indirection.h | Explore in Territory

// 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.

#pragma once

#include <stddef.h>
#include <stdint.h>

#include "xnnpack.h"
#include "xnnpack/common.h"

#ifdef __cplusplus
extern "C" {
#endif

XNN_INTERNAL void xnn_indirection_init_conv2d(
  size_t output_tile_size,
  size_t output_start,
  size_t output_end,
  const void** indirection_buffer,
  const void* input,
  const void* zero_buffer,
  size_t input_pixel_stride,
  size_t input_height,
  size_t input_width,
  size_t output_height,
  size_t output_width,
  size_t kernel_height,
  size_t kernel_width,
  size_t stride_height,
  size_t stride_width,
  size_t dilation_height,
  size_t dilation_width,
  size_t input_padding_top,
  size_t input_padding_left);

// Initialize compressed indirection buffers.
// Original indirection buffers has a row of buffer for each row of input. Compressed indirection buffers compress rows
// of input pointers that point to valid elements in the input (not padding). In this section of the indirection buffer,
// all input pointers in row n+1 are a constant K offset away from the input pointers in row n.
//
// Compressed indirection buffers are made up of 3 section:
// - Top
// - Middle
// - Bottom
//
// Top has as many rows of input pointers as padding_top / stride_height (rounded up): indirect_top_height.
// Middle is 1 row of input pointer (this is where the compression is).
// Bottom has as many rows of input pointers as padding_bottom / stride_height (rounded up): indirect_bot_height.
// (Note: padding left and right does not affect compression.)
XNN_INTERNAL void xnn_indirection_init_dwconv2d_compressed(
  size_t output_y_start,
  size_t output_y_end,
  const void** indirection_buffer,
  const void* input,
  size_t input_pixel_stride,
  const void* zero_buffer,
  size_t input_height,
  size_t input_width,
  size_t output_height,
  size_t output_width,
  size_t kernel_height,
  size_t kernel_width,
  size_t stride_height,
  size_t stride_width,
  size_t dilation_height,
  size_t dilation_width,
  size_t input_padding_top,
  size_t input_padding_left,
  size_t step_height,
  size_t step_width,
  size_t indirect_top_height,
  size_t indirect_bot_height,
  size_t primary_tile);

XNN_INTERNAL void xnn_indirection_init_dwconv2d(
  size_t output_y_start,
  size_t output_y_end,
  const void** indirection_buffer,
  const void* input,
  size_t input_pixel_stride,
  const void* zero_buffer,
  size_t input_height,
  size_t input_width,
  size_t output_height,
  size_t output_width,
  size_t kernel_height,
  size_t kernel_width,
  size_t stride_height,
  size_t stride_width,
  size_t dilation_height,
  size_t dilation_width,
  size_t input_padding_top,
  size_t input_padding_left,
  size_t step_height,
  size_t step_width,
  size_t primary_tile);

XNN_INTERNAL void xnn_indirection_init_deconv2d(
  xnn_operator_t op,
  size_t output_tile_size,
  uint32_t log2_element_size);

XNN_INTERNAL void xnn_indirection_init_subconv2d(
  xnn_operator_t op,
  size_t output_tile_size,
  uint32_t log2_element_size);

XNN_INTERNAL void xnn_indirection_init_maxpool2d(
  xnn_operator_t op,
  size_t step_height,
  size_t step_width,
  uint32_t log2_element_size);

XNN_INTERNAL void xnn_indirection_init_resize_bilinear2d_hwc_f16(
  size_t output_y_start,
  size_t output_y_end,
  size_t input_pixel_stride,
  size_t input_height,
  size_t input_width,
  size_t output_height,
  size_t output_width,
  const void* input,
  const void** indirection_buffer,
  void* packed_weights,
  bool align_corners,
  bool tensorflow_legacy);

XNN_INTERNAL void xnn_indirection_init_resize_bilinear2d_hwc_f32(
  size_t output_y_start,
  size_t output_y_end,
  size_t input_pixel_stride,
  size_t input_height,
  size_t input_width,
  size_t output_height,
  size_t output_width,
  const void* input,
  const void** indirection_buffer,
  float* packed_weights,
  bool align_corners,
  bool tensorflow_legacy);

XNN_INTERNAL void xnn_indirection_init_resize_bilinear2d_hwc_q11(
  size_t output_y_start,
  size_t output_y_end,
  size_t input_pixel_stride,
  size_t input_height,
  size_t input_width,
  size_t output_height,
  size_t output_width,
  const void* input,
  const void** indirection_buffer,
  int16_t* packed_weights,
  bool align_corners,
  bool tensorflow_legacy);

xnn_indirection_init_resize_bilinear2d_chw_fn;

XNN_INTERNAL void xnn_indirection_init_resize_bilinear2d_chw_f16(
  size_t input_pixel_stride,
  size_t input_height,
  size_t input_width,
  size_t output_height,
  size_t output_width,
  const void* input,
  const void** indirection_buffer,
  void* packed_weights,
  bool align_corners,
  bool tensorflow_legacy);

XNN_INTERNAL void xnn_indirection_init_resize_bilinear2d_chw_f32(
  size_t input_pixel_stride,
  size_t input_height,
  size_t input_width,
  size_t output_height,
  size_t output_width,
  const void* input,
  const void** indirection_buffer,
  float* packed_weights,
  bool align_corners,
  bool tensorflow_legacy);

XNN_INTERNAL void xnn_indirection_init_unpool2d(
  xnn_operator_t op,
  size_t batch_start,
  uint32_t log2_element_size);

xnn_indirection_init_pavgpool2d_fn;

XNN_INTERNAL void xnn_indirection_init_pavgpool2d_f16(
  size_t input_height,
  size_t input_width,
  size_t output_height,
  size_t output_width,
  size_t pooling_height,
  size_t pooling_width,
  size_t stride_height,
  size_t stride_width,
  size_t padding_top,
  size_t padding_left,
  uint16_t* pixelwise_buffer);

XNN_INTERNAL void xnn_indirection_init_pavgpool2d_f32(
  size_t input_height,
  size_t input_width,
  size_t output_height,
  size_t output_width,
  size_t pooling_height,
  size_t pooling_width,
  size_t stride_height,
  size_t stride_width,
  size_t padding_top,
  size_t padding_left,
  float* pixelwise_buffer);

#ifdef __cplusplus
}  // extern "C"
#endif
chromium/third_party/xnnpack/src/src/xnnpack/indirection.h
