1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 const int8_t* restrict a,
22 size_t a_stride,
23 const void* restrict w,
24 int8_t* restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29 assert(mr != 0);
30 assert(mr <= 4);
31 assert(nc != 0);
32 assert(kc != 0);
33 assert(kc % sizeof(int8_t) == 0);
34 assert(a != NULL);
35 assert(w != NULL);
36 assert(c != NULL);
37
38 kc = round_up_po2(kc, 2 * sizeof(int8_t));
39 const int8_t* a0 = a;
40 int8_t* c0 = c;
41 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
42 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
43 if XNN_UNPREDICTABLE(mr < 2) {
44 a1 = a0;
45 c1 = c0;
46 }
47 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
48 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
49 if XNN_UNPREDICTABLE(mr <= 2) {
50 a2 = a1;
51 c2 = c1;
52 }
53 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
54 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
55 if XNN_UNPREDICTABLE(mr != 4) {
56 a3 = a2;
57 c3 = c2;
58 }
59
60 do {
61 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
62 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
63 int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
64 int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
65 int32x4_t vacc1x0123 = vacc0x0123;
66 int32x4_t vacc1x4567 = vacc0x4567;
67 int32x4_t vacc1x89AB = vacc0x89AB;
68 int32x4_t vacc1xCDEF = vacc0xCDEF;
69 int32x4_t vacc2x0123 = vacc0x0123;
70 int32x4_t vacc2x4567 = vacc0x4567;
71 int32x4_t vacc2x89AB = vacc0x89AB;
72 int32x4_t vacc2xCDEF = vacc0xCDEF;
73 int32x4_t vacc3x0123 = vacc0x0123;
74 int32x4_t vacc3x4567 = vacc0x4567;
75 int32x4_t vacc3x89AB = vacc0x89AB;
76 int32x4_t vacc3xCDEF = vacc0xCDEF;
77
78 size_t k = kc;
79
80 while (k >= 16 * sizeof(int8_t)) {
81 const int16x4_t va00x0 = vld1_dup_s16((const void*)a0);
82 const int16x4_t va01x0 = vld1_dup_s16((const void*)(a0 + 2));
83 const int16x4_t va02x0 = vld1_dup_s16((const void*)(a0 + 4));
84 const int16x4_t va03x0 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
85 const int16x4_t va00x1 = vld1_dup_s16((const void*)a0);
86 const int16x4_t va01x1 = vld1_dup_s16((const void*)(a0 + 2));
87 const int16x4_t va02x1 = vld1_dup_s16((const void*)(a0 + 4));
88 const int16x4_t va03x1 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
89 const int16x4_t va10x0 = vld1_dup_s16((const void*)a1);
90 const int16x4_t va11x0 = vld1_dup_s16((const void*)(a1 + 2));
91 const int16x4_t va12x0 = vld1_dup_s16((const void*)(a1 + 4));
92 const int16x4_t va13x0 = vld1_dup_s16((const void*)(a1 + 6)); a1 += 8;
93 const int16x4_t va10x1 = vld1_dup_s16((const void*)a1);
94 const int16x4_t va11x1 = vld1_dup_s16((const void*)(a1 + 2));
95 const int16x4_t va12x1 = vld1_dup_s16((const void*)(a1 + 4));
96 const int16x4_t va13x1 = vld1_dup_s16((const void*)(a1 + 6)); a1 += 8;
97 const int16x4_t va20x0 = vld1_dup_s16((const void*)a2);
98 const int16x4_t va21x0 = vld1_dup_s16((const void*)(a2 + 2));
99 const int16x4_t va22x0 = vld1_dup_s16((const void*)(a2 + 4));
100 const int16x4_t va23x0 = vld1_dup_s16((const void*)(a2 + 6)); a2 += 8;
101 const int16x4_t va20x1 = vld1_dup_s16((const void*)a2);
102 const int16x4_t va21x1 = vld1_dup_s16((const void*)(a2 + 2));
103 const int16x4_t va22x1 = vld1_dup_s16((const void*)(a2 + 4));
104 const int16x4_t va23x1 = vld1_dup_s16((const void*)(a2 + 6)); a2 += 8;
105 const int16x4_t va30x0 = vld1_dup_s16((const void*)a3);
106 const int16x4_t va31x0 = vld1_dup_s16((const void*)(a3 + 2));
107 const int16x4_t va32x0 = vld1_dup_s16((const void*)(a3 + 4));
108 const int16x4_t va33x0 = vld1_dup_s16((const void*)(a3 + 6)); a3 += 8;
109 const int16x4_t va30x1 = vld1_dup_s16((const void*)a3);
110 const int16x4_t va31x1 = vld1_dup_s16((const void*)(a3 + 2));
111 const int16x4_t va32x1 = vld1_dup_s16((const void*)(a3 + 4));
112 const int16x4_t va33x1 = vld1_dup_s16((const void*)(a3 + 6)); a3 += 8;
113 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
114 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
115 const int8x8_t vb89ABc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
116 const int8x8_t vbCDEFc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
117 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
118 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
119 const int8x8_t vb89ABc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
120 const int8x8_t vbCDEFc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
121 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
122 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
123 const int8x8_t vb89ABc2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
124 const int8x8_t vbCDEFc2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
125 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
126 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
127 const int8x8_t vb89ABc3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
128 const int8x8_t vbCDEFc3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
129
130 const int8x8_t va0c0x0 = vreinterpret_s8_s16(va00x0);
131 const int8x8_t va0c0x1 = vreinterpret_s8_s16(va00x1);
132 const int8x8_t va1c0x0 = vreinterpret_s8_s16(va10x0);
133 const int8x8_t va1c0x1 = vreinterpret_s8_s16(va10x1);
134 const int8x8_t va2c0x0 = vreinterpret_s8_s16(va20x0);
135 const int8x8_t va2c0x1 = vreinterpret_s8_s16(va20x1);
136 const int8x8_t va3c0x0 = vreinterpret_s8_s16(va30x0);
137 const int8x8_t va3c0x1 = vreinterpret_s8_s16(va30x1);
138
139 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0c0x0);
140 int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1c0x0);
141 int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0x0, va2c0x0);
142 int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0x0, va3c0x0);
143 const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
144 vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0c0x1);
145 vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, va1c0x1);
146 vprod2x0123c0 = vmlal_s8(vprod2x0123c0, vb0123c0x1, va2c0x1);
147 vprod3x0123c0 = vmlal_s8(vprod3x0123c0, vb0123c0x1, va3c0x1);
148 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
149 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
150 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
151 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c0);
152 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0c0x0);
153 int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1c0x0);
154 int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2c0x0);
155 int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0x0, va3c0x0);
156 const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
157 vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0c0x1);
158 vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, va1c0x1);
159 vprod2x4567c0 = vmlal_s8(vprod2x4567c0, vb4567c0x1, va2c0x1);
160 vprod3x4567c0 = vmlal_s8(vprod3x4567c0, vb4567c0x1, va3c0x1);
161 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
162 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
163 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
164 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0);
165 int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, va0c0x0);
166 int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1c0x0);
167 int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0x0, va2c0x0);
168 int16x8_t vprod3x89ABc0 = vmull_s8(vb89ABc0x0, va3c0x0);
169 const int8x8_t vb89ABc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
170 vprod0x89ABc0 = vmlal_s8(vprod0x89ABc0, vb89ABc0x1, va0c0x1);
171 vprod1x89ABc0 = vmlal_s8(vprod1x89ABc0, vb89ABc0x1, va1c0x1);
172 vprod2x89ABc0 = vmlal_s8(vprod2x89ABc0, vb89ABc0x1, va2c0x1);
173 vprod3x89ABc0 = vmlal_s8(vprod3x89ABc0, vb89ABc0x1, va3c0x1);
174 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
175 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
176 vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
177 vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc0);
178 int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0x0, va0c0x0);
179 int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0x0, va1c0x0);
180 int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0x0, va2c0x0);
181 int16x8_t vprod3xCDEFc0 = vmull_s8(vbCDEFc0x0, va3c0x0);
182 const int8x8_t vbCDEFc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
183 vprod0xCDEFc0 = vmlal_s8(vprod0xCDEFc0, vbCDEFc0x1, va0c0x1);
184 vprod1xCDEFc0 = vmlal_s8(vprod1xCDEFc0, vbCDEFc0x1, va1c0x1);
185 vprod2xCDEFc0 = vmlal_s8(vprod2xCDEFc0, vbCDEFc0x1, va2c0x1);
186 vprod3xCDEFc0 = vmlal_s8(vprod3xCDEFc0, vbCDEFc0x1, va3c0x1);
187 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
188 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
189 vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
190 vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc0);
191 const int8x8_t va0c1x0 = vreinterpret_s8_s16(va01x0);
192 const int8x8_t va0c1x1 = vreinterpret_s8_s16(va01x1);
193 const int8x8_t va1c1x0 = vreinterpret_s8_s16(va11x0);
194 const int8x8_t va1c1x1 = vreinterpret_s8_s16(va11x1);
195 const int8x8_t va2c1x0 = vreinterpret_s8_s16(va21x0);
196 const int8x8_t va2c1x1 = vreinterpret_s8_s16(va21x1);
197 const int8x8_t va3c1x0 = vreinterpret_s8_s16(va31x0);
198 const int8x8_t va3c1x1 = vreinterpret_s8_s16(va31x1);
199
200 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0c1x0);
201 int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1c1x0);
202 int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2c1x0);
203 int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1x0, va3c1x0);
204 const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
205 vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0c1x1);
206 vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, va1c1x1);
207 vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, va2c1x1);
208 vprod3x0123c1 = vmlal_s8(vprod3x0123c1, vb0123c1x1, va3c1x1);
209 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
210 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
211 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
212 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1);
213 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0);
214 int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1c1x0);
215 int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1x0, va2c1x0);
216 int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1x0, va3c1x0);
217 const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
218 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1);
219 vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, va1c1x1);
220 vprod2x4567c1 = vmlal_s8(vprod2x4567c1, vb4567c1x1, va2c1x1);
221 vprod3x4567c1 = vmlal_s8(vprod3x4567c1, vb4567c1x1, va3c1x1);
222 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
223 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
224 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
225 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c1);
226 int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0c1x0);
227 int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1x0, va1c1x0);
228 int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1x0, va2c1x0);
229 int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1x0, va3c1x0);
230 const int8x8_t vb89ABc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
231 vprod0x89ABc1 = vmlal_s8(vprod0x89ABc1, vb89ABc1x1, va0c1x1);
232 vprod1x89ABc1 = vmlal_s8(vprod1x89ABc1, vb89ABc1x1, va1c1x1);
233 vprod2x89ABc1 = vmlal_s8(vprod2x89ABc1, vb89ABc1x1, va2c1x1);
234 vprod3x89ABc1 = vmlal_s8(vprod3x89ABc1, vb89ABc1x1, va3c1x1);
235 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
236 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
237 vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
238 vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc1);
239 int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0c1x0);
240 int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1c1x0);
241 int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1x0, va2c1x0);
242 int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1x0, va3c1x0);
243 const int8x8_t vbCDEFc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
244 vprod0xCDEFc1 = vmlal_s8(vprod0xCDEFc1, vbCDEFc1x1, va0c1x1);
245 vprod1xCDEFc1 = vmlal_s8(vprod1xCDEFc1, vbCDEFc1x1, va1c1x1);
246 vprod2xCDEFc1 = vmlal_s8(vprod2xCDEFc1, vbCDEFc1x1, va2c1x1);
247 vprod3xCDEFc1 = vmlal_s8(vprod3xCDEFc1, vbCDEFc1x1, va3c1x1);
248 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
249 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
250 vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
251 vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc1);
252 const int8x8_t va0c2x0 = vreinterpret_s8_s16(va02x0);
253 const int8x8_t va0c2x1 = vreinterpret_s8_s16(va02x1);
254 const int8x8_t va1c2x0 = vreinterpret_s8_s16(va12x0);
255 const int8x8_t va1c2x1 = vreinterpret_s8_s16(va12x1);
256 const int8x8_t va2c2x0 = vreinterpret_s8_s16(va22x0);
257 const int8x8_t va2c2x1 = vreinterpret_s8_s16(va22x1);
258 const int8x8_t va3c2x0 = vreinterpret_s8_s16(va32x0);
259 const int8x8_t va3c2x1 = vreinterpret_s8_s16(va32x1);
260
261 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0c2x0);
262 int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1c2x0);
263 int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2x0, va2c2x0);
264 int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2x0, va3c2x0);
265 const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
266 vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0c2x1);
267 vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, va1c2x1);
268 vprod2x0123c2 = vmlal_s8(vprod2x0123c2, vb0123c2x1, va2c2x1);
269 vprod3x0123c2 = vmlal_s8(vprod3x0123c2, vb0123c2x1, va3c2x1);
270 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
271 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
272 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
273 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c2);
274 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0c2x0);
275 int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1c2x0);
276 int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2x0, va2c2x0);
277 int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2x0, va3c2x0);
278 const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
279 vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0c2x1);
280 vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, va1c2x1);
281 vprod2x4567c2 = vmlal_s8(vprod2x4567c2, vb4567c2x1, va2c2x1);
282 vprod3x4567c2 = vmlal_s8(vprod3x4567c2, vb4567c2x1, va3c2x1);
283 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
284 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
285 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
286 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2);
287 int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2x0, va0c2x0);
288 int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2x0, va1c2x0);
289 int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2x0, va2c2x0);
290 int16x8_t vprod3x89ABc2 = vmull_s8(vb89ABc2x0, va3c2x0);
291 const int8x8_t vb89ABc2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
292 vprod0x89ABc2 = vmlal_s8(vprod0x89ABc2, vb89ABc2x1, va0c2x1);
293 vprod1x89ABc2 = vmlal_s8(vprod1x89ABc2, vb89ABc2x1, va1c2x1);
294 vprod2x89ABc2 = vmlal_s8(vprod2x89ABc2, vb89ABc2x1, va2c2x1);
295 vprod3x89ABc2 = vmlal_s8(vprod3x89ABc2, vb89ABc2x1, va3c2x1);
296 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
297 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
298 vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
299 vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc2);
300 int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0c2x0);
301 int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2x0, va1c2x0);
302 int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2x0, va2c2x0);
303 int16x8_t vprod3xCDEFc2 = vmull_s8(vbCDEFc2x0, va3c2x0);
304 const int8x8_t vbCDEFc2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
305 vprod0xCDEFc2 = vmlal_s8(vprod0xCDEFc2, vbCDEFc2x1, va0c2x1);
306 vprod1xCDEFc2 = vmlal_s8(vprod1xCDEFc2, vbCDEFc2x1, va1c2x1);
307 vprod2xCDEFc2 = vmlal_s8(vprod2xCDEFc2, vbCDEFc2x1, va2c2x1);
308 vprod3xCDEFc2 = vmlal_s8(vprod3xCDEFc2, vbCDEFc2x1, va3c2x1);
309 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
310 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
311 vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
312 vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc2);
313 const int8x8_t va0c3x0 = vreinterpret_s8_s16(va03x0);
314 const int8x8_t va0c3x1 = vreinterpret_s8_s16(va03x1);
315 const int8x8_t va1c3x0 = vreinterpret_s8_s16(va13x0);
316 const int8x8_t va1c3x1 = vreinterpret_s8_s16(va13x1);
317 const int8x8_t va2c3x0 = vreinterpret_s8_s16(va23x0);
318 const int8x8_t va2c3x1 = vreinterpret_s8_s16(va23x1);
319 const int8x8_t va3c3x0 = vreinterpret_s8_s16(va33x0);
320 const int8x8_t va3c3x1 = vreinterpret_s8_s16(va33x1);
321
322 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0c3x0);
323 int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1c3x0);
324 int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3x0, va2c3x0);
325 int16x8_t vprod3x0123c3 = vmull_s8(vb0123c3x0, va3c3x0);
326 const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
327 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1);
328 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1);
329 vprod2x0123c3 = vmlal_s8(vprod2x0123c3, vb0123c3x1, va2c3x1);
330 vprod3x0123c3 = vmlal_s8(vprod3x0123c3, vb0123c3x1, va3c3x1);
331 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
332 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
333 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
334 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c3);
335 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0c3x0);
336 int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1c3x0);
337 int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3x0, va2c3x0);
338 int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3c3x0);
339 const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
340 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1);
341 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1);
342 vprod2x4567c3 = vmlal_s8(vprod2x4567c3, vb4567c3x1, va2c3x1);
343 vprod3x4567c3 = vmlal_s8(vprod3x4567c3, vb4567c3x1, va3c3x1);
344 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
345 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
346 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
347 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3);
348 int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3x0, va0c3x0);
349 int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3x0, va1c3x0);
350 int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, va2c3x0);
351 int16x8_t vprod3x89ABc3 = vmull_s8(vb89ABc3x0, va3c3x0);
352 const int8x8_t vb89ABc3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
353 vprod0x89ABc3 = vmlal_s8(vprod0x89ABc3, vb89ABc3x1, va0c3x1);
354 vprod1x89ABc3 = vmlal_s8(vprod1x89ABc3, vb89ABc3x1, va1c3x1);
355 vprod2x89ABc3 = vmlal_s8(vprod2x89ABc3, vb89ABc3x1, va2c3x1);
356 vprod3x89ABc3 = vmlal_s8(vprod3x89ABc3, vb89ABc3x1, va3c3x1);
357 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
358 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
359 vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc3);
360 vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc3);
361 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0c3x0);
362 int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3x0, va1c3x0);
363 int16x8_t vprod2xCDEFc3 = vmull_s8(vbCDEFc3x0, va2c3x0);
364 int16x8_t vprod3xCDEFc3 = vmull_s8(vbCDEFc3x0, va3c3x0);
365 const int8x8_t vbCDEFc3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
366 vprod0xCDEFc3 = vmlal_s8(vprod0xCDEFc3, vbCDEFc3x1, va0c3x1);
367 vprod1xCDEFc3 = vmlal_s8(vprod1xCDEFc3, vbCDEFc3x1, va1c3x1);
368 vprod2xCDEFc3 = vmlal_s8(vprod2xCDEFc3, vbCDEFc3x1, va2c3x1);
369 vprod3xCDEFc3 = vmlal_s8(vprod3xCDEFc3, vbCDEFc3x1, va3c3x1);
370 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
371 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
372 vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc3);
373 vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc3);
374
375 k -= 16 * sizeof(int8_t);
376 }
377
378 if (k >= 8 * sizeof(int8_t)) {
379 const int16x4_t va00 = vld1_dup_s16((const void*)a0);
380 const int16x4_t va01 = vld1_dup_s16((const void*)(a0 + 2));
381 const int16x4_t va02 = vld1_dup_s16((const void*)(a0 + 4));
382 const int16x4_t va03 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
383 const int16x4_t va10 = vld1_dup_s16((const void*)a1);
384 const int16x4_t va11 = vld1_dup_s16((const void*)(a1 + 2));
385 const int16x4_t va12 = vld1_dup_s16((const void*)(a1 + 4));
386 const int16x4_t va13 = vld1_dup_s16((const void*)(a1 + 6)); a1 += 8;
387 const int16x4_t va20 = vld1_dup_s16((const void*)a2);
388 const int16x4_t va21 = vld1_dup_s16((const void*)(a2 + 2));
389 const int16x4_t va22 = vld1_dup_s16((const void*)(a2 + 4));
390 const int16x4_t va23 = vld1_dup_s16((const void*)(a2 + 6)); a2 += 8;
391 const int16x4_t va30 = vld1_dup_s16((const void*)a3);
392 const int16x4_t va31 = vld1_dup_s16((const void*)(a3 + 2));
393 const int16x4_t va32 = vld1_dup_s16((const void*)(a3 + 4));
394 const int16x4_t va33 = vld1_dup_s16((const void*)(a3 + 6)); a3 += 8;
395
396 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
397 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
398 const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
399 const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
400 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
401 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
402 const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
403 const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
404 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
405 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
406 const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
407 const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
408 const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
409 const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
410 const int8x8_t vb89ABc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
411 const int8x8_t vbCDEFc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
412
413 const int8x8_t va0c0 = vreinterpret_s8_s16(va00);
414 const int8x8_t va1c0 = vreinterpret_s8_s16(va10);
415 const int8x8_t va2c0 = vreinterpret_s8_s16(va20);
416 const int8x8_t va3c0 = vreinterpret_s8_s16(va30);
417
418 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
419 const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
420 const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
421 const int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0, va3c0);
422 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
423 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
424 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
425 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c0);
426 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
427 const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
428 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
429 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0);
430 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
431 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
432 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
433 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0);
434 const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
435 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0);
436 const int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0, va2c0);
437 const int16x8_t vprod3x89ABc0 = vmull_s8(vb89ABc0, va3c0);
438 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
439 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
440 vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
441 vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc0);
442 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
443 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0);
444 const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0);
445 const int16x8_t vprod3xCDEFc0 = vmull_s8(vbCDEFc0, va3c0);
446 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
447 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
448 vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
449 vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc0);
450 const int8x8_t va0c1 = vreinterpret_s8_s16(va01);
451 const int8x8_t va1c1 = vreinterpret_s8_s16(va11);
452 const int8x8_t va2c1 = vreinterpret_s8_s16(va21);
453 const int8x8_t va3c1 = vreinterpret_s8_s16(va31);
454
455 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
456 const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
457 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
458 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1);
459 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
460 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
461 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
462 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1);
463 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
464 const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
465 const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
466 const int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1, va3c1);
467 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
468 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
469 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
470 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c1);
471 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
472 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1);
473 const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1);
474 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1);
475 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
476 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
477 vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
478 vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc1);
479 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
480 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1);
481 const int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1, va2c1);
482 const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1);
483 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
484 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
485 vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
486 vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc1);
487 const int8x8_t va0c2 = vreinterpret_s8_s16(va02);
488 const int8x8_t va1c2 = vreinterpret_s8_s16(va12);
489 const int8x8_t va2c2 = vreinterpret_s8_s16(va22);
490 const int8x8_t va3c2 = vreinterpret_s8_s16(va32);
491
492 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
493 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
494 const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
495 const int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2, va3c2);
496 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
497 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
498 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
499 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c2);
500 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
501 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
502 const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
503 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2);
504 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
505 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
506 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
507 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2);
508 const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
509 const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2);
510 const int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2, va2c2);
511 const int16x8_t vprod3x89ABc2 = vmull_s8(vb89ABc2, va3c2);
512 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
513 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
514 vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
515 vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc2);
516 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
517 const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2);
518 const int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2, va2c2);
519 const int16x8_t vprod3xCDEFc2 = vmull_s8(vbCDEFc2, va3c2);
520 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
521 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
522 vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
523 vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc2);
524 const int8x8_t va0c3 = vreinterpret_s8_s16(va03);
525 const int8x8_t va1c3 = vreinterpret_s8_s16(va13);
526 const int8x8_t va2c3 = vreinterpret_s8_s16(va23);
527 const int8x8_t va3c3 = vreinterpret_s8_s16(va33);
528
529 const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
530 const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
531 const int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3, va2c3);
532 const int16x8_t vprod3x0123c3 = vmull_s8(vb0123c3, va3c3);
533 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
534 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
535 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
536 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c3);
537 const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
538 const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
539 const int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3, va2c3);
540 const int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3, va3c3);
541 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
542 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
543 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
544 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3);
545 const int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3, va0c3);
546 const int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3, va1c3);
547 const int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3, va2c3);
548 const int16x8_t vprod3x89ABc3 = vmull_s8(vb89ABc3, va3c3);
549 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
550 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
551 vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc3);
552 vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc3);
553 const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3);
554 const int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3, va1c3);
555 const int16x8_t vprod2xCDEFc3 = vmull_s8(vbCDEFc3, va2c3);
556 const int16x8_t vprod3xCDEFc3 = vmull_s8(vbCDEFc3, va3c3);
557 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
558 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
559 vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc3);
560 vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc3);
561
562 k -= 8 * sizeof(int8_t);
563 }
564
565 if XNN_UNLIKELY(k != 0) {
566 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
567 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
568 const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
569 const int8x8_t va3 = vld1_s8(a3); a3 = (const int8_t*) ((uintptr_t) a3 + k);
570
571 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
572 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
573 const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
574 const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
575
576 const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
577 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
578 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
579 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
580 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
581 const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
582 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
583 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
584 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
585 const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
586 const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
587 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
588 const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
589 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
590 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0);
591 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
592 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0);
593 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
594 const int8x8_t va2c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0));
595 const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
596 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
597 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
598 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
599 const int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0, va2c0);
600 vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
601 const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0);
602 vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
603 const int8x8_t va3c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 0));
604 const int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0, va3c0);
605 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c0);
606 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0);
607 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0);
608 const int16x8_t vprod3x89ABc0 = vmull_s8(vb89ABc0, va3c0);
609 vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc0);
610 const int16x8_t vprod3xCDEFc0 = vmull_s8(vbCDEFc0, va3c0);
611 vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc0);
612
613 if (k > 2 * sizeof(int8_t)) {
614 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
615 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
616 const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
617 const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
618
619 const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
620 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
621 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
622 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
623 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
624 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
625 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
626 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
627 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
628 const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
629 const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
630 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
631 const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
632 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
633 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1);
634 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
635 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1);
636 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
637 const int8x8_t va2c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1));
638 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
639 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
640 const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
641 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
642 const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1);
643 vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
644 const int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1, va2c1);
645 vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
646 const int8x8_t va3c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 1));
647 const int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1, va3c1);
648 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1);
649 const int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1, va3c1);
650 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c1);
651 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1);
652 vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc1);
653 const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1);
654 vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc1);
655
656 if (k > 4 * sizeof(int8_t)) {
657 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
658 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
659 const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
660 const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
661
662 const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
663 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
664 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
665 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
666 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
667 const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
668 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
669 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
670 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
671 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
672 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
673 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
674 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
675 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
676 const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2);
677 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
678 const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2);
679 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
680 const int8x8_t va2c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2));
681 const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
682 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
683 const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
684 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
685 const int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2, va2c2);
686 vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
687 const int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2, va2c2);
688 vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
689 const int8x8_t va3c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va3), 2));
690 const int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2, va3c2);
691 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c2);
692 const int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2, va3c2);
693 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2);
694 const int16x8_t vprod3x89ABc2 = vmull_s8(vb89ABc2, va3c2);
695 vacc3x89AB = vpadalq_s16(vacc3x89AB, vprod3x89ABc2);
696 const int16x8_t vprod3xCDEFc2 = vmull_s8(vbCDEFc2, va3c2);
697 vacc3xCDEF = vpadalq_s16(vacc3xCDEF, vprod3xCDEFc2);
698 }
699 }
700 }
701
702 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
703 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
704 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
705
706 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
707 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
708 vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
709 vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
710 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
711 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
712 vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
713 vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
714 vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
715 vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
716 vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
717 vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
718 vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
719 vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
720 vacc3x89AB = vqshlq_s32(vacc3x89AB, vright_pre_shift);
721 vacc3xCDEF = vqshlq_s32(vacc3xCDEF, vright_pre_shift);
722
723 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
724 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
725 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
726 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
727 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
728 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
729 vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
730 vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
731 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
732 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
733 vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
734 vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
735 vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
736 vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
737 vacc3x89AB = vqdmulhq_s32(vacc3x89AB, vmultiplier);
738 vacc3xCDEF = vqdmulhq_s32(vacc3xCDEF, vmultiplier);
739
740 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
741 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
742 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
743 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
744 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
745 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
746 vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
747 vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
748 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
749 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
750 vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
751 vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
752 vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
753 vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
754 vacc3x89AB = vrshlq_s32(vacc3x89AB, vright_post_shift);
755 vacc3xCDEF = vrshlq_s32(vacc3xCDEF, vright_post_shift);
756
757 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
758 #if XNN_ARCH_ARM64
759 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
760 int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
761 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
762 int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
763 int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
764 int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
765 int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
766 int16x8_t vacc3x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF);
767
768 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
769 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
770 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
771 vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
772 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
773 vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
774 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
775 vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
776
777 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
778 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
779 int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
780 int8x16_t vout3x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc3x01234567), vacc3x89ABCDEF);
781 #else
782 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
783 int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
784 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
785 int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
786 int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
787 int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
788 int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
789 int16x8_t vacc3x89ABCDEF = vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF));
790
791 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
792 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
793 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
794 vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
795 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
796 vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
797 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
798 vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
799
800 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
801 int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
802 int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
803 int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCDEF));
804 #endif
805
806 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
807 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
808 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
809 vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
810 vout3x0123456789ABCDEF = vmaxq_s8(vout3x0123456789ABCDEF, voutput_min);
811
812 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
813 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
814 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
815 vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
816 vout3x0123456789ABCDEF = vminq_s8(vout3x0123456789ABCDEF, voutput_max);
817
818 if (nc >= 16) {
819 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
820 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
821 vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
822 vst1q_s8(c3 + 0, vout3x0123456789ABCDEF);
823
824 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
825 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
826 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
827 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
828
829 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
830 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
831 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
832 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
833
834 nc -= 16;
835 } else {
836 // Final case where not all of the 16 columns fit in the destination.
837 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
838 int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8(vout3x0123456789ABCDEF));
839 if (nc & 8) {
840 vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
841 vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
842 vst1_s8(c2, vget_low_s8(vout2x01234567_3x01234567)); c2 += 8;
843 vst1_s8(c3, vget_high_s8(vout2x01234567_3x01234567)); c3 += 8;
844 vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
845 vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0123456789ABCDEF));
846 }
847 if (nc & 4) {
848 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
849 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
850 vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
851 vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
852 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
853 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
854 }
855 if (nc & 2) {
856 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
857 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
858 vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
859 vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
860 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
861 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
862 }
863 if (nc & 1) {
864 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
865 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
866 vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
867 vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
868 }
869
870 nc = 0;
871 }
872 } while (nc != 0);
873 }
874