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_gemminc_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 void * restrict acc,const struct xnn_f16_scaleminmax_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_f16_gemminc_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 void*restrict acc,
31 const struct xnn_f16_scaleminmax_params params[restrict XNN_MIN_ELEMENTS(1)])
32 {
33 assert(mr != 0);
34 assert(mr <= 8);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(kc % sizeof(__fp16) == 0);
38 assert(a != NULL);
39 assert(w != NULL);
40 assert(c != NULL);
41 assert(acc != NULL);
42
43 const __fp16* a0 = (const __fp16*) a;
44 __fp16* c0 = (__fp16*) c;
45 const __fp16* a1 = (const __fp16*) ((uintptr_t) a0 + a_stride);
46 __fp16* c1 = (__fp16*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr < 2) {
48 a1 = a0;
49 c1 = c0;
50 }
51 const __fp16* a2 = (const __fp16*) ((uintptr_t) a1 + a_stride);
52 __fp16* c2 = (__fp16*) ((uintptr_t) c1 + cm_stride);
53 if XNN_UNPREDICTABLE(mr <= 2) {
54 a2 = a1;
55 c2 = c1;
56 }
57 const __fp16* a3 = (const __fp16*) ((uintptr_t) a2 + a_stride);
58 __fp16* c3 = (__fp16*) ((uintptr_t) c2 + cm_stride);
59 if XNN_UNPREDICTABLE(mr < 4) {
60 a3 = a2;
61 c3 = c2;
62 }
63 const __fp16* a4 = (const __fp16*) ((uintptr_t) a3 + a_stride);
64 __fp16* c4 = (__fp16*) ((uintptr_t) c3 + cm_stride);
65 if XNN_UNPREDICTABLE(mr <= 4) {
66 a4 = a3;
67 c4 = c3;
68 }
69 const __fp16* a5 = (const __fp16*) ((uintptr_t) a4 + a_stride);
70 __fp16* c5 = (__fp16*) ((uintptr_t) c4 + cm_stride);
71 if XNN_UNPREDICTABLE(mr < 6) {
72 a5 = a4;
73 c5 = c4;
74 }
75 const __fp16* a6 = (const __fp16*) ((uintptr_t) a5 + a_stride);
76 __fp16* c6 = (__fp16*) ((uintptr_t) c5 + cm_stride);
77 if XNN_UNPREDICTABLE(mr <= 6) {
78 a6 = a5;
79 c6 = c5;
80 }
81 const __fp16* a7 = (const __fp16*) ((uintptr_t) a6 + a_stride);
82 __fp16* c7 = (__fp16*) ((uintptr_t) c6 + cm_stride);
83 if XNN_UNPREDICTABLE(mr != 8) {
84 a7 = a6;
85 c7 = c6;
86 }
87
88 do {
89 float16x8_t vacc0x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
90 float16x8_t vacc1x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
91 float16x8_t vacc2x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
92 float16x8_t vacc3x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
93 float16x8_t vacc4x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
94 float16x8_t vacc5x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
95 float16x8_t vacc6x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
96 float16x8_t vacc7x01234567 = vld1q_f16(acc); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
97
98 size_t k = kc;
99 while (k >= 4 * sizeof(__fp16)) {
100 const float16x4_t va0 = vld1_f16(a0); a0 += 4;
101 const float16x4_t va1 = vld1_f16(a1); a1 += 4;
102 const float16x4_t va2 = vld1_f16(a2); a2 += 4;
103 const float16x4_t va3 = vld1_f16(a3); a3 += 4;
104 const float16x4_t va4 = vld1_f16(a4); a4 += 4;
105 const float16x4_t va5 = vld1_f16(a5); a5 += 4;
106 const float16x4_t va6 = vld1_f16(a6); a6 += 4;
107 const float16x4_t va7 = vld1_f16(a7); a7 += 4;
108
109 const float16x8_t vb01234567c0 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
110
111 #if XNN_ARCH_ARM64
112 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c0, va0, 0);
113 vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c0, va1, 0);
114 vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c0, va2, 0);
115 vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c0, va3, 0);
116 vacc4x01234567 = vfmaq_lane_f16(vacc4x01234567, vb01234567c0, va4, 0);
117 vacc5x01234567 = vfmaq_lane_f16(vacc5x01234567, vb01234567c0, va5, 0);
118 vacc6x01234567 = vfmaq_lane_f16(vacc6x01234567, vb01234567c0, va6, 0);
119 vacc7x01234567 = vfmaq_lane_f16(vacc7x01234567, vb01234567c0, va7, 0);
120 #else
121 const float16x8_t va0c0 = vdupq_lane_f16(va0, 0);
122 const float16x8_t va1c0 = vdupq_lane_f16(va1, 0);
123 const float16x8_t va2c0 = vdupq_lane_f16(va2, 0);
124 const float16x8_t va3c0 = vdupq_lane_f16(va3, 0);
125 const float16x8_t va4c0 = vdupq_lane_f16(va4, 0);
126 const float16x8_t va5c0 = vdupq_lane_f16(va5, 0);
127 const float16x8_t va6c0 = vdupq_lane_f16(va6, 0);
128 const float16x8_t va7c0 = vdupq_lane_f16(va7, 0);
129
130 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c0, vb01234567c0);
131 vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1c0, vb01234567c0);
132 vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c0, vb01234567c0);
133 vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3c0, vb01234567c0);
134 vacc4x01234567 = vfmaq_f16(vacc4x01234567, va4c0, vb01234567c0);
135 vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5c0, vb01234567c0);
136 vacc6x01234567 = vfmaq_f16(vacc6x01234567, va6c0, vb01234567c0);
137 vacc7x01234567 = vfmaq_f16(vacc7x01234567, va7c0, vb01234567c0);
138 #endif
139 const float16x8_t vb01234567c1 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
140
141 #if XNN_ARCH_ARM64
142 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c1, va0, 1);
143 vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c1, va1, 1);
144 vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c1, va2, 1);
145 vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c1, va3, 1);
146 vacc4x01234567 = vfmaq_lane_f16(vacc4x01234567, vb01234567c1, va4, 1);
147 vacc5x01234567 = vfmaq_lane_f16(vacc5x01234567, vb01234567c1, va5, 1);
148 vacc6x01234567 = vfmaq_lane_f16(vacc6x01234567, vb01234567c1, va6, 1);
149 vacc7x01234567 = vfmaq_lane_f16(vacc7x01234567, vb01234567c1, va7, 1);
150 #else
151 const float16x8_t va0c1 = vdupq_lane_f16(va0, 1);
152 const float16x8_t va1c1 = vdupq_lane_f16(va1, 1);
153 const float16x8_t va2c1 = vdupq_lane_f16(va2, 1);
154 const float16x8_t va3c1 = vdupq_lane_f16(va3, 1);
155 const float16x8_t va4c1 = vdupq_lane_f16(va4, 1);
156 const float16x8_t va5c1 = vdupq_lane_f16(va5, 1);
157 const float16x8_t va6c1 = vdupq_lane_f16(va6, 1);
158 const float16x8_t va7c1 = vdupq_lane_f16(va7, 1);
159
160 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c1, vb01234567c1);
161 vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1c1, vb01234567c1);
162 vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c1, vb01234567c1);
163 vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3c1, vb01234567c1);
164 vacc4x01234567 = vfmaq_f16(vacc4x01234567, va4c1, vb01234567c1);
165 vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5c1, vb01234567c1);
166 vacc6x01234567 = vfmaq_f16(vacc6x01234567, va6c1, vb01234567c1);
167 vacc7x01234567 = vfmaq_f16(vacc7x01234567, va7c1, vb01234567c1);
168 #endif
169 const float16x8_t vb01234567c2 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
170
171 #if XNN_ARCH_ARM64
172 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c2, va0, 2);
173 vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c2, va1, 2);
174 vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c2, va2, 2);
175 vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c2, va3, 2);
176 vacc4x01234567 = vfmaq_lane_f16(vacc4x01234567, vb01234567c2, va4, 2);
177 vacc5x01234567 = vfmaq_lane_f16(vacc5x01234567, vb01234567c2, va5, 2);
178 vacc6x01234567 = vfmaq_lane_f16(vacc6x01234567, vb01234567c2, va6, 2);
179 vacc7x01234567 = vfmaq_lane_f16(vacc7x01234567, vb01234567c2, va7, 2);
180 #else
181 const float16x8_t va0c2 = vdupq_lane_f16(va0, 2);
182 const float16x8_t va1c2 = vdupq_lane_f16(va1, 2);
183 const float16x8_t va2c2 = vdupq_lane_f16(va2, 2);
184 const float16x8_t va3c2 = vdupq_lane_f16(va3, 2);
185 const float16x8_t va4c2 = vdupq_lane_f16(va4, 2);
186 const float16x8_t va5c2 = vdupq_lane_f16(va5, 2);
187 const float16x8_t va6c2 = vdupq_lane_f16(va6, 2);
188 const float16x8_t va7c2 = vdupq_lane_f16(va7, 2);
189
190 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c2, vb01234567c2);
191 vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1c2, vb01234567c2);
192 vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c2, vb01234567c2);
193 vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3c2, vb01234567c2);
194 vacc4x01234567 = vfmaq_f16(vacc4x01234567, va4c2, vb01234567c2);
195 vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5c2, vb01234567c2);
196 vacc6x01234567 = vfmaq_f16(vacc6x01234567, va6c2, vb01234567c2);
197 vacc7x01234567 = vfmaq_f16(vacc7x01234567, va7c2, vb01234567c2);
198 #endif
199 const float16x8_t vb01234567c3 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
200
201 #if XNN_ARCH_ARM64
202 vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c3, va0, 3);
203 vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c3, va1, 3);
204 vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c3, va2, 3);
205 vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c3, va3, 3);
206 vacc4x01234567 = vfmaq_lane_f16(vacc4x01234567, vb01234567c3, va4, 3);
207 vacc5x01234567 = vfmaq_lane_f16(vacc5x01234567, vb01234567c3, va5, 3);
208 vacc6x01234567 = vfmaq_lane_f16(vacc6x01234567, vb01234567c3, va6, 3);
209 vacc7x01234567 = vfmaq_lane_f16(vacc7x01234567, vb01234567c3, va7, 3);
210 #else
211 const float16x8_t va0c3 = vdupq_lane_f16(va0, 3);
212 const float16x8_t va1c3 = vdupq_lane_f16(va1, 3);
213 const float16x8_t va2c3 = vdupq_lane_f16(va2, 3);
214 const float16x8_t va3c3 = vdupq_lane_f16(va3, 3);
215 const float16x8_t va4c3 = vdupq_lane_f16(va4, 3);
216 const float16x8_t va5c3 = vdupq_lane_f16(va5, 3);
217 const float16x8_t va6c3 = vdupq_lane_f16(va6, 3);
218 const float16x8_t va7c3 = vdupq_lane_f16(va7, 3);
219
220 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0c3, vb01234567c3);
221 vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1c3, vb01234567c3);
222 vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c3, vb01234567c3);
223 vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3c3, vb01234567c3);
224 vacc4x01234567 = vfmaq_f16(vacc4x01234567, va4c3, vb01234567c3);
225 vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5c3, vb01234567c3);
226 vacc6x01234567 = vfmaq_f16(vacc6x01234567, va6c3, vb01234567c3);
227 vacc7x01234567 = vfmaq_f16(vacc7x01234567, va7c3, vb01234567c3);
228 #endif
229
230 k -= 4 * sizeof(__fp16);
231 }
232 if XNN_UNLIKELY(k != 0) {
233 do {
234 const float16x8_t va0 = vld1q_dup_f16(a0); a0 += 1;
235 const float16x8_t va1 = vld1q_dup_f16(a1); a1 += 1;
236 const float16x8_t va2 = vld1q_dup_f16(a2); a2 += 1;
237 const float16x8_t va3 = vld1q_dup_f16(a3); a3 += 1;
238 const float16x8_t va4 = vld1q_dup_f16(a4); a4 += 1;
239 const float16x8_t va5 = vld1q_dup_f16(a5); a5 += 1;
240 const float16x8_t va6 = vld1q_dup_f16(a6); a6 += 1;
241 const float16x8_t va7 = vld1q_dup_f16(a7); a7 += 1;
242
243 const float16x8_t vb01234567 = vld1q_f16(w); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
244
245 vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0, vb01234567);
246 vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1, vb01234567);
247 vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2, vb01234567);
248 vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3, vb01234567);
249 vacc4x01234567 = vfmaq_f16(vacc4x01234567, va4, vb01234567);
250 vacc5x01234567 = vfmaq_f16(vacc5x01234567, va5, vb01234567);
251 vacc6x01234567 = vfmaq_f16(vacc6x01234567, va6, vb01234567);
252 vacc7x01234567 = vfmaq_f16(vacc7x01234567, va7, vb01234567);
253
254 k -= sizeof(__fp16);
255 } while (k != 0);
256 }
257
258 const float16x8_t vscale = vld1q_dup_f16((const __fp16*) ¶ms->scale);
259 vacc0x01234567 = vmulq_f16(vacc0x01234567, vscale);
260 vacc1x01234567 = vmulq_f16(vacc1x01234567, vscale);
261 vacc2x01234567 = vmulq_f16(vacc2x01234567, vscale);
262 vacc3x01234567 = vmulq_f16(vacc3x01234567, vscale);
263 vacc4x01234567 = vmulq_f16(vacc4x01234567, vscale);
264 vacc5x01234567 = vmulq_f16(vacc5x01234567, vscale);
265 vacc6x01234567 = vmulq_f16(vacc6x01234567, vscale);
266 vacc7x01234567 = vmulq_f16(vacc7x01234567, vscale);
267
268 const float16x8_t vmax = vld1q_dup_f16((const __fp16*) ¶ms->max);
269 vacc0x01234567 = vminq_f16(vacc0x01234567, vmax);
270 vacc1x01234567 = vminq_f16(vacc1x01234567, vmax);
271 vacc2x01234567 = vminq_f16(vacc2x01234567, vmax);
272 vacc3x01234567 = vminq_f16(vacc3x01234567, vmax);
273 vacc4x01234567 = vminq_f16(vacc4x01234567, vmax);
274 vacc5x01234567 = vminq_f16(vacc5x01234567, vmax);
275 vacc6x01234567 = vminq_f16(vacc6x01234567, vmax);
276 vacc7x01234567 = vminq_f16(vacc7x01234567, vmax);
277
278 const float16x8_t vmin = vld1q_dup_f16((const __fp16*) ¶ms->min);
279 vacc0x01234567 = vmaxq_f16(vacc0x01234567, vmin);
280 vacc1x01234567 = vmaxq_f16(vacc1x01234567, vmin);
281 vacc2x01234567 = vmaxq_f16(vacc2x01234567, vmin);
282 vacc3x01234567 = vmaxq_f16(vacc3x01234567, vmin);
283 vacc4x01234567 = vmaxq_f16(vacc4x01234567, vmin);
284 vacc5x01234567 = vmaxq_f16(vacc5x01234567, vmin);
285 vacc6x01234567 = vmaxq_f16(vacc6x01234567, vmin);
286 vacc7x01234567 = vmaxq_f16(vacc7x01234567, vmin);
287
288 if XNN_LIKELY(nc >= 8) {
289 vst1q_f16(c0, vacc0x01234567);
290 c0 = (__fp16*) ((uintptr_t) c0 + cn_stride);
291 vst1q_f16(c1, vacc1x01234567);
292 c1 = (__fp16*) ((uintptr_t) c1 + cn_stride);
293 vst1q_f16(c2, vacc2x01234567);
294 c2 = (__fp16*) ((uintptr_t) c2 + cn_stride);
295 vst1q_f16(c3, vacc3x01234567);
296 c3 = (__fp16*) ((uintptr_t) c3 + cn_stride);
297 vst1q_f16(c4, vacc4x01234567);
298 c4 = (__fp16*) ((uintptr_t) c4 + cn_stride);
299 vst1q_f16(c5, vacc5x01234567);
300 c5 = (__fp16*) ((uintptr_t) c5 + cn_stride);
301 vst1q_f16(c6, vacc6x01234567);
302 c6 = (__fp16*) ((uintptr_t) c6 + cn_stride);
303 vst1q_f16(c7, vacc7x01234567);
304 c7 = (__fp16*) ((uintptr_t) c7 + cn_stride);
305
306 a0 = (const __fp16*) ((uintptr_t) a0 - kc);
307 a1 = (const __fp16*) ((uintptr_t) a1 - kc);
308 a2 = (const __fp16*) ((uintptr_t) a2 - kc);
309 a3 = (const __fp16*) ((uintptr_t) a3 - kc);
310 a4 = (const __fp16*) ((uintptr_t) a4 - kc);
311 a5 = (const __fp16*) ((uintptr_t) a5 - kc);
312 a6 = (const __fp16*) ((uintptr_t) a6 - kc);
313 a7 = (const __fp16*) ((uintptr_t) a7 - kc);
314
315 nc -= 8;
316 } else {
317 float16x4_t vacc0x0123 = vget_low_f16(vacc0x01234567);
318 float16x4_t vacc1x0123 = vget_low_f16(vacc1x01234567);
319 float16x4_t vacc2x0123 = vget_low_f16(vacc2x01234567);
320 float16x4_t vacc3x0123 = vget_low_f16(vacc3x01234567);
321 float16x4_t vacc4x0123 = vget_low_f16(vacc4x01234567);
322 float16x4_t vacc5x0123 = vget_low_f16(vacc5x01234567);
323 float16x4_t vacc6x0123 = vget_low_f16(vacc6x01234567);
324 float16x4_t vacc7x0123 = vget_low_f16(vacc7x01234567);
325 if (nc & 4) {
326 vst1_f16(c0, vacc0x0123); c0 += 4;
327 vst1_f16(c1, vacc1x0123); c1 += 4;
328 vst1_f16(c2, vacc2x0123); c2 += 4;
329 vst1_f16(c3, vacc3x0123); c3 += 4;
330 vst1_f16(c4, vacc4x0123); c4 += 4;
331 vst1_f16(c5, vacc5x0123); c5 += 4;
332 vst1_f16(c6, vacc6x0123); c6 += 4;
333 vst1_f16(c7, vacc7x0123); c7 += 4;
334
335 vacc0x0123 = vget_high_f16(vacc0x01234567);
336 vacc1x0123 = vget_high_f16(vacc1x01234567);
337 vacc2x0123 = vget_high_f16(vacc2x01234567);
338 vacc3x0123 = vget_high_f16(vacc3x01234567);
339 vacc4x0123 = vget_high_f16(vacc4x01234567);
340 vacc5x0123 = vget_high_f16(vacc5x01234567);
341 vacc6x0123 = vget_high_f16(vacc6x01234567);
342 vacc7x0123 = vget_high_f16(vacc7x01234567);
343 }
344 if (nc & 2) {
345 vst1_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpret_u32_f16(vacc0x0123), 0); c0 += 2;
346 vst1_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpret_u32_f16(vacc1x0123), 0); c1 += 2;
347 vst1_lane_u32(__builtin_assume_aligned(c2, 1), vreinterpret_u32_f16(vacc2x0123), 0); c2 += 2;
348 vst1_lane_u32(__builtin_assume_aligned(c3, 1), vreinterpret_u32_f16(vacc3x0123), 0); c3 += 2;
349 vst1_lane_u32(__builtin_assume_aligned(c4, 1), vreinterpret_u32_f16(vacc4x0123), 0); c4 += 2;
350 vst1_lane_u32(__builtin_assume_aligned(c5, 1), vreinterpret_u32_f16(vacc5x0123), 0); c5 += 2;
351 vst1_lane_u32(__builtin_assume_aligned(c6, 1), vreinterpret_u32_f16(vacc6x0123), 0); c6 += 2;
352 vst1_lane_u32(__builtin_assume_aligned(c7, 1), vreinterpret_u32_f16(vacc7x0123), 0); c7 += 2;
353
354 vacc0x0123 = vext_f16(vacc0x0123, vacc0x0123, 2);
355 vacc1x0123 = vext_f16(vacc1x0123, vacc1x0123, 2);
356 vacc2x0123 = vext_f16(vacc2x0123, vacc2x0123, 2);
357 vacc3x0123 = vext_f16(vacc3x0123, vacc3x0123, 2);
358 vacc4x0123 = vext_f16(vacc4x0123, vacc4x0123, 2);
359 vacc5x0123 = vext_f16(vacc5x0123, vacc5x0123, 2);
360 vacc6x0123 = vext_f16(vacc6x0123, vacc6x0123, 2);
361 vacc7x0123 = vext_f16(vacc7x0123, vacc7x0123, 2);
362 }
363 if (nc & 1) {
364 vst1_lane_f16(c0, vacc0x0123, 0);
365 vst1_lane_f16(c1, vacc1x0123, 0);
366 vst1_lane_f16(c2, vacc2x0123, 0);
367 vst1_lane_f16(c3, vacc3x0123, 0);
368 vst1_lane_f16(c4, vacc4x0123, 0);
369 vst1_lane_f16(c5, vacc5x0123, 0);
370 vst1_lane_f16(c6, vacc6x0123, 0);
371 vst1_lane_f16(c7, vacc7x0123, 0);
372 }
373
374 nc = 0;
375 }
376 } while (nc != 0);
377 }
378