shylie/llvm-project
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

AMDGPU: Custom lower fptrunc vectors for f32 -> f16 (#141883)

Browse Source 
The latest asics support v_cvt_pk_f16_f32 instruction. However current
implementation of vector fptrunc lowering fully scalarizes the vectors,
and the scalar conversions may not always be combined to generate the
packed one.
We made v2f32 -> v2f16 legal in
https://github.com/llvm/llvm-project/pull/139956. This work is an
extension to handle wider vectors. Instead of fully scalarization, we
split the vector to packs (v2f32 -> v2f16) to ensure the packed
conversion can always been generated.
This commit is contained in:
Changpeng Fang 2025-06-06 15:15:24 -07:00 committed by GitHub
parent 7b2aa02a33
commit 70e78be7dc
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
4 changed files with 295 additions and 14 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
View File

@ -1061,10 +1061,12 @@ AMDGPULegalizerInfo::AMDGPULegalizerInfo(const GCNSubtarget &ST_,
}
auto &FPTruncActions = getActionDefinitionsBuilder(G_FPTRUNC);
if (ST.hasCvtPkF16F32Inst())
FPTruncActions.legalFor({{S32, S64}, {S16, S32}, {V2S16, V2S32}});
else
if (ST.hasCvtPkF16F32Inst()) {
FPTruncActions.legalFor({{S32, S64}, {S16, S32}, {V2S16, V2S32}})
.clampMaxNumElements(0, S16, 2);
} else {
FPTruncActions.legalFor({{S32, S64}, {S16, S32}});
}
FPTruncActions.scalarize(0).lower();
getActionDefinitionsBuilder(G_FPEXT)

32
llvm/lib/Target/AMDGPU/SIISelLowering.cpp
View File

@ -926,8 +926,11 @@ SITargetLowering::SITargetLowering(const TargetMachine &TM,
setOperationAction(ISD::BUILD_VECTOR, MVT::v2bf16, Legal);
}
if (Subtarget->hasCvtPkF16F32Inst())
setOperationAction(ISD::FP_ROUND, MVT::v2f16, Custom);
if (Subtarget->hasCvtPkF16F32Inst()) {
setOperationAction(ISD::FP_ROUND,
{MVT::v2f16, MVT::v4f16, MVT::v8f16, MVT::v16f16},
Custom);
}
setTargetDAGCombine({ISD::ADD,
ISD::UADDO_CARRY,
@ -6896,14 +6899,35 @@ SDValue SITargetLowering::getFPExtOrFPRound(SelectionDAG &DAG, SDValue Op,
DAG.getTargetConstant(0, DL, MVT::i32));
}
SDValue SITargetLowering::splitFP_ROUNDVectorOp(SDValue Op,
SelectionDAG &DAG) const {
EVT DstVT = Op.getValueType();
unsigned NumElts = DstVT.getVectorNumElements();
assert(NumElts > 2 && isPowerOf2_32(NumElts));
auto [Lo, Hi] = DAG.SplitVectorOperand(Op.getNode(), 0);
SDLoc DL(Op);
unsigned Opc = Op.getOpcode();
SDValue Flags = Op.getOperand(1);
EVT HalfDstVT =
EVT::getVectorVT(*DAG.getContext(), DstVT.getScalarType(), NumElts / 2);
SDValue OpLo = DAG.getNode(Opc, DL, HalfDstVT, Lo, Flags);
SDValue OpHi = DAG.getNode(Opc, DL, HalfDstVT, Hi, Flags);
return DAG.getNode(ISD::CONCAT_VECTORS, DL, DstVT, OpLo, OpHi);
}
SDValue SITargetLowering::lowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const {
SDValue Src = Op.getOperand(0);
EVT SrcVT = Src.getValueType();
EVT DstVT = Op.getValueType();
if (DstVT == MVT::v2f16) {
if (DstVT.isVector() && DstVT.getScalarType() == MVT::f16) {
assert(Subtarget->hasCvtPkF16F32Inst() && "support v_cvt_pk_f16_f32");
return SrcVT == MVT::v2f32 ? Op : SDValue();
if (SrcVT.getScalarType() != MVT::f32)
return SDValue();
return SrcVT == MVT::v2f32 ? Op : splitFP_ROUNDVectorOp(Op, DAG);
}
if (SrcVT.getScalarType() != MVT::f64)

1
llvm/lib/Target/AMDGPU/SIISelLowering.h
View File

@ -145,6 +145,7 @@ private:
/// Custom lowering for ISD::FP_ROUND for MVT::f16.
SDValue lowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const;
SDValue splitFP_ROUNDVectorOp(SDValue Op, SelectionDAG &DAG) const;
SDValue lowerFMINNUM_FMAXNUM(SDValue Op, SelectionDAG &DAG) const;
SDValue lowerFMINIMUM_FMAXIMUM(SDValue Op, SelectionDAG &DAG) const;
SDValue lowerFLDEXP(SDValue Op, SelectionDAG &DAG) const;

268
llvm/test/CodeGen/AMDGPU/fptrunc.v2f16.no.fast.math.ll
View File

@ -12,18 +12,272 @@ define <2 x half> @v_test_cvt_v2f32_v2f16(<2 x float> %src) {
ret <2 x half> %res
}
define half @fptrunc_v2f32_v2f16_then_extract(<2 x float> %src) {
; GFX950-LABEL: fptrunc_v2f32_v2f16_then_extract:
define <3 x half> @v_test_cvt_v3f32_v3f16(<3 x float> %src) {
; GFX950-LABEL: v_test_cvt_v3f32_v3f16:
; GFX950: ; %bb.0:
; GFX950-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX950-NEXT: v_cvt_f16_f32_e32 v2, v2
; GFX950-NEXT: v_cvt_pk_f16_f32 v0, v0, v1
; GFX950-NEXT: v_mov_b32_e32 v1, v2
; GFX950-NEXT: s_setpc_b64 s[30:31]
%res = fptrunc <3 x float> %src to <3 x half>
ret <3 x half> %res
}
define <4 x half> @v_test_cvt_v4f32_v4f16(<4 x float> %src) {
; GFX950-LABEL: v_test_cvt_v4f32_v4f16:
; GFX950: ; %bb.0:
; GFX950-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX950-NEXT: v_cvt_pk_f16_f32 v0, v0, v1
; GFX950-NEXT: v_add_f16_sdwa v0, v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; GFX950-NEXT: v_cvt_pk_f16_f32 v1, v2, v3
; GFX950-NEXT: s_setpc_b64 s[30:31]
%res = fptrunc <4 x float> %src to <4 x half>
ret <4 x half> %res
}
define <8 x half> @v_test_cvt_v8f32_v2f16(<8 x float> %src) {
; GFX950-LABEL: v_test_cvt_v8f32_v2f16:
; GFX950: ; %bb.0:
; GFX950-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX950-NEXT: v_cvt_pk_f16_f32 v0, v0, v1
; GFX950-NEXT: v_cvt_pk_f16_f32 v1, v2, v3
; GFX950-NEXT: v_cvt_pk_f16_f32 v2, v4, v5
; GFX950-NEXT: v_cvt_pk_f16_f32 v3, v6, v7
; GFX950-NEXT: s_setpc_b64 s[30:31]
%res = fptrunc <8 x float> %src to <8 x half>
ret <8 x half> %res
}
define <16 x half> @v_test_cvt_v16f32_v16f16(<16 x float> %src) {
; GFX950-LABEL: v_test_cvt_v16f32_v16f16:
; GFX950: ; %bb.0:
; GFX950-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX950-NEXT: v_cvt_pk_f16_f32 v0, v0, v1
; GFX950-NEXT: v_cvt_pk_f16_f32 v1, v2, v3
; GFX950-NEXT: v_cvt_pk_f16_f32 v2, v4, v5
; GFX950-NEXT: v_cvt_pk_f16_f32 v3, v6, v7
; GFX950-NEXT: v_cvt_pk_f16_f32 v4, v8, v9
; GFX950-NEXT: v_cvt_pk_f16_f32 v5, v10, v11
; GFX950-NEXT: v_cvt_pk_f16_f32 v6, v12, v13
; GFX950-NEXT: v_cvt_pk_f16_f32 v7, v14, v15
; GFX950-NEXT: s_setpc_b64 s[30:31]
%res = fptrunc <16 x float> %src to <16 x half>
ret <16 x half> %res
}
define half @fptrunc_v2f32_v2f16_extract_uses(<2 x float> %src) {
; GFX950-LABEL: fptrunc_v2f32_v2f16_extract_uses:
; GFX950: ; %bb.0:
; GFX950-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX950-NEXT: v_cvt_pk_f16_f32 v0, v0, v1
; GFX950-NEXT: v_add_f16_sdwa v0, v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX950-NEXT: s_setpc_b64 s[30:31]
%vec_half = fptrunc <2 x float> %src to <2 x half>
%first = extractelement <2 x half> %vec_half, i64 1
%second = extractelement <2 x half> %vec_half, i64 0
%res = fadd half %first, %second
ret half %res
%f0 = extractelement <2 x half> %vec_half, i64 0
%f1 = extractelement <2 x half> %vec_half, i64 1
%rslt = fadd half %f0, %f1
ret half %rslt
}
define half @fptrunc_v3f32_v3f16_extract_uses(<3 x float> %vec_float) {
; GFX950-SDAG-LABEL: fptrunc_v3f32_v3f16_extract_uses:
; GFX950-SDAG: ; %bb.0:
; GFX950-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX950-SDAG-NEXT: v_cvt_f16_f32_e32 v1, v1
; GFX950-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0
; GFX950-SDAG-NEXT: v_cvt_f16_f32_e32 v2, v2
; GFX950-SDAG-NEXT: v_add_f16_e32 v0, v0, v1
; GFX950-SDAG-NEXT: v_add_f16_e32 v0, v2, v0
; GFX950-SDAG-NEXT: s_setpc_b64 s[30:31]
;
; GFX950-GISEL-LABEL: fptrunc_v3f32_v3f16_extract_uses:
; GFX950-GISEL: ; %bb.0:
; GFX950-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX950-GISEL-NEXT: v_cvt_f16_f32_e32 v2, v2
; GFX950-GISEL-NEXT: v_cvt_pk_f16_f32 v0, v0, v1
; GFX950-GISEL-NEXT: v_add_f16_sdwa v0, v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX950-GISEL-NEXT: v_add_f16_e32 v0, v2, v0
; GFX950-GISEL-NEXT: s_setpc_b64 s[30:31]
%vec_half = fptrunc <3 x float> %vec_float to <3 x half>
%f0 = extractelement <3 x half> %vec_half, i64 0
%f1 = extractelement <3 x half> %vec_half, i64 1
%f2 = extractelement <3 x half> %vec_half, i64 2
%sum0 = fadd half %f0, %f1
%rslt = fadd half %f2, %sum0
ret half %rslt
}
define half @fptrunc_v4f32_v4f16_extract_uses(<4 x float> %vec_float) {
; GFX950-SDAG-LABEL: fptrunc_v4f32_v4f16_extract_uses:
; GFX950-SDAG: ; %bb.0:
; GFX950-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX950-SDAG-NEXT: v_cvt_pk_f16_f32 v2, v2, v3
; GFX950-SDAG-NEXT: v_cvt_pk_f16_f32 v0, v0, v1
; GFX950-SDAG-NEXT: v_add_f16_sdwa v0, v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX950-SDAG-NEXT: v_add_f16_sdwa v1, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX950-SDAG-NEXT: v_add_f16_e32 v0, v0, v1
; GFX950-SDAG-NEXT: s_setpc_b64 s[30:31]
;
; GFX950-GISEL-LABEL: fptrunc_v4f32_v4f16_extract_uses:
; GFX950-GISEL: ; %bb.0:
; GFX950-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX950-GISEL-NEXT: v_cvt_pk_f16_f32 v0, v0, v1
; GFX950-GISEL-NEXT: v_cvt_pk_f16_f32 v1, v2, v3
; GFX950-GISEL-NEXT: v_add_f16_sdwa v0, v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX950-GISEL-NEXT: v_add_f16_sdwa v1, v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX950-GISEL-NEXT: v_add_f16_e32 v0, v0, v1
; GFX950-GISEL-NEXT: s_setpc_b64 s[30:31]
%vec_half = fptrunc <4 x float> %vec_float to <4 x half>
%f0 = extractelement <4 x half> %vec_half, i64 0
%f1 = extractelement <4 x half> %vec_half, i64 1
%f2 = extractelement <4 x half> %vec_half, i64 2
%f3 = extractelement <4 x half> %vec_half, i64 3
%sum0 = fadd half %f0, %f1
%sum1 = fadd half %f2, %f3
%rslt = fadd half %sum0, %sum1
ret half %rslt
}
define half @fptrunc_v8f32_v8f16_extract_uses(<8 x float> %vec_float) {
; GFX950-SDAG-LABEL: fptrunc_v8f32_v8f16_extract_uses:
; GFX950-SDAG: ; %bb.0:
; GFX950-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX950-SDAG-NEXT: v_cvt_pk_f16_f32 v6, v6, v7
; GFX950-SDAG-NEXT: v_cvt_pk_f16_f32 v4, v4, v5
; GFX950-SDAG-NEXT: v_cvt_pk_f16_f32 v2, v2, v3
; GFX950-SDAG-NEXT: v_cvt_pk_f16_f32 v0, v0, v1
; GFX950-SDAG-NEXT: v_add_f16_sdwa v0, v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX950-SDAG-NEXT: v_add_f16_sdwa v1, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX950-SDAG-NEXT: v_add_f16_sdwa v2, v4, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX950-SDAG-NEXT: v_add_f16_sdwa v3, v6, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX950-SDAG-NEXT: v_add_f16_e32 v0, v0, v1
; GFX950-SDAG-NEXT: v_add_f16_e32 v1, v2, v3
; GFX950-SDAG-NEXT: v_add_f16_e32 v0, v0, v1
; GFX950-SDAG-NEXT: s_setpc_b64 s[30:31]
;
; GFX950-GISEL-LABEL: fptrunc_v8f32_v8f16_extract_uses:
; GFX950-GISEL: ; %bb.0:
; GFX950-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX950-GISEL-NEXT: v_cvt_pk_f16_f32 v0, v0, v1
; GFX950-GISEL-NEXT: v_cvt_pk_f16_f32 v1, v2, v3
; GFX950-GISEL-NEXT: v_cvt_pk_f16_f32 v2, v4, v5
; GFX950-GISEL-NEXT: v_cvt_pk_f16_f32 v3, v6, v7
; GFX950-GISEL-NEXT: v_add_f16_sdwa v0, v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX950-GISEL-NEXT: v_add_f16_sdwa v1, v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX950-GISEL-NEXT: v_add_f16_sdwa v2, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX950-GISEL-NEXT: v_add_f16_sdwa v3, v3, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX950-GISEL-NEXT: v_add_f16_e32 v0, v0, v1
; GFX950-GISEL-NEXT: v_add_f16_e32 v1, v2, v3
; GFX950-GISEL-NEXT: v_add_f16_e32 v0, v0, v1
; GFX950-GISEL-NEXT: s_setpc_b64 s[30:31]
%vec_half = fptrunc <8 x float> %vec_float to <8 x half>
%f0 = extractelement <8 x half> %vec_half, i64 0
%f1 = extractelement <8 x half> %vec_half, i64 1
%f2 = extractelement <8 x half> %vec_half, i64 2
%f3 = extractelement <8 x half> %vec_half, i64 3
%f4 = extractelement <8 x half> %vec_half, i64 4
%f5 = extractelement <8 x half> %vec_half, i64 5
%f6 = extractelement <8 x half> %vec_half, i64 6
%f7 = extractelement <8 x half> %vec_half, i64 7
%sum0 = fadd half %f0, %f1
%sum1 = fadd half %f2, %f3
%sum2 = fadd half %f4, %f5
%sum3 = fadd half %f6, %f7
%sum4 = fadd half %sum0, %sum1
%sum5 = fadd half %sum2, %sum3
%rslt = fadd half %sum4, %sum5
ret half %rslt
}
define half @fptrunc_v16f32_v16f16_extract_uses(<16 x float> %vec_float) {
; GFX950-SDAG-LABEL: fptrunc_v16f32_v16f16_extract_uses:
; GFX950-SDAG: ; %bb.0:
; GFX950-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX950-SDAG-NEXT: v_cvt_pk_f16_f32 v14, v14, v15
; GFX950-SDAG-NEXT: v_cvt_pk_f16_f32 v12, v12, v13
; GFX950-SDAG-NEXT: v_cvt_pk_f16_f32 v10, v10, v11
; GFX950-SDAG-NEXT: v_cvt_pk_f16_f32 v8, v8, v9
; GFX950-SDAG-NEXT: v_cvt_pk_f16_f32 v6, v6, v7
; GFX950-SDAG-NEXT: v_cvt_pk_f16_f32 v4, v4, v5
; GFX950-SDAG-NEXT: v_cvt_pk_f16_f32 v2, v2, v3
; GFX950-SDAG-NEXT: v_cvt_pk_f16_f32 v0, v0, v1
; GFX950-SDAG-NEXT: v_add_f16_sdwa v0, v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX950-SDAG-NEXT: v_add_f16_sdwa v1, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX950-SDAG-NEXT: v_add_f16_sdwa v2, v4, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX950-SDAG-NEXT: v_add_f16_sdwa v3, v6, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX950-SDAG-NEXT: v_add_f16_sdwa v4, v8, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX950-SDAG-NEXT: v_add_f16_sdwa v5, v10, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX950-SDAG-NEXT: v_add_f16_sdwa v6, v12, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX950-SDAG-NEXT: v_add_f16_sdwa v7, v14, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX950-SDAG-NEXT: v_add_f16_e32 v0, v0, v1
; GFX950-SDAG-NEXT: v_add_f16_e32 v1, v2, v3
; GFX950-SDAG-NEXT: v_add_f16_e32 v2, v4, v5
; GFX950-SDAG-NEXT: v_add_f16_e32 v3, v6, v7
; GFX950-SDAG-NEXT: v_add_f16_e32 v0, v0, v1
; GFX950-SDAG-NEXT: v_add_f16_e32 v1, v2, v3
; GFX950-SDAG-NEXT: v_add_f16_e32 v0, v0, v1
; GFX950-SDAG-NEXT: s_setpc_b64 s[30:31]
;
; GFX950-GISEL-LABEL: fptrunc_v16f32_v16f16_extract_uses:
; GFX950-GISEL: ; %bb.0:
; GFX950-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX950-GISEL-NEXT: v_cvt_pk_f16_f32 v0, v0, v1
; GFX950-GISEL-NEXT: v_cvt_pk_f16_f32 v1, v2, v3
; GFX950-GISEL-NEXT: v_cvt_pk_f16_f32 v2, v4, v5
; GFX950-GISEL-NEXT: v_cvt_pk_f16_f32 v3, v6, v7
; GFX950-GISEL-NEXT: v_cvt_pk_f16_f32 v4, v8, v9
; GFX950-GISEL-NEXT: v_cvt_pk_f16_f32 v5, v10, v11
; GFX950-GISEL-NEXT: v_cvt_pk_f16_f32 v6, v12, v13
; GFX950-GISEL-NEXT: v_cvt_pk_f16_f32 v7, v14, v15
; GFX950-GISEL-NEXT: v_add_f16_sdwa v0, v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX950-GISEL-NEXT: v_add_f16_sdwa v1, v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX950-GISEL-NEXT: v_add_f16_sdwa v2, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX950-GISEL-NEXT: v_add_f16_sdwa v3, v3, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX950-GISEL-NEXT: v_add_f16_sdwa v4, v4, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX950-GISEL-NEXT: v_add_f16_sdwa v5, v5, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX950-GISEL-NEXT: v_add_f16_sdwa v6, v6, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX950-GISEL-NEXT: v_add_f16_sdwa v7, v7, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX950-GISEL-NEXT: v_add_f16_e32 v0, v0, v1
; GFX950-GISEL-NEXT: v_add_f16_e32 v1, v2, v3
; GFX950-GISEL-NEXT: v_add_f16_e32 v2, v4, v5
; GFX950-GISEL-NEXT: v_add_f16_e32 v3, v6, v7
; GFX950-GISEL-NEXT: v_add_f16_e32 v0, v0, v1
; GFX950-GISEL-NEXT: v_add_f16_e32 v1, v2, v3
; GFX950-GISEL-NEXT: v_add_f16_e32 v0, v0, v1
; GFX950-GISEL-NEXT: s_setpc_b64 s[30:31]
%vec_half = fptrunc <16 x float> %vec_float to <16 x half>
%f0 = extractelement <16 x half> %vec_half, i64 0
%f1 = extractelement <16 x half> %vec_half, i64 1
%f2 = extractelement <16 x half> %vec_half, i64 2
%f3 = extractelement <16 x half> %vec_half, i64 3
%f4 = extractelement <16 x half> %vec_half, i64 4
%f5 = extractelement <16 x half> %vec_half, i64 5
%f6 = extractelement <16 x half> %vec_half, i64 6
%f7 = extractelement <16 x half> %vec_half, i64 7
%f8 = extractelement <16 x half> %vec_half, i64 8
%f9 = extractelement <16 x half> %vec_half, i64 9
%f10 = extractelement <16 x half> %vec_half, i64 10
%f11 = extractelement <16 x half> %vec_half, i64 11
%f12 = extractelement <16 x half> %vec_half, i64 12
%f13 = extractelement <16 x half> %vec_half, i64 13
%f14 = extractelement <16 x half> %vec_half, i64 14
%f15 = extractelement <16 x half> %vec_half, i64 15
%sum0 = fadd half %f0, %f1
%sum1 = fadd half %f2, %f3
%sum2 = fadd half %f4, %f5
%sum3 = fadd half %f6, %f7
%sum4 = fadd half %f8, %f9
%sum5 = fadd half %f10, %f11
%sum6 = fadd half %f12, %f13
%sum7 = fadd half %f14, %f15
%sum8 = fadd half %sum0, %sum1
%sum9 = fadd half %sum2, %sum3
%sum10 = fadd half %sum4, %sum5
%sum11 = fadd half %sum6, %sum7
%sum12 = fadd half %sum8, %sum9
%sum13 = fadd half %sum10, %sum11
%rslt = fadd half %sum12, %sum13
ret half %rslt
}
define <2 x half> @v_test_cvt_v2f64_v2f16(<2 x double> %src) {