1 ; RUN: llc -march=amdgcn -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GFX9_10,GFX9 %s
2 ; RUN: llc -march=amdgcn -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GFX9_10,GFX10 %s
4 ; GCN-LABEL: {{^}}test_pk_max_f16_literal_0_1:
5 ; GFX9_10: v_pk_max_f16 v{{[0-9]+}}, v{{[0-9]+}}, 1.0 op_sel:[0,1] op_sel_hi:[1,0]{{$}}
6 define amdgpu_kernel void @test_pk_max_f16_literal_0_1(<2 x half> addrspace(1)* nocapture %arg) {
8 %tmp = tail call i32 @llvm.amdgcn.workitem.id.x()
9 %tmp1 = zext i32 %tmp to i64
10 %tmp2 = getelementptr inbounds <2 x half>, <2 x half> addrspace(1)* %arg, i64 %tmp1
11 %tmp3 = load <2 x half>, <2 x half> addrspace(1)* %tmp2, align 4
12 %tmp4 = tail call <2 x half> @llvm.maxnum.v2f16(<2 x half> %tmp3, <2 x half> <half 0xH0000, half 0xH3C00>)
13 store <2 x half> %tmp4, <2 x half> addrspace(1)* %tmp2, align 4
17 ; GCN-LABEL: {{^}}test_pk_max_f16_literal_1_0:
18 ; GFX9_10: v_pk_max_f16 v{{[0-9]+}}, v{{[0-9]+}}, 1.0{{$}}
19 define amdgpu_kernel void @test_pk_max_f16_literal_1_0(<2 x half> addrspace(1)* nocapture %arg) {
21 %tmp = tail call i32 @llvm.amdgcn.workitem.id.x()
22 %tmp1 = zext i32 %tmp to i64
23 %tmp2 = getelementptr inbounds <2 x half>, <2 x half> addrspace(1)* %arg, i64 %tmp1
24 %tmp3 = load <2 x half>, <2 x half> addrspace(1)* %tmp2, align 4
25 %tmp4 = tail call <2 x half> @llvm.maxnum.v2f16(<2 x half> %tmp3, <2 x half> <half 0xH3C00, half 0xH0000>)
26 store <2 x half> %tmp4, <2 x half> addrspace(1)* %tmp2, align 4
30 ; GCN-LABEL: {{^}}test_pk_max_f16_literal_1_1:
31 ; GFX9_10: v_pk_max_f16 v{{[0-9]+}}, v{{[0-9]+}}, 1.0 op_sel_hi:[1,0]{{$}}
32 define amdgpu_kernel void @test_pk_max_f16_literal_1_1(<2 x half> addrspace(1)* nocapture %arg) {
34 %tmp = tail call i32 @llvm.amdgcn.workitem.id.x()
35 %tmp1 = zext i32 %tmp to i64
36 %tmp2 = getelementptr inbounds <2 x half>, <2 x half> addrspace(1)* %arg, i64 %tmp1
37 %tmp3 = load <2 x half>, <2 x half> addrspace(1)* %tmp2, align 4
38 %tmp4 = tail call <2 x half> @llvm.maxnum.v2f16(<2 x half> %tmp3, <2 x half> <half 0xH3C00, half 0xH3C00>)
39 store <2 x half> %tmp4, <2 x half> addrspace(1)* %tmp2, align 4
43 ; GCN-LABEL: {{^}}test_pk_max_f16_literal_0_m1:
44 ; GFX9_10: v_pk_max_f16 v{{[0-9]+}}, v{{[0-9]+}}, -1.0 op_sel:[0,1] op_sel_hi:[1,0]{{$}}
45 define amdgpu_kernel void @test_pk_max_f16_literal_0_m1(<2 x half> addrspace(1)* nocapture %arg) {
47 %tmp = tail call i32 @llvm.amdgcn.workitem.id.x()
48 %tmp1 = zext i32 %tmp to i64
49 %tmp2 = getelementptr inbounds <2 x half>, <2 x half> addrspace(1)* %arg, i64 %tmp1
50 %tmp3 = load <2 x half>, <2 x half> addrspace(1)* %tmp2, align 4
51 %tmp4 = tail call <2 x half> @llvm.maxnum.v2f16(<2 x half> %tmp3, <2 x half> <half 0xH0000, half 0xHBC00>)
52 store <2 x half> %tmp4, <2 x half> addrspace(1)* %tmp2, align 4
56 ; GCN-LABEL: {{^}}test_pk_max_f16_literal_m1_0:
57 ; GFX9_10: v_pk_max_f16 v{{[0-9]+}}, v{{[0-9]+}}, -1.0{{$}}
58 define amdgpu_kernel void @test_pk_max_f16_literal_m1_0(<2 x half> addrspace(1)* nocapture %arg) {
60 %tmp = tail call i32 @llvm.amdgcn.workitem.id.x()
61 %tmp1 = zext i32 %tmp to i64
62 %tmp2 = getelementptr inbounds <2 x half>, <2 x half> addrspace(1)* %arg, i64 %tmp1
63 %tmp3 = load <2 x half>, <2 x half> addrspace(1)* %tmp2, align 4
64 %tmp4 = tail call <2 x half> @llvm.maxnum.v2f16(<2 x half> %tmp3, <2 x half> <half 0xHBC00, half 0xH0000>)
65 store <2 x half> %tmp4, <2 x half> addrspace(1)* %tmp2, align 4
69 ; GCN-LABEL: {{^}}test_pk_max_f16_literal_m1_m1:
70 ; GFX9_10: v_pk_max_f16 v{{[0-9]+}}, v{{[0-9]+}}, -1.0 op_sel_hi:[1,0]{{$}}
71 define amdgpu_kernel void @test_pk_max_f16_literal_m1_m1(<2 x half> addrspace(1)* nocapture %arg) {
73 %tmp = tail call i32 @llvm.amdgcn.workitem.id.x()
74 %tmp1 = zext i32 %tmp to i64
75 %tmp2 = getelementptr inbounds <2 x half>, <2 x half> addrspace(1)* %arg, i64 %tmp1
76 %tmp3 = load <2 x half>, <2 x half> addrspace(1)* %tmp2, align 4
77 %tmp4 = tail call <2 x half> @llvm.maxnum.v2f16(<2 x half> %tmp3, <2 x half> <half 0xHBC00, half 0xHBC00>)
78 store <2 x half> %tmp4, <2 x half> addrspace(1)* %tmp2, align 4
82 ; GCN-LABEL: {{^}}test_pk_max_f16_literal_0_0:
83 ; GFX9_10: v_pk_max_f16 v{{[0-9]+}}, v{{[0-9]+}}, 0{{$}}
84 define amdgpu_kernel void @test_pk_max_f16_literal_0_0(<2 x half> addrspace(1)* nocapture %arg) {
86 %tmp = tail call i32 @llvm.amdgcn.workitem.id.x()
87 %tmp1 = zext i32 %tmp to i64
88 %tmp2 = getelementptr inbounds <2 x half>, <2 x half> addrspace(1)* %arg, i64 %tmp1
89 %tmp3 = load <2 x half>, <2 x half> addrspace(1)* %tmp2, align 4
90 %tmp4 = tail call <2 x half> @llvm.maxnum.v2f16(<2 x half> %tmp3, <2 x half> <half 0xH0000, half 0xH0000>)
91 store <2 x half> %tmp4, <2 x half> addrspace(1)* %tmp2, align 4
95 ; GCN-LABEL: {{^}}test_pk_max_f16_literal_0_41c8:
96 ; GFX9: s_mov_b32 [[C:s[0-9]+]], 0x41c80000
97 ; GFX9: v_pk_max_f16 v{{[0-9]+}}, v{{[0-9]+}}, [[C]]{{$}}
98 ; GFX10: v_pk_max_f16 v{{[0-9]+}}, 0x41c8, v{{[0-9]+}} op_sel:[1,0] op_sel_hi:[0,1]{{$}}
99 define amdgpu_kernel void @test_pk_max_f16_literal_0_41c8(<2 x half> addrspace(1)* nocapture %arg) {
101 %tmp = tail call i32 @llvm.amdgcn.workitem.id.x()
102 %tmp1 = zext i32 %tmp to i64
103 %tmp2 = getelementptr inbounds <2 x half>, <2 x half> addrspace(1)* %arg, i64 %tmp1
104 %tmp3 = load <2 x half>, <2 x half> addrspace(1)* %tmp2, align 4
105 %tmp4 = tail call <2 x half> @llvm.maxnum.v2f16(<2 x half> %tmp3, <2 x half> <half 0xH0000, half 0xH41C8>)
106 store <2 x half> %tmp4, <2 x half> addrspace(1)* %tmp2, align 4
110 ; GCN-LABEL: {{^}}test_pk_max_f16_literal_41c8_0:
111 ; GFX9: s_movk_i32 [[C:s[0-9]+]], 0x41c8
112 ; GFX9: v_pk_max_f16 v{{[0-9]+}}, v{{[0-9]+}}, [[C]]{{$}}
113 ; GFX10: v_pk_max_f16 v{{[0-9]+}}, 0x41c8, v{{[0-9]+}}{{$}}
114 define amdgpu_kernel void @test_pk_max_f16_literal_41c8_0(<2 x half> addrspace(1)* nocapture %arg) {
116 %tmp = tail call i32 @llvm.amdgcn.workitem.id.x()
117 %tmp1 = zext i32 %tmp to i64
118 %tmp2 = getelementptr inbounds <2 x half>, <2 x half> addrspace(1)* %arg, i64 %tmp1
119 %tmp3 = load <2 x half>, <2 x half> addrspace(1)* %tmp2, align 4
120 %tmp4 = tail call <2 x half> @llvm.maxnum.v2f16(<2 x half> %tmp3, <2 x half> <half 0xH41C8, half 0xH0>)
121 store <2 x half> %tmp4, <2 x half> addrspace(1)* %tmp2, align 4
125 ; GCN-LABEL: {{^}}test_pk_max_f16_literal_42ca_41c8:
126 ; GFX9: s_mov_b32 [[C:s[0-9]+]], 0x41c842ca
127 ; GFX9: v_pk_max_f16 v{{[0-9]+}}, v{{[0-9]+}}, [[C]]{{$}}
128 ; GFX10: v_pk_max_f16 v{{[0-9]+}}, 0x41c842ca, v{{[0-9]+}}{{$}}
129 define amdgpu_kernel void @test_pk_max_f16_literal_42ca_41c8(<2 x half> addrspace(1)* nocapture %arg) {
131 %tmp = tail call i32 @llvm.amdgcn.workitem.id.x()
132 %tmp1 = zext i32 %tmp to i64
133 %tmp2 = getelementptr inbounds <2 x half>, <2 x half> addrspace(1)* %arg, i64 %tmp1
134 %tmp3 = load <2 x half>, <2 x half> addrspace(1)* %tmp2, align 4
135 %tmp4 = tail call <2 x half> @llvm.maxnum.v2f16(<2 x half> %tmp3, <2 x half> <half 0xH42CA, half 0xH41C8>)
136 store <2 x half> %tmp4, <2 x half> addrspace(1)* %tmp2, align 4
140 declare <2 x half> @llvm.maxnum.v2f16(<2 x half>, <2 x half>)
141 declare i32 @llvm.amdgcn.workitem.id.x()