]> git.djapps.eu Git - pkg/ggml/sources/llama.cpp/commit
overview / pkg / ggml / sources / llama.cpp / commit
summary | shortlog | log | commit | commitdiff | tree
(parent: 7faa746)
Adding a simple program to measure speed of dot products (#1041)
authorKawrakow <redacted>
Tue, 18 Apr 2023 19:00:14 +0000 (21:00 +0200)
committerGitHub <redacted>
Tue, 18 Apr 2023 19:00:14 +0000 (19:00 +0000)
commit5ecff35151156118c2df74899637ad34ee384b9b
tree7fb7a564ef23ccdb832a8c3d96f5a49b75c1d7datree
parent7faa7460f03bdd88becf1e659cf359f274055404commit | diff
Adding a simple program to measure speed of dot products (#1041)

On my Mac, the direct Q4_1 product is marginally slower
(~69 vs ~55 us for Q4_0). The SIMD-ified ggml version
is now almost 2X slower (~121 us).

On a Ryzen 7950X CPU, the direct product for Q4_1 quantization
is faster than the AVX2 implementation (~60 vs ~62 us).

---------

Co-authored-by: Iwan Kawrakow <redacted>
CMakeLists.txt diff | blob | history
Makefile diff | blob | history
pocs/CMakeLists.txt [new file with mode: 0644] blob
pocs/vdot/CMakeLists.txt [new file with mode: 0644] blob
pocs/vdot/vdot.cpp [new file with mode: 0644] blob
Packaging of ggml-org/llama.cpp