path: root/src/operators/convolution-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 <math.h>
#include <stdbool.h>
#include <stddef.h>
#include <stdint.h>
#include <stdlib.h>
#include <string.h>

#include <fp16.h>

#include <xnnpack.h>
#include <xnnpack/allocator.h>
#include <xnnpack/cache.h>
#include <xnnpack/common.h>
#include <xnnpack/compute.h>
#include <xnnpack/indirection.h>
#include <xnnpack/log.h>
#include <xnnpack/math.h>
#include <xnnpack/operator.h>
#include <xnnpack/pack.h>
#include <xnnpack/params.h>
#include <xnnpack/post-operation.h>
#include <xnnpack/microparams-init.h>

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

static inline size_t compute_output_dimension_with_tf_same_padding(
    size_t input_dimension,
    size_t subsampling_dimension)
{
  return divide_round_up(input_dimension, subsampling_dimension);
}

static inline const struct dwconv_parameters* find_dwconv_ukernel(
    size_t kernel_size,
    const struct dwconv_parameters* ukernel,
    size_t num_ukernels)
{
  while (num_ukernels-- != 0) {
    if (ukernel->primary_tile == kernel_size) {
      return ukernel;
    }
    ukernel++;
  }
  return NULL;
}

#if XNN_PLATFORM_JIT
static inline uintptr_t cached_code_at_offset(xnn_operator_t op, size_t offset)
{
  return (uintptr_t)op->code_cache->cache.code.start + offset;
}

static size_t get_generated_gemm(
    struct xnn_hmp_gemm_codegen generators,
    struct jit_gemm_params *jit_gemm_params,
    size_t mr,
    size_t group_output_channels,
    size_t nr,
    size_t group_input_channels,
    size_t log2_input_element_size,
    struct xnn_code_cache* code_cache)
{
  size_t offset = XNN_CACHE_NOT_FOUND;
  xnn_jit_gemm_code_generator_function generator = generators.function[XNN_UARCH_DEFAULT];
  if (generator == NULL) {
    goto error;
  }

  enum xnn_status status = xnn_status_success;

  status = xnn_reserve_code_memory(&code_cache->cache.code, XNN_DEFAULT_MICROKERNEL_SIZE);
  if (xnn_status_success != status) {
    xnn_log_error("failed to ensure sufficient space in the code buffer for a microkernel");
    goto error;
  }

  const size_t old_size = code_cache->cache.code.size;
  void* old_code = (uint8_t*) code_cache->cache.code.start + old_size;
  status = generator(&code_cache->cache.code, mr, group_output_channels % nr,
                     group_input_channels << log2_input_element_size,
                     jit_gemm_params);

  if (xnn_status_success != status) {
    xnn_log_error("failed to generate GEMM microkernel");
    goto error;
  }

  const size_t new_size = code_cache->cache.code.size;
  return xnn_get_or_insert_code_cache(code_cache, old_code, new_size - old_size);

error:
  return offset;
}

static void generate_gemms_up_to_max_mr(
    size_t max_mr,
    struct gemm_codegens generators,
    struct jit_gemm_params *jit_gemm_params,
    size_t group_output_channels,
    size_t nr,
    size_t group_input_channels,
    size_t log2_input_element_size,
    xnn_operator_t convolution_op)
{
  assert(XNN_MAX_MR >= max_mr);
  if (convolution_op->code_cache == NULL) {
    return;
  }
  convolution_op->ukernel.gemm.gemm_cases[0].generated_code_offset[XNN_UARCH_DEFAULT] =
      get_generated_gemm(generators.gemm1, jit_gemm_params, 1, group_output_channels, nr, group_input_channels,
                         log2_input_element_size, convolution_op->code_cache);
  for (size_t mr = 2; mr <= max_mr; mr++) {
    convolution_op->ukernel.gemm.gemm_cases[mr - 1].generated_code_offset[XNN_UARCH_DEFAULT] =
        get_generated_gemm(generators.gemm, jit_gemm_params, mr, group_output_channels, nr, group_input_channels,
                           log2_input_element_size, convolution_op->code_cache);
  }
}

static size_t get_generated_igemm(
    struct xnn_hmp_igemm_codegen generators,
    struct jit_gemm_params *jit_gemm_params,
    size_t group_output_channels,
    size_t nr,
    size_t group_input_channels,
    size_t log2_input_element_size,
    size_t kernel_size,
    size_t mr,
    struct xnn_code_cache* code_cache)
{
  size_t offset = XNN_CACHE_NOT_FOUND;
  xnn_jit_igemm_code_generator_function generator = generators.function[XNN_UARCH_DEFAULT];
  if (generator == NULL) {
    goto error;
  }
  enum xnn_status status = xnn_status_success;

  status = xnn_reserve_code_memory(&code_cache->cache.code, XNN_DEFAULT_MICROKERNEL_SIZE);
  if (xnn_status_success != status) {
    xnn_log_error("failed to ensure sufficient space in code buffer for microkernel");
    goto error;
  }

  const size_t old_size = code_cache->cache.code.size;
  void* old_code = (uint8_t*) code_cache->cache.code.start + old_size;
  status = generator(&code_cache->cache.code, mr, group_output_channels % nr,
                     group_input_channels << log2_input_element_size,
                     kernel_size * mr * sizeof(void*), jit_gemm_params);
  if (status != xnn_status_success) {
    xnn_log_error("failed to generate IGEMM microkernel");
    goto error;
  }

  const size_t new_size = code_cache->cache.code.size;
  return xnn_get_or_insert_code_cache(code_cache, old_code, new_size - old_size);

error:
  return offset;
}

static void generate_igemms_up_to_max_mr(
    size_t max_mr,
    struct gemm_codegens generators,
    struct jit_gemm_params *jit_gemm_params,
    size_t group_output_channels,
    size_t nr,
    size_t group_input_channels,
    size_t log2_input_element_size,
    size_t kernel_size,
    xnn_operator_t convolution_op)
{
  assert(XNN_MAX_MR >= max_mr);
  if (convolution_op->code_cache == NULL) {
    return;
  }
  convolution_op->ukernel.igemm.igemm_cases[0].generated_code_offset[XNN_UARCH_DEFAULT] =
      get_generated_igemm(generators.igemm1, jit_gemm_params, group_output_channels, nr, group_input_channels,
                          log2_input_element_size, kernel_size, 1, convolution_op->code_cache);
  for (size_t mr = 2; mr <= max_mr; mr++) {
    convolution_op->ukernel.igemm.igemm_cases[mr - 1].generated_code_offset[XNN_UARCH_DEFAULT] =
      get_generated_igemm(generators.igemm, jit_gemm_params, group_output_channels, nr, group_input_channels,
                          log2_input_element_size, kernel_size, mr, convolution_op->code_cache);
  }
}
#endif  // XNN_PLATFORM_JIT

static enum xnn_status create_convolution2d_nhwc(
    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,
    size_t input_channel_stride,
    size_t output_channel_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_vmulcaddc_w_function pack_vmulcaddc_w,
    xnn_pack_dwconv_hwg_w_function pack_dwconv_hwg_w,
    xnn_pack_dwconv_ghw_w_function pack_dwconv_ghw_w,
    xnn_pack_gemm_goi_w_function pack_gemm_goi_w,
    xnn_pack_conv_kgo_w_function pack_conv_kgo_w,
    xnn_pack_conv_goki_w_function pack_conv_goki_w,
    const void* packing_params,
    int input_padding_byte,
    int packed_weights_padding_byte,
    size_t extra_weights_bytes,
    xnn_init_qc8_scale_params_fn init_scale_params,
    const float* scale_params,
    const void* gemm_params,
    size_t gemm_params_size,
    const void* dwconv_params,
    size_t dwconv_params_size,
    const void* vmulcaddc_params,
    size_t vmulcaddc_params_size,
    const struct gemm_parameters* gemm_parameters,
    const struct dwconv_parameters* dwconv_ukernel,
    const struct vmulcaddc_parameters* vmulcaddc_parameters,
    struct jit_gemm_params* jit_gemm_params,
    bool linear_activation,
    bool relu_activation,
    uint32_t datatype_init_flags,
    enum xnn_operator_type operator_type,
    size_t num_post_operations,
    void* post_operation_params,
    xnn_caches_t caches,
    xnn_operator_t* convolution_op_out)
{
  xnn_operator_t convolution_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(operator_type));
    goto error;
  }

  status = xnn_status_unsupported_hardware;

  if ((xnn_params.init_flags & datatype_init_flags) != datatype_init_flags) {
    xnn_log_error(
      "failed to create %s operator: operations on data type are not supported",
      xnn_operator_type_to_string(operator_type));
    goto error;
  }

  status = xnn_status_invalid_parameter;

  if (kernel_width == 0 || kernel_height == 0) {
    xnn_log_error(
      "failed to create %s operator with %" PRIu32 "x%" PRIu32 " kernel: kernel dimensions must be non-zero",
      xnn_operator_type_to_string(operator_type), kernel_width, kernel_height);
    goto error;
  }

  if (subsampling_width == 0 || subsampling_height == 0) {
    xnn_log_error(
      "failed to create %s operator with %" PRIu32 "x%" PRIu32 " subsampling: subsampling dimensions must be non-zero",
      xnn_operator_type_to_string(operator_type), subsampling_width, subsampling_height);
    goto error;
  }

  if (dilation_width == 0 || dilation_height == 0) {
    xnn_log_error(
      "failed to create %s operator with %" PRIu32 "x%" PRIu32 " dilation: dilation dimensions must be non-zero",
      xnn_operator_type_to_string(operator_type), dilation_width, dilation_height);
    goto error;
  }

  if (groups == 0) {
    xnn_log_error(
      "failed to create %s operator with %" PRIu32 " groups: number of groups must be non-zero",
      xnn_operator_type_to_string(operator_type), groups);
    goto error;
  }

  if (group_input_channels == 0) {
    xnn_log_error(
      "failed to create %s operator with %zu input channels per group: number of channels must be non-zero",
      xnn_operator_type_to_string(operator_type), group_input_channels);
    goto error;
  }

  if (group_output_channels == 0) {
    xnn_log_error(
      "failed to create %s operator with %zu output channels per group: number of channels must be non-zero",
      xnn_operator_type_to_string(operator_type), group_output_channels);
    goto error;
  }

  const size_t input_channels = groups * group_input_channels;
  if (input_channel_stride < input_channels) {
    xnn_log_error(
      "failed to create %s operator with input channel stride of %zu: "
      "stride must be at least as large as the number of input channels (%" PRIu32 "x%zu)",
      xnn_operator_type_to_string(operator_type),
      input_channel_stride, groups, group_input_channels);
    goto error;
  }

  const size_t output_channels = groups * group_output_channels;
  if (output_channel_stride < output_channels) {
    xnn_log_error(
      "failed to create %s operator with output channel stride of %zu: "
      "stride must be at least as large as the number of output channels (%" PRIu32 "x%zu)",
      xnn_operator_type_to_string(operator_type),
      output_channel_stride, groups, group_output_channels);
    goto error;
  }

  if ((flags & XNN_FLAG_DEPTHWISE_CONVOLUTION) != 0 && group_input_channels != 1) {
    xnn_log_error(
      "failed to create depthwise %s operator with %zu input channels per group: "
      "depthwise convolution must have exactly 1 input channel per group",
      xnn_operator_type_to_string(operator_type), group_input_channels);
    goto error;
  }

  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) {
    if (any_padding) {
      xnn_log_error(
        "failed to create %s operator with %" PRIu32 "+%" PRIu32 "x%" PRIu32 "+%" PRIu32" padding: "
        "TensorFlow SAME padding can't be combined with explicit padding specification",
        xnn_operator_type_to_string(operator_type),
        input_padding_top, input_padding_left, input_padding_bottom, input_padding_right);
      goto error;
    }
  }

  status = xnn_status_out_of_memory;

  convolution_op = xnn_allocate_zero_simd_memory(sizeof(struct xnn_operator));
  if (convolution_op == NULL) {
    xnn_log_error(
      "failed to allocate %zu bytes for %s operator descriptor",
      sizeof(struct xnn_operator), xnn_operator_type_to_string(operator_type));
    goto error;
  }

  if (caches != NULL) {
    convolution_op->weights_cache = caches->weights_cache;
    convolution_op->code_cache = caches->code_cache;
  }

  const size_t kernel_size = kernel_height * kernel_width;

  enum xnn_ukernel_type ukernel_type = xnn_ukernel_type_default;
  const bool unit_subsampling = (subsampling_width | subsampling_height) == 1;
  if (group_input_channels == 1 && group_output_channels == 1 && kernel_size == 1 && unit_subsampling && !any_padding && vmulcaddc_parameters != NULL) {
    ukernel_type = xnn_ukernel_type_vmulcaddc;
  } else if (group_input_channels == 1 && group_output_channels == 1 && dwconv_ukernel != NULL)
  {
    ukernel_type = xnn_ukernel_type_dwconv;
  } else if (kernel_size == 1 && unit_subsampling && !any_padding) {
    ukernel_type = xnn_ukernel_type_gemm;
  } else {
    ukernel_type = xnn_ukernel_type_igemm;
  }
  assert(ukernel_type != xnn_ukernel_type_default);

  if (num_post_operations != 0 && ukernel_type != xnn_ukernel_type_gemm) {
    xnn_log_error(
        "convolution with post operations not support for these parameters: "
        "kernel_size: %zu unit_subsampling: %d padding: %d",
        kernel_size, unit_subsampling, any_padding);
    goto error;
  }

  size_t zero_size = 0;
  switch (ukernel_type) {
    case xnn_ukernel_type_vmulcaddc:
    {
      assert(vmulcaddc_parameters != NULL);
      assert(vmulcaddc_params != NULL);

      const size_t c_stride = round_up_po2(groups, vmulcaddc_parameters->channel_tile);
      const size_t packed_weights_size = ((UINT32_C(1) << log2_filter_element_size) + bias_element_size) * c_stride;
      size_t aligned_total_weights_size = round_up_po2(packed_weights_size, XNN_ALLOCATION_ALIGNMENT);
      void* weights_ptr = xnn_get_pointer_to_write_weights(
          convolution_op, aligned_total_weights_size, packed_weights_padding_byte);
      if (weights_ptr == NULL) {
        xnn_log_error("failed to reserve or allocated %zu bytes for %s operator vmulcaddc packed weights",
                      aligned_total_weights_size, xnn_operator_type_to_string(operator_type));
        goto error;
      }

      pack_vmulcaddc_w(
        groups, vmulcaddc_parameters->channel_tile,
        kernel, bias, weights_ptr, packing_params);

      if (use_weights_cache(convolution_op)) {
        convolution_op->packed_weights.offset = xnn_get_or_insert_weights_cache(
            convolution_op->weights_cache, weights_ptr, aligned_total_weights_size);
      }

      memcpy(&convolution_op->params, vmulcaddc_params, vmulcaddc_params_size);

      convolution_op->ukernel.vmulcaddc = (struct xnn_ukernel_vmulcaddc) {
        .function = vmulcaddc_parameters->ukernel,
        .mr = vmulcaddc_parameters->row_tile,
      };
      break;
    }
    case xnn_ukernel_type_dwconv:
    {
      assert(dwconv_ukernel != NULL);
      assert(dwconv_ukernel->primary_tile == kernel_size);

      const size_t c_stride = round_up_po2(groups, dwconv_ukernel->channel_tile);
      const size_t packed_weights_size = ((kernel_size << log2_filter_element_size) + bias_element_size + extra_weights_bytes) * c_stride;
      size_t aligned_total_weights_size = round_up_po2(packed_weights_size, XNN_ALLOCATION_ALIGNMENT);
      void* weights_ptr = xnn_get_pointer_to_write_weights(
          convolution_op, aligned_total_weights_size, packed_weights_padding_byte);
      if (weights_ptr == NULL) {
        xnn_log_error("failed to reserve or allocated %zu bytes for %s operator dwconv packed weights",
                      aligned_total_weights_size, xnn_operator_type_to_string(operator_type));
        goto error;
      }
      memcpy(&convolution_op->params, dwconv_params, dwconv_params_size);

      if (flags & XNN_FLAG_DEPTHWISE_CONVOLUTION) {
        pack_dwconv_hwg_w(
          kernel_height, kernel_width,
          groups, dwconv_ukernel->channel_tile,
          kernel, bias, weights_ptr,
          dwconv_ukernel->channel_tile * extra_weights_bytes,
          packing_params);
      } else {
        pack_dwconv_ghw_w(
          kernel_height, kernel_width,
          groups, dwconv_ukernel->channel_tile,
          kernel, bias, weights_ptr,
          dwconv_ukernel->channel_tile * extra_weights_bytes,
          packing_params);
      }

      if (scale_params != NULL) {
        assert(init_scale_params != NULL);

        init_scale_params(
          groups, dwconv_ukernel->channel_tile,
          dwconv_ukernel->channel_tile * ((kernel_size << log2_filter_element_size) + bias_element_size + extra_weights_bytes),
          scale_params,
          (void*) ((uintptr_t) weights_ptr + dwconv_ukernel->channel_tile * ((kernel_size << log2_filter_element_size) + bias_element_size)));
      }

      if (use_weights_cache(convolution_op)) {
        convolution_op->packed_weights.offset = xnn_get_or_insert_weights_cache(
            convolution_op->weights_cache, weights_ptr, aligned_total_weights_size);
      }

      const union dwconv_fused_ukernels* ukernels = &dwconv_ukernel->minmax;
      if (linear_activation && dwconv_ukernel->linear.unipass != NULL) {
        ukernels = &dwconv_ukernel->linear;
      }
      convolution_op->ukernel.dwconv = (struct xnn_ukernel_dwconv) {
        .unipass_function = ukernels->unipass,
        .primary_tile = dwconv_ukernel->primary_tile,
        .incremental_tile = dwconv_ukernel->incremental_tile,
      };

      zero_size = XNN_EXTRA_BYTES + (c_stride << log2_input_element_size);
      break;
    }
    case xnn_ukernel_type_gemm:
    case xnn_ukernel_type_igemm:
    {
      const uint32_t nr = gemm_parameters->nr;
      const uint32_t kr = UINT32_C(1) << gemm_parameters->log2_kr;
      const uint32_t sr = UINT32_C(1) << gemm_parameters->log2_sr;
      const size_t n_stride = round_up(group_output_channels, nr);
      const size_t k_stride = round_up_po2(group_input_channels, kr * sr);

      const size_t packed_group_weights_size = ((kernel_size * k_stride << log2_filter_element_size) + bias_element_size + extra_weights_bytes) * n_stride;
      const size_t aligned_total_weights_size = round_up_po2(packed_group_weights_size * groups, XNN_ALLOCATION_ALIGNMENT);
      void* weights_ptr = xnn_get_pointer_to_write_weights(
        convolution_op, aligned_total_weights_size, packed_weights_padding_byte);
      if (weights_ptr == NULL) {
        xnn_log_error("failed to reserve or allocated %zu bytes for %s operator gemm packed weights",
                      aligned_total_weights_size, xnn_operator_type_to_string(operator_type));
        goto error;
      }
      memcpy(&convolution_op->params, gemm_params, gemm_params_size);
      convolution_op->num_post_operation_params = num_post_operations;
      convolution_op->post_operation_params = post_operation_params;

      const struct gemm_fused_ukernels* gemm_ukernels = &gemm_parameters->minmax;
      const uint32_t mr = gemm_parameters->mr;
      if (linear_activation && gemm_parameters->linear.gemm[mr - 1].function[XNN_UARCH_DEFAULT] != NULL) {
        gemm_ukernels = &gemm_parameters->linear;
      } else if (relu_activation && gemm_parameters->relu.gemm[mr - 1].function[XNN_UARCH_DEFAULT] != NULL) {
        gemm_ukernels = &gemm_parameters->relu;
      }
      switch (ukernel_type) {
        case xnn_ukernel_type_gemm:
            pack_gemm_goi_w(
                groups, group_output_channels, group_input_channels,
                nr, kr, sr,
                kernel, bias, weights_ptr, gemm_parameters->nr * extra_weights_bytes, packing_params);
          convolution_op->ukernel.gemm = (struct xnn_ukernel_gemm) {
            .mr = mr,
            .nr = nr,
            .kr = kr,
            .sr = sr,
          };

          assert(XNN_MAX_MR >= mr);
          for (size_t i = 0; i < mr; i++) {
            convolution_op->ukernel.gemm.gemm_cases[i] = gemm_ukernels->gemm[i];
          }

          #if XNN_PLATFORM_JIT
            generate_gemms_up_to_max_mr(
              mr, gemm_parameters->generator, jit_gemm_params, group_output_channels, nr,
              group_input_channels, log2_input_element_size, convolution_op);
          #endif  // XNN_PLATFORM_JIT

          break;
        case xnn_ukernel_type_igemm:
          if (flags & XNN_FLAG_DEPTHWISE_CONVOLUTION) {
            pack_conv_kgo_w(
              groups, group_output_channels, kernel_size,
              nr, kr, sr,
              kernel, bias, weights_ptr, gemm_parameters->nr * extra_weights_bytes, packing_params);
          } else {
            pack_conv_goki_w(
              groups, group_output_channels, kernel_size, group_input_channels,
              nr, kr, sr,
              kernel, bias, weights_ptr, gemm_parameters->nr * extra_weights_bytes, packing_params);
          }
          convolution_op->ukernel.igemm = (struct xnn_ukernel_igemm) {
            .mr = mr,
            .nr = nr,
            .kr = kr,
            .sr = sr,
          };

          assert(XNN_MAX_MR >= mr);
          for (size_t i = 0; i < mr; i++) {
            convolution_op->ukernel.igemm.igemm_cases[i] = gemm_ukernels->igemm[i];
          }

          #if XNN_PLATFORM_JIT
            generate_igemms_up_to_max_mr(
              mr, gemm_parameters->generator, jit_gemm_params, group_output_channels, nr,
              group_input_channels, log2_input_element_size, kernel_size, convolution_op);
          #endif  // XNN_PLATFORM_JIT

          break;
        default:
          XNN_UNREACHABLE;
      }

      if (scale_params != NULL) {
        assert(init_scale_params != NULL);

        void* group_weights = (void*)
          ((uintptr_t) weights_ptr + gemm_parameters->nr * ((kernel_size * k_stride << log2_filter_element_size) + bias_element_size));
        const size_t weights_stride = (kernel_size * k_stride << log2_filter_element_size) + bias_element_size + extra_weights_bytes;
        for (uint32_t group = 0; group < groups; group++) {
          init_scale_params(
            group_output_channels, gemm_parameters->nr,
            gemm_parameters->nr * weights_stride,
            scale_params, group_weights);
          scale_params += group_output_channels;
          group_weights = (void*) ((uintptr_t) group_weights + n_stride * weights_stride);
        }
      }

      if (use_weights_cache(convolution_op)) {
        convolution_op->packed_weights.offset = xnn_get_or_insert_weights_cache(
            convolution_op->weights_cache, weights_ptr, aligned_total_weights_size);
      }

      zero_size = XNN_EXTRA_BYTES + (k_stride << log2_input_element_size);
      break;
    }
    default:
      XNN_UNREACHABLE;
  }

  const bool tf_same_padding = (flags & XNN_FLAG_TENSORFLOW_SAME_PADDING) != 0 && kernel_size != 1;
  if (any_padding || tf_same_padding) {
    convolution_op->zero_buffer = xnn_allocate_simd_memory(zero_size);
    if (convolution_op->zero_buffer == NULL) {
      xnn_log_error(
        "failed to allocate %zu bytes for %s operator zero padding",
        zero_size, xnn_operator_type_to_string(operator_type));
      goto error;
    }
    memset(convolution_op->zero_buffer, input_padding_byte, zero_size);
  }

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

  convolution_op->kernel_height = kernel_height;
  convolution_op->kernel_width = kernel_width;
  convolution_op->stride_height = subsampling_height;
  convolution_op->stride_width = subsampling_width;
  convolution_op->dilation_height = dilation_height;
  convolution_op->dilation_width = dilation_width;
  convolution_op->groups = groups;
  convolution_op->group_input_channels = group_input_channels;
  convolution_op->group_output_channels = group_output_channels;
  convolution_op->input_pixel_stride = input_channel_stride;
  convolution_op->output_pixel_stride = output_channel_stride;

  convolution_op->type = operator_type;
  convolution_op->ukernel.type = ukernel_type;
  convolution_op->flags = flags & ~XNN_FLAG_TENSORFLOW_SAME_PADDING;
  if (tf_same_padding) {
    convolution_op->flags |= XNN_FLAG_TENSORFLOW_SAME_PADDING;
  }

  convolution_op->state = xnn_run_state_invalid;

  *convolution_op_out = convolution_op;
  return xnn_status_success;

error:
  xnn_delete_operator(convolution_op);
  return status;
}

enum xnn_status xnn_create_convolution2d_nhwc_qu8(
    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,
    size_t input_channel_stride,
    size_t output_channel_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_caches_t caches,
    xnn_operator_t* convolution_op_out)
{
  if (input_scale <= 0.0f || !isnormal(input_scale)) {
    xnn_log_error(
      "failed to create %s operator with %.7g input scale: scale must be finite, normalized, and positive",
      xnn_operator_type_to_string(xnn_operator_type_convolution_nhwc_qu8), input_scale);
    return xnn_status_invalid_parameter;
  }

  if (kernel_scale <= 0.0f || !isnormal(kernel_scale)) {
    xnn_log_error(
      "failed to create %s operator with %.7g kernel scale: scale must be finite, normalized, and positive",
      xnn_operator_type_to_string(xnn_operator_type_convolution_nhwc_qu8), kernel_scale);
    return xnn_status_invalid_parameter;
  }

  if (output_scale <= 0.0f || !isnormal(output_scale)) {
    xnn_log_error(
      "failed to create %s operator with %.7g output scale: scale must be finite, normalized, and positive",
      xnn_operator_type_to_string(xnn_operator_type_convolution_nhwc_qu8), output_scale);
    return xnn_status_invalid_parameter;
  }

  if (output_min >= output_max) {
    xnn_log_error(
      "failed to create %s operator with [%" PRIu8 ", %" PRIu8 "] output range: range min must be below range max",
      xnn_operator_type_to_string(xnn_operator_type_convolution_nhwc_qu8), output_min, output_max);
    return xnn_status_invalid_parameter;
  }

  const float requantization_scale = input_scale * kernel_scale / output_scale;
  if (requantization_scale >= 256.0f) {
    xnn_log_error(
      "failed to create %s operator with %.7g input scale, %.7g kernel scale, and %.7g output scale: "
      "requantization scale %.7g is greater or equal to 256.0",
      xnn_operator_type_to_string(xnn_operator_type_convolution_nhwc_qu8),
      input_scale, kernel_scale, output_scale, requantization_scale);
    return xnn_status_unsupported_parameter;
  }

  const struct xnn_qu8_packing_params packing_params = {
    .input_zero_point = input_zero_point,
    .kernel_zero_point = kernel_zero_point,
  };


  union xnn_qu8_conv_minmax_params gemm_params;
  if XNN_LIKELY(xnn_params.qu8.gemm.init.qu8 != NULL) {
    xnn_params.qu8.gemm.init.qu8(&gemm_params,
      kernel_zero_point, requantization_scale, output_zero_point, output_min, output_max);
  }

  union xnn_qu8_conv_minmax_params dwconv_params;
  const struct dwconv_parameters* dwconv_ukernel =
    find_dwconv_ukernel(kernel_height * kernel_width, xnn_params.qu8.dwconv, XNN_MAX_QU8_DWCONV_UKERNELS);
  if XNN_LIKELY(dwconv_ukernel != NULL) {
    dwconv_ukernel->init.qu8(&dwconv_params,
      kernel_zero_point, requantization_scale, output_zero_point, output_min, output_max);
  }

  return create_convolution2d_nhwc(
    input_padding_top, input_padding_right, input_padding_bottom, input_padding_left,
    kernel_height, kernel_width,
    subsampling_height, subsampling_width,
    dilation_height, dilation_width,
    groups, group_input_channels, group_output_channels,
    input_channel_stride, output_channel_stride,
    kernel, bias, flags,
    0 /* log2(sizeof(input element)) = log2(sizeof(uint8_t)) */,
    0 /* log2(sizeof(filter element)) = log2(sizeof(uint8_t)) */,
    sizeof(int32_t) /* sizeof(bias element) */,
    (xnn_pack_vmulcaddc_w_function) NULL,
    (xnn_pack_dwconv_hwg_w_function) xnn_pack_qu8_dwconv_hwg_w,
    (xnn_pack_dwconv_ghw_w_function) xnn_pack_qu8_dwconv_ghw_w,
    (xnn_pack_gemm_goi_w_function) xnn_pack_qu8_gemm_goi_w,
    (xnn_pack_conv_kgo_w_function) xnn_pack_qu8_conv_kgo_w,
    (xnn_pack_conv_goki_w_function) xnn_pack_qu8_conv_goki_w,
    &packing_params, input_zero_point /* input padding byte */, kernel_zero_point /* packed weights padding byte */,
    0 /* extra weights bytes */, NULL /* init scale params fn */, NULL /* scale params */,
    &gemm_params, sizeof(gemm_params),
    &dwconv_params, sizeof(dwconv_params),
    NULL /* vmulcaddc params */, 0,
    &xnn_params.qu8.gemm, dwconv_ukernel, NULL /* vmulcaddc parameters */,
    NULL /* jit_gemm_params */,
    false /* linear activation */, false /* relu activation */, XNN_INIT_FLAG_QU8,
    xnn_operator_type_convolution_nhwc_qu8,
    0, NULL,
    caches,
    convolution_op_out);
}

enum xnn_status xnn_create_convolution2d_nhwc_qs8(
    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,
    size_t input_channel_stride,
    size_t output_channel_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_caches_t caches,
    xnn_operator_t* convolution_op_out)
{
  if (input_scale <= 0.0f || !isnormal(input_scale)) {
    xnn_log_error(
      "failed to create %s operator with %.7g input scale: scale must be finite, normalized, and positive",
      xnn_operator_type_to_string(xnn_operator_type_convolution_nhwc_qs8), input_scale);
    return xnn_status_invalid_parameter;
  }

  if (kernel_scale <= 0.0f || !isnormal(kernel_scale)) {
    xnn_log_error(
      "failed to create %s operator with %.7g kernel scale: scale must be finite, normalized, and positive",
      xnn_operator_type_to_string(xnn_operator_type_convolution_nhwc_qs8), kernel_scale);
    return xnn_status_invalid_parameter;
  }

  if (output_scale <= 0.0f || !isnormal(output_scale)) {
    xnn_log_error(
      "failed to create %s operator with %.7g output scale: scale must be finite, normalized, and positive",
      xnn_operator_type_to_string(xnn_operator_type_convolution_nhwc_qs8), output_scale);
    return xnn_status_invalid_parameter;
  }

  if (output_min >= output_max) {
    xnn_log_error(
      "failed to create %s operator with [%" PRId8 ", %" PRId8 "] output range: range min must be below range max",
      xnn_operator_type_to_string(xnn_operator_type_convolution_nhwc_qs8), output_min, output_max);
    return xnn_status_invalid_parameter;
  }

  const float requantization_scale = input_scale * kernel_scale / output_scale;
  if (requantization_scale >= 256.0f) {
    xnn_log_error(
      "failed to create %s operator with %.7g input scale, %.7g kernel scale, and %.7g output scale: "
      "requantization scale %.7g is greater or equal to 256.0",
      xnn_operator_type_to_string(xnn_operator_type_convolution_nhwc_qs8),
      input_scale, kernel_scale, output_scale, requantization_scale);
    return xnn_status_unsupported_parameter;
  }

  const struct xnn_qs8_packing_params packing_params = { .input_zero_point = input_zero_point, };

  union xnn_qs8_conv_minmax_params gemm_params;
  if XNN_LIKELY(xnn_params.qs8.gemm.init.qs8 != NULL) {
    xnn_params.qs8.gemm.init.qs8(&gemm_params,
      requantization_scale, output_zero_point, output_min, output_max);
  }

  union xnn_qs8_conv_minmax_params dwconv_params;
  const struct dwconv_parameters* dwconv_ukernel =
    find_dwconv_ukernel(kernel_height * kernel_width, xnn_params.qs8.dwconv, XNN_MAX_QS8_DWCONV_UKERNELS);
  if XNN_LIKELY(dwconv_ukernel != NULL) {
    dwconv_ukernel->init.qs8(&dwconv_params,
      requantization_scale, output_zero_point, output_min, output_max);
  }

  return create_convolution2d_nhwc(
    input_padding_top, input_padding_right, input_padding_bottom, input_padding_left,
    kernel_height, kernel_width,
    subsampling_height, subsampling_width,
    dilation_height, dilation_width,
    groups, group_input_channels, group_output_channels,
    input_channel_stride, output_channel_stride,
    kernel, bias, flags,
    0 /* log2(sizeof(input element)) = log2(sizeof(int8_t)) */,
    0 /* log2(sizeof(filter element)) = log2(sizeof(int8_t)) */,
    sizeof(int32_t) /* sizeof(bias element) */,
    (xnn_pack_vmulcaddc_w_function) NULL,
    (xnn_pack_dwconv_hwg_w_function) xnn_pack_qs8_dwconv_hwg_w,
    (xnn_pack_dwconv_ghw_w_function) xnn_pack_qs8_dwconv_ghw_w,
    (xnn_pack_gemm_goi_w_function) xnn_pack_qs8_gemm_goi_w,
    (xnn_pack_conv_kgo_w_function) xnn_pack_qs8_conv_kgo_w,
    (xnn_pack_conv_goki_w_function) xnn_pack_qs8_conv_goki_w,
    &packing_params, input_zero_point /* input padding byte */, 0 /* packed weights padding byte */,
    0 /* extra weights bytes */, NULL /* init scale params fn */, NULL /* scale params */,
    &gemm_params, sizeof(gemm_params),
    &dwconv_params, sizeof(dwconv_params),
    NULL /* vmulcaddc params */, 0,
    &xnn_params.qs8.gemm, dwconv_ukernel, NULL /* vmulcaddc parameters */,
    NULL /* jit_gemm_params */,
    false /* linear activation */, false /* relu activation */, XNN_INIT_FLAG_QS8,
    xnn_operator_type_convolution_nhwc_qs8,
    0, NULL,
    caches,
    convolution_op_out);
}

enum xnn_status xnn_create_convolution2d_nhwc_qc8(
    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,
    size_t input_channel_stride,
    size_t output_channel_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_caches_t caches,
    xnn_operator_t* convolution_op_out)
{
  if (input_scale <= 0.0f || !isnormal(input_scale)) {
    xnn_log_error(
      "failed to create %s operator with %.7g input scale: scale must be finite, normalized, and positive",
      xnn_operator_type_to_string(xnn_operator_type_convolution_nhwc_qc8), input_scale);
    return xnn_status_invalid_parameter;
  }

  for (size_t output_channel = 0; output_channel < groups * group_output_channels; output_channel++) {
    if (kernel_scale[output_channel] <= 0.0f || !isnormal(kernel_scale[output_channel])) {
      xnn_log_error(
        "failed to create %s operator with %.7g kernel scale in output channel #%zu: "
        "scale must be finite, normalized, and positive",
        xnn_operator_type_to_string(xnn_operator_type_convolution_nhwc_qc8), kernel_scale[output_channel],
        output_channel);
      return xnn_status_invalid_parameter;
    }
  }

  if (output_scale <= 0.0f || !isnormal(output_scale)) {
    xnn_log_error(
      "failed to create %s operator with %.7g output scale: scale must be finite, normalized, and positive",
      xnn_operator_type_to_string(xnn_operator_type_convolution_nhwc_qc8), output_scale);
    return xnn_status_invalid_parameter;
  }

  if (output_min >= output_max) {
    xnn_log_error(
      "failed to create %s operator with [%" PRId8 ", %" PRId8 "] output range: range min must be below range max",
      xnn_operator_type_to_string(xnn_operator_type_convolution_nhwc_qc8), output_min, output_max);
    return xnn_status_invalid_parameter;
  }

  float* requantization_scale = XNN_SIMD_ALLOCA(groups * group_output_channels * sizeof(float));
  for (size_t output_channel = 0; output_channel < groups * group_output_channels; output_channel++) {
    requantization_scale[output_channel] = input_scale * kernel_scale[output_channel] / output_scale;
    if (requantization_scale[output_channel] >= 256.0f) {
      xnn_log_error(
        "failed to create %s operator with %.7g input scale, %.7g kernel scale, and %.7g output scale in output channel #%zu: "
        "requantization scale %.7g is greater or equal to 256.0",
        xnn_operator_type_to_string(xnn_operator_type_convolution_nhwc_qc8),
        input_scale, kernel_scale[output_channel], output_scale,
        output_channel, requantization_scale[output_channel]);
      return xnn_status_unsupported_parameter;
    }
  }

  const struct xnn_qs8_packing_params packing_params = { .input_zero_point = input_zero_point, };

  union xnn_qc8_conv_minmax_params gemm_params;
  if XNN_LIKELY(xnn_params.qc8.gemm.init.qc8 != NULL) {
    xnn_params.qc8.gemm.init.qc8(&gemm_params,
      output_zero_point, output_min, output_max);
  }

  union xnn_qc8_conv_minmax_params dwconv_params;
  const struct dwconv_parameters* dwconv_ukernel =
    find_dwconv_ukernel(kernel_height * kernel_width, xnn_params.qc8.dwconv, XNN_MAX_QC8_DWCONV_UKERNELS);
  if XNN_LIKELY(dwconv_ukernel != NULL) {
    dwconv_ukernel->init.qc8(&dwconv_params,
      output_zero_point, output_min, output_max);
  }

  return create_convolution2d_nhwc(
    input_padding_top, input_padding_right, input_padding_bottom, input_padding_left,
    kernel_height, kernel_width,
    subsampling_height, subsampling_width,
    dilation_height, dilation_width,
    groups, group_input_channels, group_output_channels,
    input_channel_stride, output_channel_stride,
    kernel, bias, flags,
    0 /* log2(sizeof(input element)) = log2(sizeof(int8_t)) */,
    0 /* log2(sizeof(filter element)) = log2(sizeof(int8_t)) */,
    sizeof(int32_t) /* sizeof(bias element) */,
    (xnn_pack_vmulcaddc_w_function) NULL,
    (xnn_pack_dwconv_hwg_w_function) xnn_pack_qs8_dwconv_hwg_w,
    (xnn_pack_dwconv_ghw_w_function) xnn_pack_qs8_dwconv_ghw_w,
    (xnn_pack_gemm_goi_w_function) xnn_pack_qs8_gemm_goi_w,
    (xnn_pack_conv_kgo_w_function) xnn_pack_qs8_conv_kgo_w,
    (xnn_pack_conv_goki_w_function) xnn_pack_qs8_conv_goki_w,
    &packing_params, input_zero_point /* input padding byte */, 0 /* packed weights padding byte */,
    sizeof(float) /* extra weights bytes */, xnn_init_qc8_scale_fp32_params, requantization_scale,
    &gemm_params, sizeof(gemm_params),
    &dwconv_params, sizeof(dwconv_params),
    NULL /* vmulcaddc params */, 0,
    &xnn_params.qc8.gemm, dwconv_ukernel, NULL /* vmulcaddc parameters */,
    NULL /* jit_gemm_params */,
    false /* linear activation */, false /* relu activation */, XNN_INIT_FLAG_QC8,
    xnn_operator_type_convolution_nhwc_qc8,
    0, NULL,
    caches,
    convolution_op_out);
}

enum xnn_status xnn_create_convolution2d_nhwc_f16(
    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,
    size_t input_channel_stride,
    size_t output_channel_stride,
    const void* kernel,
    const void* bias,
    float output_min,
    float output_max,
    uint32_t flags,
    xnn_caches_t caches,
    xnn_operator_t* convolution_op_out)
{
  if (isnan(output_min)) {
    xnn_log_error(
      "failed to create %s operator with NaN output lower bound: lower bound must be non-NaN",
      xnn_operator_type_to_string(xnn_operator_type_convolution_nhwc_f16));
    return xnn_status_invalid_parameter;
  }

  if (isnan(output_max)) {
    xnn_log_error(
      "failed to create %s operator with NaN output upper bound: upper bound must be non-NaN",
      xnn_operator_type_to_string(xnn_operator_type_convolution_nhwc_f16));
    return xnn_status_invalid_parameter;
  }

  const uint16_t fp16_output_min = fp16_ieee_from_fp32_value(output_min);
  const uint16_t fp16_output_max = fp16_ieee_from_fp32_value(output_max);
  const float rounded_output_min = fp16_ieee_to_fp32_value(fp16_output_min);
  const float rounded_output_max = fp16_ieee_to_fp32_value(fp16_output_max);
  if (rounded_output_min >= rounded_output_max) {
    xnn_log_error(
      "failed to create %s operator with [%.7g, %.7g] output range: lower bound must be below upper bound",
      xnn_operator_type_to_string(xnn_operator_type_convolution_nhwc_f16), rounded_output_min, rounded_output_max);
    return xnn_status_invalid_parameter;
  }

  union xnn_f16_minmax_params gemm_params;
  if XNN_LIKELY(xnn_params.f16.gemm.init.f16 != NULL) {
    xnn_params.f16.gemm.init.f16(&gemm_params,
      fp16_output_min, fp16_output_max);
  }

  union xnn_f16_minmax_params dwconv_params;
  const struct dwconv_parameters* dwconv_ukernel =
    find_dwconv_ukernel(kernel_height * kernel_width, xnn_params.f16.dwconv, XNN_MAX_F16_DWCONV_UKERNELS);
  if XNN_LIKELY(dwconv_ukernel != NULL) {
    dwconv_ukernel->init.f16(&dwconv_params, fp16_output_min, fp16_output_max);
  }

  union xnn_f16_minmax_params vmulcaddc_params;
  if XNN_LIKELY(xnn_params.f16.vmulcaddc.init.f16 != NULL) {
    xnn_params.f16.vmulcaddc.init.f16(&vmulcaddc_params, fp16_output_min, fp16_output_max);
  }

  xnn_pack_vmulcaddc_w_function pack_vmulcaddc_w = (xnn_pack_vmulcaddc_w_function) xnn_pack_f16_vmulcaddc_w;
  xnn_pack_dwconv_hwg_w_function pack_dwconv_hwg_w = (xnn_pack_dwconv_hwg_w_function) xnn_pack_f16_dwconv_hwg_w;
  xnn_pack_dwconv_ghw_w_function pack_dwconv_ghw_w = (xnn_pack_dwconv_ghw_w_function) xnn_pack_f16_dwconv_ghw_w;
  xnn_pack_gemm_goi_w_function pack_gemm_goi_w = (xnn_pack_gemm_goi_w_function) xnn_pack_f16_gemm_goi_w;
  xnn_pack_conv_kgo_w_function pack_conv_kgo_w = (xnn_pack_conv_kgo_w_function) xnn_pack_f16_conv_kgo_w;
  xnn_pack_conv_goki_w_function pack_conv_goki_w = (xnn_pack_conv_goki_w_function) xnn_pack_f16_conv_goki_w;
  if (flags & XNN_FLAG_FP32_STATIC_WEIGHTS) {
    pack_vmulcaddc_w = (xnn_pack_vmulcaddc_w_function) xnn_pack_f32_to_f16_vmulcaddc_w;
    pack_dwconv_hwg_w = (xnn_pack_dwconv_hwg_w_function) xnn_pack_f32_to_f16_dwconv_hwg_w;
    pack_dwconv_ghw_w = (xnn_pack_dwconv_ghw_w_function) xnn_pack_f32_to_f16_dwconv_ghw_w;
    pack_gemm_goi_w = (xnn_pack_gemm_goi_w_function) xnn_pack_f32_to_f16_gemm_goi_w;
    pack_conv_kgo_w = (xnn_pack_conv_kgo_w_function) xnn_pack_f32_to_f16_conv_kgo_w;
    pack_conv_goki_w = (xnn_pack_conv_goki_w_function) xnn_pack_f32_to_f16_conv_goki_w;
  }

  return create_convolution2d_nhwc(
    input_padding_top, input_padding_right, input_padding_bottom, input_padding_left,
    kernel_height, kernel_width,
    subsampling_height, subsampling_width,
    dilation_height, dilation_width,
    groups, group_input_channels, group_output_channels,
    input_channel_stride, output_channel_stride,
    kernel, bias, flags,
    1 /* log2(sizeof(input element)) = log2(sizeof(uint16_t)) */,
    1 /* log2(sizeof(filter element)) = log2(sizeof(uint16_t)) */,
    sizeof(uint16_t) /* sizeof(bias element) */,
    pack_vmulcaddc_w,
    pack_dwconv_hwg_w,
    pack_dwconv_ghw_w,
    pack_gemm_goi_w,
    pack_conv_kgo_w,
    pack_conv_goki_w,
    NULL /* packing params */, 0 /* input padding byte */, 0 /* packed weights padding byte */,
    0 /* extra weights bytes */, NULL /* init scale params fn */, NULL /* scale params */,
    &gemm_params, sizeof(gemm_params),
    &dwconv_params, sizeof(dwconv_params),
    &vmulcaddc_params, sizeof(vmulcaddc_params),
    &xnn_params.f16.gemm, dwconv_ukernel, &xnn_params.f16.vmulcaddc,
    NULL /* jit_gemm_params */,
    false /* linear activation */, false /* relu activation */, XNN_INIT_FLAG_F16,
    xnn_operator_type_convolution_nhwc_f16,
    0, NULL,
    caches,
    convolution_op_out);
}

enum xnn_status xnn_create_convolution2d_nhwc_f32(
    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,
    size_t input_channel_stride,
    size_t output_channel_stride,
    const float* kernel,
    const float* bias,
    float output_min,
    float output_max,
    uint32_t flags,
    xnn_caches_t caches,
    xnn_operator_t* convolution_op_out)
{
  if (isnan(output_min)) {
    xnn_log_error(
      "failed to create %s operator with NaN output lower bound: lower bound must be non-NaN",
      xnn_operator_type_to_string(xnn_operator_type_convolution_nhwc_f32));
    return xnn_status_invalid_parameter;
  }

  if (isnan(output_max)) {
    xnn_log_error(
      "failed to create %s operator with NaN output upper bound: upper bound must be non-NaN",
      xnn_operator_type_to_string(xnn_operator_type_convolution_nhwc_f32));
    return xnn_status_invalid_parameter;
  }

  if (output_min >= output_max) {
    xnn_log_error(
      "failed to create %s operator with [%.7g, %.7g] output range: lower bound must be below upper bound",
      xnn_operator_type_to_string(xnn_operator_type_convolution_nhwc_f32), output_min, output_max);
    return xnn_status_invalid_parameter;
  }

  const bool linear_activation = (output_max == INFINITY) && (output_min == -output_max);
  const bool relu_activation = (output_max == INFINITY) && (output_min == 0.0f);

  const struct gemm_parameters* gemm_parameters = &xnn_params.f32.gemm;
  if (gemm_parameters->nr > group_output_channels) {
    // Default micro-kernel is suboptimal. Try to find a better micro-kernel.

    if (xnn_params.f32.gemm2.minmax.igemm[gemm_parameters->mr].function[XNN_UARCH_DEFAULT] != NULL) {
      gemm_parameters = &xnn_params.f32.gemm2;
    }
  }

  union xnn_f32_minmax_params gemm_params;
  if XNN_LIKELY(gemm_parameters->init.f32 != NULL) {
    gemm_parameters->init.f32(&gemm_params, output_min, output_max);
  }

  struct jit_gemm_params jit_gemm_params = {
    .f32_minmax = {
      .min = output_min,
      .max = output_max
    }
  };

  union xnn_f32_minmax_params dwconv_params;
  const struct dwconv_parameters* dwconv_ukernel =
    find_dwconv_ukernel(kernel_height * kernel_width, xnn_params.f32.dwconv, XNN_MAX_F32_DWCONV_UKERNELS);
  if XNN_LIKELY(dwconv_ukernel != NULL) {
    dwconv_ukernel->init.f32(&dwconv_params, output_min, output_max);
  }

  union xnn_f32_minmax_params vmulcaddc_params;
  if XNN_LIKELY(xnn_params.f32.vmulcaddc.init.f32 != NULL) {
    xnn_params.f32.vmulcaddc.init.f32(&vmulcaddc_params, output_min, output_max);
  }

  return create_convolution2d_nhwc(
    input_padding_top, input_padding_right, input_padding_bottom, input_padding_left,
    kernel_height, kernel_width,
    subsampling_height, subsampling_width,
    dilation_height, dilation_width,
    groups, group_input_channels, group_output_channels,
    input_channel_stride, output_channel_stride,
    kernel, bias, flags,
    2 /* log2(sizeof(input element)) = log2(sizeof(float)) */,
    2 /* log2(sizeof(filter element)) = log2(sizeof(float)) */,
    sizeof(float) /* sizeof(bias element) */,
    (xnn_pack_vmulcaddc_w_function) xnn_pack_f32_vmulcaddc_w,
    (xnn_pack_dwconv_hwg_w_function) xnn_pack_f32_dwconv_hwg_w,
    (xnn_pack_dwconv_ghw_w_function) xnn_pack_f32_dwconv_ghw_w,
    (xnn_pack_gemm_goi_w_function) xnn_pack_f32_gemm_goi_w,
    (xnn_pack_conv_kgo_w_function) xnn_pack_f32_conv_kgo_w,
    (xnn_pack_conv_goki_w_function) xnn_pack_f32_conv_goki_w,
    NULL /* packing params */, 0 /* input padding byte */, 0 /* packed weights padding byte */,
    0 /* extra weights bytes */, NULL /* init scale params fn */, NULL /* scale params */,
    &gemm_params, sizeof(gemm_params),
    &dwconv_params, sizeof(dwconv_params),
    &vmulcaddc_params, sizeof(vmulcaddc_params),
    gemm_parameters, dwconv_ukernel, &xnn_params.f32.vmulcaddc,
    &jit_gemm_params,
    linear_activation, relu_activation, XNN_INIT_FLAG_F32,
    xnn_operator_type_convolution_nhwc_f32,
    0, NULL,
    caches,
    convolution_op_out);
}

enum xnn_status xnn_create_fused_convolution2d_nhwc_f32(
    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,
    size_t input_channel_stride,
    size_t output_channel_stride,
    const float* kernel,
    const float* bias,
    size_t num_post_operations,
    struct xnn_post_operation* post_operations,
    uint32_t flags,
    xnn_caches_t caches,
    xnn_operator_t* convolution_op_out)
{
  #if !XNN_ENABLE_JIT
    xnn_log_error(
      "failed to create %s operator: convolution with post operations available only if JIT is enabled",
      xnn_operator_type_to_string(xnn_operator_type_convolution_nhwc_f32));
    return xnn_status_invalid_parameter;
  #endif

  // Convolution is specified with linear activation, any clamping should be specified as a post operator.
  const float output_max = INFINITY;
  const float output_min = -INFINITY;

  struct jit_gemm_params jit_gemm_params = {
    .f32_minmax = {
      .min = output_min,
      .max = output_max
    },
    .num_post_operations = num_post_operations,
    .post_operations = post_operations,
  };

  char* post_operation_params = allocate_and_initialize_post_operation_params(num_post_operations, post_operations);

  union xnn_f32_minmax_params gemm_params;
  if XNN_LIKELY(xnn_params.f32.gemm.init.f32 != NULL) {
    xnn_params.f32.gemm.init.f32(&gemm_params, output_min, output_max);
  }

  union xnn_f32_minmax_params dwconv_params;
  const struct dwconv_parameters* dwconv_ukernel =
    find_dwconv_ukernel(kernel_height * kernel_width, xnn_params.f32.dwconv, XNN_MAX_F32_DWCONV_UKERNELS);
  if XNN_LIKELY(dwconv_ukernel != NULL) {
    dwconv_ukernel->init.f32(&dwconv_params, output_min, output_max);
  }

  union xnn_f32_minmax_params vmulcaddc_params;
  if XNN_LIKELY(xnn_params.f32.vmulcaddc.init.f32 != NULL) {
    xnn_params.f32.vmulcaddc.init.f32(&vmulcaddc_params, output_min, output_max);
  }

  return create_convolution2d_nhwc(
    input_padding_top, input_padding_right, input_padding_bottom, input_padding_left,
    kernel_height, kernel_width,
    subsampling_height, subsampling_width,
    dilation_height, dilation_width,
    groups, group_input_channels, group_output_channels,
    input_channel_stride, output_channel_stride,
    kernel, bias, flags,
    2 /* log2(sizeof(input element)) = log2(sizeof(float)) */,
    2 /* log2(sizeof(filter element)) = log2(sizeof(float)) */,
    sizeof(float) /* sizeof(bias element) */,
    (xnn_pack_vmulcaddc_w_function) xnn_pack_f32_vmulcaddc_w,
    (xnn_pack_dwconv_hwg_w_function) xnn_pack_f32_dwconv_hwg_w,
    (xnn_pack_dwconv_ghw_w_function) xnn_pack_f32_dwconv_ghw_w,
    (xnn_pack_gemm_goi_w_function) xnn_pack_f32_gemm_goi_w,
    (xnn_pack_conv_kgo_w_function) xnn_pack_f32_conv_kgo_w,
    (xnn_pack_conv_goki_w_function) xnn_pack_f32_conv_goki_w,
    NULL /* packing params */, 0 /* input padding byte */, 0 /* packed weights padding byte */,
    0 /* extra weights bytes */, NULL /* init scale params fn */, NULL /* scale params */,
    (void*) &gemm_params, sizeof(gemm_params),
    &dwconv_params, sizeof(dwconv_params),
    &vmulcaddc_params, sizeof(vmulcaddc_params),
    &xnn_params.f32.gemm, dwconv_ukernel, &xnn_params.f32.vmulcaddc,
    &jit_gemm_params,
    true /* linear_activation */, false /* relu_activation */, XNN_INIT_FLAG_F32,
    xnn_operator_type_convolution_nhwc_f32,
    num_post_operations, post_operation_params,
    caches,
    convolution_op_out);
}

static enum xnn_status setup_convolution2d_nhwc(
  xnn_operator_t convolution_op,
  enum xnn_operator_type expected_operator_type,
  size_t batch_size,
  size_t input_height,
  size_t input_width,
  const void* input,
  void* output,
  uint32_t datatype_init_flags,
  uint32_t log2_input_element_size,
  uint32_t log2_filter_element_size,
  uint32_t extra_weights_elements_size,
  uint32_t log2_output_element_size,
  size_t num_threads)
{
  if (convolution_op->type != expected_operator_type) {
    xnn_log_error("failed to setup operator: operator type mismatch (expected %s, got %s)",
      xnn_operator_type_to_string(expected_operator_type),
      xnn_operator_type_to_string(convolution_op->type));
    return xnn_status_invalid_parameter;
  }
  convolution_op->state = xnn_run_state_invalid;

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

  if ((xnn_params.init_flags & datatype_init_flags) != datatype_init_flags) {
    xnn_log_error(
      "failed to create %s operator: operations on data type are not supported",
      xnn_operator_type_to_string(convolution_op->type));
    return xnn_status_unsupported_hardware;
  }

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

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

  if (convolution_op->weights_cache != NULL && !xnn_weights_cache_is_finalized(convolution_op->weights_cache)) {
    xnn_log_error("failed to setup %s operator: weights cache is not finalized",
      xnn_operator_type_to_string(convolution_op->type));
    return xnn_status_invalid_state;
  }

  convolution_op->batch_size = batch_size;
  convolution_op->input_height = input_height;
  convolution_op->input_width = input_width;
  convolution_op->input = input;

  if (convolution_op->flags & XNN_FLAG_TENSORFLOW_SAME_PADDING) {
    convolution_op->output_height = compute_output_dimension_with_tf_same_padding(
        input_height, convolution_op->stride_height);
    convolution_op->output_width = compute_output_dimension_with_tf_same_padding(
        input_width, convolution_op->stride_width);

    const uint32_t effective_kernel_height = (convolution_op->kernel_height - 1) * convolution_op->dilation_height + 1;
    const uint32_t effective_kernel_width = (convolution_op->kernel_width - 1) * convolution_op->dilation_width + 1;
    const size_t total_padding_height =
      (convolution_op->output_height - 1) * convolution_op->stride_height + effective_kernel_height - input_height;
    const size_t total_padding_width =
      (convolution_op->output_width - 1) * convolution_op->stride_width + effective_kernel_width - input_width;
    convolution_op->padding_top = total_padding_height / 2;
    convolution_op->padding_left = total_padding_width / 2;
    convolution_op->padding_bottom = total_padding_height - convolution_op->padding_top;
    convolution_op->padding_right = total_padding_width - convolution_op->padding_left;
  } else {
    convolution_op->output_height = xnn_compute_convolution_output_dimension(
        convolution_op->padding_top + input_height + convolution_op->padding_bottom,
        convolution_op->kernel_height,
        convolution_op->dilation_height,
        convolution_op->stride_height);
    convolution_op->output_width = xnn_compute_convolution_output_dimension(
        convolution_op->padding_left + input_width + convolution_op->padding_right,
        convolution_op->kernel_width,
        convolution_op->dilation_width,
        convolution_op->stride_width);
  }
  convolution_op->output = output;

  switch (convolution_op->ukernel.type) {
    case xnn_ukernel_type_gemm:
    {
      // Convolution maps directly to GEMM and doesn't use indirection buffer.

      const size_t output_height = convolution_op->output_height;
      const size_t output_width = convolution_op->output_width;
      const size_t output_size = output_height * output_width;
      const size_t batch_output_size = batch_size * output_size;

      const size_t groups = convolution_op->groups;
      const size_t group_input_channels = convolution_op->group_input_channels;
      const size_t w_stride = extra_weights_elements_size +
        (round_up_po2(group_input_channels, convolution_op->ukernel.gemm.kr * convolution_op->ukernel.gemm.sr) << log2_filter_element_size);
      const size_t group_output_channels = convolution_op->group_output_channels;

      uint32_t mr = convolution_op->ukernel.gemm.mr;
      const uint32_t nr = convolution_op->ukernel.gemm.nr;
      struct xnn_hmp_gemm_ukernel *gemm_cases = convolution_op->ukernel.gemm.gemm_cases;

      #if XNN_ENABLE_GEMM_M_SPECIALIZATION
        mr = xnn_get_heuristic_mr_gemm(batch_output_size, mr, nr, gemm_cases);
      #else
        if (batch_output_size == 1 && gemm_cases[0].function[XNN_UARCH_DEFAULT] != NULL) {
          mr = 1;
        }
      #endif

      #if XNN_PLATFORM_JIT
        if (convolution_op->code_cache != NULL) {
          const size_t jit_code_offset = gemm_cases[mr - 1].generated_code_offset[XNN_UARCH_DEFAULT];
          if (jit_code_offset != XNN_CACHE_NOT_FOUND) {
            gemm_cases[mr - 1].function[XNN_UARCH_DEFAULT] =
                (xnn_gemm_ukernel_function) cached_code_at_offset(convolution_op, jit_code_offset);
          }
        }
      #endif  // XNN_PLATFORM_JIT
      struct xnn_hmp_gemm_ukernel gemm_ukernel = gemm_cases[mr - 1];

      convolution_op->context.gemm = (struct gemm_context) {
          .k_scaled = group_input_channels << log2_input_element_size,
          .a = input,
          .a_stride = convolution_op->input_pixel_stride << log2_input_element_size,
          .packed_w = packed_weights(convolution_op),
          .w_stride = w_stride,
          .wg_stride = w_stride * round_up(group_output_channels, nr),
          .c = output,
          .cm_stride = convolution_op->output_pixel_stride << log2_output_element_size,
          .cn_stride = nr << log2_output_element_size,
          .cg_stride = group_output_channels << log2_output_element_size,
          .log2_csize = log2_output_element_size,
          .ukernel = gemm_ukernel,
      };
      memcpy(&convolution_op->context.gemm.params, &convolution_op->params, sizeof(convolution_op->context.gemm.params));
      if (convolution_op->num_post_operation_params == 0) {
        convolution_op->context.gemm.fused_params = &convolution_op->context.gemm.params;
      } else {
        convolution_op->context.gemm.fused_params = convolution_op->post_operation_params;
      }

      #if XNN_TEST_MODE
        const size_t nc = nr;
      #else
        size_t nc = group_output_channels;
        if (num_threads > 1) {
          const size_t num_other_tiles = groups * divide_round_up(batch_output_size, mr);
          const size_t target_tiles_per_thread = 5;
          const size_t max_nc = divide_round_up(group_output_channels * num_other_tiles, num_threads * target_tiles_per_thread);
          if (max_nc < nc) {
            nc = min(nc, divide_round_up(nc, max_nc * nr) * nr);
          }
        }
      #endif
      if (groups == 1) {
        #if XNN_MAX_UARCH_TYPES > 1
          if (xnn_is_hmp_gemm_ukernel(gemm_ukernel)) {
            convolution_op->compute.type = xnn_parallelization_type_2d_tile_2d_with_uarch;
            convolution_op->compute.task_2d_tile_2d_with_id = (pthreadpool_task_2d_tile_2d_with_id_t) xnn_compute_hmp_gemm;
          } else {
            convolution_op->compute.type = xnn_parallelization_type_2d_tile_2d;
            convolution_op->compute.task_2d_tile_2d = (pthreadpool_task_2d_tile_2d_t) xnn_compute_gemm;
          }
        #else
          convolution_op->compute.type = xnn_parallelization_type_2d_tile_2d;
          convolution_op->compute.task_2d_tile_2d = (pthreadpool_task_2d_tile_2d_t) xnn_compute_gemm;
        #endif
        convolution_op->compute.range[0] = batch_output_size;
        convolution_op->compute.range[1] = group_output_channels;
        convolution_op->compute.tile[0] = mr;
        convolution_op->compute.tile[1] = nc;
      } else {
        #if XNN_MAX_UARCH_TYPES > 1
          if (xnn_is_hmp_gemm_ukernel(gemm_ukernel)) {
            convolution_op->compute.type = xnn_parallelization_type_3d_tile_2d_with_uarch;
            convolution_op->compute.task_3d_tile_2d_with_id = (pthreadpool_task_3d_tile_2d_with_id_t) xnn_compute_hmp_grouped_gemm;
          } else {
            convolution_op->compute.type = xnn_parallelization_type_3d_tile_2d;
            convolution_op->compute.task_3d_tile_2d = (pthreadpool_task_3d_tile_2d_t) xnn_compute_grouped_gemm;
          }
        #else
          convolution_op->compute.type = xnn_parallelization_type_3d_tile_2d;
          convolution_op->compute.task_3d_tile_2d = (pthreadpool_task_3d_tile_2d_t) xnn_compute_grouped_gemm;
        #endif
        convolution_op->compute.range[0] = groups;
        convolution_op->compute.range[1] = batch_output_size;
        convolution_op->compute.range[2] = group_output_channels;
        convolution_op->compute.tile[0] = mr;
        convolution_op->compute.tile[1] = nc;
      }
      convolution_op->state = xnn_run_state_ready;

      return xnn_status_success;
    }
    case xnn_ukernel_type_igemm:
    {
      const size_t groups = convolution_op->groups;
      const size_t kernel_height = convolution_op->kernel_height;
      const size_t kernel_width = convolution_op->kernel_width;
      const size_t kernel_size = kernel_height * kernel_width;
      const size_t output_height = convolution_op->output_height;
      const size_t output_width = convolution_op->output_width;
      const size_t output_size = output_height * output_width;

      uint32_t mr = convolution_op->ukernel.igemm.mr;
      const uint32_t nr = convolution_op->ukernel.igemm.nr;
      struct xnn_hmp_igemm_ukernel* igemm_cases = convolution_op->ukernel.igemm.igemm_cases;

      #if XNN_ENABLE_GEMM_M_SPECIALIZATION
        mr = xnn_get_heuristic_mr_igemm(output_size, mr, nr, igemm_cases);
      #else
        if (output_size == 1 && igemm_cases[0].function[XNN_UARCH_DEFAULT] != NULL) {
          mr = 1;
        }
      #endif

      #if XNN_PLATFORM_JIT
        if (convolution_op->code_cache != NULL) {
          const size_t jit_code_offset = igemm_cases[mr - 1].generated_code_offset[XNN_UARCH_DEFAULT];
          if (jit_code_offset != XNN_CACHE_NOT_FOUND) {
            igemm_cases[mr - 1].function[XNN_UARCH_DEFAULT] =
                (xnn_igemm_ukernel_function) cached_code_at_offset(convolution_op, jit_code_offset);
          }
        }
      #endif  // XNN_PLATFORM_JIT
      struct xnn_hmp_igemm_ukernel igemm_ukernel = igemm_cases[mr - 1];

      const size_t tiled_output_size = round_up(output_size, mr);
      const size_t indirection_buffer_size = sizeof(void*) * kernel_size * tiled_output_size;

      if (input_height != convolution_op->last_input_height ||
          input_width != convolution_op->last_input_width)
      {
        const void** indirection_buffer = (const void**) xnn_reallocate_memory((void*) 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(convolution_op->type));
          return xnn_status_out_of_memory;
        }
        convolution_op->indirection_buffer = indirection_buffer;
        convolution_op->last_input = input;
        convolution_op->last_input_height = input_height;
        convolution_op->last_input_width = input_width;

        xnn_indirection_init_conv2d(convolution_op, mr, log2_input_element_size);
      }

      const size_t group_input_channels = convolution_op->group_input_channels;
      const size_t w_stride = extra_weights_elements_size +
        (round_up_po2(group_input_channels, convolution_op->ukernel.igemm.kr * convolution_op->ukernel.igemm.sr) * kernel_size << log2_filter_element_size);
      const size_t group_output_channels = convolution_op->group_output_channels;
      convolution_op->context.igemm = (struct igemm_context) {
          .ks = kernel_size,
          .ks_scaled = kernel_size * mr * sizeof(void*),
          .kc = group_input_channels << log2_input_element_size,
          .w_stride = w_stride,
          .indirect_a = convolution_op->indirection_buffer,
          .a_offset = (size_t) ((uintptr_t) input - (uintptr_t) convolution_op->last_input),
          .zero = convolution_op->zero_buffer,
          .packed_w = packed_weights(convolution_op),
          .c = convolution_op->output,
          .cm_stride = convolution_op->output_pixel_stride << log2_output_element_size,
          .cn_stride = nr << log2_output_element_size,
          .ga_stride = group_input_channels << log2_input_element_size,
          .gw_stride = w_stride * round_up(group_output_channels, nr),
          .gc_stride = group_output_channels << log2_output_element_size,
          .ba_stride = input_height * input_width * convolution_op->input_pixel_stride << log2_input_element_size,
          .bc_stride = output_size * convolution_op->output_pixel_stride << log2_output_element_size,
          .log2_csize = log2_output_element_size,
          .ukernel = igemm_ukernel,
      };
      memcpy(&convolution_op->context.igemm.params, &convolution_op->params, sizeof(convolution_op->context.igemm.params));

      #if XNN_TEST_MODE
        const size_t nc = nr;
      #else
        size_t nc = group_output_channels;
        if (num_threads > 1) {
          const size_t num_other_tiles = groups * batch_size * divide_round_up(output_size, mr);
          const size_t target_tiles_per_thread = 5;
          const size_t max_nc = divide_round_up(group_output_channels * num_other_tiles, num_threads * target_tiles_per_thread);
          if (max_nc < nc) {
            nc = min(nc, divide_round_up(nc, max_nc * nr) * nr);
          }
        }
      #endif
      if (groups == 1) {
        #if XNN_MAX_UARCH_TYPES > 1
          if (xnn_is_hmp_igemm_ukernel(igemm_ukernel)) {
            if (batch_size > 1) {
              convolution_op->compute.type = xnn_parallelization_type_3d_tile_2d_with_uarch;
              convolution_op->compute.task_3d_tile_2d_with_id = (pthreadpool_task_3d_tile_2d_with_id_t) xnn_compute_batch_hmp_igemm;
            } else {
              convolution_op->compute.type = xnn_parallelization_type_2d_tile_2d_with_uarch;
              convolution_op->compute.task_2d_tile_2d_with_id = (pthreadpool_task_2d_tile_2d_with_id_t) xnn_compute_hmp_igemm;
            }
          } else {
            if (batch_size > 1) {
              convolution_op->compute.type = xnn_parallelization_type_3d_tile_2d;
              convolution_op->compute.task_3d_tile_2d = (pthreadpool_task_3d_tile_2d_t) xnn_compute_batch_igemm;
            } else {
              convolution_op->compute.type = xnn_parallelization_type_2d_tile_2d;
              convolution_op->compute.task_2d_tile_2d = (pthreadpool_task_2d_tile_2d_t) xnn_compute_igemm;
            }
          }
        #else
          if (batch_size > 1) {
            convolution_op->compute.type = xnn_parallelization_type_3d_tile_2d;
            convolution_op->compute.task_3d_tile_2d = (pthreadpool_task_3d_tile_2d_t) xnn_compute_batch_igemm;
          } else {
            convolution_op->compute.type = xnn_parallelization_type_2d_tile_2d;
            convolution_op->compute.task_2d_tile_2d = (pthreadpool_task_2d_tile_2d_t) xnn_compute_igemm;
          }
        #endif
        if (batch_size > 1) {
          convolution_op->compute.range[0] = batch_size;
          convolution_op->compute.range[1] = output_size;
          convolution_op->compute.range[2] = group_output_channels;
        } else {
          convolution_op->compute.range[0] = output_size;
          convolution_op->compute.range[1] = group_output_channels;
        }
        convolution_op->compute.tile[0] = mr;
        convolution_op->compute.tile[1] = nc;
      } else {
        #if XNN_MAX_UARCH_TYPES > 1
          if (xnn_is_hmp_igemm_ukernel(igemm_ukernel)) {
            if (batch_size > 1) {
              convolution_op->compute.type = xnn_parallelization_type_4d_tile_2d_with_uarch;
              convolution_op->compute.task_4d_tile_2d_with_id = (pthreadpool_task_4d_tile_2d_with_id_t) xnn_compute_hmp_grouped_batch_igemm;
            } else {
              convolution_op->compute.type = xnn_parallelization_type_3d_tile_2d_with_uarch;
              convolution_op->compute.task_3d_tile_2d_with_id = (pthreadpool_task_3d_tile_2d_with_id_t) xnn_compute_hmp_grouped_igemm;
            }
          } else {
            if (batch_size > 1) {
              convolution_op->compute.type = xnn_parallelization_type_4d_tile_2d;
              convolution_op->compute.task_4d_tile_2d = (pthreadpool_task_4d_tile_2d_t) xnn_compute_grouped_batch_igemm;
            } else {
              convolution_op->compute.type = xnn_parallelization_type_3d_tile_2d;
              convolution_op->compute.task_3d_tile_2d = (pthreadpool_task_3d_tile_2d_t) xnn_compute_grouped_igemm;
            }
          }
        #else
          if (batch_size > 1) {
            convolution_op->compute.type = xnn_parallelization_type_4d_tile_2d;
            convolution_op->compute.task_4d_tile_2d = (pthreadpool_task_4d_tile_2d_t) xnn_compute_grouped_batch_igemm;
          } else {
            convolution_op->compute.type = xnn_parallelization_type_3d_tile_2d;
            convolution_op->compute.task_3d_tile_2d = (pthreadpool_task_3d_tile_2d_t) xnn_compute_grouped_igemm;
          }
        #endif
        if (batch_size > 1) {
          convolution_op->compute.range[0] = batch_size;
          convolution_op->compute.range[1] = groups;
          convolution_op->compute.range[2] = output_size;
          convolution_op->compute.range[3] = group_output_channels;
        } else {
          convolution_op->compute.range[0] = groups;
          convolution_op->compute.range[1] = output_size;
          convolution_op->compute.range[2] = group_output_channels;
        }
        convolution_op->compute.tile[0] = mr;
        convolution_op->compute.tile[1] = nc;
      }
      convolution_op->state = xnn_run_state_ready;

      return xnn_status_success;
    }
    case xnn_ukernel_type_dwconv:
    {
      const size_t kernel_height = convolution_op->kernel_height;
      const size_t kernel_width = convolution_op->kernel_width;
      const size_t kernel_size = kernel_height * kernel_width;
      const size_t output_height = convolution_op->output_height;
      const size_t output_width = convolution_op->output_width;
      const size_t step_width = convolution_op->dilation_width == 1 ? convolution_op->stride_width : kernel_width;
      const size_t step_height = kernel_size + (output_width - 1) * step_width * kernel_height;
      if (input_height != convolution_op->last_input_height || input_width != convolution_op->last_input_width) {
        const size_t indirection_buffer_size = sizeof(void*) * output_height * step_height;

        const void** indirection_buffer =
          (const void**) xnn_reallocate_memory(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(convolution_op->type));
          return xnn_status_out_of_memory;
        }
        convolution_op->indirection_buffer = indirection_buffer;

        xnn_indirection_init_dwconv2d(convolution_op, step_height, step_width, log2_input_element_size);

        convolution_op->last_input = input;
        convolution_op->last_input_height = input_height;
        convolution_op->last_input_width = input_width;
      }

      const size_t groups = convolution_op->groups;
      convolution_op->context.dwconv = (struct dwconv_context) {
          .indirect_input = convolution_op->indirection_buffer,
          .indirect_input_width_stride = kernel_height * step_width * sizeof(void*),
          .indirect_input_height_stride = step_height * sizeof(void*),
          .input_offset = (size_t) ((uintptr_t) input - (uintptr_t) convolution_op->last_input),
          .input_batch_stride = (input_height * input_width * convolution_op->input_pixel_stride) << log2_input_element_size,
          .packed_weights = packed_weights(convolution_op),
          .output = convolution_op->output,
          .output_batch_stride = (output_height * output_width * convolution_op->output_pixel_stride) << log2_output_element_size,
          .output_height_stride = (output_width * convolution_op->output_pixel_stride) << log2_output_element_size,
          .output_width = output_width,
          .groups = groups,
          .zero = convolution_op->zero_buffer,
          .output_increment = (convolution_op->output_pixel_stride - groups) << log2_output_element_size,
          .unipass_ukernel = convolution_op->ukernel.dwconv.unipass_function,
      };
      memcpy(&convolution_op->context.dwconv.params, &convolution_op->params, sizeof(convolution_op->context.dwconv.params));

      convolution_op->compute.type = xnn_parallelization_type_2d;
      convolution_op->compute.task_2d = (pthreadpool_task_2d_t) xnn_compute_dwconv_unipass;
      convolution_op->compute.range[0] = batch_size;
      convolution_op->compute.range[1] = output_height;
      convolution_op->state = xnn_run_state_ready;

      return xnn_status_success;
    }
    case xnn_ukernel_type_vmulcaddc:
    {
      const size_t batch_output_size = batch_size * convolution_op->output_height * convolution_op->output_width;

      convolution_op->context.vmulcaddc = (struct vmulcaddc_context) {
          .n = convolution_op->groups << log2_input_element_size,
          .x = input,
          .x_stride = convolution_op->input_pixel_stride << log2_input_element_size,
          .w = packed_weights(convolution_op),
          .y = output,
          .y_stride = convolution_op->output_pixel_stride << log2_output_element_size,
          .ukernel = convolution_op->ukernel.vmulcaddc.function,
      };
      memcpy(&convolution_op->context.vmulcaddc.params, &convolution_op->params, sizeof(convolution_op->context.vmulcaddc.params));

      #if XNN_TEST_MODE
        const size_t mc = convolution_op->ukernel.vmulcaddc.mr;
      #else
        size_t mc = batch_output_size;
        if (num_threads > 1) {
          const size_t target_tiles_per_thread = 5;
          const size_t max_mc = divide_round_up(batch_output_size, num_threads * target_tiles_per_thread);
          if (max_mc < mc) {
            const uint32_t mr = convolution_op->ukernel.vmulcaddc.mr;
            mc = min(mc, divide_round_up(mc, max_mc * mr) * mr);
          }
        }
      #endif
      convolution_op->compute.type = xnn_parallelization_type_1d_tile_1d;
      convolution_op->compute.task_1d_tile_1d = (pthreadpool_task_1d_tile_1d_t) xnn_compute_vmulcaddc;
      convolution_op->compute.range[0] = batch_output_size;
      convolution_op->compute.tile[0] = mc;
      convolution_op->state = xnn_run_state_ready;

      return xnn_status_success;
    }
    default:
      XNN_UNREACHABLE;
  }
}

enum xnn_status xnn_setup_convolution2d_nhwc_qu8(
    xnn_operator_t convolution_op,
    size_t batch_size,
    size_t input_height,
    size_t input_width,
    const uint8_t* input,
    uint8_t* output,
    pthreadpool_t threadpool)
{
  return setup_convolution2d_nhwc(
    convolution_op, xnn_operator_type_convolution_nhwc_qu8,
    batch_size, input_height, input_width,
    input, output,
    XNN_INIT_FLAG_QU8,
    0 /* log2(sizeof(input element)) = log2(sizeof(uint8_t)) */,
    0 /* log2(sizeof(filter element)) = log2(sizeof(uint8_t)) */,
    sizeof(int32_t) /* sizeof(extra weights elements) */,
    0 /* log2(sizeof(output element)) = log2(sizeof(uint8_t)) */,
    pthreadpool_get_threads_count(threadpool));
}

enum xnn_status xnn_setup_convolution2d_nhwc_qs8(
    xnn_operator_t convolution_op,
    size_t batch_size,
    size_t input_height,
    size_t input_width,
    const int8_t* input,
    int8_t* output,
    pthreadpool_t threadpool)
{
  return setup_convolution2d_nhwc(
    convolution_op, xnn_operator_type_convolution_nhwc_qs8,
    batch_size, input_height, input_width,
    input, output,
    XNN_INIT_FLAG_QS8,
    0 /* log2(sizeof(input element)) = log2(sizeof(int8_t)) */,
    0 /* log2(sizeof(filter element)) = log2(sizeof(int8_t)) */,
    sizeof(int32_t) /* sizeof(extra weights elements) */,
    0 /* log2(sizeof(output element)) = log2(sizeof(int8_t)) */,
    pthreadpool_get_threads_count(threadpool));
}

enum xnn_status xnn_setup_convolution2d_nhwc_qc8(
    xnn_operator_t convolution_op,
    size_t batch_size,
    size_t input_height,
    size_t input_width,
    const int8_t* input,
    int8_t* output,
    pthreadpool_t threadpool)
{
  return setup_convolution2d_nhwc(
    convolution_op, xnn_operator_type_convolution_nhwc_qc8,
    batch_size, input_height, input_width,
    input, output,
    XNN_INIT_FLAG_QC8,
    0 /* log2(sizeof(input element)) = log2(sizeof(int8_t)) */,
    0 /* log2(sizeof(filter element)) = log2(sizeof(int8_t)) */,
    sizeof(int32_t) + sizeof(float) /* sizeof(extra weights elements) */,
    0 /* log2(sizeof(output element)) = log2(sizeof(int8_t)) */,
    pthreadpool_get_threads_count(threadpool));
}

enum xnn_status xnn_setup_convolution2d_nhwc_f16(
    xnn_operator_t convolution_op,
    size_t batch_size,
    size_t input_height,
    size_t input_width,
    const void* input,
    void* output,
    pthreadpool_t threadpool)
{
  return setup_convolution2d_nhwc(
    convolution_op, xnn_operator_type_convolution_nhwc_f16,
    batch_size, input_height, input_width,
    input, output,
    XNN_INIT_FLAG_F16,
    1 /* log2(sizeof(input element)) = log2(sizeof(uint16_t)) */,
    1 /* log2(sizeof(filter element)) = log2(sizeof(uint16_t)) */,
    sizeof(uint16_t) /* sizeof(extra weights elements) */,
    1 /* log2(sizeof(output element)) = log2(sizeof(uint16_t)) */,
    pthreadpool_get_threads_count(threadpool));
}

enum xnn_status xnn_setup_convolution2d_nhwc_f32(
    xnn_operator_t convolution_op,
    size_t batch_size,
    size_t input_height,
    size_t input_width,
    const float* input,
    float* output,
    pthreadpool_t threadpool)
{
  return setup_convolution2d_nhwc(
    convolution_op, xnn_operator_type_convolution_nhwc_f32,
    batch_size, input_height, input_width,
    input, output,
    XNN_INIT_FLAG_F32,
    2 /* log2(sizeof(input element)) = log2(sizeof(float)) */,
    2 /* log2(sizeof(filter element)) = log2(sizeof(float)) */,
    sizeof(float) /* sizeof(extra weights elements) */,
    2 /* log2(sizeof(output element)) = log2(sizeof(float)) */,
    pthreadpool_get_threads_count(threadpool));
}