1 /*
2 * Copyright (c) 2019-2021 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
14 * all 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
21 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
22 * IN THE SOFTWARE.
23 */
24 #ifdef __aarch64__
25
26 #include "arm_gemm.hpp"
27 #include "../../utils.hpp"
28
29 #include <cassert>
30 #include <limits>
31
32 namespace arm_gemm {
33
a64_hybrid_fp32_mla_8x4(unsigned int num_strings,const unsigned int * string_lengths,IndirectInputArg<float> A_arg,size_t M,size_t N,const float * B_ptr,IndirectOutputArg<float> output_arg,const float * bias,Activation act,bool accumulate)34 void a64_hybrid_fp32_mla_8x4 (
35 unsigned int num_strings, const unsigned int *string_lengths, IndirectInputArg<float> A_arg,
36 size_t M, size_t N, const float *B_ptr, IndirectOutputArg<float> output_arg,
37 const float *bias, Activation act, bool accumulate
38 )
39 {
40 struct KernelArgs {
41 float maxval = static_cast<float>(std::numeric_limits<float>::infinity());
42 float minval = - static_cast<float>(std::numeric_limits<float>::infinity());
43 unsigned int num_strings = {};
44 const unsigned int *string_lengths = {};
45 size_t N = {};
46 const float *B_ptr = {};
47 size_t output_offset = {};
48 size_t input_initial_col = {};
49 size_t input_offset = {};
50 } ka;
51
52 unsigned long flags=0;
53 void *output_ptr;
54 void *input_ptr;
55
56 if (output_arg.is_indirect) {
57 output_ptr=(void *)(output_arg.indirect.ptr);
58 ka.output_offset=output_arg.indirect.offset;
59 flags |= 0x4;
60 } else {
61 output_ptr=(void *)(output_arg.direct.base);
62 ka.output_offset=output_arg.direct.stride;
63 }
64
65 if (A_arg.is_indirect) {
66 input_ptr=(void *)(A_arg.indirect.ptr);
67 ka.input_offset=A_arg.indirect.start_row;
68 ka.input_initial_col=A_arg.indirect.start_col;
69 flags |= 0x8;
70 } else {
71 assert(num_strings==1);
72 input_ptr=(void *)(A_arg.direct.base);
73 ka.input_offset=A_arg.direct.stride;
74 }
75 if (accumulate) {
76 flags |= 0x1;
77 }
78 ka.num_strings = num_strings;
79 ka.string_lengths = string_lengths;
80 ka.N = N;
81 ka.B_ptr = B_ptr;
82 switch(act.type) {
83 default:
84 case Activation::Type::None:
85 break;
86 case Activation::Type::BoundedReLU:
87 ka.maxval = static_cast<float>(act.param1);
88 /* fall through */
89 case Activation::Type::ReLU:
90 ka.minval = 0;
91 flags |= 0x2;
92 break;
93 }
94 __asm__ __volatile__(
95
96 "1:" // Row loop
97 "cmp %x[M], #0x8\n"
98 "bge 148f\n"
99 "cmp %x[M], #0x6\n"
100 "bgt 127f\n"
101 "beq 106f\n"
102 "cmp %x[M], #0x4\n"
103 "bgt 85f\n"
104 "beq 64f\n"
105 "cmp %x[M], #0x2\n"
106 "bgt 43f\n"
107 "beq 22f\n"
108 "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
109 "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
110 "mov x11, %x[bias]\n"
111 "mov x10, %x[output_ptr]\n"
112 "2:" // Height 1: Column loop
113 "cbz x11, 3f\n"
114 "ldr q24, [x11, #0x0]\n"
115 "add x11, x11, #0x10\n"
116 "b 8f\n"
117 "3:" // Height 1: no bias
118 "tbz %x[flags], #0, 7f\n"
119 "cmp x13, #0x4\n"
120 "bge 6f\n"
121 "tbz x13, #1, 4f\n"
122 "ldr d24, [x10], #0x8\n"
123 "mov x19, #0x8\n"
124 "tbz x13, #0, 5f\n"
125 "ld1 { v24.s }[2], [x10]\n"
126 "b 5f\n"
127 "4:" // Height 1: Partial accumulate: partial_1_0
128 "ldr s24, [x10, #0x0]\n"
129 "mov x19, #0x0\n"
130 "5:" // Height 1: Partial accumulate: Done
131 "sub x10, x10, x19\n"
132 "b 8f\n"
133 "6:" // Height 1: full accumulate
134 "ldr q24, [x10, #0x0]\n"
135 "b 8f\n"
136 "7:" // Height 1: no accumulate
137 "movi v24.16b, #0x0\n"
138 "8:" // Height 1: setup done
139 "mov x9, #0x0\n"
140 "9:" // Height 1: String loop
141 "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
142 "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n"
143 "ldr w28, [x20, x9, LSL #0x2]\n"
144 "tbz %x[flags], #3, 10f\n"
145 "ldr x20, [%x[input_ptr], x9, LSL #0x3]\n"
146 "add x20, x20, x19, LSL #3\n"
147 "ldr x27, [x20, #0x0]\n"
148 "cbnz x9, 11f\n"
149 "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
150 "add x27, x27, x19, LSL #2\n"
151 "b 11f\n"
152 "10:" // Height 1: setup direct input
153 "mov x27, %x[input_ptr]\n"
154 "11:" // Height 1: input setup done
155 "cmp x28, #0x4\n"
156 "blt 14f\n"
157 "ldr q0, [x27, #0x0]\n"
158 "ldr q8, [x12, #0x0]\n"
159 "cmp x28, #0x8\n"
160 "blt 13f\n"
161 "12:" // Height 1: Multiply loop: Main loop head
162 "fmla v24.4s, v8.4s, v0.s[0]\n"
163 "ldr q9, [x12, #0x10]\n"
164 "add x27, x27, #0x10\n"
165 "fmla v24.4s, v9.4s, v0.s[1]\n"
166 "ldr q10, [x12, #0x20]\n"
167 "sub x28, x28, #0x4\n"
168 "fmla v24.4s, v10.4s, v0.s[2]\n"
169 "ldr q11, [x12, #0x30]\n"
170 "cmp x28, #0x8\n"
171 "fmla v24.4s, v11.4s, v0.s[3]\n"
172 "prfm pldl1keep, [x27, #0x80]\n"
173 "add x12, x12, #0x40\n"
174 "ldr q0, [x27, #0x0]\n"
175 "ldr q8, [x12, #0x0]\n"
176 "bge 12b\n"
177 "13:" // Height 1: Multiply loop: Single iteration only
178 "fmla v24.4s, v8.4s, v0.s[0]\n"
179 "ldr q9, [x12, #0x10]\n"
180 "sub x28, x28, #0x4\n"
181 "fmla v24.4s, v9.4s, v0.s[1]\n"
182 "ldr q10, [x12, #0x20]\n"
183 "add x27, x27, #0x10\n"
184 "fmla v24.4s, v10.4s, v0.s[2]\n"
185 "ldr q11, [x12, #0x30]\n"
186 "add x12, x12, #0x40\n"
187 "fmla v24.4s, v11.4s, v0.s[3]\n"
188 "prfm pldl1keep, [x27, #0x80]\n"
189 "14:" // Height 1: Multiply loop: Main loop skip
190 "cbz x28, 16f\n"
191 "15:" // Height 1: Multiply loop: Odd block loop
192 "ldr s0, [x27], #0x4\n"
193 "sub x28, x28, #0x1\n"
194 "ldr q12, [x12, #0x0]\n"
195 "fmla v24.4s, v12.4s, v0.s[0]\n"
196 "add x12, x12, #0x10\n"
197 "cbnz x28, 15b\n"
198 "16:" // Height 1: Multiply loop: No odd multiplies
199 "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n"
200 "add x9, x9, #0x1\n"
201 "cmp x9, x19\n"
202 "bne 9b\n"
203 "prfm pstl1keep, [x10, #0x0]\n"
204 "tbz %x[flags], #1, 17f\n"
205 "add x19, %x[args_ptr], %[offset_min]\n"
206 "ld1r { v17.4s }, [x19]\n"
207 "add x19, %x[args_ptr], %[offset_max]\n"
208 "ld1r { v16.4s }, [x19]\n"
209 "fmin v24.4s, v24.4s, v16.4s\n"
210 "fmax v24.4s, v24.4s, v17.4s\n"
211 "17:" // Height 1: No activation
212 "cmp x13, #0x4\n"
213 "bge 20f\n"
214 "tbz x13, #1, 18f\n"
215 "str d24, [x10], #0x8\n"
216 "tbz x13, #0, 19f\n"
217 "st1 { v24.s }[2], [x10]\n"
218 "b 19f\n"
219 "18:" // Height 1: Partial direct writeback: partial_1_0
220 "str s24, [x10, #0x0]\n"
221 "19:" // Height 1: Partial direct writeback: Done
222 "b 21f\n"
223 "20:" // Height 1: Full writeback
224 "str q24, [x10, #0x0]\n"
225 "add x10, x10, #0x10\n"
226 "21:" // Height 1: Writeback done
227 "subs x13, x13, #0x4\n"
228 "bgt 2b\n"
229 "b 170f\n"
230 "22:" // Height 2
231 "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
232 "mov x11, %x[bias]\n"
233 "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
234 "mov x10, %x[output_ptr]\n"
235 "23:" // Height 2: Column loop
236 "cbz x11, 24f\n"
237 "ldr q24, [x11, #0x0]\n"
238 "mov v25.16b, v24.16b\n"
239 "add x11, x11, #0x10\n"
240 "b 29f\n"
241 "24:" // Height 2: no bias
242 "tbz %x[flags], #0, 28f\n"
243 "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
244 "cmp x13, #0x4\n"
245 "add x26, x10, x19, LSL #2\n"
246 "bge 27f\n"
247 "tbz x13, #1, 25f\n"
248 "ldr d24, [x10], #0x8\n"
249 "ldr d25, [x26], #0x8\n"
250 "mov x19, #0x8\n"
251 "tbz x13, #0, 26f\n"
252 "ld1 { v24.s }[2], [x10]\n"
253 "ld1 { v25.s }[2], [x26]\n"
254 "b 26f\n"
255 "25:" // Height 2: Partial accumulate: partial_1_0
256 "ldr s24, [x10, #0x0]\n"
257 "mov x19, #0x0\n"
258 "ldr s25, [x26, #0x0]\n"
259 "26:" // Height 2: Partial accumulate: Done
260 "sub x10, x10, x19\n"
261 "b 29f\n"
262 "27:" // Height 2: full accumulate
263 "ldr q24, [x10, #0x0]\n"
264 "ldr q25, [x26, #0x0]\n"
265 "b 29f\n"
266 "28:" // Height 2: no accumulate
267 "movi v24.16b, #0x0\n"
268 "movi v25.16b, #0x0\n"
269 "29:" // Height 2: setup done
270 "mov x9, #0x0\n"
271 "30:" // Height 2: String loop
272 "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
273 "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n"
274 "ldr w28, [x20, x9, LSL #0x2]\n"
275 "tbz %x[flags], #3, 31f\n"
276 "ldr x20, [%x[input_ptr], x9, LSL #0x3]\n"
277 "add x20, x20, x19, LSL #3\n"
278 "ldr x27, [x20, #0x0]\n"
279 "ldr x26, [x20, #0x8]\n"
280 "cbnz x9, 32f\n"
281 "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
282 "add x27, x27, x19, LSL #2\n"
283 "add x26, x26, x19, LSL #2\n"
284 "b 32f\n"
285 "31:" // Height 2: setup direct input
286 "mov x27, %x[input_ptr]\n"
287 "add x26, x27, x19, LSL #2\n"
288 "32:" // Height 2: input setup done
289 "cmp x28, #0x4\n"
290 "blt 35f\n"
291 "ldr q0, [x27, #0x0]\n"
292 "ldr q1, [x26, #0x0]\n"
293 "cmp x28, #0x8\n"
294 "ldr q8, [x12, #0x0]\n"
295 "blt 34f\n"
296 "33:" // Height 2: Multiply loop: Main loop head
297 "fmla v24.4s, v8.4s, v0.s[0]\n"
298 "ldr q9, [x12, #0x10]\n"
299 "add x27, x27, #0x10\n"
300 "fmla v25.4s, v8.4s, v1.s[0]\n"
301 "ldr q10, [x12, #0x20]\n"
302 "add x26, x26, #0x10\n"
303 "fmla v24.4s, v9.4s, v0.s[1]\n"
304 "ldr q11, [x12, #0x30]\n"
305 "sub x28, x28, #0x4\n"
306 "fmla v25.4s, v9.4s, v1.s[1]\n"
307 "prfm pldl1keep, [x27, #0x80]\n"
308 "cmp x28, #0x8\n"
309 "fmla v24.4s, v10.4s, v0.s[2]\n"
310 "prfm pldl1keep, [x26, #0x80]\n"
311 "add x12, x12, #0x40\n"
312 "fmla v25.4s, v10.4s, v1.s[2]\n"
313 "ldr q8, [x12, #0x0]\n"
314 "fmla v24.4s, v11.4s, v0.s[3]\n"
315 "ldr q0, [x27, #0x0]\n"
316 "fmla v25.4s, v11.4s, v1.s[3]\n"
317 "ldr q1, [x26, #0x0]\n"
318 "bge 33b\n"
319 "34:" // Height 2: Multiply loop: Single iteration only
320 "fmla v24.4s, v8.4s, v0.s[0]\n"
321 "ldr q9, [x12, #0x10]\n"
322 "sub x28, x28, #0x4\n"
323 "fmla v25.4s, v8.4s, v1.s[0]\n"
324 "ldr q10, [x12, #0x20]\n"
325 "add x27, x27, #0x10\n"
326 "fmla v24.4s, v9.4s, v0.s[1]\n"
327 "ldr q11, [x12, #0x30]\n"
328 "add x26, x26, #0x10\n"
329 "fmla v25.4s, v9.4s, v1.s[1]\n"
330 "prfm pldl1keep, [x27, #0x80]\n"
331 "add x12, x12, #0x40\n"
332 "fmla v24.4s, v10.4s, v0.s[2]\n"
333 "prfm pldl1keep, [x26, #0x80]\n"
334 "fmla v25.4s, v10.4s, v1.s[2]\n"
335 "fmla v24.4s, v11.4s, v0.s[3]\n"
336 "fmla v25.4s, v11.4s, v1.s[3]\n"
337 "35:" // Height 2: Multiply loop: Main loop skip
338 "cbz x28, 37f\n"
339 "36:" // Height 2: Multiply loop: Odd block loop
340 "ldr s0, [x27], #0x4\n"
341 "sub x28, x28, #0x1\n"
342 "ldr s1, [x26], #0x4\n"
343 "ldr q12, [x12, #0x0]\n"
344 "fmla v24.4s, v12.4s, v0.s[0]\n"
345 "add x12, x12, #0x10\n"
346 "fmla v25.4s, v12.4s, v1.s[0]\n"
347 "cbnz x28, 36b\n"
348 "37:" // Height 2: Multiply loop: No odd multiplies
349 "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n"
350 "add x9, x9, #0x1\n"
351 "cmp x9, x19\n"
352 "bne 30b\n"
353 "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
354 "prfm pstl1keep, [x10, #0x0]\n"
355 "add x26, x10, x19, LSL #2\n"
356 "prfm pstl1keep, [x26, #0x0]\n"
357 "tbz %x[flags], #1, 38f\n"
358 "add x19, %x[args_ptr], %[offset_min]\n"
359 "ld1r { v17.4s }, [x19]\n"
360 "add x19, %x[args_ptr], %[offset_max]\n"
361 "ld1r { v16.4s }, [x19]\n"
362 "fmin v24.4s, v24.4s, v16.4s\n"
363 "fmin v25.4s, v25.4s, v16.4s\n"
364 "fmax v24.4s, v24.4s, v17.4s\n"
365 "fmax v25.4s, v25.4s, v17.4s\n"
366 "38:" // Height 2: No activation
367 "cmp x13, #0x4\n"
368 "bge 41f\n"
369 "tbz x13, #1, 39f\n"
370 "str d24, [x10], #0x8\n"
371 "str d25, [x26], #0x8\n"
372 "tbz x13, #0, 40f\n"
373 "st1 { v24.s }[2], [x10]\n"
374 "st1 { v25.s }[2], [x26]\n"
375 "b 40f\n"
376 "39:" // Height 2: Partial direct writeback: partial_1_0
377 "str s24, [x10, #0x0]\n"
378 "str s25, [x26, #0x0]\n"
379 "40:" // Height 2: Partial direct writeback: Done
380 "b 42f\n"
381 "41:" // Height 2: Full writeback
382 "str q24, [x10, #0x0]\n"
383 "add x10, x10, #0x10\n"
384 "str q25, [x26, #0x0]\n"
385 "42:" // Height 2: Writeback done
386 "subs x13, x13, #0x4\n"
387 "bgt 23b\n"
388 "b 170f\n"
389 "43:" // Height 3
390 "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
391 "mov x11, %x[bias]\n"
392 "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
393 "mov x10, %x[output_ptr]\n"
394 "44:" // Height 3: Column loop
395 "cbz x11, 45f\n"
396 "ldr q24, [x11, #0x0]\n"
397 "mov v25.16b, v24.16b\n"
398 "add x11, x11, #0x10\n"
399 "mov v26.16b, v24.16b\n"
400 "b 50f\n"
401 "45:" // Height 3: no bias
402 "tbz %x[flags], #0, 49f\n"
403 "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
404 "cmp x13, #0x4\n"
405 "add x26, x10, x19, LSL #2\n"
406 "add x25, x26, x19, LSL #2\n"
407 "bge 48f\n"
408 "tbz x13, #1, 46f\n"
409 "ldr d24, [x10], #0x8\n"
410 "ldr d25, [x26], #0x8\n"
411 "mov x19, #0x8\n"
412 "ldr d26, [x25], #0x8\n"
413 "tbz x13, #0, 47f\n"
414 "ld1 { v24.s }[2], [x10]\n"
415 "ld1 { v25.s }[2], [x26]\n"
416 "ld1 { v26.s }[2], [x25]\n"
417 "b 47f\n"
418 "46:" // Height 3: Partial accumulate: partial_1_0
419 "ldr s24, [x10, #0x0]\n"
420 "mov x19, #0x0\n"
421 "ldr s25, [x26, #0x0]\n"
422 "ldr s26, [x25, #0x0]\n"
423 "47:" // Height 3: Partial accumulate: Done
424 "sub x10, x10, x19\n"
425 "b 50f\n"
426 "48:" // Height 3: full accumulate
427 "ldr q24, [x10, #0x0]\n"
428 "ldr q25, [x26, #0x0]\n"
429 "ldr q26, [x25, #0x0]\n"
430 "b 50f\n"
431 "49:" // Height 3: no accumulate
432 "movi v24.16b, #0x0\n"
433 "movi v25.16b, #0x0\n"
434 "movi v26.16b, #0x0\n"
435 "50:" // Height 3: setup done
436 "mov x9, #0x0\n"
437 "51:" // Height 3: String loop
438 "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
439 "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n"
440 "ldr w28, [x20, x9, LSL #0x2]\n"
441 "tbz %x[flags], #3, 52f\n"
442 "ldr x20, [%x[input_ptr], x9, LSL #0x3]\n"
443 "add x20, x20, x19, LSL #3\n"
444 "ldr x27, [x20, #0x0]\n"
445 "ldr x26, [x20, #0x8]\n"
446 "ldr x25, [x20, #0x10]\n"
447 "cbnz x9, 53f\n"
448 "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
449 "add x27, x27, x19, LSL #2\n"
450 "add x26, x26, x19, LSL #2\n"
451 "add x25, x25, x19, LSL #2\n"
452 "b 53f\n"
453 "52:" // Height 3: setup direct input
454 "mov x27, %x[input_ptr]\n"
455 "add x26, x27, x19, LSL #2\n"
456 "add x25, x26, x19, LSL #2\n"
457 "53:" // Height 3: input setup done
458 "cmp x28, #0x4\n"
459 "blt 56f\n"
460 "ldr q0, [x27, #0x0]\n"
461 "ldr q1, [x26, #0x0]\n"
462 "cmp x28, #0x8\n"
463 "ldr q2, [x25, #0x0]\n"
464 "ldr q8, [x12, #0x0]\n"
465 "blt 55f\n"
466 "54:" // Height 3: Multiply loop: Main loop head
467 "fmla v24.4s, v8.4s, v0.s[0]\n"
468 "ldr q9, [x12, #0x10]\n"
469 "add x27, x27, #0x10\n"
470 "fmla v25.4s, v8.4s, v1.s[0]\n"
471 "ldr q10, [x12, #0x20]\n"
472 "add x26, x26, #0x10\n"
473 "fmla v26.4s, v8.4s, v2.s[0]\n"
474 "ldr q11, [x12, #0x30]\n"
475 "add x25, x25, #0x10\n"
476 "fmla v24.4s, v9.4s, v0.s[1]\n"
477 "prfm pldl1keep, [x27, #0x80]\n"
478 "sub x28, x28, #0x4\n"
479 "fmla v25.4s, v9.4s, v1.s[1]\n"
480 "prfm pldl1keep, [x26, #0x80]\n"
481 "cmp x28, #0x8\n"
482 "fmla v26.4s, v9.4s, v2.s[1]\n"
483 "prfm pldl1keep, [x25, #0x80]\n"
484 "add x12, x12, #0x40\n"
485 "fmla v24.4s, v10.4s, v0.s[2]\n"
486 "ldr q8, [x12, #0x0]\n"
487 "fmla v25.4s, v10.4s, v1.s[2]\n"
488 "fmla v26.4s, v10.4s, v2.s[2]\n"
489 "fmla v24.4s, v11.4s, v0.s[3]\n"
490 "ldr q0, [x27, #0x0]\n"
491 "fmla v25.4s, v11.4s, v1.s[3]\n"
492 "ldr q1, [x26, #0x0]\n"
493 "fmla v26.4s, v11.4s, v2.s[3]\n"
494 "ldr q2, [x25, #0x0]\n"
495 "bge 54b\n"
496 "55:" // Height 3: Multiply loop: Single iteration only
497 "fmla v24.4s, v8.4s, v0.s[0]\n"
498 "ldr q9, [x12, #0x10]\n"
499 "sub x28, x28, #0x4\n"
500 "fmla v25.4s, v8.4s, v1.s[0]\n"
501 "ldr q10, [x12, #0x20]\n"
502 "add x27, x27, #0x10\n"
503 "fmla v26.4s, v8.4s, v2.s[0]\n"
504 "ldr q11, [x12, #0x30]\n"
505 "add x26, x26, #0x10\n"
506 "fmla v24.4s, v9.4s, v0.s[1]\n"
507 "prfm pldl1keep, [x27, #0x80]\n"
508 "add x25, x25, #0x10\n"
509 "fmla v25.4s, v9.4s, v1.s[1]\n"
510 "prfm pldl1keep, [x26, #0x80]\n"
511 "add x12, x12, #0x40\n"
512 "fmla v26.4s, v9.4s, v2.s[1]\n"
513 "prfm pldl1keep, [x25, #0x80]\n"
514 "fmla v24.4s, v10.4s, v0.s[2]\n"
515 "fmla v25.4s, v10.4s, v1.s[2]\n"
516 "fmla v26.4s, v10.4s, v2.s[2]\n"
517 "fmla v24.4s, v11.4s, v0.s[3]\n"
518 "fmla v25.4s, v11.4s, v1.s[3]\n"
519 "fmla v26.4s, v11.4s, v2.s[3]\n"
520 "56:" // Height 3: Multiply loop: Main loop skip
521 "cbz x28, 58f\n"
522 "57:" // Height 3: Multiply loop: Odd block loop
523 "ldr s0, [x27], #0x4\n"
524 "sub x28, x28, #0x1\n"
525 "ldr s1, [x26], #0x4\n"
526 "ldr s2, [x25], #0x4\n"
527 "ldr q12, [x12, #0x0]\n"
528 "fmla v24.4s, v12.4s, v0.s[0]\n"
529 "add x12, x12, #0x10\n"
530 "fmla v25.4s, v12.4s, v1.s[0]\n"
531 "fmla v26.4s, v12.4s, v2.s[0]\n"
532 "cbnz x28, 57b\n"
533 "58:" // Height 3: Multiply loop: No odd multiplies
534 "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n"
535 "add x9, x9, #0x1\n"
536 "cmp x9, x19\n"
537 "bne 51b\n"
538 "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
539 "prfm pstl1keep, [x10, #0x0]\n"
540 "add x26, x10, x19, LSL #2\n"
541 "prfm pstl1keep, [x26, #0x0]\n"
542 "add x25, x26, x19, LSL #2\n"
543 "prfm pstl1keep, [x25, #0x0]\n"
544 "tbz %x[flags], #1, 59f\n"
545 "add x19, %x[args_ptr], %[offset_min]\n"
546 "ld1r { v17.4s }, [x19]\n"
547 "add x19, %x[args_ptr], %[offset_max]\n"
548 "ld1r { v16.4s }, [x19]\n"
549 "fmin v24.4s, v24.4s, v16.4s\n"
550 "fmin v25.4s, v25.4s, v16.4s\n"
551 "fmin v26.4s, v26.4s, v16.4s\n"
552 "fmax v24.4s, v24.4s, v17.4s\n"
553 "fmax v25.4s, v25.4s, v17.4s\n"
554 "fmax v26.4s, v26.4s, v17.4s\n"
555 "59:" // Height 3: No activation
556 "cmp x13, #0x4\n"
557 "bge 62f\n"
558 "tbz x13, #1, 60f\n"
559 "str d24, [x10], #0x8\n"
560 "str d25, [x26], #0x8\n"
561 "str d26, [x25], #0x8\n"
562 "tbz x13, #0, 61f\n"
563 "st1 { v24.s }[2], [x10]\n"
564 "st1 { v25.s }[2], [x26]\n"
565 "st1 { v26.s }[2], [x25]\n"
566 "b 61f\n"
567 "60:" // Height 3: Partial direct writeback: partial_1_0
568 "str s24, [x10, #0x0]\n"
569 "str s25, [x26, #0x0]\n"
570 "str s26, [x25, #0x0]\n"
571 "61:" // Height 3: Partial direct writeback: Done
572 "b 63f\n"
573 "62:" // Height 3: Full writeback
574 "str q24, [x10, #0x0]\n"
575 "add x10, x10, #0x10\n"
576 "str q25, [x26, #0x0]\n"
577 "str q26, [x25, #0x0]\n"
578 "63:" // Height 3: Writeback done
579 "subs x13, x13, #0x4\n"
580 "bgt 44b\n"
581 "b 170f\n"
582 "64:" // Height 4
583 "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
584 "mov x11, %x[bias]\n"
585 "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
586 "mov x10, %x[output_ptr]\n"
587 "65:" // Height 4: Column loop
588 "cbz x11, 66f\n"
589 "ldr q24, [x11, #0x0]\n"
590 "mov v25.16b, v24.16b\n"
591 "add x11, x11, #0x10\n"
592 "mov v26.16b, v24.16b\n"
593 "mov v27.16b, v24.16b\n"
594 "b 71f\n"
595 "66:" // Height 4: no bias
596 "tbz %x[flags], #0, 70f\n"
597 "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
598 "cmp x13, #0x4\n"
599 "add x26, x10, x19, LSL #2\n"
600 "add x25, x26, x19, LSL #2\n"
601 "add x24, x25, x19, LSL #2\n"
602 "bge 69f\n"
603 "tbz x13, #1, 67f\n"
604 "ldr d24, [x10], #0x8\n"
605 "ldr d25, [x26], #0x8\n"
606 "mov x19, #0x8\n"
607 "ldr d26, [x25], #0x8\n"
608 "ldr d27, [x24], #0x8\n"
609 "tbz x13, #0, 68f\n"
610 "ld1 { v24.s }[2], [x10]\n"
611 "ld1 { v25.s }[2], [x26]\n"
612 "ld1 { v26.s }[2], [x25]\n"
613 "ld1 { v27.s }[2], [x24]\n"
614 "b 68f\n"
615 "67:" // Height 4: Partial accumulate: partial_1_0
616 "ldr s24, [x10, #0x0]\n"
617 "mov x19, #0x0\n"
618 "ldr s25, [x26, #0x0]\n"
619 "ldr s26, [x25, #0x0]\n"
620 "ldr s27, [x24, #0x0]\n"
621 "68:" // Height 4: Partial accumulate: Done
622 "sub x10, x10, x19\n"
623 "b 71f\n"
624 "69:" // Height 4: full accumulate
625 "ldr q24, [x10, #0x0]\n"
626 "ldr q25, [x26, #0x0]\n"
627 "ldr q26, [x25, #0x0]\n"
628 "ldr q27, [x24, #0x0]\n"
629 "b 71f\n"
630 "70:" // Height 4: no accumulate
631 "movi v24.16b, #0x0\n"
632 "movi v25.16b, #0x0\n"
633 "movi v26.16b, #0x0\n"
634 "movi v27.16b, #0x0\n"
635 "71:" // Height 4: setup done
636 "mov x9, #0x0\n"
637 "72:" // Height 4: String loop
638 "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
639 "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n"
640 "ldr w28, [x20, x9, LSL #0x2]\n"
641 "tbz %x[flags], #3, 73f\n"
642 "ldr x20, [%x[input_ptr], x9, LSL #0x3]\n"
643 "add x20, x20, x19, LSL #3\n"
644 "ldr x27, [x20, #0x0]\n"
645 "ldr x26, [x20, #0x8]\n"
646 "ldr x25, [x20, #0x10]\n"
647 "ldr x24, [x20, #0x18]\n"
648 "cbnz x9, 74f\n"
649 "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
650 "add x27, x27, x19, LSL #2\n"
651 "add x26, x26, x19, LSL #2\n"
652 "add x25, x25, x19, LSL #2\n"
653 "add x24, x24, x19, LSL #2\n"
654 "b 74f\n"
655 "73:" // Height 4: setup direct input
656 "mov x27, %x[input_ptr]\n"
657 "add x26, x27, x19, LSL #2\n"
658 "add x25, x26, x19, LSL #2\n"
659 "add x24, x25, x19, LSL #2\n"
660 "74:" // Height 4: input setup done
661 "cmp x28, #0x4\n"
662 "blt 77f\n"
663 "ldr q0, [x27, #0x0]\n"
664 "ldr q1, [x26, #0x0]\n"
665 "cmp x28, #0x8\n"
666 "ldr q2, [x25, #0x0]\n"
667 "ldr q3, [x24, #0x0]\n"
668 "ldr q8, [x12, #0x0]\n"
669 "blt 76f\n"
670 "75:" // Height 4: Multiply loop: Main loop head
671 "fmla v24.4s, v8.4s, v0.s[0]\n"
672 "ldr q9, [x12, #0x10]\n"
673 "add x27, x27, #0x10\n"
674 "fmla v25.4s, v8.4s, v1.s[0]\n"
675 "ldr q10, [x12, #0x20]\n"
676 "add x26, x26, #0x10\n"
677 "fmla v26.4s, v8.4s, v2.s[0]\n"
678 "ldr q11, [x12, #0x30]\n"
679 "add x25, x25, #0x10\n"
680 "fmla v27.4s, v8.4s, v3.s[0]\n"
681 "prfm pldl1keep, [x27, #0x80]\n"
682 "add x24, x24, #0x10\n"
683 "fmla v24.4s, v9.4s, v0.s[1]\n"
684 "prfm pldl1keep, [x26, #0x80]\n"
685 "sub x28, x28, #0x4\n"
686 "fmla v25.4s, v9.4s, v1.s[1]\n"
687 "prfm pldl1keep, [x25, #0x80]\n"
688 "cmp x28, #0x8\n"
689 "fmla v26.4s, v9.4s, v2.s[1]\n"
690 "prfm pldl1keep, [x24, #0x80]\n"
691 "add x12, x12, #0x40\n"
692 "fmla v27.4s, v9.4s, v3.s[1]\n"
693 "ldr q8, [x12, #0x0]\n"
694 "fmla v24.4s, v10.4s, v0.s[2]\n"
695 "fmla v25.4s, v10.4s, v1.s[2]\n"
696 "fmla v26.4s, v10.4s, v2.s[2]\n"
697 "fmla v27.4s, v10.4s, v3.s[2]\n"
698 "fmla v24.4s, v11.4s, v0.s[3]\n"
699 "ldr q0, [x27, #0x0]\n"
700 "fmla v25.4s, v11.4s, v1.s[3]\n"
701 "ldr q1, [x26, #0x0]\n"
702 "fmla v26.4s, v11.4s, v2.s[3]\n"
703 "ldr q2, [x25, #0x0]\n"
704 "fmla v27.4s, v11.4s, v3.s[3]\n"
705 "ldr q3, [x24, #0x0]\n"
706 "bge 75b\n"
707 "76:" // Height 4: Multiply loop: Single iteration only
708 "fmla v24.4s, v8.4s, v0.s[0]\n"
709 "ldr q9, [x12, #0x10]\n"
710 "sub x28, x28, #0x4\n"
711 "fmla v25.4s, v8.4s, v1.s[0]\n"
712 "ldr q10, [x12, #0x20]\n"
713 "add x27, x27, #0x10\n"
714 "fmla v26.4s, v8.4s, v2.s[0]\n"
715 "ldr q11, [x12, #0x30]\n"
716 "add x26, x26, #0x10\n"
717 "fmla v27.4s, v8.4s, v3.s[0]\n"
718 "prfm pldl1keep, [x27, #0x80]\n"
719 "add x25, x25, #0x10\n"
720 "fmla v24.4s, v9.4s, v0.s[1]\n"
721 "prfm pldl1keep, [x26, #0x80]\n"
722 "add x24, x24, #0x10\n"
723 "fmla v25.4s, v9.4s, v1.s[1]\n"
724 "prfm pldl1keep, [x25, #0x80]\n"
725 "add x12, x12, #0x40\n"
726 "fmla v26.4s, v9.4s, v2.s[1]\n"
727 "prfm pldl1keep, [x24, #0x80]\n"
728 "fmla v27.4s, v9.4s, v3.s[1]\n"
729 "fmla v24.4s, v10.4s, v0.s[2]\n"
730 "fmla v25.4s, v10.4s, v1.s[2]\n"
731 "fmla v26.4s, v10.4s, v2.s[2]\n"
732 "fmla v27.4s, v10.4s, v3.s[2]\n"
733 "fmla v24.4s, v11.4s, v0.s[3]\n"
734 "fmla v25.4s, v11.4s, v1.s[3]\n"
735 "fmla v26.4s, v11.4s, v2.s[3]\n"
736 "fmla v27.4s, v11.4s, v3.s[3]\n"
737 "77:" // Height 4: Multiply loop: Main loop skip
738 "cbz x28, 79f\n"
739 "78:" // Height 4: Multiply loop: Odd block loop
740 "ldr s0, [x27], #0x4\n"
741 "sub x28, x28, #0x1\n"
742 "ldr s1, [x26], #0x4\n"
743 "ldr s2, [x25], #0x4\n"
744 "ldr s3, [x24], #0x4\n"
745 "ldr q12, [x12, #0x0]\n"
746 "fmla v24.4s, v12.4s, v0.s[0]\n"
747 "add x12, x12, #0x10\n"
748 "fmla v25.4s, v12.4s, v1.s[0]\n"
749 "fmla v26.4s, v12.4s, v2.s[0]\n"
750 "fmla v27.4s, v12.4s, v3.s[0]\n"
751 "cbnz x28, 78b\n"
752 "79:" // Height 4: Multiply loop: No odd multiplies
753 "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n"
754 "add x9, x9, #0x1\n"
755 "cmp x9, x19\n"
756 "bne 72b\n"
757 "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
758 "prfm pstl1keep, [x10, #0x0]\n"
759 "add x26, x10, x19, LSL #2\n"
760 "prfm pstl1keep, [x26, #0x0]\n"
761 "add x25, x26, x19, LSL #2\n"
762 "prfm pstl1keep, [x25, #0x0]\n"
763 "add x24, x25, x19, LSL #2\n"
764 "prfm pstl1keep, [x24, #0x0]\n"
765 "tbz %x[flags], #1, 80f\n"
766 "add x19, %x[args_ptr], %[offset_min]\n"
767 "ld1r { v17.4s }, [x19]\n"
768 "add x19, %x[args_ptr], %[offset_max]\n"
769 "ld1r { v16.4s }, [x19]\n"
770 "fmin v24.4s, v24.4s, v16.4s\n"
771 "fmin v25.4s, v25.4s, v16.4s\n"
772 "fmin v26.4s, v26.4s, v16.4s\n"
773 "fmin v27.4s, v27.4s, v16.4s\n"
774 "fmax v24.4s, v24.4s, v17.4s\n"
775 "fmax v25.4s, v25.4s, v17.4s\n"
776 "fmax v26.4s, v26.4s, v17.4s\n"
777 "fmax v27.4s, v27.4s, v17.4s\n"
778 "80:" // Height 4: No activation
779 "cmp x13, #0x4\n"
780 "bge 83f\n"
781 "tbz x13, #1, 81f\n"
782 "str d24, [x10], #0x8\n"
783 "str d25, [x26], #0x8\n"
784 "str d26, [x25], #0x8\n"
785 "str d27, [x24], #0x8\n"
786 "tbz x13, #0, 82f\n"
787 "st1 { v24.s }[2], [x10]\n"
788 "st1 { v25.s }[2], [x26]\n"
789 "st1 { v26.s }[2], [x25]\n"
790 "st1 { v27.s }[2], [x24]\n"
791 "b 82f\n"
792 "81:" // Height 4: Partial direct writeback: partial_1_0
793 "str s24, [x10, #0x0]\n"
794 "str s25, [x26, #0x0]\n"
795 "str s26, [x25, #0x0]\n"
796 "str s27, [x24, #0x0]\n"
797 "82:" // Height 4: Partial direct writeback: Done
798 "b 84f\n"
799 "83:" // Height 4: Full writeback
800 "str q24, [x10, #0x0]\n"
801 "add x10, x10, #0x10\n"
802 "str q25, [x26, #0x0]\n"
803 "str q26, [x25, #0x0]\n"
804 "str q27, [x24, #0x0]\n"
805 "84:" // Height 4: Writeback done
806 "subs x13, x13, #0x4\n"
807 "bgt 65b\n"
808 "b 170f\n"
809 "85:" // Height 5
810 "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
811 "mov x11, %x[bias]\n"
812 "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
813 "mov x10, %x[output_ptr]\n"
814 "86:" // Height 5: Column loop
815 "cbz x11, 87f\n"
816 "ldr q24, [x11, #0x0]\n"
817 "mov v25.16b, v24.16b\n"
818 "add x11, x11, #0x10\n"
819 "mov v26.16b, v24.16b\n"
820 "mov v27.16b, v24.16b\n"
821 "mov v28.16b, v24.16b\n"
822 "b 92f\n"
823 "87:" // Height 5: no bias
824 "tbz %x[flags], #0, 91f\n"
825 "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
826 "cmp x13, #0x4\n"
827 "add x26, x10, x19, LSL #2\n"
828 "add x25, x26, x19, LSL #2\n"
829 "add x24, x25, x19, LSL #2\n"
830 "add x23, x24, x19, LSL #2\n"
831 "bge 90f\n"
832 "tbz x13, #1, 88f\n"
833 "ldr d24, [x10], #0x8\n"
834 "ldr d25, [x26], #0x8\n"
835 "mov x19, #0x8\n"
836 "ldr d26, [x25], #0x8\n"
837 "ldr d27, [x24], #0x8\n"
838 "ldr d28, [x23], #0x8\n"
839 "tbz x13, #0, 89f\n"
840 "ld1 { v24.s }[2], [x10]\n"
841 "ld1 { v25.s }[2], [x26]\n"
842 "ld1 { v26.s }[2], [x25]\n"
843 "ld1 { v27.s }[2], [x24]\n"
844 "ld1 { v28.s }[2], [x23]\n"
845 "b 89f\n"
846 "88:" // Height 5: Partial accumulate: partial_1_0
847 "ldr s24, [x10, #0x0]\n"
848 "mov x19, #0x0\n"
849 "ldr s25, [x26, #0x0]\n"
850 "ldr s26, [x25, #0x0]\n"
851 "ldr s27, [x24, #0x0]\n"
852 "ldr s28, [x23, #0x0]\n"
853 "89:" // Height 5: Partial accumulate: Done
854 "sub x10, x10, x19\n"
855 "b 92f\n"
856 "90:" // Height 5: full accumulate
857 "ldr q24, [x10, #0x0]\n"
858 "ldr q25, [x26, #0x0]\n"
859 "ldr q26, [x25, #0x0]\n"
860 "ldr q27, [x24, #0x0]\n"
861 "ldr q28, [x23, #0x0]\n"
862 "b 92f\n"
863 "91:" // Height 5: no accumulate
864 "movi v24.16b, #0x0\n"
865 "movi v25.16b, #0x0\n"
866 "movi v26.16b, #0x0\n"
867 "movi v27.16b, #0x0\n"
868 "movi v28.16b, #0x0\n"
869 "92:" // Height 5: setup done
870 "mov x9, #0x0\n"
871 "93:" // Height 5: String loop
872 "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
873 "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n"
874 "ldr w28, [x20, x9, LSL #0x2]\n"
875 "tbz %x[flags], #3, 94f\n"
876 "ldr x20, [%x[input_ptr], x9, LSL #0x3]\n"
877 "add x20, x20, x19, LSL #3\n"
878 "ldr x27, [x20, #0x0]\n"
879 "ldr x26, [x20, #0x8]\n"
880 "ldr x25, [x20, #0x10]\n"
881 "ldr x24, [x20, #0x18]\n"
882 "ldr x23, [x20, #0x20]\n"
883 "cbnz x9, 95f\n"
884 "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
885 "add x27, x27, x19, LSL #2\n"
886 "add x26, x26, x19, LSL #2\n"
887 "add x25, x25, x19, LSL #2\n"
888 "add x24, x24, x19, LSL #2\n"
889 "add x23, x23, x19, LSL #2\n"
890 "b 95f\n"
891 "94:" // Height 5: setup direct input
892 "mov x27, %x[input_ptr]\n"
893 "add x26, x27, x19, LSL #2\n"
894 "add x25, x26, x19, LSL #2\n"
895 "add x24, x25, x19, LSL #2\n"
896 "add x23, x24, x19, LSL #2\n"
897 "95:" // Height 5: input setup done
898 "cmp x28, #0x4\n"
899 "blt 98f\n"
900 "ldr q0, [x27, #0x0]\n"
901 "ldr q1, [x26, #0x0]\n"
902 "cmp x28, #0x8\n"
903 "ldr q2, [x25, #0x0]\n"
904 "ldr q3, [x24, #0x0]\n"
905 "ldr q4, [x23, #0x0]\n"
906 "ldr q8, [x12, #0x0]\n"
907 "blt 97f\n"
908 "96:" // Height 5: Multiply loop: Main loop head
909 "fmla v24.4s, v8.4s, v0.s[0]\n"
910 "ldr q9, [x12, #0x10]\n"
911 "add x27, x27, #0x10\n"
912 "fmla v25.4s, v8.4s, v1.s[0]\n"
913 "ldr q10, [x12, #0x20]\n"
914 "add x26, x26, #0x10\n"
915 "fmla v26.4s, v8.4s, v2.s[0]\n"
916 "ldr q11, [x12, #0x30]\n"
917 "add x25, x25, #0x10\n"
918 "fmla v27.4s, v8.4s, v3.s[0]\n"
919 "prfm pldl1keep, [x27, #0x80]\n"
920 "add x24, x24, #0x10\n"
921 "fmla v28.4s, v8.4s, v4.s[0]\n"
922 "prfm pldl1keep, [x26, #0x80]\n"
923 "add x23, x23, #0x10\n"
924 "fmla v24.4s, v9.4s, v0.s[1]\n"
925 "prfm pldl1keep, [x25, #0x80]\n"
926 "sub x28, x28, #0x4\n"
927 "fmla v25.4s, v9.4s, v1.s[1]\n"
928 "prfm pldl1keep, [x24, #0x80]\n"
929 "cmp x28, #0x8\n"
930 "fmla v26.4s, v9.4s, v2.s[1]\n"
931 "prfm pldl1keep, [x23, #0x80]\n"
932 "add x12, x12, #0x40\n"
933 "fmla v27.4s, v9.4s, v3.s[1]\n"
934 "ldr q8, [x12, #0x0]\n"
935 "fmla v28.4s, v9.4s, v4.s[1]\n"
936 "fmla v24.4s, v10.4s, v0.s[2]\n"
937 "fmla v25.4s, v10.4s, v1.s[2]\n"
938 "fmla v26.4s, v10.4s, v2.s[2]\n"
939 "fmla v27.4s, v10.4s, v3.s[2]\n"
940 "fmla v28.4s, v10.4s, v4.s[2]\n"
941 "fmla v24.4s, v11.4s, v0.s[3]\n"
942 "ldr q0, [x27, #0x0]\n"
943 "fmla v25.4s, v11.4s, v1.s[3]\n"
944 "ldr q1, [x26, #0x0]\n"
945 "fmla v26.4s, v11.4s, v2.s[3]\n"
946 "ldr q2, [x25, #0x0]\n"
947 "fmla v27.4s, v11.4s, v3.s[3]\n"
948 "ldr q3, [x24, #0x0]\n"
949 "fmla v28.4s, v11.4s, v4.s[3]\n"
950 "ldr q4, [x23, #0x0]\n"
951 "bge 96b\n"
952 "97:" // Height 5: Multiply loop: Single iteration only
953 "fmla v24.4s, v8.4s, v0.s[0]\n"
954 "ldr q9, [x12, #0x10]\n"
955 "sub x28, x28, #0x4\n"
956 "fmla v25.4s, v8.4s, v1.s[0]\n"
957 "ldr q10, [x12, #0x20]\n"
958 "add x27, x27, #0x10\n"
959 "fmla v26.4s, v8.4s, v2.s[0]\n"
960 "ldr q11, [x12, #0x30]\n"
961 "add x26, x26, #0x10\n"
962 "fmla v27.4s, v8.4s, v3.s[0]\n"
963 "prfm pldl1keep, [x27, #0x80]\n"
964 "add x25, x25, #0x10\n"
965 "fmla v28.4s, v8.4s, v4.s[0]\n"
966 "prfm pldl1keep, [x26, #0x80]\n"
967 "add x24, x24, #0x10\n"
968 "fmla v24.4s, v9.4s, v0.s[1]\n"
969 "prfm pldl1keep, [x25, #0x80]\n"
970 "add x23, x23, #0x10\n"
971 "fmla v25.4s, v9.4s, v1.s[1]\n"
972 "prfm pldl1keep, [x24, #0x80]\n"
973 "add x12, x12, #0x40\n"
974 "fmla v26.4s, v9.4s, v2.s[1]\n"
975 "prfm pldl1keep, [x23, #0x80]\n"
976 "fmla v27.4s, v9.4s, v3.s[1]\n"
977 "fmla v28.4s, v9.4s, v4.s[1]\n"
978 "fmla v24.4s, v10.4s, v0.s[2]\n"
979 "fmla v25.4s, v10.4s, v1.s[2]\n"
980 "fmla v26.4s, v10.4s, v2.s[2]\n"
981 "fmla v27.4s, v10.4s, v3.s[2]\n"
982 "fmla v28.4s, v10.4s, v4.s[2]\n"
983 "fmla v24.4s, v11.4s, v0.s[3]\n"
984 "fmla v25.4s, v11.4s, v1.s[3]\n"
985 "fmla v26.4s, v11.4s, v2.s[3]\n"
986 "fmla v27.4s, v11.4s, v3.s[3]\n"
987 "fmla v28.4s, v11.4s, v4.s[3]\n"
988 "98:" // Height 5: Multiply loop: Main loop skip
989 "cbz x28, 100f\n"
990 "99:" // Height 5: Multiply loop: Odd block loop
991 "ldr s0, [x27], #0x4\n"
992 "sub x28, x28, #0x1\n"
993 "ldr s1, [x26], #0x4\n"
994 "ldr s2, [x25], #0x4\n"
995 "ldr s3, [x24], #0x4\n"
996 "ldr s4, [x23], #0x4\n"
997 "ldr q12, [x12, #0x0]\n"
998 "fmla v24.4s, v12.4s, v0.s[0]\n"
999 "add x12, x12, #0x10\n"
1000 "fmla v25.4s, v12.4s, v1.s[0]\n"
1001 "fmla v26.4s, v12.4s, v2.s[0]\n"
1002 "fmla v27.4s, v12.4s, v3.s[0]\n"
1003 "fmla v28.4s, v12.4s, v4.s[0]\n"
1004 "cbnz x28, 99b\n"
1005 "100:" // Height 5: Multiply loop: No odd multiplies
1006 "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n"
1007 "add x9, x9, #0x1\n"
1008 "cmp x9, x19\n"
1009 "bne 93b\n"
1010 "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
1011 "prfm pstl1keep, [x10, #0x0]\n"
1012 "add x26, x10, x19, LSL #2\n"
1013 "prfm pstl1keep, [x26, #0x0]\n"
1014 "add x25, x26, x19, LSL #2\n"
1015 "prfm pstl1keep, [x25, #0x0]\n"
1016 "add x24, x25, x19, LSL #2\n"
1017 "prfm pstl1keep, [x24, #0x0]\n"
1018 "add x23, x24, x19, LSL #2\n"
1019 "prfm pstl1keep, [x23, #0x0]\n"
1020 "tbz %x[flags], #1, 101f\n"
1021 "add x19, %x[args_ptr], %[offset_min]\n"
1022 "ld1r { v17.4s }, [x19]\n"
1023 "add x19, %x[args_ptr], %[offset_max]\n"
1024 "ld1r { v16.4s }, [x19]\n"
1025 "fmin v24.4s, v24.4s, v16.4s\n"
1026 "fmin v25.4s, v25.4s, v16.4s\n"
1027 "fmin v26.4s, v26.4s, v16.4s\n"
1028 "fmin v27.4s, v27.4s, v16.4s\n"
1029 "fmax v24.4s, v24.4s, v17.4s\n"
1030 "fmax v25.4s, v25.4s, v17.4s\n"
1031 "fmax v26.4s, v26.4s, v17.4s\n"
1032 "fmax v27.4s, v27.4s, v17.4s\n"
1033 "fmin v28.4s, v28.4s, v16.4s\n"
1034 "fmax v28.4s, v28.4s, v17.4s\n"
1035 "101:" // Height 5: No activation
1036 "cmp x13, #0x4\n"
1037 "bge 104f\n"
1038 "tbz x13, #1, 102f\n"
1039 "str d24, [x10], #0x8\n"
1040 "str d25, [x26], #0x8\n"
1041 "str d26, [x25], #0x8\n"
1042 "str d27, [x24], #0x8\n"
1043 "str d28, [x23], #0x8\n"
1044 "tbz x13, #0, 103f\n"
1045 "st1 { v24.s }[2], [x10]\n"
1046 "st1 { v25.s }[2], [x26]\n"
1047 "st1 { v26.s }[2], [x25]\n"
1048 "st1 { v27.s }[2], [x24]\n"
1049 "st1 { v28.s }[2], [x23]\n"
1050 "b 103f\n"
1051 "102:" // Height 5: Partial direct writeback: partial_1_0
1052 "str s24, [x10, #0x0]\n"
1053 "str s25, [x26, #0x0]\n"
1054 "str s26, [x25, #0x0]\n"
1055 "str s27, [x24, #0x0]\n"
1056 "str s28, [x23, #0x0]\n"
1057 "103:" // Height 5: Partial direct writeback: Done
1058 "b 105f\n"
1059 "104:" // Height 5: Full writeback
1060 "str q24, [x10, #0x0]\n"
1061 "add x10, x10, #0x10\n"
1062 "str q25, [x26, #0x0]\n"
1063 "str q26, [x25, #0x0]\n"
1064 "str q27, [x24, #0x0]\n"
1065 "str q28, [x23, #0x0]\n"
1066 "105:" // Height 5: Writeback done
1067 "subs x13, x13, #0x4\n"
1068 "bgt 86b\n"
1069 "b 170f\n"
1070 "106:" // Height 6
1071 "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
1072 "mov x11, %x[bias]\n"
1073 "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
1074 "mov x10, %x[output_ptr]\n"
1075 "107:" // Height 6: Column loop
1076 "cbz x11, 108f\n"
1077 "ldr q24, [x11, #0x0]\n"
1078 "mov v25.16b, v24.16b\n"
1079 "add x11, x11, #0x10\n"
1080 "mov v26.16b, v24.16b\n"
1081 "mov v27.16b, v24.16b\n"
1082 "mov v28.16b, v24.16b\n"
1083 "mov v29.16b, v24.16b\n"
1084 "b 113f\n"
1085 "108:" // Height 6: no bias
1086 "tbz %x[flags], #0, 112f\n"
1087 "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
1088 "cmp x13, #0x4\n"
1089 "add x26, x10, x19, LSL #2\n"
1090 "add x25, x26, x19, LSL #2\n"
1091 "add x24, x25, x19, LSL #2\n"
1092 "add x23, x24, x19, LSL #2\n"
1093 "add x22, x23, x19, LSL #2\n"
1094 "bge 111f\n"
1095 "tbz x13, #1, 109f\n"
1096 "ldr d24, [x10], #0x8\n"
1097 "ldr d25, [x26], #0x8\n"
1098 "mov x19, #0x8\n"
1099 "ldr d26, [x25], #0x8\n"
1100 "ldr d27, [x24], #0x8\n"
1101 "ldr d28, [x23], #0x8\n"
1102 "ldr d29, [x22], #0x8\n"
1103 "tbz x13, #0, 110f\n"
1104 "ld1 { v24.s }[2], [x10]\n"
1105 "ld1 { v25.s }[2], [x26]\n"
1106 "ld1 { v26.s }[2], [x25]\n"
1107 "ld1 { v27.s }[2], [x24]\n"
1108 "ld1 { v28.s }[2], [x23]\n"
1109 "ld1 { v29.s }[2], [x22]\n"
1110 "b 110f\n"
1111 "109:" // Height 6: Partial accumulate: partial_1_0
1112 "ldr s24, [x10, #0x0]\n"
1113 "mov x19, #0x0\n"
1114 "ldr s25, [x26, #0x0]\n"
1115 "ldr s26, [x25, #0x0]\n"
1116 "ldr s27, [x24, #0x0]\n"
1117 "ldr s28, [x23, #0x0]\n"
1118 "ldr s29, [x22, #0x0]\n"
1119 "110:" // Height 6: Partial accumulate: Done
1120 "sub x10, x10, x19\n"
1121 "b 113f\n"
1122 "111:" // Height 6: full accumulate
1123 "ldr q24, [x10, #0x0]\n"
1124 "ldr q25, [x26, #0x0]\n"
1125 "ldr q26, [x25, #0x0]\n"
1126 "ldr q27, [x24, #0x0]\n"
1127 "ldr q28, [x23, #0x0]\n"
1128 "ldr q29, [x22, #0x0]\n"
1129 "b 113f\n"
1130 "112:" // Height 6: no accumulate
1131 "movi v24.16b, #0x0\n"
1132 "movi v25.16b, #0x0\n"
1133 "movi v26.16b, #0x0\n"
1134 "movi v27.16b, #0x0\n"
1135 "movi v28.16b, #0x0\n"
1136 "movi v29.16b, #0x0\n"
1137 "113:" // Height 6: setup done
1138 "mov x9, #0x0\n"
1139 "114:" // Height 6: String loop
1140 "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
1141 "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n"
1142 "ldr w28, [x20, x9, LSL #0x2]\n"
1143 "tbz %x[flags], #3, 115f\n"
1144 "ldr x20, [%x[input_ptr], x9, LSL #0x3]\n"
1145 "add x20, x20, x19, LSL #3\n"
1146 "ldr x27, [x20, #0x0]\n"
1147 "ldr x26, [x20, #0x8]\n"
1148 "ldr x25, [x20, #0x10]\n"
1149 "ldr x24, [x20, #0x18]\n"
1150 "ldr x23, [x20, #0x20]\n"
1151 "ldr x22, [x20, #0x28]\n"
1152 "cbnz x9, 116f\n"
1153 "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
1154 "add x27, x27, x19, LSL #2\n"
1155 "add x26, x26, x19, LSL #2\n"
1156 "add x25, x25, x19, LSL #2\n"
1157 "add x24, x24, x19, LSL #2\n"
1158 "add x23, x23, x19, LSL #2\n"
1159 "add x22, x22, x19, LSL #2\n"
1160 "b 116f\n"
1161 "115:" // Height 6: setup direct input
1162 "mov x27, %x[input_ptr]\n"
1163 "add x26, x27, x19, LSL #2\n"
1164 "add x25, x26, x19, LSL #2\n"
1165 "add x24, x25, x19, LSL #2\n"
1166 "add x23, x24, x19, LSL #2\n"
1167 "add x22, x23, x19, LSL #2\n"
1168 "116:" // Height 6: input setup done
1169 "cmp x28, #0x4\n"
1170 "blt 119f\n"
1171 "ldr q0, [x27, #0x0]\n"
1172 "ldr q1, [x26, #0x0]\n"
1173 "cmp x28, #0x8\n"
1174 "ldr q2, [x25, #0x0]\n"
1175 "ldr q3, [x24, #0x0]\n"
1176 "ldr q4, [x23, #0x0]\n"
1177 "ldr q5, [x22, #0x0]\n"
1178 "ldr q8, [x12, #0x0]\n"
1179 "blt 118f\n"
1180 "117:" // Height 6: Multiply loop: Main loop head
1181 "fmla v24.4s, v8.4s, v0.s[0]\n"
1182 "ldr q9, [x12, #0x10]\n"
1183 "add x27, x27, #0x10\n"
1184 "fmla v25.4s, v8.4s, v1.s[0]\n"
1185 "ldr q10, [x12, #0x20]\n"
1186 "add x26, x26, #0x10\n"
1187 "fmla v26.4s, v8.4s, v2.s[0]\n"
1188 "ldr q11, [x12, #0x30]\n"
1189 "add x25, x25, #0x10\n"
1190 "fmla v27.4s, v8.4s, v3.s[0]\n"
1191 "prfm pldl1keep, [x27, #0x80]\n"
1192 "add x24, x24, #0x10\n"
1193 "fmla v28.4s, v8.4s, v4.s[0]\n"
1194 "prfm pldl1keep, [x26, #0x80]\n"
1195 "add x23, x23, #0x10\n"
1196 "fmla v29.4s, v8.4s, v5.s[0]\n"
1197 "prfm pldl1keep, [x25, #0x80]\n"
1198 "add x22, x22, #0x10\n"
1199 "fmla v24.4s, v9.4s, v0.s[1]\n"
1200 "prfm pldl1keep, [x24, #0x80]\n"
1201 "sub x28, x28, #0x4\n"
1202 "fmla v25.4s, v9.4s, v1.s[1]\n"
1203 "prfm pldl1keep, [x23, #0x80]\n"
1204 "cmp x28, #0x8\n"
1205 "fmla v26.4s, v9.4s, v2.s[1]\n"
1206 "prfm pldl1keep, [x22, #0x80]\n"
1207 "add x12, x12, #0x40\n"
1208 "fmla v27.4s, v9.4s, v3.s[1]\n"
1209 "ldr q8, [x12, #0x0]\n"
1210 "fmla v28.4s, v9.4s, v4.s[1]\n"
1211 "fmla v29.4s, v9.4s, v5.s[1]\n"
1212 "fmla v24.4s, v10.4s, v0.s[2]\n"
1213 "fmla v25.4s, v10.4s, v1.s[2]\n"
1214 "fmla v26.4s, v10.4s, v2.s[2]\n"
1215 "fmla v27.4s, v10.4s, v3.s[2]\n"
1216 "fmla v28.4s, v10.4s, v4.s[2]\n"
1217 "fmla v29.4s, v10.4s, v5.s[2]\n"
1218 "fmla v24.4s, v11.4s, v0.s[3]\n"
1219 "ldr q0, [x27, #0x0]\n"
1220 "fmla v25.4s, v11.4s, v1.s[3]\n"
1221 "ldr q1, [x26, #0x0]\n"
1222 "fmla v26.4s, v11.4s, v2.s[3]\n"
1223 "ldr q2, [x25, #0x0]\n"
1224 "fmla v27.4s, v11.4s, v3.s[3]\n"
1225 "ldr q3, [x24, #0x0]\n"
1226 "fmla v28.4s, v11.4s, v4.s[3]\n"
1227 "ldr q4, [x23, #0x0]\n"
1228 "fmla v29.4s, v11.4s, v5.s[3]\n"
1229 "ldr q5, [x22, #0x0]\n"
1230 "bge 117b\n"
1231 "118:" // Height 6: Multiply loop: Single iteration only
1232 "fmla v24.4s, v8.4s, v0.s[0]\n"
1233 "ldr q9, [x12, #0x10]\n"
1234 "sub x28, x28, #0x4\n"
1235 "fmla v25.4s, v8.4s, v1.s[0]\n"
1236 "ldr q10, [x12, #0x20]\n"
1237 "add x27, x27, #0x10\n"
1238 "fmla v26.4s, v8.4s, v2.s[0]\n"
1239 "ldr q11, [x12, #0x30]\n"
1240 "add x26, x26, #0x10\n"
1241 "fmla v27.4s, v8.4s, v3.s[0]\n"
1242 "prfm pldl1keep, [x27, #0x80]\n"
1243 "add x25, x25, #0x10\n"
1244 "fmla v28.4s, v8.4s, v4.s[0]\n"
1245 "prfm pldl1keep, [x26, #0x80]\n"
1246 "add x24, x24, #0x10\n"
1247 "fmla v29.4s, v8.4s, v5.s[0]\n"
1248 "prfm pldl1keep, [x25, #0x80]\n"
1249 "add x23, x23, #0x10\n"
1250 "fmla v24.4s, v9.4s, v0.s[1]\n"
1251 "prfm pldl1keep, [x24, #0x80]\n"
1252 "add x22, x22, #0x10\n"
1253 "fmla v25.4s, v9.4s, v1.s[1]\n"
1254 "prfm pldl1keep, [x23, #0x80]\n"
1255 "add x12, x12, #0x40\n"
1256 "fmla v26.4s, v9.4s, v2.s[1]\n"
1257 "prfm pldl1keep, [x22, #0x80]\n"
1258 "fmla v27.4s, v9.4s, v3.s[1]\n"
1259 "fmla v28.4s, v9.4s, v4.s[1]\n"
1260 "fmla v29.4s, v9.4s, v5.s[1]\n"
1261 "fmla v24.4s, v10.4s, v0.s[2]\n"
1262 "fmla v25.4s, v10.4s, v1.s[2]\n"
1263 "fmla v26.4s, v10.4s, v2.s[2]\n"
1264 "fmla v27.4s, v10.4s, v3.s[2]\n"
1265 "fmla v28.4s, v10.4s, v4.s[2]\n"
1266 "fmla v29.4s, v10.4s, v5.s[2]\n"
1267 "fmla v24.4s, v11.4s, v0.s[3]\n"
1268 "fmla v25.4s, v11.4s, v1.s[3]\n"
1269 "fmla v26.4s, v11.4s, v2.s[3]\n"
1270 "fmla v27.4s, v11.4s, v3.s[3]\n"
1271 "fmla v28.4s, v11.4s, v4.s[3]\n"
1272 "fmla v29.4s, v11.4s, v5.s[3]\n"
1273 "119:" // Height 6: Multiply loop: Main loop skip
1274 "cbz x28, 121f\n"
1275 "120:" // Height 6: Multiply loop: Odd block loop
1276 "ldr s0, [x27], #0x4\n"
1277 "sub x28, x28, #0x1\n"
1278 "ldr s1, [x26], #0x4\n"
1279 "ldr s2, [x25], #0x4\n"
1280 "ldr s3, [x24], #0x4\n"
1281 "ldr s4, [x23], #0x4\n"
1282 "ldr s5, [x22], #0x4\n"
1283 "ldr q12, [x12, #0x0]\n"
1284 "fmla v24.4s, v12.4s, v0.s[0]\n"
1285 "add x12, x12, #0x10\n"
1286 "fmla v25.4s, v12.4s, v1.s[0]\n"
1287 "fmla v26.4s, v12.4s, v2.s[0]\n"
1288 "fmla v27.4s, v12.4s, v3.s[0]\n"
1289 "fmla v28.4s, v12.4s, v4.s[0]\n"
1290 "fmla v29.4s, v12.4s, v5.s[0]\n"
1291 "cbnz x28, 120b\n"
1292 "121:" // Height 6: Multiply loop: No odd multiplies
1293 "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n"
1294 "add x9, x9, #0x1\n"
1295 "cmp x9, x19\n"
1296 "bne 114b\n"
1297 "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
1298 "prfm pstl1keep, [x10, #0x0]\n"
1299 "add x26, x10, x19, LSL #2\n"
1300 "prfm pstl1keep, [x26, #0x0]\n"
1301 "add x25, x26, x19, LSL #2\n"
1302 "prfm pstl1keep, [x25, #0x0]\n"
1303 "add x24, x25, x19, LSL #2\n"
1304 "prfm pstl1keep, [x24, #0x0]\n"
1305 "add x23, x24, x19, LSL #2\n"
1306 "prfm pstl1keep, [x23, #0x0]\n"
1307 "add x22, x23, x19, LSL #2\n"
1308 "prfm pstl1keep, [x22, #0x0]\n"
1309 "tbz %x[flags], #1, 122f\n"
1310 "add x19, %x[args_ptr], %[offset_min]\n"
1311 "ld1r { v17.4s }, [x19]\n"
1312 "add x19, %x[args_ptr], %[offset_max]\n"
1313 "ld1r { v16.4s }, [x19]\n"
1314 "fmin v24.4s, v24.4s, v16.4s\n"
1315 "fmin v25.4s, v25.4s, v16.4s\n"
1316 "fmin v26.4s, v26.4s, v16.4s\n"
1317 "fmin v27.4s, v27.4s, v16.4s\n"
1318 "fmax v24.4s, v24.4s, v17.4s\n"
1319 "fmax v25.4s, v25.4s, v17.4s\n"
1320 "fmax v26.4s, v26.4s, v17.4s\n"
1321 "fmax v27.4s, v27.4s, v17.4s\n"
1322 "fmin v28.4s, v28.4s, v16.4s\n"
1323 "fmin v29.4s, v29.4s, v16.4s\n"
1324 "fmax v28.4s, v28.4s, v17.4s\n"
1325 "fmax v29.4s, v29.4s, v17.4s\n"
1326 "122:" // Height 6: No activation
1327 "cmp x13, #0x4\n"
1328 "bge 125f\n"
1329 "tbz x13, #1, 123f\n"
1330 "str d24, [x10], #0x8\n"
1331 "str d25, [x26], #0x8\n"
1332 "str d26, [x25], #0x8\n"
1333 "str d27, [x24], #0x8\n"
1334 "str d28, [x23], #0x8\n"
1335 "str d29, [x22], #0x8\n"
1336 "tbz x13, #0, 124f\n"
1337 "st1 { v24.s }[2], [x10]\n"
1338 "st1 { v25.s }[2], [x26]\n"
1339 "st1 { v26.s }[2], [x25]\n"
1340 "st1 { v27.s }[2], [x24]\n"
1341 "st1 { v28.s }[2], [x23]\n"
1342 "st1 { v29.s }[2], [x22]\n"
1343 "b 124f\n"
1344 "123:" // Height 6: Partial direct writeback: partial_1_0
1345 "str s24, [x10, #0x0]\n"
1346 "str s25, [x26, #0x0]\n"
1347 "str s26, [x25, #0x0]\n"
1348 "str s27, [x24, #0x0]\n"
1349 "str s28, [x23, #0x0]\n"
1350 "str s29, [x22, #0x0]\n"
1351 "124:" // Height 6: Partial direct writeback: Done
1352 "b 126f\n"
1353 "125:" // Height 6: Full writeback
1354 "str q24, [x10, #0x0]\n"
1355 "add x10, x10, #0x10\n"
1356 "str q25, [x26, #0x0]\n"
1357 "str q26, [x25, #0x0]\n"
1358 "str q27, [x24, #0x0]\n"
1359 "str q28, [x23, #0x0]\n"
1360 "str q29, [x22, #0x0]\n"
1361 "126:" // Height 6: Writeback done
1362 "subs x13, x13, #0x4\n"
1363 "bgt 107b\n"
1364 "b 170f\n"
1365 "127:" // Height 7
1366 "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
1367 "mov x11, %x[bias]\n"
1368 "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
1369 "mov x10, %x[output_ptr]\n"
1370 "128:" // Height 7: Column loop
1371 "cbz x11, 129f\n"
1372 "ldr q24, [x11, #0x0]\n"
1373 "mov v25.16b, v24.16b\n"
1374 "add x11, x11, #0x10\n"
1375 "mov v26.16b, v24.16b\n"
1376 "mov v27.16b, v24.16b\n"
1377 "mov v28.16b, v24.16b\n"
1378 "mov v29.16b, v24.16b\n"
1379 "mov v30.16b, v24.16b\n"
1380 "b 134f\n"
1381 "129:" // Height 7: no bias
1382 "tbz %x[flags], #0, 133f\n"
1383 "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
1384 "cmp x13, #0x4\n"
1385 "add x26, x10, x19, LSL #2\n"
1386 "add x25, x26, x19, LSL #2\n"
1387 "add x24, x25, x19, LSL #2\n"
1388 "add x23, x24, x19, LSL #2\n"
1389 "add x22, x23, x19, LSL #2\n"
1390 "add x21, x22, x19, LSL #2\n"
1391 "bge 132f\n"
1392 "tbz x13, #1, 130f\n"
1393 "ldr d24, [x10], #0x8\n"
1394 "ldr d25, [x26], #0x8\n"
1395 "mov x19, #0x8\n"
1396 "ldr d26, [x25], #0x8\n"
1397 "ldr d27, [x24], #0x8\n"
1398 "ldr d28, [x23], #0x8\n"
1399 "ldr d29, [x22], #0x8\n"
1400 "ldr d30, [x21], #0x8\n"
1401 "tbz x13, #0, 131f\n"
1402 "ld1 { v24.s }[2], [x10]\n"
1403 "ld1 { v25.s }[2], [x26]\n"
1404 "ld1 { v26.s }[2], [x25]\n"
1405 "ld1 { v27.s }[2], [x24]\n"
1406 "ld1 { v28.s }[2], [x23]\n"
1407 "ld1 { v29.s }[2], [x22]\n"
1408 "ld1 { v30.s }[2], [x21]\n"
1409 "b 131f\n"
1410 "130:" // Height 7: Partial accumulate: partial_1_0
1411 "ldr s24, [x10, #0x0]\n"
1412 "mov x19, #0x0\n"
1413 "ldr s25, [x26, #0x0]\n"
1414 "ldr s26, [x25, #0x0]\n"
1415 "ldr s27, [x24, #0x0]\n"
1416 "ldr s28, [x23, #0x0]\n"
1417 "ldr s29, [x22, #0x0]\n"
1418 "ldr s30, [x21, #0x0]\n"
1419 "131:" // Height 7: Partial accumulate: Done
1420 "sub x10, x10, x19\n"
1421 "b 134f\n"
1422 "132:" // Height 7: full accumulate
1423 "ldr q24, [x10, #0x0]\n"
1424 "ldr q25, [x26, #0x0]\n"
1425 "ldr q26, [x25, #0x0]\n"
1426 "ldr q27, [x24, #0x0]\n"
1427 "ldr q28, [x23, #0x0]\n"
1428 "ldr q29, [x22, #0x0]\n"
1429 "ldr q30, [x21, #0x0]\n"
1430 "b 134f\n"
1431 "133:" // Height 7: no accumulate
1432 "movi v24.16b, #0x0\n"
1433 "movi v25.16b, #0x0\n"
1434 "movi v26.16b, #0x0\n"
1435 "movi v27.16b, #0x0\n"
1436 "movi v28.16b, #0x0\n"
1437 "movi v29.16b, #0x0\n"
1438 "movi v30.16b, #0x0\n"
1439 "134:" // Height 7: setup done
1440 "mov x9, #0x0\n"
1441 "135:" // Height 7: String loop
1442 "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
1443 "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n"
1444 "ldr w28, [x20, x9, LSL #0x2]\n"
1445 "tbz %x[flags], #3, 136f\n"
1446 "ldr x20, [%x[input_ptr], x9, LSL #0x3]\n"
1447 "add x20, x20, x19, LSL #3\n"
1448 "ldr x27, [x20, #0x0]\n"
1449 "ldr x26, [x20, #0x8]\n"
1450 "ldr x25, [x20, #0x10]\n"
1451 "ldr x24, [x20, #0x18]\n"
1452 "ldr x23, [x20, #0x20]\n"
1453 "ldr x22, [x20, #0x28]\n"
1454 "ldr x21, [x20, #0x30]\n"
1455 "cbnz x9, 137f\n"
1456 "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
1457 "add x27, x27, x19, LSL #2\n"
1458 "add x26, x26, x19, LSL #2\n"
1459 "add x25, x25, x19, LSL #2\n"
1460 "add x24, x24, x19, LSL #2\n"
1461 "add x23, x23, x19, LSL #2\n"
1462 "add x22, x22, x19, LSL #2\n"
1463 "add x21, x21, x19, LSL #2\n"
1464 "b 137f\n"
1465 "136:" // Height 7: setup direct input
1466 "mov x27, %x[input_ptr]\n"
1467 "add x26, x27, x19, LSL #2\n"
1468 "add x25, x26, x19, LSL #2\n"
1469 "add x24, x25, x19, LSL #2\n"
1470 "add x23, x24, x19, LSL #2\n"
1471 "add x22, x23, x19, LSL #2\n"
1472 "add x21, x22, x19, LSL #2\n"
1473 "137:" // Height 7: input setup done
1474 "cmp x28, #0x4\n"
1475 "blt 140f\n"
1476 "ldr q0, [x27, #0x0]\n"
1477 "ldr q1, [x26, #0x0]\n"
1478 "cmp x28, #0x8\n"
1479 "ldr q2, [x25, #0x0]\n"
1480 "ldr q3, [x24, #0x0]\n"
1481 "ldr q4, [x23, #0x0]\n"
1482 "ldr q5, [x22, #0x0]\n"
1483 "ldr q6, [x21, #0x0]\n"
1484 "ldr q8, [x12, #0x0]\n"
1485 "blt 139f\n"
1486 "138:" // Height 7: Multiply loop: Main loop head
1487 "fmla v24.4s, v8.4s, v0.s[0]\n"
1488 "ldr q9, [x12, #0x10]\n"
1489 "add x27, x27, #0x10\n"
1490 "fmla v25.4s, v8.4s, v1.s[0]\n"
1491 "ldr q10, [x12, #0x20]\n"
1492 "add x26, x26, #0x10\n"
1493 "fmla v26.4s, v8.4s, v2.s[0]\n"
1494 "ldr q11, [x12, #0x30]\n"
1495 "add x25, x25, #0x10\n"
1496 "fmla v27.4s, v8.4s, v3.s[0]\n"
1497 "prfm pldl1keep, [x27, #0x80]\n"
1498 "add x24, x24, #0x10\n"
1499 "fmla v28.4s, v8.4s, v4.s[0]\n"
1500 "prfm pldl1keep, [x26, #0x80]\n"
1501 "add x23, x23, #0x10\n"
1502 "fmla v29.4s, v8.4s, v5.s[0]\n"
1503 "prfm pldl1keep, [x25, #0x80]\n"
1504 "add x22, x22, #0x10\n"
1505 "fmla v30.4s, v8.4s, v6.s[0]\n"
1506 "prfm pldl1keep, [x24, #0x80]\n"
1507 "add x21, x21, #0x10\n"
1508 "fmla v24.4s, v9.4s, v0.s[1]\n"
1509 "prfm pldl1keep, [x23, #0x80]\n"
1510 "sub x28, x28, #0x4\n"
1511 "fmla v25.4s, v9.4s, v1.s[1]\n"
1512 "prfm pldl1keep, [x22, #0x80]\n"
1513 "cmp x28, #0x8\n"
1514 "fmla v26.4s, v9.4s, v2.s[1]\n"
1515 "prfm pldl1keep, [x21, #0x80]\n"
1516 "add x12, x12, #0x40\n"
1517 "fmla v27.4s, v9.4s, v3.s[1]\n"
1518 "ldr q8, [x12, #0x0]\n"
1519 "fmla v28.4s, v9.4s, v4.s[1]\n"
1520 "fmla v29.4s, v9.4s, v5.s[1]\n"
1521 "fmla v30.4s, v9.4s, v6.s[1]\n"
1522 "fmla v24.4s, v10.4s, v0.s[2]\n"
1523 "fmla v25.4s, v10.4s, v1.s[2]\n"
1524 "fmla v26.4s, v10.4s, v2.s[2]\n"
1525 "fmla v27.4s, v10.4s, v3.s[2]\n"
1526 "fmla v28.4s, v10.4s, v4.s[2]\n"
1527 "fmla v29.4s, v10.4s, v5.s[2]\n"
1528 "fmla v30.4s, v10.4s, v6.s[2]\n"
1529 "fmla v24.4s, v11.4s, v0.s[3]\n"
1530 "ldr q0, [x27, #0x0]\n"
1531 "fmla v25.4s, v11.4s, v1.s[3]\n"
1532 "ldr q1, [x26, #0x0]\n"
1533 "fmla v26.4s, v11.4s, v2.s[3]\n"
1534 "ldr q2, [x25, #0x0]\n"
1535 "fmla v27.4s, v11.4s, v3.s[3]\n"
1536 "ldr q3, [x24, #0x0]\n"
1537 "fmla v28.4s, v11.4s, v4.s[3]\n"
1538 "ldr q4, [x23, #0x0]\n"
1539 "fmla v29.4s, v11.4s, v5.s[3]\n"
1540 "ldr q5, [x22, #0x0]\n"
1541 "fmla v30.4s, v11.4s, v6.s[3]\n"
1542 "ldr q6, [x21, #0x0]\n"
1543 "bge 138b\n"
1544 "139:" // Height 7: Multiply loop: Single iteration only
1545 "fmla v24.4s, v8.4s, v0.s[0]\n"
1546 "ldr q9, [x12, #0x10]\n"
1547 "sub x28, x28, #0x4\n"
1548 "fmla v25.4s, v8.4s, v1.s[0]\n"
1549 "ldr q10, [x12, #0x20]\n"
1550 "add x27, x27, #0x10\n"
1551 "fmla v26.4s, v8.4s, v2.s[0]\n"
1552 "ldr q11, [x12, #0x30]\n"
1553 "add x26, x26, #0x10\n"
1554 "fmla v27.4s, v8.4s, v3.s[0]\n"
1555 "prfm pldl1keep, [x27, #0x80]\n"
1556 "add x25, x25, #0x10\n"
1557 "fmla v28.4s, v8.4s, v4.s[0]\n"
1558 "prfm pldl1keep, [x26, #0x80]\n"
1559 "add x24, x24, #0x10\n"
1560 "fmla v29.4s, v8.4s, v5.s[0]\n"
1561 "prfm pldl1keep, [x25, #0x80]\n"
1562 "add x23, x23, #0x10\n"
1563 "fmla v30.4s, v8.4s, v6.s[0]\n"
1564 "prfm pldl1keep, [x24, #0x80]\n"
1565 "add x22, x22, #0x10\n"
1566 "fmla v24.4s, v9.4s, v0.s[1]\n"
1567 "prfm pldl1keep, [x23, #0x80]\n"
1568 "add x21, x21, #0x10\n"
1569 "fmla v25.4s, v9.4s, v1.s[1]\n"
1570 "prfm pldl1keep, [x22, #0x80]\n"
1571 "add x12, x12, #0x40\n"
1572 "fmla v26.4s, v9.4s, v2.s[1]\n"
1573 "prfm pldl1keep, [x21, #0x80]\n"
1574 "fmla v27.4s, v9.4s, v3.s[1]\n"
1575 "fmla v28.4s, v9.4s, v4.s[1]\n"
1576 "fmla v29.4s, v9.4s, v5.s[1]\n"
1577 "fmla v30.4s, v9.4s, v6.s[1]\n"
1578 "fmla v24.4s, v10.4s, v0.s[2]\n"
1579 "fmla v25.4s, v10.4s, v1.s[2]\n"
1580 "fmla v26.4s, v10.4s, v2.s[2]\n"
1581 "fmla v27.4s, v10.4s, v3.s[2]\n"
1582 "fmla v28.4s, v10.4s, v4.s[2]\n"
1583 "fmla v29.4s, v10.4s, v5.s[2]\n"
1584 "fmla v30.4s, v10.4s, v6.s[2]\n"
1585 "fmla v24.4s, v11.4s, v0.s[3]\n"
1586 "fmla v25.4s, v11.4s, v1.s[3]\n"
1587 "fmla v26.4s, v11.4s, v2.s[3]\n"
1588 "fmla v27.4s, v11.4s, v3.s[3]\n"
1589 "fmla v28.4s, v11.4s, v4.s[3]\n"
1590 "fmla v29.4s, v11.4s, v5.s[3]\n"
1591 "fmla v30.4s, v11.4s, v6.s[3]\n"
1592 "140:" // Height 7: Multiply loop: Main loop skip
1593 "cbz x28, 142f\n"
1594 "141:" // Height 7: Multiply loop: Odd block loop
1595 "ldr s0, [x27], #0x4\n"
1596 "sub x28, x28, #0x1\n"
1597 "ldr s1, [x26], #0x4\n"
1598 "ldr s2, [x25], #0x4\n"
1599 "ldr s3, [x24], #0x4\n"
1600 "ldr s4, [x23], #0x4\n"
1601 "ldr s5, [x22], #0x4\n"
1602 "ldr s6, [x21], #0x4\n"
1603 "ldr q12, [x12, #0x0]\n"
1604 "fmla v24.4s, v12.4s, v0.s[0]\n"
1605 "add x12, x12, #0x10\n"
1606 "fmla v25.4s, v12.4s, v1.s[0]\n"
1607 "fmla v26.4s, v12.4s, v2.s[0]\n"
1608 "fmla v27.4s, v12.4s, v3.s[0]\n"
1609 "fmla v28.4s, v12.4s, v4.s[0]\n"
1610 "fmla v29.4s, v12.4s, v5.s[0]\n"
1611 "fmla v30.4s, v12.4s, v6.s[0]\n"
1612 "cbnz x28, 141b\n"
1613 "142:" // Height 7: Multiply loop: No odd multiplies
1614 "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n"
1615 "add x9, x9, #0x1\n"
1616 "cmp x9, x19\n"
1617 "bne 135b\n"
1618 "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
1619 "prfm pstl1keep, [x10, #0x0]\n"
1620 "add x26, x10, x19, LSL #2\n"
1621 "prfm pstl1keep, [x26, #0x0]\n"
1622 "add x25, x26, x19, LSL #2\n"
1623 "prfm pstl1keep, [x25, #0x0]\n"
1624 "add x24, x25, x19, LSL #2\n"
1625 "prfm pstl1keep, [x24, #0x0]\n"
1626 "add x23, x24, x19, LSL #2\n"
1627 "prfm pstl1keep, [x23, #0x0]\n"
1628 "add x22, x23, x19, LSL #2\n"
1629 "prfm pstl1keep, [x22, #0x0]\n"
1630 "add x21, x22, x19, LSL #2\n"
1631 "prfm pstl1keep, [x21, #0x0]\n"
1632 "tbz %x[flags], #1, 143f\n"
1633 "add x19, %x[args_ptr], %[offset_min]\n"
1634 "ld1r { v17.4s }, [x19]\n"
1635 "add x19, %x[args_ptr], %[offset_max]\n"
1636 "ld1r { v16.4s }, [x19]\n"
1637 "fmin v24.4s, v24.4s, v16.4s\n"
1638 "fmin v25.4s, v25.4s, v16.4s\n"
1639 "fmin v26.4s, v26.4s, v16.4s\n"
1640 "fmin v27.4s, v27.4s, v16.4s\n"
1641 "fmax v24.4s, v24.4s, v17.4s\n"
1642 "fmax v25.4s, v25.4s, v17.4s\n"
1643 "fmax v26.4s, v26.4s, v17.4s\n"
1644 "fmax v27.4s, v27.4s, v17.4s\n"
1645 "fmin v28.4s, v28.4s, v16.4s\n"
1646 "fmin v29.4s, v29.4s, v16.4s\n"
1647 "fmin v30.4s, v30.4s, v16.4s\n"
1648 "fmax v28.4s, v28.4s, v17.4s\n"
1649 "fmax v29.4s, v29.4s, v17.4s\n"
1650 "fmax v30.4s, v30.4s, v17.4s\n"
1651 "143:" // Height 7: No activation
1652 "cmp x13, #0x4\n"
1653 "bge 146f\n"
1654 "tbz x13, #1, 144f\n"
1655 "str d24, [x10], #0x8\n"
1656 "str d25, [x26], #0x8\n"
1657 "str d26, [x25], #0x8\n"
1658 "str d27, [x24], #0x8\n"
1659 "str d28, [x23], #0x8\n"
1660 "str d29, [x22], #0x8\n"
1661 "str d30, [x21], #0x8\n"
1662 "tbz x13, #0, 145f\n"
1663 "st1 { v24.s }[2], [x10]\n"
1664 "st1 { v25.s }[2], [x26]\n"
1665 "st1 { v26.s }[2], [x25]\n"
1666 "st1 { v27.s }[2], [x24]\n"
1667 "st1 { v28.s }[2], [x23]\n"
1668 "st1 { v29.s }[2], [x22]\n"
1669 "st1 { v30.s }[2], [x21]\n"
1670 "b 145f\n"
1671 "144:" // Height 7: Partial direct writeback: partial_1_0
1672 "str s24, [x10, #0x0]\n"
1673 "str s25, [x26, #0x0]\n"
1674 "str s26, [x25, #0x0]\n"
1675 "str s27, [x24, #0x0]\n"
1676 "str s28, [x23, #0x0]\n"
1677 "str s29, [x22, #0x0]\n"
1678 "str s30, [x21, #0x0]\n"
1679 "145:" // Height 7: Partial direct writeback: Done
1680 "b 147f\n"
1681 "146:" // Height 7: Full writeback
1682 "str q24, [x10, #0x0]\n"
1683 "add x10, x10, #0x10\n"
1684 "str q25, [x26, #0x0]\n"
1685 "str q26, [x25, #0x0]\n"
1686 "str q27, [x24, #0x0]\n"
1687 "str q28, [x23, #0x0]\n"
1688 "str q29, [x22, #0x0]\n"
1689 "str q30, [x21, #0x0]\n"
1690 "147:" // Height 7: Writeback done
1691 "subs x13, x13, #0x4\n"
1692 "bgt 128b\n"
1693 "b 170f\n"
1694 "148:" // Height 8
1695 "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
1696 "mov x11, %x[bias]\n"
1697 "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n"
1698 "mov x10, %x[output_ptr]\n"
1699 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
1700 "mov x19, #0x20\n"
1701 "madd %x[output_ptr], x20, x19, %x[output_ptr]\n"
1702 "149:" // Height 8: Column loop
1703 "cbz x11, 150f\n"
1704 "ldr q24, [x11, #0x0]\n"
1705 "mov v25.16b, v24.16b\n"
1706 "add x11, x11, #0x10\n"
1707 "mov v26.16b, v24.16b\n"
1708 "mov v27.16b, v24.16b\n"
1709 "mov v28.16b, v24.16b\n"
1710 "mov v29.16b, v24.16b\n"
1711 "mov v30.16b, v24.16b\n"
1712 "mov v31.16b, v24.16b\n"
1713 "b 155f\n"
1714 "150:" // Height 8: no bias
1715 "tbz %x[flags], #0, 154f\n"
1716 "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
1717 "cmp x13, #0x4\n"
1718 "add x26, x10, x19, LSL #2\n"
1719 "add x25, x26, x19, LSL #2\n"
1720 "add x24, x25, x19, LSL #2\n"
1721 "add x23, x24, x19, LSL #2\n"
1722 "add x22, x23, x19, LSL #2\n"
1723 "add x21, x22, x19, LSL #2\n"
1724 "add x20, x21, x19, LSL #2\n"
1725 "bge 153f\n"
1726 "tbz x13, #1, 151f\n"
1727 "ldr d24, [x10], #0x8\n"
1728 "ldr d25, [x26], #0x8\n"
1729 "mov x19, #0x8\n"
1730 "ldr d26, [x25], #0x8\n"
1731 "ldr d27, [x24], #0x8\n"
1732 "ldr d28, [x23], #0x8\n"
1733 "ldr d29, [x22], #0x8\n"
1734 "ldr d30, [x21], #0x8\n"
1735 "ldr d31, [x20], #0x8\n"
1736 "tbz x13, #0, 152f\n"
1737 "ld1 { v24.s }[2], [x10]\n"
1738 "ld1 { v25.s }[2], [x26]\n"
1739 "ld1 { v26.s }[2], [x25]\n"
1740 "ld1 { v27.s }[2], [x24]\n"
1741 "ld1 { v28.s }[2], [x23]\n"
1742 "ld1 { v29.s }[2], [x22]\n"
1743 "ld1 { v30.s }[2], [x21]\n"
1744 "ld1 { v31.s }[2], [x20]\n"
1745 "b 152f\n"
1746 "151:" // Height 8: Partial accumulate: partial_1_0
1747 "ldr s24, [x10, #0x0]\n"
1748 "mov x19, #0x0\n"
1749 "ldr s25, [x26, #0x0]\n"
1750 "ldr s26, [x25, #0x0]\n"
1751 "ldr s27, [x24, #0x0]\n"
1752 "ldr s28, [x23, #0x0]\n"
1753 "ldr s29, [x22, #0x0]\n"
1754 "ldr s30, [x21, #0x0]\n"
1755 "ldr s31, [x20, #0x0]\n"
1756 "152:" // Height 8: Partial accumulate: Done
1757 "sub x10, x10, x19\n"
1758 "b 155f\n"
1759 "153:" // Height 8: full accumulate
1760 "ldr q24, [x10, #0x0]\n"
1761 "ldr q25, [x26, #0x0]\n"
1762 "ldr q26, [x25, #0x0]\n"
1763 "ldr q27, [x24, #0x0]\n"
1764 "ldr q28, [x23, #0x0]\n"
1765 "ldr q29, [x22, #0x0]\n"
1766 "ldr q30, [x21, #0x0]\n"
1767 "ldr q31, [x20, #0x0]\n"
1768 "b 155f\n"
1769 "154:" // Height 8: no accumulate
1770 "movi v24.16b, #0x0\n"
1771 "movi v25.16b, #0x0\n"
1772 "movi v26.16b, #0x0\n"
1773 "movi v27.16b, #0x0\n"
1774 "movi v28.16b, #0x0\n"
1775 "movi v29.16b, #0x0\n"
1776 "movi v30.16b, #0x0\n"
1777 "movi v31.16b, #0x0\n"
1778 "155:" // Height 8: setup done
1779 "mov x9, #0x0\n"
1780 "156:" // Height 8: String loop
1781 "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
1782 "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n"
1783 "ldr w28, [x20, x9, LSL #0x2]\n"
1784 "tbz %x[flags], #3, 157f\n"
1785 "ldr x20, [%x[input_ptr], x9, LSL #0x3]\n"
1786 "add x20, x20, x19, LSL #3\n"
1787 "ldr x27, [x20, #0x0]\n"
1788 "ldr x26, [x20, #0x8]\n"
1789 "ldr x25, [x20, #0x10]\n"
1790 "ldr x24, [x20, #0x18]\n"
1791 "ldr x23, [x20, #0x20]\n"
1792 "ldr x22, [x20, #0x28]\n"
1793 "ldr x21, [x20, #0x30]\n"
1794 "ldr x20, [x20, #0x38]\n"
1795 "cbnz x9, 158f\n"
1796 "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
1797 "add x27, x27, x19, LSL #2\n"
1798 "add x26, x26, x19, LSL #2\n"
1799 "add x25, x25, x19, LSL #2\n"
1800 "add x24, x24, x19, LSL #2\n"
1801 "add x23, x23, x19, LSL #2\n"
1802 "add x22, x22, x19, LSL #2\n"
1803 "add x21, x21, x19, LSL #2\n"
1804 "add x20, x20, x19, LSL #2\n"
1805 "b 158f\n"
1806 "157:" // Height 8: setup direct input
1807 "mov x27, %x[input_ptr]\n"
1808 "add x26, x27, x19, LSL #2\n"
1809 "add x25, x26, x19, LSL #2\n"
1810 "add x24, x25, x19, LSL #2\n"
1811 "add x23, x24, x19, LSL #2\n"
1812 "add x22, x23, x19, LSL #2\n"
1813 "add x21, x22, x19, LSL #2\n"
1814 "add x20, x21, x19, LSL #2\n"
1815 "158:" // Height 8: input setup done
1816 "cmp x28, #0x4\n"
1817 "blt 161f\n"
1818 "ldr q0, [x27, #0x0]\n"
1819 "ldr q1, [x26, #0x0]\n"
1820 "cmp x28, #0x8\n"
1821 "ldr q2, [x25, #0x0]\n"
1822 "ldr q3, [x24, #0x0]\n"
1823 "ldr q4, [x23, #0x0]\n"
1824 "ldr q5, [x22, #0x0]\n"
1825 "ldr q6, [x21, #0x0]\n"
1826 "ldr q7, [x20, #0x0]\n"
1827 "ldr q8, [x12, #0x0]\n"
1828 "blt 160f\n"
1829 "159:" // Height 8: Multiply loop: Main loop head
1830 "fmla v24.4s, v8.4s, v0.s[0]\n"
1831 "ldr q9, [x12, #0x10]\n"
1832 "add x27, x27, #0x10\n"
1833 "fmla v25.4s, v8.4s, v1.s[0]\n"
1834 "ldr q10, [x12, #0x20]\n"
1835 "add x26, x26, #0x10\n"
1836 "fmla v26.4s, v8.4s, v2.s[0]\n"
1837 "ldr q11, [x12, #0x30]\n"
1838 "add x25, x25, #0x10\n"
1839 "fmla v27.4s, v8.4s, v3.s[0]\n"
1840 "prfm pldl1keep, [x27, #0x80]\n"
1841 "add x24, x24, #0x10\n"
1842 "fmla v28.4s, v8.4s, v4.s[0]\n"
1843 "prfm pldl1keep, [x26, #0x80]\n"
1844 "add x23, x23, #0x10\n"
1845 "fmla v29.4s, v8.4s, v5.s[0]\n"
1846 "prfm pldl1keep, [x25, #0x80]\n"
1847 "add x22, x22, #0x10\n"
1848 "fmla v30.4s, v8.4s, v6.s[0]\n"
1849 "prfm pldl1keep, [x24, #0x80]\n"
1850 "add x21, x21, #0x10\n"
1851 "fmla v31.4s, v8.4s, v7.s[0]\n"
1852 "prfm pldl1keep, [x23, #0x80]\n"
1853 "add x20, x20, #0x10\n"
1854 "fmla v24.4s, v9.4s, v0.s[1]\n"
1855 "prfm pldl1keep, [x22, #0x80]\n"
1856 "sub x28, x28, #0x4\n"
1857 "fmla v25.4s, v9.4s, v1.s[1]\n"
1858 "prfm pldl1keep, [x21, #0x80]\n"
1859 "cmp x28, #0x8\n"
1860 "fmla v26.4s, v9.4s, v2.s[1]\n"
1861 "prfm pldl1keep, [x20, #0x80]\n"
1862 "add x12, x12, #0x40\n"
1863 "fmla v27.4s, v9.4s, v3.s[1]\n"
1864 "ldr q8, [x12, #0x0]\n"
1865 "fmla v28.4s, v9.4s, v4.s[1]\n"
1866 "fmla v29.4s, v9.4s, v5.s[1]\n"
1867 "fmla v30.4s, v9.4s, v6.s[1]\n"
1868 "fmla v31.4s, v9.4s, v7.s[1]\n"
1869 "fmla v24.4s, v10.4s, v0.s[2]\n"
1870 "fmla v25.4s, v10.4s, v1.s[2]\n"
1871 "fmla v26.4s, v10.4s, v2.s[2]\n"
1872 "fmla v27.4s, v10.4s, v3.s[2]\n"
1873 "fmla v28.4s, v10.4s, v4.s[2]\n"
1874 "fmla v29.4s, v10.4s, v5.s[2]\n"
1875 "fmla v30.4s, v10.4s, v6.s[2]\n"
1876 "fmla v31.4s, v10.4s, v7.s[2]\n"
1877 "fmla v24.4s, v11.4s, v0.s[3]\n"
1878 "ldr q0, [x27, #0x0]\n"
1879 "fmla v25.4s, v11.4s, v1.s[3]\n"
1880 "ldr q1, [x26, #0x0]\n"
1881 "fmla v26.4s, v11.4s, v2.s[3]\n"
1882 "ldr q2, [x25, #0x0]\n"
1883 "fmla v27.4s, v11.4s, v3.s[3]\n"
1884 "ldr q3, [x24, #0x0]\n"
1885 "fmla v28.4s, v11.4s, v4.s[3]\n"
1886 "ldr q4, [x23, #0x0]\n"
1887 "fmla v29.4s, v11.4s, v5.s[3]\n"
1888 "ldr q5, [x22, #0x0]\n"
1889 "fmla v30.4s, v11.4s, v6.s[3]\n"
1890 "ldr q6, [x21, #0x0]\n"
1891 "fmla v31.4s, v11.4s, v7.s[3]\n"
1892 "ldr q7, [x20, #0x0]\n"
1893 "bge 159b\n"
1894 "160:" // Height 8: Multiply loop: Single iteration only
1895 "fmla v24.4s, v8.4s, v0.s[0]\n"
1896 "ldr q9, [x12, #0x10]\n"
1897 "sub x28, x28, #0x4\n"
1898 "fmla v25.4s, v8.4s, v1.s[0]\n"
1899 "ldr q10, [x12, #0x20]\n"
1900 "add x27, x27, #0x10\n"
1901 "fmla v26.4s, v8.4s, v2.s[0]\n"
1902 "ldr q11, [x12, #0x30]\n"
1903 "add x26, x26, #0x10\n"
1904 "fmla v27.4s, v8.4s, v3.s[0]\n"
1905 "prfm pldl1keep, [x27, #0x80]\n"
1906 "add x25, x25, #0x10\n"
1907 "fmla v28.4s, v8.4s, v4.s[0]\n"
1908 "prfm pldl1keep, [x26, #0x80]\n"
1909 "add x24, x24, #0x10\n"
1910 "fmla v29.4s, v8.4s, v5.s[0]\n"
1911 "prfm pldl1keep, [x25, #0x80]\n"
1912 "add x23, x23, #0x10\n"
1913 "fmla v30.4s, v8.4s, v6.s[0]\n"
1914 "prfm pldl1keep, [x24, #0x80]\n"
1915 "add x22, x22, #0x10\n"
1916 "fmla v31.4s, v8.4s, v7.s[0]\n"
1917 "prfm pldl1keep, [x23, #0x80]\n"
1918 "add x21, x21, #0x10\n"
1919 "fmla v24.4s, v9.4s, v0.s[1]\n"
1920 "prfm pldl1keep, [x22, #0x80]\n"
1921 "add x20, x20, #0x10\n"
1922 "fmla v25.4s, v9.4s, v1.s[1]\n"
1923 "prfm pldl1keep, [x21, #0x80]\n"
1924 "add x12, x12, #0x40\n"
1925 "fmla v26.4s, v9.4s, v2.s[1]\n"
1926 "prfm pldl1keep, [x20, #0x80]\n"
1927 "fmla v27.4s, v9.4s, v3.s[1]\n"
1928 "fmla v28.4s, v9.4s, v4.s[1]\n"
1929 "fmla v29.4s, v9.4s, v5.s[1]\n"
1930 "fmla v30.4s, v9.4s, v6.s[1]\n"
1931 "fmla v31.4s, v9.4s, v7.s[1]\n"
1932 "fmla v24.4s, v10.4s, v0.s[2]\n"
1933 "fmla v25.4s, v10.4s, v1.s[2]\n"
1934 "fmla v26.4s, v10.4s, v2.s[2]\n"
1935 "fmla v27.4s, v10.4s, v3.s[2]\n"
1936 "fmla v28.4s, v10.4s, v4.s[2]\n"
1937 "fmla v29.4s, v10.4s, v5.s[2]\n"
1938 "fmla v30.4s, v10.4s, v6.s[2]\n"
1939 "fmla v31.4s, v10.4s, v7.s[2]\n"
1940 "fmla v24.4s, v11.4s, v0.s[3]\n"
1941 "fmla v25.4s, v11.4s, v1.s[3]\n"
1942 "fmla v26.4s, v11.4s, v2.s[3]\n"
1943 "fmla v27.4s, v11.4s, v3.s[3]\n"
1944 "fmla v28.4s, v11.4s, v4.s[3]\n"
1945 "fmla v29.4s, v11.4s, v5.s[3]\n"
1946 "fmla v30.4s, v11.4s, v6.s[3]\n"
1947 "fmla v31.4s, v11.4s, v7.s[3]\n"
1948 "161:" // Height 8: Multiply loop: Main loop skip
1949 "cbz x28, 163f\n"
1950 "162:" // Height 8: Multiply loop: Odd block loop
1951 "ldr s0, [x27], #0x4\n"
1952 "sub x28, x28, #0x1\n"
1953 "ldr s1, [x26], #0x4\n"
1954 "ldr s2, [x25], #0x4\n"
1955 "ldr s3, [x24], #0x4\n"
1956 "ldr s4, [x23], #0x4\n"
1957 "ldr s5, [x22], #0x4\n"
1958 "ldr s6, [x21], #0x4\n"
1959 "ldr s7, [x20], #0x4\n"
1960 "ldr q12, [x12, #0x0]\n"
1961 "fmla v24.4s, v12.4s, v0.s[0]\n"
1962 "add x12, x12, #0x10\n"
1963 "fmla v25.4s, v12.4s, v1.s[0]\n"
1964 "fmla v26.4s, v12.4s, v2.s[0]\n"
1965 "fmla v27.4s, v12.4s, v3.s[0]\n"
1966 "fmla v28.4s, v12.4s, v4.s[0]\n"
1967 "fmla v29.4s, v12.4s, v5.s[0]\n"
1968 "fmla v30.4s, v12.4s, v6.s[0]\n"
1969 "fmla v31.4s, v12.4s, v7.s[0]\n"
1970 "cbnz x28, 162b\n"
1971 "163:" // Height 8: Multiply loop: No odd multiplies
1972 "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n"
1973 "add x9, x9, #0x1\n"
1974 "cmp x9, x19\n"
1975 "bne 156b\n"
1976 "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
1977 "prfm pstl1keep, [x10, #0x0]\n"
1978 "add x26, x10, x19, LSL #2\n"
1979 "prfm pstl1keep, [x26, #0x0]\n"
1980 "add x25, x26, x19, LSL #2\n"
1981 "prfm pstl1keep, [x25, #0x0]\n"
1982 "add x24, x25, x19, LSL #2\n"
1983 "prfm pstl1keep, [x24, #0x0]\n"
1984 "add x23, x24, x19, LSL #2\n"
1985 "prfm pstl1keep, [x23, #0x0]\n"
1986 "add x22, x23, x19, LSL #2\n"
1987 "prfm pstl1keep, [x22, #0x0]\n"
1988 "add x21, x22, x19, LSL #2\n"
1989 "prfm pstl1keep, [x21, #0x0]\n"
1990 "add x20, x21, x19, LSL #2\n"
1991 "prfm pstl1keep, [x20, #0x0]\n"
1992 "tbz %x[flags], #1, 164f\n"
1993 "add x19, %x[args_ptr], %[offset_min]\n"
1994 "ld1r { v17.4s }, [x19]\n"
1995 "add x19, %x[args_ptr], %[offset_max]\n"
1996 "ld1r { v16.4s }, [x19]\n"
1997 "fmin v24.4s, v24.4s, v16.4s\n"
1998 "fmin v25.4s, v25.4s, v16.4s\n"
1999 "fmin v26.4s, v26.4s, v16.4s\n"
2000 "fmin v27.4s, v27.4s, v16.4s\n"
2001 "fmax v24.4s, v24.4s, v17.4s\n"
2002 "fmax v25.4s, v25.4s, v17.4s\n"
2003 "fmax v26.4s, v26.4s, v17.4s\n"
2004 "fmax v27.4s, v27.4s, v17.4s\n"
2005 "fmin v28.4s, v28.4s, v16.4s\n"
2006 "fmin v29.4s, v29.4s, v16.4s\n"
2007 "fmin v30.4s, v30.4s, v16.4s\n"
2008 "fmax v28.4s, v28.4s, v17.4s\n"
2009 "fmax v29.4s, v29.4s, v17.4s\n"
2010 "fmax v30.4s, v30.4s, v17.4s\n"
2011 "fmin v31.4s, v31.4s, v16.4s\n"
2012 "fmax v31.4s, v31.4s, v17.4s\n"
2013 "164:" // Height 8: No activation
2014 "cmp x13, #0x4\n"
2015 "bge 167f\n"
2016 "tbz x13, #1, 165f\n"
2017 "str d24, [x10], #0x8\n"
2018 "str d25, [x26], #0x8\n"
2019 "str d26, [x25], #0x8\n"
2020 "str d27, [x24], #0x8\n"
2021 "str d28, [x23], #0x8\n"
2022 "str d29, [x22], #0x8\n"
2023 "str d30, [x21], #0x8\n"
2024 "str d31, [x20], #0x8\n"
2025 "tbz x13, #0, 166f\n"
2026 "st1 { v24.s }[2], [x10]\n"
2027 "st1 { v25.s }[2], [x26]\n"
2028 "st1 { v26.s }[2], [x25]\n"
2029 "st1 { v27.s }[2], [x24]\n"
2030 "st1 { v28.s }[2], [x23]\n"
2031 "st1 { v29.s }[2], [x22]\n"
2032 "st1 { v30.s }[2], [x21]\n"
2033 "st1 { v31.s }[2], [x20]\n"
2034 "b 166f\n"
2035 "165:" // Height 8: Partial direct writeback: partial_1_0
2036 "str s24, [x10, #0x0]\n"
2037 "str s25, [x26, #0x0]\n"
2038 "str s26, [x25, #0x0]\n"
2039 "str s27, [x24, #0x0]\n"
2040 "str s28, [x23, #0x0]\n"
2041 "str s29, [x22, #0x0]\n"
2042 "str s30, [x21, #0x0]\n"
2043 "str s31, [x20, #0x0]\n"
2044 "166:" // Height 8: Partial direct writeback: Done
2045 "b 168f\n"
2046 "167:" // Height 8: Full writeback
2047 "str q24, [x10, #0x0]\n"
2048 "add x10, x10, #0x10\n"
2049 "str q25, [x26, #0x0]\n"
2050 "str q26, [x25, #0x0]\n"
2051 "str q27, [x24, #0x0]\n"
2052 "str q28, [x23, #0x0]\n"
2053 "str q29, [x22, #0x0]\n"
2054 "str q30, [x21, #0x0]\n"
2055 "str q31, [x20, #0x0]\n"
2056 "168:" // Height 8: Writeback done
2057 "subs x13, x13, #0x4\n"
2058 "bgt 149b\n"
2059 "subs %x[M], %x[M], #0x8\n"
2060 "beq 170f\n"
2061 "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n"
2062 "tbz %x[flags], #3, 169f\n"
2063 "add x20, x20, #0x8\n"
2064 "str x20, [%x[args_ptr], %[offsetof_input_offset]]\n"
2065 "b 1b\n"
2066 "169:" // Update direct input
2067 "mov x19, #0x20\n"
2068 "madd %x[input_ptr], x19, x20, %x[input_ptr]\n"
2069 "b 1b\n"
2070 "170:" // Exit
2071
2072 : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
2073 : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths))
2074 : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v16", "v17", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
2075 );
2076 }
2077
2078 } // namespace arm_gemm
2079 #endif // __aarch64__
2080