Skip to content

Commit 673e5e3

Browse files
fhahnfhahn
committed
[Matrix] Add dedicated tests for transpose lifting.
Add extra test coverage for transpose lifting using -matrix-print-after-transpose-opt. The added tests show a mis-compile.
1 parent 1da7f41 commit 673e5e3

File tree

1 file changed

+115
-0
lines changed

1 file changed

+115
-0
lines changed

‎llvm/test/Transforms/LowerMatrixIntrinsics/transpose-opts-lifting.ll

Lines changed: 115 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,115 @@
1+
; RUN: opt -p lower-matrix-intrinsics -matrix-print-after-transpose-opt -disable-output -S %s 2>&1 | FileCheck %s
2+
3+
; REQUIRES: asserts
4+
5+
target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
6+
7+
; FIXME: Lifted transpose dimensions are incorrect.
8+
define <6 x double> @lift_through_add_matching_transpose_dimensions(<6 x double> %a, <6 x double> %b) {
9+
; CHECK-LABEL: define <6 x double> @lift_through_add_matching_transpose_dimensions(<6 x double> %a, <6 x double> %b) {
10+
; CHECK-NEXT: entry:
11+
; CHECK-NEXT: [[A:%.+]] = fadd <6 x double> %a, %b
12+
; CHECK-NEXT: [[T:%.+]] = call <6 x double> @llvm.matrix.transpose.v6f64(<6 x double> [[A]], i32 2, i32 3)
13+
; CHECK-NEXT: ret <6 x double> [[T]]
14+
;
15+
entry:
16+
%a.t = call <6 x double> @llvm.matrix.transpose.v6f64(<6 x double> %a, i32 3, i32 2)
17+
%b.t = call <6 x double> @llvm.matrix.transpose.v6f64(<6 x double> %b, i32 3, i32 2)
18+
%add = fadd <6 x double> %a.t, %b.t
19+
ret <6 x double> %add
20+
}
21+
22+
define <6 x double> @lift_through_add_matching_transpose_dimensions_ops_also_have_shape_info(ptr %a.ptr, ptr %b.ptr) {
23+
; CHECK-LABEL: define <6 x double> @lift_through_add_matching_transpose_dimensions_ops_also_have_shape_info(ptr %a.ptr, ptr %b.ptr)
24+
; CHECK-NEXT: entry:
25+
; CHECK-NEXT: [[A:%.+]] = load <6 x double>, ptr %a.ptr
26+
; CHECK-NEXT: [[B:%.+]] = load <6 x double>, ptr %b.ptr
27+
; CHECK-NEXT: [[ADD:%.+]] = fadd <6 x double> [[A]], [[B]]
28+
; CHECK-NEXT: [[T:%.+]] = call <6 x double> @llvm.matrix.transpose.v6f64(<6 x double> [[ADD]], i32 2, i32 3)
29+
; CHECK-NEXT: ret <6 x double> [[T]]
30+
;
31+
entry:
32+
%a = load <6 x double>, ptr %a.ptr
33+
%b = load <6 x double>, ptr %b.ptr
34+
%a.t = call <6 x double> @llvm.matrix.transpose.v6f64(<6 x double> %a, i32 3, i32 2)
35+
%b.t = call <6 x double> @llvm.matrix.transpose.v6f64(<6 x double> %b, i32 3, i32 2)
36+
%add = fadd <6 x double> %a.t, %b.t
37+
ret <6 x double> %add
38+
}
39+
40+
define <6 x double> @lift_through_add_mismatching_dimensions_1(<6 x double> %a, <6 x double> %b) {
41+
; CHECK-LABEL: define <6 x double> @lift_through_add_mismatching_dimensions_1(<6 x double> %a, <6 x double> %b) {
42+
; CHECK-NEXT: entry:
43+
; CHECK-NEXT: [[A:%.+]] = fadd <6 x double> %a, %b
44+
; CHECK-NEXT: [[T:%.+]] = call <6 x double> @llvm.matrix.transpose.v6f64(<6 x double> [[A]], i32 2, i32 3)
45+
; CHECK-NEXT: ret <6 x double> [[T]]
46+
;
47+
entry:
48+
%a.t = call <6 x double> @llvm.matrix.transpose.v6f64(<6 x double> %a, i32 1, i32 6)
49+
%b.t = call <6 x double> @llvm.matrix.transpose.v6f64(<6 x double> %b, i32 3, i32 2)
50+
%add = fadd <6 x double> %a.t, %b.t
51+
ret <6 x double> %add
52+
}
53+
54+
define <6 x double> @lift_through_add_mismatching_dimensions_2(<6 x double> %a, <6 x double> %b) {
55+
; CHECK-LABEL: define <6 x double> @lift_through_add_mismatching_dimensions_2(<6 x double> %a, <6 x double> %b) {
56+
; CHECK-NEXT: entry:
57+
; CHECK-NEXT: [[A:%.+]] = fadd <6 x double> %a, %b
58+
; CHECK-NEXT: [[T:%.+]] = call <6 x double> @llvm.matrix.transpose.v6f64(<6 x double> [[A]], i32 1, i32 6)
59+
; CHECK-NEXT: ret <6 x double> [[T]]
60+
;
61+
62+
entry:
63+
%a.t = call <6 x double> @llvm.matrix.transpose.v6f64(<6 x double> %a, i32 3, i32 2)
64+
%b.t = call <6 x double> @llvm.matrix.transpose.v6f64(<6 x double> %b, i32 6, i32 1)
65+
%add = fadd <6 x double> %a.t, %b.t
66+
ret <6 x double> %add
67+
}
68+
69+
define <9 x double> @lift_through_multiply(<6 x double> %a, <6 x double> %b) {
70+
; CHECK-LABEL: define <9 x double> @lift_through_multiply(<6 x double> %a, <6 x double> %b) {
71+
; CHECK-NEXT: entry:
72+
; CHECK-NEXT: [[MUL:%.+]] = call <9 x double> @llvm.matrix.multiply.v9f64.v6f64.v6f64(<6 x double> %b, <6 x double> %a, i32 3, i32 2, i32 3)
73+
; CHECK-NEXT: [[T:%.+]] = call <9 x double> @llvm.matrix.transpose.v9f64(<9 x double> [[MUL]], i32 3, i32 3)
74+
; CHECK-NEXT: ret <9 x double> [[T]]
75+
;
76+
entry:
77+
%a.t = call <6 x double> @llvm.matrix.transpose.v6f64(<6 x double> %a, i32 3, i32 2)
78+
%b.t = call <6 x double> @llvm.matrix.transpose.v6f64(<6 x double> %b, i32 2, i32 3)
79+
%mul = call <9 x double> @llvm.matrix.multiply.v9f64.v6f64(<6 x double> %a.t, <6 x double> %b.t, i32 3, i32 2 , i32 3)
80+
ret <9 x double> %mul
81+
}
82+
83+
define <6 x double> @lift_through_multiply_2(<6 x double> %a, <4 x double> %b) {
84+
; CHECK-LABEL: define <6 x double> @lift_through_multiply_2(<6 x double> %a, <4 x double> %b) {
85+
; CHECK-NEXT: entry:
86+
; CHECK-NEXT: [[MUL:%.+]] = call <6 x double> @llvm.matrix.multiply.v6f64.v4f64.v6f64(<4 x double> %b, <6 x double> %a, i32 2, i32 2, i32 3)
87+
; CHECK-NEXT: [[T:%.+]] = call <6 x double> @llvm.matrix.transpose.v6f64(<6 x double> [[MUL]], i32 2, i32 3)
88+
; CHECK-NEXT: ret <6 x double> [[T]]
89+
;
90+
entry:
91+
%a.t = call <6 x double> @llvm.matrix.transpose.v6f64(<6 x double> %a, i32 3, i32 2)
92+
%b.t = call <4 x double> @llvm.matrix.transpose.v4f64(<4 x double> %b, i32 2, i32 2)
93+
%mul = call <6 x double> @llvm.matrix.multiply.v6f64.v6f64.v4f64(<6 x double> %a.t, <4 x double> %b.t, i32 3, i32 2 , i32 2)
94+
ret <6 x double> %mul
95+
}
96+
97+
define <6 x double> @lift_through_multiply_3(<4 x double> %a, <6 x double> %b) {
98+
; CHECK-LABEL: define <6 x double> @lift_through_multiply_3(<4 x double> %a, <6 x double> %b) {
99+
; CHECK-NEXT: entry:
100+
; CHECK-NEXT: [[MUL:%.+]] = call <6 x double> @llvm.matrix.multiply.v6f64.v6f64.v4f64(<6 x double> %b, <4 x double> %a, i32 3, i32 2, i32 2)
101+
; CHECK-NEXT: [[T:%.+]] = call <6 x double> @llvm.matrix.transpose.v6f64(<6 x double> [[MUL]], i32 3, i32 2)
102+
; CHECK-NEXT: ret <6 x double> [[T]]
103+
;
104+
entry:
105+
%a.t = call <4 x double> @llvm.matrix.transpose.v4f64(<4 x double> %a, i32 2, i32 2)
106+
%b.t = call <6 x double> @llvm.matrix.transpose.v6f64(<6 x double> %b, i32 2, i32 3)
107+
%mul = call <6 x double> @llvm.matrix.multiply.v6f64.v4f64.v6f64(<4 x double> %a.t, <6 x double> %b.t, i32 2, i32 2 , i32 3)
108+
ret <6 x double> %mul
109+
}
110+
111+
declare <6 x double> @llvm.matrix.transpose.v6f64.v6f64(<6 x double>, i32, i32)
112+
declare <4 x double> @llvm.matrix.transpose.v4f64.v4f64(<4 x double>, i32, i32)
113+
declare <9 x double> @llvm.matrix.multiply.v9f64.v6f64(<6 x double>, <6 x double>, i32, i32, i32)
114+
declare <6 x double> @llvm.matrix.multiply.v6f64.v6f64.v4f64(<6 x double>, <4 x double>, i32, i32, i32)
115+
declare <6 x double> @llvm.matrix.multiply.v6f64.v6f64.v6f64(<6 x double>, <4 x double>, i32, i32, i32)

0 commit comments

Comments
 (0)