1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c4-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_4x16c4s2__neon_mull(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_4x16c4s2__neon_mull(
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 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 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
53 if XNN_UNPREDICTABLE(mr != 4) {
54 c3 = c2;
55 }
56
57 kc = round_up_po2(kc, 8 * sizeof(int8_t));
58 do {
59 int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
60 int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
61 int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
62 int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
63 int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
64 int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
65 int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
66 int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
67 int32x4_t vacc1x01 = vacc0x01;
68 int32x4_t vacc1x23 = vacc0x23;
69 int32x4_t vacc1x45 = vacc0x45;
70 int32x4_t vacc1x67 = vacc0x67;
71 int32x4_t vacc1x89 = vacc0x89;
72 int32x4_t vacc1xAB = vacc0xAB;
73 int32x4_t vacc1xCD = vacc0xCD;
74 int32x4_t vacc1xEF = vacc0xEF;
75 int32x4_t vacc2x01 = vacc0x01;
76 int32x4_t vacc2x23 = vacc0x23;
77 int32x4_t vacc2x45 = vacc0x45;
78 int32x4_t vacc2x67 = vacc0x67;
79 int32x4_t vacc2x89 = vacc0x89;
80 int32x4_t vacc2xAB = vacc0xAB;
81 int32x4_t vacc2xCD = vacc0xCD;
82 int32x4_t vacc2xEF = vacc0xEF;
83 int32x4_t vacc3x01 = vacc0x01;
84 int32x4_t vacc3x23 = vacc0x23;
85 int32x4_t vacc3x45 = vacc0x45;
86 int32x4_t vacc3x67 = vacc0x67;
87 int32x4_t vacc3x89 = vacc0x89;
88 int32x4_t vacc3xAB = vacc0xAB;
89 int32x4_t vacc3xCD = vacc0xCD;
90 int32x4_t vacc3xEF = vacc0xEF;
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 do {
114 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
115 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
116 int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
117 int8x8_t va3x0 = vld1_s8(a3); a3 += 8;
118
119 const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
120 const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
121 const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
122 const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
123 const int8x8_t vb89c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
124 const int8x8_t vbABc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
125 const int8x8_t vbCDc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
126 const int8x8_t vbEFc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
127 const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
128 const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
129 const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
130 const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
131 const int8x8_t vb89c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
132 const int8x8_t vbABc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
133 const int8x8_t vbCDc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
134 const int8x8_t vbEFc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
135
136 int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
137 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1x0);
138 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2x0);
139 int16x8_t vprod3x01c0 = vmull_s8(vb01c0x0, va3x0);
140 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
141 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
142 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
143 vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0);
144 int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
145 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1x0);
146 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2x0);
147 int16x8_t vprod3x23c0 = vmull_s8(vb23c0x0, va3x0);
148 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
149 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
150 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
151 vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0);
152 int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
153 int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1x0);
154 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2x0);
155 int16x8_t vprod3x45c0 = vmull_s8(vb45c0x0, va3x0);
156 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
157 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
158 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
159 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0);
160 int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
161 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1x0);
162 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2x0);
163 int16x8_t vprod3x67c0 = vmull_s8(vb67c0x0, va3x0);
164 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
165 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
166 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
167 vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0);
168 int16x8_t vprod0x89c0 = vmull_s8(vb89c0x0, va0x0);
169 int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1x0);
170 int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2x0);
171 int16x8_t vprod3x89c0 = vmull_s8(vb89c0x0, va3x0);
172 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
173 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
174 vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c0);
175 vacc3x89 = vpadalq_s16(vacc3x89, vprod3x89c0);
176 int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0x0);
177 int16x8_t vprod1xABc0 = vmull_s8(vbABc0x0, va1x0);
178 int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2x0);
179 int16x8_t vprod3xABc0 = vmull_s8(vbABc0x0, va3x0);
180 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
181 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
182 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0);
183 vacc3xAB = vpadalq_s16(vacc3xAB, vprod3xABc0);
184 int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0x0);
185 int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0x0, va1x0);
186 int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2x0);
187 int16x8_t vprod3xCDc0 = vmull_s8(vbCDc0x0, va3x0);
188 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
189 vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
190 vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc0);
191 vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc0);
192 int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0x0, va0x0);
193 int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0x0, va1x0);
194 int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0x0, va2x0);
195 int16x8_t vprod3xEFc0 = vmull_s8(vbEFc0x0, va3x0);
196 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
197 vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
198 vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc0);
199 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc0);
200 va0x0 = vext_s8(va0x0, va0x0, 4);
201 va1x0 = vext_s8(va1x0, va1x0, 4);
202 va2x0 = vext_s8(va2x0, va2x0, 4);
203 va3x0 = vext_s8(va3x0, va3x0, 4);
204 int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
205 int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1x0);
206 int16x8_t vprod2x01c1 = vmull_s8(vb01c1x0, va2x0);
207 int16x8_t vprod3x01c1 = vmull_s8(vb01c1x0, va3x0);
208 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
209 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
210 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
211 vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c1);
212 int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
213 int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1x0);
214 int16x8_t vprod2x23c1 = vmull_s8(vb23c1x0, va2x0);
215 int16x8_t vprod3x23c1 = vmull_s8(vb23c1x0, va3x0);
216 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
217 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
218 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
219 vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c1);
220 int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
221 int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1x0);
222 int16x8_t vprod2x45c1 = vmull_s8(vb45c1x0, va2x0);
223 int16x8_t vprod3x45c1 = vmull_s8(vb45c1x0, va3x0);
224 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
225 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
226 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
227 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c1);
228 int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
229 int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1x0);
230 int16x8_t vprod2x67c1 = vmull_s8(vb67c1x0, va2x0);
231 int16x8_t vprod3x67c1 = vmull_s8(vb67c1x0, va3x0);
232 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
233 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
234 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
235 vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c1);
236 int16x8_t vprod0x89c1 = vmull_s8(vb89c1x0, va0x0);
237 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0);
238 int16x8_t vprod2x89c1 = vmull_s8(vb89c1x0, va2x0);
239 int16x8_t vprod3x89c1 = vmull_s8(vb89c1x0, va3x0);
240 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
241 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
242 vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c1);
243 vacc3x89 = vpadalq_s16(vacc3x89, vprod3x89c1);
244 int16x8_t vprod0xABc1 = vmull_s8(vbABc1x0, va0x0);
245 int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1x0);
246 int16x8_t vprod2xABc1 = vmull_s8(vbABc1x0, va2x0);
247 int16x8_t vprod3xABc1 = vmull_s8(vbABc1x0, va3x0);
248 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
249 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
250 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc1);
251 vacc3xAB = vpadalq_s16(vacc3xAB, vprod3xABc1);
252 int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1x0, va0x0);
253 int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1x0);
254 int16x8_t vprod2xCDc1 = vmull_s8(vbCDc1x0, va2x0);
255 int16x8_t vprod3xCDc1 = vmull_s8(vbCDc1x0, va3x0);
256 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
257 vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
258 vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc1);
259 vacc3xCD = vpadalq_s16(vacc3xCD, vprod3xCDc1);
260 int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1x0, va0x0);
261 int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1x0, va1x0);
262 int16x8_t vprod2xEFc1 = vmull_s8(vbEFc1x0, va2x0);
263 int16x8_t vprod3xEFc1 = vmull_s8(vbEFc1x0, va3x0);
264 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
265 vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
266 vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc1);
267 vacc3xEF = vpadalq_s16(vacc3xEF, vprod3xEFc1);
268
269 k -= 8 * sizeof(int8_t);
270 } while (k != 0);
271
272 p -= 4 * sizeof(void*);
273 } while (p != 0);
274
275 #if XNN_ARCH_ARM64
276 int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
277 int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
278 int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB);
279 int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF);
280 int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
281 int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
282 int32x4_t vacc1x89AB = vpaddq_s32(vacc1x89, vacc1xAB);
283 int32x4_t vacc1xCDEF = vpaddq_s32(vacc1xCD, vacc1xEF);
284 int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23);
285 int32x4_t vacc2x4567 = vpaddq_s32(vacc2x45, vacc2x67);
286 int32x4_t vacc2x89AB = vpaddq_s32(vacc2x89, vacc2xAB);
287 int32x4_t vacc2xCDEF = vpaddq_s32(vacc2xCD, vacc2xEF);
288 int32x4_t vacc3x0123 = vpaddq_s32(vacc3x01, vacc3x23);
289 int32x4_t vacc3x4567 = vpaddq_s32(vacc3x45, vacc3x67);
290 int32x4_t vacc3x89AB = vpaddq_s32(vacc3x89, vacc3xAB);
291 int32x4_t vacc3xCDEF = vpaddq_s32(vacc3xCD, vacc3xEF);
292 #else
293 const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
294 const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
295 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
296 const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
297 const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
298 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
299 const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89));
300 const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB));
301 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB);
302 const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD));
303 const int32x2_t vsum0xEF = vpadd_s32(vget_low_s32(vacc0xEF), vget_high_s32(vacc0xEF));
304 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF);
305 const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
306 const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
307 int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
308 const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
309 const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
310 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
311 const int32x2_t vsum1x89 = vpadd_s32(vget_low_s32(vacc1x89), vget_high_s32(vacc1x89));
312 const int32x2_t vsum1xAB = vpadd_s32(vget_low_s32(vacc1xAB), vget_high_s32(vacc1xAB));
313 int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB);
314 const int32x2_t vsum1xCD = vpadd_s32(vget_low_s32(vacc1xCD), vget_high_s32(vacc1xCD));
315 const int32x2_t vsum1xEF = vpadd_s32(vget_low_s32(vacc1xEF), vget_high_s32(vacc1xEF));
316 int32x4_t vacc1xCDEF = vcombine_s32(vsum1xCD, vsum1xEF);
317 const int32x2_t vsum2x01 = vpadd_s32(vget_low_s32(vacc2x01), vget_high_s32(vacc2x01));
318 const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23));
319 int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23);
320 const int32x2_t vsum2x45 = vpadd_s32(vget_low_s32(vacc2x45), vget_high_s32(vacc2x45));
321 const int32x2_t vsum2x67 = vpadd_s32(vget_low_s32(vacc2x67), vget_high_s32(vacc2x67));
322 int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67);
323 const int32x2_t vsum2x89 = vpadd_s32(vget_low_s32(vacc2x89), vget_high_s32(vacc2x89));
324 const int32x2_t vsum2xAB = vpadd_s32(vget_low_s32(vacc2xAB), vget_high_s32(vacc2xAB));
325 int32x4_t vacc2x89AB = vcombine_s32(vsum2x89, vsum2xAB);
326 const int32x2_t vsum2xCD = vpadd_s32(vget_low_s32(vacc2xCD), vget_high_s32(vacc2xCD));
327 const int32x2_t vsum2xEF = vpadd_s32(vget_low_s32(vacc2xEF), vget_high_s32(vacc2xEF));
328 int32x4_t vacc2xCDEF = vcombine_s32(vsum2xCD, vsum2xEF);
329 const int32x2_t vsum3x01 = vpadd_s32(vget_low_s32(vacc3x01), vget_high_s32(vacc3x01));
330 const int32x2_t vsum3x23 = vpadd_s32(vget_low_s32(vacc3x23), vget_high_s32(vacc3x23));
331 int32x4_t vacc3x0123 = vcombine_s32(vsum3x01, vsum3x23);
332 const int32x2_t vsum3x45 = vpadd_s32(vget_low_s32(vacc3x45), vget_high_s32(vacc3x45));
333 const int32x2_t vsum3x67 = vpadd_s32(vget_low_s32(vacc3x67), vget_high_s32(vacc3x67));
334 int32x4_t vacc3x4567 = vcombine_s32(vsum3x45, vsum3x67);
335 const int32x2_t vsum3x89 = vpadd_s32(vget_low_s32(vacc3x89), vget_high_s32(vacc3x89));
336 const int32x2_t vsum3xAB = vpadd_s32(vget_low_s32(vacc3xAB), vget_high_s32(vacc3xAB));
337 int32x4_t vacc3x89AB = vcombine_s32(vsum3x89, vsum3xAB);
338 const int32x2_t vsum3xCD = vpadd_s32(vget_low_s32(vacc3xCD), vget_high_s32(vacc3xCD));
339 const int32x2_t vsum3xEF = vpadd_s32(vget_low_s32(vacc3xEF), vget_high_s32(vacc3xEF));
340 int32x4_t vacc3xCDEF = vcombine_s32(vsum3xCD, vsum3xEF);
341 #endif
342
343 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
344 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
345 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
346
347 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
348 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
349 vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
350 vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
351 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
352 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
353 vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
354 vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
355 vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
356 vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
357 vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
358 vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
359 vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
360 vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
361 vacc3x89AB = vqshlq_s32(vacc3x89AB, vright_pre_shift);
362 vacc3xCDEF = vqshlq_s32(vacc3xCDEF, vright_pre_shift);
363
364 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
365 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
366 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
367 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
368 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
369 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
370 vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
371 vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
372 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
373 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
374 vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
375 vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
376 vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
377 vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
378 vacc3x89AB = vqdmulhq_s32(vacc3x89AB, vmultiplier);
379 vacc3xCDEF = vqdmulhq_s32(vacc3xCDEF, vmultiplier);
380
381 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
382 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
383 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
384 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
385 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
386 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
387 vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
388 vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
389 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
390 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
391 vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
392 vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
393 vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
394 vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
395 vacc3x89AB = vrshlq_s32(vacc3x89AB, vright_post_shift);
396 vacc3xCDEF = vrshlq_s32(vacc3xCDEF, vright_post_shift);
397
398 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
399 #if XNN_ARCH_ARM64
400 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
401 int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
402 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
403 int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
404 int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
405 int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
406 int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
407 int16x8_t vacc3x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF);
408
409 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
410 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
411 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
412 vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
413 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
414 vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
415 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
416 vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
417
418 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
419 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
420 int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
421 int8x16_t vout3x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc3x01234567), vacc3x89ABCDEF);
422 #else
423 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
424 int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
425 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
426 int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
427 int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
428 int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
429 int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
430 int16x8_t vacc3x89ABCDEF = vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF));
431
432 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
433 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
434 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
435 vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
436 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
437 vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
438 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
439 vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
440
441 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
442 int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
443 int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
444 int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCDEF));
445 #endif
446
447 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
448 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
449 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
450 vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
451 vout3x0123456789ABCDEF = vmaxq_s8(vout3x0123456789ABCDEF, voutput_min);
452
453 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
454 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
455 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
456 vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
457 vout3x0123456789ABCDEF = vminq_s8(vout3x0123456789ABCDEF, voutput_max);
458
459 if (nc >= 16) {
460 vst1q_s8(c3 + 0, vout3x0123456789ABCDEF);
461 vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
462 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
463 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
464
465 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
466 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
467 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
468 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
469
470 a = (const int8_t**restrict) ((uintptr_t) a - ks);
471
472 nc -= 16;
473 } else {
474 int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8(vout3x0123456789ABCDEF));
475 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
476 if (nc & 8) {
477 vst1_s8(c3, vget_high_s8(vout2x01234567_3x01234567)); c3 += 8;
478 vst1_s8(c2, vget_low_s8(vout2x01234567_3x01234567)); c2 += 8;
479 vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
480 vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
481 vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0123456789ABCDEF));
482 vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
483 }
484 if (nc & 4) {
485 vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
486 vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
487 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
488 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
489 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
490 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
491 }
492 if (nc & 2) {
493 vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
494 vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
495 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
496 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
497 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
498 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
499 }
500 if (nc & 1) {
501 vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
502 vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
503 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
504 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
505 }
506
507 nc = 0;
508 }
509 } while (nc != 0);
510 }
511