24.04
sme_interleave1VL_bf16_bf16.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, 1, VLType::SME, false>(
29
bfloat16
* &out,
const
bfloat16
*
const
*in,
30
size_t
width,
size_t
height,
size_t
row_offset,
bool
31
)
32
{
33
__asm__ __volatile__(
34
".inst 0xd503477f // SMSTART ZA\n"
35
"mov x21, %x[width]\n"
36
"inch x21\n"
37
"cnth x11\n"
38
"sub x21, x21, #0x1\n"
39
"udiv x21, x21, x11\n"
// n_passes = ceildiv(width, VL<T>)
40
"mov x20, %x[width]\n"
41
"sub x10, x11, #0x1\n"
42
"sub x9, x21, #0x1\n"
43
"ands x10, x20, x10\n"
44
"sub x28, x11, #0x2\n"
45
"lsl x20, %x[height], #0x1\n"
// height * 2
46
"mov x27, #0x0\n"
47
"mov x26, %x[in]\n"
48
"lsr x9, x9, #0x1\n"
// n_loops = (n_passes - 1) / 2
49
"ldr x25, [x26, #0x0]\n"
50
"and x24, x21, #0x1\n"
// odd_tail = bool(n_passes & 0x1)
51
"csel x10, x10, x11, NE\n"
52
"ldr x23, [x26, #0x8]\n"
53
"ptrue p11.h\n"
54
"whilelt p10.h, XZR, x20\n"
55
"mov x22, %x[row_offset]\n"
56
"mov x21, %x[out]\n"
57
"whilelt p9.h, x27, %x[width]\n"
58
"whilelt p8.h, x27, %x[width]\n"
59
"add x26, x26, #0x10\n"
60
"mov x12, #0x0\n"
61
"cbz x28, 2f\n"
62
"1:"
// K loop: Charge: Loop
63
".inst 0x25286140 // psel p0.h, p8.h/Z, p10.h[w12]\n"
64
".inst 0xe0560320 // ld1h { za0h.h[x12] }, p0/Z, [x25, x22, LSL #1]\n"
65
".inst 0x25386140 // psel p0.h, p8.h/Z, p10.h[w12, #1]\n"
66
"ldr x25, [x26, #0x0]\n"
67
".inst 0xe05602e1 // ld1h { za0h.h[x12, #1] }, p0/Z, [x23, x22, LSL #1]\n"
68
"add x12, x12, #0x2\n"
69
"cmp x12, x28\n"
70
"ldr x23, [x26, #0x8]\n"
71
"add x26, x26, #0x10\n"
72
"blt 1b\n"
73
"2:"
// K loop: Charge: End
74
".inst 0x25286140 // psel p0.h, p8.h/Z, p10.h[w12]\n"
75
".inst 0xe0560320 // ld1h { za0h.h[x12] }, p0/Z, [x25, x22, LSL #1]\n"
76
".inst 0x25386140 // psel p0.h, p8.h/Z, p10.h[w12, #1]\n"
77
"mov x26, %x[in]\n"
78
".inst 0xe05602e1 // ld1h { za0h.h[x12, #1] }, p0/Z, [x23, x22, LSL #1]\n"
79
"ldr x25, [x26, #0x0]\n"
80
"inch x22\n"
81
"ldr x23, [x26, #0x8]\n"
82
"add x26, x26, #0x10\n"
83
"inch x27\n"
84
"cbz x9, 8f\n"
85
"mov x20, x9\n"
86
"3:"
// K loop: Main loop
87
"whilelt p8.h, x27, %x[width]\n"
88
"mov x12, #0x0\n"
89
"cbz x28, 5f\n"
90
"4:"
// K loop: Main loop: First: Loop
91
".inst 0x25286140 // psel p0.h, p8.h/Z, p10.h[w12]\n"
92
".inst 0xe0560328 // ld1h { za1h.h[x12] }, p0/Z, [x25, x22, LSL #1]\n"
93
".inst 0x25386141 // psel p1.h, p8.h/Z, p10.h[w12, #1]\n"
94
"ldr x25, [x26, #0x0]\n"
95
".inst 0x25286d20 // psel p0.h, p11.h/Z, p9.h[w12]\n"
96
".inst 0xe05606e9 // ld1h { za1h.h[x12, #1] }, p1/Z, [x23, x22, LSL #1]\n"
97
"ldr x23, [x26, #0x8]\n"
98
".inst 0xe07f82a0 // st1h { za0v.h[x12] }, p0/Z, [x21, XZR, LSL #1]\n"
99
".inst 0x25386d20 // psel p0.h, p11.h/Z, p9.h[w12, #1]\n"
100
".inst 0xe06b82a1 // st1h { za0v.h[x12, #1] }, p0/Z, [x21, x11, LSL #1]\n"
101
"add x12, x12, #0x2\n"
102
"cmp x12, x28\n"
103
"add x26, x26, #0x10\n"
104
"addvl x21, x21, #2\n"
105
"blt 4b\n"
106
"5:"
// K loop: Main loop: First: Tail
107
".inst 0x25286140 // psel p0.h, p8.h/Z, p10.h[w12]\n"
108
".inst 0xe0560328 // ld1h { za1h.h[x12] }, p0/Z, [x25, x22, LSL #1]\n"
109
"mov x26, %x[in]\n"
110
"ldr x25, [x26, #0x0]\n"
111
".inst 0x25386141 // psel p1.h, p8.h/Z, p10.h[w12, #1]\n"
112
".inst 0x25286d20 // psel p0.h, p11.h/Z, p9.h[w12]\n"
113
".inst 0xe05606e9 // ld1h { za1h.h[x12, #1] }, p1/Z, [x23, x22, LSL #1]\n"
114
"ldr x23, [x26, #0x8]\n"
115
".inst 0xe07f82a0 // st1h { za0v.h[x12] }, p0/Z, [x21, XZR, LSL #1]\n"
116
".inst 0x25386d20 // psel p0.h, p11.h/Z, p9.h[w12, #1]\n"
117
"whilelt p9.h, x27, %x[width]\n"
118
"inch x27\n"
119
"add x26, x26, #0x10\n"
120
".inst 0xe06b82a1 // st1h { za0v.h[x12, #1] }, p0/Z, [x21, x11, LSL #1]\n"
121
"addvl x21, x21, #2\n"
122
"inch x22\n"
123
"whilelt p8.h, x27, %x[width]\n"
124
"mov x12, #0x0\n"
125
"cbz x28, 7f\n"
126
"6:"
// K loop: Main loop: Second: Loop
127
".inst 0x25286140 // psel p0.h, p8.h/Z, p10.h[w12]\n"
128
".inst 0xe0560320 // ld1h { za0h.h[x12] }, p0/Z, [x25, x22, LSL #1]\n"
129
".inst 0x25386141 // psel p1.h, p8.h/Z, p10.h[w12, #1]\n"
130
"ldr x25, [x26, #0x0]\n"
131
".inst 0x25286d20 // psel p0.h, p11.h/Z, p9.h[w12]\n"
132
".inst 0xe05606e1 // ld1h { za0h.h[x12, #1] }, p1/Z, [x23, x22, LSL #1]\n"
133
"ldr x23, [x26, #0x8]\n"
134
".inst 0xe07f82a8 // st1h { za1v.h[x12] }, p0/Z, [x21, XZR, LSL #1]\n"
135
".inst 0x25386d20 // psel p0.h, p11.h/Z, p9.h[w12, #1]\n"
136
".inst 0xe06b82a9 // st1h { za1v.h[x12, #1] }, p0/Z, [x21, x11, LSL #1]\n"
137
"add x12, x12, #0x2\n"
138
"cmp x12, x28\n"
139
"add x26, x26, #0x10\n"
140
"addvl x21, x21, #2\n"
141
"blt 6b\n"
142
"7:"
// K loop: Main loop: Second: Tail
143
".inst 0x25286140 // psel p0.h, p8.h/Z, p10.h[w12]\n"
144
".inst 0xe0560320 // ld1h { za0h.h[x12] }, p0/Z, [x25, x22, LSL #1]\n"
145
"mov x26, %x[in]\n"
146
"ldr x25, [x26, #0x0]\n"
147
".inst 0x25386141 // psel p1.h, p8.h/Z, p10.h[w12, #1]\n"
148
".inst 0x25286d20 // psel p0.h, p11.h/Z, p9.h[w12]\n"
149
".inst 0xe05606e1 // ld1h { za0h.h[x12, #1] }, p1/Z, [x23, x22, LSL #1]\n"
150
"ldr x23, [x26, #0x8]\n"
151
".inst 0xe07f82a8 // st1h { za1v.h[x12] }, p0/Z, [x21, XZR, LSL #1]\n"
152
".inst 0x25386d20 // psel p0.h, p11.h/Z, p9.h[w12, #1]\n"
153
"whilelt p9.h, x27, %x[width]\n"
154
"subs x20, x20, #0x1\n"
155
"add x26, x26, #0x10\n"
156
".inst 0xe06b82a9 // st1h { za1v.h[x12, #1] }, p0/Z, [x21, x11, LSL #1]\n"
157
"addvl x21, x21, #2\n"
158
"inch x27\n"
159
"inch x22\n"
160
"bgt 3b\n"
161
"8:"
// K loop: Tails
162
"cbnz x24, 11f\n"
163
"mov x26, %x[in]\n"
164
"whilelt p8.h, x27, %x[width]\n"
165
"mov x12, #0x0\n"
166
"9:"
// K loop: Tails: Even: First
167
".inst 0x25286d20 // psel p0.h, p11.h/Z, p9.h[w12]\n"
168
".inst 0xe07f82a0 // st1h { za0v.h[x12] }, p0/Z, [x21, XZR, LSL #1]\n"
169
"ldr x20, [x26, #0x0]\n"
170
".inst 0x25286140 // psel p0.h, p8.h/Z, p10.h[w12]\n"
171
".inst 0xe0560288 // ld1h { za1h.h[x12] }, p0/Z, [x20, x22, LSL #1]\n"
172
"add x12, x12, #0x1\n"
173
"cmp x12, x11\n"
174
"add x26, x26, #0x8\n"
175
"addvl x21, x21, #1\n"
176
"blt 9b\n"
177
"whilelt p9.h, x27, %x[width]\n"
178
"whilelt p8.h, x27, %x[width]\n"
179
"mov x12, #0x0\n"
180
"10:"
// K loop: Tails: Even: Second
181
".inst 0x25286d20 // psel p0.h, p11.h/Z, p9.h[w12]\n"
182
".inst 0xe07f82a8 // st1h { za1v.h[x12] }, p0/Z, [x21, XZR, LSL #1]\n"
183
"add x12, x12, #0x1\n"
184
"cmp x12, x10\n"
185
"addvl x21, x21, #1\n"
186
"blt 10b\n"
187
"whilelt p8.h, x27, %x[width]\n"
188
"b 13f\n"
189
"11:"
// K loop: Tails: Odd
190
"mov x12, #0x0\n"
191
"12:"
// K loop: Tails: Odd: Loop
192
".inst 0x25286d20 // psel p0.h, p11.h/Z, p9.h[w12]\n"
193
".inst 0xe07f82a0 // st1h { za0v.h[x12] }, p0/Z, [x21, XZR, LSL #1]\n"
194
"add x12, x12, #0x1\n"
195
"cmp x12, x10\n"
196
"addvl x21, x21, #1\n"
197
"blt 12b\n"
198
"13:"
// K loop: End
199
"mov %x[out], x21\n"
200
".inst 0xd503467f // SMSTOP\n"
201
: [out]
"+&r"
(out)
202
: [height]
"r"
(height), [in]
"r"
(in), [row_offset]
"r"
(row_offset), [width]
"r"
(width)
203
:
"cc"
,
"memory"
,
"p0"
,
"p1"
,
"p2"
,
"p3"
,
"p4"
,
"p5"
,
"p6"
,
"p7"
,
"p8"
,
"p9"
,
"p10"
,
"p11"
,
"p12"
,
"p13"
,
"p14"
,
"p15"
,
"x9"
,
"x10"
,
"x11"
,
"x12"
,
"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"
204
);
205
}
206
207
#endif // defined(ARM_COMPUTE_ENABLE_SME)
arm_gemm::bfloat16
arm_compute::bfloat16 bfloat16
Definition:
bfloat.hpp:30
src
core
NEON
kernels
arm_gemm
indirect-interleaves
sme_interleave1VL_bf16_bf16.hpp
Generated on Mon Apr 29 2024 10:53:55 for Compute Library by
1.8.17