mirror/whisper.cpp
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

ggml-opencl, llama: using reserve() if count already known (llama/7272)

Browse Source
This commit is contained in:
Herman Semenov 2024-05-20 07:33:21 +00:00 committed by Georgi Gerganov
parent eb26f55b40
commit cc1ae10989
1 changed files with 5 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
ggml-opencl.cpp
View File

@ -1,4 +1,4 @@
#include "ggml.h" #include "ggml.h"
#include "ggml-opencl.h" #include "ggml-opencl.h"
#include "ggml-backend-impl.h" #include "ggml-backend-impl.h"
@ -1835,7 +1835,10 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
CL_CHECK(clEnqueueNDRangeKernel(queue, *to_fp32_cl, 1, &offset, &global, local > 0 ? &local : NULL, events.size(), !events.empty() ? events.data() : NULL, NULL)); CL_CHECK(clEnqueueNDRangeKernel(queue, *to_fp32_cl, 1, &offset, &global, local > 0 ? &local : NULL, events.size(), !events.empty() ? events.data() : NULL, NULL));
} }
for (int64_t i12 = i02 * r2, e12 = i12 + r2; i12 < e12; i12++) { int64_t i12 = i02 * r2;
int64_t e12 = i12 + r2;
events.reserve(e12 - i12);
for (; i12 < e12; i12++) {
if (mul_mat_vec) { // specialized dequantize_mul_mat_vec kernel if (mul_mat_vec) { // specialized dequantize_mul_mat_vec kernel
// copy src1 to device // copy src1 to device
events.emplace_back(); events.emplace_back();