 + ], + "dims": [ + 40, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_373", + "type": "internal" + } + ], + "layer_order": 238 + }, + "Split_375_Duplicated#1": { + "SliceHCWC8Adf": { + "aie_arch": "aie2p", + "axis_letter": "C", + "dim_c": 80, + "dim_h": 23, + "dim_w": 40, + "dtype": "bfloat16", + "end": 80, + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "num_ifm_shim_ch": 2, + "num_ofm_shim_ch": 2, + "output_l3_format": "HCWN_C8", + "start": 40, + "step": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 80, + 23, + 40 + ], + "dims": [ + 80, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 40, + 23, + 40 + ], + "dims": [ + 40, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_373", + "type": "internal" + } + ], + "layer_order": 239 + }, + "Concat_376": { + "Concat": { + "axis": 1, + "dims_and_bounds": [ + { + "bounds": [ + 40, + 23, + 40 + ], + "dims": [ + 40, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 40, + 23, + 40 + ], + "dims": [ + 40, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 80, + 23, + 40 + ], + "dims": [ + 80, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "dtype": "bfloat16" + }, + "input_layer_names": [ + { + "name": "Split_375_Duplicated#1", + "type": "internal" + }, + { + "name": "Input.3", + "type": "external" + } + ], + "layer_order": 240 + }, + "Conv_377": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 0, + "data_io.ofm.dimensions.width": 48, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 1, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 80, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 5, + "ifm_height": 23, + "ifm_sign": 0, + "ifm_sv_depth": 16, + "ifm_sv_height": 5, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 50, + "ifm_width": 40, + "ifm_x_iter": 1, + "ifm_y_iter": 2, + "kernel_height": 3, + "kernel_width": 3, + "ksize.height": 3, + "ksize.width": 3, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 5, + "num_ifm_height_iters": 2, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 2, + "ofm_depth": 80, + "ofm_depth_iter": 2, + "ofm_height": 23, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 16, + "ofm_sv_height": 3, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 48, + "ofm_width": 40, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 1, + "pad_left": 1, + "pad_right": 1, + "pad_top": 1, + "pad_value": 0, + "padding_sv_height": 5, + "padding_sv_width": 50, + "psum_buff_offset_scaled": 3, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 80, + 23, + 40 + ], + "dims": [ + 80, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 80, + 23, + 40 + ], + "dims": [ + 128, + 24, + 48 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Concat_376", + "type": "internal" + } + ], + "layer_order": 241, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 288, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 5, + "ifm_depth_padded": 80, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 3, + "kernel_width": 3, + "krnl_w_div": 1, + "num_bias_sub_groups": 2, + "ofm_depth_iter": 2, + "ofm_depth_padded": 128, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 19388160, + "wts_no_of_reads": 3456 + } + }, + "Sigmoid_378": { + "SigmoidTemplatedBf16": { + "ENABLE_FP16_AS_BF16": 0, + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 24, + "ifmsv_height": 2, + "ifmsv_size": 1920, + "ifmsv_width": 40, + "num_kernel_iters": 3, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 80, + 23, + 40 + ], + "dims": [ + 128, + 24, + 48 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 80, + 23, + 40 + ], + "dims": [ + 96, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_377", + "type": "internal" + } + ], + "layer_order": 242 + }, + "Split_379_Duplicated#1": { + "SliceHCWC8Adf": { + "aie_arch": "aie2p", + "axis_letter": "C", + "dim_c": 80, + "dim_h": 23, + "dim_w": 40, + "dtype": "bfloat16", + "end": 80, + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "num_ifm_shim_ch": 2, + "num_ofm_shim_ch": 2, + "output_l3_format": "HCWN_C8", + "start": 40, + "step": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 80, + 23, + 40 + ], + "dims": [ + 80, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 40, + 23, + 40 + ], + "dims": [ + 40, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Sigmoid_378", + "type": "internal" + } + ], + "layer_order": 243 + }, + "Sub_385": { + "SubBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 6, + "ifmsv_size": 960, + "ifmsv_width": 10, + "num_elems": 920, + "num_kernel_iters": 4, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "constants_reader": { + "Const.3": { + "bytes_to_skip": 20764800, + "channel_stride": 8, + "depth": 1, + "dtype": "bfloat16", + "height": 23, + "nibbles": 4, + "num_reads": 1380, + "rank": 4, + "reads_per_line": 3, + "width": 40 + } + }, + "dims_and_bounds": [ + { + "bounds": [ + 40, + 23, + 40 + ], + "dims": [ + 40, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 40, + 23, + 40 + ], + "dims": [ + 40, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 40, + 23, + 40 + ], + "dims": [ + 64, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Const.3", + "type": "constant" + }, + { + "name": "Split_379_Duplicated#1", + "type": "internal" + } + ], + "layer_order": 244 + }, + "Mul_386": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 6, + "ifmsv_size": 960, + "ifmsv_width": 10, + "num_elems": 920, + "num_kernel_iters": 4, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 40, + 23, + 40 + ], + "dims": [ + 64, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 40, + 23, + 40 + ], + "dims": [ + 40, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 40, + 23, + 40 + ], + "dims": [ + 64, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Sub_385", + "type": "internal" + }, + { + "name": "Input.3", + "type": "external" + } + ], + "layer_order": 245 + }, + "Split_379_Duplicated#0": { + "SliceHCWC8Adf": { + "aie_arch": "aie2p", + "axis_letter": "C", + "dim_c": 80, + "dim_h": 23, + "dim_w": 40, + "dtype": "bfloat16", + "end": 40, + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "num_ifm_shim_ch": 2, + "num_ofm_shim_ch": 2, + "output_l3_format": "HCWN_C8", + "start": 0, + "step": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 80, + 23, + 40 + ], + "dims": [ + 80, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 40, + 23, + 40 + ], + "dims": [ + 40, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Sigmoid_378", + "type": "internal" + } + ], + "layer_order": 246 + }, + "Mul_380": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 6, + "ifmsv_size": 960, + "ifmsv_width": 10, + "num_elems": 920, + "num_kernel_iters": 4, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 40, + 23, + 40 + ], + "dims": [ + 40, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 40, + 23, + 40 + ], + "dims": [ + 40, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 40, + 23, + 40 + ], + "dims": [ + 64, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Split_379_Duplicated#0", + "type": "internal" + }, + { + "name": "Input.3", + "type": "external" + } + ], + "layer_order": 247 + }, + "Concat_381": { + "Concat": { + "axis": 1, + "dims_and_bounds": [ + { + "bounds": [ + 40, + 23, + 40 + ], + "dims": [ + 40, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 40, + 23, + 40 + ], + "dims": [ + 40, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 80, + 23, + 40 + ], + "dims": [ + 80, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "dtype": "bfloat16" + }, + "input_layer_names": [ + { + "name": "Split_375_Duplicated#1", + "type": "internal" + }, + { + "name": "Mul_380", + "type": "internal" + } + ], + "layer_order": 248 + }, + "Conv_382": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 0, + "data_io.ofm.dimensions.width": 48, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 1, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 80, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 5, + "ifm_height": 23, + "ifm_sign": 0, + "ifm_sv_depth": 16, + "ifm_sv_height": 5, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 50, + "ifm_width": 40, + "ifm_x_iter": 1, + "ifm_y_iter": 2, + "kernel_height": 3, + "kernel_width": 3, + "ksize.height": 3, + "ksize.width": 3, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 5, + "num_ifm_height_iters": 2, + "num_ifm_width_iters": 1, + "num_ofm_depth_iters": 1, + "ofm_depth": 40, + "ofm_depth_iter": 1, + "ofm_height": 23, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 16, + "ofm_sv_height": 3, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 48, + "ofm_width": 40, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 1, + "pad_left": 1, + "pad_right": 1, + "pad_top": 1, + "pad_value": 0, + "padding_sv_height": 5, + "padding_sv_width": 50, + "psum_buff_offset_scaled": 3, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 80, + 23, + 40 + ], + "dims": [ + 80, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 40, + 23, + 40 + ], + "dims": [ + 64, + 24, + 48 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Concat_381", + "type": "internal" + } + ], + "layer_order": 249, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 288, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 5, + "ifm_depth_padded": 80, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 3, + "kernel_width": 3, + "krnl_w_div": 1, + "num_bias_sub_groups": 2, + "ofm_depth_iter": 1, + "ofm_depth_padded": 64, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 19814400, + "wts_no_of_reads": 3456 + } + }, + "Tanh_383": { + "TanhTemplatedBf16": { + "ENABLE_FP16_AS_BF16": 0, + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 6, + "ifmsv_size": 1920, + "ifmsv_width": 20, + "num_kernel_iters": 2, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 40, + 23, + 40 + ], + "dims": [ + 64, + 24, + 48 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 40, + 23, + 40 + ], + "dims": [ + 64, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_382", + "type": "internal" + } + ], + "layer_order": 250 + }, + "Mul_387": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 6, + "ifmsv_size": 960, + "ifmsv_width": 10, + "num_elems": 920, + "num_kernel_iters": 4, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 40, + 23, + 40 + ], + "dims": [ + 40, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 40, + 23, + 40 + ], + "dims": [ + 64, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 40, + 23, + 40 + ], + "dims": [ + 64, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Split_379_Duplicated#1", + "type": "internal" + }, + { + "name": "Tanh_383", + "type": "internal" + } + ], + "layer_order": 251 + }, + "Add_388": { + "AddBf16": { + "act": 0, + "act_type": "LINEAR", + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 6, + "ifmsv_size": 960, + "ifmsv_width": 10, + "num_elems": 920, + "num_kernel_iters": 4, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 40, + 23, + 40 + ], + "dims": [ + 64, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 40, + 23, + 40 + ], + "dims": [ + 64, + 24, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 40, + 23, + 40 + ], + "dims": [ + 40, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "external_output_layer": 2, + "input_layer_names": [ + { + "name": "Mul_386", + "type": "internal" + }, + { + "name": "Mul_387", + "type": "internal" + } + ], + "layer_order": 252 + }, + "Concat_389": { + "Concat": { + "axis": 1, + "dims_and_bounds": [ + { + "bounds": [ + 40, + 23, + 40 + ], + "dims": [ + 40, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 40, + 23, + 40 + ], + "dims": [ + 40, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 80, + 23, + 40 + ], + "dims": [ + 80, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "dtype": "bfloat16" + }, + "input_layer_names": [ + { + "name": "Split_375_Duplicated#0", + "type": "internal" + }, + { + "name": "Add_388", + "type": "internal" + } + ], + "layer_order": 253 + }, + "Resize_391": { + "ResizeAdf": { + "co_trans_mode": 1, + "dim_0": 1, + "dim_1": 80, + "dim_2": 23, + "dim_3": 40, + "dtype": "bfloat16", + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "mode": 1, + "nearest_mode": 0, + "num_ifm_shim_ch": 2, + "num_ofm_shim_ch": 2, + "output_H": 46, + "output_W": 80, + "output_l3_format": "HCWN_C8" + }, + "dims_and_bounds": [ + { + "bounds": [ + 80, + 23, + 40 + ], + "dims": [ + 80, + 23, + 40 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 80, + 46, + 80 + ], + "dims": [ + 80, + 46, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Concat_389", + "type": "internal" + } + ], + "layer_order": 254 + }, + "Slice_397": { + "SliceHCWC8Adf": { + "aie_arch": "aie2p", + "axis_letter": "H", + "dim_c": 80, + "dim_h": 46, + "dim_w": 80, + "dtype": "bfloat16", + "end": 45, + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "num_ifm_shim_ch": 2, + "num_ofm_shim_ch": 2, + "output_l3_format": "HCWN_C8", + "start": 0, + "step": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 80, + 46, + 80 + ], + "dims": [ + 80, + 46, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 80, + 45, + 80 + ], + "dims": [ + 80, + 45, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Resize_391", + "type": "internal" + } + ], + "layer_order": 255 + }, + "Concat_398": { + "Concat": { + "axis": 1, + "dims_and_bounds": [ + { + "bounds": [ + 80, + 45, + 80 + ], + "dims": [ + 80, + 45, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 24, + 45, + 80 + ], + "dims": [ + 24, + 45, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 3, + 45, + 80 + ], + "dims": [ + 8, + 45, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 107, + 45, + 80 + ], + "dims": [ + 112, + 45, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "dtype": "bfloat16" + }, + "input_layer_names": [ + { + "name": "Slice_397", + "type": "internal" + }, + { + "name": "Conv_39", + "type": "internal" + }, + { + "name": "AveragePool_347", + "type": "internal" + } + ], + "layer_order": 256 + }, + "Conv_399": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 1, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 0, + "data_io.ofm.dimensions.width": 16, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 1, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 112, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 7, + "ifm_height": 45, + "ifm_sign": 0, + "ifm_sv_depth": 16, + "ifm_sv_height": 14, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 18, + "ifm_width": 80, + "ifm_x_iter": 5, + "ifm_y_iter": 1, + "kernel_height": 3, + "kernel_width": 3, + "ksize.height": 3, + "ksize.width": 3, + "lrelu_alpha": 0, + "lrelu_alpha_kernel": 0, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 7, + "num_ifm_height_iters": 1, + "num_ifm_width_iters": 5, + "num_ofm_depth_iters": 1, + "ofm_depth": 40, + "ofm_depth_iter": 1, + "ofm_height": 45, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 16, + "ofm_sv_height": 12, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 16, + "ofm_width": 80, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 1, + "pad_left": 1, + "pad_right": 1, + "pad_top": 1, + "pad_value": 0, + "padding_sv_height": 14, + "padding_sv_width": 18, + "psum_buff_offset_scaled": 1, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 107, + 45, + 80 + ], + "dims": [ + 112, + 45, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 40, + 45, + 80 + ], + "dims": [ + 64, + 48, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Concat_398", + "type": "internal" + } + ], + "layer_order": 257, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 288, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 7, + "ifm_depth_padded": 112, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 3, + "kernel_width": 3, + "krnl_w_div": 1, + "num_bias_sub_groups": 2, + "ofm_depth_iter": 1, + "ofm_depth_padded": 64, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 20027520, + "wts_no_of_reads": 3456 + } + }, + "Split_401_Duplicated#0": { + "SliceHCWC8Adf": { + "aie_arch": "aie2p", + "axis_letter": "C", + "dim_c": 40, + "dim_h": 45, + "dim_w": 80, + "dtype": "bfloat16", + "end": 20, + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "num_ifm_shim_ch": 2, + "num_ofm_shim_ch": 2, + "output_l3_format": "HCWN_C8", + "start": 0, + "step": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 40, + 45, + 80 + ], + "dims": [ + 40, + 45, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 20, + 45, + 80 + ], + "dims": [ + 24, + 45, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_399", + "type": "internal" + } + ], + "layer_order": 258 + }, + "Split_401_Duplicated#1": { + "SliceHCWC8Adf": { + "aie_arch": "aie2p", + "axis_letter": "C", + "dim_c": 40, + "dim_h": 45, + "dim_w": 80, + "dtype": "bfloat16", + "end": 40, + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "num_ifm_shim_ch": 2, + "num_ofm_shim_ch": 2, + "output_l3_format": "HCWN_C8", + "start": 20, + "step": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 40, + 45, + 80 + ], + "dims": [ + 40, + 45, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 20, + 45, + 80 + ], + "dims": [ + 24, + 45, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_399", + "type": "internal" + } + ], + "layer_order": 259 + }, + "Concat_402": { + "ConcatC8Adf": { + "aie_arch": "aie2p", + "dtype": "bfloat16", + "in1_dim_c": 24, + "in1_dim_h": 45, + "in1_dim_w": 80, + "in2_dim_c": 24, + "in2_dim_h": 45, + "in2_dim_w": 80, + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "num_eff_concat_input0_size": 20, + "num_eff_concat_input0_start": 0, + "num_eff_concat_input1_size": 20, + "num_eff_concat_input1_start": 0, + "output_l3_format": "HCWN_C8" + }, + "dims_and_bounds": [ + { + "bounds": [ + 20, + 45, + 80 + ], + "dims": [ + 24, + 45, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 20, + 45, + 80 + ], + "dims": [ + 24, + 45, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 40, + 45, + 80 + ], + "dims": [ + 40, + 45, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Split_401_Duplicated#1", + "type": "internal" + }, + { + "name": "Input.2", + "type": "external" + } + ], + "layer_order": 260 + }, + "Conv_403": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 0, + "data_io.ofm.dimensions.width": 16, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 1, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 40, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 3, + "ifm_height": 45, + "ifm_sign": 0, + "ifm_sv_depth": 16, + "ifm_sv_height": 14, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 18, + "ifm_width": 80, + "ifm_x_iter": 5, + "ifm_y_iter": 1, + "kernel_height": 3, + "kernel_width": 3, + "ksize.height": 3, + "ksize.width": 3, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 3, + "num_ifm_height_iters": 1, + "num_ifm_width_iters": 5, + "num_ofm_depth_iters": 1, + "ofm_depth": 40, + "ofm_depth_iter": 1, + "ofm_height": 45, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 16, + "ofm_sv_height": 12, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 16, + "ofm_width": 80, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 1, + "pad_left": 1, + "pad_right": 1, + "pad_top": 1, + "pad_value": 0, + "padding_sv_height": 14, + "padding_sv_width": 18, + "psum_buff_offset_scaled": 1, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 40, + 45, + 80 + ], + "dims": [ + 40, + 45, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 40, + 45, + 80 + ], + "dims": [ + 64, + 48, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Concat_402", + "type": "internal" + } + ], + "layer_order": 261, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 288, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 3, + "ifm_depth_padded": 48, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 3, + "kernel_width": 3, + "krnl_w_div": 1, + "num_bias_sub_groups": 2, + "ofm_depth_iter": 1, + "ofm_depth_padded": 64, + "output_stride": 16, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 20325888, + "wts_no_of_reads": 3456 + } + }, + "Sigmoid_404": { + "SigmoidTemplatedBf16": { + "ENABLE_FP16_AS_BF16": 0, + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 16, + "ifmsv_height": 12, + "ifmsv_size": 1920, + "ifmsv_width": 10, + "num_kernel_iters": 8, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 40, + 45, + 80 + ], + "dims": [ + 64, + 48, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 40, + 45, + 80 + ], + "dims": [ + 64, + 48, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_403", + "type": "internal" + } + ], + "layer_order": 262 + }, + "Split_405_Duplicated#1": { + "SliceHCWC8Adf": { + "aie_arch": "aie2p", + "axis_letter": "C", + "dim_c": 40, + "dim_h": 45, + "dim_w": 80, + "dtype": "bfloat16", + "end": 40, + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "num_ifm_shim_ch": 2, + "num_ofm_shim_ch": 2, + "output_l3_format": "HCWN_C8", + "start": 20, + "step": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 40, + 45, + 80 + ], + "dims": [ + 40, + 45, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 20, + 45, + 80 + ], + "dims": [ + 24, + 45, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Sigmoid_404", + "type": "internal" + } + ], + "layer_order": 263 + }, + "Sub_411": { + "SubBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 8, + "ifmsv_height": 12, + "ifmsv_size": 960, + "ifmsv_width": 10, + "num_elems": 3600, + "num_kernel_iters": 8, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "constants_reader": { + "Const.4": { + "bytes_to_skip": 20764800, + "channel_stride": 8, + "depth": 1, + "dtype": "bfloat16", + "height": 45, + "nibbles": 4, + "num_reads": 5400, + "rank": 4, + "reads_per_line": 3, + "width": 80 + } + }, + "dims_and_bounds": [ + { + "bounds": [ + 20, + 45, + 80 + ], + "dims": [ + 24, + 45, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 20, + 45, + 80 + ], + "dims": [ + 24, + 45, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 20, + 45, + 80 + ], + "dims": [ + 32, + 48, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Const.4", + "type": "constant" + }, + { + "name": "Split_405_Duplicated#1", + "type": "internal" + } + ], + "layer_order": 264 + }, + "Mul_412": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 8, + "ifmsv_height": 12, + "ifmsv_size": 960, + "ifmsv_width": 10, + "num_elems": 3600, + "num_kernel_iters": 8, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 20, + 45, + 80 + ], + "dims": [ + 32, + 48, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 20, + 45, + 80 + ], + "dims": [ + 24, + 45, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 20, + 45, + 80 + ], + "dims": [ + 32, + 48, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Sub_411", + "type": "internal" + }, + { + "name": "Input.2", + "type": "external" + } + ], + "layer_order": 265 + }, + "Split_405_Duplicated#0": { + "SliceHCWC8Adf": { + "aie_arch": "aie2p", + "axis_letter": "C", + "dim_c": 40, + "dim_h": 45, + "dim_w": 80, + "dtype": "bfloat16", + "end": 20, + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "num_ifm_shim_ch": 2, + "num_ofm_shim_ch": 2, + "output_l3_format": "HCWN_C8", + "start": 0, + "step": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 40, + 45, + 80 + ], + "dims": [ + 40, + 45, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 20, + 45, + 80 + ], + "dims": [ + 24, + 45, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Sigmoid_404", + "type": "internal" + } + ], + "layer_order": 266 + }, + "Mul_406": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 8, + "ifmsv_height": 12, + "ifmsv_size": 960, + "ifmsv_width": 10, + "num_elems": 3600, + "num_kernel_iters": 8, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 20, + 45, + 80 + ], + "dims": [ + 24, + 45, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 20, + 45, + 80 + ], + "dims": [ + 24, + 45, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 20, + 45, + 80 + ], + "dims": [ + 32, + 48, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Split_405_Duplicated#0", + "type": "internal" + }, + { + "name": "Input.2", + "type": "external" + } + ], + "layer_order": 267 + }, + "Concat_407": { + "ConcatC8Adf": { + "aie_arch": "aie2p", + "dtype": "bfloat16", + "in1_dim_c": 24, + "in1_dim_h": 45, + "in1_dim_w": 80, + "in2_dim_c": 24, + "in2_dim_h": 45, + "in2_dim_w": 80, + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "num_eff_concat_input0_size": 20, + "num_eff_concat_input0_start": 0, + "num_eff_concat_input1_size": 20, + "num_eff_concat_input1_start": 0, + "output_l3_format": "HCWN_C8" + }, + "dims_and_bounds": [ + { + "bounds": [ + 20, + 45, + 80 + ], + "dims": [ + 24, + 45, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 20, + 45, + 80 + ], + "dims": [ + 24, + 45, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 40, + 45, + 80 + ], + "dims": [ + 40, + 45, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Split_401_Duplicated#1", + "type": "internal" + }, + { + "name": "Mul_406", + "type": "internal" + } + ], + "layer_order": 268 + }, + "Conv_408": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 64, + "data_io.ofm.dimensions.width": 32, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 1, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 40, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 5, + "ifm_height": 45, + "ifm_sign": 0, + "ifm_sv_depth": 8, + "ifm_sv_height": 14, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 34, + "ifm_width": 80, + "ifm_x_iter": 3, + "ifm_y_iter": 1, + "kernel_height": 3, + "kernel_width": 3, + "ksize.height": 3, + "ksize.width": 3, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 5, + "num_ifm_height_iters": 1, + "num_ifm_width_iters": 3, + "num_ofm_depth_iters": 1, + "ofm_depth": 24, + "ofm_depth_iter": 1, + "ofm_height": 45, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 8, + "ofm_sv_height": 12, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 32, + "ofm_width": 80, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 1, + "pad_left": 1, + "pad_right": 1, + "pad_top": 1, + "pad_value": 0, + "padding_sv_height": 14, + "padding_sv_width": 34, + "psum_buff_offset_scaled": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 40, + 45, + 80 + ], + "dims": [ + 40, + 45, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 20, + 45, + 80 + ], + "dims": [ + 32, + 48, + 96 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Concat_407", + "type": "internal" + } + ], + "layer_order": 269, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 144, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 5, + "ifm_depth_padded": 40, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 3, + "kernel_width": 3, + "krnl_w_div": 1, + "num_bias_sub_groups": 1, + "ofm_depth_iter": 1, + "ofm_depth_padded": 32, + "output_stride": 8, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 20453760, + "wts_no_of_reads": 864 + } + }, + "Tanh_409": { + "TanhTemplatedBf16": { + "ENABLE_FP16_AS_BF16": 0, + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 8, + "ifmsv_height": 12, + "ifmsv_size": 1920, + "ifmsv_width": 20, + "num_kernel_iters": 4, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 20, + 45, + 80 + ], + "dims": [ + 32, + 48, + 96 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 20, + 45, + 80 + ], + "dims": [ + 32, + 48, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_408", + "type": "internal" + } + ], + "layer_order": 270 + }, + "Mul_413": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 8, + "ifmsv_height": 12, + "ifmsv_size": 960, + "ifmsv_width": 10, + "num_elems": 3600, + "num_kernel_iters": 8, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 20, + 45, + 80 + ], + "dims": [ + 24, + 45, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 20, + 45, + 80 + ], + "dims": [ + 32, + 48, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 20, + 45, + 80 + ], + "dims": [ + 32, + 48, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Split_405_Duplicated#1", + "type": "internal" + }, + { + "name": "Tanh_409", + "type": "internal" + } + ], + "layer_order": 271 + }, + "Add_414": { + "AddBf16": { + "act": 0, + "act_type": "LINEAR", + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 8, + "ifmsv_height": 12, + "ifmsv_size": 960, + "ifmsv_width": 10, + "num_elems": 3600, + "num_kernel_iters": 8, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 20, + 45, + 80 + ], + "dims": [ + 32, + 48, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 20, + 45, + 80 + ], + "dims": [ + 32, + 48, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 20, + 45, + 80 + ], + "dims": [ + 24, + 45, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "external_output_layer": 1, + "input_layer_names": [ + { + "name": "Mul_412", + "type": "internal" + }, + { + "name": "Mul_413", + "type": "internal" + } + ], + "layer_order": 272 + }, + "Concat_415": { + "ConcatC8Adf": { + "aie_arch": "aie2p", + "dtype": "bfloat16", + "in1_dim_c": 24, + "in1_dim_h": 45, + "in1_dim_w": 80, + "in2_dim_c": 24, + "in2_dim_h": 45, + "in2_dim_w": 80, + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "num_eff_concat_input0_size": 20, + "num_eff_concat_input0_start": 0, + "num_eff_concat_input1_size": 20, + "num_eff_concat_input1_start": 0, + "output_l3_format": "HCWN_C8" + }, + "dims_and_bounds": [ + { + "bounds": [ + 20, + 45, + 80 + ], + "dims": [ + 24, + 45, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 20, + 45, + 80 + ], + "dims": [ + 24, + 45, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 40, + 45, + 80 + ], + "dims": [ + 40, + 45, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Split_401_Duplicated#0", + "type": "internal" + }, + { + "name": "Add_414", + "type": "internal" + } + ], + "layer_order": 273 + }, + "Resize_417": { + "ResizeAdf": { + "co_trans_mode": 1, + "dim_0": 1, + "dim_1": 40, + "dim_2": 45, + "dim_3": 80, + "dtype": "bfloat16", + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "mode": 1, + "nearest_mode": 0, + "num_ifm_shim_ch": 2, + "num_ofm_shim_ch": 2, + "output_H": 90, + "output_W": 160, + "output_l3_format": "HCWN_C8" + }, + "dims_and_bounds": [ + { + "bounds": [ + 40, + 45, + 80 + ], + "dims": [ + 40, + 45, + 80 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 40, + 90, + 160 + ], + "dims": [ + 40, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Concat_415", + "type": "internal" + } + ], + "layer_order": 274 + }, + "Concat_418": { + "Concat": { + "axis": 1, + "dims_and_bounds": [ + { + "bounds": [ + 40, + 90, + 160 + ], + "dims": [ + 40, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 3, + 90, + 160 + ], + "dims": [ + 8, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 59, + 90, + 160 + ], + "dims": [ + 64, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "dtype": "bfloat16" + }, + "input_layer_names": [ + { + "name": "Resize_417", + "type": "internal" + }, + { + "name": "Conv_28", + "type": "internal" + }, + { + "name": "AveragePool_346", + "type": "internal" + } + ], + "layer_order": 275 + }, + "Conv_419": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 1, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 64, + "data_io.ofm.dimensions.width": 32, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 1, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 64, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 4, + "ifm_height": 90, + "ifm_sign": 0, + "ifm_sv_depth": 16, + "ifm_sv_height": 7, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 34, + "ifm_width": 160, + "ifm_x_iter": 5, + "ifm_y_iter": 5, + "kernel_height": 3, + "kernel_width": 3, + "ksize.height": 3, + "ksize.width": 3, + "lrelu_alpha": 0, + "lrelu_alpha_kernel": 0, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 4, + "num_ifm_height_iters": 5, + "num_ifm_width_iters": 5, + "num_ofm_depth_iters": 1, + "ofm_depth": 32, + "ofm_depth_iter": 1, + "ofm_height": 90, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 8, + "ofm_sv_height": 5, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 32, + "ofm_width": 160, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 1, + "pad_left": 1, + "pad_right": 1, + "pad_top": 1, + "pad_value": 0, + "padding_sv_height": 7, + "padding_sv_width": 34, + "psum_buff_offset_scaled": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 59, + 90, + 160 + ], + "dims": [ + 64, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 32, + 90, + 160 + ], + "dims": [ + 32, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Concat_418", + "type": "internal" + } + ], + "layer_order": 276, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 144, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 4, + "ifm_depth_padded": 64, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 3, + "kernel_width": 3, + "krnl_w_div": 1, + "num_bias_sub_groups": 1, + "ofm_depth_iter": 1, + "ofm_depth_padded": 32, + "output_stride": 8, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 20508480, + "wts_no_of_reads": 1728 + } + }, + "Split_421_Duplicated#0": { + "SliceHCWC8Adf": { + "aie_arch": "aie2p", + "axis_letter": "C", + "dim_c": 32, + "dim_h": 90, + "dim_w": 160, + "dtype": "bfloat16", + "end": 16, + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "num_ifm_shim_ch": 2, + "num_ofm_shim_ch": 2, + "output_l3_format": "HCWN_C8", + "start": 0, + "step": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 32, + 90, + 160 + ], + "dims": [ + 32, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_419", + "type": "internal" + } + ], + "layer_order": 277 + }, + "Split_421_Duplicated#1": { + "SliceHCWC8Adf": { + "aie_arch": "aie2p", + "axis_letter": "C", + "dim_c": 32, + "dim_h": 90, + "dim_w": 160, + "dtype": "bfloat16", + "end": 32, + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "num_ifm_shim_ch": 2, + "num_ofm_shim_ch": 2, + "output_l3_format": "HCWN_C8", + "start": 16, + "step": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 32, + 90, + 160 + ], + "dims": [ + 32, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_419", + "type": "internal" + } + ], + "layer_order": 278 + }, + "Concat_422": { + "Concat": { + "axis": 1, + "dims_and_bounds": [ + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 32, + 90, + 160 + ], + "dims": [ + 32, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "dtype": "bfloat16" + }, + "input_layer_names": [ + { + "name": "Split_421_Duplicated#1", + "type": "internal" + }, + { + "name": "Input.1", + "type": "external" + } + ], + "layer_order": 279 + }, + "Conv_423": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 64, + "data_io.ofm.dimensions.width": 32, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 1, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 32, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 2, + "ifm_height": 90, + "ifm_sign": 0, + "ifm_sv_depth": 16, + "ifm_sv_height": 6, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 34, + "ifm_width": 160, + "ifm_x_iter": 5, + "ifm_y_iter": 6, + "kernel_height": 3, + "kernel_width": 3, + "ksize.height": 3, + "ksize.width": 3, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 2, + "num_ifm_height_iters": 6, + "num_ifm_width_iters": 5, + "num_ofm_depth_iters": 1, + "ofm_depth": 32, + "ofm_depth_iter": 1, + "ofm_height": 90, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 8, + "ofm_sv_height": 4, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 32, + "ofm_width": 160, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 1, + "pad_left": 1, + "pad_right": 1, + "pad_top": 1, + "pad_value": 0, + "padding_sv_height": 6, + "padding_sv_width": 34, + "psum_buff_offset_scaled": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 32, + 90, + 160 + ], + "dims": [ + 32, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 32, + 90, + 160 + ], + "dims": [ + 32, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Concat_422", + "type": "internal" + } + ], + "layer_order": 280, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 144, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 2, + "ifm_depth_padded": 32, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 3, + "kernel_width": 3, + "krnl_w_div": 1, + "num_bias_sub_groups": 1, + "ofm_depth_iter": 1, + "ofm_depth_padded": 32, + "output_stride": 8, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 20593728, + "wts_no_of_reads": 1728 + } + }, + "Sigmoid_424": { + "SigmoidTemplatedBf16": { + "ENABLE_FP16_AS_BF16": 0, + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 8, + "ifmsv_height": 8, + "ifmsv_size": 2048, + "ifmsv_width": 32, + "num_kernel_iters": 15, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 32, + 90, + 160 + ], + "dims": [ + 32, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 32, + 90, + 160 + ], + "dims": [ + 32, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_423", + "type": "internal" + } + ], + "layer_order": 281 + }, + "Split_425_Duplicated#1": { + "SliceHCWC8Adf": { + "aie_arch": "aie2p", + "axis_letter": "C", + "dim_c": 32, + "dim_h": 90, + "dim_w": 160, + "dtype": "bfloat16", + "end": 32, + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "num_ifm_shim_ch": 2, + "num_ofm_shim_ch": 2, + "output_l3_format": "HCWN_C8", + "start": 16, + "step": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 32, + 90, + 160 + ], + "dims": [ + 32, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Sigmoid_424", + "type": "internal" + } + ], + "layer_order": 282 + }, + "Sub_431": { + "SubBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 8, + "ifmsv_height": 8, + "ifmsv_size": 1024, + "ifmsv_width": 16, + "num_elems": 14400, + "num_kernel_iters": 30, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "constants_reader": { + "Const.5": { + "bytes_to_skip": 20764800, + "channel_stride": 8, + "depth": 1, + "dtype": "bfloat16", + "height": 90, + "nibbles": 4, + "num_reads": 21600, + "rank": 4, + "reads_per_line": 3, + "width": 160 + } + }, + "dims_and_bounds": [ + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Const.5", + "type": "constant" + }, + { + "name": "Split_425_Duplicated#1", + "type": "internal" + } + ], + "layer_order": 283 + }, + "Mul_432": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 8, + "ifmsv_height": 8, + "ifmsv_size": 1024, + "ifmsv_width": 16, + "num_elems": 14400, + "num_kernel_iters": 30, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Sub_431", + "type": "internal" + }, + { + "name": "Input.1", + "type": "external" + } + ], + "layer_order": 284 + }, + "Split_425_Duplicated#0": { + "SliceHCWC8Adf": { + "aie_arch": "aie2p", + "axis_letter": "C", + "dim_c": 32, + "dim_h": 90, + "dim_w": 160, + "dtype": "bfloat16", + "end": 16, + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "num_ifm_shim_ch": 2, + "num_ofm_shim_ch": 2, + "output_l3_format": "HCWN_C8", + "start": 0, + "step": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 32, + 90, + 160 + ], + "dims": [ + 32, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Sigmoid_424", + "type": "internal" + } + ], + "layer_order": 285 + }, + "Mul_426": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 8, + "ifmsv_height": 8, + "ifmsv_size": 1024, + "ifmsv_width": 16, + "num_elems": 14400, + "num_kernel_iters": 30, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Split_425_Duplicated#0", + "type": "internal" + }, + { + "name": "Input.1", + "type": "external" + } + ], + "layer_order": 286 + }, + "Concat_427": { + "Concat": { + "axis": 1, + "dims_and_bounds": [ + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 32, + 90, + 160 + ], + "dims": [ + 32, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "dtype": "bfloat16" + }, + "input_layer_names": [ + { + "name": "Split_421_Duplicated#1", + "type": "internal" + }, + { + "name": "Mul_426", + "type": "internal" + } + ], + "layer_order": 287 + }, + "Conv_428": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 64, + "data_io.ofm.dimensions.width": 32, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 1, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 32, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 2, + "ifm_height": 90, + "ifm_sign": 0, + "ifm_sv_depth": 16, + "ifm_sv_height": 6, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 34, + "ifm_width": 160, + "ifm_x_iter": 5, + "ifm_y_iter": 6, + "kernel_height": 3, + "kernel_width": 3, + "ksize.height": 3, + "ksize.width": 3, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 2, + "num_ifm_height_iters": 6, + "num_ifm_width_iters": 5, + "num_ofm_depth_iters": 1, + "ofm_depth": 16, + "ofm_depth_iter": 1, + "ofm_height": 90, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 8, + "ofm_sv_height": 4, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 32, + "ofm_width": 160, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 1, + "pad_left": 1, + "pad_right": 1, + "pad_top": 1, + "pad_value": 0, + "padding_sv_height": 6, + "padding_sv_width": 34, + "psum_buff_offset_scaled": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 32, + 90, + 160 + ], + "dims": [ + 32, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Concat_427", + "type": "internal" + } + ], + "layer_order": 288, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 144, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 2, + "ifm_depth_padded": 32, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 3, + "kernel_width": 3, + "krnl_w_div": 1, + "num_bias_sub_groups": 1, + "ofm_depth_iter": 1, + "ofm_depth_padded": 32, + "output_stride": 8, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 20636352, + "wts_no_of_reads": 1728 + } + }, + "Tanh_429": { + "TanhTemplatedBf16": { + "ENABLE_FP16_AS_BF16": 0, + "aie_arch": "aie2p", + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 8, + "ifmsv_height": 23, + "ifmsv_size": 1472, + "ifmsv_width": 8, + "num_kernel_iters": 20, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 32, + 92, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_428", + "type": "internal" + } + ], + "layer_order": 289 + }, + "Mul_433": { + "MulBf16": { + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 8, + "ifmsv_height": 8, + "ifmsv_size": 1024, + "ifmsv_width": 16, + "num_elems": 14400, + "num_kernel_iters": 30, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Split_425_Duplicated#1", + "type": "internal" + }, + { + "name": "Tanh_429", + "type": "internal" + } + ], + "layer_order": 290 + }, + "Add_434": { + "AddBf16": { + "act": 0, + "act_type": "LINEAR", + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 8, + "ifmsv_height": 8, + "ifmsv_size": 1024, + "ifmsv_width": 16, + "num_elems": 14400, + "num_kernel_iters": 30, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "external_output_layer": 0, + "input_layer_names": [ + { + "name": "Mul_432", + "type": "internal" + }, + { + "name": "Mul_433", + "type": "internal" + } + ], + "layer_order": 291 + }, + "Concat_435": { + "Concat": { + "axis": 1, + "dims_and_bounds": [ + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 16, + 90, + 160 + ], + "dims": [ + 16, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 32, + 90, + 160 + ], + "dims": [ + 32, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "dtype": "bfloat16" + }, + "input_layer_names": [ + { + "name": "Split_421_Duplicated#0", + "type": "internal" + }, + { + "name": "Add_434", + "type": "internal" + } + ], + "layer_order": 292 + }, + "Resize_437": { + "ResizeAdf": { + "co_trans_mode": 1, + "dim_0": 1, + "dim_1": 32, + "dim_2": 90, + "dim_3": 160, + "dtype": "bfloat16", + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "mode": 1, + "nearest_mode": 0, + "num_ifm_shim_ch": 2, + "num_ofm_shim_ch": 2, + "output_H": 180, + "output_W": 320, + "output_l3_format": "HCWN_C8" + }, + "dims_and_bounds": [ + { + "bounds": [ + 32, + 90, + 160 + ], + "dims": [ + 32, + 90, + 160 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 32, + 180, + 320 + ], + "dims": [ + 32, + 180, + 320 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Concat_435", + "type": "internal" + } + ], + "layer_order": 293 + }, + "Concat_438": { + "Concat": { + "axis": 1, + "dims_and_bounds": [ + { + "bounds": [ + 32, + 180, + 320 + ], + "dims": [ + 32, + 180, + 320 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 3, + 180, + 320 + ], + "dims": [ + 8, + 180, + 320 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 35, + 180, + 320 + ], + "dims": [ + 40, + 180, + 320 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "dtype": "bfloat16" + }, + "input_layer_names": [ + { + "name": "Resize_437", + "type": "internal" + }, + { + "name": "Generated-#4", + "type": "internal" + } + ], + "layer_order": 294 + }, + "Conv_439": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 1, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 64, + "data_io.ofm.dimensions.width": 64, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 1, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 40, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 5, + "ifm_height": 180, + "ifm_sign": 0, + "ifm_sv_depth": 8, + "ifm_sv_height": 6, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 66, + "ifm_width": 320, + "ifm_x_iter": 5, + "ifm_y_iter": 12, + "kernel_height": 3, + "kernel_width": 3, + "ksize.height": 3, + "ksize.width": 3, + "lrelu_alpha": 0, + "lrelu_alpha_kernel": 0, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 5, + "num_ifm_height_iters": 12, + "num_ifm_width_iters": 5, + "num_ofm_depth_iters": 1, + "ofm_depth": 16, + "ofm_depth_iter": 1, + "ofm_height": 180, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 8, + "ofm_sv_height": 4, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 64, + "ofm_width": 320, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 1, + "pad_left": 1, + "pad_right": 1, + "pad_top": 1, + "pad_value": 0, + "padding_sv_height": 6, + "padding_sv_width": 66, + "psum_buff_offset_scaled": 4, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 35, + 180, + 320 + ], + "dims": [ + 40, + 180, + 320 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 16, + 180, + 320 + ], + "dims": [ + 16, + 180, + 320 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Concat_438", + "type": "internal" + } + ], + "layer_order": 295, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 144, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 5, + "ifm_depth_padded": 40, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 3, + "kernel_width": 3, + "krnl_w_div": 1, + "num_bias_sub_groups": 1, + "ofm_depth_iter": 1, + "ofm_depth_padded": 32, + "output_stride": 8, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 20678976, + "wts_no_of_reads": 864 + } + }, + "Conv_441": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 1, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 64, + "data_io.ofm.dimensions.width": 32, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 1, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 16, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 1, + "ifm_height": 180, + "ifm_sign": 0, + "ifm_sv_depth": 16, + "ifm_sv_height": 7, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 34, + "ifm_width": 320, + "ifm_x_iter": 10, + "ifm_y_iter": 9, + "kernel_height": 3, + "kernel_width": 3, + "ksize.height": 3, + "ksize.width": 3, + "lrelu_alpha": 0, + "lrelu_alpha_kernel": 0, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 1, + "num_ifm_height_iters": 9, + "num_ifm_width_iters": 10, + "num_ofm_depth_iters": 1, + "ofm_depth": 16, + "ofm_depth_iter": 1, + "ofm_height": 180, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 8, + "ofm_sv_height": 5, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 32, + "ofm_width": 320, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 1, + "pad_left": 1, + "pad_right": 1, + "pad_top": 1, + "pad_value": 0, + "padding_sv_height": 7, + "padding_sv_width": 34, + "psum_buff_offset_scaled": 2, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 16, + 180, + 320 + ], + "dims": [ + 16, + 180, + 320 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 16, + 180, + 320 + ], + "dims": [ + 16, + 180, + 320 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_439", + "type": "internal" + } + ], + "layer_order": 296, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 144, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 16, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 3, + "kernel_width": 3, + "krnl_w_div": 1, + "num_bias_sub_groups": 1, + "ofm_depth_iter": 1, + "ofm_depth_padded": 32, + "output_stride": 8, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 20733696, + "wts_no_of_reads": 1728 + } + }, + "Conv_443": { + "Conv2DBf16": { + "AIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16": 1, + "act": 0, + "act_type": "RELU", + "aie_arch": "aie2p", + "alpha": 0, + "alpha_scaled": 3277, + "batch_size": 1, + "compiler": "chess", + "conv2d_ofm_pad": 0, + "conv_type": 64, + "data_io.ofm.dimensions.width": 64, + "dtype_ifm": "bfloat16", + "dtype_ofm": "bfloat16", + "dtype_wts": "bfloat16", + "enable_padding": 0, + "force_och_gran_8": 1, + "func_type": 0, + "ifm_depth": 16, + "ifm_depth_concat_extend": 0, + "ifm_depth_iter": 1, + "ifm_height": 180, + "ifm_sign": 0, + "ifm_sv_depth": 64, + "ifm_sv_height": 1, + "ifm_sv_height.padding": 0, + "ifm_sv_width": 64, + "ifm_width": 320, + "ifm_x_iter": 5, + "ifm_y_iter": 45, + "kernel_height": 1, + "kernel_width": 1, + "ksize.height": 1, + "ksize.width": 1, + "lrelu_alpha": 1, + "lrelu_alpha_kernel": 1, + "num_adf_cols": 4, + "num_adf_rows": 4, + "num_ifm_depth_iters": 1, + "num_ifm_height_iters": 45, + "num_ifm_width_iters": 5, + "num_ofm_depth_iters": 1, + "ofm_depth": 8, + "ofm_depth_iter": 1, + "ofm_height": 180, + "ofm_offset_packed.left": 0, + "ofm_offset_packed.top": 0, + "ofm_sv_depth": 8, + "ofm_sv_height": 1, + "ofm_sv_overlap_height": 0, + "ofm_sv_overlap_width": 0, + "ofm_sv_width": 64, + "ofm_width": 320, + "ofm_width_pad": 0, + "op_type": 0, + "out_mode": 0, + "pad_bottom": 0, + "pad_left": 0, + "pad_right": 0, + "pad_top": 0, + "pad_value": 0, + "padding_sv_height": 1, + "padding_sv_width": 64, + "psum_buff_offset_scaled": 4, + "relu_int8_enable": 0, + "shift_bias_init": 0, + "shift_lrelu_alpha": 0, + "shift_lrelu_in": 0, + "shift_lrelu_out": 0, + "shift_out": 0, + "shift_psum": 0, + "shift_psum_in": 0, + "signed_value": 0, + "stride_h": 1, + "stride_height": 1, + "stride_log2.stride_log2_height": 0, + "stride_log2.stride_log2_width": 0, + "stride_w": 1, + "stride_width": 1, + "zp_en": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 16, + 180, + 320 + ], + "dims": [ + 16, + 180, + 320 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 4, + 180, + 320 + ], + "dims": [ + 8, + 180, + 320 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_441", + "type": "internal" + } + ], + "layer_order": 297, + "split_width": false, + "wts_reader": { + "BFP16MMUL": true, + "bias_bytes_to_skip": 144, + "bias_rep": 4, + "bias_sg_bytes_to_skip": 36, + "bias_sub_group_num_reads": 12, + "cstride": 8, + "ifm_depth_iter": 1, + "ifm_depth_padded": 64, + "is_dwc": false, + "is_zero_padded": false, + "istride": 8, + "kernel_height": 1, + "kernel_width": 1, + "krnl_w_div": 1, + "num_bias_sub_groups": 1, + "ofm_depth_iter": 1, + "ofm_depth_padded": 32, + "output_stride": 8, + "rtp_bytes_to_skip": 144, + "total_layer_bytes_to_skip": 20755008, + "wts_no_of_reads": 768 + } + }, + "Split_444_Duplicated#1": { + "SliceHCWC8Adf": { + "aie_arch": "aie2p", + "axis_letter": "C", + "dim_c": 8, + "dim_h": 180, + "dim_w": 320, + "dtype": "bfloat16", + "end": 4, + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "num_ifm_shim_ch": 2, + "num_ofm_shim_ch": 2, + "output_l3_format": "HCWN_C8", + "start": 3, + "step": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 4, + 180, + 320 + ], + "dims": [ + 8, + 180, + 320 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 1, + 180, + 320 + ], + "dims": [ + 8, + 180, + 320 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_443", + "type": "internal" + } + ], + "layer_order": 298 + }, + "Clip_447": { + "ClipBf16": { + "aie_arch": "aie2p", + "clamp_max": 1, + "clamp_min": 0, + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 8, + "ifmsv_height": 5, + "ifmsv_size": 1600, + "ifmsv_width": 40, + "num_kernel_iters": 72, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 1, + 180, + 320 + ], + "dims": [ + 8, + 180, + 320 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 1, + 180, + 320 + ], + "dims": [ + 8, + 180, + 320 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "external_output_layer": 5, + "input_layer_names": [ + { + "name": "Split_444_Duplicated#1", + "type": "internal" + } + ], + "layer_order": 299 + }, + "Split_444_Duplicated#0": { + "SliceHCWC8Adf": { + "aie_arch": "aie2p", + "axis_letter": "C", + "dim_c": 8, + "dim_h": 180, + "dim_w": 320, + "dtype": "bfloat16", + "end": 3, + "input_l3_format": "HCWN_C8", + "kernel_col": 4, + "num_ifm_shim_ch": 2, + "num_ofm_shim_ch": 2, + "output_l3_format": "HCWN_C8", + "start": 0, + "step": 1 + }, + "dims_and_bounds": [ + { + "bounds": [ + 4, + 180, + 320 + ], + "dims": [ + 8, + 180, + 320 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 3, + 180, + 320 + ], + "dims": [ + 8, + 180, + 320 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Conv_443", + "type": "internal" + } + ], + "layer_order": 300 + }, + "Add_445": { + "AddBf16": { + "act": 0, + "act_type": "LINEAR", + "aie_arch": "aie2p", + "compiler": "chess", + "dtype": "bfloat16", + "ifm1_shift": 0, + "ifm1_shift_biased": 0, + "ifm2_shift": 0, + "ifm2_shift_biased": 0, + "ifmsv_depth": 8, + "ifmsv_height": 5, + "ifmsv_size": 640, + "ifmsv_width": 16, + "num_elems": 57600, + "num_kernel_iters": 180, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 3, + 180, + 320 + ], + "dims": [ + 8, + 180, + 320 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 3, + 180, + 320 + ], + "dims": [ + 8, + 180, + 320 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 3, + 180, + 320 + ], + "dims": [ + 8, + 180, + 320 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "input_layer_names": [ + { + "name": "Split_444_Duplicated#0", + "type": "internal" + }, + { + "name": "Generated-#4", + "type": "internal" + } + ], + "layer_order": 301 + }, + "Clip_446": { + "ClipBf16": { + "aie_arch": "aie2p", + "clamp_max": 1, + "clamp_min": 0, + "compiler": "chess", + "ifm_shift": 0, + "ifm_shift_biased": 0, + "ifmsv_depth": 8, + "ifmsv_height": 5, + "ifmsv_size": 1600, + "ifmsv_width": 40, + "num_kernel_iters": 72, + "ofm_shift": 0, + "ofm_shift_biased": 0 + }, + "dims_and_bounds": [ + { + "bounds": [ + 3, + 180, + 320 + ], + "dims": [ + 8, + 180, + 320 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + }, + { + "bounds": [ + 3, + 180, + 320 + ], + "dims": [ + 8, + 180, + 320 + ], + "dtype": "bfloat16", + "keys": [ + "depth", + "height", + "width" + ], + "padding_at_start": { + "cin": 0, + "cout": 0, + "height": 0, + "width": 0 + } + } + ], + "external_output_layer": 4, + "input_layer_names": [ + { + "name": "Add_445", + "type": "internal" + } + ], + "layer_order": 302 + } +}