Reenable "[ESIMD] Remove one of the uses on __SYCL_EXPLICIT_SIMD__ (#3242) #3311
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
…__ (intel#3242)" This patch is a part of the efforts for allowing ESIMD and regular SYCL kernels to coexist in the same translation unit and in the same program. Previously, in ESIMD device code we had calls to SPIRV intrinsics that didn't have definitions. With the change in spirv_vars.hpp, SYCL optimization passes convert calls to SPIRV intrinsics into loads from globals (SPIRV builtins). Thus, there is a need to change the implementation of LowerESIMD pass to lower such new constructs. Example: // @__spirv_BuiltInGlobalInvocationId = external dso_local local_unnamed_addr addrspace(1) constant <3 x i64>, align 32 // ... // %0 = load <3 x i64>, <3 x i64> addrspace(4)* addrspacecast // (<3 x i64> addrspace(1)* @__spirv_BuiltInLocalInvocationId // to <3 x i64> addrspace(4)*), align 32 // %1 = extractelement <3 x i64> %0, i64 0 // // => // // %.esimd = call <3 x i32> @llvm.genx.local.id.v3i32() // %local_id.x = extractelement <3 x i32> %.esimd, i32 0 // %local_id.x.cast.ty = zext i32 %local_id.x to i64 Current tests in sycl/test/esimd/spirv_intrins_trans.cpp check that there is no regression in how we lower SPRIV intrinsics into GenX counterparts. But also, I added some more tests.
kbobrovs
reviewed
Mar 6, 2021
| // this is ESIMD intrinsic - record for later translation | ||
| ESIMDIntrCalls.push_back(CI); | ||
| // Translate loads from SPIRV builtin globals into GenX intrinsics | ||
| auto *LI = dyn_cast<LoadInst>(&I); |
There was a problem hiding this comment.
I think the following would be more efficient from compile-time and code simplicity perspective and slightly more reliable:
- iterating through M.global_begin()/M.global_end() and finding SPIRV globals among them
- iterating through each global's uses and applying the esimd translation t-form.
This can be considered a Nit. But please add a TODO if you decide not to implement.
There was a problem hiding this comment.
That's a good point. I agree to reimplement it, but I would prefer to do it in a separate patch if you don't mind. We talked about splitting the LowerESIMD pass into two: ModulePass and FunctionPass. I think it would be appropriate to do those two changes together.
llvm/lib/SYCLLowerIR/LowerESIMD.cpp
Outdated
Comment on lines
1361
to
1362
| if (isa<GlobalVariable>(SpirvGlobal) && | ||
| SpirvGlobal->getName().startswith(SPIRV_INTRIN_PREF)) { |
There was a problem hiding this comment.
Suggested change
| if (isa<GlobalVariable>(SpirvGlobal) && | |
| SpirvGlobal->getName().startswith(SPIRV_INTRIN_PREF)) { | |
| if (!isa<GlobalVariable>(SpirvGlobal) || | |
| !SpirvGlobal->getName().startswith(SPIRV_INTRIN_PREF)) | |
| continue; |
llvm/lib/SYCLLowerIR/LowerESIMD.cpp
Outdated
| Value *TranslatedVal = translateSpirvIntrinsic( | ||
| EEI, SpirvGlobal->getName().drop_front(PrefLen)); | ||
|
|
||
| if (TranslatedVal) { |
There was a problem hiding this comment.
should this be rather an assert(TranslatedVal)? Failure to translate a load from a SPIR-V intrinsic var seems like an error.
There was a problem hiding this comment.
Yes. Will change.
romanovvlad
approved these changes
Mar 8, 2021
kbobrovs
approved these changes
Mar 9, 2021
| @@ -859,7 +859,7 @@ static Value *translateSpirvIntrinsic(ExtractElementInst *EEI, | |||
| else if (IndexC->equalsInt(2)) | |||
| Suff = 'z'; | |||
| else | |||
| return nullptr; | |||
| assert(false && "Extract element index should be either 0, 1, or 2"); | |||
There was a problem hiding this comment.
Nit: llvm_unrecheable seems a better fit for the purpose
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 patch is a part of the efforts for allowing ESIMD and
regular SYCL kernels to coexist in the same translation unit and
in the same program.
Previously, in ESIMD device code we had calls to SPIRV intrinsics
that didn't have definitions. With the change in spirv_vars.hpp,
SYCL optimization passes convert calls to SPIRV intrinsics into
loads from globals (SPIRV builtins). Thus, there is a need to change
the implementation of LowerESIMD pass to lower such new constructs.
Example:
Current tests in sycl/test/esimd/spirv_intrins_trans.cpp check that
there is no regression in how we lower SPRIV intrinsics into GenX
counterparts. But also, I added some more tests.