24.04
sme_interleave1VL_block4_u8_u8_summing.hpp
Go to the documentation of this file.
1
/*
2
* Copyright (c) 2022-2023 Arm Limited.
3
*
4
* SPDX-License-Identifier: MIT
5
*
6
* Permission is hereby granted, free of charge, to any person obtaining a copy
7
* of this software and associated documentation files (the "Software"), to
8
* deal in the Software without restriction, including without limitation the
9
* rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10
* sell copies of the Software, and to permit persons to whom the Software is
11
* furnished to do so, subject to the following conditions:
12
*
13
* The above copyright notice and this permission notice shall be included in all
14
* copies or substantial portions of the Software.
15
*
16
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22
* SOFTWARE.
23
*/
24
25
#if defined(ARM_COMPUTE_ENABLE_SME)
26
27
template
<>
28
void
interleave_block<1, 4, VLType::SME, true>(
29
uint8_t * &out,
const
uint8_t *
const
*in,
30
size_t
width,
size_t
height,
size_t
row_offset,
bool
first
31
)
32
{
33
__asm__ __volatile__(
34
".inst 0xd503477f // SMSTART ZA\n"
35
"cntb x21\n"
36
"mov x23, %x[width]\n"
37
"mov z18.b, #0x1\n"
38
"incb x23\n"
39
"mov x20, %x[width]\n"
40
"mov z17.s, #0x0\n"
41
"sub x10, x21, #0x1\n"
42
"cntw x9\n"
43
"sub x23, x23, #0x1\n"
44
"ands x10, x20, x10\n"
45
"udiv x23, x23, x21\n"
// n_passes = ceildiv(width, VL<T>)
46
"csel x10, x10, x21, NE\n"
47
"lsl x22, %x[height], #0x1\n"
// height * 2
48
"lsl x21, x9, #0x1\n"
49
"sub x20, x23, #0x1\n"
50
"add x10, x10, #0x3\n"
51
"whilelt p9.b, XZR, x22\n"
52
"whilelt p8.b, x21, x22\n"
53
"mov x28, #0x0\n"
54
"ptrue p2.b\n"
55
"lsr x20, x20, #0x1\n"
// n_loops = (n_passes - 1) / 2
56
"and x27, x23, #0x1\n"
// odd_tail = bool(n_passes & 0x1)
57
"lsr x10, x10, #0x2\n"
58
"sub x26, x9, #0x2\n"
59
"ptrue p11.s\n"
60
"zip1 p10.b, p9.b, p8.b\n"
61
"mov x25, %x[row_offset]\n"
62
"mov x24, %x[out]\n"
63
"whilelt p9.b, x28, %x[width]\n"
64
"whilelt p8.b, x28, %x[width]\n"
65
"cbnz %x[first], 1f\n"
66
"addvl x24, x24, #-1\n"
67
"ld1w { z17.s }, p2/Z, [x24]\n"
68
"1:"
// K loop: Load row sums: End
69
"mov x23, %x[in]\n"
70
"ldr x22, [x23, #0x0]\n"
71
"mov x12, #0x0\n"
72
"ldr x21, [x23, #0x8]\n"
73
"add x23, x23, #0x10\n"
74
"cbz x26, 3f\n"
75
"2:"
// K loop: Charge: Loop
76
".inst 0x25246140 // psel p0.b, p8.b/Z, p10.b[w12]\n"
77
".inst 0xe01902c0 // ld1b { za0h.b[x12] }, p0/Z, [x22, x25]\n"
78
".inst 0x25646140 // psel p0.b, p8.b/Z, p10.b[w12, #4]\n"
79
"ldr x22, [x23, #0x0]\n"
80
".inst 0xe01902a4 // ld1b { za0h.b[x12, #4] }, p0/Z, [x21, x25]\n"
81
"add x12, x12, #0x8\n"
82
"cmp x12, x26, LSL #2\n"
83
"ldr x21, [x23, #0x8]\n"
84
"add x23, x23, #0x10\n"
85
"blt 2b\n"
86
"3:"
// K loop: Charge: End
87
".inst 0x25246140 // psel p0.b, p8.b/Z, p10.b[w12]\n"
88
".inst 0xe01902c0 // ld1b { za0h.b[x12] }, p0/Z, [x22, x25]\n"
89
".inst 0x25646140 // psel p0.b, p8.b/Z, p10.b[w12, #4]\n"
90
"mov x23, %x[in]\n"
91
".inst 0xe01902a4 // ld1b { za0h.b[x12, #4] }, p0/Z, [x21, x25]\n"
92
"ldr x22, [x23, #0x0]\n"
93
"incb x25\n"
94
"ldr x21, [x23, #0x8]\n"
95
"add x23, x23, #0x10\n"
96
"incb x28\n"
97
"cbz x20, 9f\n"
98
"mov x20, x20\n"
99
"4:"
// K loop: Main loop
100
"whilelt p8.b, x28, %x[width]\n"
101
"mov x13, #0x0\n"
102
"mov x12, #0x0\n"
103
"cbz x26, 6f\n"
104
"5:"
// K loop: Main loop: First: Loop
105
".inst 0x25356140 // psel p0.b, p8.b/Z, p10.b[w13, #2]\n"
106
".inst 0xe01922c2 // ld1b { za0h.b[x13, #2] }, p0/Z, [x22, x25]\n"
107
".inst 0x25756140 // psel p0.b, p8.b/Z, p10.b[w13, #6]\n"
108
"ldr x22, [x23, #0x0]\n"
109
".inst 0xe01922a6 // ld1b { za0h.b[x13, #6] }, p0/Z, [x21, x25]\n"
110
".inst 0xc0828810 // mova z16.s, p2/M, za0v.s[x12]\n"
111
".inst 0x25306d20 // psel p0.s, p11.s/Z, p9.s[w12]\n"
112
"udot z17.s, z16.b, z18.b\n"
113
"ldr x21, [x23, #0x8]\n"
114
".inst 0xe0bf8300 // st1w { za0v.s[x12] }, p0/Z, [x24, XZR, LSL #2]\n"
115
".inst 0x25706d20 // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
116
".inst 0xc0828830 // mova z16.s, p2/M, za0v.s[x12, #1]\n"
117
".inst 0xe0a98301 // st1w { za0v.s[x12, #1] }, p0/Z, [x24, x9, LSL #2]\n"
118
"add x12, x12, #0x2\n"
119
"cmp x12, x26\n"
120
"udot z17.s, z16.b, z18.b\n"
121
"add x23, x23, #0x10\n"
122
"addvl x24, x24, #2\n"
123
"add x13, x13, #0x8\n"
124
"blt 5b\n"
125
"6:"
// K loop: Main loop: First: Tail
126
".inst 0x25356140 // psel p0.b, p8.b/Z, p10.b[w13, #2]\n"
127
".inst 0xe01922c2 // ld1b { za0h.b[x13, #2] }, p0/Z, [x22, x25]\n"
128
".inst 0x25756140 // psel p0.b, p8.b/Z, p10.b[w13, #6]\n"
129
".inst 0xe01922a6 // ld1b { za0h.b[x13, #6] }, p0/Z, [x21, x25]\n"
130
".inst 0xc0828810 // mova z16.s, p2/M, za0v.s[x12]\n"
131
"udot z17.s, z16.b, z18.b\n"
132
"mov x23, %x[in]\n"
133
".inst 0x25306d21 // psel p1.s, p11.s/Z, p9.s[w12]\n"
134
"ldr x22, [x23, #0x0]\n"
135
".inst 0x25706d20 // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
136
".inst 0xc0828830 // mova z16.s, p2/M, za0v.s[x12, #1]\n"
137
"ldr x21, [x23, #0x8]\n"
138
".inst 0xe0bf8700 // st1w { za0v.s[x12] }, p1/Z, [x24, XZR, LSL #2]\n"
139
"whilelt p9.b, x28, %x[width]\n"
140
"incb x28\n"
141
"add x23, x23, #0x10\n"
142
".inst 0xe0a98301 // st1w { za0v.s[x12, #1] }, p0/Z, [x24, x9, LSL #2]\n"
143
"udot z17.s, z16.b, z18.b\n"
144
"addvl x24, x24, #2\n"
145
"incb x25\n"
146
"whilelt p8.b, x28, %x[width]\n"
147
"mov x13, #0x0\n"
148
"mov x12, #0x0\n"
149
"cbz x26, 8f\n"
150
"7:"
// K loop: Main loop: Second: Loop
151
".inst 0x25256140 // psel p0.b, p8.b/Z, p10.b[w13]\n"
152
".inst 0xe01922c0 // ld1b { za0h.b[x13] }, p0/Z, [x22, x25]\n"
153
".inst 0x25656140 // psel p0.b, p8.b/Z, p10.b[w13, #4]\n"
154
"ldr x22, [x23, #0x0]\n"
155
".inst 0xe01922a4 // ld1b { za0h.b[x13, #4] }, p0/Z, [x21, x25]\n"
156
".inst 0xc0828910 // mova z16.s, p2/M, za2v.s[x12]\n"
157
".inst 0x25306d20 // psel p0.s, p11.s/Z, p9.s[w12]\n"
158
"udot z17.s, z16.b, z18.b\n"
159
"ldr x21, [x23, #0x8]\n"
160
".inst 0xe0bf8308 // st1w { za2v.s[x12] }, p0/Z, [x24, XZR, LSL #2]\n"
161
".inst 0x25706d20 // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
162
".inst 0xc0828930 // mova z16.s, p2/M, za2v.s[x12, #1]\n"
163
".inst 0xe0a98309 // st1w { za2v.s[x12, #1] }, p0/Z, [x24, x9, LSL #2]\n"
164
"add x12, x12, #0x2\n"
165
"cmp x12, x26\n"
166
"udot z17.s, z16.b, z18.b\n"
167
"add x23, x23, #0x10\n"
168
"addvl x24, x24, #2\n"
169
"add x13, x13, #0x8\n"
170
"blt 7b\n"
171
"8:"
// K loop: Main loop: Second: Tail
172
".inst 0x25256140 // psel p0.b, p8.b/Z, p10.b[w13]\n"
173
".inst 0xe01922c0 // ld1b { za0h.b[x13] }, p0/Z, [x22, x25]\n"
174
".inst 0x25656140 // psel p0.b, p8.b/Z, p10.b[w13, #4]\n"
175
".inst 0xe01922a4 // ld1b { za0h.b[x13, #4] }, p0/Z, [x21, x25]\n"
176
".inst 0xc0828910 // mova z16.s, p2/M, za2v.s[x12]\n"
177
"udot z17.s, z16.b, z18.b\n"
178
"mov x23, %x[in]\n"
179
".inst 0x25306d21 // psel p1.s, p11.s/Z, p9.s[w12]\n"
180
"ldr x22, [x23, #0x0]\n"
181
".inst 0x25706d20 // psel p0.s, p11.s/Z, p9.s[w12, #1]\n"
182
".inst 0xc0828930 // mova z16.s, p2/M, za2v.s[x12, #1]\n"
183
"ldr x21, [x23, #0x8]\n"
184
".inst 0xe0bf8708 // st1w { za2v.s[x12] }, p1/Z, [x24, XZR, LSL #2]\n"
185
"whilelt p9.b, x28, %x[width]\n"
186
"subs x20, x20, #0x1\n"
187
"add x23, x23, #0x10\n"
188
".inst 0xe0a98309 // st1w { za2v.s[x12, #1] }, p0/Z, [x24, x9, LSL #2]\n"
189
"udot z17.s, z16.b, z18.b\n"
190
"addvl x24, x24, #2\n"
191
"incb x28\n"
192
"incb x25\n"
193
"bgt 4b\n"
194
"9:"
// K loop: Tails
195
"cbnz x27, 12f\n"
196
"mov x23, %x[in]\n"
197
"whilelt p8.b, x28, %x[width]\n"
198
"mov x13, #0x0\n"
199
"mov x12, #0x0\n"
200
"10:"
// K loop: Tails: Even: First
201
".inst 0x25306d20 // psel p0.s, p11.s/Z, p9.s[w12]\n"
202
".inst 0xe0bf8300 // st1w { za0v.s[x12] }, p0/Z, [x24, XZR, LSL #2]\n"
203
"ldr x20, [x23, #0x0]\n"
204
".inst 0xc0828810 // mova z16.s, p2/M, za0v.s[x12]\n"
205
"add x12, x12, #0x1\n"
206
".inst 0x25356140 // psel p0.b, p8.b/Z, p10.b[w13, #2]\n"
207
"udot z17.s, z16.b, z18.b\n"
208
".inst 0xe0192282 // ld1b { za0h.b[x13, #2] }, p0/Z, [x20, x25]\n"
209
"cmp x12, x9\n"
210
"add x23, x23, #0x8\n"
211
"addvl x24, x24, #1\n"
212
"add x13, x13, #0x4\n"
213
"blt 10b\n"
214
"whilelt p9.b, x28, %x[width]\n"
215
"whilelt p8.b, x28, %x[width]\n"
216
"mov x20, #0x0\n"
217
"mov x12, #0x0\n"
218
"11:"
// K loop: Tails: Even: Second
219
".inst 0x25306d20 // psel p0.s, p11.s/Z, p9.s[w12]\n"
220
".inst 0xe0bf8308 // st1w { za2v.s[x12] }, p0/Z, [x24, XZR, LSL #2]\n"
221
".inst 0xc0828910 // mova z16.s, p2/M, za2v.s[x12]\n"
222
"add x12, x12, #0x1\n"
223
"cmp x12, x10\n"
224
"udot z17.s, z16.b, z18.b\n"
225
"addvl x24, x24, #1\n"
226
"add x20, x20, #0x4\n"
227
"blt 11b\n"
228
"whilelt p8.b, x28, %x[width]\n"
229
"b 14f\n"
230
"12:"
// K loop: Tails: Odd
231
"mov x12, #0x0\n"
232
"13:"
// K loop: Tails: Odd: Loop
233
".inst 0x25306d20 // psel p0.s, p11.s/Z, p9.s[w12]\n"
234
".inst 0xe0bf8300 // st1w { za0v.s[x12] }, p0/Z, [x24, XZR, LSL #2]\n"
235
".inst 0xc0828810 // mova z16.s, p2/M, za0v.s[x12]\n"
236
"add x12, x12, #0x1\n"
237
"cmp x12, x10\n"
238
"udot z17.s, z16.b, z18.b\n"
239
"addvl x24, x24, #1\n"
240
"blt 13b\n"
241
"14:"
// K loop: End
242
"st1w { z17.s }, p2, [x24]\n"
243
"addvl x24, x24, #1\n"
244
"mov %x[out], x24\n"
245
".inst 0xd503467f // SMSTOP\n"
246
: [out]
"+&r"
(out)
247
: [first]
"r"
(first), [height]
"r"
(height), [in]
"r"
(in), [row_offset]
"r"
(row_offset), [width]
"r"
(width)
248
:
"cc"
,
"memory"
,
"p0"
,
"p1"
,
"p2"
,
"p3"
,
"p4"
,
"p5"
,
"p6"
,
"p7"
,
"p8"
,
"p9"
,
"p10"
,
"p11"
,
"p12"
,
"p13"
,
"p14"
,
"p15"
,
"x9"
,
"x10"
,
"x12"
,
"x13"
,
"x20"
,
"x21"
,
"x22"
,
"x23"
,
"x24"
,
"x25"
,
"x26"
,
"x27"
,
"x28"
,
"z0"
,
"z1"
,
"z2"
,
"z3"
,
"z4"
,
"z5"
,
"z6"
,
"z7"
,
"z8"
,
"z9"
,
"z10"
,
"z11"
,
"z12"
,
"z13"
,
"z14"
,
"z15"
,
"z16"
,
"z17"
,
"z18"
,
"z19"
,
"z20"
,
"z21"
,
"z22"
,
"z23"
,
"z24"
,
"z25"
,
"z26"
,
"z27"
,
"z28"
,
"z29"
,
"z30"
,
"z31"
249
);
250
}
251
252
#endif // defined(ARM_COMPUTE_ENABLE_SME)
src
core
NEON
kernels
arm_gemm
indirect-interleaves
sme_interleave1VL_block4_u8_u8_summing.hpp
Generated on Mon Apr 29 2024 10:53:55 for Compute Library by
1.8.17