117 lines
4.8 KiB
C++
117 lines
4.8 KiB
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 <algorithm>
|
|
#include <array>
|
|
#include <cstddef>
|
|
#include <cstdint>
|
|
#include <limits>
|
|
#include <memory>
|
|
#include <random>
|
|
|
|
#include <xnnpack.h>
|
|
#include <xnnpack/node-type.h>
|
|
#include <xnnpack/operator.h>
|
|
#include <xnnpack/subgraph.h>
|
|
|
|
#include "subgraph-unary-tester.h"
|
|
#include <gtest/gtest.h>
|
|
|
|
using SoftmaxTestF32 = UnaryTest<float, /*OutputType=*/float, /*min_dim=*/1>;
|
|
|
|
TEST_F(SoftmaxTestF32, 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=*/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, dims.size(), 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, dims.size(), 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_softmax(subgraph, input_id, output_id, /*flags=*/0));
|
|
|
|
ASSERT_EQ(subgraph->num_nodes, 1);
|
|
const struct xnn_node* node = &subgraph->nodes[0];
|
|
ASSERT_EQ(node->type, xnn_node_type_softmax);
|
|
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(SoftmaxTestF32, matches_operator_api)
|
|
{
|
|
// Choose such range that expf(x[i]) overflows, but expf(x[i] - x_max) doesn't.
|
|
// However, the range is still narrow enough that single-precision exp doesn't overflow.
|
|
std::uniform_real_distribution<float> f32dist(90.0f, 100.0f);
|
|
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_softmax_nc_f32(channels, channels, channels, /*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_softmax_nc_f32(op, batch_size, 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, dims.size(), 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, dims.size(), 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_softmax(subgraph, input_id, output_id, /*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);
|
|
}
|