diff options
| author |  Antonio Sanchez <cantonios@google.com> | 2020-12-22 22:49:06 -0800 |
|---|---|---|
| committer | Antonio Sanchez <cantonios@google.com> | 2020-12-22 23:25:23 -0800 |
| commit | 070d303d56d46d2e018a58214da24ca629ea454f (patch) | |
| tree | 3dfa72bf48ffdca0a67bd794596e4e452d50ed19 /unsupported | |
| parent | fdf2ee62c5174441076fb64c9737d89bbe102759 (diff) | |
Add CUDA complex sqrt.
This is to support scalar `sqrt` of complex numbers `std::complex<T>` on
device, requested by Tensorflow folks.
Technically `std::complex` is not supported by NVCC on device
(though it is by clang), so the default `sqrt(std::complex<T>)` function only
works on the host. Here we create an overload to add back the
functionality.
Also modified the CMake file to add `--relaxed-constexpr` (or
equivalent) flag for NVCC to allow calling constexpr functions from
device functions, and added support for specifying compute architecture for
NVCC (was already available for clang).
Diffstat (limited to 'unsupported')
| -rw-r--r-- | unsupported/Eigen/CXX11/src/Tensor/TensorGpuHipCudaDefines.h | 8 |
1 files changed, 7 insertions, 1 deletions
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorGpuHipCudaDefines.h b/unsupported/Eigen/CXX11/src/Tensor/TensorGpuHipCudaDefines.h index f32ce27e9..cb53ce298 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorGpuHipCudaDefines.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorGpuHipCudaDefines.h @@ -16,7 +16,7 @@ // for some reason gets sent to the gcc/host compiler instead of the gpu/nvcc/hipcc compiler // When compiling such files, gcc will end up trying to pick up the CUDA headers by // default (see the code within "unsupported/Eigen/CXX11/Tensor" that is guarded by EIGEN_USE_GPU) -// This will obsviously not work when trying to compile tensorflow on a system with no CUDA +// This will obviously not work when trying to compile tensorflow on a system with no CUDA // To work around this issue for HIP systems (and leave the default behaviour intact), the // HIP tensorflow build defines EIGEN_USE_HIP when compiling all source files, and // "unsupported/Eigen/CXX11/Tensor" has been updated to use HIP header when EIGEN_USE_HIP is @@ -30,6 +30,9 @@ #define gpuSuccess hipSuccess #define gpuErrorNotReady hipErrorNotReady #define gpuGetDeviceCount hipGetDeviceCount +#define gpuGetLastError hipGetLastError +#define gpuPeekAtLastError hipPeekAtLastError +#define gpuGetErrorName hipGetErrorName #define gpuGetErrorString hipGetErrorString #define gpuGetDeviceProperties hipGetDeviceProperties #define gpuStreamDefault hipStreamDefault @@ -57,6 +60,9 @@ #define gpuSuccess cudaSuccess #define gpuErrorNotReady cudaErrorNotReady #define gpuGetDeviceCount cudaGetDeviceCount +#define gpuGetLastError cudaGetLastError +#define gpuPeekAtLastError cudaPeekAtLastError +#define gpuGetErrorName cudaGetErrorName #define gpuGetErrorString cudaGetErrorString #define gpuGetDeviceProperties cudaGetDeviceProperties #define gpuStreamDefault cudaStreamDefault |