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
| ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=kaveri -mattr=-code-object-v3 -disable-promote-alloca-to-vector < %s | FileCheck -check-prefix=GCN %s
; This shows that the amount of LDS estimate is sensitive to the order
; of the LDS globals.
; Both of these functions use the same amount of LDS, but the total
; changes depending on the visit order of first use.
; The one with the suboptimal order resulting in extra padding exceeds
; the desired limit
; The padding estimate heuristic used by the promote alloca pass
; is mostly determined by the order of the globals,
; Raw usage = 1060 bytes
; Rounded usage:
; 292 + (4 pad) + 256 + (8 pad) + 512 = 1072
; 512 + (0 pad) + 256 + (0 pad) + 292 = 1060
; At default occupancy guess of 7, 2340 bytes available total.
; 1280 need to be left to promote alloca
; optimally packed, this requires
@lds0 = internal unnamed_addr addrspace(3) global [32 x <4 x i32>] undef, align 16
@lds2 = internal unnamed_addr addrspace(3) global [32 x i64] undef, align 8
@lds1 = internal unnamed_addr addrspace(3) global [73 x i32] undef, align 4
; GCN-LABEL: {{^}}promote_alloca_size_order_0:
; GCN: workgroup_group_segment_byte_size = 2340
define amdgpu_kernel void @promote_alloca_size_order_0(i32 addrspace(1)* nocapture %out, i32 addrspace(1)* nocapture %in, i32 %idx) #0 {
entry:
%stack = alloca [5 x i32], align 4, addrspace(5)
%tmp0 = load i32, i32 addrspace(1)* %in, align 4
%arrayidx1 = getelementptr inbounds [5 x i32], [5 x i32] addrspace(5)* %stack, i32 0, i32 %tmp0
store i32 4, i32 addrspace(5)* %arrayidx1, align 4
%arrayidx2 = getelementptr inbounds i32, i32 addrspace(1)* %in, i32 1
%tmp1 = load i32, i32 addrspace(1)* %arrayidx2, align 4
%arrayidx3 = getelementptr inbounds [5 x i32], [5 x i32] addrspace(5)* %stack, i32 0, i32 %tmp1
store i32 5, i32 addrspace(5)* %arrayidx3, align 4
%arrayidx10 = getelementptr inbounds [5 x i32], [5 x i32] addrspace(5)* %stack, i32 0, i32 0
%tmp2 = load i32, i32 addrspace(5)* %arrayidx10, align 4
store i32 %tmp2, i32 addrspace(1)* %out, align 4
%arrayidx12 = getelementptr inbounds [5 x i32], [5 x i32] addrspace(5)* %stack, i32 0, i32 1
%tmp3 = load i32, i32 addrspace(5)* %arrayidx12
%arrayidx13 = getelementptr inbounds i32, i32 addrspace(1)* %out, i32 1
store i32 %tmp3, i32 addrspace(1)* %arrayidx13
%gep.lds1 = getelementptr inbounds [73 x i32], [73 x i32] addrspace(3)* @lds1, i32 0, i32 %idx
store volatile i32 0, i32 addrspace(3)* %gep.lds1, align 4
%gep.lds2 = getelementptr inbounds [32 x i64], [32 x i64] addrspace(3)* @lds2, i32 0, i32 %idx
store volatile i64 0, i64 addrspace(3)* %gep.lds2, align 8
%gep.lds0 = getelementptr inbounds [32 x <4 x i32>], [32 x <4 x i32>] addrspace(3)* @lds0, i32 0, i32 %idx
store volatile <4 x i32> zeroinitializer, <4 x i32> addrspace(3)* %gep.lds0, align 16
ret void
}
; GCN-LABEL: {{^}}promote_alloca_size_order_1:
; GCN: workgroup_group_segment_byte_size = 2352
define amdgpu_kernel void @promote_alloca_size_order_1(i32 addrspace(1)* nocapture %out, i32 addrspace(1)* nocapture %in, i32 %idx) #0 {
entry:
%stack = alloca [5 x i32], align 4, addrspace(5)
%tmp0 = load i32, i32 addrspace(1)* %in, align 4
%arrayidx1 = getelementptr inbounds [5 x i32], [5 x i32] addrspace(5)* %stack, i32 0, i32 %tmp0
store i32 4, i32 addrspace(5)* %arrayidx1, align 4
%arrayidx2 = getelementptr inbounds i32, i32 addrspace(1)* %in, i32 1
%tmp1 = load i32, i32 addrspace(1)* %arrayidx2, align 4
%arrayidx3 = getelementptr inbounds [5 x i32], [5 x i32] addrspace(5)* %stack, i32 0, i32 %tmp1
store i32 5, i32 addrspace(5)* %arrayidx3, align 4
%arrayidx10 = getelementptr inbounds [5 x i32], [5 x i32] addrspace(5)* %stack, i32 0, i32 0
%tmp2 = load i32, i32 addrspace(5)* %arrayidx10, align 4
store i32 %tmp2, i32 addrspace(1)* %out, align 4
%arrayidx12 = getelementptr inbounds [5 x i32], [5 x i32] addrspace(5)* %stack, i32 0, i32 1
%tmp3 = load i32, i32 addrspace(5)* %arrayidx12
%arrayidx13 = getelementptr inbounds i32, i32 addrspace(1)* %out, i32 1
store i32 %tmp3, i32 addrspace(1)* %arrayidx13
%gep.lds0 = getelementptr inbounds [32 x <4 x i32>], [32 x <4 x i32>] addrspace(3)* @lds0, i32 0, i32 %idx
store volatile <4 x i32> zeroinitializer, <4 x i32> addrspace(3)* %gep.lds0, align 16
%gep.lds2 = getelementptr inbounds [32 x i64], [32 x i64] addrspace(3)* @lds2, i32 0, i32 %idx
store volatile i64 0, i64 addrspace(3)* %gep.lds2, align 8
%gep.lds1 = getelementptr inbounds [73 x i32], [73 x i32] addrspace(3)* @lds1, i32 0, i32 %idx
store volatile i32 0, i32 addrspace(3)* %gep.lds1, align 4
ret void
}
@lds3 = internal unnamed_addr addrspace(3) global [13 x i32] undef, align 4
@lds4 = internal unnamed_addr addrspace(3) global [63 x <4 x i32>] undef, align 16
; The guess from the alignment padding pushes this over the determined
; size limit, so it isn't promoted
; GCN-LABEL: {{^}}promote_alloca_align_pad_guess_over_limit:
; GCN: workgroup_group_segment_byte_size = 1060
define amdgpu_kernel void @promote_alloca_align_pad_guess_over_limit(i32 addrspace(1)* nocapture %out, i32 addrspace(1)* nocapture %in, i32 %idx) #0 {
entry:
%stack = alloca [5 x i32], align 4, addrspace(5)
%tmp0 = load i32, i32 addrspace(1)* %in, align 4
%arrayidx1 = getelementptr inbounds [5 x i32], [5 x i32] addrspace(5)* %stack, i32 0, i32 %tmp0
store i32 4, i32 addrspace(5)* %arrayidx1, align 4
%arrayidx2 = getelementptr inbounds i32, i32 addrspace(1)* %in, i32 1
%tmp1 = load i32, i32 addrspace(1)* %arrayidx2, align 4
%arrayidx3 = getelementptr inbounds [5 x i32], [5 x i32] addrspace(5)* %stack, i32 0, i32 %tmp1
store i32 5, i32 addrspace(5)* %arrayidx3, align 4
%arrayidx10 = getelementptr inbounds [5 x i32], [5 x i32] addrspace(5)* %stack, i32 0, i32 0
%tmp2 = load i32, i32 addrspace(5)* %arrayidx10, align 4
store i32 %tmp2, i32 addrspace(1)* %out, align 4
%arrayidx12 = getelementptr inbounds [5 x i32], [5 x i32] addrspace(5)* %stack, i32 0, i32 1
%tmp3 = load i32, i32 addrspace(5)* %arrayidx12
%arrayidx13 = getelementptr inbounds i32, i32 addrspace(1)* %out, i32 1
store i32 %tmp3, i32 addrspace(1)* %arrayidx13
%gep.lds3 = getelementptr inbounds [13 x i32], [13 x i32] addrspace(3)* @lds3, i32 0, i32 %idx
store volatile i32 0, i32 addrspace(3)* %gep.lds3, align 4
%gep.lds4 = getelementptr inbounds [63 x <4 x i32>], [63 x <4 x i32>] addrspace(3)* @lds4, i32 0, i32 %idx
store volatile <4 x i32> zeroinitializer, <4 x i32> addrspace(3)* %gep.lds4, align 16
ret void
}
attributes #0 = { nounwind "amdgpu-flat-work-group-size"="64,64" "amdgpu-waves-per-eu"="1,7" }
|