[LLVM][GPU][+refactoring] Replacement of math intrinsics with library calls #835
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
|
Logfiles from GitLab pipeline #44990 (:white_check_mark:) have been uploaded here! Status and direct links: |
8 tasks
iomaganaris
approved these changes
Apr 5, 2022
| DISPATCH("llvm.pow.f64", "_ZGVeN8vv_pow", FIXED(8)) | ||
| // clang-format on | ||
| }; | ||
| #undef DISPATCH |
There was a problem hiding this comment.
FIXED should also be undefined?
Suggested change
| #undef DISPATCH | |
| #undef DISPATCH | |
| #undef FIXED |
Comment on lines
113
to
121
| // Add vectorizable functions to the target library info. | ||
| switch (library->second) { | ||
| case VecLib::LIBMVEC_X86: | ||
| if (!triple.isX86() || !triple.isArch64Bit()) | ||
| break; | ||
| default: | ||
| tli.addVectorizableFunctionsFromVecLib(library->second); | ||
| break; | ||
| } |
There was a problem hiding this comment.
Just a personal opinion, not sure what should be the proper way to do it or what could be the benefit of the switch but I think it would be more understandable to write this like:
Suggested change
| // Add vectorizable functions to the target library info. | |
| switch (library->second) { | |
| case VecLib::LIBMVEC_X86: | |
| if (!triple.isX86() || !triple.isArch64Bit()) | |
| break; | |
| default: | |
| tli.addVectorizableFunctionsFromVecLib(library->second); | |
| break; | |
| } | |
| if (library->second != VecLib::LIBMVEC_X86 || (triple.isX86() && triple.isArch64Bit())) { | |
| tli.addVectorizableFunctionsFromVecLib(library->second); | |
| } |
Feel free to keep this as you prefer
|
|
||
| // Map of supported replacements. For now it is only exp. | ||
| static const std::map<std::string, std::string> libdevice_name = { | ||
| {"llvm.exp.f32", "__nv_expf"}, |
There was a problem hiding this comment.
I think it would be good to add also pow and maybe look for any other math function commonly used in the mod files and also add it to the x86 and aarch64 maps. I can look at this also the next days
|
Logfiles from GitLab pipeline #47044 (:no_entry:) have been uploaded here! Status and direct links: |
|
Logfiles from GitLab pipeline #47063 (:white_check_mark:) have been uploaded here! Status and direct links: |
|
Logfiles from GitLab pipeline #47062 (:white_check_mark:) have been uploaded here! Status and direct links: |
|
Logfiles from GitLab pipeline #47154 (:white_check_mark:) have been uploaded here! Status and direct links: |
|
Logfiles from GitLab pipeline #47153 (:white_check_mark:) have been uploaded here! Status and direct links: |
iomaganaris
added a commit
that referenced
this pull request
May 10, 2022
… calls (#835) Added an LLVM pass that replaces math intrinsics with calls to math library. In particular: * Functionality of replacement with SIMD functions is factored out into a separate file and LLVM version dependencies are dropped (LLVM 13 is already used anyway). * A pass to replace intrinsics with libdevice calls when targeting CUDA platforms has been added. So far only `exp` and `pow` are supported (single and double precision). * Added a test to check the replacement Co-authored-by: Ioannis Magkanaris <iomagkanaris@gmail.com>
iomaganaris
added a commit
that referenced
this pull request
May 12, 2022
… calls (#835) Added an LLVM pass that replaces math intrinsics with calls to math library. In particular: * Functionality of replacement with SIMD functions is factored out into a separate file and LLVM version dependencies are dropped (LLVM 13 is already used anyway). * A pass to replace intrinsics with libdevice calls when targeting CUDA platforms has been added. So far only `exp` and `pow` are supported (single and double precision). * Added a test to check the replacement Co-authored-by: Ioannis Magkanaris <iomagkanaris@gmail.com>
iomaganaris
added a commit
that referenced
this pull request
Sep 15, 2022
… calls (#835) Added an LLVM pass that replaces math intrinsics with calls to math library. In particular: * Functionality of replacement with SIMD functions is factored out into a separate file and LLVM version dependencies are dropped (LLVM 13 is already used anyway). * A pass to replace intrinsics with libdevice calls when targeting CUDA platforms has been added. So far only `exp` and `pow` are supported (single and double precision). * Added a test to check the replacement Co-authored-by: Ioannis Magkanaris <iomagkanaris@gmail.com>
iomaganaris
added a commit
that referenced
this pull request
Sep 15, 2022
… calls (#835) Added an LLVM pass that replaces math intrinsics with calls to math library. In particular: * Functionality of replacement with SIMD functions is factored out into a separate file and LLVM version dependencies are dropped (LLVM 13 is already used anyway). * A pass to replace intrinsics with libdevice calls when targeting CUDA platforms has been added. So far only `exp` and `pow` are supported (single and double precision). * Added a test to check the replacement Co-authored-by: Ioannis Magkanaris <iomagkanaris@gmail.com>
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.
This PR adds a LLVM pass that replaces math intrinsics
with calls to math library. In particular:
Functionality of replacement with SIMD functions is factored
out into a separate file and LLVM version dependencies are
dropped (we use LLVM 13 already anyway).
A pass to replace intrinsics with libdevice calls when targeting
CUDA platforms has been added. So far only
expis supported(single and double precision)
Added a test to check the replacement
Note: factoring replacement functionality into a separate file
allows us to completely drop the dependency on target information
inside
LLVMCodegenVisitor😊