diff options
| author | Nicolas Patry <patry.nicolas@protonmail.com> | 2023-08-10 16:23:17 +0200 |
|---|---|---|
| committer | GitHub <noreply@github.com> | 2023-08-10 16:23:17 +0200 |
| commit | 0b0fa56978b9fe722250d52c87cf39f9d8b2acdb (patch) | |
| tree | eed0d5d77bff6a26d756d387b0a14f282eaa109b | |
| parent | 385f0d261c9199b83373fc10b7b52a1fef67a9d0 (diff) | |
| parent | 66d1c093e029e789ab66acc0b747643cb7b05ffa (diff) | |
| download | candle-0b0fa56978b9fe722250d52c87cf39f9d8b2acdb.tar.gz candle-0b0fa56978b9fe722250d52c87cf39f9d8b2acdb.tar.bz2 candle-0b0fa56978b9fe722250d52c87cf39f9d8b2acdb.zip | |
Merge pull request #386 from huggingface/enabling_61_maybe
This is duplicated code on Cuda 12.2.
| -rw-r--r-- | candle-kernels/src/compatibility.cuh | 18 |
1 files changed, 0 insertions, 18 deletions
diff --git a/candle-kernels/src/compatibility.cuh b/candle-kernels/src/compatibility.cuh index 2df8e921..5a22f4bc 100644 --- a/candle-kernels/src/compatibility.cuh +++ b/candle-kernels/src/compatibility.cuh @@ -6,24 +6,6 @@ // FIXME: the minimum compute capabilities are just guesses since the table is not specific enough -// #if __CUDA_ARCH__ < 600 -// __device__ __forceinline__ __half __hmax(__half a, __half b) { -// return __float2half(fmaxf(__half2float(a), __half2float(b))); -// } -// __device__ __forceinline__ __half __hmin(__half a, __half b) { -// return __float2half(fminf(__half2float(a), __half2float(b))); -// } -// #endif - -#if __CUDA_ARCH__ < 800 -__device__ __forceinline__ __half __hmax_nan(__half a, __half b) { - // return __hisnan(a) ? a : (__hisnan(b) ? b : __hmax(a, b)); -} -__device__ __forceinline__ __half __hmin_nan(__half a, __half b) { - // return __hisnan(a) ? a : (__hisnan(b) ? b : __hmin(a, b)); -} -#endif - #if __CUDA_ARCH__ < 600 // Copied from https://docs.nvidia.com/cuda/cuda-c-programming-guide/#atomic-functions __device__ double atomicAdd(double* address, double val) { |