CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-30 08:42:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

ggml-opencl : store GPU buffer in ggml_tensor::extra (#2994)

Browse Source
This commit is contained in:
slaren
2023-09-04 14:59:52 +02:00
committed by GitHub
parent 3103568144
commit bd33e5ab92
1 changed files with 6 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
ggml-opencl.cpp
View File

@@ -1334,7 +1334,7 @@ void ggml_cl_free_data(const struct ggml_tensor* tensor) {
return; return;
} }
cl_mem mem = (cl_mem)tensor->data; cl_mem mem = (cl_mem)tensor->extra;
clReleaseMemObject(mem); clReleaseMemObject(mem);
} }
@@ -1393,7 +1393,7 @@ static void ggml_cl_mul_f32(const ggml_tensor * src0, const ggml_tensor * src1,
size_t d_size; size_t d_size;
cl_mem d_X = ggml_cl_pool_malloc(ne0 * sizeof(float), &x_size); // src0 cl_mem d_X = ggml_cl_pool_malloc(ne0 * sizeof(float), &x_size); // src0
cl_mem d_Y = (cl_mem) src1->data; // src1 is already on device, broadcasted. cl_mem d_Y = (cl_mem) src1->extra; // src1 is already on device, broadcasted.
cl_mem d_D = ggml_cl_pool_malloc(ne0 * sizeof(float), &d_size); // dst cl_mem d_D = ggml_cl_pool_malloc(ne0 * sizeof(float), &d_size); // dst
@@ -1491,7 +1491,7 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
size_t d_size; size_t d_size;
cl_mem d_X; cl_mem d_X;
if (src0->backend == GGML_BACKEND_GPU) { // NOLINT if (src0->backend == GGML_BACKEND_GPU) { // NOLINT
d_X = (cl_mem) src0->data; d_X = (cl_mem) src0->extra;
} else { } else {
d_X = ggml_cl_pool_malloc(sizeof(float) * x_ne, &x_size); d_X = ggml_cl_pool_malloc(sizeof(float) * x_ne, &x_size);
} }
@@ -1567,7 +1567,7 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
size_t d_size; size_t d_size;
cl_mem d_X; cl_mem d_X;
if (src0->backend == GGML_BACKEND_GPU) { // NOLINT if (src0->backend == GGML_BACKEND_GPU) { // NOLINT
d_X = (cl_mem) src0->data; d_X = (cl_mem) src0->extra;
} else { } else {
d_X = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * x_ne, &x_size); d_X = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * x_ne, &x_size);
} }
@@ -1697,7 +1697,7 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
events.emplace_back(); events.emplace_back();
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Q, 0, src0, i03, i02, events.data() + ev_idx++)); CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Q, 0, src0, i03, i02, events.data() + ev_idx++));
} else if (src0->backend == GGML_BACKEND_GPU) { } else if (src0->backend == GGML_BACKEND_GPU) {
d_Q = (cl_mem) src0->data; d_Q = (cl_mem) src0->extra;
} else { } else {
GGML_ASSERT(false); GGML_ASSERT(false);
} }
@@ -1860,6 +1860,6 @@ void ggml_cl_transform_tensor(void * data, ggml_tensor * tensor) {
CL_CHECK(clFinish(queue)); CL_CHECK(clFinish(queue));
tensor->data = dst; tensor->extra = dst;
GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU); GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
} }