1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c8-neon-mull.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/igemm.h>
15 #include <xnnpack/math.h>
16
17
xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__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_4x8c8__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 <= 4);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (4 * 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 kc = round_up_po2(kc, 8 * sizeof(int8_t));
44 int8_t* c0 = c;
45 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr < 2) {
47 c1 = c0;
48 }
49 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 c2 = c1;
52 }
53 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
54 if XNN_UNPREDICTABLE(mr != 4) {
55 c3 = c2;
56 }
57
58 do {
59 int32x4_t vacc0x0 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
60 int32x4_t vacc0x1 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
61 int32x4_t vacc0x2 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
62 int32x4_t vacc0x3 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
63 int32x4_t vacc0x4 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
64 int32x4_t vacc0x5 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
65 int32x4_t vacc0x6 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
66 int32x4_t vacc0x7 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
67 int32x4_t vacc1x0 = vacc0x0;
68 int32x4_t vacc1x1 = vacc0x1;
69 int32x4_t vacc1x2 = vacc0x2;
70 int32x4_t vacc1x3 = vacc0x3;
71 int32x4_t vacc1x4 = vacc0x4;
72 int32x4_t vacc1x5 = vacc0x5;
73 int32x4_t vacc1x6 = vacc0x6;
74 int32x4_t vacc1x7 = vacc0x7;
75 int32x4_t vacc2x0 = vacc0x0;
76 int32x4_t vacc2x1 = vacc0x1;
77 int32x4_t vacc2x2 = vacc0x2;
78 int32x4_t vacc2x3 = vacc0x3;
79 int32x4_t vacc2x4 = vacc0x4;
80 int32x4_t vacc2x5 = vacc0x5;
81 int32x4_t vacc2x6 = vacc0x6;
82 int32x4_t vacc2x7 = vacc0x7;
83 int32x4_t vacc3x0 = vacc0x0;
84 int32x4_t vacc3x1 = vacc0x1;
85 int32x4_t vacc3x2 = vacc0x2;
86 int32x4_t vacc3x3 = vacc0x3;
87 int32x4_t vacc3x4 = vacc0x4;
88 int32x4_t vacc3x5 = vacc0x5;
89 int32x4_t vacc3x6 = vacc0x6;
90 int32x4_t vacc3x7 = vacc0x7;
91
92 size_t p = ks;
93 do {
94 const int8_t* restrict a0 = a[0];
95 if XNN_UNPREDICTABLE(a0 != zero) {
96 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
97 }
98 const int8_t* restrict a1 = a[1];
99 if XNN_UNPREDICTABLE(a1 != zero) {
100 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
101 }
102 const int8_t* restrict a2 = a[2];
103 if XNN_UNPREDICTABLE(a2 != zero) {
104 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
105 }
106 const int8_t* restrict a3 = a[3];
107 if XNN_UNPREDICTABLE(a3 != zero) {
108 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
109 }
110 a += 4;
111
112 size_t k = kc;
113 // 2x partial unrolled loop to load 16 bytes at a time using MLA.
114 while (k >= 16 * sizeof(int8_t)) {
115 const int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
116 const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
117 const int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
118 const int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
119 const int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
120 const int8x8_t va2x1 = vld1_s8(a2); a2 += 8;
121 const int8x8_t va3x0 = vld1_s8(a3); a3 += 8;
122 const int8x8_t va3x1 = vld1_s8(a3); a3 += 8;
123
124 const int8x8_t vb0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
125 const int8x8_t vb1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
126 const int8x8_t vb2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
127 const int8x8_t vb3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
128 const int8x8_t vb4x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
129 const int8x8_t vb5x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
130 const int8x8_t vb6x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
131 const int8x8_t vb7x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
132
133 const int8x8_t vb0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
134 int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0);
135 int16x8_t vprod1x0 = vmull_s8(vb0x0, va1x0);
136 int16x8_t vprod2x0 = vmull_s8(vb0x0, va2x0);
137 int16x8_t vprod3x0 = vmull_s8(vb0x0, va3x0);
138 vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1);
139 vprod1x0 = vmlal_s8(vprod1x0, vb0x1, va1x1);
140 vprod2x0 = vmlal_s8(vprod2x0, vb0x1, va2x1);
141 vprod3x0 = vmlal_s8(vprod3x0, vb0x1, va3x1);
142 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
143 vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
144 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0);
145 vacc3x0 = vpadalq_s16(vacc3x0, vprod3x0);
146 const int8x8_t vb1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
147 int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0);
148 int16x8_t vprod1x1 = vmull_s8(vb1x0, va1x0);
149 int16x8_t vprod2x1 = vmull_s8(vb1x0, va2x0);
150 int16x8_t vprod3x1 = vmull_s8(vb1x0, va3x0);
151 vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1);
152 vprod1x1 = vmlal_s8(vprod1x1, vb1x1, va1x1);
153 vprod2x1 = vmlal_s8(vprod2x1, vb1x1, va2x1);
154 vprod3x1 = vmlal_s8(vprod3x1, vb1x1, va3x1);
155 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
156 vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
157 vacc2x1 = vpadalq_s16(vacc2x1, vprod2x1);
158 vacc3x1 = vpadalq_s16(vacc3x1, vprod3x1);
159 const int8x8_t vb2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
160 int16x8_t vprod0x2 = vmull_s8(vb2x0, va0x0);
161 int16x8_t vprod1x2 = vmull_s8(vb2x0, va1x0);
162 int16x8_t vprod2x2 = vmull_s8(vb2x0, va2x0);
163 int16x8_t vprod3x2 = vmull_s8(vb2x0, va3x0);
164 vprod0x2 = vmlal_s8(vprod0x2, vb2x1, va0x1);
165 vprod1x2 = vmlal_s8(vprod1x2, vb2x1, va1x1);
166 vprod2x2 = vmlal_s8(vprod2x2, vb2x1, va2x1);
167 vprod3x2 = vmlal_s8(vprod3x2, vb2x1, va3x1);
168 vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
169 vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
170 vacc2x2 = vpadalq_s16(vacc2x2, vprod2x2);
171 vacc3x2 = vpadalq_s16(vacc3x2, vprod3x2);
172 const int8x8_t vb3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
173 int16x8_t vprod0x3 = vmull_s8(vb3x0, va0x0);
174 int16x8_t vprod1x3 = vmull_s8(vb3x0, va1x0);
175 int16x8_t vprod2x3 = vmull_s8(vb3x0, va2x0);
176 int16x8_t vprod3x3 = vmull_s8(vb3x0, va3x0);
177 vprod0x3 = vmlal_s8(vprod0x3, vb3x1, va0x1);
178 vprod1x3 = vmlal_s8(vprod1x3, vb3x1, va1x1);
179 vprod2x3 = vmlal_s8(vprod2x3, vb3x1, va2x1);
180 vprod3x3 = vmlal_s8(vprod3x3, vb3x1, va3x1);
181 vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
182 vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
183 vacc2x3 = vpadalq_s16(vacc2x3, vprod2x3);
184 vacc3x3 = vpadalq_s16(vacc3x3, vprod3x3);
185 const int8x8_t vb4x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
186 int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0);
187 int16x8_t vprod1x4 = vmull_s8(vb4x0, va1x0);
188 int16x8_t vprod2x4 = vmull_s8(vb4x0, va2x0);
189 int16x8_t vprod3x4 = vmull_s8(vb4x0, va3x0);
190 vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1);
191 vprod1x4 = vmlal_s8(vprod1x4, vb4x1, va1x1);
192 vprod2x4 = vmlal_s8(vprod2x4, vb4x1, va2x1);
193 vprod3x4 = vmlal_s8(vprod3x4, vb4x1, va3x1);
194 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
195 vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
196 vacc2x4 = vpadalq_s16(vacc2x4, vprod2x4);
197 vacc3x4 = vpadalq_s16(vacc3x4, vprod3x4);
198 const int8x8_t vb5x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
199 int16x8_t vprod0x5 = vmull_s8(vb5x0, va0x0);
200 int16x8_t vprod1x5 = vmull_s8(vb5x0, va1x0);
201 int16x8_t vprod2x5 = vmull_s8(vb5x0, va2x0);
202 int16x8_t vprod3x5 = vmull_s8(vb5x0, va3x0);
203 vprod0x5 = vmlal_s8(vprod0x5, vb5x1, va0x1);
204 vprod1x5 = vmlal_s8(vprod1x5, vb5x1, va1x1);
205 vprod2x5 = vmlal_s8(vprod2x5, vb5x1, va2x1);
206 vprod3x5 = vmlal_s8(vprod3x5, vb5x1, va3x1);
207 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
208 vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
209 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5);
210 vacc3x5 = vpadalq_s16(vacc3x5, vprod3x5);
211 const int8x8_t vb6x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
212 int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0);
213 int16x8_t vprod1x6 = vmull_s8(vb6x0, va1x0);
214 int16x8_t vprod2x6 = vmull_s8(vb6x0, va2x0);
215 int16x8_t vprod3x6 = vmull_s8(vb6x0, va3x0);
216 vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1);
217 vprod1x6 = vmlal_s8(vprod1x6, vb6x1, va1x1);
218 vprod2x6 = vmlal_s8(vprod2x6, vb6x1, va2x1);
219 vprod3x6 = vmlal_s8(vprod3x6, vb6x1, va3x1);
220 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
221 vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
222 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6);
223 vacc3x6 = vpadalq_s16(vacc3x6, vprod3x6);
224 const int8x8_t vb7x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
225 int16x8_t vprod0x7 = vmull_s8(vb7x0, va0x0);
226 int16x8_t vprod1x7 = vmull_s8(vb7x0, va1x0);
227 int16x8_t vprod2x7 = vmull_s8(vb7x0, va2x0);
228 int16x8_t vprod3x7 = vmull_s8(vb7x0, va3x0);
229 vprod0x7 = vmlal_s8(vprod0x7, vb7x1, va0x1);
230 vprod1x7 = vmlal_s8(vprod1x7, vb7x1, va1x1);
231 vprod2x7 = vmlal_s8(vprod2x7, vb7x1, va2x1);
232 vprod3x7 = vmlal_s8(vprod3x7, vb7x1, va3x1);
233 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
234 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
235 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7);
236 vacc3x7 = vpadalq_s16(vacc3x7, vprod3x7);
237
238 k -= 16 * sizeof(int8_t);
239 }
240
241 // Handle 8 bytes at a time using MUL.
242 if (k != 0) {
243 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
244 const int8x8_t va1 = vld1_s8(a1); a1 += 8;
245 const int8x8_t va2 = vld1_s8(a2); a2 += 8;
246 const int8x8_t va3 = vld1_s8(a3); a3 += 8;
247
248 const int8x8_t vb0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
249 const int16x8_t vprod0x0 = vmull_s8(vb0, va0);
250 const int16x8_t vprod1x0 = vmull_s8(vb0, va1);
251 const int16x8_t vprod2x0 = vmull_s8(vb0, va2);
252 const int16x8_t vprod3x0 = vmull_s8(vb0, va3);
253 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
254 vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
255 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0);
256 vacc3x0 = vpadalq_s16(vacc3x0, vprod3x0);
257 const int8x8_t vb1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
258 const int16x8_t vprod0x1 = vmull_s8(vb1, va0);
259 const int16x8_t vprod1x1 = vmull_s8(vb1, va1);
260 const int16x8_t vprod2x1 = vmull_s8(vb1, va2);
261 const int16x8_t vprod3x1 = vmull_s8(vb1, va3);
262 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
263 vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
264 vacc2x1 = vpadalq_s16(vacc2x1, vprod2x1);
265 vacc3x1 = vpadalq_s16(vacc3x1, vprod3x1);
266 const int8x8_t vb2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
267 const int16x8_t vprod0x2 = vmull_s8(vb2, va0);
268 const int16x8_t vprod1x2 = vmull_s8(vb2, va1);
269 const int16x8_t vprod2x2 = vmull_s8(vb2, va2);
270 const int16x8_t vprod3x2 = vmull_s8(vb2, va3);
271 vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
272 vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
273 vacc2x2 = vpadalq_s16(vacc2x2, vprod2x2);
274 vacc3x2 = vpadalq_s16(vacc3x2, vprod3x2);
275 const int8x8_t vb3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
276 const int16x8_t vprod0x3 = vmull_s8(vb3, va0);
277 const int16x8_t vprod1x3 = vmull_s8(vb3, va1);
278 const int16x8_t vprod2x3 = vmull_s8(vb3, va2);
279 const int16x8_t vprod3x3 = vmull_s8(vb3, va3);
280 vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
281 vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
282 vacc2x3 = vpadalq_s16(vacc2x3, vprod2x3);
283 vacc3x3 = vpadalq_s16(vacc3x3, vprod3x3);
284 const int8x8_t vb4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
285 const int16x8_t vprod0x4 = vmull_s8(vb4, va0);
286 const int16x8_t vprod1x4 = vmull_s8(vb4, va1);
287 const int16x8_t vprod2x4 = vmull_s8(vb4, va2);
288 const int16x8_t vprod3x4 = vmull_s8(vb4, va3);
289 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
290 vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
291 vacc2x4 = vpadalq_s16(vacc2x4, vprod2x4);
292 vacc3x4 = vpadalq_s16(vacc3x4, vprod3x4);
293 const int8x8_t vb5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
294 const int16x8_t vprod0x5 = vmull_s8(vb5, va0);
295 const int16x8_t vprod1x5 = vmull_s8(vb5, va1);
296 const int16x8_t vprod2x5 = vmull_s8(vb5, va2);
297 const int16x8_t vprod3x5 = vmull_s8(vb5, va3);
298 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
299 vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
300 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5);
301 vacc3x5 = vpadalq_s16(vacc3x5, vprod3x5);
302 const int8x8_t vb6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
303 const int16x8_t vprod0x6 = vmull_s8(vb6, va0);
304 const int16x8_t vprod1x6 = vmull_s8(vb6, va1);
305 const int16x8_t vprod2x6 = vmull_s8(vb6, va2);
306 const int16x8_t vprod3x6 = vmull_s8(vb6, va3);
307 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
308 vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
309 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6);
310 vacc3x6 = vpadalq_s16(vacc3x6, vprod3x6);
311 const int8x8_t vb7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
312 const int16x8_t vprod0x7 = vmull_s8(vb7, va0);
313 const int16x8_t vprod1x7 = vmull_s8(vb7, va1);
314 const int16x8_t vprod2x7 = vmull_s8(vb7, va2);
315 const int16x8_t vprod3x7 = vmull_s8(vb7, va3);
316 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
317 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
318 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7);
319 vacc3x7 = vpadalq_s16(vacc3x7, vprod3x7);
320
321 k -= 8 * sizeof(int8_t);
322 }
323
324 p -= 4 * sizeof(void*);
325 } while (p != 0);
326
327 #if XNN_ARCH_ARM64
328 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1);
329 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3);
330 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5);
331 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7);
332 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1);
333 const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3);
334 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5);
335 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7);
336 const int32x4_t vsum2x01 = vpaddq_s32(vacc2x0, vacc2x1);
337 const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3);
338 const int32x4_t vsum2x45 = vpaddq_s32(vacc2x4, vacc2x5);
339 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7);
340 const int32x4_t vsum3x01 = vpaddq_s32(vacc3x0, vacc3x1);
341 const int32x4_t vsum3x23 = vpaddq_s32(vacc3x2, vacc3x3);
342 const int32x4_t vsum3x45 = vpaddq_s32(vacc3x4, vacc3x5);
343 const int32x4_t vsum3x67 = vpaddq_s32(vacc3x6, vacc3x7);
344
345 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23);
346 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67);
347 int32x4_t vacc1x0123 = vpaddq_s32(vsum1x01, vsum1x23);
348 int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67);
349 int32x4_t vacc2x0123 = vpaddq_s32(vsum2x01, vsum2x23);
350 int32x4_t vacc2x4567 = vpaddq_s32(vsum2x45, vsum2x67);
351 int32x4_t vacc3x0123 = vpaddq_s32(vsum3x01, vsum3x23);
352 int32x4_t vacc3x4567 = vpaddq_s32(vsum3x45, vsum3x67);
353 #else
354 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0));
355 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1));
356 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2));
357 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3));
358 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1);
359 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3);
360 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 );
361 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4));
362 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5));
363 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6));
364 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7));
365 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5);
366 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7);
367 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 );
368 const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0));
369 const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1));
370 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2));
371 const int32x2_t vpsum1x3 = vadd_s32(vget_low_s32(vacc1x3), vget_high_s32(vacc1x3));
372 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1);
373 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3);
374 int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23 );
375 const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4));
376 const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(vacc1x5));
377 const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6));
378 const int32x2_t vpsum1x7 = vadd_s32(vget_low_s32(vacc1x7), vget_high_s32(vacc1x7));
379 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5);
380 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7);
381 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 );
382 const int32x2_t vpsum2x0 = vadd_s32(vget_low_s32(vacc2x0), vget_high_s32(vacc2x0));
383 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1));
384 const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2));
385 const int32x2_t vpsum2x3 = vadd_s32(vget_low_s32(vacc2x3), vget_high_s32(vacc2x3));
386 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1);
387 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3);
388 int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23 );
389 const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4));
390 const int32x2_t vpsum2x5 = vadd_s32(vget_low_s32(vacc2x5), vget_high_s32(vacc2x5));
391 const int32x2_t vpsum2x6 = vadd_s32(vget_low_s32(vacc2x6), vget_high_s32(vacc2x6));
392 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7));
393 const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5);
394 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7);
395 int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67 );
396 const int32x2_t vpsum3x0 = vadd_s32(vget_low_s32(vacc3x0), vget_high_s32(vacc3x0));
397 const int32x2_t vpsum3x1 = vadd_s32(vget_low_s32(vacc3x1), vget_high_s32(vacc3x1));
398 const int32x2_t vpsum3x2 = vadd_s32(vget_low_s32(vacc3x2), vget_high_s32(vacc3x2));
399 const int32x2_t vpsum3x3 = vadd_s32(vget_low_s32(vacc3x3), vget_high_s32(vacc3x3));
400 const int32x2_t vsum3x01 = vpadd_s32(vpsum3x0, vpsum3x1);
401 const int32x2_t vsum3x23 = vpadd_s32(vpsum3x2, vpsum3x3);
402 int32x4_t vacc3x0123 = vcombine_s32(vsum3x01, vsum3x23 );
403 const int32x2_t vpsum3x4 = vadd_s32(vget_low_s32(vacc3x4), vget_high_s32(vacc3x4));
404 const int32x2_t vpsum3x5 = vadd_s32(vget_low_s32(vacc3x5), vget_high_s32(vacc3x5));
405 const int32x2_t vpsum3x6 = vadd_s32(vget_low_s32(vacc3x6), vget_high_s32(vacc3x6));
406 const int32x2_t vpsum3x7 = vadd_s32(vget_low_s32(vacc3x7), vget_high_s32(vacc3x7));
407 const int32x2_t vsum3x45 = vpadd_s32(vpsum3x4, vpsum3x5);
408 const int32x2_t vsum3x67 = vpadd_s32(vpsum3x6, vpsum3x7);
409 int32x4_t vacc3x4567 = vcombine_s32(vsum3x45, vsum3x67 );
410 #endif
411
412 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
413 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
414 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
415
416 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
417 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
418 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
419 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
420 vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
421 vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
422 vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
423 vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
424
425 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
426 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
427 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
428 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
429 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
430 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
431 vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
432 vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
433
434 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
435 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
436 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
437 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
438 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
439 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
440 vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
441 vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
442
443 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
444 #if XNN_ARCH_ARM64
445 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
446 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
447 int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
448 int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
449
450 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
451 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
452 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
453 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
454
455 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
456 int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567);
457 #else
458 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
459 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
460 int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
461 int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
462
463 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
464 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
465 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
466 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
467
468 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
469 int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567));
470 #endif
471
472 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
473 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
474 vout2x01234567_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, voutput_min);
475
476 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
477 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
478 vout2x01234567_3x01234567 = vminq_s8(vout2x01234567_3x01234567, voutput_max);
479
480 if (nc >= 8) {
481 vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567));
482 vst1_s8(c2 + 0, vget_low_s8(vout2x01234567_3x01234567));
483 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
484 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
485
486 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
487 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
488 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
489 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
490
491 a = (const int8_t**restrict) ((uintptr_t) a - ks);
492
493 nc -= 8;
494 } else {
495 if (nc & 4) {
496 vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
497 vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
498 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
499 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
500 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
501 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
502 }
503 if (nc & 2) {
504 vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
505 vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
506 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
507 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
508 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
509 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
510 }
511 if (nc & 1) {
512 vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
513 vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
514 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
515 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
516 }
517
518 nc = 0;
519 }
520 } while (nc != 0);
521 }
522