Skip to content

Commit 2c920a1

Browse files
rampitecrampitec
authored
[AMDGPU] wmma_scale* IR verification (#155493)
1 parent d42f5eb commit 2c920a1

File tree

2 files changed

+127
-1
lines changed

2 files changed

+127
-1
lines changed

‎llvm/lib/IR/Verifier.cpp

Lines changed: 3 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -6724,7 +6724,9 @@ void Verifier::visitIntrinsicCall(Intrinsic::ID ID, CallBase &Call) {
67246724
"invalid vector type for format", &Call, Src1, Call.getArgOperand(5));
67256725
break;
67266726
}
6727-
case Intrinsic::amdgcn_wmma_f32_16x16x128_f8f6f4: {
6727+
case Intrinsic::amdgcn_wmma_f32_16x16x128_f8f6f4:
6728+
case Intrinsic::amdgcn_wmma_scale_f32_16x16x128_f8f6f4:
6729+
case Intrinsic::amdgcn_wmma_scale16_f32_16x16x128_f8f6f4: {
67286730
Value *Src0 = Call.getArgOperand(1);
67296731
Value *Src1 = Call.getArgOperand(3);
67306732

‎llvm/test/Verifier/AMDGPU/wmma-f8f6f4.ll

Lines changed: 124 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -24,6 +24,46 @@ bb:
2424
ret void
2525
}
2626

27+
; CHECK: operand 1 must be 8, 12 or 16 element i32 vector
28+
; CHECK-NEXT: %res = call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v16i64.v16i32(i32 0, <16 x i64> %A, i32 0, <16 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)
29+
; CHECK-NEXT: <16 x i64> %A
30+
define amdgpu_ps void @test_wmma_scale_f32_16x16x128_f8f6f4___v16i64_fp8___v16i32_fp8(<16 x i64> %A, <16 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {
31+
bb:
32+
%res = call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v16i64.v16i32(i32 0, <16 x i64> %A, i32 0, <16 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)
33+
store <8 x float> %res, ptr addrspace(1) %out
34+
ret void
35+
}
36+
37+
; CHECK: operand 3 must be 8, 12 or 16 element i32 vector
38+
; CHECK-NEXT: %res = call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i64(i32 0, <16 x i32> %A, i32 0, <16 x i64> %B, i16 0, <8 x float> %C, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)
39+
; CHECK-NEXT: <16 x i64> %B
40+
define amdgpu_ps void @test_wmma_scale_f32_16x16x128_f8f6f4___v16i32_fp8___v16i64_fp8(<16 x i32> %A, <16 x i64> %B, <8 x float> %C, ptr addrspace(1) %out) {
41+
bb:
42+
%res = call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i64(i32 0, <16 x i32> %A, i32 0, <16 x i64> %B, i16 0, <8 x float> %C, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)
43+
store <8 x float> %res, ptr addrspace(1) %out
44+
ret void
45+
}
46+
47+
; CHECK: operand 1 must be 8, 12 or 16 element i32 vector
48+
; CHECK-NEXT: %res = call <8 x float> @llvm.amdgcn.wmma.scale16.f32.16x16x128.f8f6f4.v8f32.v16i64.v16i32(i32 0, <16 x i64> %A, i32 0, <16 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i64 0, i32 0, i32 0, i64 0, i1 false, i1 false)
49+
; CHECK-NEXT: <16 x i64> %A
50+
define amdgpu_ps void @test_wmma_scale16_f32_16x16x128_f8f6f4___v16i64_fp8___v16i32_fp8(<16 x i64> %A, <16 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {
51+
bb:
52+
%res = call <8 x float> @llvm.amdgcn.wmma.scale16.f32.16x16x128.f8f6f4.v8f32.v16i64.v16i32(i32 0, <16 x i64> %A, i32 0, <16 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i64 0, i32 0, i32 0, i64 0, i1 false, i1 false)
53+
store <8 x float> %res, ptr addrspace(1) %out
54+
ret void
55+
}
56+
57+
; CHECK: operand 3 must be 8, 12 or 16 element i32 vector
58+
; CHECK-NEXT: %res = call <8 x float> @llvm.amdgcn.wmma.scale16.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i64(i32 0, <16 x i32> %A, i32 0, <16 x i64> %B, i16 0, <8 x float> %C, i32 0, i32 0, i64 0, i32 0, i32 0, i64 0, i1 false, i1 false)
59+
; CHECK-NEXT: <16 x i64> %B
60+
define amdgpu_ps void @test_wmma_scale16_f32_16x16x128_f8f6f4___v16i32_fp8___v16i64_fp8(<16 x i32> %A, <16 x i64> %B, <8 x float> %C, ptr addrspace(1) %out) {
61+
bb:
62+
%res = call <8 x float> @llvm.amdgcn.wmma.scale16.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i64(i32 0, <16 x i32> %A, i32 0, <16 x i64> %B, i16 0, <8 x float> %C, i32 0, i32 0, i64 0, i32 0, i32 0, i64 0, i1 false, i1 false)
63+
store <8 x float> %res, ptr addrspace(1) %out
64+
ret void
65+
}
66+
2767
; --------------------------------------------------------------------
2868
; Impossible vector types
2969
; --------------------------------------------------------------------
@@ -48,6 +88,26 @@ bb:
4888
ret void
4989
}
5090

91+
; CHECK: operand 1 must be 8, 12 or 16 element i32 vector
92+
; CHECK-NEXT: call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v15i32.v16i32(i32 0, <15 x i32> %A, i32 0, <16 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)
93+
; CHECK-NEXT: <15 x i32> %A
94+
define amdgpu_ps void @test_wmma_scale_f32_16x16x128_f8f6f4___v15i32_fp8___v16i32_fp8(<15 x i32> %A, <16 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {
95+
bb:
96+
%res = call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v15i32.v16i32(i32 0, <15 x i32> %A, i32 0, <16 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)
97+
store <8 x float> %res, ptr addrspace(1) %out
98+
ret void
99+
}
100+
101+
; CHECK: operand 3 must be 8, 12 or 16 element i32 vector
102+
; CHECK-NEXT: call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v16i32.v15i32(i32 0, <16 x i32> %A, i32 0, <15 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)
103+
; CHECK-NEXT: <15 x i32> %B
104+
define amdgpu_ps void @test_wmma_scale_f32_16x16x128_f8f6f4___v16i32_fp8___v15i32_fp8(<16 x i32> %A, <15 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {
105+
bb:
106+
%res = call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v16i32.v15i32(i32 0, <16 x i32> %A, i32 0, <15 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)
107+
store <8 x float> %res, ptr addrspace(1) %out
108+
ret void
109+
}
110+
51111
; --------------------------------------------------------------------
52112
; Out of bounds format
53113
; --------------------------------------------------------------------
@@ -72,6 +132,26 @@ bb:
72132
ret void
73133
}
74134

135+
; CHECK: invalid value for matrix format
136+
; CHECK-NEXT: %res = call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 5, <16 x i32> %A, i32 0, <16 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)
137+
; CHECK-NEXT: i32 5
138+
define amdgpu_ps void @test_wmma_scale_f32_16x16x128_f8f6f4___v16i32_invalid0___v16i32_fp8(<16 x i32> %A, <16 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {
139+
bb:
140+
%res = call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 5, <16 x i32> %A, i32 0, <16 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)
141+
store <8 x float> %res, ptr addrspace(1) %out
142+
ret void
143+
}
144+
145+
; CHECK: invalid value for matrix format
146+
; CHECK-NEXT: %res = call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 0, <16 x i32> %A, i32 5, <16 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)
147+
; CHECK-NEXT: i32 5
148+
define amdgpu_ps void @test_wmma_scale_f32_16x16x128_f8f6f4___v16i32_fp8___v16i32_invalid1(<16 x i32> %A, <16 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {
149+
bb:
150+
%res = call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 0, <16 x i32> %A, i32 5, <16 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)
151+
store <8 x float> %res, ptr addrspace(1) %out
152+
ret void
153+
}
154+
75155
; --------------------------------------------------------------------
76156
; Incorrect signature for format cases (IR vector too small)
77157
; --------------------------------------------------------------------
@@ -163,3 +243,47 @@ bb:
163243
store <8 x float> %res, ptr addrspace(1) %out
164244
ret void
165245
}
246+
247+
; CHECK: invalid vector type for format
248+
; CHECK-NEXT: %res = call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v8i32.v16i32(i32 2, <8 x i32> %A, i32 0, <16 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)
249+
; CHECK-NEXT: <8 x i32> %A
250+
; CHECK-NEXT: i32 2
251+
define amdgpu_ps void @test_wmma_scale_f32_16x16x128_f8f6f4___v8i32_fp6___v16i32_fp8(<8 x i32> %A, <16 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {
252+
bb:
253+
%res = call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v8i32.v16i32(i32 2, <8 x i32> %A, i32 0, <16 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)
254+
store <8 x float> %res, ptr addrspace(1) %out
255+
ret void
256+
}
257+
258+
; CHECK: invalid vector type for format
259+
; CHECK-NEXT: %res = call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v16i32.v8i32(i32 0, <16 x i32> %A, i32 2, <8 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)
260+
; CHECK-NEXT: <8 x i32> %B
261+
; CHECK-NEXT: i32 2
262+
define amdgpu_ps void @test_wmma_scale_f32_16x16x128_f8f6f4___v16i32_fp8___v8i32_fp6(<16 x i32> %A, <8 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {
263+
bb:
264+
%res = call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v8i32.v16i32(i32 0, <16 x i32> %A, i32 2, <8 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)
265+
store <8 x float> %res, ptr addrspace(1) %out
266+
ret void
267+
}
268+
269+
; CHECK: invalid vector type for format
270+
; CHECK-NEXT: %res = call <8 x float> @llvm.amdgcn.wmma.scale16.f32.16x16x128.f8f6f4.v8f32.v8i32.v16i32(i32 3, <8 x i32> %A, i32 0, <16 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i64 0, i32 0, i32 0, i64 0, i1 false, i1 false)
271+
; CHECK-NEXT: <8 x i32> %A
272+
; CHECK-NEXT: i32 3
273+
define amdgpu_ps void @test_wmma_scale16_f32_16x16x128_f8f6f4___v8i32_bf6___v16i32_fp8(<8 x i32> %A, <16 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {
274+
bb:
275+
%res = call <8 x float> @llvm.amdgcn.wmma.scale16.f32.16x16x128.f8f6f4.v8f32.v8i32.v16i32(i32 3, <8 x i32> %A, i32 0, <16 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i64 0, i32 0, i32 0, i64 0, i1 false, i1 false)
276+
store <8 x float> %res, ptr addrspace(1) %out
277+
ret void
278+
}
279+
280+
; CHECK: invalid vector type for format
281+
; CHECK-NEXT: %res = call <8 x float> @llvm.amdgcn.wmma.scale16.f32.16x16x128.f8f6f4.v8f32.v16i32.v8i32(i32 0, <16 x i32> %A, i32 3, <8 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i64 0, i32 0, i32 0, i64 0, i1 false, i1 false)
282+
; CHECK-NEXT: <8 x i32> %B
283+
; CHECK-NEXT: i32 3
284+
define amdgpu_ps void @test_wmma_scale_f32_16x16x128_f8f6f4___v16i32_fp8___v8i32_bf6(<16 x i32> %A, <8 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {
285+
bb:
286+
%res = call <8 x float> @llvm.amdgcn.wmma.scale16.f32.16x16x128.f8f6f4.v8f32.v8i32.v16i32(i32 0, <16 x i32> %A, i32 3, <8 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i64 0, i32 0, i32 0, i64 0, i1 false, i1 false)
287+
store <8 x float> %res, ptr addrspace(1) %out
288+
ret void
289+
}

0 commit comments

Comments
 (0)