From 66d1c093e029e789ab66acc0b747643cb7b05ffa Mon Sep 17 00:00:00 2001 From: Nicolas Patry Date: Thu, 10 Aug 2023 09:20:18 +0200 Subject: [PATCH] This is duplicated code on Cuda 12.2. Without it we can compile for 52 (but I get Operation Not supported when actually trying to use those kernels). --- candle-kernels/src/compatibility.cuh | 18 ------------------ 1 file changed, 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) {