mirror of
https://github.com/pytorch/pytorch.git
synced 2025-10-20 21:14:14 +08:00
e04c9195b7
Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17670 Update math::Transpose to support tensor with size > 2G i-am-not-moving-c2-to-c10 Differential Revision: D14313624 fbshipit-source-id: 0b4a85b913972e5a8981f0d40d0c539407b98f30
432 lines
12 KiB
C++
432 lines
12 KiB
C++
#include <array>
|
|
#include <iostream>
|
|
#include <memory>
|
|
#include <vector>
|
|
|
|
#include <gtest/gtest.h>
|
|
|
|
#include "caffe2/core/context.h"
|
|
#include "caffe2/core/context_gpu.h"
|
|
#include "caffe2/core/flags.h"
|
|
#include "caffe2/operators/utility_ops.h"
|
|
#include "caffe2/utils/math.h"
|
|
|
|
C10_DECLARE_string(caffe_test_root);
|
|
|
|
namespace caffe2 {
|
|
|
|
void executeGpuBinaryOpTest(
|
|
int shapex0,
|
|
int shapex1,
|
|
int shapey,
|
|
std::function<float(int)> input0,
|
|
std::function<float(int)> input1,
|
|
std::function<void(
|
|
int N0,
|
|
int N1,
|
|
const float* src0,
|
|
const float* src1,
|
|
float* dst,
|
|
CUDAContext* context)> operation,
|
|
std::function<float(int)> correct_output) {
|
|
if (!HasCudaGPU())
|
|
return;
|
|
Workspace ws;
|
|
DeviceOption option;
|
|
option.set_device_type(PROTO_CUDA);
|
|
CUDAContext context(option);
|
|
|
|
Blob* blobx0 = ws.CreateBlob("X0");
|
|
Blob* blobx1 = ws.CreateBlob("X1");
|
|
Blob* bloby = ws.CreateBlob("Y");
|
|
Blob* bloby_host = ws.CreateBlob("Y_host");
|
|
|
|
auto* tensorx0 = BlobGetMutableTensor(blobx0, CUDA);
|
|
auto* tensorx1 = BlobGetMutableTensor(blobx1, CUDA);
|
|
auto* tensory = BlobGetMutableTensor(bloby, CUDA);
|
|
|
|
vector<int> shapex0_vector{shapex0};
|
|
vector<int> shapex1_vector{shapex1};
|
|
vector<int> shapey_vector{shapey};
|
|
|
|
tensorx0->Resize(shapex0_vector);
|
|
tensorx1->Resize(shapex1_vector);
|
|
tensory->Resize(shapey_vector);
|
|
|
|
for (int i = 0; i < shapex0; i++) {
|
|
math::Set<float, CUDAContext>(
|
|
1, input0(i), tensorx0->mutable_data<float>() + i, &context);
|
|
}
|
|
for (int i = 0; i < shapex1; i++) {
|
|
math::Set<float, CUDAContext>(
|
|
1, input1(i), tensorx1->mutable_data<float>() + i, &context);
|
|
}
|
|
operation(
|
|
shapex0,
|
|
shapex1,
|
|
tensorx0->template data<float>(),
|
|
tensorx1->template data<float>(),
|
|
tensory->mutable_data<float>(),
|
|
&context);
|
|
context.FinishDeviceComputation();
|
|
|
|
// Copy result to CPU so we can inspect it
|
|
auto* tensory_host = BlobGetMutableTensor(bloby_host, CPU);
|
|
tensory_host->CopyFrom(*tensory);
|
|
|
|
for (int i = 0; i < shapey; ++i) {
|
|
EXPECT_EQ(tensory_host->data<float>()[i], correct_output(i));
|
|
}
|
|
}
|
|
|
|
TEST(MathUtilGPUTest, testAddStripedBatch) {
|
|
if (!HasCudaGPU())
|
|
return;
|
|
Workspace ws;
|
|
DeviceOption option;
|
|
option.set_device_type(PROTO_CUDA);
|
|
CUDAContext context(option);
|
|
Blob* blobx = ws.CreateBlob("X");
|
|
Blob* bloby = ws.CreateBlob("Y");
|
|
Blob* bloby_host = ws.CreateBlob("Y_host");
|
|
|
|
vector<int> shapex{33 * 9, 25};
|
|
vector<int> shapey{33, 25};
|
|
|
|
auto* tensorx = BlobGetMutableTensor(blobx, CUDA);
|
|
tensorx->Resize(shapex);
|
|
int stripe = 33 * 25;
|
|
vector<float> tot(33, 0.0);
|
|
for (int j = 0; j < 9; j++) {
|
|
// Have different values for each line
|
|
for (int k = 0; k < 33; k++) {
|
|
math::Set<float, CUDAContext>(
|
|
33,
|
|
1.0 + j + k,
|
|
tensorx->mutable_data<float>() + j * stripe + k * 25,
|
|
&context);
|
|
tot[k] += 1.0 + j + k;
|
|
}
|
|
}
|
|
|
|
auto* tensory = BlobGetMutableTensor(bloby, CUDA);
|
|
tensory->Resize(shapey);
|
|
math::Set<float, CUDAContext>(
|
|
stripe, 0.0, tensory->mutable_data<float>(), &context);
|
|
|
|
math::AddStripedBatch<float, CUDAContext>(
|
|
stripe,
|
|
tensorx->template data<float>(),
|
|
tensory->mutable_data<float>(),
|
|
stripe,
|
|
9,
|
|
&context);
|
|
context.FinishDeviceComputation();
|
|
|
|
// Copy result to CPU so we can inspect it
|
|
auto* tensory_host = BlobGetMutableTensor(bloby_host, CPU);
|
|
tensory_host->CopyFrom(*tensory);
|
|
|
|
for (int k = 0; k < 33; k++) {
|
|
for (int i = 0; i < 25; i++) {
|
|
EXPECT_EQ(tensory_host->data<float>()[k * 25 + i], tot[k]);
|
|
}
|
|
}
|
|
}
|
|
|
|
TEST(MathUtilGPUTest, testReduceMin) {
|
|
executeGpuBinaryOpTest(
|
|
6,
|
|
1,
|
|
1,
|
|
[](int /*i*/) { return 11.0f; },
|
|
[](int /*i*/) { return 0.0f; },
|
|
[](int N0,
|
|
int /*N1*/,
|
|
const float* src0,
|
|
const float* /*src1*/,
|
|
float* dst,
|
|
CUDAContext* context) {
|
|
Tensor aux(CUDA);
|
|
math::ReduceMin<float, CUDAContext>(N0, src0, dst, &aux, context);
|
|
},
|
|
[](int /*i*/) { return 11.0f; });
|
|
executeGpuBinaryOpTest(
|
|
6,
|
|
1,
|
|
1,
|
|
[](int i) { return i == 3 ? 11.0f : 17.0f; },
|
|
[](int /*i*/) { return 0.0f; },
|
|
[](int N0,
|
|
int /*N1*/,
|
|
const float* src0,
|
|
const float* /*src1*/,
|
|
float* dst,
|
|
CUDAContext* context) {
|
|
Tensor aux(CUDA);
|
|
math::ReduceMin<float, CUDAContext>(N0, src0, dst, &aux, context);
|
|
},
|
|
[](int /*i*/) { return 11.0f; });
|
|
}
|
|
|
|
TEST(MathUtilGPUTest, testReduceMax) {
|
|
executeGpuBinaryOpTest(
|
|
6,
|
|
1,
|
|
1,
|
|
[](int /*i*/) { return 11.0f; },
|
|
[](int /*i*/) { return 0.0f; },
|
|
[](int N0,
|
|
int /*N1*/,
|
|
const float* src0,
|
|
const float* /*src1*/,
|
|
float* dst,
|
|
CUDAContext* context) {
|
|
Tensor aux(CUDA);
|
|
math::ReduceMax<float, CUDAContext>(N0, src0, dst, &aux, context);
|
|
},
|
|
[](int /*i*/) { return 11.0f; });
|
|
executeGpuBinaryOpTest(
|
|
6,
|
|
1,
|
|
1,
|
|
[](int i) { return i == 3 ? 17.0f : 11.0f; },
|
|
[](int /*i*/) { return 0.0f; },
|
|
[](int N0,
|
|
int /*N1*/,
|
|
const float* src0,
|
|
const float* /*src1*/,
|
|
float* dst,
|
|
CUDAContext* context) {
|
|
Tensor aux(CUDA);
|
|
math::ReduceMax<float, CUDAContext>(N0, src0, dst, &aux, context);
|
|
},
|
|
[](int /*i*/) { return 17.0f; });
|
|
}
|
|
|
|
TEST(MathUtilGPUTest, testCopyVector) {
|
|
executeGpuBinaryOpTest(
|
|
6,
|
|
1,
|
|
6,
|
|
[](int i) { return 5.0f - i; },
|
|
[](int /*i*/) { return 0.0f; },
|
|
[](int N0,
|
|
int /*N1*/,
|
|
const float* src0,
|
|
const float* /*src1*/,
|
|
float* dst,
|
|
CUDAContext* context) {
|
|
math::CopyVector<float, CUDAContext>(N0, src0, dst, context);
|
|
},
|
|
[](int i) { return 5.0f - i; });
|
|
}
|
|
|
|
namespace {
|
|
|
|
constexpr float kEps = 1e-5;
|
|
|
|
class GemmBatchedGPUTest
|
|
: public testing::TestWithParam<testing::tuple<bool, bool>> {
|
|
protected:
|
|
void SetUp() override {
|
|
if (!HasCudaGPU()) {
|
|
return;
|
|
}
|
|
option_.set_device_type(PROTO_CUDA);
|
|
cuda_context_ = make_unique<CUDAContext>(option_);
|
|
Blob* X_blob = ws_.CreateBlob("X");
|
|
Blob* W_blob = ws_.CreateBlob("W");
|
|
Blob* Y_blob = ws_.CreateBlob("Y");
|
|
X_ = BlobGetMutableTensor(X_blob, CUDA);
|
|
W_ = BlobGetMutableTensor(W_blob, CUDA);
|
|
Y_ = BlobGetMutableTensor(Y_blob, CUDA);
|
|
X_->Resize(std::vector<int64_t>{3, 5, 10});
|
|
W_->Resize(std::vector<int64_t>{3, 6, 10});
|
|
Y_->Resize(std::vector<int64_t>{3, 5, 6});
|
|
math::Set<float, CUDAContext>(
|
|
X_->numel(), 1.0f, X_->mutable_data<float>(), cuda_context_.get());
|
|
math::Set<float, CUDAContext>(
|
|
W_->numel(), 1.0f, W_->mutable_data<float>(), cuda_context_.get());
|
|
trans_X_ = std::get<0>(GetParam());
|
|
trans_W_ = std::get<1>(GetParam());
|
|
}
|
|
|
|
void RunGemmBatched(const float alpha, const float beta) {
|
|
const float* X_data = X_->template data<float>();
|
|
const float* W_data = W_->template data<float>();
|
|
float* Y_data = Y_->template mutable_data<float>();
|
|
const int X_stride = 5 * 10;
|
|
const int W_stride = 6 * 10;
|
|
const int Y_stride = 5 * 6;
|
|
std::array<const float*, 3> X_array = {
|
|
X_data, X_data + X_stride, X_data + 2 * X_stride};
|
|
std::array<const float*, 3> W_array = {
|
|
W_data, W_data + W_stride, W_data + 2 * W_stride};
|
|
std::array<float*, 3> Y_array = {
|
|
Y_data, Y_data + Y_stride, Y_data + 2 * Y_stride};
|
|
math::GemmBatched<float, CUDAContext>(
|
|
trans_X_ ? CblasTrans : CblasNoTrans,
|
|
trans_W_ ? CblasTrans : CblasNoTrans,
|
|
3,
|
|
5,
|
|
6,
|
|
10,
|
|
alpha,
|
|
X_array.data(),
|
|
W_array.data(),
|
|
beta,
|
|
Y_array.data(),
|
|
cuda_context_.get());
|
|
}
|
|
|
|
void RunGemmStridedBatched(const float alpha, const float beta) {
|
|
const float* X_data = X_->template data<float>();
|
|
const float* W_data = W_->template data<float>();
|
|
float* Y_data = Y_->template mutable_data<float>();
|
|
const int X_stride = 5 * 10;
|
|
const int W_stride = 6 * 10;
|
|
const int Y_stride = 5 * 6;
|
|
math::GemmStridedBatched<float, CUDAContext>(
|
|
trans_X_ ? CblasTrans : CblasNoTrans,
|
|
trans_W_ ? CblasTrans : CblasNoTrans,
|
|
3,
|
|
5,
|
|
6,
|
|
10,
|
|
alpha,
|
|
X_data,
|
|
X_stride,
|
|
W_data,
|
|
W_stride,
|
|
beta,
|
|
Y_data,
|
|
Y_stride,
|
|
cuda_context_.get());
|
|
}
|
|
|
|
void VerifyOutput(const float value) const {
|
|
Tensor Y_cpu(*Y_, CPU);
|
|
for (int i = 0; i < Y_cpu.numel(); ++i) {
|
|
EXPECT_FLOAT_EQ(value, Y_cpu.template data<float>()[i]);
|
|
}
|
|
}
|
|
|
|
Workspace ws_;
|
|
DeviceOption option_;
|
|
std::unique_ptr<CUDAContext> cuda_context_;
|
|
Tensor* X_ = nullptr;
|
|
Tensor* W_ = nullptr;
|
|
Tensor* Y_ = nullptr;
|
|
bool trans_X_;
|
|
bool trans_W_;
|
|
};
|
|
|
|
TEST_P(GemmBatchedGPUTest, GemmBatchedGPUFloatTest) {
|
|
if (!HasCudaGPU()) {
|
|
return;
|
|
}
|
|
RunGemmBatched(1.0f, 0.0f);
|
|
VerifyOutput(10.0f);
|
|
RunGemmBatched(1.0f, 0.5f);
|
|
VerifyOutput(15.0f);
|
|
RunGemmBatched(0.5f, 1.0f);
|
|
VerifyOutput(20.0f);
|
|
}
|
|
|
|
TEST_P(GemmBatchedGPUTest, GemmStridedBatchedGPUFloatTest) {
|
|
if (!HasCudaGPU()) {
|
|
return;
|
|
}
|
|
RunGemmStridedBatched(1.0f, 0.0f);
|
|
VerifyOutput(10.0f);
|
|
RunGemmStridedBatched(1.0f, 0.5f);
|
|
VerifyOutput(15.0f);
|
|
RunGemmStridedBatched(0.5f, 1.0f);
|
|
VerifyOutput(20.0f);
|
|
}
|
|
|
|
INSTANTIATE_TEST_CASE_P(
|
|
GemmBatchedGPUTrans,
|
|
GemmBatchedGPUTest,
|
|
testing::Combine(testing::Bool(), testing::Bool()));
|
|
|
|
class BroadcastGPUTest : public testing::Test {
|
|
protected:
|
|
void SetUp() override {
|
|
if (!HasCudaGPU()) {
|
|
return;
|
|
}
|
|
option_.set_device_type(PROTO_CUDA);
|
|
cuda_context_ = make_unique<CUDAContext>(option_);
|
|
Blob* blob_x = ws_.CreateBlob("X");
|
|
Blob* blob_y = ws_.CreateBlob("Y");
|
|
X_ = BlobGetMutableTensor(blob_x, CUDA);
|
|
Y_ = BlobGetMutableTensor(blob_y, CUDA);
|
|
}
|
|
|
|
void SetUpData(
|
|
const std::vector<int>& X_dims,
|
|
const std::vector<int>& Y_dims,
|
|
const std::vector<float>& X_data) {
|
|
X_->Resize(X_dims);
|
|
Y_->Resize(Y_dims);
|
|
ASSERT_EQ(X_data.size(), X_->numel());
|
|
cuda_context_->CopyFromCPU<float>(
|
|
X_data.size(), X_data.data(), X_->mutable_data<float>());
|
|
}
|
|
|
|
void VerifyResult(const std::vector<float>& expected_output) {
|
|
Blob* blob_y_host = ws_.CreateBlob("Y_host");
|
|
auto* Y_host = BlobGetMutableTensor(blob_y_host, CPU);
|
|
Y_host->CopyFrom(*Y_);
|
|
ASSERT_EQ(expected_output.size(), Y_host->numel());
|
|
for (std::size_t i = 0; i < expected_output.size(); ++i) {
|
|
EXPECT_FLOAT_EQ(expected_output[i], Y_host->data<float>()[i]);
|
|
}
|
|
}
|
|
|
|
void RunBroadcastTest(
|
|
const std::vector<int>& X_dims,
|
|
const std::vector<int>& Y_dims,
|
|
const std::vector<float>& X_data,
|
|
const std::vector<float>& Y_data) {
|
|
SetUpData(X_dims, Y_dims, X_data);
|
|
math::Broadcast<float, CUDAContext>(
|
|
X_dims.size(),
|
|
X_dims.data(),
|
|
Y_dims.size(),
|
|
Y_dims.data(),
|
|
1.0f,
|
|
X_->data<float>(),
|
|
Y_->mutable_data<float>(),
|
|
cuda_context_.get());
|
|
VerifyResult(Y_data);
|
|
}
|
|
|
|
Workspace ws_;
|
|
DeviceOption option_;
|
|
std::unique_ptr<CUDAContext> cuda_context_;
|
|
Tensor* X_ = nullptr;
|
|
Tensor* Y_ = nullptr;
|
|
};
|
|
|
|
TEST_F(BroadcastGPUTest, BroadcastGPUFloatTest) {
|
|
if (!HasCudaGPU()) {
|
|
return;
|
|
}
|
|
RunBroadcastTest({2}, {2}, {1.0f, 2.0f}, {1.0f, 2.0f});
|
|
RunBroadcastTest({1}, {2}, {1.0f}, {1.0f, 1.0f});
|
|
RunBroadcastTest({1}, {2, 2}, {1.0f}, {1.0f, 1.0f, 1.0f, 1.0f});
|
|
RunBroadcastTest({2, 1}, {2, 2}, {1.0f, 2.0f}, {1.0f, 1.0f, 2.0f, 2.0f});
|
|
RunBroadcastTest(
|
|
{2, 1},
|
|
{2, 2, 2},
|
|
{1.0f, 2.0f},
|
|
{1.0f, 1.0f, 2.0f, 2.0f, 1.0f, 1.0f, 2.0f, 2.0f});
|
|
}
|
|
|
|
} // namespace
|
|
|
|
} // namespace caffe2
|