test/even-split2.cc - platform/external/XNNPACK - Git at Google

 // 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 <algorithm>
 #include <array>
 #include <cstddef>
 #include <cstdint>
 #include <limits>
 #include <memory>
 #include <numeric>
 #include <random>

 #include <xnnpack.h>
 #include <xnnpack/node-type.h>
 #include <xnnpack/operator.h>
 #include <xnnpack/subgraph.h>

 #include <gtest/gtest.h>

 template <typename T> class EvenSplit2Test : public ::testing::Test {
 protected:
   EvenSplit2Test()
   {
     random_device = std::unique_ptr<std::random_device>(new std::random_device());
     rng = std::mt19937((*random_device)());
     shape_dist = std::uniform_int_distribution<size_t>(1, XNN_MAX_TENSOR_DIMS);
     dim_dist = std::uniform_int_distribution<size_t>(1, 9);
     f32dist = std::uniform_real_distribution<float>();
     i8dist =
       std::uniform_int_distribution<int32_t>(std::numeric_limits<int8_t>::min(), std::numeric_limits<int8_t>::max());
     u8dist =
       std::uniform_int_distribution<int32_t>(std::numeric_limits<uint8_t>::min(), std::numeric_limits<uint8_t>::max());
     scale_dist = std::uniform_real_distribution<float>(0.1f, 5.0f);

     output1_dims = RandomShape();
     axis = RandomAxis(output1_dims);
     output2_dims = output1_dims;
     input_dims = output1_dims;
     input_dims[axis] = output1_dims[axis] + output2_dims[axis];

     input = std::vector<T>(NumElements(input_dims));
     operator_output1 = std::vector<T>(NumElements(output1_dims));
     operator_output2 = std::vector<T>(NumElements(output2_dims));
     subgraph_output1 = std::vector<T>(NumElements(output1_dims));
     subgraph_output2 = std::vector<T>(NumElements(output2_dims));

     signed_zero_point = i8dist(rng);
     unsigned_zero_point = u8dist(rng);
     scale = scale_dist(rng);

     batch_size = 1;
     input_stride = 1;
     for (size_t i = 0; i < axis; i++) {
       batch_size *= input_dims[i];
     }

     for (size_t i = axis; i < input_dims.size(); i++) {
       input_stride *= input_dims[i];
     }
     channels = input_stride / 2;
   }

   std::vector<size_t> RandomShape()
   {
     std::vector<size_t> dims(shape_dist(rng));
     std::generate(dims.begin(), dims.end(), [&] { return dim_dist(rng); });
     return dims;
   }

   size_t RandomAxis(const std::vector<size_t>& dims)
   {
     return std::uniform_int_distribution<size_t>(0, dims.size() - 1)(rng);
   }

   size_t NumElements(const std::vector<size_t>& dims)
   {
     return std::accumulate(dims.begin(), dims.end(), size_t(1), std::multiplies<size_t>());
   }

   std::unique_ptr<std::random_device> random_device;
   std::mt19937 rng;
   std::uniform_int_distribution<size_t> shape_dist;
   std::uniform_int_distribution<size_t> dim_dist;
   std::uniform_real_distribution<float> f32dist;
   std::uniform_int_distribution<int32_t> i8dist;
   std::uniform_int_distribution<int32_t> u8dist;
   std::uniform_real_distribution<float> scale_dist;

   uint32_t output1_id;
   uint32_t output2_id;
   uint32_t input_id;

   std::vector<size_t> output1_dims;
   std::vector<size_t> output2_dims;
   std::vector<size_t> input_dims;

   size_t axis;
   size_t batch_size;
   size_t channels;
   size_t input_stride;

   int32_t signed_zero_point;
   int32_t unsigned_zero_point;
   float scale;

   std::vector<T> operator_output1;
   std::vector<T> operator_output2;
   std::vector<T> subgraph_output1;
   std::vector<T> subgraph_output2;
   std::vector<T> input;
 };

 using EvenSplit2TestQS8 = EvenSplit2Test<int8_t>;
 using EvenSplit2TestQU8 = EvenSplit2Test<uint8_t>;
 using EvenSplit2TestF32 = EvenSplit2Test<float>;

 TEST_F(EvenSplit2TestQS8, define)
 {
   ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));

   xnn_subgraph_t subgraph = nullptr;
   ASSERT_EQ(xnn_status_success, xnn_create_subgraph(/*external_value_ids=*/3, /*flags=*/0, &subgraph));
   std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

   input_id = XNN_INVALID_NODE_ID;
   ASSERT_EQ(
     xnn_status_success,
     xnn_define_quantized_tensor_value(
       subgraph, xnn_datatype_qint8, signed_zero_point, scale, input_dims.size(), input_dims.data(), nullptr, 0,
       /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
   ASSERT_NE(input_id, XNN_INVALID_NODE_ID);

   output1_id = XNN_INVALID_NODE_ID;
   ASSERT_EQ(
     xnn_status_success,
     xnn_define_quantized_tensor_value(
       subgraph, xnn_datatype_qint8, signed_zero_point, scale, output1_dims.size(), output1_dims.data(), nullptr, 1,
       /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output1_id));
   ASSERT_NE(output1_id, XNN_INVALID_NODE_ID);

   output2_id = XNN_INVALID_NODE_ID;
   ASSERT_EQ(
     xnn_status_success,
     xnn_define_quantized_tensor_value(
       subgraph, xnn_datatype_qint8, signed_zero_point, scale, output2_dims.size(), output2_dims.data(), nullptr, 2,
       /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output2_id));
   ASSERT_NE(output2_id, XNN_INVALID_NODE_ID);

   ASSERT_EQ(xnn_status_success, xnn_define_even_split2(subgraph, axis, input_id, output1_id, output2_id, /*flags=*/0));

   ASSERT_EQ(subgraph->num_nodes, 1);
   const struct xnn_node* node = &subgraph->nodes[0];
   ASSERT_EQ(node->type, xnn_node_type_even_split2);
   ASSERT_EQ(node->compute_type, xnn_compute_type_qs8);
   ASSERT_EQ(node->params.even_split.axis, axis);
   ASSERT_EQ(node->num_inputs, 1);
   ASSERT_EQ(node->inputs[0], input_id);
   ASSERT_EQ(node->num_outputs, 2);
   ASSERT_EQ(node->outputs[0], output1_id);
   ASSERT_EQ(node->outputs[1], output2_id);
   ASSERT_EQ(node->flags, 0);
 }

 TEST_F(EvenSplit2TestQU8, define)
 {
   ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));

   xnn_subgraph_t subgraph = nullptr;
   ASSERT_EQ(xnn_status_success, xnn_create_subgraph(/*external_value_ids=*/3, /*flags=*/0, &subgraph));
   std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

   input_id = XNN_INVALID_NODE_ID;
   ASSERT_EQ(
     xnn_status_success,
     xnn_define_quantized_tensor_value(
       subgraph, xnn_datatype_quint8, unsigned_zero_point, scale, input_dims.size(), input_dims.data(), nullptr, 0,
       /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
   ASSERT_NE(input_id, XNN_INVALID_NODE_ID);

   output1_id = XNN_INVALID_NODE_ID;
   ASSERT_EQ(
     xnn_status_success,
     xnn_define_quantized_tensor_value(
       subgraph, xnn_datatype_quint8, unsigned_zero_point, scale, output1_dims.size(), output1_dims.data(), nullptr, 1,
       /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output1_id));
   ASSERT_NE(output1_id, XNN_INVALID_NODE_ID);

   output2_id = XNN_INVALID_NODE_ID;
   ASSERT_EQ(
     xnn_status_success,
     xnn_define_quantized_tensor_value(
       subgraph, xnn_datatype_quint8, unsigned_zero_point, scale, output2_dims.size(), output2_dims.data(), nullptr, 2,
       /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output2_id));
   ASSERT_NE(output2_id, XNN_INVALID_NODE_ID);

   ASSERT_EQ(xnn_status_success, xnn_define_even_split2(subgraph, axis, input_id, output1_id, output2_id, /*flags=*/0));

   ASSERT_EQ(subgraph->num_nodes, 1);
   const struct xnn_node* node = &subgraph->nodes[0];
   ASSERT_EQ(node->type, xnn_node_type_even_split2);
   ASSERT_EQ(node->compute_type, xnn_compute_type_qu8);
   ASSERT_EQ(node->params.even_split.axis, axis);
   ASSERT_EQ(node->num_inputs, 1);
   ASSERT_EQ(node->inputs[0], input_id);
   ASSERT_EQ(node->num_outputs, 2);
   ASSERT_EQ(node->outputs[0], output1_id);
   ASSERT_EQ(node->outputs[1], output2_id);
   ASSERT_EQ(node->flags, 0);
 }

 TEST_F(EvenSplit2TestF32, define)
 {
   ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));

   xnn_subgraph_t subgraph = nullptr;
   ASSERT_EQ(xnn_status_success, xnn_create_subgraph(/*external_value_ids=*/3, /*flags=*/0, &subgraph));
   std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

   input_id = XNN_INVALID_NODE_ID;
   ASSERT_EQ(
     xnn_status_success, xnn_define_tensor_value(
                           subgraph, xnn_datatype_fp32, input_dims.size(), input_dims.data(), nullptr, 0,
                           /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
   ASSERT_NE(input_id, XNN_INVALID_NODE_ID);

   output1_id = XNN_INVALID_NODE_ID;
   ASSERT_EQ(
     xnn_status_success, xnn_define_tensor_value(
                           subgraph, xnn_datatype_fp32, output1_dims.size(), output1_dims.data(), nullptr, 1,
                           /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output1_id));
   ASSERT_NE(output1_id, XNN_INVALID_NODE_ID);

   output2_id = XNN_INVALID_NODE_ID;
   ASSERT_EQ(
     xnn_status_success, xnn_define_tensor_value(
                           subgraph, xnn_datatype_fp32, output2_dims.size(), output2_dims.data(), nullptr, 2,
                           /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output2_id));
   ASSERT_NE(output2_id, XNN_INVALID_NODE_ID);
   ASSERT_EQ(xnn_status_success, xnn_define_even_split2(subgraph, axis, input_id, output1_id, output2_id, /*flags=*/0));

   ASSERT_EQ(subgraph->num_nodes, 1);
   const struct xnn_node* node = &subgraph->nodes[0];
   ASSERT_EQ(node->type, xnn_node_type_even_split2);
   ASSERT_EQ(node->compute_type, xnn_compute_type_fp32);
   ASSERT_EQ(node->params.even_split.axis, axis);
   ASSERT_EQ(node->num_inputs, 1);
   ASSERT_EQ(node->inputs[0], input_id);
   ASSERT_EQ(node->num_outputs, 2);
   ASSERT_EQ(node->outputs[0], output1_id);
   ASSERT_EQ(node->outputs[1], output2_id);
   ASSERT_EQ(node->flags, 0);
 }

 TEST_F(EvenSplit2TestQS8, matches_operator_api)
 {
   std::generate(input.begin(), input.end(), [&]() { return i8dist(rng); });
   std::fill(operator_output1.begin(), operator_output1.end(), INT8_C(0xA5));
   std::fill(operator_output2.begin(), operator_output2.end(), INT8_C(0xA5));
   std::fill(subgraph_output1.begin(), subgraph_output1.end(), INT8_C(0xA5));
   std::fill(subgraph_output2.begin(), subgraph_output2.end(), INT8_C(0xA5));

   ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));

   xnn_operator_t op1 = nullptr;
   xnn_operator_t op2 = nullptr;

   // Call operator API.
   ASSERT_EQ(xnn_status_success, xnn_create_copy_nc_x8(channels, input_stride, channels, /*flags=*/0, &op1));
   std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op1(op1, xnn_delete_operator);
   ASSERT_EQ(xnn_status_success, xnn_create_copy_nc_x8(channels, input_stride, channels, /*flags=*/0, &op2));
   std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op2(op2, xnn_delete_operator);

   ASSERT_EQ(
     xnn_status_success,
     xnn_setup_copy_nc_x8(op1, batch_size, input.data(), operator_output1.data(), nullptr /* thread pool */));
   ASSERT_EQ(
     xnn_status_success,
     xnn_setup_copy_nc_x8(
       op2, batch_size, (uint8_t*) input.data() + op1->channels, operator_output2.data(), nullptr /* thread pool */));

   ASSERT_EQ(xnn_status_success, xnn_run_operator(op1, nullptr /* thread pool */));
   ASSERT_EQ(xnn_status_success, xnn_run_operator(op2, nullptr /* thread pool */));

   // Call subgraph API.
   xnn_subgraph_t subgraph = nullptr;
   ASSERT_EQ(xnn_status_success, xnn_create_subgraph(/*external_value_ids=*/3, /*flags=*/0, &subgraph));
   std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

   input_id = XNN_INVALID_NODE_ID;
   ASSERT_EQ(
     xnn_status_success,
     xnn_define_quantized_tensor_value(
       subgraph, xnn_datatype_qint8, signed_zero_point, scale, input_dims.size(), input_dims.data(), nullptr, 0,
       /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
   ASSERT_NE(input_id, XNN_INVALID_NODE_ID);

   output1_id = XNN_INVALID_NODE_ID;
   ASSERT_EQ(
     xnn_status_success,
     xnn_define_quantized_tensor_value(
       subgraph, xnn_datatype_qint8, signed_zero_point, scale, output1_dims.size(), output1_dims.data(), nullptr, 1,
       /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output1_id));
   ASSERT_NE(output1_id, XNN_INVALID_NODE_ID);

   output2_id = XNN_INVALID_NODE_ID;
   ASSERT_EQ(
     xnn_status_success,
     xnn_define_quantized_tensor_value(
       subgraph, xnn_datatype_qint8, signed_zero_point, scale, output2_dims.size(), output2_dims.data(), nullptr, 2,
       /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output2_id));
   ASSERT_NE(output2_id, XNN_INVALID_NODE_ID);

   ASSERT_EQ(xnn_status_success, xnn_define_even_split2(subgraph, axis, input_id, output1_id, output2_id, /*flags=*/0));

   xnn_runtime_t runtime = nullptr;
   ASSERT_EQ(xnn_status_success, xnn_create_runtime_v3(subgraph, nullptr, nullptr, /*flags=*/0, &runtime));
   ASSERT_NE(nullptr, runtime);
   std::unique_ptr<xnn_runtime, decltype(&xnn_delete_runtime)> auto_runtime(runtime, xnn_delete_runtime);
   std::array<xnn_external_value, 3> external = {
     xnn_external_value{input_id, input.data()},
     xnn_external_value{output1_id, subgraph_output1.data()},
     xnn_external_value{output2_id, subgraph_output2.data()},
   };
   ASSERT_EQ(xnn_status_success, xnn_setup_runtime(runtime, external.size(), external.data()));
   ASSERT_EQ(xnn_status_success, xnn_invoke_runtime(runtime));

   ASSERT_EQ(subgraph_output1, operator_output1);
   ASSERT_EQ(subgraph_output2, operator_output2);
 }

 TEST_F(EvenSplit2TestQU8, matches_operator_api)
 {
   std::generate(input.begin(), input.end(), [&]() { return u8dist(rng); });
   std::fill(operator_output1.begin(), operator_output1.end(), UINT8_C(0xA5));
   std::fill(operator_output2.begin(), operator_output2.end(), UINT8_C(0xA5));
   std::fill(subgraph_output1.begin(), subgraph_output1.end(), UINT8_C(0xA5));
   std::fill(subgraph_output2.begin(), subgraph_output2.end(), UINT8_C(0xA5));

   ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));

   xnn_operator_t op1 = nullptr;
   xnn_operator_t op2 = nullptr;

   // Call operator API.
   ASSERT_EQ(xnn_status_success, xnn_create_copy_nc_x8(channels, input_stride, channels, /*flags=*/0, &op1));
   std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op1(op1, xnn_delete_operator);
   ASSERT_EQ(xnn_status_success, xnn_create_copy_nc_x8(channels, input_stride, channels, /*flags=*/0, &op2));
   std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op2(op2, xnn_delete_operator);

   ASSERT_EQ(
     xnn_status_success,
     xnn_setup_copy_nc_x8(op1, batch_size, input.data(), operator_output1.data(), nullptr /* thread pool */));
   ASSERT_EQ(
     xnn_status_success,
     xnn_setup_copy_nc_x8(
       op2, batch_size, (uint8_t*) input.data() + op1->channels, operator_output2.data(), nullptr /* thread pool */));

   ASSERT_EQ(xnn_status_success, xnn_run_operator(op1, nullptr /* thread pool */));
   ASSERT_EQ(xnn_status_success, xnn_run_operator(op2, nullptr /* thread pool */));

   // Call subgraph API.
   xnn_subgraph_t subgraph = nullptr;
   ASSERT_EQ(xnn_status_success, xnn_create_subgraph(/*external_value_ids=*/3, /*flags=*/0, &subgraph));
   std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

   input_id = XNN_INVALID_NODE_ID;
   ASSERT_EQ(
     xnn_status_success,
     xnn_define_quantized_tensor_value(
       subgraph, xnn_datatype_quint8, unsigned_zero_point, scale, input_dims.size(), input_dims.data(), nullptr, 0,
       /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
   ASSERT_NE(input_id, XNN_INVALID_NODE_ID);

   output1_id = XNN_INVALID_NODE_ID;
   ASSERT_EQ(
     xnn_status_success,
     xnn_define_quantized_tensor_value(
       subgraph, xnn_datatype_quint8, unsigned_zero_point, scale, output1_dims.size(), output1_dims.data(), nullptr, 1,
       /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output1_id));
   ASSERT_NE(output1_id, XNN_INVALID_NODE_ID);

   output2_id = XNN_INVALID_NODE_ID;
   ASSERT_EQ(
     xnn_status_success,
     xnn_define_quantized_tensor_value(
       subgraph, xnn_datatype_quint8, unsigned_zero_point, scale, output2_dims.size(), output2_dims.data(), nullptr, 2,
       /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output2_id));
   ASSERT_NE(output2_id, XNN_INVALID_NODE_ID);

   ASSERT_EQ(xnn_status_success, xnn_define_even_split2(subgraph, axis, input_id, output1_id, output2_id, /*flags=*/0));

   xnn_runtime_t runtime = nullptr;
   ASSERT_EQ(xnn_status_success, xnn_create_runtime_v3(subgraph, nullptr, nullptr, /*flags=*/0, &runtime));
   ASSERT_NE(nullptr, runtime);
   std::unique_ptr<xnn_runtime, decltype(&xnn_delete_runtime)> auto_runtime(runtime, xnn_delete_runtime);
   std::array<xnn_external_value, 3> external = {
     xnn_external_value{input_id, input.data()},
     xnn_external_value{output1_id, subgraph_output1.data()},
     xnn_external_value{output2_id, subgraph_output2.data()},
   };
   ASSERT_EQ(xnn_status_success, xnn_setup_runtime(runtime, external.size(), external.data()));
   ASSERT_EQ(xnn_status_success, xnn_invoke_runtime(runtime));

   ASSERT_EQ(subgraph_output1, operator_output1);
   ASSERT_EQ(subgraph_output2, operator_output2);
 }

 TEST_F(EvenSplit2TestF32, matches_operator_api)
 {
   std::generate(input.begin(), input.end(), [&]() { return f32dist(rng); });
   std::fill(operator_output1.begin(), operator_output1.end(), std::nanf(""));
   std::fill(operator_output2.begin(), operator_output2.end(), std::nanf(""));
   std::fill(subgraph_output1.begin(), subgraph_output1.end(), std::nanf(""));
   std::fill(subgraph_output2.begin(), subgraph_output2.end(), std::nanf(""));

   ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));

   xnn_operator_t op1 = nullptr;
   xnn_operator_t op2 = nullptr;

   // Call operator API.
   ASSERT_EQ(xnn_status_success, xnn_create_copy_nc_x32(channels, input_stride, channels, /*flags=*/0, &op1));
   std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op1(op1, xnn_delete_operator);
   ASSERT_EQ(xnn_status_success, xnn_create_copy_nc_x32(channels, input_stride, channels, /*flags=*/0, &op2));
   std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op2(op2, xnn_delete_operator);

   ASSERT_EQ(
     xnn_status_success,
     xnn_setup_copy_nc_x32(op1, batch_size, input.data(), operator_output1.data(), nullptr /* thread pool */));
   ASSERT_EQ(
     xnn_status_success,
     xnn_setup_copy_nc_x32(
       op2, batch_size, (uint32_t*) input.data() + op1->channels, operator_output2.data(), nullptr /* thread pool */));

   ASSERT_EQ(xnn_status_success, xnn_run_operator(op1, nullptr /* thread pool */));
   ASSERT_EQ(xnn_status_success, xnn_run_operator(op2, nullptr /* thread pool */));

   // Call subgraph API.
   xnn_subgraph_t subgraph = nullptr;
   ASSERT_EQ(xnn_status_success, xnn_create_subgraph(/*external_value_ids=*/3, /*flags=*/0, &subgraph));
   std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

   input_id = XNN_INVALID_NODE_ID;
   ASSERT_EQ(
     xnn_status_success, xnn_define_tensor_value(
                           subgraph, xnn_datatype_fp32, input_dims.size(), input_dims.data(), nullptr, 0,
                           /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
   ASSERT_NE(input_id, XNN_INVALID_NODE_ID);

   output1_id = XNN_INVALID_NODE_ID;
   ASSERT_EQ(
     xnn_status_success, xnn_define_tensor_value(
                           subgraph, xnn_datatype_fp32, output1_dims.size(), output1_dims.data(), nullptr, 1,
                           /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output1_id));
   ASSERT_NE(output1_id, XNN_INVALID_NODE_ID);

   output2_id = XNN_INVALID_NODE_ID;
   ASSERT_EQ(
     xnn_status_success, xnn_define_tensor_value(
                           subgraph, xnn_datatype_fp32, output2_dims.size(), output2_dims.data(), nullptr, 2,
                           /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output2_id));
   ASSERT_NE(output2_id, XNN_INVALID_NODE_ID);

   ASSERT_EQ(xnn_status_success, xnn_define_even_split2(subgraph, axis, input_id, output1_id, output2_id, /*flags=*/0));

   xnn_runtime_t runtime = nullptr;
   ASSERT_EQ(xnn_status_success, xnn_create_runtime_v3(subgraph, nullptr, nullptr, /*flags=*/0, &runtime));
   ASSERT_NE(nullptr, runtime);
   std::unique_ptr<xnn_runtime, decltype(&xnn_delete_runtime)> auto_runtime(runtime, xnn_delete_runtime);
   std::array<xnn_external_value, 3> external = {
     xnn_external_value{input_id, input.data()},
     xnn_external_value{output1_id, subgraph_output1.data()},
     xnn_external_value{output2_id, subgraph_output2.data()},
   };
   ASSERT_EQ(xnn_status_success, xnn_setup_runtime(runtime, external.size(), external.data()));
   ASSERT_EQ(xnn_status_success, xnn_invoke_runtime(runtime));

   ASSERT_EQ(subgraph_output1, operator_output1);
   ASSERT_EQ(subgraph_output2, operator_output2);
 }
	// 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 <algorithm>
	#include <array>
	#include <cstddef>
	#include <cstdint>
	#include <limits>
	#include <memory>
	#include <numeric>
	#include <random>

	#include <xnnpack.h>
	#include <xnnpack/node-type.h>
	#include <xnnpack/operator.h>
	#include <xnnpack/subgraph.h>

	#include <gtest/gtest.h>

	template <typename T> class EvenSplit2Test : public ::testing::Test {
	protected:
	EvenSplit2Test()
	{
	random_device = std::unique_ptr<std::random_device>(new std::random_device());
	rng = std::mt19937((*random_device)());
	shape_dist = std::uniform_int_distribution<size_t>(1, XNN_MAX_TENSOR_DIMS);
	dim_dist = std::uniform_int_distribution<size_t>(1, 9);
	f32dist = std::uniform_real_distribution<float>();
	i8dist =
	std::uniform_int_distribution<int32_t>(std::numeric_limits<int8_t>::min(), std::numeric_limits<int8_t>::max());
	u8dist =
	std::uniform_int_distribution<int32_t>(std::numeric_limits<uint8_t>::min(), std::numeric_limits<uint8_t>::max());
	scale_dist = std::uniform_real_distribution<float>(0.1f, 5.0f);

	output1_dims = RandomShape();
	axis = RandomAxis(output1_dims);
	output2_dims = output1_dims;
	input_dims = output1_dims;
	input_dims[axis] = output1_dims[axis] + output2_dims[axis];

	input = std::vector<T>(NumElements(input_dims));
	operator_output1 = std::vector<T>(NumElements(output1_dims));
	operator_output2 = std::vector<T>(NumElements(output2_dims));
	subgraph_output1 = std::vector<T>(NumElements(output1_dims));
	subgraph_output2 = std::vector<T>(NumElements(output2_dims));

	signed_zero_point = i8dist(rng);
	unsigned_zero_point = u8dist(rng);
	scale = scale_dist(rng);

	batch_size = 1;
	input_stride = 1;
	for (size_t i = 0; i < axis; i++) {
	batch_size *= input_dims[i];
	}

	for (size_t i = axis; i < input_dims.size(); i++) {
	input_stride *= input_dims[i];
	}
	channels = input_stride / 2;
	}

	std::vector<size_t> RandomShape()
	{
	std::vector<size_t> dims(shape_dist(rng));
	std::generate(dims.begin(), dims.end(), [&] { return dim_dist(rng); });
	return dims;
	}

	size_t RandomAxis(const std::vector<size_t>& dims)
	{
	return std::uniform_int_distribution<size_t>(0, dims.size() - 1)(rng);
	}

	size_t NumElements(const std::vector<size_t>& dims)
	{
	return std::accumulate(dims.begin(), dims.end(), size_t(1), std::multiplies<size_t>());
	}

	std::unique_ptr<std::random_device> random_device;
	std::mt19937 rng;
	std::uniform_int_distribution<size_t> shape_dist;
	std::uniform_int_distribution<size_t> dim_dist;
	std::uniform_real_distribution<float> f32dist;
	std::uniform_int_distribution<int32_t> i8dist;
	std::uniform_int_distribution<int32_t> u8dist;
	std::uniform_real_distribution<float> scale_dist;

	uint32_t output1_id;
	uint32_t output2_id;
	uint32_t input_id;

	std::vector<size_t> output1_dims;
	std::vector<size_t> output2_dims;
	std::vector<size_t> input_dims;

	size_t axis;
	size_t batch_size;
	size_t channels;
	size_t input_stride;

	int32_t signed_zero_point;
	int32_t unsigned_zero_point;
	float scale;

	std::vector<T> operator_output1;
	std::vector<T> operator_output2;
	std::vector<T> subgraph_output1;
	std::vector<T> subgraph_output2;
	std::vector<T> input;
	};

	using EvenSplit2TestQS8 = EvenSplit2Test<int8_t>;
	using EvenSplit2TestQU8 = EvenSplit2Test<uint8_t>;
	using EvenSplit2TestF32 = EvenSplit2Test<float>;

	TEST_F(EvenSplit2TestQS8, define)
	{
	ASSERT_EQ(xnn_status_success, xnn_initialize(/allocator=/nullptr));

	xnn_subgraph_t subgraph = nullptr;
	ASSERT_EQ(xnn_status_success, xnn_create_subgraph(/external_value_ids=/3, /flags=/0, &subgraph));
	std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

	input_id = XNN_INVALID_NODE_ID;
	ASSERT_EQ(
	xnn_status_success,
	xnn_define_quantized_tensor_value(
	subgraph, xnn_datatype_qint8, signed_zero_point, scale, input_dims.size(), input_dims.data(), nullptr, 0,
	/flags=/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
	ASSERT_NE(input_id, XNN_INVALID_NODE_ID);

	output1_id = XNN_INVALID_NODE_ID;
	ASSERT_EQ(
	xnn_status_success,
	xnn_define_quantized_tensor_value(
	subgraph, xnn_datatype_qint8, signed_zero_point, scale, output1_dims.size(), output1_dims.data(), nullptr, 1,
	/flags=/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output1_id));
	ASSERT_NE(output1_id, XNN_INVALID_NODE_ID);

	output2_id = XNN_INVALID_NODE_ID;
	ASSERT_EQ(
	xnn_status_success,
	xnn_define_quantized_tensor_value(
	subgraph, xnn_datatype_qint8, signed_zero_point, scale, output2_dims.size(), output2_dims.data(), nullptr, 2,
	/flags=/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output2_id));
	ASSERT_NE(output2_id, XNN_INVALID_NODE_ID);

	ASSERT_EQ(xnn_status_success, xnn_define_even_split2(subgraph, axis, input_id, output1_id, output2_id, /flags=/0));

	ASSERT_EQ(subgraph->num_nodes, 1);
	const struct xnn_node* node = &subgraph->nodes[0];
	ASSERT_EQ(node->type, xnn_node_type_even_split2);
	ASSERT_EQ(node->compute_type, xnn_compute_type_qs8);
	ASSERT_EQ(node->params.even_split.axis, axis);
	ASSERT_EQ(node->num_inputs, 1);
	ASSERT_EQ(node->inputs[0], input_id);
	ASSERT_EQ(node->num_outputs, 2);
	ASSERT_EQ(node->outputs[0], output1_id);
	ASSERT_EQ(node->outputs[1], output2_id);
	ASSERT_EQ(node->flags, 0);
	}

	TEST_F(EvenSplit2TestQU8, define)
	{
	ASSERT_EQ(xnn_status_success, xnn_initialize(/allocator=/nullptr));

	xnn_subgraph_t subgraph = nullptr;
	ASSERT_EQ(xnn_status_success, xnn_create_subgraph(/external_value_ids=/3, /flags=/0, &subgraph));
	std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

	input_id = XNN_INVALID_NODE_ID;
	ASSERT_EQ(
	xnn_status_success,
	xnn_define_quantized_tensor_value(
	subgraph, xnn_datatype_quint8, unsigned_zero_point, scale, input_dims.size(), input_dims.data(), nullptr, 0,
	/flags=/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
	ASSERT_NE(input_id, XNN_INVALID_NODE_ID);

	output1_id = XNN_INVALID_NODE_ID;
	ASSERT_EQ(
	xnn_status_success,
	xnn_define_quantized_tensor_value(
	subgraph, xnn_datatype_quint8, unsigned_zero_point, scale, output1_dims.size(), output1_dims.data(), nullptr, 1,
	/flags=/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output1_id));
	ASSERT_NE(output1_id, XNN_INVALID_NODE_ID);

	output2_id = XNN_INVALID_NODE_ID;
	ASSERT_EQ(
	xnn_status_success,
	xnn_define_quantized_tensor_value(
	subgraph, xnn_datatype_quint8, unsigned_zero_point, scale, output2_dims.size(), output2_dims.data(), nullptr, 2,
	/flags=/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output2_id));
	ASSERT_NE(output2_id, XNN_INVALID_NODE_ID);

	ASSERT_EQ(xnn_status_success, xnn_define_even_split2(subgraph, axis, input_id, output1_id, output2_id, /flags=/0));

	ASSERT_EQ(subgraph->num_nodes, 1);
	const struct xnn_node* node = &subgraph->nodes[0];
	ASSERT_EQ(node->type, xnn_node_type_even_split2);
	ASSERT_EQ(node->compute_type, xnn_compute_type_qu8);
	ASSERT_EQ(node->params.even_split.axis, axis);
	ASSERT_EQ(node->num_inputs, 1);
	ASSERT_EQ(node->inputs[0], input_id);
	ASSERT_EQ(node->num_outputs, 2);
	ASSERT_EQ(node->outputs[0], output1_id);
	ASSERT_EQ(node->outputs[1], output2_id);
	ASSERT_EQ(node->flags, 0);
	}

	TEST_F(EvenSplit2TestF32, define)
	{
	ASSERT_EQ(xnn_status_success, xnn_initialize(/allocator=/nullptr));

	xnn_subgraph_t subgraph = nullptr;
	ASSERT_EQ(xnn_status_success, xnn_create_subgraph(/external_value_ids=/3, /flags=/0, &subgraph));
	std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

	input_id = XNN_INVALID_NODE_ID;
	ASSERT_EQ(
	xnn_status_success, xnn_define_tensor_value(
	subgraph, xnn_datatype_fp32, input_dims.size(), input_dims.data(), nullptr, 0,
	/flags=/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
	ASSERT_NE(input_id, XNN_INVALID_NODE_ID);

	output1_id = XNN_INVALID_NODE_ID;
	ASSERT_EQ(
	xnn_status_success, xnn_define_tensor_value(
	subgraph, xnn_datatype_fp32, output1_dims.size(), output1_dims.data(), nullptr, 1,
	/flags=/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output1_id));
	ASSERT_NE(output1_id, XNN_INVALID_NODE_ID);

	output2_id = XNN_INVALID_NODE_ID;
	ASSERT_EQ(
	xnn_status_success, xnn_define_tensor_value(
	subgraph, xnn_datatype_fp32, output2_dims.size(), output2_dims.data(), nullptr, 2,
	/flags=/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output2_id));
	ASSERT_NE(output2_id, XNN_INVALID_NODE_ID);
	ASSERT_EQ(xnn_status_success, xnn_define_even_split2(subgraph, axis, input_id, output1_id, output2_id, /flags=/0));

	ASSERT_EQ(subgraph->num_nodes, 1);
	const struct xnn_node* node = &subgraph->nodes[0];
	ASSERT_EQ(node->type, xnn_node_type_even_split2);
	ASSERT_EQ(node->compute_type, xnn_compute_type_fp32);
	ASSERT_EQ(node->params.even_split.axis, axis);
	ASSERT_EQ(node->num_inputs, 1);
	ASSERT_EQ(node->inputs[0], input_id);
	ASSERT_EQ(node->num_outputs, 2);
	ASSERT_EQ(node->outputs[0], output1_id);
	ASSERT_EQ(node->outputs[1], output2_id);
	ASSERT_EQ(node->flags, 0);
	}

	TEST_F(EvenSplit2TestQS8, matches_operator_api)
	{
	std::generate(input.begin(), input.end(), [&]() { return i8dist(rng); });
	std::fill(operator_output1.begin(), operator_output1.end(), INT8_C(0xA5));
	std::fill(operator_output2.begin(), operator_output2.end(), INT8_C(0xA5));
	std::fill(subgraph_output1.begin(), subgraph_output1.end(), INT8_C(0xA5));
	std::fill(subgraph_output2.begin(), subgraph_output2.end(), INT8_C(0xA5));

	ASSERT_EQ(xnn_status_success, xnn_initialize(/allocator=/nullptr));

	xnn_operator_t op1 = nullptr;
	xnn_operator_t op2 = nullptr;

	// Call operator API.
	ASSERT_EQ(xnn_status_success, xnn_create_copy_nc_x8(channels, input_stride, channels, /flags=/0, &op1));
	std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op1(op1, xnn_delete_operator);
	ASSERT_EQ(xnn_status_success, xnn_create_copy_nc_x8(channels, input_stride, channels, /flags=/0, &op2));
	std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op2(op2, xnn_delete_operator);

	ASSERT_EQ(
	xnn_status_success,
	xnn_setup_copy_nc_x8(op1, batch_size, input.data(), operator_output1.data(), nullptr /* thread pool */));
	ASSERT_EQ(
	xnn_status_success,
	xnn_setup_copy_nc_x8(
	op2, batch_size, (uint8_t) input.data() + op1->channels, operator_output2.data(), nullptr / thread pool */));

	ASSERT_EQ(xnn_status_success, xnn_run_operator(op1, nullptr /* thread pool */));
	ASSERT_EQ(xnn_status_success, xnn_run_operator(op2, nullptr /* thread pool */));

	// Call subgraph API.
	xnn_subgraph_t subgraph = nullptr;
	ASSERT_EQ(xnn_status_success, xnn_create_subgraph(/external_value_ids=/3, /flags=/0, &subgraph));
	std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

	input_id = XNN_INVALID_NODE_ID;
	ASSERT_EQ(
	xnn_status_success,
	xnn_define_quantized_tensor_value(
	subgraph, xnn_datatype_qint8, signed_zero_point, scale, input_dims.size(), input_dims.data(), nullptr, 0,
	/flags=/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
	ASSERT_NE(input_id, XNN_INVALID_NODE_ID);

	output1_id = XNN_INVALID_NODE_ID;
	ASSERT_EQ(
	xnn_status_success,
	xnn_define_quantized_tensor_value(
	subgraph, xnn_datatype_qint8, signed_zero_point, scale, output1_dims.size(), output1_dims.data(), nullptr, 1,
	/flags=/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output1_id));
	ASSERT_NE(output1_id, XNN_INVALID_NODE_ID);

	output2_id = XNN_INVALID_NODE_ID;
	ASSERT_EQ(
	xnn_status_success,
	xnn_define_quantized_tensor_value(
	subgraph, xnn_datatype_qint8, signed_zero_point, scale, output2_dims.size(), output2_dims.data(), nullptr, 2,
	/flags=/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output2_id));
	ASSERT_NE(output2_id, XNN_INVALID_NODE_ID);

	ASSERT_EQ(xnn_status_success, xnn_define_even_split2(subgraph, axis, input_id, output1_id, output2_id, /flags=/0));

	xnn_runtime_t runtime = nullptr;
	ASSERT_EQ(xnn_status_success, xnn_create_runtime_v3(subgraph, nullptr, nullptr, /flags=/0, &runtime));
	ASSERT_NE(nullptr, runtime);
	std::unique_ptr<xnn_runtime, decltype(&xnn_delete_runtime)> auto_runtime(runtime, xnn_delete_runtime);
	std::array<xnn_external_value, 3> external = {
	xnn_external_value{input_id, input.data()},
	xnn_external_value{output1_id, subgraph_output1.data()},
	xnn_external_value{output2_id, subgraph_output2.data()},
	};
	ASSERT_EQ(xnn_status_success, xnn_setup_runtime(runtime, external.size(), external.data()));
	ASSERT_EQ(xnn_status_success, xnn_invoke_runtime(runtime));

	ASSERT_EQ(subgraph_output1, operator_output1);
	ASSERT_EQ(subgraph_output2, operator_output2);
	}

	TEST_F(EvenSplit2TestQU8, matches_operator_api)
	{
	std::generate(input.begin(), input.end(), [&]() { return u8dist(rng); });
	std::fill(operator_output1.begin(), operator_output1.end(), UINT8_C(0xA5));
	std::fill(operator_output2.begin(), operator_output2.end(), UINT8_C(0xA5));
	std::fill(subgraph_output1.begin(), subgraph_output1.end(), UINT8_C(0xA5));
	std::fill(subgraph_output2.begin(), subgraph_output2.end(), UINT8_C(0xA5));

	ASSERT_EQ(xnn_status_success, xnn_initialize(/allocator=/nullptr));

	xnn_operator_t op1 = nullptr;
	xnn_operator_t op2 = nullptr;

	// Call operator API.
	ASSERT_EQ(xnn_status_success, xnn_create_copy_nc_x8(channels, input_stride, channels, /flags=/0, &op1));
	std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op1(op1, xnn_delete_operator);
	ASSERT_EQ(xnn_status_success, xnn_create_copy_nc_x8(channels, input_stride, channels, /flags=/0, &op2));
	std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op2(op2, xnn_delete_operator);

	ASSERT_EQ(
	xnn_status_success,
	xnn_setup_copy_nc_x8(op1, batch_size, input.data(), operator_output1.data(), nullptr /* thread pool */));
	ASSERT_EQ(
	xnn_status_success,
	xnn_setup_copy_nc_x8(
	op2, batch_size, (uint8_t) input.data() + op1->channels, operator_output2.data(), nullptr / thread pool */));

	ASSERT_EQ(xnn_status_success, xnn_run_operator(op1, nullptr /* thread pool */));
	ASSERT_EQ(xnn_status_success, xnn_run_operator(op2, nullptr /* thread pool */));

	// Call subgraph API.
	xnn_subgraph_t subgraph = nullptr;
	ASSERT_EQ(xnn_status_success, xnn_create_subgraph(/external_value_ids=/3, /flags=/0, &subgraph));
	std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

	input_id = XNN_INVALID_NODE_ID;
	ASSERT_EQ(
	xnn_status_success,
	xnn_define_quantized_tensor_value(
	subgraph, xnn_datatype_quint8, unsigned_zero_point, scale, input_dims.size(), input_dims.data(), nullptr, 0,
	/flags=/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
	ASSERT_NE(input_id, XNN_INVALID_NODE_ID);

	output1_id = XNN_INVALID_NODE_ID;
	ASSERT_EQ(
	xnn_status_success,
	xnn_define_quantized_tensor_value(
	subgraph, xnn_datatype_quint8, unsigned_zero_point, scale, output1_dims.size(), output1_dims.data(), nullptr, 1,
	/flags=/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output1_id));
	ASSERT_NE(output1_id, XNN_INVALID_NODE_ID);

	output2_id = XNN_INVALID_NODE_ID;
	ASSERT_EQ(
	xnn_status_success,
	xnn_define_quantized_tensor_value(
	subgraph, xnn_datatype_quint8, unsigned_zero_point, scale, output2_dims.size(), output2_dims.data(), nullptr, 2,
	/flags=/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output2_id));
	ASSERT_NE(output2_id, XNN_INVALID_NODE_ID);

	ASSERT_EQ(xnn_status_success, xnn_define_even_split2(subgraph, axis, input_id, output1_id, output2_id, /flags=/0));

	xnn_runtime_t runtime = nullptr;
	ASSERT_EQ(xnn_status_success, xnn_create_runtime_v3(subgraph, nullptr, nullptr, /flags=/0, &runtime));
	ASSERT_NE(nullptr, runtime);
	std::unique_ptr<xnn_runtime, decltype(&xnn_delete_runtime)> auto_runtime(runtime, xnn_delete_runtime);
	std::array<xnn_external_value, 3> external = {
	xnn_external_value{input_id, input.data()},
	xnn_external_value{output1_id, subgraph_output1.data()},
	xnn_external_value{output2_id, subgraph_output2.data()},
	};
	ASSERT_EQ(xnn_status_success, xnn_setup_runtime(runtime, external.size(), external.data()));
	ASSERT_EQ(xnn_status_success, xnn_invoke_runtime(runtime));

	ASSERT_EQ(subgraph_output1, operator_output1);
	ASSERT_EQ(subgraph_output2, operator_output2);
	}

	TEST_F(EvenSplit2TestF32, matches_operator_api)
	{
	std::generate(input.begin(), input.end(), [&]() { return f32dist(rng); });
	std::fill(operator_output1.begin(), operator_output1.end(), std::nanf(""));
	std::fill(operator_output2.begin(), operator_output2.end(), std::nanf(""));
	std::fill(subgraph_output1.begin(), subgraph_output1.end(), std::nanf(""));
	std::fill(subgraph_output2.begin(), subgraph_output2.end(), std::nanf(""));

	ASSERT_EQ(xnn_status_success, xnn_initialize(/allocator=/nullptr));

	xnn_operator_t op1 = nullptr;
	xnn_operator_t op2 = nullptr;

	// Call operator API.
	ASSERT_EQ(xnn_status_success, xnn_create_copy_nc_x32(channels, input_stride, channels, /flags=/0, &op1));
	std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op1(op1, xnn_delete_operator);
	ASSERT_EQ(xnn_status_success, xnn_create_copy_nc_x32(channels, input_stride, channels, /flags=/0, &op2));
	std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op2(op2, xnn_delete_operator);

	ASSERT_EQ(
	xnn_status_success,
	xnn_setup_copy_nc_x32(op1, batch_size, input.data(), operator_output1.data(), nullptr /* thread pool */));
	ASSERT_EQ(
	xnn_status_success,
	xnn_setup_copy_nc_x32(
	op2, batch_size, (uint32_t) input.data() + op1->channels, operator_output2.data(), nullptr / thread pool */));

	ASSERT_EQ(xnn_status_success, xnn_run_operator(op1, nullptr /* thread pool */));
	ASSERT_EQ(xnn_status_success, xnn_run_operator(op2, nullptr /* thread pool */));

	// Call subgraph API.
	xnn_subgraph_t subgraph = nullptr;
	ASSERT_EQ(xnn_status_success, xnn_create_subgraph(/external_value_ids=/3, /flags=/0, &subgraph));
	std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

	input_id = XNN_INVALID_NODE_ID;
	ASSERT_EQ(
	xnn_status_success, xnn_define_tensor_value(
	subgraph, xnn_datatype_fp32, input_dims.size(), input_dims.data(), nullptr, 0,
	/flags=/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
	ASSERT_NE(input_id, XNN_INVALID_NODE_ID);

	output1_id = XNN_INVALID_NODE_ID;
	ASSERT_EQ(
	xnn_status_success, xnn_define_tensor_value(
	subgraph, xnn_datatype_fp32, output1_dims.size(), output1_dims.data(), nullptr, 1,
	/flags=/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output1_id));
	ASSERT_NE(output1_id, XNN_INVALID_NODE_ID);

	output2_id = XNN_INVALID_NODE_ID;
	ASSERT_EQ(
	xnn_status_success, xnn_define_tensor_value(
	subgraph, xnn_datatype_fp32, output2_dims.size(), output2_dims.data(), nullptr, 2,
	/flags=/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output2_id));
	ASSERT_NE(output2_id, XNN_INVALID_NODE_ID);

	ASSERT_EQ(xnn_status_success, xnn_define_even_split2(subgraph, axis, input_id, output1_id, output2_id, /flags=/0));

	xnn_runtime_t runtime = nullptr;
	ASSERT_EQ(xnn_status_success, xnn_create_runtime_v3(subgraph, nullptr, nullptr, /flags=/0, &runtime));
	ASSERT_NE(nullptr, runtime);
	std::unique_ptr<xnn_runtime, decltype(&xnn_delete_runtime)> auto_runtime(runtime, xnn_delete_runtime);
	std::array<xnn_external_value, 3> external = {
	xnn_external_value{input_id, input.data()},
	xnn_external_value{output1_id, subgraph_output1.data()},
	xnn_external_value{output2_id, subgraph_output2.data()},
	};
	ASSERT_EQ(xnn_status_success, xnn_setup_runtime(runtime, external.size(), external.data()));
	ASSERT_EQ(xnn_status_success, xnn_invoke_runtime(runtime));

	ASSERT_EQ(subgraph_output1, operator_output1);
	ASSERT_EQ(subgraph_output2, operator_output2);
	}