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_3x16c2__neon_mlal_ld2r(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_3x16c2__neon_mlal_ld2r(
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 <= 3);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(ks != 0);
36 assert(ks % (3 * 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
53 do {
54 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
55 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
56 int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
57 int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
58 int32x4_t vacc1x0123 = vacc0x0123;
59 int32x4_t vacc1x4567 = vacc0x4567;
60 int32x4_t vacc1x89AB = vacc0x89AB;
61 int32x4_t vacc1xCDEF = vacc0xCDEF;
62 int32x4_t vacc2x0123 = vacc0x0123;
63 int32x4_t vacc2x4567 = vacc0x4567;
64 int32x4_t vacc2x89AB = vacc0x89AB;
65 int32x4_t vacc2xCDEF = vacc0xCDEF;
66
67 size_t p = ks;
68 do {
69 const int8_t* restrict a0 = a[0];
70 if XNN_UNPREDICTABLE(a0 != zero) {
71 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
72 }
73 const int8_t* restrict a1 = a[1];
74 if XNN_UNPREDICTABLE(a1 != zero) {
75 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
76 }
77 const int8_t* restrict a2 = a[2];
78 if XNN_UNPREDICTABLE(a2 != zero) {
79 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
80 }
81 a += 3;
82
83 size_t k = kc;
84
85 while (k >= 16 * sizeof(int8_t)) {
86 const int16x4x2_t va00x0 = vld2_dup_s16((const void*)a0);
87 const int16x4x2_t va01x0 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
88 const int16x4x2_t va00x1 = vld2_dup_s16((const void*)a0);
89 const int16x4x2_t va01x1 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
90 const int16x4x2_t va10x0 = vld2_dup_s16((const void*)a1);
91 const int16x4x2_t va11x0 = vld2_dup_s16((const void*)(a1 + 4)); a1 += 8;
92 const int16x4x2_t va10x1 = vld2_dup_s16((const void*)a1);
93 const int16x4x2_t va11x1 = vld2_dup_s16((const void*)(a1 + 4)); a1 += 8;
94 const int16x4x2_t va20x0 = vld2_dup_s16((const void*)a2);
95 const int16x4x2_t va21x0 = vld2_dup_s16((const void*)(a2 + 4)); a2 += 8;
96 const int16x4x2_t va20x1 = vld2_dup_s16((const void*)a2);
97 const int16x4x2_t va21x1 = vld2_dup_s16((const void*)(a2 + 4)); a2 += 8;
98
99 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
100 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
101 const int8x8_t vb89ABc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
102 const int8x8_t vbCDEFc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
103 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
104 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
105 const int8x8_t vb89ABc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
106 const int8x8_t vbCDEFc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
107 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
108 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
109 const int8x8_t vb89ABc2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
110 const int8x8_t vbCDEFc2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
111 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
112 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
113 const int8x8_t vb89ABc3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
114 const int8x8_t vbCDEFc3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
115
116 const int8x8_t va0c0x0 = vreinterpret_s8_s16(va00x0.val[0]);
117 const int8x8_t va0c0x1 = vreinterpret_s8_s16(va00x1.val[0]);
118 const int8x8_t va1c0x0 = vreinterpret_s8_s16(va10x0.val[0]);
119 const int8x8_t va1c0x1 = vreinterpret_s8_s16(va10x1.val[0]);
120 const int8x8_t va2c0x0 = vreinterpret_s8_s16(va20x0.val[0]);
121 const int8x8_t va2c0x1 = vreinterpret_s8_s16(va20x1.val[0]);
122
123 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0c0x0);
124 int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1c0x0);
125 int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0x0, va2c0x0);
126 const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
127 vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0c0x1);
128 vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, va1c0x1);
129 vprod2x0123c0 = vmlal_s8(vprod2x0123c0, vb0123c0x1, va2c0x1);
130 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
131 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
132 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
133 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0c0x0);
134 int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1c0x0);
135 int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2c0x0);
136 const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
137 vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0c0x1);
138 vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, va1c0x1);
139 vprod2x4567c0 = vmlal_s8(vprod2x4567c0, vb4567c0x1, va2c0x1);
140 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
141 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
142 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
143 int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, va0c0x0);
144 int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1c0x0);
145 int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0x0, va2c0x0);
146 const int8x8_t vb89ABc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
147 vprod0x89ABc0 = vmlal_s8(vprod0x89ABc0, vb89ABc0x1, va0c0x1);
148 vprod1x89ABc0 = vmlal_s8(vprod1x89ABc0, vb89ABc0x1, va1c0x1);
149 vprod2x89ABc0 = vmlal_s8(vprod2x89ABc0, vb89ABc0x1, va2c0x1);
150 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
151 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
152 vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
153 int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0x0, va0c0x0);
154 int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0x0, va1c0x0);
155 int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0x0, va2c0x0);
156 const int8x8_t vbCDEFc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
157 vprod0xCDEFc0 = vmlal_s8(vprod0xCDEFc0, vbCDEFc0x1, va0c0x1);
158 vprod1xCDEFc0 = vmlal_s8(vprod1xCDEFc0, vbCDEFc0x1, va1c0x1);
159 vprod2xCDEFc0 = vmlal_s8(vprod2xCDEFc0, vbCDEFc0x1, va2c0x1);
160 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
161 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
162 vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
163 const int8x8_t va0c1x0 = vreinterpret_s8_s16(va00x0.val[1]);
164 const int8x8_t va0c1x1 = vreinterpret_s8_s16(va00x1.val[1]);
165 const int8x8_t va1c1x0 = vreinterpret_s8_s16(va10x0.val[1]);
166 const int8x8_t va1c1x1 = vreinterpret_s8_s16(va10x1.val[1]);
167 const int8x8_t va2c1x0 = vreinterpret_s8_s16(va20x0.val[1]);
168 const int8x8_t va2c1x1 = vreinterpret_s8_s16(va20x1.val[1]);
169
170 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0c1x0);
171 int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1c1x0);
172 int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2c1x0);
173 const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
174 vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0c1x1);
175 vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, va1c1x1);
176 vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, va2c1x1);
177 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
178 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
179 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
180 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0);
181 int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1c1x0);
182 int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1x0, va2c1x0);
183 const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
184 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1);
185 vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, va1c1x1);
186 vprod2x4567c1 = vmlal_s8(vprod2x4567c1, vb4567c1x1, va2c1x1);
187 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
188 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
189 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
190 int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0c1x0);
191 int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1x0, va1c1x0);
192 int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1x0, va2c1x0);
193 const int8x8_t vb89ABc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
194 vprod0x89ABc1 = vmlal_s8(vprod0x89ABc1, vb89ABc1x1, va0c1x1);
195 vprod1x89ABc1 = vmlal_s8(vprod1x89ABc1, vb89ABc1x1, va1c1x1);
196 vprod2x89ABc1 = vmlal_s8(vprod2x89ABc1, vb89ABc1x1, va2c1x1);
197 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
198 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
199 vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
200 int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0c1x0);
201 int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1c1x0);
202 int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1x0, va2c1x0);
203 const int8x8_t vbCDEFc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
204 vprod0xCDEFc1 = vmlal_s8(vprod0xCDEFc1, vbCDEFc1x1, va0c1x1);
205 vprod1xCDEFc1 = vmlal_s8(vprod1xCDEFc1, vbCDEFc1x1, va1c1x1);
206 vprod2xCDEFc1 = vmlal_s8(vprod2xCDEFc1, vbCDEFc1x1, va2c1x1);
207 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
208 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
209 vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
210 const int8x8_t va0c2x0 = vreinterpret_s8_s16(va01x0.val[0]);
211 const int8x8_t va0c2x1 = vreinterpret_s8_s16(va01x1.val[0]);
212 const int8x8_t va1c2x0 = vreinterpret_s8_s16(va11x0.val[0]);
213 const int8x8_t va1c2x1 = vreinterpret_s8_s16(va11x1.val[0]);
214 const int8x8_t va2c2x0 = vreinterpret_s8_s16(va21x0.val[0]);
215 const int8x8_t va2c2x1 = vreinterpret_s8_s16(va21x1.val[0]);
216
217 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0c2x0);
218 int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1c2x0);
219 int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2x0, va2c2x0);
220 const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
221 vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0c2x1);
222 vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, va1c2x1);
223 vprod2x0123c2 = vmlal_s8(vprod2x0123c2, vb0123c2x1, va2c2x1);
224 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
225 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
226 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
227 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0c2x0);
228 int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1c2x0);
229 int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2x0, va2c2x0);
230 const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
231 vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0c2x1);
232 vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, va1c2x1);
233 vprod2x4567c2 = vmlal_s8(vprod2x4567c2, vb4567c2x1, va2c2x1);
234 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
235 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
236 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
237 int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2x0, va0c2x0);
238 int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2x0, va1c2x0);
239 int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2x0, va2c2x0);
240 const int8x8_t vb89ABc2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
241 vprod0x89ABc2 = vmlal_s8(vprod0x89ABc2, vb89ABc2x1, va0c2x1);
242 vprod1x89ABc2 = vmlal_s8(vprod1x89ABc2, vb89ABc2x1, va1c2x1);
243 vprod2x89ABc2 = vmlal_s8(vprod2x89ABc2, vb89ABc2x1, va2c2x1);
244 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
245 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
246 vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
247 int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0c2x0);
248 int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2x0, va1c2x0);
249 int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2x0, va2c2x0);
250 const int8x8_t vbCDEFc2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
251 vprod0xCDEFc2 = vmlal_s8(vprod0xCDEFc2, vbCDEFc2x1, va0c2x1);
252 vprod1xCDEFc2 = vmlal_s8(vprod1xCDEFc2, vbCDEFc2x1, va1c2x1);
253 vprod2xCDEFc2 = vmlal_s8(vprod2xCDEFc2, vbCDEFc2x1, va2c2x1);
254 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
255 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
256 vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
257 const int8x8_t va0c3x0 = vreinterpret_s8_s16(va01x0.val[1]);
258 const int8x8_t va0c3x1 = vreinterpret_s8_s16(va01x1.val[1]);
259 const int8x8_t va1c3x0 = vreinterpret_s8_s16(va11x0.val[1]);
260 const int8x8_t va1c3x1 = vreinterpret_s8_s16(va11x1.val[1]);
261 const int8x8_t va2c3x0 = vreinterpret_s8_s16(va21x0.val[1]);
262 const int8x8_t va2c3x1 = vreinterpret_s8_s16(va21x1.val[1]);
263
264 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0c3x0);
265 int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1c3x0);
266 int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3x0, va2c3x0);
267 const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
268 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1);
269 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1);
270 vprod2x0123c3 = vmlal_s8(vprod2x0123c3, vb0123c3x1, va2c3x1);
271 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
272 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
273 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
274 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0c3x0);
275 int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1c3x0);
276 int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3x0, va2c3x0);
277 const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
278 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1);
279 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1);
280 vprod2x4567c3 = vmlal_s8(vprod2x4567c3, vb4567c3x1, va2c3x1);
281 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
282 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
283 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
284 int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3x0, va0c3x0);
285 int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3x0, va1c3x0);
286 int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, va2c3x0);
287 const int8x8_t vb89ABc3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
288 vprod0x89ABc3 = vmlal_s8(vprod0x89ABc3, vb89ABc3x1, va0c3x1);
289 vprod1x89ABc3 = vmlal_s8(vprod1x89ABc3, vb89ABc3x1, va1c3x1);
290 vprod2x89ABc3 = vmlal_s8(vprod2x89ABc3, vb89ABc3x1, va2c3x1);
291 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
292 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
293 vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc3);
294 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0c3x0);
295 int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3x0, va1c3x0);
296 int16x8_t vprod2xCDEFc3 = vmull_s8(vbCDEFc3x0, va2c3x0);
297 const int8x8_t vbCDEFc3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
298 vprod0xCDEFc3 = vmlal_s8(vprod0xCDEFc3, vbCDEFc3x1, va0c3x1);
299 vprod1xCDEFc3 = vmlal_s8(vprod1xCDEFc3, vbCDEFc3x1, va1c3x1);
300 vprod2xCDEFc3 = vmlal_s8(vprod2xCDEFc3, vbCDEFc3x1, va2c3x1);
301 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
302 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
303 vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc3);
304
305 k -= 16 * sizeof(int8_t);
306 }
307
308 if (k >= 8 * sizeof(int8_t)) {
309 const int16x4x2_t va00 = vld2_dup_s16((const void*)a0);
310 const int16x4x2_t va01 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
311 const int16x4x2_t va10 = vld2_dup_s16((const void*)a1);
312 const int16x4x2_t va11 = vld2_dup_s16((const void*)(a1 + 4)); a1 += 8;
313 const int16x4x2_t va20 = vld2_dup_s16((const void*)a2);
314 const int16x4x2_t va21 = vld2_dup_s16((const void*)(a2 + 4)); a2 += 8;
315
316 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
317 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
318 const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
319 const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
320 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
321 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
322 const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
323 const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
324 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
325 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
326 const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
327 const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
328 const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
329 const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
330 const int8x8_t vb89ABc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
331 const int8x8_t vbCDEFc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
332
333 const int8x8_t va0c0 = vreinterpret_s8_s16(va00.val[0]);
334 const int8x8_t va1c0 = vreinterpret_s8_s16(va10.val[0]);
335 const int8x8_t va2c0 = vreinterpret_s8_s16(va20.val[0]);
336
337 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
338 const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
339 const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
340 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
341 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
342 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
343 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
344 const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
345 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
346 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
347 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
348 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
349 const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
350 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0);
351 const int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0, va2c0);
352 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
353 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
354 vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
355 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
356 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0);
357 const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0);
358 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
359 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
360 vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
361 const int8x8_t va0c1 = vreinterpret_s8_s16(va00.val[1]);
362 const int8x8_t va1c1 = vreinterpret_s8_s16(va10.val[1]);
363 const int8x8_t va2c1 = vreinterpret_s8_s16(va20.val[1]);
364
365 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
366 const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
367 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
368 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
369 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
370 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
371 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
372 const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
373 const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
374 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
375 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
376 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
377 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
378 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1);
379 const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1);
380 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
381 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
382 vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
383 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
384 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1);
385 const int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1, va2c1);
386 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
387 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
388 vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
389 const int8x8_t va0c2 = vreinterpret_s8_s16(va01.val[0]);
390 const int8x8_t va1c2 = vreinterpret_s8_s16(va11.val[0]);
391 const int8x8_t va2c2 = vreinterpret_s8_s16(va21.val[0]);
392
393 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
394 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
395 const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
396 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
397 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
398 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
399 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
400 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
401 const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
402 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
403 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
404 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
405 const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
406 const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2);
407 const int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2, va2c2);
408 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
409 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
410 vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
411 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
412 const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2);
413 const int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2, va2c2);
414 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
415 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
416 vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
417 const int8x8_t va0c3 = vreinterpret_s8_s16(va01.val[1]);
418 const int8x8_t va1c3 = vreinterpret_s8_s16(va11.val[1]);
419 const int8x8_t va2c3 = vreinterpret_s8_s16(va21.val[1]);
420
421 const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
422 const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
423 const int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3, va2c3);
424 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
425 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
426 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
427 const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
428 const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
429 const int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3, va2c3);
430 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
431 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
432 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
433 const int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3, va0c3);
434 const int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3, va1c3);
435 const int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3, va2c3);
436 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
437 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
438 vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc3);
439 const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3);
440 const int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3, va1c3);
441 const int16x8_t vprod2xCDEFc3 = vmull_s8(vbCDEFc3, va2c3);
442 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
443 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
444 vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc3);
445
446 k -= 8 * sizeof(int8_t);
447 }
448
449 if XNN_UNLIKELY(k != 0) {
450 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
451 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
452 const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
453
454 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
455 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
456 const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
457 const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
458
459 const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
460 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
461 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
462 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
463 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
464 const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
465 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
466 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
467 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
468 const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
469 const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
470 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
471 const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
472 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
473 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0);
474 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
475 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0);
476 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
477 const int8x8_t va2c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0));
478 const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
479 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
480 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
481 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
482 const int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0, va2c0);
483 vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
484 const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0);
485 vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
486
487 if (k > 2 * sizeof(int8_t)) {
488 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
489 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
490 const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
491 const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
492
493 const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
494 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
495 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
496 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
497 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
498 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
499 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
500 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
501 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
502 const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
503 const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
504 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
505 const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
506 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
507 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1);
508 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
509 const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1);
510 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
511 const int8x8_t va2c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1));
512 const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
513 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
514 const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
515 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
516 const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1);
517 vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
518 const int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1, va2c1);
519 vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
520
521 if (k > 4 * sizeof(int8_t)) {
522 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
523 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
524 const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
525 const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
526
527 const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
528 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
529 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
530 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
531 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
532 const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
533 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
534 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
535 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
536 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
537 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
538 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
539 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
540 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
541 const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2);
542 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
543 const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2);
544 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
545 const int8x8_t va2c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2));
546 const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
547 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
548 const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
549 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
550 const int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2, va2c2);
551 vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
552 const int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2, va2c2);
553 vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
554 }
555 }
556 }
557 p -= 3 * sizeof(void*);
558 } while (p != 0);
559
560 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
561 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
562 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
563
564 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
565 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
566 vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
567 vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
568 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
569 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
570 vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
571 vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
572 vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
573 vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
574 vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
575 vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
576
577 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
578 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
579 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
580 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
581 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
582 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
583 vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
584 vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
585 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
586 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
587 vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
588 vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
589
590 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
591 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
592 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
593 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
594 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
595 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
596 vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
597 vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
598 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
599 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
600 vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
601 vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
602
603 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
604 #if XNN_ARCH_ARM64
605 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
606 int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
607 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
608 int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
609 int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
610 int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
611
612 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
613 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
614 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
615 vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
616 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
617 vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
618
619 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
620 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
621 int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
622 #else
623 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
624 int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
625 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
626 int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
627 int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
628 int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
629
630 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
631 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
632 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
633 vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
634 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
635 vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
636
637 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
638 int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
639 int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
640 #endif
641
642 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
643 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
644 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
645 vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
646
647 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
648 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
649 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
650 vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
651
652 if (nc >= 16) {
653 vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
654 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
655 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
656
657 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
658 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
659 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
660
661 a = (const int8_t**restrict) ((uintptr_t) a - ks);
662
663 nc -= 16;
664 } else {
665 int8x8_t vout2x01234567 = vget_low_s8(vout2x0123456789ABCDEF);
666 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
667 if (nc & 8) {
668 vst1_s8(c2, vout2x01234567); c2 += 8;
669 vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
670 vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
671 vout2x01234567 = vget_high_s8(vout2x0123456789ABCDEF);
672 vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
673 }
674 if (nc & 4) {
675 vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
676 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
677 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
678 vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
679 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
680 }
681 if (nc & 2) {
682 vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
683 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
684 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
685 vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
686 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
687 }
688 if (nc & 1) {
689 vst1_lane_s8(c2, vout2x01234567, 0);
690 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
691 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
692 }
693
694 nc = 0;
695 }
696 } while (nc != 0);
697 }
698