cuda : tweak mm stride to double perf on P40 + GTX 970

Author: cebtenzzre

ggml-cuda.cu

@@ -467,7 +467,7 @@ static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUA
 #define GGML_CUDA_PEER_MAX_BATCH_SIZE 128
 #endif // GGML_CUDA_PEER_MAX_BATCH_SIZE
 
-#define MUL_MAT_SRC1_COL_STRIDE 128
+#define MUL_MAT_SRC1_COL_STRIDE 4096
 
 #define MAX_STREAMS 8
 static cudaStream_t g_cudaStreams[GGML_CUDA_MAX_DEVICES][MAX_STREAMS] = { { nullptr } };