1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c2-neon-mull-dup.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
xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup(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)])17 void xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 size_t ks,
22 const int8_t** restrict a,
23 const void* restrict w,
24 int8_t* restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 size_t a_offset,
28 const int8_t* zero,
29 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 4);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(ks != 0);
36 assert(ks % (4 * sizeof(void*)) == 0);
37 assert(a_offset % sizeof(int8_t) == 0);
38 assert(a != NULL);
39 assert(w != NULL);
40 assert(c != NULL);
41
42 kc = round_up_po2(kc, 2 * sizeof(int8_t));
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 do {
58 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
59 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
60 int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
61 int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
62 int32x4_t vacc1x0123 = vacc0x0123;
63 int32x4_t vacc1x4567 = vacc0x4567;
64 int32x4_t vacc1x89AB = vacc0x89AB;
65 int32x4_t vacc1xCDEF = vacc0xCDEF;
66 int32x4_t vacc2x0123 = vacc0x0123;
67 int32x4_t vacc2x4567 = vacc0x4567;
68 int32x4_t vacc2x89AB = vacc0x89AB;
69 int32x4_t vacc2xCDEF = vacc0xCDEF;
70 int32x4_t vacc3x0123 = vacc0x0123;
71 int32x4_t vacc3x4567 = vacc0x4567;
72 int32x4_t vacc3x89AB = vacc0x89AB;
73 int32x4_t vacc3xCDEF = vacc0xCDEF;
74
75 size_t p = ks;
76 do {
77 const int8_t* restrict a0 = a[0];
78 if XNN_UNPREDICTABLE(a0 != zero) {
79 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
80 }
81 const int8_t* restrict a1 = a[1];
82 if XNN_UNPREDICTABLE(a1 != zero) {
83 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
84 }
85 const int8_t* restrict a2 = a[2];
86 if XNN_UNPREDICTABLE(a2 != zero) {
87 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
88 }
89 const int8_t* restrict a3 = a[3];
90 if XNN_UNPREDICTABLE(a3 != zero) {
91 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
92 }
93 a += 4;
94
95 size_t k = kc;
96
97 while (k >= 16 * sizeof(int8_t)) {
98 const int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
99 const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
100 const int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
101 const int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
102 const int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
103 const int8x8_t va2x1 = vld1_s8(a2); a2 += 8;
104 const int8x8_t va3x0 = vld1_s8(a3); a3 += 8;
105 const int8x8_t va3x1 = vld1_s8(a3); a3 += 8;
106
107 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
108 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
109 const int8x8_t vb89ABc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
110 const int8x8_t vbCDEFc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
111 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
112 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
113 const int8x8_t vb89ABc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
114 const int8x8_t vbCDEFc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
115 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
116 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
117 const int8x8_t vb89ABc2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
118 const int8x8_t vbCDEFc2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
119 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
120 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
121 const int8x8_t vb89ABc3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
122 const int8x8_t vbCDEFc3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
123
124 const int8x8_t va0c0x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 0));
125 const int8x8_t va0c0x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 0));
126 const int8x8_t va1c0x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 0));
127 const int8x8_t va1c0x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 0));
128 const int8x8_t va2c0x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 0));
129 const int8x8_t va2c0x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 0));
130 const int8x8_t va3c0x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3x0), 0));
131 const int8x8_t va3c0x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3x1), 0));
132
133 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0c0x0);
134 int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1c0x0);
135 int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0x0, va2c0x0);
136 int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0x0, va3c0x0);
137 const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
138 vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0c0x1);
139 vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, va1c0x1);
140 vprod2x0123c0 = vmlal_s8(vprod2x0123c0, vb0123c0x1, va2c0x1);
141 vprod3x0123c0 = vmlal_s8(vprod3x0123c0, vb0123c0x1, va3c0x1);
142 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
143 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
144 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
145 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c0);
146 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0c0x0);
147 int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1c0x0);
148 int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2c0x0);
149 int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0x0, va3c0x0);
150 const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
151 vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0c0x1);
152 vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, va1c0x1);
153 vprod2x4567c0 = vmlal_s8(vprod2x4567c0, vb4567c0x1, va2c0x1);
154 vprod3x4567c0 = vmlal_s8(vprod3x4567c0, vb4567c0x1, va3c0x1);
155 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
156 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
157 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
158 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0);
159 int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, va0c0x0);
160 int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1c0x0);
161 int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0x0, va2c0x0);
162 int16x8_t vprod3x89ABc0 = vmull_s8(vb89ABc0x0, va3c0x0);
163 const int8x8_t vb89ABc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
164 vprod0x89ABc0 = vmlal_s8(vprod0x89ABc0, vb89ABc0x1, va0c0x1);
165 vprod1x89ABc0 = vmlal_s8(vprod1x89ABc0, vb89ABc0x1, va1c0x1);
166 vprod2x89ABc0 = vmlal_s8(vprod2x89ABc0, vb89ABc0x1, va2c0x1);
167 vprod3x89ABc0 = vmlal_s8(vprod3x89ABc0, vb89ABc0x1, va3c0x1);
168 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
169 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
170 vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
171 vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc0);
172 int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0x0, va0c0x0);
173 int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0x0, va1c0x0);
174 int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0x0, va2c0x0);
175 int16x8_t vprod3xCDEFc0 = vmull_s8(vbCDEFc0x0, va3c0x0);
176 const int8x8_t vbCDEFc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
177 vprod0xCDEFc0 = vmlal_s8(vprod0xCDEFc0, vbCDEFc0x1, va0c0x1);
178 vprod1xCDEFc0 = vmlal_s8(vprod1xCDEFc0, vbCDEFc0x1, va1c0x1);
179 vprod2xCDEFc0 = vmlal_s8(vprod2xCDEFc0, vbCDEFc0x1, va2c0x1);
180 vprod3xCDEFc0 = vmlal_s8(vprod3xCDEFc0, vbCDEFc0x1, va3c0x1);
181 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
182 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
183 vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
184 vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc0);
185 const int8x8_t va0c1x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 1));
186 const int8x8_t va0c1x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 1));
187 const int8x8_t va1c1x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 1));
188 const int8x8_t va1c1x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 1));
189 const int8x8_t va2c1x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 1));
190 const int8x8_t va2c1x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 1));
191 const int8x8_t va3c1x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3x0), 1));
192 const int8x8_t va3c1x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3x1), 1));
193
194 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0c1x0);
195 int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1c1x0);
196 int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2c1x0);
197 int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1x0, va3c1x0);
198 const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
199 vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0c1x1);
200 vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, va1c1x1);
201 vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, va2c1x1);
202 vprod3x0123c1 = vmlal_s8(vprod3x0123c1, vb0123c1x1, va3c1x1);
203 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
204 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
205 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
206 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1);
207 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0);
208 int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1c1x0);
209 int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1x0, va2c1x0);
210 int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1x0, va3c1x0);
211 const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
212 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1);
213 vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, va1c1x1);
214 vprod2x4567c1 = vmlal_s8(vprod2x4567c1, vb4567c1x1, va2c1x1);
215 vprod3x4567c1 = vmlal_s8(vprod3x4567c1, vb4567c1x1, va3c1x1);
216 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
217 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
218 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
219 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c1);
220 int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0c1x0);
221 int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1x0, va1c1x0);
222 int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1x0, va2c1x0);
223 int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1x0, va3c1x0);
224 const int8x8_t vb89ABc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
225 vprod0x89ABc1 = vmlal_s8(vprod0x89ABc1, vb89ABc1x1, va0c1x1);
226 vprod1x89ABc1 = vmlal_s8(vprod1x89ABc1, vb89ABc1x1, va1c1x1);
227 vprod2x89ABc1 = vmlal_s8(vprod2x89ABc1, vb89ABc1x1, va2c1x1);
228 vprod3x89ABc1 = vmlal_s8(vprod3x89ABc1, vb89ABc1x1, va3c1x1);
229 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
230 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
231 vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
232 vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc1);
233 int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0c1x0);
234 int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1c1x0);
235 int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1x0, va2c1x0);
236 int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1x0, va3c1x0);
237 const int8x8_t vbCDEFc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
238 vprod0xCDEFc1 = vmlal_s8(vprod0xCDEFc1, vbCDEFc1x1, va0c1x1);
239 vprod1xCDEFc1 = vmlal_s8(vprod1xCDEFc1, vbCDEFc1x1, va1c1x1);
240 vprod2xCDEFc1 = vmlal_s8(vprod2xCDEFc1, vbCDEFc1x1, va2c1x1);
241 vprod3xCDEFc1 = vmlal_s8(vprod3xCDEFc1, vbCDEFc1x1, va3c1x1);
242 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
243 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
244 vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
245 vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc1);
246 const int8x8_t va0c2x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 2));
247 const int8x8_t va0c2x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 2));
248 const int8x8_t va1c2x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 2));
249 const int8x8_t va1c2x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 2));
250 const int8x8_t va2c2x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 2));
251 const int8x8_t va2c2x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 2));
252 const int8x8_t va3c2x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3x0), 2));
253 const int8x8_t va3c2x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3x1), 2));
254
255 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0c2x0);
256 int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1c2x0);
257 int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2x0, va2c2x0);
258 int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2x0, va3c2x0);
259 const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
260 vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0c2x1);
261 vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, va1c2x1);
262 vprod2x0123c2 = vmlal_s8(vprod2x0123c2, vb0123c2x1, va2c2x1);
263 vprod3x0123c2 = vmlal_s8(vprod3x0123c2, vb0123c2x1, va3c2x1);
264 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
265 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
266 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
267 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c2);
268 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0c2x0);
269 int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1c2x0);
270 int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2x0, va2c2x0);
271 int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2x0, va3c2x0);
272 const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
273 vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0c2x1);
274 vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, va1c2x1);
275 vprod2x4567c2 = vmlal_s8(vprod2x4567c2, vb4567c2x1, va2c2x1);
276 vprod3x4567c2 = vmlal_s8(vprod3x4567c2, vb4567c2x1, va3c2x1);
277 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
278 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
279 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
280 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2);
281 int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2x0, va0c2x0);
282 int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2x0, va1c2x0);
283 int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2x0, va2c2x0);
284 int16x8_t vprod3x89ABc2 = vmull_s8(vb89ABc2x0, va3c2x0);
285 const int8x8_t vb89ABc2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
286 vprod0x89ABc2 = vmlal_s8(vprod0x89ABc2, vb89ABc2x1, va0c2x1);
287 vprod1x89ABc2 = vmlal_s8(vprod1x89ABc2, vb89ABc2x1, va1c2x1);
288 vprod2x89ABc2 = vmlal_s8(vprod2x89ABc2, vb89ABc2x1, va2c2x1);
289 vprod3x89ABc2 = vmlal_s8(vprod3x89ABc2, vb89ABc2x1, va3c2x1);
290 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
291 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
292 vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
293 vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc2);
294 int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0c2x0);
295 int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2x0, va1c2x0);
296 int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2x0, va2c2x0);
297 int16x8_t vprod3xCDEFc2 = vmull_s8(vbCDEFc2x0, va3c2x0);
298 const int8x8_t vbCDEFc2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
299 vprod0xCDEFc2 = vmlal_s8(vprod0xCDEFc2, vbCDEFc2x1, va0c2x1);
300 vprod1xCDEFc2 = vmlal_s8(vprod1xCDEFc2, vbCDEFc2x1, va1c2x1);
301 vprod2xCDEFc2 = vmlal_s8(vprod2xCDEFc2, vbCDEFc2x1, va2c2x1);
302 vprod3xCDEFc2 = vmlal_s8(vprod3xCDEFc2, vbCDEFc2x1, va3c2x1);
303 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
304 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
305 vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
306 vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc2);
307 const int8x8_t va0c3x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 3));
308 const int8x8_t va0c3x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 3));
309 const int8x8_t va1c3x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 3));
310 const int8x8_t va1c3x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 3));
311 const int8x8_t va2c3x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 3));
312 const int8x8_t va2c3x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 3));
313 const int8x8_t va3c3x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3x0), 3));
314 const int8x8_t va3c3x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3x1), 3));
315
316 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0c3x0);
317 int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1c3x0);
318 int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3x0, va2c3x0);
319 int16x8_t vprod3x0123c3 = vmull_s8(vb0123c3x0, va3c3x0);
320 const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
321 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1);
322 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1);
323 vprod2x0123c3 = vmlal_s8(vprod2x0123c3, vb0123c3x1, va2c3x1);
324 vprod3x0123c3 = vmlal_s8(vprod3x0123c3, vb0123c3x1, va3c3x1);
325 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
326 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
327 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
328 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c3);
329 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0c3x0);
330 int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1c3x0);
331 int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3x0, va2c3x0);
332 int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3c3x0);
333 const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
334 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1);
335 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1);
336 vprod2x4567c3 = vmlal_s8(vprod2x4567c3, vb4567c3x1, va2c3x1);
337 vprod3x4567c3 = vmlal_s8(vprod3x4567c3, vb4567c3x1, va3c3x1);
338 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
339 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
340 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
341 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3);
342 int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3x0, va0c3x0);
343 int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3x0, va1c3x0);
344 int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, va2c3x0);
345 int16x8_t vprod3x89ABc3 = vmull_s8(vb89ABc3x0, va3c3x0);
346 const int8x8_t vb89ABc3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
347 vprod0x89ABc3 = vmlal_s8(vprod0x89ABc3, vb89ABc3x1, va0c3x1);
348 vprod1x89ABc3 = vmlal_s8(vprod1x89ABc3, vb89ABc3x1, va1c3x1);
349 vprod2x89ABc3 = vmlal_s8(vprod2x89ABc3, vb89ABc3x1, va2c3x1);
350 vprod3x89ABc3 = vmlal_s8(vprod3x89ABc3, vb89ABc3x1, va3c3x1);
351 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
352 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
353 vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc3);
354 vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc3);
355 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0c3x0);
356 int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3x0, va1c3x0);
357 int16x8_t vprod2xCDEFc3 = vmull_s8(vbCDEFc3x0, va2c3x0);
358 int16x8_t vprod3xCDEFc3 = vmull_s8(vbCDEFc3x0, va3c3x0);
359 const int8x8_t vbCDEFc3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
360 vprod0xCDEFc3 = vmlal_s8(vprod0xCDEFc3, vbCDEFc3x1, va0c3x1);
361 vprod1xCDEFc3 = vmlal_s8(vprod1xCDEFc3, vbCDEFc3x1, va1c3x1);
362 vprod2xCDEFc3 = vmlal_s8(vprod2xCDEFc3, vbCDEFc3x1, va2c3x1);
363 vprod3xCDEFc3 = vmlal_s8(vprod3xCDEFc3, vbCDEFc3x1, va3c3x1);
364 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
365 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
366 vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc3);
367 vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc3);
368
369 k -= 16 * sizeof(int8_t);
370 }
371
372 if (k >= 8 * sizeof(int8_t)) {
373 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
374 const int8x8_t va1 = vld1_s8(a1); a1 += 8;
375 const int8x8_t va2 = vld1_s8(a2); a2 += 8;
376 const int8x8_t va3 = vld1_s8(a3); a3 += 8;
377
378 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
379 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
380 const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
381 const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
382 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
383 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
384 const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
385 const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
386 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
387 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
388 const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
389 const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
390 const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
391 const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
392 const int8x8_t vb89ABc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
393 const int8x8_t vbCDEFc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
394
395 const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
396 const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
397 const int8x8_t va2c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0));
398 const int8x8_t va3c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 0));
399
400 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
401 const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
402 const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
403 const int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0, va3c0);
404 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
405 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
406 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
407 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c0);
408 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
409 const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
410 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
411 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0);
412 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
413 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
414 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
415 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0);
416 const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
417 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0);
418 const int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0, va2c0);
419 const int16x8_t vprod3x89ABc0 = vmull_s8(vb89ABc0, va3c0);
420 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
421 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
422 vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
423 vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc0);
424 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
425 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0);
426 const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0);
427 const int16x8_t vprod3xCDEFc0 = vmull_s8(vbCDEFc0, va3c0);
428 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
429 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
430 vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
431 vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc0);
432 const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
433 const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
434 const int8x8_t va2c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1));
435 const int8x8_t va3c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 1));
436
437 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
438 const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
439 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
440 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1);
441 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
442 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
443 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
444 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1);
445 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
446 const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
447 const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
448 const int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1, va3c1);
449 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
450 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
451 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
452 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c1);
453 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
454 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1);
455 const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1);
456 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1);
457 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
458 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
459 vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
460 vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc1);
461 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
462 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1);
463 const int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1, va2c1);
464 const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1);
465 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
466 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
467 vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
468 vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc1);
469 const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
470 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
471 const int8x8_t va2c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2));
472 const int8x8_t va3c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 2));
473
474 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
475 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
476 const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
477 const int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2, va3c2);
478 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
479 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
480 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
481 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c2);
482 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
483 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
484 const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
485 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2);
486 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
487 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
488 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
489 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2);
490 const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
491 const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2);
492 const int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2, va2c2);
493 const int16x8_t vprod3x89ABc2 = vmull_s8(vb89ABc2, va3c2);
494 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
495 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
496 vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
497 vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc2);
498 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
499 const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2);
500 const int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2, va2c2);
501 const int16x8_t vprod3xCDEFc2 = vmull_s8(vbCDEFc2, va3c2);
502 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
503 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
504 vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
505 vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc2);
506 const int8x8_t va0c3 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 3));
507 const int8x8_t va1c3 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 3));
508 const int8x8_t va2c3 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 3));
509 const int8x8_t va3c3 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 3));
510
511 const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
512 const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
513 const int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3, va2c3);
514 const int16x8_t vprod3x0123c3 = vmull_s8(vb0123c3, va3c3);
515 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
516 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
517 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
518 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c3);
519 const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
520 const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
521 const int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3, va2c3);
522 const int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3, va3c3);
523 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
524 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
525 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
526 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3);
527 const int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3, va0c3);
528 const int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3, va1c3);
529 const int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3, va2c3);
530 const int16x8_t vprod3x89ABc3 = vmull_s8(vb89ABc3, va3c3);
531 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
532 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
533 vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc3);
534 vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc3);
535 const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3);
536 const int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3, va1c3);
537 const int16x8_t vprod2xCDEFc3 = vmull_s8(vbCDEFc3, va2c3);
538 const int16x8_t vprod3xCDEFc3 = vmull_s8(vbCDEFc3, va3c3);
539 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
540 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
541 vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc3);
542 vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc3);
543
544 k -= 8 * sizeof(int8_t);
545 }
546
547 if XNN_UNLIKELY(k != 0) {
548 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
549 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
550 const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
551 const int8x8_t va3 = vld1_s8(a3); a3 = (const int8_t*) ((uintptr_t) a3 + k);
552
553 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
554 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
555 const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
556 const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
557
558 const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
559 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
560 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
561 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
562 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
563 const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
564 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
565 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
566 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
567 const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
568 const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
569 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
570 const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
571 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
572 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0);
573 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
574 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0);
575 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
576 const int8x8_t va2c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0));
577 const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
578 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
579 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
580 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
581 const int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0, va2c0);
582 vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
583 const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0);
584 vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
585 const int8x8_t va3c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 0));
586 const int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0, va3c0);
587 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c0);
588 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0);
589 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0);
590 const int16x8_t vprod3x89ABc0 = vmull_s8(vb89ABc0, va3c0);
591 vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc0);
592 const int16x8_t vprod3xCDEFc0 = vmull_s8(vbCDEFc0, va3c0);
593 vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc0);
594
595 if (k > 2 * sizeof(int8_t)) {
596 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
597 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
598 const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
599 const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
600
601 const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
602 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
603 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
604 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
605 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
606 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
607 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
608 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
609 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
610 const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
611 const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
612 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
613 const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
614 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
615 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1);
616 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
617 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1);
618 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
619 const int8x8_t va2c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1));
620 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
621 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
622 const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
623 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
624 const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1);
625 vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
626 const int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1, va2c1);
627 vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
628 const int8x8_t va3c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 1));
629 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1);
630 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1);
631 const int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1, va3c1);
632 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c1);
633 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1);
634 vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc1);
635 const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1);
636 vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc1);
637
638 if (k > 4 * sizeof(int8_t)) {
639 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
640 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
641 const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
642 const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
643
644 const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
645 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
646 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
647 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
648 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
649 const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
650 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
651 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
652 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
653 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
654 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
655 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
656 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
657 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
658 const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2);
659 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
660 const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2);
661 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
662 const int8x8_t va2c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2));
663 const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
664 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
665 const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
666 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
667 const int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2, va2c2);
668 vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
669 const int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2, va2c2);
670 vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
671 const int8x8_t va3c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 2));
672 const int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2, va3c2);
673 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c2);
674 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2);
675 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2);
676 const int16x8_t vprod3x89ABc2 = vmull_s8(vb89ABc2, va3c2);
677 vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc2);
678 const int16x8_t vprod3xCDEFc2 = vmull_s8(vbCDEFc2, va3c2);
679 vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc2);
680 }
681 }
682 }
683 p -= 4 * sizeof(void*);
684 } while (p != 0);
685
686 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
687 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
688 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
689
690 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
691 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
692 vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
693 vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
694 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
695 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
696 vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
697 vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
698 vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
699 vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
700 vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
701 vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
702 vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
703 vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
704 vacc3x89AB = vqshlq_s32(vacc3x89AB, vright_pre_shift);
705 vacc3xCDEF = vqshlq_s32(vacc3xCDEF, vright_pre_shift);
706
707 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
708 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
709 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
710 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
711 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
712 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
713 vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
714 vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
715 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
716 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
717 vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
718 vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
719 vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
720 vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
721 vacc3x89AB = vqdmulhq_s32(vacc3x89AB, vmultiplier);
722 vacc3xCDEF = vqdmulhq_s32(vacc3xCDEF, vmultiplier);
723
724 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
725 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
726 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
727 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
728 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
729 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
730 vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
731 vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
732 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
733 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
734 vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
735 vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
736 vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
737 vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
738 vacc3x89AB = vrshlq_s32(vacc3x89AB, vright_post_shift);
739 vacc3xCDEF = vrshlq_s32(vacc3xCDEF, vright_post_shift);
740
741 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
742 #if XNN_ARCH_ARM64
743 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
744 int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
745 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
746 int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
747 int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
748 int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
749 int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
750 int16x8_t vacc3x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF);
751
752 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
753 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
754 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
755 vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
756 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
757 vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
758 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
759 vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
760
761 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
762 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
763 int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
764 int8x16_t vout3x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc3x01234567), vacc3x89ABCDEF);
765 #else
766 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
767 int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
768 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
769 int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
770 int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
771 int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
772 int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
773 int16x8_t vacc3x89ABCDEF = vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF));
774
775 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
776 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
777 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
778 vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
779 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
780 vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
781 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
782 vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
783
784 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
785 int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
786 int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
787 int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCDEF));
788 #endif
789
790 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
791 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
792 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
793 vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
794 vout3x0123456789ABCDEF = vmaxq_s8(vout3x0123456789ABCDEF, voutput_min);
795
796 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
797 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
798 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
799 vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
800 vout3x0123456789ABCDEF = vminq_s8(vout3x0123456789ABCDEF, voutput_max);
801
802 if (nc >= 16) {
803 vst1q_s8(c3 + 0, vout3x0123456789ABCDEF);
804 vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
805 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
806 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
807
808 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
809 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
810 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
811 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
812
813 a = (const int8_t**restrict) ((uintptr_t) a - ks);
814
815 nc -= 16;
816 } else {
817 int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8(vout3x0123456789ABCDEF));
818 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
819 if (nc & 8) {
820 vst1_s8(c3, vget_high_s8(vout2x01234567_3x01234567)); c3 += 8;
821 vst1_s8(c2, vget_low_s8(vout2x01234567_3x01234567)); c2 += 8;
822 vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
823 vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
824 vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0123456789ABCDEF));
825 vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
826 }
827 if (nc & 4) {
828 vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
829 vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
830 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
831 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
832 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
833 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
834 }
835 if (nc & 2) {
836 vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
837 vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
838 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
839 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
840 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
841 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
842 }
843 if (nc & 1) {
844 vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
845 vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
846 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
847 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
848 }
849
850 nc = 0;
851 }
852 } while (nc != 0);
853 }
854