Faster AVX2 matrix multiplications for lgacy quants #405
Conversation
There was a problem hiding this comment.
Looks good to me. Approved. Could you sync to head please? I needed to change the way your earlier contribution is compiled in an effort to make room in the binary size for flash attention. I basically just renamed a file and added an if statement which uses X86_HAVE(AVX2) to do dispatching at runtime. That helped me get your first iteration into a release and I can cut another once this is merged too. Thanks!
Somehow memcpy is kind of slow, so for getting 4 bytes from 2-byte-aligned data it is faster to just do or on two consecutive 16-bit entries.
However, the way it is currently, we have lost the zen4-tuned version.
ikawrakow
force-pushed
the
ik/new_legacy_mul_mat
branch
from
1baed32
to
a3bca82
Compare
|
@jart I adapted to head. But to get back the Ryzen-7950X performance I had to make two separate |
I'd encourage you to work your magic on Cosmopolitan's memcpy() function. https://github.com/jart/cosmopolitan/blob/master/libc/intrin/memmove.c You can run the tests by either running |
|
Also, did you notice this? https://www.phoronix.com/news/Llamafile-0.8.2-More-AVX2 Congrats! |
It seems some people still use the
ggmllegacy qunatsQ4_0, Q4_1, Q5_0andQ5_1, so here is a PR that improves matrix multiplication performance for these quants on AVX2. The gains forQ4_1, Q5_0andQ5_1, which do not have tiniBLAS implementation are very significant, but evenQ4_0is faster than tinyBLAS (see table below).I have gone for a templated implementation. This costs 2-3% in performance but reduces the code by at least a factor of 2.
The implementation requires at least a C++14 compiler because I have used
autofor the return type of two functions. Is this a problem?Prompt processing speed for a 512-token prompt (PP-512) for a 7B LLaMA model
The PR can also help with token generation (TG) speed. On my system TG is fully memory bound for more than 4-8 threads (depending on quantization type). So, to have a better illustration of the performance differences, here are the TG-128 results with just 2 threads on a Ryzen-7950X for a 7B LLaMA model: