Add fast math flag translation for OpenCL std lib #2762
Conversation
Such possibility was added in SPIR-V 1.6. This patch also introduces limited translation of nofpclass LLVM parameter attribute. Signed-off-by: Sidorov, Dmitry <[email protected]>
lib/SPIRV/SPIRVWriter.cpp
Outdated
| // Handle nofpclass attribute. Nothing to do if fast math flag is already | ||
| // set. | ||
| if ((BV->isExtInst() && static_cast<SPIRVExtInst *>( | ||
| BV)->getExtSetKind() == SPIRVEIS_OpenCL) && |
There was a problem hiding this comment.
Ideally we should also check, if fast math flags are applicable to a particular function from OpenCL std lib according to environmental spec, but I beg your pardon for not doing that.
Signed-off-by: Sidorov, Dmitry <[email protected]>
| @@ -3038,7 +3038,8 @@ bool LLVMToSPIRVBase::transDecoration(Value *V, SPIRVValue *BV) { | |||
| if (Opcode == Instruction::FAdd || Opcode == Instruction::FSub || | |||
| Opcode == Instruction::FMul || Opcode == Instruction::FDiv || | |||
| Opcode == Instruction::FRem || | |||
| ((Opcode == Instruction::FNeg || Opcode == Instruction::FCmp) && | |||
| ((Opcode == Instruction::FNeg || Opcode == Instruction::FCmp || | |||
There was a problem hiding this comment.
@MrSidims, could you check if SPIRVReader handles fast-math flags for FNeg and FCmp correctly, please?
If not, please, file a tracker to follow-up.
Unfortunately, SPIR-V format is not able to translate LLVM semantics correctly. @bashbaug, SPIR-V doesn't allow applying fast-math flags to OpSelect and OpCall instructions, but LLVM IR can have fast math flags attached to select and call instructions. Do you know if this is going to be addressed by future SPIR-V versions?
There was a problem hiding this comment.
SPIR-V doesn't allow applying fast-math flags to OpSelect and OpCall instructions, but LLVM IR can have fast math flags attached to select and call instructions
This is a part of SPV_KHR_float_controls2:
Replace the text following "Only valid on …" with:
All core instructions which use any floating-point type for either operands or result.
OpExInst extended instructions, where expressly permitted by the extended instruction set in use.
It may be a bit too much and some instructions can or should be excluded from the list, but it is what it is.
There was a problem hiding this comment.
Given:
OpExInst extended instructions, where expressly permitted by the extended instruction set in use.
Do we need some updates to the OpenCL Extended Instruction set, or perhaps the OpenCL SPIR-V environment spec? Feel free to file a SPIR-V or OpenCL issue for tracking, if so, so this doesn't slip through the cracks.
We should also add some targeted CTS tests for this functionality, even if they aren't exhaustive.
There was a problem hiding this comment.
could you check if SPIRVReader handles fast-math flags for FNeg and FCmp correctly, please
@bader according to passing tests https://github.com/KhronosGroup/SPIRV-LLVM-Translator/blob/main/test/transcoding/fcmp.ll and https://github.com/KhronosGroup/SPIRV-LLVM-Translator/blob/main/test/transcoding/fneg.ll it handles correctly. Manual run also doesn't show anything fishy.
There was a problem hiding this comment.
Do we need some updates to the OpenCL Extended Instruction set, or perhaps the OpenCL SPIR-V environment spec?
@bashbaug no, https://registry.khronos.org/SPIR-V/specs/unified1/OpenCL.ExtendedInstructionSet.100.pdf already contains the required lines like:
For environments that allow use of FPFastMathMode decorations on OpExtInst instructions,
FPFastMathMode decorations may be applied to the math instructions.
and there are other several references about it.
But I believe Alexey had another question related to select and call instructions.
We should also add some targeted CTS tests for this functionality, even if they aren't exhaustive.
We should, https://github.com/KhronosGroup/OpenCL-CTS/pulls?q=is%3Apr+is%3Aopen+FPFastMathMode search shows nothing
Signed-off-by: Sidorov, Dmitry <[email protected]>
| @@ -3038,7 +3038,8 @@ bool LLVMToSPIRVBase::transDecoration(Value *V, SPIRVValue *BV) { | |||
| if (Opcode == Instruction::FAdd || Opcode == Instruction::FSub || | |||
| Opcode == Instruction::FMul || Opcode == Instruction::FDiv || | |||
| Opcode == Instruction::FRem || | |||
| ((Opcode == Instruction::FNeg || Opcode == Instruction::FCmp) && | |||
| ((Opcode == Instruction::FNeg || Opcode == Instruction::FCmp || | |||
Such possibility was added in SPIR-V 1.6.
This patch also introduces limited translation of nofpclass LLVM parameter attribute.