xref: /aosp_15_r20/external/XNNPACK/src/f32-gemm/gen-inc/6x8s4inc-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_6x8s4__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_6x8s4__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 <= 6);
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 
74   do {
75     float32x4_t vacc0x0123 = vld1q_f32(acc); acc += 4;
76     float32x4_t vacc0x4567 = vld1q_f32(acc); acc += 4;
77     float32x4_t vacc1x0123 = vld1q_f32(acc); acc += 4;
78     float32x4_t vacc1x4567 = vld1q_f32(acc); acc += 4;
79     float32x4_t vacc2x0123 = vld1q_f32(acc); acc += 4;
80     float32x4_t vacc2x4567 = vld1q_f32(acc); acc += 4;
81     float32x4_t vacc3x0123 = vld1q_f32(acc); acc += 4;
82     float32x4_t vacc3x4567 = vld1q_f32(acc); acc += 4;
83     float32x4_t vacc4x0123 = vld1q_f32(acc); acc += 4;
84     float32x4_t vacc4x4567 = vld1q_f32(acc); acc += 4;
85     float32x4_t vacc5x0123 = vld1q_f32(acc); acc += 4;
86     float32x4_t vacc5x4567 = vld1q_f32(acc); acc += 4;
87 
88     size_t k = kc;
89     while (k >= 4 * sizeof(float)) {
90       float32x4_t va0 = vld1q_f32(a0); a0 += 4;
91       float32x4_t va1 = vld1q_f32(a1); a1 += 4;
92       float32x4_t va2 = vld1q_f32(a2); a2 += 4;
93       float32x4_t va3 = vld1q_f32(a3); a3 += 4;
94       float32x4_t va4 = vld1q_f32(a4); a4 += 4;
95       float32x4_t va5 = vld1q_f32(a5); a5 += 4;
96 
97 
98       const float32x4_t vb0123c0 = vld1q_f32(w); w += 4;
99       const float32x4_t vb4567c0 = vld1q_f32(w); w += 4;
100 
101       vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123c0);
102       vacc1x0123 = vmlaq_f32(vacc1x0123, va1, vb0123c0);
103       vacc2x0123 = vmlaq_f32(vacc2x0123, va2, vb0123c0);
104       vacc3x0123 = vmlaq_f32(vacc3x0123, va3, vb0123c0);
105       vacc4x0123 = vmlaq_f32(vacc4x0123, va4, vb0123c0);
106       vacc5x0123 = vmlaq_f32(vacc5x0123, va5, vb0123c0);
107       vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567c0);
108       vacc1x4567 = vmlaq_f32(vacc1x4567, va1, vb4567c0);
109       vacc2x4567 = vmlaq_f32(vacc2x4567, va2, vb4567c0);
110       vacc3x4567 = vmlaq_f32(vacc3x4567, va3, vb4567c0);
111       vacc4x4567 = vmlaq_f32(vacc4x4567, va4, vb4567c0);
112       vacc5x4567 = vmlaq_f32(vacc5x4567, va5, vb4567c0);
113 
114       va0 = vextq_f32(va0, va0, 1);
115       va1 = vextq_f32(va1, va1, 1);
116       va2 = vextq_f32(va2, va2, 1);
117       va3 = vextq_f32(va3, va3, 1);
118       va4 = vextq_f32(va4, va4, 1);
119       va5 = vextq_f32(va5, va5, 1);
120 
121       const float32x4_t vb0123c1 = vld1q_f32(w); w += 4;
122       const float32x4_t vb4567c1 = vld1q_f32(w); w += 4;
123 
124       vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123c1);
125       vacc1x0123 = vmlaq_f32(vacc1x0123, va1, vb0123c1);
126       vacc2x0123 = vmlaq_f32(vacc2x0123, va2, vb0123c1);
127       vacc3x0123 = vmlaq_f32(vacc3x0123, va3, vb0123c1);
128       vacc4x0123 = vmlaq_f32(vacc4x0123, va4, vb0123c1);
129       vacc5x0123 = vmlaq_f32(vacc5x0123, va5, vb0123c1);
130       vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567c1);
131       vacc1x4567 = vmlaq_f32(vacc1x4567, va1, vb4567c1);
132       vacc2x4567 = vmlaq_f32(vacc2x4567, va2, vb4567c1);
133       vacc3x4567 = vmlaq_f32(vacc3x4567, va3, vb4567c1);
134       vacc4x4567 = vmlaq_f32(vacc4x4567, va4, vb4567c1);
135       vacc5x4567 = vmlaq_f32(vacc5x4567, va5, vb4567c1);
136 
137       va0 = vextq_f32(va0, va0, 1);
138       va1 = vextq_f32(va1, va1, 1);
139       va2 = vextq_f32(va2, va2, 1);
140       va3 = vextq_f32(va3, va3, 1);
141       va4 = vextq_f32(va4, va4, 1);
142       va5 = vextq_f32(va5, va5, 1);
143 
144       const float32x4_t vb0123c2 = vld1q_f32(w); w += 4;
145       const float32x4_t vb4567c2 = vld1q_f32(w); w += 4;
146 
147       vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123c2);
148       vacc1x0123 = vmlaq_f32(vacc1x0123, va1, vb0123c2);
149       vacc2x0123 = vmlaq_f32(vacc2x0123, va2, vb0123c2);
150       vacc3x0123 = vmlaq_f32(vacc3x0123, va3, vb0123c2);
151       vacc4x0123 = vmlaq_f32(vacc4x0123, va4, vb0123c2);
152       vacc5x0123 = vmlaq_f32(vacc5x0123, va5, vb0123c2);
153       vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567c2);
154       vacc1x4567 = vmlaq_f32(vacc1x4567, va1, vb4567c2);
155       vacc2x4567 = vmlaq_f32(vacc2x4567, va2, vb4567c2);
156       vacc3x4567 = vmlaq_f32(vacc3x4567, va3, vb4567c2);
157       vacc4x4567 = vmlaq_f32(vacc4x4567, va4, vb4567c2);
158       vacc5x4567 = vmlaq_f32(vacc5x4567, va5, vb4567c2);
159 
160       va0 = vextq_f32(va0, va0, 1);
161       va1 = vextq_f32(va1, va1, 1);
162       va2 = vextq_f32(va2, va2, 1);
163       va3 = vextq_f32(va3, va3, 1);
164       va4 = vextq_f32(va4, va4, 1);
165       va5 = vextq_f32(va5, va5, 1);
166 
167       const float32x4_t vb0123c3 = vld1q_f32(w); w += 4;
168       const float32x4_t vb4567c3 = vld1q_f32(w); w += 4;
169 
170       vacc0x0123 = vmlaq_f32(vacc0x0123, va0, vb0123c3);
171       vacc1x0123 = vmlaq_f32(vacc1x0123, va1, vb0123c3);
172       vacc2x0123 = vmlaq_f32(vacc2x0123, va2, vb0123c3);
173       vacc3x0123 = vmlaq_f32(vacc3x0123, va3, vb0123c3);
174       vacc4x0123 = vmlaq_f32(vacc4x0123, va4, vb0123c3);
175       vacc5x0123 = vmlaq_f32(vacc5x0123, va5, vb0123c3);
176       vacc0x4567 = vmlaq_f32(vacc0x4567, va0, vb4567c3);
177       vacc1x4567 = vmlaq_f32(vacc1x4567, va1, vb4567c3);
178       vacc2x4567 = vmlaq_f32(vacc2x4567, va2, vb4567c3);
179       vacc3x4567 = vmlaq_f32(vacc3x4567, va3, vb4567c3);
180       vacc4x4567 = vmlaq_f32(vacc4x4567, va4, vb4567c3);
181       vacc5x4567 = vmlaq_f32(vacc5x4567, va5, vb4567c3);
182 
183 
184       k -= 4 * sizeof(float);
185     }
186     if XNN_UNLIKELY(k != 0) {
187       float32x4_t va0 = vld1q_f32(a0); a0 = (const float*) ((uintptr_t) a0 + k);
188       float32x4_t va1 = vld1q_f32(a1); a1 = (const float*) ((uintptr_t) a1 + k);
189       float32x4_t va2 = vld1q_f32(a2); a2 = (const float*) ((uintptr_t) a2 + k);
190       float32x4_t va3 = vld1q_f32(a3); a3 = (const float*) ((uintptr_t) a3 + k);
191       float32x4_t va4 = vld1q_f32(a4); a4 = (const float*) ((uintptr_t) a4 + k);
192       float32x4_t va5 = vld1q_f32(a5); a5 = (const float*) ((uintptr_t) a5 + k);
193 
194 
195       const float32x4_t vb0123c0 = vld1q_f32(w); w += 4;
196       const float32x4_t vb4567c0 = vld1q_f32(w); w += 4;
197 
198       const float32x4_t vmska0x0123c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb0123c0, vmovq_n_f32(0.0f))));
199       vacc0x0123 = vmlaq_f32(vacc0x0123, vmska0x0123c0, vb0123c0);
200       const float32x4_t vmska1x0123c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb0123c0, vmovq_n_f32(0.0f))));
201       vacc1x0123 = vmlaq_f32(vacc1x0123, vmska1x0123c0, vb0123c0);
202       const float32x4_t vmska2x0123c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb0123c0, vmovq_n_f32(0.0f))));
203       vacc2x0123 = vmlaq_f32(vacc2x0123, vmska2x0123c0, vb0123c0);
204       const float32x4_t vmska3x0123c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb0123c0, vmovq_n_f32(0.0f))));
205       vacc3x0123 = vmlaq_f32(vacc3x0123, vmska3x0123c0, vb0123c0);
206       const float32x4_t vmska4x0123c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va4), vceqq_f32(vb0123c0, vmovq_n_f32(0.0f))));
207       vacc4x0123 = vmlaq_f32(vacc4x0123, vmska4x0123c0, vb0123c0);
208       const float32x4_t vmska5x0123c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va5), vceqq_f32(vb0123c0, vmovq_n_f32(0.0f))));
209       vacc5x0123 = vmlaq_f32(vacc5x0123, vmska5x0123c0, vb0123c0);
210       const float32x4_t vmska0x4567c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb4567c0, vmovq_n_f32(0.0f))));
211       vacc0x4567 = vmlaq_f32(vacc0x4567, vmska0x4567c0, vb4567c0);
212       const float32x4_t vmska1x4567c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb4567c0, vmovq_n_f32(0.0f))));
213       vacc1x4567 = vmlaq_f32(vacc1x4567, vmska1x4567c0, vb4567c0);
214       const float32x4_t vmska2x4567c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb4567c0, vmovq_n_f32(0.0f))));
215       vacc2x4567 = vmlaq_f32(vacc2x4567, vmska2x4567c0, vb4567c0);
216       const float32x4_t vmska3x4567c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb4567c0, vmovq_n_f32(0.0f))));
217       vacc3x4567 = vmlaq_f32(vacc3x4567, vmska3x4567c0, vb4567c0);
218       const float32x4_t vmska4x4567c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va4), vceqq_f32(vb4567c0, vmovq_n_f32(0.0f))));
219       vacc4x4567 = vmlaq_f32(vacc4x4567, vmska4x4567c0, vb4567c0);
220       const float32x4_t vmska5x4567c0 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va5), vceqq_f32(vb4567c0, vmovq_n_f32(0.0f))));
221       vacc5x4567 = vmlaq_f32(vacc5x4567, vmska5x4567c0, vb4567c0);
222 
223       va0 = vextq_f32(va0, va0, 1);
224       va1 = vextq_f32(va1, va1, 1);
225       va2 = vextq_f32(va2, va2, 1);
226       va3 = vextq_f32(va3, va3, 1);
227       va4 = vextq_f32(va4, va4, 1);
228       va5 = vextq_f32(va5, va5, 1);
229 
230       const float32x4_t vb0123c1 = vld1q_f32(w); w += 4;
231       const float32x4_t vb4567c1 = vld1q_f32(w); w += 4;
232 
233       const float32x4_t vmska0x0123c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb0123c1, vmovq_n_f32(0.0f))));
234       vacc0x0123 = vmlaq_f32(vacc0x0123, vmska0x0123c1, vb0123c1);
235       const float32x4_t vmska1x0123c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb0123c1, vmovq_n_f32(0.0f))));
236       vacc1x0123 = vmlaq_f32(vacc1x0123, vmska1x0123c1, vb0123c1);
237       const float32x4_t vmska2x0123c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb0123c1, vmovq_n_f32(0.0f))));
238       vacc2x0123 = vmlaq_f32(vacc2x0123, vmska2x0123c1, vb0123c1);
239       const float32x4_t vmska3x0123c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb0123c1, vmovq_n_f32(0.0f))));
240       vacc3x0123 = vmlaq_f32(vacc3x0123, vmska3x0123c1, vb0123c1);
241       const float32x4_t vmska4x0123c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va4), vceqq_f32(vb0123c1, vmovq_n_f32(0.0f))));
242       vacc4x0123 = vmlaq_f32(vacc4x0123, vmska4x0123c1, vb0123c1);
243       const float32x4_t vmska5x0123c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va5), vceqq_f32(vb0123c1, vmovq_n_f32(0.0f))));
244       vacc5x0123 = vmlaq_f32(vacc5x0123, vmska5x0123c1, vb0123c1);
245       const float32x4_t vmska0x4567c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb4567c1, vmovq_n_f32(0.0f))));
246       vacc0x4567 = vmlaq_f32(vacc0x4567, vmska0x4567c1, vb4567c1);
247       const float32x4_t vmska1x4567c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb4567c1, vmovq_n_f32(0.0f))));
248       vacc1x4567 = vmlaq_f32(vacc1x4567, vmska1x4567c1, vb4567c1);
249       const float32x4_t vmska2x4567c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb4567c1, vmovq_n_f32(0.0f))));
250       vacc2x4567 = vmlaq_f32(vacc2x4567, vmska2x4567c1, vb4567c1);
251       const float32x4_t vmska3x4567c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb4567c1, vmovq_n_f32(0.0f))));
252       vacc3x4567 = vmlaq_f32(vacc3x4567, vmska3x4567c1, vb4567c1);
253       const float32x4_t vmska4x4567c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va4), vceqq_f32(vb4567c1, vmovq_n_f32(0.0f))));
254       vacc4x4567 = vmlaq_f32(vacc4x4567, vmska4x4567c1, vb4567c1);
255       const float32x4_t vmska5x4567c1 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va5), vceqq_f32(vb4567c1, vmovq_n_f32(0.0f))));
256       vacc5x4567 = vmlaq_f32(vacc5x4567, vmska5x4567c1, vb4567c1);
257 
258       va0 = vextq_f32(va0, va0, 1);
259       va1 = vextq_f32(va1, va1, 1);
260       va2 = vextq_f32(va2, va2, 1);
261       va3 = vextq_f32(va3, va3, 1);
262       va4 = vextq_f32(va4, va4, 1);
263       va5 = vextq_f32(va5, va5, 1);
264 
265       const float32x4_t vb0123c2 = vld1q_f32(w); w += 4;
266       const float32x4_t vb4567c2 = vld1q_f32(w); w += 4;
267 
268       const float32x4_t vmska0x0123c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb0123c2, vmovq_n_f32(0.0f))));
269       vacc0x0123 = vmlaq_f32(vacc0x0123, vmska0x0123c2, vb0123c2);
270       const float32x4_t vmska1x0123c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb0123c2, vmovq_n_f32(0.0f))));
271       vacc1x0123 = vmlaq_f32(vacc1x0123, vmska1x0123c2, vb0123c2);
272       const float32x4_t vmska2x0123c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb0123c2, vmovq_n_f32(0.0f))));
273       vacc2x0123 = vmlaq_f32(vacc2x0123, vmska2x0123c2, vb0123c2);
274       const float32x4_t vmska3x0123c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb0123c2, vmovq_n_f32(0.0f))));
275       vacc3x0123 = vmlaq_f32(vacc3x0123, vmska3x0123c2, vb0123c2);
276       const float32x4_t vmska4x0123c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va4), vceqq_f32(vb0123c2, vmovq_n_f32(0.0f))));
277       vacc4x0123 = vmlaq_f32(vacc4x0123, vmska4x0123c2, vb0123c2);
278       const float32x4_t vmska5x0123c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va5), vceqq_f32(vb0123c2, vmovq_n_f32(0.0f))));
279       vacc5x0123 = vmlaq_f32(vacc5x0123, vmska5x0123c2, vb0123c2);
280       const float32x4_t vmska0x4567c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb4567c2, vmovq_n_f32(0.0f))));
281       vacc0x4567 = vmlaq_f32(vacc0x4567, vmska0x4567c2, vb4567c2);
282       const float32x4_t vmska1x4567c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb4567c2, vmovq_n_f32(0.0f))));
283       vacc1x4567 = vmlaq_f32(vacc1x4567, vmska1x4567c2, vb4567c2);
284       const float32x4_t vmska2x4567c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb4567c2, vmovq_n_f32(0.0f))));
285       vacc2x4567 = vmlaq_f32(vacc2x4567, vmska2x4567c2, vb4567c2);
286       const float32x4_t vmska3x4567c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb4567c2, vmovq_n_f32(0.0f))));
287       vacc3x4567 = vmlaq_f32(vacc3x4567, vmska3x4567c2, vb4567c2);
288       const float32x4_t vmska4x4567c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va4), vceqq_f32(vb4567c2, vmovq_n_f32(0.0f))));
289       vacc4x4567 = vmlaq_f32(vacc4x4567, vmska4x4567c2, vb4567c2);
290       const float32x4_t vmska5x4567c2 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va5), vceqq_f32(vb4567c2, vmovq_n_f32(0.0f))));
291       vacc5x4567 = vmlaq_f32(vacc5x4567, vmska5x4567c2, vb4567c2);
292 
293       va0 = vextq_f32(va0, va0, 1);
294       va1 = vextq_f32(va1, va1, 1);
295       va2 = vextq_f32(va2, va2, 1);
296       va3 = vextq_f32(va3, va3, 1);
297       va4 = vextq_f32(va4, va4, 1);
298       va5 = vextq_f32(va5, va5, 1);
299 
300       const float32x4_t vb0123c3 = vld1q_f32(w); w += 4;
301       const float32x4_t vb4567c3 = vld1q_f32(w); w += 4;
302 
303       const float32x4_t vmska0x0123c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb0123c3, vmovq_n_f32(0.0f))));
304       vacc0x0123 = vmlaq_f32(vacc0x0123, vmska0x0123c3, vb0123c3);
305       const float32x4_t vmska1x0123c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb0123c3, vmovq_n_f32(0.0f))));
306       vacc1x0123 = vmlaq_f32(vacc1x0123, vmska1x0123c3, vb0123c3);
307       const float32x4_t vmska2x0123c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb0123c3, vmovq_n_f32(0.0f))));
308       vacc2x0123 = vmlaq_f32(vacc2x0123, vmska2x0123c3, vb0123c3);
309       const float32x4_t vmska3x0123c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb0123c3, vmovq_n_f32(0.0f))));
310       vacc3x0123 = vmlaq_f32(vacc3x0123, vmska3x0123c3, vb0123c3);
311       const float32x4_t vmska4x0123c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va4), vceqq_f32(vb0123c3, vmovq_n_f32(0.0f))));
312       vacc4x0123 = vmlaq_f32(vacc4x0123, vmska4x0123c3, vb0123c3);
313       const float32x4_t vmska5x0123c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va5), vceqq_f32(vb0123c3, vmovq_n_f32(0.0f))));
314       vacc5x0123 = vmlaq_f32(vacc5x0123, vmska5x0123c3, vb0123c3);
315       const float32x4_t vmska0x4567c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va0), vceqq_f32(vb4567c3, vmovq_n_f32(0.0f))));
316       vacc0x4567 = vmlaq_f32(vacc0x4567, vmska0x4567c3, vb4567c3);
317       const float32x4_t vmska1x4567c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va1), vceqq_f32(vb4567c3, vmovq_n_f32(0.0f))));
318       vacc1x4567 = vmlaq_f32(vacc1x4567, vmska1x4567c3, vb4567c3);
319       const float32x4_t vmska2x4567c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va2), vceqq_f32(vb4567c3, vmovq_n_f32(0.0f))));
320       vacc2x4567 = vmlaq_f32(vacc2x4567, vmska2x4567c3, vb4567c3);
321       const float32x4_t vmska3x4567c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va3), vceqq_f32(vb4567c3, vmovq_n_f32(0.0f))));
322       vacc3x4567 = vmlaq_f32(vacc3x4567, vmska3x4567c3, vb4567c3);
323       const float32x4_t vmska4x4567c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va4), vceqq_f32(vb4567c3, vmovq_n_f32(0.0f))));
324       vacc4x4567 = vmlaq_f32(vacc4x4567, vmska4x4567c3, vb4567c3);
325       const float32x4_t vmska5x4567c3 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(va5), vceqq_f32(vb4567c3, vmovq_n_f32(0.0f))));
326       vacc5x4567 = vmlaq_f32(vacc5x4567, vmska5x4567c3, vb4567c3);
327 
328     }
329     const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max);
330     vacc0x0123 = vminq_f32(vacc0x0123, vmax);
331     vacc1x0123 = vminq_f32(vacc1x0123, vmax);
332     vacc2x0123 = vminq_f32(vacc2x0123, vmax);
333     vacc3x0123 = vminq_f32(vacc3x0123, vmax);
334     vacc4x0123 = vminq_f32(vacc4x0123, vmax);
335     vacc5x0123 = vminq_f32(vacc5x0123, vmax);
336     vacc0x4567 = vminq_f32(vacc0x4567, vmax);
337     vacc1x4567 = vminq_f32(vacc1x4567, vmax);
338     vacc2x4567 = vminq_f32(vacc2x4567, vmax);
339     vacc3x4567 = vminq_f32(vacc3x4567, vmax);
340     vacc4x4567 = vminq_f32(vacc4x4567, vmax);
341     vacc5x4567 = vminq_f32(vacc5x4567, vmax);
342 
343     const float32x4_t vmin = vld1q_dup_f32(&params->scalar.min);
344     vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
345     vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
346     vacc2x0123 = vmaxq_f32(vacc2x0123, vmin);
347     vacc3x0123 = vmaxq_f32(vacc3x0123, vmin);
348     vacc4x0123 = vmaxq_f32(vacc4x0123, vmin);
349     vacc5x0123 = vmaxq_f32(vacc5x0123, vmin);
350     vacc0x4567 = vmaxq_f32(vacc0x4567, vmin);
351     vacc1x4567 = vmaxq_f32(vacc1x4567, vmin);
352     vacc2x4567 = vmaxq_f32(vacc2x4567, vmin);
353     vacc3x4567 = vmaxq_f32(vacc3x4567, vmin);
354     vacc4x4567 = vmaxq_f32(vacc4x4567, vmin);
355     vacc5x4567 = vmaxq_f32(vacc5x4567, vmin);
356 
357     if XNN_LIKELY(nc >= 8) {
358       vst1q_f32(c5, vacc5x0123);
359       vst1q_f32(c5 + 4, vacc5x4567);
360       c5 = (float*) ((uintptr_t) c5 + cn_stride);
361       vst1q_f32(c4, vacc4x0123);
362       vst1q_f32(c4 + 4, vacc4x4567);
363       c4 = (float*) ((uintptr_t) c4 + cn_stride);
364       vst1q_f32(c3, vacc3x0123);
365       vst1q_f32(c3 + 4, vacc3x4567);
366       c3 = (float*) ((uintptr_t) c3 + cn_stride);
367       vst1q_f32(c2, vacc2x0123);
368       vst1q_f32(c2 + 4, vacc2x4567);
369       c2 = (float*) ((uintptr_t) c2 + cn_stride);
370       vst1q_f32(c1, vacc1x0123);
371       vst1q_f32(c1 + 4, vacc1x4567);
372       c1 = (float*) ((uintptr_t) c1 + cn_stride);
373       vst1q_f32(c0, vacc0x0123);
374       vst1q_f32(c0 + 4, vacc0x4567);
375       c0 = (float*) ((uintptr_t) c0 + cn_stride);
376 
377       a5 = (const float*) ((uintptr_t) a5 - kc);
378       a4 = (const float*) ((uintptr_t) a4 - kc);
379       a3 = (const float*) ((uintptr_t) a3 - kc);
380       a2 = (const float*) ((uintptr_t) a2 - kc);
381       a1 = (const float*) ((uintptr_t) a1 - kc);
382       a0 = (const float*) ((uintptr_t) a0 - kc);
383 
384       nc -= 8;
385 
386     } else {
387       if (nc & 4) {
388         vst1q_f32(c5, vacc5x0123); c5 += 4;
389         vst1q_f32(c4, vacc4x0123); c4 += 4;
390         vst1q_f32(c3, vacc3x0123); c3 += 4;
391         vst1q_f32(c2, vacc2x0123); c2 += 4;
392         vst1q_f32(c1, vacc1x0123); c1 += 4;
393         vst1q_f32(c0, vacc0x0123); c0 += 4;
394 
395         vacc5x0123 = vacc5x4567;
396         vacc4x0123 = vacc4x4567;
397         vacc3x0123 = vacc3x4567;
398         vacc2x0123 = vacc2x4567;
399         vacc1x0123 = vacc1x4567;
400         vacc0x0123 = vacc0x4567;
401       }
402       float32x2_t vacc5x01 = vget_low_f32(vacc5x0123);
403       float32x2_t vacc4x01 = vget_low_f32(vacc4x0123);
404       float32x2_t vacc3x01 = vget_low_f32(vacc3x0123);
405       float32x2_t vacc2x01 = vget_low_f32(vacc2x0123);
406       float32x2_t vacc1x01 = vget_low_f32(vacc1x0123);
407       float32x2_t vacc0x01 = vget_low_f32(vacc0x0123);
408       if (nc & 2) {
409         vst1_f32(c5, vacc5x01); c5 += 2;
410         vst1_f32(c4, vacc4x01); c4 += 2;
411         vst1_f32(c3, vacc3x01); c3 += 2;
412         vst1_f32(c2, vacc2x01); c2 += 2;
413         vst1_f32(c1, vacc1x01); c1 += 2;
414         vst1_f32(c0, vacc0x01); c0 += 2;
415 
416         vacc5x01 = vget_high_f32(vacc5x0123);
417         vacc4x01 = vget_high_f32(vacc4x0123);
418         vacc3x01 = vget_high_f32(vacc3x0123);
419         vacc2x01 = vget_high_f32(vacc2x0123);
420         vacc1x01 = vget_high_f32(vacc1x0123);
421         vacc0x01 = vget_high_f32(vacc0x0123);
422       }
423       if (nc & 1) {
424         vst1_lane_f32(c5, vacc5x01, 0);
425         vst1_lane_f32(c4, vacc4x01, 0);
426         vst1_lane_f32(c3, vacc3x01, 0);
427         vst1_lane_f32(c2, vacc2x01, 0);
428         vst1_lane_f32(c1, vacc1x01, 0);
429         vst1_lane_f32(c0, vacc0x01, 0);
430       }
431 
432       nc = 0;
433     }
434   } while (nc != 0);
435 }
436