1 /*
2 * Copyright (c) 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 #include "../../bfloat.hpp"
29
30 #include <cassert>
31 #include <limits>
32
33 namespace arm_gemm {
34
a64_hybrid_fp32bf16fp32_mmla_4x24(unsigned int num_strings,const unsigned int * string_lengths,IndirectInputArg<float> A_arg,size_t M,size_t N,const bfloat16 * B_ptr,IndirectOutputArg<float> output_arg,const float * bias,Activation act,bool accumulate)35 void a64_hybrid_fp32bf16fp32_mmla_4x24 (
36 unsigned int num_strings, const unsigned int *string_lengths, IndirectInputArg<float> A_arg,
37 size_t M, size_t N, const bfloat16 *B_ptr, IndirectOutputArg<float> output_arg,
38 const float *bias, Activation act, bool accumulate
39 )
40 {
41 struct KernelArgs {
42 float maxval = static_cast<float>(std::numeric_limits<float>::infinity());
43 float minval = - static_cast<float>(std::numeric_limits<float>::infinity());
44 unsigned int num_strings = {};
45 const unsigned int *string_lengths = {};
46 size_t N = {};
47 const bfloat16 *B_ptr = {};
48 size_t output_offset = {};
49 size_t input_initial_col = {};
50 size_t input_offset = {};
51 } ka;
52
53 unsigned long flags=0;
54 void *output_ptr;
55 void *input_ptr;
56
57 if (output_arg.is_indirect) {
58 output_ptr=(void *)(output_arg.indirect.ptr);
59 ka.output_offset=output_arg.indirect.offset;
60 flags |= 0x4;
61 } else {
62 output_ptr=(void *)(output_arg.direct.base);
63 ka.output_offset=output_arg.direct.stride;
64 }
65
66 if (A_arg.is_indirect) {
67 input_ptr=(void *)(A_arg.indirect.ptr);
68 ka.input_offset=A_arg.indirect.start_row;
69 ka.input_initial_col=A_arg.indirect.start_col;
70 flags |= 0x8;
71 } else {
72 assert(num_strings==1);
73 input_ptr=(void *)(A_arg.direct.base);
74 ka.input_offset=A_arg.direct.stride;
75 }
76 if (accumulate) {
77 flags |= 0x1;
78 }
79 ka.num_strings = num_strings;
80 ka.string_lengths = string_lengths;
81 ka.N = N;
82 ka.B_ptr = B_ptr;
83 switch(act.type) {
84 default:
85 case Activation::Type::None:
86 break;
87 case Activation::Type::BoundedReLU:
88 ka.maxval = static_cast<float>(act.param1);
89 /* fall through */
90 case Activation::Type::ReLU:
91 ka.minval = 0;
92 flags |= 0x2;
93 break;
94 }
95 __asm__ __volatile__(
96
97 "1:" // Row loop
98 "cmp %x[M], #0x4\n"
99 "bge 130f\n"
100 "cmp %x[M], #0x2\n"
101 "bgt 87f\n"
102 "beq 44f\n"
103 "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
104 "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
105 "mov x27, %x[bias]\n"
106 "mov x26, %x[output_ptr]\n"
107 "2:" // Height 1: Column loop
108 "cbz x27, 3f\n"
109 "ldr q8, [x27, #0x0]\n"
110 "zip2 v14.2d, v8.2d, v8.2d\n"
111 "ldr q9, [x27, #0x10]\n"
112 "zip1 v8.2d, v8.2d, v8.2d\n"
113 "ldr q10, [x27, #0x20]\n"
114 "ldr q11, [x27, #0x30]\n"
115 "zip2 v15.2d, v9.2d, v9.2d\n"
116 "ldr q12, [x27, #0x40]\n"
117 "zip1 v9.2d, v9.2d, v9.2d\n"
118 "ldr q13, [x27, #0x50]\n"
119 "add x27, x27, #0x60\n"
120 "zip2 v16.2d, v10.2d, v10.2d\n"
121 "zip1 v10.2d, v10.2d, v10.2d\n"
122 "zip2 v17.2d, v11.2d, v11.2d\n"
123 "zip1 v11.2d, v11.2d, v11.2d\n"
124 "zip2 v18.2d, v12.2d, v12.2d\n"
125 "zip1 v12.2d, v12.2d, v12.2d\n"
126 "zip2 v19.2d, v13.2d, v13.2d\n"
127 "zip1 v13.2d, v13.2d, v13.2d\n"
128 "b 19f\n"
129 "3:" // Height 1: no bias
130 "tbz %x[flags], #0, 18f\n"
131 "cmp x9, #0x18\n"
132 "bge 16f\n"
133 "tbz x9, #4, 7f\n"
134 "ld1 { v9.4s }, [x26], #0x10\n"
135 "ld1 { v10.4s }, [x26], #0x10\n"
136 "ld1 { v11.4s }, [x26], #0x10\n"
137 "ld1 { v12.4s }, [x26], #0x10\n"
138 "tbz x9, #2, 5f\n"
139 "ld1 { v13.4s }, [x26], #0x10\n"
140 "tbz x9, #1, 4f\n"
141 "mov x19, #0x58\n"
142 "ldr d20, [x26], #0x8\n"
143 "tbz x9, #0, 15f\n"
144 "ld1 { v20.s }[2], [x26]\n"
145 "b 15f\n"
146 "4:" // Height 1: Partial accumulate: partial_1_20
147 "mov x19, #0x50\n"
148 "tbz x9, #0, 15f\n"
149 "ldr s20, [x26, #0x0]\n"
150 "b 15f\n"
151 "5:" // Height 1: Partial accumulate: partial_2_16
152 "tbz x9, #1, 6f\n"
153 "ldr d13, [x26], #0x8\n"
154 "mov x19, #0x48\n"
155 "tbz x9, #0, 15f\n"
156 "ld1 { v13.s }[2], [x26]\n"
157 "b 15f\n"
158 "6:" // Height 1: Partial accumulate: partial_1_16
159 "mov x19, #0x40\n"
160 "tbz x9, #0, 15f\n"
161 "ldr s13, [x26, #0x0]\n"
162 "b 15f\n"
163 "7:" // Height 1: Partial accumulate: partial_8_0
164 "tbz x9, #3, 11f\n"
165 "ld1 { v9.4s }, [x26], #0x10\n"
166 "ld1 { v10.4s }, [x26], #0x10\n"
167 "tbz x9, #2, 9f\n"
168 "ld1 { v11.4s }, [x26], #0x10\n"
169 "tbz x9, #1, 8f\n"
170 "mov x19, #0x38\n"
171 "ldr d12, [x26], #0x8\n"
172 "tbz x9, #0, 15f\n"
173 "ld1 { v12.s }[2], [x26]\n"
174 "b 15f\n"
175 "8:" // Height 1: Partial accumulate: partial_1_12
176 "mov x19, #0x30\n"
177 "tbz x9, #0, 15f\n"
178 "ldr s12, [x26, #0x0]\n"
179 "b 15f\n"
180 "9:" // Height 1: Partial accumulate: partial_2_8
181 "tbz x9, #1, 10f\n"
182 "ldr d11, [x26], #0x8\n"
183 "mov x19, #0x28\n"
184 "tbz x9, #0, 15f\n"
185 "ld1 { v11.s }[2], [x26]\n"
186 "b 15f\n"
187 "10:" // Height 1: Partial accumulate: partial_1_8
188 "mov x19, #0x20\n"
189 "tbz x9, #0, 15f\n"
190 "ldr s11, [x26, #0x0]\n"
191 "b 15f\n"
192 "11:" // Height 1: Partial accumulate: partial_4_0
193 "tbz x9, #2, 13f\n"
194 "ld1 { v9.4s }, [x26], #0x10\n"
195 "tbz x9, #1, 12f\n"
196 "ldr d10, [x26], #0x8\n"
197 "mov x19, #0x18\n"
198 "tbz x9, #0, 15f\n"
199 "ld1 { v10.s }[2], [x26]\n"
200 "b 15f\n"
201 "12:" // Height 1: Partial accumulate: partial_1_4
202 "mov x19, #0x10\n"
203 "tbz x9, #0, 15f\n"
204 "ldr s10, [x26, #0x0]\n"
205 "b 15f\n"
206 "13:" // Height 1: Partial accumulate: partial_2_0
207 "tbz x9, #1, 14f\n"
208 "ldr d9, [x26], #0x8\n"
209 "mov x19, #0x8\n"
210 "tbz x9, #0, 15f\n"
211 "ld1 { v9.s }[2], [x26]\n"
212 "b 15f\n"
213 "14:" // Height 1: Partial accumulate: partial_1_0
214 "ldr s9, [x26, #0x0]\n"
215 "mov x19, #0x0\n"
216 "15:" // Height 1: Partial accumulate: Done
217 "sub x26, x26, x19\n"
218 "b 17f\n"
219 "16:" // Height 1: full accumulate
220 "ldr q9, [x26, #0x0]\n"
221 "ldr q10, [x26, #0x10]\n"
222 "ldr q11, [x26, #0x20]\n"
223 "ldr q12, [x26, #0x30]\n"
224 "ldr q13, [x26, #0x40]\n"
225 "ldr q20, [x26, #0x50]\n"
226 "17:" // Height 1: MMLA fixup
227 "zip1 v8.2d, v9.2d, v14.2d\n"
228 "zip2 v14.2d, v9.2d, v14.2d\n"
229 "zip1 v9.2d, v10.2d, v15.2d\n"
230 "zip2 v15.2d, v10.2d, v15.2d\n"
231 "zip1 v10.2d, v11.2d, v16.2d\n"
232 "zip2 v16.2d, v11.2d, v16.2d\n"
233 "zip1 v11.2d, v12.2d, v17.2d\n"
234 "zip2 v17.2d, v12.2d, v17.2d\n"
235 "zip1 v12.2d, v13.2d, v18.2d\n"
236 "zip2 v18.2d, v13.2d, v18.2d\n"
237 "zip1 v13.2d, v20.2d, v19.2d\n"
238 "zip2 v19.2d, v20.2d, v19.2d\n"
239 "b 19f\n"
240 "18:" // Height 1: no accumulate
241 "movi v8.16b, #0x0\n"
242 "movi v9.16b, #0x0\n"
243 "movi v10.16b, #0x0\n"
244 "movi v11.16b, #0x0\n"
245 "movi v12.16b, #0x0\n"
246 "movi v13.16b, #0x0\n"
247 "movi v14.16b, #0x0\n"
248 "movi v15.16b, #0x0\n"
249 "movi v16.16b, #0x0\n"
250 "movi v17.16b, #0x0\n"
251 "movi v18.16b, #0x0\n"
252 "movi v19.16b, #0x0\n"
253 "19:" // Height 1: setup done
254 "mov x25, #0x0\n"
255 "20:" // Height 1: String loop
256 "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
257 "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n"
258 "ldr w24, [x20, x25, LSL #0x2]\n"
259 "tbz %x[flags], #3, 21f\n"
260 "ldr x20, [%x[input_ptr], x25, LSL #0x3]\n"
261 "add x20, x20, x19, LSL #3\n"
262 "ldr x23, [x20, #0x0]\n"
263 "cbnz x25, 22f\n"
264 "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
265 "add x23, x23, x19, LSL #2\n"
266 "b 22f\n"
267 "21:" // Height 1: setup direct input
268 "mov x23, %x[input_ptr]\n"
269 "22:" // Height 1: input setup done
270 "cmp x24, #0x4\n"
271 "blt 25f\n"
272 "ld1 { v0.4s }, [x23], #0x10\n"
273 "cmp x24, #0x8\n"
274 "blt 24f\n"
275 "23:" // Height 1: Multiply loop: Main loop head
276 ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n"
277 "ldr q4, [x28, #0x0]\n"
278 "sub x24, x24, #0x4\n"
279 ".inst 0x6e44ec08 // bfmmla v8.4s, v0.8h, v4.8h\n"
280 "ldr q5, [x28, #0x10]\n"
281 "cmp x24, #0x8\n"
282 ".inst 0x6e45ec0e // bfmmla v14.4s, v0.8h, v5.8h\n"
283 "ldr q6, [x28, #0x20]\n"
284 "ldr q7, [x28, #0x30]\n"
285 ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n"
286 "ldr q4, [x28, #0x40]\n"
287 ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n"
288 "ldr q5, [x28, #0x50]\n"
289 "ldr q6, [x28, #0x60]\n"
290 ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n"
291 "ldr q7, [x28, #0x70]\n"
292 "ldr q4, [x28, #0x80]\n"
293 ".inst 0x6e45ec10 // bfmmla v16.4s, v0.8h, v5.8h\n"
294 ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n"
295 "ldr q5, [x28, #0x90]\n"
296 "ldr q6, [x28, #0xa0]\n"
297 ".inst 0x6e47ec11 // bfmmla v17.4s, v0.8h, v7.8h\n"
298 "ldr q7, [x28, #0xb0]\n"
299 "add x28, x28, #0xc0\n"
300 ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n"
301 "prfm pldl1keep, [x23, #0x80]\n"
302 ".inst 0x6e45ec12 // bfmmla v18.4s, v0.8h, v5.8h\n"
303 ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n"
304 ".inst 0x6e47ec13 // bfmmla v19.4s, v0.8h, v7.8h\n"
305 "ld1 { v0.4s }, [x23], #0x10\n"
306 "bge 23b\n"
307 "24:" // Height 1: Multiply loop: Single iteration only
308 ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n"
309 "ldr q4, [x28, #0x0]\n"
310 "sub x24, x24, #0x4\n"
311 ".inst 0x6e44ec08 // bfmmla v8.4s, v0.8h, v4.8h\n"
312 "ldr q5, [x28, #0x10]\n"
313 "ldr q6, [x28, #0x20]\n"
314 ".inst 0x6e45ec0e // bfmmla v14.4s, v0.8h, v5.8h\n"
315 "ldr q7, [x28, #0x30]\n"
316 ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n"
317 "ldr q4, [x28, #0x40]\n"
318 "ldr q5, [x28, #0x50]\n"
319 ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n"
320 "ldr q6, [x28, #0x60]\n"
321 "ldr q7, [x28, #0x70]\n"
322 ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n"
323 ".inst 0x6e45ec10 // bfmmla v16.4s, v0.8h, v5.8h\n"
324 "ldr q4, [x28, #0x80]\n"
325 "ldr q5, [x28, #0x90]\n"
326 ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n"
327 "ldr q6, [x28, #0xa0]\n"
328 ".inst 0x6e47ec11 // bfmmla v17.4s, v0.8h, v7.8h\n"
329 "ldr q7, [x28, #0xb0]\n"
330 "add x28, x28, #0xc0\n"
331 ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n"
332 "prfm pldl1keep, [x23, #0x80]\n"
333 ".inst 0x6e45ec12 // bfmmla v18.4s, v0.8h, v5.8h\n"
334 ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n"
335 ".inst 0x6e47ec13 // bfmmla v19.4s, v0.8h, v7.8h\n"
336 "25:" // Height 1: Multiply loop: Main loop skip
337 "cbz x24, 28f\n"
338 "cbz x24, 28f\n"
339 "tbz x24, #1, 26f\n"
340 "ldr d0, [x23], #0x8\n"
341 "tbz x24, #0, 27f\n"
342 "ld1 { v0.s }[2], [x23]\n"
343 "b 27f\n"
344 "26:" // Height 1: Multiply loop: Ragged operand read: partial_1_0
345 "ldr s0, [x23, #0x0]\n"
346 "27:" // Height 1: Multiply loop: Ragged operand read: Done
347 ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n"
348 "ldr q4, [x28, #0x0]\n"
349 "ldr q5, [x28, #0x10]\n"
350 ".inst 0x6e44ec08 // bfmmla v8.4s, v0.8h, v4.8h\n"
351 "ldr q6, [x28, #0x20]\n"
352 ".inst 0x6e45ec0e // bfmmla v14.4s, v0.8h, v5.8h\n"
353 "ldr q7, [x28, #0x30]\n"
354 "ldr q4, [x28, #0x40]\n"
355 ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n"
356 "ldr q5, [x28, #0x50]\n"
357 "ldr q6, [x28, #0x60]\n"
358 ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n"
359 ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n"
360 "ldr q7, [x28, #0x70]\n"
361 "ldr q4, [x28, #0x80]\n"
362 ".inst 0x6e45ec10 // bfmmla v16.4s, v0.8h, v5.8h\n"
363 "ldr q5, [x28, #0x90]\n"
364 ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n"
365 "ldr q6, [x28, #0xa0]\n"
366 ".inst 0x6e47ec11 // bfmmla v17.4s, v0.8h, v7.8h\n"
367 "ldr q7, [x28, #0xb0]\n"
368 "add x28, x28, #0xc0\n"
369 ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n"
370 ".inst 0x6e45ec12 // bfmmla v18.4s, v0.8h, v5.8h\n"
371 ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n"
372 ".inst 0x6e47ec13 // bfmmla v19.4s, v0.8h, v7.8h\n"
373 "28:" // Height 1: Multiply loop: No odd multiplies
374 "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n"
375 "add x25, x25, #0x1\n"
376 "cmp x25, x19\n"
377 "bne 20b\n"
378 "uzp1 v8.2d, v8.2d, v14.2d\n"
379 "prfm pstl1keep, [x26, #0x0]\n"
380 "uzp1 v9.2d, v9.2d, v15.2d\n"
381 "uzp1 v10.2d, v10.2d, v16.2d\n"
382 "uzp1 v11.2d, v11.2d, v17.2d\n"
383 "uzp1 v12.2d, v12.2d, v18.2d\n"
384 "uzp1 v13.2d, v13.2d, v19.2d\n"
385 "tbz %x[flags], #1, 29f\n"
386 "add x19, %x[args_ptr], %[offset_min]\n"
387 "ld1r { v1.4s }, [x19]\n"
388 "add x19, %x[args_ptr], %[offset_max]\n"
389 "ld1r { v0.4s }, [x19]\n"
390 "fmin v8.4s, v8.4s, v0.4s\n"
391 "fmin v9.4s, v9.4s, v0.4s\n"
392 "fmin v10.4s, v10.4s, v0.4s\n"
393 "fmin v11.4s, v11.4s, v0.4s\n"
394 "fmax v8.4s, v8.4s, v1.4s\n"
395 "fmax v9.4s, v9.4s, v1.4s\n"
396 "fmax v10.4s, v10.4s, v1.4s\n"
397 "fmax v11.4s, v11.4s, v1.4s\n"
398 "fmin v12.4s, v12.4s, v0.4s\n"
399 "fmin v13.4s, v13.4s, v0.4s\n"
400 "fmax v12.4s, v12.4s, v1.4s\n"
401 "fmax v13.4s, v13.4s, v1.4s\n"
402 "29:" // Height 1: No activation
403 "cmp x9, #0x18\n"
404 "bge 42f\n"
405 "tbz x9, #4, 33f\n"
406 "st1 { v8.4s }, [x26], #0x10\n"
407 "st1 { v9.4s }, [x26], #0x10\n"
408 "st1 { v10.4s }, [x26], #0x10\n"
409 "st1 { v11.4s }, [x26], #0x10\n"
410 "tbz x9, #2, 31f\n"
411 "st1 { v12.4s }, [x26], #0x10\n"
412 "tbz x9, #1, 30f\n"
413 "str d13, [x26], #0x8\n"
414 "tbz x9, #0, 41f\n"
415 "st1 { v13.s }[2], [x26]\n"
416 "b 41f\n"
417 "30:" // Height 1: Partial direct writeback: partial_1_20
418 "tbz x9, #0, 41f\n"
419 "str s13, [x26, #0x0]\n"
420 "b 41f\n"
421 "31:" // Height 1: Partial direct writeback: partial_2_16
422 "tbz x9, #1, 32f\n"
423 "str d12, [x26], #0x8\n"
424 "tbz x9, #0, 41f\n"
425 "st1 { v12.s }[2], [x26]\n"
426 "b 41f\n"
427 "32:" // Height 1: Partial direct writeback: partial_1_16
428 "tbz x9, #0, 41f\n"
429 "str s12, [x26, #0x0]\n"
430 "b 41f\n"
431 "33:" // Height 1: Partial direct writeback: partial_8_0
432 "tbz x9, #3, 37f\n"
433 "st1 { v8.4s }, [x26], #0x10\n"
434 "st1 { v9.4s }, [x26], #0x10\n"
435 "tbz x9, #2, 35f\n"
436 "st1 { v10.4s }, [x26], #0x10\n"
437 "tbz x9, #1, 34f\n"
438 "str d11, [x26], #0x8\n"
439 "tbz x9, #0, 41f\n"
440 "st1 { v11.s }[2], [x26]\n"
441 "b 41f\n"
442 "34:" // Height 1: Partial direct writeback: partial_1_12
443 "tbz x9, #0, 41f\n"
444 "str s11, [x26, #0x0]\n"
445 "b 41f\n"
446 "35:" // Height 1: Partial direct writeback: partial_2_8
447 "tbz x9, #1, 36f\n"
448 "str d10, [x26], #0x8\n"
449 "tbz x9, #0, 41f\n"
450 "st1 { v10.s }[2], [x26]\n"
451 "b 41f\n"
452 "36:" // Height 1: Partial direct writeback: partial_1_8
453 "tbz x9, #0, 41f\n"
454 "str s10, [x26, #0x0]\n"
455 "b 41f\n"
456 "37:" // Height 1: Partial direct writeback: partial_4_0
457 "tbz x9, #2, 39f\n"
458 "st1 { v8.4s }, [x26], #0x10\n"
459 "tbz x9, #1, 38f\n"
460 "str d9, [x26], #0x8\n"
461 "tbz x9, #0, 41f\n"
462 "st1 { v9.s }[2], [x26]\n"
463 "b 41f\n"
464 "38:" // Height 1: Partial direct writeback: partial_1_4
465 "tbz x9, #0, 41f\n"
466 "str s9, [x26, #0x0]\n"
467 "b 41f\n"
468 "39:" // Height 1: Partial direct writeback: partial_2_0
469 "tbz x9, #1, 40f\n"
470 "str d8, [x26], #0x8\n"
471 "tbz x9, #0, 41f\n"
472 "st1 { v8.s }[2], [x26]\n"
473 "b 41f\n"
474 "40:" // Height 1: Partial direct writeback: partial_1_0
475 "str s8, [x26, #0x0]\n"
476 "41:" // Height 1: Partial direct writeback: Done
477 "b 43f\n"
478 "42:" // Height 1: Full writeback
479 "str q8, [x26, #0x0]\n"
480 "str q9, [x26, #0x10]\n"
481 "str q10, [x26, #0x20]\n"
482 "str q11, [x26, #0x30]\n"
483 "str q12, [x26, #0x40]\n"
484 "str q13, [x26, #0x50]\n"
485 "add x26, x26, #0x60\n"
486 "43:" // Height 1: Writeback done
487 "subs x9, x9, #0x18\n"
488 "bgt 2b\n"
489 "b 174f\n"
490 "44:" // Height 2
491 "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
492 "mov x27, %x[bias]\n"
493 "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
494 "mov x26, %x[output_ptr]\n"
495 "45:" // Height 2: Column loop
496 "cbz x27, 46f\n"
497 "ldr q8, [x27, #0x0]\n"
498 "zip2 v14.2d, v8.2d, v8.2d\n"
499 "ldr q9, [x27, #0x10]\n"
500 "zip1 v8.2d, v8.2d, v8.2d\n"
501 "ldr q10, [x27, #0x20]\n"
502 "ldr q11, [x27, #0x30]\n"
503 "zip2 v15.2d, v9.2d, v9.2d\n"
504 "ldr q12, [x27, #0x40]\n"
505 "zip1 v9.2d, v9.2d, v9.2d\n"
506 "ldr q13, [x27, #0x50]\n"
507 "add x27, x27, #0x60\n"
508 "zip2 v16.2d, v10.2d, v10.2d\n"
509 "zip1 v10.2d, v10.2d, v10.2d\n"
510 "zip2 v17.2d, v11.2d, v11.2d\n"
511 "zip1 v11.2d, v11.2d, v11.2d\n"
512 "zip2 v18.2d, v12.2d, v12.2d\n"
513 "zip1 v12.2d, v12.2d, v12.2d\n"
514 "zip2 v19.2d, v13.2d, v13.2d\n"
515 "zip1 v13.2d, v13.2d, v13.2d\n"
516 "b 62f\n"
517 "46:" // Height 2: no bias
518 "tbz %x[flags], #0, 61f\n"
519 "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
520 "cmp x9, #0x18\n"
521 "add x22, x26, x19, LSL #2\n"
522 "bge 59f\n"
523 "tbz x9, #4, 50f\n"
524 "ld1 { v9.4s }, [x26], #0x10\n"
525 "ld1 { v14.4s }, [x22], #0x10\n"
526 "ld1 { v10.4s }, [x26], #0x10\n"
527 "ld1 { v15.4s }, [x22], #0x10\n"
528 "ld1 { v11.4s }, [x26], #0x10\n"
529 "ld1 { v16.4s }, [x22], #0x10\n"
530 "ld1 { v12.4s }, [x26], #0x10\n"
531 "ld1 { v17.4s }, [x22], #0x10\n"
532 "tbz x9, #2, 48f\n"
533 "ld1 { v13.4s }, [x26], #0x10\n"
534 "ld1 { v18.4s }, [x22], #0x10\n"
535 "tbz x9, #1, 47f\n"
536 "mov x19, #0x58\n"
537 "ldr d20, [x26], #0x8\n"
538 "ldr d19, [x22], #0x8\n"
539 "tbz x9, #0, 58f\n"
540 "ld1 { v20.s }[2], [x26]\n"
541 "ld1 { v19.s }[2], [x22]\n"
542 "b 58f\n"
543 "47:" // Height 2: Partial accumulate: partial_1_20
544 "mov x19, #0x50\n"
545 "tbz x9, #0, 58f\n"
546 "ldr s20, [x26, #0x0]\n"
547 "ldr s19, [x22, #0x0]\n"
548 "b 58f\n"
549 "48:" // Height 2: Partial accumulate: partial_2_16
550 "tbz x9, #1, 49f\n"
551 "ldr d13, [x26], #0x8\n"
552 "ldr d18, [x22], #0x8\n"
553 "mov x19, #0x48\n"
554 "tbz x9, #0, 58f\n"
555 "ld1 { v13.s }[2], [x26]\n"
556 "ld1 { v18.s }[2], [x22]\n"
557 "b 58f\n"
558 "49:" // Height 2: Partial accumulate: partial_1_16
559 "mov x19, #0x40\n"
560 "tbz x9, #0, 58f\n"
561 "ldr s13, [x26, #0x0]\n"
562 "ldr s18, [x22, #0x0]\n"
563 "b 58f\n"
564 "50:" // Height 2: Partial accumulate: partial_8_0
565 "tbz x9, #3, 54f\n"
566 "ld1 { v9.4s }, [x26], #0x10\n"
567 "ld1 { v14.4s }, [x22], #0x10\n"
568 "ld1 { v10.4s }, [x26], #0x10\n"
569 "ld1 { v15.4s }, [x22], #0x10\n"
570 "tbz x9, #2, 52f\n"
571 "ld1 { v11.4s }, [x26], #0x10\n"
572 "ld1 { v16.4s }, [x22], #0x10\n"
573 "tbz x9, #1, 51f\n"
574 "mov x19, #0x38\n"
575 "ldr d12, [x26], #0x8\n"
576 "ldr d17, [x22], #0x8\n"
577 "tbz x9, #0, 58f\n"
578 "ld1 { v12.s }[2], [x26]\n"
579 "ld1 { v17.s }[2], [x22]\n"
580 "b 58f\n"
581 "51:" // Height 2: Partial accumulate: partial_1_12
582 "mov x19, #0x30\n"
583 "tbz x9, #0, 58f\n"
584 "ldr s12, [x26, #0x0]\n"
585 "ldr s17, [x22, #0x0]\n"
586 "b 58f\n"
587 "52:" // Height 2: Partial accumulate: partial_2_8
588 "tbz x9, #1, 53f\n"
589 "ldr d11, [x26], #0x8\n"
590 "ldr d16, [x22], #0x8\n"
591 "mov x19, #0x28\n"
592 "tbz x9, #0, 58f\n"
593 "ld1 { v11.s }[2], [x26]\n"
594 "ld1 { v16.s }[2], [x22]\n"
595 "b 58f\n"
596 "53:" // Height 2: Partial accumulate: partial_1_8
597 "mov x19, #0x20\n"
598 "tbz x9, #0, 58f\n"
599 "ldr s11, [x26, #0x0]\n"
600 "ldr s16, [x22, #0x0]\n"
601 "b 58f\n"
602 "54:" // Height 2: Partial accumulate: partial_4_0
603 "tbz x9, #2, 56f\n"
604 "ld1 { v9.4s }, [x26], #0x10\n"
605 "ld1 { v14.4s }, [x22], #0x10\n"
606 "tbz x9, #1, 55f\n"
607 "mov x19, #0x18\n"
608 "ldr d10, [x26], #0x8\n"
609 "ldr d15, [x22], #0x8\n"
610 "tbz x9, #0, 58f\n"
611 "ld1 { v10.s }[2], [x26]\n"
612 "ld1 { v15.s }[2], [x22]\n"
613 "b 58f\n"
614 "55:" // Height 2: Partial accumulate: partial_1_4
615 "mov x19, #0x10\n"
616 "tbz x9, #0, 58f\n"
617 "ldr s10, [x26, #0x0]\n"
618 "ldr s15, [x22, #0x0]\n"
619 "b 58f\n"
620 "56:" // Height 2: Partial accumulate: partial_2_0
621 "tbz x9, #1, 57f\n"
622 "ldr d9, [x26], #0x8\n"
623 "ldr d14, [x22], #0x8\n"
624 "mov x19, #0x8\n"
625 "tbz x9, #0, 58f\n"
626 "ld1 { v9.s }[2], [x26]\n"
627 "ld1 { v14.s }[2], [x22]\n"
628 "b 58f\n"
629 "57:" // Height 2: Partial accumulate: partial_1_0
630 "ldr s9, [x26, #0x0]\n"
631 "mov x19, #0x0\n"
632 "ldr s14, [x22, #0x0]\n"
633 "58:" // Height 2: Partial accumulate: Done
634 "sub x26, x26, x19\n"
635 "b 60f\n"
636 "59:" // Height 2: full accumulate
637 "ldr q9, [x26, #0x0]\n"
638 "ldr q10, [x26, #0x10]\n"
639 "ldr q11, [x26, #0x20]\n"
640 "ldr q12, [x26, #0x30]\n"
641 "ldr q13, [x26, #0x40]\n"
642 "ldr q20, [x26, #0x50]\n"
643 "ldr q14, [x22, #0x0]\n"
644 "ldr q15, [x22, #0x10]\n"
645 "ldr q16, [x22, #0x20]\n"
646 "ldr q17, [x22, #0x30]\n"
647 "ldr q18, [x22, #0x40]\n"
648 "ldr q19, [x22, #0x50]\n"
649 "60:" // Height 2: MMLA fixup
650 "zip1 v8.2d, v9.2d, v14.2d\n"
651 "zip2 v14.2d, v9.2d, v14.2d\n"
652 "zip1 v9.2d, v10.2d, v15.2d\n"
653 "zip2 v15.2d, v10.2d, v15.2d\n"
654 "zip1 v10.2d, v11.2d, v16.2d\n"
655 "zip2 v16.2d, v11.2d, v16.2d\n"
656 "zip1 v11.2d, v12.2d, v17.2d\n"
657 "zip2 v17.2d, v12.2d, v17.2d\n"
658 "zip1 v12.2d, v13.2d, v18.2d\n"
659 "zip2 v18.2d, v13.2d, v18.2d\n"
660 "zip1 v13.2d, v20.2d, v19.2d\n"
661 "zip2 v19.2d, v20.2d, v19.2d\n"
662 "b 62f\n"
663 "61:" // Height 2: no accumulate
664 "movi v8.16b, #0x0\n"
665 "movi v9.16b, #0x0\n"
666 "movi v10.16b, #0x0\n"
667 "movi v11.16b, #0x0\n"
668 "movi v12.16b, #0x0\n"
669 "movi v13.16b, #0x0\n"
670 "movi v14.16b, #0x0\n"
671 "movi v15.16b, #0x0\n"
672 "movi v16.16b, #0x0\n"
673 "movi v17.16b, #0x0\n"
674 "movi v18.16b, #0x0\n"
675 "movi v19.16b, #0x0\n"
676 "62:" // Height 2: setup done
677 "mov x25, #0x0\n"
678 "63:" // Height 2: String loop
679 "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
680 "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n"
681 "ldr w24, [x20, x25, LSL #0x2]\n"
682 "tbz %x[flags], #3, 64f\n"
683 "ldr x20, [%x[input_ptr], x25, LSL #0x3]\n"
684 "add x20, x20, x19, LSL #3\n"
685 "ldr x23, [x20, #0x0]\n"
686 "ldr x22, [x20, #0x8]\n"
687 "cbnz x25, 65f\n"
688 "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
689 "add x23, x23, x19, LSL #2\n"
690 "add x22, x22, x19, LSL #2\n"
691 "b 65f\n"
692 "64:" // Height 2: setup direct input
693 "mov x23, %x[input_ptr]\n"
694 "add x22, x23, x19, LSL #2\n"
695 "65:" // Height 2: input setup done
696 "cmp x24, #0x4\n"
697 "blt 68f\n"
698 "ld1 { v0.4s }, [x23], #0x10\n"
699 "cmp x24, #0x8\n"
700 "blt 67f\n"
701 "66:" // Height 2: Multiply loop: Main loop head
702 ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n"
703 "ld1 { v1.4s }, [x22], #0x10\n"
704 "sub x24, x24, #0x4\n"
705 ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n"
706 "ldr q4, [x28, #0x0]\n"
707 "cmp x24, #0x8\n"
708 ".inst 0x6e44ec08 // bfmmla v8.4s, v0.8h, v4.8h\n"
709 "ldr q5, [x28, #0x10]\n"
710 "ldr q6, [x28, #0x20]\n"
711 ".inst 0x6e45ec0e // bfmmla v14.4s, v0.8h, v5.8h\n"
712 "ldr q7, [x28, #0x30]\n"
713 ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n"
714 "ldr q4, [x28, #0x40]\n"
715 "ldr q5, [x28, #0x50]\n"
716 ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n"
717 "ldr q6, [x28, #0x60]\n"
718 "ldr q7, [x28, #0x70]\n"
719 ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n"
720 ".inst 0x6e45ec10 // bfmmla v16.4s, v0.8h, v5.8h\n"
721 "ldr q4, [x28, #0x80]\n"
722 "ldr q5, [x28, #0x90]\n"
723 ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n"
724 "ldr q6, [x28, #0xa0]\n"
725 ".inst 0x6e47ec11 // bfmmla v17.4s, v0.8h, v7.8h\n"
726 "ldr q7, [x28, #0xb0]\n"
727 "add x28, x28, #0xc0\n"
728 ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n"
729 "prfm pldl1keep, [x23, #0x80]\n"
730 ".inst 0x6e45ec12 // bfmmla v18.4s, v0.8h, v5.8h\n"
731 "prfm pldl1keep, [x22, #0x80]\n"
732 ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n"
733 ".inst 0x6e47ec13 // bfmmla v19.4s, v0.8h, v7.8h\n"
734 "ld1 { v0.4s }, [x23], #0x10\n"
735 "bge 66b\n"
736 "67:" // Height 2: Multiply loop: Single iteration only
737 ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n"
738 "ld1 { v1.4s }, [x22], #0x10\n"
739 "sub x24, x24, #0x4\n"
740 ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n"
741 "ldr q4, [x28, #0x0]\n"
742 "ldr q5, [x28, #0x10]\n"
743 ".inst 0x6e44ec08 // bfmmla v8.4s, v0.8h, v4.8h\n"
744 "ldr q6, [x28, #0x20]\n"
745 ".inst 0x6e45ec0e // bfmmla v14.4s, v0.8h, v5.8h\n"
746 "ldr q7, [x28, #0x30]\n"
747 "ldr q4, [x28, #0x40]\n"
748 ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n"
749 "ldr q5, [x28, #0x50]\n"
750 "ldr q6, [x28, #0x60]\n"
751 ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n"
752 ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n"
753 "ldr q7, [x28, #0x70]\n"
754 "ldr q4, [x28, #0x80]\n"
755 ".inst 0x6e45ec10 // bfmmla v16.4s, v0.8h, v5.8h\n"
756 "ldr q5, [x28, #0x90]\n"
757 ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n"
758 "ldr q6, [x28, #0xa0]\n"
759 "prfm pldl1keep, [x23, #0x80]\n"
760 ".inst 0x6e47ec11 // bfmmla v17.4s, v0.8h, v7.8h\n"
761 ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n"
762 "ldr q7, [x28, #0xb0]\n"
763 "add x28, x28, #0xc0\n"
764 ".inst 0x6e45ec12 // bfmmla v18.4s, v0.8h, v5.8h\n"
765 "prfm pldl1keep, [x22, #0x80]\n"
766 ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n"
767 ".inst 0x6e47ec13 // bfmmla v19.4s, v0.8h, v7.8h\n"
768 "68:" // Height 2: Multiply loop: Main loop skip
769 "cbz x24, 71f\n"
770 "cbz x24, 71f\n"
771 "tbz x24, #1, 69f\n"
772 "ldr d0, [x23], #0x8\n"
773 "ldr d1, [x22], #0x8\n"
774 "tbz x24, #0, 70f\n"
775 "ld1 { v0.s }[2], [x23]\n"
776 "ld1 { v1.s }[2], [x22]\n"
777 "b 70f\n"
778 "69:" // Height 2: Multiply loop: Ragged operand read: partial_1_0
779 "ldr s0, [x23, #0x0]\n"
780 "ldr s1, [x22, #0x0]\n"
781 "70:" // Height 2: Multiply loop: Ragged operand read: Done
782 ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n"
783 "ldr q4, [x28, #0x0]\n"
784 "ldr q5, [x28, #0x10]\n"
785 ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n"
786 "ldr q6, [x28, #0x20]\n"
787 "ldr q7, [x28, #0x30]\n"
788 ".inst 0x6e44ec08 // bfmmla v8.4s, v0.8h, v4.8h\n"
789 "ldr q4, [x28, #0x40]\n"
790 ".inst 0x6e45ec0e // bfmmla v14.4s, v0.8h, v5.8h\n"
791 "ldr q5, [x28, #0x50]\n"
792 ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n"
793 "ldr q6, [x28, #0x60]\n"
794 ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n"
795 "ldr q7, [x28, #0x70]\n"
796 ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n"
797 "ldr q4, [x28, #0x80]\n"
798 ".inst 0x6e45ec10 // bfmmla v16.4s, v0.8h, v5.8h\n"
799 "ldr q5, [x28, #0x90]\n"
800 ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n"
801 "ldr q6, [x28, #0xa0]\n"
802 ".inst 0x6e47ec11 // bfmmla v17.4s, v0.8h, v7.8h\n"
803 "ldr q7, [x28, #0xb0]\n"
804 "add x28, x28, #0xc0\n"
805 ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n"
806 ".inst 0x6e45ec12 // bfmmla v18.4s, v0.8h, v5.8h\n"
807 ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n"
808 ".inst 0x6e47ec13 // bfmmla v19.4s, v0.8h, v7.8h\n"
809 "71:" // Height 2: Multiply loop: No odd multiplies
810 "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n"
811 "add x25, x25, #0x1\n"
812 "cmp x25, x19\n"
813 "bne 63b\n"
814 "uzp1 v4.2d, v8.2d, v14.2d\n"
815 "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
816 "uzp2 v8.2d, v8.2d, v14.2d\n"
817 "prfm pstl1keep, [x26, #0x0]\n"
818 "add x22, x26, x19, LSL #2\n"
819 "uzp1 v14.2d, v9.2d, v15.2d\n"
820 "prfm pstl1keep, [x22, #0x0]\n"
821 "uzp2 v9.2d, v9.2d, v15.2d\n"
822 "uzp1 v15.2d, v10.2d, v16.2d\n"
823 "uzp2 v10.2d, v10.2d, v16.2d\n"
824 "uzp1 v16.2d, v11.2d, v17.2d\n"
825 "uzp2 v11.2d, v11.2d, v17.2d\n"
826 "uzp1 v17.2d, v12.2d, v18.2d\n"
827 "uzp2 v12.2d, v12.2d, v18.2d\n"
828 "uzp1 v18.2d, v13.2d, v19.2d\n"
829 "uzp2 v13.2d, v13.2d, v19.2d\n"
830 "tbz %x[flags], #1, 72f\n"
831 "add x19, %x[args_ptr], %[offset_min]\n"
832 "ld1r { v1.4s }, [x19]\n"
833 "add x19, %x[args_ptr], %[offset_max]\n"
834 "ld1r { v0.4s }, [x19]\n"
835 "fmin v4.4s, v4.4s, v0.4s\n"
836 "fmin v14.4s, v14.4s, v0.4s\n"
837 "fmin v15.4s, v15.4s, v0.4s\n"
838 "fmin v16.4s, v16.4s, v0.4s\n"
839 "fmax v4.4s, v4.4s, v1.4s\n"
840 "fmax v14.4s, v14.4s, v1.4s\n"
841 "fmax v15.4s, v15.4s, v1.4s\n"
842 "fmax v16.4s, v16.4s, v1.4s\n"
843 "fmin v17.4s, v17.4s, v0.4s\n"
844 "fmin v18.4s, v18.4s, v0.4s\n"
845 "fmin v8.4s, v8.4s, v0.4s\n"
846 "fmax v17.4s, v17.4s, v1.4s\n"
847 "fmax v18.4s, v18.4s, v1.4s\n"
848 "fmax v8.4s, v8.4s, v1.4s\n"
849 "fmin v9.4s, v9.4s, v0.4s\n"
850 "fmin v10.4s, v10.4s, v0.4s\n"
851 "fmin v11.4s, v11.4s, v0.4s\n"
852 "fmax v9.4s, v9.4s, v1.4s\n"
853 "fmax v10.4s, v10.4s, v1.4s\n"
854 "fmax v11.4s, v11.4s, v1.4s\n"
855 "fmin v12.4s, v12.4s, v0.4s\n"
856 "fmin v13.4s, v13.4s, v0.4s\n"
857 "fmax v12.4s, v12.4s, v1.4s\n"
858 "fmax v13.4s, v13.4s, v1.4s\n"
859 "72:" // Height 2: No activation
860 "cmp x9, #0x18\n"
861 "bge 85f\n"
862 "tbz x9, #4, 76f\n"
863 "st1 { v4.4s }, [x26], #0x10\n"
864 "st1 { v14.4s }, [x26], #0x10\n"
865 "st1 { v15.4s }, [x26], #0x10\n"
866 "st1 { v16.4s }, [x26], #0x10\n"
867 "st1 { v8.4s }, [x22], #0x10\n"
868 "st1 { v9.4s }, [x22], #0x10\n"
869 "st1 { v10.4s }, [x22], #0x10\n"
870 "st1 { v11.4s }, [x22], #0x10\n"
871 "tbz x9, #2, 74f\n"
872 "st1 { v17.4s }, [x26], #0x10\n"
873 "st1 { v12.4s }, [x22], #0x10\n"
874 "tbz x9, #1, 73f\n"
875 "str d18, [x26], #0x8\n"
876 "str d13, [x22], #0x8\n"
877 "tbz x9, #0, 84f\n"
878 "st1 { v18.s }[2], [x26]\n"
879 "st1 { v13.s }[2], [x22]\n"
880 "b 84f\n"
881 "73:" // Height 2: Partial direct writeback: partial_1_20
882 "tbz x9, #0, 84f\n"
883 "str s18, [x26, #0x0]\n"
884 "str s13, [x22, #0x0]\n"
885 "b 84f\n"
886 "74:" // Height 2: Partial direct writeback: partial_2_16
887 "tbz x9, #1, 75f\n"
888 "str d17, [x26], #0x8\n"
889 "str d12, [x22], #0x8\n"
890 "tbz x9, #0, 84f\n"
891 "st1 { v17.s }[2], [x26]\n"
892 "st1 { v12.s }[2], [x22]\n"
893 "b 84f\n"
894 "75:" // Height 2: Partial direct writeback: partial_1_16
895 "tbz x9, #0, 84f\n"
896 "str s17, [x26, #0x0]\n"
897 "str s12, [x22, #0x0]\n"
898 "b 84f\n"
899 "76:" // Height 2: Partial direct writeback: partial_8_0
900 "tbz x9, #3, 80f\n"
901 "st1 { v4.4s }, [x26], #0x10\n"
902 "st1 { v14.4s }, [x26], #0x10\n"
903 "st1 { v8.4s }, [x22], #0x10\n"
904 "st1 { v9.4s }, [x22], #0x10\n"
905 "tbz x9, #2, 78f\n"
906 "st1 { v15.4s }, [x26], #0x10\n"
907 "st1 { v10.4s }, [x22], #0x10\n"
908 "tbz x9, #1, 77f\n"
909 "str d16, [x26], #0x8\n"
910 "str d11, [x22], #0x8\n"
911 "tbz x9, #0, 84f\n"
912 "st1 { v16.s }[2], [x26]\n"
913 "st1 { v11.s }[2], [x22]\n"
914 "b 84f\n"
915 "77:" // Height 2: Partial direct writeback: partial_1_12
916 "tbz x9, #0, 84f\n"
917 "str s16, [x26, #0x0]\n"
918 "str s11, [x22, #0x0]\n"
919 "b 84f\n"
920 "78:" // Height 2: Partial direct writeback: partial_2_8
921 "tbz x9, #1, 79f\n"
922 "str d15, [x26], #0x8\n"
923 "str d10, [x22], #0x8\n"
924 "tbz x9, #0, 84f\n"
925 "st1 { v15.s }[2], [x26]\n"
926 "st1 { v10.s }[2], [x22]\n"
927 "b 84f\n"
928 "79:" // Height 2: Partial direct writeback: partial_1_8
929 "tbz x9, #0, 84f\n"
930 "str s15, [x26, #0x0]\n"
931 "str s10, [x22, #0x0]\n"
932 "b 84f\n"
933 "80:" // Height 2: Partial direct writeback: partial_4_0
934 "tbz x9, #2, 82f\n"
935 "st1 { v4.4s }, [x26], #0x10\n"
936 "st1 { v8.4s }, [x22], #0x10\n"
937 "tbz x9, #1, 81f\n"
938 "str d14, [x26], #0x8\n"
939 "str d9, [x22], #0x8\n"
940 "tbz x9, #0, 84f\n"
941 "st1 { v14.s }[2], [x26]\n"
942 "st1 { v9.s }[2], [x22]\n"
943 "b 84f\n"
944 "81:" // Height 2: Partial direct writeback: partial_1_4
945 "tbz x9, #0, 84f\n"
946 "str s14, [x26, #0x0]\n"
947 "str s9, [x22, #0x0]\n"
948 "b 84f\n"
949 "82:" // Height 2: Partial direct writeback: partial_2_0
950 "tbz x9, #1, 83f\n"
951 "str d4, [x26], #0x8\n"
952 "str d8, [x22], #0x8\n"
953 "tbz x9, #0, 84f\n"
954 "st1 { v4.s }[2], [x26]\n"
955 "st1 { v8.s }[2], [x22]\n"
956 "b 84f\n"
957 "83:" // Height 2: Partial direct writeback: partial_1_0
958 "str s4, [x26, #0x0]\n"
959 "str s8, [x22, #0x0]\n"
960 "84:" // Height 2: Partial direct writeback: Done
961 "b 86f\n"
962 "85:" // Height 2: Full writeback
963 "str q4, [x26, #0x0]\n"
964 "str q14, [x26, #0x10]\n"
965 "str q15, [x26, #0x20]\n"
966 "str q16, [x26, #0x30]\n"
967 "str q17, [x26, #0x40]\n"
968 "str q18, [x26, #0x50]\n"
969 "add x26, x26, #0x60\n"
970 "str q8, [x22, #0x0]\n"
971 "str q9, [x22, #0x10]\n"
972 "str q10, [x22, #0x20]\n"
973 "str q11, [x22, #0x30]\n"
974 "str q12, [x22, #0x40]\n"
975 "str q13, [x22, #0x50]\n"
976 "86:" // Height 2: Writeback done
977 "subs x9, x9, #0x18\n"
978 "bgt 45b\n"
979 "b 174f\n"
980 "87:" // Height 3
981 "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
982 "mov x27, %x[bias]\n"
983 "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
984 "mov x26, %x[output_ptr]\n"
985 "88:" // Height 3: Column loop
986 "cbz x27, 89f\n"
987 "ldr q8, [x27, #0x0]\n"
988 "zip2 v14.2d, v8.2d, v8.2d\n"
989 "ldr q9, [x27, #0x10]\n"
990 "zip1 v8.2d, v8.2d, v8.2d\n"
991 "ldr q10, [x27, #0x20]\n"
992 "mov v20.16b, v8.16b\n"
993 "ldr q11, [x27, #0x30]\n"
994 "mov v26.16b, v14.16b\n"
995 "ldr q12, [x27, #0x40]\n"
996 "ldr q13, [x27, #0x50]\n"
997 "zip2 v15.2d, v9.2d, v9.2d\n"
998 "add x27, x27, #0x60\n"
999 "zip1 v9.2d, v9.2d, v9.2d\n"
1000 "zip2 v16.2d, v10.2d, v10.2d\n"
1001 "zip1 v10.2d, v10.2d, v10.2d\n"
1002 "zip2 v17.2d, v11.2d, v11.2d\n"
1003 "zip1 v11.2d, v11.2d, v11.2d\n"
1004 "zip2 v18.2d, v12.2d, v12.2d\n"
1005 "zip1 v12.2d, v12.2d, v12.2d\n"
1006 "zip2 v19.2d, v13.2d, v13.2d\n"
1007 "zip1 v13.2d, v13.2d, v13.2d\n"
1008 "mov v21.16b, v9.16b\n"
1009 "mov v27.16b, v15.16b\n"
1010 "mov v22.16b, v10.16b\n"
1011 "mov v28.16b, v16.16b\n"
1012 "mov v23.16b, v11.16b\n"
1013 "mov v29.16b, v17.16b\n"
1014 "mov v24.16b, v12.16b\n"
1015 "mov v30.16b, v18.16b\n"
1016 "mov v25.16b, v13.16b\n"
1017 "mov v31.16b, v19.16b\n"
1018 "b 105f\n"
1019 "89:" // Height 3: no bias
1020 "tbz %x[flags], #0, 104f\n"
1021 "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
1022 "cmp x9, #0x18\n"
1023 "add x22, x26, x19, LSL #2\n"
1024 "add x21, x22, x19, LSL #2\n"
1025 "bge 102f\n"
1026 "tbz x9, #4, 93f\n"
1027 "ld1 { v9.4s }, [x26], #0x10\n"
1028 "ld1 { v14.4s }, [x22], #0x10\n"
1029 "ld1 { v21.4s }, [x21], #0x10\n"
1030 "ld1 { v10.4s }, [x26], #0x10\n"
1031 "ld1 { v15.4s }, [x22], #0x10\n"
1032 "ld1 { v22.4s }, [x21], #0x10\n"
1033 "ld1 { v11.4s }, [x26], #0x10\n"
1034 "ld1 { v16.4s }, [x22], #0x10\n"
1035 "ld1 { v23.4s }, [x21], #0x10\n"
1036 "ld1 { v12.4s }, [x26], #0x10\n"
1037 "ld1 { v17.4s }, [x22], #0x10\n"
1038 "ld1 { v24.4s }, [x21], #0x10\n"
1039 "tbz x9, #2, 91f\n"
1040 "ld1 { v13.4s }, [x26], #0x10\n"
1041 "ld1 { v18.4s }, [x22], #0x10\n"
1042 "ld1 { v25.4s }, [x21], #0x10\n"
1043 "tbz x9, #1, 90f\n"
1044 "mov x19, #0x58\n"
1045 "ldr d20, [x26], #0x8\n"
1046 "ldr d19, [x22], #0x8\n"
1047 "ldr d4, [x21], #0x8\n"
1048 "tbz x9, #0, 101f\n"
1049 "ld1 { v20.s }[2], [x26]\n"
1050 "ld1 { v19.s }[2], [x22]\n"
1051 "ld1 { v4.s }[2], [x21]\n"
1052 "b 101f\n"
1053 "90:" // Height 3: Partial accumulate: partial_1_20
1054 "mov x19, #0x50\n"
1055 "tbz x9, #0, 101f\n"
1056 "ldr s20, [x26, #0x0]\n"
1057 "ldr s19, [x22, #0x0]\n"
1058 "ldr s4, [x21, #0x0]\n"
1059 "b 101f\n"
1060 "91:" // Height 3: Partial accumulate: partial_2_16
1061 "tbz x9, #1, 92f\n"
1062 "ldr d13, [x26], #0x8\n"
1063 "ldr d18, [x22], #0x8\n"
1064 "mov x19, #0x48\n"
1065 "ldr d25, [x21], #0x8\n"
1066 "tbz x9, #0, 101f\n"
1067 "ld1 { v13.s }[2], [x26]\n"
1068 "ld1 { v18.s }[2], [x22]\n"
1069 "ld1 { v25.s }[2], [x21]\n"
1070 "b 101f\n"
1071 "92:" // Height 3: Partial accumulate: partial_1_16
1072 "mov x19, #0x40\n"
1073 "tbz x9, #0, 101f\n"
1074 "ldr s13, [x26, #0x0]\n"
1075 "ldr s18, [x22, #0x0]\n"
1076 "ldr s25, [x21, #0x0]\n"
1077 "b 101f\n"
1078 "93:" // Height 3: Partial accumulate: partial_8_0
1079 "tbz x9, #3, 97f\n"
1080 "ld1 { v9.4s }, [x26], #0x10\n"
1081 "ld1 { v14.4s }, [x22], #0x10\n"
1082 "ld1 { v21.4s }, [x21], #0x10\n"
1083 "ld1 { v10.4s }, [x26], #0x10\n"
1084 "ld1 { v15.4s }, [x22], #0x10\n"
1085 "ld1 { v22.4s }, [x21], #0x10\n"
1086 "tbz x9, #2, 95f\n"
1087 "ld1 { v11.4s }, [x26], #0x10\n"
1088 "ld1 { v16.4s }, [x22], #0x10\n"
1089 "ld1 { v23.4s }, [x21], #0x10\n"
1090 "tbz x9, #1, 94f\n"
1091 "mov x19, #0x38\n"
1092 "ldr d12, [x26], #0x8\n"
1093 "ldr d17, [x22], #0x8\n"
1094 "ldr d24, [x21], #0x8\n"
1095 "tbz x9, #0, 101f\n"
1096 "ld1 { v12.s }[2], [x26]\n"
1097 "ld1 { v17.s }[2], [x22]\n"
1098 "ld1 { v24.s }[2], [x21]\n"
1099 "b 101f\n"
1100 "94:" // Height 3: Partial accumulate: partial_1_12
1101 "mov x19, #0x30\n"
1102 "tbz x9, #0, 101f\n"
1103 "ldr s12, [x26, #0x0]\n"
1104 "ldr s17, [x22, #0x0]\n"
1105 "ldr s24, [x21, #0x0]\n"
1106 "b 101f\n"
1107 "95:" // Height 3: Partial accumulate: partial_2_8
1108 "tbz x9, #1, 96f\n"
1109 "ldr d11, [x26], #0x8\n"
1110 "ldr d16, [x22], #0x8\n"
1111 "mov x19, #0x28\n"
1112 "ldr d23, [x21], #0x8\n"
1113 "tbz x9, #0, 101f\n"
1114 "ld1 { v11.s }[2], [x26]\n"
1115 "ld1 { v16.s }[2], [x22]\n"
1116 "ld1 { v23.s }[2], [x21]\n"
1117 "b 101f\n"
1118 "96:" // Height 3: Partial accumulate: partial_1_8
1119 "mov x19, #0x20\n"
1120 "tbz x9, #0, 101f\n"
1121 "ldr s11, [x26, #0x0]\n"
1122 "ldr s16, [x22, #0x0]\n"
1123 "ldr s23, [x21, #0x0]\n"
1124 "b 101f\n"
1125 "97:" // Height 3: Partial accumulate: partial_4_0
1126 "tbz x9, #2, 99f\n"
1127 "ld1 { v9.4s }, [x26], #0x10\n"
1128 "ld1 { v14.4s }, [x22], #0x10\n"
1129 "ld1 { v21.4s }, [x21], #0x10\n"
1130 "tbz x9, #1, 98f\n"
1131 "mov x19, #0x18\n"
1132 "ldr d10, [x26], #0x8\n"
1133 "ldr d15, [x22], #0x8\n"
1134 "ldr d22, [x21], #0x8\n"
1135 "tbz x9, #0, 101f\n"
1136 "ld1 { v10.s }[2], [x26]\n"
1137 "ld1 { v15.s }[2], [x22]\n"
1138 "ld1 { v22.s }[2], [x21]\n"
1139 "b 101f\n"
1140 "98:" // Height 3: Partial accumulate: partial_1_4
1141 "mov x19, #0x10\n"
1142 "tbz x9, #0, 101f\n"
1143 "ldr s10, [x26, #0x0]\n"
1144 "ldr s15, [x22, #0x0]\n"
1145 "ldr s22, [x21, #0x0]\n"
1146 "b 101f\n"
1147 "99:" // Height 3: Partial accumulate: partial_2_0
1148 "tbz x9, #1, 100f\n"
1149 "ldr d9, [x26], #0x8\n"
1150 "ldr d14, [x22], #0x8\n"
1151 "mov x19, #0x8\n"
1152 "ldr d21, [x21], #0x8\n"
1153 "tbz x9, #0, 101f\n"
1154 "ld1 { v9.s }[2], [x26]\n"
1155 "ld1 { v14.s }[2], [x22]\n"
1156 "ld1 { v21.s }[2], [x21]\n"
1157 "b 101f\n"
1158 "100:" // Height 3: Partial accumulate: partial_1_0
1159 "ldr s9, [x26, #0x0]\n"
1160 "mov x19, #0x0\n"
1161 "ldr s14, [x22, #0x0]\n"
1162 "ldr s21, [x21, #0x0]\n"
1163 "101:" // Height 3: Partial accumulate: Done
1164 "sub x26, x26, x19\n"
1165 "b 103f\n"
1166 "102:" // Height 3: full accumulate
1167 "ldr q9, [x26, #0x0]\n"
1168 "ldr q10, [x26, #0x10]\n"
1169 "ldr q11, [x26, #0x20]\n"
1170 "ldr q12, [x26, #0x30]\n"
1171 "ldr q13, [x26, #0x40]\n"
1172 "ldr q20, [x26, #0x50]\n"
1173 "ldr q14, [x22, #0x0]\n"
1174 "ldr q15, [x22, #0x10]\n"
1175 "ldr q16, [x22, #0x20]\n"
1176 "ldr q17, [x22, #0x30]\n"
1177 "ldr q18, [x22, #0x40]\n"
1178 "ldr q19, [x22, #0x50]\n"
1179 "ldr q21, [x21, #0x0]\n"
1180 "ldr q22, [x21, #0x10]\n"
1181 "ldr q23, [x21, #0x20]\n"
1182 "ldr q24, [x21, #0x30]\n"
1183 "ldr q25, [x21, #0x40]\n"
1184 "ldr q4, [x21, #0x50]\n"
1185 "103:" // Height 3: MMLA fixup
1186 "zip1 v8.2d, v9.2d, v14.2d\n"
1187 "zip2 v14.2d, v9.2d, v14.2d\n"
1188 "zip1 v9.2d, v10.2d, v15.2d\n"
1189 "zip2 v15.2d, v10.2d, v15.2d\n"
1190 "zip1 v10.2d, v11.2d, v16.2d\n"
1191 "zip2 v16.2d, v11.2d, v16.2d\n"
1192 "zip1 v11.2d, v12.2d, v17.2d\n"
1193 "zip2 v17.2d, v12.2d, v17.2d\n"
1194 "zip1 v12.2d, v13.2d, v18.2d\n"
1195 "zip2 v18.2d, v13.2d, v18.2d\n"
1196 "zip1 v13.2d, v20.2d, v19.2d\n"
1197 "zip2 v19.2d, v20.2d, v19.2d\n"
1198 "zip1 v20.2d, v21.2d, v26.2d\n"
1199 "zip2 v26.2d, v21.2d, v26.2d\n"
1200 "zip1 v21.2d, v22.2d, v27.2d\n"
1201 "zip2 v27.2d, v22.2d, v27.2d\n"
1202 "zip1 v22.2d, v23.2d, v28.2d\n"
1203 "zip2 v28.2d, v23.2d, v28.2d\n"
1204 "zip1 v23.2d, v24.2d, v29.2d\n"
1205 "zip2 v29.2d, v24.2d, v29.2d\n"
1206 "zip1 v24.2d, v25.2d, v30.2d\n"
1207 "zip2 v30.2d, v25.2d, v30.2d\n"
1208 "zip1 v25.2d, v4.2d, v31.2d\n"
1209 "zip2 v31.2d, v4.2d, v31.2d\n"
1210 "b 105f\n"
1211 "104:" // Height 3: no accumulate
1212 "movi v8.16b, #0x0\n"
1213 "movi v9.16b, #0x0\n"
1214 "movi v10.16b, #0x0\n"
1215 "movi v11.16b, #0x0\n"
1216 "movi v12.16b, #0x0\n"
1217 "movi v13.16b, #0x0\n"
1218 "movi v14.16b, #0x0\n"
1219 "movi v15.16b, #0x0\n"
1220 "movi v16.16b, #0x0\n"
1221 "movi v17.16b, #0x0\n"
1222 "movi v18.16b, #0x0\n"
1223 "movi v19.16b, #0x0\n"
1224 "movi v20.16b, #0x0\n"
1225 "movi v21.16b, #0x0\n"
1226 "movi v22.16b, #0x0\n"
1227 "movi v23.16b, #0x0\n"
1228 "movi v24.16b, #0x0\n"
1229 "movi v25.16b, #0x0\n"
1230 "movi v26.16b, #0x0\n"
1231 "movi v27.16b, #0x0\n"
1232 "movi v28.16b, #0x0\n"
1233 "movi v29.16b, #0x0\n"
1234 "movi v30.16b, #0x0\n"
1235 "movi v31.16b, #0x0\n"
1236 "105:" // Height 3: setup done
1237 "mov x25, #0x0\n"
1238 "106:" // Height 3: String loop
1239 "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
1240 "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n"
1241 "ldr w24, [x20, x25, LSL #0x2]\n"
1242 "tbz %x[flags], #3, 107f\n"
1243 "ldr x20, [%x[input_ptr], x25, LSL #0x3]\n"
1244 "add x20, x20, x19, LSL #3\n"
1245 "ldr x23, [x20, #0x0]\n"
1246 "ldr x22, [x20, #0x8]\n"
1247 "ldr x21, [x20, #0x10]\n"
1248 "cbnz x25, 108f\n"
1249 "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
1250 "add x23, x23, x19, LSL #2\n"
1251 "add x22, x22, x19, LSL #2\n"
1252 "add x21, x21, x19, LSL #2\n"
1253 "b 108f\n"
1254 "107:" // Height 3: setup direct input
1255 "mov x23, %x[input_ptr]\n"
1256 "add x22, x23, x19, LSL #2\n"
1257 "add x21, x22, x19, LSL #2\n"
1258 "108:" // Height 3: input setup done
1259 "cmp x24, #0x4\n"
1260 "blt 111f\n"
1261 "ld1 { v0.4s }, [x23], #0x10\n"
1262 "cmp x24, #0x8\n"
1263 "blt 110f\n"
1264 "109:" // Height 3: Multiply loop: Main loop head
1265 ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n"
1266 "ld1 { v1.4s }, [x22], #0x10\n"
1267 "sub x24, x24, #0x4\n"
1268 ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n"
1269 "ld1 { v2.4s }, [x21], #0x10\n"
1270 "cmp x24, #0x8\n"
1271 ".inst 0x0ea16842 // bfcvtn v2.4h, v2.4s\n"
1272 "ldr q4, [x28, #0x0]\n"
1273 "ldr q5, [x28, #0x10]\n"
1274 ".inst 0x6e44ec08 // bfmmla v8.4s, v0.8h, v4.8h\n"
1275 "ldr q6, [x28, #0x20]\n"
1276 ".inst 0x6e44ec54 // bfmmla v20.4s, v2.8h, v4.8h\n"
1277 "ldr q7, [x28, #0x30]\n"
1278 ".inst 0x6e45ec0e // bfmmla v14.4s, v0.8h, v5.8h\n"
1279 "ldr q4, [x28, #0x40]\n"
1280 ".inst 0x6e45ec5a // bfmmla v26.4s, v2.8h, v5.8h\n"
1281 "ldr q5, [x28, #0x50]\n"
1282 "prfm pldl1keep, [x23, #0x80]\n"
1283 ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n"
1284 ".inst 0x6e46ec55 // bfmmla v21.4s, v2.8h, v6.8h\n"
1285 "ldr q6, [x28, #0x60]\n"
1286 ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n"
1287 "prfm pldl1keep, [x22, #0x80]\n"
1288 ".inst 0x6e47ec5b // bfmmla v27.4s, v2.8h, v7.8h\n"
1289 "ldr q7, [x28, #0x70]\n"
1290 ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n"
1291 "prfm pldl1keep, [x21, #0x80]\n"
1292 ".inst 0x6e44ec56 // bfmmla v22.4s, v2.8h, v4.8h\n"
1293 "ldr q4, [x28, #0x80]\n"
1294 ".inst 0x6e45ec10 // bfmmla v16.4s, v0.8h, v5.8h\n"
1295 ".inst 0x6e45ec5c // bfmmla v28.4s, v2.8h, v5.8h\n"
1296 "ldr q5, [x28, #0x90]\n"
1297 ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n"
1298 ".inst 0x6e46ec57 // bfmmla v23.4s, v2.8h, v6.8h\n"
1299 "ldr q6, [x28, #0xa0]\n"
1300 ".inst 0x6e47ec11 // bfmmla v17.4s, v0.8h, v7.8h\n"
1301 ".inst 0x6e47ec5d // bfmmla v29.4s, v2.8h, v7.8h\n"
1302 "ldr q7, [x28, #0xb0]\n"
1303 "add x28, x28, #0xc0\n"
1304 ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n"
1305 ".inst 0x6e44ec58 // bfmmla v24.4s, v2.8h, v4.8h\n"
1306 ".inst 0x6e45ec12 // bfmmla v18.4s, v0.8h, v5.8h\n"
1307 ".inst 0x6e45ec5e // bfmmla v30.4s, v2.8h, v5.8h\n"
1308 ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n"
1309 ".inst 0x6e46ec59 // bfmmla v25.4s, v2.8h, v6.8h\n"
1310 ".inst 0x6e47ec13 // bfmmla v19.4s, v0.8h, v7.8h\n"
1311 "ld1 { v0.4s }, [x23], #0x10\n"
1312 ".inst 0x6e47ec5f // bfmmla v31.4s, v2.8h, v7.8h\n"
1313 "bge 109b\n"
1314 "110:" // Height 3: Multiply loop: Single iteration only
1315 ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n"
1316 "ld1 { v1.4s }, [x22], #0x10\n"
1317 "sub x24, x24, #0x4\n"
1318 ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n"
1319 "ld1 { v2.4s }, [x21], #0x10\n"
1320 "ldr q4, [x28, #0x0]\n"
1321 ".inst 0x0ea16842 // bfcvtn v2.4h, v2.4s\n"
1322 "ldr q5, [x28, #0x10]\n"
1323 ".inst 0x6e44ec08 // bfmmla v8.4s, v0.8h, v4.8h\n"
1324 "ldr q6, [x28, #0x20]\n"
1325 "ldr q7, [x28, #0x30]\n"
1326 ".inst 0x6e44ec54 // bfmmla v20.4s, v2.8h, v4.8h\n"
1327 "ldr q4, [x28, #0x40]\n"
1328 "prfm pldl1keep, [x23, #0x80]\n"
1329 ".inst 0x6e45ec0e // bfmmla v14.4s, v0.8h, v5.8h\n"
1330 ".inst 0x6e45ec5a // bfmmla v26.4s, v2.8h, v5.8h\n"
1331 "ldr q5, [x28, #0x50]\n"
1332 ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n"
1333 "prfm pldl1keep, [x22, #0x80]\n"
1334 ".inst 0x6e46ec55 // bfmmla v21.4s, v2.8h, v6.8h\n"
1335 "ldr q6, [x28, #0x60]\n"
1336 ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n"
1337 "prfm pldl1keep, [x21, #0x80]\n"
1338 ".inst 0x6e47ec5b // bfmmla v27.4s, v2.8h, v7.8h\n"
1339 "ldr q7, [x28, #0x70]\n"
1340 ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n"
1341 ".inst 0x6e44ec56 // bfmmla v22.4s, v2.8h, v4.8h\n"
1342 "ldr q4, [x28, #0x80]\n"
1343 ".inst 0x6e45ec10 // bfmmla v16.4s, v0.8h, v5.8h\n"
1344 ".inst 0x6e45ec5c // bfmmla v28.4s, v2.8h, v5.8h\n"
1345 "ldr q5, [x28, #0x90]\n"
1346 ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n"
1347 ".inst 0x6e46ec57 // bfmmla v23.4s, v2.8h, v6.8h\n"
1348 "ldr q6, [x28, #0xa0]\n"
1349 ".inst 0x6e47ec11 // bfmmla v17.4s, v0.8h, v7.8h\n"
1350 ".inst 0x6e47ec5d // bfmmla v29.4s, v2.8h, v7.8h\n"
1351 "ldr q7, [x28, #0xb0]\n"
1352 "add x28, x28, #0xc0\n"
1353 ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n"
1354 ".inst 0x6e44ec58 // bfmmla v24.4s, v2.8h, v4.8h\n"
1355 ".inst 0x6e45ec12 // bfmmla v18.4s, v0.8h, v5.8h\n"
1356 ".inst 0x6e45ec5e // bfmmla v30.4s, v2.8h, v5.8h\n"
1357 ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n"
1358 ".inst 0x6e46ec59 // bfmmla v25.4s, v2.8h, v6.8h\n"
1359 ".inst 0x6e47ec13 // bfmmla v19.4s, v0.8h, v7.8h\n"
1360 ".inst 0x6e47ec5f // bfmmla v31.4s, v2.8h, v7.8h\n"
1361 "111:" // Height 3: Multiply loop: Main loop skip
1362 "cbz x24, 114f\n"
1363 "cbz x24, 114f\n"
1364 "tbz x24, #1, 112f\n"
1365 "ldr d0, [x23], #0x8\n"
1366 "ldr d1, [x22], #0x8\n"
1367 "ldr d2, [x21], #0x8\n"
1368 "tbz x24, #0, 113f\n"
1369 "ld1 { v0.s }[2], [x23]\n"
1370 "ld1 { v1.s }[2], [x22]\n"
1371 "ld1 { v2.s }[2], [x21]\n"
1372 "b 113f\n"
1373 "112:" // Height 3: Multiply loop: Ragged operand read: partial_1_0
1374 "ldr s0, [x23, #0x0]\n"
1375 "ldr s1, [x22, #0x0]\n"
1376 "ldr s2, [x21, #0x0]\n"
1377 "113:" // Height 3: Multiply loop: Ragged operand read: Done
1378 ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n"
1379 "ldr q4, [x28, #0x0]\n"
1380 ".inst 0x0ea16842 // bfcvtn v2.4h, v2.4s\n"
1381 "ldr q5, [x28, #0x10]\n"
1382 "ldr q6, [x28, #0x20]\n"
1383 ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n"
1384 "ldr q7, [x28, #0x30]\n"
1385 ".inst 0x6e44ec54 // bfmmla v20.4s, v2.8h, v4.8h\n"
1386 ".inst 0x6e45ec5a // bfmmla v26.4s, v2.8h, v5.8h\n"
1387 ".inst 0x6e46ec55 // bfmmla v21.4s, v2.8h, v6.8h\n"
1388 ".inst 0x6e44ec08 // bfmmla v8.4s, v0.8h, v4.8h\n"
1389 "ldr q4, [x28, #0x40]\n"
1390 ".inst 0x6e45ec0e // bfmmla v14.4s, v0.8h, v5.8h\n"
1391 "ldr q5, [x28, #0x50]\n"
1392 ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n"
1393 "ldr q6, [x28, #0x60]\n"
1394 ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n"
1395 ".inst 0x6e47ec5b // bfmmla v27.4s, v2.8h, v7.8h\n"
1396 "ldr q7, [x28, #0x70]\n"
1397 ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n"
1398 ".inst 0x6e44ec56 // bfmmla v22.4s, v2.8h, v4.8h\n"
1399 "ldr q4, [x28, #0x80]\n"
1400 ".inst 0x6e45ec10 // bfmmla v16.4s, v0.8h, v5.8h\n"
1401 ".inst 0x6e45ec5c // bfmmla v28.4s, v2.8h, v5.8h\n"
1402 "ldr q5, [x28, #0x90]\n"
1403 ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n"
1404 ".inst 0x6e46ec57 // bfmmla v23.4s, v2.8h, v6.8h\n"
1405 "ldr q6, [x28, #0xa0]\n"
1406 ".inst 0x6e47ec11 // bfmmla v17.4s, v0.8h, v7.8h\n"
1407 ".inst 0x6e47ec5d // bfmmla v29.4s, v2.8h, v7.8h\n"
1408 "ldr q7, [x28, #0xb0]\n"
1409 "add x28, x28, #0xc0\n"
1410 ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n"
1411 ".inst 0x6e44ec58 // bfmmla v24.4s, v2.8h, v4.8h\n"
1412 ".inst 0x6e45ec12 // bfmmla v18.4s, v0.8h, v5.8h\n"
1413 ".inst 0x6e45ec5e // bfmmla v30.4s, v2.8h, v5.8h\n"
1414 ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n"
1415 ".inst 0x6e46ec59 // bfmmla v25.4s, v2.8h, v6.8h\n"
1416 ".inst 0x6e47ec13 // bfmmla v19.4s, v0.8h, v7.8h\n"
1417 ".inst 0x6e47ec5f // bfmmla v31.4s, v2.8h, v7.8h\n"
1418 "114:" // Height 3: Multiply loop: No odd multiplies
1419 "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n"
1420 "add x25, x25, #0x1\n"
1421 "cmp x25, x19\n"
1422 "bne 106b\n"
1423 "uzp1 v4.2d, v8.2d, v14.2d\n"
1424 "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
1425 "uzp2 v8.2d, v8.2d, v14.2d\n"
1426 "prfm pstl1keep, [x26, #0x0]\n"
1427 "add x22, x26, x19, LSL #2\n"
1428 "uzp1 v14.2d, v9.2d, v15.2d\n"
1429 "prfm pstl1keep, [x22, #0x0]\n"
1430 "uzp2 v9.2d, v9.2d, v15.2d\n"
1431 "add x21, x22, x19, LSL #2\n"
1432 "uzp1 v15.2d, v10.2d, v16.2d\n"
1433 "prfm pstl1keep, [x21, #0x0]\n"
1434 "uzp2 v10.2d, v10.2d, v16.2d\n"
1435 "uzp1 v16.2d, v11.2d, v17.2d\n"
1436 "uzp2 v11.2d, v11.2d, v17.2d\n"
1437 "uzp1 v17.2d, v12.2d, v18.2d\n"
1438 "uzp2 v12.2d, v12.2d, v18.2d\n"
1439 "uzp1 v18.2d, v13.2d, v19.2d\n"
1440 "uzp2 v13.2d, v13.2d, v19.2d\n"
1441 "uzp1 v20.2d, v20.2d, v26.2d\n"
1442 "uzp1 v21.2d, v21.2d, v27.2d\n"
1443 "uzp1 v22.2d, v22.2d, v28.2d\n"
1444 "uzp1 v23.2d, v23.2d, v29.2d\n"
1445 "uzp1 v24.2d, v24.2d, v30.2d\n"
1446 "uzp1 v25.2d, v25.2d, v31.2d\n"
1447 "tbz %x[flags], #1, 115f\n"
1448 "add x19, %x[args_ptr], %[offset_min]\n"
1449 "ld1r { v1.4s }, [x19]\n"
1450 "add x19, %x[args_ptr], %[offset_max]\n"
1451 "ld1r { v0.4s }, [x19]\n"
1452 "fmin v4.4s, v4.4s, v0.4s\n"
1453 "fmin v14.4s, v14.4s, v0.4s\n"
1454 "fmin v15.4s, v15.4s, v0.4s\n"
1455 "fmin v16.4s, v16.4s, v0.4s\n"
1456 "fmax v4.4s, v4.4s, v1.4s\n"
1457 "fmax v14.4s, v14.4s, v1.4s\n"
1458 "fmax v15.4s, v15.4s, v1.4s\n"
1459 "fmax v16.4s, v16.4s, v1.4s\n"
1460 "fmin v17.4s, v17.4s, v0.4s\n"
1461 "fmin v18.4s, v18.4s, v0.4s\n"
1462 "fmin v8.4s, v8.4s, v0.4s\n"
1463 "fmax v17.4s, v17.4s, v1.4s\n"
1464 "fmax v18.4s, v18.4s, v1.4s\n"
1465 "fmax v8.4s, v8.4s, v1.4s\n"
1466 "fmin v9.4s, v9.4s, v0.4s\n"
1467 "fmin v10.4s, v10.4s, v0.4s\n"
1468 "fmin v11.4s, v11.4s, v0.4s\n"
1469 "fmax v9.4s, v9.4s, v1.4s\n"
1470 "fmax v10.4s, v10.4s, v1.4s\n"
1471 "fmax v11.4s, v11.4s, v1.4s\n"
1472 "fmin v12.4s, v12.4s, v0.4s\n"
1473 "fmin v13.4s, v13.4s, v0.4s\n"
1474 "fmin v20.4s, v20.4s, v0.4s\n"
1475 "fmax v12.4s, v12.4s, v1.4s\n"
1476 "fmax v13.4s, v13.4s, v1.4s\n"
1477 "fmax v20.4s, v20.4s, v1.4s\n"
1478 "fmin v21.4s, v21.4s, v0.4s\n"
1479 "fmin v22.4s, v22.4s, v0.4s\n"
1480 "fmin v23.4s, v23.4s, v0.4s\n"
1481 "fmax v21.4s, v21.4s, v1.4s\n"
1482 "fmax v22.4s, v22.4s, v1.4s\n"
1483 "fmax v23.4s, v23.4s, v1.4s\n"
1484 "fmin v24.4s, v24.4s, v0.4s\n"
1485 "fmin v25.4s, v25.4s, v0.4s\n"
1486 "fmax v24.4s, v24.4s, v1.4s\n"
1487 "fmax v25.4s, v25.4s, v1.4s\n"
1488 "115:" // Height 3: No activation
1489 "cmp x9, #0x18\n"
1490 "bge 128f\n"
1491 "tbz x9, #4, 119f\n"
1492 "st1 { v4.4s }, [x26], #0x10\n"
1493 "st1 { v14.4s }, [x26], #0x10\n"
1494 "st1 { v15.4s }, [x26], #0x10\n"
1495 "st1 { v16.4s }, [x26], #0x10\n"
1496 "st1 { v8.4s }, [x22], #0x10\n"
1497 "st1 { v9.4s }, [x22], #0x10\n"
1498 "st1 { v10.4s }, [x22], #0x10\n"
1499 "st1 { v11.4s }, [x22], #0x10\n"
1500 "st1 { v20.4s }, [x21], #0x10\n"
1501 "st1 { v21.4s }, [x21], #0x10\n"
1502 "st1 { v22.4s }, [x21], #0x10\n"
1503 "st1 { v23.4s }, [x21], #0x10\n"
1504 "tbz x9, #2, 117f\n"
1505 "st1 { v17.4s }, [x26], #0x10\n"
1506 "st1 { v12.4s }, [x22], #0x10\n"
1507 "st1 { v24.4s }, [x21], #0x10\n"
1508 "tbz x9, #1, 116f\n"
1509 "str d18, [x26], #0x8\n"
1510 "str d13, [x22], #0x8\n"
1511 "str d25, [x21], #0x8\n"
1512 "tbz x9, #0, 127f\n"
1513 "st1 { v18.s }[2], [x26]\n"
1514 "st1 { v13.s }[2], [x22]\n"
1515 "st1 { v25.s }[2], [x21]\n"
1516 "b 127f\n"
1517 "116:" // Height 3: Partial direct writeback: partial_1_20
1518 "tbz x9, #0, 127f\n"
1519 "str s18, [x26, #0x0]\n"
1520 "str s13, [x22, #0x0]\n"
1521 "str s25, [x21, #0x0]\n"
1522 "b 127f\n"
1523 "117:" // Height 3: Partial direct writeback: partial_2_16
1524 "tbz x9, #1, 118f\n"
1525 "str d17, [x26], #0x8\n"
1526 "str d12, [x22], #0x8\n"
1527 "str d24, [x21], #0x8\n"
1528 "tbz x9, #0, 127f\n"
1529 "st1 { v17.s }[2], [x26]\n"
1530 "st1 { v12.s }[2], [x22]\n"
1531 "st1 { v24.s }[2], [x21]\n"
1532 "b 127f\n"
1533 "118:" // Height 3: Partial direct writeback: partial_1_16
1534 "tbz x9, #0, 127f\n"
1535 "str s17, [x26, #0x0]\n"
1536 "str s12, [x22, #0x0]\n"
1537 "str s24, [x21, #0x0]\n"
1538 "b 127f\n"
1539 "119:" // Height 3: Partial direct writeback: partial_8_0
1540 "tbz x9, #3, 123f\n"
1541 "st1 { v4.4s }, [x26], #0x10\n"
1542 "st1 { v14.4s }, [x26], #0x10\n"
1543 "st1 { v8.4s }, [x22], #0x10\n"
1544 "st1 { v9.4s }, [x22], #0x10\n"
1545 "st1 { v20.4s }, [x21], #0x10\n"
1546 "st1 { v21.4s }, [x21], #0x10\n"
1547 "tbz x9, #2, 121f\n"
1548 "st1 { v15.4s }, [x26], #0x10\n"
1549 "st1 { v10.4s }, [x22], #0x10\n"
1550 "st1 { v22.4s }, [x21], #0x10\n"
1551 "tbz x9, #1, 120f\n"
1552 "str d16, [x26], #0x8\n"
1553 "str d11, [x22], #0x8\n"
1554 "str d23, [x21], #0x8\n"
1555 "tbz x9, #0, 127f\n"
1556 "st1 { v16.s }[2], [x26]\n"
1557 "st1 { v11.s }[2], [x22]\n"
1558 "st1 { v23.s }[2], [x21]\n"
1559 "b 127f\n"
1560 "120:" // Height 3: Partial direct writeback: partial_1_12
1561 "tbz x9, #0, 127f\n"
1562 "str s16, [x26, #0x0]\n"
1563 "str s11, [x22, #0x0]\n"
1564 "str s23, [x21, #0x0]\n"
1565 "b 127f\n"
1566 "121:" // Height 3: Partial direct writeback: partial_2_8
1567 "tbz x9, #1, 122f\n"
1568 "str d15, [x26], #0x8\n"
1569 "str d10, [x22], #0x8\n"
1570 "str d22, [x21], #0x8\n"
1571 "tbz x9, #0, 127f\n"
1572 "st1 { v15.s }[2], [x26]\n"
1573 "st1 { v10.s }[2], [x22]\n"
1574 "st1 { v22.s }[2], [x21]\n"
1575 "b 127f\n"
1576 "122:" // Height 3: Partial direct writeback: partial_1_8
1577 "tbz x9, #0, 127f\n"
1578 "str s15, [x26, #0x0]\n"
1579 "str s10, [x22, #0x0]\n"
1580 "str s22, [x21, #0x0]\n"
1581 "b 127f\n"
1582 "123:" // Height 3: Partial direct writeback: partial_4_0
1583 "tbz x9, #2, 125f\n"
1584 "st1 { v4.4s }, [x26], #0x10\n"
1585 "st1 { v8.4s }, [x22], #0x10\n"
1586 "st1 { v20.4s }, [x21], #0x10\n"
1587 "tbz x9, #1, 124f\n"
1588 "str d14, [x26], #0x8\n"
1589 "str d9, [x22], #0x8\n"
1590 "str d21, [x21], #0x8\n"
1591 "tbz x9, #0, 127f\n"
1592 "st1 { v14.s }[2], [x26]\n"
1593 "st1 { v9.s }[2], [x22]\n"
1594 "st1 { v21.s }[2], [x21]\n"
1595 "b 127f\n"
1596 "124:" // Height 3: Partial direct writeback: partial_1_4
1597 "tbz x9, #0, 127f\n"
1598 "str s14, [x26, #0x0]\n"
1599 "str s9, [x22, #0x0]\n"
1600 "str s21, [x21, #0x0]\n"
1601 "b 127f\n"
1602 "125:" // Height 3: Partial direct writeback: partial_2_0
1603 "tbz x9, #1, 126f\n"
1604 "str d4, [x26], #0x8\n"
1605 "str d8, [x22], #0x8\n"
1606 "str d20, [x21], #0x8\n"
1607 "tbz x9, #0, 127f\n"
1608 "st1 { v4.s }[2], [x26]\n"
1609 "st1 { v8.s }[2], [x22]\n"
1610 "st1 { v20.s }[2], [x21]\n"
1611 "b 127f\n"
1612 "126:" // Height 3: Partial direct writeback: partial_1_0
1613 "str s4, [x26, #0x0]\n"
1614 "str s8, [x22, #0x0]\n"
1615 "str s20, [x21, #0x0]\n"
1616 "127:" // Height 3: Partial direct writeback: Done
1617 "b 129f\n"
1618 "128:" // Height 3: Full writeback
1619 "str q4, [x26, #0x0]\n"
1620 "str q14, [x26, #0x10]\n"
1621 "str q15, [x26, #0x20]\n"
1622 "str q16, [x26, #0x30]\n"
1623 "str q17, [x26, #0x40]\n"
1624 "str q18, [x26, #0x50]\n"
1625 "add x26, x26, #0x60\n"
1626 "str q8, [x22, #0x0]\n"
1627 "str q9, [x22, #0x10]\n"
1628 "str q10, [x22, #0x20]\n"
1629 "str q11, [x22, #0x30]\n"
1630 "str q12, [x22, #0x40]\n"
1631 "str q13, [x22, #0x50]\n"
1632 "str q20, [x21, #0x0]\n"
1633 "str q21, [x21, #0x10]\n"
1634 "str q22, [x21, #0x20]\n"
1635 "str q23, [x21, #0x30]\n"
1636 "str q24, [x21, #0x40]\n"
1637 "str q25, [x21, #0x50]\n"
1638 "129:" // Height 3: Writeback done
1639 "subs x9, x9, #0x18\n"
1640 "bgt 88b\n"
1641 "b 174f\n"
1642 "130:" // Height 4
1643 "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
1644 "mov x27, %x[bias]\n"
1645 "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
1646 "mov x26, %x[output_ptr]\n"
1647 "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
1648 "mov x19, #0x10\n"
1649 "madd %x[output_ptr], x20, x19, %x[output_ptr]\n"
1650 "131:" // Height 4: Column loop
1651 "cbz x27, 132f\n"
1652 "ldr q8, [x27, #0x0]\n"
1653 "zip2 v14.2d, v8.2d, v8.2d\n"
1654 "ldr q9, [x27, #0x10]\n"
1655 "zip1 v8.2d, v8.2d, v8.2d\n"
1656 "ldr q10, [x27, #0x20]\n"
1657 "mov v20.16b, v8.16b\n"
1658 "ldr q11, [x27, #0x30]\n"
1659 "mov v26.16b, v14.16b\n"
1660 "ldr q12, [x27, #0x40]\n"
1661 "ldr q13, [x27, #0x50]\n"
1662 "zip2 v15.2d, v9.2d, v9.2d\n"
1663 "add x27, x27, #0x60\n"
1664 "zip1 v9.2d, v9.2d, v9.2d\n"
1665 "zip2 v16.2d, v10.2d, v10.2d\n"
1666 "zip1 v10.2d, v10.2d, v10.2d\n"
1667 "zip2 v17.2d, v11.2d, v11.2d\n"
1668 "zip1 v11.2d, v11.2d, v11.2d\n"
1669 "zip2 v18.2d, v12.2d, v12.2d\n"
1670 "zip1 v12.2d, v12.2d, v12.2d\n"
1671 "zip2 v19.2d, v13.2d, v13.2d\n"
1672 "zip1 v13.2d, v13.2d, v13.2d\n"
1673 "mov v21.16b, v9.16b\n"
1674 "mov v27.16b, v15.16b\n"
1675 "mov v22.16b, v10.16b\n"
1676 "mov v28.16b, v16.16b\n"
1677 "mov v23.16b, v11.16b\n"
1678 "mov v29.16b, v17.16b\n"
1679 "mov v24.16b, v12.16b\n"
1680 "mov v30.16b, v18.16b\n"
1681 "mov v25.16b, v13.16b\n"
1682 "mov v31.16b, v19.16b\n"
1683 "b 148f\n"
1684 "132:" // Height 4: no bias
1685 "tbz %x[flags], #0, 147f\n"
1686 "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
1687 "cmp x9, #0x18\n"
1688 "add x22, x26, x19, LSL #2\n"
1689 "add x21, x22, x19, LSL #2\n"
1690 "add x20, x21, x19, LSL #2\n"
1691 "bge 145f\n"
1692 "tbz x9, #4, 136f\n"
1693 "ld1 { v9.4s }, [x26], #0x10\n"
1694 "ld1 { v14.4s }, [x22], #0x10\n"
1695 "ld1 { v21.4s }, [x21], #0x10\n"
1696 "ld1 { v26.4s }, [x20], #0x10\n"
1697 "ld1 { v10.4s }, [x26], #0x10\n"
1698 "ld1 { v15.4s }, [x22], #0x10\n"
1699 "ld1 { v22.4s }, [x21], #0x10\n"
1700 "ld1 { v27.4s }, [x20], #0x10\n"
1701 "ld1 { v11.4s }, [x26], #0x10\n"
1702 "ld1 { v16.4s }, [x22], #0x10\n"
1703 "ld1 { v23.4s }, [x21], #0x10\n"
1704 "ld1 { v28.4s }, [x20], #0x10\n"
1705 "ld1 { v12.4s }, [x26], #0x10\n"
1706 "ld1 { v17.4s }, [x22], #0x10\n"
1707 "ld1 { v24.4s }, [x21], #0x10\n"
1708 "ld1 { v29.4s }, [x20], #0x10\n"
1709 "tbz x9, #2, 134f\n"
1710 "ld1 { v13.4s }, [x26], #0x10\n"
1711 "ld1 { v18.4s }, [x22], #0x10\n"
1712 "ld1 { v25.4s }, [x21], #0x10\n"
1713 "ld1 { v30.4s }, [x20], #0x10\n"
1714 "tbz x9, #1, 133f\n"
1715 "mov x19, #0x58\n"
1716 "ldr d20, [x26], #0x8\n"
1717 "ldr d19, [x22], #0x8\n"
1718 "ldr d4, [x21], #0x8\n"
1719 "ldr d31, [x20], #0x8\n"
1720 "tbz x9, #0, 144f\n"
1721 "ld1 { v20.s }[2], [x26]\n"
1722 "ld1 { v19.s }[2], [x22]\n"
1723 "ld1 { v4.s }[2], [x21]\n"
1724 "ld1 { v31.s }[2], [x20]\n"
1725 "b 144f\n"
1726 "133:" // Height 4: Partial accumulate: partial_1_20
1727 "mov x19, #0x50\n"
1728 "tbz x9, #0, 144f\n"
1729 "ldr s20, [x26, #0x0]\n"
1730 "ldr s19, [x22, #0x0]\n"
1731 "ldr s4, [x21, #0x0]\n"
1732 "ldr s31, [x20, #0x0]\n"
1733 "b 144f\n"
1734 "134:" // Height 4: Partial accumulate: partial_2_16
1735 "tbz x9, #1, 135f\n"
1736 "ldr d13, [x26], #0x8\n"
1737 "ldr d18, [x22], #0x8\n"
1738 "mov x19, #0x48\n"
1739 "ldr d25, [x21], #0x8\n"
1740 "ldr d30, [x20], #0x8\n"
1741 "tbz x9, #0, 144f\n"
1742 "ld1 { v13.s }[2], [x26]\n"
1743 "ld1 { v18.s }[2], [x22]\n"
1744 "ld1 { v25.s }[2], [x21]\n"
1745 "ld1 { v30.s }[2], [x20]\n"
1746 "b 144f\n"
1747 "135:" // Height 4: Partial accumulate: partial_1_16
1748 "mov x19, #0x40\n"
1749 "tbz x9, #0, 144f\n"
1750 "ldr s13, [x26, #0x0]\n"
1751 "ldr s18, [x22, #0x0]\n"
1752 "ldr s25, [x21, #0x0]\n"
1753 "ldr s30, [x20, #0x0]\n"
1754 "b 144f\n"
1755 "136:" // Height 4: Partial accumulate: partial_8_0
1756 "tbz x9, #3, 140f\n"
1757 "ld1 { v9.4s }, [x26], #0x10\n"
1758 "ld1 { v14.4s }, [x22], #0x10\n"
1759 "ld1 { v21.4s }, [x21], #0x10\n"
1760 "ld1 { v26.4s }, [x20], #0x10\n"
1761 "ld1 { v10.4s }, [x26], #0x10\n"
1762 "ld1 { v15.4s }, [x22], #0x10\n"
1763 "ld1 { v22.4s }, [x21], #0x10\n"
1764 "ld1 { v27.4s }, [x20], #0x10\n"
1765 "tbz x9, #2, 138f\n"
1766 "ld1 { v11.4s }, [x26], #0x10\n"
1767 "ld1 { v16.4s }, [x22], #0x10\n"
1768 "ld1 { v23.4s }, [x21], #0x10\n"
1769 "ld1 { v28.4s }, [x20], #0x10\n"
1770 "tbz x9, #1, 137f\n"
1771 "mov x19, #0x38\n"
1772 "ldr d12, [x26], #0x8\n"
1773 "ldr d17, [x22], #0x8\n"
1774 "ldr d24, [x21], #0x8\n"
1775 "ldr d29, [x20], #0x8\n"
1776 "tbz x9, #0, 144f\n"
1777 "ld1 { v12.s }[2], [x26]\n"
1778 "ld1 { v17.s }[2], [x22]\n"
1779 "ld1 { v24.s }[2], [x21]\n"
1780 "ld1 { v29.s }[2], [x20]\n"
1781 "b 144f\n"
1782 "137:" // Height 4: Partial accumulate: partial_1_12
1783 "mov x19, #0x30\n"
1784 "tbz x9, #0, 144f\n"
1785 "ldr s12, [x26, #0x0]\n"
1786 "ldr s17, [x22, #0x0]\n"
1787 "ldr s24, [x21, #0x0]\n"
1788 "ldr s29, [x20, #0x0]\n"
1789 "b 144f\n"
1790 "138:" // Height 4: Partial accumulate: partial_2_8
1791 "tbz x9, #1, 139f\n"
1792 "ldr d11, [x26], #0x8\n"
1793 "ldr d16, [x22], #0x8\n"
1794 "mov x19, #0x28\n"
1795 "ldr d23, [x21], #0x8\n"
1796 "ldr d28, [x20], #0x8\n"
1797 "tbz x9, #0, 144f\n"
1798 "ld1 { v11.s }[2], [x26]\n"
1799 "ld1 { v16.s }[2], [x22]\n"
1800 "ld1 { v23.s }[2], [x21]\n"
1801 "ld1 { v28.s }[2], [x20]\n"
1802 "b 144f\n"
1803 "139:" // Height 4: Partial accumulate: partial_1_8
1804 "mov x19, #0x20\n"
1805 "tbz x9, #0, 144f\n"
1806 "ldr s11, [x26, #0x0]\n"
1807 "ldr s16, [x22, #0x0]\n"
1808 "ldr s23, [x21, #0x0]\n"
1809 "ldr s28, [x20, #0x0]\n"
1810 "b 144f\n"
1811 "140:" // Height 4: Partial accumulate: partial_4_0
1812 "tbz x9, #2, 142f\n"
1813 "ld1 { v9.4s }, [x26], #0x10\n"
1814 "ld1 { v14.4s }, [x22], #0x10\n"
1815 "ld1 { v21.4s }, [x21], #0x10\n"
1816 "ld1 { v26.4s }, [x20], #0x10\n"
1817 "tbz x9, #1, 141f\n"
1818 "mov x19, #0x18\n"
1819 "ldr d10, [x26], #0x8\n"
1820 "ldr d15, [x22], #0x8\n"
1821 "ldr d22, [x21], #0x8\n"
1822 "ldr d27, [x20], #0x8\n"
1823 "tbz x9, #0, 144f\n"
1824 "ld1 { v10.s }[2], [x26]\n"
1825 "ld1 { v15.s }[2], [x22]\n"
1826 "ld1 { v22.s }[2], [x21]\n"
1827 "ld1 { v27.s }[2], [x20]\n"
1828 "b 144f\n"
1829 "141:" // Height 4: Partial accumulate: partial_1_4
1830 "mov x19, #0x10\n"
1831 "tbz x9, #0, 144f\n"
1832 "ldr s10, [x26, #0x0]\n"
1833 "ldr s15, [x22, #0x0]\n"
1834 "ldr s22, [x21, #0x0]\n"
1835 "ldr s27, [x20, #0x0]\n"
1836 "b 144f\n"
1837 "142:" // Height 4: Partial accumulate: partial_2_0
1838 "tbz x9, #1, 143f\n"
1839 "ldr d9, [x26], #0x8\n"
1840 "ldr d14, [x22], #0x8\n"
1841 "mov x19, #0x8\n"
1842 "ldr d21, [x21], #0x8\n"
1843 "ldr d26, [x20], #0x8\n"
1844 "tbz x9, #0, 144f\n"
1845 "ld1 { v9.s }[2], [x26]\n"
1846 "ld1 { v14.s }[2], [x22]\n"
1847 "ld1 { v21.s }[2], [x21]\n"
1848 "ld1 { v26.s }[2], [x20]\n"
1849 "b 144f\n"
1850 "143:" // Height 4: Partial accumulate: partial_1_0
1851 "ldr s9, [x26, #0x0]\n"
1852 "mov x19, #0x0\n"
1853 "ldr s14, [x22, #0x0]\n"
1854 "ldr s21, [x21, #0x0]\n"
1855 "ldr s26, [x20, #0x0]\n"
1856 "144:" // Height 4: Partial accumulate: Done
1857 "sub x26, x26, x19\n"
1858 "b 146f\n"
1859 "145:" // Height 4: full accumulate
1860 "ldr q9, [x26, #0x0]\n"
1861 "ldr q10, [x26, #0x10]\n"
1862 "ldr q11, [x26, #0x20]\n"
1863 "ldr q12, [x26, #0x30]\n"
1864 "ldr q13, [x26, #0x40]\n"
1865 "ldr q20, [x26, #0x50]\n"
1866 "ldr q14, [x22, #0x0]\n"
1867 "ldr q15, [x22, #0x10]\n"
1868 "ldr q16, [x22, #0x20]\n"
1869 "ldr q17, [x22, #0x30]\n"
1870 "ldr q18, [x22, #0x40]\n"
1871 "ldr q19, [x22, #0x50]\n"
1872 "ldr q21, [x21, #0x0]\n"
1873 "ldr q22, [x21, #0x10]\n"
1874 "ldr q23, [x21, #0x20]\n"
1875 "ldr q24, [x21, #0x30]\n"
1876 "ldr q25, [x21, #0x40]\n"
1877 "ldr q4, [x21, #0x50]\n"
1878 "ldr q26, [x20, #0x0]\n"
1879 "ldr q27, [x20, #0x10]\n"
1880 "ldr q28, [x20, #0x20]\n"
1881 "ldr q29, [x20, #0x30]\n"
1882 "ldr q30, [x20, #0x40]\n"
1883 "ldr q31, [x20, #0x50]\n"
1884 "146:" // Height 4: MMLA fixup
1885 "zip1 v8.2d, v9.2d, v14.2d\n"
1886 "zip2 v14.2d, v9.2d, v14.2d\n"
1887 "zip1 v9.2d, v10.2d, v15.2d\n"
1888 "zip2 v15.2d, v10.2d, v15.2d\n"
1889 "zip1 v10.2d, v11.2d, v16.2d\n"
1890 "zip2 v16.2d, v11.2d, v16.2d\n"
1891 "zip1 v11.2d, v12.2d, v17.2d\n"
1892 "zip2 v17.2d, v12.2d, v17.2d\n"
1893 "zip1 v12.2d, v13.2d, v18.2d\n"
1894 "zip2 v18.2d, v13.2d, v18.2d\n"
1895 "zip1 v13.2d, v20.2d, v19.2d\n"
1896 "zip2 v19.2d, v20.2d, v19.2d\n"
1897 "zip1 v20.2d, v21.2d, v26.2d\n"
1898 "zip2 v26.2d, v21.2d, v26.2d\n"
1899 "zip1 v21.2d, v22.2d, v27.2d\n"
1900 "zip2 v27.2d, v22.2d, v27.2d\n"
1901 "zip1 v22.2d, v23.2d, v28.2d\n"
1902 "zip2 v28.2d, v23.2d, v28.2d\n"
1903 "zip1 v23.2d, v24.2d, v29.2d\n"
1904 "zip2 v29.2d, v24.2d, v29.2d\n"
1905 "zip1 v24.2d, v25.2d, v30.2d\n"
1906 "zip2 v30.2d, v25.2d, v30.2d\n"
1907 "zip1 v25.2d, v4.2d, v31.2d\n"
1908 "zip2 v31.2d, v4.2d, v31.2d\n"
1909 "b 148f\n"
1910 "147:" // Height 4: no accumulate
1911 "movi v8.16b, #0x0\n"
1912 "movi v9.16b, #0x0\n"
1913 "movi v10.16b, #0x0\n"
1914 "movi v11.16b, #0x0\n"
1915 "movi v12.16b, #0x0\n"
1916 "movi v13.16b, #0x0\n"
1917 "movi v14.16b, #0x0\n"
1918 "movi v15.16b, #0x0\n"
1919 "movi v16.16b, #0x0\n"
1920 "movi v17.16b, #0x0\n"
1921 "movi v18.16b, #0x0\n"
1922 "movi v19.16b, #0x0\n"
1923 "movi v20.16b, #0x0\n"
1924 "movi v21.16b, #0x0\n"
1925 "movi v22.16b, #0x0\n"
1926 "movi v23.16b, #0x0\n"
1927 "movi v24.16b, #0x0\n"
1928 "movi v25.16b, #0x0\n"
1929 "movi v26.16b, #0x0\n"
1930 "movi v27.16b, #0x0\n"
1931 "movi v28.16b, #0x0\n"
1932 "movi v29.16b, #0x0\n"
1933 "movi v30.16b, #0x0\n"
1934 "movi v31.16b, #0x0\n"
1935 "148:" // Height 4: setup done
1936 "mov x25, #0x0\n"
1937 "149:" // Height 4: String loop
1938 "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
1939 "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n"
1940 "ldr w24, [x20, x25, LSL #0x2]\n"
1941 "tbz %x[flags], #3, 150f\n"
1942 "ldr x20, [%x[input_ptr], x25, LSL #0x3]\n"
1943 "add x20, x20, x19, LSL #3\n"
1944 "ldr x23, [x20, #0x0]\n"
1945 "ldr x22, [x20, #0x8]\n"
1946 "ldr x21, [x20, #0x10]\n"
1947 "ldr x20, [x20, #0x18]\n"
1948 "cbnz x25, 151f\n"
1949 "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
1950 "add x23, x23, x19, LSL #2\n"
1951 "add x22, x22, x19, LSL #2\n"
1952 "add x21, x21, x19, LSL #2\n"
1953 "add x20, x20, x19, LSL #2\n"
1954 "b 151f\n"
1955 "150:" // Height 4: setup direct input
1956 "mov x23, %x[input_ptr]\n"
1957 "add x22, x23, x19, LSL #2\n"
1958 "add x21, x22, x19, LSL #2\n"
1959 "add x20, x21, x19, LSL #2\n"
1960 "151:" // Height 4: input setup done
1961 "cmp x24, #0x4\n"
1962 "blt 154f\n"
1963 "ld1 { v0.4s }, [x23], #0x10\n"
1964 "cmp x24, #0x8\n"
1965 "blt 153f\n"
1966 "152:" // Height 4: Multiply loop: Main loop head
1967 ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n"
1968 "ld1 { v1.4s }, [x22], #0x10\n"
1969 "sub x24, x24, #0x4\n"
1970 ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n"
1971 "ld1 { v2.4s }, [x21], #0x10\n"
1972 "cmp x24, #0x8\n"
1973 ".inst 0x0ea16842 // bfcvtn v2.4h, v2.4s\n"
1974 "ld1 { v3.4s }, [x20], #0x10\n"
1975 "ldr q4, [x28, #0x0]\n"
1976 ".inst 0x4ea16862 // bfcvtn2 v2.8h, v3.4s\n"
1977 "ldr q5, [x28, #0x10]\n"
1978 ".inst 0x6e44ec08 // bfmmla v8.4s, v0.8h, v4.8h\n"
1979 "ldr q6, [x28, #0x20]\n"
1980 "ldr q7, [x28, #0x30]\n"
1981 ".inst 0x6e44ec54 // bfmmla v20.4s, v2.8h, v4.8h\n"
1982 "ldr q4, [x28, #0x40]\n"
1983 "prfm pldl1keep, [x23, #0x80]\n"
1984 ".inst 0x6e45ec0e // bfmmla v14.4s, v0.8h, v5.8h\n"
1985 ".inst 0x6e45ec5a // bfmmla v26.4s, v2.8h, v5.8h\n"
1986 "ldr q5, [x28, #0x50]\n"
1987 ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n"
1988 "prfm pldl1keep, [x22, #0x80]\n"
1989 ".inst 0x6e46ec55 // bfmmla v21.4s, v2.8h, v6.8h\n"
1990 "ldr q6, [x28, #0x60]\n"
1991 ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n"
1992 "prfm pldl1keep, [x21, #0x80]\n"
1993 ".inst 0x6e47ec5b // bfmmla v27.4s, v2.8h, v7.8h\n"
1994 "ldr q7, [x28, #0x70]\n"
1995 ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n"
1996 "prfm pldl1keep, [x20, #0x80]\n"
1997 ".inst 0x6e44ec56 // bfmmla v22.4s, v2.8h, v4.8h\n"
1998 "ldr q4, [x28, #0x80]\n"
1999 ".inst 0x6e45ec10 // bfmmla v16.4s, v0.8h, v5.8h\n"
2000 ".inst 0x6e45ec5c // bfmmla v28.4s, v2.8h, v5.8h\n"
2001 "ldr q5, [x28, #0x90]\n"
2002 ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n"
2003 ".inst 0x6e46ec57 // bfmmla v23.4s, v2.8h, v6.8h\n"
2004 "ldr q6, [x28, #0xa0]\n"
2005 ".inst 0x6e47ec11 // bfmmla v17.4s, v0.8h, v7.8h\n"
2006 ".inst 0x6e47ec5d // bfmmla v29.4s, v2.8h, v7.8h\n"
2007 "ldr q7, [x28, #0xb0]\n"
2008 "add x28, x28, #0xc0\n"
2009 ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n"
2010 ".inst 0x6e44ec58 // bfmmla v24.4s, v2.8h, v4.8h\n"
2011 ".inst 0x6e45ec12 // bfmmla v18.4s, v0.8h, v5.8h\n"
2012 ".inst 0x6e45ec5e // bfmmla v30.4s, v2.8h, v5.8h\n"
2013 ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n"
2014 ".inst 0x6e46ec59 // bfmmla v25.4s, v2.8h, v6.8h\n"
2015 ".inst 0x6e47ec13 // bfmmla v19.4s, v0.8h, v7.8h\n"
2016 "ld1 { v0.4s }, [x23], #0x10\n"
2017 ".inst 0x6e47ec5f // bfmmla v31.4s, v2.8h, v7.8h\n"
2018 "bge 152b\n"
2019 "153:" // Height 4: Multiply loop: Single iteration only
2020 ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n"
2021 "ld1 { v1.4s }, [x22], #0x10\n"
2022 "sub x24, x24, #0x4\n"
2023 ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n"
2024 "ld1 { v2.4s }, [x21], #0x10\n"
2025 "ld1 { v3.4s }, [x20], #0x10\n"
2026 ".inst 0x0ea16842 // bfcvtn v2.4h, v2.4s\n"
2027 "ldr q4, [x28, #0x0]\n"
2028 "ldr q5, [x28, #0x10]\n"
2029 ".inst 0x4ea16862 // bfcvtn2 v2.8h, v3.4s\n"
2030 "ldr q6, [x28, #0x20]\n"
2031 "ldr q7, [x28, #0x30]\n"
2032 ".inst 0x6e44ec08 // bfmmla v8.4s, v0.8h, v4.8h\n"
2033 ".inst 0x6e45ec0e // bfmmla v14.4s, v0.8h, v5.8h\n"
2034 "prfm pldl1keep, [x23, #0x80]\n"
2035 "prfm pldl1keep, [x22, #0x80]\n"
2036 ".inst 0x6e44ec54 // bfmmla v20.4s, v2.8h, v4.8h\n"
2037 ".inst 0x6e45ec5a // bfmmla v26.4s, v2.8h, v5.8h\n"
2038 "ldr q4, [x28, #0x40]\n"
2039 ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n"
2040 "ldr q5, [x28, #0x50]\n"
2041 ".inst 0x6e46ec55 // bfmmla v21.4s, v2.8h, v6.8h\n"
2042 "ldr q6, [x28, #0x60]\n"
2043 ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n"
2044 "prfm pldl1keep, [x21, #0x80]\n"
2045 ".inst 0x6e47ec5b // bfmmla v27.4s, v2.8h, v7.8h\n"
2046 "ldr q7, [x28, #0x70]\n"
2047 ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n"
2048 "prfm pldl1keep, [x20, #0x80]\n"
2049 ".inst 0x6e44ec56 // bfmmla v22.4s, v2.8h, v4.8h\n"
2050 "ldr q4, [x28, #0x80]\n"
2051 ".inst 0x6e45ec10 // bfmmla v16.4s, v0.8h, v5.8h\n"
2052 ".inst 0x6e45ec5c // bfmmla v28.4s, v2.8h, v5.8h\n"
2053 "ldr q5, [x28, #0x90]\n"
2054 ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n"
2055 ".inst 0x6e46ec57 // bfmmla v23.4s, v2.8h, v6.8h\n"
2056 "ldr q6, [x28, #0xa0]\n"
2057 ".inst 0x6e47ec11 // bfmmla v17.4s, v0.8h, v7.8h\n"
2058 ".inst 0x6e47ec5d // bfmmla v29.4s, v2.8h, v7.8h\n"
2059 "ldr q7, [x28, #0xb0]\n"
2060 "add x28, x28, #0xc0\n"
2061 ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n"
2062 ".inst 0x6e44ec58 // bfmmla v24.4s, v2.8h, v4.8h\n"
2063 ".inst 0x6e45ec12 // bfmmla v18.4s, v0.8h, v5.8h\n"
2064 ".inst 0x6e45ec5e // bfmmla v30.4s, v2.8h, v5.8h\n"
2065 ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n"
2066 ".inst 0x6e46ec59 // bfmmla v25.4s, v2.8h, v6.8h\n"
2067 ".inst 0x6e47ec13 // bfmmla v19.4s, v0.8h, v7.8h\n"
2068 ".inst 0x6e47ec5f // bfmmla v31.4s, v2.8h, v7.8h\n"
2069 "154:" // Height 4: Multiply loop: Main loop skip
2070 "cbz x24, 157f\n"
2071 "cbz x24, 157f\n"
2072 "tbz x24, #1, 155f\n"
2073 "ldr d0, [x23], #0x8\n"
2074 "ldr d1, [x22], #0x8\n"
2075 "ldr d2, [x21], #0x8\n"
2076 "ldr d3, [x20], #0x8\n"
2077 "tbz x24, #0, 156f\n"
2078 "ld1 { v0.s }[2], [x23]\n"
2079 "ld1 { v1.s }[2], [x22]\n"
2080 "ld1 { v2.s }[2], [x21]\n"
2081 "ld1 { v3.s }[2], [x20]\n"
2082 "b 156f\n"
2083 "155:" // Height 4: Multiply loop: Ragged operand read: partial_1_0
2084 "ldr s0, [x23, #0x0]\n"
2085 "ldr s1, [x22, #0x0]\n"
2086 "ldr s2, [x21, #0x0]\n"
2087 "ldr s3, [x20, #0x0]\n"
2088 "156:" // Height 4: Multiply loop: Ragged operand read: Done
2089 ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n"
2090 "ldr q4, [x28, #0x0]\n"
2091 ".inst 0x0ea16842 // bfcvtn v2.4h, v2.4s\n"
2092 "ldr q5, [x28, #0x10]\n"
2093 "ldr q6, [x28, #0x20]\n"
2094 ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n"
2095 "ldr q7, [x28, #0x30]\n"
2096 ".inst 0x4ea16862 // bfcvtn2 v2.8h, v3.4s\n"
2097 ".inst 0x6e44ec08 // bfmmla v8.4s, v0.8h, v4.8h\n"
2098 ".inst 0x6e44ec54 // bfmmla v20.4s, v2.8h, v4.8h\n"
2099 "ldr q4, [x28, #0x40]\n"
2100 ".inst 0x6e45ec0e // bfmmla v14.4s, v0.8h, v5.8h\n"
2101 ".inst 0x6e45ec5a // bfmmla v26.4s, v2.8h, v5.8h\n"
2102 "ldr q5, [x28, #0x50]\n"
2103 ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n"
2104 ".inst 0x6e46ec55 // bfmmla v21.4s, v2.8h, v6.8h\n"
2105 "ldr q6, [x28, #0x60]\n"
2106 ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n"
2107 ".inst 0x6e47ec5b // bfmmla v27.4s, v2.8h, v7.8h\n"
2108 "ldr q7, [x28, #0x70]\n"
2109 ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n"
2110 ".inst 0x6e44ec56 // bfmmla v22.4s, v2.8h, v4.8h\n"
2111 "ldr q4, [x28, #0x80]\n"
2112 ".inst 0x6e45ec10 // bfmmla v16.4s, v0.8h, v5.8h\n"
2113 ".inst 0x6e45ec5c // bfmmla v28.4s, v2.8h, v5.8h\n"
2114 "ldr q5, [x28, #0x90]\n"
2115 ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n"
2116 ".inst 0x6e46ec57 // bfmmla v23.4s, v2.8h, v6.8h\n"
2117 "ldr q6, [x28, #0xa0]\n"
2118 ".inst 0x6e47ec11 // bfmmla v17.4s, v0.8h, v7.8h\n"
2119 ".inst 0x6e47ec5d // bfmmla v29.4s, v2.8h, v7.8h\n"
2120 "ldr q7, [x28, #0xb0]\n"
2121 "add x28, x28, #0xc0\n"
2122 ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n"
2123 ".inst 0x6e44ec58 // bfmmla v24.4s, v2.8h, v4.8h\n"
2124 ".inst 0x6e45ec12 // bfmmla v18.4s, v0.8h, v5.8h\n"
2125 ".inst 0x6e45ec5e // bfmmla v30.4s, v2.8h, v5.8h\n"
2126 ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n"
2127 ".inst 0x6e46ec59 // bfmmla v25.4s, v2.8h, v6.8h\n"
2128 ".inst 0x6e47ec13 // bfmmla v19.4s, v0.8h, v7.8h\n"
2129 ".inst 0x6e47ec5f // bfmmla v31.4s, v2.8h, v7.8h\n"
2130 "157:" // Height 4: Multiply loop: No odd multiplies
2131 "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n"
2132 "add x25, x25, #0x1\n"
2133 "cmp x25, x19\n"
2134 "bne 149b\n"
2135 "uzp1 v4.2d, v8.2d, v14.2d\n"
2136 "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
2137 "uzp2 v8.2d, v8.2d, v14.2d\n"
2138 "prfm pstl1keep, [x26, #0x0]\n"
2139 "add x22, x26, x19, LSL #2\n"
2140 "uzp1 v14.2d, v9.2d, v15.2d\n"
2141 "prfm pstl1keep, [x22, #0x0]\n"
2142 "uzp2 v9.2d, v9.2d, v15.2d\n"
2143 "add x21, x22, x19, LSL #2\n"
2144 "uzp1 v15.2d, v10.2d, v16.2d\n"
2145 "prfm pstl1keep, [x21, #0x0]\n"
2146 "add x20, x21, x19, LSL #2\n"
2147 "uzp2 v10.2d, v10.2d, v16.2d\n"
2148 "prfm pstl1keep, [x20, #0x0]\n"
2149 "uzp1 v16.2d, v11.2d, v17.2d\n"
2150 "uzp2 v11.2d, v11.2d, v17.2d\n"
2151 "uzp1 v17.2d, v12.2d, v18.2d\n"
2152 "uzp2 v12.2d, v12.2d, v18.2d\n"
2153 "uzp1 v18.2d, v13.2d, v19.2d\n"
2154 "uzp2 v13.2d, v13.2d, v19.2d\n"
2155 "uzp1 v19.2d, v20.2d, v26.2d\n"
2156 "uzp2 v20.2d, v20.2d, v26.2d\n"
2157 "uzp1 v26.2d, v21.2d, v27.2d\n"
2158 "uzp2 v21.2d, v21.2d, v27.2d\n"
2159 "uzp1 v27.2d, v22.2d, v28.2d\n"
2160 "uzp2 v22.2d, v22.2d, v28.2d\n"
2161 "uzp1 v28.2d, v23.2d, v29.2d\n"
2162 "uzp2 v23.2d, v23.2d, v29.2d\n"
2163 "uzp1 v29.2d, v24.2d, v30.2d\n"
2164 "uzp2 v24.2d, v24.2d, v30.2d\n"
2165 "uzp1 v30.2d, v25.2d, v31.2d\n"
2166 "uzp2 v25.2d, v25.2d, v31.2d\n"
2167 "tbz %x[flags], #1, 158f\n"
2168 "add x19, %x[args_ptr], %[offset_min]\n"
2169 "ld1r { v1.4s }, [x19]\n"
2170 "add x19, %x[args_ptr], %[offset_max]\n"
2171 "ld1r { v0.4s }, [x19]\n"
2172 "fmin v4.4s, v4.4s, v0.4s\n"
2173 "fmin v14.4s, v14.4s, v0.4s\n"
2174 "fmin v15.4s, v15.4s, v0.4s\n"
2175 "fmin v16.4s, v16.4s, v0.4s\n"
2176 "fmax v4.4s, v4.4s, v1.4s\n"
2177 "fmax v14.4s, v14.4s, v1.4s\n"
2178 "fmax v15.4s, v15.4s, v1.4s\n"
2179 "fmax v16.4s, v16.4s, v1.4s\n"
2180 "fmin v17.4s, v17.4s, v0.4s\n"
2181 "fmin v18.4s, v18.4s, v0.4s\n"
2182 "fmin v8.4s, v8.4s, v0.4s\n"
2183 "fmax v17.4s, v17.4s, v1.4s\n"
2184 "fmax v18.4s, v18.4s, v1.4s\n"
2185 "fmax v8.4s, v8.4s, v1.4s\n"
2186 "fmin v9.4s, v9.4s, v0.4s\n"
2187 "fmin v10.4s, v10.4s, v0.4s\n"
2188 "fmin v11.4s, v11.4s, v0.4s\n"
2189 "fmax v9.4s, v9.4s, v1.4s\n"
2190 "fmax v10.4s, v10.4s, v1.4s\n"
2191 "fmax v11.4s, v11.4s, v1.4s\n"
2192 "fmin v12.4s, v12.4s, v0.4s\n"
2193 "fmin v13.4s, v13.4s, v0.4s\n"
2194 "fmin v19.4s, v19.4s, v0.4s\n"
2195 "fmax v12.4s, v12.4s, v1.4s\n"
2196 "fmax v13.4s, v13.4s, v1.4s\n"
2197 "fmax v19.4s, v19.4s, v1.4s\n"
2198 "fmin v26.4s, v26.4s, v0.4s\n"
2199 "fmin v27.4s, v27.4s, v0.4s\n"
2200 "fmin v28.4s, v28.4s, v0.4s\n"
2201 "fmax v26.4s, v26.4s, v1.4s\n"
2202 "fmax v27.4s, v27.4s, v1.4s\n"
2203 "fmax v28.4s, v28.4s, v1.4s\n"
2204 "fmin v29.4s, v29.4s, v0.4s\n"
2205 "fmin v30.4s, v30.4s, v0.4s\n"
2206 "fmin v20.4s, v20.4s, v0.4s\n"
2207 "fmax v29.4s, v29.4s, v1.4s\n"
2208 "fmax v30.4s, v30.4s, v1.4s\n"
2209 "fmax v20.4s, v20.4s, v1.4s\n"
2210 "fmin v21.4s, v21.4s, v0.4s\n"
2211 "fmin v22.4s, v22.4s, v0.4s\n"
2212 "fmin v23.4s, v23.4s, v0.4s\n"
2213 "fmax v21.4s, v21.4s, v1.4s\n"
2214 "fmax v22.4s, v22.4s, v1.4s\n"
2215 "fmax v23.4s, v23.4s, v1.4s\n"
2216 "fmin v24.4s, v24.4s, v0.4s\n"
2217 "fmin v25.4s, v25.4s, v0.4s\n"
2218 "fmax v24.4s, v24.4s, v1.4s\n"
2219 "fmax v25.4s, v25.4s, v1.4s\n"
2220 "158:" // Height 4: No activation
2221 "cmp x9, #0x18\n"
2222 "bge 171f\n"
2223 "tbz x9, #4, 162f\n"
2224 "st1 { v4.4s }, [x26], #0x10\n"
2225 "st1 { v14.4s }, [x26], #0x10\n"
2226 "st1 { v15.4s }, [x26], #0x10\n"
2227 "st1 { v16.4s }, [x26], #0x10\n"
2228 "st1 { v8.4s }, [x22], #0x10\n"
2229 "st1 { v9.4s }, [x22], #0x10\n"
2230 "st1 { v10.4s }, [x22], #0x10\n"
2231 "st1 { v11.4s }, [x22], #0x10\n"
2232 "st1 { v19.4s }, [x21], #0x10\n"
2233 "st1 { v26.4s }, [x21], #0x10\n"
2234 "st1 { v27.4s }, [x21], #0x10\n"
2235 "st1 { v28.4s }, [x21], #0x10\n"
2236 "st1 { v20.4s }, [x20], #0x10\n"
2237 "st1 { v21.4s }, [x20], #0x10\n"
2238 "st1 { v22.4s }, [x20], #0x10\n"
2239 "st1 { v23.4s }, [x20], #0x10\n"
2240 "tbz x9, #2, 160f\n"
2241 "st1 { v17.4s }, [x26], #0x10\n"
2242 "st1 { v12.4s }, [x22], #0x10\n"
2243 "st1 { v29.4s }, [x21], #0x10\n"
2244 "st1 { v24.4s }, [x20], #0x10\n"
2245 "tbz x9, #1, 159f\n"
2246 "str d18, [x26], #0x8\n"
2247 "str d13, [x22], #0x8\n"
2248 "str d30, [x21], #0x8\n"
2249 "str d25, [x20], #0x8\n"
2250 "tbz x9, #0, 170f\n"
2251 "st1 { v18.s }[2], [x26]\n"
2252 "st1 { v13.s }[2], [x22]\n"
2253 "st1 { v30.s }[2], [x21]\n"
2254 "st1 { v25.s }[2], [x20]\n"
2255 "b 170f\n"
2256 "159:" // Height 4: Partial direct writeback: partial_1_20
2257 "tbz x9, #0, 170f\n"
2258 "str s18, [x26, #0x0]\n"
2259 "str s13, [x22, #0x0]\n"
2260 "str s30, [x21, #0x0]\n"
2261 "str s25, [x20, #0x0]\n"
2262 "b 170f\n"
2263 "160:" // Height 4: Partial direct writeback: partial_2_16
2264 "tbz x9, #1, 161f\n"
2265 "str d17, [x26], #0x8\n"
2266 "str d12, [x22], #0x8\n"
2267 "str d29, [x21], #0x8\n"
2268 "str d24, [x20], #0x8\n"
2269 "tbz x9, #0, 170f\n"
2270 "st1 { v17.s }[2], [x26]\n"
2271 "st1 { v12.s }[2], [x22]\n"
2272 "st1 { v29.s }[2], [x21]\n"
2273 "st1 { v24.s }[2], [x20]\n"
2274 "b 170f\n"
2275 "161:" // Height 4: Partial direct writeback: partial_1_16
2276 "tbz x9, #0, 170f\n"
2277 "str s17, [x26, #0x0]\n"
2278 "str s12, [x22, #0x0]\n"
2279 "str s29, [x21, #0x0]\n"
2280 "str s24, [x20, #0x0]\n"
2281 "b 170f\n"
2282 "162:" // Height 4: Partial direct writeback: partial_8_0
2283 "tbz x9, #3, 166f\n"
2284 "st1 { v4.4s }, [x26], #0x10\n"
2285 "st1 { v14.4s }, [x26], #0x10\n"
2286 "st1 { v8.4s }, [x22], #0x10\n"
2287 "st1 { v9.4s }, [x22], #0x10\n"
2288 "st1 { v19.4s }, [x21], #0x10\n"
2289 "st1 { v26.4s }, [x21], #0x10\n"
2290 "st1 { v20.4s }, [x20], #0x10\n"
2291 "st1 { v21.4s }, [x20], #0x10\n"
2292 "tbz x9, #2, 164f\n"
2293 "st1 { v15.4s }, [x26], #0x10\n"
2294 "st1 { v10.4s }, [x22], #0x10\n"
2295 "st1 { v27.4s }, [x21], #0x10\n"
2296 "st1 { v22.4s }, [x20], #0x10\n"
2297 "tbz x9, #1, 163f\n"
2298 "str d16, [x26], #0x8\n"
2299 "str d11, [x22], #0x8\n"
2300 "str d28, [x21], #0x8\n"
2301 "str d23, [x20], #0x8\n"
2302 "tbz x9, #0, 170f\n"
2303 "st1 { v16.s }[2], [x26]\n"
2304 "st1 { v11.s }[2], [x22]\n"
2305 "st1 { v28.s }[2], [x21]\n"
2306 "st1 { v23.s }[2], [x20]\n"
2307 "b 170f\n"
2308 "163:" // Height 4: Partial direct writeback: partial_1_12
2309 "tbz x9, #0, 170f\n"
2310 "str s16, [x26, #0x0]\n"
2311 "str s11, [x22, #0x0]\n"
2312 "str s28, [x21, #0x0]\n"
2313 "str s23, [x20, #0x0]\n"
2314 "b 170f\n"
2315 "164:" // Height 4: Partial direct writeback: partial_2_8
2316 "tbz x9, #1, 165f\n"
2317 "str d15, [x26], #0x8\n"
2318 "str d10, [x22], #0x8\n"
2319 "str d27, [x21], #0x8\n"
2320 "str d22, [x20], #0x8\n"
2321 "tbz x9, #0, 170f\n"
2322 "st1 { v15.s }[2], [x26]\n"
2323 "st1 { v10.s }[2], [x22]\n"
2324 "st1 { v27.s }[2], [x21]\n"
2325 "st1 { v22.s }[2], [x20]\n"
2326 "b 170f\n"
2327 "165:" // Height 4: Partial direct writeback: partial_1_8
2328 "tbz x9, #0, 170f\n"
2329 "str s15, [x26, #0x0]\n"
2330 "str s10, [x22, #0x0]\n"
2331 "str s27, [x21, #0x0]\n"
2332 "str s22, [x20, #0x0]\n"
2333 "b 170f\n"
2334 "166:" // Height 4: Partial direct writeback: partial_4_0
2335 "tbz x9, #2, 168f\n"
2336 "st1 { v4.4s }, [x26], #0x10\n"
2337 "st1 { v8.4s }, [x22], #0x10\n"
2338 "st1 { v19.4s }, [x21], #0x10\n"
2339 "st1 { v20.4s }, [x20], #0x10\n"
2340 "tbz x9, #1, 167f\n"
2341 "str d14, [x26], #0x8\n"
2342 "str d9, [x22], #0x8\n"
2343 "str d26, [x21], #0x8\n"
2344 "str d21, [x20], #0x8\n"
2345 "tbz x9, #0, 170f\n"
2346 "st1 { v14.s }[2], [x26]\n"
2347 "st1 { v9.s }[2], [x22]\n"
2348 "st1 { v26.s }[2], [x21]\n"
2349 "st1 { v21.s }[2], [x20]\n"
2350 "b 170f\n"
2351 "167:" // Height 4: Partial direct writeback: partial_1_4
2352 "tbz x9, #0, 170f\n"
2353 "str s14, [x26, #0x0]\n"
2354 "str s9, [x22, #0x0]\n"
2355 "str s26, [x21, #0x0]\n"
2356 "str s21, [x20, #0x0]\n"
2357 "b 170f\n"
2358 "168:" // Height 4: Partial direct writeback: partial_2_0
2359 "tbz x9, #1, 169f\n"
2360 "str d4, [x26], #0x8\n"
2361 "str d8, [x22], #0x8\n"
2362 "str d19, [x21], #0x8\n"
2363 "str d20, [x20], #0x8\n"
2364 "tbz x9, #0, 170f\n"
2365 "st1 { v4.s }[2], [x26]\n"
2366 "st1 { v8.s }[2], [x22]\n"
2367 "st1 { v19.s }[2], [x21]\n"
2368 "st1 { v20.s }[2], [x20]\n"
2369 "b 170f\n"
2370 "169:" // Height 4: Partial direct writeback: partial_1_0
2371 "str s4, [x26, #0x0]\n"
2372 "str s8, [x22, #0x0]\n"
2373 "str s19, [x21, #0x0]\n"
2374 "str s20, [x20, #0x0]\n"
2375 "170:" // Height 4: Partial direct writeback: Done
2376 "b 172f\n"
2377 "171:" // Height 4: Full writeback
2378 "str q4, [x26, #0x0]\n"
2379 "str q14, [x26, #0x10]\n"
2380 "str q15, [x26, #0x20]\n"
2381 "str q16, [x26, #0x30]\n"
2382 "str q17, [x26, #0x40]\n"
2383 "str q18, [x26, #0x50]\n"
2384 "add x26, x26, #0x60\n"
2385 "str q8, [x22, #0x0]\n"
2386 "str q9, [x22, #0x10]\n"
2387 "str q10, [x22, #0x20]\n"
2388 "str q11, [x22, #0x30]\n"
2389 "str q12, [x22, #0x40]\n"
2390 "str q13, [x22, #0x50]\n"
2391 "str q19, [x21, #0x0]\n"
2392 "str q26, [x21, #0x10]\n"
2393 "str q27, [x21, #0x20]\n"
2394 "str q28, [x21, #0x30]\n"
2395 "str q29, [x21, #0x40]\n"
2396 "str q30, [x21, #0x50]\n"
2397 "str q20, [x20, #0x0]\n"
2398 "str q21, [x20, #0x10]\n"
2399 "str q22, [x20, #0x20]\n"
2400 "str q23, [x20, #0x30]\n"
2401 "str q24, [x20, #0x40]\n"
2402 "str q25, [x20, #0x50]\n"
2403 "172:" // Height 4: Writeback done
2404 "subs x9, x9, #0x18\n"
2405 "bgt 131b\n"
2406 "subs %x[M], %x[M], #0x4\n"
2407 "beq 174f\n"
2408 "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n"
2409 "tbz %x[flags], #3, 173f\n"
2410 "add x20, x20, #0x4\n"
2411 "str x20, [%x[args_ptr], %[offsetof_input_offset]]\n"
2412 "b 1b\n"
2413 "173:" // Update direct input
2414 "mov x19, #0x10\n"
2415 "madd %x[input_ptr], x19, x20, %x[input_ptr]\n"
2416 "b 1b\n"
2417 "174:" // Exit
2418
2419 : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
2420 : [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))
2421 : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
2422 );
2423 }
2424
2425 } // namespace arm_gemm
2426 #endif // __aarch64__
2427