aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_u8u32_dot_8x3VL/a64fx.cpp
blob: 79e794a834bc0f5463d1ce8f3b4d04f3a7289cc8 (plain)
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
/*
 * Copyright (c) 2021, 2023 Arm Limited.
 *
 * SPDX-License-Identifier: MIT
 *
 * Permission is hereby granted, free of charge, to any person obtaining a copy
 * of this software and associated documentation files (the "Software"), to
 * deal in the Software without restriction, including without limitation the
 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
 * sell copies of the Software, and to permit persons to whom the Software is
 * furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice shall be included in all
 * copies or substantial portions of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 * SOFTWARE.
 */
#ifdef ARM_COMPUTE_ENABLE_SVE

#include <cstddef>
#include <cstdint>

namespace arm_gemm {

void sve_interleaved_u8u32_dot_8x3VL_a64fx(
    const uint8_t *Apanel,
    const uint8_t *Bpanel,
    uint32_t *Cpanel,
    int ablocks,
    int bblocks,
    int K) {

    struct KernelArgs {
        size_t K = {};
        const uint8_t *Bpanel = {};
        size_t bblocks = {};
    } ka;

    ka.K = (K/4) - 1;
    ka.Bpanel = Bpanel;
    ka.bblocks = bblocks;

    __asm__ __volatile__(
      "ptrue p0.b\n"
      "1:"  // Height loop
      "ldr x23, [%x[args_ptr], %[offsetof_bblocks]]\n"
      "ldr x22, [%x[args_ptr], %[offsetof_Bpanel]]\n"
      "mov x21, %x[Apanel]\n"
      "2:"  // Width loop
      "ldr x20, [%x[args_ptr], %[offsetof_K]]\n"
      "mov %x[Apanel], x21\n"
      "cmp x20, #0x2\n"
      "mov z8.s, #0x0\n"
      "mov z9.s, #0x0\n"
      "ld1b { z0.b }, p0/Z, [x22]\n"
      "mov z10.s, #0x0\n"
      "mov z11.s, #0x0\n"
      "ld1b { z1.b }, p0/Z, [x22, #1, MUL VL]\n"
      "mov z12.s, #0x0\n"
      "mov z13.s, #0x0\n"
      "ld1b { z2.b }, p0/Z, [x22, #2, MUL VL]\n"
      "mov z14.s, #0x0\n"
      "mov z15.s, #0x0\n"
      "ld1rw { z3.s }, p0/Z, [%x[Apanel]]\n"
      "mov z16.s, #0x0\n"
      "mov z17.s, #0x0\n"
      "ld1rw { z4.s }, p0/Z, [%x[Apanel], #4]\n"
      "mov z18.s, #0x0\n"
      "mov z19.s, #0x0\n"
      "ld1rw { z5.s }, p0/Z, [%x[Apanel], #8]\n"
      "mov z20.s, #0x0\n"
      "mov z21.s, #0x0\n"
      "ld1rw { z6.s }, p0/Z, [%x[Apanel], #12]\n"
      "mov z22.s, #0x0\n"
      "mov z23.s, #0x0\n"
      "mov z24.s, #0x0\n"
      "mov z25.s, #0x0\n"
      "mov z26.s, #0x0\n"
      "mov z27.s, #0x0\n"
      "mov z28.s, #0x0\n"
      "mov z29.s, #0x0\n"
      "mov z30.s, #0x0\n"
      "mov z31.s, #0x0\n"
      "blt 4f\n"
      "3:"  // main loop head
      "udot z8.s, z0.b, z3.b\n"
      "udot z9.s, z1.b, z3.b\n"
      "sub x20, x20, #0x2\n"
      "udot z10.s, z2.b, z3.b\n"
      "ld1rw { z7.s }, p0/Z, [%x[Apanel], #16]\n"
      "udot z11.s, z0.b, z4.b\n"
      "udot z12.s, z1.b, z4.b\n"
      "udot z13.s, z2.b, z4.b\n"
      "ld1rw { z4.s }, p0/Z, [%x[Apanel], #20]\n"
      "udot z14.s, z0.b, z5.b\n"
      "udot z15.s, z1.b, z5.b\n"
      "cmp x20, #0x2\n"
      "udot z16.s, z2.b, z5.b\n"
      "ld1rw { z3.s }, p0/Z, [%x[Apanel], #24]\n"
      "udot z17.s, z0.b, z6.b\n"
      "udot z18.s, z1.b, z6.b\n"
      "udot z19.s, z2.b, z6.b\n"
      "ld1rw { z5.s }, p0/Z, [%x[Apanel], #28]\n"
      "udot z20.s, z0.b, z7.b\n"
      "udot z21.s, z1.b, z7.b\n"
      "udot z22.s, z2.b, z7.b\n"
      "ld1rw { z7.s }, p0/Z, [%x[Apanel], #32]\n"
      "udot z23.s, z0.b, z4.b\n"
      "udot z24.s, z1.b, z4.b\n"
      "udot z25.s, z2.b, z4.b\n"
      "ld1rw { z4.s }, p0/Z, [%x[Apanel], #36]\n"
      "udot z26.s, z0.b, z3.b\n"
      "udot z27.s, z1.b, z3.b\n"
      "udot z28.s, z2.b, z3.b\n"
      "ld1rw { z3.s }, p0/Z, [%x[Apanel], #40]\n"
      "udot z29.s, z0.b, z5.b\n"
      "ld1b { z6.b }, p0/Z, [x22, #3, MUL VL]\n"
      "udot z30.s, z1.b, z5.b\n"
      "udot z31.s, z2.b, z5.b\n"
      "ld1b { z2.b }, p0/Z, [x22, #4, MUL VL]\n"
      "ld1b { z5.b }, p0/Z, [x22, #5, MUL VL]\n"
      "udot z8.s, z6.b, z7.b\n"
      "ld1rw { z1.s }, p0/Z, [%x[Apanel], #44]\n"
      "udot z9.s, z2.b, z7.b\n"
      "udot z10.s, z5.b, z7.b\n"
      "udot z11.s, z6.b, z4.b\n"
      "ld1rw { z7.s }, p0/Z, [%x[Apanel], #48]\n"
      "udot z12.s, z2.b, z4.b\n"
      "udot z13.s, z5.b, z4.b\n"
      "ld1rw { z4.s }, p0/Z, [%x[Apanel], #52]\n"
      "udot z14.s, z6.b, z3.b\n"
      "udot z15.s, z2.b, z3.b\n"
      "addvl x22, x22, #6\n"
      "udot z16.s, z5.b, z3.b\n"
      "ld1rw { z0.s }, p0/Z, [%x[Apanel], #56]\n"
      "udot z17.s, z6.b, z1.b\n"
      "udot z18.s, z2.b, z1.b\n"
      "udot z19.s, z5.b, z1.b\n"
      "ld1rw { z1.s }, p0/Z, [%x[Apanel], #60]\n"
      "add %x[Apanel], %x[Apanel], #0x40\n"
      "udot z20.s, z6.b, z7.b\n"
      "udot z21.s, z2.b, z7.b\n"
      "udot z22.s, z5.b, z7.b\n"
      "udot z23.s, z6.b, z4.b\n"
      "ld1rw { z3.s }, p0/Z, [%x[Apanel]]\n"
      "udot z24.s, z2.b, z4.b\n"
      "udot z25.s, z5.b, z4.b\n"
      "ld1rw { z4.s }, p0/Z, [%x[Apanel], #4]\n"
      "udot z26.s, z6.b, z0.b\n"
      "udot z27.s, z2.b, z0.b\n"
      "udot z28.s, z5.b, z0.b\n"
      "udot z29.s, z6.b, z1.b\n"
      "ld1b { z0.b }, p0/Z, [x22]\n"
      "udot z30.s, z2.b, z1.b\n"
      "udot z31.s, z5.b, z1.b\n"
      "ld1b { z1.b }, p0/Z, [x22, #1, MUL VL]\n"
      "ld1b { z2.b }, p0/Z, [x22, #2, MUL VL]\n"
      "ld1rw { z5.s }, p0/Z, [%x[Apanel], #8]\n"
      "ld1rw { z6.s }, p0/Z, [%x[Apanel], #12]\n"
      "bge 3b\n"
      "4:"  // main loop skip
      "udot z8.s, z0.b, z3.b\n"
      "udot z9.s, z1.b, z3.b\n"
      "addvl x22, x22, #3\n"
      "udot z10.s, z2.b, z3.b\n"
      "ld1rw { z7.s }, p0/Z, [%x[Apanel], #16]\n"
      "udot z11.s, z0.b, z4.b\n"
      "udot z12.s, z1.b, z4.b\n"
      "udot z13.s, z2.b, z4.b\n"
      "ld1rw { z4.s }, p0/Z, [%x[Apanel], #20]\n"
      "udot z14.s, z0.b, z5.b\n"
      "udot z15.s, z1.b, z5.b\n"
      "udot z16.s, z2.b, z5.b\n"
      "ld1rw { z5.s }, p0/Z, [%x[Apanel], #24]\n"
      "udot z17.s, z0.b, z6.b\n"
      "udot z18.s, z1.b, z6.b\n"
      "udot z19.s, z2.b, z6.b\n"
      "ld1rw { z3.s }, p0/Z, [%x[Apanel], #28]\n"
      "udot z20.s, z0.b, z7.b\n"
      "udot z21.s, z1.b, z7.b\n"
      "add %x[Apanel], %x[Apanel], #0x20\n"
      "udot z22.s, z2.b, z7.b\n"
      "udot z23.s, z0.b, z4.b\n"
      "udot z24.s, z1.b, z4.b\n"
      "udot z25.s, z2.b, z4.b\n"
      "udot z26.s, z0.b, z5.b\n"
      "udot z27.s, z1.b, z5.b\n"
      "udot z28.s, z2.b, z5.b\n"
      "udot z29.s, z0.b, z3.b\n"
      "udot z30.s, z1.b, z3.b\n"
      "udot z31.s, z2.b, z3.b\n"
      "cbz x20, 5f\n"
      "ld1b { z6.b }, p0/Z, [x22]\n"
      "ld1b { z5.b }, p0/Z, [x22, #1, MUL VL]\n"
      "ld1b { z4.b }, p0/Z, [x22, #2, MUL VL]\n"
      "ld1rw { z3.s }, p0/Z, [%x[Apanel]]\n"
      "udot z8.s, z6.b, z3.b\n"
      "ld1rw { z2.s }, p0/Z, [%x[Apanel], #4]\n"
      "ld1rw { z1.s }, p0/Z, [%x[Apanel], #8]\n"
      "udot z9.s, z5.b, z3.b\n"
      "ld1rw { z0.s }, p0/Z, [%x[Apanel], #12]\n"
      "udot z10.s, z4.b, z3.b\n"
      "udot z11.s, z6.b, z2.b\n"
      "udot z12.s, z5.b, z2.b\n"
      "udot z13.s, z4.b, z2.b\n"
      "ld1rw { z3.s }, p0/Z, [%x[Apanel], #16]\n"
      "udot z14.s, z6.b, z1.b\n"
      "udot z15.s, z5.b, z1.b\n"
      "ld1rw { z2.s }, p0/Z, [%x[Apanel], #20]\n"
      "udot z16.s, z4.b, z1.b\n"
      "udot z17.s, z6.b, z0.b\n"
      "ld1rw { z1.s }, p0/Z, [%x[Apanel], #24]\n"
      "udot z18.s, z5.b, z0.b\n"
      "udot z19.s, z4.b, z0.b\n"
      "ld1rw { z0.s }, p0/Z, [%x[Apanel], #28]\n"
      "udot z20.s, z6.b, z3.b\n"
      "udot z21.s, z5.b, z3.b\n"
      "addvl x22, x22, #3\n"
      "udot z22.s, z4.b, z3.b\n"
      "udot z23.s, z6.b, z2.b\n"
      "add %x[Apanel], %x[Apanel], #0x20\n"
      "udot z24.s, z5.b, z2.b\n"
      "udot z25.s, z4.b, z2.b\n"
      "udot z26.s, z6.b, z1.b\n"
      "udot z27.s, z5.b, z1.b\n"
      "udot z28.s, z4.b, z1.b\n"
      "udot z29.s, z6.b, z0.b\n"
      "udot z30.s, z5.b, z0.b\n"
      "udot z31.s, z4.b, z0.b\n"
      "5:"  // multiply loop done
      "st1w { z8.s }, p0, [%x[Cpanel]]\n"
      "subs x23, x23, #0x1\n"
      "st1w { z9.s }, p0, [%x[Cpanel], #1, MUL VL]\n"
      "st1w { z10.s }, p0, [%x[Cpanel], #2, MUL VL]\n"
      "st1w { z11.s }, p0, [%x[Cpanel], #3, MUL VL]\n"
      "st1w { z12.s }, p0, [%x[Cpanel], #4, MUL VL]\n"
      "st1w { z13.s }, p0, [%x[Cpanel], #5, MUL VL]\n"
      "st1w { z14.s }, p0, [%x[Cpanel], #6, MUL VL]\n"
      "st1w { z15.s }, p0, [%x[Cpanel], #7, MUL VL]\n"
      "addvl %x[Cpanel], %x[Cpanel], #16\n"
      "st1w { z16.s }, p0, [%x[Cpanel], #-8, MUL VL]\n"
      "st1w { z17.s }, p0, [%x[Cpanel], #-7, MUL VL]\n"
      "st1w { z18.s }, p0, [%x[Cpanel], #-6, MUL VL]\n"
      "st1w { z19.s }, p0, [%x[Cpanel], #-5, MUL VL]\n"
      "st1w { z20.s }, p0, [%x[Cpanel], #-4, MUL VL]\n"
      "st1w { z21.s }, p0, [%x[Cpanel], #-3, MUL VL]\n"
      "st1w { z22.s }, p0, [%x[Cpanel], #-2, MUL VL]\n"
      "st1w { z23.s }, p0, [%x[Cpanel], #-1, MUL VL]\n"
      "st1w { z24.s }, p0, [%x[Cpanel]]\n"
      "st1w { z25.s }, p0, [%x[Cpanel], #1, MUL VL]\n"
      "st1w { z26.s }, p0, [%x[Cpanel], #2, MUL VL]\n"
      "st1w { z27.s }, p0, [%x[Cpanel], #3, MUL VL]\n"
      "st1w { z28.s }, p0, [%x[Cpanel], #4, MUL VL]\n"
      "st1w { z29.s }, p0, [%x[Cpanel], #5, MUL VL]\n"
      "st1w { z30.s }, p0, [%x[Cpanel], #6, MUL VL]\n"
      "st1w { z31.s }, p0, [%x[Cpanel], #7, MUL VL]\n"
      "addvl %x[Cpanel], %x[Cpanel], #8\n"
      "bgt 2b\n"
      "subs %x[ablocks], %x[ablocks], #0x1\n"
      "bne 1b\n"
      : [Apanel] "+&r" (Apanel), [Cpanel] "+&r" (Cpanel), [ablocks] "+&r" (ablocks)
      : [args_ptr] "r" (&ka), [offsetof_Bpanel] "I" (offsetof(KernelArgs, Bpanel)), [offsetof_K] "I" (offsetof(KernelArgs, K)), [offsetof_bblocks] "I" (offsetof(KernelArgs, bblocks))
      : "cc", "memory", "p0", "x20", "x21", "x22", "x23", "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"
    );
}

} // namespace arm_gemm
#endif // ARM_COMPUTE_ENABLE_SVE