Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Add fast math flag translation for OpenCL std lib #2762

Open
wants to merge 3 commits into
base: main
Choose a base branch
from
+153 −4
Open

Add fast math flag translation for OpenCL std lib #2762

Show file tree
Hide file tree
Changes from 2 commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
ce69ac7
Add fast math flag translation for OpenCL std lib
MrSidims Oct 10, 2024
0de7ccb
format
MrSidims Oct 16, 2024
f562615
Apply the suggestion
MrSidims Oct 18, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
3 changes: 3 additions & 0 deletions lib/SPIRV/SPIRVBuiltinHelper.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -106,6 +106,9 @@ Value *BuiltinCallMutator::doConversion() {
NewCall->copyMetadata(*CI);
NewCall->setAttributes(CallAttrs);
NewCall->setTailCall(CI->isTailCall());
if (isa<FPMathOperator>(CI))
NewCall->setFastMathFlags(CI->getFastMathFlags());

if (CI->hasFnAttr("fpbuiltin-max-error")) {
auto Attr = CI->getFnAttr("fpbuiltin-max-error");
NewCall->addFnAttr(Attr);
Expand Down
7 changes: 5 additions & 2 deletions lib/SPIRV/SPIRVReader.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2474,8 +2474,11 @@ Value *SPIRVToLLVM::transValueWithoutDecoration(SPIRVValue *BV, Function *F,
case OpExtInst: {
auto *ExtInst = static_cast<SPIRVExtInst *>(BV);
switch (ExtInst->getExtSetKind()) {
case SPIRVEIS_OpenCL:
return mapValue(BV, transOCLBuiltinFromExtInst(ExtInst, BB));
case SPIRVEIS_OpenCL: {
auto *V = mapValue(BV, transOCLBuiltinFromExtInst(ExtInst, BB));
applyFPFastMathModeDecorations(BV, static_cast<Instruction *>(V));
return V;
}
case SPIRVEIS_Debug:
case SPIRVEIS_OpenCL_DebugInfo_100:
case SPIRVEIS_NonSemantic_Shader_DebugInfo_100:
Expand Down
46 changes: 44 additions & 2 deletions lib/SPIRV/SPIRVWriter.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -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 ||
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@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?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

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.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

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.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

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.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks!

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

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

BV->isExtInst()) &&
BM->isAllowedToUseVersion(VersionNumber::SPIRV_1_6))) {
FastMathFlags FMF = BVF->getFastMathFlags();
SPIRVWord M{0};
Expand All @@ -3065,8 +3066,49 @@ bool LLVMToSPIRVBase::transDecoration(Value *V, SPIRVValue *BV) {
}
}
}
if (M != 0)
// Handle nofpclass attribute. Nothing to do if fast math flag is already
// set.
if ((BV->isExtInst() &&
static_cast<SPIRVExtInst *>(BV)->getExtSetKind() ==
SPIRVEIS_OpenCL) &&
BM->isAllowedToUseVersion(VersionNumber::SPIRV_1_6) &&
!(M & FPFastMathModeFastMask)) {
auto *F = cast<CallInst>(V)->getCalledFunction();
auto FAttrs = F->getAttributes();
AttributeSet RetAttrs = FAttrs.getRetAttrs();
if (RetAttrs.hasAttribute(Attribute::NoFPClass)) {
FPClassTest RetTest =
RetAttrs.getAttribute(Attribute::NoFPClass).getNoFPClass();
AttributeSet RetAttrs = FAttrs.getRetAttrs();
// Only Nan and Inf tests are representable in SPIR-V now.
bool ToAddNoNan = RetTest & fcNan;
bool ToAddNoInf = RetTest & fcInf;
if (ToAddNoNan || ToAddNoInf) {
const size_t NumParams = F->getFunctionType()->getNumParams();
for (size_t I = 0; I != NumParams; ++I) {
if (!F->hasParamAttribute(I, Attribute::NoFPClass)) {
bader marked this conversation as resolved.
Show resolved Hide resolved
ToAddNoNan = false;
ToAddNoInf = false;
break;
}
FPClassTest ArgTest =
FAttrs.getParamAttr(I, Attribute::NoFPClass).getNoFPClass();
ToAddNoNan = ToAddNoNan && static_cast<bool>(ArgTest & fcNan);
ToAddNoInf = ToAddNoInf && static_cast<bool>(ArgTest & fcInf);
}
}
if (ToAddNoNan)
M |= FPFastMathModeNotNaNMask;
if (ToAddNoInf)
M |= FPFastMathModeNotInfMask;
}
}
if (M != 0) {
BV->setFPFastMathMode(M);
if (Opcode == Instruction::FNeg || Opcode == Instruction::FCmp ||
BV->isExtInst())
BM->setMinSPIRVVersion(VersionNumber::SPIRV_1_6);
}
}
}
if (Instruction *Inst = dyn_cast<Instruction>(V)) {
Expand Down
78 changes: 78 additions & 0 deletions test/transcoding/fast-math-opencl-builtins.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,78 @@
; RUN: llvm-as %s -o %t.bc
; RUN: llvm-spirv -spirv-text %t.bc -o - | FileCheck %s --check-prefix=CHECK-SPIRV
; RUN: llvm-spirv %t.bc -o %t.spv
; RUN: spirv-val %t.spv
; RUN: llvm-spirv -r %t.spv -o - | llvm-dis -o - | FileCheck %s --check-prefix=CHECK-LLVM-OCL
; RUN: llvm-spirv -r --spirv-target-env=SPV-IR %t.spv -o - | llvm-dis -o - | FileCheck %s --check-prefix=CHECK-LLVM-SPV

; RUN: llvm-spirv -spirv-text --spirv-max-version=1.5 %t.bc -o - | FileCheck %s --check-prefix=CHECK-SPIRV-NEG

; CHECK-SPIRV: Decorate [[#FPDec1:]] FPFastMathMode 3
; CHECK-SPIRV: Decorate [[#FPDec2:]] FPFastMathMode 2
; CHECK-SPIRV: Decorate [[#FPDec3:]] FPFastMathMode 16
; CHECK-SPIRV: ExtInst [[#]] [[#FPDec1]] [[#]] fmax [[#]] [[#]]
; CHECK-SPIRV: ExtInst [[#]] [[#FPDec2]] [[#]] fmin [[#]] [[#]]
; CHECK-SPIRV: ExtInst [[#]] [[#FPDec3]] [[#]] fmax [[#]] [[#]]

; CHECK-SPIRV-NEG-NOT: Decorate [[#]] FPFastMathMode [[#]]

; CHECK-LLVM-OCL: call nnan ninf spir_func float @_Z4fmaxff(float %[[#]], float %[[#]])
; CHECK-LLVM-OCL: call ninf spir_func float @_Z4fminff(float %[[#]], float %[[#]])
; CHECK-LLVM-OCL: call fast spir_func float @_Z4fmaxff(float %[[#]], float %[[#]])

; CHECK-LLVM-SPV: call nnan ninf spir_func float @_Z16__spirv_ocl_fmaxff(float %[[#]], float %[[#]])
; CHECK-LLVM-SPV: call ninf spir_func float @_Z16__spirv_ocl_fminff(float %[[#]], float %[[#]])
; CHECK-LLVM-SPV: call fast spir_func float @_Z16__spirv_ocl_fmaxff(float %[[#]], float %[[#]])

; ModuleID = 'test.bc'
source_filename = "test.cpp"
target datalayout = "e-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024-n8:16:32:64"
target triple = "spir64-unknown-unknown"

@__spirv_BuiltInGlobalInvocationId = external dso_local local_unnamed_addr addrspace(1) constant <3 x i64>, align 32

declare dso_local spir_func noundef nofpclass(nan inf) float @_Z16__spirv_ocl_fmaxff(float noundef nofpclass(nan inf), float noundef nofpclass(nan inf)) local_unnamed_addr

declare dso_local spir_func noundef nofpclass(nan inf) float @_Z16__spirv_ocl_fminff(float noundef nofpclass(inf), float noundef nofpclass(nan inf)) local_unnamed_addr

define weak_odr dso_local spir_kernel void @nofpclass_all(ptr addrspace(1) noundef align 4 %_arg_data, ptr addrspace(1) noundef align 4 %_arg_dat1, ptr addrspace(1) noundef align 4 %_arg_dat2) local_unnamed_addr {
entry:
%0 = load i64, ptr addrspace(1) @__spirv_BuiltInGlobalInvocationId, align 32
%arrayidx.i = getelementptr inbounds float, ptr addrspace(1) %_arg_data, i64 %0
%arrayidx3.i = getelementptr inbounds float, ptr addrspace(1) %_arg_dat1, i64 %0
%cmp.i = icmp ult i64 %0, 2147483648
%arrayidx5.i = getelementptr inbounds float, ptr addrspace(1) %_arg_dat2, i64 %0
%1 = load float, ptr addrspace(1) %arrayidx3.i, align 4
%2 = load float, ptr addrspace(1) %arrayidx5.i, align 4
%call.i.i = tail call spir_func noundef nofpclass(nan inf) float @_Z16__spirv_ocl_fmaxff(float noundef nofpclass(nan inf) %1, float noundef nofpclass(nan inf) %2)
store float %call.i.i, ptr addrspace(1) %arrayidx.i, align 4
ret void
}

define weak_odr dso_local spir_kernel void @nofpclass_part(ptr addrspace(1) noundef align 4 %_arg_data, ptr addrspace(1) noundef align 4 %_arg_dat1, ptr addrspace(1) noundef align 4 %_arg_dat2) local_unnamed_addr {
entry:
%0 = load i64, ptr addrspace(1) @__spirv_BuiltInGlobalInvocationId, align 32
%arrayidx.i = getelementptr inbounds float, ptr addrspace(1) %_arg_data, i64 %0
%arrayidx3.i = getelementptr inbounds float, ptr addrspace(1) %_arg_dat1, i64 %0
%cmp.i = icmp ult i64 %0, 2147483648
%arrayidx5.i = getelementptr inbounds float, ptr addrspace(1) %_arg_dat2, i64 %0
%1 = load float, ptr addrspace(1) %arrayidx3.i, align 4
%2 = load float, ptr addrspace(1) %arrayidx5.i, align 4
%call.i.i = tail call spir_func noundef nofpclass(nan inf) float @_Z16__spirv_ocl_fminff(float noundef nofpclass(inf) %1, float noundef nofpclass(nan inf) %2)
store float %call.i.i, ptr addrspace(1) %arrayidx.i, align 4
ret void
}

define weak_odr dso_local spir_kernel void @nofpclass_fast(ptr addrspace(1) noundef align 4 %_arg_data, ptr addrspace(1) noundef align 4 %_arg_dat1, ptr addrspace(1) noundef align 4 %_arg_dat2) local_unnamed_addr {
entry:
%0 = load i64, ptr addrspace(1) @__spirv_BuiltInGlobalInvocationId, align 32
%arrayidx.i = getelementptr inbounds float, ptr addrspace(1) %_arg_data, i64 %0
%arrayidx3.i = getelementptr inbounds float, ptr addrspace(1) %_arg_dat1, i64 %0
%cmp.i = icmp ult i64 %0, 2147483648
%arrayidx5.i = getelementptr inbounds float, ptr addrspace(1) %_arg_dat2, i64 %0
%1 = load float, ptr addrspace(1) %arrayidx3.i, align 4
%2 = load float, ptr addrspace(1) %arrayidx5.i, align 4
%call.i.i = tail call fast spir_func noundef nofpclass(nan inf) float @_Z16__spirv_ocl_fmaxff(float noundef nofpclass(nan inf) %1, float noundef nofpclass(nan inf) %2)
store float %call.i.i, ptr addrspace(1) %arrayidx.i, align 4
ret void
}
Loading