1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c2-neon-mull-shuffle.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 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 #include <assert.h>
11
12 #include <arm_neon.h>
13
14 #include <xnnpack/gemm.h>
15 #include <xnnpack/math.h>
16
17
xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2s4__neon_mlal(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2s4__neon_mlal(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const int8_t** restrict a,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 3);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (3 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 int8_t* c0 = c;
44 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 c1 = c0;
47 }
48 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
49 if XNN_UNPREDICTABLE(mr <= 2) {
50 c2 = c1;
51 }
52
53 kc = round_up_po2(kc, 8 * sizeof(int8_t));
54 do {
55 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
56 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
57 int32x4_t vacc1x0123 = vacc0x0123;
58 int32x4_t vacc1x4567 = vacc0x4567;
59 int32x4_t vacc2x0123 = vacc0x0123;
60 int32x4_t vacc2x4567 = vacc0x4567;
61
62 size_t p = ks;
63 do {
64 const int8_t* restrict a0 = a[0];
65 if XNN_UNPREDICTABLE(a0 != zero) {
66 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
67 }
68 const int8_t* restrict a1 = a[1];
69 if XNN_UNPREDICTABLE(a1 != zero) {
70 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
71 }
72 const int8_t* restrict a2 = a[2];
73 if XNN_UNPREDICTABLE(a2 != zero) {
74 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
75 }
76 a += 3;
77
78 size_t k = kc;
79 while (k >= 16 * sizeof(int8_t)) {
80 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
81 int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
82 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
83 int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
84 int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
85 int8x8_t va2x1 = vld1_s8(a2); a2 += 8;
86
87 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
88 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
89 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
90 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
91 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
92 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
93 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
94 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
95
96 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
97 int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1x0);
98 int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0x0, va2x0);
99 const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
100 vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0x1);
101 vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, va1x1);
102 vprod2x0123c0 = vmlal_s8(vprod2x0123c0, vb0123c0x1, va2x1);
103 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
104 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
105 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
106 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
107 int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1x0);
108 int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2x0);
109 const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
110 vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0x1);
111 vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, va1x1);
112 vprod2x4567c0 = vmlal_s8(vprod2x4567c0, vb4567c0x1, va2x1);
113 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
114 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
115 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
116 va0x0 = vext_s8(va0x0, va0x0, 2);
117 va0x1 = vext_s8(va0x1, va0x1, 2);
118 va1x0 = vext_s8(va1x0, va1x0, 2);
119 va1x1 = vext_s8(va1x1, va1x1, 2);
120 va2x0 = vext_s8(va2x0, va2x0, 2);
121 va2x1 = vext_s8(va2x1, va2x1, 2);
122 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
123 int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1x0);
124 int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2x0);
125 const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
126 vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0x1);
127 vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, va1x1);
128 vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, va2x1);
129 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
130 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
131 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
132 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
133 int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1x0);
134 int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1x0, va2x0);
135 const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
136 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0x1);
137 vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, va1x1);
138 vprod2x4567c1 = vmlal_s8(vprod2x4567c1, vb4567c1x1, va2x1);
139 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
140 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
141 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
142 va0x0 = vext_s8(va0x0, va0x0, 2);
143 va0x1 = vext_s8(va0x1, va0x1, 2);
144 va1x0 = vext_s8(va1x0, va1x0, 2);
145 va1x1 = vext_s8(va1x1, va1x1, 2);
146 va2x0 = vext_s8(va2x0, va2x0, 2);
147 va2x1 = vext_s8(va2x1, va2x1, 2);
148 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
149 int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1x0);
150 int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2x0, va2x0);
151 const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
152 vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0x1);
153 vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, va1x1);
154 vprod2x0123c2 = vmlal_s8(vprod2x0123c2, vb0123c2x1, va2x1);
155 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
156 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
157 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
158 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
159 int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1x0);
160 int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2x0, va2x0);
161 const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
162 vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0x1);
163 vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, va1x1);
164 vprod2x4567c2 = vmlal_s8(vprod2x4567c2, vb4567c2x1, va2x1);
165 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
166 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
167 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
168 va0x0 = vext_s8(va0x0, va0x0, 2);
169 va0x1 = vext_s8(va0x1, va0x1, 2);
170 va1x0 = vext_s8(va1x0, va1x0, 2);
171 va1x1 = vext_s8(va1x1, va1x1, 2);
172 va2x0 = vext_s8(va2x0, va2x0, 2);
173 va2x1 = vext_s8(va2x1, va2x1, 2);
174 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
175 int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1x0);
176 int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3x0, va2x0);
177 const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
178 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0x1);
179 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1x1);
180 vprod2x0123c3 = vmlal_s8(vprod2x0123c3, vb0123c3x1, va2x1);
181 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
182 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
183 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
184 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
185 int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1x0);
186 int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3x0, va2x0);
187 const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
188 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0x1);
189 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1x1);
190 vprod2x4567c3 = vmlal_s8(vprod2x4567c3, vb4567c3x1, va2x1);
191 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
192 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
193 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
194
195 k -= 16 * sizeof(int8_t);
196 }
197 if (k != 0) {
198 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
199 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
200 int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
201
202 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
203 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
204 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
205 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
206 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
207 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
208 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
209 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
210
211 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
212 int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1x0);
213 int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0x0, va2x0);
214 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
215 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
216 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
217 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
218 int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1x0);
219 int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2x0);
220 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
221 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
222 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
223 va0x0 = vext_s8(va0x0, va0x0, 2);
224 va1x0 = vext_s8(va1x0, va1x0, 2);
225 va2x0 = vext_s8(va2x0, va2x0, 2);
226 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
227 int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1x0);
228 int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2x0);
229 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
230 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
231 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
232 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
233 int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1x0);
234 int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1x0, va2x0);
235 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
236 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
237 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
238 va0x0 = vext_s8(va0x0, va0x0, 2);
239 va1x0 = vext_s8(va1x0, va1x0, 2);
240 va2x0 = vext_s8(va2x0, va2x0, 2);
241 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
242 int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1x0);
243 int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2x0, va2x0);
244 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
245 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
246 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
247 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
248 int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1x0);
249 int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2x0, va2x0);
250 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
251 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
252 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
253 va0x0 = vext_s8(va0x0, va0x0, 2);
254 va1x0 = vext_s8(va1x0, va1x0, 2);
255 va2x0 = vext_s8(va2x0, va2x0, 2);
256 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
257 int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1x0);
258 int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3x0, va2x0);
259 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
260 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
261 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
262 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
263 int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1x0);
264 int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3x0, va2x0);
265 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
266 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
267 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
268
269 }
270
271 p -= 3 * sizeof(void*);
272 } while (p != 0);
273
274 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
275 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
276 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
277
278 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
279 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
280 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
281 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
282 vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
283 vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
284
285 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
286 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
287 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
288 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
289 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
290 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
291
292 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
293 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
294 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
295 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
296 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
297 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
298
299 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
300 #if XNN_ARCH_ARM64
301 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
302 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
303 int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
304
305 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
306 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
307 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
308
309 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
310 int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
311 #else
312 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
313 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
314 int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
315
316 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
317 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
318 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
319
320 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
321 int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
322 #endif
323
324 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
325 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
326 vout2x01234567 = vmax_s8(vout2x01234567, vget_low_s8(voutput_min));
327
328 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
329 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
330 vout2x01234567 = vmin_s8(vout2x01234567, vget_low_s8(voutput_max));
331
332 if (nc >= 8) {
333 vst1_s8(c2 + 0, vout2x01234567);
334 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
335 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
336
337 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
338 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
339 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
340
341 a = (const int8_t**restrict) ((uintptr_t) a - ks);
342
343 nc -= 8;
344 } else {
345 if (nc & 4) {
346 vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
347 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
348 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
349 vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
350 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
351 }
352 if (nc & 2) {
353 vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
354 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
355 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
356 vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
357 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
358 }
359 if (nc & 1) {
360 vst1_lane_s8(c2, vout2x01234567, 0);
361 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
362 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
363 }
364
365 nc = 0;
366 }
367 } while (nc != 0);
368 }
369