File size: 18,882 Bytes
8b7c501
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
// Copyright 2019 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.

#include <xnnpack/assembly.h>

# void xnn_f32_gemm${"inc" if INC else ""}_minmax_ukernel_4x8__asm_aarch64_neonfma_cortex_a53${"_prfm" if PREFETCH else ""}(
#     size_t mr,                x0
#     size_t nc,                x1
#     size_t kc,                x2 / x0
#     const float* a,           x3
#     size_t a_stride,          x4
#     const float* w,           x5
#     float* c,                 x6
#     size_t cm_stride,         x7
#     size_t cn_stride,         [sp] -> (x0)
$if INC:
  #     const float* acc,         [sp + 8] -> x15
  #     const xnn_f32_minmax_params* params)  [sp + 16] -> (x8)
$else:
  #     const xnn_f32_minmax_params* params)  [sp + 8] -> (x8)

# d8-d15, x19-x30 need to be preserved if used. x18 is reserved by the OS.

# Register usage
# A0  x3  v0     v3
# A1  x9  v0[1]  v3[1]
# A2  x10 v1     v4
# A3  x11 v1[1]  v4[1]
# B   x5  v12 v13 v14 v15 second set of B
# B       v16 v17 v18 v19 first set
# C   x6  v20 v21
# C   x16 v22 v23
# C   x17 v24 v25
# C   x14 v26 v27
# Clamp v6 v7
# temporary vector shadow register x4

# unused A   v8 v9 v10 v11
# x12 a4
# x13 c4
#  x7 c5
# A4  v2     v5
# A5  v2[1]  v5[1]
# C   v28 v29
# C   v30 v31

BEGIN_FUNCTION xnn_f32_gemm${"inc" if INC else ""}_minmax_ukernel_4x8__asm_aarch64_neonfma_cortex_a53${"_prfm" if PREFETCH else ""}

        $if INC:
          # Load acc, params pointer
          LDP         x15, x8, [sp, 8]
        $else:
          # Load params pointer
          LDR         x8, [sp, 8]

        # Clamp A and C pointers
        CMP         x0, 2                   // if mr < 2
        ADD         x9, x3, x4              // a1 = a0 + a_stride
        ADD         x16, x6, x7             // c1 = c0 + cm_stride
        CSEL        x9, x3, x9, LO          //   a1 = a0
        CSEL        x16, x6, x16, LO        //   c1 = c0

        ADD         x10, x9, x4             // a2 = a1 + a_stride
        ADD         x17, x16, x7            // c2 = c1 + cm_stride
                                            // if mr <= 2
        CSEL        x10, x9, x10, LS        //   a2 = a1
        CSEL        x17, x16, x17, LS       //   c2 = c1

        CMP         x0, 4                   // if mr < 4
        ADD         x11, x10, x4            // a3 = a2 + a_stride
        ADD         x14, x17, x7            // c3 = c2 + cm_stride
        CSEL        x11, x10, x11, LO       //   a3 = a2
        CSEL        x14, x17, x14, LO       //   c3 = c2

        # Load min/max values
        LD2R        {v6.4s, v7.4s}, [x8]

        # Save d12-d15 on stack
        STP         d12, d13, [sp, -32]!
        STP         d14, d15, [sp, 16]

0:
        $if INC:
          # Load initial accumulators
          LDP         q20, q21, [x15], 32
          LDP         q22, q23, [x15], 32
          LDP         q24, q25, [x15], 32
          LDP         q26, q27, [x15], 32
          $if PREFETCH:
            PRFM        PLDL1KEEP,  [x3,  0]    // Prefetch A
          $if PREFETCH:
            PRFM        PLDL1KEEP,  [x3, 64]
          $if PREFETCH:
            PRFM        PLDL1KEEP,  [x9,  0]
          $if PREFETCH:
            PRFM        PLDL1KEEP,  [x9, 64]
          $if PREFETCH:
            PRFM        PLDL1KEEP, [x10,  0]
          $if PREFETCH:
            PRFM        PLDL1KEEP, [x10, 64]
          $if PREFETCH:
            PRFM        PLDL1KEEP, [x11,  0]
          $if PREFETCH:
            PRFM        PLDL1KEEP, [x11, 64]
          $if PREFETCH:
            PRFM        PLDL1KEEP, [x5,   0]    // Prefetch B
          $if PREFETCH:
            PRFM        PLDL1KEEP, [x5,  64]
          $if PREFETCH:
            PRFM        PLDL1KEEP, [x5, 128]
          $if PREFETCH:
            PRFM        PLDL1KEEP, [x5, 192]
        $else:
          # Load initial bias from w into accumulators
          LDP         q20, q21, [x5], 32
          MOV         v22.16b, v20.16b
          $if PREFETCH:
            PRFM        PLDL1KEEP,  [x3,  0]    // Prefetch A
          $if PREFETCH:
            PRFM        PLDL1KEEP,  [x3, 64]
          MOV         v23.16b, v21.16b
          $if PREFETCH:
            PRFM        PLDL1KEEP,  [x9,  0]
          $if PREFETCH:
            PRFM        PLDL1KEEP,  [x9, 64]
          MOV         v24.16b, v20.16b
          $if PREFETCH:
            PRFM        PLDL1KEEP, [x10,  0]
          $if PREFETCH:
            PRFM        PLDL1KEEP, [x10, 64]
          MOV         v25.16b, v21.16b
          $if PREFETCH:
            PRFM        PLDL1KEEP, [x11,  0]
          $if PREFETCH:
            PRFM        PLDL1KEEP, [x11, 64]
          MOV         v26.16b, v20.16b
          $if PREFETCH:
            PRFM        PLDL1KEEP, [x5,   0]    // Prefetch B
          MOV         v27.16b, v21.16b
          $if PREFETCH:
            PRFM        PLDL1KEEP, [x5,  64]
          $if PREFETCH:
            PRFM        PLDL1KEEP, [x5, 128]
          $if PREFETCH:
            PRFM        PLDL1KEEP, [x5, 192]

        # Is there at least 4 floats (16 bytes) for prologue + epilogue?
        SUBS        x0, x2, 16              // k = kc - 16
        B.LO        4f

        # Prologue - First group loads, no FMA
        LDR         d0, [x3], 8                 // a0
        LDP         q16, q17, [x5], 32          // b
        LDR         d1, [x10], 8                // a2
        LD1         {v0.d}[1],  [x9], 8         // a1
        LD1         {v1.d}[1], [x11], 8         // a3
        SUBS        x0, x0, 16
        LDR         q18, [x5], 16
        LDR         d19, [x5], 8
        LDR         x4, [x5], 8             // ins is in BLOCK 0

        # Is there at least 4 floats (16 bytes) for main loop?
        B.LO        2f

        # Main loop - 4 floats of A (16 bytes)
        # 32 FMA + 8 LD64 A + 8 LDR B
1:
        # First group of 16 FMA, Second group loads
        # BLOCK 0
        LDR         d3, [x3], 8                 // a0
        INS         v19.d[1], x4                // b from second group
        FMLA        v20.4s, v16.4s,  v0.s[0]
        LDR         x4, [x9], 8                 // a1
        FMLA        v22.4s, v16.4s,  v0.s[2]
        FMLA        v24.4s, v16.4s,  v1.s[0]

        # BLOCK 1
        LDR         d12, [x5]
        INS         v3.d[1], x4                 // a1 ins
        FMLA        v26.4s, v16.4s,  v1.s[2]
        LDR         x4, [x5, 8]             // b
        FMLA        v21.4s, v17.4s,  v0.s[0]
        FMLA        v23.4s, v17.4s,  v0.s[2]

        # BLOCK 2
        LDR         d4, [x10], 8                // a2
        INS         v12.d[1], x4            // b  ins
        FMLA        v25.4s, v17.4s,  v1.s[0]
        LDR         x4, [x11], 8                // a3
        FMLA        v27.4s, v17.4s,  v1.s[2]
        FMLA        v20.4s, v18.4s,  v0.s[1]

        # BLOCK 3
        LDR         d13, [x5, 16]
        INS         v4.d[1], x4                 // a3 ins
        FMLA        v22.4s, v18.4s,  v0.s[3]
        LDR         x4, [x5, 24]
        FMLA        v24.4s, v18.4s,  v1.s[1]
        FMLA        v26.4s, v18.4s,  v1.s[3]

        # BLOCK 4
        LDR         d14, [x5, 32]
        INS         v13.d[1], x4            // b
        FMLA        v21.4s, v19.4s,  v0.s[1]
        LDR         x4, [x5, 40]
        FMLA        v23.4s, v19.4s,  v0.s[3]
        FMLA        v25.4s, v19.4s,  v1.s[1]

        # BLOCK 5
        # NOPs to ensure 4 cycle LDR lands on next LDR
        LDR         d15, [x5, 48]
        INS         v14.d[1], x4            // b from previous
        FMLA        v27.4s, v19.4s,  v1.s[3]
        LDR         x4, [x5, 56]
        NOP
        NOP
        NOP
        NOP

        # Second group of 16 FMA, First group of loads
        # BLOCK 0
        LDR         d0, [x3], 8                 // a0
        INS         v15.d[1], x4            // b from previous
        FMLA        v20.4s, v12.4s,  v3.s[0]
        LDR         x4, [x9], 8                 // a1
        FMLA        v22.4s, v12.4s,  v3.s[2]
        FMLA        v24.4s, v12.4s,  v4.s[0]
        $if PREFETCH:
          PRFM        PLDL1KEEP, [x3, 128]        // Prefetch A0

        # BLOCK 1
        LDR         d16, [x5, 64]
        INS         v0.d[1], x4                 // a1 ins
        FMLA        v26.4s, v12.4s,  v4.s[2]
        LDR         x4, [x5, 72]            // b
        FMLA        v21.4s, v13.4s,  v3.s[0]
        FMLA        v23.4s, v13.4s,  v3.s[2]
        $if PREFETCH:
          PRFM        PLDL1KEEP, [x9, 128]        // Prefetch A1

        # BLOCK 2
        LDR         d1, [x10], 8                // a2
        INS         v16.d[1], x4            // b
        FMLA        v25.4s, v13.4s,  v4.s[0]
        LDR         x4, [x11], 8                // a3
        FMLA        v27.4s, v13.4s,  v4.s[2]
        FMLA        v20.4s, v14.4s,  v3.s[1]
        $if PREFETCH:
          PRFM        PLDL1KEEP, [x10, 128]       // Prefetch A2

        # BLOCK 3
        LDR         d17, [x5, 80]
        INS         v1.d[1], x4                 // a3 ins
        FMLA        v22.4s, v14.4s,  v3.s[3]
        LDR         x4, [x5, 88]
        FMLA        v24.4s, v14.4s,  v4.s[1]
        FMLA        v26.4s, v14.4s,  v4.s[3]
        $if PREFETCH:
          PRFM        PLDL1KEEP, [x11, 128]       // Prefetch A3

        # BLOCK 4
        LDR         d18, [x5, 96]
        INS         v17.d[1], x4            // b
        FMLA        v21.4s, v15.4s,  v3.s[1]
        LDR         x4, [x5, 104]
        FMLA        v23.4s, v15.4s,  v3.s[3]
        FMLA        v25.4s, v15.4s,  v4.s[1]
        $if PREFETCH:
          PRFM        PLDL1KEEP, [x5, 192]        // Prefetch B

        # BLOCK 5
        # NOTE that block needs to be 4 cycles for LDR not to stall
        LDR         d19, [x5, 112]
        INS         v18.d[1], x4
        FMLA        v27.4s, v15.4s,  v4.s[3]
        LDR         x4, [x5, 120]
        SUBS        x0, x0, 16
        $if PREFETCH:
          PRFM        PLDL1KEEP, [x5, 256]        // Prefetch B
        ADD         x5, x5, 128
        B.HS        1b

        # Epilogue - 4 floats of A (16 bytes)
        # 32 FMA + 8 LD64 A + 8 LDR B
2:
        # First group of 16 FMA, Second group loads
        # BLOCK 0
        LDR         d3, [x3], 8                 // a0
        INS         v19.d[1], x4                // b from second group
        FMLA        v20.4s, v16.4s,  v0.s[0]
        LDR         x4, [x9], 8                 // a1
        FMLA        v22.4s, v16.4s,  v0.s[2]
        FMLA        v24.4s, v16.4s,  v1.s[0]

        # BLOCK 1
        LDR         d12, [x5]
        INS         v3.d[1], x4                 // a1 ins
        FMLA        v26.4s, v16.4s,  v1.s[2]
        LDR         x4, [x5, 8]             // b
        FMLA        v21.4s, v17.4s,  v0.s[0]
        FMLA        v23.4s, v17.4s,  v0.s[2]

        # BLOCK 2
        LDR         d4, [x10], 8                // a2
        INS         v12.d[1], x4            // b  ins
        FMLA        v25.4s, v17.4s,  v1.s[0]
        LDR         x4, [x11], 8                // a3
        FMLA        v27.4s, v17.4s,  v1.s[2]
        FMLA        v20.4s, v18.4s,  v0.s[1]

        # BLOCK 3
        LDR         d13, [x5, 16]
        INS         v4.d[1], x4                 // a3 ins
        FMLA        v22.4s, v18.4s,  v0.s[3]
        LDR         x4, [x5, 24]
        FMLA        v24.4s, v18.4s,  v1.s[1]
        FMLA        v26.4s, v18.4s,  v1.s[3]

        # BLOCK 4
        LDR         d14, [x5, 32]
        INS         v13.d[1], x4            // b
        FMLA        v21.4s, v19.4s,  v0.s[1]
        LDR         x4, [x5, 40]
        FMLA        v23.4s, v19.4s,  v0.s[3]
        FMLA        v25.4s, v19.4s,  v1.s[1]

        # BLOCK 5
        # NOPs to ensure 4 cycle LDR lands on next LDR
        LDR         d15, [x5, 48]
        INS         v14.d[1], x4
        FMLA        v27.4s, v19.4s,  v1.s[3]
        LDR         x4, [x5, 56]
        NOP         // fma
        NOP
        NOP         // fma
        NOP

        # Second group of 16 FMA, no loads
        # BLOCK 0
        INS         v15.d[1], x4            // b from previous
        FMLA        v20.4s, v12.4s,  v3.s[0]
        FMLA        v22.4s, v12.4s,  v3.s[2]
        FMLA        v24.4s, v12.4s,  v4.s[0]

        # BLOCK 1
        FMLA        v26.4s, v12.4s,  v4.s[2]
        FMLA        v21.4s, v13.4s,  v3.s[0]
        FMLA        v23.4s, v13.4s,  v3.s[2]

        # BLOCK 2
        FMLA        v25.4s, v13.4s,  v4.s[0]
        FMLA        v27.4s, v13.4s,  v4.s[2]
        FMLA        v20.4s, v14.4s,  v3.s[1]

        # BLOCK 3
        FMLA        v22.4s, v14.4s,  v3.s[3]
        FMLA        v24.4s, v14.4s,  v4.s[1]
        FMLA        v26.4s, v14.4s,  v4.s[3]
        TST         x0, 15

        # BLOCK 4
        FMLA        v21.4s, v15.4s,  v3.s[1]
        FMLA        v23.4s, v15.4s,  v3.s[3]
        FMLA        v25.4s, v15.4s,  v4.s[1]
        ADD         x5, x5, 64

        # BLOCK 5
        FMLA        v27.4s, v15.4s,  v4.s[3]

        # Is there a remainder?- 2 floats of A (8 bytes) or less
        B.NE        4f

3:
        # Clamp
        FMAX        v20.4s, v20.4s, v6.4s
        # Load cn_stride
        LDR         x0, [sp, 32]
        FMAX        v21.4s, v21.4s, v6.4s
        FMAX        v22.4s, v22.4s, v6.4s
        FMAX        v23.4s, v23.4s, v6.4s
        FMAX        v24.4s, v24.4s, v6.4s
        FMAX        v25.4s, v25.4s, v6.4s
        FMAX        v26.4s, v26.4s, v6.4s
        FMAX        v27.4s, v27.4s, v6.4s
        SUBS        x1, x1, 8
        FMIN        v20.4s, v20.4s, v7.4s
        FMIN        v21.4s, v21.4s, v7.4s
        FMIN        v22.4s, v22.4s, v7.4s
        FMIN        v23.4s, v23.4s, v7.4s
        FMIN        v24.4s, v24.4s, v7.4s
        FMIN        v25.4s, v25.4s, v7.4s
        FMIN        v26.4s, v26.4s, v7.4s
        FMIN        v27.4s, v27.4s, v7.4s

        # Store full 4 x 8
        B.LO        6f

        $if INC:
          ST1         {v26.16b, v27.16b}, [x14], x0
          SUB         x3,  x3, x2             // a0 -= kc
          ST1         {v24.16b, v25.16b}, [x17], x0
          SUB         x9,  x9, x2             // a1 -= kc
          ST1         {v22.16b, v23.16b}, [x16], x0
          SUB         x10, x10, x2            // a2 -= kc
          ST1         {v20.16b, v21.16b},  [x6], x0
          SUB         x11, x11, x2            // a3 -= kc
        $else:
          ST1         {v20.16b, v21.16b},  [x6], x0
          SUB         x3,  x3, x2             // a0 -= kc
          ST1         {v22.16b, v23.16b}, [x16], x0
          SUB         x9,  x9, x2             // a1 -= kc
          ST1         {v24.16b, v25.16b}, [x17], x0
          SUB         x10, x10, x2            // a2 -= kc
          ST1         {v26.16b, v27.16b}, [x14], x0
          SUB         x11, x11, x2            // a3 -= kc

        B.HI        0b

        # Restore d12-d15 from stack
        LDP         d14, d15, [sp, 16]
        LDP         d12, d13, [sp], 32
        RET

4:
        # Is there a remainder?- 2 floats of A (8 bytes)
        TBZ         x0, 3, 5f

        # Remainder- 2 floats of A (8 bytes)
        LDR         d0,  [x3], 8
        LDR         q16, [x5], 16
        LD1         {v0.d}[1], [x9], 8
        LDR         d1, [x10], 8
        LD1         {v1.d}[1], [x11], 8
        LDR         q17, [x5], 16
        LDR         q18, [x5], 16
        LDR         q19, [x5], 16
        FMLA        v20.4s, v16.4s,  v0.s[0]
        FMLA        v22.4s, v16.4s,  v0.s[2]
        FMLA        v24.4s, v16.4s,  v1.s[0]
        FMLA        v26.4s, v16.4s,  v1.s[2]
        FMLA        v21.4s, v17.4s,  v0.s[0]
        FMLA        v23.4s, v17.4s,  v0.s[2]
        FMLA        v25.4s, v17.4s,  v1.s[0]
        FMLA        v27.4s, v17.4s,  v1.s[2]

        FMLA        v20.4s, v18.4s,  v0.s[1]
        FMLA        v22.4s, v18.4s,  v0.s[3]
        FMLA        v24.4s, v18.4s,  v1.s[1]
        FMLA        v26.4s, v18.4s,  v1.s[3]
        FMLA        v21.4s, v19.4s,  v0.s[1]
        FMLA        v23.4s, v19.4s,  v0.s[3]
        FMLA        v25.4s, v19.4s,  v1.s[1]
        FMLA        v27.4s, v19.4s,  v1.s[3]

        # Is there a remainder?- 1 float of A (4 bytes)
        TBZ         x0, 2, 3b

5:
        # Remainder- 1 float of A (4 bytes)
        LDR         s0,  [x3], 4
        LDR         q16, [x5], 16
        LD1         {v0.s}[2], [x9], 4
        LDR         s1, [x10], 4
        LD1         {v1.s}[2], [x11], 4
        LDR         q17, [x5], 16

        FMLA        v20.4s, v16.4s,  v0.s[0]
        FMLA        v22.4s, v16.4s,  v0.s[2]
        FMLA        v24.4s, v16.4s,  v1.s[0]
        FMLA        v26.4s, v16.4s,  v1.s[2]
        FMLA        v21.4s, v17.4s,  v0.s[0]
        FMLA        v23.4s, v17.4s,  v0.s[2]
        FMLA        v25.4s, v17.4s,  v1.s[0]
        FMLA        v27.4s, v17.4s,  v1.s[2]
        B           3b

        # Store odd width
6:
        TBZ         x1, 2, 7f
        $if INC:
          STR         q26, [x14], 16
          MOV         v26.16b, v27.16b
          STR         q24, [x17], 16
          MOV         v24.16b, v25.16b
          STR         q22, [x16], 16
          MOV         v22.16b, v23.16b
          STR         q20,  [x6], 16
          MOV         v20.16b, v21.16b
        $else:
          STR         q20,  [x6], 16
          MOV         v20.16b, v21.16b
          STR         q22, [x16], 16
          MOV         v22.16b, v23.16b
          STR         q24, [x17], 16
          MOV         v24.16b, v25.16b
          STR         q26, [x14], 16
          MOV         v26.16b, v27.16b

7:
        TBZ         x1, 1, 8f
        $if INC:
          STR         d26, [x14], 8
          STR         d24, [x17], 8
          DUP         d26, v26.d[1]
          DUP         d24, v24.d[1]
          STR         d22, [x16], 8
          STR         d20,  [x6], 8
          DUP         d22, v22.d[1]
          DUP         d20, v20.d[1]
        $else:
          STR         d20,  [x6], 8
          STR         d22, [x16], 8
          DUP         d20, v20.d[1]
          DUP         d22, v22.d[1]
          STR         d24, [x17], 8
          STR         d26, [x14], 8
          DUP         d24, v24.d[1]
          DUP         d26, v26.d[1]

8:
        TBZ         x1, 0, 9f
        $if INC:
          STR         s26, [x14]
          STR         s24, [x17]
          STR         s22, [x16]
          STR         s20,  [x6]
        $else:
          STR         s20,  [x6]
          STR         s22, [x16]
          STR         s24, [x17]
          STR         s26, [x14]
9:
        # Restore d12-d15 from stack
        LDP         d14, d15, [sp, 16]
        LDP         d12, d13, [sp], 32
        RET

END_FUNCTION xnn_f32_gemm${"inc" if INC else ""}_minmax_ukernel_4x8__asm_aarch64_neonfma_cortex_a53${"_prfm" if PREFETCH else ""}

#ifdef __ELF__
.section ".note.GNU-stack","",%progbits
#endif