src/subgraph/concatenate.c

// Copyright 2022 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 <stdint.h>

#include <xnnpack.h>
#include <xnnpack/log.h>
#include <xnnpack/operator.h>
#include <xnnpack/params.h>
#include <xnnpack/subgraph.h>
#include <xnnpack/subgraph-validation.h>

static enum xnn_status create_concatenate_operator_helper(
  const struct xnn_node *node,
  size_t channels,
  size_t input_stride,
  size_t output_stride,
  struct xnn_operator_data *opdata,
  size_t index)
{
  switch (node->compute_type) {
#ifndef XNN_NO_F16_OPERATORS
    case xnn_compute_type_fp16: {
      return xnn_create_copy_nc_x16(channels, input_stride, output_stride, node->flags, &opdata->operator_objects[index]);
    }
#endif  // !defined(XNN_NO_F16_OPERATORS)
    case xnn_compute_type_fp32: {
      return xnn_create_copy_nc_x32(channels, input_stride, output_stride, node->flags, &opdata->operator_objects[index]);
    }
#ifndef XNN_NO_QS8_OPERATORS
    case xnn_compute_type_qs8:
#endif  // !defined(XNN_NO_QS8_OPERATORS)
#ifndef XNN_NO_QU8_OPERATORS
    case xnn_compute_type_qu8:
#endif  // !defined(XNN_NO_QU8_OPERATORS)
#if !defined(XNN_NO_QS8_OPERATORS) || !defined(XNN_NO_QU8_OPERATORS)
    {
      return xnn_create_copy_nc_x8(channels, input_stride, output_stride, node->flags, &opdata->operator_objects[index]);
    }
#endif  // !defined(XNN_NO_QS8_OPERATORS) || !defined(XNN_NO_QU8_OPERATORS)
    default:
      XNN_UNREACHABLE;
  }
}

static enum xnn_status create_concatenate2_operator(
  const struct xnn_node* node,
  const struct xnn_value* values,
  size_t num_values,
  struct xnn_operator_data* opdata,
  const struct xnn_caches* caches)
{
  assert(node->num_inputs == 2);
  const uint32_t input1_id = node->inputs[0];
  assert(input1_id != XNN_INVALID_VALUE_ID);
  assert(input1_id < num_values);
  const uint32_t input2_id = node->inputs[1];
  assert(input2_id != XNN_INVALID_VALUE_ID);
  assert(input2_id < num_values);

  assert(node->num_outputs == 1);
  const uint32_t output_id = node->outputs[0];
  assert(output_id != XNN_INVALID_VALUE_ID);
  assert(output_id < num_values);

  const size_t axis = node->params.concatenate.axis;
  size_t batch_size = 1, channels_1 = 1, channels_2 = 1;
  for (size_t i = 0; i < axis; i++) {
    batch_size *= values[output_id].shape.dim[i];
  }

  for (size_t i = axis; i < values[input1_id].shape.num_dims; i++) {
    channels_1 *= values[input1_id].shape.dim[i];
    channels_2 *= values[input2_id].shape.dim[i];
  }
  const size_t output_stride = channels_1 + channels_2;

  enum xnn_status status;
  status = create_concatenate_operator_helper(node, channels_1, channels_1, output_stride, opdata, 0);
  if (status != xnn_status_success) {
    return status;
  }
  status = create_concatenate_operator_helper(node, channels_2, channels_2, output_stride, opdata, 1);
  if (status != xnn_status_success) {
    return status;
  }

  opdata->inputs[0] = input1_id;
  opdata->inputs[1] = input2_id;
  opdata->outputs[0] = output_id;
  opdata->batch_size = batch_size;

  return status;
}

static enum xnn_status create_concatenate3_operator(
  const struct xnn_node* node,
  const struct xnn_value* values,
  size_t num_values,
  struct xnn_operator_data* opdata,
  const struct xnn_caches* caches)
{
  assert(node->num_inputs == 3);
  const uint32_t input1_id = node->inputs[0];
  assert(input1_id != XNN_INVALID_VALUE_ID);
  assert(input1_id < num_values);
  const uint32_t input2_id = node->inputs[1];
  assert(input2_id != XNN_INVALID_VALUE_ID);
  assert(input2_id < num_values);
  const uint32_t input3_id = node->inputs[2];
  assert(input3_id != XNN_INVALID_VALUE_ID);
  assert(input3_id < num_values);

  assert(node->num_outputs == 1);
  const uint32_t output_id = node->outputs[0];
  assert(output_id != XNN_INVALID_VALUE_ID);
  assert(output_id < num_values);

  const size_t axis = node->params.concatenate.axis;
  size_t batch_size = 1, channels_1 = 1, channels_2 = 1, channels_3 = 1;
  for (size_t i = 0; i < axis; i++) {
    batch_size *= values[output_id].shape.dim[i];
  }

  for (size_t i = axis; i < values[input1_id].shape.num_dims; i++) {
    channels_1 *= values[input1_id].shape.dim[i];
    channels_2 *= values[input2_id].shape.dim[i];
    channels_3 *= values[input3_id].shape.dim[i];
  }
  const size_t output_stride = channels_1 + channels_2 + channels_3;

  enum xnn_status status;
  status = create_concatenate_operator_helper(node, channels_1, channels_1, output_stride, opdata, 0);
  if (status != xnn_status_success) {
    return status;
  }
  status = create_concatenate_operator_helper(node, channels_2, channels_2, output_stride, opdata, 1);
  if (status != xnn_status_success) {
    return status;
  }
  status = create_concatenate_operator_helper(node, channels_3, channels_3, output_stride, opdata, 2);
  if (status != xnn_status_success) {
    return status;
  }

  opdata->inputs[0] = input1_id;
  opdata->inputs[1] = input2_id;
  opdata->inputs[2] = input3_id;
  opdata->outputs[0] = output_id;
  opdata->batch_size = batch_size;

  return status;
}

static enum xnn_status create_concatenate4_operator(
  const struct xnn_node* node,
  const struct xnn_value* values,
  size_t num_values,
  struct xnn_operator_data* opdata,
  const struct xnn_caches* caches)
{
  assert(node->num_inputs == 4);
  const uint32_t input1_id = node->inputs[0];
  assert(input1_id != XNN_INVALID_VALUE_ID);
  assert(input1_id < num_values);
  const uint32_t input2_id = node->inputs[1];
  assert(input2_id != XNN_INVALID_VALUE_ID);
  assert(input2_id < num_values);
  const uint32_t input3_id = node->inputs[2];
  assert(input3_id != XNN_INVALID_VALUE_ID);
  assert(input3_id < num_values);
  const uint32_t input4_id = node->inputs[3];
  assert(input4_id != XNN_INVALID_VALUE_ID);
  assert(input4_id < num_values);

  assert(node->num_outputs == 1);
  const uint32_t output_id = node->outputs[0];
  assert(output_id != XNN_INVALID_VALUE_ID);
  assert(output_id < num_values);

  const size_t axis = node->params.concatenate.axis;
  size_t batch_size = 1, channels_1 = 1, channels_2 = 1, channels_3 = 1, channels_4 = 1;
  for (size_t i = 0; i < axis; i++) {
    batch_size *= values[output_id].shape.dim[i];
  }

  for (size_t i = axis; i < values[input1_id].shape.num_dims; i++) {
    channels_1 *= values[input1_id].shape.dim[i];
    channels_2 *= values[input2_id].shape.dim[i];
    channels_3 *= values[input3_id].shape.dim[i];
    channels_4 *= values[input4_id].shape.dim[i];
  }
  const size_t output_stride = channels_1 + channels_2 + channels_3 + channels_4;

  enum xnn_status status;
  status = create_concatenate_operator_helper(node, channels_1, channels_1, output_stride, opdata, 0);
  if (status != xnn_status_success) {
    return status;
  }
  status = create_concatenate_operator_helper(node, channels_2, channels_2, output_stride, opdata, 1);
  if (status != xnn_status_success) {
    return status;
  }
  status = create_concatenate_operator_helper(node, channels_3, channels_3, output_stride, opdata, 2);
  if (status != xnn_status_success) {
    return status;
  }
  status = create_concatenate_operator_helper(node, channels_4, channels_4, output_stride, opdata, 3);
  if (status != xnn_status_success) {
    return status;
  }

  opdata->inputs[0] = input1_id;
  opdata->inputs[1] = input2_id;
  opdata->inputs[2] = input3_id;
  opdata->inputs[3] = input4_id;
  opdata->outputs[0] = output_id;
  opdata->batch_size = batch_size;

  return status;
}

static enum xnn_status setup_concatenate_operator_helper(
  const void* input_data,
  void* output_data,
  const struct xnn_operator_data *opdata,
  size_t index,
  pthreadpool_t threadpool)
{
  // The output pointer of this operator is the sum of all channels of the earlier operators.
  size_t channels = 0;
  for (size_t i = 0; i < index; i++) {
    channels += opdata->operator_objects[i]->channels;
  }

  switch (opdata->operator_objects[index]->type) {
#ifndef XNN_NO_F16_OPERATORS
    case xnn_operator_type_copy_nc_x16: {
      return xnn_setup_copy_nc_x16(
        opdata->operator_objects[index],
        opdata->batch_size,
        input_data,
        (uint16_t*) output_data + channels,
        threadpool);
    }
#endif  // !defined(XNN_NO_F16_OPERATORS)
    case xnn_operator_type_copy_nc_x32: {
      return xnn_setup_copy_nc_x32(
        opdata->operator_objects[index],
        opdata->batch_size,
        input_data,
        (uint32_t*) output_data + channels,
        threadpool);
    }
#if !defined(XNN_NO_QS8_OPERATORS) || !defined(XNN_NO_QU8_OPERATORS)
    case xnn_operator_type_copy_nc_x8: {
      return xnn_setup_copy_nc_x8(
        opdata->operator_objects[index],
        opdata->batch_size,
        input_data,
        (uint8_t*) output_data + channels,
        threadpool);
    }
#endif  // !defined(XNN_NO_QS8_OPERATORS) || !defined(XNN_NO_QU8_OPERATORS)
    default:
      XNN_UNREACHABLE;
  }
}

static enum xnn_status setup_concatenate2_operator(
  const struct xnn_operator_data* opdata,
  const struct xnn_blob* blobs,
  size_t num_blobs,
  pthreadpool_t threadpool)
{
  const uint32_t input1_id = opdata->inputs[0];
  assert(input1_id != XNN_INVALID_VALUE_ID);
  assert(input1_id < num_blobs);

  const uint32_t input2_id = opdata->inputs[1];
  assert(input2_id != XNN_INVALID_VALUE_ID);
  assert(input2_id < num_blobs);

  const uint32_t output_id = opdata->outputs[0];
  assert(output_id != XNN_INVALID_VALUE_ID);
  assert(output_id < num_blobs);

  const struct xnn_blob* input1_blob = blobs + input1_id;
  const void* input1_data = input1_blob->data;
  assert(input1_data != NULL);

  const struct xnn_blob* input2_blob = blobs + input2_id;
  const void* input2_data = input2_blob->data;
  assert(input2_data != NULL);

  const struct xnn_blob* output_blob = blobs + output_id;
  void* output_data = output_blob->data;
  assert(output_data != NULL);

  enum xnn_status status;

  status = setup_concatenate_operator_helper(input1_data, output_data, opdata, 0, threadpool);
  if (status != xnn_status_success) {
    return status;
  }
  return setup_concatenate_operator_helper(input2_data, output_data, opdata, 1, threadpool);
}

static enum xnn_status setup_concatenate3_operator(
  const struct xnn_operator_data* opdata,
  const struct xnn_blob* blobs,
  size_t num_blobs,
  pthreadpool_t threadpool)
{
  const uint32_t input1_id = opdata->inputs[0];
  assert(input1_id != XNN_INVALID_VALUE_ID);
  assert(input1_id < num_blobs);

  const uint32_t input2_id = opdata->inputs[1];
  assert(input2_id != XNN_INVALID_VALUE_ID);
  assert(input2_id < num_blobs);

  const uint32_t input3_id = opdata->inputs[2];
  assert(input3_id != XNN_INVALID_VALUE_ID);
  assert(input3_id < num_blobs);

  const uint32_t output_id = opdata->outputs[0];
  assert(output_id != XNN_INVALID_VALUE_ID);
  assert(output_id < num_blobs);

  const struct xnn_blob* input1_blob = blobs + input1_id;
  const void* input1_data = input1_blob->data;
  assert(input1_data != NULL);

  const struct xnn_blob* input2_blob = blobs + input2_id;
  const void* input2_data = input2_blob->data;
  assert(input2_data != NULL);

  const struct xnn_blob* input3_blob = blobs + input3_id;
  const void* input3_data = input3_blob->data;
  assert(input3_data != NULL);

  const struct xnn_blob* output_blob = blobs + output_id;
  void* output_data = output_blob->data;
  assert(output_data != NULL);

  enum xnn_status status;

  status = setup_concatenate_operator_helper(input1_data, output_data, opdata, 0, threadpool);
  if (status != xnn_status_success) {
    return status;
  }
  status = setup_concatenate_operator_helper(input2_data, output_data, opdata, 1, threadpool);
  if (status != xnn_status_success) {
    return status;
  }
  return setup_concatenate_operator_helper(input3_data, output_data, opdata, 2, threadpool);
}

static enum xnn_status setup_concatenate4_operator(
  const struct xnn_operator_data* opdata,
  const struct xnn_blob* blobs,
  size_t num_blobs,
  pthreadpool_t threadpool)
{
  const uint32_t input1_id = opdata->inputs[0];
  assert(input1_id != XNN_INVALID_VALUE_ID);
  assert(input1_id < num_blobs);

  const uint32_t input2_id = opdata->inputs[1];
  assert(input2_id != XNN_INVALID_VALUE_ID);
  assert(input2_id < num_blobs);

  const uint32_t input3_id = opdata->inputs[2];
  assert(input3_id != XNN_INVALID_VALUE_ID);
  assert(input3_id < num_blobs);

  const uint32_t input4_id = opdata->inputs[3];
  assert(input4_id != XNN_INVALID_VALUE_ID);
  assert(input4_id < num_blobs);

  const uint32_t output_id = opdata->outputs[0];
  assert(output_id != XNN_INVALID_VALUE_ID);
  assert(output_id < num_blobs);

  const struct xnn_blob* input1_blob = blobs + input1_id;
  const void* input1_data = input1_blob->data;
  assert(input1_data != NULL);

  const struct xnn_blob* input2_blob = blobs + input2_id;
  const void* input2_data = input2_blob->data;
  assert(input2_data != NULL);

  const struct xnn_blob* input3_blob = blobs + input3_id;
  const void* input3_data = input3_blob->data;
  assert(input3_data != NULL);

  const struct xnn_blob* input4_blob = blobs + input4_id;
  const void* input4_data = input4_blob->data;
  assert(input4_data != NULL);

  const struct xnn_blob* output_blob = blobs + output_id;
  void* output_data = output_blob->data;
  assert(output_data != NULL);

  enum xnn_status status;

  status = setup_concatenate_operator_helper(input1_data, output_data, opdata, 0, threadpool);
  if (status != xnn_status_success) {
    return status;
  }
  status = setup_concatenate_operator_helper(input2_data, output_data, opdata, 1, threadpool);
  if (status != xnn_status_success) {
    return status;
  }
  status = setup_concatenate_operator_helper(input3_data, output_data, opdata, 2, threadpool);
  if (status != xnn_status_success) {
    return status;
  }
  return setup_concatenate_operator_helper(input4_data, output_data, opdata, 3, threadpool);
}

enum xnn_status check_input_value(
  xnn_subgraph_t subgraph,
  size_t axis,
  uint32_t input_id,
  uint32_t output_id,
  size_t nth,
  enum xnn_node_type node_type)
{
  enum xnn_status status;
  if ((status = xnn_subgraph_check_nth_input_node_id(node_type, input_id, subgraph->num_values, nth)) !=
      xnn_status_success) {
    return status;
  }

  const struct xnn_value* input_value = &subgraph->values[input_id];
  status = xnn_subgraph_check_input_type_dense(node_type, input_id, input_value);
  if (status != xnn_status_success) {
    return status;
  }

  const struct xnn_value* output_value = &subgraph->values[output_id];
  if (input_value->shape.num_dims != output_value->shape.num_dims) {
    xnn_log_error(
      "failed to define %s operator with input %zu ID #%" PRIu32
      ": mismatch number of dimensions, input %zu has %zu, output has %zu",
      xnn_node_type_to_string(node_type), nth, input_id, nth, input_value->shape.num_dims,
      output_value->shape.num_dims);
    return xnn_status_invalid_parameter;
  }

  for (size_t i = 0; i < input_value->shape.num_dims; i++) {
    if (i != axis && input_value->shape.dim[i] != output_value->shape.dim[i]) {
      xnn_log_error(
        "failed to define %s operator with input ID #%" PRIu32
        ": mismatch dimension %zu, input %zu has %zu, output has %zu",
        xnn_node_type_to_string(node_type), input_id, i, nth, input_value->shape.dim[i], output_value->shape.dim[i]);
      return xnn_status_invalid_parameter;
    }
  }

  status = xnn_subgraph_check_datatype_matches(node_type, input_id, input_value, output_id, output_value);
  if (status != xnn_status_success) {
    return status;
  }

  return xnn_status_success;
}

#if !defined(XNN_NO_QS8_OPERATORS) || !defined(XNN_NO_QU8_OPERATORS)
enum xnn_status check_input_compute_type(
  xnn_subgraph_t subgraph,
  uint32_t input_id,
  uint32_t output_id,
  const char* nth,
  enum xnn_node_type node_type)
{
  const struct xnn_value* input_value = &subgraph->values[input_id];
  const struct xnn_value* output_value = &subgraph->values[output_id];
  if (input_value->quantization.zero_point != output_value->quantization.zero_point) {
    xnn_log_error(
        "failed to define %s operator with input ID #%" PRIu32 " and output ID #%" PRIu32
        ": mismatching quantization zero point across the %s input (%" PRId32 ") and the output (%" PRId32 ")",
        xnn_node_type_to_string(node_type), input_id, output_id,
        nth, input_value->quantization.zero_point, output_value->quantization.zero_point);
    return xnn_status_invalid_parameter;
  }
  if (input_value->quantization.scale != output_value->quantization.scale) {
    xnn_log_error(
        "failed to define %s operator with input ID #%" PRIu32 " and output ID #%" PRIu32
        ": mismatching quantization scale across the %s input (%.7g) and the output (%.7g)",
        xnn_node_type_to_string(node_type), input_id, output_id,
        nth, input_value->quantization.scale, output_value->quantization.scale);
    return xnn_status_invalid_parameter;
  }
  return xnn_status_success;
}
#endif  // !defined( XNN_NO_QS8_OPERATORS) || !defined(XNN_NO_QU8_OPERATORS)

enum xnn_status xnn_define_concatenate_n(
  enum xnn_node_type node_type,
  xnn_subgraph_t subgraph,
  size_t axis,
  size_t num_inputs,
  uint32_t* input_ids,
  uint32_t output_id,
  uint32_t flags)
{
  assert(num_inputs >= 2);
  assert(num_inputs <= 4);

  enum xnn_status status;
  if ((status = xnn_subgraph_check_xnnpack_initialized(node_type)) != xnn_status_success) {
    return status;
  }

  status = xnn_subgraph_check_output_node_id(node_type, output_id, subgraph->num_values);
  if (status != xnn_status_success) {
    return status;
  }

  const struct xnn_value* output_value = &subgraph->values[output_id];

  status = xnn_subgraph_check_output_type_dense(node_type, output_id, output_value);
  if (status != xnn_status_success) {
    return status;
  }

  if (axis >= output_value->shape.num_dims) {
    xnn_log_error(
      "failed to define %s operator with the output ID #%" PRIu32
      ": axis (%zu) exceeds the number of dimensions (%zu)",
      xnn_node_type_to_string(node_type), output_id, axis, output_value->shape.num_dims);
    return xnn_status_invalid_parameter;
  }

  for (size_t i = 0; i < num_inputs; i++) {
    status = check_input_value(subgraph, axis, input_ids[i], output_id, i+1, node_type);
    if (status != xnn_status_success) {
      return status;
    }
  }

  size_t input_axis_dimensions_sum = 0;
  for (size_t i = 0; i < num_inputs; i++) {
    const struct xnn_value* input_value = &subgraph->values[input_ids[i]];
    input_axis_dimensions_sum += input_value->shape.dim[axis];
  }

  if (output_value->shape.dim[axis] != input_axis_dimensions_sum) {
    xnn_log_error(
      "failed to define %s operator with output ID #%" PRIu32
      ": mismatch axis dimension %zu, output has %zu, sum of input dimensions is %zu",
      xnn_node_type_to_string(node_type), output_id, axis, output_value->shape.dim[axis], input_axis_dimensions_sum);
    return xnn_status_invalid_parameter;
  }

  enum xnn_compute_type compute_type = xnn_compute_type_invalid;
  switch (output_value->datatype) {
#ifndef XNN_NO_F16_OPERATORS
    case xnn_datatype_fp16:
      compute_type = xnn_compute_type_fp16;
      break;
#endif  // !defined(XNN_NO_F16_OPERATORS)
    case xnn_datatype_fp32:
      compute_type = xnn_compute_type_fp32;
      break;
#ifndef XNN_NO_QS8_OPERATORS
    case xnn_datatype_qint8:
      compute_type = xnn_compute_type_qs8;
      break;
#endif  // !defined(XNN_NO_QS8_OPERATORS)
#ifndef XNN_NO_QU8_OPERATORS
    case xnn_datatype_quint8:
      compute_type = xnn_compute_type_qu8;
      break;
#endif  // !defined(XNN_NO_QU8_OPERATORS)
    default:
      xnn_log_error(
        "failed to define %s operator with output ID #%" PRIu32 ": unsupported Value datatype %s (%d)",
        xnn_node_type_to_string(node_type), output_id,
        xnn_datatype_to_string(output_value->datatype), output_value->datatype);
      return xnn_status_invalid_parameter;
  }

  #if !defined(XNN_NO_QS8_OPERATORS) || !defined(XNN_NO_QU8_OPERATORS)
    if (compute_type == xnn_compute_type_qs8 || compute_type == xnn_compute_type_qu8) {
      check_input_compute_type(subgraph, input_ids[0], output_id, "first", node_type);
      check_input_compute_type(subgraph, input_ids[1], output_id, "second", node_type);
    }
    if (num_inputs > 2) {
      check_input_compute_type(subgraph, input_ids[2], output_id, "third", node_type);
    }
    if (num_inputs > 3) {
      check_input_compute_type(subgraph, input_ids[3], output_id, "fourth", node_type);
    }
  #endif  // !defined( XNN_NO_QS8_OPERATORS) || !defined(XNN_NO_QU8_OPERATORS)

  struct xnn_node* node = xnn_subgraph_new_node(subgraph);
  if (node == NULL) {
    return xnn_status_out_of_memory;
  }

  node->params.concatenate.axis = axis;
  node->type = node_type;
  node->compute_type = compute_type;
  node->num_inputs = num_inputs;
  node->inputs[0] = input_ids[0];
  node->inputs[1] = input_ids[1];
  node->num_outputs = 1;
  node->outputs[0] = output_id;
  node->flags = flags;

  switch (num_inputs) {
    case 2:
      node->create = create_concatenate2_operator;
      node->setup = setup_concatenate2_operator;
      break;
    case 3:
      node->create = create_concatenate3_operator;
      node->setup = setup_concatenate3_operator;
      node->inputs[2] = input_ids[2];
      break;
    case 4:
      node->create = create_concatenate4_operator;
      node->setup = setup_concatenate4_operator;
      node->inputs[2] = input_ids[2];
      node->inputs[3] = input_ids[3];
      break;
    default:
      XNN_UNREACHABLE;
  }

  return xnn_status_success;
}

enum xnn_status xnn_define_concatenate2(
  xnn_subgraph_t subgraph,
  size_t axis,
  uint32_t input1_id,
  uint32_t input2_id,
  uint32_t output_id,
  uint32_t flags)
{
  uint32_t input_ids[2] = { input1_id, input2_id };
  return xnn_define_concatenate_n(
    xnn_node_type_concatenate2, subgraph, axis, XNN_COUNT_OF(input_ids), input_ids, output_id, flags);
}

enum xnn_status xnn_define_concatenate3(
  xnn_subgraph_t subgraph,
  size_t axis,
  uint32_t input1_id,
  uint32_t input2_id,
  uint32_t input3_id,
  uint32_t output_id,
  uint32_t flags)
{
  uint32_t input_ids[3] = { input1_id, input2_id, input3_id };
  return xnn_define_concatenate_n(
    xnn_node_type_concatenate3, subgraph, axis, XNN_COUNT_OF(input_ids), input_ids, output_id, flags);
}

enum xnn_status xnn_define_concatenate4(
  xnn_subgraph_t subgraph,
  size_t axis,
  uint32_t input1_id,
  uint32_t input2_id,
  uint32_t input3_id,
  uint32_t input4_id,
  uint32_t output_id,
  uint32_t flags)
{
  uint32_t input_ids[4] = { input1_id, input2_id, input3_id, input4_id };
  return xnn_define_concatenate_n(
    xnn_node_type_concatenate4, subgraph, axis, XNN_COUNT_OF(input_ids), input_ids, output_id, flags);
}