backend: (x86) Extend x86 AVX instructions to single-precision #3887
Conversation
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## main #3887 +/- ##
==========================================
+ Coverage 91.25% 91.26% +0.01%
==========================================
Files 466 466
Lines 57945 57981 +36
Branches 5575 5575
==========================================
+ Hits 52879 52919 +40
+ Misses 3635 3626 -9
- Partials 1431 1436 +5 ☔ View full report in Codecov by Sentry. |
xdsl/dialects/x86/ops.py
Outdated
There was a problem hiding this comment.
could you please open a separate PR for these changes, with a change to the x86 roundtrip and assembly emission tests for this?
There was a problem hiding this comment.
after merging the other PRs I suggested in the other comments, could you please move this to tests/filecheck/projects/libxsmm? My feeling is that we don't need the rountrip test for the matmul since each individual op should already be handled by the dialect rountrip filecheck test.
| %c3_tmp0 = x86.rm.vmovups %ymm7, %rcx, 96 : (!x86.avx2reg<ymm7>, !x86.reg<rcx>) -> !x86.avx2reg<ymm7> | ||
| // Load column 0 of B | ||
| %b_col_0 = x86.rm.vbroadcastss %ymm8, %rsi, 0 : (!x86.avx2reg<ymm8>, !x86.reg<rsi>) -> !x86.avx2reg<ymm8> | ||
| // Go brrr |
There was a problem hiding this comment.
Haha it describes the magic of a well-formed fmas sequence which is analoguous to the one of a money printer (should I say a performance printer ?). Sorry I used it when writing the assembly. Will remove when submitting it again.
|
We use the convention of letting the original commenter resolve the comments on PRs, to make it easier to track conversations, to make sure that the point has been addressed. |
maybe something to add in the PR template? |
Oh ok |
qaco
force-pushed
the
hugo/instructions_for_matmul
branch
from
86668e9 to
dec2af3
Compare
No description provided.