1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const int8_t** restrict a,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 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 vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
60 int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
61 int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
62 int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
63 int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
64 int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
65 int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
66 int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
67 int32x4_t vacc1x01 = vacc0x01;
68 int32x4_t vacc1x23 = vacc0x23;
69 int32x4_t vacc1x45 = vacc0x45;
70 int32x4_t vacc1x67 = vacc0x67;
71 int32x4_t vacc1x89 = vacc0x89;
72 int32x4_t vacc1xAB = vacc0xAB;
73 int32x4_t vacc1xCD = vacc0xCD;
74 int32x4_t vacc1xEF = vacc0xEF;
75 int32x4_t vacc2x01 = vacc0x01;
76 int32x4_t vacc2x23 = vacc0x23;
77 int32x4_t vacc2x45 = vacc0x45;
78 int32x4_t vacc2x67 = vacc0x67;
79 int32x4_t vacc2x89 = vacc0x89;
80 int32x4_t vacc2xAB = vacc0xAB;
81 int32x4_t vacc2xCD = vacc0xCD;
82 int32x4_t vacc2xEF = vacc0xEF;
83 int32x4_t vacc3x01 = vacc0x01;
84 int32x4_t vacc3x23 = vacc0x23;
85 int32x4_t vacc3x45 = vacc0x45;
86 int32x4_t vacc3x67 = vacc0x67;
87 int32x4_t vacc3x89 = vacc0x89;
88 int32x4_t vacc3xAB = vacc0xAB;
89 int32x4_t vacc3xCD = vacc0xCD;
90 int32x4_t vacc3xEF = vacc0xEF;
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 while (k >= 16 * sizeof(int8_t)) {
114 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
115 int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
116 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
117 int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
118 int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
119 int8x8_t va2x1 = vld1_s8(a2); a2 += 8;
120 int8x8_t va3x0 = vld1_s8(a3); a3 += 8;
121 int8x8_t va3x1 = vld1_s8(a3); a3 += 8;
122
123 const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
124 const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
125 const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
126 const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
127 const int8x8_t vb89c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
128 const int8x8_t vbABc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
129 const int8x8_t vbCDc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
130 const int8x8_t vbEFc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
131 const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
132 const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
133 const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
134 const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
135 const int8x8_t vb89c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
136 const int8x8_t vbABc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
137 const int8x8_t vbCDc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
138 const int8x8_t vbEFc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
139
140 int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
141 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1x0);
142 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2x0);
143 int16x8_t vprod3x01c0 = vmull_s8(vb01c0x0, va3x0);
144 const int8x8_t vb01c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
145 vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0x1);
146 vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1x1);
147 vprod2x01c0 = vmlal_s8(vprod2x01c0, vb01c0x1, va2x1);
148 vprod3x01c0 = vmlal_s8(vprod3x01c0, vb01c0x1, va3x1);
149 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
150 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
151 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
152 vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0);
153 int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
154 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1x0);
155 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2x0);
156 int16x8_t vprod3x23c0 = vmull_s8(vb23c0x0, va3x0);
157 const int8x8_t vb23c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
158 vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0x1);
159 vprod1x23c0 = vmlal_s8(vprod1x23c0, vb23c0x1, va1x1);
160 vprod2x23c0 = vmlal_s8(vprod2x23c0, vb23c0x1, va2x1);
161 vprod3x23c0 = vmlal_s8(vprod3x23c0, vb23c0x1, va3x1);
162 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
163 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
164 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
165 vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0);
166 int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
167 int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1x0);
168 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2x0);
169 int16x8_t vprod3x45c0 = vmull_s8(vb45c0x0, va3x0);
170 const int8x8_t vb45c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
171 vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0x1);
172 vprod1x45c0 = vmlal_s8(vprod1x45c0, vb45c0x1, va1x1);
173 vprod2x45c0 = vmlal_s8(vprod2x45c0, vb45c0x1, va2x1);
174 vprod3x45c0 = vmlal_s8(vprod3x45c0, vb45c0x1, va3x1);
175 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
176 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
177 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
178 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0);
179 int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
180 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1x0);
181 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2x0);
182 int16x8_t vprod3x67c0 = vmull_s8(vb67c0x0, va3x0);
183 const int8x8_t vb67c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
184 vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0x1);
185 vprod1x67c0 = vmlal_s8(vprod1x67c0, vb67c0x1, va1x1);
186 vprod2x67c0 = vmlal_s8(vprod2x67c0, vb67c0x1, va2x1);
187 vprod3x67c0 = vmlal_s8(vprod3x67c0, vb67c0x1, va3x1);
188 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
189 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
190 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
191 vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0);
192 int16x8_t vprod0x89c0 = vmull_s8(vb89c0x0, va0x0);
193 int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1x0);
194 int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2x0);
195 int16x8_t vprod3x89c0 = vmull_s8(vb89c0x0, va3x0);
196 const int8x8_t vb89c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
197 vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0x1);
198 vprod1x89c0 = vmlal_s8(vprod1x89c0, vb89c0x1, va1x1);
199 vprod2x89c0 = vmlal_s8(vprod2x89c0, vb89c0x1, va2x1);
200 vprod3x89c0 = vmlal_s8(vprod3x89c0, vb89c0x1, va3x1);
201 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
202 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
203 vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c0);
204 vacc3x89 = vpadalq_s16(vacc3x89, vprod3x89c0);
205 int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0x0);
206 int16x8_t vprod1xABc0 = vmull_s8(vbABc0x0, va1x0);
207 int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2x0);
208 int16x8_t vprod3xABc0 = vmull_s8(vbABc0x0, va3x0);
209 const int8x8_t vbABc0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
210 vprod0xABc0 = vmlal_s8(vprod0xABc0, vbABc0x1, va0x1);
211 vprod1xABc0 = vmlal_s8(vprod1xABc0, vbABc0x1, va1x1);
212 vprod2xABc0 = vmlal_s8(vprod2xABc0, vbABc0x1, va2x1);
213 vprod3xABc0 = vmlal_s8(vprod3xABc0, vbABc0x1, va3x1);
214 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
215 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
216 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0);
217 vacc3xAB = vpadalq_s16(vacc3xAB, vprod3xABc0);
218 int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0x0);
219 int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0x0, va1x0);
220 int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2x0);
221 int16x8_t vprod3xCDc0 = vmull_s8(vbCDc0x0, va3x0);
222 const int8x8_t vbCDc0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
223 vprod0xCDc0 = vmlal_s8(vprod0xCDc0, vbCDc0x1, va0x1);
224 vprod1xCDc0 = vmlal_s8(vprod1xCDc0, vbCDc0x1, va1x1);
225 vprod2xCDc0 = vmlal_s8(vprod2xCDc0, vbCDc0x1, va2x1);
226 vprod3xCDc0 = vmlal_s8(vprod3xCDc0, vbCDc0x1, va3x1);
227 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
228 vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
229 vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc0);
230 vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc0);
231 int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0x0, va0x0);
232 int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0x0, va1x0);
233 int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0x0, va2x0);
234 int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0x0, va3x0);
235 const int8x8_t vbEFc0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
236 vprod0xEFc0 = vmlal_s8(vprod0xEFc0, vbEFc0x1, va0x1);
237 vprod1xEFc0 = vmlal_s8(vprod1xEFc0, vbEFc0x1, va1x1);
238 vprod2xEFc0 = vmlal_s8(vprod2xEFc0, vbEFc0x1, va2x1);
239 vprod3xEFc0 = vmlal_s8(vprod3xEFc0, vbEFc0x1, va3x1);
240 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
241 vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
242 vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc0);
243 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0);
244 va0x0 = vext_s8(va0x0, va0x0, 4);
245 va0x1 = vext_s8(va0x1, va0x1, 4);
246 va1x0 = vext_s8(va1x0, va1x0, 4);
247 va1x1 = vext_s8(va1x1, va1x1, 4);
248 va2x0 = vext_s8(va2x0, va2x0, 4);
249 va2x1 = vext_s8(va2x1, va2x1, 4);
250 va3x0 = vext_s8(va3x0, va3x0, 4);
251 va3x1 = vext_s8(va3x1, va3x1, 4);
252 int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
253 int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1x0);
254 int16x8_t vprod2x01c1 = vmull_s8(vb01c1x0, va2x0);
255 int16x8_t vprod3x01c1 = vmull_s8(vb01c1x0, va3x0);
256 const int8x8_t vb01c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
257 vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0x1);
258 vprod1x01c1 = vmlal_s8(vprod1x01c1, vb01c1x1, va1x1);
259 vprod2x01c1 = vmlal_s8(vprod2x01c1, vb01c1x1, va2x1);
260 vprod3x01c1 = vmlal_s8(vprod3x01c1, vb01c1x1, va3x1);
261 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
262 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
263 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
264 vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c1);
265 int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
266 int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1x0);
267 int16x8_t vprod2x23c1 = vmull_s8(vb23c1x0, va2x0);
268 int16x8_t vprod3x23c1 = vmull_s8(vb23c1x0, va3x0);
269 const int8x8_t vb23c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
270 vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0x1);
271 vprod1x23c1 = vmlal_s8(vprod1x23c1, vb23c1x1, va1x1);
272 vprod2x23c1 = vmlal_s8(vprod2x23c1, vb23c1x1, va2x1);
273 vprod3x23c1 = vmlal_s8(vprod3x23c1, vb23c1x1, va3x1);
274 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
275 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
276 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
277 vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c1);
278 int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
279 int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1x0);
280 int16x8_t vprod2x45c1 = vmull_s8(vb45c1x0, va2x0);
281 int16x8_t vprod3x45c1 = vmull_s8(vb45c1x0, va3x0);
282 const int8x8_t vb45c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
283 vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0x1);
284 vprod1x45c1 = vmlal_s8(vprod1x45c1, vb45c1x1, va1x1);
285 vprod2x45c1 = vmlal_s8(vprod2x45c1, vb45c1x1, va2x1);
286 vprod3x45c1 = vmlal_s8(vprod3x45c1, vb45c1x1, va3x1);
287 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
288 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
289 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
290 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c1);
291 int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
292 int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1x0);
293 int16x8_t vprod2x67c1 = vmull_s8(vb67c1x0, va2x0);
294 int16x8_t vprod3x67c1 = vmull_s8(vb67c1x0, va3x0);
295 const int8x8_t vb67c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
296 vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0x1);
297 vprod1x67c1 = vmlal_s8(vprod1x67c1, vb67c1x1, va1x1);
298 vprod2x67c1 = vmlal_s8(vprod2x67c1, vb67c1x1, va2x1);
299 vprod3x67c1 = vmlal_s8(vprod3x67c1, vb67c1x1, va3x1);
300 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
301 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
302 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
303 vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c1);
304 int16x8_t vprod0x89c1 = vmull_s8(vb89c1x0, va0x0);
305 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0);
306 int16x8_t vprod2x89c1 = vmull_s8(vb89c1x0, va2x0);
307 int16x8_t vprod3x89c1 = vmull_s8(vb89c1x0, va3x0);
308 const int8x8_t vb89c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
309 vprod0x89c1 = vmlal_s8(vprod0x89c1, vb89c1x1, va0x1);
310 vprod1x89c1 = vmlal_s8(vprod1x89c1, vb89c1x1, va1x1);
311 vprod2x89c1 = vmlal_s8(vprod2x89c1, vb89c1x1, va2x1);
312 vprod3x89c1 = vmlal_s8(vprod3x89c1, vb89c1x1, va3x1);
313 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
314 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
315 vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c1);
316 vacc3x89 = vpadalq_s16(vacc3x89, vprod3x89c1);
317 int16x8_t vprod0xABc1 = vmull_s8(vbABc1x0, va0x0);
318 int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1x0);
319 int16x8_t vprod2xABc1 = vmull_s8(vbABc1x0, va2x0);
320 int16x8_t vprod3xABc1 = vmull_s8(vbABc1x0, va3x0);
321 const int8x8_t vbABc1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
322 vprod0xABc1 = vmlal_s8(vprod0xABc1, vbABc1x1, va0x1);
323 vprod1xABc1 = vmlal_s8(vprod1xABc1, vbABc1x1, va1x1);
324 vprod2xABc1 = vmlal_s8(vprod2xABc1, vbABc1x1, va2x1);
325 vprod3xABc1 = vmlal_s8(vprod3xABc1, vbABc1x1, va3x1);
326 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
327 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
328 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc1);
329 vacc3xAB = vpadalq_s16(vacc3xAB, vprod3xABc1);
330 int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1x0, va0x0);
331 int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1x0);
332 int16x8_t vprod2xCDc1 = vmull_s8(vbCDc1x0, va2x0);
333 int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1x0, va3x0);
334 const int8x8_t vbCDc1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
335 vprod0xCDc1 = vmlal_s8(vprod0xCDc1, vbCDc1x1, va0x1);
336 vprod1xCDc1 = vmlal_s8(vprod1xCDc1, vbCDc1x1, va1x1);
337 vprod2xCDc1 = vmlal_s8(vprod2xCDc1, vbCDc1x1, va2x1);
338 vprod3xCDc1 = vmlal_s8(vprod3xCDc1, vbCDc1x1, va3x1);
339 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
340 vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
341 vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc1);
342 vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1);
343 int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1x0, va0x0);
344 int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1x0, va1x0);
345 int16x8_t vprod2xEFc1 = vmull_s8(vbEFc1x0, va2x0);
346 int16x8_t vprod3xEFc1 = vmull_s8(vbEFc1x0, va3x0);
347 const int8x8_t vbEFc1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
348 vprod0xEFc1 = vmlal_s8(vprod0xEFc1, vbEFc1x1, va0x1);
349 vprod1xEFc1 = vmlal_s8(vprod1xEFc1, vbEFc1x1, va1x1);
350 vprod2xEFc1 = vmlal_s8(vprod2xEFc1, vbEFc1x1, va2x1);
351 vprod3xEFc1 = vmlal_s8(vprod3xEFc1, vbEFc1x1, va3x1);
352 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
353 vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
354 vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc1);
355 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc1);
356
357 k -= 16 * sizeof(int8_t);
358 }
359 if (k != 0) {
360 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
361 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
362 int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
363 int8x8_t va3x0 = vld1_s8(a3); a3 += 8;
364
365 const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
366 const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
367 const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
368 const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
369 const int8x8_t vb89c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
370 const int8x8_t vbABc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
371 const int8x8_t vbCDc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
372 const int8x8_t vbEFc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
373 const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
374 const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
375 const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
376 const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
377 const int8x8_t vb89c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
378 const int8x8_t vbABc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
379 const int8x8_t vbCDc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
380 const int8x8_t vbEFc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
381
382 int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
383 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1x0);
384 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2x0);
385 int16x8_t vprod3x01c0 = vmull_s8(vb01c0x0, va3x0);
386 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
387 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
388 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
389 vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0);
390 int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
391 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1x0);
392 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2x0);
393 int16x8_t vprod3x23c0 = vmull_s8(vb23c0x0, va3x0);
394 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
395 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
396 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
397 vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0);
398 int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
399 int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1x0);
400 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2x0);
401 int16x8_t vprod3x45c0 = vmull_s8(vb45c0x0, va3x0);
402 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
403 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
404 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
405 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0);
406 int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
407 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1x0);
408 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2x0);
409 int16x8_t vprod3x67c0 = vmull_s8(vb67c0x0, va3x0);
410 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
411 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
412 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
413 vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0);
414 int16x8_t vprod0x89c0 = vmull_s8(vb89c0x0, va0x0);
415 int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1x0);
416 int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2x0);
417 int16x8_t vprod3x89c0 = vmull_s8(vb89c0x0, va3x0);
418 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
419 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
420 vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c0);
421 vacc3x89 = vpadalq_s16(vacc3x89, vprod3x89c0);
422 int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0x0);
423 int16x8_t vprod1xABc0 = vmull_s8(vbABc0x0, va1x0);
424 int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2x0);
425 int16x8_t vprod3xABc0 = vmull_s8(vbABc0x0, va3x0);
426 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
427 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
428 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0);
429 vacc3xAB = vpadalq_s16(vacc3xAB, vprod3xABc0);
430 int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0x0);
431 int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0x0, va1x0);
432 int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2x0);
433 int16x8_t vprod3xCDc0 = vmull_s8(vbCDc0x0, va3x0);
434 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
435 vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
436 vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc0);
437 vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc0);
438 int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0x0, va0x0);
439 int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0x0, va1x0);
440 int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0x0, va2x0);
441 int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0x0, va3x0);
442 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
443 vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
444 vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc0);
445 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0);
446 va0x0 = vext_s8(va0x0, va0x0, 4);
447 va1x0 = vext_s8(va1x0, va1x0, 4);
448 va2x0 = vext_s8(va2x0, va2x0, 4);
449 va3x0 = vext_s8(va3x0, va3x0, 4);
450 int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
451 int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1x0);
452 int16x8_t vprod2x01c1 = vmull_s8(vb01c1x0, va2x0);
453 int16x8_t vprod3x01c1 = vmull_s8(vb01c1x0, va3x0);
454 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
455 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
456 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
457 vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c1);
458 int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
459 int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1x0);
460 int16x8_t vprod2x23c1 = vmull_s8(vb23c1x0, va2x0);
461 int16x8_t vprod3x23c1 = vmull_s8(vb23c1x0, va3x0);
462 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
463 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
464 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
465 vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c1);
466 int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
467 int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1x0);
468 int16x8_t vprod2x45c1 = vmull_s8(vb45c1x0, va2x0);
469 int16x8_t vprod3x45c1 = vmull_s8(vb45c1x0, va3x0);
470 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
471 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
472 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
473 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c1);
474 int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
475 int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1x0);
476 int16x8_t vprod2x67c1 = vmull_s8(vb67c1x0, va2x0);
477 int16x8_t vprod3x67c1 = vmull_s8(vb67c1x0, va3x0);
478 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
479 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
480 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
481 vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c1);
482 int16x8_t vprod0x89c1 = vmull_s8(vb89c1x0, va0x0);
483 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0);
484 int16x8_t vprod2x89c1 = vmull_s8(vb89c1x0, va2x0);
485 int16x8_t vprod3x89c1 = vmull_s8(vb89c1x0, va3x0);
486 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
487 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
488 vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c1);
489 vacc3x89 = vpadalq_s16(vacc3x89, vprod3x89c1);
490 int16x8_t vprod0xABc1 = vmull_s8(vbABc1x0, va0x0);
491 int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1x0);
492 int16x8_t vprod2xABc1 = vmull_s8(vbABc1x0, va2x0);
493 int16x8_t vprod3xABc1 = vmull_s8(vbABc1x0, va3x0);
494 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
495 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
496 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc1);
497 vacc3xAB = vpadalq_s16(vacc3xAB, vprod3xABc1);
498 int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1x0, va0x0);
499 int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1x0);
500 int16x8_t vprod2xCDc1 = vmull_s8(vbCDc1x0, va2x0);
501 int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1x0, va3x0);
502 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
503 vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
504 vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc1);
505 vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1);
506 int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1x0, va0x0);
507 int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1x0, va1x0);
508 int16x8_t vprod2xEFc1 = vmull_s8(vbEFc1x0, va2x0);
509 int16x8_t vprod3xEFc1 = vmull_s8(vbEFc1x0, va3x0);
510 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
511 vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
512 vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc1);
513 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc1);
514
515 }
516
517 p -= 4 * sizeof(void*);
518 } while (p != 0);
519
520 #if XNN_ARCH_ARM64
521 int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
522 int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
523 int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB);
524 int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF);
525 int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
526 int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
527 int32x4_t vacc1x89AB = vpaddq_s32(vacc1x89, vacc1xAB);
528 int32x4_t vacc1xCDEF = vpaddq_s32(vacc1xCD, vacc1xEF);
529 int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23);
530 int32x4_t vacc2x4567 = vpaddq_s32(vacc2x45, vacc2x67);
531 int32x4_t vacc2x89AB = vpaddq_s32(vacc2x89, vacc2xAB);
532 int32x4_t vacc2xCDEF = vpaddq_s32(vacc2xCD, vacc2xEF);
533 int32x4_t vacc3x0123 = vpaddq_s32(vacc3x01, vacc3x23);
534 int32x4_t vacc3x4567 = vpaddq_s32(vacc3x45, vacc3x67);
535 int32x4_t vacc3x89AB = vpaddq_s32(vacc3x89, vacc3xAB);
536 int32x4_t vacc3xCDEF = vpaddq_s32(vacc3xCD, vacc3xEF);
537 #else
538 const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
539 const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
540 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
541 const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
542 const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
543 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
544 const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89));
545 const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB));
546 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB);
547 const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD));
548 const int32x2_t vsum0xEF = vpadd_s32(vget_low_s32(vacc0xEF), vget_high_s32(vacc0xEF));
549 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF);
550 const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
551 const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
552 int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
553 const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
554 const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
555 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
556 const int32x2_t vsum1x89 = vpadd_s32(vget_low_s32(vacc1x89), vget_high_s32(vacc1x89));
557 const int32x2_t vsum1xAB = vpadd_s32(vget_low_s32(vacc1xAB), vget_high_s32(vacc1xAB));
558 int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB);
559 const int32x2_t vsum1xCD = vpadd_s32(vget_low_s32(vacc1xCD), vget_high_s32(vacc1xCD));
560 const int32x2_t vsum1xEF = vpadd_s32(vget_low_s32(vacc1xEF), vget_high_s32(vacc1xEF));
561 int32x4_t vacc1xCDEF = vcombine_s32(vsum1xCD, vsum1xEF);
562 const int32x2_t vsum2x01 = vpadd_s32(vget_low_s32(vacc2x01), vget_high_s32(vacc2x01));
563 const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23));
564 int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23);
565 const int32x2_t vsum2x45 = vpadd_s32(vget_low_s32(vacc2x45), vget_high_s32(vacc2x45));
566 const int32x2_t vsum2x67 = vpadd_s32(vget_low_s32(vacc2x67), vget_high_s32(vacc2x67));
567 int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67);
568 const int32x2_t vsum2x89 = vpadd_s32(vget_low_s32(vacc2x89), vget_high_s32(vacc2x89));
569 const int32x2_t vsum2xAB = vpadd_s32(vget_low_s32(vacc2xAB), vget_high_s32(vacc2xAB));
570 int32x4_t vacc2x89AB = vcombine_s32(vsum2x89, vsum2xAB);
571 const int32x2_t vsum2xCD = vpadd_s32(vget_low_s32(vacc2xCD), vget_high_s32(vacc2xCD));
572 const int32x2_t vsum2xEF = vpadd_s32(vget_low_s32(vacc2xEF), vget_high_s32(vacc2xEF));
573 int32x4_t vacc2xCDEF = vcombine_s32(vsum2xCD, vsum2xEF);
574 const int32x2_t vsum3x01 = vpadd_s32(vget_low_s32(vacc3x01), vget_high_s32(vacc3x01));
575 const int32x2_t vsum3x23 = vpadd_s32(vget_low_s32(vacc3x23), vget_high_s32(vacc3x23));
576 int32x4_t vacc3x0123 = vcombine_s32(vsum3x01, vsum3x23);
577 const int32x2_t vsum3x45 = vpadd_s32(vget_low_s32(vacc3x45), vget_high_s32(vacc3x45));
578 const int32x2_t vsum3x67 = vpadd_s32(vget_low_s32(vacc3x67), vget_high_s32(vacc3x67));
579 int32x4_t vacc3x4567 = vcombine_s32(vsum3x45, vsum3x67);
580 const int32x2_t vsum3x89 = vpadd_s32(vget_low_s32(vacc3x89), vget_high_s32(vacc3x89));
581 const int32x2_t vsum3xAB = vpadd_s32(vget_low_s32(vacc3xAB), vget_high_s32(vacc3xAB));
582 int32x4_t vacc3x89AB = vcombine_s32(vsum3x89, vsum3xAB);
583 const int32x2_t vsum3xCD = vpadd_s32(vget_low_s32(vacc3xCD), vget_high_s32(vacc3xCD));
584 const int32x2_t vsum3xEF = vpadd_s32(vget_low_s32(vacc3xEF), vget_high_s32(vacc3xEF));
585 int32x4_t vacc3xCDEF = vcombine_s32(vsum3xCD, vsum3xEF);
586 #endif
587
588 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
589 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
590 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
591
592 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
593 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
594 vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
595 vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
596 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
597 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
598 vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
599 vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
600 vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
601 vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
602 vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
603 vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
604 vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
605 vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
606 vacc3x89AB = vqshlq_s32(vacc3x89AB, vright_pre_shift);
607 vacc3xCDEF = vqshlq_s32(vacc3xCDEF, vright_pre_shift);
608
609 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
610 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
611 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
612 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
613 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
614 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
615 vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
616 vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
617 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
618 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
619 vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
620 vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
621 vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
622 vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
623 vacc3x89AB = vqdmulhq_s32(vacc3x89AB, vmultiplier);
624 vacc3xCDEF = vqdmulhq_s32(vacc3xCDEF, vmultiplier);
625
626 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
627 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
628 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
629 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
630 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
631 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
632 vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
633 vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
634 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
635 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
636 vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
637 vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
638 vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
639 vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
640 vacc3x89AB = vrshlq_s32(vacc3x89AB, vright_post_shift);
641 vacc3xCDEF = vrshlq_s32(vacc3xCDEF, vright_post_shift);
642
643 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
644 #if XNN_ARCH_ARM64
645 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
646 int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
647 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
648 int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
649 int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
650 int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
651 int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
652 int16x8_t vacc3x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF);
653
654 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
655 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
656 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
657 vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
658 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
659 vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
660 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
661 vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
662
663 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
664 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
665 int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
666 int8x16_t vout3x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc3x01234567), vacc3x89ABCDEF);
667 #else
668 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
669 int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
670 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
671 int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
672 int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
673 int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
674 int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
675 int16x8_t vacc3x89ABCDEF = vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF));
676
677 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
678 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
679 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
680 vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
681 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
682 vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
683 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
684 vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
685
686 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
687 int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
688 int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
689 int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCDEF));
690 #endif
691
692 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
693 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
694 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
695 vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
696 vout3x0123456789ABCDEF = vmaxq_s8(vout3x0123456789ABCDEF, voutput_min);
697
698 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
699 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
700 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
701 vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
702 vout3x0123456789ABCDEF = vminq_s8(vout3x0123456789ABCDEF, voutput_max);
703
704 if (nc >= 16) {
705 vst1q_s8(c3 + 0, vout3x0123456789ABCDEF);
706 vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
707 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
708 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
709
710 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
711 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
712 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
713 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
714
715 a = (const int8_t**restrict) ((uintptr_t) a - ks);
716
717 nc -= 16;
718 } else {
719 int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8(vout3x0123456789ABCDEF));
720 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
721 if (nc & 8) {
722 vst1_s8(c3, vget_high_s8(vout2x01234567_3x01234567)); c3 += 8;
723 vst1_s8(c2, vget_low_s8(vout2x01234567_3x01234567)); c2 += 8;
724 vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
725 vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
726 vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0123456789ABCDEF));
727 vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
728 }
729 if (nc & 4) {
730 vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
731 vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
732 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
733 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
734 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
735 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
736 }
737 if (nc & 2) {
738 vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
739 vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
740 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
741 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
742 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
743 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
744 }
745 if (nc & 1) {
746 vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
747 vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
748 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
749 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
750 }
751
752 nc = 0;
753 }
754 } while (nc != 0);
755 }
756