Skip to content

Commit 5c722a6

Browse files
rampitecrampitec
committed
[AMDGPU] Prioritize allocation of low 256 VGPR classes
If we have 1024 VGPRs available we need to give priority to the allocation of these registers where operands can only use low 256. That is noteably scale operands of V_WMMA_SCALE instructions. Otherwise large tuples will be allocated first and take all low registers, so we would have to spill to get a room for these scale registers. Allocation priority itself does not eliminate spilling completely in large kernels, although helps to some degree. Increasing spill weight of a restricted class on top of it helps.
1 parent 4be0ab6 commit 5c722a6

File tree

3 files changed

+142
-1
lines changed

3 files changed

+142
-1
lines changed

‎llvm/lib/Target/AMDGPU/SIRegisterInfo.h‎

Lines changed: 11 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -496,6 +496,17 @@ class SIRegisterInfo final : public AMDGPUGenRegisterInfo {
496496

497497
SmallVector<StringLiteral>
498498
getVRegFlagsOfReg(Register Reg, const MachineFunction &MF) const override;
499+
500+
float
501+
getSpillWeightScaleFactor(const TargetRegisterClass *RC) const override {
502+
// Prioritize VGPR_32_Lo256 over other classes which may occupy registers
503+
// beyond v256.
504+
return AMDGPUGenRegisterInfo::getSpillWeightScaleFactor(RC) *
505+
((RC == &AMDGPU::VGPR_32_Lo256RegClass ||
506+
RC == &AMDGPU::VReg_64_Lo256_Align2RegClass)
507+
? 2.0
508+
: 1.0);
509+
}
499510
};
500511

501512
namespace AMDGPU {

‎llvm/lib/Target/AMDGPU/SIRegisterInfo.td‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -644,7 +644,7 @@ def VGPR_32_Lo128 : SIRegisterClass<"AMDGPU", !listconcat(Reg32Types.types, Reg1
644644
// Identical to VGPR_32 except it only contains the low 256 (Lo256) registers.
645645
def VGPR_32_Lo256 : SIRegisterClass<"AMDGPU", !listconcat(Reg32Types.types, Reg16Types.types), 32,
646646
(add (sequence "VGPR%u", 0, 255))> {
647-
let AllocationPriority = 0;
647+
let AllocationPriority = !add(3, !mul(BaseClassPriority, BaseClassScaleFactor));
648648
let GeneratePressureSet = 0;
649649
let Size = 32;
650650
let Weight = 1;

‎llvm/test/CodeGen/AMDGPU/regalloc-spill-wmma-scale.ll‎

Lines changed: 130 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,130 @@
1+
; RUN: llc -mtriple=amdgcn -mcpu=gfx1250 < %s | FileCheck %s
2+
3+
; Scale operands of WMMA are limited to low 256 VGPRs
4+
; Make sure we do not spill scale operands because of the low 256 restriction.
5+
; CHECK: ; ScratchSize: 0
6+
; CHECK: ; Occupancy: 1
7+
8+
define amdgpu_kernel void @spill_scale_test(float %arg, float %arg1, float %arg2, float %arg3, float %arg4, float %arg5, float %arg6, float %arg7, <16 x i32> %arg8, float %arg9, <16 x i32> %arg10, float %arg11, <16 x i8> %arg12) #0 {
9+
bb:
10+
%i = shufflevector <16 x i8> %arg12, <16 x i8> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
11+
tail call void @llvm.amdgcn.global.load.async.to.lds.b32(ptr addrspace(1) null, ptr addrspace(3) null, i32 0, i32 0)
12+
%i13 = bitcast <64 x i8> %i to <16 x i32>
13+
tail call void @llvm.amdgcn.global.load.async.to.lds.b32(ptr addrspace(1) null, ptr addrspace(3) null, i32 0, i32 0)
14+
%i14 = tail call <2 x i32> @llvm.amdgcn.ds.load.tr8.b64.v2i32(ptr addrspace(3) null)
15+
%i15 = bitcast <2 x i32> %i14 to <8 x i8>
16+
%i16 = tail call <2 x i32> @llvm.amdgcn.ds.load.tr8.b64.v2i32(ptr addrspace(3) null)
17+
%i17 = shufflevector <8 x i8> %i15, <8 x i8> zeroinitializer, <64 x i32> <i32 0, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
18+
%i18 = shufflevector <64 x i8> zeroinitializer, <64 x i8> %i17, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 64, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
19+
%i19 = insertelement <64 x i8> %i18, i8 0, i64 57
20+
%i20 = bitcast <64 x i8> %i19 to <16 x i32>
21+
%.extract2214 = extractelement <2 x i32> %i16, i64 0
22+
%i21 = tail call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 0, <16 x i32> %i20, i32 0, <16 x i32> %i13, i16 0, <8 x float> zeroinitializer, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)
23+
%i22 = extractelement <8 x float> %i21, i64 0
24+
%i23 = tail call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 0, <16 x i32> zeroinitializer, i32 0, <16 x i32> zeroinitializer, i16 0, <8 x float> zeroinitializer, i32 0, i32 0, i32 %.extract2214, i32 0, i32 0, i32 0, i1 false, i1 false)
25+
%i24 = extractelement <8 x float> %i23, i64 0
26+
%i25 = tail call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 0, <16 x i32> %arg8, i32 0, <16 x i32> zeroinitializer, i16 0, <8 x float> zeroinitializer, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)
27+
%i26 = extractelement <8 x float> %i25, i64 0
28+
%i27 = fmul float %i22, 0.000000e+00
29+
%i28 = fmul float %i24, 0.000000e+00
30+
%i29 = insertelement <2 x float> zeroinitializer, float %i26, i64 1
31+
%i30 = insertelement <2 x float> zeroinitializer, float %i28, i64 0
32+
%i31 = insertelement <2 x float> zeroinitializer, float %arg11, i64 0
33+
%i32 = fadd <2 x float> %i31, %i30
34+
%i33 = insertelement <2 x float> zeroinitializer, float %arg9, i64 0
35+
%i34 = fadd <2 x float> %i33, %i32
36+
%i35 = insertelement <2 x float> zeroinitializer, float %arg7, i64 0
37+
%i36 = fadd <2 x float> %i35, %i34
38+
%i37 = insertelement <2 x float> zeroinitializer, float %arg1, i64 0
39+
%i38 = fadd <2 x float> %i37, %i36
40+
%i39 = insertelement <2 x float> zeroinitializer, float %arg6, i64 0
41+
%i40 = fadd <2 x float> %i39, %i38
42+
%i41 = insertelement <2 x float> zeroinitializer, float %arg4, i64 0
43+
%i42 = fadd <2 x float> %i41, %i40
44+
%i43 = insertelement <2 x float> zeroinitializer, float %arg5, i64 0
45+
%i44 = fadd <2 x float> %i43, %i42
46+
%i45 = insertelement <2 x float> zeroinitializer, float %arg3, i64 0
47+
%i46 = fadd <2 x float> %i45, %i44
48+
%i47 = insertelement <2 x float> zeroinitializer, float %arg, i64 0
49+
%i48 = insertelement <2 x float> zeroinitializer, float %arg2, i64 0
50+
%i49 = fadd <2 x float> %i48, %i46
51+
%i50 = fadd <2 x float> %i29, %i49
52+
%i51 = fadd <2 x float> %i47, %i50
53+
%i52 = insertelement <8 x float> zeroinitializer, float %i27, i64 0
54+
%i53 = tail call <2 x i32> @llvm.amdgcn.cvt.scalef32.pk8.fp8.f32(<8 x float> %i52, float 0.000000e+00)
55+
%i54 = tail call <2 x i32> @llvm.amdgcn.cvt.scalef32.pk8.fp8.f32(<8 x float> splat (float 0x7FF8000000000000), float 0.000000e+00)
56+
%i55 = tail call <2 x i32> @llvm.amdgcn.cvt.scalef32.pk8.fp8.f32(<8 x float> splat (float 1.000000e+00), float 0.000000e+00)
57+
%.extract1415 = extractelement <2 x i32> %i53, i64 0
58+
%.extract1416 = extractelement <2 x i32> %i54, i64 0
59+
%.extract1424 = extractelement <2 x i32> %i55, i64 0
60+
%i56 = tail call i32 @llvm.amdgcn.ds.swizzle(i32 %.extract1416, i32 0)
61+
%i57 = bitcast i32 %.extract1415 to <4 x i8>
62+
%i58 = shufflevector <4 x i8> %i57, <4 x i8> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
63+
%i59 = bitcast i32 %i56 to <4 x i8>
64+
%i60 = bitcast i32 %.extract1424 to <4 x i8>
65+
%i61 = shufflevector <4 x i8> %i60, <4 x i8> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
66+
%i62 = tail call <2 x i32> @llvm.amdgcn.ds.load.tr8.b64.v2i32(ptr addrspace(3) null)
67+
%i63 = bitcast <2 x i32> %i62 to <8 x i8>
68+
%i64 = shufflevector <8 x i8> %i63, <8 x i8> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
69+
%i65 = tail call <2 x i32> @llvm.amdgcn.ds.load.tr8.b64.v2i32(ptr addrspace(3) null)
70+
%i66 = bitcast <2 x i32> %i65 to <8 x i8>
71+
%i67 = shufflevector <8 x i8> %i66, <8 x i8> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
72+
%i68 = tail call <2 x i32> @llvm.amdgcn.ds.load.tr8.b64.v2i32(ptr addrspace(3) null)
73+
%i69 = bitcast <2 x i32> %i68 to <8 x i8>
74+
%i70 = shufflevector <8 x i8> %i69, <8 x i8> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
75+
%i71 = tail call <2 x i32> @llvm.amdgcn.ds.load.tr8.b64.v2i32(ptr addrspace(3) getelementptr (i8, ptr addrspace(3) null, i32 75232))
76+
%i72 = shufflevector <64 x i8> zeroinitializer, <64 x i8> %i58, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 64, i32 65, i32 66, i32 67>
77+
%i73 = bitcast <64 x i8> %i72 to <16 x i32>
78+
%i74 = shufflevector <4 x i8> %i59, <4 x i8> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
79+
%i75 = shufflevector <64 x i8> %i74, <64 x i8> %i61, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 64, i32 65, i32 66, i32 67, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
80+
%i76 = bitcast <64 x i8> %i75 to <16 x i32>
81+
%i77 = shufflevector <64 x i8> zeroinitializer, <64 x i8> %i64, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
82+
%i78 = bitcast <64 x i8> %i77 to <16 x i32>
83+
%i79 = bitcast <64 x i8> %i67 to <16 x i32>
84+
%i80 = shufflevector <64 x i8> zeroinitializer, <64 x i8> %i70, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 64, i32 65, i32 66, i32 67, i32 68, i32 69, i32 70, i32 71>
85+
%i81 = bitcast <64 x i8> %i80 to <16 x i32>
86+
%.extract1434 = extractelement <2 x i32> %i71, i64 0
87+
%i82 = tail call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 0, <16 x i32> %i78, i32 0, <16 x i32> zeroinitializer, i16 0, <8 x float> zeroinitializer, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)
88+
%i83 = tail call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 0, <16 x i32> %arg10, i32 0, <16 x i32> %i73, i16 0, <8 x float> zeroinitializer, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)
89+
%i84 = tail call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 0, <16 x i32> %i79, i32 0, <16 x i32> %i73, i16 0, <8 x float> zeroinitializer, i32 0, i32 0, i32 0, i32 0, i32 0, i32 2139062143, i1 false, i1 false)
90+
%i85 = tail call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 0, <16 x i32> %i81, i32 0, <16 x i32> %arg8, i16 0, <8 x float> zeroinitializer, i32 0, i32 0, i32 0, i32 0, i32 0, i32 2139062143, i1 false, i1 false)
91+
%i86 = tail call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 0, <16 x i32> splat (i32 16843009), i32 0, <16 x i32> %arg10, i16 0, <8 x float> zeroinitializer, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)
92+
%i87 = tail call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 0, <16 x i32> zeroinitializer, i32 0, <16 x i32> zeroinitializer, i16 0, <8 x float> zeroinitializer, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)
93+
%i88 = tail call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 0, <16 x i32> splat (i32 1), i32 0, <16 x i32> %i76, i16 0, <8 x float> zeroinitializer, i32 0, i32 0, i32 %.extract1434, i32 0, i32 0, i32 0, i1 false, i1 false)
94+
%i89 = fdiv <8 x float> %i82, zeroinitializer
95+
%i90 = fcmp uno <8 x float> %i89, zeroinitializer
96+
%i91 = select <8 x i1> %i90, <8 x bfloat> splat (bfloat 0xR3F80), <8 x bfloat> zeroinitializer
97+
%i92 = bitcast <8 x bfloat> %i91 to <4 x i32>
98+
tail call void @llvm.amdgcn.raw.ptr.buffer.store.v4i32(<4 x i32> %i92, ptr addrspace(8) null, i32 0, i32 0, i32 0)
99+
%i93 = fdiv <8 x float> %i83, zeroinitializer
100+
%i94 = fcmp uno <8 x float> %i93, zeroinitializer
101+
%i95 = select <8 x i1> %i94, <8 x bfloat> splat (bfloat 0xR3F80), <8 x bfloat> zeroinitializer
102+
%i96 = bitcast <8 x bfloat> %i95 to <4 x i32>
103+
tail call void @llvm.amdgcn.raw.ptr.buffer.store.v4i32(<4 x i32> %i96, ptr addrspace(8) null, i32 0, i32 0, i32 0)
104+
%i97 = fcmp uno <8 x float> %i84, zeroinitializer
105+
%i98 = select <8 x i1> %i97, <8 x bfloat> splat (bfloat 0xR3F80), <8 x bfloat> zeroinitializer
106+
%i99 = bitcast <8 x bfloat> %i98 to <4 x i32>
107+
tail call void @llvm.amdgcn.raw.ptr.buffer.store.v4i32(<4 x i32> %i99, ptr addrspace(8) null, i32 0, i32 0, i32 0)
108+
%i100 = fcmp uno <8 x float> %i85, zeroinitializer
109+
%i101 = select <8 x i1> %i100, <8 x bfloat> splat (bfloat 0xR3F80), <8 x bfloat> zeroinitializer
110+
%i102 = bitcast <8 x bfloat> %i101 to <4 x i32>
111+
tail call void @llvm.amdgcn.raw.ptr.buffer.store.v4i32(<4 x i32> %i102, ptr addrspace(8) null, i32 0, i32 0, i32 0)
112+
%i103 = fcmp uno <8 x float> %i86, zeroinitializer
113+
%i104 = select <8 x i1> %i103, <8 x bfloat> splat (bfloat 0xR3F80), <8 x bfloat> zeroinitializer
114+
%i105 = bitcast <8 x bfloat> %i104 to <4 x i32>
115+
tail call void @llvm.amdgcn.raw.ptr.buffer.store.v4i32(<4 x i32> %i105, ptr addrspace(8) null, i32 0, i32 0, i32 0)
116+
%i106 = fcmp uno <8 x float> %i87, zeroinitializer
117+
%i107 = select <8 x i1> %i106, <8 x bfloat> splat (bfloat 0xR3F80), <8 x bfloat> zeroinitializer
118+
%i108 = bitcast <8 x bfloat> %i107 to <4 x i32>
119+
tail call void @llvm.amdgcn.raw.ptr.buffer.store.v4i32(<4 x i32> %i108, ptr addrspace(8) null, i32 0, i32 0, i32 0)
120+
%i109 = shufflevector <2 x float> %i51, <2 x float> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 0, i32 1>
121+
%i110 = shufflevector <4 x float> %i109, <4 x float> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
122+
%i111 = fmul <8 x float> %i88, %i110
123+
%i112 = fcmp uno <8 x float> %i111, zeroinitializer
124+
%i113 = select <8 x i1> %i112, <8 x bfloat> splat (bfloat 0xR3F80), <8 x bfloat> zeroinitializer
125+
%i114 = bitcast <8 x bfloat> %i113 to <4 x i32>
126+
tail call void @llvm.amdgcn.raw.ptr.buffer.store.v4i32(<4 x i32> %i114, ptr addrspace(8) null, i32 0, i32 0, i32 0)
127+
ret void
128+
}
129+
130+
attributes #0 = { "amdgpu-flat-work-group-size"="1,128" "amdgpu-waves-per-eu"="1,1" }

0 commit comments

Comments
 (0)