• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f16-gemm/neonfp16arith-ld64.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2020 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/common.h>
16 
17 #include <xnnpack/gemm.h>
18 
19 
xnn_f16_gemm_minmax_ukernel_8x8__neonfp16arith_ld64(size_t mr,size_t nc,size_t kc,const void * restrict a,size_t a_stride,const void * restrict w,void * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_f16_gemm_minmax_ukernel_8x8__neonfp16arith_ld64(
21     size_t mr,
22     size_t nc,
23     size_t kc,
24     const void* restrict a,
25     size_t a_stride,
26     const void* restrict w,
27     void* restrict c,
28     size_t cm_stride,
29     size_t cn_stride,
30     const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
31 {
32   assert(mr != 0);
33   assert(mr <= 8);
34   assert(nc != 0);
35   assert(kc != 0);
36   assert(kc % sizeof(__fp16) == 0);
37   assert(a != NULL);
38   assert(w != NULL);
39   assert(c != NULL);
40 
41   const __fp16* a0 = (const __fp16*) a;
42   __fp16* c0 = (__fp16*) c;
43   const __fp16* a1 = (const __fp16*) ((uintptr_t) a0 + a_stride);
44   __fp16* c1 = (__fp16*) ((uintptr_t) c0 + cm_stride);
45   if XNN_UNPREDICTABLE(mr < 2) {
46     a1 = a0;
47     c1 = c0;
48   }
49   const __fp16* a2 = (const __fp16*) ((uintptr_t) a1 + a_stride);
50   __fp16* c2 = (__fp16*) ((uintptr_t) c1 + cm_stride);
51   if XNN_UNPREDICTABLE(mr <= 2) {
52     a2 = a1;
53     c2 = c1;
54   }
55   const __fp16* a3 = (const __fp16*) ((uintptr_t) a2 + a_stride);
56   __fp16* c3 = (__fp16*) ((uintptr_t) c2 + cm_stride);
57   if XNN_UNPREDICTABLE(mr < 4) {
58     a3 = a2;
59     c3 = c2;
60   }
61   const __fp16* a4 = (const __fp16*) ((uintptr_t) a3 + a_stride);
62   __fp16* c4 = (__fp16*) ((uintptr_t) c3 + cm_stride);
63   if XNN_UNPREDICTABLE(mr <= 4) {
64     a4 = a3;
65     c4 = c3;
66   }
67   const __fp16* a5 = (const __fp16*) ((uintptr_t) a4 + a_stride);
68   __fp16* c5 = (__fp16*) ((uintptr_t) c4 + cm_stride);
69   if XNN_UNPREDICTABLE(mr < 6) {
70     a5 = a4;
71     c5 = c4;
72   }
73   const __fp16* a6 = (const __fp16*) ((uintptr_t) a5 + a_stride);
74   __fp16* c6 = (__fp16*) ((uintptr_t) c5 + cm_stride);
75   if XNN_UNPREDICTABLE(mr <= 6) {
76     a6 = a5;
77     c6 = c5;
78   }
79   const __fp16* a7 = (const __fp16*) ((uintptr_t) a6 + a_stride);
80   __fp16* c7 = (__fp16*) ((uintptr_t) c6 + cm_stride);
81   if XNN_UNPREDICTABLE(mr != 8) {
82     a7 = a6;
83     c7 = c6;
84   }
85 
86   do {
87     float16x8_t vacc0x01234567 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
88     float16x8_t vacc1x01234567 = vacc0x01234567;
89     float16x8_t vacc2x01234567 = vacc0x01234567;
90     float16x8_t vacc3x01234567 = vacc0x01234567;
91     float16x8_t vacc4x01234567 = vacc0x01234567;
92     float16x8_t vacc5x01234567 = vacc0x01234567;
93     float16x8_t vacc6x01234567 = vacc0x01234567;
94     float16x8_t vacc7x01234567 = vacc0x01234567;
95 
96     size_t k = kc;
97     while (k >= 4 * sizeof(__fp16)) {
98       const float16x4_t va0 = vld1_f16(a0); a0 += 4;
99       const float16x4_t va1 = vld1_f16(a1); a1 += 4;
100       const float16x4_t va2 = vld1_f16(a2); a2 += 4;
101       const float16x4_t va3 = vld1_f16(a3); a3 += 4;
102       const float16x4_t va4 = vld1_f16(a4); a4 += 4;
103       const float16x4_t va5 = vld1_f16(a5); a5 += 4;
104       const float16x4_t va6 = vld1_f16(a6); a6 += 4;
105       const float16x4_t va7 = vld1_f16(a7); a7 += 4;
106 
107       const float16x8_t vb01234567c0 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
108 
109       #if XNN_ARCH_ARM64
110         vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c0, va0, 0);
111         vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c0, va1, 0);
112         vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c0, va2, 0);
113         vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c0, va3, 0);
114         vacc4x01234567 = vfmaq_lane_f16(vacc4x01234567, vb01234567c0, va4, 0);
115         vacc5x01234567 = vfmaq_lane_f16(vacc5x01234567, vb01234567c0, va5, 0);
116         vacc6x01234567 = vfmaq_lane_f16(vacc6x01234567, vb01234567c0, va6, 0);
117         vacc7x01234567 = vfmaq_lane_f16(vacc7x01234567, vb01234567c0, va7, 0);
118       #else
119         const float16x8_t va0c0 = vdupq_lane_f16(va0, 0);
120         const float16x8_t va1c0 = vdupq_lane_f16(va1, 0);
121         const float16x8_t va2c0 = vdupq_lane_f16(va2, 0);
122         const float16x8_t va3c0 = vdupq_lane_f16(va3, 0);
123         const float16x8_t va4c0 = vdupq_lane_f16(va4, 0);
124         const float16x8_t va5c0 = vdupq_lane_f16(va5, 0);
125         const float16x8_t va6c0 = vdupq_lane_f16(va6, 0);
126         const float16x8_t va7c0 = vdupq_lane_f16(va7, 0);
127 
128         vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c0, vb01234567c0);
129         vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1c0, vb01234567c0);
130         vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c0, vb01234567c0);
131         vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3c0, vb01234567c0);
132         vacc4x01234567 = vfmaq_f16(vacc4x01234567, va4c0, vb01234567c0);
133         vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5c0, vb01234567c0);
134         vacc6x01234567 = vfmaq_f16(vacc6x01234567, va6c0, vb01234567c0);
135         vacc7x01234567 = vfmaq_f16(vacc7x01234567, va7c0, vb01234567c0);
136       #endif
137       const float16x8_t vb01234567c1 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
138 
139       #if XNN_ARCH_ARM64
140         vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c1, va0, 1);
141         vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c1, va1, 1);
142         vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c1, va2, 1);
143         vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c1, va3, 1);
144         vacc4x01234567 = vfmaq_lane_f16(vacc4x01234567, vb01234567c1, va4, 1);
145         vacc5x01234567 = vfmaq_lane_f16(vacc5x01234567, vb01234567c1, va5, 1);
146         vacc6x01234567 = vfmaq_lane_f16(vacc6x01234567, vb01234567c1, va6, 1);
147         vacc7x01234567 = vfmaq_lane_f16(vacc7x01234567, vb01234567c1, va7, 1);
148       #else
149         const float16x8_t va0c1 = vdupq_lane_f16(va0, 1);
150         const float16x8_t va1c1 = vdupq_lane_f16(va1, 1);
151         const float16x8_t va2c1 = vdupq_lane_f16(va2, 1);
152         const float16x8_t va3c1 = vdupq_lane_f16(va3, 1);
153         const float16x8_t va4c1 = vdupq_lane_f16(va4, 1);
154         const float16x8_t va5c1 = vdupq_lane_f16(va5, 1);
155         const float16x8_t va6c1 = vdupq_lane_f16(va6, 1);
156         const float16x8_t va7c1 = vdupq_lane_f16(va7, 1);
157 
158         vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c1, vb01234567c1);
159         vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1c1, vb01234567c1);
160         vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c1, vb01234567c1);
161         vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3c1, vb01234567c1);
162         vacc4x01234567 = vfmaq_f16(vacc4x01234567, va4c1, vb01234567c1);
163         vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5c1, vb01234567c1);
164         vacc6x01234567 = vfmaq_f16(vacc6x01234567, va6c1, vb01234567c1);
165         vacc7x01234567 = vfmaq_f16(vacc7x01234567, va7c1, vb01234567c1);
166       #endif
167       const float16x8_t vb01234567c2 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
168 
169       #if XNN_ARCH_ARM64
170         vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c2, va0, 2);
171         vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c2, va1, 2);
172         vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c2, va2, 2);
173         vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c2, va3, 2);
174         vacc4x01234567 = vfmaq_lane_f16(vacc4x01234567, vb01234567c2, va4, 2);
175         vacc5x01234567 = vfmaq_lane_f16(vacc5x01234567, vb01234567c2, va5, 2);
176         vacc6x01234567 = vfmaq_lane_f16(vacc6x01234567, vb01234567c2, va6, 2);
177         vacc7x01234567 = vfmaq_lane_f16(vacc7x01234567, vb01234567c2, va7, 2);
178       #else
179         const float16x8_t va0c2 = vdupq_lane_f16(va0, 2);
180         const float16x8_t va1c2 = vdupq_lane_f16(va1, 2);
181         const float16x8_t va2c2 = vdupq_lane_f16(va2, 2);
182         const float16x8_t va3c2 = vdupq_lane_f16(va3, 2);
183         const float16x8_t va4c2 = vdupq_lane_f16(va4, 2);
184         const float16x8_t va5c2 = vdupq_lane_f16(va5, 2);
185         const float16x8_t va6c2 = vdupq_lane_f16(va6, 2);
186         const float16x8_t va7c2 = vdupq_lane_f16(va7, 2);
187 
188         vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c2, vb01234567c2);
189         vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1c2, vb01234567c2);
190         vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c2, vb01234567c2);
191         vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3c2, vb01234567c2);
192         vacc4x01234567 = vfmaq_f16(vacc4x01234567, va4c2, vb01234567c2);
193         vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5c2, vb01234567c2);
194         vacc6x01234567 = vfmaq_f16(vacc6x01234567, va6c2, vb01234567c2);
195         vacc7x01234567 = vfmaq_f16(vacc7x01234567, va7c2, vb01234567c2);
196       #endif
197       const float16x8_t vb01234567c3 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
198 
199       #if XNN_ARCH_ARM64
200         vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c3, va0, 3);
201         vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c3, va1, 3);
202         vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c3, va2, 3);
203         vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c3, va3, 3);
204         vacc4x01234567 = vfmaq_lane_f16(vacc4x01234567, vb01234567c3, va4, 3);
205         vacc5x01234567 = vfmaq_lane_f16(vacc5x01234567, vb01234567c3, va5, 3);
206         vacc6x01234567 = vfmaq_lane_f16(vacc6x01234567, vb01234567c3, va6, 3);
207         vacc7x01234567 = vfmaq_lane_f16(vacc7x01234567, vb01234567c3, va7, 3);
208       #else
209         const float16x8_t va0c3 = vdupq_lane_f16(va0, 3);
210         const float16x8_t va1c3 = vdupq_lane_f16(va1, 3);
211         const float16x8_t va2c3 = vdupq_lane_f16(va2, 3);
212         const float16x8_t va3c3 = vdupq_lane_f16(va3, 3);
213         const float16x8_t va4c3 = vdupq_lane_f16(va4, 3);
214         const float16x8_t va5c3 = vdupq_lane_f16(va5, 3);
215         const float16x8_t va6c3 = vdupq_lane_f16(va6, 3);
216         const float16x8_t va7c3 = vdupq_lane_f16(va7, 3);
217 
218         vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c3, vb01234567c3);
219         vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1c3, vb01234567c3);
220         vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c3, vb01234567c3);
221         vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3c3, vb01234567c3);
222         vacc4x01234567 = vfmaq_f16(vacc4x01234567, va4c3, vb01234567c3);
223         vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5c3, vb01234567c3);
224         vacc6x01234567 = vfmaq_f16(vacc6x01234567, va6c3, vb01234567c3);
225         vacc7x01234567 = vfmaq_f16(vacc7x01234567, va7c3, vb01234567c3);
226       #endif
227 
228       k -= 4 * sizeof(__fp16);
229     }
230     if XNN_UNLIKELY(k != 0) {
231       do {
232         const float16x8_t va0 = vld1q_dup_f16(a0); a0 += 1;
233         const float16x8_t va1 = vld1q_dup_f16(a1); a1 += 1;
234         const float16x8_t va2 = vld1q_dup_f16(a2); a2 += 1;
235         const float16x8_t va3 = vld1q_dup_f16(a3); a3 += 1;
236         const float16x8_t va4 = vld1q_dup_f16(a4); a4 += 1;
237         const float16x8_t va5 = vld1q_dup_f16(a5); a5 += 1;
238         const float16x8_t va6 = vld1q_dup_f16(a6); a6 += 1;
239         const float16x8_t va7 = vld1q_dup_f16(a7); a7 += 1;
240 
241         const float16x8_t vb01234567 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
242 
243         vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0, vb01234567);
244         vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1, vb01234567);
245         vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2, vb01234567);
246         vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3, vb01234567);
247         vacc4x01234567 = vfmaq_f16(vacc4x01234567, va4, vb01234567);
248         vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5, vb01234567);
249         vacc6x01234567 = vfmaq_f16(vacc6x01234567, va6, vb01234567);
250         vacc7x01234567 = vfmaq_f16(vacc7x01234567, va7, vb01234567);
251 
252         k -= sizeof(__fp16);
253       } while (k != 0);
254     }
255 
256 
257     const float16x8_t vmax = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neon.max));
258     vacc0x01234567 = vminq_f16(vacc0x01234567, vmax);
259     vacc1x01234567 = vminq_f16(vacc1x01234567, vmax);
260     vacc2x01234567 = vminq_f16(vacc2x01234567, vmax);
261     vacc3x01234567 = vminq_f16(vacc3x01234567, vmax);
262     vacc4x01234567 = vminq_f16(vacc4x01234567, vmax);
263     vacc5x01234567 = vminq_f16(vacc5x01234567, vmax);
264     vacc6x01234567 = vminq_f16(vacc6x01234567, vmax);
265     vacc7x01234567 = vminq_f16(vacc7x01234567, vmax);
266 
267     const float16x8_t vmin = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neon.min));
268     vacc0x01234567 = vmaxq_f16(vacc0x01234567, vmin);
269     vacc1x01234567 = vmaxq_f16(vacc1x01234567, vmin);
270     vacc2x01234567 = vmaxq_f16(vacc2x01234567, vmin);
271     vacc3x01234567 = vmaxq_f16(vacc3x01234567, vmin);
272     vacc4x01234567 = vmaxq_f16(vacc4x01234567, vmin);
273     vacc5x01234567 = vmaxq_f16(vacc5x01234567, vmin);
274     vacc6x01234567 = vmaxq_f16(vacc6x01234567, vmin);
275     vacc7x01234567 = vmaxq_f16(vacc7x01234567, vmin);
276 
277     if XNN_LIKELY(nc >= 8) {
278       vst1q_f16(c0, vacc0x01234567);
279       c0 = (__fp16*) ((uintptr_t) c0 + cn_stride);
280       vst1q_f16(c1, vacc1x01234567);
281       c1 = (__fp16*) ((uintptr_t) c1 + cn_stride);
282       vst1q_f16(c2, vacc2x01234567);
283       c2 = (__fp16*) ((uintptr_t) c2 + cn_stride);
284       vst1q_f16(c3, vacc3x01234567);
285       c3 = (__fp16*) ((uintptr_t) c3 + cn_stride);
286       vst1q_f16(c4, vacc4x01234567);
287       c4 = (__fp16*) ((uintptr_t) c4 + cn_stride);
288       vst1q_f16(c5, vacc5x01234567);
289       c5 = (__fp16*) ((uintptr_t) c5 + cn_stride);
290       vst1q_f16(c6, vacc6x01234567);
291       c6 = (__fp16*) ((uintptr_t) c6 + cn_stride);
292       vst1q_f16(c7, vacc7x01234567);
293       c7 = (__fp16*) ((uintptr_t) c7 + cn_stride);
294 
295       a0 = (const __fp16*) ((uintptr_t) a0 - kc);
296       a1 = (const __fp16*) ((uintptr_t) a1 - kc);
297       a2 = (const __fp16*) ((uintptr_t) a2 - kc);
298       a3 = (const __fp16*) ((uintptr_t) a3 - kc);
299       a4 = (const __fp16*) ((uintptr_t) a4 - kc);
300       a5 = (const __fp16*) ((uintptr_t) a5 - kc);
301       a6 = (const __fp16*) ((uintptr_t) a6 - kc);
302       a7 = (const __fp16*) ((uintptr_t) a7 - kc);
303 
304       nc -= 8;
305     } else {
306       float16x4_t vacc0x0123 = vget_low_f16(vacc0x01234567);
307       float16x4_t vacc1x0123 = vget_low_f16(vacc1x01234567);
308       float16x4_t vacc2x0123 = vget_low_f16(vacc2x01234567);
309       float16x4_t vacc3x0123 = vget_low_f16(vacc3x01234567);
310       float16x4_t vacc4x0123 = vget_low_f16(vacc4x01234567);
311       float16x4_t vacc5x0123 = vget_low_f16(vacc5x01234567);
312       float16x4_t vacc6x0123 = vget_low_f16(vacc6x01234567);
313       float16x4_t vacc7x0123 = vget_low_f16(vacc7x01234567);
314       if (nc & 4) {
315         vst1_f16(c0, vacc0x0123); c0 += 4;
316         vst1_f16(c1, vacc1x0123); c1 += 4;
317         vst1_f16(c2, vacc2x0123); c2 += 4;
318         vst1_f16(c3, vacc3x0123); c3 += 4;
319         vst1_f16(c4, vacc4x0123); c4 += 4;
320         vst1_f16(c5, vacc5x0123); c5 += 4;
321         vst1_f16(c6, vacc6x0123); c6 += 4;
322         vst1_f16(c7, vacc7x0123); c7 += 4;
323 
324         vacc0x0123 = vget_high_f16(vacc0x01234567);
325         vacc1x0123 = vget_high_f16(vacc1x01234567);
326         vacc2x0123 = vget_high_f16(vacc2x01234567);
327         vacc3x0123 = vget_high_f16(vacc3x01234567);
328         vacc4x0123 = vget_high_f16(vacc4x01234567);
329         vacc5x0123 = vget_high_f16(vacc5x01234567);
330         vacc6x0123 = vget_high_f16(vacc6x01234567);
331         vacc7x0123 = vget_high_f16(vacc7x01234567);
332       }
333       if (nc & 2) {
334         vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vacc0x0123), 0); c0 += 2;
335         vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vacc1x0123), 0); c1 += 2;
336         vst1_lane_u32((void*) c2, vreinterpret_u32_f16(vacc2x0123), 0); c2 += 2;
337         vst1_lane_u32((void*) c3, vreinterpret_u32_f16(vacc3x0123), 0); c3 += 2;
338         vst1_lane_u32((void*) c4, vreinterpret_u32_f16(vacc4x0123), 0); c4 += 2;
339         vst1_lane_u32((void*) c5, vreinterpret_u32_f16(vacc5x0123), 0); c5 += 2;
340         vst1_lane_u32((void*) c6, vreinterpret_u32_f16(vacc6x0123), 0); c6 += 2;
341         vst1_lane_u32((void*) c7, vreinterpret_u32_f16(vacc7x0123), 0); c7 += 2;
342 
343         vacc0x0123 = vext_f16(vacc0x0123, vacc0x0123, 2);
344         vacc1x0123 = vext_f16(vacc1x0123, vacc1x0123, 2);
345         vacc2x0123 = vext_f16(vacc2x0123, vacc2x0123, 2);
346         vacc3x0123 = vext_f16(vacc3x0123, vacc3x0123, 2);
347         vacc4x0123 = vext_f16(vacc4x0123, vacc4x0123, 2);
348         vacc5x0123 = vext_f16(vacc5x0123, vacc5x0123, 2);
349         vacc6x0123 = vext_f16(vacc6x0123, vacc6x0123, 2);
350         vacc7x0123 = vext_f16(vacc7x0123, vacc7x0123, 2);
351       }
352       if (nc & 1) {
353         vst1_lane_f16(c0, vacc0x0123, 0);
354         vst1_lane_f16(c1, vacc1x0123, 0);
355         vst1_lane_f16(c2, vacc2x0123, 0);
356         vst1_lane_f16(c3, vacc3x0123, 0);
357         vst1_lane_f16(c4, vacc4x0123, 0);
358         vst1_lane_f16(c5, vacc5x0123, 0);
359         vst1_lane_f16(c6, vacc6x0123, 0);
360         vst1_lane_f16(c7, vacc7x0123, 0);
361       }
362 
363       nc = 0;
364     }
365   } while (nc != 0);
366 }
367