use johannes implementation instead

LostRuins · May 1, 2024 · f5e6709 · f5e6709
1 parent e9978bf
commit f5e6709
Show file tree

Hide file tree

Showing 2 changed files with 24 additions and 27 deletions.
diff --git a/ggml-cuda/common.cuh b/ggml-cuda/common.cuh
@@ -137,7 +137,8 @@
 #define STRINGIZE(...) STRINGIZE_IMPL(__VA_ARGS__)
 
 #define WARP_SIZE 32
-#define CUDART_HMAX     11070 // CUDA 11.7, min. ver. for which __hmax and __hmax2 are known to work (may be higher than needed)
+#define CUDART_HMAX   11070 // CUDA 11.7, min. ver. for which __hmax and __hmax2 are known to work (may be higher than needed)
+#define CUDART_HMASK  12000 // CUDA 12.0, min. ver. for half2 -> uint mask comparisons
 
 #define CC_PASCAL     600
 #define MIN_CC_DP4A   610 // minimum compute capability for __dp4a, an intrinsic for byte-wise dot products
@@ -293,8 +294,27 @@ static __device__ __forceinline__ float warp_reduce_max(float x) {
     return x;
 }
 
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
+static __device__ __forceinline__ half hmax(const half a, const half b) {
+    return __half2float(a) > __half2float(b) ? a : b;
+}
+static __device__ __forceinline__ half2 hmax2(const half2 a, const half2 b) {
+    half2 ret;
+    reinterpret_cast<half&>(ret.x) =  __low2float(a) >  __low2float(b) ?  __low2half(a) :  __low2half(b);
+    reinterpret_cast<half&>(ret.y) = __high2float(a) > __high2float(b) ? __high2half(a) : __high2half(b);
+    return ret;
+}
+#else
+static __device__ __inline__ __half hmax(const __half a, const __half b) {
+    return __hmax(a,b);
+}
+static __device__ __inline__ __half2 hmax2(const __half2 a, const __half2 b) {
+    return __hmax2(a,b);
+}
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
+
 static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL && CUDART_VERSION >= CUDART_HMAX
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
 #pragma unroll
    for (int mask = 16; mask > 0; mask >>= 1) {
        x = __hmax2(x, __shfl_xor_sync(0xffffffff, x, mask, 32));
@@ -303,10 +323,10 @@ static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
 #else
    GGML_UNUSED(x);
    NO_DEVICE_CODE;
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL && CUDART_VERSION >= CUDART_HMAX
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
 }
 
-#if CUDART_VERSION < 12000
+#if CUDART_VERSION < CUDART_HMASK
 static __device__ __forceinline__ uint32_t __hgt2_mask(const half2 a, const half2 b) {
     const uint32_t mask_low  = 0x0000FFFF * (float( __low2half(a)) > float( __low2half(b)));
     const uint32_t mask_high = 0xFFFF0000 * (float(__high2half(a)) > float(__high2half(b)));

diff --git a/ggml-cuda/fattn.cu b/ggml-cuda/fattn.cu
@@ -11,29 +11,6 @@
 #define HALF_MAX_HALF         __float2half(65504.0f/2) // Use neg. of this instead of -INFINITY to initialize KQ max vals to avoid NaN upon subtraction.
 #define SOFTMAX_FTZ_THRESHOLD -20.0f                   // Softmax exp. of values smaller than this are flushed to zero to avoid NaNs.
 
-//hack: polyfill hmax and hmax2 for older cuda version
-#if CUDART_VERSION < CUDART_HMAX
-__device__ __inline__ __half hmax(const __half a, const __half b) {
-    const float fa = __half2float(a);
-    const float fb = __half2float(b);
-    return __float2half(fa > fb ? fa : fb);
-}
-__device__ __inline__ __half2 hmax2(const __half2 a, const __half2 b) {
-    __half2 result;
-    result.x = hmax(a.x, b.x);
-    result.y = hmax(a.y, b.y);
-    return result;
-}
-#else
-__device__ __inline__ __half hmax(const __half a, const __half b) {
-    return __hmax(a,b);
-}
-__device__ __inline__ __half2 hmax2(const __half2 a, const __half2 b) {
-    return __hmax2(a,b);
-}
-#endif
-
-
 template<int D, int parallel_blocks> // D == head size
 __launch_bounds__(((D + WARP_SIZE - 1) / WARP_SIZE)*WARP_SIZE, 1)
 static __global__ void flash_attn_vec_ext_f16(