// Copyright 2020 Google LLC // // This source code is licensed under the BSD-style license found in the // LICENSE file in the root directory of this source tree. #include #include #include #include #include "xnnpack.h" #include "xnnpack/common.h" #include "xnnpack/config-types.h" #include "xnnpack/config.h" #include "xnnpack/internal.h" #include "xnnpack/log.h" #include "xnnpack/node-type.h" #include "xnnpack/operator-type.h" #include "xnnpack/operator.h" #include "xnnpack/requantization.h" #include "xnnpack/subgraph-validation.h" #include "xnnpack/subgraph.h" #include "pthreadpool.h" static enum xnn_status create_convolution_operator( const struct xnn_node* node, const struct xnn_value* values, size_t num_values, struct xnn_operator_data* opdata, struct xnn_code_cache* code_cache, xnn_weights_cache_t weights_cache) { assert(node->num_inputs >= 2); assert(node->num_inputs <= 3); const uint32_t input_id = node->inputs[0]; assert(input_id != XNN_INVALID_VALUE_ID); assert(input_id < num_values); const uint32_t filter_id = node->inputs[1]; assert(filter_id != XNN_INVALID_VALUE_ID); assert(filter_id < num_values); assert(node->num_outputs == 1); const uint32_t output_id = node->outputs[0]; assert(output_id != XNN_INVALID_VALUE_ID); assert(output_id < num_values); const void* filter_data = values[filter_id].fp32_data != NULL ? values[filter_id].fp32_data : values[filter_id].data; assert(filter_data != NULL); const void* bias_data = NULL; uint32_t bias_id = XNN_INVALID_VALUE_ID; if (node->num_inputs > 2) { bias_id = node->inputs[2]; assert(bias_id != XNN_INVALID_VALUE_ID); assert(bias_id < num_values); bias_data = values[bias_id].fp32_data != NULL ? values[bias_id].fp32_data : values[bias_id].data; assert(bias_data != NULL); } enum xnn_status status; const enum xnn_datatype input_datatype = values[input_id].datatype; const enum xnn_datatype filter_datatype = values[filter_id].datatype; const enum xnn_datatype bias_datatype = bias_id != XNN_INVALID_VALUE_ID ? values[filter_id].datatype : xnn_datatype_invalid; const enum xnn_datatype output_datatype = values[output_id].datatype; if (values[output_id].layout == xnn_layout_type_nchw) { switch (output_datatype) { case xnn_datatype_fp16: status = xnn_create_convolution2d_nchw_f16( node->params.convolution_2d.input_padding_top, node->params.convolution_2d.input_padding_right, node->params.convolution_2d.input_padding_bottom, node->params.convolution_2d.input_padding_left, node->params.convolution_2d.kernel_height, node->params.convolution_2d.kernel_width, node->params.convolution_2d.subsampling_height, node->params.convolution_2d.subsampling_width, node->params.convolution_2d.dilation_height, node->params.convolution_2d.dilation_width, node->params.convolution_2d.groups, node->params.convolution_2d.group_input_channels, node->params.convolution_2d.group_output_channels, node->params.convolution_2d.group_input_channels * node->params.convolution_2d.groups /* input_pixel_stride */, node->params.convolution_2d.group_output_channels * node->params.convolution_2d.groups /* output_pixel_stride */, filter_data, bias_data, node->activation.output_min, node->activation.output_max, node->flags | (values[input_id].layout == xnn_layout_type_nhwc ? XNN_FLAG_INPUT_NHWC : 0) | XNN_FLAG_FP32_STATIC_WEIGHTS, code_cache, weights_cache, &opdata->operator_objects[0]); break; case xnn_datatype_fp32: status = xnn_create_convolution2d_nchw_f32( node->params.convolution_2d.input_padding_top, node->params.convolution_2d.input_padding_right, node->params.convolution_2d.input_padding_bottom, node->params.convolution_2d.input_padding_left, node->params.convolution_2d.kernel_height, node->params.convolution_2d.kernel_width, node->params.convolution_2d.subsampling_height, node->params.convolution_2d.subsampling_width, node->params.convolution_2d.dilation_height, node->params.convolution_2d.dilation_width, node->params.convolution_2d.groups, node->params.convolution_2d.group_input_channels, node->params.convolution_2d.group_output_channels, node->params.convolution_2d.group_input_channels * node->params.convolution_2d.groups /* input_pixel_stride */, node->params.convolution_2d.group_output_channels * node->params.convolution_2d.groups /* output_pixel_stride */, filter_data, bias_data, node->activation.output_min, node->activation.output_max, node->flags | (values[input_id].layout == xnn_layout_type_nhwc ? XNN_FLAG_INPUT_NHWC : 0), code_cache, weights_cache, &opdata->operator_objects[0]); break; default: XNN_UNREACHABLE; } } else { assert(values[input_id].layout == xnn_layout_type_nhwc); assert(values[output_id].layout == xnn_layout_type_nhwc); switch (output_datatype) { case xnn_datatype_fp32: switch (filter_datatype) { case xnn_datatype_fp16: { uint32_t flags = node->flags; if (bias_datatype == xnn_datatype_fp32) { flags |= XNN_FLAG_FP32_STATIC_BIASES; } switch (input_datatype) { case xnn_datatype_fp32: status = xnn_create_convolution2d_nhwc_f32_f16( node->params.convolution_2d.input_padding_top, node->params.convolution_2d.input_padding_right, node->params.convolution_2d.input_padding_bottom, node->params.convolution_2d.input_padding_left, node->params.convolution_2d.kernel_height, node->params.convolution_2d.kernel_width, node->params.convolution_2d.subsampling_height, node->params.convolution_2d.subsampling_width, node->params.convolution_2d.dilation_height, node->params.convolution_2d.dilation_width, node->params.convolution_2d.groups, node->params.convolution_2d.group_input_channels, node->params.convolution_2d.group_output_channels, node->params.convolution_2d.group_input_channels * node->params.convolution_2d .groups /* input_pixel_stride */, node->params.convolution_2d.group_output_channels * node->params.convolution_2d .groups /* output_pixel_stride */, filter_data, bias_data, node->activation.output_min, node->activation.output_max, flags, code_cache, weights_cache, &opdata->operator_objects[0]); break; default: XNN_UNREACHABLE; } break; } case xnn_datatype_fp32: switch (input_datatype) { case xnn_datatype_fp32: status = xnn_create_convolution2d_nhwc_f32( node->params.convolution_2d.input_padding_top, node->params.convolution_2d.input_padding_right, node->params.convolution_2d.input_padding_bottom, node->params.convolution_2d.input_padding_left, node->params.convolution_2d.kernel_height, node->params.convolution_2d.kernel_width, node->params.convolution_2d.subsampling_height, node->params.convolution_2d.subsampling_width, node->params.convolution_2d.dilation_height, node->params.convolution_2d.dilation_width, node->params.convolution_2d.groups, node->params.convolution_2d.group_input_channels, node->params.convolution_2d.group_output_channels, node->params.convolution_2d.group_input_channels * node->params.convolution_2d.groups /* input_pixel_stride */, node->params.convolution_2d.group_output_channels * node->params.convolution_2d.groups /* output_pixel_stride */, filter_data, bias_data, node->activation.output_min, node->activation.output_max, node->flags, code_cache, weights_cache, &opdata->operator_objects[0]); break; case xnn_datatype_pfp32: status = xnn_create_convolution2d_nhwc_pf32( node->params.convolution_2d.input_padding_top, node->params.convolution_2d.input_padding_right, node->params.convolution_2d.input_padding_bottom, node->params.convolution_2d.input_padding_left, node->params.convolution_2d.kernel_height, node->params.convolution_2d.kernel_width, node->params.convolution_2d.subsampling_height, node->params.convolution_2d.subsampling_width, node->params.convolution_2d.dilation_height, node->params.convolution_2d.dilation_width, node->params.convolution_2d.groups, node->params.convolution_2d.group_input_channels, node->params.convolution_2d.group_output_channels, node->params.convolution_2d.group_input_channels * node->params.convolution_2d.groups /* input_pixel_stride */, node->params.convolution_2d.group_output_channels * node->params.convolution_2d.groups /* output_pixel_stride */, filter_data, bias_data, node->activation.output_min, node->activation.output_max, node->flags, code_cache, weights_cache, &opdata->operator_objects[0]); break; default: XNN_UNREACHABLE; } break; case xnn_datatype_qcint8: switch (input_datatype) { case xnn_datatype_qdint8: status = xnn_create_convolution2d_nhwc_qd8_f32_qc8w( node->params.convolution_2d.input_padding_top, node->params.convolution_2d.input_padding_right, node->params.convolution_2d.input_padding_bottom, node->params.convolution_2d.input_padding_left, node->params.convolution_2d.kernel_height, node->params.convolution_2d.kernel_width, node->params.convolution_2d.subsampling_height, node->params.convolution_2d.subsampling_width, node->params.convolution_2d.dilation_height, node->params.convolution_2d.dilation_width, node->params.convolution_2d.groups, node->params.convolution_2d.group_input_channels, node->params.convolution_2d.group_output_channels, /*input_channel_stride=*/node->params.convolution_2d.group_input_channels * node->params.convolution_2d.groups, /*output_channel_stride=*/node->params.convolution_2d.group_output_channels * node->params.convolution_2d.groups, values[filter_id].quantization.channelwise_scale, filter_data, bias_data, node->activation.output_min, node->activation.output_max, node->flags, code_cache, weights_cache, &opdata->operator_objects[0]); break; case xnn_datatype_qduint8: status = xnn_create_convolution2d_nhwc_qdu8_f32_qc8w( node->params.convolution_2d.input_padding_top, node->params.convolution_2d.input_padding_right, node->params.convolution_2d.input_padding_bottom, node->params.convolution_2d.input_padding_left, node->params.convolution_2d.kernel_height, node->params.convolution_2d.kernel_width, node->params.convolution_2d.subsampling_height, node->params.convolution_2d.subsampling_width, node->params.convolution_2d.dilation_height, node->params.convolution_2d.dilation_width, node->params.convolution_2d.groups, node->params.convolution_2d.group_input_channels, node->params.convolution_2d.group_output_channels, /*input_channel_stride=*/node->params.convolution_2d.group_input_channels * node->params.convolution_2d.groups, /*output_channel_stride=*/node->params.convolution_2d.group_output_channels * node->params.convolution_2d.groups, values[filter_id].quantization.channelwise_scale, filter_data, bias_data, node->activation.output_min, node->activation.output_max, node->flags, code_cache, weights_cache, &opdata->operator_objects[0]); break; default: XNN_UNREACHABLE; } break; default: XNN_UNREACHABLE; } break; case xnn_datatype_fp16: { uint32_t flags = node->flags; if (filter_datatype == xnn_datatype_fp32) { flags |= XNN_FLAG_FP32_STATIC_WEIGHTS; } if (bias_datatype == xnn_datatype_fp32) { flags |= XNN_FLAG_FP32_STATIC_BIASES; } switch (filter_datatype) { case xnn_datatype_fp16: case xnn_datatype_fp32: status = xnn_create_convolution2d_nhwc_f16( node->params.convolution_2d.input_padding_top, node->params.convolution_2d.input_padding_right, node->params.convolution_2d.input_padding_bottom, node->params.convolution_2d.input_padding_left, node->params.convolution_2d.kernel_height, node->params.convolution_2d.kernel_width, node->params.convolution_2d.subsampling_height, node->params.convolution_2d.subsampling_width, node->params.convolution_2d.dilation_height, node->params.convolution_2d.dilation_width, node->params.convolution_2d.groups, node->params.convolution_2d.group_input_channels, node->params.convolution_2d.group_output_channels, node->params.convolution_2d.group_input_channels * node->params.convolution_2d.groups /* input_pixel_stride */, node->params.convolution_2d.group_output_channels * node->params.convolution_2d.groups /* output_pixel_stride */, filter_data, bias_data, node->activation.output_min, node->activation.output_max, flags, code_cache, weights_cache, &opdata->operator_objects[0]); break; case xnn_datatype_qcint8: switch (input_datatype) { case xnn_datatype_qdint8: status = xnn_create_convolution2d_nhwc_qd8_f16_qc8w( node->params.convolution_2d.input_padding_top, node->params.convolution_2d.input_padding_right, node->params.convolution_2d.input_padding_bottom, node->params.convolution_2d.input_padding_left, node->params.convolution_2d.kernel_height, node->params.convolution_2d.kernel_width, node->params.convolution_2d.subsampling_height, node->params.convolution_2d.subsampling_width, node->params.convolution_2d.dilation_height, node->params.convolution_2d.dilation_width, node->params.convolution_2d.groups, node->params.convolution_2d.group_input_channels, node->params.convolution_2d.group_output_channels, /*input_channel_stride=*/node->params.convolution_2d.group_input_channels * node->params.convolution_2d.groups, /*output_channel_stride=*/node->params.convolution_2d.group_output_channels * node->params.convolution_2d.groups, values[filter_id].quantization.channelwise_scale, filter_data, bias_data, node->activation.output_min, node->activation.output_max, node->flags, code_cache, weights_cache, &opdata->operator_objects[0]); break; case xnn_datatype_qduint8: status = xnn_create_convolution2d_nhwc_qdu8_f16_qc8w( node->params.convolution_2d.input_padding_top, node->params.convolution_2d.input_padding_right, node->params.convolution_2d.input_padding_bottom, node->params.convolution_2d.input_padding_left, node->params.convolution_2d.kernel_height, node->params.convolution_2d.kernel_width, node->params.convolution_2d.subsampling_height, node->params.convolution_2d.subsampling_width, node->params.convolution_2d.dilation_height, node->params.convolution_2d.dilation_width, node->params.convolution_2d.groups, node->params.convolution_2d.group_input_channels, node->params.convolution_2d.group_output_channels, /*input_channel_stride=*/node->params.convolution_2d.group_input_channels * node->params.convolution_2d.groups, /*output_channel_stride=*/node->params.convolution_2d.group_output_channels * node->params.convolution_2d.groups, values[filter_id].quantization.channelwise_scale, filter_data, bias_data, node->activation.output_min, node->activation.output_max, node->flags, code_cache, weights_cache, &opdata->operator_objects[0]); break; default: XNN_UNREACHABLE; } break; default: XNN_UNREACHABLE; } break; } case xnn_datatype_qint8: switch (filter_datatype) { case xnn_datatype_qint8: { const float output_scale = values[output_id].quantization.scale; const int32_t output_zero_point = values[output_id].quantization.zero_point; const int8_t output_min = xnn_qs8_quantize(node->activation.output_min, output_scale, output_zero_point); const int8_t output_max = xnn_qs8_quantize(node->activation.output_max, output_scale, output_zero_point); status = xnn_create_convolution2d_nhwc_qs8( node->params.convolution_2d.input_padding_top, node->params.convolution_2d.input_padding_right, node->params.convolution_2d.input_padding_bottom, node->params.convolution_2d.input_padding_left, node->params.convolution_2d.kernel_height, node->params.convolution_2d.kernel_width, node->params.convolution_2d.subsampling_height, node->params.convolution_2d.subsampling_width, node->params.convolution_2d.dilation_height, node->params.convolution_2d.dilation_width, node->params.convolution_2d.groups, node->params.convolution_2d.group_input_channels, node->params.convolution_2d.group_output_channels, node->params.convolution_2d.group_input_channels * node->params.convolution_2d.groups /* input_pixel_stride */, node->params.convolution_2d.group_output_channels * node->params.convolution_2d.groups /* output_pixel_stride */, (int8_t) values[input_id].quantization.zero_point, values[input_id].quantization.scale, values[filter_id].quantization.scale, filter_data, bias_data, (int8_t) output_zero_point, output_scale, output_min, output_max, node->flags, code_cache, weights_cache, &opdata->operator_objects[0]); break; } case xnn_datatype_qcint8: { const float output_scale = values[output_id].quantization.scale; const int32_t output_zero_point = values[output_id].quantization.zero_point; const int8_t output_min = xnn_qs8_quantize(node->activation.output_min, output_scale, output_zero_point); const int8_t output_max = xnn_qs8_quantize(node->activation.output_max, output_scale, output_zero_point); status = xnn_create_convolution2d_nhwc_qs8_qc8w( node->params.convolution_2d.input_padding_top, node->params.convolution_2d.input_padding_right, node->params.convolution_2d.input_padding_bottom, node->params.convolution_2d.input_padding_left, node->params.convolution_2d.kernel_height, node->params.convolution_2d.kernel_width, node->params.convolution_2d.subsampling_height, node->params.convolution_2d.subsampling_width, node->params.convolution_2d.dilation_height, node->params.convolution_2d.dilation_width, node->params.convolution_2d.groups, node->params.convolution_2d.group_input_channels, node->params.convolution_2d.group_output_channels, node->params.convolution_2d.group_input_channels * node->params.convolution_2d.groups /* input_pixel_stride */, node->params.convolution_2d.group_output_channels * node->params.convolution_2d.groups /* output_pixel_stride */, (int8_t) values[input_id].quantization.zero_point, values[input_id].quantization.scale, values[filter_id].quantization.channelwise_scale, filter_data, bias_data, (int8_t) output_zero_point, output_scale, output_min, output_max, node->flags, code_cache, weights_cache, &opdata->operator_objects[0]); break; } default: XNN_UNREACHABLE; } break; case xnn_datatype_quint8: { const float output_scale = values[output_id].quantization.scale; const int32_t output_zero_point = values[output_id].quantization.zero_point; const uint8_t output_min = xnn_qu8_quantize(node->activation.output_min, output_scale, output_zero_point); const uint8_t output_max = xnn_qu8_quantize(node->activation.output_max, output_scale, output_zero_point); status = xnn_create_convolution2d_nhwc_qu8( node->params.convolution_2d.input_padding_top, node->params.convolution_2d.input_padding_right, node->params.convolution_2d.input_padding_bottom, node->params.convolution_2d.input_padding_left, node->params.convolution_2d.kernel_height, node->params.convolution_2d.kernel_width, node->params.convolution_2d.subsampling_height, node->params.convolution_2d.subsampling_width, node->params.convolution_2d.dilation_height, node->params.convolution_2d.dilation_width, node->params.convolution_2d.groups, node->params.convolution_2d.group_input_channels, node->params.convolution_2d.group_output_channels, node->params.convolution_2d.group_input_channels * node->params.convolution_2d.groups /* input_pixel_stride */, node->params.convolution_2d.group_output_channels * node->params.convolution_2d.groups /* output_pixel_stride */, (uint8_t) values[input_id].quantization.zero_point, values[input_id].quantization.scale, (uint8_t) values[filter_id].quantization.zero_point, values[filter_id].quantization.scale, filter_data, bias_data, (uint8_t) output_zero_point, output_scale, output_min, output_max, node->flags, code_cache, weights_cache, &opdata->operator_objects[0]); break; } default: XNN_UNREACHABLE; } } return status; } static enum xnn_status reshape_convolution_operator( struct xnn_operator_data* opdata, struct xnn_value* values, size_t num_values, pthreadpool_t threadpool) { const uint32_t input_id = opdata->inputs[0]; assert(input_id < num_values); const uint32_t output_id = opdata->outputs[0]; assert(output_id < num_values); const size_t batch_size = values[input_id].shape.dim[0]; const size_t input_height = values[input_id].shape.dim[1]; const size_t input_width = values[input_id].shape.dim[2]; size_t output_height, output_width; enum xnn_status status = xnn_status_invalid_state; const size_t old_workspace_size = opdata->workspace_size; switch (opdata->operator_objects[0]->type) { case xnn_operator_type_convolution_nchw_f16: status = xnn_reshape_convolution2d_nchw_f16( opdata->operator_objects[0], batch_size, input_height, input_width, &output_height, &output_width, threadpool); break; case xnn_operator_type_convolution_nchw_f32: status = xnn_reshape_convolution2d_nchw_f32( opdata->operator_objects[0], batch_size, input_height, input_width, &output_height, &output_width, threadpool); break; case xnn_operator_type_convolution_nhwc_f32: status = xnn_reshape_convolution2d_nhwc_f32( opdata->operator_objects[0], batch_size, input_height, input_width, &opdata->workspace_size, &opdata->workspace_alignment, &output_height, &output_width, threadpool); break; case xnn_operator_type_convolution_nhwc_f16: status = xnn_reshape_convolution2d_nhwc_f16( opdata->operator_objects[0], batch_size, input_height, input_width, &opdata->workspace_size, &opdata->workspace_alignment, &output_height, &output_width, threadpool); break; case xnn_operator_type_convolution_nhwc_qd8_f16_qc8w: status = xnn_reshape_convolution2d_nhwc_qd8_f16_qc8w( opdata->operator_objects[0], batch_size, input_height, input_width, &opdata->workspace_size, &opdata->workspace_alignment, &output_height, &output_width, threadpool); break; case xnn_operator_type_convolution_nhwc_qdu8_f16_qc8w: status = xnn_reshape_convolution2d_nhwc_qdu8_f16_qc8w( opdata->operator_objects[0], batch_size, input_height, input_width, &opdata->workspace_size, &opdata->workspace_alignment, &output_height, &output_width, threadpool); break; case xnn_operator_type_convolution_nhwc_qd8_f32_qc8w: status = xnn_reshape_convolution2d_nhwc_qd8_f32_qc8w( opdata->operator_objects[0], batch_size, input_height, input_width, &opdata->workspace_size, &opdata->workspace_alignment, &output_height, &output_width, threadpool); break; case xnn_operator_type_convolution_nhwc_qdu8_f32_qc8w: status = xnn_reshape_convolution2d_nhwc_qdu8_f32_qc8w( opdata->operator_objects[0], batch_size, input_height, input_width, &opdata->workspace_size, &opdata->workspace_alignment, &output_height, &output_width, threadpool); break; case xnn_operator_type_convolution_nhwc_qc8: status = xnn_reshape_convolution2d_nhwc_qs8_qc8w( opdata->operator_objects[0], batch_size, input_height, input_width, &opdata->workspace_size, &opdata->workspace_alignment, &output_height, &output_width, threadpool); break; case xnn_operator_type_convolution_nhwc_qs8: status = xnn_reshape_convolution2d_nhwc_qs8( opdata->operator_objects[0], batch_size, input_height, input_width, &opdata->workspace_size, &opdata->workspace_alignment, &output_height, &output_width, threadpool); break; case xnn_operator_type_convolution_nhwc_qu8: status = xnn_reshape_convolution2d_nhwc_qu8( opdata->operator_objects[0], batch_size, input_height, input_width, &opdata->workspace_size, &opdata->workspace_alignment, &output_height, &output_width, threadpool); break; default: XNN_UNREACHABLE; } if (status != xnn_status_success) { return status; } const size_t output_pixel_stride = opdata->operator_objects[0]->output_pixel_stride; struct xnn_value* output_value = values + output_id; output_value->shape.dim[0] = batch_size; output_value->shape.dim[1] = output_height; output_value->shape.dim[2] = output_width; output_value->shape.dim[3] = output_pixel_stride; output_value->shape.num_dims = 4; const size_t new_size = xnn_tensor_get_size(output_value); if (new_size > output_value->size || opdata->workspace_size > old_workspace_size) { output_value->size = new_size; return xnn_status_reallocation_required; } return xnn_status_success; } static enum xnn_status setup_convolution_operator( const struct xnn_operator_data* opdata, const struct xnn_value* values, size_t num_values, pthreadpool_t threadpool) { const uint32_t input_id = opdata->inputs[0]; assert(input_id != XNN_INVALID_VALUE_ID); assert(input_id < num_values); const uint32_t output_id = opdata->outputs[0]; assert(output_id != XNN_INVALID_VALUE_ID); assert(output_id < num_values); const struct xnn_value* input_value = values + input_id; const void* input_data = input_value->data; assert(input_data != NULL); const struct xnn_value* output_value = values + output_id; void* output_data = output_value->data; assert(output_data != NULL); switch (opdata->operator_objects[0]->type) { case xnn_operator_type_convolution_nchw_f16: return xnn_setup_convolution2d_nchw_f16( opdata->operator_objects[0], input_data, output_data); break; case xnn_operator_type_convolution_nchw_f32: return xnn_setup_convolution2d_nchw_f32( opdata->operator_objects[0], input_data, output_data); break; case xnn_operator_type_convolution_nhwc_f32: return xnn_setup_convolution2d_nhwc_f32( opdata->operator_objects[0], opdata->workspace, input_data, output_data); break; case xnn_operator_type_convolution_nhwc_f16: return xnn_setup_convolution2d_nhwc_f16( opdata->operator_objects[0], opdata->workspace, input_data, output_data); break; case xnn_operator_type_convolution_nhwc_qc8: return xnn_setup_convolution2d_nhwc_qs8_qc8w( opdata->operator_objects[0], opdata->workspace, input_data, output_data); break; case xnn_operator_type_convolution_nhwc_qd8_f16_qc8w: { const void* quantization_params = input_value->quantization.dynamic_params; assert(quantization_params != NULL); return xnn_setup_convolution2d_nhwc_qd8_f16_qc8w( opdata->operator_objects[0], opdata->workspace, input_data, output_data, quantization_params); } break; case xnn_operator_type_convolution_nhwc_qdu8_f16_qc8w: { const void* quantization_params = input_value->quantization.dynamic_params; assert(quantization_params != NULL); return xnn_setup_convolution2d_nhwc_qdu8_f16_qc8w( opdata->operator_objects[0], opdata->workspace, input_data, output_data, quantization_params); } break; case xnn_operator_type_convolution_nhwc_qd8_f32_qc8w: { const void* quantization_params = input_value->quantization.dynamic_params; assert(quantization_params != NULL); return xnn_setup_convolution2d_nhwc_qd8_f32_qc8w( opdata->operator_objects[0], opdata->workspace, input_data, output_data, quantization_params); } break; case xnn_operator_type_convolution_nhwc_qdu8_f32_qc8w: { const void* quantization_params = input_value->quantization.dynamic_params; assert(quantization_params != NULL); return xnn_setup_convolution2d_nhwc_qdu8_f32_qc8w( opdata->operator_objects[0], opdata->workspace, input_data, output_data, quantization_params); } break; case xnn_operator_type_convolution_nhwc_qs8: return xnn_setup_convolution2d_nhwc_qs8( opdata->operator_objects[0], opdata->workspace, input_data, output_data); break; case xnn_operator_type_convolution_nhwc_qu8: return xnn_setup_convolution2d_nhwc_qu8( opdata->operator_objects[0], opdata->workspace, input_data, output_data); break; default: XNN_UNREACHABLE; } } static inline bool validate_datatypes_with_bias( enum xnn_datatype input_datatype, enum xnn_datatype filter_datatype, enum xnn_datatype bias_datatype, enum xnn_datatype output_datatype) { switch (filter_datatype) { case xnn_datatype_fp32: if (input_datatype == xnn_datatype_fp32 && bias_datatype == xnn_datatype_fp32 && output_datatype == xnn_datatype_fp32) { return true; } else if (input_datatype == xnn_datatype_fp16 && bias_datatype == xnn_datatype_fp32 && output_datatype == xnn_datatype_fp16) { // Flag: XNN_FLAG_FP32_STATIC_WEIGHTS return true; } break; case xnn_datatype_fp16: if (input_datatype == xnn_datatype_fp32 && bias_datatype == xnn_datatype_fp16 && output_datatype == xnn_datatype_fp32) { return true; } else if (input_datatype == xnn_datatype_fp32 && bias_datatype == xnn_datatype_fp32 && output_datatype == xnn_datatype_fp32) { // Flag: XNN_FLAG_FP32_STATIC_BIASES return true; } break; case xnn_datatype_qint8: if (input_datatype == xnn_datatype_qint8 && bias_datatype == xnn_datatype_qint32 && output_datatype == xnn_datatype_qint8) { return true; } break; case xnn_datatype_qcint8: if (input_datatype == xnn_datatype_qint8 && bias_datatype == xnn_datatype_qcint32 && output_datatype == xnn_datatype_qint8) { return true; } else if (input_datatype == xnn_datatype_qdint8 && bias_datatype == xnn_datatype_fp32 && output_datatype == xnn_datatype_fp32) { return true; } else if (input_datatype == xnn_datatype_qdint8 && bias_datatype == xnn_datatype_fp32 && output_datatype == xnn_datatype_fp16) { return true; } break; case xnn_datatype_quint8: if (input_datatype == xnn_datatype_quint8 && bias_datatype == xnn_datatype_qint32 && output_datatype == xnn_datatype_quint8) { return true; } break; default: XNN_UNREACHABLE; } return false; } static inline bool validate_datatypes_without_bias( enum xnn_datatype input_datatype, enum xnn_datatype filter_datatype, enum xnn_datatype output_datatype) { switch (filter_datatype) { case xnn_datatype_fp32: if (input_datatype == xnn_datatype_fp32 && output_datatype == xnn_datatype_fp32) { return true; } else if (input_datatype == xnn_datatype_fp16 && output_datatype == xnn_datatype_fp16) { // Flag: XNN_FLAG_FP32_STATIC_WEIGHTS return true; } break; case xnn_datatype_fp16: if (input_datatype == xnn_datatype_fp32 && output_datatype == xnn_datatype_fp32) { return true; } break; case xnn_datatype_qint8: if (input_datatype == xnn_datatype_qint8 && output_datatype == xnn_datatype_qint8) { return true; } break; case xnn_datatype_qcint8: if (input_datatype == xnn_datatype_qint8 && output_datatype == xnn_datatype_qint8) { return true; } else if (input_datatype == xnn_datatype_qdint8 && output_datatype == xnn_datatype_fp32) { return true; } else if (input_datatype == xnn_datatype_qdint8 && output_datatype == xnn_datatype_fp16) { return true; } break; case xnn_datatype_quint8: if (input_datatype == xnn_datatype_quint8 && output_datatype == xnn_datatype_quint8) { return true; } break; default: XNN_UNREACHABLE; } return false; } enum xnn_status xnn_define_convolution_2d( xnn_subgraph_t subgraph, uint32_t input_padding_top, uint32_t input_padding_right, uint32_t input_padding_bottom, uint32_t input_padding_left, uint32_t kernel_height, uint32_t kernel_width, uint32_t subsampling_height, uint32_t subsampling_width, uint32_t dilation_height, uint32_t dilation_width, uint32_t groups, size_t group_input_channels, size_t group_output_channels, float output_min, float output_max, uint32_t input_id, uint32_t filter_id, uint32_t bias_id, uint32_t output_id, uint32_t flags) { enum xnn_status status; if ((status = xnn_subgraph_check_xnnpack_initialized(xnn_node_type_convolution_2d)) != xnn_status_success) { return status; } if (kernel_width == 0 || kernel_height == 0) { xnn_log_error( "failed to define %s operator with %" PRIu32 "x%" PRIu32 " kernel: kernel dimensions must be non-zero", xnn_node_type_to_string(xnn_node_type_convolution_2d), kernel_width, kernel_height); return xnn_status_invalid_parameter; } if (subsampling_width == 0 || subsampling_height == 0) { xnn_log_error( "failed to define %s operator with %" PRIu32 "x%" PRIu32 " subsampling: subsampling dimensions must be non-zero", xnn_node_type_to_string(xnn_node_type_convolution_2d), subsampling_width, subsampling_height); return xnn_status_invalid_parameter; } if (dilation_width == 0 || dilation_height == 0) { xnn_log_error( "failed to define %s operator with %" PRIu32 "x%" PRIu32 " dilation: dilation dimensions must be non-zero", xnn_node_type_to_string(xnn_node_type_convolution_2d), dilation_width, dilation_height); return xnn_status_invalid_parameter; } if (groups == 0) { xnn_log_error( "failed to define %s operator with %" PRIu32 " groups: number of groups must be non-zero", xnn_node_type_to_string(xnn_node_type_convolution_2d), groups); return xnn_status_invalid_parameter; } if (group_input_channels == 0) { xnn_log_error( "failed to define %s operator with %zu input channels per group: number of channels must be non-zero", xnn_node_type_to_string(xnn_node_type_convolution_2d), group_input_channels); return xnn_status_invalid_parameter; } if (group_output_channels == 0) { xnn_log_error( "failed to define %s operator with %zu output channels per group: number of channels must be non-zero", xnn_node_type_to_string(xnn_node_type_convolution_2d), group_output_channels); return xnn_status_invalid_parameter; } status = xnn_subgraph_check_output_min_max(xnn_node_type_convolution_2d, output_min, output_max); if (status != xnn_status_success) { return status; } const uint32_t supported_flags = XNN_FLAG_TENSORFLOW_SAME_PADDING | XNN_FLAG_TRANSIENT_INDIRECTION_BUFFER; const uint32_t invalid_flags = flags & ~supported_flags; if (invalid_flags != 0) { xnn_log_error( "failed to define %s operator with 0x%08" PRIx32 " flags: invalid flags 0x%08" PRIx32, xnn_node_type_to_string(xnn_node_type_convolution_2d), flags, invalid_flags); return xnn_status_invalid_parameter; } const bool any_padding = (input_padding_left | input_padding_top | input_padding_right | input_padding_bottom) != 0; if ((flags & XNN_FLAG_TENSORFLOW_SAME_PADDING) != 0 && any_padding) { xnn_log_error( "failed to define %s operator with %" PRIu32 "+%" PRIu32 "x%" PRIu32 "+%" PRIu32" padding: " "TensorFlow SAME padding can't be combined with explicit padding specification", xnn_node_type_to_string(xnn_node_type_convolution_2d), input_padding_top, input_padding_left, input_padding_bottom, input_padding_right); return xnn_status_invalid_parameter; } // Convert TensorFlow SAME padding to explicit padding specification whenever possible if ((flags & XNN_FLAG_TENSORFLOW_SAME_PADDING) != 0 && (subsampling_height | subsampling_width) == 1) { flags &= ~XNN_FLAG_TENSORFLOW_SAME_PADDING; const uint32_t padding_height = (kernel_height - 1) * dilation_height; const uint32_t padding_width = (kernel_width - 1) * dilation_width; input_padding_left = padding_width / 2; input_padding_top = padding_height / 2; input_padding_right = padding_width - input_padding_left; input_padding_bottom = padding_height - input_padding_top; } if ((status = xnn_subgraph_check_input_node_id(xnn_node_type_convolution_2d, input_id, subgraph->num_values)) != xnn_status_success) { return status; } const struct xnn_value* input_value = &subgraph->values[input_id]; status = xnn_subgraph_check_input_type_dense(xnn_node_type_convolution_2d, input_id, input_value); if (status != xnn_status_success) { return status; } switch (input_value->datatype) { case xnn_datatype_fp16: case xnn_datatype_fp32: case xnn_datatype_qint8: case xnn_datatype_quint8: break; case xnn_datatype_qdint8: if (input_value->quantization.num_nonbatch_dims >= input_value->shape.num_dims) { xnn_log_error( "failed to define %s operator with input ID #%" PRIu32 ": num_nonbatch_dims (%zu) must be " "< num_dims (%zu)", xnn_node_type_to_string(xnn_node_type_convolution_2d), input_id, input_value->quantization.num_nonbatch_dims, input_value->shape.num_dims); return xnn_status_invalid_parameter; } break; default: xnn_log_error( "failed to define %s operator with input ID #%" PRIu32 ": unsupported Value datatype %s (%d)", xnn_node_type_to_string(xnn_node_type_convolution_2d), input_id, xnn_datatype_to_string(input_value->datatype), input_value->datatype); return xnn_status_invalid_parameter; } if (filter_id >= subgraph->num_values) { xnn_log_error( "failed to define %s operator with filter ID #%" PRIu32 ": invalid Value ID", xnn_node_type_to_string(xnn_node_type_convolution_2d), filter_id); return xnn_status_invalid_parameter; } const struct xnn_value* filter_value = &subgraph->values[filter_id]; if (filter_value->type != xnn_value_type_dense_tensor) { xnn_log_error( "failed to define %s operator with filter ID #%" PRIu32 ": unsupported Value type %d (expected dense tensor)", xnn_node_type_to_string(xnn_node_type_convolution_2d), filter_id, filter_value->type); return xnn_status_invalid_parameter; } if (filter_value->data == NULL) { xnn_log_error( "failed to define %s operator with filter ID #%" PRIu32 ": non-static Value", xnn_node_type_to_string(xnn_node_type_convolution_2d), filter_id); return xnn_status_invalid_parameter; } switch (filter_value->datatype) { case xnn_datatype_fp32: case xnn_datatype_fp16: break; case xnn_datatype_qint8: if (filter_value->quantization.zero_point != 0) { xnn_log_error( "failed to define %s operator with filter ID #%" PRIu32 ": unsupported quantization zero point %" PRId32 " for datatype %s", xnn_node_type_to_string(xnn_node_type_convolution_2d), filter_id, filter_value->quantization.zero_point, xnn_datatype_to_string(filter_value->datatype)); return xnn_status_invalid_parameter; } break; case xnn_datatype_qcint8: break; case xnn_datatype_quint8: break; default: xnn_log_error( "failed to define %s operator with filter ID #%" PRIu32 ": unsupported Value datatype %s (%d)", xnn_node_type_to_string(xnn_node_type_convolution_2d), filter_id, xnn_datatype_to_string(filter_value->datatype), filter_value->datatype); return xnn_status_invalid_parameter; } const struct xnn_value* bias_value = NULL; if (bias_id != XNN_INVALID_VALUE_ID) { if (bias_id >= subgraph->num_values) { xnn_log_error( "failed to define %s operator with bias ID #%" PRIu32 ": invalid Value ID", xnn_node_type_to_string(xnn_node_type_convolution_2d), bias_id); return xnn_status_invalid_parameter; } bias_value = &subgraph->values[bias_id]; if (bias_value->type != xnn_value_type_dense_tensor) { xnn_log_error( "failed to define %s operator with bias ID #%" PRIu32 ": unsupported Value type %d (expected dense tensor)", xnn_node_type_to_string(xnn_node_type_convolution_2d), bias_id, bias_value->type); return xnn_status_invalid_parameter; } if (bias_value->data == NULL) { xnn_log_error( "failed to define %s operator with bias ID #%" PRIu32 ": non-static Value", xnn_node_type_to_string(xnn_node_type_convolution_2d), bias_id); return xnn_status_invalid_parameter; } switch (bias_value->datatype) { case xnn_datatype_fp16: case xnn_datatype_fp32: case xnn_datatype_qint32: case xnn_datatype_qcint32: break; default: xnn_log_error( "failed to define %s operator with bias ID #%" PRIu32 ": unsupported Value datatype %s (%d)", xnn_node_type_to_string(xnn_node_type_convolution_2d), bias_id, xnn_datatype_to_string(bias_value->datatype), bias_value->datatype); return xnn_status_invalid_parameter; } } status = xnn_subgraph_check_output_node_id(xnn_node_type_convolution_2d, output_id, subgraph->num_values); if (status != xnn_status_success) { return status; } const struct xnn_value* output_value = &subgraph->values[output_id]; status = xnn_subgraph_check_output_type_dense(xnn_node_type_convolution_2d, output_id, output_value); if (status != xnn_status_success) { return status; } if (filter_value->shape.dim[0] != group_output_channels * groups) { xnn_log_error( "failed to define %s operator with filter output channels %zu, groups #%" PRIu32 " and group_output_channels %zu:" "mismatching shapes, filter output channels must be equal to groups * group_output_channels.", xnn_node_type_to_string(xnn_node_type_convolution_2d), filter_value->shape.dim[0], groups, group_output_channels); return xnn_status_invalid_parameter; } switch (output_value->datatype) { case xnn_datatype_fp16: case xnn_datatype_fp32: case xnn_datatype_qint8: case xnn_datatype_quint8: break; default: xnn_log_error( "failed to define %s operator with output ID #%" PRIu32 ": unsupported Value datatype %s (%d)", xnn_node_type_to_string(xnn_node_type_convolution_2d), output_id, xnn_datatype_to_string(output_value->datatype), output_value->datatype); return xnn_status_invalid_parameter; } if (bias_value != NULL) { if (!validate_datatypes_with_bias( input_value->datatype, filter_value->datatype, bias_value->datatype, output_value->datatype)) { xnn_log_error( "failed to define %s operator with input ID #%" PRIu32 ", filter ID #%" PRIu32 ", bias ID #%" PRIu32 ", and output ID #%" PRIu32 ": mismatching datatypes across input (%s), filter (%s), bias (%s), and output (%s)", xnn_node_type_to_string(xnn_node_type_convolution_2d), input_id, filter_id, bias_id, output_id, xnn_datatype_to_string(input_value->datatype), xnn_datatype_to_string(filter_value->datatype), xnn_datatype_to_string(bias_value->datatype), xnn_datatype_to_string(output_value->datatype)); return xnn_status_invalid_parameter; } } else { if (!validate_datatypes_without_bias( input_value->datatype, filter_value->datatype, output_value->datatype)) { xnn_log_error( "failed to define %s operator with input ID #%" PRIu32 ", filter ID #%" PRIu32 ", and output ID #%" PRIu32 ": mismatching datatypes across input (%s), filter (%s), and output (%s)", xnn_node_type_to_string(xnn_node_type_convolution_2d), input_id, filter_id, output_id, xnn_datatype_to_string(input_value->datatype), xnn_datatype_to_string(filter_value->datatype), xnn_datatype_to_string(output_value->datatype)); return xnn_status_invalid_parameter; } } if (filter_value->datatype == xnn_datatype_qcint8) { if (filter_value->quantization.channel_dimension != 0) { xnn_log_error( "failed to define %s operator with filter ID #%" PRIu32 ": invalid channel dimension %zu", xnn_node_type_to_string(xnn_node_type_convolution_2d), input_id, filter_value->quantization.channel_dimension); return xnn_status_invalid_parameter; } if (bias_value != NULL) { assert(bias_value->datatype == xnn_datatype_qcint32 || bias_value->datatype == xnn_datatype_fp32); if (bias_value->datatype == xnn_datatype_qcint32 && bias_value->quantization.channel_dimension != 0) { xnn_log_error( "failed to define %s operator with bias ID #%" PRIu32 ": invalid channel dimension %zu", xnn_node_type_to_string(xnn_node_type_convolution_2d), bias_id, bias_value->quantization.channel_dimension); return xnn_status_invalid_parameter; } } } if (input_value->datatype == xnn_datatype_fp32 && output_value->datatype == xnn_datatype_fp32 && (!bias_value || bias_value->datatype == xnn_datatype_fp32)) { const bool unit_subsampling = (subsampling_width | subsampling_height) == 1; const size_t kernel_size = kernel_height * kernel_width; if (groups == 1 && kernel_size == 1 && unit_subsampling && !any_padding) { const struct xnn_gemm_config* gemm_config = xnn_init_pf32_gemm_config(); if (gemm_config != NULL && gemm_config->init.f32 != NULL) { // Insert a node to pack the LHS. uint32_t new_id = XNN_INVALID_VALUE_ID; status = xnn_insert_pack_lh_node(subgraph, input_value, input_id, &new_id); if (status != xnn_status_success) { return status; } input_id = new_id; } } } struct xnn_node* node = xnn_subgraph_new_node(subgraph); if (node == NULL) { return xnn_status_out_of_memory; } node->type = xnn_node_type_convolution_2d; node->params.convolution_2d.input_padding_top = input_padding_top; node->params.convolution_2d.input_padding_right = input_padding_right; node->params.convolution_2d.input_padding_bottom = input_padding_bottom; node->params.convolution_2d.input_padding_left = input_padding_left; node->params.convolution_2d.kernel_height = kernel_height; node->params.convolution_2d.kernel_width = kernel_width; node->params.convolution_2d.subsampling_height = subsampling_height; node->params.convolution_2d.subsampling_width = subsampling_width; node->params.convolution_2d.dilation_height = dilation_height; node->params.convolution_2d.dilation_width = dilation_width; node->params.convolution_2d.groups = groups; node->params.convolution_2d.group_input_channels = group_input_channels; node->params.convolution_2d.group_output_channels = group_output_channels; node->activation.output_min = output_min; node->activation.output_max = output_max; node->num_inputs = 2 + (size_t) (bias_id != XNN_INVALID_VALUE_ID); node->inputs[0] = input_id; node->inputs[1] = filter_id; node->inputs[2] = bias_id; node->num_outputs = 1; node->outputs[0] = output_id; node->flags = flags; node->create = create_convolution_operator; node->reshape = reshape_convolution_operator; node->setup = setup_convolution_operator; return xnn_status_success; };