aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_4x4/generic.cpp
blob: 0a881ffde33e41c7205db90ab9eb4963bfc3980f (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
/*
 * Copyright (c) 2017-2018 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 __aarch64__

#include <arm_neon.h>

#include "../../asmlib.hpp"

namespace arm_gemm
{
void a64_gemm_u8_4x4(const uint8_t *Apanel, const uint8_t *Bpanel, uint32_t *Cpanel, int ablocks, int bblocks, int K)
{
    const uint8_t *a_ptr = Apanel;
    uint32_t      *c_ptr = Cpanel;
    K /= 16;

    for(int yb = 0; yb < ablocks; yb++)
    {
        const uint8_t *a_ptr0 = a_ptr;
        const uint8_t *b_ptr  = Bpanel;

        for(int xb = 0; xb < bblocks; xb++)
        {
            a_ptr = a_ptr0;

            int k = K - 1;

            register uint8x16_t b0 asm("v4");
            register uint8x16_t b1 asm("v5");
            register uint8x16_t b2 asm("v6");
            register uint8x16_t b3 asm("v7");

            __asm __volatile(
                "movi    v16.4s, #0x0\n"
                "ldr    q0, [%[a_ptr]]\n"
                "movi    v17.4s, #0x0\n"
                "ldr    %q[b0], [%[b_ptr]]\n"
                "movi    v18.4s, #0x0\n"
                "ldr    %q[b1], [%[b_ptr], #16]\n"
                "movi    v19.4s, #0x0\n"
                "ldr    %q[b2], [%[b_ptr], #32]\n"
                "movi    v20.4s, #0x0\n"
                "ldr    %q[b3], [%[b_ptr], #48]\n"
                "movi    v21.4s, #0x0\n"
                "ldr    q1, [%[a_ptr], #16]\n"
                "movi    v22.4s, #0x0\n"
                "ldr    q2, [%[a_ptr], #32]\n"
                "movi    v23.4s, #0x0\n"
                "ldr    q3, [%[a_ptr], #48]\n"
                "movi    v24.4s, #0x0\n" ASM_PREFETCH("[%[b_ptr], #64]") "movi    v25.4s, #0x0\n" ASM_PREFETCH("[%[a_ptr], #64]") "movi    v26.4s, #0x0\n" ASM_PREFETCH("[%[b_ptr], #128]") "movi    v27.4s, #0x0\n"
                ASM_PREFETCH("[%[a_ptr], #128]") "movi    v28.4s, #0x0\n" ASM_PREFETCH("[%[b_ptr], #192]") "movi    v29.4s, #0x0\n" ASM_PREFETCH("[%[a_ptr], #192]") "movi    v30.4s, #0x0\n"
                ASM_PREFETCH("[%[b_ptr], #256]") "movi    v31.4s, #0x0\n" ASM_PREFETCH("[%[a_ptr], #256]")

                "umull    v12.8h, v0.8b, %[b0].8b\n"
                "add    %[a_ptr], %[a_ptr], #64\n"
                "umull    v13.8h, v0.8b, %[b1].8b\n"
                "umull    v14.8h, v0.8b, %[b2].8b\n"
                "add    %[b_ptr], %[b_ptr], #64\n"
                "umull    v15.8h, v0.8b, %[b3].8b\n"

                // Skip loop if we are doing zero iterations of it.
                "cbz    %w[k], 2f\n"

                "1:\n"
                "uadalp    v16.4s, v12.8h\n"
                "umull2    v12.8h, v0.16b, %[b0].16b\n"
                "uadalp    v17.4s, v13.8h\n"
                "umull2    v13.8h, v0.16b, %[b1].16b\n"
                "uadalp    v18.4s, v14.8h\n"
                "umull2    v14.8h, v0.16b, %[b2].16b\n"
                "uadalp    v19.4s, v15.8h\n"
                "umull2    v15.8h, v0.16b, %[b3].16b\n"
                "ldr     q0, [%[a_ptr]]\n"

                "uadalp    v16.4s, v12.8h\n"
                "umull    v12.8h, v1.8b, %[b0].8b\n"
                "uadalp    v17.4s, v13.8h\n"
                "umull    v13.8h, v1.8b, %[b1].8b\n"
                "subs    %w[k], %w[k], #1\n"
                "uadalp    v18.4s, v14.8h\n"
                "umull    v14.8h, v1.8b, %[b2].8b\n"
                "uadalp    v19.4s, v15.8h\n"
                "umull    v15.8h, v1.8b, %[b3].8b\n"

                "uadalp    v20.4s, v12.8h\n"
                "umull2    v12.8h, v1.16b, %[b0].16b\n"
                "uadalp    v21.4s, v13.8h\n"
                "umull2    v13.8h, v1.16b, %[b1].16b\n" ASM_PREFETCH("[%[a_ptr], #256]")
                "uadalp    v22.4s, v14.8h\n"
                "umull2    v14.8h, v1.16b, %[b2].16b\n"
                "uadalp    v23.4s, v15.8h\n"
                "umull2    v15.8h, v1.16b, %[b3].16b\n"
                "ldr     q1, [%[a_ptr], #16]\n"

                "uadalp    v20.4s, v12.8h\n"
                "umull    v12.8h, v2.8b, %[b0].8b\n"
                "uadalp    v21.4s, v13.8h\n"
                "umull    v13.8h, v2.8b, %[b1].8b\n" ASM_PREFETCH("[%[b_ptr], #256]")
                "uadalp    v22.4s, v14.8h\n"
                "umull    v14.8h, v2.8b, %[b2].8b\n"
                "uadalp    v23.4s, v15.8h\n"
                "umull    v15.8h, v2.8b, %[b3].8b\n"

                "uadalp    v24.4s, v12.8h\n"
                "umull2    v12.8h, v2.16b, %[b0].16b\n"
                "uadalp    v25.4s, v13.8h\n"
                "umull2    v13.8h, v2.16b, %[b1].16b\n"
                "uadalp    v26.4s, v14.8h\n"
                "umull2    v14.8h, v2.16b, %[b2].16b\n"
                "uadalp    v27.4s, v15.8h\n"
                "umull2    v15.8h, v2.16b, %[b3].16b\n"
                "ldr    q2, [%[a_ptr], #32]\n"

                "uadalp    v24.4s, v12.8h\n"
                "umull    v12.8h, v3.8b, %[b0].8b\n"
                "uadalp    v25.4s, v13.8h\n"
                "umull    v13.8h, v3.8b, %[b1].8b\n"
                "uadalp    v26.4s, v14.8h\n"
                "umull    v14.8h, v3.8b, %[b2].8b\n"
                "uadalp    v27.4s, v15.8h\n"
                "umull    v15.8h, v3.8b, %[b3].8b\n"

                "uadalp    v28.4s, v12.8h\n"
                "umull2    v12.8h, v3.16b, %[b0].16b\n"
                "ldr    %q[b0], [%[b_ptr]]\n"
                "uadalp    v29.4s, v13.8h\n"
                "umull2    v13.8h, v3.16b, %[b1].16b\n"
                "ldr    %q[b1], [%[b_ptr], #16]\n"
                "uadalp    v30.4s, v14.8h\n"
                "umull2    v14.8h, v3.16b, %[b2].16b\n"
                "ldr    %q[b2], [%[b_ptr], #32]\n"
                "uadalp    v31.4s, v15.8h\n"
                "umull2    v15.8h, v3.16b, %[b3].16b\n"
                "ldr    %q[b3], [%[b_ptr], #48]\n"

                "uadalp    v28.4s, v12.8h\n"
                "umull    v12.8h, v0.8b, %[b0].8b\n"
                "add    %[b_ptr], %[b_ptr], #64\n"
                "uadalp    v29.4s, v13.8h\n"
                "umull    v13.8h, v0.8b, %[b1].8b\n"
                "ldr    q3, [%[a_ptr], #48]\n"
                "uadalp    v30.4s, v14.8h\n"
                "umull    v14.8h, v0.8b, %[b2].8b\n"
                "add    %[a_ptr], %[a_ptr], #64\n"
                "uadalp    v31.4s, v15.8h\n"
                "umull    v15.8h, v0.8b, %[b3].8b\n"
                "bne    1b\n"

                // Branch target
                "2:\n"
                "uadalp    v16.4s, v12.8h\n"
                "umull2    v12.8h, v0.16b, %[b0].16b\n"
                "uadalp    v17.4s, v13.8h\n"
                "umull2    v13.8h, v0.16b, %[b1].16b\n"
                "uadalp    v18.4s, v14.8h\n"
                "umull2    v14.8h, v0.16b, %[b2].16b\n"
                "uadalp    v19.4s, v15.8h\n"
                "umull2    v15.8h, v0.16b, %[b3].16b\n"

                "uadalp    v16.4s, v12.8h\n"
                "umull    v12.8h, v1.8b, %[b0].8b\n"
                "uadalp    v17.4s, v13.8h\n"
                "umull    v13.8h, v1.8b, %[b1].8b\n"
                "uadalp    v18.4s, v14.8h\n"
                "umull    v14.8h, v1.8b, %[b2].8b\n"
                "uadalp    v19.4s, v15.8h\n"
                "umull    v15.8h, v1.8b, %[b3].8b\n"

                "uadalp    v20.4s, v12.8h\n"
                "umull2    v12.8h, v1.16b, %[b0].16b\n"
                "uadalp    v21.4s, v13.8h\n"
                "umull2    v13.8h, v1.16b, %[b1].16b\n"
                "uadalp    v22.4s, v14.8h\n"
                "umull2    v14.8h, v1.16b, %[b2].16b\n"
                "uadalp    v23.4s, v15.8h\n"
                "umull2    v15.8h, v1.16b, %[b3].16b\n"

                "uadalp    v20.4s, v12.8h\n"
                "umull    v12.8h, v2.8b, %[b0].8b\n"
                "uadalp    v21.4s, v13.8h\n"
                "umull    v13.8h, v2.8b, %[b1].8b\n"
                "uadalp    v22.4s, v14.8h\n"
                "umull    v14.8h, v2.8b, %[b2].8b\n"
                "uadalp    v23.4s, v15.8h\n"
                "umull    v15.8h, v2.8b, %[b3].8b\n"

                "uadalp    v24.4s, v12.8h\n"
                "umull2    v12.8h, v2.16b, %[b0].16b\n"
                "uadalp    v25.4s, v13.8h\n"
                "umull2    v13.8h, v2.16b, %[b1].16b\n"
                "uadalp    v26.4s, v14.8h\n"
                "umull2    v14.8h, v2.16b, %[b2].16b\n"
                "uadalp    v27.4s, v15.8h\n"
                "umull2    v15.8h, v2.16b, %[b3].16b\n"

                "uadalp    v24.4s, v12.8h\n"
                "umull    v12.8h, v3.8b, %[b0].8b\n"
                "uadalp    v25.4s, v13.8h\n"
                "umull    v13.8h, v3.8b, %[b1].8b\n"
                "uadalp    v26.4s, v14.8h\n"
                "umull    v14.8h, v3.8b, %[b2].8b\n"
                "uadalp    v27.4s, v15.8h\n"
                "umull    v15.8h, v3.8b, %[b3].8b\n"

                "uadalp    v28.4s, v12.8h\n"
                "umull2    v12.8h, v3.16b, %[b0].16b\n"
                "uadalp    v29.4s, v13.8h\n"
                "umull2    v13.8h, v3.16b, %[b1].16b\n"
                "uadalp    v30.4s, v14.8h\n"
                "umull2    v14.8h, v3.16b, %[b2].16b\n"
                "uadalp    v31.4s, v15.8h\n"
                "umull2    v15.8h, v3.16b, %[b3].16b\n"

                "uadalp    v28.4s, v12.8h\n"
                "uadalp    v29.4s, v13.8h\n"
                "uadalp    v30.4s, v14.8h\n"
                "uadalp    v31.4s, v15.8h\n"

                "addp    v16.4s, v16.4s, v17.4s\n"
                "addp    v17.4s, v18.4s, v19.4s\n"
                "addp    v18.4s, v20.4s, v21.4s\n"
                "addp    v19.4s, v22.4s, v23.4s\n"
                "addp    v20.4s, v24.4s, v25.4s\n"
                "addp    v21.4s, v26.4s, v27.4s\n"
                "addp    v22.4s, v28.4s, v29.4s\n"
                "addp    v23.4s, v30.4s, v31.4s\n"

                "addp    v16.4s, v16.4s, v17.4s\n"
                "addp    v17.4s, v18.4s, v19.4s\n"
                "addp    v18.4s, v20.4s, v21.4s\n"
                "addp    v19.4s, v22.4s, v23.4s\n"

                "str    q16, [%[c_ptr]]\n"
                "str    q17, [%[c_ptr], #16]\n"
                "str    q18, [%[c_ptr], #32]\n"
                "str    q19, [%[c_ptr], #48]\n"
                "add    %[c_ptr], %[c_ptr], #64\n"

                :
                [a_ptr] "+r"(a_ptr), [b_ptr] "+r"(b_ptr), [c_ptr] "+r"(c_ptr),
                [b0] "+w"(b0), [b1] "+w"(b1), [b2] "+w"(b2), [b3] "+w"(b3),
                [k] "+r"(k)
                :
                : "x20", "x21", "v0", "v1", "v2", "v3", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19",
                "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "cc");
        }
    }
}

} // namespace arm_gemm

#endif // __aarch64__