1 /*
2 * Copyright (c) 2022 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 ARM_COMPUTE_ENABLE_SVE
25
26 #include "arm_gemm.hpp"
27 #include "../../utils.hpp"
28
29 #include <cassert>
30 #include <limits>
31
32 namespace arm_gemm {
33
sve_ffhybrid_fp32_mla_6x4VL_a64fx(unsigned int num_strings,const unsigned int * string_lengths,IndirectInputArg<float> A_arg,size_t M,size_t N,const float * B_ptr,size_t B_stride,IndirectOutputArg<float> output_arg,const float * bias,Activation act,bool accumulate)34 void sve_ffhybrid_fp32_mla_6x4VL_a64fx (
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, size_t B_stride, 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 const float *cur_B_ptr = {};
48 size_t B_stride = {};
49 size_t output_offset = {};
50 size_t input_initial_col = {};
51 size_t input_offset = {};
52 } ka;
53
54 unsigned long flags=0;
55 void *output_ptr;
56 void *input_ptr;
57
58 if (output_arg.is_indirect) {
59 output_ptr=(void *)(output_arg.indirect.ptr);
60 ka.output_offset=output_arg.indirect.offset;
61 flags |= 0x4;
62 } else {
63 output_ptr=(void *)(output_arg.direct.base);
64 ka.output_offset=output_arg.direct.stride;
65 }
66
67 if (A_arg.is_indirect) {
68 input_ptr=(void *)(A_arg.indirect.ptr);
69 ka.input_offset=A_arg.indirect.start_row;
70 ka.input_initial_col=A_arg.indirect.start_col;
71 flags |= 0x8;
72 } else {
73 assert(num_strings==1);
74 input_ptr=(void *)(A_arg.direct.base);
75 ka.input_offset=A_arg.direct.stride;
76 }
77 if (accumulate) {
78 flags |= 0x1;
79 }
80 ka.num_strings = num_strings;
81 ka.string_lengths = string_lengths;
82 ka.N = N;
83 ka.B_ptr = B_ptr;
84 ka.B_stride = B_stride;
85 switch(act.type) {
86 default:
87 case Activation::Type::None:
88 break;
89 case Activation::Type::BoundedReLU:
90 ka.maxval = static_cast<float>(act.param1);
91 /* fall through */
92 case Activation::Type::ReLU:
93 ka.minval = 0;
94 flags |= 0x2;
95 break;
96 }
97 __asm__ __volatile__(
98 "ptrue p4.b\n"
99 "1:" // Row loop
100 "cmp %x[M], #0x6\n"
101 "bge 66f\n"
102 "cmp %x[M], #0x4\n"
103 "bgt 53f\n"
104 "beq 40f\n"
105 "cmp %x[M], #0x2\n"
106 "bgt 27f\n"
107 "beq 14f\n"
108 "ldr x19, [%x[args_ptr], %[offsetof_B_ptr]]\n"
109 "mov x14, %x[bias]\n"
110 "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
111 "str x19, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
112 "mov x12, %x[output_ptr]\n"
113 "2:" // Height 1: Column loop
114 "ldr x11, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
115 "ldr x19, [%x[args_ptr], %[offsetof_B_stride]]\n"
116 "add x10, x11, x19, LSL #2\n"
117 "cntw x20, ALL, MUL #3\n"
118 "add x9, x10, x19, LSL #2\n"
119 "add x28, x9, x19, LSL #2\n"
120 "add x19, x28, x19, LSL #2\n"
121 "cmp x13, x20\n"
122 "str x19, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
123 "bgt 3f\n"
124 "decw x20\n"
125 "cmp x13, x20\n"
126 "mov x28, x11\n"
127 "bgt 3f\n"
128 "decw x20\n"
129 "cmp x13, x20\n"
130 "mov x9, x11\n"
131 "bgt 3f\n"
132 "mov x10, x11\n"
133 "3:" // Height 1: B setup done
134 "mov x19, #0x0\n"
135 "whilelt p3.s, x19, x13\n"
136 "incw x19\n"
137 "whilelt p2.s, x19, x13\n"
138 "incw x19\n"
139 "whilelt p1.s, x19, x13\n"
140 "incw x19\n"
141 "whilelt p0.s, x19, x13\n"
142 "cbz x14, 4f\n"
143 "ld1w { z8.s }, p4/Z, [x14]\n"
144 "ld1w { z9.s }, p4/Z, [x14, #1, MUL VL]\n"
145 "ld1w { z10.s }, p4/Z, [x14, #2, MUL VL]\n"
146 "ld1w { z11.s }, p4/Z, [x14, #3, MUL VL]\n"
147 "addvl x14, x14, #4\n"
148 "b 6f\n"
149 "4:" // Height 1: no bias
150 "tbz %x[flags], #0, 5f\n"
151 "ld1w { z8.s }, p3/Z, [x12]\n"
152 "ld1w { z9.s }, p2/Z, [x12, #1, MUL VL]\n"
153 "ld1w { z10.s }, p1/Z, [x12, #2, MUL VL]\n"
154 "ld1w { z11.s }, p0/Z, [x12, #3, MUL VL]\n"
155 "b 6f\n"
156 "5:" // Height 1: no accumulate
157 "mov z8.b, #0x0\n"
158 "mov z9.b, #0x0\n"
159 "mov z10.b, #0x0\n"
160 "mov z11.b, #0x0\n"
161 "6:" // Height 1: setup done
162 "mov x27, #0x0\n"
163 "7:" // Height 1: String loop
164 "ldr x19, [%x[args_ptr], %[offsetof_string_lengths]]\n"
165 "ldr w26, [x19, x27, LSL #0x2]\n"
166 "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n"
167 "tbz %x[flags], #3, 8f\n"
168 "ldr x20, [%x[input_ptr], x27, LSL #0x3]\n"
169 "add x20, x20, x19, LSL #3\n"
170 "ldr x25, [x20, #0x0]\n"
171 "cbnz x27, 9f\n"
172 "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
173 "add x25, x25, x19, LSL #2\n"
174 "b 9f\n"
175 "8:" // Height 1: setup direct input
176 "mov x25, %x[input_ptr]\n"
177 "9:" // Height 1: input setup done
178 "subs x26, x26, #0x1\n"
179 "ld1rw { z0.s }, p4/Z, [x25]\n"
180 "ld1w { z6.s }, p4/Z, [x11]\n"
181 "ld1w { z7.s }, p4/Z, [x10]\n"
182 "ble 11f\n"
183 "10:" // Height 1: Multiply loop: Main loop
184 "fmla z8.s, p4/M, z6.s, z0.s\n"
185 "fmla z9.s, p4/M, z7.s, z0.s\n"
186 "ld1w { z6.s }, p4/Z, [x9]\n"
187 "ld1w { z7.s }, p4/Z, [x28]\n"
188 "addvl x11, x11, #1\n"
189 "addvl x10, x10, #1\n"
190 "fmla z10.s, p4/M, z6.s, z0.s\n"
191 "fmla z11.s, p4/M, z7.s, z0.s\n"
192 "add x25, x25, #0x4\n"
193 "subs x26, x26, #0x1\n"
194 "ld1rw { z0.s }, p4/Z, [x25]\n"
195 "ld1w { z6.s }, p4/Z, [x11]\n"
196 "addvl x9, x9, #1\n"
197 "addvl x28, x28, #1\n"
198 "ld1w { z7.s }, p4/Z, [x10]\n"
199 "bgt 10b\n"
200 "11:" // Height 1: Multiply loop: Main loop skip
201 "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n"
202 "fmla z8.s, p4/M, z6.s, z0.s\n"
203 "fmla z9.s, p4/M, z7.s, z0.s\n"
204 "ld1w { z6.s }, p4/Z, [x9]\n"
205 "ld1w { z7.s }, p4/Z, [x28]\n"
206 "add x27, x27, #0x1\n"
207 "cmp x27, x19\n"
208 "fmla z10.s, p4/M, z6.s, z0.s\n"
209 "fmla z11.s, p4/M, z7.s, z0.s\n"
210 "addvl x11, x11, #1\n"
211 "addvl x10, x10, #1\n"
212 "addvl x9, x9, #1\n"
213 "addvl x28, x28, #1\n"
214 "bne 7b\n"
215 "tbz %x[flags], #1, 12f\n"
216 "add x19, %x[args_ptr], %[offset_max]\n"
217 "ld1rw { z1.s }, p4/Z, [x19]\n"
218 "add x19, %x[args_ptr], %[offset_min]\n"
219 "ld1rw { z0.s }, p4/Z, [x19]\n"
220 "fmin z8.s, p4/M, z8.s, z1.s\n"
221 "fmin z9.s, p4/M, z9.s, z1.s\n"
222 "fmin z10.s, p4/M, z10.s, z1.s\n"
223 "fmin z11.s, p4/M, z11.s, z1.s\n"
224 "fmax z8.s, p4/M, z8.s, z0.s\n"
225 "fmax z9.s, p4/M, z9.s, z0.s\n"
226 "fmax z10.s, p4/M, z10.s, z0.s\n"
227 "fmax z11.s, p4/M, z11.s, z0.s\n"
228 "12:" // Height 1: No activation
229 "st1w { z8.s }, p3, [x12]\n"
230 "st1w { z9.s }, p2, [x12, #1, MUL VL]\n"
231 "st1w { z10.s }, p1, [x12, #2, MUL VL]\n"
232 "st1w { z11.s }, p0, [x12, #3, MUL VL]\n"
233 "addvl x12, x12, #4\n"
234 "13:" // Height 1: Writeback done
235 "decw x13, ALL, MUL #4\n"
236 "cmp x13, XZR\n"
237 "bgt 2b\n"
238 "b 80f\n"
239 "14:" // Height 2
240 "ldr x19, [%x[args_ptr], %[offsetof_B_ptr]]\n"
241 "mov x14, %x[bias]\n"
242 "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
243 "str x19, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
244 "mov x12, %x[output_ptr]\n"
245 "15:" // Height 2: Column loop
246 "ldr x11, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
247 "ldr x19, [%x[args_ptr], %[offsetof_B_stride]]\n"
248 "add x10, x11, x19, LSL #2\n"
249 "cntw x20, ALL, MUL #3\n"
250 "add x9, x10, x19, LSL #2\n"
251 "add x28, x9, x19, LSL #2\n"
252 "add x19, x28, x19, LSL #2\n"
253 "cmp x13, x20\n"
254 "str x19, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
255 "bgt 16f\n"
256 "decw x20\n"
257 "cmp x13, x20\n"
258 "mov x28, x11\n"
259 "bgt 16f\n"
260 "decw x20\n"
261 "cmp x13, x20\n"
262 "mov x9, x11\n"
263 "bgt 16f\n"
264 "mov x10, x11\n"
265 "16:" // Height 2: B setup done
266 "mov x19, #0x0\n"
267 "whilelt p3.s, x19, x13\n"
268 "incw x19\n"
269 "whilelt p2.s, x19, x13\n"
270 "incw x19\n"
271 "whilelt p1.s, x19, x13\n"
272 "incw x19\n"
273 "whilelt p0.s, x19, x13\n"
274 "cbz x14, 17f\n"
275 "ld1w { z8.s }, p4/Z, [x14]\n"
276 "ld1w { z9.s }, p4/Z, [x14, #1, MUL VL]\n"
277 "mov z12.d, z8.d\n"
278 "mov z13.d, z9.d\n"
279 "ld1w { z10.s }, p4/Z, [x14, #2, MUL VL]\n"
280 "ld1w { z11.s }, p4/Z, [x14, #3, MUL VL]\n"
281 "mov z14.d, z10.d\n"
282 "mov z15.d, z11.d\n"
283 "addvl x14, x14, #4\n"
284 "b 19f\n"
285 "17:" // Height 2: no bias
286 "tbz %x[flags], #0, 18f\n"
287 "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
288 "add x24, x12, x19, LSL #2\n"
289 "ld1w { z8.s }, p3/Z, [x12]\n"
290 "ld1w { z9.s }, p2/Z, [x12, #1, MUL VL]\n"
291 "ld1w { z10.s }, p1/Z, [x12, #2, MUL VL]\n"
292 "ld1w { z11.s }, p0/Z, [x12, #3, MUL VL]\n"
293 "ld1w { z12.s }, p3/Z, [x24]\n"
294 "ld1w { z13.s }, p2/Z, [x24, #1, MUL VL]\n"
295 "ld1w { z14.s }, p1/Z, [x24, #2, MUL VL]\n"
296 "ld1w { z15.s }, p0/Z, [x24, #3, MUL VL]\n"
297 "b 19f\n"
298 "18:" // Height 2: no accumulate
299 "mov z8.b, #0x0\n"
300 "mov z9.b, #0x0\n"
301 "mov z10.b, #0x0\n"
302 "mov z11.b, #0x0\n"
303 "mov z12.b, #0x0\n"
304 "mov z13.b, #0x0\n"
305 "mov z14.b, #0x0\n"
306 "mov z15.b, #0x0\n"
307 "19:" // Height 2: setup done
308 "mov x27, #0x0\n"
309 "20:" // Height 2: String loop
310 "ldr x19, [%x[args_ptr], %[offsetof_string_lengths]]\n"
311 "ldr w26, [x19, x27, LSL #0x2]\n"
312 "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n"
313 "tbz %x[flags], #3, 21f\n"
314 "ldr x20, [%x[input_ptr], x27, LSL #0x3]\n"
315 "add x20, x20, x19, LSL #3\n"
316 "ldr x25, [x20, #0x0]\n"
317 "ldr x24, [x20, #0x8]\n"
318 "cbnz x27, 22f\n"
319 "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
320 "add x25, x25, x19, LSL #2\n"
321 "add x24, x24, x19, LSL #2\n"
322 "b 22f\n"
323 "21:" // Height 2: setup direct input
324 "mov x25, %x[input_ptr]\n"
325 "add x24, x25, x19, LSL #2\n"
326 "22:" // Height 2: input setup done
327 "subs x26, x26, #0x1\n"
328 "ld1rw { z0.s }, p4/Z, [x25]\n"
329 "ld1rw { z1.s }, p4/Z, [x24]\n"
330 "ld1w { z6.s }, p4/Z, [x11]\n"
331 "ld1w { z7.s }, p4/Z, [x10]\n"
332 "ble 24f\n"
333 "23:" // Height 2: Multiply loop: Main loop
334 "fmla z8.s, p4/M, z6.s, z0.s\n"
335 "fmla z12.s, p4/M, z6.s, z1.s\n"
336 "ld1w { z6.s }, p4/Z, [x9]\n"
337 "addvl x11, x11, #1\n"
338 "fmla z9.s, p4/M, z7.s, z0.s\n"
339 "fmla z13.s, p4/M, z7.s, z1.s\n"
340 "ld1w { z7.s }, p4/Z, [x28]\n"
341 "addvl x10, x10, #1\n"
342 "add x25, x25, #0x4\n"
343 "subs x26, x26, #0x1\n"
344 "fmla z10.s, p4/M, z6.s, z0.s\n"
345 "fmla z14.s, p4/M, z6.s, z1.s\n"
346 "add x24, x24, #0x4\n"
347 "fmla z11.s, p4/M, z7.s, z0.s\n"
348 "fmla z15.s, p4/M, z7.s, z1.s\n"
349 "addvl x9, x9, #1\n"
350 "ld1rw { z0.s }, p4/Z, [x25]\n"
351 "ld1rw { z1.s }, p4/Z, [x24]\n"
352 "addvl x28, x28, #1\n"
353 "ld1w { z6.s }, p4/Z, [x11]\n"
354 "ld1w { z7.s }, p4/Z, [x10]\n"
355 "bgt 23b\n"
356 "24:" // Height 2: Multiply loop: Main loop skip
357 "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n"
358 "fmla z8.s, p4/M, z6.s, z0.s\n"
359 "fmla z12.s, p4/M, z6.s, z1.s\n"
360 "ld1w { z6.s }, p4/Z, [x9]\n"
361 "fmla z9.s, p4/M, z7.s, z0.s\n"
362 "fmla z13.s, p4/M, z7.s, z1.s\n"
363 "ld1w { z7.s }, p4/Z, [x28]\n"
364 "add x27, x27, #0x1\n"
365 "cmp x27, x19\n"
366 "fmla z10.s, p4/M, z6.s, z0.s\n"
367 "fmla z14.s, p4/M, z6.s, z1.s\n"
368 "addvl x11, x11, #1\n"
369 "addvl x10, x10, #1\n"
370 "fmla z11.s, p4/M, z7.s, z0.s\n"
371 "fmla z15.s, p4/M, z7.s, z1.s\n"
372 "addvl x9, x9, #1\n"
373 "addvl x28, x28, #1\n"
374 "bne 20b\n"
375 "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
376 "add x24, x12, x19, LSL #2\n"
377 "tbz %x[flags], #1, 25f\n"
378 "add x19, %x[args_ptr], %[offset_max]\n"
379 "ld1rw { z1.s }, p4/Z, [x19]\n"
380 "add x19, %x[args_ptr], %[offset_min]\n"
381 "ld1rw { z0.s }, p4/Z, [x19]\n"
382 "fmin z8.s, p4/M, z8.s, z1.s\n"
383 "fmin z9.s, p4/M, z9.s, z1.s\n"
384 "fmin z10.s, p4/M, z10.s, z1.s\n"
385 "fmin z11.s, p4/M, z11.s, z1.s\n"
386 "fmin z12.s, p4/M, z12.s, z1.s\n"
387 "fmin z13.s, p4/M, z13.s, z1.s\n"
388 "fmin z14.s, p4/M, z14.s, z1.s\n"
389 "fmin z15.s, p4/M, z15.s, z1.s\n"
390 "fmax z8.s, p4/M, z8.s, z0.s\n"
391 "fmax z9.s, p4/M, z9.s, z0.s\n"
392 "fmax z10.s, p4/M, z10.s, z0.s\n"
393 "fmax z11.s, p4/M, z11.s, z0.s\n"
394 "fmax z12.s, p4/M, z12.s, z0.s\n"
395 "fmax z13.s, p4/M, z13.s, z0.s\n"
396 "fmax z14.s, p4/M, z14.s, z0.s\n"
397 "fmax z15.s, p4/M, z15.s, z0.s\n"
398 "25:" // Height 2: No activation
399 "st1w { z8.s }, p3, [x12]\n"
400 "st1w { z9.s }, p2, [x12, #1, MUL VL]\n"
401 "st1w { z10.s }, p1, [x12, #2, MUL VL]\n"
402 "st1w { z11.s }, p0, [x12, #3, MUL VL]\n"
403 "addvl x12, x12, #4\n"
404 "st1w { z12.s }, p3, [x24]\n"
405 "st1w { z13.s }, p2, [x24, #1, MUL VL]\n"
406 "st1w { z14.s }, p1, [x24, #2, MUL VL]\n"
407 "st1w { z15.s }, p0, [x24, #3, MUL VL]\n"
408 "26:" // Height 2: Writeback done
409 "decw x13, ALL, MUL #4\n"
410 "cmp x13, XZR\n"
411 "bgt 15b\n"
412 "b 80f\n"
413 "27:" // Height 3
414 "ldr x19, [%x[args_ptr], %[offsetof_B_ptr]]\n"
415 "mov x14, %x[bias]\n"
416 "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
417 "str x19, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
418 "mov x12, %x[output_ptr]\n"
419 "28:" // Height 3: Column loop
420 "ldr x11, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
421 "ldr x19, [%x[args_ptr], %[offsetof_B_stride]]\n"
422 "add x10, x11, x19, LSL #2\n"
423 "cntw x20, ALL, MUL #3\n"
424 "add x9, x10, x19, LSL #2\n"
425 "add x28, x9, x19, LSL #2\n"
426 "add x19, x28, x19, LSL #2\n"
427 "cmp x13, x20\n"
428 "str x19, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
429 "bgt 29f\n"
430 "decw x20\n"
431 "cmp x13, x20\n"
432 "mov x28, x11\n"
433 "bgt 29f\n"
434 "decw x20\n"
435 "cmp x13, x20\n"
436 "mov x9, x11\n"
437 "bgt 29f\n"
438 "mov x10, x11\n"
439 "29:" // Height 3: B setup done
440 "mov x19, #0x0\n"
441 "whilelt p3.s, x19, x13\n"
442 "incw x19\n"
443 "whilelt p2.s, x19, x13\n"
444 "incw x19\n"
445 "whilelt p1.s, x19, x13\n"
446 "incw x19\n"
447 "whilelt p0.s, x19, x13\n"
448 "cbz x14, 30f\n"
449 "ld1w { z8.s }, p4/Z, [x14]\n"
450 "ld1w { z9.s }, p4/Z, [x14, #1, MUL VL]\n"
451 "mov z12.d, z8.d\n"
452 "mov z13.d, z9.d\n"
453 "ld1w { z10.s }, p4/Z, [x14, #2, MUL VL]\n"
454 "ld1w { z11.s }, p4/Z, [x14, #3, MUL VL]\n"
455 "mov z14.d, z10.d\n"
456 "mov z15.d, z11.d\n"
457 "mov z16.d, z8.d\n"
458 "mov z17.d, z9.d\n"
459 "addvl x14, x14, #4\n"
460 "mov z18.d, z10.d\n"
461 "mov z19.d, z11.d\n"
462 "b 32f\n"
463 "30:" // Height 3: no bias
464 "tbz %x[flags], #0, 31f\n"
465 "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
466 "add x24, x12, x19, LSL #2\n"
467 "add x23, x24, x19, LSL #2\n"
468 "ld1w { z8.s }, p3/Z, [x12]\n"
469 "ld1w { z9.s }, p2/Z, [x12, #1, MUL VL]\n"
470 "ld1w { z10.s }, p1/Z, [x12, #2, MUL VL]\n"
471 "ld1w { z11.s }, p0/Z, [x12, #3, MUL VL]\n"
472 "ld1w { z12.s }, p3/Z, [x24]\n"
473 "ld1w { z13.s }, p2/Z, [x24, #1, MUL VL]\n"
474 "ld1w { z14.s }, p1/Z, [x24, #2, MUL VL]\n"
475 "ld1w { z15.s }, p0/Z, [x24, #3, MUL VL]\n"
476 "ld1w { z16.s }, p3/Z, [x23]\n"
477 "ld1w { z17.s }, p2/Z, [x23, #1, MUL VL]\n"
478 "ld1w { z18.s }, p1/Z, [x23, #2, MUL VL]\n"
479 "ld1w { z19.s }, p0/Z, [x23, #3, MUL VL]\n"
480 "b 32f\n"
481 "31:" // Height 3: no accumulate
482 "mov z8.b, #0x0\n"
483 "mov z9.b, #0x0\n"
484 "mov z10.b, #0x0\n"
485 "mov z11.b, #0x0\n"
486 "mov z12.b, #0x0\n"
487 "mov z13.b, #0x0\n"
488 "mov z14.b, #0x0\n"
489 "mov z15.b, #0x0\n"
490 "mov z16.b, #0x0\n"
491 "mov z17.b, #0x0\n"
492 "mov z18.b, #0x0\n"
493 "mov z19.b, #0x0\n"
494 "32:" // Height 3: setup done
495 "mov x27, #0x0\n"
496 "33:" // Height 3: String loop
497 "ldr x19, [%x[args_ptr], %[offsetof_string_lengths]]\n"
498 "ldr w26, [x19, x27, LSL #0x2]\n"
499 "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n"
500 "tbz %x[flags], #3, 34f\n"
501 "ldr x20, [%x[input_ptr], x27, LSL #0x3]\n"
502 "add x20, x20, x19, LSL #3\n"
503 "ldr x25, [x20, #0x0]\n"
504 "ldr x24, [x20, #0x8]\n"
505 "ldr x23, [x20, #0x10]\n"
506 "cbnz x27, 35f\n"
507 "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
508 "add x25, x25, x19, LSL #2\n"
509 "add x24, x24, x19, LSL #2\n"
510 "add x23, x23, x19, LSL #2\n"
511 "b 35f\n"
512 "34:" // Height 3: setup direct input
513 "mov x25, %x[input_ptr]\n"
514 "add x24, x25, x19, LSL #2\n"
515 "add x23, x24, x19, LSL #2\n"
516 "35:" // Height 3: input setup done
517 "subs x26, x26, #0x1\n"
518 "ld1rw { z0.s }, p4/Z, [x25]\n"
519 "ld1rw { z1.s }, p4/Z, [x24]\n"
520 "ld1rw { z2.s }, p4/Z, [x23]\n"
521 "ld1w { z6.s }, p4/Z, [x11]\n"
522 "ld1w { z7.s }, p4/Z, [x10]\n"
523 "ble 37f\n"
524 "36:" // Height 3: Multiply loop: Main loop
525 "fmla z8.s, p4/M, z6.s, z0.s\n"
526 "fmla z12.s, p4/M, z6.s, z1.s\n"
527 "addvl x11, x11, #1\n"
528 "addvl x10, x10, #1\n"
529 "fmla z16.s, p4/M, z6.s, z2.s\n"
530 "fmla z9.s, p4/M, z7.s, z0.s\n"
531 "ld1w { z6.s }, p4/Z, [x9]\n"
532 "add x25, x25, #0x4\n"
533 "fmla z13.s, p4/M, z7.s, z1.s\n"
534 "fmla z17.s, p4/M, z7.s, z2.s\n"
535 "ld1w { z7.s }, p4/Z, [x28]\n"
536 "subs x26, x26, #0x1\n"
537 "add x24, x24, #0x4\n"
538 "add x23, x23, #0x4\n"
539 "fmla z10.s, p4/M, z6.s, z0.s\n"
540 "fmla z14.s, p4/M, z6.s, z1.s\n"
541 "fmla z18.s, p4/M, z6.s, z2.s\n"
542 "fmla z11.s, p4/M, z7.s, z0.s\n"
543 "addvl x9, x9, #1\n"
544 "addvl x28, x28, #1\n"
545 "fmla z15.s, p4/M, z7.s, z1.s\n"
546 "fmla z19.s, p4/M, z7.s, z2.s\n"
547 "ld1rw { z0.s }, p4/Z, [x25]\n"
548 "ld1rw { z1.s }, p4/Z, [x24]\n"
549 "ld1rw { z2.s }, p4/Z, [x23]\n"
550 "ld1w { z6.s }, p4/Z, [x11]\n"
551 "ld1w { z7.s }, p4/Z, [x10]\n"
552 "bgt 36b\n"
553 "37:" // Height 3: Multiply loop: Main loop skip
554 "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n"
555 "fmla z8.s, p4/M, z6.s, z0.s\n"
556 "fmla z12.s, p4/M, z6.s, z1.s\n"
557 "add x27, x27, #0x1\n"
558 "fmla z16.s, p4/M, z6.s, z2.s\n"
559 "fmla z9.s, p4/M, z7.s, z0.s\n"
560 "ld1w { z6.s }, p4/Z, [x9]\n"
561 "cmp x27, x19\n"
562 "fmla z13.s, p4/M, z7.s, z1.s\n"
563 "fmla z17.s, p4/M, z7.s, z2.s\n"
564 "ld1w { z7.s }, p4/Z, [x28]\n"
565 "addvl x11, x11, #1\n"
566 "fmla z10.s, p4/M, z6.s, z0.s\n"
567 "fmla z14.s, p4/M, z6.s, z1.s\n"
568 "addvl x10, x10, #1\n"
569 "addvl x9, x9, #1\n"
570 "fmla z18.s, p4/M, z6.s, z2.s\n"
571 "fmla z11.s, p4/M, z7.s, z0.s\n"
572 "addvl x28, x28, #1\n"
573 "fmla z15.s, p4/M, z7.s, z1.s\n"
574 "fmla z19.s, p4/M, z7.s, z2.s\n"
575 "bne 33b\n"
576 "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
577 "add x24, x12, x19, LSL #2\n"
578 "add x23, x24, x19, LSL #2\n"
579 "tbz %x[flags], #1, 38f\n"
580 "add x19, %x[args_ptr], %[offset_max]\n"
581 "ld1rw { z1.s }, p4/Z, [x19]\n"
582 "add x19, %x[args_ptr], %[offset_min]\n"
583 "ld1rw { z0.s }, p4/Z, [x19]\n"
584 "fmin z8.s, p4/M, z8.s, z1.s\n"
585 "fmin z9.s, p4/M, z9.s, z1.s\n"
586 "fmin z10.s, p4/M, z10.s, z1.s\n"
587 "fmin z11.s, p4/M, z11.s, z1.s\n"
588 "fmin z12.s, p4/M, z12.s, z1.s\n"
589 "fmin z13.s, p4/M, z13.s, z1.s\n"
590 "fmin z14.s, p4/M, z14.s, z1.s\n"
591 "fmin z15.s, p4/M, z15.s, z1.s\n"
592 "fmin z16.s, p4/M, z16.s, z1.s\n"
593 "fmin z17.s, p4/M, z17.s, z1.s\n"
594 "fmin z18.s, p4/M, z18.s, z1.s\n"
595 "fmin z19.s, p4/M, z19.s, z1.s\n"
596 "fmax z8.s, p4/M, z8.s, z0.s\n"
597 "fmax z9.s, p4/M, z9.s, z0.s\n"
598 "fmax z10.s, p4/M, z10.s, z0.s\n"
599 "fmax z11.s, p4/M, z11.s, z0.s\n"
600 "fmax z12.s, p4/M, z12.s, z0.s\n"
601 "fmax z13.s, p4/M, z13.s, z0.s\n"
602 "fmax z14.s, p4/M, z14.s, z0.s\n"
603 "fmax z15.s, p4/M, z15.s, z0.s\n"
604 "fmax z16.s, p4/M, z16.s, z0.s\n"
605 "fmax z17.s, p4/M, z17.s, z0.s\n"
606 "fmax z18.s, p4/M, z18.s, z0.s\n"
607 "fmax z19.s, p4/M, z19.s, z0.s\n"
608 "38:" // Height 3: No activation
609 "st1w { z8.s }, p3, [x12]\n"
610 "st1w { z9.s }, p2, [x12, #1, MUL VL]\n"
611 "st1w { z10.s }, p1, [x12, #2, MUL VL]\n"
612 "st1w { z11.s }, p0, [x12, #3, MUL VL]\n"
613 "addvl x12, x12, #4\n"
614 "st1w { z12.s }, p3, [x24]\n"
615 "st1w { z13.s }, p2, [x24, #1, MUL VL]\n"
616 "st1w { z14.s }, p1, [x24, #2, MUL VL]\n"
617 "st1w { z15.s }, p0, [x24, #3, MUL VL]\n"
618 "st1w { z16.s }, p3, [x23]\n"
619 "st1w { z17.s }, p2, [x23, #1, MUL VL]\n"
620 "st1w { z18.s }, p1, [x23, #2, MUL VL]\n"
621 "st1w { z19.s }, p0, [x23, #3, MUL VL]\n"
622 "39:" // Height 3: Writeback done
623 "decw x13, ALL, MUL #4\n"
624 "cmp x13, XZR\n"
625 "bgt 28b\n"
626 "b 80f\n"
627 "40:" // Height 4
628 "ldr x19, [%x[args_ptr], %[offsetof_B_ptr]]\n"
629 "mov x14, %x[bias]\n"
630 "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
631 "str x19, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
632 "mov x12, %x[output_ptr]\n"
633 "41:" // Height 4: Column loop
634 "ldr x11, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
635 "ldr x19, [%x[args_ptr], %[offsetof_B_stride]]\n"
636 "add x10, x11, x19, LSL #2\n"
637 "cntw x20, ALL, MUL #3\n"
638 "add x9, x10, x19, LSL #2\n"
639 "add x28, x9, x19, LSL #2\n"
640 "add x19, x28, x19, LSL #2\n"
641 "cmp x13, x20\n"
642 "str x19, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
643 "bgt 42f\n"
644 "decw x20\n"
645 "cmp x13, x20\n"
646 "mov x28, x11\n"
647 "bgt 42f\n"
648 "decw x20\n"
649 "cmp x13, x20\n"
650 "mov x9, x11\n"
651 "bgt 42f\n"
652 "mov x10, x11\n"
653 "42:" // Height 4: B setup done
654 "mov x19, #0x0\n"
655 "whilelt p3.s, x19, x13\n"
656 "incw x19\n"
657 "whilelt p2.s, x19, x13\n"
658 "incw x19\n"
659 "whilelt p1.s, x19, x13\n"
660 "incw x19\n"
661 "whilelt p0.s, x19, x13\n"
662 "cbz x14, 43f\n"
663 "ld1w { z8.s }, p4/Z, [x14]\n"
664 "ld1w { z9.s }, p4/Z, [x14, #1, MUL VL]\n"
665 "mov z12.d, z8.d\n"
666 "mov z13.d, z9.d\n"
667 "ld1w { z10.s }, p4/Z, [x14, #2, MUL VL]\n"
668 "ld1w { z11.s }, p4/Z, [x14, #3, MUL VL]\n"
669 "mov z14.d, z10.d\n"
670 "mov z15.d, z11.d\n"
671 "mov z16.d, z8.d\n"
672 "mov z17.d, z9.d\n"
673 "addvl x14, x14, #4\n"
674 "mov z18.d, z10.d\n"
675 "mov z19.d, z11.d\n"
676 "mov z20.d, z8.d\n"
677 "mov z21.d, z9.d\n"
678 "mov z22.d, z10.d\n"
679 "mov z23.d, z11.d\n"
680 "b 45f\n"
681 "43:" // Height 4: no bias
682 "tbz %x[flags], #0, 44f\n"
683 "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
684 "add x24, x12, x19, LSL #2\n"
685 "add x23, x24, x19, LSL #2\n"
686 "ld1w { z8.s }, p3/Z, [x12]\n"
687 "add x22, x23, x19, LSL #2\n"
688 "ld1w { z9.s }, p2/Z, [x12, #1, MUL VL]\n"
689 "ld1w { z10.s }, p1/Z, [x12, #2, MUL VL]\n"
690 "ld1w { z11.s }, p0/Z, [x12, #3, MUL VL]\n"
691 "ld1w { z12.s }, p3/Z, [x24]\n"
692 "ld1w { z13.s }, p2/Z, [x24, #1, MUL VL]\n"
693 "ld1w { z14.s }, p1/Z, [x24, #2, MUL VL]\n"
694 "ld1w { z15.s }, p0/Z, [x24, #3, MUL VL]\n"
695 "ld1w { z16.s }, p3/Z, [x23]\n"
696 "ld1w { z17.s }, p2/Z, [x23, #1, MUL VL]\n"
697 "ld1w { z18.s }, p1/Z, [x23, #2, MUL VL]\n"
698 "ld1w { z19.s }, p0/Z, [x23, #3, MUL VL]\n"
699 "ld1w { z20.s }, p3/Z, [x22]\n"
700 "ld1w { z21.s }, p2/Z, [x22, #1, MUL VL]\n"
701 "ld1w { z22.s }, p1/Z, [x22, #2, MUL VL]\n"
702 "ld1w { z23.s }, p0/Z, [x22, #3, MUL VL]\n"
703 "b 45f\n"
704 "44:" // Height 4: no accumulate
705 "mov z8.b, #0x0\n"
706 "mov z9.b, #0x0\n"
707 "mov z10.b, #0x0\n"
708 "mov z11.b, #0x0\n"
709 "mov z12.b, #0x0\n"
710 "mov z13.b, #0x0\n"
711 "mov z14.b, #0x0\n"
712 "mov z15.b, #0x0\n"
713 "mov z16.b, #0x0\n"
714 "mov z17.b, #0x0\n"
715 "mov z18.b, #0x0\n"
716 "mov z19.b, #0x0\n"
717 "mov z20.b, #0x0\n"
718 "mov z21.b, #0x0\n"
719 "mov z22.b, #0x0\n"
720 "mov z23.b, #0x0\n"
721 "45:" // Height 4: setup done
722 "mov x27, #0x0\n"
723 "46:" // Height 4: String loop
724 "ldr x19, [%x[args_ptr], %[offsetof_string_lengths]]\n"
725 "ldr w26, [x19, x27, LSL #0x2]\n"
726 "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n"
727 "tbz %x[flags], #3, 47f\n"
728 "ldr x20, [%x[input_ptr], x27, LSL #0x3]\n"
729 "add x20, x20, x19, LSL #3\n"
730 "ldr x25, [x20, #0x0]\n"
731 "ldr x24, [x20, #0x8]\n"
732 "ldr x23, [x20, #0x10]\n"
733 "ldr x22, [x20, #0x18]\n"
734 "cbnz x27, 48f\n"
735 "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
736 "add x25, x25, x19, LSL #2\n"
737 "add x24, x24, x19, LSL #2\n"
738 "add x23, x23, x19, LSL #2\n"
739 "add x22, x22, x19, LSL #2\n"
740 "b 48f\n"
741 "47:" // Height 4: setup direct input
742 "mov x25, %x[input_ptr]\n"
743 "add x24, x25, x19, LSL #2\n"
744 "add x23, x24, x19, LSL #2\n"
745 "add x22, x23, x19, LSL #2\n"
746 "48:" // Height 4: input setup done
747 "subs x26, x26, #0x1\n"
748 "ld1rw { z0.s }, p4/Z, [x25]\n"
749 "ld1rw { z1.s }, p4/Z, [x24]\n"
750 "ld1rw { z2.s }, p4/Z, [x23]\n"
751 "ld1rw { z3.s }, p4/Z, [x22]\n"
752 "ld1w { z6.s }, p4/Z, [x11]\n"
753 "ld1w { z7.s }, p4/Z, [x10]\n"
754 "ble 50f\n"
755 "49:" // Height 4: Multiply loop: Main loop
756 "fmla z8.s, p4/M, z6.s, z0.s\n"
757 "fmla z12.s, p4/M, z6.s, z1.s\n"
758 "addvl x11, x11, #1\n"
759 "addvl x10, x10, #1\n"
760 "fmla z16.s, p4/M, z6.s, z2.s\n"
761 "fmla z20.s, p4/M, z6.s, z3.s\n"
762 "ld1w { z6.s }, p4/Z, [x9]\n"
763 "add x25, x25, #0x4\n"
764 "fmla z9.s, p4/M, z7.s, z0.s\n"
765 "fmla z13.s, p4/M, z7.s, z1.s\n"
766 "subs x26, x26, #0x1\n"
767 "add x24, x24, #0x4\n"
768 "fmla z17.s, p4/M, z7.s, z2.s\n"
769 "fmla z21.s, p4/M, z7.s, z3.s\n"
770 "ld1w { z7.s }, p4/Z, [x28]\n"
771 "add x23, x23, #0x4\n"
772 "add x22, x22, #0x4\n"
773 "fmla z10.s, p4/M, z6.s, z0.s\n"
774 "fmla z14.s, p4/M, z6.s, z1.s\n"
775 "addvl x9, x9, #1\n"
776 "fmla z18.s, p4/M, z6.s, z2.s\n"
777 "fmla z22.s, p4/M, z6.s, z3.s\n"
778 "addvl x28, x28, #1\n"
779 "ld1w { z6.s }, p4/Z, [x11]\n"
780 "fmla z11.s, p4/M, z7.s, z0.s\n"
781 "fmla z15.s, p4/M, z7.s, z1.s\n"
782 "ld1rw { z0.s }, p4/Z, [x25]\n"
783 "ld1rw { z1.s }, p4/Z, [x24]\n"
784 "fmla z19.s, p4/M, z7.s, z2.s\n"
785 "fmla z23.s, p4/M, z7.s, z3.s\n"
786 "ld1rw { z2.s }, p4/Z, [x23]\n"
787 "ld1rw { z3.s }, p4/Z, [x22]\n"
788 "ld1w { z7.s }, p4/Z, [x10]\n"
789 "bgt 49b\n"
790 "50:" // Height 4: Multiply loop: Main loop skip
791 "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n"
792 "fmla z8.s, p4/M, z6.s, z0.s\n"
793 "fmla z12.s, p4/M, z6.s, z1.s\n"
794 "add x27, x27, #0x1\n"
795 "fmla z16.s, p4/M, z6.s, z2.s\n"
796 "fmla z20.s, p4/M, z6.s, z3.s\n"
797 "ld1w { z6.s }, p4/Z, [x9]\n"
798 "cmp x27, x19\n"
799 "fmla z9.s, p4/M, z7.s, z0.s\n"
800 "fmla z13.s, p4/M, z7.s, z1.s\n"
801 "addvl x11, x11, #1\n"
802 "addvl x10, x10, #1\n"
803 "fmla z17.s, p4/M, z7.s, z2.s\n"
804 "fmla z21.s, p4/M, z7.s, z3.s\n"
805 "ld1w { z7.s }, p4/Z, [x28]\n"
806 "addvl x9, x9, #1\n"
807 "fmla z10.s, p4/M, z6.s, z0.s\n"
808 "fmla z14.s, p4/M, z6.s, z1.s\n"
809 "addvl x28, x28, #1\n"
810 "fmla z18.s, p4/M, z6.s, z2.s\n"
811 "fmla z22.s, p4/M, z6.s, z3.s\n"
812 "fmla z11.s, p4/M, z7.s, z0.s\n"
813 "fmla z15.s, p4/M, z7.s, z1.s\n"
814 "fmla z19.s, p4/M, z7.s, z2.s\n"
815 "fmla z23.s, p4/M, z7.s, z3.s\n"
816 "bne 46b\n"
817 "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
818 "add x24, x12, x19, LSL #2\n"
819 "add x23, x24, x19, LSL #2\n"
820 "add x22, x23, x19, LSL #2\n"
821 "tbz %x[flags], #1, 51f\n"
822 "add x19, %x[args_ptr], %[offset_max]\n"
823 "ld1rw { z1.s }, p4/Z, [x19]\n"
824 "add x19, %x[args_ptr], %[offset_min]\n"
825 "ld1rw { z0.s }, p4/Z, [x19]\n"
826 "fmin z8.s, p4/M, z8.s, z1.s\n"
827 "fmin z9.s, p4/M, z9.s, z1.s\n"
828 "fmin z10.s, p4/M, z10.s, z1.s\n"
829 "fmin z11.s, p4/M, z11.s, z1.s\n"
830 "fmin z12.s, p4/M, z12.s, z1.s\n"
831 "fmin z13.s, p4/M, z13.s, z1.s\n"
832 "fmin z14.s, p4/M, z14.s, z1.s\n"
833 "fmin z15.s, p4/M, z15.s, z1.s\n"
834 "fmin z16.s, p4/M, z16.s, z1.s\n"
835 "fmin z17.s, p4/M, z17.s, z1.s\n"
836 "fmin z18.s, p4/M, z18.s, z1.s\n"
837 "fmin z19.s, p4/M, z19.s, z1.s\n"
838 "fmin z20.s, p4/M, z20.s, z1.s\n"
839 "fmin z21.s, p4/M, z21.s, z1.s\n"
840 "fmin z22.s, p4/M, z22.s, z1.s\n"
841 "fmin z23.s, p4/M, z23.s, z1.s\n"
842 "fmax z8.s, p4/M, z8.s, z0.s\n"
843 "fmax z9.s, p4/M, z9.s, z0.s\n"
844 "fmax z10.s, p4/M, z10.s, z0.s\n"
845 "fmax z11.s, p4/M, z11.s, z0.s\n"
846 "fmax z12.s, p4/M, z12.s, z0.s\n"
847 "fmax z13.s, p4/M, z13.s, z0.s\n"
848 "fmax z14.s, p4/M, z14.s, z0.s\n"
849 "fmax z15.s, p4/M, z15.s, z0.s\n"
850 "fmax z16.s, p4/M, z16.s, z0.s\n"
851 "fmax z17.s, p4/M, z17.s, z0.s\n"
852 "fmax z18.s, p4/M, z18.s, z0.s\n"
853 "fmax z19.s, p4/M, z19.s, z0.s\n"
854 "fmax z20.s, p4/M, z20.s, z0.s\n"
855 "fmax z21.s, p4/M, z21.s, z0.s\n"
856 "fmax z22.s, p4/M, z22.s, z0.s\n"
857 "fmax z23.s, p4/M, z23.s, z0.s\n"
858 "51:" // Height 4: No activation
859 "st1w { z8.s }, p3, [x12]\n"
860 "st1w { z9.s }, p2, [x12, #1, MUL VL]\n"
861 "st1w { z10.s }, p1, [x12, #2, MUL VL]\n"
862 "st1w { z11.s }, p0, [x12, #3, MUL VL]\n"
863 "addvl x12, x12, #4\n"
864 "st1w { z12.s }, p3, [x24]\n"
865 "st1w { z13.s }, p2, [x24, #1, MUL VL]\n"
866 "st1w { z14.s }, p1, [x24, #2, MUL VL]\n"
867 "st1w { z15.s }, p0, [x24, #3, MUL VL]\n"
868 "st1w { z16.s }, p3, [x23]\n"
869 "st1w { z17.s }, p2, [x23, #1, MUL VL]\n"
870 "st1w { z18.s }, p1, [x23, #2, MUL VL]\n"
871 "st1w { z19.s }, p0, [x23, #3, MUL VL]\n"
872 "st1w { z20.s }, p3, [x22]\n"
873 "st1w { z21.s }, p2, [x22, #1, MUL VL]\n"
874 "st1w { z22.s }, p1, [x22, #2, MUL VL]\n"
875 "st1w { z23.s }, p0, [x22, #3, MUL VL]\n"
876 "52:" // Height 4: Writeback done
877 "decw x13, ALL, MUL #4\n"
878 "cmp x13, XZR\n"
879 "bgt 41b\n"
880 "b 80f\n"
881 "53:" // Height 5
882 "ldr x19, [%x[args_ptr], %[offsetof_B_ptr]]\n"
883 "mov x14, %x[bias]\n"
884 "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
885 "str x19, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
886 "mov x12, %x[output_ptr]\n"
887 "54:" // Height 5: Column loop
888 "ldr x11, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
889 "ldr x19, [%x[args_ptr], %[offsetof_B_stride]]\n"
890 "add x10, x11, x19, LSL #2\n"
891 "cntw x20, ALL, MUL #3\n"
892 "add x9, x10, x19, LSL #2\n"
893 "add x28, x9, x19, LSL #2\n"
894 "add x19, x28, x19, LSL #2\n"
895 "cmp x13, x20\n"
896 "str x19, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
897 "bgt 55f\n"
898 "decw x20\n"
899 "cmp x13, x20\n"
900 "mov x28, x11\n"
901 "bgt 55f\n"
902 "decw x20\n"
903 "cmp x13, x20\n"
904 "mov x9, x11\n"
905 "bgt 55f\n"
906 "mov x10, x11\n"
907 "55:" // Height 5: B setup done
908 "mov x19, #0x0\n"
909 "whilelt p3.s, x19, x13\n"
910 "incw x19\n"
911 "whilelt p2.s, x19, x13\n"
912 "incw x19\n"
913 "whilelt p1.s, x19, x13\n"
914 "incw x19\n"
915 "whilelt p0.s, x19, x13\n"
916 "cbz x14, 56f\n"
917 "ld1w { z8.s }, p4/Z, [x14]\n"
918 "ld1w { z9.s }, p4/Z, [x14, #1, MUL VL]\n"
919 "mov z12.d, z8.d\n"
920 "mov z13.d, z9.d\n"
921 "ld1w { z10.s }, p4/Z, [x14, #2, MUL VL]\n"
922 "ld1w { z11.s }, p4/Z, [x14, #3, MUL VL]\n"
923 "mov z14.d, z10.d\n"
924 "mov z15.d, z11.d\n"
925 "mov z16.d, z8.d\n"
926 "mov z17.d, z9.d\n"
927 "addvl x14, x14, #4\n"
928 "mov z18.d, z10.d\n"
929 "mov z19.d, z11.d\n"
930 "mov z20.d, z8.d\n"
931 "mov z21.d, z9.d\n"
932 "mov z22.d, z10.d\n"
933 "mov z23.d, z11.d\n"
934 "mov z24.d, z8.d\n"
935 "mov z25.d, z9.d\n"
936 "mov z26.d, z10.d\n"
937 "mov z27.d, z11.d\n"
938 "b 58f\n"
939 "56:" // Height 5: no bias
940 "tbz %x[flags], #0, 57f\n"
941 "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
942 "add x24, x12, x19, LSL #2\n"
943 "add x23, x24, x19, LSL #2\n"
944 "ld1w { z8.s }, p3/Z, [x12]\n"
945 "add x22, x23, x19, LSL #2\n"
946 "add x21, x22, x19, LSL #2\n"
947 "ld1w { z9.s }, p2/Z, [x12, #1, MUL VL]\n"
948 "ld1w { z10.s }, p1/Z, [x12, #2, MUL VL]\n"
949 "ld1w { z11.s }, p0/Z, [x12, #3, MUL VL]\n"
950 "ld1w { z12.s }, p3/Z, [x24]\n"
951 "ld1w { z13.s }, p2/Z, [x24, #1, MUL VL]\n"
952 "ld1w { z14.s }, p1/Z, [x24, #2, MUL VL]\n"
953 "ld1w { z15.s }, p0/Z, [x24, #3, MUL VL]\n"
954 "ld1w { z16.s }, p3/Z, [x23]\n"
955 "ld1w { z17.s }, p2/Z, [x23, #1, MUL VL]\n"
956 "ld1w { z18.s }, p1/Z, [x23, #2, MUL VL]\n"
957 "ld1w { z19.s }, p0/Z, [x23, #3, MUL VL]\n"
958 "ld1w { z20.s }, p3/Z, [x22]\n"
959 "ld1w { z21.s }, p2/Z, [x22, #1, MUL VL]\n"
960 "ld1w { z22.s }, p1/Z, [x22, #2, MUL VL]\n"
961 "ld1w { z23.s }, p0/Z, [x22, #3, MUL VL]\n"
962 "ld1w { z24.s }, p3/Z, [x21]\n"
963 "ld1w { z25.s }, p2/Z, [x21, #1, MUL VL]\n"
964 "ld1w { z26.s }, p1/Z, [x21, #2, MUL VL]\n"
965 "ld1w { z27.s }, p0/Z, [x21, #3, MUL VL]\n"
966 "b 58f\n"
967 "57:" // Height 5: no accumulate
968 "mov z8.b, #0x0\n"
969 "mov z9.b, #0x0\n"
970 "mov z10.b, #0x0\n"
971 "mov z11.b, #0x0\n"
972 "mov z12.b, #0x0\n"
973 "mov z13.b, #0x0\n"
974 "mov z14.b, #0x0\n"
975 "mov z15.b, #0x0\n"
976 "mov z16.b, #0x0\n"
977 "mov z17.b, #0x0\n"
978 "mov z18.b, #0x0\n"
979 "mov z19.b, #0x0\n"
980 "mov z20.b, #0x0\n"
981 "mov z21.b, #0x0\n"
982 "mov z22.b, #0x0\n"
983 "mov z23.b, #0x0\n"
984 "mov z24.b, #0x0\n"
985 "mov z25.b, #0x0\n"
986 "mov z26.b, #0x0\n"
987 "mov z27.b, #0x0\n"
988 "58:" // Height 5: setup done
989 "mov x27, #0x0\n"
990 "59:" // Height 5: String loop
991 "ldr x19, [%x[args_ptr], %[offsetof_string_lengths]]\n"
992 "ldr w26, [x19, x27, LSL #0x2]\n"
993 "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n"
994 "tbz %x[flags], #3, 60f\n"
995 "ldr x20, [%x[input_ptr], x27, LSL #0x3]\n"
996 "add x20, x20, x19, LSL #3\n"
997 "ldr x25, [x20, #0x0]\n"
998 "ldr x24, [x20, #0x8]\n"
999 "ldr x23, [x20, #0x10]\n"
1000 "ldr x22, [x20, #0x18]\n"
1001 "ldr x21, [x20, #0x20]\n"
1002 "cbnz x27, 61f\n"
1003 "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
1004 "add x25, x25, x19, LSL #2\n"
1005 "add x24, x24, x19, LSL #2\n"
1006 "add x23, x23, x19, LSL #2\n"
1007 "add x22, x22, x19, LSL #2\n"
1008 "add x21, x21, x19, LSL #2\n"
1009 "b 61f\n"
1010 "60:" // Height 5: setup direct input
1011 "mov x25, %x[input_ptr]\n"
1012 "add x24, x25, x19, LSL #2\n"
1013 "add x23, x24, x19, LSL #2\n"
1014 "add x22, x23, x19, LSL #2\n"
1015 "add x21, x22, x19, LSL #2\n"
1016 "61:" // Height 5: input setup done
1017 "subs x26, x26, #0x1\n"
1018 "ld1rw { z0.s }, p4/Z, [x25]\n"
1019 "ld1rw { z1.s }, p4/Z, [x24]\n"
1020 "ld1rw { z2.s }, p4/Z, [x23]\n"
1021 "ld1rw { z3.s }, p4/Z, [x22]\n"
1022 "ld1rw { z4.s }, p4/Z, [x21]\n"
1023 "ld1w { z6.s }, p4/Z, [x11]\n"
1024 "ld1w { z7.s }, p4/Z, [x10]\n"
1025 "ble 63f\n"
1026 "62:" // Height 5: Multiply loop: Main loop
1027 "fmla z8.s, p4/M, z6.s, z0.s\n"
1028 "fmla z12.s, p4/M, z6.s, z1.s\n"
1029 "addvl x11, x11, #1\n"
1030 "addvl x10, x10, #1\n"
1031 "fmla z16.s, p4/M, z6.s, z2.s\n"
1032 "fmla z20.s, p4/M, z6.s, z3.s\n"
1033 "add x25, x25, #0x4\n"
1034 "subs x26, x26, #0x1\n"
1035 "fmla z24.s, p4/M, z6.s, z4.s\n"
1036 "fmla z9.s, p4/M, z7.s, z0.s\n"
1037 "ld1w { z6.s }, p4/Z, [x9]\n"
1038 "add x24, x24, #0x4\n"
1039 "fmla z13.s, p4/M, z7.s, z1.s\n"
1040 "fmla z17.s, p4/M, z7.s, z2.s\n"
1041 "add x23, x23, #0x4\n"
1042 "add x22, x22, #0x4\n"
1043 "fmla z21.s, p4/M, z7.s, z3.s\n"
1044 "fmla z25.s, p4/M, z7.s, z4.s\n"
1045 "ld1w { z7.s }, p4/Z, [x28]\n"
1046 "add x21, x21, #0x4\n"
1047 "fmla z10.s, p4/M, z6.s, z0.s\n"
1048 "fmla z14.s, p4/M, z6.s, z1.s\n"
1049 "addvl x9, x9, #1\n"
1050 "addvl x28, x28, #1\n"
1051 "fmla z18.s, p4/M, z6.s, z2.s\n"
1052 "fmla z22.s, p4/M, z6.s, z3.s\n"
1053 "fmla z26.s, p4/M, z6.s, z4.s\n"
1054 "fmla z11.s, p4/M, z7.s, z0.s\n"
1055 "ld1rw { z0.s }, p4/Z, [x25]\n"
1056 "ld1w { z6.s }, p4/Z, [x11]\n"
1057 "fmla z15.s, p4/M, z7.s, z1.s\n"
1058 "fmla z19.s, p4/M, z7.s, z2.s\n"
1059 "ld1rw { z1.s }, p4/Z, [x24]\n"
1060 "ld1rw { z2.s }, p4/Z, [x23]\n"
1061 "fmla z23.s, p4/M, z7.s, z3.s\n"
1062 "fmla z27.s, p4/M, z7.s, z4.s\n"
1063 "ld1rw { z3.s }, p4/Z, [x22]\n"
1064 "ld1rw { z4.s }, p4/Z, [x21]\n"
1065 "ld1w { z7.s }, p4/Z, [x10]\n"
1066 "bgt 62b\n"
1067 "63:" // Height 5: Multiply loop: Main loop skip
1068 "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n"
1069 "fmla z8.s, p4/M, z6.s, z0.s\n"
1070 "fmla z12.s, p4/M, z6.s, z1.s\n"
1071 "add x27, x27, #0x1\n"
1072 "fmla z16.s, p4/M, z6.s, z2.s\n"
1073 "fmla z20.s, p4/M, z6.s, z3.s\n"
1074 "cmp x27, x19\n"
1075 "addvl x11, x11, #1\n"
1076 "fmla z24.s, p4/M, z6.s, z4.s\n"
1077 "fmla z9.s, p4/M, z7.s, z0.s\n"
1078 "ld1w { z6.s }, p4/Z, [x9]\n"
1079 "addvl x10, x10, #1\n"
1080 "fmla z13.s, p4/M, z7.s, z1.s\n"
1081 "fmla z17.s, p4/M, z7.s, z2.s\n"
1082 "addvl x9, x9, #1\n"
1083 "fmla z21.s, p4/M, z7.s, z3.s\n"
1084 "fmla z25.s, p4/M, z7.s, z4.s\n"
1085 "ld1w { z7.s }, p4/Z, [x28]\n"
1086 "addvl x28, x28, #1\n"
1087 "fmla z10.s, p4/M, z6.s, z0.s\n"
1088 "fmla z14.s, p4/M, z6.s, z1.s\n"
1089 "fmla z18.s, p4/M, z6.s, z2.s\n"
1090 "fmla z22.s, p4/M, z6.s, z3.s\n"
1091 "fmla z26.s, p4/M, z6.s, z4.s\n"
1092 "fmla z11.s, p4/M, z7.s, z0.s\n"
1093 "fmla z15.s, p4/M, z7.s, z1.s\n"
1094 "fmla z19.s, p4/M, z7.s, z2.s\n"
1095 "fmla z23.s, p4/M, z7.s, z3.s\n"
1096 "fmla z27.s, p4/M, z7.s, z4.s\n"
1097 "bne 59b\n"
1098 "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
1099 "add x24, x12, x19, LSL #2\n"
1100 "add x23, x24, x19, LSL #2\n"
1101 "add x22, x23, x19, LSL #2\n"
1102 "add x21, x22, x19, LSL #2\n"
1103 "tbz %x[flags], #1, 64f\n"
1104 "add x19, %x[args_ptr], %[offset_max]\n"
1105 "ld1rw { z1.s }, p4/Z, [x19]\n"
1106 "add x19, %x[args_ptr], %[offset_min]\n"
1107 "ld1rw { z0.s }, p4/Z, [x19]\n"
1108 "fmin z8.s, p4/M, z8.s, z1.s\n"
1109 "fmin z9.s, p4/M, z9.s, z1.s\n"
1110 "fmin z10.s, p4/M, z10.s, z1.s\n"
1111 "fmin z11.s, p4/M, z11.s, z1.s\n"
1112 "fmin z12.s, p4/M, z12.s, z1.s\n"
1113 "fmin z13.s, p4/M, z13.s, z1.s\n"
1114 "fmin z14.s, p4/M, z14.s, z1.s\n"
1115 "fmin z15.s, p4/M, z15.s, z1.s\n"
1116 "fmin z16.s, p4/M, z16.s, z1.s\n"
1117 "fmin z17.s, p4/M, z17.s, z1.s\n"
1118 "fmin z18.s, p4/M, z18.s, z1.s\n"
1119 "fmin z19.s, p4/M, z19.s, z1.s\n"
1120 "fmin z20.s, p4/M, z20.s, z1.s\n"
1121 "fmin z21.s, p4/M, z21.s, z1.s\n"
1122 "fmin z22.s, p4/M, z22.s, z1.s\n"
1123 "fmin z23.s, p4/M, z23.s, z1.s\n"
1124 "fmin z24.s, p4/M, z24.s, z1.s\n"
1125 "fmin z25.s, p4/M, z25.s, z1.s\n"
1126 "fmin z26.s, p4/M, z26.s, z1.s\n"
1127 "fmin z27.s, p4/M, z27.s, z1.s\n"
1128 "fmax z8.s, p4/M, z8.s, z0.s\n"
1129 "fmax z9.s, p4/M, z9.s, z0.s\n"
1130 "fmax z10.s, p4/M, z10.s, z0.s\n"
1131 "fmax z11.s, p4/M, z11.s, z0.s\n"
1132 "fmax z12.s, p4/M, z12.s, z0.s\n"
1133 "fmax z13.s, p4/M, z13.s, z0.s\n"
1134 "fmax z14.s, p4/M, z14.s, z0.s\n"
1135 "fmax z15.s, p4/M, z15.s, z0.s\n"
1136 "fmax z16.s, p4/M, z16.s, z0.s\n"
1137 "fmax z17.s, p4/M, z17.s, z0.s\n"
1138 "fmax z18.s, p4/M, z18.s, z0.s\n"
1139 "fmax z19.s, p4/M, z19.s, z0.s\n"
1140 "fmax z20.s, p4/M, z20.s, z0.s\n"
1141 "fmax z21.s, p4/M, z21.s, z0.s\n"
1142 "fmax z22.s, p4/M, z22.s, z0.s\n"
1143 "fmax z23.s, p4/M, z23.s, z0.s\n"
1144 "fmax z24.s, p4/M, z24.s, z0.s\n"
1145 "fmax z25.s, p4/M, z25.s, z0.s\n"
1146 "fmax z26.s, p4/M, z26.s, z0.s\n"
1147 "fmax z27.s, p4/M, z27.s, z0.s\n"
1148 "64:" // Height 5: No activation
1149 "st1w { z8.s }, p3, [x12]\n"
1150 "st1w { z9.s }, p2, [x12, #1, MUL VL]\n"
1151 "st1w { z10.s }, p1, [x12, #2, MUL VL]\n"
1152 "st1w { z11.s }, p0, [x12, #3, MUL VL]\n"
1153 "addvl x12, x12, #4\n"
1154 "st1w { z12.s }, p3, [x24]\n"
1155 "st1w { z13.s }, p2, [x24, #1, MUL VL]\n"
1156 "st1w { z14.s }, p1, [x24, #2, MUL VL]\n"
1157 "st1w { z15.s }, p0, [x24, #3, MUL VL]\n"
1158 "st1w { z16.s }, p3, [x23]\n"
1159 "st1w { z17.s }, p2, [x23, #1, MUL VL]\n"
1160 "st1w { z18.s }, p1, [x23, #2, MUL VL]\n"
1161 "st1w { z19.s }, p0, [x23, #3, MUL VL]\n"
1162 "st1w { z20.s }, p3, [x22]\n"
1163 "st1w { z21.s }, p2, [x22, #1, MUL VL]\n"
1164 "st1w { z22.s }, p1, [x22, #2, MUL VL]\n"
1165 "st1w { z23.s }, p0, [x22, #3, MUL VL]\n"
1166 "st1w { z24.s }, p3, [x21]\n"
1167 "st1w { z25.s }, p2, [x21, #1, MUL VL]\n"
1168 "st1w { z26.s }, p1, [x21, #2, MUL VL]\n"
1169 "st1w { z27.s }, p0, [x21, #3, MUL VL]\n"
1170 "65:" // Height 5: Writeback done
1171 "decw x13, ALL, MUL #4\n"
1172 "cmp x13, XZR\n"
1173 "bgt 54b\n"
1174 "b 80f\n"
1175 "66:" // Height 6
1176 "ldr x19, [%x[args_ptr], %[offsetof_B_ptr]]\n"
1177 "ldr x13, [%x[args_ptr], %[offsetof_N]]\n"
1178 "str x19, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
1179 "mov x20, #0x18\n"
1180 "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
1181 "mov x14, %x[bias]\n"
1182 "mov x12, %x[output_ptr]\n"
1183 "madd %x[output_ptr], x19, x20, %x[output_ptr]\n"
1184 "67:" // Height 6: Column loop
1185 "ldr x11, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
1186 "ldr x19, [%x[args_ptr], %[offsetof_B_stride]]\n"
1187 "add x10, x11, x19, LSL #2\n"
1188 "cntw x20, ALL, MUL #3\n"
1189 "add x9, x10, x19, LSL #2\n"
1190 "add x28, x9, x19, LSL #2\n"
1191 "add x19, x28, x19, LSL #2\n"
1192 "cmp x13, x20\n"
1193 "str x19, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n"
1194 "bgt 68f\n"
1195 "decw x20\n"
1196 "cmp x13, x20\n"
1197 "mov x28, x11\n"
1198 "bgt 68f\n"
1199 "decw x20\n"
1200 "cmp x13, x20\n"
1201 "mov x9, x11\n"
1202 "bgt 68f\n"
1203 "mov x10, x11\n"
1204 "68:" // Height 6: B setup done
1205 "mov x19, #0x0\n"
1206 "whilelt p3.s, x19, x13\n"
1207 "incw x19\n"
1208 "whilelt p2.s, x19, x13\n"
1209 "incw x19\n"
1210 "whilelt p1.s, x19, x13\n"
1211 "incw x19\n"
1212 "whilelt p0.s, x19, x13\n"
1213 "cbz x14, 69f\n"
1214 "ld1w { z8.s }, p4/Z, [x14]\n"
1215 "ld1w { z9.s }, p4/Z, [x14, #1, MUL VL]\n"
1216 "mov z12.d, z8.d\n"
1217 "mov z13.d, z9.d\n"
1218 "ld1w { z10.s }, p4/Z, [x14, #2, MUL VL]\n"
1219 "ld1w { z11.s }, p4/Z, [x14, #3, MUL VL]\n"
1220 "mov z14.d, z10.d\n"
1221 "mov z15.d, z11.d\n"
1222 "mov z16.d, z8.d\n"
1223 "mov z17.d, z9.d\n"
1224 "addvl x14, x14, #4\n"
1225 "mov z18.d, z10.d\n"
1226 "mov z19.d, z11.d\n"
1227 "mov z20.d, z8.d\n"
1228 "mov z21.d, z9.d\n"
1229 "mov z22.d, z10.d\n"
1230 "mov z23.d, z11.d\n"
1231 "mov z24.d, z8.d\n"
1232 "mov z25.d, z9.d\n"
1233 "mov z26.d, z10.d\n"
1234 "mov z27.d, z11.d\n"
1235 "mov z28.d, z8.d\n"
1236 "mov z29.d, z9.d\n"
1237 "mov z30.d, z10.d\n"
1238 "mov z31.d, z11.d\n"
1239 "b 71f\n"
1240 "69:" // Height 6: no bias
1241 "tbz %x[flags], #0, 70f\n"
1242 "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
1243 "add x24, x12, x19, LSL #2\n"
1244 "add x23, x24, x19, LSL #2\n"
1245 "ld1w { z8.s }, p3/Z, [x12]\n"
1246 "add x22, x23, x19, LSL #2\n"
1247 "add x21, x22, x19, LSL #2\n"
1248 "ld1w { z9.s }, p2/Z, [x12, #1, MUL VL]\n"
1249 "ld1w { z10.s }, p1/Z, [x12, #2, MUL VL]\n"
1250 "add x20, x21, x19, LSL #2\n"
1251 "ld1w { z11.s }, p0/Z, [x12, #3, MUL VL]\n"
1252 "ld1w { z12.s }, p3/Z, [x24]\n"
1253 "ld1w { z13.s }, p2/Z, [x24, #1, MUL VL]\n"
1254 "ld1w { z14.s }, p1/Z, [x24, #2, MUL VL]\n"
1255 "ld1w { z15.s }, p0/Z, [x24, #3, MUL VL]\n"
1256 "ld1w { z16.s }, p3/Z, [x23]\n"
1257 "ld1w { z17.s }, p2/Z, [x23, #1, MUL VL]\n"
1258 "ld1w { z18.s }, p1/Z, [x23, #2, MUL VL]\n"
1259 "ld1w { z19.s }, p0/Z, [x23, #3, MUL VL]\n"
1260 "ld1w { z20.s }, p3/Z, [x22]\n"
1261 "ld1w { z21.s }, p2/Z, [x22, #1, MUL VL]\n"
1262 "ld1w { z22.s }, p1/Z, [x22, #2, MUL VL]\n"
1263 "ld1w { z23.s }, p0/Z, [x22, #3, MUL VL]\n"
1264 "ld1w { z24.s }, p3/Z, [x21]\n"
1265 "ld1w { z25.s }, p2/Z, [x21, #1, MUL VL]\n"
1266 "ld1w { z26.s }, p1/Z, [x21, #2, MUL VL]\n"
1267 "ld1w { z27.s }, p0/Z, [x21, #3, MUL VL]\n"
1268 "ld1w { z28.s }, p3/Z, [x20]\n"
1269 "ld1w { z29.s }, p2/Z, [x20, #1, MUL VL]\n"
1270 "ld1w { z30.s }, p1/Z, [x20, #2, MUL VL]\n"
1271 "ld1w { z31.s }, p0/Z, [x20, #3, MUL VL]\n"
1272 "b 71f\n"
1273 "70:" // Height 6: no accumulate
1274 "mov z8.b, #0x0\n"
1275 "mov z9.b, #0x0\n"
1276 "mov z10.b, #0x0\n"
1277 "mov z11.b, #0x0\n"
1278 "mov z12.b, #0x0\n"
1279 "mov z13.b, #0x0\n"
1280 "mov z14.b, #0x0\n"
1281 "mov z15.b, #0x0\n"
1282 "mov z16.b, #0x0\n"
1283 "mov z17.b, #0x0\n"
1284 "mov z18.b, #0x0\n"
1285 "mov z19.b, #0x0\n"
1286 "mov z20.b, #0x0\n"
1287 "mov z21.b, #0x0\n"
1288 "mov z22.b, #0x0\n"
1289 "mov z23.b, #0x0\n"
1290 "mov z24.b, #0x0\n"
1291 "mov z25.b, #0x0\n"
1292 "mov z26.b, #0x0\n"
1293 "mov z27.b, #0x0\n"
1294 "mov z28.b, #0x0\n"
1295 "mov z29.b, #0x0\n"
1296 "mov z30.b, #0x0\n"
1297 "mov z31.b, #0x0\n"
1298 "71:" // Height 6: setup done
1299 "mov x27, #0x0\n"
1300 "72:" // Height 6: String loop
1301 "ldr x19, [%x[args_ptr], %[offsetof_string_lengths]]\n"
1302 "ldr w26, [x19, x27, LSL #0x2]\n"
1303 "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n"
1304 "tbz %x[flags], #3, 73f\n"
1305 "ldr x20, [%x[input_ptr], x27, LSL #0x3]\n"
1306 "add x20, x20, x19, LSL #3\n"
1307 "ldr x25, [x20, #0x0]\n"
1308 "ldr x24, [x20, #0x8]\n"
1309 "ldr x23, [x20, #0x10]\n"
1310 "ldr x22, [x20, #0x18]\n"
1311 "ldr x21, [x20, #0x20]\n"
1312 "ldr x20, [x20, #0x28]\n"
1313 "cbnz x27, 74f\n"
1314 "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
1315 "add x25, x25, x19, LSL #2\n"
1316 "add x24, x24, x19, LSL #2\n"
1317 "add x23, x23, x19, LSL #2\n"
1318 "add x22, x22, x19, LSL #2\n"
1319 "add x21, x21, x19, LSL #2\n"
1320 "add x20, x20, x19, LSL #2\n"
1321 "b 74f\n"
1322 "73:" // Height 6: setup direct input
1323 "mov x25, %x[input_ptr]\n"
1324 "add x24, x25, x19, LSL #2\n"
1325 "add x23, x24, x19, LSL #2\n"
1326 "add x22, x23, x19, LSL #2\n"
1327 "add x21, x22, x19, LSL #2\n"
1328 "add x20, x21, x19, LSL #2\n"
1329 "74:" // Height 6: input setup done
1330 "subs x26, x26, #0x1\n"
1331 "ld1rw { z0.s }, p4/Z, [x25]\n"
1332 "ld1rw { z1.s }, p4/Z, [x24]\n"
1333 "ld1rw { z2.s }, p4/Z, [x23]\n"
1334 "ld1rw { z3.s }, p4/Z, [x22]\n"
1335 "ld1rw { z4.s }, p4/Z, [x21]\n"
1336 "ld1rw { z5.s }, p4/Z, [x20]\n"
1337 "ld1w { z6.s }, p4/Z, [x11]\n"
1338 "ld1w { z7.s }, p4/Z, [x10]\n"
1339 "ble 76f\n"
1340 "75:" // Height 6: Multiply loop: Main loop
1341 "fmla z8.s, p4/M, z6.s, z0.s\n"
1342 "fmla z12.s, p4/M, z6.s, z1.s\n"
1343 "addvl x11, x11, #1\n"
1344 "addvl x10, x10, #1\n"
1345 "fmla z16.s, p4/M, z6.s, z2.s\n"
1346 "fmla z20.s, p4/M, z6.s, z3.s\n"
1347 "add x25, x25, #0x4\n"
1348 "subs x26, x26, #0x1\n"
1349 "fmla z24.s, p4/M, z6.s, z4.s\n"
1350 "fmla z28.s, p4/M, z6.s, z5.s\n"
1351 "ld1w { z6.s }, p4/Z, [x9]\n"
1352 "add x24, x24, #0x4\n"
1353 "fmla z9.s, p4/M, z7.s, z0.s\n"
1354 "fmla z13.s, p4/M, z7.s, z1.s\n"
1355 "add x23, x23, #0x4\n"
1356 "add x22, x22, #0x4\n"
1357 "fmla z17.s, p4/M, z7.s, z2.s\n"
1358 "fmla z21.s, p4/M, z7.s, z3.s\n"
1359 "add x21, x21, #0x4\n"
1360 "add x20, x20, #0x4\n"
1361 "fmla z25.s, p4/M, z7.s, z4.s\n"
1362 "fmla z29.s, p4/M, z7.s, z5.s\n"
1363 "ld1w { z7.s }, p4/Z, [x28]\n"
1364 "addvl x9, x9, #1\n"
1365 "fmla z10.s, p4/M, z6.s, z0.s\n"
1366 "fmla z14.s, p4/M, z6.s, z1.s\n"
1367 "addvl x28, x28, #1\n"
1368 "fmla z18.s, p4/M, z6.s, z2.s\n"
1369 "fmla z22.s, p4/M, z6.s, z3.s\n"
1370 "fmla z26.s, p4/M, z6.s, z4.s\n"
1371 "fmla z30.s, p4/M, z6.s, z5.s\n"
1372 "ld1w { z6.s }, p4/Z, [x11]\n"
1373 "fmla z11.s, p4/M, z7.s, z0.s\n"
1374 "fmla z15.s, p4/M, z7.s, z1.s\n"
1375 "ld1rw { z0.s }, p4/Z, [x25]\n"
1376 "ld1rw { z1.s }, p4/Z, [x24]\n"
1377 "fmla z19.s, p4/M, z7.s, z2.s\n"
1378 "fmla z23.s, p4/M, z7.s, z3.s\n"
1379 "ld1rw { z2.s }, p4/Z, [x23]\n"
1380 "ld1rw { z3.s }, p4/Z, [x22]\n"
1381 "fmla z27.s, p4/M, z7.s, z4.s\n"
1382 "fmla z31.s, p4/M, z7.s, z5.s\n"
1383 "ld1rw { z4.s }, p4/Z, [x21]\n"
1384 "ld1rw { z5.s }, p4/Z, [x20]\n"
1385 "ld1w { z7.s }, p4/Z, [x10]\n"
1386 "bgt 75b\n"
1387 "76:" // Height 6: Multiply loop: Main loop skip
1388 "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n"
1389 "fmla z8.s, p4/M, z6.s, z0.s\n"
1390 "fmla z12.s, p4/M, z6.s, z1.s\n"
1391 "add x27, x27, #0x1\n"
1392 "fmla z16.s, p4/M, z6.s, z2.s\n"
1393 "fmla z20.s, p4/M, z6.s, z3.s\n"
1394 "cmp x27, x19\n"
1395 "addvl x11, x11, #1\n"
1396 "fmla z24.s, p4/M, z6.s, z4.s\n"
1397 "fmla z28.s, p4/M, z6.s, z5.s\n"
1398 "ld1w { z6.s }, p4/Z, [x9]\n"
1399 "addvl x10, x10, #1\n"
1400 "fmla z9.s, p4/M, z7.s, z0.s\n"
1401 "fmla z13.s, p4/M, z7.s, z1.s\n"
1402 "addvl x9, x9, #1\n"
1403 "fmla z17.s, p4/M, z7.s, z2.s\n"
1404 "fmla z21.s, p4/M, z7.s, z3.s\n"
1405 "fmla z25.s, p4/M, z7.s, z4.s\n"
1406 "fmla z29.s, p4/M, z7.s, z5.s\n"
1407 "ld1w { z7.s }, p4/Z, [x28]\n"
1408 "addvl x28, x28, #1\n"
1409 "fmla z10.s, p4/M, z6.s, z0.s\n"
1410 "fmla z14.s, p4/M, z6.s, z1.s\n"
1411 "fmla z18.s, p4/M, z6.s, z2.s\n"
1412 "fmla z22.s, p4/M, z6.s, z3.s\n"
1413 "fmla z26.s, p4/M, z6.s, z4.s\n"
1414 "fmla z30.s, p4/M, z6.s, z5.s\n"
1415 "fmla z11.s, p4/M, z7.s, z0.s\n"
1416 "fmla z15.s, p4/M, z7.s, z1.s\n"
1417 "fmla z19.s, p4/M, z7.s, z2.s\n"
1418 "fmla z23.s, p4/M, z7.s, z3.s\n"
1419 "fmla z27.s, p4/M, z7.s, z4.s\n"
1420 "fmla z31.s, p4/M, z7.s, z5.s\n"
1421 "bne 72b\n"
1422 "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
1423 "add x24, x12, x19, LSL #2\n"
1424 "add x23, x24, x19, LSL #2\n"
1425 "add x22, x23, x19, LSL #2\n"
1426 "add x21, x22, x19, LSL #2\n"
1427 "add x20, x21, x19, LSL #2\n"
1428 "tbz %x[flags], #1, 77f\n"
1429 "add x19, %x[args_ptr], %[offset_max]\n"
1430 "ld1rw { z1.s }, p4/Z, [x19]\n"
1431 "add x19, %x[args_ptr], %[offset_min]\n"
1432 "ld1rw { z0.s }, p4/Z, [x19]\n"
1433 "fmin z8.s, p4/M, z8.s, z1.s\n"
1434 "fmin z9.s, p4/M, z9.s, z1.s\n"
1435 "fmin z10.s, p4/M, z10.s, z1.s\n"
1436 "fmin z11.s, p4/M, z11.s, z1.s\n"
1437 "fmin z12.s, p4/M, z12.s, z1.s\n"
1438 "fmin z13.s, p4/M, z13.s, z1.s\n"
1439 "fmin z14.s, p4/M, z14.s, z1.s\n"
1440 "fmin z15.s, p4/M, z15.s, z1.s\n"
1441 "fmin z16.s, p4/M, z16.s, z1.s\n"
1442 "fmin z17.s, p4/M, z17.s, z1.s\n"
1443 "fmin z18.s, p4/M, z18.s, z1.s\n"
1444 "fmin z19.s, p4/M, z19.s, z1.s\n"
1445 "fmin z20.s, p4/M, z20.s, z1.s\n"
1446 "fmin z21.s, p4/M, z21.s, z1.s\n"
1447 "fmin z22.s, p4/M, z22.s, z1.s\n"
1448 "fmin z23.s, p4/M, z23.s, z1.s\n"
1449 "fmin z24.s, p4/M, z24.s, z1.s\n"
1450 "fmin z25.s, p4/M, z25.s, z1.s\n"
1451 "fmin z26.s, p4/M, z26.s, z1.s\n"
1452 "fmin z27.s, p4/M, z27.s, z1.s\n"
1453 "fmin z28.s, p4/M, z28.s, z1.s\n"
1454 "fmin z29.s, p4/M, z29.s, z1.s\n"
1455 "fmin z30.s, p4/M, z30.s, z1.s\n"
1456 "fmin z31.s, p4/M, z31.s, z1.s\n"
1457 "fmax z8.s, p4/M, z8.s, z0.s\n"
1458 "fmax z9.s, p4/M, z9.s, z0.s\n"
1459 "fmax z10.s, p4/M, z10.s, z0.s\n"
1460 "fmax z11.s, p4/M, z11.s, z0.s\n"
1461 "fmax z12.s, p4/M, z12.s, z0.s\n"
1462 "fmax z13.s, p4/M, z13.s, z0.s\n"
1463 "fmax z14.s, p4/M, z14.s, z0.s\n"
1464 "fmax z15.s, p4/M, z15.s, z0.s\n"
1465 "fmax z16.s, p4/M, z16.s, z0.s\n"
1466 "fmax z17.s, p4/M, z17.s, z0.s\n"
1467 "fmax z18.s, p4/M, z18.s, z0.s\n"
1468 "fmax z19.s, p4/M, z19.s, z0.s\n"
1469 "fmax z20.s, p4/M, z20.s, z0.s\n"
1470 "fmax z21.s, p4/M, z21.s, z0.s\n"
1471 "fmax z22.s, p4/M, z22.s, z0.s\n"
1472 "fmax z23.s, p4/M, z23.s, z0.s\n"
1473 "fmax z24.s, p4/M, z24.s, z0.s\n"
1474 "fmax z25.s, p4/M, z25.s, z0.s\n"
1475 "fmax z26.s, p4/M, z26.s, z0.s\n"
1476 "fmax z27.s, p4/M, z27.s, z0.s\n"
1477 "fmax z28.s, p4/M, z28.s, z0.s\n"
1478 "fmax z29.s, p4/M, z29.s, z0.s\n"
1479 "fmax z30.s, p4/M, z30.s, z0.s\n"
1480 "fmax z31.s, p4/M, z31.s, z0.s\n"
1481 "77:" // Height 6: No activation
1482 "st1w { z8.s }, p3, [x12]\n"
1483 "st1w { z9.s }, p2, [x12, #1, MUL VL]\n"
1484 "st1w { z10.s }, p1, [x12, #2, MUL VL]\n"
1485 "st1w { z11.s }, p0, [x12, #3, MUL VL]\n"
1486 "addvl x12, x12, #4\n"
1487 "st1w { z12.s }, p3, [x24]\n"
1488 "st1w { z13.s }, p2, [x24, #1, MUL VL]\n"
1489 "st1w { z14.s }, p1, [x24, #2, MUL VL]\n"
1490 "st1w { z15.s }, p0, [x24, #3, MUL VL]\n"
1491 "st1w { z16.s }, p3, [x23]\n"
1492 "st1w { z17.s }, p2, [x23, #1, MUL VL]\n"
1493 "st1w { z18.s }, p1, [x23, #2, MUL VL]\n"
1494 "st1w { z19.s }, p0, [x23, #3, MUL VL]\n"
1495 "st1w { z20.s }, p3, [x22]\n"
1496 "st1w { z21.s }, p2, [x22, #1, MUL VL]\n"
1497 "st1w { z22.s }, p1, [x22, #2, MUL VL]\n"
1498 "st1w { z23.s }, p0, [x22, #3, MUL VL]\n"
1499 "st1w { z24.s }, p3, [x21]\n"
1500 "st1w { z25.s }, p2, [x21, #1, MUL VL]\n"
1501 "st1w { z26.s }, p1, [x21, #2, MUL VL]\n"
1502 "st1w { z27.s }, p0, [x21, #3, MUL VL]\n"
1503 "st1w { z28.s }, p3, [x20]\n"
1504 "st1w { z29.s }, p2, [x20, #1, MUL VL]\n"
1505 "st1w { z30.s }, p1, [x20, #2, MUL VL]\n"
1506 "st1w { z31.s }, p0, [x20, #3, MUL VL]\n"
1507 "78:" // Height 6: Writeback done
1508 "decw x13, ALL, MUL #4\n"
1509 "cmp x13, XZR\n"
1510 "bgt 67b\n"
1511 "subs %x[M], %x[M], #0x6\n"
1512 "beq 80f\n"
1513 "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n"
1514 "tbz %x[flags], #3, 79f\n"
1515 "add x20, x20, #0x6\n"
1516 "str x20, [%x[args_ptr], %[offsetof_input_offset]]\n"
1517 "b 1b\n"
1518 "79:" // Update direct input
1519 "mov x19, #0x18\n"
1520 "madd %x[input_ptr], x19, x20, %x[input_ptr]\n"
1521 "b 1b\n"
1522 "80:" // Exit
1523 : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
1524 : [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_B_stride] "I" (offsetof(KernelArgs, B_stride)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_cur_B_ptr] "I" (offsetof(KernelArgs, cur_B_ptr)), [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))
1525 : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "x9", "x10", "x11", "x12", "x13", "x14", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
1526 );
1527 }
1528
1529 } // namespace arm_gemm
1530 #endif // ARM_COMPUTE_ENABLE_SVE
1531