This is an archive of the discontinued LLVM Phabricator instance.

Differential D50376

AMDGPU: Fix enabling denormals by default on pre-VI targets
ClosedPublic

Authored by arsenm on Aug 7 2018, 2:46 AM.

Details

Reviewers
b-sumner
kzhuravl
Summary

Fast FMAF is not a sufficient condition to enable denormals.
Before VI, enabling denormals caused F32 instructions to
run at F64 speeds.

Diff Detail

Event Timeline

arsenm created this revision.Aug 7 2018, 2:46 AM
Herald added subscribers: t-tye, tpr, dstuttard and 3 others. · View Herald TranscriptAug 7 2018, 2:46 AM
b-sumner added a comment.Aug 7 2018, 7:39 AM

This approach seems fine to me.

kzhuravl accepted this revision.Aug 8 2018, 6:04 AM

LGTM

This revision is now accepted and ready to land.Aug 8 2018, 6:04 AM
arsenm closed this revision.Aug 8 2018, 10:49 AM

r339278

Revision Contents

PathSize
lib/
Basic/
Targets/
131 lines
3 lines
test/
CodeGenOpenCL/
36 lines

Diff 159471

lib/Basic/Targets/AMDGPU.h

Show First 20 Lines • Show All 88 Lines • ▼ Show 20 Lines struct GPUInfo {
llvm::StringLiteral Name; llvm::StringLiteral Name;
llvm::StringLiteral CanonicalName; llvm::StringLiteral CanonicalName;
AMDGPUTargetInfo::GPUKind Kind; AMDGPUTargetInfo::GPUKind Kind;
bool HasFMAF; bool HasFMAF;
bool HasFastFMAF; bool HasFastFMAF;
bool HasLDEXPF; bool HasLDEXPF;
bool HasFP64; bool HasFP64;
bool HasFastFMA; bool HasFastFMA;
bool HasFullRateF32Denorms;
}; };
static constexpr GPUInfo InvalidGPU = static constexpr GPUInfo InvalidGPU =
{{""}, {""}, GK_NONE, false, false, false, false, false}; {{""}, {""}, GK_NONE, false, false, false, false, false, false};
static constexpr GPUInfo R600GPUs[26] = { static constexpr GPUInfo R600GPUs[26] = {
// Name Canonical Kind Has Has Has Has Has // Name Canonical Kind Has Has Has Has Has Has
// Name FMAF Fast LDEXPF FP64 Fast // Name FMAF Fast LDEXPF FP64 Fast Fast
// FMAF FMA // FMAF FMA Denorm
{{"r600"}, {"r600"}, GK_R600, false, false, false, false, false}, {{"r600"}, {"r600"}, GK_R600, false, false, false, false, false, false},
{{"rv630"}, {"r600"}, GK_R600, false, false, false, false, false}, {{"rv630"}, {"r600"}, GK_R600, false, false, false, false, false, false},
{{"rv635"}, {"r600"}, GK_R600, false, false, false, false, false}, {{"rv635"}, {"r600"}, GK_R600, false, false, false, false, false, false},
{{"r630"}, {"r630"}, GK_R630, false, false, false, false, false}, {{"r630"}, {"r630"}, GK_R630, false, false, false, false, false, false},
{{"rs780"}, {"rs880"}, GK_RS880, false, false, false, false, false}, {{"rs780"}, {"rs880"}, GK_RS880, false, false, false, false, false, false},
{{"rs880"}, {"rs880"}, GK_RS880, false, false, false, false, false}, {{"rs880"}, {"rs880"}, GK_RS880, false, false, false, false, false, false},
{{"rv610"}, {"rs880"}, GK_RS880, false, false, false, false, false}, {{"rv610"}, {"rs880"}, GK_RS880, false, false, false, false, false, false},
{{"rv620"}, {"rs880"}, GK_RS880, false, false, false, false, false}, {{"rv620"}, {"rs880"}, GK_RS880, false, false, false, false, false, false},
{{"rv670"}, {"rv670"}, GK_RV670, false, false, false, false, false}, {{"rv670"}, {"rv670"}, GK_RV670, false, false, false, false, false, false},
{{"rv710"}, {"rv710"}, GK_RV710, false, false, false, false, false}, {{"rv710"}, {"rv710"}, GK_RV710, false, false, false, false, false, false},
{{"rv730"}, {"rv730"}, GK_RV730, false, false, false, false, false}, {{"rv730"}, {"rv730"}, GK_RV730, false, false, false, false, false, false},
{{"rv740"}, {"rv770"}, GK_RV770, false, false, false, false, false}, {{"rv740"}, {"rv770"}, GK_RV770, false, false, false, false, false, false},
{{"rv770"}, {"rv770"}, GK_RV770, false, false, false, false, false}, {{"rv770"}, {"rv770"}, GK_RV770, false, false, false, false, false, false},
{{"cedar"}, {"cedar"}, GK_CEDAR, false, false, false, false, false}, {{"cedar"}, {"cedar"}, GK_CEDAR, false, false, false, false, false, false},
{{"palm"}, {"cedar"}, GK_CEDAR, false, false, false, false, false}, {{"palm"}, {"cedar"}, GK_CEDAR, false, false, false, false, false, false},
{{"cypress"}, {"cypress"}, GK_CYPRESS, true, false, false, false, false}, {{"cypress"}, {"cypress"}, GK_CYPRESS, true, false, false, false, false, false},
{{"hemlock"}, {"cypress"}, GK_CYPRESS, true, false, false, false, false}, {{"hemlock"}, {"cypress"}, GK_CYPRESS, true, false, false, false, false, false},
{{"juniper"}, {"juniper"}, GK_JUNIPER, false, false, false, false, false}, {{"juniper"}, {"juniper"}, GK_JUNIPER, false, false, false, false, false, false},
{{"redwood"}, {"redwood"}, GK_REDWOOD, false, false, false, false, false}, {{"redwood"}, {"redwood"}, GK_REDWOOD, false, false, false, false, false, false},
{{"sumo"}, {"sumo"}, GK_SUMO, false, false, false, false, false}, {{"sumo"}, {"sumo"}, GK_SUMO, false, false, false, false, false, false},
{{"sumo2"}, {"sumo"}, GK_SUMO, false, false, false, false, false}, {{"sumo2"}, {"sumo"}, GK_SUMO, false, false, false, false, false, false},
{{"barts"}, {"barts"}, GK_BARTS, false, false, false, false, false}, {{"barts"}, {"barts"}, GK_BARTS, false, false, false, false, false, false},
{{"caicos"}, {"caicos"}, GK_BARTS, false, false, false, false, false}, {{"caicos"}, {"caicos"}, GK_BARTS, false, false, false, false, false, false},
{{"aruba"}, {"cayman"}, GK_CAYMAN, true, false, false, false, false}, {{"aruba"}, {"cayman"}, GK_CAYMAN, true, false, false, false, false, false},
{{"cayman"}, {"cayman"}, GK_CAYMAN, true, false, false, false, false}, {{"cayman"}, {"cayman"}, GK_CAYMAN, true, false, false, false, false, false},
{{"turks"}, {"turks"}, GK_TURKS, false, false, false, false, false}, {{"turks"}, {"turks"}, GK_TURKS, false, false, false, false, false, false},
}; };
static constexpr GPUInfo AMDGCNGPUs[32] = { static constexpr GPUInfo AMDGCNGPUs[32] = {
// Name Canonical Kind Has Has Has Has Has // Name Canonical Kind Has Has Has Has Has Has
// Name FMAF Fast LDEXPF FP64 Fast // Name FMAF Fast LDEXPF FP64 Fast Fast
// FMAF FMA // FMAF FMA Denorm
{{"gfx600"}, {"gfx600"}, GK_GFX600, true, true, true, true, true}, {{"gfx600"}, {"gfx600"}, GK_GFX600, true, true, true, true, true, false},
{{"tahiti"}, {"gfx600"}, GK_GFX600, true, true, true, true, true}, {{"tahiti"}, {"gfx600"}, GK_GFX600, true, true, true, true, true, false},
{{"gfx601"}, {"gfx601"}, GK_GFX601, true, false, true, true, true}, {{"gfx601"}, {"gfx601"}, GK_GFX601, true, false, true, true, true, false},
{{"hainan"}, {"gfx601"}, GK_GFX601, true, false, true, true, true}, {{"hainan"}, {"gfx601"}, GK_GFX601, true, false, true, true, true, false},
{{"oland"}, {"gfx601"}, GK_GFX601, true, false, true, true, true}, {{"oland"}, {"gfx601"}, GK_GFX601, true, false, true, true, true, false},
{{"pitcairn"}, {"gfx601"}, GK_GFX601, true, false, true, true, true}, {{"pitcairn"}, {"gfx601"}, GK_GFX601, true, false, true, true, true, false},
{{"verde"}, {"gfx601"}, GK_GFX601, true, false, true, true, true}, {{"verde"}, {"gfx601"}, GK_GFX601, true, false, true, true, true, false},
{{"gfx700"}, {"gfx700"}, GK_GFX700, true, false, true, true, true}, {{"gfx700"}, {"gfx700"}, GK_GFX700, true, false, true, true, true, false},
{{"kaveri"}, {"gfx700"}, GK_GFX700, true, false, true, true, true}, {{"kaveri"}, {"gfx700"}, GK_GFX700, true, false, true, true, true, false},
{{"gfx701"}, {"gfx701"}, GK_GFX701, true, true, true, true, true}, {{"gfx701"}, {"gfx701"}, GK_GFX701, true, true, true, true, true, false},
{{"hawaii"}, {"gfx701"}, GK_GFX701, true, true, true, true, true}, {{"hawaii"}, {"gfx701"}, GK_GFX701, true, true, true, true, true, false},
{{"gfx702"}, {"gfx702"}, GK_GFX702, true, true, true, true, true}, {{"gfx702"}, {"gfx702"}, GK_GFX702, true, true, true, true, true, false},
{{"gfx703"}, {"gfx703"}, GK_GFX703, true, false, true, true, true}, {{"gfx703"}, {"gfx703"}, GK_GFX703, true, false, true, true, true, false},
{{"kabini"}, {"gfx703"}, GK_GFX703, true, false, true, true, true}, {{"kabini"}, {"gfx703"}, GK_GFX703, true, false, true, true, true, false},
{{"mullins"}, {"gfx703"}, GK_GFX703, true, false, true, true, true}, {{"mullins"}, {"gfx703"}, GK_GFX703, true, false, true, true, true, false},
{{"gfx704"}, {"gfx704"}, GK_GFX704, true, false, true, true, true}, {{"gfx704"}, {"gfx704"}, GK_GFX704, true, false, true, true, true, false},
{{"bonaire"}, {"gfx704"}, GK_GFX704, true, false, true, true, true}, {{"bonaire"}, {"gfx704"}, GK_GFX704, true, false, true, true, true, false},
{{"gfx801"}, {"gfx801"}, GK_GFX801, true, true, true, true, true}, {{"gfx801"}, {"gfx801"}, GK_GFX801, true, true, true, true, true, true},
{{"carrizo"}, {"gfx801"}, GK_GFX801, true, true, true, true, true}, {{"carrizo"}, {"gfx801"}, GK_GFX801, true, true, true, true, true, true},
{{"gfx802"}, {"gfx802"}, GK_GFX802, true, false, true, true, true}, {{"gfx802"}, {"gfx802"}, GK_GFX802, true, false, true, true, true, true},
{{"iceland"}, {"gfx802"}, GK_GFX802, true, false, true, true, true}, {{"iceland"}, {"gfx802"}, GK_GFX802, true, false, true, true, true, true},
{{"tonga"}, {"gfx802"}, GK_GFX802, true, false, true, true, true}, {{"tonga"}, {"gfx802"}, GK_GFX802, true, false, true, true, true, true},
{{"gfx803"}, {"gfx803"}, GK_GFX803, true, false, true, true, true}, {{"gfx803"}, {"gfx803"}, GK_GFX803, true, false, true, true, true, true},
{{"fiji"}, {"gfx803"}, GK_GFX803, true, false, true, true, true}, {{"fiji"}, {"gfx803"}, GK_GFX803, true, false, true, true, true, true},
{{"polaris10"}, {"gfx803"}, GK_GFX803, true, false, true, true, true}, {{"polaris10"}, {"gfx803"}, GK_GFX803, true, false, true, true, true, true},
{{"polaris11"}, {"gfx803"}, GK_GFX803, true, false, true, true, true}, {{"polaris11"}, {"gfx803"}, GK_GFX803, true, false, true, true, true, true},
{{"gfx810"}, {"gfx810"}, GK_GFX810, true, false, true, true, true}, {{"gfx810"}, {"gfx810"}, GK_GFX810, true, false, true, true, true, true},
{{"stoney"}, {"gfx810"}, GK_GFX810, true, false, true, true, true}, {{"stoney"}, {"gfx810"}, GK_GFX810, true, false, true, true, true, true},
{{"gfx900"}, {"gfx900"}, GK_GFX900, true, true, true, true, true}, {{"gfx900"}, {"gfx900"}, GK_GFX900, true, true, true, true, true, true},
{{"gfx902"}, {"gfx902"}, GK_GFX900, true, true, true, true, true}, {{"gfx902"}, {"gfx902"}, GK_GFX900, true, true, true, true, true, true},
{{"gfx904"}, {"gfx904"}, GK_GFX904, true, true, true, true, true}, {{"gfx904"}, {"gfx904"}, GK_GFX904, true, true, true, true, true, true},
{{"gfx906"}, {"gfx906"}, GK_GFX906, true, true, true, true, true}, {{"gfx906"}, {"gfx906"}, GK_GFX906, true, true, true, true, true, true},
}; };
static GPUInfo parseR600Name(StringRef Name); static GPUInfo parseR600Name(StringRef Name);
static GPUInfo parseAMDGCNName(StringRef Name); static GPUInfo parseAMDGCNName(StringRef Name);
GPUInfo parseGPUName(StringRef Name) const; GPUInfo parseGPUName(StringRef Name) const;
▲ Show 20 Lines • Show All 276 Lines • Show Last 20 Lines

lib/Basic/Targets/AMDGPU.cpp

Show First 20 Lines • Show All 204 Lines • ▼ Show 20 Linesvoid AMDGPUTargetInfo::adjustTargetOptions(const CodeGenOptions &CGOpts,
for (auto &I : TargetOpts.FeaturesAsWritten) { for (auto &I : TargetOpts.FeaturesAsWritten) {
if (I == "+fp32-denormals" || I == "-fp32-denormals") if (I == "+fp32-denormals" || I == "-fp32-denormals")
hasFP32Denormals = true; hasFP32Denormals = true;
if (I == "+fp64-fp16-denormals" || I == "-fp64-fp16-denormals") if (I == "+fp64-fp16-denormals" || I == "-fp64-fp16-denormals")
hasFP64Denormals = true; hasFP64Denormals = true;
} }
if (!hasFP32Denormals) if (!hasFP32Denormals)
TargetOpts.Features.push_back( TargetOpts.Features.push_back(
(Twine(CGOptsGPU.HasFastFMAF && !CGOpts.FlushDenorm (Twine(CGOptsGPU.HasFastFMAF && CGOptsGPU.HasFullRateF32Denorms &&
!CGOpts.FlushDenorm
? '+' ? '+'
: '-') + : '-') +
Twine("fp32-denormals")) Twine("fp32-denormals"))
.str()); .str());
// Always do not flush fp64 or fp16 denorms. // Always do not flush fp64 or fp16 denorms.
if (!hasFP64Denormals && CGOptsGPU.HasFP64) if (!hasFP64Denormals && CGOptsGPU.HasFP64)
TargetOpts.Features.push_back("+fp64-fp16-denormals"); TargetOpts.Features.push_back("+fp64-fp16-denormals");
} }
▲ Show 20 Lines • Show All 108 Lines • Show Last 20 Lines

test/CodeGenOpenCL/denorms-are-zero.cl

// RUN: %clang_cc1 -emit-llvm -o - %s | FileCheck %s // RUN: %clang_cc1 -emit-llvm -o - %s | FileCheck %s
// RUN: %clang_cc1 -emit-llvm -cl-denorms-are-zero -o - %s | FileCheck %s --check-prefix=DENORM-ZERO // RUN: %clang_cc1 -emit-llvm -cl-denorms-are-zero -o - %s | FileCheck -check-prefix=DENORM-ZERO %s
// RUN: %clang_cc1 -emit-llvm -cl-denorms-are-zero -o - -triple amdgcn--amdhsa -target-cpu fiji %s | FileCheck %s --check-prefix=AMDGCN
// RUN: %clang_cc1 -emit-llvm -o - -triple amdgcn--amdhsa -target-cpu fiji %s | FileCheck %s --check-prefix=AMDGCN-DENORM // Slow FMAF and slow f32 denormals
// RUN: %clang_cc1 -emit-llvm -target-feature +fp32-denormals -target-feature -fp64-fp16-denormals -cl-denorms-are-zero -o - -triple amdgcn--amdhsa -target-cpu fiji %s | FileCheck --check-prefix=AMDGCN-FEATURE %s // RUN: %clang_cc1 -emit-llvm -o - -triple amdgcn--amdhsa -target-cpu pitcairn %s | FileCheck -check-prefixes=AMDGCN,AMDGCN-FLUSH %s
// RUN: %clang_cc1 -emit-llvm -cl-denorms-are-zero -o - -triple amdgcn--amdhsa -target-cpu pitcairn %s | FileCheck -check-prefixes=AMDGCN,AMDGCN-FLUSH-OPT %s
// Fast FMAF, but slow f32 denormals
// RUN: %clang_cc1 -emit-llvm -o - -triple amdgcn--amdhsa -target-cpu tahiti %s | FileCheck -check-prefixes=AMDGCN,AMDGCN-FLUSH %s
// RUN: %clang_cc1 -emit-llvm -cl-denorms-are-zero -o - -triple amdgcn--amdhsa -target-cpu tahiti %s | FileCheck -check-prefixes=AMDGCN,AMDGCN-FLUSH-OPT %s
// Fast F32 denormals, but slow FMAF
// RUN: %clang_cc1 -emit-llvm -o - -triple amdgcn--amdhsa -target-cpu fiji %s | FileCheck -check-prefixes=AMDGCN,AMDGCN-FLUSH %s
// RUN: %clang_cc1 -emit-llvm -cl-denorms-are-zero -o - -triple amdgcn--amdhsa -target-cpu fiji %s | FileCheck -check-prefixes=AMDGCN,AMDGCN-FLUSH-OPT %s
// Fast F32 denormals and fast FMAF
// RUN: %clang_cc1 -emit-llvm -o - -triple amdgcn--amdhsa -target-cpu gfx900 %s | FileCheck -check-prefixes=AMDGCN,AMDGCN-DENORM %s
// RUN: %clang_cc1 -emit-llvm -cl-denorms-are-zero -o - -triple amdgcn--amdhsa -target-cpu gfx900 %s | FileCheck -check-prefixes=AMDGCN,AMDGCN-FLUSH-OPT %s
// RUN: %clang_cc1 -emit-llvm -target-feature +fp32-denormals -target-feature -fp64-fp16-denormals -cl-denorms-are-zero -o - -triple amdgcn--amdhsa -target-cpu fiji %s | FileCheck -check-prefixes=AMDGCN,AMDGCN-FEATURE %s
// RUN: %clang_cc1 -emit-llvm -target-feature +fp32-denormals -target-feature -fp64-fp16-denormals -cl-denorms-are-zero -o - -triple amdgcn--amdhsa -target-cpu pitcairn %s | FileCheck -check-prefixes=AMDGCN,AMDGCN-FEATURE %s
// For all targets 'denorms-are-zero' attribute is set to 'true' // For all targets 'denorms-are-zero' attribute is set to 'true'
// if '-cl-denorms-are-zero' was specified and to 'false' otherwise. // if '-cl-denorms-are-zero' was specified and to 'false' otherwise.
// CHECK-LABEL: define {{(dso_local )?}}void @f() // CHECK-LABEL: define {{(dso_local )?}}void @f()
// CHECK: attributes #{{[0-9]*}} = {{{[^}]*}} "denorms-are-zero"="false" // CHECK: attributes #{{[0-9]*}} = {{{[^}]*}} "denorms-are-zero"="false"
// //
// DENORM-ZERO-LABEL: define {{(dso_local )?}}void @f() // DENORM-ZERO-LABEL: define {{(dso_local )?}}void @f()
// DENORM-ZERO: attributes #{{[0-9]*}} = {{{[^}]*}} "denorms-are-zero"="true" // DENORM-ZERO: attributes #{{[0-9]*}} = {{{[^}]*}} "denorms-are-zero"="true"
// For amdgcn target cpu fiji, fp32 should be flushed since fiji does not support fp32 denormals, unless +fp32-denormals is // For amdgcn target cpu fiji, fp32 should be flushed since fiji does not support fp32 denormals, unless +fp32-denormals is
// explicitly set. amdgcn target always do not flush fp64 denormals. The control for fp64 and fp16 denormals is the same. // explicitly set. amdgcn target always do not flush fp64 denormals. The control for fp64 and fp16 denormals is the same.
// AMDGCN-LABEL: define void @f() // AMDGCN-LABEL: define void @f()
// AMDGCN: attributes #{{[0-9]*}} = {{{[^}]*}} "denorms-are-zero"="true" {{.*}} "target-features"="{{[^"]*}}+fp64-fp16-denormals,{{[^"]*}}-fp32-denormals{{[^"]*}}"
// AMDGCN-DENORM-LABEL: define void @f() // AMDGCN-FLUSH: attributes #{{[0-9]*}} = {{{[^}]*}} "denorms-are-zero"="false" {{.*}} "target-features"="{{[^"]*}}+fp64-fp16-denormals,{{[^"]*}}-fp32-denormals{{[^"]*}}"
// AMDGCN-DENORM: attributes #{{[0-9]*}} = {{{[^}]*}} "denorms-are-zero"="false" {{.*}} "target-features"="{{[^"]*}}+fp64-fp16-denormals,{{[^"]*}}-fp32-denormals{{[^"]*}}" // AMDGCN-FLUSH-OPT: attributes #{{[0-9]*}} = {{{[^}]*}} "denorms-are-zero"="true" {{.*}} "target-features"="{{[^"]*}}+fp64-fp16-denormals,{{[^"]*}}-fp32-denormals{{[^"]*}}"
// AMDGCN-FEATURE-LABEL: define void @f()
// AMDGCN-DENORM: attributes #{{[0-9]*}} = {{{[^}]*}} "denorms-are-zero"="false" {{.*}} "target-features"="{{[^"]*}}+fp32-denormals,{{[^"]*}}+fp64-fp16-denormals{{[^"]*}}"
// AMDGCN-FEATURE: attributes #{{[0-9]*}} = {{{[^}]*}} "denorms-are-zero"="true" {{.*}} "target-features"="{{[^"]*}}+fp32-denormals,{{[^"]*}}-fp64-fp16-denormals{{[^"]*}}" // AMDGCN-FEATURE: attributes #{{[0-9]*}} = {{{[^}]*}} "denorms-are-zero"="true" {{.*}} "target-features"="{{[^"]*}}+fp32-denormals,{{[^"]*}}-fp64-fp16-denormals{{[^"]*}}"
void f() {} void f() {}