A few more MMAIntrinsics #19099
Merged
A few more MMAIntrinsics #19099
+133
−8
Conversation
This file contains 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
bjacob
force-pushed
the
users/bjacob/more-mma-intrinsics
branch
from
a831afb
to
50c222c
Compare
This was referenced Nov 11, 2024
bjacob
force-pushed
the
users/bjacob/more-mma-intrinsics
branch
from
50c222c
to
ffc4d58
Compare
bjacob
requested review from
kuhar,
antiagainst and
qedawkins
as code owners
kuhar
approved these changes
Nov 11, 2024
bjacob
force-pushed
the
users/bjacob/refactor-mma-intrinsics
branch
from
a44474a
to
17dc529
Compare
bjacob
force-pushed
the
users/bjacob/more-mma-intrinsics
branch
from
ffc4d58
to
09d91a5
Compare
kuhar
approved these changes
Nov 11, 2024
Yes :-) |
bjacob
added a commit
that referenced
this pull request
Nov 12, 2024
This adds some e2e matmul tests on CPU for f64 (double precision). Motivation: - This is technically supported by IREE. - This was already > 50% implemented in the e2e matmul testing framework, and the partly implemented status was not a good steady state to be in. - My real motivation is #19099, I'm doing a round of AMD MFMA / data tiling improvements, thought we should support a denser set of MFMA intrinsics, noticed this has been supported since CDNA2 so we shouldn't be prevented from supporting it just because our e2e matmul tests can't cover it. Signed-off-by: Benoit Jacob <[email protected]>
bjacob
added a commit
that referenced
this pull request
Nov 12, 2024
Some code simplifications in `IREEGPUAttrs.cpp`, and some shuffling of the MMAIntrinsic enum, making it easier to add more MMAIntrinsics in the future (see #19099). --------- Signed-off-by: Benoit Jacob <[email protected]>
Base automatically changed from
users/bjacob/refactor-mma-intrinsics
to
main
November 12, 2024 16:20
Signed-off-by: Benoit Jacob <[email protected]>
bjacob
force-pushed
the
users/bjacob/more-mma-intrinsics
branch
from
09d91a5
to
ab2aa4c
Compare
Merged
ScottTodd
reviewed
Nov 13, 2024
Comment on lines
+277
to
+282
| case MMAIntrinsic::WMMA_F32_16x16x16_BF16: { | ||
| return {bf16, bf16, f32}; | ||
| } | ||
| case MMAIntrinsic::WMMA_BF16_16x16x16_BF16: { | ||
| return {bf16, bf16, bf16}; | ||
| } |
There was a problem hiding this comment.
gcc build error somewhere in this function: https://github.com/iree-org/iree/actions/runs/11814370974/job/32913293915#step:4:7077
[6444/8163] Building CXX object compiler/src/iree/compiler/Codegen/Dialect/GPU/IR/CMakeFiles/iree_compiler_Codegen_Dialect_GPU_IR_IREEGPUDialect.objects.dir/IREEGPUAttrs.cpp.o
FAILED: compiler/src/iree/compiler/Codegen/Dialect/GPU/IR/CMakeFiles/iree_compiler_Codegen_Dialect_GPU_IR_IREEGPUDialect.objects.dir/IREEGPUAttrs.cpp.o
/usr/local/bin/ccache /usr/bin/g++-9 -I/__w/iree/iree -I/__w/iree/iree/build-gcc -I/__w/iree/iree/third_party/llvm-project/llvm/include -I/__w/iree/iree/build-gcc/llvm-project/include -I/__w/iree/iree/third_party/llvm-project/mlir/include -I/__w/iree/iree/build-gcc/llvm-project/tools/mlir/include -I/__w/iree/iree/third_party/llvm-project/lld/include -I/__w/iree/iree/build-gcc/llvm-project/tools/lld/include -I/__w/iree/iree/compiler/src -I/__w/iree/iree/build-gcc/compiler/src -I/__w/iree/iree/runtime/src -I/__w/iree/iree/build-gcc/runtime/src -I/__w/iree/iree/build-gcc/runtime/src/iree/base/internal/flatcc -I/__w/iree/iree/compiler/bindings/c/. -I/__w/iree/iree/build-gcc/compiler/bindings/c/. -isystem /__w/iree/iree/third_party/flatcc/include -O3 -fPIC -fvisibility=hidden -fno-rtti -fno-exceptions -Wall -Werror -Wno-error=deprecated-declarations -Wno-address -Wno-address-of-packed-member -Wno-comment -Wno-format-zero-length -Wno-uninitialized -Wno-overloaded-virtual -Wno-invalid-offsetof -Wno-sign-compare -Wno-unused-function -Wno-unknown-pragmas -Wno-unused-but-set-variable -Wno-misleading-indentation -fmacro-prefix-map=/__w/iree/iree=iree -I/__w/iree/iree/third_party/flatcc/include/ -I/__w/iree/iree/third_party/flatcc/include/flatcc/reflection/ -std=gnu++17 -MD -MT compiler/src/iree/compiler/Codegen/Dialect/GPU/IR/CMakeFiles/iree_compiler_Codegen_Dialect_GPU_IR_IREEGPUDialect.objects.dir/IREEGPUAttrs.cpp.o -MF compiler/src/iree/compiler/Codegen/Dialect/GPU/IR/CMakeFiles/iree_compiler_Codegen_Dialect_GPU_IR_IREEGPUDialect.objects.dir/IREEGPUAttrs.cpp.o.d -o compiler/src/iree/compiler/Codegen/Dialect/GPU/IR/CMakeFiles/iree_compiler_Codegen_Dialect_GPU_IR_IREEGPUDialect.objects.dir/IREEGPUAttrs.cpp.o -c /__w/iree/iree/compiler/src/iree/compiler/Codegen/Dialect/GPU/IR/IREEGPUAttrs.cpp
/__w/iree/iree/compiler/src/iree/compiler/Codegen/Dialect/GPU/IR/IREEGPUAttrs.cpp: In function 'std::tuple<mlir::Type, mlir::Type, mlir::Type> mlir::iree_compiler::IREE::GPU::getABCElementTypes(mlir::MLIRContext*, mlir::iree_compiler::IREE::GPU::MMAIntrinsic)':
/__w/iree/iree/compiler/src/iree/compiler/Codegen/Dialect/GPU/IR/IREEGPUAttrs.cpp:293:1: error: control reaches end of non-void function [-Werror=return-type]
293 | }
| ^
cc1plus: all warnings being treated as errors
unhandled switch case?
There was a problem hiding this comment.
Thanks for reporting! I'll fix asap (just bad combination right now of: I'm WFH, the server i've been dev on is down, and my workstation in the office is turned off).
There was a problem hiding this comment.
And this is not a missing switch case (Clang would complain in that case) but GCC rightfully pointing out that even though we have handled all switch cases, it is still unsafe to assume that one of them is taken. See llvm/llvm-project#115345.
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.
Supporting some MMAIntrinsics that were omitted for no particular reason, and that seemed like they could be useful. In particular, supporting the f64 intrinsic present in CDNA2/3, adding mixed-f8-types support in CDNA3, and adding bf16 support to CDNA2 and RDNA3.