1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal(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)])18 void xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const int8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(mr != 0);
31 assert(mr <= 4);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(int8_t) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38
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 kc = round_up_po2(kc, 8 * sizeof(int8_t));
61 do {
62 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
63 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
64 int32x4_t vacc1x0123 = vacc0x0123;
65 int32x4_t vacc1x4567 = vacc0x4567;
66 int32x4_t vacc2x0123 = vacc0x0123;
67 int32x4_t vacc2x4567 = vacc0x4567;
68 int32x4_t vacc3x0123 = vacc0x0123;
69 int32x4_t vacc3x4567 = vacc0x4567;
70
71 size_t k = kc;
72 while (k >= 16 * sizeof(int8_t)) {
73 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
74 int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
75 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
76 int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
77 int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
78 int8x8_t va2x1 = vld1_s8(a2); a2 += 8;
79 int8x8_t va3x0 = vld1_s8(a3); a3 += 8;
80 int8x8_t va3x1 = vld1_s8(a3); a3 += 8;
81
82 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
83 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
84 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
85 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
86 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
87 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
88 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
89 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
90
91 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
92 int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1x0);
93 int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0x0, va2x0);
94 int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0x0, va3x0);
95 const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
96 vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0x1);
97 vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, va1x1);
98 vprod2x0123c0 = vmlal_s8(vprod2x0123c0, vb0123c0x1, va2x1);
99 vprod3x0123c0 = vmlal_s8(vprod3x0123c0, vb0123c0x1, va3x1);
100 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
101 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
102 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
103 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c0);
104 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
105 int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1x0);
106 int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2x0);
107 int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0x0, va3x0);
108 const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
109 vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0x1);
110 vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, va1x1);
111 vprod2x4567c0 = vmlal_s8(vprod2x4567c0, vb4567c0x1, va2x1);
112 vprod3x4567c0 = vmlal_s8(vprod3x4567c0, vb4567c0x1, va3x1);
113 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
114 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
115 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
116 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0);
117 va0x0 = vext_s8(va0x0, va0x0, 2);
118 va0x1 = vext_s8(va0x1, va0x1, 2);
119 va1x0 = vext_s8(va1x0, va1x0, 2);
120 va1x1 = vext_s8(va1x1, va1x1, 2);
121 va2x0 = vext_s8(va2x0, va2x0, 2);
122 va2x1 = vext_s8(va2x1, va2x1, 2);
123 va3x0 = vext_s8(va3x0, va3x0, 2);
124 va3x1 = vext_s8(va3x1, va3x1, 2);
125 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
126 int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1x0);
127 int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2x0);
128 int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1x0, va3x0);
129 const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
130 vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0x1);
131 vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, va1x1);
132 vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, va2x1);
133 vprod3x0123c1 = vmlal_s8(vprod3x0123c1, vb0123c1x1, va3x1);
134 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
135 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
136 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
137 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1);
138 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
139 int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1x0);
140 int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1x0, va2x0);
141 int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1x0, va3x0);
142 const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
143 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0x1);
144 vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, va1x1);
145 vprod2x4567c1 = vmlal_s8(vprod2x4567c1, vb4567c1x1, va2x1);
146 vprod3x4567c1 = vmlal_s8(vprod3x4567c1, vb4567c1x1, va3x1);
147 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
148 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
149 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
150 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c1);
151 va0x0 = vext_s8(va0x0, va0x0, 2);
152 va0x1 = vext_s8(va0x1, va0x1, 2);
153 va1x0 = vext_s8(va1x0, va1x0, 2);
154 va1x1 = vext_s8(va1x1, va1x1, 2);
155 va2x0 = vext_s8(va2x0, va2x0, 2);
156 va2x1 = vext_s8(va2x1, va2x1, 2);
157 va3x0 = vext_s8(va3x0, va3x0, 2);
158 va3x1 = vext_s8(va3x1, va3x1, 2);
159 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
160 int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1x0);
161 int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2x0, va2x0);
162 int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2x0, va3x0);
163 const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
164 vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0x1);
165 vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, va1x1);
166 vprod2x0123c2 = vmlal_s8(vprod2x0123c2, vb0123c2x1, va2x1);
167 vprod3x0123c2 = vmlal_s8(vprod3x0123c2, vb0123c2x1, va3x1);
168 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
169 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
170 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
171 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c2);
172 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
173 int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1x0);
174 int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2x0, va2x0);
175 int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2x0, va3x0);
176 const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
177 vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0x1);
178 vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, va1x1);
179 vprod2x4567c2 = vmlal_s8(vprod2x4567c2, vb4567c2x1, va2x1);
180 vprod3x4567c2 = vmlal_s8(vprod3x4567c2, vb4567c2x1, va3x1);
181 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
182 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
183 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
184 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2);
185 va0x0 = vext_s8(va0x0, va0x0, 2);
186 va0x1 = vext_s8(va0x1, va0x1, 2);
187 va1x0 = vext_s8(va1x0, va1x0, 2);
188 va1x1 = vext_s8(va1x1, va1x1, 2);
189 va2x0 = vext_s8(va2x0, va2x0, 2);
190 va2x1 = vext_s8(va2x1, va2x1, 2);
191 va3x0 = vext_s8(va3x0, va3x0, 2);
192 va3x1 = vext_s8(va3x1, va3x1, 2);
193 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
194 int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1x0);
195 int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3x0, va2x0);
196 int16x8_t vprod3x0123c3 = vmull_s8(vb0123c3x0, va3x0);
197 const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
198 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0x1);
199 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1x1);
200 vprod2x0123c3 = vmlal_s8(vprod2x0123c3, vb0123c3x1, va2x1);
201 vprod3x0123c3 = vmlal_s8(vprod3x0123c3, vb0123c3x1, va3x1);
202 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
203 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
204 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
205 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c3);
206 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
207 int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1x0);
208 int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3x0, va2x0);
209 int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3x0);
210 const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
211 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0x1);
212 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1x1);
213 vprod2x4567c3 = vmlal_s8(vprod2x4567c3, vb4567c3x1, va2x1);
214 vprod3x4567c3 = vmlal_s8(vprod3x4567c3, vb4567c3x1, va3x1);
215 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
216 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
217 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
218 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3);
219
220 k -= 16 * sizeof(int8_t);
221 }
222 if (k != 0) {
223 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
224 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
225 int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
226 int8x8_t va3x0 = vld1_s8(a3); a3 += 8;
227
228 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
229 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
230 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
231 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
232 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
233 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
234 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
235 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
236
237 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
238 int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1x0);
239 int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0x0, va2x0);
240 int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0x0, va3x0);
241 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
242 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
243 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
244 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c0);
245 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
246 int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1x0);
247 int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2x0);
248 int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0x0, va3x0);
249 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
250 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
251 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
252 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c0);
253 va0x0 = vext_s8(va0x0, va0x0, 2);
254 va1x0 = vext_s8(va1x0, va1x0, 2);
255 va2x0 = vext_s8(va2x0, va2x0, 2);
256 va3x0 = vext_s8(va3x0, va3x0, 2);
257 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
258 int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1x0);
259 int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2x0);
260 int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1x0, va3x0);
261 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
262 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
263 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
264 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c1);
265 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
266 int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1x0);
267 int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1x0, va2x0);
268 int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1x0, va3x0);
269 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
270 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
271 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
272 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c1);
273 va0x0 = vext_s8(va0x0, va0x0, 2);
274 va1x0 = vext_s8(va1x0, va1x0, 2);
275 va2x0 = vext_s8(va2x0, va2x0, 2);
276 va3x0 = vext_s8(va3x0, va3x0, 2);
277 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
278 int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1x0);
279 int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2x0, va2x0);
280 int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2x0, va3x0);
281 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
282 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
283 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
284 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c2);
285 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
286 int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1x0);
287 int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2x0, va2x0);
288 int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2x0, va3x0);
289 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
290 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
291 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
292 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c2);
293 va0x0 = vext_s8(va0x0, va0x0, 2);
294 va1x0 = vext_s8(va1x0, va1x0, 2);
295 va2x0 = vext_s8(va2x0, va2x0, 2);
296 va3x0 = vext_s8(va3x0, va3x0, 2);
297 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
298 int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1x0);
299 int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3x0, va2x0);
300 int16x8_t vprod3x0123c3 = vmull_s8(vb0123c3x0, va3x0);
301 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
302 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
303 vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
304 vacc3x0123 = vpadalq_s16(vacc3x0123, vprod3x0123c3);
305 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
306 int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1x0);
307 int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3x0, va2x0);
308 int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3x0);
309 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
310 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
311 vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
312 vacc3x4567 = vpadalq_s16(vacc3x4567, vprod3x4567c3);
313
314 }
315
316 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
317 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
318 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
319
320 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
321 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
322 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
323 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
324 vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
325 vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
326 vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
327 vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
328
329 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
330 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
331 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
332 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
333 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
334 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
335 vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
336 vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
337
338 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
339 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
340 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
341 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
342 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
343 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
344 vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
345 vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
346
347 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
348 #if XNN_ARCH_ARM64
349 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
350 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
351 int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
352 int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
353
354 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
355 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
356 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
357 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
358
359 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
360 int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567);
361 #else
362 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
363 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
364 int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
365 int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
366
367 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
368 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
369 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
370 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
371
372 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
373 int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567));
374 #endif
375
376 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
377 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
378 vout2x01234567_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, voutput_min);
379
380 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
381 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
382 vout2x01234567_3x01234567 = vminq_s8(vout2x01234567_3x01234567, voutput_max);
383
384 if (nc >= 8) {
385 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
386 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
387 vst1_s8(c2 + 0, vget_low_s8(vout2x01234567_3x01234567));
388 vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567));
389
390 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
391 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
392 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
393 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
394
395 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
396 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
397 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
398 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
399
400 nc -= 8;
401 } else {
402 // Final case where not all of the 8 columns fit in the destination.
403 if (nc & 4) {
404 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
405 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
406 vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
407 vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
408 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
409 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
410 }
411 if (nc & 2) {
412 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
413 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
414 vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
415 vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
416 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
417 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
418 }
419 if (nc & 1) {
420 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
421 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
422 vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
423 vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
424 }
425
426 nc = 0;
427 }
428 } while (nc != 0);
429 }
430