1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c8-neon-mull.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/igemm.h>
15 #include <xnnpack/math.h>
16
17
xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__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_4x8c8__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 kc = round_up_po2(kc, 8 * 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
58 do {
59 int32x4_t vacc0x0 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
60 int32x4_t vacc0x1 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
61 int32x4_t vacc0x2 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
62 int32x4_t vacc0x3 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
63 int32x4_t vacc0x4 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
64 int32x4_t vacc0x5 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
65 int32x4_t vacc0x6 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
66 int32x4_t vacc0x7 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
67 int32x4_t vacc1x0 = vacc0x0;
68 int32x4_t vacc1x1 = vacc0x1;
69 int32x4_t vacc1x2 = vacc0x2;
70 int32x4_t vacc1x3 = vacc0x3;
71 int32x4_t vacc1x4 = vacc0x4;
72 int32x4_t vacc1x5 = vacc0x5;
73 int32x4_t vacc1x6 = vacc0x6;
74 int32x4_t vacc1x7 = vacc0x7;
75 int32x4_t vacc2x0 = vacc0x0;
76 int32x4_t vacc2x1 = vacc0x1;
77 int32x4_t vacc2x2 = vacc0x2;
78 int32x4_t vacc2x3 = vacc0x3;
79 int32x4_t vacc2x4 = vacc0x4;
80 int32x4_t vacc2x5 = vacc0x5;
81 int32x4_t vacc2x6 = vacc0x6;
82 int32x4_t vacc2x7 = vacc0x7;
83 int32x4_t vacc3x0 = vacc0x0;
84 int32x4_t vacc3x1 = vacc0x1;
85 int32x4_t vacc3x2 = vacc0x2;
86 int32x4_t vacc3x3 = vacc0x3;
87 int32x4_t vacc3x4 = vacc0x4;
88 int32x4_t vacc3x5 = vacc0x5;
89 int32x4_t vacc3x6 = vacc0x6;
90 int32x4_t vacc3x7 = vacc0x7;
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 a += 4;
111
112 size_t k = kc;
113
114 // Handle 8 bytes at a time using MUL.
115 while (k != 0) {
116 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
117 const int8x8_t va1 = vld1_s8(a1); a1 += 8;
118 const int8x8_t va2 = vld1_s8(a2); a2 += 8;
119 const int8x8_t va3 = vld1_s8(a3); a3 += 8;
120
121 const int8x8_t vb0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
122 const int16x8_t vprod0x0 = vmull_s8(vb0, va0);
123 const int16x8_t vprod1x0 = vmull_s8(vb0, va1);
124 const int16x8_t vprod2x0 = vmull_s8(vb0, va2);
125 const int16x8_t vprod3x0 = vmull_s8(vb0, va3);
126 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
127 vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
128 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0);
129 vacc3x0 = vpadalq_s16(vacc3x0, vprod3x0);
130 const int8x8_t vb1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
131 const int16x8_t vprod0x1 = vmull_s8(vb1, va0);
132 const int16x8_t vprod1x1 = vmull_s8(vb1, va1);
133 const int16x8_t vprod2x1 = vmull_s8(vb1, va2);
134 const int16x8_t vprod3x1 = vmull_s8(vb1, va3);
135 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
136 vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
137 vacc2x1 = vpadalq_s16(vacc2x1, vprod2x1);
138 vacc3x1 = vpadalq_s16(vacc3x1, vprod3x1);
139 const int8x8_t vb2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
140 const int16x8_t vprod0x2 = vmull_s8(vb2, va0);
141 const int16x8_t vprod1x2 = vmull_s8(vb2, va1);
142 const int16x8_t vprod2x2 = vmull_s8(vb2, va2);
143 const int16x8_t vprod3x2 = vmull_s8(vb2, va3);
144 vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
145 vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
146 vacc2x2 = vpadalq_s16(vacc2x2, vprod2x2);
147 vacc3x2 = vpadalq_s16(vacc3x2, vprod3x2);
148 const int8x8_t vb3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
149 const int16x8_t vprod0x3 = vmull_s8(vb3, va0);
150 const int16x8_t vprod1x3 = vmull_s8(vb3, va1);
151 const int16x8_t vprod2x3 = vmull_s8(vb3, va2);
152 const int16x8_t vprod3x3 = vmull_s8(vb3, va3);
153 vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
154 vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
155 vacc2x3 = vpadalq_s16(vacc2x3, vprod2x3);
156 vacc3x3 = vpadalq_s16(vacc3x3, vprod3x3);
157 const int8x8_t vb4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
158 const int16x8_t vprod0x4 = vmull_s8(vb4, va0);
159 const int16x8_t vprod1x4 = vmull_s8(vb4, va1);
160 const int16x8_t vprod2x4 = vmull_s8(vb4, va2);
161 const int16x8_t vprod3x4 = vmull_s8(vb4, va3);
162 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
163 vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
164 vacc2x4 = vpadalq_s16(vacc2x4, vprod2x4);
165 vacc3x4 = vpadalq_s16(vacc3x4, vprod3x4);
166 const int8x8_t vb5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
167 const int16x8_t vprod0x5 = vmull_s8(vb5, va0);
168 const int16x8_t vprod1x5 = vmull_s8(vb5, va1);
169 const int16x8_t vprod2x5 = vmull_s8(vb5, va2);
170 const int16x8_t vprod3x5 = vmull_s8(vb5, va3);
171 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
172 vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
173 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5);
174 vacc3x5 = vpadalq_s16(vacc3x5, vprod3x5);
175 const int8x8_t vb6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
176 const int16x8_t vprod0x6 = vmull_s8(vb6, va0);
177 const int16x8_t vprod1x6 = vmull_s8(vb6, va1);
178 const int16x8_t vprod2x6 = vmull_s8(vb6, va2);
179 const int16x8_t vprod3x6 = vmull_s8(vb6, va3);
180 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
181 vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
182 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6);
183 vacc3x6 = vpadalq_s16(vacc3x6, vprod3x6);
184 const int8x8_t vb7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
185 const int16x8_t vprod0x7 = vmull_s8(vb7, va0);
186 const int16x8_t vprod1x7 = vmull_s8(vb7, va1);
187 const int16x8_t vprod2x7 = vmull_s8(vb7, va2);
188 const int16x8_t vprod3x7 = vmull_s8(vb7, va3);
189 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
190 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
191 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7);
192 vacc3x7 = vpadalq_s16(vacc3x7, vprod3x7);
193
194 k -= 8 * sizeof(int8_t);
195 }
196
197 p -= 4 * sizeof(void*);
198 } while (p != 0);
199
200 #if XNN_ARCH_ARM64
201 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1);
202 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3);
203 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5);
204 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7);
205 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1);
206 const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3);
207 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5);
208 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7);
209 const int32x4_t vsum2x01 = vpaddq_s32(vacc2x0, vacc2x1);
210 const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3);
211 const int32x4_t vsum2x45 = vpaddq_s32(vacc2x4, vacc2x5);
212 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7);
213 const int32x4_t vsum3x01 = vpaddq_s32(vacc3x0, vacc3x1);
214 const int32x4_t vsum3x23 = vpaddq_s32(vacc3x2, vacc3x3);
215 const int32x4_t vsum3x45 = vpaddq_s32(vacc3x4, vacc3x5);
216 const int32x4_t vsum3x67 = vpaddq_s32(vacc3x6, vacc3x7);
217
218 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23);
219 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67);
220 int32x4_t vacc1x0123 = vpaddq_s32(vsum1x01, vsum1x23);
221 int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67);
222 int32x4_t vacc2x0123 = vpaddq_s32(vsum2x01, vsum2x23);
223 int32x4_t vacc2x4567 = vpaddq_s32(vsum2x45, vsum2x67);
224 int32x4_t vacc3x0123 = vpaddq_s32(vsum3x01, vsum3x23);
225 int32x4_t vacc3x4567 = vpaddq_s32(vsum3x45, vsum3x67);
226 #else
227 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0));
228 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1));
229 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2));
230 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3));
231 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1);
232 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3);
233 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 );
234 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4));
235 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5));
236 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6));
237 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7));
238 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5);
239 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7);
240 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 );
241 const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0));
242 const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1));
243 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2));
244 const int32x2_t vpsum1x3 = vadd_s32(vget_low_s32(vacc1x3), vget_high_s32(vacc1x3));
245 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1);
246 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3);
247 int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23 );
248 const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4));
249 const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(vacc1x5));
250 const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6));
251 const int32x2_t vpsum1x7 = vadd_s32(vget_low_s32(vacc1x7), vget_high_s32(vacc1x7));
252 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5);
253 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7);
254 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 );
255 const int32x2_t vpsum2x0 = vadd_s32(vget_low_s32(vacc2x0), vget_high_s32(vacc2x0));
256 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1));
257 const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2));
258 const int32x2_t vpsum2x3 = vadd_s32(vget_low_s32(vacc2x3), vget_high_s32(vacc2x3));
259 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1);
260 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3);
261 int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23 );
262 const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4));
263 const int32x2_t vpsum2x5 = vadd_s32(vget_low_s32(vacc2x5), vget_high_s32(vacc2x5));
264 const int32x2_t vpsum2x6 = vadd_s32(vget_low_s32(vacc2x6), vget_high_s32(vacc2x6));
265 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7));
266 const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5);
267 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7);
268 int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67 );
269 const int32x2_t vpsum3x0 = vadd_s32(vget_low_s32(vacc3x0), vget_high_s32(vacc3x0));
270 const int32x2_t vpsum3x1 = vadd_s32(vget_low_s32(vacc3x1), vget_high_s32(vacc3x1));
271 const int32x2_t vpsum3x2 = vadd_s32(vget_low_s32(vacc3x2), vget_high_s32(vacc3x2));
272 const int32x2_t vpsum3x3 = vadd_s32(vget_low_s32(vacc3x3), vget_high_s32(vacc3x3));
273 const int32x2_t vsum3x01 = vpadd_s32(vpsum3x0, vpsum3x1);
274 const int32x2_t vsum3x23 = vpadd_s32(vpsum3x2, vpsum3x3);
275 int32x4_t vacc3x0123 = vcombine_s32(vsum3x01, vsum3x23 );
276 const int32x2_t vpsum3x4 = vadd_s32(vget_low_s32(vacc3x4), vget_high_s32(vacc3x4));
277 const int32x2_t vpsum3x5 = vadd_s32(vget_low_s32(vacc3x5), vget_high_s32(vacc3x5));
278 const int32x2_t vpsum3x6 = vadd_s32(vget_low_s32(vacc3x6), vget_high_s32(vacc3x6));
279 const int32x2_t vpsum3x7 = vadd_s32(vget_low_s32(vacc3x7), vget_high_s32(vacc3x7));
280 const int32x2_t vsum3x45 = vpadd_s32(vpsum3x4, vpsum3x5);
281 const int32x2_t vsum3x67 = vpadd_s32(vpsum3x6, vpsum3x7);
282 int32x4_t vacc3x4567 = vcombine_s32(vsum3x45, vsum3x67 );
283 #endif
284
285 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
286 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
287 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
288
289 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
290 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
291 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
292 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
293 vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
294 vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
295 vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
296 vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
297
298 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
299 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
300 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
301 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
302 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
303 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
304 vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
305 vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
306
307 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
308 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
309 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
310 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
311 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
312 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
313 vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
314 vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
315
316 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
317 #if XNN_ARCH_ARM64
318 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
319 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
320 int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
321 int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
322
323 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
324 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
325 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
326 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
327
328 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
329 int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567);
330 #else
331 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
332 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
333 int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
334 int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
335
336 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
337 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
338 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
339 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
340
341 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
342 int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567));
343 #endif
344
345 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
346 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
347 vout2x01234567_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, voutput_min);
348
349 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
350 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
351 vout2x01234567_3x01234567 = vminq_s8(vout2x01234567_3x01234567, voutput_max);
352
353 if (nc >= 8) {
354 vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567));
355 vst1_s8(c2 + 0, vget_low_s8(vout2x01234567_3x01234567));
356 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
357 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
358
359 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
360 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
361 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
362 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
363
364 a = (const int8_t**restrict) ((uintptr_t) a - ks);
365
366 nc -= 8;
367 } else {
368 if (nc & 4) {
369 vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
370 vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
371 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
372 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
373 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
374 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
375 }
376 if (nc & 2) {
377 vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
378 vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
379 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
380 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
381 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
382 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
383 }
384 if (nc & 1) {
385 vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
386 vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
387 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
388 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
389 }
390
391 nc = 0;
392 }
393 } while (nc != 0);
394 }
395