Factor out reduce_rows_f32 from common.cuh

This increases iteration cycle speed by not having to recompile
every kernel all the time

Author: ORippler

ggml/src/ggml-cuda/common.cuh

@@ -411,26 +411,6 @@ static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
 #endif // FP16_AVAILABLE
 }
 
-// Row reduction kernel template - compute sum (norm=false) or mean (norm=true)
-template<bool norm>
-static __global__ void reduce_rows_f32(const float * x, float * dst, const int ncols) {
-    const int row = blockIdx.x;
-    const int col = threadIdx.x;
-
-    float sum = 0.0f;
-    for (int i = col; i < ncols; i += blockDim.x) {
-        sum += x[row * ncols + i];
-    }
-
-    sum = warp_reduce_sum(sum);
-
-    if (col != 0) {
-        return;
-    }
-
-    dst[row] = norm ? sum / ncols : sum;
-}
-
 template<int width = WARP_SIZE>
 static __device__ __forceinline__ int warp_reduce_all(int x) {
 #ifdef GGML_USE_HIP

ggml/src/ggml-cuda/mean.cu

@@ -1,4 +1,5 @@
 #include "mean.cuh"
+#include "reduce_rows.cuh"
 
 void ggml_cuda_op_mean(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0   = dst->src[0];

ggml/src/ggml-cuda/reduce_rows.cuh

@@ -0,0 +1,21 @@
+#include "common.cuh"
+
+// Row reduction kernel template - compute sum (norm=false) or mean (norm=true)
+template<bool norm>
+static __global__ void reduce_rows_f32(const float * x, float * dst, const int ncols) {
+    const int row = blockIdx.x;
+    const int col = threadIdx.x;
+
+    float sum = 0.0f;
+    for (int i = col; i < ncols; i += blockDim.x) {
+        sum += x[row * ncols + i];
+    }
+
+    sum = warp_reduce_sum(sum);
+
+    if (col != 0) {
+        return;
+    }
+
+    dst[row] = norm ? sum / ncols : sum;
+}

ggml/src/ggml-cuda/sumrows.cu

@@ -1,4 +1,5 @@
 #include "sumrows.cuh"
+#include "reduce_rows.cuh"
 
 void sum_rows_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     const dim3 block_dims(WARP_SIZE, 1, 1);