24.04
generic.cpp
Go to the documentation of this file.
1
/*
2
* Copyright (c) 2017-2018 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
#ifdef __aarch64__
25
26
#include <arm_neon.h>
27
28
#include "../../asmlib.hpp"
29
30
// Kernel implementation.
31
//
32
// Assume that "Apanel" points to a chunk of A blocks (each size 8xK) in read-order.
33
// Assume that "Bpanel" points to a chunk of B blocks (each size 12xK) in read-order.
34
// Assume that "Cpanel" points to a chunk of C output blocks (each size
35
// 12x8), the chunks being arranged in a row major fashion.
36
//
37
// Note that the intent of this is that either ablocks or bblocks will be 1
38
// - this construction allows the output loop to proceed in either order.
39
40
namespace
arm_gemm
{
41
42
void
a64_sgemm_asimd_8x12(
const
float
*Apanel,
const
float
*Bpanel,
float
*Cpanel,
int
ablocks,
int
bblocks,
int
K
) {
43
const
float
*a_ptr = Apanel;
44
float
*c_ptr = Cpanel;
45
46
for
(
int
yb=0; yb<ablocks; yb++) {
47
const
float
*a_ptr0 = a_ptr;
48
const
float
*b_ptr = Bpanel;
49
50
for
(
int
xb=0; xb<bblocks; xb++) {
51
a_ptr = a_ptr0;
52
// Fix up for odd lengths - set a flag if K is odd, but make
53
// sure we round up the iteration count.
54
int
oddk = (
K
& 1);
55
int
k = ((
K
+1)/2) - 1;
56
57
register
float32x4_t a0
asm
(
"v0"
);
58
register
float32x4_t a1
asm
(
"v1"
);
59
register
float32x4_t b0
asm
(
"v2"
);
60
register
float32x4_t b1
asm
(
"v3"
);
61
register
float32x4_t b2
asm
(
"v4"
);
62
register
float32x4_t a0a
asm
(
"v5"
);
63
register
float32x4_t a1a
asm
(
"v6"
);
64
65
__asm __volatile (
66
// Initialize result registers, load initial operands, prime prefetches.
67
"movi v8.4s, #0x0\n"
68
"ldr %q[a0], [%[a_ptr]]\n"
69
"movi v9.4s, #0x0\n"
70
"ldr %q[b0], [%[b_ptr]]\n"
71
"movi v10.4s, #0x0\n"
72
"ldr %q[a1], [%[a_ptr], #16]\n"
73
"movi v11.4s, #0x0\n"
74
"ldr %q[b1], [%[b_ptr], #16]\n"
75
"movi v12.4s, #0x0\n"
76
ASM_PREFETCH
(
"[%[b_ptr], #64]"
)
77
"movi v13.4s, #0x0\n"
78
ASM_PREFETCH
(
"[%[a_ptr], #64]"
)
79
"movi v14.4s, #0x0\n"
80
ASM_PREFETCH
(
"[%[b_ptr], #128]"
)
81
"movi v15.4s, #0x0\n"
82
ASM_PREFETCH
(
"[%[a_ptr], #128]"
)
83
"movi v16.4s, #0x0\n"
84
ASM_PREFETCH
(
"[%[b_ptr], #192]"
)
85
"movi v17.4s, #0x0\n"
86
ASM_PREFETCH
(
"[%[b_ptr], #256]"
)
87
"movi v18.4s, #0x0\n"
88
ASM_PREFETCH
(
"[%[a_ptr], #192]"
)
89
"movi v19.4s, #0x0\n"
90
ASM_PREFETCH
(
"[%[b_ptr], #320]"
)
91
"movi v20.4s, #0x0\n"
92
ASM_PREFETCH
(
"[%[a_ptr], #256]"
)
93
"movi v21.4s, #0x0\n"
94
ASM_PREFETCH
(
"[%[b_ptr], #384]"
)
95
"movi v22.4s, #0x0\n"
96
"movi v23.4s, #0x0\n"
97
"movi v24.4s, #0x0\n"
98
"movi v25.4s, #0x0\n"
99
"movi v26.4s, #0x0\n"
100
"movi v27.4s, #0x0\n"
101
"movi v28.4s, #0x0\n"
102
"movi v29.4s, #0x0\n"
103
"movi v30.4s, #0x0\n"
104
"movi v31.4s, #0x0\n"
105
106
// Skip loop if we are doing zero iterations of it.
107
"cbz %w[k], 4f\n"
108
109
// Loop proper
110
"1:\n"
111
"fmla v8.4s , %[b0].4s, %[a0].s[0]\n"
112
"fmla v9.4s , %[b0].4s, %[a0].s[1]\n"
113
"ldr %q[b2], [%[b_ptr], #32]\n"
114
"fmla v10.4s, %[b0].4s, %[a0].s[2]\n"
115
"fmla v11.4s, %[b0].4s, %[a0].s[3]\n"
116
"ldr %q[a0a], [%[a_ptr], #32]\n"
117
"fmla v12.4s, %[b0].4s, %[a1].s[0]\n"
118
"fmla v13.4s, %[b0].4s, %[a1].s[1]\n"
119
"ldr %q[a1a], [%[a_ptr], #48]\n"
120
"fmla v14.4s, %[b0].4s, %[a1].s[2]\n"
121
"fmla v15.4s, %[b0].4s, %[a1].s[3]\n"
122
"ldr %q[b0], [%[b_ptr], #48]\n"
123
124
"fmla v16.4s, %[b1].4s, %[a0].s[0]\n"
125
"fmla v17.4s, %[b1].4s, %[a0].s[1]\n"
126
ASM_PREFETCH
(
"[%[a_ptr], #320]"
)
127
"fmla v18.4s, %[b1].4s, %[a0].s[2]\n"
128
"fmla v19.4s, %[b1].4s, %[a0].s[3]\n"
129
"fmla v20.4s, %[b1].4s, %[a1].s[0]\n"
130
"fmla v21.4s, %[b1].4s, %[a1].s[1]\n"
131
"fmla v22.4s, %[b1].4s, %[a1].s[2]\n"
132
"fmla v23.4s, %[b1].4s, %[a1].s[3]\n"
133
"ldr %q[b1], [%[b_ptr], #64]\n"
134
135
"fmla v24.4s, %[b2].4s, %[a0].s[0]\n"
136
"fmla v25.4s, %[b2].4s, %[a0].s[1]\n"
137
ASM_PREFETCH
(
"[%[b_ptr], #448]"
)
138
"fmla v26.4s, %[b2].4s, %[a0].s[2]\n"
139
"fmla v27.4s, %[b2].4s, %[a0].s[3]\n"
140
"fmla v28.4s, %[b2].4s, %[a1].s[0]\n"
141
"fmla v29.4s, %[b2].4s, %[a1].s[1]\n"
142
"fmla v30.4s, %[b2].4s, %[a1].s[2]\n"
143
"fmla v31.4s, %[b2].4s, %[a1].s[3]\n"
144
"ldr %q[b2], [%[b_ptr], #80]\n"
145
146
"fmla v8.4s , %[b0].4s, %[a0a].s[0]\n"
147
"fmla v9.4s , %[b0].4s, %[a0a].s[1]\n"
148
"ldr %q[a0], [%[a_ptr], #64]\n"
149
"fmla v10.4s, %[b0].4s, %[a0a].s[2]\n"
150
"fmla v11.4s, %[b0].4s, %[a0a].s[3]\n"
151
"fmla v12.4s, %[b0].4s, %[a1a].s[0]\n"
152
"ldr %q[a1], [%[a_ptr], #80]\n"
153
"fmla v13.4s, %[b0].4s, %[a1a].s[1]\n"
154
"fmla v14.4s, %[b0].4s, %[a1a].s[2]\n"
155
"fmla v15.4s, %[b0].4s, %[a1a].s[3]\n"
156
"ldr %q[b0], [%[b_ptr], #96]\n"
157
158
"fmla v16.4s, %[b1].4s, %[a0a].s[0]\n"
159
"fmla v17.4s, %[b1].4s, %[a0a].s[1]\n"
160
ASM_PREFETCH
(
"[%[b_ptr], #512]"
)
161
"fmla v18.4s, %[b1].4s, %[a0a].s[2]\n"
162
"fmla v19.4s, %[b1].4s, %[a0a].s[3]\n"
163
"fmla v20.4s, %[b1].4s, %[a1a].s[0]\n"
164
"fmla v21.4s, %[b1].4s, %[a1a].s[1]\n"
165
"fmla v22.4s, %[b1].4s, %[a1a].s[2]\n"
166
"fmla v23.4s, %[b1].4s, %[a1a].s[3]\n"
167
"ldr %q[b1], [%[b_ptr], #112]\n"
168
169
"fmla v24.4s, %[b2].4s, %[a0a].s[0]\n"
170
"fmla v25.4s, %[b2].4s, %[a0a].s[1]\n"
171
"add %[a_ptr], %[a_ptr], #64\n"
172
"fmla v26.4s, %[b2].4s, %[a0a].s[2]\n"
173
"fmla v27.4s, %[b2].4s, %[a0a].s[3]\n"
174
"add %[b_ptr], %[b_ptr], #96\n"
175
"fmla v28.4s, %[b2].4s, %[a1a].s[0]\n"
176
"fmla v29.4s, %[b2].4s, %[a1a].s[1]\n"
177
"subs %w[k], %w[k], #1\n"
178
"fmla v30.4s, %[b2].4s, %[a1a].s[2]\n"
179
"fmla v31.4s, %[b2].4s, %[a1a].s[3]\n"
180
"bne 1b\n"
181
182
// Target to use when K is 1 or 2 (i.e. zero iterations of main loop)
183
"4:\n"
184
185
// Branch to alternative tail for odd K
186
"cbnz %w[oddk], 2f\n"
187
188
// Detached final iteration (even K)
189
"fmla v8.4s , %[b0].4s, %[a0].s[0]\n"
190
"fmla v9.4s , %[b0].4s, %[a0].s[1]\n"
191
"ldr %q[b2], [%[b_ptr], #32]\n"
192
"fmla v10.4s, %[b0].4s, %[a0].s[2]\n"
193
"fmla v11.4s, %[b0].4s, %[a0].s[3]\n"
194
"ldr %q[a0a], [%[a_ptr], #32]\n"
195
"fmla v12.4s, %[b0].4s, %[a1].s[0]\n"
196
"fmla v13.4s, %[b0].4s, %[a1].s[1]\n"
197
"ldr %q[a1a], [%[a_ptr], #48]\n"
198
"fmla v14.4s, %[b0].4s, %[a1].s[2]\n"
199
"fmla v15.4s, %[b0].4s, %[a1].s[3]\n"
200
"ldr %q[b0], [%[b_ptr], #48]\n"
201
202
"fmla v16.4s, %[b1].4s, %[a0].s[0]\n"
203
"fmla v17.4s, %[b1].4s, %[a0].s[1]\n"
204
"fmla v18.4s, %[b1].4s, %[a0].s[2]\n"
205
"fmla v19.4s, %[b1].4s, %[a0].s[3]\n"
206
"fmla v20.4s, %[b1].4s, %[a1].s[0]\n"
207
"fmla v21.4s, %[b1].4s, %[a1].s[1]\n"
208
"fmla v22.4s, %[b1].4s, %[a1].s[2]\n"
209
"fmla v23.4s, %[b1].4s, %[a1].s[3]\n"
210
"ldr %q[b1], [%[b_ptr], #64]\n"
211
212
"fmla v24.4s, %[b2].4s, %[a0].s[0]\n"
213
"fmla v25.4s, %[b2].4s, %[a0].s[1]\n"
214
"add %[a_ptr], %[a_ptr], #64\n"
215
"fmla v26.4s, %[b2].4s, %[a0].s[2]\n"
216
"fmla v27.4s, %[b2].4s, %[a0].s[3]\n"
217
"fmla v28.4s, %[b2].4s, %[a1].s[0]\n"
218
"fmla v29.4s, %[b2].4s, %[a1].s[1]\n"
219
"fmla v30.4s, %[b2].4s, %[a1].s[2]\n"
220
"fmla v31.4s, %[b2].4s, %[a1].s[3]\n"
221
"ldr %q[b2], [%[b_ptr], #80]\n"
222
223
"fmla v8.4s , %[b0].4s, %[a0a].s[0]\n"
224
"fmla v16.4s, %[b1].4s, %[a0a].s[0]\n"
225
"add %[b_ptr], %[b_ptr], #96\n"
226
"fmla v9.4s , %[b0].4s, %[a0a].s[1]\n"
227
"str q8, [%[c_ptr], #0]\n"
228
"fmla v17.4s, %[b1].4s, %[a0a].s[1]\n"
229
"str q16, [%[c_ptr], #16]\n"
230
"fmla v24.4s, %[b2].4s, %[a0a].s[0]\n"
231
"str q24, [%[c_ptr], #32]\n"
232
233
"fmla v25.4s, %[b2].4s, %[a0a].s[1]\n"
234
"str q9, [%[c_ptr], #48]\n"
235
"fmla v10.4s, %[b0].4s, %[a0a].s[2]\n"
236
"str q17, [%[c_ptr], #64]\n"
237
"fmla v18.4s, %[b1].4s, %[a0a].s[2]\n"
238
"str q25, [%[c_ptr], #80]\n"
239
"fmla v26.4s, %[b2].4s, %[a0a].s[2]\n"
240
"str q10, [%[c_ptr], #96]\n"
241
242
"fmla v11.4s, %[b0].4s, %[a0a].s[3]\n"
243
"str q18, [%[c_ptr], #112]\n"
244
"fmla v19.4s, %[b1].4s, %[a0a].s[3]\n"
245
"str q26, [%[c_ptr], #128]\n"
246
"fmla v27.4s, %[b2].4s, %[a0a].s[3]\n"
247
"str q11, [%[c_ptr], #144]\n"
248
249
"fmla v12.4s, %[b0].4s, %[a1a].s[0]\n"
250
"str q19, [%[c_ptr], #160]\n"
251
"fmla v20.4s, %[b1].4s, %[a1a].s[0]\n"
252
"str q27, [%[c_ptr], #176]\n"
253
"fmla v28.4s, %[b2].4s, %[a1a].s[0]\n"
254
"str q12, [%[c_ptr], #192]\n"
255
256
"fmla v13.4s, %[b0].4s, %[a1a].s[1]\n"
257
"str q20, [%[c_ptr], #208]\n"
258
"fmla v21.4s, %[b1].4s, %[a1a].s[1]\n"
259
"str q28, [%[c_ptr], #224]\n"
260
"fmla v29.4s, %[b2].4s, %[a1a].s[1]\n"
261
"str q13, [%[c_ptr], #240]\n"
262
263
"fmla v14.4s, %[b0].4s, %[a1a].s[2]\n"
264
"str q21, [%[c_ptr], #256]\n"
265
"fmla v22.4s, %[b1].4s, %[a1a].s[2]\n"
266
"str q29, [%[c_ptr], #272]\n"
267
"fmla v30.4s, %[b2].4s, %[a1a].s[2]\n"
268
"str q14, [%[c_ptr], #288]\n"
269
270
"fmla v15.4s, %[b0].4s, %[a1a].s[3]\n"
271
"str q22, [%[c_ptr], #304]\n"
272
"fmla v23.4s, %[b1].4s, %[a1a].s[3]\n"
273
"str q30, [%[c_ptr], #320]\n"
274
"fmla v31.4s, %[b2].4s, %[a1a].s[3]\n"
275
"str q15, [%[c_ptr], #336]\n"
276
277
"b 3f\n"
278
279
// Detached final iteration (odd K)
280
"2:\n"
281
"fmla v8.4s , %[b0].4s, %[a0].s[0]\n"
282
"ldr %q[b2], [%[b_ptr], #32]\n"
283
"fmla v16.4s, %[b1].4s, %[a0].s[0]\n"
284
"fmla v9.4s , %[b0].4s, %[a0].s[1]\n"
285
"str q8, [%[c_ptr], #0]\n"
286
"fmla v17.4s, %[b1].4s, %[a0].s[1]\n"
287
"str q16, [%[c_ptr], #16]\n"
288
"fmla v24.4s, %[b2].4s, %[a0].s[0]\n"
289
"add %[b_ptr], %[b_ptr], #48\n"
290
"add %[a_ptr], %[a_ptr], #32\n"
291
"str q24, [%[c_ptr], #32]\n"
292
"fmla v25.4s, %[b2].4s, %[a0].s[1]\n"
293
"str q9, [%[c_ptr], #48]\n"
294
295
"fmla v10.4s, %[b0].4s, %[a0].s[2]\n"
296
"str q17, [%[c_ptr], #64]\n"
297
"fmla v18.4s, %[b1].4s, %[a0].s[2]\n"
298
"str q25, [%[c_ptr], #80]\n"
299
"fmla v26.4s, %[b2].4s, %[a0].s[2]\n"
300
"str q10, [%[c_ptr], #96]\n"
301
302
"fmla v11.4s, %[b0].4s, %[a0].s[3]\n"
303
"str q18, [%[c_ptr], #112]\n"
304
"fmla v19.4s, %[b1].4s, %[a0].s[3]\n"
305
"str q26, [%[c_ptr], #128]\n"
306
"fmla v27.4s, %[b2].4s, %[a0].s[3]\n"
307
"str q11, [%[c_ptr], #144]\n"
308
309
"fmla v12.4s, %[b0].4s, %[a1].s[0]\n"
310
"str q19, [%[c_ptr], #160]\n"
311
"fmla v20.4s, %[b1].4s, %[a1].s[0]\n"
312
"str q27, [%[c_ptr], #176]\n"
313
"fmla v28.4s, %[b2].4s, %[a1].s[0]\n"
314
"str q12, [%[c_ptr], #192]\n"
315
316
"fmla v13.4s, %[b0].4s, %[a1].s[1]\n"
317
"str q20, [%[c_ptr], #208]\n"
318
"fmla v21.4s, %[b1].4s, %[a1].s[1]\n"
319
"str q28, [%[c_ptr], #224]\n"
320
"fmla v29.4s, %[b2].4s, %[a1].s[1]\n"
321
"str q13, [%[c_ptr], #240]\n"
322
323
"fmla v14.4s, %[b0].4s, %[a1].s[2]\n"
324
"str q21, [%[c_ptr], #256]\n"
325
"fmla v22.4s, %[b1].4s, %[a1].s[2]\n"
326
"str q29, [%[c_ptr], #272]\n"
327
"fmla v30.4s, %[b2].4s, %[a1].s[2]\n"
328
"str q14, [%[c_ptr], #288]\n"
329
330
"fmla v15.4s, %[b0].4s, %[a1].s[3]\n"
331
"str q22, [%[c_ptr], #304]\n"
332
"fmla v23.4s, %[b1].4s, %[a1].s[3]\n"
333
"str q30, [%[c_ptr], #320]\n"
334
"fmla v31.4s, %[b2].4s, %[a1].s[3]\n"
335
"str q15, [%[c_ptr], #336]\n"
336
337
// Common tail
338
"3:\n"
339
"str q23, [%[c_ptr], #352]\n"
340
"str q31, [%[c_ptr], #368]\n"
341
"add %[c_ptr], %[c_ptr], #384\n"
342
:
343
[a_ptr]
"+r"
(a_ptr), [b_ptr]
"+r"
(b_ptr), [c_ptr]
"+r"
(c_ptr),
344
[a0]
"+w"
(a0), [a1]
"+w"
(a1), [a0a]
"+w"
(a0a), [a1a]
"+w"
(a1a),
345
[b0]
"+w"
(b0), [b1]
"+w"
(b1), [b2]
"+w"
(b2), [k]
"+r"
(k)
346
: [oddk]
"r"
(oddk)
347
:
"x20"
,
"x21"
,
"v8"
,
"v9"
,
"v10"
,
"v11"
,
"v12"
,
"v13"
,
"v14"
,
"v15"
,
"v16"
,
"v17"
,
"v18"
,
348
"v19"
,
"v20"
,
"v21"
,
"v22"
,
"v23"
,
"v24"
,
"v25"
,
"v26"
,
"v27"
,
"v28"
,
"v29"
,
"v30"
,
"v31"
,
"cc"
349
);
350
}
351
}
352
}
353
354
}
// namespace arm_gemm
355
356
#endif
ASM_PREFETCH
#define ASM_PREFETCH(address)
Definition:
asmlib.hpp:45
arm_gemm
Definition:
barrier.hpp:30
K
unsigned int K
Definition:
CpuGemmAssemblyDispatch.cpp:106
src
core
NEON
kernels
arm_gemm
kernels
a64_sgemm_8x12
generic.cpp
Generated on Mon Apr 29 2024 10:53:55 for Compute Library by
1.8.17