vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Fix tensor load to device

Browse source 
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
This commit is contained in:
0cc4m 2023-05-16 18:49:49 +02:00
parent b3ff66d87f
commit de10afa80f
1 changed files with 10 additions and 4 deletions
Download patch file Download diff file Expand all files Collapse all files

14
ggml-opencl.cpp
View file

@ -919,13 +919,19 @@ void ggml_cl_transform_tensor(ggml_tensor * tensor) {
const size_t q_sz = ggml_type_size(type) * ne0 * ne1 * ne2 * ne3 / ggml_blck_size(type); const size_t q_sz = ggml_type_size(type) * ne0 * ne1 * ne2 * ne3 / ggml_blck_size(type);
size_t q_size; size_t q_size;
cl_mem* d_Q = (cl_mem*) malloc(sizeof(cl_mem)); cl_mem* dst = (cl_mem*) malloc(sizeof(cl_mem));
*d_Q = ggml_cl_pool_malloc(q_sz, &q_size, CL_MEM_READ_ONLY); *dst = ggml_cl_pool_malloc(q_sz, &q_size, CL_MEM_READ_ONLY);
// copy tensor to device // copy tensor to device
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, *d_Q, 0, tensor, 0, 0, NULL), "ggml_cl_h2d_tensor_2d"); for (int64_t i3 = 0; i3 < ne3; i3++) {
for (int64_t i2 = 0; i2 < ne2; i2++) {
int i = i3*ne2 + i2;
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, *dst, i*ne0*ne1, tensor, i3, i2, NULL), "ggml_cl_h2d_tensor_2d");
}
}
CL_CHECK(clFinish(queue), "clFinish"); CL_CHECK(clFinish(queue), "clFinish");
tensor->data = d_Q; tensor->data = dst;
tensor->backend = GGML_BACKEND_CL; tensor->backend = GGML_BACKEND_CL;
} }