[CPU] Populate to_elements unrolling patterns in LLVM conversion.

[hanhanW]

It is a follow-up to #21851. The CPU backend also needs the patterns, which is not surprising. It was not detected in the first place because a test did not exist. It only happens in dynamic gather+attention cases.

The test is failing on VMVX and SPIR-V backend. Thus, the test is only added to CPU and hip for now.

Fixes #22007

[amd-eochoalo]

Should we also add it to

compiler/src/iree/compiler/Codegen/LLVMGPU/ConvertToNVVM.cpp

diff --git a/compiler/src/iree/compiler/Codegen/LLVMGPU/ConvertToNVVM.cpp b/compiler/src/iree/compiler/Codegen/LLVMGPU/ConvertToNVVM.cpp
index 8636d514a7..6bc6e27b5f 100644
--- a/compiler/src/iree/compiler/Codegen/LLVMGPU/ConvertToNVVM.cpp
+++ b/compiler/src/iree/compiler/Codegen/LLVMGPU/ConvertToNVVM.cpp
@@ -105,6 +105,7 @@ struct ConvertToNVVMPass final
       vector::populateVectorGatherLoweringPatterns(patterns);
       vector::populateVectorMaskOpLoweringPatterns(patterns);
       vector::populateVectorFromElementsLoweringPatterns(patterns);
+      vector::populateVectorToElementsLoweringPatterns(patterns);
       // We currently always use 64 bit indices, thus ensure the bit width of
       // the mask compare is consistent.
       vector::populateVectorMaskMaterializationPatterns(

[hanhanW]

Should we also add it to

compiler/src/iree/compiler/Codegen/LLVMGPU/ConvertToNVVM.cpp

diff --git a/compiler/src/iree/compiler/Codegen/LLVMGPU/ConvertToNVVM.cpp b/compiler/src/iree/compiler/Codegen/LLVMGPU/ConvertToNVVM.cpp
index 8636d514a7..6bc6e27b5f 100644
--- a/compiler/src/iree/compiler/Codegen/LLVMGPU/ConvertToNVVM.cpp
+++ b/compiler/src/iree/compiler/Codegen/LLVMGPU/ConvertToNVVM.cpp
@@ -105,6 +105,7 @@ struct ConvertToNVVMPass final
       vector::populateVectorGatherLoweringPatterns(patterns);
       vector::populateVectorMaskOpLoweringPatterns(patterns);
       vector::populateVectorFromElementsLoweringPatterns(patterns);
+      vector::populateVectorToElementsLoweringPatterns(patterns);
       // We currently always use 64 bit indices, thus ensure the bit width of
       // the mask compare is consistent.
       vector::populateVectorMaskMaterializationPatterns(

hmm, that is a mess and in maintenance mode; I don't really want to touch the code if not needed. I even don't know if we have CI for NVGPU or not. A solution may be having a utility that populates all the vector unrolling patterns, but I can see that the backend can diverge at some point. People should revamp the NVGPU codegen if they care about it. Adding new code to those non-active passes may be tech debt; nobody cares. (If someone cares, they'd raise a PR like this or file an issue.)

[amd-eochoalo]

@hanhanW thanks for the answer! No worries, I was just double checking in case it was necessary :)

[hanhanW]

@hanhanW thanks for the answer! No worries, I was just double checking in case it was necessary :)

No worries, thanks for asking! I'm happy to share my understanding with reviewers and developers.