1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c2-neon-mull-shuffle.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <arm_neon.h>
13
14 #include <xnnpack/gemm.h>
15 #include <xnnpack/math.h>
16
17
xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mlal(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mlal(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const int8_t** restrict a,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 2);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (2 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 int8_t* c0 = c;
44 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr != 2) {
46 c1 = c0;
47 }
48
49 kc = round_up_po2(kc, 8 * sizeof(int8_t));
50 do {
51 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
52 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
53 int32x4_t vacc0x89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
54 int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
55 int32x4_t vacc1x0123 = vacc0x0123;
56 int32x4_t vacc1x4567 = vacc0x4567;
57 int32x4_t vacc1x89AB = vacc0x89AB;
58 int32x4_t vacc1xCDEF = vacc0xCDEF;
59
60 size_t p = ks;
61 do {
62 const int8_t* restrict a0 = a[0];
63 if XNN_UNPREDICTABLE(a0 != zero) {
64 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
65 }
66 const int8_t* restrict a1 = a[1];
67 if XNN_UNPREDICTABLE(a1 != zero) {
68 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
69 }
70 a += 2;
71
72 size_t k = kc;
73 while (k >= 16 * sizeof(int8_t)) {
74 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
75 int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
76 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
77 int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
78
79 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
80 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
81 const int8x8_t vb89ABc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
82 const int8x8_t vbCDEFc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
83 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
84 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
85 const int8x8_t vb89ABc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
86 const int8x8_t vbCDEFc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
87 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
88 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
89 const int8x8_t vb89ABc2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
90 const int8x8_t vbCDEFc2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
91 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
92 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
93 const int8x8_t vb89ABc3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
94 const int8x8_t vbCDEFc3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
95
96 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
97 int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1x0);
98 const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
99 vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0x1);
100 vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, va1x1);
101 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
102 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
103 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
104 int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1x0);
105 const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
106 vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0x1);
107 vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, va1x1);
108 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
109 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
110 int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, va0x0);
111 int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1x0);
112 const int8x8_t vb89ABc0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
113 vprod0x89ABc0 = vmlal_s8(vprod0x89ABc0, vb89ABc0x1, va0x1);
114 vprod1x89ABc0 = vmlal_s8(vprod1x89ABc0, vb89ABc0x1, va1x1);
115 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
116 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
117 int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0x0, va0x0);
118 int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0x0, va1x0);
119 const int8x8_t vbCDEFc0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
120 vprod0xCDEFc0 = vmlal_s8(vprod0xCDEFc0, vbCDEFc0x1, va0x1);
121 vprod1xCDEFc0 = vmlal_s8(vprod1xCDEFc0, vbCDEFc0x1, va1x1);
122 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
123 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
124 va0x0 = vext_s8(va0x0, va0x0, 2);
125 va0x1 = vext_s8(va0x1, va0x1, 2);
126 va1x0 = vext_s8(va1x0, va1x0, 2);
127 va1x1 = vext_s8(va1x1, va1x1, 2);
128 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
129 int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1x0);
130 const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
131 vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0x1);
132 vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, va1x1);
133 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
134 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
135 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
136 int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1x0);
137 const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
138 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0x1);
139 vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, va1x1);
140 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
141 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
142 int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0x0);
143 int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1x0, va1x0);
144 const int8x8_t vb89ABc1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
145 vprod0x89ABc1 = vmlal_s8(vprod0x89ABc1, vb89ABc1x1, va0x1);
146 vprod1x89ABc1 = vmlal_s8(vprod1x89ABc1, vb89ABc1x1, va1x1);
147 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
148 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
149 int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0x0);
150 int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1x0);
151 const int8x8_t vbCDEFc1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
152 vprod0xCDEFc1 = vmlal_s8(vprod0xCDEFc1, vbCDEFc1x1, va0x1);
153 vprod1xCDEFc1 = vmlal_s8(vprod1xCDEFc1, vbCDEFc1x1, va1x1);
154 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
155 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
156 va0x0 = vext_s8(va0x0, va0x0, 2);
157 va0x1 = vext_s8(va0x1, va0x1, 2);
158 va1x0 = vext_s8(va1x0, va1x0, 2);
159 va1x1 = vext_s8(va1x1, va1x1, 2);
160 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
161 int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1x0);
162 const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
163 vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0x1);
164 vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, va1x1);
165 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
166 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
167 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
168 int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1x0);
169 const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
170 vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0x1);
171 vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, va1x1);
172 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
173 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
174 int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2x0, va0x0);
175 int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2x0, va1x0);
176 const int8x8_t vb89ABc2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
177 vprod0x89ABc2 = vmlal_s8(vprod0x89ABc2, vb89ABc2x1, va0x1);
178 vprod1x89ABc2 = vmlal_s8(vprod1x89ABc2, vb89ABc2x1, va1x1);
179 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
180 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
181 int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0x0);
182 int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2x0, va1x0);
183 const int8x8_t vbCDEFc2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
184 vprod0xCDEFc2 = vmlal_s8(vprod0xCDEFc2, vbCDEFc2x1, va0x1);
185 vprod1xCDEFc2 = vmlal_s8(vprod1xCDEFc2, vbCDEFc2x1, va1x1);
186 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
187 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
188 va0x0 = vext_s8(va0x0, va0x0, 2);
189 va0x1 = vext_s8(va0x1, va0x1, 2);
190 va1x0 = vext_s8(va1x0, va1x0, 2);
191 va1x1 = vext_s8(va1x1, va1x1, 2);
192 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
193 int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1x0);
194 const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
195 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0x1);
196 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1x1);
197 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
198 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
199 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
200 int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1x0);
201 const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
202 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0x1);
203 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1x1);
204 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
205 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
206 int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3x0, va0x0);
207 int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3x0, va1x0);
208 const int8x8_t vb89ABc3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
209 vprod0x89ABc3 = vmlal_s8(vprod0x89ABc3, vb89ABc3x1, va0x1);
210 vprod1x89ABc3 = vmlal_s8(vprod1x89ABc3, vb89ABc3x1, va1x1);
211 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
212 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
213 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0x0);
214 int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3x0, va1x0);
215 const int8x8_t vbCDEFc3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
216 vprod0xCDEFc3 = vmlal_s8(vprod0xCDEFc3, vbCDEFc3x1, va0x1);
217 vprod1xCDEFc3 = vmlal_s8(vprod1xCDEFc3, vbCDEFc3x1, va1x1);
218 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
219 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
220
221 k -= 16 * sizeof(int8_t);
222 }
223 if (k != 0) {
224 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
225 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
226
227 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
228 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
229 const int8x8_t vb89ABc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
230 const int8x8_t vbCDEFc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
231 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
232 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
233 const int8x8_t vb89ABc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
234 const int8x8_t vbCDEFc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
235 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
236 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
237 const int8x8_t vb89ABc2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
238 const int8x8_t vbCDEFc2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
239 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
240 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
241 const int8x8_t vb89ABc3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
242 const int8x8_t vbCDEFc3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
243
244 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
245 int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1x0);
246 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
247 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
248 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
249 int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1x0);
250 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
251 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
252 int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, va0x0);
253 int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1x0);
254 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
255 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
256 int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0x0, va0x0);
257 int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0x0, va1x0);
258 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
259 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
260 va0x0 = vext_s8(va0x0, va0x0, 2);
261 va1x0 = vext_s8(va1x0, va1x0, 2);
262 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
263 int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1x0);
264 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
265 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
266 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
267 int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1x0);
268 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
269 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
270 int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0x0);
271 int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1x0, va1x0);
272 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
273 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
274 int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0x0);
275 int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1x0);
276 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
277 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
278 va0x0 = vext_s8(va0x0, va0x0, 2);
279 va1x0 = vext_s8(va1x0, va1x0, 2);
280 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
281 int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1x0);
282 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
283 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
284 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
285 int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1x0);
286 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
287 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
288 int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2x0, va0x0);
289 int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2x0, va1x0);
290 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
291 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
292 int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0x0);
293 int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2x0, va1x0);
294 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
295 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
296 va0x0 = vext_s8(va0x0, va0x0, 2);
297 va1x0 = vext_s8(va1x0, va1x0, 2);
298 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
299 int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1x0);
300 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
301 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
302 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
303 int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1x0);
304 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
305 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
306 int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3x0, va0x0);
307 int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3x0, va1x0);
308 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
309 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
310 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0x0);
311 int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3x0, va1x0);
312 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
313 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
314
315 }
316
317 p -= 2 * sizeof(void*);
318 } while (p != 0);
319
320 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
321 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
322 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
323
324 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
325 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
326 vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
327 vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
328 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
329 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
330 vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
331 vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
332
333 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
334 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
335 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
336 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
337 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
338 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
339 vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
340 vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
341
342 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
343 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
344 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
345 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
346 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
347 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
348 vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
349 vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
350
351 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
352 #if XNN_ARCH_ARM64
353 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
354 int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
355 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
356 int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
357
358 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
359 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
360 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
361 vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
362
363 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
364 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
365 #else
366 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
367 int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
368 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
369 int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
370
371 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
372 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
373 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
374 vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
375
376 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
377 int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
378 #endif
379
380 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
381 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
382 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
383
384 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
385 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
386 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
387
388 if (nc >= 16) {
389 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
390 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
391
392 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
393 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
394
395 a = (const int8_t**restrict) ((uintptr_t) a - ks);
396
397 nc -= 16;
398 } else {
399 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
400 if (nc & 8) {
401 vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
402 vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
403 vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
404 }
405 if (nc & 4) {
406 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
407 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
408 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
409 }
410 if (nc & 2) {
411 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
412 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
413 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
414 }
415 if (nc & 1) {
416 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
417 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
418 }
419
420 nc = 0;
421 }
422 } while (nc != 0);
423 }
424