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_4x16c2s4__neon_mull(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_4x16c2s4__neon_mull(
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 <= 4);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (4 * 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 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
53 if XNN_UNPREDICTABLE(mr != 4) {
54 c3 = c2;
55 }
56
57 kc = round_up_po2(kc, 8 * sizeof(int8_t));
58 do {
59 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
60 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
61 int32x4_t vacc0x89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
62 int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
63 int32x4_t vacc1x0123 = vacc0x0123;
64 int32x4_t vacc1x4567 = vacc0x4567;
65 int32x4_t vacc1x89AB = vacc0x89AB;
66 int32x4_t vacc1xCDEF = vacc0xCDEF;
67 int32x4_t vacc2x0123 = vacc0x0123;
68 int32x4_t vacc2x4567 = vacc0x4567;
69 int32x4_t vacc2x89AB = vacc0x89AB;
70 int32x4_t vacc2xCDEF = vacc0xCDEF;
71 int32x4_t vacc3x0123 = vacc0x0123;
72 int32x4_t vacc3x4567 = vacc0x4567;
73 int32x4_t vacc3x89AB = vacc0x89AB;
74 int32x4_t vacc3xCDEF = vacc0xCDEF;
75
76 size_t p = ks;
77 do {
78 const int8_t* restrict a0 = a[0];
79 if XNN_UNPREDICTABLE(a0 != zero) {
80 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
81 }
82 const int8_t* restrict a1 = a[1];
83 if XNN_UNPREDICTABLE(a1 != zero) {
84 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
85 }
86 const int8_t* restrict a2 = a[2];
87 if XNN_UNPREDICTABLE(a2 != zero) {
88 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
89 }
90 const int8_t* restrict a3 = a[3];
91 if XNN_UNPREDICTABLE(a3 != zero) {
92 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
93 }
94 a += 4;
95
96 size_t k = kc;
97 do {
98 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
99 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
100 int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
101 int8x8_t va3x0 = vld1_s8(a3); a3 += 8;
102
103 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
104 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
105 const int8x8_t vb89ABc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
106 const int8x8_t vbCDEFc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
107 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
108 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
109 const int8x8_t vb89ABc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
110 const int8x8_t vbCDEFc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
111 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
112 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
113 const int8x8_t vb89ABc2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
114 const int8x8_t vbCDEFc2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
115 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
116 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
117 const int8x8_t vb89ABc3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
118 const int8x8_t vbCDEFc3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
119
120 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
121 int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1x0);
122 int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0x0, va2x0);
123 int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0x0, va3x0);
124 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
125 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
126 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
127 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c0);
128 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
129 int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1x0);
130 int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2x0);
131 int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0x0, va3x0);
132 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
133 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
134 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
135 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0);
136 int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, va0x0);
137 int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1x0);
138 int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0x0, va2x0);
139 int16x8_t vprod3x89ABc0 = vmull_s8(vb89ABc0x0, va3x0);
140 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
141 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
142 vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
143 vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc0);
144 int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0x0, va0x0);
145 int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0x0, va1x0);
146 int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0x0, va2x0);
147 int16x8_t vprod3xCDEFc0 = vmull_s8(vbCDEFc0x0, va3x0);
148 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
149 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
150 vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
151 vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc0);
152 va0x0 = vext_s8(va0x0, va0x0, 2);
153 va1x0 = vext_s8(va1x0, va1x0, 2);
154 va2x0 = vext_s8(va2x0, va2x0, 2);
155 va3x0 = vext_s8(va3x0, va3x0, 2);
156 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
157 int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1x0);
158 int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2x0);
159 int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1x0, va3x0);
160 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
161 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
162 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
163 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1);
164 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
165 int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1x0);
166 int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1x0, va2x0);
167 int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1x0, va3x0);
168 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
169 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
170 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
171 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c1);
172 int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0x0);
173 int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1x0, va1x0);
174 int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1x0, va2x0);
175 int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1x0, va3x0);
176 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
177 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
178 vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
179 vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc1);
180 int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0x0);
181 int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1x0);
182 int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1x0, va2x0);
183 int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1x0, va3x0);
184 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
185 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
186 vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
187 vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc1);
188 va0x0 = vext_s8(va0x0, va0x0, 2);
189 va1x0 = vext_s8(va1x0, va1x0, 2);
190 va2x0 = vext_s8(va2x0, va2x0, 2);
191 va3x0 = vext_s8(va3x0, va3x0, 2);
192 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
193 int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1x0);
194 int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2x0, va2x0);
195 int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2x0, va3x0);
196 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
197 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
198 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
199 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c2);
200 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
201 int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1x0);
202 int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2x0, va2x0);
203 int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2x0, va3x0);
204 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
205 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
206 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
207 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2);
208 int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2x0, va0x0);
209 int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2x0, va1x0);
210 int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2x0, va2x0);
211 int16x8_t vprod3x89ABc2 = vmull_s8(vb89ABc2x0, va3x0);
212 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
213 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
214 vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
215 vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc2);
216 int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0x0);
217 int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2x0, va1x0);
218 int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2x0, va2x0);
219 int16x8_t vprod3xCDEFc2 = vmull_s8(vbCDEFc2x0, va3x0);
220 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
221 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
222 vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
223 vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc2);
224 va0x0 = vext_s8(va0x0, va0x0, 2);
225 va1x0 = vext_s8(va1x0, va1x0, 2);
226 va2x0 = vext_s8(va2x0, va2x0, 2);
227 va3x0 = vext_s8(va3x0, va3x0, 2);
228 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
229 int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1x0);
230 int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3x0, va2x0);
231 int16x8_t vprod3x0123c3 = vmull_s8(vb0123c3x0, va3x0);
232 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
233 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
234 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
235 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c3);
236 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
237 int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1x0);
238 int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3x0, va2x0);
239 int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3x0);
240 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
241 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
242 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
243 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3);
244 int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3x0, va0x0);
245 int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3x0, va1x0);
246 int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, va2x0);
247 int16x8_t vprod3x89ABc3 = vmull_s8(vb89ABc3x0, va3x0);
248 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
249 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
250 vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc3);
251 vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc3);
252 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0x0);
253 int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3x0, va1x0);
254 int16x8_t vprod2xCDEFc3 = vmull_s8(vbCDEFc3x0, va2x0);
255 int16x8_t vprod3xCDEFc3 = vmull_s8(vbCDEFc3x0, va3x0);
256 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
257 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
258 vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc3);
259 vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc3);
260
261 k -= 8 * sizeof(int8_t);
262 } while (k != 0);
263
264 p -= 4 * sizeof(void*);
265 } while (p != 0);
266
267 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
268 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
269 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
270
271 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
272 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
273 vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
274 vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
275 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
276 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
277 vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
278 vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
279 vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
280 vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
281 vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
282 vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
283 vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
284 vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
285 vacc3x89AB = vqshlq_s32(vacc3x89AB, vright_pre_shift);
286 vacc3xCDEF = vqshlq_s32(vacc3xCDEF, vright_pre_shift);
287
288 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
289 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
290 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
291 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
292 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
293 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
294 vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
295 vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
296 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
297 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
298 vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
299 vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
300 vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
301 vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
302 vacc3x89AB = vqdmulhq_s32(vacc3x89AB, vmultiplier);
303 vacc3xCDEF = vqdmulhq_s32(vacc3xCDEF, vmultiplier);
304
305 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
306 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
307 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
308 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
309 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
310 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
311 vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
312 vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
313 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
314 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
315 vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
316 vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
317 vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
318 vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
319 vacc3x89AB = vrshlq_s32(vacc3x89AB, vright_post_shift);
320 vacc3xCDEF = vrshlq_s32(vacc3xCDEF, vright_post_shift);
321
322 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
323 #if XNN_ARCH_ARM64
324 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
325 int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
326 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
327 int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
328 int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
329 int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
330 int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
331 int16x8_t vacc3x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF);
332
333 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
334 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
335 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
336 vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
337 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
338 vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
339 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
340 vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
341
342 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
343 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
344 int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
345 int8x16_t vout3x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc3x01234567), vacc3x89ABCDEF);
346 #else
347 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
348 int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
349 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
350 int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
351 int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
352 int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
353 int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
354 int16x8_t vacc3x89ABCDEF = vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF));
355
356 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
357 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
358 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
359 vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
360 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
361 vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
362 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
363 vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
364
365 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
366 int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
367 int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
368 int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCDEF));
369 #endif
370
371 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
372 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
373 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
374 vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
375 vout3x0123456789ABCDEF = vmaxq_s8(vout3x0123456789ABCDEF, voutput_min);
376
377 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
378 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
379 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
380 vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
381 vout3x0123456789ABCDEF = vminq_s8(vout3x0123456789ABCDEF, voutput_max);
382
383 if (nc >= 16) {
384 vst1q_s8(c3 + 0, vout3x0123456789ABCDEF);
385 vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
386 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
387 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
388
389 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
390 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
391 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
392 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
393
394 a = (const int8_t**restrict) ((uintptr_t) a - ks);
395
396 nc -= 16;
397 } else {
398 int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8(vout3x0123456789ABCDEF));
399 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
400 if (nc & 8) {
401 vst1_s8(c3, vget_high_s8(vout2x01234567_3x01234567)); c3 += 8;
402 vst1_s8(c2, vget_low_s8(vout2x01234567_3x01234567)); c2 += 8;
403 vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
404 vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
405 vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0123456789ABCDEF));
406 vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
407 }
408 if (nc & 4) {
409 vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
410 vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
411 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
412 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
413 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
414 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
415 }
416 if (nc & 2) {
417 vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
418 vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
419 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
420 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
421 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
422 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
423 }
424 if (nc & 1) {
425 vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
426 vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
427 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
428 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
429 }
430
431 nc = 0;
432 }
433 } while (nc != 0);
434 }
435