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
|
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
; RUN: llc < %s -mtriple=nvptx64 -mcpu=sm_90 -mattr=+ptx80| FileCheck --check-prefixes=CHECK,CHECK-PTX64 %s
; RUN: llc < %s -mtriple=nvptx64 -mcpu=sm_90 -mattr=+ptx80 --nvptx-short-ptr| FileCheck --check-prefixes=CHECK,CHECK-PTX-SHARED32 %s
; RUN: %if ptxas-12.3 %{ llc < %s -mtriple=nvptx64 -mcpu=sm_90 -mattr=+ptx80| %ptxas-verify -arch=sm_90 %}
; RUN: %if ptxas-12.3 %{ llc < %s -mtriple=nvptx64 -mcpu=sm_90 -mattr=+ptx80 --nvptx-short-ptr| %ptxas-verify -arch=sm_90 %}
target triple = "nvptx64-nvidia-cuda"
declare void @llvm.nvvm.cp.async.bulk.global.to.shared.cluster(ptr addrspace(7), ptr addrspace(3), ptr addrspace(1), i32, i16, i64, i1, i1)
declare void @llvm.nvvm.cp.async.bulk.shared.cta.to.global(ptr addrspace(1), ptr addrspace(3), i32, i64, i1)
declare void @llvm.nvvm.cp.async.bulk.shared.cta.to.cluster(ptr addrspace(7), ptr addrspace(3), ptr addrspace(3), i32)
declare void @llvm.nvvm.cp.async.bulk.prefetch.L2(ptr addrspace(1), i32, i64, i1)
define void @cp_async_bulk_g2s(ptr addrspace(1) %src, ptr addrspace(3) %bar, ptr addrspace(7) %dst, i32 %size, i16 %mc, i64 %ch) {
; CHECK-PTX64-LABEL: cp_async_bulk_g2s(
; CHECK-PTX64: {
; CHECK-PTX64-NEXT: .reg .b16 %rs<2>;
; CHECK-PTX64-NEXT: .reg .b32 %r<2>;
; CHECK-PTX64-NEXT: .reg .b64 %rd<5>;
; CHECK-PTX64-EMPTY:
; CHECK-PTX64-NEXT: // %bb.0:
; CHECK-PTX64-NEXT: ld.param.b64 %rd1, [cp_async_bulk_g2s_param_0];
; CHECK-PTX64-NEXT: ld.param.b64 %rd2, [cp_async_bulk_g2s_param_1];
; CHECK-PTX64-NEXT: ld.param.b64 %rd3, [cp_async_bulk_g2s_param_2];
; CHECK-PTX64-NEXT: ld.param.b32 %r1, [cp_async_bulk_g2s_param_3];
; CHECK-PTX64-NEXT: ld.param.b16 %rs1, [cp_async_bulk_g2s_param_4];
; CHECK-PTX64-NEXT: ld.param.b64 %rd4, [cp_async_bulk_g2s_param_5];
; CHECK-PTX64-NEXT: cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes [%rd3], [%rd1], %r1, [%rd2];
; CHECK-PTX64-NEXT: cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes.L2::cache_hint [%rd3], [%rd1], %r1, [%rd2], %rd4;
; CHECK-PTX64-NEXT: cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes.multicast::cluster [%rd3], [%rd1], %r1, [%rd2], %rs1;
; CHECK-PTX64-NEXT: cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes.multicast::cluster.L2::cache_hint [%rd3], [%rd1], %r1, [%rd2], %rs1, %rd4;
; CHECK-PTX64-NEXT: ret;
;
; CHECK-PTX-SHARED32-LABEL: cp_async_bulk_g2s(
; CHECK-PTX-SHARED32: {
; CHECK-PTX-SHARED32-NEXT: .reg .b16 %rs<2>;
; CHECK-PTX-SHARED32-NEXT: .reg .b32 %r<4>;
; CHECK-PTX-SHARED32-NEXT: .reg .b64 %rd<3>;
; CHECK-PTX-SHARED32-EMPTY:
; CHECK-PTX-SHARED32-NEXT: // %bb.0:
; CHECK-PTX-SHARED32-NEXT: ld.param.b64 %rd1, [cp_async_bulk_g2s_param_0];
; CHECK-PTX-SHARED32-NEXT: ld.param.b32 %r1, [cp_async_bulk_g2s_param_1];
; CHECK-PTX-SHARED32-NEXT: ld.param.b32 %r2, [cp_async_bulk_g2s_param_2];
; CHECK-PTX-SHARED32-NEXT: ld.param.b32 %r3, [cp_async_bulk_g2s_param_3];
; CHECK-PTX-SHARED32-NEXT: ld.param.b16 %rs1, [cp_async_bulk_g2s_param_4];
; CHECK-PTX-SHARED32-NEXT: ld.param.b64 %rd2, [cp_async_bulk_g2s_param_5];
; CHECK-PTX-SHARED32-NEXT: cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes [%r2], [%rd1], %r3, [%r1];
; CHECK-PTX-SHARED32-NEXT: cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes.L2::cache_hint [%r2], [%rd1], %r3, [%r1], %rd2;
; CHECK-PTX-SHARED32-NEXT: cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes.multicast::cluster [%r2], [%rd1], %r3, [%r1], %rs1;
; CHECK-PTX-SHARED32-NEXT: cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes.multicast::cluster.L2::cache_hint [%r2], [%rd1], %r3, [%r1], %rs1, %rd2;
; CHECK-PTX-SHARED32-NEXT: ret;
tail call void @llvm.nvvm.cp.async.bulk.global.to.shared.cluster(ptr addrspace(7) %dst, ptr addrspace(3) %bar, ptr addrspace(1) %src, i32 %size, i16 %mc, i64 %ch, i1 0, i1 0)
tail call void @llvm.nvvm.cp.async.bulk.global.to.shared.cluster(ptr addrspace(7) %dst, ptr addrspace(3) %bar, ptr addrspace(1) %src, i32 %size, i16 %mc, i64 %ch, i1 0, i1 1)
tail call void @llvm.nvvm.cp.async.bulk.global.to.shared.cluster(ptr addrspace(7) %dst, ptr addrspace(3) %bar, ptr addrspace(1) %src, i32 %size, i16 %mc, i64 %ch, i1 1, i1 0)
tail call void @llvm.nvvm.cp.async.bulk.global.to.shared.cluster(ptr addrspace(7) %dst, ptr addrspace(3) %bar, ptr addrspace(1) %src, i32 %size, i16 %mc, i64 %ch, i1 1, i1 1)
ret void
}
; Tests to verify that the immediate values for "mc and ch" land correctly in PTX.
; The values of 16 and 64 are arbitrary and do not have any significance.
define void @cp_async_bulk_g2s_imm_mc_ch(ptr addrspace(1) %src, ptr addrspace(3) %bar, ptr addrspace(7) %dst, i32 %size, i16 %mc, i64 %ch) {
; CHECK-PTX64-LABEL: cp_async_bulk_g2s_imm_mc_ch(
; CHECK-PTX64: {
; CHECK-PTX64-NEXT: .reg .b16 %rs<3>;
; CHECK-PTX64-NEXT: .reg .b32 %r<2>;
; CHECK-PTX64-NEXT: .reg .b64 %rd<6>;
; CHECK-PTX64-EMPTY:
; CHECK-PTX64-NEXT: // %bb.0:
; CHECK-PTX64-NEXT: ld.param.b64 %rd1, [cp_async_bulk_g2s_imm_mc_ch_param_0];
; CHECK-PTX64-NEXT: ld.param.b64 %rd2, [cp_async_bulk_g2s_imm_mc_ch_param_1];
; CHECK-PTX64-NEXT: ld.param.b64 %rd3, [cp_async_bulk_g2s_imm_mc_ch_param_2];
; CHECK-PTX64-NEXT: ld.param.b32 %r1, [cp_async_bulk_g2s_imm_mc_ch_param_3];
; CHECK-PTX64-NEXT: mov.b64 %rd4, 64;
; CHECK-PTX64-NEXT: mov.b16 %rs1, 16;
; CHECK-PTX64-NEXT: cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes.multicast::cluster.L2::cache_hint [%rd3], [%rd1], %r1, [%rd2], %rs1, %rd4;
; CHECK-PTX64-NEXT: ld.param.b64 %rd5, [cp_async_bulk_g2s_imm_mc_ch_param_5];
; CHECK-PTX64-NEXT: cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes.multicast::cluster [%rd3], [%rd1], %r1, [%rd2], %rs1;
; CHECK-PTX64-NEXT: ld.param.b16 %rs2, [cp_async_bulk_g2s_imm_mc_ch_param_4];
; CHECK-PTX64-NEXT: cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes.L2::cache_hint [%rd3], [%rd1], %r1, [%rd2], %rd4;
; CHECK-PTX64-NEXT: ret;
;
; CHECK-PTX-SHARED32-LABEL: cp_async_bulk_g2s_imm_mc_ch(
; CHECK-PTX-SHARED32: {
; CHECK-PTX-SHARED32-NEXT: .reg .b16 %rs<3>;
; CHECK-PTX-SHARED32-NEXT: .reg .b32 %r<4>;
; CHECK-PTX-SHARED32-NEXT: .reg .b64 %rd<4>;
; CHECK-PTX-SHARED32-EMPTY:
; CHECK-PTX-SHARED32-NEXT: // %bb.0:
; CHECK-PTX-SHARED32-NEXT: ld.param.b64 %rd1, [cp_async_bulk_g2s_imm_mc_ch_param_0];
; CHECK-PTX-SHARED32-NEXT: ld.param.b32 %r1, [cp_async_bulk_g2s_imm_mc_ch_param_1];
; CHECK-PTX-SHARED32-NEXT: ld.param.b32 %r2, [cp_async_bulk_g2s_imm_mc_ch_param_2];
; CHECK-PTX-SHARED32-NEXT: ld.param.b32 %r3, [cp_async_bulk_g2s_imm_mc_ch_param_3];
; CHECK-PTX-SHARED32-NEXT: mov.b64 %rd2, 64;
; CHECK-PTX-SHARED32-NEXT: mov.b16 %rs1, 16;
; CHECK-PTX-SHARED32-NEXT: cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes.multicast::cluster.L2::cache_hint [%r2], [%rd1], %r3, [%r1], %rs1, %rd2;
; CHECK-PTX-SHARED32-NEXT: ld.param.b64 %rd3, [cp_async_bulk_g2s_imm_mc_ch_param_5];
; CHECK-PTX-SHARED32-NEXT: cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes.multicast::cluster [%r2], [%rd1], %r3, [%r1], %rs1;
; CHECK-PTX-SHARED32-NEXT: ld.param.b16 %rs2, [cp_async_bulk_g2s_imm_mc_ch_param_4];
; CHECK-PTX-SHARED32-NEXT: cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes.L2::cache_hint [%r2], [%rd1], %r3, [%r1], %rd2;
; CHECK-PTX-SHARED32-NEXT: ret;
tail call void @llvm.nvvm.cp.async.bulk.global.to.shared.cluster(ptr addrspace(7) %dst, ptr addrspace(3) %bar, ptr addrspace(1) %src, i32 %size, i16 16, i64 64, i1 1, i1 1)
tail call void @llvm.nvvm.cp.async.bulk.global.to.shared.cluster(ptr addrspace(7) %dst, ptr addrspace(3) %bar, ptr addrspace(1) %src, i32 %size, i16 16, i64 %ch, i1 1, i1 0)
tail call void @llvm.nvvm.cp.async.bulk.global.to.shared.cluster(ptr addrspace(7) %dst, ptr addrspace(3) %bar, ptr addrspace(1) %src, i32 %size, i16 %mc, i64 64, i1 0, i1 1)
ret void
}
define void @cp_async_bulk_s2g(ptr addrspace(3) %src, ptr addrspace(1) %dst, i32 %size, i64 %ch) {
; CHECK-PTX64-LABEL: cp_async_bulk_s2g(
; CHECK-PTX64: {
; CHECK-PTX64-NEXT: .reg .b32 %r<2>;
; CHECK-PTX64-NEXT: .reg .b64 %rd<5>;
; CHECK-PTX64-EMPTY:
; CHECK-PTX64-NEXT: // %bb.0:
; CHECK-PTX64-NEXT: ld.param.b64 %rd1, [cp_async_bulk_s2g_param_0];
; CHECK-PTX64-NEXT: ld.param.b64 %rd2, [cp_async_bulk_s2g_param_1];
; CHECK-PTX64-NEXT: ld.param.b32 %r1, [cp_async_bulk_s2g_param_2];
; CHECK-PTX64-NEXT: mov.b64 %rd3, 64;
; CHECK-PTX64-NEXT: cp.async.bulk.global.shared::cta.bulk_group.L2::cache_hint [%rd2], [%rd1], %r1, %rd3;
; CHECK-PTX64-NEXT: ld.param.b64 %rd4, [cp_async_bulk_s2g_param_3];
; CHECK-PTX64-NEXT: cp.async.bulk.global.shared::cta.bulk_group [%rd2], [%rd1], %r1;
; CHECK-PTX64-NEXT: cp.async.bulk.global.shared::cta.bulk_group.L2::cache_hint [%rd2], [%rd1], %r1, %rd4;
; CHECK-PTX64-NEXT: ret;
;
; CHECK-PTX-SHARED32-LABEL: cp_async_bulk_s2g(
; CHECK-PTX-SHARED32: {
; CHECK-PTX-SHARED32-NEXT: .reg .b32 %r<3>;
; CHECK-PTX-SHARED32-NEXT: .reg .b64 %rd<4>;
; CHECK-PTX-SHARED32-EMPTY:
; CHECK-PTX-SHARED32-NEXT: // %bb.0:
; CHECK-PTX-SHARED32-NEXT: ld.param.b32 %r1, [cp_async_bulk_s2g_param_0];
; CHECK-PTX-SHARED32-NEXT: ld.param.b64 %rd1, [cp_async_bulk_s2g_param_1];
; CHECK-PTX-SHARED32-NEXT: ld.param.b32 %r2, [cp_async_bulk_s2g_param_2];
; CHECK-PTX-SHARED32-NEXT: mov.b64 %rd2, 64;
; CHECK-PTX-SHARED32-NEXT: cp.async.bulk.global.shared::cta.bulk_group.L2::cache_hint [%rd1], [%r1], %r2, %rd2;
; CHECK-PTX-SHARED32-NEXT: ld.param.b64 %rd3, [cp_async_bulk_s2g_param_3];
; CHECK-PTX-SHARED32-NEXT: cp.async.bulk.global.shared::cta.bulk_group [%rd1], [%r1], %r2;
; CHECK-PTX-SHARED32-NEXT: cp.async.bulk.global.shared::cta.bulk_group.L2::cache_hint [%rd1], [%r1], %r2, %rd3;
; CHECK-PTX-SHARED32-NEXT: ret;
tail call void @llvm.nvvm.cp.async.bulk.shared.cta.to.global(ptr addrspace(1) %dst, ptr addrspace(3) %src, i32 %size, i64 64, i1 1)
tail call void @llvm.nvvm.cp.async.bulk.shared.cta.to.global(ptr addrspace(1) %dst, ptr addrspace(3) %src, i32 %size, i64 %ch, i1 0)
tail call void @llvm.nvvm.cp.async.bulk.shared.cta.to.global(ptr addrspace(1) %dst, ptr addrspace(3) %src, i32 %size, i64 %ch, i1 1)
ret void
}
define void @cp_async_bulk_cta_to_cluster(ptr addrspace(3) %src, ptr addrspace(3) %bar, ptr addrspace(7) %dst, i32 %size) {
; CHECK-PTX64-LABEL: cp_async_bulk_cta_to_cluster(
; CHECK-PTX64: {
; CHECK-PTX64-NEXT: .reg .b32 %r<2>;
; CHECK-PTX64-NEXT: .reg .b64 %rd<4>;
; CHECK-PTX64-EMPTY:
; CHECK-PTX64-NEXT: // %bb.0:
; CHECK-PTX64-NEXT: ld.param.b64 %rd1, [cp_async_bulk_cta_to_cluster_param_0];
; CHECK-PTX64-NEXT: ld.param.b64 %rd2, [cp_async_bulk_cta_to_cluster_param_1];
; CHECK-PTX64-NEXT: ld.param.b64 %rd3, [cp_async_bulk_cta_to_cluster_param_2];
; CHECK-PTX64-NEXT: ld.param.b32 %r1, [cp_async_bulk_cta_to_cluster_param_3];
; CHECK-PTX64-NEXT: cp.async.bulk.shared::cluster.shared::cta.mbarrier::complete_tx::bytes [%rd3], [%rd1], %r1, [%rd2];
; CHECK-PTX64-NEXT: ret;
;
; CHECK-PTX-SHARED32-LABEL: cp_async_bulk_cta_to_cluster(
; CHECK-PTX-SHARED32: {
; CHECK-PTX-SHARED32-NEXT: .reg .b32 %r<5>;
; CHECK-PTX-SHARED32-EMPTY:
; CHECK-PTX-SHARED32-NEXT: // %bb.0:
; CHECK-PTX-SHARED32-NEXT: ld.param.b32 %r1, [cp_async_bulk_cta_to_cluster_param_0];
; CHECK-PTX-SHARED32-NEXT: ld.param.b32 %r2, [cp_async_bulk_cta_to_cluster_param_1];
; CHECK-PTX-SHARED32-NEXT: ld.param.b32 %r3, [cp_async_bulk_cta_to_cluster_param_2];
; CHECK-PTX-SHARED32-NEXT: ld.param.b32 %r4, [cp_async_bulk_cta_to_cluster_param_3];
; CHECK-PTX-SHARED32-NEXT: cp.async.bulk.shared::cluster.shared::cta.mbarrier::complete_tx::bytes [%r3], [%r1], %r4, [%r2];
; CHECK-PTX-SHARED32-NEXT: ret;
tail call void @llvm.nvvm.cp.async.bulk.shared.cta.to.cluster(ptr addrspace(7) %dst, ptr addrspace(3) %bar, ptr addrspace(3) %src, i32 %size)
ret void
}
define void @cp_async_bulk_prefetch(ptr addrspace(1) %src, i32 %size, i64 %ch) {
; CHECK-LABEL: cp_async_bulk_prefetch(
; CHECK: {
; CHECK-NEXT: .reg .b32 %r<2>;
; CHECK-NEXT: .reg .b64 %rd<4>;
; CHECK-EMPTY:
; CHECK-NEXT: // %bb.0:
; CHECK-NEXT: ld.param.b64 %rd1, [cp_async_bulk_prefetch_param_0];
; CHECK-NEXT: ld.param.b32 %r1, [cp_async_bulk_prefetch_param_1];
; CHECK-NEXT: mov.b64 %rd2, 64;
; CHECK-NEXT: cp.async.bulk.prefetch.L2.global.L2::cache_hint [%rd1], %r1, %rd2;
; CHECK-NEXT: ld.param.b64 %rd3, [cp_async_bulk_prefetch_param_2];
; CHECK-NEXT: cp.async.bulk.prefetch.L2.global.L2::cache_hint [%rd1], %r1, %rd3;
; CHECK-NEXT: cp.async.bulk.prefetch.L2.global [%rd1], %r1;
; CHECK-NEXT: ret;
tail call void @llvm.nvvm.cp.async.bulk.prefetch.L2(ptr addrspace(1) %src, i32 %size, i64 64, i1 1)
tail call void @llvm.nvvm.cp.async.bulk.prefetch.L2(ptr addrspace(1) %src, i32 %size, i64 %ch, i1 1)
tail call void @llvm.nvvm.cp.async.bulk.prefetch.L2(ptr addrspace(1) %src, i32 %size, i64 %ch, i1 0)
ret void
}
|
