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
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
|
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
; RUN: llc -mattr=+sve,+bf16 < %s | FileCheck %s -check-prefixes=CHECK,SVE
; RUN: llc -mattr=+sme -force-streaming < %s | FileCheck %s -check-prefixes=CHECK,SME
target triple = "aarch64-unknown-linux-gnu"
; FADDV
define bfloat @faddv_nxv2bf16(<vscale x 2 x bfloat> %a) {
; CHECK-LABEL: faddv_nxv2bf16:
; CHECK: // %bb.0:
; CHECK-NEXT: lsl z0.s, z0.s, #16
; CHECK-NEXT: ptrue p0.d
; CHECK-NEXT: faddv s0, p0, z0.s
; CHECK-NEXT: bfcvt h0, s0
; CHECK-NEXT: ret
%res = call fast bfloat @llvm.vector.reduce.fadd.nxv2bf16(bfloat zeroinitializer, <vscale x 2 x bfloat> %a)
ret bfloat %res
}
define bfloat @faddv_nxv4bf16(<vscale x 4 x bfloat> %a) {
; CHECK-LABEL: faddv_nxv4bf16:
; CHECK: // %bb.0:
; CHECK-NEXT: lsl z0.s, z0.s, #16
; CHECK-NEXT: ptrue p0.s
; CHECK-NEXT: faddv s0, p0, z0.s
; CHECK-NEXT: bfcvt h0, s0
; CHECK-NEXT: ret
%res = call fast bfloat @llvm.vector.reduce.fadd.nxv4bf16(bfloat zeroinitializer, <vscale x 4 x bfloat> %a)
ret bfloat %res
}
define bfloat @faddv_nxv8bf16(<vscale x 8 x bfloat> %a) {
; CHECK-LABEL: faddv_nxv8bf16:
; CHECK: // %bb.0:
; CHECK-NEXT: uunpkhi z1.s, z0.h
; CHECK-NEXT: uunpklo z0.s, z0.h
; CHECK-NEXT: ptrue p0.s
; CHECK-NEXT: lsl z1.s, z1.s, #16
; CHECK-NEXT: lsl z0.s, z0.s, #16
; CHECK-NEXT: fadd z0.s, z0.s, z1.s
; CHECK-NEXT: faddv s0, p0, z0.s
; CHECK-NEXT: bfcvt h0, s0
; CHECK-NEXT: ret
%res = call fast bfloat @llvm.vector.reduce.fadd.nxv8bf16(bfloat zeroinitializer, <vscale x 8 x bfloat> %a)
ret bfloat %res
}
; FMAXNMV
define bfloat @fmaxv_nxv2bf16(<vscale x 2 x bfloat> %a) {
; CHECK-LABEL: fmaxv_nxv2bf16:
; CHECK: // %bb.0:
; CHECK-NEXT: lsl z0.s, z0.s, #16
; CHECK-NEXT: ptrue p0.d
; CHECK-NEXT: fmaxnmv s0, p0, z0.s
; CHECK-NEXT: bfcvt h0, s0
; CHECK-NEXT: ret
%res = call bfloat @llvm.vector.reduce.fmax.nxv2bf16(<vscale x 2 x bfloat> %a)
ret bfloat %res
}
define bfloat @fmaxv_nxv4bf16(<vscale x 4 x bfloat> %a) {
; CHECK-LABEL: fmaxv_nxv4bf16:
; CHECK: // %bb.0:
; CHECK-NEXT: lsl z0.s, z0.s, #16
; CHECK-NEXT: ptrue p0.s
; CHECK-NEXT: fmaxnmv s0, p0, z0.s
; CHECK-NEXT: bfcvt h0, s0
; CHECK-NEXT: ret
%res = call bfloat @llvm.vector.reduce.fmax.nxv4bf16(<vscale x 4 x bfloat> %a)
ret bfloat %res
}
define bfloat @fmaxv_nxv8bf16(<vscale x 8 x bfloat> %a) {
; CHECK-LABEL: fmaxv_nxv8bf16:
; CHECK: // %bb.0:
; CHECK-NEXT: uunpkhi z1.s, z0.h
; CHECK-NEXT: uunpklo z0.s, z0.h
; CHECK-NEXT: ptrue p0.s
; CHECK-NEXT: lsl z1.s, z1.s, #16
; CHECK-NEXT: lsl z0.s, z0.s, #16
; CHECK-NEXT: fmaxnm z0.s, p0/m, z0.s, z1.s
; CHECK-NEXT: fmaxnmv s0, p0, z0.s
; CHECK-NEXT: bfcvt h0, s0
; CHECK-NEXT: ret
%res = call bfloat @llvm.vector.reduce.fmax.nxv8bf16(<vscale x 8 x bfloat> %a)
ret bfloat %res
}
; FMINNMV
define bfloat @fminv_nxv2bf16(<vscale x 2 x bfloat> %a) {
; CHECK-LABEL: fminv_nxv2bf16:
; CHECK: // %bb.0:
; CHECK-NEXT: lsl z0.s, z0.s, #16
; CHECK-NEXT: ptrue p0.d
; CHECK-NEXT: fminnmv s0, p0, z0.s
; CHECK-NEXT: bfcvt h0, s0
; CHECK-NEXT: ret
%res = call bfloat @llvm.vector.reduce.fmin.nxv2bf16(<vscale x 2 x bfloat> %a)
ret bfloat %res
}
define bfloat @fminv_nxv4bf16(<vscale x 4 x bfloat> %a) {
; CHECK-LABEL: fminv_nxv4bf16:
; CHECK: // %bb.0:
; CHECK-NEXT: lsl z0.s, z0.s, #16
; CHECK-NEXT: ptrue p0.s
; CHECK-NEXT: fminnmv s0, p0, z0.s
; CHECK-NEXT: bfcvt h0, s0
; CHECK-NEXT: ret
%res = call bfloat @llvm.vector.reduce.fmin.nxv4bf16(<vscale x 4 x bfloat> %a)
ret bfloat %res
}
define bfloat @fminv_nxv8bf16(<vscale x 8 x bfloat> %a) {
; CHECK-LABEL: fminv_nxv8bf16:
; CHECK: // %bb.0:
; CHECK-NEXT: uunpkhi z1.s, z0.h
; CHECK-NEXT: uunpklo z0.s, z0.h
; CHECK-NEXT: ptrue p0.s
; CHECK-NEXT: lsl z1.s, z1.s, #16
; CHECK-NEXT: lsl z0.s, z0.s, #16
; CHECK-NEXT: fminnm z0.s, p0/m, z0.s, z1.s
; CHECK-NEXT: fminnmv s0, p0, z0.s
; CHECK-NEXT: bfcvt h0, s0
; CHECK-NEXT: ret
%res = call bfloat @llvm.vector.reduce.fmin.nxv8bf16(<vscale x 8 x bfloat> %a)
ret bfloat %res
}
; FMAXV
define bfloat @fmaximumv_nxv2bf16(<vscale x 2 x bfloat> %a) {
; CHECK-LABEL: fmaximumv_nxv2bf16:
; CHECK: // %bb.0:
; CHECK-NEXT: lsl z0.s, z0.s, #16
; CHECK-NEXT: ptrue p0.d
; CHECK-NEXT: fmaxv s0, p0, z0.s
; CHECK-NEXT: bfcvt h0, s0
; CHECK-NEXT: ret
%res = call bfloat @llvm.vector.reduce.fmaximum.nxv2bf16(<vscale x 2 x bfloat> %a)
ret bfloat %res
}
define bfloat @fmaximumv_nxv4bf16(<vscale x 4 x bfloat> %a) {
; CHECK-LABEL: fmaximumv_nxv4bf16:
; CHECK: // %bb.0:
; CHECK-NEXT: lsl z0.s, z0.s, #16
; CHECK-NEXT: ptrue p0.s
; CHECK-NEXT: fmaxv s0, p0, z0.s
; CHECK-NEXT: bfcvt h0, s0
; CHECK-NEXT: ret
%res = call bfloat @llvm.vector.reduce.fmaximum.nxv4bf16(<vscale x 4 x bfloat> %a)
ret bfloat %res
}
define bfloat @fmaximumv_nxv8bf16(<vscale x 8 x bfloat> %a) {
; CHECK-LABEL: fmaximumv_nxv8bf16:
; CHECK: // %bb.0:
; CHECK-NEXT: uunpkhi z1.s, z0.h
; CHECK-NEXT: uunpklo z0.s, z0.h
; CHECK-NEXT: ptrue p0.s
; CHECK-NEXT: lsl z1.s, z1.s, #16
; CHECK-NEXT: lsl z0.s, z0.s, #16
; CHECK-NEXT: fmax z0.s, p0/m, z0.s, z1.s
; CHECK-NEXT: fmaxv s0, p0, z0.s
; CHECK-NEXT: bfcvt h0, s0
; CHECK-NEXT: ret
%res = call bfloat @llvm.vector.reduce.fmaximum.nxv8bf16(<vscale x 8 x bfloat> %a)
ret bfloat %res
}
; FMINV
define bfloat @fminimumv_nxv2bf16(<vscale x 2 x bfloat> %a) {
; CHECK-LABEL: fminimumv_nxv2bf16:
; CHECK: // %bb.0:
; CHECK-NEXT: lsl z0.s, z0.s, #16
; CHECK-NEXT: ptrue p0.d
; CHECK-NEXT: fminv s0, p0, z0.s
; CHECK-NEXT: bfcvt h0, s0
; CHECK-NEXT: ret
%res = call bfloat @llvm.vector.reduce.fminimum.nxv2bf16(<vscale x 2 x bfloat> %a)
ret bfloat %res
}
define bfloat @fminimumv_nxv4bf16(<vscale x 4 x bfloat> %a) {
; CHECK-LABEL: fminimumv_nxv4bf16:
; CHECK: // %bb.0:
; CHECK-NEXT: lsl z0.s, z0.s, #16
; CHECK-NEXT: ptrue p0.s
; CHECK-NEXT: fminv s0, p0, z0.s
; CHECK-NEXT: bfcvt h0, s0
; CHECK-NEXT: ret
%res = call bfloat @llvm.vector.reduce.fminimum.nxv4bf16(<vscale x 4 x bfloat> %a)
ret bfloat %res
}
define bfloat @fminimumv_nxv8bf16(<vscale x 8 x bfloat> %a) {
; CHECK-LABEL: fminimumv_nxv8bf16:
; CHECK: // %bb.0:
; CHECK-NEXT: uunpkhi z1.s, z0.h
; CHECK-NEXT: uunpklo z0.s, z0.h
; CHECK-NEXT: ptrue p0.s
; CHECK-NEXT: lsl z1.s, z1.s, #16
; CHECK-NEXT: lsl z0.s, z0.s, #16
; CHECK-NEXT: fmin z0.s, p0/m, z0.s, z1.s
; CHECK-NEXT: fminv s0, p0, z0.s
; CHECK-NEXT: bfcvt h0, s0
; CHECK-NEXT: ret
%res = call bfloat @llvm.vector.reduce.fminimum.nxv8bf16(<vscale x 8 x bfloat> %a)
ret bfloat %res
}
; The reduction is performed at a higher precision. Because add operations
; can utilise that precision, its result must be rounded even if it's then
; promoted.
define float @promoted_fadd(<vscale x 4 x bfloat> %a) {
; SVE-LABEL: promoted_fadd:
; SVE: // %bb.0:
; SVE-NEXT: lsl z0.s, z0.s, #16
; SVE-NEXT: ptrue p0.s
; SVE-NEXT: faddv s0, p0, z0.s
; SVE-NEXT: bfcvt h0, s0
; SVE-NEXT: shll v0.4s, v0.4h, #16
; SVE-NEXT: // kill: def $s0 killed $s0 killed $q0
; SVE-NEXT: ret
;
; SME-LABEL: promoted_fadd:
; SME: // %bb.0:
; SME-NEXT: lsl z0.s, z0.s, #16
; SME-NEXT: ptrue p0.s
; SME-NEXT: faddv s0, p0, z0.s
; SME-NEXT: bfcvt h0, s0
; SME-NEXT: fmov w8, s0
; SME-NEXT: lsl w8, w8, #16
; SME-NEXT: fmov s0, w8
; SME-NEXT: ret
%rdx = call fast bfloat @llvm.vector.reduce.fadd.nxv4bf16(bfloat zeroinitializer, <vscale x 4 x bfloat> %a)
%res = fpext bfloat %rdx to float
ret float %res
}
; The reduction is performed at a higher precision. Because min/max operations
; don't utilise that precision, its result can be used directly.
define float @promoted_fmax(<vscale x 4 x bfloat> %a) {
; CHECK-LABEL: promoted_fmax:
; CHECK: // %bb.0:
; CHECK-NEXT: lsl z0.s, z0.s, #16
; CHECK-NEXT: ptrue p0.s
; CHECK-NEXT: fmaxnmv s0, p0, z0.s
; CHECK-NEXT: // kill: def $s0 killed $s0 killed $z0
; CHECK-NEXT: ret
%rdx = call bfloat @llvm.vector.reduce.fmax.nxv4bf16(<vscale x 4 x bfloat> %a)
%res = fpext bfloat %rdx to float
ret float %res
}
declare bfloat @llvm.vector.reduce.fadd.nxv2bf16(bfloat, <vscale x 2 x bfloat>)
declare bfloat @llvm.vector.reduce.fadd.nxv4bf16(bfloat, <vscale x 4 x bfloat>)
declare bfloat @llvm.vector.reduce.fadd.nxv8bf16(bfloat, <vscale x 8 x bfloat>)
declare bfloat @llvm.vector.reduce.fmax.nxv2bf16(<vscale x 2 x bfloat>)
declare bfloat @llvm.vector.reduce.fmax.nxv4bf16(<vscale x 4 x bfloat>)
declare bfloat @llvm.vector.reduce.fmax.nxv8bf16(<vscale x 8 x bfloat>)
declare bfloat @llvm.vector.reduce.fmin.nxv2bf16(<vscale x 2 x bfloat>)
declare bfloat @llvm.vector.reduce.fmin.nxv4bf16(<vscale x 4 x bfloat>)
declare bfloat @llvm.vector.reduce.fmin.nxv8bf16(<vscale x 8 x bfloat>)
declare bfloat @llvm.vector.reduce.fmaximum.nxv2bf16(<vscale x 2 x bfloat>)
declare bfloat @llvm.vector.reduce.fmaximum.nxv4bf16(<vscale x 4 x bfloat>)
declare bfloat @llvm.vector.reduce.fmaximum.nxv8bf16(<vscale x 8 x bfloat>)
declare bfloat @llvm.vector.reduce.fminimum.nxv2bf16(<vscale x 2 x bfloat>)
declare bfloat @llvm.vector.reduce.fminimum.nxv4bf16(<vscale x 4 x bfloat>)
declare bfloat @llvm.vector.reduce.fminimum.nxv8bf16(<vscale x 8 x bfloat>)
|
