xref: /aosp_15_r20/external/llvm/test/CodeGen/NVPTX/shfl.ll (revision 9880d6810fe72a1726cb53787c6711e909410d58)
1*9880d681SAndroid Build Coastguard Worker; RUN: llc < %s -march=nvptx64 -mcpu=sm_30 -disable-nvptx-favor-non-generic | FileCheck %s
2*9880d681SAndroid Build Coastguard Worker
3*9880d681SAndroid Build Coastguard Workerdeclare i32 @llvm.nvvm.shfl.down.i32(i32, i32, i32)
4*9880d681SAndroid Build Coastguard Workerdeclare float @llvm.nvvm.shfl.down.f32(float, i32, i32)
5*9880d681SAndroid Build Coastguard Workerdeclare i32 @llvm.nvvm.shfl.up.i32(i32, i32, i32)
6*9880d681SAndroid Build Coastguard Workerdeclare float @llvm.nvvm.shfl.up.f32(float, i32, i32)
7*9880d681SAndroid Build Coastguard Workerdeclare i32 @llvm.nvvm.shfl.bfly.i32(i32, i32, i32)
8*9880d681SAndroid Build Coastguard Workerdeclare float @llvm.nvvm.shfl.bfly.f32(float, i32, i32)
9*9880d681SAndroid Build Coastguard Workerdeclare i32 @llvm.nvvm.shfl.idx.i32(i32, i32, i32)
10*9880d681SAndroid Build Coastguard Workerdeclare float @llvm.nvvm.shfl.idx.f32(float, i32, i32)
11*9880d681SAndroid Build Coastguard Worker
12*9880d681SAndroid Build Coastguard Worker; Try all four permutations of register and immediate parameters with
13*9880d681SAndroid Build Coastguard Worker; shfl.down.
14*9880d681SAndroid Build Coastguard Worker
15*9880d681SAndroid Build Coastguard Worker; CHECK-LABEL: .func{{.*}}shfl.down1
16*9880d681SAndroid Build Coastguard Workerdefine i32 @shfl.down1(i32 %in) {
17*9880d681SAndroid Build Coastguard Worker  ; CHECK: ld.param.u32 [[IN:%r[0-9]+]]
18*9880d681SAndroid Build Coastguard Worker  ; CHECK: shfl.down.b32 [[OUT:%r[0-9]+]], [[IN]], 1, 2;
19*9880d681SAndroid Build Coastguard Worker  ; CHECK: st.param.{{.}}32 {{.*}}, [[OUT]]
20*9880d681SAndroid Build Coastguard Worker  %val = call i32 @llvm.nvvm.shfl.down.i32(i32 %in, i32 1, i32 2)
21*9880d681SAndroid Build Coastguard Worker  ret i32 %val
22*9880d681SAndroid Build Coastguard Worker}
23*9880d681SAndroid Build Coastguard Worker
24*9880d681SAndroid Build Coastguard Worker; CHECK-LABEL: .func{{.*}}shfl.down2
25*9880d681SAndroid Build Coastguard Workerdefine i32 @shfl.down2(i32 %in, i32 %width) {
26*9880d681SAndroid Build Coastguard Worker  ; CHECK: ld.param.u32 [[IN1:%r[0-9]+]]
27*9880d681SAndroid Build Coastguard Worker  ; CHECK: ld.param.u32 [[IN2:%r[0-9]+]]
28*9880d681SAndroid Build Coastguard Worker  ; CHECK: shfl.down.{{.}}32 %r{{[0-9]+}}, [[IN1]], [[IN2]], 3;
29*9880d681SAndroid Build Coastguard Worker  %val = call i32 @llvm.nvvm.shfl.down.i32(i32 %in, i32 %width, i32 3)
30*9880d681SAndroid Build Coastguard Worker  ret i32 %val
31*9880d681SAndroid Build Coastguard Worker}
32*9880d681SAndroid Build Coastguard Worker
33*9880d681SAndroid Build Coastguard Worker; CHECK-LABEL: .func{{.*}}shfl.down3
34*9880d681SAndroid Build Coastguard Workerdefine i32 @shfl.down3(i32 %in, i32 %mask) {
35*9880d681SAndroid Build Coastguard Worker  ; CHECK: ld.param.u32 [[IN1:%r[0-9]+]]
36*9880d681SAndroid Build Coastguard Worker  ; CHECK: ld.param.u32 [[IN2:%r[0-9]+]]
37*9880d681SAndroid Build Coastguard Worker  ; CHECK: shfl.down.{{.}}32 %r{{[0-9]+}}, [[IN1]], 4, [[IN2]];
38*9880d681SAndroid Build Coastguard Worker  %val = call i32 @llvm.nvvm.shfl.down.i32(i32 %in, i32 4, i32 %mask)
39*9880d681SAndroid Build Coastguard Worker  ret i32 %val
40*9880d681SAndroid Build Coastguard Worker}
41*9880d681SAndroid Build Coastguard Worker
42*9880d681SAndroid Build Coastguard Worker; CHECK-LABEL: .func{{.*}}shfl.down4
43*9880d681SAndroid Build Coastguard Workerdefine i32 @shfl.down4(i32 %in, i32 %width, i32 %mask) {
44*9880d681SAndroid Build Coastguard Worker  ; CHECK: ld.param.u32 [[IN1:%r[0-9]+]]
45*9880d681SAndroid Build Coastguard Worker  ; CHECK: ld.param.u32 [[IN2:%r[0-9]+]]
46*9880d681SAndroid Build Coastguard Worker  ; CHECK: ld.param.u32 [[IN3:%r[0-9]+]]
47*9880d681SAndroid Build Coastguard Worker  ; CHECK: shfl.down.{{.}}32 %r{{[0-9]+}}, [[IN1]], [[IN2]], [[IN3]];
48*9880d681SAndroid Build Coastguard Worker  %val = call i32 @llvm.nvvm.shfl.down.i32(i32 %in, i32 %width, i32 %mask)
49*9880d681SAndroid Build Coastguard Worker  ret i32 %val
50*9880d681SAndroid Build Coastguard Worker}
51*9880d681SAndroid Build Coastguard Worker
52*9880d681SAndroid Build Coastguard Worker; Try shfl.down with floating-point params.
53*9880d681SAndroid Build Coastguard Worker; CHECK-LABEL: .func{{.*}}shfl.down.float
54*9880d681SAndroid Build Coastguard Workerdefine float @shfl.down.float(float %in) {
55*9880d681SAndroid Build Coastguard Worker  ; CHECK: ld.param.f32 [[IN:%f[0-9]+]]
56*9880d681SAndroid Build Coastguard Worker  ; CHECK: shfl.down.b32 [[OUT:%f[0-9]+]], [[IN]], 5, 6;
57*9880d681SAndroid Build Coastguard Worker  ; CHECK: st.param.{{.}}32 {{.*}}, [[OUT]]
58*9880d681SAndroid Build Coastguard Worker  %out = call float @llvm.nvvm.shfl.down.f32(float %in, i32 5, i32 6)
59*9880d681SAndroid Build Coastguard Worker  ret float %out
60*9880d681SAndroid Build Coastguard Worker}
61*9880d681SAndroid Build Coastguard Worker
62*9880d681SAndroid Build Coastguard Worker; Try the rest of the shfl modes.  Hopefully they're declared in such a way
63*9880d681SAndroid Build Coastguard Worker; that if shfl.down works correctly, they also work correctly.
64*9880d681SAndroid Build Coastguard Workerdefine void @shfl.rest(i32 %in_i32, float %in_float, i32* %out_i32, float* %out_float) {
65*9880d681SAndroid Build Coastguard Worker  ; CHECK: shfl.up.b32 %r{{[0-9]+}}, %r{{[0-9]+}}, 1, 2;
66*9880d681SAndroid Build Coastguard Worker  %up_i32 = call i32 @llvm.nvvm.shfl.up.i32(i32 %in_i32, i32 1, i32 2)
67*9880d681SAndroid Build Coastguard Worker  store i32 %up_i32, i32* %out_i32
68*9880d681SAndroid Build Coastguard Worker
69*9880d681SAndroid Build Coastguard Worker  ; CHECK: shfl.up.b32 %f{{[0-9]+}}, %f{{[0-9]+}}, 3, 4;
70*9880d681SAndroid Build Coastguard Worker  %up_float = call float @llvm.nvvm.shfl.up.f32(float %in_float, i32 3, i32 4)
71*9880d681SAndroid Build Coastguard Worker  store float %up_float, float* %out_float
72*9880d681SAndroid Build Coastguard Worker
73*9880d681SAndroid Build Coastguard Worker  ; CHECK: shfl.bfly.b32 %r{{[0-9]+}}, %r{{[0-9]+}}, 5, 6;
74*9880d681SAndroid Build Coastguard Worker  %bfly_i32 = call i32 @llvm.nvvm.shfl.bfly.i32(i32 %in_i32, i32 5, i32 6)
75*9880d681SAndroid Build Coastguard Worker  store i32 %bfly_i32, i32* %out_i32
76*9880d681SAndroid Build Coastguard Worker
77*9880d681SAndroid Build Coastguard Worker  ; CHECK: shfl.bfly.b32 %f{{[0-9]+}}, %f{{[0-9]+}}, 7, 8;
78*9880d681SAndroid Build Coastguard Worker  %bfly_float = call float @llvm.nvvm.shfl.bfly.f32(float %in_float, i32 7, i32 8)
79*9880d681SAndroid Build Coastguard Worker  store float %bfly_float, float* %out_float
80*9880d681SAndroid Build Coastguard Worker
81*9880d681SAndroid Build Coastguard Worker  ; CHECK: shfl.idx.b32 %r{{[0-9]+}}, %r{{[0-9]+}}, 9, 10;
82*9880d681SAndroid Build Coastguard Worker  %idx_i32 = call i32 @llvm.nvvm.shfl.idx.i32(i32 %in_i32, i32 9, i32 10)
83*9880d681SAndroid Build Coastguard Worker  store i32 %idx_i32, i32* %out_i32
84*9880d681SAndroid Build Coastguard Worker
85*9880d681SAndroid Build Coastguard Worker  ; CHECK: shfl.idx.b32 %f{{[0-9]+}}, %f{{[0-9]+}}, 11, 12;
86*9880d681SAndroid Build Coastguard Worker  %idx_float = call float @llvm.nvvm.shfl.idx.f32(float %in_float, i32 11, i32 12)
87*9880d681SAndroid Build Coastguard Worker  store float %idx_float, float* %out_float
88*9880d681SAndroid Build Coastguard Worker
89*9880d681SAndroid Build Coastguard Worker  ret void
90*9880d681SAndroid Build Coastguard Worker}
91