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

Made half floats usable on hardware that doesn't support them natively.

Browse Source
This commit is contained in:
Benoit Steiner
2016-03-11 17:21:42 -08:00
parent b72ffcb05e
commit 048c4d6efd
4 changed files with 343 additions and 58 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
unsupported/test/cxx11_tensor_cast_float16_cuda.cu
View File

@@ -19,30 +19,28 @@
using Eigen::Tensor;
#ifdef EIGEN_HAS_CUDA_FP16
void test_cuda_conversion() {
Eigen::CudaStreamDevice stream;
Eigen::GpuDevice gpu_device(&stream);
int num_elem = 101;
Tensor<float, 1> floats(num_elem);
floats.setRandom();
float* d_float = (float*)gpu_device.allocate(num_elem * sizeof(float));
half* d_half = (half*)gpu_device.allocate(num_elem * sizeof(half));
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<half, 1>, Eigen::Aligned> gpu_half(
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_device.memcpyHostToDevice(d_float, floats.data(), num_elem*sizeof(float));
gpu_half.device(gpu_device) = gpu_float.cast<half>();
gpu_half.device(gpu_device) = gpu_float.cast<Eigen::half>();
gpu_conv.device(gpu_device) = gpu_half.cast<float>();
Tensor<float, 1> initial(num_elem);
@@ -60,14 +58,8 @@ void test_cuda_conversion() {
gpu_device.deallocate(d_conv);
}
#endif
void test_cxx11_tensor_cast_float16_cuda()
{
#ifdef EIGEN_HAS_CUDA_FP16
CALL_SUBTEST(test_cuda_conversion());
#else
std::cout << "Half floats are not supported by this version of cuda: skipping the test" << std::endl;
#endif
}