1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/c4-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_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const int8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(mr != 0);
31 assert(mr <= 3);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(int8_t) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38
39 const int8_t* a0 = a;
40 int8_t* c0 = c;
41 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
42 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
43 if XNN_UNPREDICTABLE(mr < 2) {
44 a1 = a0;
45 c1 = c0;
46 }
47 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
48 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
49 if XNN_UNPREDICTABLE(mr <= 2) {
50 a2 = a1;
51 c2 = c1;
52 }
53
54 kc = round_up_po2(kc, 8 * sizeof(int8_t));
55 do {
56 int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
57 int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
58 int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
59 int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
60 int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
61 int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
62 int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
63 int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
64 int32x4_t vacc1x01 = vacc0x01;
65 int32x4_t vacc1x23 = vacc0x23;
66 int32x4_t vacc1x45 = vacc0x45;
67 int32x4_t vacc1x67 = vacc0x67;
68 int32x4_t vacc1x89 = vacc0x89;
69 int32x4_t vacc1xAB = vacc0xAB;
70 int32x4_t vacc1xCD = vacc0xCD;
71 int32x4_t vacc1xEF = vacc0xEF;
72 int32x4_t vacc2x01 = vacc0x01;
73 int32x4_t vacc2x23 = vacc0x23;
74 int32x4_t vacc2x45 = vacc0x45;
75 int32x4_t vacc2x67 = vacc0x67;
76 int32x4_t vacc2x89 = vacc0x89;
77 int32x4_t vacc2xAB = vacc0xAB;
78 int32x4_t vacc2xCD = vacc0xCD;
79 int32x4_t vacc2xEF = vacc0xEF;
80
81 size_t k = kc;
82 while (k >= 16 * sizeof(int8_t)) {
83 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
84 int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
85 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
86 int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
87 int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
88 int8x8_t va2x1 = vld1_s8(a2); a2 += 8;
89
90 const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
91 const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
92 const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
93 const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
94 const int8x8_t vb89c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
95 const int8x8_t vbABc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
96 const int8x8_t vbCDc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
97 const int8x8_t vbEFc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
98 const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
99 const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
100 const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
101 const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
102 const int8x8_t vb89c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
103 const int8x8_t vbABc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
104 const int8x8_t vbCDc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
105 const int8x8_t vbEFc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
106
107 int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
108 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1x0);
109 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2x0);
110 const int8x8_t vb01c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
111 vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0x1);
112 vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1x1);
113 vprod2x01c0 = vmlal_s8(vprod2x01c0, vb01c0x1, va2x1);
114 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
115 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
116 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
117 int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
118 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1x0);
119 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2x0);
120 const int8x8_t vb23c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
121 vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0x1);
122 vprod1x23c0 = vmlal_s8(vprod1x23c0, vb23c0x1, va1x1);
123 vprod2x23c0 = vmlal_s8(vprod2x23c0, vb23c0x1, va2x1);
124 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
125 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
126 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
127 int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
128 int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1x0);
129 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2x0);
130 const int8x8_t vb45c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
131 vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0x1);
132 vprod1x45c0 = vmlal_s8(vprod1x45c0, vb45c0x1, va1x1);
133 vprod2x45c0 = vmlal_s8(vprod2x45c0, vb45c0x1, va2x1);
134 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
135 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
136 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
137 int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
138 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1x0);
139 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2x0);
140 const int8x8_t vb67c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
141 vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0x1);
142 vprod1x67c0 = vmlal_s8(vprod1x67c0, vb67c0x1, va1x1);
143 vprod2x67c0 = vmlal_s8(vprod2x67c0, vb67c0x1, va2x1);
144 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
145 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
146 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
147 int16x8_t vprod0x89c0 = vmull_s8(vb89c0x0, va0x0);
148 int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1x0);
149 int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2x0);
150 const int8x8_t vb89c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
151 vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0x1);
152 vprod1x89c0 = vmlal_s8(vprod1x89c0, vb89c0x1, va1x1);
153 vprod2x89c0 = vmlal_s8(vprod2x89c0, vb89c0x1, va2x1);
154 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
155 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
156 vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c0);
157 int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0x0);
158 int16x8_t vprod1xABc0 = vmull_s8(vbABc0x0, va1x0);
159 int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2x0);
160 const int8x8_t vbABc0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
161 vprod0xABc0 = vmlal_s8(vprod0xABc0, vbABc0x1, va0x1);
162 vprod1xABc0 = vmlal_s8(vprod1xABc0, vbABc0x1, va1x1);
163 vprod2xABc0 = vmlal_s8(vprod2xABc0, vbABc0x1, va2x1);
164 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
165 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
166 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0);
167 int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0x0);
168 int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0x0, va1x0);
169 int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2x0);
170 const int8x8_t vbCDc0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
171 vprod0xCDc0 = vmlal_s8(vprod0xCDc0, vbCDc0x1, va0x1);
172 vprod1xCDc0 = vmlal_s8(vprod1xCDc0, vbCDc0x1, va1x1);
173 vprod2xCDc0 = vmlal_s8(vprod2xCDc0, vbCDc0x1, va2x1);
174 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
175 vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
176 vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc0);
177 int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0x0, va0x0);
178 int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0x0, va1x0);
179 int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0x0, va2x0);
180 const int8x8_t vbEFc0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
181 vprod0xEFc0 = vmlal_s8(vprod0xEFc0, vbEFc0x1, va0x1);
182 vprod1xEFc0 = vmlal_s8(vprod1xEFc0, vbEFc0x1, va1x1);
183 vprod2xEFc0 = vmlal_s8(vprod2xEFc0, vbEFc0x1, va2x1);
184 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
185 vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
186 vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc0);
187 va0x0 = vext_s8(va0x0, va0x0, 4);
188 va0x1 = vext_s8(va0x1, va0x1, 4);
189 va1x0 = vext_s8(va1x0, va1x0, 4);
190 va1x1 = vext_s8(va1x1, va1x1, 4);
191 va2x0 = vext_s8(va2x0, va2x0, 4);
192 va2x1 = vext_s8(va2x1, va2x1, 4);
193 int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
194 int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1x0);
195 int16x8_t vprod2x01c1 = vmull_s8(vb01c1x0, va2x0);
196 const int8x8_t vb01c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
197 vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0x1);
198 vprod1x01c1 = vmlal_s8(vprod1x01c1, vb01c1x1, va1x1);
199 vprod2x01c1 = vmlal_s8(vprod2x01c1, vb01c1x1, va2x1);
200 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
201 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
202 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
203 int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
204 int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1x0);
205 int16x8_t vprod2x23c1 = vmull_s8(vb23c1x0, va2x0);
206 const int8x8_t vb23c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
207 vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0x1);
208 vprod1x23c1 = vmlal_s8(vprod1x23c1, vb23c1x1, va1x1);
209 vprod2x23c1 = vmlal_s8(vprod2x23c1, vb23c1x1, va2x1);
210 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
211 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
212 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
213 int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
214 int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1x0);
215 int16x8_t vprod2x45c1 = vmull_s8(vb45c1x0, va2x0);
216 const int8x8_t vb45c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
217 vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0x1);
218 vprod1x45c1 = vmlal_s8(vprod1x45c1, vb45c1x1, va1x1);
219 vprod2x45c1 = vmlal_s8(vprod2x45c1, vb45c1x1, va2x1);
220 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
221 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
222 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
223 int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
224 int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1x0);
225 int16x8_t vprod2x67c1 = vmull_s8(vb67c1x0, va2x0);
226 const int8x8_t vb67c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
227 vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0x1);
228 vprod1x67c1 = vmlal_s8(vprod1x67c1, vb67c1x1, va1x1);
229 vprod2x67c1 = vmlal_s8(vprod2x67c1, vb67c1x1, va2x1);
230 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
231 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
232 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
233 int16x8_t vprod0x89c1 = vmull_s8(vb89c1x0, va0x0);
234 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0);
235 int16x8_t vprod2x89c1 = vmull_s8(vb89c1x0, va2x0);
236 const int8x8_t vb89c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
237 vprod0x89c1 = vmlal_s8(vprod0x89c1, vb89c1x1, va0x1);
238 vprod1x89c1 = vmlal_s8(vprod1x89c1, vb89c1x1, va1x1);
239 vprod2x89c1 = vmlal_s8(vprod2x89c1, vb89c1x1, va2x1);
240 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
241 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
242 vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c1);
243 int16x8_t vprod0xABc1 = vmull_s8(vbABc1x0, va0x0);
244 int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1x0);
245 int16x8_t vprod2xABc1 = vmull_s8(vbABc1x0, va2x0);
246 const int8x8_t vbABc1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
247 vprod0xABc1 = vmlal_s8(vprod0xABc1, vbABc1x1, va0x1);
248 vprod1xABc1 = vmlal_s8(vprod1xABc1, vbABc1x1, va1x1);
249 vprod2xABc1 = vmlal_s8(vprod2xABc1, vbABc1x1, va2x1);
250 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
251 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
252 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc1);
253 int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1x0, va0x0);
254 int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1x0);
255 int16x8_t vprod2xCDc1 = vmull_s8(vbCDc1x0, va2x0);
256 const int8x8_t vbCDc1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
257 vprod0xCDc1 = vmlal_s8(vprod0xCDc1, vbCDc1x1, va0x1);
258 vprod1xCDc1 = vmlal_s8(vprod1xCDc1, vbCDc1x1, va1x1);
259 vprod2xCDc1 = vmlal_s8(vprod2xCDc1, vbCDc1x1, va2x1);
260 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
261 vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
262 vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc1);
263 int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1x0, va0x0);
264 int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1x0, va1x0);
265 int16x8_t vprod2xEFc1 = vmull_s8(vbEFc1x0, va2x0);
266 const int8x8_t vbEFc1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
267 vprod0xEFc1 = vmlal_s8(vprod0xEFc1, vbEFc1x1, va0x1);
268 vprod1xEFc1 = vmlal_s8(vprod1xEFc1, vbEFc1x1, va1x1);
269 vprod2xEFc1 = vmlal_s8(vprod2xEFc1, vbEFc1x1, va2x1);
270 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
271 vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
272 vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc1);
273
274 k -= 16 * sizeof(int8_t);
275 }
276 if (k != 0) {
277 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
278 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
279 int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
280
281 const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
282 const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
283 const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
284 const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
285 const int8x8_t vb89c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
286 const int8x8_t vbABc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
287 const int8x8_t vbCDc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
288 const int8x8_t vbEFc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
289 const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
290 const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
291 const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
292 const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
293 const int8x8_t vb89c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
294 const int8x8_t vbABc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
295 const int8x8_t vbCDc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
296 const int8x8_t vbEFc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
297
298 int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
299 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1x0);
300 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2x0);
301 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
302 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
303 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
304 int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
305 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1x0);
306 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2x0);
307 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
308 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
309 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
310 int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
311 int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1x0);
312 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2x0);
313 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
314 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
315 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
316 int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
317 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1x0);
318 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2x0);
319 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
320 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
321 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
322 int16x8_t vprod0x89c0 = vmull_s8(vb89c0x0, va0x0);
323 int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1x0);
324 int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2x0);
325 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
326 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
327 vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c0);
328 int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0x0);
329 int16x8_t vprod1xABc0 = vmull_s8(vbABc0x0, va1x0);
330 int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2x0);
331 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
332 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
333 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0);
334 int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0x0);
335 int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0x0, va1x0);
336 int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2x0);
337 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
338 vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
339 vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc0);
340 int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0x0, va0x0);
341 int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0x0, va1x0);
342 int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0x0, va2x0);
343 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
344 vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
345 vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc0);
346 va0x0 = vext_s8(va0x0, va0x0, 4);
347 va1x0 = vext_s8(va1x0, va1x0, 4);
348 va2x0 = vext_s8(va2x0, va2x0, 4);
349 int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
350 int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1x0);
351 int16x8_t vprod2x01c1 = vmull_s8(vb01c1x0, va2x0);
352 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
353 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
354 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
355 int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
356 int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1x0);
357 int16x8_t vprod2x23c1 = vmull_s8(vb23c1x0, va2x0);
358 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
359 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
360 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
361 int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
362 int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1x0);
363 int16x8_t vprod2x45c1 = vmull_s8(vb45c1x0, va2x0);
364 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
365 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
366 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
367 int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
368 int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1x0);
369 int16x8_t vprod2x67c1 = vmull_s8(vb67c1x0, va2x0);
370 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
371 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
372 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
373 int16x8_t vprod0x89c1 = vmull_s8(vb89c1x0, va0x0);
374 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0);
375 int16x8_t vprod2x89c1 = vmull_s8(vb89c1x0, va2x0);
376 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
377 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
378 vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c1);
379 int16x8_t vprod0xABc1 = vmull_s8(vbABc1x0, va0x0);
380 int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1x0);
381 int16x8_t vprod2xABc1 = vmull_s8(vbABc1x0, va2x0);
382 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
383 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
384 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc1);
385 int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1x0, va0x0);
386 int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1x0);
387 int16x8_t vprod2xCDc1 = vmull_s8(vbCDc1x0, va2x0);
388 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
389 vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
390 vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc1);
391 int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1x0, va0x0);
392 int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1x0, va1x0);
393 int16x8_t vprod2xEFc1 = vmull_s8(vbEFc1x0, va2x0);
394 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
395 vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
396 vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc1);
397
398 }
399
400 #if XNN_ARCH_ARM64
401 int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
402 int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
403 int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB);
404 int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF);
405 int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
406 int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
407 int32x4_t vacc1x89AB = vpaddq_s32(vacc1x89, vacc1xAB);
408 int32x4_t vacc1xCDEF = vpaddq_s32(vacc1xCD, vacc1xEF);
409 int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23);
410 int32x4_t vacc2x4567 = vpaddq_s32(vacc2x45, vacc2x67);
411 int32x4_t vacc2x89AB = vpaddq_s32(vacc2x89, vacc2xAB);
412 int32x4_t vacc2xCDEF = vpaddq_s32(vacc2xCD, vacc2xEF);
413 #else
414 const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
415 const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
416 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
417 const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
418 const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
419 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
420 const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89));
421 const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB));
422 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB);
423 const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD));
424 const int32x2_t vsum0xEF = vpadd_s32(vget_low_s32(vacc0xEF), vget_high_s32(vacc0xEF));
425 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF);
426 const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
427 const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
428 int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
429 const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
430 const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
431 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
432 const int32x2_t vsum1x89 = vpadd_s32(vget_low_s32(vacc1x89), vget_high_s32(vacc1x89));
433 const int32x2_t vsum1xAB = vpadd_s32(vget_low_s32(vacc1xAB), vget_high_s32(vacc1xAB));
434 int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB);
435 const int32x2_t vsum1xCD = vpadd_s32(vget_low_s32(vacc1xCD), vget_high_s32(vacc1xCD));
436 const int32x2_t vsum1xEF = vpadd_s32(vget_low_s32(vacc1xEF), vget_high_s32(vacc1xEF));
437 int32x4_t vacc1xCDEF = vcombine_s32(vsum1xCD, vsum1xEF);
438 const int32x2_t vsum2x01 = vpadd_s32(vget_low_s32(vacc2x01), vget_high_s32(vacc2x01));
439 const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23));
440 int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23);
441 const int32x2_t vsum2x45 = vpadd_s32(vget_low_s32(vacc2x45), vget_high_s32(vacc2x45));
442 const int32x2_t vsum2x67 = vpadd_s32(vget_low_s32(vacc2x67), vget_high_s32(vacc2x67));
443 int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67);
444 const int32x2_t vsum2x89 = vpadd_s32(vget_low_s32(vacc2x89), vget_high_s32(vacc2x89));
445 const int32x2_t vsum2xAB = vpadd_s32(vget_low_s32(vacc2xAB), vget_high_s32(vacc2xAB));
446 int32x4_t vacc2x89AB = vcombine_s32(vsum2x89, vsum2xAB);
447 const int32x2_t vsum2xCD = vpadd_s32(vget_low_s32(vacc2xCD), vget_high_s32(vacc2xCD));
448 const int32x2_t vsum2xEF = vpadd_s32(vget_low_s32(vacc2xEF), vget_high_s32(vacc2xEF));
449 int32x4_t vacc2xCDEF = vcombine_s32(vsum2xCD, vsum2xEF);
450 #endif
451
452 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
453 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
454 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
455
456 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
457 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
458 vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
459 vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
460 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
461 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
462 vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
463 vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
464 vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
465 vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
466 vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
467 vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
468
469 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
470 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
471 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
472 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
473 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
474 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
475 vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
476 vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
477 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
478 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
479 vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
480 vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
481
482 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
483 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
484 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
485 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
486 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
487 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
488 vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
489 vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
490 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
491 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
492 vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
493 vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
494
495 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
496 #if XNN_ARCH_ARM64
497 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
498 int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
499 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
500 int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
501 int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
502 int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
503
504 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
505 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
506 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
507 vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
508 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
509 vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
510
511 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
512 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
513 int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
514 #else
515 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
516 int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
517 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
518 int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
519 int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
520 int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
521
522 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
523 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
524 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
525 vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
526 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
527 vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
528
529 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
530 int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
531 int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
532 #endif
533
534 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
535 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
536 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
537 vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
538
539 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
540 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
541 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
542 vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
543
544 if (nc >= 16) {
545 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
546 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
547 vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
548
549 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
550 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
551 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
552
553 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
554 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
555 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
556
557 nc -= 16;
558 } else {
559 // Final case where not all of the 16 columns fit in the destination.
560 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
561 int8x8_t vout2x01234567 = vget_low_s8(vout2x0123456789ABCDEF);
562 if (nc & 8) {
563 vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
564 vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
565 vst1_s8(c2, vout2x01234567); c2 += 8;
566 vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
567 vout2x01234567 = vget_high_s8(vout2x0123456789ABCDEF);
568 }
569 if (nc & 4) {
570 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
571 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
572 vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
573 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
574 vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
575 }
576 if (nc & 2) {
577 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
578 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
579 vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
580 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
581 vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
582 }
583 if (nc & 1) {
584 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
585 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
586 vst1_lane_s8(c2, vout2x01234567, 0);
587 }
588
589 nc = 0;
590 }
591 } while (nc != 0);
592 }
593