fix NVCC version on Makefile, __halves2half2 -> make_half2

ggml-org · Jul 15, 2023 · 6bfbdf8 · 6bfbdf8
1 parent 83595ec
commit 6bfbdf8
Show file tree

Hide file tree

Showing 2 changed files with 2 additions and 1 deletion.
diff --git a/Makefile b/Makefile
@@ -170,6 +170,7 @@ ifdef LLAMA_CUDA
 	OBJS      += ggml-cuda.o
 	NVCC      = nvcc
 	NVCCFLAGS = --forward-unknown-to-host-compiler
+	NVCCV 	  := $(shell $(NVCC) --version | tail -n 1)
 ifdef LLAMA_DEBUG
 	NVCCFLAGS += -lineinfo
 endif # LLAMA_DEBUG

diff --git a/ggml-cuda-kern.h b/ggml-cuda-kern.h
@@ -14,7 +14,7 @@ template<> struct vec2_t_impl<float>  { typedef float2 type; };
 template<typename T> using vec2_t = typename vec2_t_impl<T>::type;
 
 template<typename T> inline __host__ __device__ vec2_t<T> make_vec2_t(const T & x, const T & y);
-template<> inline __host__ __device__ vec2_t<half>  make_vec2_t(const  half & x, const  half & y) { return __halves2half2(x, y); }
+template<> inline __host__ __device__ vec2_t<half>  make_vec2_t(const  half & x, const  half & y) { return make_half2 (x, y); }
 template<> inline __host__ __device__ vec2_t<float> make_vec2_t(const float & x, const float & y) { return make_float2(x, y); }
 
 // the cuda headers define operators for half2, but not for float2