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

Added support for simple coefficient wise tensor expression using half floats on CUDA devices

Browse Source
This commit is contained in:
Benoit Steiner
2016-02-19 08:19:12 +00:00
parent 0606a0a39b
commit ac5d706a94
3 changed files with 264 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

43
unsupported/test/cxx11_tensor_of_float16_cuda.cu
View File

@@ -20,6 +20,7 @@
using Eigen::Tensor;
#ifdef EIGEN_HAS_CUDA_FP16
void test_cuda_conversion() {
Eigen::CudaStreamDevice stream;
Eigen::GpuDevice gpu_device(&stream);
@@ -53,11 +54,53 @@ void test_cuda_conversion() {
gpu_device.deallocate(d_half);
gpu_device.deallocate(d_conv);
}
void test_cuda_elementwise() {
Eigen::CudaStreamDevice stream;
Eigen::GpuDevice gpu_device(&stream);
int num_elem = 101;
float* d_float1 = (float*)gpu_device.allocate(num_elem * sizeof(float));
float* d_float2 = (float*)gpu_device.allocate(num_elem * sizeof(float));
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);
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<half>() + gpu_float2.cast<half>()) * gpu_float1.cast<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));
for (int i = 0; i < num_elem; ++i) {
VERIFY_IS_APPROX(full_prec(i), half_prec(i));
}
gpu_device.deallocate(d_float1);
gpu_device.deallocate(d_float2);
gpu_device.deallocate(d_res_half);
gpu_device.deallocate(d_res_float);
}
#endif
void test_cxx11_tensor_of_float16_cuda()
{
#ifdef EIGEN_HAS_CUDA_FP16
CALL_SUBTEST_1(test_cuda_conversion());
CALL_SUBTEST_1(test_cuda_element_wise());
#endif
}