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

[SYCL] refactor (#6408)

Browse source 
* seperate lower precision GEMM from the main files

* fix workgroup size hardcode
This commit is contained in:
Meng, Hengyu 2024-06-19 09:11:51 +08:00 • committed by GitHub
parent 37bef89433
commit 623494a478
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
13 changed files with 7600 additions and 6997 deletions
Download patch file Download diff file Expand all files Collapse all files

2
ggml-sycl/presets.hpp
View file

@ -18,8 +18,6 @@
#define GGML_SYCL_MAX_DEVICES 48
#define GGML_SYCL_NAME "SYCL"
// FIXME: 1024 from cuda
#define GROUP_SIZE 1024
#define WARP_SIZE 32
#define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses