src/operators/unpooling-nhwc.c

// 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 <stddef.h>
#include <stdint.h>
#include <stdlib.h>
#include <string.h>

#include "include/xnnpack.h"
#include "src/xnnpack/allocator.h"
#include "src/xnnpack/common.h"
#include "src/xnnpack/compute.h"
#include "src/xnnpack/config-types.h"
#include "src/xnnpack/config.h"
#include "src/xnnpack/indirection.h"
#include "src/xnnpack/log.h"
#include "src/xnnpack/math.h"
#include "src/xnnpack/operator-type.h"
#include "src/xnnpack/operator-utils.h"
#include "src/xnnpack/operator.h"
#include "src/xnnpack/params.h"
#include <pthreadpool.h>

enum xnn_status xnn_create_unpooling2d_nhwc_x32(
    uint32_t input_padding_top,
    uint32_t input_padding_right,
    uint32_t input_padding_bottom,
    uint32_t input_padding_left,
    uint32_t pooling_height,
    uint32_t pooling_width,
    uint32_t flags,
    xnn_operator_t* unpooling_op_out)
{
  xnn_operator_t unpooling_op = NULL;
  enum xnn_status status = xnn_status_uninitialized;

  if ((xnn_params.init_flags & XNN_INIT_FLAG_XNNPACK) == 0) {
    xnn_log_error("failed to create %s operator: XNNPACK is not initialized",
      xnn_operator_type_to_string(xnn_operator_type_unpooling_nhwc_x32));
    goto error;
  }

  status = xnn_status_invalid_parameter;

  const uint32_t pooling_size = pooling_height * pooling_width;
  if (pooling_size == 0) {
    xnn_log_error(
        "failed to create %s operator with %" PRIu32 "x%" PRIu32
        " pooling size: pooling size dimensions must be non-zero",
        xnn_operator_type_to_string(xnn_operator_type_unpooling_nhwc_x32),
        pooling_width, pooling_height);
    goto error;
  }

  status = xnn_status_out_of_memory;

  unpooling_op = xnn_allocate_zero_simd_memory(sizeof(struct xnn_operator));
  if (unpooling_op == NULL) {
    xnn_log_error(
      "failed to allocate %zu bytes for %s operator descriptor",
      sizeof(struct xnn_operator), xnn_operator_type_to_string(xnn_operator_type_unpooling_nhwc_x32));
    goto error;
  }
  unpooling_op->compute = xnn_allocate_zero_memory(sizeof(struct compute_parameters));
  if (unpooling_op->compute == NULL) {
    xnn_log_error("failed to allocate %zu bytes for %s operator descriptor",
                  sizeof(struct compute_parameters),
                  xnn_operator_type_to_string(xnn_operator_type_unpooling_nhwc_x32));
    goto error;
  }
  unpooling_op->num_compute_invocations = 1;
  unpooling_op->convolution_op = xnn_allocate_zero_memory(sizeof(struct xnn_convolution_operator));
  if (unpooling_op->convolution_op == NULL) {
    xnn_log_error("failed to allocate %zu bytes for %s operator descriptor",
                  sizeof(struct xnn_convolution_operator),
                  xnn_operator_type_to_string(xnn_operator_type_unpooling_nhwc_x32));
    goto error;
  }

  const struct xnn_unpool_config* unpool_config = xnn_init_x32_unpool_config();
  if (unpool_config == NULL) {
    xnn_log_error(
      "failed to create %s operator: unsupported hardware configuration",
      xnn_operator_type_to_string(xnn_operator_type_unpooling_nhwc_x32));
    return xnn_status_unsupported_hardware;
  }

  unpooling_op->convolution_op->padding_top = input_padding_top;
  unpooling_op->convolution_op->padding_right = input_padding_right;
  unpooling_op->convolution_op->padding_bottom = input_padding_bottom;
  unpooling_op->convolution_op->padding_left = input_padding_left;

  unpooling_op->convolution_op->kernel_height = pooling_height;
  unpooling_op->convolution_op->kernel_width = pooling_width;

  unpooling_op->type = xnn_operator_type_unpooling_nhwc_x32;
  unpooling_op->flags = flags;
  unpooling_op->unpool_config = unpool_config;

  unpooling_op->state = xnn_run_state_invalid;

  *unpooling_op_out = unpooling_op;
  return xnn_status_success;

error:
  xnn_delete_operator(unpooling_op);
  return status;
}

enum xnn_status xnn_reshape_unpooling2d_nhwc_x32(
  xnn_operator_t unpooling_op,
  size_t batch_size,
  size_t input_height,
  size_t input_width,
  size_t channels,
  size_t input_pixel_stride,
  size_t output_pixel_stride,
  size_t* output_height_out,
  size_t* output_width_out,
  pthreadpool_t threadpool)
{
  if (unpooling_op->type != xnn_operator_type_unpooling_nhwc_x32) {
    xnn_log_error(
        "failed to reshape operator: operator type mismatch (expected %s, got "
        "%s)",
        xnn_operator_type_to_string(xnn_operator_type_unpooling_nhwc_x32),
        xnn_operator_type_to_string_v2(unpooling_op));
    return xnn_status_invalid_parameter;
  }
  unpooling_op->state = xnn_run_state_invalid;

  if ((xnn_params.init_flags & XNN_INIT_FLAG_XNNPACK) == 0) {
    xnn_log_error("failed to reshape %s operator: XNNPACK is not initialized",
      xnn_operator_type_to_string(xnn_operator_type_unpooling_nhwc_x32));
    return xnn_status_uninitialized;
  }

  if (input_width == 0 || input_height == 0) {
    xnn_log_error(
      "failed to reshape %s operator with %zux%zu input: input dimensions must be non-zero",
      xnn_operator_type_to_string(xnn_operator_type_unpooling_nhwc_x32), input_width, input_height);
    return xnn_status_invalid_parameter;
  }

  if (batch_size == 0) {
    unpooling_op->state = xnn_run_state_skip;
    return xnn_status_success;
  }

  if (channels == 0) {
    xnn_log_error(
      "failed to reshape %s operator with %zu channels: number of channels must be non-zero",
      xnn_operator_type_to_string(xnn_operator_type_unpooling_nhwc_x32), channels);
    return xnn_status_invalid_parameter;
  }

  if (input_pixel_stride < channels) {
    xnn_log_error(
        "failed to reshape %s operator with input pixel stride of %zu: stride "
        "must be at least as large as the number of channels (%zu)",
        xnn_operator_type_to_string(xnn_operator_type_unpooling_nhwc_x32),
        input_pixel_stride, channels);
    return xnn_status_invalid_parameter;
  }

  if (output_pixel_stride < channels) {
    xnn_log_error(
        "failed to reshape %s operator with output pixel stride of %zu: stride "
        "must be at least as large as the number of channels (%zu)",
        xnn_operator_type_to_string(xnn_operator_type_unpooling_nhwc_x32),
        output_pixel_stride, channels);
    return xnn_status_invalid_parameter;
  }

  unpooling_op->batch_size = batch_size;
  unpooling_op->input_pixel_stride = input_pixel_stride;
  unpooling_op->output_pixel_stride = output_pixel_stride;
  unpooling_op->convolution_op->input_height = input_height;
  unpooling_op->convolution_op->input_width = input_width;
  unpooling_op->channels = channels;

  unpooling_op->convolution_op->output_height = xnn_compute_unpooling_output_dimension(
    input_height, unpooling_op->convolution_op->padding_top + unpooling_op->convolution_op->padding_bottom,
    unpooling_op->convolution_op->kernel_height);
  unpooling_op->convolution_op->output_width = xnn_compute_unpooling_output_dimension(
    input_width, unpooling_op->convolution_op->padding_left + unpooling_op->convolution_op->padding_right,
    unpooling_op->convolution_op->kernel_width);

  if (output_height_out != NULL) {
    *output_height_out = unpooling_op->convolution_op->output_height;
  }
  if (output_width_out != NULL) {
    *output_width_out = unpooling_op->convolution_op->output_width;
  }

  // Dummy output for initializing indirection buffers. Output needs to be earlier output due to valid_batch_size
  // optimization, where the smaller batch sizes are not re-initialized if we setup with different output.
  unpooling_op->convolution_op->output = unpooling_op->convolution_op->last_output;

  size_t valid_batch_size = 0;
  if (input_height == unpooling_op->convolution_op->last_input_height &&
      input_width == unpooling_op->convolution_op->last_input_width)
  {
    valid_batch_size = unpooling_op->convolution_op->valid_batch_size;
    if (batch_size <= valid_batch_size) {
      unpooling_op->compute[0].range[0] = batch_size * input_height;
      unpooling_op->state = xnn_run_state_needs_setup;
      return xnn_status_success;
    }
  }

  const size_t pooling_height = unpooling_op->convolution_op->kernel_height;
  const size_t pooling_width = unpooling_op->convolution_op->kernel_width;
  const size_t pooling_size = pooling_height * pooling_width;

  const size_t indirection_buffer_size = sizeof(void*) * (batch_size * input_height * input_width * pooling_size);
  const void** indirection_buffer = (const void**) xnn_reallocate_memory(unpooling_op->convolution_op->indirection_buffer, indirection_buffer_size);
  if (indirection_buffer == NULL) {
    xnn_log_error(
      "failed to allocate %zu bytes for %s operator indirection buffer",
      indirection_buffer_size, xnn_operator_type_to_string(xnn_operator_type_unpooling_nhwc_x32));
    return xnn_status_out_of_memory;
  }
  unpooling_op->convolution_op->indirection_buffer = indirection_buffer;
  xnn_log_debug("allocated %zu bytes for indirection buffer in %s operator",
    indirection_buffer_size, xnn_operator_type_to_string(xnn_operator_type_unpooling_nhwc_x32));

  xnn_indirection_init_unpool2d(
    unpooling_op->convolution_op->indirection_buffer,
    unpooling_op->convolution_op->output,
    unpooling_op->output_pixel_stride << XNN_LOG2_SIZEOF_FLOAT,
    unpooling_op->batch_size,
    unpooling_op->convolution_op->input_height,
    unpooling_op->convolution_op->input_width,
    unpooling_op->convolution_op->output_height,
    unpooling_op->convolution_op->output_width,
    unpooling_op->convolution_op->kernel_height,
    unpooling_op->convolution_op->kernel_width,
    unpooling_op->convolution_op->padding_top,
    unpooling_op->convolution_op->padding_left,
    valid_batch_size);

  const size_t input_pixel_stride_in_bytes = unpooling_op->input_pixel_stride * sizeof(float);
  unpooling_op->context.unpooling = (struct unpooling_context) {
    .input_height_stride = input_width * input_pixel_stride_in_bytes,
    .input_width_stride = input_pixel_stride_in_bytes,
    .index_height_stride = input_width * channels * sizeof(uint32_t),
    .index_width_stride = channels * sizeof(uint32_t),
    .indirect_output = indirection_buffer,
    .indirect_output_height_stride = input_width * pooling_size * sizeof(void*),
    .indirect_output_width_stride = pooling_size * sizeof(void*),
    .pooling_size = pooling_size,
    .channels = channels,
    .fill_value = 0,
    .ukernel = unpooling_op->unpool_config->unpool,
  };
  unpooling_op->compute[0].type = xnn_parallelization_type_2d;
  unpooling_op->compute[0].task_2d = (pthreadpool_task_2d_t) xnn_compute_unpooling;
  unpooling_op->compute[0].range[0] = batch_size * input_height;
  unpooling_op->compute[0].range[1] = input_width;
  unpooling_op->state = xnn_run_state_needs_setup;

  unpooling_op->convolution_op->last_input_height = input_height;
  unpooling_op->convolution_op->last_input_width = input_width;
  unpooling_op->convolution_op->valid_batch_size = max(valid_batch_size, batch_size);

  return xnn_status_success;
}

enum xnn_status xnn_setup_unpooling2d_nhwc_x32(
    xnn_operator_t unpooling_op,
    const void* input,
    const uint32_t* index,
    void* output)
{
  if (unpooling_op->type != xnn_operator_type_unpooling_nhwc_x32) {
    xnn_log_error(
        "failed to setup operator: operator type mismatch (expected %s, got "
        "%s)",
        xnn_operator_type_to_string(xnn_operator_type_unpooling_nhwc_x32),
        xnn_operator_type_to_string_v2(unpooling_op));
    return xnn_status_invalid_parameter;
  }

  switch (unpooling_op->state) {
    case xnn_run_state_skip:
      return xnn_status_success;
    case xnn_run_state_invalid:
      xnn_log_error(
          "failed to setup %s operator: operator has not been reshaped yet",
          xnn_operator_type_to_string_v2(unpooling_op));
      return xnn_status_invalid_state;
    case xnn_run_state_needs_setup:
      // Operator has been reshaped, but not setup, continue with setup.
    case xnn_run_state_ready:
      // Operator has been reshaped, and we are setting up with different pointers.
      break;
  }

  const size_t pooling_height = unpooling_op->convolution_op->kernel_height;
  const size_t pooling_width = unpooling_op->convolution_op->kernel_width;
  const size_t pooling_size = pooling_height * pooling_width;
  const size_t batch_size = unpooling_op->convolution_op->valid_batch_size;
  const size_t input_height = unpooling_op->convolution_op->input_height;
  const size_t input_width = unpooling_op->convolution_op->input_width;

  const size_t indirection_buffer_num_elements = batch_size * input_height * input_width * pooling_size;
  for (size_t i = 0; i < indirection_buffer_num_elements; i++) {
    unpooling_op->context.unpooling.indirect_output[i] =
      (void*) ((uintptr_t) unpooling_op->context.unpooling.indirect_output[i] +
               ((uintptr_t) output - (uintptr_t) unpooling_op->convolution_op->last_output));
  }

  unpooling_op->context.unpooling.input = input;
  unpooling_op->context.unpooling.index = index;

  unpooling_op->state = xnn_run_state_ready;

  unpooling_op->convolution_op->last_output = output;

  return xnn_status_success;
}