1 ; RUN: llc -march=amdgcn -mcpu=tahiti -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN -check-prefix=SI %s
2 ; RUN: llc -march=amdgcn -mcpu=bonaire -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN -check-prefix=SI %s
3 ; RUN: llc -march=amdgcn -mcpu=fiji -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN -check-prefix=VI %s
4 ; RUN: llc -march=amdgcn -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN -check-prefix=VI %s
6 ; GCN-LABEL: {{^}}s_cvt_pknorm_u16_f32:
7 ; GCN-DAG: s_load_dwordx2 s{{\[}}[[SX:[0-9]+]]:[[SY:[0-9]+]]{{\]}}, s[0:1], 0x{{b|2c}}
8 ; GCN: v_mov_b32_e32 [[VY:v[0-9]+]], s[[SY]]
9 ; SI: v_cvt_pknorm_u16_f32_e32 v{{[0-9]+}}, s[[SX]], [[VY]]
10 ; VI: v_cvt_pknorm_u16_f32 v{{[0-9]+}}, s[[SX]], [[VY]]
11 define amdgpu_kernel void @s_cvt_pknorm_u16_f32(i32 addrspace(1)* %out, float %x, float %y) #0 {
12 %result = call <2 x i16> @llvm.amdgcn.cvt.pknorm.u16(float %x, float %y)
13 %r = bitcast <2 x i16> %result to i32
14 store i32 %r, i32 addrspace(1)* %out
18 ; GCN-LABEL: {{^}}s_cvt_pknorm_u16_samereg_f32:
19 ; GCN: s_load_dword [[X:s[0-9]+]]
20 ; GCN: v_cvt_pknorm_u16_f32{{(_e64)*}} v{{[0-9]+}}, [[X]], [[X]]
21 define amdgpu_kernel void @s_cvt_pknorm_u16_samereg_f32(i32 addrspace(1)* %out, float %x) #0 {
22 %result = call <2 x i16> @llvm.amdgcn.cvt.pknorm.u16(float %x, float %x)
23 %r = bitcast <2 x i16> %result to i32
24 store i32 %r, i32 addrspace(1)* %out
28 ; GCN-LABEL: {{^}}v_cvt_pknorm_u16_f32:
29 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]]
30 ; GCN: {{buffer|flat|global}}_load_dword [[B:v[0-9]+]]
31 ; SI: v_cvt_pknorm_u16_f32_e32 v{{[0-9]+}}, [[A]], [[B]]
32 ; VI: v_cvt_pknorm_u16_f32 v{{[0-9]+}}, [[A]], [[B]]
33 define amdgpu_kernel void @v_cvt_pknorm_u16_f32(i32 addrspace(1)* %out, float addrspace(1)* %a.ptr, float addrspace(1)* %b.ptr) #0 {
34 %tid = call i32 @llvm.amdgcn.workitem.id.x()
35 %tid.ext = sext i32 %tid to i64
36 %a.gep = getelementptr inbounds float, float addrspace(1)* %a.ptr, i64 %tid.ext
37 %b.gep = getelementptr inbounds float, float addrspace(1)* %b.ptr, i64 %tid.ext
38 %out.gep = getelementptr inbounds i32, i32 addrspace(1)* %out, i64 %tid.ext
39 %a = load volatile float, float addrspace(1)* %a.gep
40 %b = load volatile float, float addrspace(1)* %b.gep
41 %cvt = call <2 x i16> @llvm.amdgcn.cvt.pknorm.u16(float %a, float %b)
42 %r = bitcast <2 x i16> %cvt to i32
43 store i32 %r, i32 addrspace(1)* %out.gep
47 ; GCN-LABEL: {{^}}v_cvt_pknorm_u16_f32_reg_imm:
48 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]]
49 ; GCN: v_cvt_pknorm_u16_f32{{(_e64)*}} v{{[0-9]+}}, [[A]], 1.0
50 define amdgpu_kernel void @v_cvt_pknorm_u16_f32_reg_imm(i32 addrspace(1)* %out, float addrspace(1)* %a.ptr) #0 {
51 %tid = call i32 @llvm.amdgcn.workitem.id.x()
52 %tid.ext = sext i32 %tid to i64
53 %a.gep = getelementptr inbounds float, float addrspace(1)* %a.ptr, i64 %tid.ext
54 %out.gep = getelementptr inbounds i32, i32 addrspace(1)* %out, i64 %tid.ext
55 %a = load volatile float, float addrspace(1)* %a.gep
56 %cvt = call <2 x i16> @llvm.amdgcn.cvt.pknorm.u16(float %a, float 1.0)
57 %r = bitcast <2 x i16> %cvt to i32
58 store i32 %r, i32 addrspace(1)* %out.gep
62 ; GCN-LABEL: {{^}}v_cvt_pknorm_u16_f32_imm_reg:
63 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]]
64 ; SI: v_cvt_pknorm_u16_f32_e32 v{{[0-9]+}}, 1.0, [[A]]
65 ; VI: v_cvt_pknorm_u16_f32 v{{[0-9]+}}, 1.0, [[A]]
66 define amdgpu_kernel void @v_cvt_pknorm_u16_f32_imm_reg(i32 addrspace(1)* %out, float addrspace(1)* %a.ptr) #0 {
67 %tid = call i32 @llvm.amdgcn.workitem.id.x()
68 %tid.ext = sext i32 %tid to i64
69 %a.gep = getelementptr inbounds float, float addrspace(1)* %a.ptr, i64 %tid.ext
70 %out.gep = getelementptr inbounds i32, i32 addrspace(1)* %out, i64 %tid.ext
71 %a = load volatile float, float addrspace(1)* %a.gep
72 %cvt = call <2 x i16> @llvm.amdgcn.cvt.pknorm.u16(float 1.0, float %a)
73 %r = bitcast <2 x i16> %cvt to i32
74 store i32 %r, i32 addrspace(1)* %out.gep
78 ; GCN-LABEL: {{^}}v_cvt_pknorm_u16_f32_fneg_lo:
79 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]]
80 ; GCN: {{buffer|flat|global}}_load_dword [[B:v[0-9]+]]
81 ; GCN: v_cvt_pknorm_u16_f32{{(_e64)*}} v{{[0-9]+}}, -[[A]], [[B]]
82 define amdgpu_kernel void @v_cvt_pknorm_u16_f32_fneg_lo(i32 addrspace(1)* %out, float addrspace(1)* %a.ptr, float addrspace(1)* %b.ptr) #0 {
83 %tid = call i32 @llvm.amdgcn.workitem.id.x()
84 %tid.ext = sext i32 %tid to i64
85 %a.gep = getelementptr inbounds float, float addrspace(1)* %a.ptr, i64 %tid.ext
86 %b.gep = getelementptr inbounds float, float addrspace(1)* %b.ptr, i64 %tid.ext
87 %out.gep = getelementptr inbounds i32, i32 addrspace(1)* %out, i64 %tid.ext
88 %a = load volatile float, float addrspace(1)* %a.gep
89 %b = load volatile float, float addrspace(1)* %b.gep
90 %neg.a = fsub float -0.0, %a
91 %cvt = call <2 x i16> @llvm.amdgcn.cvt.pknorm.u16(float %neg.a, float %b)
92 %r = bitcast <2 x i16> %cvt to i32
93 store i32 %r, i32 addrspace(1)* %out.gep
97 ; GCN-LABEL: {{^}}v_cvt_pknorm_u16_f32_fneg_hi:
98 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]]
99 ; GCN: {{buffer|flat|global}}_load_dword [[B:v[0-9]+]]
100 ; GCN: v_cvt_pknorm_u16_f32{{(_e64)*}} v{{[0-9]+}}, [[A]], -[[B]]
101 define amdgpu_kernel void @v_cvt_pknorm_u16_f32_fneg_hi(i32 addrspace(1)* %out, float addrspace(1)* %a.ptr, float addrspace(1)* %b.ptr) #0 {
102 %tid = call i32 @llvm.amdgcn.workitem.id.x()
103 %tid.ext = sext i32 %tid to i64
104 %a.gep = getelementptr inbounds float, float addrspace(1)* %a.ptr, i64 %tid.ext
105 %b.gep = getelementptr inbounds float, float addrspace(1)* %b.ptr, i64 %tid.ext
106 %out.gep = getelementptr inbounds i32, i32 addrspace(1)* %out, i64 %tid.ext
107 %a = load volatile float, float addrspace(1)* %a.gep
108 %b = load volatile float, float addrspace(1)* %b.gep
109 %neg.b = fsub float -0.0, %b
110 %cvt = call <2 x i16> @llvm.amdgcn.cvt.pknorm.u16(float %a, float %neg.b)
111 %r = bitcast <2 x i16> %cvt to i32
112 store i32 %r, i32 addrspace(1)* %out.gep
116 ; GCN-LABEL: {{^}}v_cvt_pknorm_u16_f32_fneg_lo_hi:
117 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]]
118 ; GCN: {{buffer|flat|global}}_load_dword [[B:v[0-9]+]]
119 ; GCN: v_cvt_pknorm_u16_f32{{(_e64)*}} v{{[0-9]+}}, -[[A]], -[[B]]
120 define amdgpu_kernel void @v_cvt_pknorm_u16_f32_fneg_lo_hi(i32 addrspace(1)* %out, float addrspace(1)* %a.ptr, float addrspace(1)* %b.ptr) #0 {
121 %tid = call i32 @llvm.amdgcn.workitem.id.x()
122 %tid.ext = sext i32 %tid to i64
123 %a.gep = getelementptr inbounds float, float addrspace(1)* %a.ptr, i64 %tid.ext
124 %b.gep = getelementptr inbounds float, float addrspace(1)* %b.ptr, i64 %tid.ext
125 %out.gep = getelementptr inbounds i32, i32 addrspace(1)* %out, i64 %tid.ext
126 %a = load volatile float, float addrspace(1)* %a.gep
127 %b = load volatile float, float addrspace(1)* %b.gep
128 %neg.a = fsub float -0.0, %a
129 %neg.b = fsub float -0.0, %b
130 %cvt = call <2 x i16> @llvm.amdgcn.cvt.pknorm.u16(float %neg.a, float %neg.b)
131 %r = bitcast <2 x i16> %cvt to i32
132 store i32 %r, i32 addrspace(1)* %out.gep
136 ; GCN-LABEL: {{^}}v_cvt_pknorm_u16_f32_fneg_fabs_lo_fneg_hi:
137 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]]
138 ; GCN: {{buffer|flat|global}}_load_dword [[B:v[0-9]+]]
139 ; GCN: v_cvt_pknorm_u16_f32{{(_e64)*}} v{{[0-9]+}}, -|[[A]]|, -[[B]]
140 define amdgpu_kernel void @v_cvt_pknorm_u16_f32_fneg_fabs_lo_fneg_hi(i32 addrspace(1)* %out, float addrspace(1)* %a.ptr, float addrspace(1)* %b.ptr) #0 {
141 %tid = call i32 @llvm.amdgcn.workitem.id.x()
142 %tid.ext = sext i32 %tid to i64
143 %a.gep = getelementptr inbounds float, float addrspace(1)* %a.ptr, i64 %tid.ext
144 %b.gep = getelementptr inbounds float, float addrspace(1)* %b.ptr, i64 %tid.ext
145 %out.gep = getelementptr inbounds i32, i32 addrspace(1)* %out, i64 %tid.ext
146 %a = load volatile float, float addrspace(1)* %a.gep
147 %b = load volatile float, float addrspace(1)* %b.gep
148 %fabs.a = call float @llvm.fabs.f32(float %a)
149 %neg.fabs.a = fsub float -0.0, %fabs.a
150 %neg.b = fsub float -0.0, %b
151 %cvt = call <2 x i16> @llvm.amdgcn.cvt.pknorm.u16(float %neg.fabs.a, float %neg.b)
152 %r = bitcast <2 x i16> %cvt to i32
153 store i32 %r, i32 addrspace(1)* %out.gep
157 declare <2 x i16> @llvm.amdgcn.cvt.pknorm.u16(float, float) #1
158 declare float @llvm.fabs.f32(float) #1
159 declare i32 @llvm.amdgcn.workitem.id.x() #1
162 attributes #0 = { nounwind }
163 attributes #1 = { nounwind readnone }