Track all customizations that are made in llvm-spirv in intel/llvm

[MrSidims]

The goal is to keep intel/llvm's llvm-spirv on par with https://github.com/KhronosGroup/SPIRV-LLVM-Translator as much as possible. I'm not expecting it to be always on par, since some patches in https://github.com/KhronosGroup/SPIRV-LLVM-Translator requires appropriate update on devices' drivers in CI here, so the list should be updated.

Following patches are missing in this repository in comparison to https://github.com/KhronosGroup/SPIRV-LLVM-Translator :

• 1. Add SPIR-V 1.4 checks. It's going to be resolved by [SPIR-V] Cherry-pick of "Add SPIR-V 1.4 checks" #7493 , but it's blocked by GPU driver updates on windows.
• 2. Add an entry point wrapper around functions (llvm pass) . It depends on a patch 1.
• 3. Translate function attributes (incl. parameter attrs) for entry point kernels . It depends on a patch 2.
• 4. Fix the collection of entry point interfaces It depends on a patch 2.

This repository adds extra code:

• 1. In SPIRVRegularizeLLVM.cpp we have a 'handling' of freeze instruction. We should either upstream W/A or implement a proper handling of the instruction in Khronos. See Unsupported instruction: freeze KhronosGroup/SPIRV-LLVM-Translator#1140 for reference.
• 2. This PR adds a W/A to replace GEP on ID-related builtin GVs with extractelement instruction. Need to remove it, when we have a fix in SPIR-V consumer propagated to intel/llvm's CI as part of driver updates.
• 3. Added disable-spirv-tools LIT parameter to llvm-spirv tests to explicitly disable using spirv-tools in intel/llvm CI ([CI] Temporarily disable tests requiring spirv-tools in CI #16743). These changes will be reverted once the test failures are fixed.

This repository removes some code:

• 1. This PR adds W/A that removes FunctionParameterAttributeNoReadWrite, need to return it.
• 2. This PR Removes binding between FPFastMathModeINTEL capability and SPV_INTEL_fp_fast_math_mode extension.

[MrSidims]

@vmaksimo please take a look and add notes if I'm missing something.

[bader]

The goal is to keep intel/llvm's llvm-spirv on pair with https://github.com/KhronosGroup/SPIRV-LLVM-Translator as much as possible. I'm not expecting it to be always on pair, since some patches in https://github.com/KhronosGroup/SPIRV-LLVM-Translator requires appropriate update on devices' drivers in CI here, so the list should be updated.

In my opinion we should remove llvm-spirv sources from intel/llvm repository. It difficult to track the difference between sources in Khronos and intel repositories. Custom layout makes it hard to sync and maintain. I suggest we consider:

1. Add an intel branch in Khronos repository which should aim to be "always on par" with the main development branch but can be used to keep patches for Intel device. Non-SYCL compilers targeting Intel devices should benefit from this branch as well. If Khronos is against vendor branches in their repository, we can consider using a fork under intel organization.
2. Keep set of patches on top of Khronos main branch in intel/llvm repository instead of the copy of sources. Keeping track of changes should be a no-brainer.

[vmaksimo]

please take a look and add notes if I'm missing something.

In this repo we also have DecorationFuncParamDescINTEL and DecorationFuncParamKindINTEL, which are not present in Khronos. Other things mentioned are the same as in #3686 issue, I'll close it as a duplicate.

[asudarsa]

The goal is to keep intel/llvm's llvm-spirv on par with https://github.com/KhronosGroup/SPIRV-LLVM-Translator as much as possible. I'm not expecting it to be always on par, since some patches in https://github.com/KhronosGroup/SPIRV-LLVM-Translator requires appropriate update on devices' drivers in CI here, so the list should be updated.

Following patches are missing in this repository in comparison to https://github.com/KhronosGroup/SPIRV-LLVM-Translator :

• 1. Add SPIR-V 1.4 checks. It's going to be resolved by [SPIR-V] Cherry-pick of "Add SPIR-V 1.4 checks" #7493 , but it's blocked by GPU driver updates on windows.
• 2. Add an entry point wrapper around functions (llvm pass) . It depends on a patch 1.
• 3. Translate function attributes (incl. parameter attrs) for entry point kernels . It depends on a patch 2.
• 4. Fix the collection of entry point interfaces It depends on a patch 2.

This repository adds extra code:

• 1. In SPIRVRegularizeLLVM.cpp we have a 'handling' of freeze instruction. We should either upstream W/A or implement a proper handling of the instruction in Khronos. See Unsupported instruction: freeze KhronosGroup/SPIRV-LLVM-Translator#1140 for reference.
• 2. This PR adds a W/A to replace GEP on ID-related builtin GVs with extractelement instruction. Need to remove it, when we have a fix in SPIR-V consumer propagated to intel/llvm's CI as part of driver updates.

This repository removes some code:

• 1. This PR adds W/A that removes FunctionParameterAttributeNoReadWrite, need to return it.

Hi @MrSidims

I see PR in #1 (Add SPIR-V 1.4 checks) has been merged. Can we proceed with the remaining pointers here?

Thanks

[bader]

@MrSidims, @asudarsa, #11186 - it seems like we missed a commit made 3 years ago. Can you check if anything else is missing, please?

[AlexeySachkov]

Linking pulldown PR #11920 in here and in particular this comment and this commit - it should be reverted once KhronosGroup/SPIRV-LLVM-Translator#2224 is resolved. @jsji, @MrSidims: FYI

[jsji]

KhronosGroup/SPIRV-LLVM-Translator@64fe75e388dc3b7 is reverted in #12868 due to failures .
Failed Tests (3):
SYCL :: NonUniformGroups/ballot_group_algorithms.cpp
SYCL :: NonUniformGroups/opportunistic_group_algorithms.cpp
SYCL :: NonUniformGroups/tangle_group_algorithms.cpp

"[7:34 PM] Sidorov, Dmitry
I was checking if we indeed pass runtime value and it's not some sort of missing constant propagation. Unfortunately id passed to OpGroupNonUniformBroadcast is always runtime value for SYCL headers. Hence we need to revert the patch. "

[asudarsa]

One more:
bbf3800

[jsji]

One more: bbf3800

Yes, this is the one mentioned above #7592 (comment)

[asudarsa]

Ah..Thanks. I missed that. Will be addressing these soon.

[maarquitos14]

#15298 reverts the W/A that removes FunctionParameterAttributeNoReadWrite. Marked as solved in the description the corresponding item.

[MrSidims]

Another customization was removed in #15882

[uditagarwal97]

Added disable-spirv-tools LIT parameter to llvm-spirv tests to explicitly disable using spirv-tools in intel/llvm CI (#16743). These changes will be reverted once the test failures are fixed.