1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c4-neondot.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 #include <assert.h>
11
12 #include <arm_neon.h>
13
14 #include <xnnpack/igemm.h>
15 #include <xnnpack/math.h>
16
17
xnn_qs8_igemm_minmax_rndnu_ukernel_8x8c4__neondot(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_8x8c4__neondot(
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 <= 8);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (8 * 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 kc = round_up_po2(kc, 4 * sizeof(int8_t));
44 int8_t* c0 = c;
45 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr < 2) {
47 c1 = c0;
48 }
49 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 c2 = c1;
52 }
53 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
54 if XNN_UNPREDICTABLE(mr < 4) {
55 c3 = c2;
56 }
57 int8_t* c4 = (int8_t*) ((uintptr_t) c3 + cm_stride);
58 if XNN_UNPREDICTABLE(mr <= 4) {
59 c4 = c3;
60 }
61 int8_t* c5 = (int8_t*) ((uintptr_t) c4 + cm_stride);
62 if XNN_UNPREDICTABLE(mr < 6) {
63 c5 = c4;
64 }
65 int8_t* c6 = (int8_t*) ((uintptr_t) c5 + cm_stride);
66 if XNN_UNPREDICTABLE(mr <= 6) {
67 c6 = c5;
68 }
69 int8_t* c7 = (int8_t*) ((uintptr_t) c6 + cm_stride);
70 if XNN_UNPREDICTABLE(mr != 8) {
71 c7 = c6;
72 }
73
74 do {
75 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
76 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
77 int32x4_t vacc1x0123 = vacc0x0123;
78 int32x4_t vacc1x4567 = vacc0x4567;
79 int32x4_t vacc2x0123 = vacc0x0123;
80 int32x4_t vacc2x4567 = vacc0x4567;
81 int32x4_t vacc3x0123 = vacc0x0123;
82 int32x4_t vacc3x4567 = vacc0x4567;
83 int32x4_t vacc4x0123 = vacc0x0123;
84 int32x4_t vacc4x4567 = vacc0x4567;
85 int32x4_t vacc5x0123 = vacc0x0123;
86 int32x4_t vacc5x4567 = vacc0x4567;
87 int32x4_t vacc6x0123 = vacc0x0123;
88 int32x4_t vacc6x4567 = vacc0x4567;
89 int32x4_t vacc7x0123 = vacc0x0123;
90 int32x4_t vacc7x4567 = vacc0x4567;
91
92 size_t p = ks;
93 do {
94 const int8_t* restrict a0 = a[0];
95 if XNN_UNPREDICTABLE(a0 != zero) {
96 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
97 }
98 const int8_t* restrict a1 = a[1];
99 if XNN_UNPREDICTABLE(a1 != zero) {
100 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
101 }
102 const int8_t* restrict a2 = a[2];
103 if XNN_UNPREDICTABLE(a2 != zero) {
104 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
105 }
106 const int8_t* restrict a3 = a[3];
107 if XNN_UNPREDICTABLE(a3 != zero) {
108 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
109 }
110 const int8_t* restrict a4 = a[4];
111 if XNN_UNPREDICTABLE(a4 != zero) {
112 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
113 }
114 const int8_t* restrict a5 = a[5];
115 if XNN_UNPREDICTABLE(a5 != zero) {
116 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
117 }
118 const int8_t* restrict a6 = a[6];
119 if XNN_UNPREDICTABLE(a6 != zero) {
120 a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
121 }
122 const int8_t* restrict a7 = a[7];
123 if XNN_UNPREDICTABLE(a7 != zero) {
124 a7 = (const int8_t*) ((uintptr_t) a7 + a_offset);
125 }
126 a += 8;
127
128 // Inner accumulation loop along the 8 columns.
129 size_t k = kc;
130 // 2x partial unrolled loop to load 8 bytes at a time.
131 while (k >= 8 * sizeof(int8_t)) {
132 // Load a 8x8 block of activations.
133 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
134 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 8;
135 const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 8;
136 const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 8;
137 const int8x8_t va4x01234567 = vld1_s8(a4); a4 += 8;
138 const int8x8_t va5x01234567 = vld1_s8(a5); a5 += 8;
139 const int8x8_t va6x01234567 = vld1_s8(a6); a6 += 8;
140 const int8x8_t va7x01234567 = vld1_s8(a7); a7 += 8;
141
142 // Load a 8x8 block of weights.
143 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
144 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
145 const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
146 const int8x16_t vb4567x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
147
148 // Multiply-accumulate: 8x8 * 8x8 --> 8x8.
149 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
150 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
151 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
152 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
153 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
154 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
155 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
156 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
157 vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0);
158 vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0);
159 vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb0123x0123, va5x01234567, 0);
160 vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb0123x4567, va5x01234567, 0);
161 vacc6x0123 = vdotq_lane_s32(vacc6x0123, vb0123x0123, va6x01234567, 0);
162 vacc6x4567 = vdotq_lane_s32(vacc6x4567, vb0123x4567, va6x01234567, 0);
163 vacc7x0123 = vdotq_lane_s32(vacc7x0123, vb0123x0123, va7x01234567, 0);
164 vacc7x4567 = vdotq_lane_s32(vacc7x4567, vb0123x4567, va7x01234567, 0);
165 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
166 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
167 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1);
168 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1);
169 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1);
170 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb4567x4567, va2x01234567, 1);
171 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1);
172 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb4567x4567, va3x01234567, 1);
173 vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb4567x0123, va4x01234567, 1);
174 vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb4567x4567, va4x01234567, 1);
175 vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb4567x0123, va5x01234567, 1);
176 vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb4567x4567, va5x01234567, 1);
177 vacc6x0123 = vdotq_lane_s32(vacc6x0123, vb4567x0123, va6x01234567, 1);
178 vacc6x4567 = vdotq_lane_s32(vacc6x4567, vb4567x4567, va6x01234567, 1);
179 vacc7x0123 = vdotq_lane_s32(vacc7x0123, vb4567x0123, va7x01234567, 1);
180 vacc7x4567 = vdotq_lane_s32(vacc7x4567, vb4567x4567, va7x01234567, 1);
181
182 k -= 8 * sizeof(int8_t);
183 }
184 // Handle up to 4 final positions of `k`
185 if XNN_UNLIKELY(k != 0) {
186 // Load a 8x4 block of activations.
187 const int8x8_t va0x01234567 = vld1_s8(a0);
188 const int8x8_t va1x01234567 = vld1_s8(a1);
189 const int8x8_t va2x01234567 = vld1_s8(a2);
190 const int8x8_t va3x01234567 = vld1_s8(a3);
191 const int8x8_t va4x01234567 = vld1_s8(a4);
192 const int8x8_t va5x01234567 = vld1_s8(a5);
193 const int8x8_t va6x01234567 = vld1_s8(a6);
194 const int8x8_t va7x01234567 = vld1_s8(a7);
195
196 // Load a 4x8 block of weights.
197 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
198 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
199
200 // Multiply-accumulate: 8x4 * 4x8 --> 8x8.
201 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
202 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
203 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
204 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
205 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
206 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
207 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
208 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
209 vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0);
210 vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0);
211 vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb0123x0123, va5x01234567, 0);
212 vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb0123x4567, va5x01234567, 0);
213 vacc6x0123 = vdotq_lane_s32(vacc6x0123, vb0123x0123, va6x01234567, 0);
214 vacc6x4567 = vdotq_lane_s32(vacc6x4567, vb0123x4567, va6x01234567, 0);
215 vacc7x0123 = vdotq_lane_s32(vacc7x0123, vb0123x0123, va7x01234567, 0);
216 vacc7x4567 = vdotq_lane_s32(vacc7x4567, vb0123x4567, va7x01234567, 0);
217 }
218 p -= 8 * sizeof(void*);
219 } while (p != 0);
220
221 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
222 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
223 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
224
225 vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
226 vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
227 vacc1x0123 = vshlq_s32(vacc1x0123, vright_pre_shift);
228 vacc1x4567 = vshlq_s32(vacc1x4567, vright_pre_shift);
229 vacc2x0123 = vshlq_s32(vacc2x0123, vright_pre_shift);
230 vacc2x4567 = vshlq_s32(vacc2x4567, vright_pre_shift);
231 vacc3x0123 = vshlq_s32(vacc3x0123, vright_pre_shift);
232 vacc3x4567 = vshlq_s32(vacc3x4567, vright_pre_shift);
233 vacc4x0123 = vshlq_s32(vacc4x0123, vright_pre_shift);
234 vacc4x4567 = vshlq_s32(vacc4x4567, vright_pre_shift);
235 vacc5x0123 = vshlq_s32(vacc5x0123, vright_pre_shift);
236 vacc5x4567 = vshlq_s32(vacc5x4567, vright_pre_shift);
237 vacc6x0123 = vshlq_s32(vacc6x0123, vright_pre_shift);
238 vacc6x4567 = vshlq_s32(vacc6x4567, vright_pre_shift);
239 vacc7x0123 = vshlq_s32(vacc7x0123, vright_pre_shift);
240 vacc7x4567 = vshlq_s32(vacc7x4567, vright_pre_shift);
241
242 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
243 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
244 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
245 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
246 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
247 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
248 vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
249 vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
250 vacc4x0123 = vqdmulhq_s32(vacc4x0123, vmultiplier);
251 vacc4x4567 = vqdmulhq_s32(vacc4x4567, vmultiplier);
252 vacc5x0123 = vqdmulhq_s32(vacc5x0123, vmultiplier);
253 vacc5x4567 = vqdmulhq_s32(vacc5x4567, vmultiplier);
254 vacc6x0123 = vqdmulhq_s32(vacc6x0123, vmultiplier);
255 vacc6x4567 = vqdmulhq_s32(vacc6x4567, vmultiplier);
256 vacc7x0123 = vqdmulhq_s32(vacc7x0123, vmultiplier);
257 vacc7x4567 = vqdmulhq_s32(vacc7x4567, vmultiplier);
258
259 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
260 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
261 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
262 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
263 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
264 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
265 vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
266 vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
267 vacc4x0123 = vrshlq_s32(vacc4x0123, vright_post_shift);
268 vacc4x4567 = vrshlq_s32(vacc4x4567, vright_post_shift);
269 vacc5x0123 = vrshlq_s32(vacc5x0123, vright_post_shift);
270 vacc5x4567 = vrshlq_s32(vacc5x4567, vright_post_shift);
271 vacc6x0123 = vrshlq_s32(vacc6x0123, vright_post_shift);
272 vacc6x4567 = vrshlq_s32(vacc6x4567, vright_post_shift);
273 vacc7x0123 = vrshlq_s32(vacc7x0123, vright_post_shift);
274 vacc7x4567 = vrshlq_s32(vacc7x4567, vright_post_shift);
275
276 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
277 #if XNN_ARCH_ARM64
278 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
279 const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
280 const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
281 const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point);
282 const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), voutput_zero_point);
283 const int16x8_t vacc5x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc5x0123), vacc5x4567), voutput_zero_point);
284 const int16x8_t vacc6x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc6x0123), vacc6x4567), voutput_zero_point);
285 const int16x8_t vacc7x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc7x0123), vacc7x4567), voutput_zero_point);
286
287 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
288 int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567);
289 int8x16_t vout4x01234567_5x01234567 = vqmovn_high_s16(vqmovn_s16(vacc4x01234567), vacc5x01234567);
290 int8x16_t vout6x01234567_7x01234567 = vqmovn_high_s16(vqmovn_s16(vacc6x01234567), vacc7x01234567);
291 #else
292 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
293 const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
294 const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
295 const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point);
296 const int16x8_t vacc4x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x4567)), voutput_zero_point);
297 const int16x8_t vacc5x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc5x0123), vqmovn_s32(vacc5x4567)), voutput_zero_point);
298 const int16x8_t vacc6x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc6x0123), vqmovn_s32(vacc6x4567)), voutput_zero_point);
299 const int16x8_t vacc7x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc7x0123), vqmovn_s32(vacc7x4567)), voutput_zero_point);
300
301 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
302 int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567));
303 int8x16_t vout4x01234567_5x01234567 = vcombine_s8(vqmovn_s16(vacc4x01234567), vqmovn_s16(vacc5x01234567));
304 int8x16_t vout6x01234567_7x01234567 = vcombine_s8(vqmovn_s16(vacc6x01234567), vqmovn_s16(vacc7x01234567));
305 #endif
306 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
307 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
308
309 vout6x01234567_7x01234567 = vmaxq_s8(vout6x01234567_7x01234567, voutput_min);
310 vout4x01234567_5x01234567 = vmaxq_s8(vout4x01234567_5x01234567, voutput_min);
311 vout2x01234567_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, voutput_min);
312 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
313
314 vout6x01234567_7x01234567 = vminq_s8(vout6x01234567_7x01234567, voutput_max);
315 vout4x01234567_5x01234567 = vminq_s8(vout4x01234567_5x01234567, voutput_max);
316 vout2x01234567_3x01234567 = vminq_s8(vout2x01234567_3x01234567, voutput_max);
317 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
318
319 if (nc >= 8) {
320 vst1_s8(c7 + 0, vget_high_s8(vout6x01234567_7x01234567));
321 vst1_s8(c6 + 0, vget_low_s8(vout6x01234567_7x01234567));
322 vst1_s8(c5 + 0, vget_high_s8(vout4x01234567_5x01234567));
323 vst1_s8(c4 + 0, vget_low_s8(vout4x01234567_5x01234567));
324 vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567));
325 vst1_s8(c2 + 0, vget_low_s8(vout2x01234567_3x01234567));
326 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
327 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
328
329 c7 = (int8_t*) ((uintptr_t) c7 + cn_stride);
330 c6 = (int8_t*) ((uintptr_t) c6 + cn_stride);
331 c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);
332 c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
333 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
334 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
335 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
336 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
337
338 a = (const int8_t**restrict) ((uintptr_t) a - ks);
339
340 nc -= 8;
341 } else {
342 if (nc & 4) {
343 vst1q_lane_u32((void*) c7, vreinterpretq_u32_s8(vout6x01234567_7x01234567), 2); c7 += 4;
344 vst1q_lane_u32((void*) c6, vreinterpretq_u32_s8(vout6x01234567_7x01234567), 0); c6 += 4;
345 vst1q_lane_u32((void*) c5, vreinterpretq_u32_s8(vout4x01234567_5x01234567), 2); c5 += 4;
346 vst1q_lane_u32((void*) c4, vreinterpretq_u32_s8(vout4x01234567_5x01234567), 0); c4 += 4;
347 vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
348 vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
349 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
350 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
351 vout6x01234567_7x01234567 = vextq_s8(vout6x01234567_7x01234567, vout6x01234567_7x01234567, 4);
352 vout4x01234567_5x01234567 = vextq_s8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 4);
353 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
354 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
355 }
356 if (nc & 2) {
357 vst1q_lane_u16((void*) c7, vreinterpretq_u16_s8(vout6x01234567_7x01234567), 4); c7 += 2;
358 vst1q_lane_u16((void*) c6, vreinterpretq_u16_s8(vout6x01234567_7x01234567), 0); c6 += 2;
359 vst1q_lane_u16((void*) c5, vreinterpretq_u16_s8(vout4x01234567_5x01234567), 4); c5 += 2;
360 vst1q_lane_u16((void*) c4, vreinterpretq_u16_s8(vout4x01234567_5x01234567), 0); c4 += 2;
361 vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
362 vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
363 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
364 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
365 vout6x01234567_7x01234567 = vextq_s8(vout6x01234567_7x01234567, vout6x01234567_7x01234567, 2);
366 vout4x01234567_5x01234567 = vextq_s8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 2);
367 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
368 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
369 }
370 if (nc & 1) {
371 vst1q_lane_s8(c7, vout6x01234567_7x01234567, 8);
372 vst1q_lane_s8(c6, vout6x01234567_7x01234567, 0);
373 vst1q_lane_s8(c5, vout4x01234567_5x01234567, 8);
374 vst1q_lane_s8(c4, vout4x01234567_5x01234567, 0);
375 vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
376 vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
377 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
378 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
379 }
380
381 nc = 0;
382 }
383 } while (nc != 0);
384 }
385