|
121
|
1 ; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,SI %s
|
|
|
2 ; RUN: llc -march=amdgcn -mcpu=fiji -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GFX89 %s
|
|
|
3 ; RUN: llc -march=amdgcn -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GFX89,GFX9 %s
|
|
|
4
|
|
|
5 ; GCN-LABEL: {{^}}v_clamp_add_src_f32:
|
|
|
6 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]]
|
|
|
7 ; GCN-NOT: [[A]]
|
|
|
8 ; GCN: v_add_f32_e64 v{{[0-9]+}}, [[A]], 1.0 clamp{{$}}
|
|
|
9 define amdgpu_kernel void @v_clamp_add_src_f32(float addrspace(1)* %out, float addrspace(1)* %aptr) #0 {
|
|
|
10 %tid = call i32 @llvm.amdgcn.workitem.id.x()
|
|
|
11 %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
|
|
|
12 %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
|
|
|
13 %a = load float, float addrspace(1)* %gep0
|
|
|
14 %add = fadd float %a, 1.0
|
|
|
15 %max = call float @llvm.maxnum.f32(float %add, float 0.0)
|
|
|
16 %clamp = call float @llvm.minnum.f32(float %max, float 1.0)
|
|
|
17 store float %clamp, float addrspace(1)* %out.gep
|
|
|
18 ret void
|
|
|
19 }
|
|
|
20
|
|
|
21 ; GCN-LABEL: {{^}}v_clamp_multi_use_src_f32:
|
|
|
22 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]]
|
|
|
23 ; GCN: v_add_f32_e32 [[ADD:v[0-9]+]], 1.0, [[A]]{{$}}
|
|
|
24 ; GCN: v_max_f32_e64 v{{[0-9]+}}, [[ADD]], [[ADD]] clamp{{$}}
|
|
|
25 define amdgpu_kernel void @v_clamp_multi_use_src_f32(float addrspace(1)* %out, float addrspace(1)* %aptr) #0 {
|
|
|
26 %tid = call i32 @llvm.amdgcn.workitem.id.x()
|
|
|
27 %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
|
|
|
28 %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
|
|
|
29 %a = load float, float addrspace(1)* %gep0
|
|
|
30 %add = fadd float %a, 1.0
|
|
|
31 %max = call float @llvm.maxnum.f32(float %add, float 0.0)
|
|
|
32 %clamp = call float @llvm.minnum.f32(float %max, float 1.0)
|
|
|
33 store float %clamp, float addrspace(1)* %out.gep
|
|
|
34 store volatile float %add, float addrspace(1)* undef
|
|
|
35 ret void
|
|
|
36 }
|
|
|
37
|
|
|
38 ; GCN-LABEL: {{^}}v_clamp_dbg_use_src_f32:
|
|
|
39 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]]
|
|
|
40 ; GCN-NOT: [[A]]
|
|
|
41 ; GCN: v_add_f32_e64 v{{[0-9]+}}, [[A]], 1.0 clamp{{$}}
|
|
|
42 define amdgpu_kernel void @v_clamp_dbg_use_src_f32(float addrspace(1)* %out, float addrspace(1)* %aptr) #0 {
|
|
|
43 %tid = call i32 @llvm.amdgcn.workitem.id.x()
|
|
|
44 %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
|
|
|
45 %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
|
|
|
46 %a = load float, float addrspace(1)* %gep0
|
|
|
47 %add = fadd float %a, 1.0
|
|
|
48 call void @llvm.dbg.value(metadata float %add, i64 0, metadata !4, metadata !9), !dbg !10
|
|
|
49 %max = call float @llvm.maxnum.f32(float %add, float 0.0)
|
|
|
50 %clamp = call float @llvm.minnum.f32(float %max, float 1.0)
|
|
|
51 store float %clamp, float addrspace(1)* %out.gep
|
|
|
52 ret void
|
|
|
53 }
|
|
|
54
|
|
|
55 ; GCN-LABEL: {{^}}v_clamp_add_neg_src_f32:
|
|
|
56 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]]
|
|
|
57 ; GCN: v_floor_f32_e32 [[FLOOR:v[0-9]+]], [[A]]
|
|
|
58 ; GCN: v_max_f32_e64 v{{[0-9]+}}, -[[FLOOR]], -[[FLOOR]] clamp{{$}}
|
|
|
59 define amdgpu_kernel void @v_clamp_add_neg_src_f32(float addrspace(1)* %out, float addrspace(1)* %aptr) #0 {
|
|
|
60 %tid = call i32 @llvm.amdgcn.workitem.id.x()
|
|
|
61 %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
|
|
|
62 %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
|
|
|
63 %a = load float, float addrspace(1)* %gep0
|
|
|
64 %floor = call float @llvm.floor.f32(float %a)
|
|
|
65 %neg.floor = fsub float -0.0, %floor
|
|
|
66 %max = call float @llvm.maxnum.f32(float %neg.floor, float 0.0)
|
|
|
67 %clamp = call float @llvm.minnum.f32(float %max, float 1.0)
|
|
|
68 store float %clamp, float addrspace(1)* %out.gep
|
|
|
69 ret void
|
|
|
70 }
|
|
|
71
|
|
|
72 ; GCN-LABEL: {{^}}v_non_clamp_max_f32:
|
|
|
73 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]]
|
|
|
74 ; GCN: v_add_f32_e32 [[ADD:v[0-9]+]], 1.0, [[A]]{{$}}
|
|
|
75 ; GCN: v_max_f32_e32 v{{[0-9]+}}, 0, [[ADD]]{{$}}
|
|
|
76 define amdgpu_kernel void @v_non_clamp_max_f32(float addrspace(1)* %out, float addrspace(1)* %aptr) #0 {
|
|
|
77 %tid = call i32 @llvm.amdgcn.workitem.id.x()
|
|
|
78 %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
|
|
|
79 %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
|
|
|
80 %a = load float, float addrspace(1)* %gep0
|
|
|
81 %add = fadd float %a, 1.0
|
|
|
82 %max = call float @llvm.maxnum.f32(float %add, float 0.0)
|
|
|
83 store float %max, float addrspace(1)* %out.gep
|
|
|
84 ret void
|
|
|
85 }
|
|
|
86
|
|
|
87 ; GCN-LABEL: {{^}}v_clamp_add_src_f32_denormals:
|
|
|
88 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]]
|
|
|
89 ; GCN: v_add_f32_e64 [[ADD:v[0-9]+]], [[A]], 1.0 clamp{{$}}
|
|
|
90 define amdgpu_kernel void @v_clamp_add_src_f32_denormals(float addrspace(1)* %out, float addrspace(1)* %aptr) #2 {
|
|
|
91 %tid = call i32 @llvm.amdgcn.workitem.id.x()
|
|
|
92 %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
|
|
|
93 %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
|
|
|
94 %a = load float, float addrspace(1)* %gep0
|
|
|
95 %add = fadd float %a, 1.0
|
|
|
96 %max = call float @llvm.maxnum.f32(float %add, float 0.0)
|
|
|
97 %clamp = call float @llvm.minnum.f32(float %max, float 1.0)
|
|
|
98 store float %clamp, float addrspace(1)* %out.gep
|
|
|
99 ret void
|
|
|
100 }
|
|
|
101
|
|
|
102 ; GCN-LABEL: {{^}}v_clamp_add_src_f16_denorm:
|
|
|
103 ; GCN: {{buffer|flat|global}}_load_ushort [[A:v[0-9]+]]
|
|
|
104 ; GFX89: v_add_f16_e64 [[ADD:v[0-9]+]], [[A]], 1.0 clamp{{$}}
|
|
|
105
|
|
|
106 ; SI: v_cvt_f32_f16_e32 [[CVT:v[0-9]+]], [[A]]
|
|
|
107 ; SI: v_add_f32_e64 [[ADD:v[0-9]+]], [[CVT]], 1.0 clamp{{$}}
|
|
|
108 ; SI: v_cvt_f16_f32_e32 v{{[0-9]+}}, [[ADD]]
|
|
|
109 define amdgpu_kernel void @v_clamp_add_src_f16_denorm(half addrspace(1)* %out, half addrspace(1)* %aptr) #0 {
|
|
|
110 %tid = call i32 @llvm.amdgcn.workitem.id.x()
|
|
|
111 %gep0 = getelementptr half, half addrspace(1)* %aptr, i32 %tid
|
|
|
112 %out.gep = getelementptr half, half addrspace(1)* %out, i32 %tid
|
|
|
113 %a = load half, half addrspace(1)* %gep0
|
|
|
114 %add = fadd half %a, 1.0
|
|
|
115 %max = call half @llvm.maxnum.f16(half %add, half 0.0)
|
|
|
116 %clamp = call half @llvm.minnum.f16(half %max, half 1.0)
|
|
|
117 store half %clamp, half addrspace(1)* %out.gep
|
|
|
118 ret void
|
|
|
119 }
|
|
|
120
|
|
|
121 ; GCN-LABEL: {{^}}v_clamp_add_src_f16_no_denormals:
|
|
|
122 ; GCN: {{buffer|flat|global}}_load_ushort [[A:v[0-9]+]]
|
|
|
123 ; GFX89-NOT: [[A]]
|
|
|
124 ; GFX89: v_add_f16_e64 v{{[0-9]+}}, [[A]], 1.0 clamp{{$}}
|
|
|
125
|
|
|
126 ; SI: v_cvt_f32_f16_e32 [[CVT:v[0-9]+]], [[A]]
|
|
|
127 ; SI: v_add_f32_e64 [[ADD:v[0-9]+]], [[CVT]], 1.0 clamp{{$}}
|
|
|
128 ; SI: v_cvt_f16_f32_e32 v{{[0-9]+}}, [[ADD]]
|
|
|
129 define amdgpu_kernel void @v_clamp_add_src_f16_no_denormals(half addrspace(1)* %out, half addrspace(1)* %aptr) #3 {
|
|
|
130 %tid = call i32 @llvm.amdgcn.workitem.id.x()
|
|
|
131 %gep0 = getelementptr half, half addrspace(1)* %aptr, i32 %tid
|
|
|
132 %out.gep = getelementptr half, half addrspace(1)* %out, i32 %tid
|
|
|
133 %a = load half, half addrspace(1)* %gep0
|
|
|
134 %add = fadd half %a, 1.0
|
|
|
135 %max = call half @llvm.maxnum.f16(half %add, half 0.0)
|
|
|
136 %clamp = call half @llvm.minnum.f16(half %max, half 1.0)
|
|
|
137 store half %clamp, half addrspace(1)* %out.gep
|
|
|
138 ret void
|
|
|
139 }
|
|
|
140
|
|
|
141 ; GCN-LABEL: {{^}}v_clamp_add_src_v2f32:
|
|
|
142 ; GCN: {{buffer|flat|global}}_load_dwordx2 v{{\[}}[[A:[0-9]+]]:[[B:[0-9]+]]{{\]}}
|
|
|
143 ; GCN-DAG: v_add_f32_e64 v{{[0-9]+}}, v[[A]], 1.0 clamp{{$}}
|
|
|
144 ; GCN-DAG: v_add_f32_e64 v{{[0-9]+}}, v[[B]], 1.0 clamp{{$}}
|
|
|
145 define amdgpu_kernel void @v_clamp_add_src_v2f32(<2 x float> addrspace(1)* %out, <2 x float> addrspace(1)* %aptr) #0 {
|
|
|
146 %tid = call i32 @llvm.amdgcn.workitem.id.x()
|
|
|
147 %gep0 = getelementptr <2 x float>, <2 x float> addrspace(1)* %aptr, i32 %tid
|
|
|
148 %out.gep = getelementptr <2 x float>, <2 x float> addrspace(1)* %out, i32 %tid
|
|
|
149 %a = load <2 x float>, <2 x float> addrspace(1)* %gep0
|
|
|
150 %add = fadd <2 x float> %a, <float 1.0, float 1.0>
|
|
|
151 %max = call <2 x float> @llvm.maxnum.v2f32(<2 x float> %add, <2 x float> zeroinitializer)
|
|
|
152 %clamp = call <2 x float> @llvm.minnum.v2f32(<2 x float> %max, <2 x float> <float 1.0, float 1.0>)
|
|
|
153 store <2 x float> %clamp, <2 x float> addrspace(1)* %out.gep
|
|
|
154 ret void
|
|
|
155 }
|
|
|
156
|
|
|
157 ; GCN-LABEL: {{^}}v_clamp_add_src_f64:
|
|
|
158 ; GCN: {{buffer|flat|global}}_load_dwordx2 [[A:v\[[0-9]+:[0-9]+\]]]
|
|
|
159 ; GCN: v_add_f64 v{{\[[0-9]+:[0-9]+\]}}, [[A]], 1.0 clamp{{$}}
|
|
|
160 define amdgpu_kernel void @v_clamp_add_src_f64(double addrspace(1)* %out, double addrspace(1)* %aptr) #0 {
|
|
|
161 %tid = call i32 @llvm.amdgcn.workitem.id.x()
|
|
|
162 %gep0 = getelementptr double, double addrspace(1)* %aptr, i32 %tid
|
|
|
163 %out.gep = getelementptr double, double addrspace(1)* %out, i32 %tid
|
|
|
164 %a = load double, double addrspace(1)* %gep0
|
|
|
165 %add = fadd double %a, 1.0
|
|
|
166 %max = call double @llvm.maxnum.f64(double %add, double 0.0)
|
|
|
167 %clamp = call double @llvm.minnum.f64(double %max, double 1.0)
|
|
|
168 store double %clamp, double addrspace(1)* %out.gep
|
|
|
169 ret void
|
|
|
170 }
|
|
|
171
|
|
|
172 ; GCN-LABEL: {{^}}v_clamp_mac_to_mad:
|
|
|
173 ; GCN: v_mad_f32 v{{[0-9]+}}, s{{[0-9]+}}, s{{[0-9]+}}, v{{[0-9]}} clamp{{$}}
|
|
|
174 define amdgpu_kernel void @v_clamp_mac_to_mad(float addrspace(1)* %out, float addrspace(1)* %aptr, float %a) #0 {
|
|
|
175 %tid = call i32 @llvm.amdgcn.workitem.id.x()
|
|
|
176 %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
|
|
|
177 %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
|
|
|
178 %b = load float, float addrspace(1)* %gep0
|
|
|
179
|
|
|
180 %mul = fmul float %a, %a
|
|
|
181 %add = fadd float %mul, %b
|
|
|
182 %max = call float @llvm.maxnum.f32(float %add, float 0.0)
|
|
|
183 %clamp = call float @llvm.minnum.f32(float %max, float 1.0)
|
|
|
184 %res = fadd float %clamp, %b
|
|
|
185 store float %res, float addrspace(1)* %out.gep
|
|
|
186 ret void
|
|
|
187 }
|
|
|
188
|
|
|
189
|
|
|
190 ; GCN-LABEL: {{^}}v_clamp_add_src_v2f16_denorm:
|
|
|
191 ; GCN-DAG: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]]
|
|
|
192 ; GFX9-DAG: s_mov_b32 [[ONE:s[0-9]+]], 0x3c003c00
|
|
|
193 ; GFX9: v_pk_add_f16 [[ADD:v[0-9]+]], [[A]], [[ONE]] clamp{{$}}
|
|
|
194 define amdgpu_kernel void @v_clamp_add_src_v2f16_denorm(<2 x half> addrspace(1)* %out, <2 x half> addrspace(1)* %aptr) #0 {
|
|
|
195 %tid = call i32 @llvm.amdgcn.workitem.id.x()
|
|
|
196 %gep0 = getelementptr <2 x half>, <2 x half> addrspace(1)* %aptr, i32 %tid
|
|
|
197 %out.gep = getelementptr <2 x half>, <2 x half> addrspace(1)* %out, i32 %tid
|
|
|
198 %a = load <2 x half>, <2 x half> addrspace(1)* %gep0
|
|
|
199 %add = fadd <2 x half> %a, <half 1.0, half 1.0>
|
|
|
200 %max = call <2 x half> @llvm.maxnum.v2f16(<2 x half> %add, <2 x half> zeroinitializer)
|
|
|
201 %clamp = call <2 x half> @llvm.minnum.v2f16(<2 x half> %max, <2 x half> <half 1.0, half 1.0>)
|
|
|
202 store <2 x half> %clamp, <2 x half> addrspace(1)* %out.gep
|
|
|
203 ret void
|
|
|
204 }
|
|
|
205
|
|
|
206 ; GCN-LABEL: {{^}}v_clamp_add_src_v2f16_no_denormals:
|
|
|
207 ; GCN-DAG: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]]
|
|
|
208 ; GFX9-DAG: s_mov_b32 [[ONE:s[0-9]+]], 0x3c003c00
|
|
|
209 ; GFX9: v_pk_add_f16 [[ADD:v[0-9]+]], [[A]], [[ONE]] clamp{{$}}
|
|
|
210 define amdgpu_kernel void @v_clamp_add_src_v2f16_no_denormals(<2 x half> addrspace(1)* %out, <2 x half> addrspace(1)* %aptr) #3 {
|
|
|
211 %tid = call i32 @llvm.amdgcn.workitem.id.x()
|
|
|
212 %gep0 = getelementptr <2 x half>, <2 x half> addrspace(1)* %aptr, i32 %tid
|
|
|
213 %out.gep = getelementptr <2 x half>, <2 x half> addrspace(1)* %out, i32 %tid
|
|
|
214 %a = load <2 x half>, <2 x half> addrspace(1)* %gep0
|
|
|
215 %add = fadd <2 x half> %a, <half 1.0, half 1.0>
|
|
|
216 %max = call <2 x half> @llvm.maxnum.v2f16(<2 x half> %add, <2 x half> zeroinitializer)
|
|
|
217 %clamp = call <2 x half> @llvm.minnum.v2f16(<2 x half> %max, <2 x half> <half 1.0, half 1.0>)
|
|
|
218 store <2 x half> %clamp, <2 x half> addrspace(1)* %out.gep
|
|
|
219 ret void
|
|
|
220 }
|
|
|
221
|
|
|
222 ; GCN-LABEL: {{^}}v_clamp_add_src_v2f16_denorm_neg:
|
|
|
223 ; GCN-DAG: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]]
|
|
|
224 ; GFX9-DAG: s_mov_b32 [[ONE:s[0-9]+]], 0x3c003c00
|
|
|
225 ; GFX9: v_pk_add_f16 [[ADD:v[0-9]+]], [[A]], [[ONE]]{{$}}
|
|
|
226 ; GFX9: v_pk_max_f16 [[MAX:v[0-9]+]], [[ADD]], [[ADD]] neg_lo:[1,1] neg_hi:[1,1] clamp{{$}}
|
|
|
227 define amdgpu_kernel void @v_clamp_add_src_v2f16_denorm_neg(<2 x half> addrspace(1)* %out, <2 x half> addrspace(1)* %aptr) #0 {
|
|
|
228 %tid = call i32 @llvm.amdgcn.workitem.id.x()
|
|
|
229 %gep0 = getelementptr <2 x half>, <2 x half> addrspace(1)* %aptr, i32 %tid
|
|
|
230 %out.gep = getelementptr <2 x half>, <2 x half> addrspace(1)* %out, i32 %tid
|
|
|
231 %a = load <2 x half>, <2 x half> addrspace(1)* %gep0
|
|
|
232 %add = fadd <2 x half> %a, <half 1.0, half 1.0>
|
|
|
233 %neg.add = fsub <2 x half> <half -0.0, half -0.0>, %add
|
|
|
234 %max = call <2 x half> @llvm.maxnum.v2f16(<2 x half> %neg.add, <2 x half> zeroinitializer)
|
|
|
235 %clamp = call <2 x half> @llvm.minnum.v2f16(<2 x half> %max, <2 x half> <half 1.0, half 1.0>)
|
|
|
236 store <2 x half> %clamp, <2 x half> addrspace(1)* %out.gep
|
|
|
237 ret void
|
|
|
238 }
|
|
|
239
|
|
|
240 ; GCN-LABEL: {{^}}v_clamp_add_src_v2f16_denorm_neg_lo:
|
|
|
241 ; GCN-DAG: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]]
|
|
|
242 ; GFX9-DAG: s_mov_b32 [[ONE:s[0-9]+]], 0x3c003c00
|
|
|
243 ; GFX9: v_pk_add_f16 [[ADD:v[0-9]+]], [[A]], [[ONE]]{{$}}
|
|
|
244 ; GFX9: v_pk_max_f16 [[MAX:v[0-9]+]], [[ADD]], [[ADD]] neg_lo:[1,1] clamp{{$}}
|
|
|
245 define amdgpu_kernel void @v_clamp_add_src_v2f16_denorm_neg_lo(<2 x half> addrspace(1)* %out, <2 x half> addrspace(1)* %aptr) #0 {
|
|
|
246 %tid = call i32 @llvm.amdgcn.workitem.id.x()
|
|
|
247 %gep0 = getelementptr <2 x half>, <2 x half> addrspace(1)* %aptr, i32 %tid
|
|
|
248 %out.gep = getelementptr <2 x half>, <2 x half> addrspace(1)* %out, i32 %tid
|
|
|
249 %a = load <2 x half>, <2 x half> addrspace(1)* %gep0
|
|
|
250 %add = fadd <2 x half> %a, <half 1.0, half 1.0>
|
|
|
251 %lo = extractelement <2 x half> %add, i32 0
|
|
|
252 %neg.lo = fsub half -0.0, %lo
|
|
|
253 %neg.lo.add = insertelement <2 x half> %add, half %neg.lo, i32 0
|
|
|
254 %max = call <2 x half> @llvm.maxnum.v2f16(<2 x half> %neg.lo.add, <2 x half> zeroinitializer)
|
|
|
255 %clamp = call <2 x half> @llvm.minnum.v2f16(<2 x half> %max, <2 x half> <half 1.0, half 1.0>)
|
|
|
256 store <2 x half> %clamp, <2 x half> addrspace(1)* %out.gep
|
|
|
257 ret void
|
|
|
258 }
|
|
|
259
|
|
|
260 ; GCN-LABEL: {{^}}v_clamp_add_src_v2f16_denorm_neg_hi:
|
|
|
261 ; GCN-DAG: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]]
|
|
|
262 ; GFX9-DAG: s_mov_b32 [[ONE:s[0-9]+]], 0x3c003c00
|
|
|
263 ; GFX9: v_pk_add_f16 [[ADD:v[0-9]+]], [[A]], [[ONE]]{{$}}
|
|
|
264 ; GFX9: v_pk_max_f16 [[MAX:v[0-9]+]], [[ADD]], [[ADD]] neg_hi:[1,1] clamp{{$}}
|
|
|
265 define amdgpu_kernel void @v_clamp_add_src_v2f16_denorm_neg_hi(<2 x half> addrspace(1)* %out, <2 x half> addrspace(1)* %aptr) #0 {
|
|
|
266 %tid = call i32 @llvm.amdgcn.workitem.id.x()
|
|
|
267 %gep0 = getelementptr <2 x half>, <2 x half> addrspace(1)* %aptr, i32 %tid
|
|
|
268 %out.gep = getelementptr <2 x half>, <2 x half> addrspace(1)* %out, i32 %tid
|
|
|
269 %a = load <2 x half>, <2 x half> addrspace(1)* %gep0
|
|
|
270 %add = fadd <2 x half> %a, <half 1.0, half 1.0>
|
|
|
271 %hi = extractelement <2 x half> %add, i32 1
|
|
|
272 %neg.hi = fsub half -0.0, %hi
|
|
|
273 %neg.hi.add = insertelement <2 x half> %add, half %neg.hi, i32 1
|
|
|
274 %max = call <2 x half> @llvm.maxnum.v2f16(<2 x half> %neg.hi.add, <2 x half> zeroinitializer)
|
|
|
275 %clamp = call <2 x half> @llvm.minnum.v2f16(<2 x half> %max, <2 x half> <half 1.0, half 1.0>)
|
|
|
276 store <2 x half> %clamp, <2 x half> addrspace(1)* %out.gep
|
|
|
277 ret void
|
|
|
278 }
|
|
|
279
|
|
|
280 ; GCN-LABEL: {{^}}v_clamp_add_src_v2f16_denorm_shuf:
|
|
|
281 ; GCN-DAG: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]]
|
|
|
282 ; GFX9-DAG: s_mov_b32 [[ONE:s[0-9]+]], 0x3c003c00
|
|
|
283 ; GFX9: v_pk_add_f16 [[ADD:v[0-9]+]], [[A]], [[ONE]]{{$}}
|
|
|
284 ; GFX9: v_pk_max_f16 [[MAX:v[0-9]+]], [[ADD]], [[ADD]] op_sel:[1,1] op_sel_hi:[0,0] clamp{{$}}
|
|
|
285 define amdgpu_kernel void @v_clamp_add_src_v2f16_denorm_shuf(<2 x half> addrspace(1)* %out, <2 x half> addrspace(1)* %aptr) #0 {
|
|
|
286 %tid = call i32 @llvm.amdgcn.workitem.id.x()
|
|
|
287 %gep0 = getelementptr <2 x half>, <2 x half> addrspace(1)* %aptr, i32 %tid
|
|
|
288 %out.gep = getelementptr <2 x half>, <2 x half> addrspace(1)* %out, i32 %tid
|
|
|
289 %a = load <2 x half>, <2 x half> addrspace(1)* %gep0
|
|
|
290 %add = fadd <2 x half> %a, <half 1.0, half 1.0>
|
|
|
291 %shuf = shufflevector <2 x half> %add, <2 x half> undef, <2 x i32> <i32 1, i32 0>
|
|
|
292
|
|
|
293 %max = call <2 x half> @llvm.maxnum.v2f16(<2 x half> %shuf, <2 x half> zeroinitializer)
|
|
|
294 %clamp = call <2 x half> @llvm.minnum.v2f16(<2 x half> %max, <2 x half> <half 1.0, half 1.0>)
|
|
|
295 store <2 x half> %clamp, <2 x half> addrspace(1)* %out.gep
|
|
|
296 ret void
|
|
|
297 }
|
|
|
298
|
|
|
299 ; GCN-LABEL: {{^}}v_no_clamp_add_src_v2f16_f32_src:
|
|
|
300 ; GCN-DAG: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]]
|
|
|
301 ; GFX9: v_add_f32_e32 [[ADD:v[0-9]+]], 1.0, [[A]]{{$}}
|
|
|
302 ; GFX9: v_pk_max_f16 [[CLAMP:v[0-9]+]], [[ADD]], [[ADD]] clamp{{$}}
|
|
|
303 define amdgpu_kernel void @v_no_clamp_add_src_v2f16_f32_src(<2 x half> addrspace(1)* %out, <2 x half> addrspace(1)* %aptr) #0 {
|
|
|
304 %tid = call i32 @llvm.amdgcn.workitem.id.x()
|
|
|
305 %gep0 = getelementptr <2 x half>, <2 x half> addrspace(1)* %aptr, i32 %tid
|
|
|
306 %out.gep = getelementptr <2 x half>, <2 x half> addrspace(1)* %out, i32 %tid
|
|
|
307 %a = load <2 x half>, <2 x half> addrspace(1)* %gep0
|
|
|
308 %bc = bitcast <2 x half> %a to float
|
|
|
309 %f32.op = fadd float %bc, 1.0
|
|
|
310 %f32.op.cast = bitcast float %f32.op to <2 x half>
|
|
|
311 %max = call <2 x half> @llvm.maxnum.v2f16(<2 x half> %f32.op.cast, <2 x half> zeroinitializer)
|
|
|
312 %clamp = call <2 x half> @llvm.minnum.v2f16(<2 x half> %max, <2 x half> <half 1.0, half 1.0>)
|
|
|
313 store <2 x half> %clamp, <2 x half> addrspace(1)* %out.gep
|
|
|
314 ret void
|
|
|
315 }
|
|
|
316
|
|
|
317 ; GCN-LABEL: {{^}}v_no_clamp_add_packed_src_f32:
|
|
|
318 ; GCN-DAG: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]]
|
|
|
319 ; GFX9-DAG: s_mov_b32 [[ONE:s[0-9]+]], 0x3c003c00
|
|
|
320 ; GFX9: v_pk_add_f16 [[ADD:v[0-9]+]], [[A]], [[ONE]]{{$}}
|
|
|
321 ; GFX9: v_max_f32_e64 [[CLAMP:v[0-9]+]], [[ADD]], [[ADD]] clamp{{$}}
|
|
|
322 define amdgpu_kernel void @v_no_clamp_add_packed_src_f32(float addrspace(1)* %out, <2 x half> addrspace(1)* %aptr) #0 {
|
|
|
323 %tid = call i32 @llvm.amdgcn.workitem.id.x()
|
|
|
324 %gep0 = getelementptr <2 x half>, <2 x half> addrspace(1)* %aptr, i32 %tid
|
|
|
325 %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
|
|
|
326 %a = load <2 x half>, <2 x half> addrspace(1)* %gep0
|
|
|
327 %add = fadd <2 x half> %a, <half 1.0, half 1.0>
|
|
|
328 %bc.add = bitcast <2 x half> %add to float
|
|
|
329 %max = call float @llvm.maxnum.f32(float %bc.add, float 0.0)
|
|
|
330 %clamp = call float @llvm.minnum.f32(float %max, float 1.0)
|
|
|
331 store float %clamp, float addrspace(1)* %out.gep
|
|
|
332 ret void
|
|
|
333 }
|
|
|
334
|
|
|
335 ; Since the high bits are zeroed, it probably would be OK in this case
|
|
|
336 ; to use clamp.
|
|
|
337 ; GCN-LABEL: {{^}}v_no_clamp_add_src_v2f16_f16_src:
|
|
|
338 ; GCN-DAG: {{buffer|flat|global}}_load_ushort [[A:v[0-9]+]]
|
|
|
339 ; GFX9: v_add_f16_e32 [[ADD:v[0-9]+]], 1.0, [[A]]{{$}}
|
|
|
340 ; GFX9: v_pk_max_f16 [[CLAMP:v[0-9]+]], [[ADD]], [[ADD]] clamp{{$}}
|
|
|
341 define amdgpu_kernel void @v_no_clamp_add_src_v2f16_f16_src(<2 x half> addrspace(1)* %out, half addrspace(1)* %aptr) #0 {
|
|
|
342 %tid = call i32 @llvm.amdgcn.workitem.id.x()
|
|
|
343 %gep0 = getelementptr half, half addrspace(1)* %aptr, i32 %tid
|
|
|
344 %out.gep = getelementptr <2 x half>, <2 x half> addrspace(1)* %out, i32 %tid
|
|
|
345 %a = load half, half addrspace(1)* %gep0
|
|
|
346 %add = fadd half %a, 1.0
|
|
|
347 %bc = bitcast half %add to i16
|
|
|
348 %zext = zext i16 %bc to i32
|
|
|
349 %v2f16 = bitcast i32 %zext to <2 x half>
|
|
|
350 %max = call <2 x half> @llvm.maxnum.v2f16(<2 x half> %v2f16, <2 x half> zeroinitializer)
|
|
|
351 %clamp = call <2 x half> @llvm.minnum.v2f16(<2 x half> %max, <2 x half> <half 1.0, half 1.0>)
|
|
|
352 store <2 x half> %clamp, <2 x half> addrspace(1)* %out.gep
|
|
|
353 ret void
|
|
|
354 }
|
|
|
355
|
|
|
356 declare i32 @llvm.amdgcn.workitem.id.x() #1
|
|
|
357 declare float @llvm.fabs.f32(float) #1
|
|
|
358 declare float @llvm.floor.f32(float) #1
|
|
|
359 declare float @llvm.minnum.f32(float, float) #1
|
|
|
360 declare float @llvm.maxnum.f32(float, float) #1
|
|
|
361 declare float @llvm.amdgcn.fmed3.f32(float, float, float) #1
|
|
|
362 declare double @llvm.fabs.f64(double) #1
|
|
|
363 declare double @llvm.minnum.f64(double, double) #1
|
|
|
364 declare double @llvm.maxnum.f64(double, double) #1
|
|
|
365 declare half @llvm.fabs.f16(half) #1
|
|
|
366 declare half @llvm.minnum.f16(half, half) #1
|
|
|
367 declare half @llvm.maxnum.f16(half, half) #1
|
|
|
368 declare <2 x half> @llvm.minnum.v2f16(<2 x half>, <2 x half>) #1
|
|
|
369 declare <2 x half> @llvm.maxnum.v2f16(<2 x half>, <2 x half>) #1
|
|
|
370 declare <2 x float> @llvm.minnum.v2f32(<2 x float>, <2 x float>) #1
|
|
|
371 declare <2 x float> @llvm.maxnum.v2f32(<2 x float>, <2 x float>) #1
|
|
|
372
|
|
|
373
|
|
|
374 declare void @llvm.dbg.value(metadata, i64, metadata, metadata) #1
|
|
|
375
|
|
|
376 attributes #0 = { nounwind }
|
|
|
377 attributes #1 = { nounwind readnone }
|
|
|
378 attributes #2 = { nounwind "target-features"="+fp32-denormals" }
|
|
|
379 attributes #3 = { nounwind "target-features"="-fp64-fp16-denormals" }
|
|
|
380
|
|
|
381 !llvm.dbg.cu = !{!0}
|
|
|
382 !llvm.module.flags = !{!2, !3}
|
|
|
383
|
|
|
384 !0 = distinct !DICompileUnit(language: DW_LANG_C99, file: !1, isOptimized: true, runtimeVersion: 0, emissionKind: NoDebug)
|
|
|
385 !1 = !DIFile(filename: "/tmp/foo.cl", directory: "/dev/null")
|
|
|
386 !2 = !{i32 2, !"Dwarf Version", i32 4}
|
|
|
387 !3 = !{i32 2, !"Debug Info Version", i32 3}
|
|
|
388 !4 = !DILocalVariable(name: "add", arg: 1, scope: !5, file: !1, line: 1)
|
|
|
389 !5 = distinct !DISubprogram(name: "foo", scope: !1, file: !1, line: 1, type: !6, isLocal: false, isDefinition: true, scopeLine: 2, flags: DIFlagPrototyped, isOptimized: true, unit: !0)
|
|
|
390 !6 = !DISubroutineType(types: !7)
|
|
|
391 !7 = !{null, !8}
|
|
|
392 !8 = !DIBasicType(name: "float", size: 32, align: 32)
|
|
|
393 !9 = !DIExpression()
|
|
|
394 !10 = !DILocation(line: 1, column: 42, scope: !5)
|
