blob: 4efa3b77c26e1ef00044b5f66e897f3a72fa4110 [file] [log] [blame] [edit]
// 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 <cassert>
#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 DepthToSpaceTest : public ::testing::Test {
protected:
DepthToSpaceTest()
{
random_device = std::unique_ptr<std::random_device>(new std::random_device());
rng = std::mt19937((*random_device)());
dim_dist = std::uniform_int_distribution<size_t>(1, 9);
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, 10.0f);
f32dist = std::uniform_real_distribution<float>(0.01f, 1.0f);
input_dims = RandomShape(4);
block_size = std::uniform_int_distribution<uint32_t>(2, 10)(rng);
uint32_t output_channels = dim_dist(rng);
output_dims = {input_dims[0], input_dims[1] * block_size, input_dims[2] * block_size, output_channels};
input_dims[3] = block_size * block_size * output_channels;
size_t num_output_elements = NumElements(output_dims);
input = std::vector<T>(NumElements(input_dims) + XNN_EXTRA_BYTES / sizeof(T));
operator_output = std::vector<T>(num_output_elements);
subgraph_output = std::vector<T>(num_output_elements);
}
size_t NumElements(std::vector<size_t>& dims)
{
return std::accumulate(dims.begin(), dims.end(), size_t(1), std::multiplies<size_t>());
}
std::vector<size_t> RandomShape(size_t num_dims)
{
std::vector<size_t> dims(num_dims);
std::generate(dims.begin(), dims.end(), [&] { return dim_dist(rng); });
return dims;
}
size_t batch_size()
{
assert(input_dims[0] == output_dims[0]);
return input_dims[0];
}
size_t input_height() { return input_dims[1]; }
size_t input_width() { return input_dims[2]; }
size_t input_channel() { return input_dims[3]; }
size_t output_channel() { return output_dims[3]; }
std::unique_ptr<std::random_device> random_device;
std::mt19937 rng;
std::uniform_int_distribution<size_t> dim_dist;
std::uniform_real_distribution<float> scale_dist;
std::uniform_int_distribution<int32_t> i8dist;
std::uniform_int_distribution<int32_t> u8dist;
std::uniform_real_distribution<float> f32dist;
std::vector<size_t> input_dims;
std::vector<size_t> output_dims;
std::vector<T> input;
std::vector<T> operator_output;
std::vector<T> subgraph_output;
uint32_t block_size;
uint32_t input_id;
uint32_t output_id;
};
using DepthToSpaceTestQS8 = DepthToSpaceTest<int8_t>;
using DepthToSpaceTestQU8 = DepthToSpaceTest<uint8_t>;
using DepthToSpaceTestF32 = DepthToSpaceTest<float>;
TEST_F(DepthToSpaceTestQS8, define)
{
const int32_t input_zero_point = i8dist(rng);
const float input_scale = scale_dist(rng);
const int32_t output_zero_point = input_zero_point;
const float output_scale = input_scale;
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=*/2, /*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, input_zero_point, input_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);
output_id = XNN_INVALID_NODE_ID;
ASSERT_EQ(
xnn_status_success, xnn_define_quantized_tensor_value(
subgraph, xnn_datatype_qint8, output_zero_point, output_scale, output_dims.size(),
output_dims.data(), nullptr, 1, /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output_id));
ASSERT_NE(output_id, XNN_INVALID_NODE_ID);
ASSERT_EQ(xnn_status_success, xnn_define_depth_to_space(subgraph, input_id, output_id, block_size, /*flags=*/0));
ASSERT_EQ(subgraph->num_nodes, 1);
const struct xnn_node* node = &subgraph->nodes[0];
ASSERT_EQ(node->type, xnn_node_type_depth_to_space);
ASSERT_EQ(node->compute_type, xnn_compute_type_qs8);
ASSERT_EQ(node->num_inputs, 1);
ASSERT_EQ(node->inputs[0], input_id);
ASSERT_EQ(node->num_outputs, 1);
ASSERT_EQ(node->outputs[0], output_id);
ASSERT_EQ(node->flags, 0);
}
TEST_F(DepthToSpaceTestQU8, define)
{
const int32_t input_zero_point = u8dist(rng);
const float input_scale = scale_dist(rng);
const int32_t output_zero_point = input_zero_point;
const float output_scale = input_scale;
uint32_t block_size = std::uniform_int_distribution<uint32_t>(2, 10)(rng);
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=*/2, /*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, input_zero_point, input_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);
output_id = XNN_INVALID_NODE_ID;
ASSERT_EQ(
xnn_status_success, xnn_define_quantized_tensor_value(
subgraph, xnn_datatype_quint8, output_zero_point, output_scale, output_dims.size(),
output_dims.data(), nullptr, 1, /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output_id));
ASSERT_NE(output_id, XNN_INVALID_NODE_ID);
ASSERT_EQ(xnn_status_success, xnn_define_depth_to_space(subgraph, input_id, output_id, block_size, /*flags=*/0));
ASSERT_EQ(subgraph->num_nodes, 1);
const struct xnn_node* node = &subgraph->nodes[0];
ASSERT_EQ(node->type, xnn_node_type_depth_to_space);
ASSERT_EQ(node->compute_type, xnn_compute_type_qu8);
ASSERT_EQ(node->num_inputs, 1);
ASSERT_EQ(node->inputs[0], input_id);
ASSERT_EQ(node->num_outputs, 1);
ASSERT_EQ(node->outputs[0], output_id);
ASSERT_EQ(node->flags, 0);
}
TEST_F(DepthToSpaceTestF32, define)
{
uint32_t block_size = std::uniform_int_distribution<uint32_t>(2, 10)(rng);
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=*/2, /*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);
output_id = XNN_INVALID_NODE_ID;
ASSERT_EQ(
xnn_status_success, xnn_define_tensor_value(
subgraph, xnn_datatype_fp32, output_dims.size(), output_dims.data(), nullptr, 1,
/*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output_id));
ASSERT_NE(output_id, XNN_INVALID_NODE_ID);
ASSERT_EQ(xnn_status_success, xnn_define_depth_to_space(subgraph, input_id, output_id, block_size, /*flags=*/0));
ASSERT_EQ(subgraph->num_nodes, 1);
const struct xnn_node* node = &subgraph->nodes[0];
ASSERT_EQ(node->type, xnn_node_type_depth_to_space);
ASSERT_EQ(node->compute_type, xnn_compute_type_fp32);
ASSERT_EQ(node->num_inputs, 1);
ASSERT_EQ(node->inputs[0], input_id);
ASSERT_EQ(node->num_outputs, 1);
ASSERT_EQ(node->outputs[0], output_id);
ASSERT_EQ(node->flags, 0);
}
TEST_F(DepthToSpaceTestQS8, matches_operator_api)
{
const int32_t input_zero_point = i8dist(rng);
const float input_scale = scale_dist(rng);
const int32_t output_zero_point = input_zero_point;
const float output_scale = input_scale;
std::generate(input.begin(), input.end(), [&]() { return i8dist(rng); });
std::fill(operator_output.begin(), operator_output.end(), INT8_C(0xA5));
std::fill(subgraph_output.begin(), subgraph_output.end(), INT8_C(0xA5));
ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));
// Call operator API.
xnn_operator_t op = nullptr;
const xnn_status status = xnn_create_depth_to_space_nhwc_x8(
output_channel(), input_channel(), output_channel(), block_size, /*flags=*/0, &op);
if (status == xnn_status_unsupported_hardware) {
GTEST_SKIP();
}
ASSERT_EQ(xnn_status_success, status);
ASSERT_NE(nullptr, op);
std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op(op, xnn_delete_operator);
ASSERT_EQ(
xnn_status_success,
xnn_setup_depth_to_space_nhwc_x8(
op, batch_size(), input_height(), input_width(), input.data(), operator_output.data(), /*threadpool=*/nullptr));
ASSERT_EQ(xnn_status_success, xnn_run_operator(op, /*threadpool=*/nullptr));
// Call subgraph API.
xnn_subgraph_t subgraph = nullptr;
ASSERT_EQ(xnn_status_success, xnn_create_subgraph(/*external_value_ids=*/2, /*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, input_zero_point, input_scale, input_dims.size(), input_dims.data(), nullptr,
/*external_id=*/0, /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
ASSERT_NE(input_id, XNN_INVALID_NODE_ID);
output_id = XNN_INVALID_NODE_ID;
ASSERT_EQ(
xnn_status_success,
xnn_define_quantized_tensor_value(
subgraph, xnn_datatype_qint8, output_zero_point, output_scale, output_dims.size(), output_dims.data(), nullptr,
/*external_id=*/1, /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output_id));
ASSERT_NE(output_id, XNN_INVALID_NODE_ID);
ASSERT_EQ(xnn_status_success, xnn_define_depth_to_space(subgraph, input_id, output_id, block_size, /*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, 2> external = {
xnn_external_value{input_id, input.data()}, xnn_external_value{output_id, subgraph_output.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_output, operator_output);
}
TEST_F(DepthToSpaceTestQU8, matches_operator_api)
{
const int32_t input_zero_point = u8dist(rng);
const float input_scale = scale_dist(rng);
const int32_t output_zero_point = input_zero_point;
const float output_scale = input_scale;
std::generate(input.begin(), input.end(), [&]() { return u8dist(rng); });
std::fill(operator_output.begin(), operator_output.end(), UINT8_C(0xA5));
std::fill(subgraph_output.begin(), subgraph_output.end(), UINT8_C(0xA5));
ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));
// Call operator API.
xnn_operator_t op = nullptr;
const xnn_status status = xnn_create_depth_to_space_nhwc_x8(
output_channel(), input_channel(), output_channel(), block_size, /*flags=*/0, &op);
if (status == xnn_status_unsupported_hardware) {
GTEST_SKIP();
}
ASSERT_EQ(xnn_status_success, status);
ASSERT_NE(nullptr, op);
std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op(op, xnn_delete_operator);
ASSERT_EQ(
xnn_status_success,
xnn_setup_depth_to_space_nhwc_x8(
op, batch_size(), input_height(), input_width(), input.data(), operator_output.data(), /*threadpool=*/nullptr));
ASSERT_EQ(xnn_status_success, xnn_run_operator(op, /*threadpool=*/nullptr));
// Call subgraph API.
xnn_subgraph_t subgraph = nullptr;
ASSERT_EQ(xnn_status_success, xnn_create_subgraph(/*external_value_ids=*/2, /*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, input_zero_point, input_scale, input_dims.size(), input_dims.data(), nullptr,
/*external_id=*/0, /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
ASSERT_NE(input_id, XNN_INVALID_NODE_ID);
output_id = XNN_INVALID_NODE_ID;
ASSERT_EQ(
xnn_status_success,
xnn_define_quantized_tensor_value(
subgraph, xnn_datatype_quint8, output_zero_point, output_scale, output_dims.size(), output_dims.data(), nullptr,
/*external_id=*/1, /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output_id));
ASSERT_NE(output_id, XNN_INVALID_NODE_ID);
ASSERT_EQ(xnn_status_success, xnn_define_depth_to_space(subgraph, input_id, output_id, block_size, /*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, 2> external = {
xnn_external_value{input_id, input.data()}, xnn_external_value{output_id, subgraph_output.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_output, operator_output);
}
TEST_F(DepthToSpaceTestF32, matches_operator_api)
{
std::uniform_real_distribution<float> f32dist(-255.0f, 255.0f);
std::generate(input.begin(), input.end(), [&]() { return f32dist(rng); });
std::fill(operator_output.begin(), operator_output.end(), nanf(""));
std::fill(subgraph_output.begin(), subgraph_output.end(), nanf(""));
ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));
// Call operator API.
xnn_operator_t op = nullptr;
const xnn_status status = xnn_create_depth_to_space_nhwc_x32(
output_channel(), input_channel(), output_channel(), block_size, /*flags=*/0, &op);
if (status == xnn_status_unsupported_hardware) {
GTEST_SKIP();
}
ASSERT_EQ(xnn_status_success, status);
ASSERT_NE(nullptr, op);
std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op(op, xnn_delete_operator);
ASSERT_EQ(
xnn_status_success,
xnn_setup_depth_to_space_nhwc_x32(
op, batch_size(), input_height(), input_width(), input.data(), operator_output.data(), /*threadpool=*/nullptr));
ASSERT_EQ(xnn_status_success, xnn_run_operator(op, /*threadpool=*/nullptr));
// Call subgraph API.
xnn_subgraph_t subgraph = nullptr;
ASSERT_EQ(xnn_status_success, xnn_create_subgraph(/*external_value_ids=*/2, /*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, /*external_id=*/0,
/*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
ASSERT_NE(input_id, XNN_INVALID_NODE_ID);
output_id = XNN_INVALID_NODE_ID;
ASSERT_EQ(
xnn_status_success,
xnn_define_tensor_value(
subgraph, xnn_datatype_fp32, output_dims.size(), output_dims.data(), nullptr, /*external_id=*/1,
/*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output_id));
ASSERT_NE(output_id, XNN_INVALID_NODE_ID);
xnn_runtime_t runtime = nullptr;
ASSERT_EQ(xnn_status_success, xnn_define_depth_to_space(subgraph, input_id, output_id, block_size, /*flags=*/0));
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, 2> external = {
xnn_external_value{input_id, input.data()}, xnn_external_value{output_id, subgraph_output.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_output, operator_output);
}