Skip to content

[AMDGPU] Fix vector legalization for bf16 valu ops #158439

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

Merged
merged 5 commits into from
Sep 25, 2025
Merged

[AMDGPU] Fix vector legalization for bf16 valu ops #158439

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
8ca54e6
[AMDGPU] Fix vector legalization for bf16 valu ops
giuseros Sep 12, 2025
3f99c0c
Add testing coverage - part I
giuseros Sep 19, 2025
32d932f
Fix CostModel for bf16
giuseros Sep 23, 2025
0b36339
Latest fixes
giuseros Sep 24, 2025
dff4a35
Merge branch 'main' into add_pk_valu
giuseros Sep 25, 2025
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
8 changes: 6 additions & 2 deletions llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -607,13 +607,15 @@ InstructionCost GCNTTIImpl::getArithmeticInstrCost(
case ISD::FSUB:
if (ST->hasPackedFP32Ops() && SLT == MVT::f32)
NElts = (NElts + 1) / 2;
if (ST->hasBF16PackedInsts() && SLT == MVT::bf16)
NElts = (NElts + 1) / 2;
if (SLT == MVT::f64)
return LT.first * NElts * get64BitInstrCost(CostKind);

if (ST->has16BitInsts() && SLT == MVT::f16)
NElts = (NElts + 1) / 2;

if (SLT == MVT::f32 || SLT == MVT::f16)
if (SLT == MVT::f32 || SLT == MVT::f16 || SLT == MVT::bf16)
return LT.first * NElts * getFullRateInstrCost();
break;
case ISD::FDIV:
Expand Down Expand Up @@ -746,7 +748,9 @@ GCNTTIImpl::getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA,

MVT::SimpleValueType SLT = LT.second.getScalarType().SimpleTy;

if ((ST->hasVOP3PInsts() && (SLT == MVT::f16 || SLT == MVT::i16)) ||
if ((ST->hasVOP3PInsts() &&
(SLT == MVT::f16 || SLT == MVT::i16 ||
(SLT == MVT::bf16 && ST->hasBF16PackedInsts()))) ||
(ST->hasPackedFP32Ops() && SLT == MVT::f32))
NElts = (NElts + 1) / 2;

Expand Down
17 changes: 13 additions & 4 deletions llvm/lib/Target/AMDGPU/SIISelLowering.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -851,6 +851,13 @@ SITargetLowering::SITargetLowering(const TargetMachine &TM,
setOperationAction(ISD::SELECT, {MVT::v4i16, MVT::v4f16, MVT::v4bf16},
Custom);

if (Subtarget->hasBF16PackedInsts()) {
for (MVT VT : {MVT::v4bf16, MVT::v8bf16, MVT::v16bf16, MVT::v32bf16})
// Split vector operations.
setOperationAction({ISD::FADD, ISD::FMUL, ISD::FMA, ISD::FCANONICALIZE},
VT, Custom);
}

if (Subtarget->hasPackedFP32Ops()) {
setOperationAction({ISD::FADD, ISD::FMUL, ISD::FMA, ISD::FNEG},
MVT::v2f32, Legal);
Expand Down Expand Up @@ -6621,10 +6628,12 @@ SDValue SITargetLowering::splitUnaryVectorOp(SDValue Op,
SelectionDAG &DAG) const {
unsigned Opc = Op.getOpcode();
EVT VT = Op.getValueType();
assert(VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4f32 ||
VT == MVT::v8i16 || VT == MVT::v8f16 || VT == MVT::v16i16 ||
VT == MVT::v16f16 || VT == MVT::v8f32 || VT == MVT::v16f32 ||
VT == MVT::v32f32 || VT == MVT::v32i16 || VT == MVT::v32f16);
assert(VT == MVT::v4i16 || VT == MVT::v4f16 || VT == MVT::v4bf16 ||
VT == MVT::v4f32 || VT == MVT::v8i16 || VT == MVT::v8f16 ||
VT == MVT::v8bf16 || VT == MVT::v16i16 || VT == MVT::v16f16 ||
VT == MVT::v16bf16 || VT == MVT::v8f32 || VT == MVT::v16f32 ||
VT == MVT::v32f32 || VT == MVT::v32i16 || VT == MVT::v32f16 ||
VT == MVT::v32bf16);

auto [Lo, Hi] = DAG.SplitVectorOperand(Op.getNode(), 0);

Expand Down
68 changes: 62 additions & 6 deletions llvm/test/Analysis/CostModel/AMDGPU/canonicalize.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3,11 +3,13 @@
; RUN: opt < %s -mtriple=amdgcn-unknown-amdhsa -mcpu=gfx803 -passes="print<cost-model>" -cost-kind=throughput 2>&1 -disable-output | FileCheck -check-prefixes=ALL,GFX8 %s
; RUN: opt < %s -mtriple=amdgcn-unknown-amdhsa -mcpu=gfx900 -passes="print<cost-model>" -cost-kind=throughput 2>&1 -disable-output | FileCheck -check-prefixes=ALL,GFX9 %s
; RUN: opt < %s -mtriple=amdgcn-unknown-amdhsa -mcpu=gfx1030 -passes="print<cost-model>" -cost-kind=throughput 2>&1 -disable-output | FileCheck -check-prefixes=ALL,GFX10 %s
; RUN: opt < %s -mtriple=amdgcn-unknown-amdhsa -mcpu=gfx1250 -passes="print<cost-model>" -cost-kind=throughput 2>&1 -disable-output | FileCheck -check-prefixes=GFX1250 %s

; RUN: opt < %s -mtriple=amdgcn-unknown-amdhsa -passes="print<cost-model>" -cost-kind=code-size 2>&1 -disable-output | FileCheck -check-prefixes=ALL-SIZE,BASE-SIZE %s
; RUN: opt < %s -mtriple=amdgcn-unknown-amdhsa -mcpu=gfx803 -passes="print<cost-model>" -cost-kind=code-size 2>&1 -disable-output | FileCheck -check-prefixes=ALL-SIZE,GFX8-SIZE %s
; RUN: opt < %s -mtriple=amdgcn-unknown-amdhsa -mcpu=gfx900 -passes="print<cost-model>" -cost-kind=code-size 2>&1 -disable-output | FileCheck -check-prefixes=ALL-SIZE,GFX9-SIZE %s
; RUN: opt < %s -mtriple=amdgcn-unknown-amdhsa -mcpu=gfx1030 -passes="print<cost-model>" -cost-kind=code-size 2>&1 -disable-output | FileCheck -check-prefixes=ALL-SIZE,GFX10-SIZE %s
; RUN: opt < %s -mtriple=amdgcn-unknown-amdhsa -mcpu=gfx1250 -passes="print<cost-model>" -cost-kind=code-size 2>&1 -disable-output | FileCheck -check-prefixes=GFX1250-SIZE %s

define void @canonicalize_f16() {
; BASE-LABEL: 'canonicalize_f16'
Expand Down Expand Up @@ -141,6 +143,16 @@ define void @canonicalize_bf16() {
; GFX10-NEXT: Cost Model: Found an estimated cost of 96 for instruction: %v17bf16 = call <17 x bfloat> @llvm.canonicalize.v17bf16(<17 x bfloat> undef)
; GFX10-NEXT: Cost Model: Found an estimated cost of 10 for instruction: ret void
;
; GFX1250-LABEL: 'canonicalize_bf16'
; GFX1250-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %bf16 = call bfloat @llvm.canonicalize.bf16(bfloat undef)
; GFX1250-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %v2bf16 = call <2 x bfloat> @llvm.canonicalize.v2bf16(<2 x bfloat> undef)
; GFX1250-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %v3bf16 = call <3 x bfloat> @llvm.canonicalize.v3bf16(<3 x bfloat> undef)
; GFX1250-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %v4bf16 = call <4 x bfloat> @llvm.canonicalize.v4bf16(<4 x bfloat> undef)
; GFX1250-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %v5bf16 = call <5 x bfloat> @llvm.canonicalize.v5bf16(<5 x bfloat> undef)
; GFX1250-NEXT: Cost Model: Found an estimated cost of 8 for instruction: %v16bf16 = call <16 x bfloat> @llvm.canonicalize.v16bf16(<16 x bfloat> undef)
; GFX1250-NEXT: Cost Model: Found an estimated cost of 48 for instruction: %v17bf16 = call <17 x bfloat> @llvm.canonicalize.v17bf16(<17 x bfloat> undef)
; GFX1250-NEXT: Cost Model: Found an estimated cost of 10 for instruction: ret void
;
; BASE-SIZE-LABEL: 'canonicalize_bf16'
; BASE-SIZE-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %bf16 = call bfloat @llvm.canonicalize.bf16(bfloat undef)
; BASE-SIZE-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %v2bf16 = call <2 x bfloat> @llvm.canonicalize.v2bf16(<2 x bfloat> undef)
Expand Down Expand Up @@ -181,6 +193,15 @@ define void @canonicalize_bf16() {
; GFX10-SIZE-NEXT: Cost Model: Found an estimated cost of 96 for instruction: %v17bf16 = call <17 x bfloat> @llvm.canonicalize.v17bf16(<17 x bfloat> undef)
; GFX10-SIZE-NEXT: Cost Model: Found an estimated cost of 1 for instruction: ret void
;
; GFX1250-SIZE-LABEL: 'canonicalize_bf16'
; GFX1250-SIZE-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %bf16 = call bfloat @llvm.canonicalize.bf16(bfloat undef)
; GFX1250-SIZE-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %v2bf16 = call <2 x bfloat> @llvm.canonicalize.v2bf16(<2 x bfloat> undef)
; GFX1250-SIZE-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %v3bf16 = call <3 x bfloat> @llvm.canonicalize.v3bf16(<3 x bfloat> undef)
; GFX1250-SIZE-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %v4bf16 = call <4 x bfloat> @llvm.canonicalize.v4bf16(<4 x bfloat> undef)
; GFX1250-SIZE-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %v5bf16 = call <5 x bfloat> @llvm.canonicalize.v5bf16(<5 x bfloat> undef)
; GFX1250-SIZE-NEXT: Cost Model: Found an estimated cost of 8 for instruction: %v16bf16 = call <16 x bfloat> @llvm.canonicalize.v16bf16(<16 x bfloat> undef)
; GFX1250-SIZE-NEXT: Cost Model: Found an estimated cost of 48 for instruction: %v17bf16 = call <17 x bfloat> @llvm.canonicalize.v17bf16(<17 x bfloat> undef)
; GFX1250-SIZE-NEXT: Cost Model: Found an estimated cost of 1 for instruction: ret void
%bf16 = call bfloat @llvm.canonicalize.bf16(bfloat undef) #1
%v2bf16 = call <2 x bfloat> @llvm.canonicalize.v2bf16(<2 x bfloat> undef) #1
%v3bf16 = call <3 x bfloat> @llvm.canonicalize.v3bf16(<3 x bfloat> undef) #1
Expand All @@ -203,6 +224,17 @@ define void @canonicalize_f32() {
; ALL-NEXT: Cost Model: Found an estimated cost of 48 for instruction: %v16f32 = call <16 x float> @llvm.canonicalize.v16f32(<16 x float> undef)
; ALL-NEXT: Cost Model: Found an estimated cost of 10 for instruction: ret void
;
; GFX1250-LABEL: 'canonicalize_f32'
; GFX1250-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %f32 = call float @llvm.canonicalize.f32(float undef)
; GFX1250-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %v2f32 = call <2 x float> @llvm.canonicalize.v2f32(<2 x float> undef)
; GFX1250-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %v3f32 = call <3 x float> @llvm.canonicalize.v3f32(<3 x float> undef)
; GFX1250-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %v4f32 = call <4 x float> @llvm.canonicalize.v4f32(<4 x float> undef)
; GFX1250-NEXT: Cost Model: Found an estimated cost of 3 for instruction: %v5f32 = call <5 x float> @llvm.canonicalize.v5f32(<5 x float> undef)
; GFX1250-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %v8f32 = call <8 x float> @llvm.canonicalize.v8f32(<8 x float> undef)
; GFX1250-NEXT: Cost Model: Found an estimated cost of 15 for instruction: %v9f32 = call <9 x float> @llvm.canonicalize.v9f32(<9 x float> undef)
; GFX1250-NEXT: Cost Model: Found an estimated cost of 24 for instruction: %v16f32 = call <16 x float> @llvm.canonicalize.v16f32(<16 x float> undef)
; GFX1250-NEXT: Cost Model: Found an estimated cost of 10 for instruction: ret void
;
; ALL-SIZE-LABEL: 'canonicalize_f32'
; ALL-SIZE-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %f32 = call float @llvm.canonicalize.f32(float undef)
; ALL-SIZE-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %v2f32 = call <2 x float> @llvm.canonicalize.v2f32(<2 x float> undef)
Expand All @@ -214,6 +246,16 @@ define void @canonicalize_f32() {
; ALL-SIZE-NEXT: Cost Model: Found an estimated cost of 48 for instruction: %v16f32 = call <16 x float> @llvm.canonicalize.v16f32(<16 x float> undef)
; ALL-SIZE-NEXT: Cost Model: Found an estimated cost of 1 for instruction: ret void
;
; GFX1250-SIZE-LABEL: 'canonicalize_f32':
; GFX1250-SIZE-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %f32 = call float @llvm.canonicalize.f32(float undef)
; GFX1250-SIZE-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %v2f32 = call <2 x float> @llvm.canonicalize.v2f32(<2 x float> undef)
; GFX1250-SIZE-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %v3f32 = call <3 x float> @llvm.canonicalize.v3f32(<3 x float> undef)
; GFX1250-SIZE-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %v4f32 = call <4 x float> @llvm.canonicalize.v4f32(<4 x float> undef)
; GFX1250-SIZE-NEXT: Cost Model: Found an estimated cost of 3 for instruction: %v5f32 = call <5 x float> @llvm.canonicalize.v5f32(<5 x float> undef)
; GFX1250-SIZE-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %v8f32 = call <8 x float> @llvm.canonicalize.v8f32(<8 x float> undef)
; GFX1250-SIZE-NEXT: Cost Model: Found an estimated cost of 15 for instruction: %v9f32 = call <9 x float> @llvm.canonicalize.v9f32(<9 x float> undef)
; GFX1250-SIZE-NEXT: Cost Model: Found an estimated cost of 24 for instruction: %v16f32 = call <16 x float> @llvm.canonicalize.v16f32(<16 x float> undef)
; GFX1250-SIZE-NEXT: Cost Model: Found an estimated cost of 1 for instruction: ret void
%f32 = call float @llvm.canonicalize.f32(float undef) #1
%v2f32 = call <2 x float> @llvm.canonicalize.v2f32(<2 x float> undef) #1
%v3f32 = call <3 x float> @llvm.canonicalize.v3f32(<3 x float> undef) #1
Expand All @@ -236,6 +278,16 @@ define void @canonicalize_f64() {
; ALL-NEXT: Cost Model: Found an estimated cost of 320 for instruction: %v16f64 = call <16 x double> @llvm.canonicalize.v16f64(<16 x double> undef)
; ALL-NEXT: Cost Model: Found an estimated cost of 10 for instruction: ret void
;
; GFX1250-LABEL: 'canonicalize_f64'
; GFX1250-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %f64 = call double @llvm.canonicalize.f64(double undef)
; GFX1250-NEXT: Cost Model: Found an estimated cost of 8 for instruction: %v2f64 = call <2 x double> @llvm.canonicalize.v2f64(<2 x double> undef)
; GFX1250-NEXT: Cost Model: Found an estimated cost of 12 for instruction: %v3f64 = call <3 x double> @llvm.canonicalize.v3f64(<3 x double> undef)
; GFX1250-NEXT: Cost Model: Found an estimated cost of 16 for instruction: %v4f64 = call <4 x double> @llvm.canonicalize.v4f64(<4 x double> undef)
; GFX1250-NEXT: Cost Model: Found an estimated cost of 96 for instruction: %v5f64 = call <5 x double> @llvm.canonicalize.v5f64(<5 x double> undef)
; GFX1250-NEXT: Cost Model: Found an estimated cost of 96 for instruction: %v8f64 = call <8 x double> @llvm.canonicalize.v8f64(<8 x double> undef)
; GFX1250-NEXT: Cost Model: Found an estimated cost of 320 for instruction: %v16f64 = call <16 x double> @llvm.canonicalize.v16f64(<16 x double> undef)
; GFX1250-NEXT: Cost Model: Found an estimated cost of 10 for instruction: ret void
;
; ALL-SIZE-LABEL: 'canonicalize_f64'
; ALL-SIZE-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %f64 = call double @llvm.canonicalize.f64(double undef)
; ALL-SIZE-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %v2f64 = call <2 x double> @llvm.canonicalize.v2f64(<2 x double> undef)
Expand All @@ -245,6 +297,16 @@ define void @canonicalize_f64() {
; ALL-SIZE-NEXT: Cost Model: Found an estimated cost of 48 for instruction: %v8f64 = call <8 x double> @llvm.canonicalize.v8f64(<8 x double> undef)
; ALL-SIZE-NEXT: Cost Model: Found an estimated cost of 160 for instruction: %v16f64 = call <16 x double> @llvm.canonicalize.v16f64(<16 x double> undef)
; ALL-SIZE-NEXT: Cost Model: Found an estimated cost of 1 for instruction: ret void
;
; GFX1250-SIZE-LABEL: 'canonicalize_f64'
; GFX1250-SIZE-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %f64 = call double @llvm.canonicalize.f64(double undef)
; GFX1250-SIZE-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %v2f64 = call <2 x double> @llvm.canonicalize.v2f64(<2 x double> undef)
; GFX1250-SIZE-NEXT: Cost Model: Found an estimated cost of 6 for instruction: %v3f64 = call <3 x double> @llvm.canonicalize.v3f64(<3 x double> undef)
; GFX1250-SIZE-NEXT: Cost Model: Found an estimated cost of 8 for instruction: %v4f64 = call <4 x double> @llvm.canonicalize.v4f64(<4 x double> undef)
; GFX1250-SIZE-NEXT: Cost Model: Found an estimated cost of 48 for instruction: %v5f64 = call <5 x double> @llvm.canonicalize.v5f64(<5 x double> undef)
; GFX1250-SIZE-NEXT: Cost Model: Found an estimated cost of 48 for instruction: %v8f64 = call <8 x double> @llvm.canonicalize.v8f64(<8 x double> undef)
; GFX1250-SIZE-NEXT: Cost Model: Found an estimated cost of 160 for instruction: %v16f64 = call <16 x double> @llvm.canonicalize.v16f64(<16 x double> undef)
; GFX1250-SIZE-NEXT: Cost Model: Found an estimated cost of 1 for instruction: ret void
;
%f64 = call double @llvm.canonicalize.f64(double undef) #1
%v2f64 = call <2 x double> @llvm.canonicalize.v2f64(<2 x double> undef) #1
Expand All @@ -255,9 +317,3 @@ define void @canonicalize_f64() {
%v16f64 = call <16 x double> @llvm.canonicalize.v16f64(<16 x double> undef) #1
ret void
}






Loading
Loading