enable splitk for mixed precision gemm #339
Open
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
taozha2
requested review from
joeatodd,
tdeng5,
mehdi-goli,
jiyang1011 and
aacostadiaz
taozha2
force-pushed
the
zt/splitK_MixedPrecision
branch
from
c8d2bfa to
f9855bd
Compare
taozha2
force-pushed
the
zt/splitK_MixedPrecision
branch
from
f9855bd to
24b6043
Compare
|
INT4 is very important for us and we need the peak performance for it. |
rolandschulz
reviewed
May 2, 2025
| for (int i = 0; i < decltype(size(tCrA_mma))::value; i++) { | ||
| tCrA_mma[i] = static_cast<DstType>(tCrA_load[i].get()); | ||
| } | ||
| convert_int_subbyte_to_half(out, in); |
There was a problem hiding this comment.
What's the point of adding in and out?
Suggested change
| convert_int_subbyte_to_half(out, in); | |
| convert_int_subbyte_to_half(tCrA_mma, tCrA_load); |
| if constexpr (sizeof_bits_v<SrcType> < 8) { | ||
| // TODO (Codeplay): Current NumericArrayConverter doesn't work for int4 on intel Xe, just workaround and |
There was a problem hiding this comment.
You didn't address the TODO. You added a different helper function rather than making the existing one work correctly.
| static constexpr auto v_cnt = decltype(size(out))::value / scalar / loop_cnt; | ||
| static constexpr auto is_src_signed = is_signed<SrcType>::value; | ||
|
|
||
| auto src_ptr = reinterpret_cast<const format_type*>(raw_pointer_cast(&(in.data()[0]))); |
There was a problem hiding this comment.
Why the &(...)[0]?
Suggested change
| auto src_ptr = reinterpret_cast<const format_type*>(raw_pointer_cast(&(in.data()[0]))); | |
| auto src_ptr = reinterpret_cast<const format_type*>(raw_pointer_cast(in.data())); |
It isn't obvious why this isn't UB and why the reinterpret_cast is needed. Both should be explained in a comment (also next line).
| static constexpr auto is_src_signed = is_signed<SrcType>::value; | ||
|
|
||
| auto src_ptr = reinterpret_cast<const format_type*>(raw_pointer_cast(&(in.data()[0]))); | ||
| auto&& dst_ptr = *(cute::intel::vector_t<ushort, decltype(size(out))::value>*)(out.data()); |
There was a problem hiding this comment.
Why auto&&? We shouldn't use C-style cast. dst_ptr isn't a pointer, is it?
Suggested change
| auto&& dst_ptr = *(cute::intel::vector_t<ushort, decltype(size(out))::value>*)(out.data()); | |
| auto& dst = *reinterpret_cast<cute::intel::vector_t<ushort, decltype(size(out))::value>*>(out.data()); |
Why does this not need a raw_pointer_cast?
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
No description provided.