1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
|
; RUN: llc -march=amdgcn -mattr=+mad-mac-f32-insts -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=SI %s
; XUN: llc -march=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VI %s
; FIXME: None of these trigger madmk emission anymore. It is still
; possible, but requires the correct registers to be used which is
; hard to trigger.
declare i32 @llvm.amdgcn.workitem.id.x() nounwind readnone
declare float @llvm.fabs.f32(float) nounwind readnone
; GCN-LABEL: {{^}}madmk_f32:
; GCN-DAG: buffer_load_dword [[VA:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 glc{{$}}
; GCN-DAG: buffer_load_dword [[VB:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:4
; GCN: v_mac_f32_e32 [[VB]], 0x41200000, [[VA]]
define amdgpu_kernel void @madmk_f32(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %in) #0 {
%tid = tail call i32 @llvm.amdgcn.workitem.id.x() nounwind readnone
%gep.0 = getelementptr float, ptr addrspace(1) %in, i32 %tid
%gep.1 = getelementptr float, ptr addrspace(1) %gep.0, i32 1
%out.gep = getelementptr float, ptr addrspace(1) %out, i32 %tid
%a = load volatile float, ptr addrspace(1) %gep.0, align 4
%b = load volatile float, ptr addrspace(1) %gep.1, align 4
%mul = fmul float %a, 10.0
%madmk = fadd float %mul, %b
store float %madmk, ptr addrspace(1) %out.gep, align 4
ret void
}
; GCN-LABEL: {{^}}madmk_2_use_f32:
; GCN-DAG: buffer_load_dword [[VA:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 glc{{$}}
; GCN-DAG: buffer_load_dword [[VB:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:4
; GCN-DAG: buffer_load_dword [[VC:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8
; GCN-DAG: v_mac_f32_e32 [[VB]], 0x41200000, [[VA]]
; GCN-DAG: v_mac_f32_e32 [[VC]], 0x41200000, [[VA]]
; GCN: s_endpgm
define amdgpu_kernel void @madmk_2_use_f32(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %in) #0 {
%tid = tail call i32 @llvm.amdgcn.workitem.id.x() nounwind readnone
%in.gep.0 = getelementptr float, ptr addrspace(1) %in, i32 %tid
%in.gep.1 = getelementptr float, ptr addrspace(1) %in.gep.0, i32 1
%in.gep.2 = getelementptr float, ptr addrspace(1) %in.gep.0, i32 2
%out.gep.0 = getelementptr float, ptr addrspace(1) %out, i32 %tid
%out.gep.1 = getelementptr float, ptr addrspace(1) %in.gep.0, i32 1
%a = load volatile float, ptr addrspace(1) %in.gep.0, align 4
%b = load volatile float, ptr addrspace(1) %in.gep.1, align 4
%c = load volatile float, ptr addrspace(1) %in.gep.2, align 4
%mul0 = fmul float %a, 10.0
%mul1 = fmul float %a, 10.0
%madmk0 = fadd float %mul0, %b
%madmk1 = fadd float %mul1, %c
store float %madmk0, ptr addrspace(1) %out.gep.0, align 4
store float %madmk1, ptr addrspace(1) %out.gep.1, align 4
ret void
}
; We don't get any benefit if the constant is an inline immediate.
; GCN-LABEL: {{^}}madmk_inline_imm_f32:
; GCN-DAG: buffer_load_dword [[VA:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 glc{{$}}
; GCN-DAG: buffer_load_dword [[VB:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:4
; GCN: v_mac_f32_e32 [[VB]], 4.0, [[VA]]
define amdgpu_kernel void @madmk_inline_imm_f32(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %in) #0 {
%tid = tail call i32 @llvm.amdgcn.workitem.id.x() nounwind readnone
%gep.0 = getelementptr float, ptr addrspace(1) %in, i32 %tid
%gep.1 = getelementptr float, ptr addrspace(1) %gep.0, i32 1
%out.gep = getelementptr float, ptr addrspace(1) %out, i32 %tid
%a = load volatile float, ptr addrspace(1) %gep.0, align 4
%b = load volatile float, ptr addrspace(1) %gep.1, align 4
%mul = fmul float %a, 4.0
%madmk = fadd float %mul, %b
store float %madmk, ptr addrspace(1) %out.gep, align 4
ret void
}
; GCN-LABEL: {{^}}s_s_madmk_f32:
; GCN-NOT: v_madmk_f32
; GCN: v_mac_f32_e32
; GCN: s_endpgm
define amdgpu_kernel void @s_s_madmk_f32(ptr addrspace(1) noalias %out, [8 x i32], float %a, [8 x i32], float %b) #0 {
%tid = tail call i32 @llvm.amdgcn.workitem.id.x() nounwind readnone
%out.gep = getelementptr float, ptr addrspace(1) %out, i32 %tid
%mul = fmul float %a, 10.0
%madmk = fadd float %mul, %b
store float %madmk, ptr addrspace(1) %out.gep, align 4
ret void
}
; GCN-LABEL: {{^}}v_s_madmk_f32:
; GCN-DAG: s_load_dword [[SREG:s[0-9]+]]
; GCN-DAG: buffer_load_dword [[VREG1:v[0-9]+]]
; GCN: v_mov_b32_e32 [[VREG2:v[0-9]+]], [[SREG]]
; GCN: v_mac_f32_e32 [[VREG2]], 0x41200000, [[VREG1]]
; GCN: s_endpgm
define amdgpu_kernel void @v_s_madmk_f32(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %in, float %b) #0 {
%tid = tail call i32 @llvm.amdgcn.workitem.id.x() nounwind readnone
%gep.0 = getelementptr float, ptr addrspace(1) %in, i32 %tid
%out.gep = getelementptr float, ptr addrspace(1) %out, i32 %tid
%a = load float, ptr addrspace(1) %gep.0, align 4
%mul = fmul float %a, 10.0
%madmk = fadd float %mul, %b
store float %madmk, ptr addrspace(1) %out.gep, align 4
ret void
}
; GCN-LABEL: {{^}}scalar_vector_madmk_f32:
; GCN-NOT: v_madmk_f32
; GCN: v_mac_f32_e32
; GCN: s_endpgm
define amdgpu_kernel void @scalar_vector_madmk_f32(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %in, float %a) #0 {
%tid = tail call i32 @llvm.amdgcn.workitem.id.x() nounwind readnone
%gep.0 = getelementptr float, ptr addrspace(1) %in, i32 %tid
%out.gep = getelementptr float, ptr addrspace(1) %out, i32 %tid
%b = load float, ptr addrspace(1) %gep.0, align 4
%mul = fmul float %a, 10.0
%madmk = fadd float %mul, %b
store float %madmk, ptr addrspace(1) %out.gep, align 4
ret void
}
; GCN-LABEL: {{^}}no_madmk_src0_modifier_f32:
; GCN-DAG: buffer_load_dword [[VA:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 glc{{$}}
; GCN-DAG: buffer_load_dword [[VB:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:4
; GCN-DAG: s_mov_b32 [[SK:s[0-9]+]], 0x41200000
; GCN: v_mad_f32 {{v[0-9]+}}, |[[VA]]|, [[SK]], [[VB]]
define amdgpu_kernel void @no_madmk_src0_modifier_f32(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %in) #0 {
%tid = tail call i32 @llvm.amdgcn.workitem.id.x() nounwind readnone
%gep.0 = getelementptr float, ptr addrspace(1) %in, i32 %tid
%gep.1 = getelementptr float, ptr addrspace(1) %gep.0, i32 1
%out.gep = getelementptr float, ptr addrspace(1) %out, i32 %tid
%a = load volatile float, ptr addrspace(1) %gep.0, align 4
%b = load volatile float, ptr addrspace(1) %gep.1, align 4
%a.fabs = call float @llvm.fabs.f32(float %a) nounwind readnone
%mul = fmul float %a.fabs, 10.0
%madmk = fadd float %mul, %b
store float %madmk, ptr addrspace(1) %out.gep, align 4
ret void
}
; GCN-LABEL: {{^}}no_madmk_src2_modifier_f32:
; GCN-DAG: buffer_load_dword [[VA:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 glc{{$}}
; GCN-DAG: buffer_load_dword [[VB:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:4
; GCN: v_mad_f32 {{v[0-9]+}}, {{v[0-9]+}}, {{[sv][0-9]+}}, |{{v[0-9]+}}|
define amdgpu_kernel void @no_madmk_src2_modifier_f32(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %in) #0 {
%tid = tail call i32 @llvm.amdgcn.workitem.id.x() nounwind readnone
%gep.0 = getelementptr float, ptr addrspace(1) %in, i32 %tid
%gep.1 = getelementptr float, ptr addrspace(1) %gep.0, i32 1
%out.gep = getelementptr float, ptr addrspace(1) %out, i32 %tid
%a = load volatile float, ptr addrspace(1) %gep.0, align 4
%b = load volatile float, ptr addrspace(1) %gep.1, align 4
%b.fabs = call float @llvm.fabs.f32(float %b) nounwind readnone
%mul = fmul float %a, 10.0
%madmk = fadd float %mul, %b.fabs
store float %madmk, ptr addrspace(1) %out.gep, align 4
ret void
}
; GCN-LABEL: {{^}}madmk_add_inline_imm_f32:
; GCN: buffer_load_dword [[A:v[0-9]+]]
; GCN: s_mov_b32 [[SK:s[0-9]+]], 0x41200000
; GCN: v_mad_f32 {{v[0-9]+}}, [[A]], [[SK]], 2.0
define amdgpu_kernel void @madmk_add_inline_imm_f32(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %in) #0 {
%tid = tail call i32 @llvm.amdgcn.workitem.id.x() nounwind readnone
%gep.0 = getelementptr float, ptr addrspace(1) %in, i32 %tid
%out.gep = getelementptr float, ptr addrspace(1) %out, i32 %tid
%a = load float, ptr addrspace(1) %gep.0, align 4
%mul = fmul float %a, 10.0
%madmk = fadd float %mul, 2.0
store float %madmk, ptr addrspace(1) %out.gep, align 4
ret void
}
; SI-LABEL: {{^}}kill_madmk_verifier_error:
; SI: s_or_b64
; SI: s_xor_b64
; SI: v_mac_f32_e32 {{v[0-9]+}}, 0x472aee8c, {{v[0-9]+}}
define amdgpu_kernel void @kill_madmk_verifier_error() #0 {
bb:
br label %bb2
bb1: ; preds = %bb2
ret void
bb2: ; preds = %bb6, %bb
%tmp = phi float [ undef, %bb ], [ %tmp8, %bb6 ]
%tid = call i32 @llvm.amdgcn.mbcnt.lo(i32 -1, i32 0) #1
%f_tid = bitcast i32 %tid to float
%tmp3 = fsub float %f_tid, %tmp
%tmp5 = fcmp oeq float %tmp3, 1.000000e+04
br i1 %tmp5, label %bb1, label %bb6
bb6: ; preds = %bb2
%tmp7 = fmul float %tmp, 0x40E55DD180000000
%tmp8 = fadd float %tmp7, %tmp
br label %bb2
}
declare i32 @llvm.amdgcn.mbcnt.lo(i32, i32) #1
attributes #0 = { nounwind "denormal-fp-math-f32"="preserve-sign,preserve-sign" }
attributes #1 = { nounwind readnone }
|
