// 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/allocation-type.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/subgraph.h" #include "xnnpack/requantization.h" #include "xnnpack/subgraph-validation.h" #include "xnnpack/subgraph.h" #include "pthreadpool.h" // Format is input_type, weights type, output type, (dynamic)? enum fully_connected_op_type { fc_type_invalid = 0, fc_type_f16_f16_f16 = 1, fc_type_f16_f16_f16_dynamic = 2, fc_type_f16_f32_f16 = 3, fc_type_f16_f32_f16_dynamic = 4, fc_type_qd8_f16_qc4w = 5, fc_type_qd8_f16_qb4w = 6, fc_type_qd8_f16_qc8w = 7, fc_type_f32_f32_f32 = 8, fc_type_f32_f32_f32_dynamic = 9, fc_type_qd8_f32_qb4w = 10, fc_type_f32_f32_qc4w = 11, fc_type_qd8_f32_qc4w = 12, fc_type_qp8_f32_qc4w = 13, fc_type_f32_f32_qc8w = 14, fc_type_qd8_f32_qc8w = 15, fc_type_qs8_qs8_qc8w = 16, fc_type_qs8_qs8_qs8 = 17, fc_type_qu8_qu8_qu8 = 18, fc_type_qp8_f32_qb4w = 19, fc_type_pf32_f32_f32 = 20, fc_type_f32_f16_f32 = 21, fc_type_qdu8_f16_qc8w = 22, fc_type_qdu8_f32_qc8w = 23, fc_type_qdu8_f32_qc4w = 24, fc_type_qdu8_f32_qb4w = 26, fc_type_qdu8_f16_qc4w = 27, }; enum fully_connected_op_type get_fully_connected_op_type( const struct xnn_value* input_value, const struct xnn_value* filter_value, const struct xnn_value* bias_value, const struct xnn_value* output_value) { const void* filter_data = filter_value->fp32_data != NULL ? filter_value->fp32_data : filter_value->data; bool has_non_static_weights = (filter_data == NULL); if (bias_value) { const void* bias_data = bias_value->fp32_data != NULL ? bias_value->fp32_data : bias_value->data; has_non_static_weights |= (bias_data == NULL); } const enum xnn_datatype input_datatype = input_value->datatype; const enum xnn_datatype filter_datatype = filter_value->datatype; const enum xnn_datatype output_datatype = output_value->datatype; switch (output_datatype) { case xnn_datatype_fp16: switch (filter_datatype) { case xnn_datatype_fp16: if (has_non_static_weights) { return fc_type_f16_f16_f16_dynamic; } else { return fc_type_f16_f16_f16; } case xnn_datatype_fp32: if (has_non_static_weights) { return fc_type_f16_f32_f16_dynamic; } else { return fc_type_f16_f32_f16; } case xnn_datatype_qcint4: switch (input_datatype) { case xnn_datatype_qdint8: return fc_type_qd8_f16_qc4w; case xnn_datatype_qduint8: return fc_type_qdu8_f16_qc4w; default: XNN_UNREACHABLE; } break; case xnn_datatype_qbint4: return fc_type_qd8_f16_qb4w; case xnn_datatype_qcint8: switch (input_datatype) { case xnn_datatype_qdint8: return fc_type_qd8_f16_qc8w; case xnn_datatype_qduint8: return fc_type_qdu8_f16_qc8w; default: XNN_UNREACHABLE; } break; default: XNN_UNREACHABLE; } break; case xnn_datatype_fp32: switch (filter_datatype) { case xnn_datatype_fp16: switch (input_datatype) { case xnn_datatype_fp32: return fc_type_f32_f16_f32; default: XNN_UNREACHABLE; } case xnn_datatype_fp32: if (has_non_static_weights) { return fc_type_f32_f32_f32_dynamic; } else { switch (input_datatype) { case xnn_datatype_fp32: return fc_type_f32_f32_f32; case xnn_datatype_pfp32: return fc_type_pf32_f32_f32; default: XNN_UNREACHABLE; } } case xnn_datatype_qbint4: switch (input_datatype) { case xnn_datatype_qdint8: return fc_type_qd8_f32_qb4w; case xnn_datatype_qduint8: return fc_type_qdu8_f32_qb4w; case xnn_datatype_qpint8: return fc_type_qp8_f32_qb4w; default: XNN_UNREACHABLE; } case xnn_datatype_qcint4: switch (input_datatype) { case xnn_datatype_fp32: return fc_type_f32_f32_qc4w; case xnn_datatype_qdint8: return fc_type_qd8_f32_qc4w; case xnn_datatype_qduint8: return fc_type_qdu8_f32_qc4w; case xnn_datatype_qpint8: return fc_type_qp8_f32_qc4w; default: XNN_UNREACHABLE; } break; case xnn_datatype_qcint8: switch (input_datatype) { case xnn_datatype_fp32: return fc_type_f32_f32_qc8w; case xnn_datatype_qdint8: return fc_type_qd8_f32_qc8w; case xnn_datatype_qduint8: return fc_type_qdu8_f32_qc8w; default: XNN_UNREACHABLE; } break; default: XNN_UNREACHABLE; } break; case xnn_datatype_qint8: switch (filter_datatype) { case xnn_datatype_qcint8: return fc_type_qs8_qs8_qc8w; case xnn_datatype_qint8: return fc_type_qs8_qs8_qs8; default: XNN_UNREACHABLE; } break; case xnn_datatype_quint8: return fc_type_qu8_qu8_qu8; default: XNN_UNREACHABLE; } } static enum xnn_status create_fully_connected_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); size_t output_channels, input_channels; if (node->flags & XNN_FLAG_TRANSPOSE_WEIGHTS) { input_channels = values[node->inputs[1]].shape.dim[0]; output_channels = values[node->inputs[1]].shape.dim[1]; } else { output_channels = values[node->inputs[1]].shape.dim[0]; input_channels = values[node->inputs[1]].shape.dim[1]; } const void* kernel_data = values[filter_id].fp32_data != NULL ? values[filter_id].fp32_data : values[filter_id].data; bool has_non_static_weights = (kernel_data == NULL); const void* bias_data = NULL; const struct xnn_value* bias_value = NULL; if (node->num_inputs > 2) { const uint32_t 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; has_non_static_weights |= (bias_data == NULL); bias_value = &values[bias_id]; } enum xnn_status status; enum fully_connected_op_type op_type = get_fully_connected_op_type( &values[input_id], &values[filter_id], bias_value, &values[output_id]); switch (op_type) { case fc_type_f16_f16_f16_dynamic: status = xnn_create_dynamic_fully_connected_nc_f16( node->activation.output_min, node->activation.output_max, /*flags=*/node->flags, &opdata->operator_objects[0]); break; case fc_type_f16_f16_f16: status = xnn_create_fully_connected_nc_f16( input_channels, output_channels, /*input_stride=*/input_channels, /*output_stride=*/output_channels, kernel_data, bias_data, node->activation.output_min, node->activation.output_max, node->flags, code_cache, weights_cache, &opdata->operator_objects[0]); break; case fc_type_f16_f32_f16_dynamic: status = xnn_create_dynamic_fully_connected_nc_f16( node->activation.output_min, node->activation.output_max, node->flags | XNN_FLAG_FP32_STATIC_WEIGHTS, &opdata->operator_objects[0]); break; case fc_type_f16_f32_f16: status = xnn_create_fully_connected_nc_f16( input_channels, output_channels, /*input_stride=*/input_channels, /*output_stride=*/output_channels, kernel_data, bias_data, node->activation.output_min, node->activation.output_max, node->flags | XNN_FLAG_FP32_STATIC_WEIGHTS, code_cache, weights_cache, &opdata->operator_objects[0]); break; case fc_type_qd8_f16_qc4w: status = xnn_create_fully_connected_nc_qd8_f16_qc4w( input_channels, output_channels, /*input_stride=*/input_channels, /*output_stride=*/output_channels, /*kernel_zero_point=*/values[filter_id].quantization.zero_point, values[filter_id].quantization.channelwise_scale, kernel_data, bias_data, node->activation.output_min, node->activation.output_max, node->flags, code_cache, weights_cache, &opdata->operator_objects[0]); break; case fc_type_qdu8_f16_qc4w: status = xnn_create_fully_connected_nc_qdu8_f16_qc4w( input_channels, output_channels, /*input_stride=*/input_channels, /*output_stride=*/output_channels, /*kernel_zero_point=*/values[filter_id].quantization.zero_point, values[filter_id].quantization.channelwise_scale, kernel_data, bias_data, node->activation.output_min, node->activation.output_max, node->flags, code_cache, weights_cache, &opdata->operator_objects[0]); break; case fc_type_qd8_f16_qb4w: status = xnn_create_fully_connected_nc_qd8_f16_qb4w( input_channels, output_channels, /*input_stride=*/input_channels, /*output_stride=*/output_channels, /*block_size=*/values[filter_id].quantization.block_size, /*kernel_zero_point=*/values[filter_id].quantization.zero_point, (const uint16_t*)values[filter_id].quantization.blockwise_scale, kernel_data, bias_data, node->activation.output_min, node->activation.output_max, node->flags, code_cache, weights_cache, &opdata->operator_objects[0]); break; case fc_type_qd8_f16_qc8w: status = xnn_create_fully_connected_nc_qd8_f16_qc8w( input_channels, output_channels, /*input_stride=*/input_channels, /*output_stride=*/output_channels, values[filter_id].quantization.channelwise_scale, kernel_data, bias_data, node->activation.output_min, node->activation.output_max, node->flags, code_cache, weights_cache, &opdata->operator_objects[0]); break; case fc_type_qdu8_f16_qc8w: status = xnn_create_fully_connected_nc_qdu8_f16_qc8w( input_channels, output_channels, /*input_stride=*/input_channels, /*output_stride=*/output_channels, values[filter_id].quantization.channelwise_scale, kernel_data, bias_data, node->activation.output_min, node->activation.output_max, node->flags, code_cache, weights_cache, &opdata->operator_objects[0]); break; case fc_type_f32_f32_f32_dynamic: status = xnn_create_dynamic_fully_connected_nc_f32( node->activation.output_min, node->activation.output_max, /*flags=*/node->flags, &opdata->operator_objects[0]); break; case fc_type_f32_f32_f32: status = xnn_create_fully_connected_nc_f32( input_channels, output_channels, /*input_stride=*/input_channels, /*output_stride=*/output_channels, kernel_data, bias_data, node->activation.output_min, node->activation.output_max, /*flags=*/node->flags, code_cache, weights_cache, &opdata->operator_objects[0]); break; case fc_type_pf32_f32_f32: status = xnn_create_fully_connected_nc_pf32( input_channels, output_channels, /*input_stride=*/input_channels, /*output_stride=*/output_channels, kernel_data, bias_data, node->activation.output_min, node->activation.output_max, /*flags=*/node->flags, code_cache, weights_cache, &opdata->operator_objects[0]); break; case fc_type_qd8_f32_qb4w: status = xnn_create_fully_connected_nc_qd8_f32_qb4w( input_channels, output_channels, /*input_stride=*/input_channels, /*output_stride=*/output_channels, /*block_size=*/values[filter_id].quantization.block_size, /*kernel_zero_point=*/values[filter_id].quantization.zero_point, (const uint16_t*)values[filter_id].quantization.blockwise_scale, kernel_data, bias_data, node->activation.output_min, node->activation.output_max, node->flags, code_cache, weights_cache, &opdata->operator_objects[0]); break; case fc_type_qdu8_f32_qb4w: status = xnn_create_fully_connected_nc_qdu8_f32_qb4w( input_channels, output_channels, /*input_stride=*/input_channels, /*output_stride=*/output_channels, /*block_size=*/values[filter_id].quantization.block_size, /*kernel_zero_point=*/values[filter_id].quantization.zero_point, (const uint16_t*)values[filter_id].quantization.blockwise_scale, kernel_data, bias_data, node->activation.output_min, node->activation.output_max, node->flags, code_cache, weights_cache, &opdata->operator_objects[0]); break; case fc_type_f32_f16_f32: { uint32_t flags = node->flags; if (bias_value != NULL && bias_value->datatype == xnn_datatype_fp32) { flags |= XNN_FLAG_FP32_STATIC_BIASES; } status = xnn_create_fully_connected_nc_f32_f16( input_channels, output_channels, /*input_stride=*/input_channels, /*output_stride=*/output_channels, kernel_data, bias_data, node->activation.output_min, node->activation.output_max, flags, code_cache, weights_cache, &opdata->operator_objects[0]); break; } case fc_type_qp8_f32_qb4w: status = xnn_create_fully_connected_nc_qp8_f32_qb4w( input_channels, output_channels, /*input_stride=*/input_channels, /*output_stride=*/output_channels, /*block_size=*/values[filter_id].quantization.block_size, /*kernel_zero_point=*/values[filter_id].quantization.zero_point, (const uint16_t*)values[filter_id].quantization.blockwise_scale, kernel_data, bias_data, node->activation.output_min, node->activation.output_max, node->flags, code_cache, weights_cache, &opdata->operator_objects[0]); break; case fc_type_f32_f32_qc4w: status = xnn_create_fully_connected_nc_f32_qc4w( input_channels, output_channels, /*input_stride=*/input_channels, /*output_stride=*/output_channels, values[filter_id].quantization.zero_point, values[filter_id].quantization.channelwise_scale, kernel_data, bias_data, node->activation.output_min, node->activation.output_max, /*flags=*/node->flags, code_cache, weights_cache, &opdata->operator_objects[0]); break; case fc_type_qd8_f32_qc4w: status = xnn_create_fully_connected_nc_qd8_f32_qc4w( input_channels, output_channels, /*input_stride=*/input_channels, /*output_stride=*/output_channels, /*kernel_zero_point=*/values[filter_id].quantization.zero_point, values[filter_id].quantization.channelwise_scale, kernel_data, bias_data, node->activation.output_min, node->activation.output_max, node->flags, code_cache, weights_cache, &opdata->operator_objects[0]); break; case fc_type_qdu8_f32_qc4w: status = xnn_create_fully_connected_nc_qdu8_f32_qc4w( input_channels, output_channels, /*input_stride=*/input_channels, /*output_stride=*/output_channels, /*kernel_zero_point=*/values[filter_id].quantization.zero_point, values[filter_id].quantization.channelwise_scale, kernel_data, bias_data, node->activation.output_min, node->activation.output_max, node->flags, code_cache, weights_cache, &opdata->operator_objects[0]); break; case fc_type_qp8_f32_qc4w: status = xnn_create_fully_connected_nc_qp8_f32_qc4w( input_channels, output_channels, /*input_stride=*/input_channels, /*output_stride=*/output_channels, /*kernel_zero_point=*/values[filter_id].quantization.zero_point, values[filter_id].quantization.channelwise_scale, kernel_data, bias_data, node->activation.output_min, node->activation.output_max, node->flags, code_cache, weights_cache, &opdata->operator_objects[0]); break; case fc_type_f32_f32_qc8w: status = xnn_create_fully_connected_nc_f32_qc8w( input_channels, output_channels, /*input_stride=*/input_channels, /*output_stride=*/output_channels, values[filter_id].quantization.channelwise_scale, kernel_data, bias_data, node->activation.output_min, node->activation.output_max, /*flags=*/node->flags, code_cache, weights_cache, &opdata->operator_objects[0]); break; case fc_type_qd8_f32_qc8w: status = xnn_create_fully_connected_nc_qd8_f32_qc8w( input_channels, output_channels, /*input_stride=*/input_channels, /*output_stride=*/output_channels, values[filter_id].quantization.channelwise_scale, kernel_data, bias_data, node->activation.output_min, node->activation.output_max, node->flags, code_cache, weights_cache, &opdata->operator_objects[0]); break; case fc_type_qdu8_f32_qc8w: status = xnn_create_fully_connected_nc_qdu8_f32_qc8w( input_channels, output_channels, /*input_stride=*/input_channels, /*output_stride=*/output_channels, values[filter_id].quantization.channelwise_scale, kernel_data, bias_data, node->activation.output_min, node->activation.output_max, node->flags, code_cache, weights_cache, &opdata->operator_objects[0]); break; case fc_type_qs8_qs8_qc8w: assert(!has_non_static_weights); assert(kernel_data != NULL); assert(values[filter_id].datatype == 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_fully_connected_nc_qs8_qc8w( input_channels, output_channels, /*input_stride=*/input_channels, /*output_stride=*/output_channels, (int8_t)values[input_id].quantization.zero_point, values[input_id].quantization.scale, values[filter_id].quantization.channelwise_scale, kernel_data, bias_data, (int8_t)output_zero_point, output_scale, output_min, output_max, /*flags=*/node->flags, code_cache, weights_cache, &opdata->operator_objects[0]); break; case fc_type_qs8_qs8_qs8: { assert(!has_non_static_weights); assert(kernel_data != NULL); 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_fully_connected_nc_qs8( input_channels, output_channels, /*input_stride=*/input_channels, /*output_stride=*/output_channels, (int8_t)values[input_id].quantization.zero_point, values[input_id].quantization.scale, values[filter_id].quantization.scale, kernel_data, bias_data, (int8_t)output_zero_point, output_scale, output_min, output_max, /*flags=*/node->flags, code_cache, weights_cache, &opdata->operator_objects[0]); } break; case fc_type_qu8_qu8_qu8: { assert(!has_non_static_weights); assert(kernel_data != NULL); 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_fully_connected_nc_qu8( input_channels, output_channels, /*input_stride=*/input_channels, /*output_stride=*/output_channels, (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, kernel_data, bias_data, (uint8_t)output_zero_point, output_scale, output_min, output_max, /*flags=*/node->flags, code_cache, weights_cache, &opdata->operator_objects[0]); break; } default: XNN_UNREACHABLE; } return status; } enum xnn_status resize_fully_connected_output_tensor( const struct xnn_operator_data* opdata, struct xnn_value* values, size_t num_values, size_t old_workspace_size, pthreadpool_t threadpool) { const uint32_t filter_id = opdata->inputs[1]; const struct xnn_value* filter = &values[filter_id]; const uint32_t output_id = opdata->outputs[0]; struct xnn_value* output = (struct xnn_value*)&values[output_id]; const uint32_t input_id = opdata->inputs[0]; const struct xnn_value* input = &values[input_id]; bool reshape_2d = opdata->flags & XNN_FLAG_TENSORFLOW_RESHAPE_2D; if (reshape_2d) { output->shape.num_dims = 2; } else { output->shape.num_dims = input->shape.num_dims; } // Infer output channels. const uint32_t filter_output_channel_index = (opdata->flags & XNN_FLAG_TRANSPOSE_WEIGHTS) ? 1 : 0; output->shape.dim[output->shape.num_dims - 1] = filter->shape.dim[filter_output_channel_index]; if (reshape_2d) { const uint32_t filter_input_channel_index = (opdata->flags & XNN_FLAG_TRANSPOSE_WEIGHTS) ? 0 : 1; const size_t num_input_elements = xnn_shape_multiply_all_dims(&input->shape); // propogate the input shape to output. output->shape.dim[0] = num_input_elements / filter->shape.dim[filter_input_channel_index]; } else { // Propagate input shape to output. for (size_t cur_dim = 0; cur_dim < input->shape.num_dims - 1; cur_dim++) { output->shape.dim[cur_dim] = input->shape.dim[cur_dim]; } } const size_t new_size = xnn_tensor_get_size(output); if (new_size > output->size || old_workspace_size < opdata->workspace_size) { output->size = new_size; return xnn_status_reallocation_required; } return xnn_status_success; } static enum xnn_status reshape_fully_connected_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 filter_id = opdata->inputs[1]; assert(filter_id < num_values); const size_t num_input_elements = xnn_shape_multiply_all_dims(&values[input_id].shape); size_t output_channels, input_channels; if (opdata->flags & XNN_FLAG_TRANSPOSE_WEIGHTS) { input_channels = values[filter_id].shape.dim[0]; output_channels = values[filter_id].shape.dim[1]; } else { output_channels = values[filter_id].shape.dim[0]; input_channels = values[filter_id].shape.dim[1]; } const size_t batch_size = num_input_elements / input_channels; const size_t old_workspace_size = opdata->workspace_size; enum xnn_status status = xnn_status_invalid_state; switch (opdata->operator_objects[0]->type) { case xnn_operator_type_dynamic_fully_connected_nc_f16: status = xnn_reshape_dynamic_fully_connected_nc_f16( opdata->operator_objects[0], batch_size, input_channels, output_channels, input_channels, output_channels, &opdata->workspace_size, &opdata->workspace_alignment, threadpool); break; case xnn_operator_type_dynamic_fully_connected_nc_f32: status = xnn_reshape_dynamic_fully_connected_nc_f32( opdata->operator_objects[0], batch_size, input_channels, output_channels, input_channels, output_channels, &opdata->workspace_size, &opdata->workspace_alignment, threadpool); break; case xnn_operator_type_fully_connected_nc_f16: status = xnn_reshape_fully_connected_nc_f16(opdata->operator_objects[0], batch_size, threadpool); break; case xnn_operator_type_fully_connected_nc_f32: status = xnn_reshape_fully_connected_nc_f32(opdata->operator_objects[0], batch_size, threadpool); break; case xnn_operator_type_fully_connected_nc_f32_qc4w: status = xnn_reshape_fully_connected_nc_f32_qc4w( opdata->operator_objects[0], batch_size, threadpool); break; case xnn_operator_type_fully_connected_nc_f32_qc8w: status = xnn_reshape_fully_connected_nc_f32_qc8w( opdata->operator_objects[0], batch_size, threadpool); break; case xnn_operator_type_fully_connected_nc_qd8_f32_qc4w: status = xnn_reshape_fully_connected_nc_qd8_f32_qc4w( opdata->operator_objects[0], batch_size, threadpool); break; case xnn_operator_type_fully_connected_nc_qdu8_f32_qc4w: status = xnn_reshape_fully_connected_nc_qdu8_f32_qc4w( opdata->operator_objects[0], batch_size, threadpool); break; case xnn_operator_type_fully_connected_nc_qd8_f16_qc4w: status = xnn_reshape_fully_connected_nc_qd8_f16_qc4w( opdata->operator_objects[0], batch_size, threadpool); break; case xnn_operator_type_fully_connected_nc_qdu8_f16_qc4w: status = xnn_reshape_fully_connected_nc_qdu8_f16_qc4w( opdata->operator_objects[0], batch_size, threadpool); break; case xnn_operator_type_fully_connected_nc_qd8_f16_qb4w: status = xnn_reshape_fully_connected_nc_qd8_f16_qb4w( opdata->operator_objects[0], batch_size, threadpool); break; case xnn_operator_type_fully_connected_nc_qd8_f32_qb4w: status = xnn_reshape_fully_connected_nc_qd8_f32_qb4w( opdata->operator_objects[0], batch_size, threadpool); break; case xnn_operator_type_fully_connected_nc_qdu8_f32_qb4w: status = xnn_reshape_fully_connected_nc_qdu8_f32_qb4w( opdata->operator_objects[0], batch_size, threadpool); break; case xnn_operator_type_fully_connected_nc_qd8_f16_qc8w: status = xnn_reshape_fully_connected_nc_qd8_f16_qc8w( opdata->operator_objects[0], batch_size, threadpool); break; case xnn_operator_type_fully_connected_nc_qdu8_f16_qc8w: status = xnn_reshape_fully_connected_nc_qdu8_f16_qc8w( opdata->operator_objects[0], batch_size, threadpool); break; case xnn_operator_type_fully_connected_nc_qd8_f32_qc8w: status = xnn_reshape_fully_connected_nc_qd8_f32_qc8w( opdata->operator_objects[0], batch_size, threadpool); break; case xnn_operator_type_fully_connected_nc_qdu8_f32_qc8w: status = xnn_reshape_fully_connected_nc_qdu8_f32_qc8w( opdata->operator_objects[0], batch_size, threadpool); break; case xnn_operator_type_fully_connected_nc_qp8_f32_qc4w: status = xnn_reshape_fully_connected_nc_qp8_f32_qc4w( opdata->operator_objects[0], batch_size, threadpool); break; case xnn_operator_type_fully_connected_nc_qp8_f32_qb4w: status = xnn_reshape_fully_connected_nc_qp8_f32_qb4w( opdata->operator_objects[0], batch_size, threadpool); break; case xnn_operator_type_fully_connected_nc_qs8: status = xnn_reshape_fully_connected_nc_qs8(opdata->operator_objects[0], batch_size, threadpool); break; case xnn_operator_type_fully_connected_nc_qs8_qc8w: status = xnn_reshape_fully_connected_nc_qs8_qc8w( opdata->operator_objects[0], batch_size, threadpool); break; case xnn_operator_type_fully_connected_nc_qu8: status = xnn_reshape_fully_connected_nc_qu8(opdata->operator_objects[0], batch_size, threadpool); break; default: XNN_UNREACHABLE; } if (status != xnn_status_success) { return status; } return resize_fully_connected_output_tensor(opdata, values, num_values, old_workspace_size, threadpool); } static enum xnn_status setup_fully_connected_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 filter_id = opdata->inputs[1]; assert(filter_id != XNN_INVALID_VALUE_ID); assert(filter_id < num_values); const uint32_t bias_id = opdata->inputs[2]; 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* kernel_value = values + filter_id; bool has_dynamic_weights = kernel_value->allocation_type != xnn_allocation_type_static; const void* kernel_data = kernel_value->allocation_type == xnn_allocation_type_static ? NULL : kernel_value->data; const void* bias_data = NULL; if (opdata->num_inputs > 2) { assert(bias_id != XNN_INVALID_VALUE_ID); assert(bias_id < num_values); const struct xnn_value* bias_value = values + bias_id; has_dynamic_weights |= bias_value->allocation_type != xnn_allocation_type_static; if (has_dynamic_weights) { kernel_data = kernel_value->data; bias_data = bias_value->data; } } 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_dynamic_fully_connected_nc_f16: assert(kernel_data != NULL); return xnn_setup_dynamic_fully_connected_nc_f16( opdata->operator_objects[0], opdata->workspace, input_data, kernel_data, bias_data, output_data); case xnn_operator_type_dynamic_fully_connected_nc_f32: assert(kernel_data != NULL); return xnn_setup_dynamic_fully_connected_nc_f32( opdata->operator_objects[0], opdata->workspace, input_data, kernel_data, bias_data, output_data); case xnn_operator_type_fully_connected_nc_f16: assert(kernel_data == NULL); assert(bias_data == NULL); return xnn_setup_fully_connected_nc_f16(opdata->operator_objects[0], input_data, output_data); case xnn_operator_type_fully_connected_nc_f32: assert(kernel_data == NULL); assert(bias_data == NULL); return xnn_setup_fully_connected_nc_f32(opdata->operator_objects[0], input_data, output_data); case xnn_operator_type_fully_connected_nc_f32_qc4w: assert(kernel_data == NULL); assert(bias_data == NULL); return xnn_setup_fully_connected_nc_f32_qc4w(opdata->operator_objects[0], input_data, output_data); case xnn_operator_type_fully_connected_nc_f32_qc8w: assert(kernel_data == NULL); assert(bias_data == NULL); return xnn_setup_fully_connected_nc_f32_qc8w(opdata->operator_objects[0], input_data, output_data); case xnn_operator_type_fully_connected_nc_qd8_f32_qc4w: { const void* quantization_params = input_value->quantization.dynamic_params; assert(kernel_data == NULL); assert(bias_data == NULL); assert(quantization_params != NULL); return xnn_setup_fully_connected_nc_qd8_f32_qc4w( opdata->operator_objects[0], input_data, output_data, quantization_params); } case xnn_operator_type_fully_connected_nc_qdu8_f32_qc4w: { const void* quantization_params = input_value->quantization.dynamic_params; assert(kernel_data == NULL); assert(bias_data == NULL); assert(quantization_params != NULL); return xnn_setup_fully_connected_nc_qdu8_f32_qc4w( opdata->operator_objects[0], input_data, output_data, quantization_params); } case xnn_operator_type_fully_connected_nc_qd8_f16_qc4w: { const void* quantization_params = input_value->quantization.dynamic_params; assert(kernel_data == NULL); assert(bias_data == NULL); assert(quantization_params != NULL); return xnn_setup_fully_connected_nc_qd8_f16_qc4w( opdata->operator_objects[0], input_data, output_data, quantization_params); } case xnn_operator_type_fully_connected_nc_qdu8_f16_qc4w: { const void* quantization_params = input_value->quantization.dynamic_params; assert(kernel_data == NULL); assert(bias_data == NULL); assert(quantization_params != NULL); return xnn_setup_fully_connected_nc_qdu8_f16_qc4w( opdata->operator_objects[0], input_data, output_data, quantization_params); } case xnn_operator_type_fully_connected_nc_qd8_f32_qb4w: { const void* quantization_params = input_value->quantization.dynamic_params; assert(kernel_data == NULL); assert(bias_data == NULL); assert(quantization_params != NULL); return xnn_setup_fully_connected_nc_qd8_f32_qb4w( opdata->operator_objects[0], input_data, output_data, quantization_params); } case xnn_operator_type_fully_connected_nc_qdu8_f32_qb4w: { const void* quantization_params = input_value->quantization.dynamic_params; assert(kernel_data == NULL); assert(bias_data == NULL); assert(quantization_params != NULL); return xnn_setup_fully_connected_nc_qdu8_f32_qb4w( opdata->operator_objects[0], input_data, output_data, quantization_params); } case xnn_operator_type_fully_connected_nc_qd8_f16_qb4w: { const void* quantization_params = input_value->quantization.dynamic_params; assert(kernel_data == NULL); assert(bias_data == NULL); assert(quantization_params != NULL); return xnn_setup_fully_connected_nc_qd8_f16_qb4w( opdata->operator_objects[0], input_data, output_data, quantization_params); } case xnn_operator_type_fully_connected_nc_qd8_f16_qc8w: { const void* quantization_params = input_value->quantization.dynamic_params; assert(kernel_data == NULL); assert(bias_data == NULL); assert(quantization_params != NULL); return xnn_setup_fully_connected_nc_qd8_f16_qc8w( opdata->operator_objects[0], input_data, output_data, quantization_params); } case xnn_operator_type_fully_connected_nc_qdu8_f16_qc8w: { const void* quantization_params = input_value->quantization.dynamic_params; assert(kernel_data == NULL); assert(bias_data == NULL); assert(quantization_params != NULL); return xnn_setup_fully_connected_nc_qdu8_f16_qc8w( opdata->operator_objects[0], input_data, output_data, quantization_params); } case xnn_operator_type_fully_connected_nc_qd8_f32_qc8w: { const void* quantization_params = input_value->quantization.dynamic_params; assert(kernel_data == NULL); assert(bias_data == NULL); assert(quantization_params != NULL); return xnn_setup_fully_connected_nc_qd8_f32_qc8w( opdata->operator_objects[0], input_data, output_data, quantization_params); } case xnn_operator_type_fully_connected_nc_qdu8_f32_qc8w: { const void* quantization_params = input_value->quantization.dynamic_params; assert(kernel_data == NULL); assert(bias_data == NULL); assert(quantization_params != NULL); return xnn_setup_fully_connected_nc_qdu8_f32_qc8w( opdata->operator_objects[0], input_data, output_data, quantization_params); } case xnn_operator_type_fully_connected_nc_qp8_f32_qc4w: { assert(kernel_data == NULL); assert(bias_data == NULL); return xnn_setup_fully_connected_nc_qp8_f32_qc4w( opdata->operator_objects[0], input_data, output_data); } case xnn_operator_type_fully_connected_nc_qp8_f32_qb4w: { assert(kernel_data == NULL); assert(bias_data == NULL); return xnn_setup_fully_connected_nc_qp8_f32_qb4w( opdata->operator_objects[0], input_data, output_data); } case xnn_operator_type_fully_connected_nc_qs8: assert(kernel_data == NULL); assert(bias_data == NULL); return xnn_setup_fully_connected_nc_qs8(opdata->operator_objects[0], input_data, output_data); case xnn_operator_type_fully_connected_nc_qs8_qc8w: assert(kernel_data == NULL); assert(bias_data == NULL); return xnn_setup_fully_connected_nc_qs8_qc8w(opdata->operator_objects[0], input_data, output_data); case xnn_operator_type_fully_connected_nc_qu8: assert(kernel_data == NULL); assert(bias_data == NULL); return xnn_setup_fully_connected_nc_qu8(opdata->operator_objects[0], input_data, output_data); default: XNN_UNREACHABLE; } } static inline bool validate_datatypes_with_bias( enum xnn_datatype input_datatype, enum xnn_datatype kernel_datatype, enum xnn_datatype bias_datatype, enum xnn_datatype output_datatype) { switch (kernel_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_fp16 && bias_datatype == xnn_datatype_fp16 && output_datatype == xnn_datatype_fp16) { return true; } else 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_qcint4: if (input_datatype == xnn_datatype_fp32 && 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_fp32) { return true; } else if (input_datatype == xnn_datatype_qpint8 && 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_qbint4: 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; } else if (input_datatype == xnn_datatype_qpint8 && bias_datatype == xnn_datatype_fp32 && output_datatype == xnn_datatype_fp32) { return true; } break; case xnn_datatype_qcint8: if (input_datatype == xnn_datatype_fp32 && 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_fp32) { return true; } else if (input_datatype == xnn_datatype_qpint8 && 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; } else if (input_datatype == xnn_datatype_qint8 && bias_datatype == xnn_datatype_qcint32 && output_datatype == xnn_datatype_qint8) { 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_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 kernel_datatype, enum xnn_datatype output_datatype) { switch (kernel_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_fp16 && output_datatype == xnn_datatype_fp16) { return true; } else if (input_datatype == xnn_datatype_fp32 && output_datatype == xnn_datatype_fp32) { return true; } break; case xnn_datatype_qcint4: if (input_datatype == xnn_datatype_fp32 && output_datatype == xnn_datatype_fp32) { return true; } else if (input_datatype == xnn_datatype_qdint8 && output_datatype == xnn_datatype_fp32) { return true; } else if (input_datatype == xnn_datatype_qpint8 && 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_qbint4: 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; } else if (input_datatype == xnn_datatype_qpint8 && output_datatype == xnn_datatype_fp32) { return true; } break; case xnn_datatype_qcint8: if (input_datatype == xnn_datatype_fp32 && output_datatype == xnn_datatype_fp32) { return true; } else if (input_datatype == xnn_datatype_qdint8 && output_datatype == xnn_datatype_fp32) { return true; } else if (input_datatype == xnn_datatype_qpint8 && output_datatype == xnn_datatype_fp32) { return true; } else if (input_datatype == xnn_datatype_qdint8 && output_datatype == xnn_datatype_fp16) { return true; } else if (input_datatype == xnn_datatype_qint8 && output_datatype == xnn_datatype_qint8) { return true; } break; case xnn_datatype_qint8: if (input_datatype == xnn_datatype_qint8 && output_datatype == xnn_datatype_qint8) { 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_fully_connected(xnn_subgraph_t subgraph, 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_fully_connected)) != xnn_status_success) { return status; } status = xnn_subgraph_check_output_min_max(xnn_node_type_fully_connected, output_min, output_max); if (status != xnn_status_success) { return status; } if ((status = xnn_subgraph_check_input_node_id( xnn_node_type_fully_connected, 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_fully_connected, 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: case xnn_datatype_qpint8: 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_fully_connected), 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_fully_connected), 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_fully_connected), filter_id); return xnn_status_invalid_parameter; } const struct xnn_value* kernel_value = &subgraph->values[filter_id]; if (kernel_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_fully_connected), filter_id, kernel_value->type); return xnn_status_invalid_parameter; } // Non-static kernel is supported, but only for some data types switch (kernel_value->datatype) { case xnn_datatype_fp32: break; // non-static kernel is supported default: if (kernel_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_fully_connected), filter_id); return xnn_status_invalid_parameter; } break; } // Non-static kernel is supported, but only for some data types switch (kernel_value->datatype) { case xnn_datatype_fp16: case xnn_datatype_fp32: break; case xnn_datatype_qbint4: case xnn_datatype_qcint4: if (kernel_value->quantization.zero_point != 8 && kernel_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, must be equals to 8 (unsigned " "weights) or 0 (signed weights) ", xnn_node_type_to_string(xnn_node_type_fully_connected), filter_id, kernel_value->quantization.zero_point, xnn_datatype_to_string(kernel_value->datatype)); return xnn_status_invalid_parameter; } break; case xnn_datatype_qcint8: break; case xnn_datatype_qint8: if (kernel_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_fully_connected), filter_id, kernel_value->quantization.zero_point, xnn_datatype_to_string(kernel_value->datatype)); } 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_fully_connected), filter_id, xnn_datatype_to_string(kernel_value->datatype), kernel_value->datatype); return xnn_status_invalid_parameter; } const bool is_channelwise_quantized = kernel_value->datatype == xnn_datatype_qcint8 || kernel_value->datatype == xnn_datatype_qcint4; if (is_channelwise_quantized) { const size_t output_channels_dim = ((flags & XNN_FLAG_TRANSPOSE_WEIGHTS) != 0) ? 1 : 0; if (kernel_value->quantization.channel_dimension != output_channels_dim) { xnn_log_error("failed to define %s operator with filter ID #%" PRIu32 ": invalid channel dimension %zu", xnn_node_type_to_string(xnn_node_type_fully_connected), input_id, kernel_value->quantization.channel_dimension); return xnn_status_invalid_parameter; } } const bool is_blockwise_quantized = kernel_value->datatype == xnn_datatype_qbint4; if (is_blockwise_quantized) { // TODO: Unsupported features assert((flags & XNN_FLAG_TRANSPOSE_WEIGHTS) == 0); const size_t input_channels_dim = ((flags & XNN_FLAG_TRANSPOSE_WEIGHTS) != 0) ? 0 : 1; const size_t output_channels_dim = ((flags & XNN_FLAG_TRANSPOSE_WEIGHTS) != 0) ? 1 : 0; if (kernel_value->quantization.channel_dimension_blockwise != output_channels_dim) { xnn_log_error("failed to define %s operator with filter ID #%" PRIu32 ": invalid channel dimension %zu", xnn_node_type_to_string(xnn_node_type_fully_connected), input_id, kernel_value->quantization.channel_dimension_blockwise); return xnn_status_invalid_parameter; } const size_t input_channels = kernel_value->shape.dim[input_channels_dim]; if (input_channels % kernel_value->quantization.block_size) { xnn_log_error("failed to define %s operator with filter ID #%" PRIu32 ": invalid block size %zu, input_channels %zu", xnn_node_type_to_string(xnn_node_type_fully_connected), input_id, kernel_value->quantization.block_size, input_channels); 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_fully_connected), 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_fully_connected), bias_id, bias_value->type); return xnn_status_invalid_parameter; } // Non-static bias is supported, but only for some data types switch (bias_value->datatype) { case xnn_datatype_fp32: if (is_channelwise_quantized && 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_fully_connected), bias_id); return xnn_status_invalid_parameter; } break; // non-static bias is supported default: 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_fully_connected), bias_id); return xnn_status_invalid_parameter; } break; } 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_fully_connected), 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_fully_connected, 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_fully_connected, output_id, output_value); if (status != xnn_status_success) { return status; } 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_fully_connected), 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, kernel_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_fully_connected), input_id, filter_id, bias_id, output_id, xnn_datatype_to_string(input_value->datatype), xnn_datatype_to_string(kernel_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, kernel_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_fully_connected), input_id, filter_id, output_id, xnn_datatype_to_string(input_value->datatype), xnn_datatype_to_string(kernel_value->datatype), xnn_datatype_to_string(output_value->datatype)); 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 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_fully_connected; 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_fully_connected_operator; node->reshape = reshape_fully_connected_operator; node->setup = setup_fully_connected_operator; return xnn_status_success; }