sglang_v0.5.2/pytorch_2.8.0/third_party/XNNPACK/test/f32-vscaleexpminusmax.cc

143 lines
7.8 KiB
C++

// 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.
//
// Specification: f32-vscaleexpminusmax
#include <algorithm>
#include <cassert>
#include <cmath>
#include <cstddef>
#include <cstdlib>
#include <functional>
#include <random>
#include <vector>
#include <gtest/gtest.h>
#include "replicable_random_device.h"
#include "xnnpack.h"
#include "xnnpack/buffer.h"
#include "xnnpack/common.h"
#include "xnnpack/isa-checks.h"
#include "xnnpack/microfnptr.h"
#include "xnnpack/vscaleexpminusmax.h"
class VScaleExpMinusMaxMicrokernelTester {
public:
VScaleExpMinusMaxMicrokernelTester& elements(size_t elements)
{
assert(elements != 0);
this->elements_ = elements;
return *this;
}
size_t elements() const { return this->elements_; }
VScaleExpMinusMaxMicrokernelTester& scale(float scale)
{
assert(std::isfinite(scale));
assert(scale > 0);
this->scale_ = scale;
return *this;
}
float scale() const { return this->scale_; }
VScaleExpMinusMaxMicrokernelTester& iterations(size_t iterations)
{
this->iterations_ = iterations;
return *this;
}
size_t iterations() const { return this->iterations_; }
void Test(xnn_f32_vscaleexpminusmax_ukernel_fn vscaleexpminusmax) const
{
xnnpack::ReplicableRandomDevice rng;
// Choose such range that expf(x[i]) overflows, but expf(x[i] - x_max) doesn't.
// However, the range is still narrow enough that double-precision exp doesn't overflow.
auto f32rng = [&rng]() { return std::uniform_real_distribution<float>(90.0f, 100.0f)(rng); };
xnnpack::Buffer<float> x(elements() + XNN_EXTRA_BYTES / sizeof(float));
xnnpack::Buffer<float> y(elements());
xnnpack::Buffer<double> y_ref(elements());
// TODO(b/372792254): This is hiding a possible msan bug in the microkernels tested here.
std::fill(y.begin(), y.end(), 0.0f);
for (size_t iteration = 0; iteration < iterations(); iteration++) {
std::generate(x.begin(), x.end(), std::ref(f32rng));
// Compute reference results.
const float x_max = *std::max_element(x.begin(), x.begin() + elements());
for (size_t i = 0; i < elements(); i++) {
y_ref[i] = double(scale()) * exp(double(x[i]) - double(x_max));
}
// Call optimized micro-kernel.
vscaleexpminusmax(elements() * sizeof(float), x.data(), y.data(), scale(), x_max);
// Verify results.
for (size_t i = 0; i < elements(); i++) {
EXPECT_NEAR(y_ref[i], y[i], std::abs(y_ref[i]) * 1.0e-6)
<< "elements = " << elements() << ", scale = " << scale() << ", x_max = " << x_max;
}
}
}
private:
size_t elements_{1};
float scale_{1.0f};
size_t iterations_{15};
};
#define XNN_TEST_VSCALEEXPMINUSMAX_ELEMENT_EQ(ukernel, arch_flags, element_tile, ...) \
TEST(ukernel, element_eq) \
{ \
TEST_REQUIRES_ARCH_FLAGS(arch_flags); \
VScaleExpMinusMaxMicrokernelTester().elements(element_tile).Test(ukernel); \
}
#define XNN_TEST_VSCALEEXPMINUSMAX_ELEMENT_GT(ukernel, arch_flags, element_tile, ...) \
TEST(ukernel, element_gt) \
{ \
TEST_REQUIRES_ARCH_FLAGS(arch_flags); \
for (size_t element_size = element_tile + 1; element_size < ((element_tile == 1) ? 10 : element_tile * 2); \
element_size++) { \
VScaleExpMinusMaxMicrokernelTester().elements(element_size).Test(ukernel); \
} \
}
#define XNN_TEST_VSCALEEXPMINUSMAX_ELEMENT_LT(ukernel, arch_flags, element_tile, ...) \
TEST(ukernel, element_lt) \
{ \
TEST_REQUIRES_ARCH_FLAGS(arch_flags); \
for (size_t element_size = 1; element_size < element_tile; element_size++) { \
VScaleExpMinusMaxMicrokernelTester().elements(element_size).Test(ukernel); \
} \
}
#define XNN_TEST_VSCALEEXPMINUSMAX_ELEMENT_DIV(ukernel, arch_flags, element_tile, ...) \
TEST(ukernel, element_div) \
{ \
TEST_REQUIRES_ARCH_FLAGS(arch_flags); \
for (size_t element_size = 2 * element_tile; element_size < 10 * element_tile; element_size += element_tile) { \
VScaleExpMinusMaxMicrokernelTester().elements(element_size).Test(ukernel); \
} \
}
#define XNN_TEST_VSCALEEXPMINUSMAX_SCALE(ukernel, arch_flags, element_tile, ...) \
TEST(ukernel, scale) \
{ \
TEST_REQUIRES_ARCH_FLAGS(arch_flags); \
for (size_t element_size = 1; element_size <= 5 * element_tile; element_size += max(1, element_tile - 1)) { \
VScaleExpMinusMaxMicrokernelTester().elements(element_size).scale(0.01f).Test(ukernel); \
VScaleExpMinusMaxMicrokernelTester().elements(element_size).scale(100.0f).Test(ukernel); \
} \
}
#define XNN_UKERNEL_WITH_PARAMS(arch_flags, ukernel, element_tile, datatype, params_type, init_params) XNN_TEST_VSCALEEXPMINUSMAX_ELEMENT_EQ(ukernel,arch_flags, element_tile, init_params); \
XNN_TEST_VSCALEEXPMINUSMAX_ELEMENT_DIV(ukernel,arch_flags, element_tile, init_params); \
XNN_TEST_VSCALEEXPMINUSMAX_ELEMENT_LT(ukernel,arch_flags, element_tile, init_params); \
XNN_TEST_VSCALEEXPMINUSMAX_ELEMENT_GT(ukernel,arch_flags, element_tile, init_params); \
XNN_TEST_VSCALEEXPMINUSMAX_SCALE(ukernel,arch_flags, element_tile, init_params);
#include "f32-vscaleexpminusmax/f32-vscaleexpminusmax.h"
#undef XNN_UKERNEL_WITH_PARAMS