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
|
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: opt -S -passes=always-inline -o %t.bc %s
; RUN: llc -march=amdgcn -mcpu=gfx1102 -mattr=+wavefrontsize32 -misched=gcn-max-occupancy -verify-machineinstrs < %t.bc | FileCheck %s --check-prefixes=CHECK
; Provide a long sequence of 32 vec4 load/store pairs that ought to be fully
; overlapped for latency hiding. Doing so requires using (at least) 128 VGPRs,
; which (incorrectly) used to look to the scheduler like an occupancy reduction.
; 6 kB of LDS, allows 10 workgroups
@lds = internal addrspace(3) global [384 x <4 x i32>] undef
define internal amdgpu_gfx void @copy(ptr addrspace(1) %src, i32 %ofs) alwaysinline {
%src.gep = getelementptr <4 x i32>, ptr addrspace(1) %src, i32 %ofs
%ld = load <4 x i32>, ptr addrspace(1) %src.gep
%dst.gep = getelementptr <4 x i32>, ptr addrspace(3) @lds, i32 %ofs
store <4 x i32> %ld, ptr addrspace(3) %dst.gep
ret void
}
define amdgpu_cs void @test(ptr addrspace(1) %src) "amdgpu-flat-work-group-size"="32,32" {
; CHECK-LABEL: test:
; CHECK: ; %bb.0:
; CHECK-NEXT: s_clause 0x1f
; CHECK-NEXT: global_load_b128 v[2:5], v[0:1], off
; CHECK-NEXT: global_load_b128 v[6:9], v[0:1], off offset:16
; CHECK-NEXT: global_load_b128 v[10:13], v[0:1], off offset:32
; CHECK-NEXT: global_load_b128 v[14:17], v[0:1], off offset:48
; CHECK-NEXT: global_load_b128 v[18:21], v[0:1], off offset:64
; CHECK-NEXT: global_load_b128 v[22:25], v[0:1], off offset:80
; CHECK-NEXT: global_load_b128 v[26:29], v[0:1], off offset:96
; CHECK-NEXT: global_load_b128 v[30:33], v[0:1], off offset:112
; CHECK-NEXT: global_load_b128 v[34:37], v[0:1], off offset:128
; CHECK-NEXT: global_load_b128 v[38:41], v[0:1], off offset:144
; CHECK-NEXT: global_load_b128 v[42:45], v[0:1], off offset:160
; CHECK-NEXT: global_load_b128 v[46:49], v[0:1], off offset:176
; CHECK-NEXT: global_load_b128 v[50:53], v[0:1], off offset:192
; CHECK-NEXT: global_load_b128 v[54:57], v[0:1], off offset:208
; CHECK-NEXT: global_load_b128 v[58:61], v[0:1], off offset:224
; CHECK-NEXT: global_load_b128 v[62:65], v[0:1], off offset:240
; CHECK-NEXT: global_load_b128 v[66:69], v[0:1], off offset:256
; CHECK-NEXT: global_load_b128 v[70:73], v[0:1], off offset:272
; CHECK-NEXT: global_load_b128 v[74:77], v[0:1], off offset:288
; CHECK-NEXT: global_load_b128 v[78:81], v[0:1], off offset:304
; CHECK-NEXT: global_load_b128 v[82:85], v[0:1], off offset:320
; CHECK-NEXT: global_load_b128 v[86:89], v[0:1], off offset:336
; CHECK-NEXT: global_load_b128 v[90:93], v[0:1], off offset:352
; CHECK-NEXT: global_load_b128 v[94:97], v[0:1], off offset:368
; CHECK-NEXT: global_load_b128 v[98:101], v[0:1], off offset:384
; CHECK-NEXT: global_load_b128 v[102:105], v[0:1], off offset:400
; CHECK-NEXT: global_load_b128 v[106:109], v[0:1], off offset:416
; CHECK-NEXT: global_load_b128 v[110:113], v[0:1], off offset:432
; CHECK-NEXT: global_load_b128 v[114:117], v[0:1], off offset:448
; CHECK-NEXT: global_load_b128 v[118:121], v[0:1], off offset:464
; CHECK-NEXT: global_load_b128 v[122:125], v[0:1], off offset:480
; CHECK-NEXT: global_load_b128 v[126:129], v[0:1], off offset:496
; CHECK-NEXT: v_mov_b32_e32 v0, 0
; CHECK-NEXT: s_waitcnt vmcnt(31)
; CHECK-NEXT: ds_store_b128 v0, v[2:5]
; CHECK-NEXT: s_waitcnt vmcnt(30)
; CHECK-NEXT: ds_store_b128 v0, v[6:9] offset:16
; CHECK-NEXT: s_waitcnt vmcnt(29)
; CHECK-NEXT: ds_store_b128 v0, v[10:13] offset:32
; CHECK-NEXT: s_waitcnt vmcnt(28)
; CHECK-NEXT: ds_store_b128 v0, v[14:17] offset:48
; CHECK-NEXT: s_waitcnt vmcnt(27)
; CHECK-NEXT: ds_store_b128 v0, v[18:21] offset:64
; CHECK-NEXT: s_waitcnt vmcnt(26)
; CHECK-NEXT: ds_store_b128 v0, v[22:25] offset:80
; CHECK-NEXT: s_waitcnt vmcnt(25)
; CHECK-NEXT: ds_store_b128 v0, v[26:29] offset:96
; CHECK-NEXT: s_waitcnt vmcnt(24)
; CHECK-NEXT: ds_store_b128 v0, v[30:33] offset:112
; CHECK-NEXT: s_waitcnt vmcnt(23)
; CHECK-NEXT: ds_store_b128 v0, v[34:37] offset:128
; CHECK-NEXT: s_waitcnt vmcnt(22)
; CHECK-NEXT: ds_store_b128 v0, v[38:41] offset:144
; CHECK-NEXT: s_waitcnt vmcnt(21)
; CHECK-NEXT: ds_store_b128 v0, v[42:45] offset:160
; CHECK-NEXT: s_waitcnt vmcnt(20)
; CHECK-NEXT: ds_store_b128 v0, v[46:49] offset:176
; CHECK-NEXT: s_waitcnt vmcnt(19)
; CHECK-NEXT: ds_store_b128 v0, v[50:53] offset:192
; CHECK-NEXT: s_waitcnt vmcnt(18)
; CHECK-NEXT: ds_store_b128 v0, v[54:57] offset:208
; CHECK-NEXT: s_waitcnt vmcnt(17)
; CHECK-NEXT: ds_store_b128 v0, v[58:61] offset:224
; CHECK-NEXT: s_waitcnt vmcnt(16)
; CHECK-NEXT: ds_store_b128 v0, v[62:65] offset:240
; CHECK-NEXT: s_waitcnt vmcnt(15)
; CHECK-NEXT: ds_store_b128 v0, v[66:69] offset:256
; CHECK-NEXT: s_waitcnt vmcnt(14)
; CHECK-NEXT: ds_store_b128 v0, v[70:73] offset:272
; CHECK-NEXT: s_waitcnt vmcnt(13)
; CHECK-NEXT: ds_store_b128 v0, v[74:77] offset:288
; CHECK-NEXT: s_waitcnt vmcnt(12)
; CHECK-NEXT: ds_store_b128 v0, v[78:81] offset:304
; CHECK-NEXT: s_waitcnt vmcnt(11)
; CHECK-NEXT: ds_store_b128 v0, v[82:85] offset:320
; CHECK-NEXT: s_waitcnt vmcnt(10)
; CHECK-NEXT: ds_store_b128 v0, v[86:89] offset:336
; CHECK-NEXT: s_waitcnt vmcnt(9)
; CHECK-NEXT: ds_store_b128 v0, v[90:93] offset:352
; CHECK-NEXT: s_waitcnt vmcnt(8)
; CHECK-NEXT: ds_store_b128 v0, v[94:97] offset:368
; CHECK-NEXT: s_waitcnt vmcnt(7)
; CHECK-NEXT: ds_store_b128 v0, v[98:101] offset:384
; CHECK-NEXT: s_waitcnt vmcnt(6)
; CHECK-NEXT: ds_store_b128 v0, v[102:105] offset:400
; CHECK-NEXT: s_waitcnt vmcnt(5)
; CHECK-NEXT: ds_store_b128 v0, v[106:109] offset:416
; CHECK-NEXT: s_waitcnt vmcnt(4)
; CHECK-NEXT: ds_store_b128 v0, v[110:113] offset:432
; CHECK-NEXT: s_waitcnt vmcnt(3)
; CHECK-NEXT: ds_store_b128 v0, v[114:117] offset:448
; CHECK-NEXT: s_waitcnt vmcnt(2)
; CHECK-NEXT: ds_store_b128 v0, v[118:121] offset:464
; CHECK-NEXT: s_waitcnt vmcnt(1)
; CHECK-NEXT: ds_store_b128 v0, v[122:125] offset:480
; CHECK-NEXT: s_waitcnt vmcnt(0)
; CHECK-NEXT: ds_store_b128 v0, v[126:129] offset:496
; CHECK-NEXT: s_endpgm
call amdgpu_gfx void @copy(ptr addrspace(1) %src, i32 0)
call amdgpu_gfx void @copy(ptr addrspace(1) %src, i32 1)
call amdgpu_gfx void @copy(ptr addrspace(1) %src, i32 2)
call amdgpu_gfx void @copy(ptr addrspace(1) %src, i32 3)
call amdgpu_gfx void @copy(ptr addrspace(1) %src, i32 4)
call amdgpu_gfx void @copy(ptr addrspace(1) %src, i32 5)
call amdgpu_gfx void @copy(ptr addrspace(1) %src, i32 6)
call amdgpu_gfx void @copy(ptr addrspace(1) %src, i32 7)
call amdgpu_gfx void @copy(ptr addrspace(1) %src, i32 8)
call amdgpu_gfx void @copy(ptr addrspace(1) %src, i32 9)
call amdgpu_gfx void @copy(ptr addrspace(1) %src, i32 10)
call amdgpu_gfx void @copy(ptr addrspace(1) %src, i32 11)
call amdgpu_gfx void @copy(ptr addrspace(1) %src, i32 12)
call amdgpu_gfx void @copy(ptr addrspace(1) %src, i32 13)
call amdgpu_gfx void @copy(ptr addrspace(1) %src, i32 14)
call amdgpu_gfx void @copy(ptr addrspace(1) %src, i32 15)
call amdgpu_gfx void @copy(ptr addrspace(1) %src, i32 16)
call amdgpu_gfx void @copy(ptr addrspace(1) %src, i32 17)
call amdgpu_gfx void @copy(ptr addrspace(1) %src, i32 18)
call amdgpu_gfx void @copy(ptr addrspace(1) %src, i32 19)
call amdgpu_gfx void @copy(ptr addrspace(1) %src, i32 20)
call amdgpu_gfx void @copy(ptr addrspace(1) %src, i32 21)
call amdgpu_gfx void @copy(ptr addrspace(1) %src, i32 22)
call amdgpu_gfx void @copy(ptr addrspace(1) %src, i32 23)
call amdgpu_gfx void @copy(ptr addrspace(1) %src, i32 24)
call amdgpu_gfx void @copy(ptr addrspace(1) %src, i32 25)
call amdgpu_gfx void @copy(ptr addrspace(1) %src, i32 26)
call amdgpu_gfx void @copy(ptr addrspace(1) %src, i32 27)
call amdgpu_gfx void @copy(ptr addrspace(1) %src, i32 28)
call amdgpu_gfx void @copy(ptr addrspace(1) %src, i32 29)
call amdgpu_gfx void @copy(ptr addrspace(1) %src, i32 30)
call amdgpu_gfx void @copy(ptr addrspace(1) %src, i32 31)
ret void
}
|
