|
121
|
1 ; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN -check-prefix=SI %s
|
|
|
2 ; RUN: llc -march=amdgcn -mcpu=fiji -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN -check-prefix=GFX89 -check-prefix=VI %s
|
|
134
|
3 ; RUN: llc -march=amdgcn -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN -check-prefix=GFX89 -check-prefix=GFX9 %s
|
|
121
|
4
|
|
|
5 ; GCN-LABEL: {{^}}s_cvt_pkrtz_v2f16_f32:
|
|
|
6 ; GCN-DAG: s_load_dword [[X:s[0-9]+]], s[0:1], 0x{{b|2c}}
|
|
|
7 ; GCN-DAG: s_load_dword [[SY:s[0-9]+]], s[0:1], 0x{{c|30}}
|
|
|
8 ; GCN: v_mov_b32_e32 [[VY:v[0-9]+]], [[SY]]
|
|
|
9 ; SI: v_cvt_pkrtz_f16_f32_e32 v{{[0-9]+}}, [[X]], [[VY]]
|
|
|
10 ; GFX89: v_cvt_pkrtz_f16_f32 v{{[0-9]+}}, [[X]], [[VY]]
|
|
|
11 define amdgpu_kernel void @s_cvt_pkrtz_v2f16_f32(<2 x half> addrspace(1)* %out, float %x, float %y) #0 {
|
|
|
12 %result = call <2 x half> @llvm.amdgcn.cvt.pkrtz(float %x, float %y)
|
|
|
13 store <2 x half> %result, <2 x half> addrspace(1)* %out
|
|
|
14 ret void
|
|
|
15 }
|
|
|
16
|
|
|
17 ; GCN-LABEL: {{^}}s_cvt_pkrtz_samereg_v2f16_f32:
|
|
|
18 ; GCN: s_load_dword [[X:s[0-9]+]]
|
|
|
19 ; GCN: v_cvt_pkrtz_f16_f32{{(_e64)*}} v{{[0-9]+}}, [[X]], [[X]]
|
|
|
20 define amdgpu_kernel void @s_cvt_pkrtz_samereg_v2f16_f32(<2 x half> addrspace(1)* %out, float %x) #0 {
|
|
|
21 %result = call <2 x half> @llvm.amdgcn.cvt.pkrtz(float %x, float %x)
|
|
|
22 store <2 x half> %result, <2 x half> addrspace(1)* %out
|
|
|
23 ret void
|
|
|
24 }
|
|
|
25
|
|
|
26 ; FIXME: Folds to 0 on gfx9
|
|
|
27 ; GCN-LABEL: {{^}}s_cvt_pkrtz_undef_undef:
|
|
134
|
28 ; GCN-NEXT: ; %bb.0
|
|
121
|
29 ; SI-NEXT: s_endpgm
|
|
|
30 ; VI-NEXT: s_endpgm
|
|
|
31 ; GFX9: v_mov_b32_e32 v{{[0-9]+}}, 0{{$}}
|
|
|
32 define amdgpu_kernel void @s_cvt_pkrtz_undef_undef(<2 x half> addrspace(1)* %out) #0 {
|
|
|
33 %result = call <2 x half> @llvm.amdgcn.cvt.pkrtz(float undef, float undef)
|
|
|
34 store <2 x half> %result, <2 x half> addrspace(1)* %out
|
|
|
35 ret void
|
|
|
36 }
|
|
|
37
|
|
|
38 ; GCN-LABEL: {{^}}v_cvt_pkrtz_v2f16_f32:
|
|
|
39 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]]
|
|
|
40 ; GCN: {{buffer|flat|global}}_load_dword [[B:v[0-9]+]]
|
|
|
41 ; SI: v_cvt_pkrtz_f16_f32_e32 v{{[0-9]+}}, [[A]], [[B]]
|
|
|
42 ; GFX89: v_cvt_pkrtz_f16_f32 v{{[0-9]+}}, [[A]], [[B]]
|
|
|
43 define amdgpu_kernel void @v_cvt_pkrtz_v2f16_f32(<2 x half> addrspace(1)* %out, float addrspace(1)* %a.ptr, float addrspace(1)* %b.ptr) #0 {
|
|
|
44 %tid = call i32 @llvm.amdgcn.workitem.id.x()
|
|
|
45 %tid.ext = sext i32 %tid to i64
|
|
|
46 %a.gep = getelementptr inbounds float, float addrspace(1)* %a.ptr, i64 %tid.ext
|
|
|
47 %b.gep = getelementptr inbounds float, float addrspace(1)* %b.ptr, i64 %tid.ext
|
|
|
48 %out.gep = getelementptr inbounds <2 x half>, <2 x half> addrspace(1)* %out, i64 %tid.ext
|
|
|
49 %a = load volatile float, float addrspace(1)* %a.gep
|
|
|
50 %b = load volatile float, float addrspace(1)* %b.gep
|
|
|
51 %cvt = call <2 x half> @llvm.amdgcn.cvt.pkrtz(float %a, float %b)
|
|
|
52 store <2 x half> %cvt, <2 x half> addrspace(1)* %out.gep
|
|
|
53 ret void
|
|
|
54 }
|
|
|
55
|
|
|
56 ; GCN-LABEL: {{^}}v_cvt_pkrtz_v2f16_f32_reg_imm:
|
|
|
57 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]]
|
|
|
58 ; GCN: v_cvt_pkrtz_f16_f32{{(_e64)*}} v{{[0-9]+}}, [[A]], 1.0
|
|
|
59 define amdgpu_kernel void @v_cvt_pkrtz_v2f16_f32_reg_imm(<2 x half> addrspace(1)* %out, float addrspace(1)* %a.ptr) #0 {
|
|
|
60 %tid = call i32 @llvm.amdgcn.workitem.id.x()
|
|
|
61 %tid.ext = sext i32 %tid to i64
|
|
|
62 %a.gep = getelementptr inbounds float, float addrspace(1)* %a.ptr, i64 %tid.ext
|
|
|
63 %out.gep = getelementptr inbounds <2 x half>, <2 x half> addrspace(1)* %out, i64 %tid.ext
|
|
|
64 %a = load volatile float, float addrspace(1)* %a.gep
|
|
|
65 %cvt = call <2 x half> @llvm.amdgcn.cvt.pkrtz(float %a, float 1.0)
|
|
|
66 store <2 x half> %cvt, <2 x half> addrspace(1)* %out.gep
|
|
|
67 ret void
|
|
|
68 }
|
|
|
69
|
|
|
70 ; GCN-LABEL: {{^}}v_cvt_pkrtz_v2f16_f32_imm_reg:
|
|
|
71 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]]
|
|
|
72 ; SI: v_cvt_pkrtz_f16_f32_e32 v{{[0-9]+}}, 1.0, [[A]]
|
|
|
73 ; GFX89: v_cvt_pkrtz_f16_f32 v{{[0-9]+}}, 1.0, [[A]]
|
|
|
74 define amdgpu_kernel void @v_cvt_pkrtz_v2f16_f32_imm_reg(<2 x half> addrspace(1)* %out, float addrspace(1)* %a.ptr) #0 {
|
|
|
75 %tid = call i32 @llvm.amdgcn.workitem.id.x()
|
|
|
76 %tid.ext = sext i32 %tid to i64
|
|
|
77 %a.gep = getelementptr inbounds float, float addrspace(1)* %a.ptr, i64 %tid.ext
|
|
|
78 %out.gep = getelementptr inbounds <2 x half>, <2 x half> addrspace(1)* %out, i64 %tid.ext
|
|
|
79 %a = load volatile float, float addrspace(1)* %a.gep
|
|
|
80 %cvt = call <2 x half> @llvm.amdgcn.cvt.pkrtz(float 1.0, float %a)
|
|
|
81 store <2 x half> %cvt, <2 x half> addrspace(1)* %out.gep
|
|
|
82 ret void
|
|
|
83 }
|
|
|
84
|
|
|
85 ; GCN-LABEL: {{^}}v_cvt_pkrtz_v2f16_f32_fneg_lo:
|
|
|
86 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]]
|
|
|
87 ; GCN: {{buffer|flat|global}}_load_dword [[B:v[0-9]+]]
|
|
|
88 ; GCN: v_cvt_pkrtz_f16_f32{{(_e64)*}} v{{[0-9]+}}, -[[A]], [[B]]
|
|
|
89 define amdgpu_kernel void @v_cvt_pkrtz_v2f16_f32_fneg_lo(<2 x half> addrspace(1)* %out, float addrspace(1)* %a.ptr, float addrspace(1)* %b.ptr) #0 {
|
|
|
90 %tid = call i32 @llvm.amdgcn.workitem.id.x()
|
|
|
91 %tid.ext = sext i32 %tid to i64
|
|
|
92 %a.gep = getelementptr inbounds float, float addrspace(1)* %a.ptr, i64 %tid.ext
|
|
|
93 %b.gep = getelementptr inbounds float, float addrspace(1)* %b.ptr, i64 %tid.ext
|
|
|
94 %out.gep = getelementptr inbounds <2 x half>, <2 x half> addrspace(1)* %out, i64 %tid.ext
|
|
|
95 %a = load volatile float, float addrspace(1)* %a.gep
|
|
|
96 %b = load volatile float, float addrspace(1)* %b.gep
|
|
|
97 %neg.a = fsub float -0.0, %a
|
|
|
98 %cvt = call <2 x half> @llvm.amdgcn.cvt.pkrtz(float %neg.a, float %b)
|
|
|
99 store <2 x half> %cvt, <2 x half> addrspace(1)* %out.gep
|
|
|
100 ret void
|
|
|
101 }
|
|
|
102
|
|
|
103 ; GCN-LABEL: {{^}}v_cvt_pkrtz_v2f16_f32_fneg_hi:
|
|
|
104 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]]
|
|
|
105 ; GCN: {{buffer|flat|global}}_load_dword [[B:v[0-9]+]]
|
|
|
106 ; GCN: v_cvt_pkrtz_f16_f32{{(_e64)*}} v{{[0-9]+}}, [[A]], -[[B]]
|
|
|
107 define amdgpu_kernel void @v_cvt_pkrtz_v2f16_f32_fneg_hi(<2 x half> addrspace(1)* %out, float addrspace(1)* %a.ptr, float addrspace(1)* %b.ptr) #0 {
|
|
|
108 %tid = call i32 @llvm.amdgcn.workitem.id.x()
|
|
|
109 %tid.ext = sext i32 %tid to i64
|
|
|
110 %a.gep = getelementptr inbounds float, float addrspace(1)* %a.ptr, i64 %tid.ext
|
|
|
111 %b.gep = getelementptr inbounds float, float addrspace(1)* %b.ptr, i64 %tid.ext
|
|
|
112 %out.gep = getelementptr inbounds <2 x half>, <2 x half> addrspace(1)* %out, i64 %tid.ext
|
|
|
113 %a = load volatile float, float addrspace(1)* %a.gep
|
|
|
114 %b = load volatile float, float addrspace(1)* %b.gep
|
|
|
115 %neg.b = fsub float -0.0, %b
|
|
|
116 %cvt = call <2 x half> @llvm.amdgcn.cvt.pkrtz(float %a, float %neg.b)
|
|
|
117 store <2 x half> %cvt, <2 x half> addrspace(1)* %out.gep
|
|
|
118 ret void
|
|
|
119 }
|
|
|
120
|
|
|
121 ; GCN-LABEL: {{^}}v_cvt_pkrtz_v2f16_f32_fneg_lo_hi:
|
|
|
122 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]]
|
|
|
123 ; GCN: {{buffer|flat|global}}_load_dword [[B:v[0-9]+]]
|
|
|
124 ; GCN: v_cvt_pkrtz_f16_f32{{(_e64)*}} v{{[0-9]+}}, -[[A]], -[[B]]
|
|
|
125 define amdgpu_kernel void @v_cvt_pkrtz_v2f16_f32_fneg_lo_hi(<2 x half> addrspace(1)* %out, float addrspace(1)* %a.ptr, float addrspace(1)* %b.ptr) #0 {
|
|
|
126 %tid = call i32 @llvm.amdgcn.workitem.id.x()
|
|
|
127 %tid.ext = sext i32 %tid to i64
|
|
|
128 %a.gep = getelementptr inbounds float, float addrspace(1)* %a.ptr, i64 %tid.ext
|
|
|
129 %b.gep = getelementptr inbounds float, float addrspace(1)* %b.ptr, i64 %tid.ext
|
|
|
130 %out.gep = getelementptr inbounds <2 x half>, <2 x half> addrspace(1)* %out, i64 %tid.ext
|
|
|
131 %a = load volatile float, float addrspace(1)* %a.gep
|
|
|
132 %b = load volatile float, float addrspace(1)* %b.gep
|
|
|
133 %neg.a = fsub float -0.0, %a
|
|
|
134 %neg.b = fsub float -0.0, %b
|
|
|
135 %cvt = call <2 x half> @llvm.amdgcn.cvt.pkrtz(float %neg.a, float %neg.b)
|
|
|
136 store <2 x half> %cvt, <2 x half> addrspace(1)* %out.gep
|
|
|
137 ret void
|
|
|
138 }
|
|
|
139
|
|
|
140 ; GCN-LABEL: {{^}}v_cvt_pkrtz_v2f16_f32_fneg_fabs_lo_fneg_hi:
|
|
|
141 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]]
|
|
|
142 ; GCN: {{buffer|flat|global}}_load_dword [[B:v[0-9]+]]
|
|
|
143 ; GCN: v_cvt_pkrtz_f16_f32{{(_e64)*}} v{{[0-9]+}}, -|[[A]]|, -[[B]]
|
|
|
144 define amdgpu_kernel void @v_cvt_pkrtz_v2f16_f32_fneg_fabs_lo_fneg_hi(<2 x half> addrspace(1)* %out, float addrspace(1)* %a.ptr, float addrspace(1)* %b.ptr) #0 {
|
|
|
145 %tid = call i32 @llvm.amdgcn.workitem.id.x()
|
|
|
146 %tid.ext = sext i32 %tid to i64
|
|
|
147 %a.gep = getelementptr inbounds float, float addrspace(1)* %a.ptr, i64 %tid.ext
|
|
|
148 %b.gep = getelementptr inbounds float, float addrspace(1)* %b.ptr, i64 %tid.ext
|
|
|
149 %out.gep = getelementptr inbounds <2 x half>, <2 x half> addrspace(1)* %out, i64 %tid.ext
|
|
|
150 %a = load volatile float, float addrspace(1)* %a.gep
|
|
|
151 %b = load volatile float, float addrspace(1)* %b.gep
|
|
|
152 %fabs.a = call float @llvm.fabs.f32(float %a)
|
|
|
153 %neg.fabs.a = fsub float -0.0, %fabs.a
|
|
|
154 %neg.b = fsub float -0.0, %b
|
|
|
155 %cvt = call <2 x half> @llvm.amdgcn.cvt.pkrtz(float %neg.fabs.a, float %neg.b)
|
|
|
156 store <2 x half> %cvt, <2 x half> addrspace(1)* %out.gep
|
|
|
157 ret void
|
|
|
158 }
|
|
|
159
|
|
|
160 declare <2 x half> @llvm.amdgcn.cvt.pkrtz(float, float) #1
|
|
|
161 declare float @llvm.fabs.f32(float) #1
|
|
|
162 declare i32 @llvm.amdgcn.workitem.id.x() #1
|
|
|
163
|
|
|
164
|
|
|
165 attributes #0 = { nounwind }
|
|
|
166 attributes #1 = { nounwind readnone }
|
