devtools/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2026-04-10 11:34:33 +08:00
Code Issues Packages Projects Releases Wiki Activity

Clang-format tests, examples, libraries, benchmarks, etc.

Browse Source
This commit is contained in:
Antonio Sánchez
2023-12-05 21:22:55 +00:00
committed by Rasmus Munk Larsen
parent 3252ecc7a4
commit 46e9cdb7fe
876 changed files with 33453 additions and 37795 deletions
Download Patch File Download Diff File Expand all files Collapse all files

192
unsupported/test/cxx11_tensor_of_float16_gpu.cu
View File

@@ -16,10 +16,9 @@
#include "main.h"
#include <unsupported/Eigen/CXX11/Tensor>
using Eigen::Tensor;
template<typename>
template <typename>
void test_gpu_numext() {
Eigen::GpuStreamDevice stream;
Eigen::GpuDevice gpu_device(&stream);
@@ -29,21 +28,19 @@ void test_gpu_numext() {
bool* d_res_half = (bool*)gpu_device.allocate(num_elem * sizeof(bool));
bool* d_res_float = (bool*)gpu_device.allocate(num_elem * sizeof(bool));
Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float(
d_float, num_elem);
Eigen::TensorMap<Eigen::Tensor<bool, 1>, Eigen::Aligned> gpu_res_half(
d_res_half, num_elem);
Eigen::TensorMap<Eigen::Tensor<bool, 1>, Eigen::Aligned> gpu_res_float(
d_res_float, num_elem);
Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float(d_float, num_elem);
Eigen::TensorMap<Eigen::Tensor<bool, 1>, Eigen::Aligned> gpu_res_half(d_res_half, num_elem);
Eigen::TensorMap<Eigen::Tensor<bool, 1>, Eigen::Aligned> gpu_res_float(d_res_float, num_elem);
gpu_float.device(gpu_device) = gpu_float.random() - gpu_float.constant(0.5f);
gpu_res_float.device(gpu_device) = gpu_float.unaryExpr(Eigen::internal::scalar_isnan_op<float>());
gpu_res_half.device(gpu_device) = gpu_float.cast<Eigen::half>().unaryExpr(Eigen::internal::scalar_isnan_op<Eigen::half>());
gpu_res_half.device(gpu_device) =
gpu_float.cast<Eigen::half>().unaryExpr(Eigen::internal::scalar_isnan_op<Eigen::half>());
Tensor<bool, 1> half_prec(num_elem);
Tensor<bool, 1> full_prec(num_elem);
gpu_device.memcpyDeviceToHost(half_prec.data(), d_res_half, num_elem*sizeof(bool));
gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, num_elem*sizeof(bool));
gpu_device.memcpyDeviceToHost(half_prec.data(), d_res_half, num_elem * sizeof(bool));
gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, num_elem * sizeof(bool));
gpu_device.synchronize();
for (int i = 0; i < num_elem; ++i) {
@@ -56,10 +53,9 @@ void test_gpu_numext() {
gpu_device.deallocate(d_res_float);
}
#ifdef EIGEN_HAS_GPU_FP16
template<typename>
template <typename>
void test_gpu_conversion() {
Eigen::GpuStreamDevice stream;
Eigen::GpuDevice gpu_device(&stream);
@@ -69,12 +65,9 @@ void test_gpu_conversion() {
Eigen::half* d_half = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half));
float* d_conv = (float*)gpu_device.allocate(num_elem * sizeof(float));
Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float(
d_float, num_elem);
Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_half(
d_half, num_elem);
Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_conv(
d_conv, num_elem);
Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float(d_float, num_elem);
Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_half(d_half, num_elem);
Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_conv(d_conv, num_elem);
gpu_float.device(gpu_device) = gpu_float.random();
gpu_half.device(gpu_device) = gpu_float.cast<Eigen::half>();
@@ -82,8 +75,8 @@ void test_gpu_conversion() {
Tensor<float, 1> initial(num_elem);
Tensor<float, 1> final(num_elem);
gpu_device.memcpyDeviceToHost(initial.data(), d_float, num_elem*sizeof(float));
gpu_device.memcpyDeviceToHost(final.data(), d_conv, num_elem*sizeof(float));
gpu_device.memcpyDeviceToHost(initial.data(), d_float, num_elem * sizeof(float));
gpu_device.memcpyDeviceToHost(final.data(), d_conv, num_elem * sizeof(float));
for (int i = 0; i < num_elem; ++i) {
VERIFY_IS_APPROX(initial(i), final(i));
@@ -94,7 +87,7 @@ void test_gpu_conversion() {
gpu_device.deallocate(d_conv);
}
template<typename>
template <typename>
void test_gpu_unary() {
Eigen::GpuStreamDevice stream;
Eigen::GpuDevice gpu_device(&stream);
@@ -104,12 +97,9 @@ void test_gpu_unary() {
float* d_res_half = (float*)gpu_device.allocate(num_elem * sizeof(float));
float* d_res_float = (float*)gpu_device.allocate(num_elem * sizeof(float));
Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float(
d_float, num_elem);
Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_res_half(
d_res_half, num_elem);
Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_res_float(
d_res_float, num_elem);
Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float(d_float, num_elem);
Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_res_half(d_res_half, num_elem);
Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_res_float(d_res_float, num_elem);
gpu_float.device(gpu_device) = gpu_float.random() - gpu_float.constant(0.5f);
gpu_res_float.device(gpu_device) = gpu_float.abs();
@@ -117,8 +107,8 @@ void test_gpu_unary() {
Tensor<float, 1> half_prec(num_elem);
Tensor<float, 1> full_prec(num_elem);
gpu_device.memcpyDeviceToHost(half_prec.data(), d_res_half, num_elem*sizeof(float));
gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, num_elem*sizeof(float));
gpu_device.memcpyDeviceToHost(half_prec.data(), d_res_half, num_elem * sizeof(float));
gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, num_elem * sizeof(float));
gpu_device.synchronize();
for (int i = 0; i < num_elem; ++i) {
@@ -131,7 +121,7 @@ void test_gpu_unary() {
gpu_device.deallocate(d_res_float);
}
template<typename>
template <typename>
void test_gpu_elementwise() {
Eigen::GpuStreamDevice stream;
Eigen::GpuDevice gpu_device(&stream);
@@ -142,28 +132,27 @@ void test_gpu_elementwise() {
float* d_res_half = (float*)gpu_device.allocate(num_elem * sizeof(float));
float* d_res_float = (float*)gpu_device.allocate(num_elem * sizeof(float));
Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float1(
d_float1, num_elem);
Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float2(
d_float2, num_elem);
Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_res_half(
d_res_half, num_elem);
Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_res_float(
d_res_float, num_elem);
Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float1(d_float1, num_elem);
Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float2(d_float2, num_elem);
Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_res_half(d_res_half, num_elem);
Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_res_float(d_res_float, num_elem);
gpu_float1.device(gpu_device) = gpu_float1.random();
gpu_float2.device(gpu_device) = gpu_float2.random();
gpu_res_float.device(gpu_device) = (gpu_float1 + gpu_float2) * gpu_float1;
gpu_res_half.device(gpu_device) = ((gpu_float1.cast<Eigen::half>() + gpu_float2.cast<Eigen::half>()) * gpu_float1.cast<Eigen::half>()).cast<float>();
gpu_res_half.device(gpu_device) =
((gpu_float1.cast<Eigen::half>() + gpu_float2.cast<Eigen::half>()) * gpu_float1.cast<Eigen::half>())
.cast<float>();
Tensor<float, 1> half_prec(num_elem);
Tensor<float, 1> full_prec(num_elem);
gpu_device.memcpyDeviceToHost(half_prec.data(), d_res_half, num_elem*sizeof(float));
gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, num_elem*sizeof(float));
gpu_device.memcpyDeviceToHost(half_prec.data(), d_res_half, num_elem * sizeof(float));
gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, num_elem * sizeof(float));
gpu_device.synchronize();
for (int i = 0; i < num_elem; ++i) {
std::cout << "Checking elemwise " << i << ": full prec = " << full_prec(i) << " vs half prec = " << half_prec(i) << std::endl;
std::cout << "Checking elemwise " << i << ": full prec = " << full_prec(i) << " vs half prec = " << half_prec(i)
<< std::endl;
VERIFY_IS_APPROX(static_cast<Eigen::half>(full_prec(i)), static_cast<Eigen::half>(half_prec(i)));
}
@@ -173,7 +162,7 @@ void test_gpu_elementwise() {
gpu_device.deallocate(d_res_float);
}
template<typename>
template <typename>
void test_gpu_trancendental() {
Eigen::GpuStreamDevice stream;
Eigen::GpuDevice gpu_device(&stream);
@@ -230,30 +219,33 @@ void test_gpu_trancendental() {
Tensor<float, 1> input3(num_elem);
Tensor<Eigen::half, 1> half_prec3(num_elem);
Tensor<Eigen::half, 1> full_prec3(num_elem);
gpu_device.memcpyDeviceToHost(input1.data(), d_float1, num_elem*sizeof(float));
gpu_device.memcpyDeviceToHost(input2.data(), d_float2, num_elem*sizeof(float));
gpu_device.memcpyDeviceToHost(input3.data(), d_float3, num_elem*sizeof(float));
gpu_device.memcpyDeviceToHost(half_prec1.data(), d_res1_half, num_elem*sizeof(Eigen::half));
gpu_device.memcpyDeviceToHost(full_prec1.data(), d_res1_float, num_elem*sizeof(Eigen::half));
gpu_device.memcpyDeviceToHost(half_prec2.data(), d_res2_half, num_elem*sizeof(Eigen::half));
gpu_device.memcpyDeviceToHost(full_prec2.data(), d_res2_float, num_elem*sizeof(Eigen::half));
gpu_device.memcpyDeviceToHost(half_prec3.data(), d_res3_half, num_elem*sizeof(Eigen::half));
gpu_device.memcpyDeviceToHost(full_prec3.data(), d_res3_float, num_elem*sizeof(Eigen::half));
gpu_device.memcpyDeviceToHost(input1.data(), d_float1, num_elem * sizeof(float));
gpu_device.memcpyDeviceToHost(input2.data(), d_float2, num_elem * sizeof(float));
gpu_device.memcpyDeviceToHost(input3.data(), d_float3, num_elem * sizeof(float));
gpu_device.memcpyDeviceToHost(half_prec1.data(), d_res1_half, num_elem * sizeof(Eigen::half));
gpu_device.memcpyDeviceToHost(full_prec1.data(), d_res1_float, num_elem * sizeof(Eigen::half));
gpu_device.memcpyDeviceToHost(half_prec2.data(), d_res2_half, num_elem * sizeof(Eigen::half));
gpu_device.memcpyDeviceToHost(full_prec2.data(), d_res2_float, num_elem * sizeof(Eigen::half));
gpu_device.memcpyDeviceToHost(half_prec3.data(), d_res3_half, num_elem * sizeof(Eigen::half));
gpu_device.memcpyDeviceToHost(full_prec3.data(), d_res3_float, num_elem * sizeof(Eigen::half));
gpu_device.synchronize();
for (int i = 0; i < num_elem; ++i) {
std::cout << "Checking elemwise exp " << i << " input = " << input1(i) << " full = " << full_prec1(i) << " half = " << half_prec1(i) << std::endl;
std::cout << "Checking elemwise exp " << i << " input = " << input1(i) << " full = " << full_prec1(i)
<< " half = " << half_prec1(i) << std::endl;
VERIFY_IS_APPROX(full_prec1(i), half_prec1(i));
}
for (int i = 0; i < num_elem; ++i) {
std::cout << "Checking elemwise log " << i << " input = " << input2(i) << " full = " << full_prec2(i) << " half = " << half_prec2(i) << std::endl;
if(std::abs(input2(i)-1.f)<0.05f) // log lacks accuracy nearby 1
VERIFY_IS_APPROX(full_prec2(i)+Eigen::half(0.1f), half_prec2(i)+Eigen::half(0.1f));
std::cout << "Checking elemwise log " << i << " input = " << input2(i) << " full = " << full_prec2(i)
<< " half = " << half_prec2(i) << std::endl;
if (std::abs(input2(i) - 1.f) < 0.05f) // log lacks accuracy nearby 1
VERIFY_IS_APPROX(full_prec2(i) + Eigen::half(0.1f), half_prec2(i) + Eigen::half(0.1f));
else
VERIFY_IS_APPROX(full_prec2(i), half_prec2(i));
}
for (int i = 0; i < num_elem; ++i) {
std::cout << "Checking elemwise plog1 " << i << " input = " << input3(i) << " full = " << full_prec3(i) << " half = " << half_prec3(i) << std::endl;
std::cout << "Checking elemwise plog1 " << i << " input = " << input3(i) << " full = " << full_prec3(i)
<< " half = " << half_prec3(i) << std::endl;
VERIFY_IS_APPROX(full_prec3(i), half_prec3(i));
}
gpu_device.deallocate(d_float1);
@@ -267,27 +259,23 @@ void test_gpu_trancendental() {
gpu_device.deallocate(d_res3_half);
}
template<typename>
template <typename>
void test_gpu_contractions() {
Eigen::GpuStreamDevice stream;
Eigen::GpuDevice gpu_device(&stream);
int rows = 23;
int cols = 23;
int num_elem = rows*cols;
int num_elem = rows * cols;
float* d_float1 = (float*)gpu_device.allocate(num_elem * sizeof(float));
float* d_float2 = (float*)gpu_device.allocate(num_elem * sizeof(float));
Eigen::half* d_res_half = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half));
Eigen::half* d_res_float = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half));
Eigen::TensorMap<Eigen::Tensor<float, 2>, Eigen::Aligned> gpu_float1(
d_float1, rows, cols);
Eigen::TensorMap<Eigen::Tensor<float, 2>, Eigen::Aligned> gpu_float2(
d_float2, rows, cols);
Eigen::TensorMap<Eigen::Tensor<Eigen::half, 2>, Eigen::Aligned> gpu_res_half(
d_res_half, rows, cols);
Eigen::TensorMap<Eigen::Tensor<Eigen::half, 2>, Eigen::Aligned> gpu_res_float(
d_res_float, rows, cols);
Eigen::TensorMap<Eigen::Tensor<float, 2>, Eigen::Aligned> gpu_float1(d_float1, rows, cols);
Eigen::TensorMap<Eigen::Tensor<float, 2>, Eigen::Aligned> gpu_float2(d_float2, rows, cols);
Eigen::TensorMap<Eigen::Tensor<Eigen::half, 2>, Eigen::Aligned> gpu_res_half(d_res_half, rows, cols);
Eigen::TensorMap<Eigen::Tensor<Eigen::half, 2>, Eigen::Aligned> gpu_res_float(d_res_float, rows, cols);
gpu_float1.device(gpu_device) = gpu_float1.random() - gpu_float1.constant(0.5f);
gpu_float2.device(gpu_device) = gpu_float2.random() - gpu_float2.constant(0.5f);
@@ -299,8 +287,8 @@ void test_gpu_contractions() {
Tensor<Eigen::half, 2> half_prec(rows, cols);
Tensor<Eigen::half, 2> full_prec(rows, cols);
gpu_device.memcpyDeviceToHost(half_prec.data(), d_res_half, num_elem*sizeof(Eigen::half));
gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, num_elem*sizeof(Eigen::half));
gpu_device.memcpyDeviceToHost(half_prec.data(), d_res_half, num_elem * sizeof(Eigen::half));
gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, num_elem * sizeof(Eigen::half));
gpu_device.synchronize();
for (int i = 0; i < rows; ++i) {
@@ -318,27 +306,22 @@ void test_gpu_contractions() {
gpu_device.deallocate(d_res_float);
}
template<typename>
template <typename>
void test_gpu_reductions(int size1, int size2, int redux) {
std::cout << "Reducing " << size1 << " by " << size2
<< " tensor along dim " << redux << std::endl;
std::cout << "Reducing " << size1 << " by " << size2 << " tensor along dim " << redux << std::endl;
Eigen::GpuStreamDevice stream;
Eigen::GpuDevice gpu_device(&stream);
int num_elem = size1*size2;
int num_elem = size1 * size2;
int result_size = (redux == 1 ? size1 : size2);
float* d_float = (float*)gpu_device.allocate(num_elem * sizeof(float));
Eigen::half* d_res_half = (Eigen::half*)gpu_device.allocate(result_size * sizeof(Eigen::half));
Eigen::half* d_res_float = (Eigen::half*)gpu_device.allocate(result_size * sizeof(Eigen::half));
Eigen::TensorMap<Eigen::Tensor<float, 2>, Eigen::Aligned> gpu_float(
d_float, size1, size2);
Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res_half(
d_res_half, result_size);
Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res_float(
d_res_float, result_size);
Eigen::TensorMap<Eigen::Tensor<float, 2>, Eigen::Aligned> gpu_float(d_float, size1, size2);
Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res_half(d_res_half, result_size);
Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res_float(d_res_float, result_size);
gpu_float.device(gpu_device) = gpu_float.random() * 2.0f;
@@ -348,8 +331,8 @@ void test_gpu_reductions(int size1, int size2, int redux) {
Tensor<Eigen::half, 1> half_prec(result_size);
Tensor<Eigen::half, 1> full_prec(result_size);
gpu_device.memcpyDeviceToHost(half_prec.data(), d_res_half, result_size*sizeof(Eigen::half));
gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, result_size*sizeof(Eigen::half));
gpu_device.memcpyDeviceToHost(half_prec.data(), d_res_half, result_size * sizeof(Eigen::half));
gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, result_size * sizeof(Eigen::half));
gpu_device.synchronize();
for (int i = 0; i < result_size; ++i) {
@@ -362,7 +345,7 @@ void test_gpu_reductions(int size1, int size2, int redux) {
gpu_device.deallocate(d_res_float);
}
template<typename>
template <typename>
void test_gpu_reductions() {
test_gpu_reductions<void>(13, 13, 0);
test_gpu_reductions<void>(13, 13, 1);
@@ -374,23 +357,20 @@ void test_gpu_reductions() {
test_gpu_reductions<void>(36, 35, 1);
}
template<typename>
template <typename>
void test_gpu_full_reductions() {
Eigen::GpuStreamDevice stream;
Eigen::GpuDevice gpu_device(&stream);
int size = 13;
int num_elem = size*size;
int num_elem = size * size;
float* d_float = (float*)gpu_device.allocate(num_elem * sizeof(float));
Eigen::half* d_res_half = (Eigen::half*)gpu_device.allocate(1 * sizeof(Eigen::half));
Eigen::half* d_res_float = (Eigen::half*)gpu_device.allocate(1 * sizeof(Eigen::half));
Eigen::TensorMap<Eigen::Tensor<float, 2>, Eigen::Aligned> gpu_float(
d_float, size, size);
Eigen::TensorMap<Eigen::Tensor<Eigen::half, 0>, Eigen::Aligned> gpu_res_half(
d_res_half);
Eigen::TensorMap<Eigen::Tensor<Eigen::half, 0>, Eigen::Aligned> gpu_res_float(
d_res_float);
Eigen::TensorMap<Eigen::Tensor<float, 2>, Eigen::Aligned> gpu_float(d_float, size, size);
Eigen::TensorMap<Eigen::Tensor<Eigen::half, 0>, Eigen::Aligned> gpu_res_half(d_res_half);
Eigen::TensorMap<Eigen::Tensor<Eigen::half, 0>, Eigen::Aligned> gpu_res_float(d_res_float);
gpu_float.device(gpu_device) = gpu_float.random();
@@ -418,9 +398,8 @@ void test_gpu_full_reductions() {
gpu_device.deallocate(d_res_float);
}
template<typename>
template <typename>
void test_gpu_forced_evals() {
Eigen::GpuStreamDevice stream;
Eigen::GpuDevice gpu_device(&stream);
int num_elem = 101;
@@ -430,14 +409,10 @@ void test_gpu_forced_evals() {
float* d_res_half2 = (float*)gpu_device.allocate(num_elem * sizeof(float));
float* d_res_float = (float*)gpu_device.allocate(num_elem * sizeof(float));
Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float(
d_float, num_elem);
Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_res_half1(
d_res_half1, num_elem);
Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Unaligned> gpu_res_half2(
d_res_half2, num_elem);
Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_res_float(
d_res_float, num_elem);
Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float(d_float, num_elem);
Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_res_half1(d_res_half1, num_elem);
Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Unaligned> gpu_res_half2(d_res_half2, num_elem);
Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_res_float(d_res_float, num_elem);
Eigen::array<int, 1> no_bcast;
no_bcast[0] = 1;
@@ -450,13 +425,14 @@ void test_gpu_forced_evals() {
Tensor<float, 1> half_prec1(num_elem);
Tensor<float, 1> half_prec2(num_elem);
Tensor<float, 1> full_prec(num_elem);
gpu_device.memcpyDeviceToHost(half_prec1.data(), d_res_half1, num_elem*sizeof(float));
gpu_device.memcpyDeviceToHost(half_prec2.data(), d_res_half2, num_elem*sizeof(float));
gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, num_elem*sizeof(float));
gpu_device.memcpyDeviceToHost(half_prec1.data(), d_res_half1, num_elem * sizeof(float));
gpu_device.memcpyDeviceToHost(half_prec2.data(), d_res_half2, num_elem * sizeof(float));
gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, num_elem * sizeof(float));
gpu_device.synchronize();
for (int i = 0; i < num_elem; ++i) {
std::cout << "Checking forced eval " << i << full_prec(i) << " vs " << half_prec1(i) << " vs " << half_prec2(i) << std::endl;
std::cout << "Checking forced eval " << i << full_prec(i) << " vs " << half_prec1(i) << " vs " << half_prec2(i)
<< std::endl;
VERIFY_IS_APPROX(full_prec(i), half_prec1(i));
VERIFY_IS_APPROX(full_prec(i), half_prec2(i));
}
@@ -468,9 +444,7 @@ void test_gpu_forced_evals() {
}
#endif
EIGEN_DECLARE_TEST(cxx11_tensor_of_float16_gpu)
{
EIGEN_DECLARE_TEST(cxx11_tensor_of_float16_gpu) {
CALL_SUBTEST_1(test_gpu_numext<void>());
#ifdef EIGEN_HAS_GPU_FP16