chromium/third_party/xnnpack/src/src/operators/deconvolution-nhwc.c

// Copyright (c) Facebook, Inc. and its affiliates.
// All rights reserved.
//
// Copyright 2019 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 <assert.h>
#include <inttypes.h>
#include <math.h>
#include <stddef.h>
#include <stdint.h>
#include <string.h>

#include <fp16/fp16.h>
#include "xnnpack.h"
#include "xnnpack/allocator.h"
#include "xnnpack/cache.h"
#include "xnnpack/common.h"
#include "xnnpack/compute.h"
#include "xnnpack/config.h"
#include "xnnpack/indirection.h"
#include "xnnpack/log.h"
#include "xnnpack/math.h"
#include "xnnpack/microfnptr.h"
#include "xnnpack/microkernel-type.h"
#include "xnnpack/microparams-init.h"
#include "xnnpack/microparams.h"
#include "xnnpack/operator-type.h"
#include "xnnpack/operator-utils.h"
#include "xnnpack/operator.h"
#include "xnnpack/pack.h"
#include "xnnpack/params.h"
#include "pthreadpool.h"

#ifndef XNN_ENABLE_GEMM_M_SPECIALIZATION
#error "XNN_ENABLE_GEMM_M_SPECIALIZATION is not defined"
#endif

static enum xnn_status create_deconvolution2d_nhwc(
    uint32_t output_padding_top,
    uint32_t output_padding_right,
    uint32_t output_padding_bottom,
    uint32_t output_padding_left,
    uint32_t kernel_height,
    uint32_t kernel_width,
    uint32_t stride_height,
    uint32_t stride_width,
    uint32_t dilation_height,
    uint32_t dilation_width,
    uint32_t groups,
    size_t group_input_channels,
    size_t group_output_channels,
    size_t input_pixel_stride,
    size_t output_pixel_stride,
    const void* kernel,
    const void* bias,
    uint32_t flags,
    uint32_t log2_input_element_size,
    uint32_t log2_filter_element_size,
    uint32_t bias_element_size,
    xnn_pack_conv_goki_w_fn pack_conv_goki_w,
    xnn_pack_deconv_goki_w_fn pack_deconv_goki_w,
    const void* packing_params,
    int input_padding_byte,
    int packed_weights_padding_byte,
    size_t extra_weights_bytes,
    xnn_init_qs8_qc8w_scale_params_fn init_scale_params,
    const float* scale_params,
    xnn_init_qs8_qc8w_scale_params_fn init_kernel_scale_params,
    const float* kernel_scale_params,
    const void* params,
    size_t params_size,
    const struct jit_gemm_params* jit_gemm_params,
    const struct xnn_gemm_config* gemm_config,
    const struct gemm_fused_ukernels* gemm_ukernels,
    enum xnn_operator_type operator_type,
    bool dynamic_quantization,
    xnn_code_cache_t code_cache,
    xnn_weights_cache_t weights_cache,
    xnn_operator_t* deconvolution_op_out)
{ … }

enum xnn_status create_deconvolution2d_nhwc_qs8_qc8w(
    enum xnn_operator_type operator_type,
    uint32_t output_padding_top,
    uint32_t output_padding_right,
    uint32_t output_padding_bottom,
    uint32_t output_padding_left,
    uint32_t kernel_height,
    uint32_t kernel_width,
    uint32_t stride_height,
    uint32_t stride_width,
    uint32_t dilation_height,
    uint32_t dilation_width,
    uint32_t groups,
    size_t group_input_channels,
    size_t group_output_channels,
    size_t input_pixel_stride,
    size_t output_pixel_stride,
    int8_t input_zero_point,
    float input_scale,
    const float* kernel_scale,
    const int8_t* kernel,
    const int32_t* bias,
    int8_t output_zero_point,
    float output_scale,
    int8_t output_min,
    int8_t output_max,
    uint32_t flags,
    xnn_code_cache_t code_cache,
    xnn_weights_cache_t weights_cache,
    xnn_operator_t* deconvolution_op_out)
{ … }

enum xnn_status xnn_create_deconvolution2d_nhwc_qs8_qc8w(
    uint32_t output_padding_top,
    uint32_t output_padding_right,
    uint32_t output_padding_bottom,
    uint32_t output_padding_left,
    uint32_t kernel_height,
    uint32_t kernel_width,
    uint32_t stride_height,
    uint32_t stride_width,
    uint32_t dilation_height,
    uint32_t dilation_width,
    uint32_t groups,
    size_t group_input_channels,
    size_t group_output_channels,
    size_t input_pixel_stride,
    size_t output_pixel_stride,
    int8_t input_zero_point,
    float input_scale,
    const float* kernel_scale,
    const int8_t* kernel,
    const int32_t* bias,
    int8_t output_zero_point,
    float output_scale,
    int8_t output_min,
    int8_t output_max,
    uint32_t flags,
    xnn_code_cache_t code_cache,
    xnn_weights_cache_t weights_cache,
    xnn_operator_t* deconvolution_op_out)
{ … }

enum xnn_status xnn_create_deconvolution2d_nhwc_qs8(
    uint32_t output_padding_top,
    uint32_t output_padding_right,
    uint32_t output_padding_bottom,
    uint32_t output_padding_left,
    uint32_t kernel_height,
    uint32_t kernel_width,
    uint32_t stride_height,
    uint32_t stride_width,
    uint32_t dilation_height,
    uint32_t dilation_width,
    uint32_t groups,
    size_t group_input_channels,
    size_t group_output_channels,
    size_t input_pixel_stride,
    size_t output_pixel_stride,
    int8_t input_zero_point,
    float input_scale,
    float kernel_scale,
    const int8_t* kernel,
    const int32_t* bias,
    int8_t output_zero_point,
    float output_scale,
    int8_t output_min,
    int8_t output_max,
    uint32_t flags,
    xnn_code_cache_t code_cache,
    xnn_weights_cache_t weights_cache,
    xnn_operator_t* deconvolution_op_out)
{ … }

enum xnn_status xnn_create_deconvolution2d_nhwc_qu8(
    uint32_t output_padding_top,
    uint32_t output_padding_right,
    uint32_t output_padding_bottom,
    uint32_t output_padding_left,
    uint32_t kernel_height,
    uint32_t kernel_width,
    uint32_t stride_height,
    uint32_t stride_width,
    uint32_t dilation_height,
    uint32_t dilation_width,
    uint32_t groups,
    size_t group_input_channels,
    size_t group_output_channels,
    size_t input_pixel_stride,
    size_t output_pixel_stride,
    uint8_t input_zero_point,
    float input_scale,
    uint8_t kernel_zero_point,
    float kernel_scale,
    const uint8_t* kernel,
    const int32_t* bias,
    uint8_t output_zero_point,
    float output_scale,
    uint8_t output_min,
    uint8_t output_max,
    uint32_t flags,
    xnn_code_cache_t code_cache,
    xnn_weights_cache_t weights_cache,
    xnn_operator_t* deconvolution_op_out)
{ … }

enum xnn_status xnn_create_deconvolution2d_nhwc_f16(
    uint32_t output_padding_top,
    uint32_t output_padding_right,
    uint32_t output_padding_bottom,
    uint32_t output_padding_left,
    uint32_t kernel_height,
    uint32_t kernel_width,
    uint32_t stride_height,
    uint32_t stride_width,
    uint32_t dilation_height,
    uint32_t dilation_width,
    uint32_t groups,
    size_t group_input_channels,
    size_t group_output_channels,
    size_t input_pixel_stride,
    size_t output_pixel_stride,
    const void* kernel,
    const void* bias,
    float output_min,
    float output_max,
    uint32_t flags,
    xnn_code_cache_t code_cache,
    xnn_weights_cache_t weights_cache,
    xnn_operator_t* deconvolution_op_out)
{ … }

enum xnn_status xnn_create_deconvolution2d_nhwc_qd8_f32_qc8w(
    uint32_t output_padding_top,
    uint32_t output_padding_right,
    uint32_t output_padding_bottom,
    uint32_t output_padding_left,
    uint32_t kernel_height,
    uint32_t kernel_width,
    uint32_t stride_height,
    uint32_t stride_width,
    uint32_t dilation_height,
    uint32_t dilation_width,
    uint32_t groups,
    size_t group_input_channels,
    size_t group_output_channels,
    size_t input_pixel_stride,
    size_t output_pixel_stride,
    const float* kernel_scale,
    const int8_t* kernel,
    const float* bias,
    float output_min,
    float output_max,
    uint32_t flags,
    xnn_code_cache_t code_cache,
    xnn_weights_cache_t weights_cache,
    xnn_operator_t* deconvolution_op_out)
{ … }

enum xnn_status xnn_create_deconvolution2d_nhwc_f32(
    uint32_t output_padding_top,
    uint32_t output_padding_right,
    uint32_t output_padding_bottom,
    uint32_t output_padding_left,
    uint32_t kernel_height,
    uint32_t kernel_width,
    uint32_t stride_height,
    uint32_t stride_width,
    uint32_t dilation_height,
    uint32_t dilation_width,
    uint32_t groups,
    size_t group_input_channels,
    size_t group_output_channels,
    size_t input_pixel_stride,
    size_t output_pixel_stride,
    const float* kernel,
    const float* bias,
    float output_min,
    float output_max,
    uint32_t flags,
    xnn_code_cache_t code_cache,
    xnn_weights_cache_t weights_cache,
    xnn_operator_t* deconvolution_op_out)
{ … }

static enum xnn_status reshape_conv_path(
  xnn_operator_t deconvolution_op,
  size_t batch_size,
  uint32_t log2_input_element_size,
  uint32_t log2_filter_element_size,
  uint32_t extra_weights_element_size,
  uint32_t log2_output_element_size,
  bool dynamic_quantization,
  const void* params,
  size_t params_size,
  size_t num_threads)
{ … }

static enum xnn_status reshape_subconv2d_path(
  xnn_operator_t deconvolution_op,
  size_t batch_size,
  uint32_t log2_input_element_size,
  uint32_t log2_filter_element_size,
  uint32_t extra_weights_element_size,
  uint32_t log2_output_element_size,
  bool dynamic_quantization,
  const void* params,
  size_t params_size,
  size_t num_threads)
{ … }

static enum xnn_status reshape_deconvolution2d_nhwc(
  xnn_operator_t deconvolution_op,
  size_t batch_size,
  size_t input_height,
  size_t input_width,
  uint32_t adjustment_height,
  uint32_t adjustment_width,
  uint32_t log2_input_element_size,
  uint32_t log2_filter_element_size,
  uint32_t extra_weights_element_size,
  uint32_t log2_output_element_size,
  bool dynamic_quantization,
  const void* params,
  size_t params_size,
  size_t* output_height_out,
  size_t* output_width_out,
  pthreadpool_t threadpool)
{ … }

enum xnn_status xnn_reshape_deconvolution2d_nhwc_qs8(
  xnn_operator_t deconvolution_op,
  size_t batch_size,
  size_t input_height,
  size_t input_width,
  uint32_t adjustment_height,
  uint32_t adjustment_width,
  size_t* output_height_out,
  size_t* output_width_out,
  pthreadpool_t threadpool)
{ … }

enum xnn_status xnn_reshape_deconvolution2d_nhwc_qs8_qc8w(
  xnn_operator_t deconvolution_op,
  size_t batch_size,
  size_t input_height,
  size_t input_width,
  uint32_t adjustment_height,
  uint32_t adjustment_width,
  size_t* output_height_out,
  size_t* output_width_out,
  pthreadpool_t threadpool)
{ … }

enum xnn_status xnn_reshape_deconvolution2d_nhwc_qu8(
  xnn_operator_t deconvolution_op,
  size_t batch_size,
  size_t input_height,
  size_t input_width,
  uint32_t adjustment_height,
  uint32_t adjustment_width,
  size_t* output_height_out,
  size_t* output_width_out,
  pthreadpool_t threadpool)
{ … }

enum xnn_status xnn_reshape_deconvolution2d_nhwc_f16(
  xnn_operator_t deconvolution_op,
  size_t batch_size,
  size_t input_height,
  size_t input_width,
  uint32_t adjustment_height,
  uint32_t adjustment_width,
  size_t* output_height_out,
  size_t* output_width_out,
  pthreadpool_t threadpool)
{ … }

enum xnn_status xnn_reshape_deconvolution2d_nhwc_qd8_f32_qc8w(
  xnn_operator_t deconvolution_op,
  size_t batch_size,
  size_t input_height,
  size_t input_width,
  uint32_t adjustment_height,
  uint32_t adjustment_width,
  size_t* output_height_out,
  size_t* output_width_out,
  pthreadpool_t threadpool)
{ … }

enum xnn_status xnn_reshape_deconvolution2d_nhwc_f32(
  xnn_operator_t deconvolution_op,
  size_t batch_size,
  size_t input_height,
  size_t input_width,
  uint32_t adjustment_height,
  uint32_t adjustment_width,
  size_t* output_height_out,
  size_t* output_width_out,
  pthreadpool_t threadpool)
{ … }

static enum xnn_status setup_conv_path(
  xnn_operator_t deconvolution_op,
  const void* input,
  void* output)
{ … }

static enum xnn_status setup_subconv2d_path(
  xnn_operator_t deconvolution_op,
  const void* input,
  void* output)
{ … }

static enum xnn_status setup_deconvolution2d_nhwc(
  xnn_operator_t deconvolution_op,
  enum xnn_operator_type expected_operator_type,
  const void* input,
  const void* quantization_params,
  void* output)
{ … }

enum xnn_status xnn_setup_deconvolution2d_nhwc_qs8(
    xnn_operator_t deconvolution_op,
    const int8_t* input,
    int8_t* output)
{ … }

enum xnn_status xnn_setup_deconvolution2d_nhwc_qs8_qc8w(
    xnn_operator_t deconvolution_op,
    const int8_t* input,
    int8_t* output)
{ … }

enum xnn_status xnn_setup_deconvolution2d_nhwc_qu8(
    xnn_operator_t deconvolution_op,
    const uint8_t* input,
    uint8_t* output)
{ … }

enum xnn_status xnn_setup_deconvolution2d_nhwc_f16(
    xnn_operator_t deconvolution_op,
    const void* input,
    void* output)
{ … }

enum xnn_status xnn_setup_deconvolution2d_nhwc_qd8_f32_qc8w(
    xnn_operator_t deconvolution_op,
    const int8_t* input,
    float* output,
    const struct xnn_dynamic_quantization_params* quantization_params)
{ … }

enum xnn_status xnn_setup_deconvolution2d_nhwc_f32(
    xnn_operator_t deconvolution_op,
    const float* input,
    float* output)
{ … }