xref: /aosp_15_r20/external/XNNPACK/src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-gemm/neon-shuffle.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2019 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9 
10 
11 #include <assert.h>
12 
13 #include <arm_neon.h>
14 
15 #include <xnnpack/gemm.h>
16 
17 
xnn_f32_gemminc_minmax_ukernel_8x8s4__neon(size_t mr,size_t nc,size_t kc,const float * restrict a,size_t a_stride,const float * restrict w,float * restrict c,size_t cm_stride,size_t cn_stride,const float * restrict acc,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_gemminc_minmax_ukernel_8x8s4__neon(
19     size_t mr,
20     size_t nc,
21     size_t kc,
22     const float* restrict a,
23     size_t a_stride,
24     const float* restrict w,
25     float* restrict c,
26     size_t cm_stride,
27     size_t cn_stride,
28     const float*restrict acc,
29     const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
30 {
31   assert(mr != 0);
32   assert(mr <= 8);
33   assert(nc != 0);
34   assert(kc != 0);
35   assert(kc % sizeof(float) == 0);
36   assert(a != NULL);
37   assert(w != NULL);
38   assert(c != NULL);
39   assert(acc != NULL);
40 
41   const float* a0 = a;
42   float* c0 = c;
43   const float* a1 = (const float*) ((uintptr_t) a0 + a_stride);
44   float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
45   if XNN_UNPREDICTABLE(mr < 2) {
46     a1 = a0;
47     c1 = c0;
48   }
49   const float* a2 = (const float*) ((uintptr_t) a1 + a_stride);
50   float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
51   if XNN_UNPREDICTABLE(mr <= 2) {
52     a2 = a1;
53     c2 = c1;
54   }
55   const float* a3 = (const float*) ((uintptr_t) a2 + a_stride);
56   float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
57   if XNN_UNPREDICTABLE(mr < 4) {
58     a3 = a2;
59     c3 = c2;
60   }
61   const float* a4 = (const float*) ((uintptr_t) a3 + a_stride);
62   float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
63   if XNN_UNPREDICTABLE(mr <= 4) {
64     a4 = a3;
65     c4 = c3;
66   }
67   const float* a5 = (const float*) ((uintptr_t) a4 + a_stride);
68   float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
69   if XNN_UNPREDICTABLE(mr < 6) {
70     a5 = a4;
71     c5 = c4;
72   }
73   const float* a6 = (const float*) ((uintptr_t) a5 + a_stride);
74   float* c6 = (float*) ((uintptr_t) c5 + cm_stride);
75   if XNN_UNPREDICTABLE(mr <= 6) {
76     a6 = a5;
77     c6 = c5;
78   }
79   const float* a7 = (const float*) ((uintptr_t) a6 + a_stride);
80   float* c7 = (float*) ((uintptr_t) c6 + cm_stride);
81   if XNN_UNPREDICTABLE(mr != 8) {
82     a7 = a6;
83     c7 = c6;
84   }
85 
86   do {
87     float32x4_t vacc0x0123 = vld1q_f32(acc); acc += 4;
88     float32x4_t vacc0x4567 = vld1q_f32(acc); acc += 4;
89     float32x4_t vacc1x0123 = vld1q_f32(acc); acc += 4;
90     float32x4_t vacc1x4567 = vld1q_f32(acc); acc += 4;
91     float32x4_t vacc2x0123 = vld1q_f32(acc); acc += 4;
92     float32x4_t vacc2x4567 = vld1q_f32(acc); acc += 4;
93     float32x4_t vacc3x0123 = vld1q_f32(acc); acc += 4;
94     float32x4_t vacc3x4567 = vld1q_f32(acc); acc += 4;
95     float32x4_t vacc4x0123 = vld1q_f32(acc); acc += 4;
96     float32x4_t vacc4x4567 = vld1q_f32(acc); acc += 4;
97     float32x4_t vacc5x0123 = vld1q_f32(acc); acc += 4;
98     float32x4_t vacc5x4567 = vld1q_f32(acc); acc += 4;
99     float32x4_t vacc6x0123 = vld1q_f32(acc); acc += 4;
100     float32x4_t vacc6x4567 = vld1q_f32(acc); acc += 4;
101     float32x4_t vacc7x0123 = vld1q_f32(acc); acc += 4;
102     float32x4_t vacc7x4567 = vld1q_f32(acc); acc += 4;
103 
104     size_t k = kc;
105     while (k >= 4 * sizeof(float)) {
106       float32x4_t va0 = vld1q_f32(a0); a0 += 4;
107       float32x4_t va1 = vld1q_f32(a1); a1 += 4;
108       float32x4_t va2 = vld1q_f32(a2); a2 += 4;
109       float32x4_t va3 = vld1q_f32(a3); a3 += 4;
110       float32x4_t va4 = vld1q_f32(a4); a4 += 4;
111       float32x4_t va5 = vld1q_f32(a5); a5 += 4;
112       float32x4_t va6 = vld1q_f32(a6); a6 += 4;
113       float32x4_t va7 = vld1q_f32(a7); a7 += 4;
114 
115 
116       const float32x4_t vb0123c0 = vld1q_f32(w); w += 4;
117       const float32x4_t vb4567c0 = vld1q_f32(w); w += 4;
118 
119       vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123c0);
120       vacc1x0123 = vmlaq_f32(vacc1x0123, va1, vb0123c0);
121       vacc2x0123 = vmlaq_f32(vacc2x0123, va2, vb0123c0);
122       vacc3x0123 = vmlaq_f32(vacc3x0123, va3, vb0123c0);
123       vacc4x0123 = vmlaq_f32(vacc4x0123, va4, vb0123c0);
124       vacc5x0123 = vmlaq_f32(vacc5x0123, va5, vb0123c0);
125       vacc6x0123 = vmlaq_f32(vacc6x0123, va6, vb0123c0);
126       vacc7x0123 = vmlaq_f32(vacc7x0123, va7, vb0123c0);
127       vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567c0);
128       vacc1x4567 = vmlaq_f32(vacc1x4567, va1, vb4567c0);
129       vacc2x4567 = vmlaq_f32(vacc2x4567, va2, vb4567c0);
130       vacc3x4567 = vmlaq_f32(vacc3x4567, va3, vb4567c0);
131       vacc4x4567 = vmlaq_f32(vacc4x4567, va4, vb4567c0);
132       vacc5x4567 = vmlaq_f32(vacc5x4567, va5, vb4567c0);
133       vacc6x4567 = vmlaq_f32(vacc6x4567, va6, vb4567c0);
134       vacc7x4567 = vmlaq_f32(vacc7x4567, va7, vb4567c0);
135 
136       va0 = vextq_f32(va0, va0, 1);
137       va1 = vextq_f32(va1, va1, 1);
138       va2 = vextq_f32(va2, va2, 1);
139       va3 = vextq_f32(va3, va3, 1);
140       va4 = vextq_f32(va4, va4, 1);
141       va5 = vextq_f32(va5, va5, 1);
142       va6 = vextq_f32(va6, va6, 1);
143       va7 = vextq_f32(va7, va7, 1);
144 
145       const float32x4_t vb0123c1 = vld1q_f32(w); w += 4;
146       const float32x4_t vb4567c1 = vld1q_f32(w); w += 4;
147 
148       vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123c1);
149       vacc1x0123 = vmlaq_f32(vacc1x0123, va1, vb0123c1);
150       vacc2x0123 = vmlaq_f32(vacc2x0123, va2, vb0123c1);
151       vacc3x0123 = vmlaq_f32(vacc3x0123, va3, vb0123c1);
152       vacc4x0123 = vmlaq_f32(vacc4x0123, va4, vb0123c1);
153       vacc5x0123 = vmlaq_f32(vacc5x0123, va5, vb0123c1);
154       vacc6x0123 = vmlaq_f32(vacc6x0123, va6, vb0123c1);
155       vacc7x0123 = vmlaq_f32(vacc7x0123, va7, vb0123c1);
156       vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567c1);
157       vacc1x4567 = vmlaq_f32(vacc1x4567, va1, vb4567c1);
158       vacc2x4567 = vmlaq_f32(vacc2x4567, va2, vb4567c1);
159       vacc3x4567 = vmlaq_f32(vacc3x4567, va3, vb4567c1);
160       vacc4x4567 = vmlaq_f32(vacc4x4567, va4, vb4567c1);
161       vacc5x4567 = vmlaq_f32(vacc5x4567, va5, vb4567c1);
162       vacc6x4567 = vmlaq_f32(vacc6x4567, va6, vb4567c1);
163       vacc7x4567 = vmlaq_f32(vacc7x4567, va7, vb4567c1);
164 
165       va0 = vextq_f32(va0, va0, 1);
166       va1 = vextq_f32(va1, va1, 1);
167       va2 = vextq_f32(va2, va2, 1);
168       va3 = vextq_f32(va3, va3, 1);
169       va4 = vextq_f32(va4, va4, 1);
170       va5 = vextq_f32(va5, va5, 1);
171       va6 = vextq_f32(va6, va6, 1);
172       va7 = vextq_f32(va7, va7, 1);
173 
174       const float32x4_t vb0123c2 = vld1q_f32(w); w += 4;
175       const float32x4_t vb4567c2 = vld1q_f32(w); w += 4;
176 
177       vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123c2);
178       vacc1x0123 = vmlaq_f32(vacc1x0123, va1, vb0123c2);
179       vacc2x0123 = vmlaq_f32(vacc2x0123, va2, vb0123c2);
180       vacc3x0123 = vmlaq_f32(vacc3x0123, va3, vb0123c2);
181       vacc4x0123 = vmlaq_f32(vacc4x0123, va4, vb0123c2);
182       vacc5x0123 = vmlaq_f32(vacc5x0123, va5, vb0123c2);
183       vacc6x0123 = vmlaq_f32(vacc6x0123, va6, vb0123c2);
184       vacc7x0123 = vmlaq_f32(vacc7x0123, va7, vb0123c2);
185       vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567c2);
186       vacc1x4567 = vmlaq_f32(vacc1x4567, va1, vb4567c2);
187       vacc2x4567 = vmlaq_f32(vacc2x4567, va2, vb4567c2);
188       vacc3x4567 = vmlaq_f32(vacc3x4567, va3, vb4567c2);
189       vacc4x4567 = vmlaq_f32(vacc4x4567, va4, vb4567c2);
190       vacc5x4567 = vmlaq_f32(vacc5x4567, va5, vb4567c2);
191       vacc6x4567 = vmlaq_f32(vacc6x4567, va6, vb4567c2);
192       vacc7x4567 = vmlaq_f32(vacc7x4567, va7, vb4567c2);
193 
194       va0 = vextq_f32(va0, va0, 1);
195       va1 = vextq_f32(va1, va1, 1);
196       va2 = vextq_f32(va2, va2, 1);
197       va3 = vextq_f32(va3, va3, 1);
198       va4 = vextq_f32(va4, va4, 1);
199       va5 = vextq_f32(va5, va5, 1);
200       va6 = vextq_f32(va6, va6, 1);
201       va7 = vextq_f32(va7, va7, 1);
202 
203       const float32x4_t vb0123c3 = vld1q_f32(w); w += 4;
204       const float32x4_t vb4567c3 = vld1q_f32(w); w += 4;
205 
206       vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123c3);
207       vacc1x0123 = vmlaq_f32(vacc1x0123, va1, vb0123c3);
208       vacc2x0123 = vmlaq_f32(vacc2x0123, va2, vb0123c3);
209       vacc3x0123 = vmlaq_f32(vacc3x0123, va3, vb0123c3);
210       vacc4x0123 = vmlaq_f32(vacc4x0123, va4, vb0123c3);
211       vacc5x0123 = vmlaq_f32(vacc5x0123, va5, vb0123c3);
212       vacc6x0123 = vmlaq_f32(vacc6x0123, va6, vb0123c3);
213       vacc7x0123 = vmlaq_f32(vacc7x0123, va7, vb0123c3);
214       vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567c3);
215       vacc1x4567 = vmlaq_f32(vacc1x4567, va1, vb4567c3);
216       vacc2x4567 = vmlaq_f32(vacc2x4567, va2, vb4567c3);
217       vacc3x4567 = vmlaq_f32(vacc3x4567, va3, vb4567c3);
218       vacc4x4567 = vmlaq_f32(vacc4x4567, va4, vb4567c3);
219       vacc5x4567 = vmlaq_f32(vacc5x4567, va5, vb4567c3);
220       vacc6x4567 = vmlaq_f32(vacc6x4567, va6, vb4567c3);
221       vacc7x4567 = vmlaq_f32(vacc7x4567, va7, vb4567c3);
222 
223 
224       k -= 4 * sizeof(float);
225     }
226     if XNN_UNLIKELY(k != 0) {
227       float32x4_t va0 = vld1q_f32(a0); a0 = (const float*) ((uintptr_t) a0 + k);
228       float32x4_t va1 = vld1q_f32(a1); a1 = (const float*) ((uintptr_t) a1 + k);
229       float32x4_t va2 = vld1q_f32(a2); a2 = (const float*) ((uintptr_t) a2 + k);
230       float32x4_t va3 = vld1q_f32(a3); a3 = (const float*) ((uintptr_t) a3 + k);
231       float32x4_t va4 = vld1q_f32(a4); a4 = (const float*) ((uintptr_t) a4 + k);
232       float32x4_t va5 = vld1q_f32(a5); a5 = (const float*) ((uintptr_t) a5 + k);
233       float32x4_t va6 = vld1q_f32(a6); a6 = (const float*) ((uintptr_t) a6 + k);
234       float32x4_t va7 = vld1q_f32(a7); a7 = (const float*) ((uintptr_t) a7 + k);
235 
236 
237       const float32x4_t vb0123c0 = vld1q_f32(w); w += 4;
238       const float32x4_t vb4567c0 = vld1q_f32(w); w += 4;
239 
240       const float32x4_t vmska0x0123c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb0123c0, vmovq_n_f32(0.0f))));
241       vacc0x0123 = vmlaq_f32(vacc0x0123, vmska0x0123c0, vb0123c0);
242       const float32x4_t vmska1x0123c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb0123c0, vmovq_n_f32(0.0f))));
243       vacc1x0123 = vmlaq_f32(vacc1x0123, vmska1x0123c0, vb0123c0);
244       const float32x4_t vmska2x0123c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb0123c0, vmovq_n_f32(0.0f))));
245       vacc2x0123 = vmlaq_f32(vacc2x0123, vmska2x0123c0, vb0123c0);
246       const float32x4_t vmska3x0123c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb0123c0, vmovq_n_f32(0.0f))));
247       vacc3x0123 = vmlaq_f32(vacc3x0123, vmska3x0123c0, vb0123c0);
248       const float32x4_t vmska4x0123c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va4), vceqq_f32(vb0123c0, vmovq_n_f32(0.0f))));
249       vacc4x0123 = vmlaq_f32(vacc4x0123, vmska4x0123c0, vb0123c0);
250       const float32x4_t vmska5x0123c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va5), vceqq_f32(vb0123c0, vmovq_n_f32(0.0f))));
251       vacc5x0123 = vmlaq_f32(vacc5x0123, vmska5x0123c0, vb0123c0);
252       const float32x4_t vmska6x0123c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va6), vceqq_f32(vb0123c0, vmovq_n_f32(0.0f))));
253       vacc6x0123 = vmlaq_f32(vacc6x0123, vmska6x0123c0, vb0123c0);
254       const float32x4_t vmska7x0123c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va7), vceqq_f32(vb0123c0, vmovq_n_f32(0.0f))));
255       vacc7x0123 = vmlaq_f32(vacc7x0123, vmska7x0123c0, vb0123c0);
256       const float32x4_t vmska0x4567c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb4567c0, vmovq_n_f32(0.0f))));
257       vacc0x4567 = vmlaq_f32(vacc0x4567, vmska0x4567c0, vb4567c0);
258       const float32x4_t vmska1x4567c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb4567c0, vmovq_n_f32(0.0f))));
259       vacc1x4567 = vmlaq_f32(vacc1x4567, vmska1x4567c0, vb4567c0);
260       const float32x4_t vmska2x4567c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb4567c0, vmovq_n_f32(0.0f))));
261       vacc2x4567 = vmlaq_f32(vacc2x4567, vmska2x4567c0, vb4567c0);
262       const float32x4_t vmska3x4567c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb4567c0, vmovq_n_f32(0.0f))));
263       vacc3x4567 = vmlaq_f32(vacc3x4567, vmska3x4567c0, vb4567c0);
264       const float32x4_t vmska4x4567c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va4), vceqq_f32(vb4567c0, vmovq_n_f32(0.0f))));
265       vacc4x4567 = vmlaq_f32(vacc4x4567, vmska4x4567c0, vb4567c0);
266       const float32x4_t vmska5x4567c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va5), vceqq_f32(vb4567c0, vmovq_n_f32(0.0f))));
267       vacc5x4567 = vmlaq_f32(vacc5x4567, vmska5x4567c0, vb4567c0);
268       const float32x4_t vmska6x4567c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va6), vceqq_f32(vb4567c0, vmovq_n_f32(0.0f))));
269       vacc6x4567 = vmlaq_f32(vacc6x4567, vmska6x4567c0, vb4567c0);
270       const float32x4_t vmska7x4567c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va7), vceqq_f32(vb4567c0, vmovq_n_f32(0.0f))));
271       vacc7x4567 = vmlaq_f32(vacc7x4567, vmska7x4567c0, vb4567c0);
272 
273       va0 = vextq_f32(va0, va0, 1);
274       va1 = vextq_f32(va1, va1, 1);
275       va2 = vextq_f32(va2, va2, 1);
276       va3 = vextq_f32(va3, va3, 1);
277       va4 = vextq_f32(va4, va4, 1);
278       va5 = vextq_f32(va5, va5, 1);
279       va6 = vextq_f32(va6, va6, 1);
280       va7 = vextq_f32(va7, va7, 1);
281 
282       const float32x4_t vb0123c1 = vld1q_f32(w); w += 4;
283       const float32x4_t vb4567c1 = vld1q_f32(w); w += 4;
284 
285       const float32x4_t vmska0x0123c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb0123c1, vmovq_n_f32(0.0f))));
286       vacc0x0123 = vmlaq_f32(vacc0x0123, vmska0x0123c1, vb0123c1);
287       const float32x4_t vmska1x0123c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb0123c1, vmovq_n_f32(0.0f))));
288       vacc1x0123 = vmlaq_f32(vacc1x0123, vmska1x0123c1, vb0123c1);
289       const float32x4_t vmska2x0123c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb0123c1, vmovq_n_f32(0.0f))));
290       vacc2x0123 = vmlaq_f32(vacc2x0123, vmska2x0123c1, vb0123c1);
291       const float32x4_t vmska3x0123c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb0123c1, vmovq_n_f32(0.0f))));
292       vacc3x0123 = vmlaq_f32(vacc3x0123, vmska3x0123c1, vb0123c1);
293       const float32x4_t vmska4x0123c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va4), vceqq_f32(vb0123c1, vmovq_n_f32(0.0f))));
294       vacc4x0123 = vmlaq_f32(vacc4x0123, vmska4x0123c1, vb0123c1);
295       const float32x4_t vmska5x0123c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va5), vceqq_f32(vb0123c1, vmovq_n_f32(0.0f))));
296       vacc5x0123 = vmlaq_f32(vacc5x0123, vmska5x0123c1, vb0123c1);
297       const float32x4_t vmska6x0123c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va6), vceqq_f32(vb0123c1, vmovq_n_f32(0.0f))));
298       vacc6x0123 = vmlaq_f32(vacc6x0123, vmska6x0123c1, vb0123c1);
299       const float32x4_t vmska7x0123c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va7), vceqq_f32(vb0123c1, vmovq_n_f32(0.0f))));
300       vacc7x0123 = vmlaq_f32(vacc7x0123, vmska7x0123c1, vb0123c1);
301       const float32x4_t vmska0x4567c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb4567c1, vmovq_n_f32(0.0f))));
302       vacc0x4567 = vmlaq_f32(vacc0x4567, vmska0x4567c1, vb4567c1);
303       const float32x4_t vmska1x4567c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb4567c1, vmovq_n_f32(0.0f))));
304       vacc1x4567 = vmlaq_f32(vacc1x4567, vmska1x4567c1, vb4567c1);
305       const float32x4_t vmska2x4567c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb4567c1, vmovq_n_f32(0.0f))));
306       vacc2x4567 = vmlaq_f32(vacc2x4567, vmska2x4567c1, vb4567c1);
307       const float32x4_t vmska3x4567c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb4567c1, vmovq_n_f32(0.0f))));
308       vacc3x4567 = vmlaq_f32(vacc3x4567, vmska3x4567c1, vb4567c1);
309       const float32x4_t vmska4x4567c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va4), vceqq_f32(vb4567c1, vmovq_n_f32(0.0f))));
310       vacc4x4567 = vmlaq_f32(vacc4x4567, vmska4x4567c1, vb4567c1);
311       const float32x4_t vmska5x4567c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va5), vceqq_f32(vb4567c1, vmovq_n_f32(0.0f))));
312       vacc5x4567 = vmlaq_f32(vacc5x4567, vmska5x4567c1, vb4567c1);
313       const float32x4_t vmska6x4567c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va6), vceqq_f32(vb4567c1, vmovq_n_f32(0.0f))));
314       vacc6x4567 = vmlaq_f32(vacc6x4567, vmska6x4567c1, vb4567c1);
315       const float32x4_t vmska7x4567c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va7), vceqq_f32(vb4567c1, vmovq_n_f32(0.0f))));
316       vacc7x4567 = vmlaq_f32(vacc7x4567, vmska7x4567c1, vb4567c1);
317 
318       va0 = vextq_f32(va0, va0, 1);
319       va1 = vextq_f32(va1, va1, 1);
320       va2 = vextq_f32(va2, va2, 1);
321       va3 = vextq_f32(va3, va3, 1);
322       va4 = vextq_f32(va4, va4, 1);
323       va5 = vextq_f32(va5, va5, 1);
324       va6 = vextq_f32(va6, va6, 1);
325       va7 = vextq_f32(va7, va7, 1);
326 
327       const float32x4_t vb0123c2 = vld1q_f32(w); w += 4;
328       const float32x4_t vb4567c2 = vld1q_f32(w); w += 4;
329 
330       const float32x4_t vmska0x0123c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb0123c2, vmovq_n_f32(0.0f))));
331       vacc0x0123 = vmlaq_f32(vacc0x0123, vmska0x0123c2, vb0123c2);
332       const float32x4_t vmska1x0123c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb0123c2, vmovq_n_f32(0.0f))));
333       vacc1x0123 = vmlaq_f32(vacc1x0123, vmska1x0123c2, vb0123c2);
334       const float32x4_t vmska2x0123c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb0123c2, vmovq_n_f32(0.0f))));
335       vacc2x0123 = vmlaq_f32(vacc2x0123, vmska2x0123c2, vb0123c2);
336       const float32x4_t vmska3x0123c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb0123c2, vmovq_n_f32(0.0f))));
337       vacc3x0123 = vmlaq_f32(vacc3x0123, vmska3x0123c2, vb0123c2);
338       const float32x4_t vmska4x0123c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va4), vceqq_f32(vb0123c2, vmovq_n_f32(0.0f))));
339       vacc4x0123 = vmlaq_f32(vacc4x0123, vmska4x0123c2, vb0123c2);
340       const float32x4_t vmska5x0123c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va5), vceqq_f32(vb0123c2, vmovq_n_f32(0.0f))));
341       vacc5x0123 = vmlaq_f32(vacc5x0123, vmska5x0123c2, vb0123c2);
342       const float32x4_t vmska6x0123c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va6), vceqq_f32(vb0123c2, vmovq_n_f32(0.0f))));
343       vacc6x0123 = vmlaq_f32(vacc6x0123, vmska6x0123c2, vb0123c2);
344       const float32x4_t vmska7x0123c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va7), vceqq_f32(vb0123c2, vmovq_n_f32(0.0f))));
345       vacc7x0123 = vmlaq_f32(vacc7x0123, vmska7x0123c2, vb0123c2);
346       const float32x4_t vmska0x4567c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb4567c2, vmovq_n_f32(0.0f))));
347       vacc0x4567 = vmlaq_f32(vacc0x4567, vmska0x4567c2, vb4567c2);
348       const float32x4_t vmska1x4567c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb4567c2, vmovq_n_f32(0.0f))));
349       vacc1x4567 = vmlaq_f32(vacc1x4567, vmska1x4567c2, vb4567c2);
350       const float32x4_t vmska2x4567c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb4567c2, vmovq_n_f32(0.0f))));
351       vacc2x4567 = vmlaq_f32(vacc2x4567, vmska2x4567c2, vb4567c2);
352       const float32x4_t vmska3x4567c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb4567c2, vmovq_n_f32(0.0f))));
353       vacc3x4567 = vmlaq_f32(vacc3x4567, vmska3x4567c2, vb4567c2);
354       const float32x4_t vmska4x4567c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va4), vceqq_f32(vb4567c2, vmovq_n_f32(0.0f))));
355       vacc4x4567 = vmlaq_f32(vacc4x4567, vmska4x4567c2, vb4567c2);
356       const float32x4_t vmska5x4567c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va5), vceqq_f32(vb4567c2, vmovq_n_f32(0.0f))));
357       vacc5x4567 = vmlaq_f32(vacc5x4567, vmska5x4567c2, vb4567c2);
358       const float32x4_t vmska6x4567c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va6), vceqq_f32(vb4567c2, vmovq_n_f32(0.0f))));
359       vacc6x4567 = vmlaq_f32(vacc6x4567, vmska6x4567c2, vb4567c2);
360       const float32x4_t vmska7x4567c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va7), vceqq_f32(vb4567c2, vmovq_n_f32(0.0f))));
361       vacc7x4567 = vmlaq_f32(vacc7x4567, vmska7x4567c2, vb4567c2);
362 
363       va0 = vextq_f32(va0, va0, 1);
364       va1 = vextq_f32(va1, va1, 1);
365       va2 = vextq_f32(va2, va2, 1);
366       va3 = vextq_f32(va3, va3, 1);
367       va4 = vextq_f32(va4, va4, 1);
368       va5 = vextq_f32(va5, va5, 1);
369       va6 = vextq_f32(va6, va6, 1);
370       va7 = vextq_f32(va7, va7, 1);
371 
372       const float32x4_t vb0123c3 = vld1q_f32(w); w += 4;
373       const float32x4_t vb4567c3 = vld1q_f32(w); w += 4;
374 
375       const float32x4_t vmska0x0123c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb0123c3, vmovq_n_f32(0.0f))));
376       vacc0x0123 = vmlaq_f32(vacc0x0123, vmska0x0123c3, vb0123c3);
377       const float32x4_t vmska1x0123c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb0123c3, vmovq_n_f32(0.0f))));
378       vacc1x0123 = vmlaq_f32(vacc1x0123, vmska1x0123c3, vb0123c3);
379       const float32x4_t vmska2x0123c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb0123c3, vmovq_n_f32(0.0f))));
380       vacc2x0123 = vmlaq_f32(vacc2x0123, vmska2x0123c3, vb0123c3);
381       const float32x4_t vmska3x0123c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb0123c3, vmovq_n_f32(0.0f))));
382       vacc3x0123 = vmlaq_f32(vacc3x0123, vmska3x0123c3, vb0123c3);
383       const float32x4_t vmska4x0123c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va4), vceqq_f32(vb0123c3, vmovq_n_f32(0.0f))));
384       vacc4x0123 = vmlaq_f32(vacc4x0123, vmska4x0123c3, vb0123c3);
385       const float32x4_t vmska5x0123c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va5), vceqq_f32(vb0123c3, vmovq_n_f32(0.0f))));
386       vacc5x0123 = vmlaq_f32(vacc5x0123, vmska5x0123c3, vb0123c3);
387       const float32x4_t vmska6x0123c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va6), vceqq_f32(vb0123c3, vmovq_n_f32(0.0f))));
388       vacc6x0123 = vmlaq_f32(vacc6x0123, vmska6x0123c3, vb0123c3);
389       const float32x4_t vmska7x0123c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va7), vceqq_f32(vb0123c3, vmovq_n_f32(0.0f))));
390       vacc7x0123 = vmlaq_f32(vacc7x0123, vmska7x0123c3, vb0123c3);
391       const float32x4_t vmska0x4567c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb4567c3, vmovq_n_f32(0.0f))));
392       vacc0x4567 = vmlaq_f32(vacc0x4567, vmska0x4567c3, vb4567c3);
393       const float32x4_t vmska1x4567c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb4567c3, vmovq_n_f32(0.0f))));
394       vacc1x4567 = vmlaq_f32(vacc1x4567, vmska1x4567c3, vb4567c3);
395       const float32x4_t vmska2x4567c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb4567c3, vmovq_n_f32(0.0f))));
396       vacc2x4567 = vmlaq_f32(vacc2x4567, vmska2x4567c3, vb4567c3);
397       const float32x4_t vmska3x4567c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb4567c3, vmovq_n_f32(0.0f))));
398       vacc3x4567 = vmlaq_f32(vacc3x4567, vmska3x4567c3, vb4567c3);
399       const float32x4_t vmska4x4567c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va4), vceqq_f32(vb4567c3, vmovq_n_f32(0.0f))));
400       vacc4x4567 = vmlaq_f32(vacc4x4567, vmska4x4567c3, vb4567c3);
401       const float32x4_t vmska5x4567c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va5), vceqq_f32(vb4567c3, vmovq_n_f32(0.0f))));
402       vacc5x4567 = vmlaq_f32(vacc5x4567, vmska5x4567c3, vb4567c3);
403       const float32x4_t vmska6x4567c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va6), vceqq_f32(vb4567c3, vmovq_n_f32(0.0f))));
404       vacc6x4567 = vmlaq_f32(vacc6x4567, vmska6x4567c3, vb4567c3);
405       const float32x4_t vmska7x4567c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va7), vceqq_f32(vb4567c3, vmovq_n_f32(0.0f))));
406       vacc7x4567 = vmlaq_f32(vacc7x4567, vmska7x4567c3, vb4567c3);
407 
408     }
409     const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max);
410     vacc0x0123 = vminq_f32(vacc0x0123, vmax);
411     vacc1x0123 = vminq_f32(vacc1x0123, vmax);
412     vacc2x0123 = vminq_f32(vacc2x0123, vmax);
413     vacc3x0123 = vminq_f32(vacc3x0123, vmax);
414     vacc4x0123 = vminq_f32(vacc4x0123, vmax);
415     vacc5x0123 = vminq_f32(vacc5x0123, vmax);
416     vacc6x0123 = vminq_f32(vacc6x0123, vmax);
417     vacc7x0123 = vminq_f32(vacc7x0123, vmax);
418     vacc0x4567 = vminq_f32(vacc0x4567, vmax);
419     vacc1x4567 = vminq_f32(vacc1x4567, vmax);
420     vacc2x4567 = vminq_f32(vacc2x4567, vmax);
421     vacc3x4567 = vminq_f32(vacc3x4567, vmax);
422     vacc4x4567 = vminq_f32(vacc4x4567, vmax);
423     vacc5x4567 = vminq_f32(vacc5x4567, vmax);
424     vacc6x4567 = vminq_f32(vacc6x4567, vmax);
425     vacc7x4567 = vminq_f32(vacc7x4567, vmax);
426 
427     const float32x4_t vmin = vld1q_dup_f32(&params->scalar.min);
428     vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
429     vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
430     vacc2x0123 = vmaxq_f32(vacc2x0123, vmin);
431     vacc3x0123 = vmaxq_f32(vacc3x0123, vmin);
432     vacc4x0123 = vmaxq_f32(vacc4x0123, vmin);
433     vacc5x0123 = vmaxq_f32(vacc5x0123, vmin);
434     vacc6x0123 = vmaxq_f32(vacc6x0123, vmin);
435     vacc7x0123 = vmaxq_f32(vacc7x0123, vmin);
436     vacc0x4567 = vmaxq_f32(vacc0x4567, vmin);
437     vacc1x4567 = vmaxq_f32(vacc1x4567, vmin);
438     vacc2x4567 = vmaxq_f32(vacc2x4567, vmin);
439     vacc3x4567 = vmaxq_f32(vacc3x4567, vmin);
440     vacc4x4567 = vmaxq_f32(vacc4x4567, vmin);
441     vacc5x4567 = vmaxq_f32(vacc5x4567, vmin);
442     vacc6x4567 = vmaxq_f32(vacc6x4567, vmin);
443     vacc7x4567 = vmaxq_f32(vacc7x4567, vmin);
444 
445     if XNN_LIKELY(nc >= 8) {
446       vst1q_f32(c7, vacc7x0123);
447       vst1q_f32(c7 + 4, vacc7x4567);
448       c7 = (float*) ((uintptr_t) c7 + cn_stride);
449       vst1q_f32(c6, vacc6x0123);
450       vst1q_f32(c6 + 4, vacc6x4567);
451       c6 = (float*) ((uintptr_t) c6 + cn_stride);
452       vst1q_f32(c5, vacc5x0123);
453       vst1q_f32(c5 + 4, vacc5x4567);
454       c5 = (float*) ((uintptr_t) c5 + cn_stride);
455       vst1q_f32(c4, vacc4x0123);
456       vst1q_f32(c4 + 4, vacc4x4567);
457       c4 = (float*) ((uintptr_t) c4 + cn_stride);
458       vst1q_f32(c3, vacc3x0123);
459       vst1q_f32(c3 + 4, vacc3x4567);
460       c3 = (float*) ((uintptr_t) c3 + cn_stride);
461       vst1q_f32(c2, vacc2x0123);
462       vst1q_f32(c2 + 4, vacc2x4567);
463       c2 = (float*) ((uintptr_t) c2 + cn_stride);
464       vst1q_f32(c1, vacc1x0123);
465       vst1q_f32(c1 + 4, vacc1x4567);
466       c1 = (float*) ((uintptr_t) c1 + cn_stride);
467       vst1q_f32(c0, vacc0x0123);
468       vst1q_f32(c0 + 4, vacc0x4567);
469       c0 = (float*) ((uintptr_t) c0 + cn_stride);
470 
471       a7 = (const float*) ((uintptr_t) a7 - kc);
472       a6 = (const float*) ((uintptr_t) a6 - kc);
473       a5 = (const float*) ((uintptr_t) a5 - kc);
474       a4 = (const float*) ((uintptr_t) a4 - kc);
475       a3 = (const float*) ((uintptr_t) a3 - kc);
476       a2 = (const float*) ((uintptr_t) a2 - kc);
477       a1 = (const float*) ((uintptr_t) a1 - kc);
478       a0 = (const float*) ((uintptr_t) a0 - kc);
479 
480       nc -= 8;
481 
482     } else {
483       if (nc & 4) {
484         vst1q_f32(c7, vacc7x0123); c7 += 4;
485         vst1q_f32(c6, vacc6x0123); c6 += 4;
486         vst1q_f32(c5, vacc5x0123); c5 += 4;
487         vst1q_f32(c4, vacc4x0123); c4 += 4;
488         vst1q_f32(c3, vacc3x0123); c3 += 4;
489         vst1q_f32(c2, vacc2x0123); c2 += 4;
490         vst1q_f32(c1, vacc1x0123); c1 += 4;
491         vst1q_f32(c0, vacc0x0123); c0 += 4;
492 
493         vacc7x0123 = vacc7x4567;
494         vacc6x0123 = vacc6x4567;
495         vacc5x0123 = vacc5x4567;
496         vacc4x0123 = vacc4x4567;
497         vacc3x0123 = vacc3x4567;
498         vacc2x0123 = vacc2x4567;
499         vacc1x0123 = vacc1x4567;
500         vacc0x0123 = vacc0x4567;
501       }
502       float32x2_t vacc7x01 = vget_low_f32(vacc7x0123);
503       float32x2_t vacc6x01 = vget_low_f32(vacc6x0123);
504       float32x2_t vacc5x01 = vget_low_f32(vacc5x0123);
505       float32x2_t vacc4x01 = vget_low_f32(vacc4x0123);
506       float32x2_t vacc3x01 = vget_low_f32(vacc3x0123);
507       float32x2_t vacc2x01 = vget_low_f32(vacc2x0123);
508       float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);
509       float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
510       if (nc & 2) {
511         vst1_f32(c7, vacc7x01); c7 += 2;
512         vst1_f32(c6, vacc6x01); c6 += 2;
513         vst1_f32(c5, vacc5x01); c5 += 2;
514         vst1_f32(c4, vacc4x01); c4 += 2;
515         vst1_f32(c3, vacc3x01); c3 += 2;
516         vst1_f32(c2, vacc2x01); c2 += 2;
517         vst1_f32(c1, vacc1x01); c1 += 2;
518         vst1_f32(c0, vacc0x01); c0 += 2;
519 
520         vacc7x01 = vget_high_f32(vacc7x0123);
521         vacc6x01 = vget_high_f32(vacc6x0123);
522         vacc5x01 = vget_high_f32(vacc5x0123);
523         vacc4x01 = vget_high_f32(vacc4x0123);
524         vacc3x01 = vget_high_f32(vacc3x0123);
525         vacc2x01 = vget_high_f32(vacc2x0123);
526         vacc1x01 = vget_high_f32(vacc1x0123);
527         vacc0x01 = vget_high_f32(vacc0x0123);
528       }
529       if (nc & 1) {
530         vst1_lane_f32(c7, vacc7x01, 0);
531         vst1_lane_f32(c6, vacc6x01, 0);
532         vst1_lane_f32(c5, vacc5x01, 0);
533         vst1_lane_f32(c4, vacc4x01, 0);
534         vst1_lane_f32(c3, vacc3x01, 0);
535         vst1_lane_f32(c2, vacc2x01, 0);
536         vst1_lane_f32(c1, vacc1x01, 0);
537         vst1_lane_f32(c0, vacc0x01, 0);
538       }
539 
540       nc = 0;
541     }
542   } while (nc != 0);
543 }
544