Q8: use int8_t, AVX/AVX2 optimizations #972
Merged
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Note: this is not destined for master, but for the branch in PR #951
Change
block_q8_0
to useint8_t
and remove the additions and subtractions of 128. I can't test on ARM NEON and might have broken something.quantize_row_q8_0
NEON had+128.5
which was intended as aroundf
substitute, but I think that's wrong for negative numbers.Apart from that, this is mainly AVX and AVX2 optimizations.
With AVX2, text generation is roughly the same speed as current master.
BTW there seems to be something wrong with the Makefile dependencies. Switching between this branch and master with just
make
instead ofmake clean
makesmain
crash with a FPE.