[CodeGen][CUDA] Vectorization for intrinsics #5101
Conversation
|
Fixing missing features exposed by #4968 |
wpan11nv
force-pushed
the
intrinsic_vectorize
branch
from
b3be9bd to
6f21588
Compare
|
This seems like a great change! Have you done any tests on how it affects performance? I'd love to know how much this speeds things up. |
I measured vectorization benefits on a vector add micro-benchmark in this PR #4968 The speedup could be as high as 20+%. This is a missing feature to enable PR4968. |
wpan11nv
force-pushed
the
intrinsic_vectorize
branch
from
6f21588 to
56a61d1
Compare
wpan11nv
force-pushed
the
intrinsic_vectorize
branch
from
56a61d1 to
e1563b4
Compare
| // | ||
| // Emit an unsupported vector call | ||
| // | ||
| // v = intrin_f((float4*)A[0], (float4*)B[0]) |
There was a problem hiding this comment.
do you mean intrin_f(((float4*)A)[0], ((float4*)B)[0])?
There was a problem hiding this comment.
that is the CallNode representation, not supported in CUDA. We are going to emit a few scalar calls instead here.
- This allows to emit vectorized loads/stores for CUDA math intrinsics. - A few intrinsics should be lowered as CUDAMath not CUDAFastMath ones. - Fixed the code block identation.
wpan11nv
force-pushed
the
intrinsic_vectorize
branch
from
e1563b4 to
9937bea
Compare
|
Thanks @wpan11nv this is merged |
- This allows to emit vectorized loads/stores for CUDA math intrinsics. - A few intrinsics should be lowered as CUDAMath not CUDAFastMath ones. - Fixed the code block identation.
- This allows to emit vectorized loads/stores for CUDA math intrinsics. - A few intrinsics should be lowered as CUDAMath not CUDAFastMath ones. - Fixed the code block identation.
This allows to emit vectorized loads/stores
for CUDA math intrinsics.
Fixed a few intrinsics that should be lowered as
CUDAMath not CUDAFastMath ones.
Thanks for contributing to TVM! Please refer to guideline https://docs.tvm.ai/contribute/ for useful information and tips. After the pull request is submitted, please request code reviews from Reviewers by @ them in the pull request thread.