ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-20 22:49:31 +00:00
Code Issues Packages Projects Releases Wiki Activity

Better concat for contiguous tensors

Browse Source 
If all the op does is to concatenate the second tensor
to the first, why would we want to have a loop?
This commit is contained in:
Iwan Kawrakow
2025-02-28 19:32:36 +02:00
parent 285b97b6bb
commit 84853b9a9b
1 changed files with 20 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
ggml/src/ggml.c
View File

@@ -12627,6 +12627,26 @@ static void ggml_compute_forward_concat_f32(
GGML_ASSERT(dim >= 0 && dim < 4);
if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && ggml_is_contiguous(dst) &&
(dim == 3 || (dim == 2 && dst->ne[3] == 1) || (dim == 1 && dst->ne[2]*dst->ne[3] == 1))) {
// simply copy the data
const int64_t size_src_0 = ggml_nbytes(src0);
const int64_t size_src_1 = ggml_nbytes(src1);
const int64_t block_size = 4096;
const int64_t num_blocks = (size_src_0 + size_src_1 + block_size - 1)/block_size;
for (int64_t i_block = ith; i_block < num_blocks; i_block += nth) {
const int64_t start = i_block*block_size;
if (start < size_src_0) {
int64_t copy_size = MIN(block_size, size_src_0 - start);
memcpy((char *)dst->data + start, (char *)src0->data + start, copy_size);
} else {
int64_t copy_size = MIN(block_size, size_src_0 + size_src_1 - start);
memcpy((char *)dst->data + start, (char *)src1->data + start - size_src_0, copy_size);
}
}
return;
}
int64_t o[4] = {0, 0, 0, 0};
o[dim] = src0->ne[dim];