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

SYCL: Introducing memory host pool (#11251)

Browse source 
* Implement host pool for matrix_info

Creating a new memory pool on the host to store memory location for
matrix_info needed to launch gemm_batch from oneMKL/oneMath.
Removing complex support in gemm_batch since it is not used in llama.cpp

* Remove unnecessary headers and cast

* Reorder member variable to avoid warning on initialization

* Formatting

* Remove unused variable

* Address PR review feedback - remove warning

---------

Signed-off-by: nscipione <nicolo.scipione@codeplay.com>
This commit is contained in:
Nicolò Scipione 2025-01-19 14:33:34 +01:00 • committed by GitHub
parent a1649cc13f
commit 99487b57d4
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
3 changed files with 137 additions and 103 deletions
Download patch file Download diff file Expand all files Collapse all files

13
ggml/src/ggml-sycl/common.hpp
View file

@ -333,8 +333,12 @@ struct ggml_backend_sycl_context {
// pool
std::unique_ptr<ggml_sycl_pool> pools[GGML_SYCL_MAX_DEVICES];
std::unique_ptr<ggml_sycl_pool> host_pools[GGML_SYCL_MAX_DEVICES];
static std::unique_ptr<ggml_sycl_pool> new_pool_for_device(queue_ptr qptr, int device);
static std::unique_ptr<ggml_sycl_pool> new_pool_for_host(queue_ptr qptr, int device);
ggml_sycl_pool & pool(int device) {
if (pools[device] == nullptr) {
pools[device] = new_pool_for_device(stream(device,0), device);
@ -345,6 +349,15 @@ struct ggml_backend_sycl_context {
ggml_sycl_pool & pool() {
return pool(device);
}
ggml_sycl_pool & host_pool(int device) {
if (host_pools[device] == nullptr) {
host_pools[device] = new_pool_for_host(stream(device, 0), device);
}
return *host_pools[device];
}
ggml_sycl_pool & host_pool() { return host_pool(device); }
};
// common device functions