CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-21 12:16:57 +00:00
Code Issues Packages Projects Releases Wiki Activity

CUDA: add mean operation (#14313)

Browse Source 
* CUDA: add mean operation

* add back sum_rows_f32_cuda

* Review: early exit if col!=0
This commit is contained in:
Aman Gupta
2025-06-22 12:39:54 +08:00
committed by GitHub
parent aa0ef5c578
commit aa064b2eb7
7 changed files with 54 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files

23
ggml/src/ggml-cuda/sumrows.cu
View File

@@ -1,25 +1,9 @@
#include "sumrows.cuh"
static __global__ void k_sum_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) {
dst[row] = sum;
}
}
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);
const dim3 block_nums(nrows, 1, 1);
k_sum_rows_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
reduce_rows_f32</*norm*/false><<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
}
void ggml_cuda_op_sum_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
@@ -35,5 +19,8 @@ void ggml_cuda_op_sum_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const int64_t ncols = src0->ne[0];
const int64_t nrows = ggml_nrows(src0);
sum_rows_f32_cuda(src0_d, dst_d, ncols, nrows, stream);
const dim3 block_dims(WARP_SIZE, 1, 1);
const dim3 block_nums(nrows, 1, 1);
reduce_rows_f32</*norm=*/false><<<block_nums, block_dims, 0, stream>>>(src0_d, dst_d, ncols);
}