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_4x8c4s2__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_4x8c4s2__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 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 vacc1x01 = vacc0x01;
64 int32x4_t vacc1x23 = vacc0x23;
65 int32x4_t vacc1x45 = vacc0x45;
66 int32x4_t vacc1x67 = vacc0x67;
67 int32x4_t vacc2x01 = vacc0x01;
68 int32x4_t vacc2x23 = vacc0x23;
69 int32x4_t vacc2x45 = vacc0x45;
70 int32x4_t vacc2x67 = vacc0x67;
71 int32x4_t vacc3x01 = vacc0x01;
72 int32x4_t vacc3x23 = vacc0x23;
73 int32x4_t vacc3x45 = vacc0x45;
74 int32x4_t vacc3x67 = vacc0x67;
75
76 size_t p = ks;
77 do {
78 const int8_t* restrict a0 = a[0];
79 if XNN_UNPREDICTABLE(a0 != zero) {
80 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
81 }
82 const int8_t* restrict a1 = a[1];
83 if XNN_UNPREDICTABLE(a1 != zero) {
84 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
85 }
86 const int8_t* restrict a2 = a[2];
87 if XNN_UNPREDICTABLE(a2 != zero) {
88 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
89 }
90 const int8_t* restrict a3 = a[3];
91 if XNN_UNPREDICTABLE(a3 != zero) {
92 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
93 }
94 a += 4;
95
96 size_t k = kc;
97 while (k >= 16 * sizeof(int8_t)) {
98 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
99 int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
100 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
101 int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
102 int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
103 int8x8_t va2x1 = vld1_s8(a2); a2 += 8;
104 int8x8_t va3x0 = vld1_s8(a3); a3 += 8;
105 int8x8_t va3x1 = vld1_s8(a3); a3 += 8;
106
107 const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
108 const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
109 const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
110 const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
111 const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
112 const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
113 const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
114 const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
115
116 int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
117 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1x0);
118 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2x0);
119 int16x8_t vprod3x01c0 = vmull_s8(vb01c0x0, va3x0);
120 const int8x8_t vb01c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
121 vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0x1);
122 vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1x1);
123 vprod2x01c0 = vmlal_s8(vprod2x01c0, vb01c0x1, va2x1);
124 vprod3x01c0 = vmlal_s8(vprod3x01c0, vb01c0x1, va3x1);
125 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
126 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
127 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
128 vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0);
129 int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
130 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1x0);
131 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2x0);
132 int16x8_t vprod3x23c0 = vmull_s8(vb23c0x0, va3x0);
133 const int8x8_t vb23c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
134 vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0x1);
135 vprod1x23c0 = vmlal_s8(vprod1x23c0, vb23c0x1, va1x1);
136 vprod2x23c0 = vmlal_s8(vprod2x23c0, vb23c0x1, va2x1);
137 vprod3x23c0 = vmlal_s8(vprod3x23c0, vb23c0x1, va3x1);
138 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
139 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
140 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
141 vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0);
142 int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
143 int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1x0);
144 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2x0);
145 int16x8_t vprod3x45c0 = vmull_s8(vb45c0x0, va3x0);
146 const int8x8_t vb45c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
147 vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0x1);
148 vprod1x45c0 = vmlal_s8(vprod1x45c0, vb45c0x1, va1x1);
149 vprod2x45c0 = vmlal_s8(vprod2x45c0, vb45c0x1, va2x1);
150 vprod3x45c0 = vmlal_s8(vprod3x45c0, vb45c0x1, va3x1);
151 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
152 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
153 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
154 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0);
155 int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
156 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1x0);
157 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2x0);
158 int16x8_t vprod3x67c0 = vmull_s8(vb67c0x0, va3x0);
159 const int8x8_t vb67c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
160 vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0x1);
161 vprod1x67c0 = vmlal_s8(vprod1x67c0, vb67c0x1, va1x1);
162 vprod2x67c0 = vmlal_s8(vprod2x67c0, vb67c0x1, va2x1);
163 vprod3x67c0 = vmlal_s8(vprod3x67c0, vb67c0x1, va3x1);
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 va0x0 = vext_s8(va0x0, va0x0, 4);
169 va0x1 = vext_s8(va0x1, va0x1, 4);
170 va1x0 = vext_s8(va1x0, va1x0, 4);
171 va1x1 = vext_s8(va1x1, va1x1, 4);
172 va2x0 = vext_s8(va2x0, va2x0, 4);
173 va2x1 = vext_s8(va2x1, va2x1, 4);
174 va3x0 = vext_s8(va3x0, va3x0, 4);
175 va3x1 = vext_s8(va3x1, va3x1, 4);
176 int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
177 int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1x0);
178 int16x8_t vprod2x01c1 = vmull_s8(vb01c1x0, va2x0);
179 int16x8_t vprod3x01c1 = vmull_s8(vb01c1x0, va3x0);
180 const int8x8_t vb01c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
181 vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0x1);
182 vprod1x01c1 = vmlal_s8(vprod1x01c1, vb01c1x1, va1x1);
183 vprod2x01c1 = vmlal_s8(vprod2x01c1, vb01c1x1, va2x1);
184 vprod3x01c1 = vmlal_s8(vprod3x01c1, vb01c1x1, va3x1);
185 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
186 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
187 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
188 vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c1);
189 int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
190 int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1x0);
191 int16x8_t vprod2x23c1 = vmull_s8(vb23c1x0, va2x0);
192 int16x8_t vprod3x23c1 = vmull_s8(vb23c1x0, va3x0);
193 const int8x8_t vb23c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
194 vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0x1);
195 vprod1x23c1 = vmlal_s8(vprod1x23c1, vb23c1x1, va1x1);
196 vprod2x23c1 = vmlal_s8(vprod2x23c1, vb23c1x1, va2x1);
197 vprod3x23c1 = vmlal_s8(vprod3x23c1, vb23c1x1, va3x1);
198 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
199 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
200 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
201 vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c1);
202 int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
203 int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1x0);
204 int16x8_t vprod2x45c1 = vmull_s8(vb45c1x0, va2x0);
205 int16x8_t vprod3x45c1 = vmull_s8(vb45c1x0, va3x0);
206 const int8x8_t vb45c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
207 vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0x1);
208 vprod1x45c1 = vmlal_s8(vprod1x45c1, vb45c1x1, va1x1);
209 vprod2x45c1 = vmlal_s8(vprod2x45c1, vb45c1x1, va2x1);
210 vprod3x45c1 = vmlal_s8(vprod3x45c1, vb45c1x1, va3x1);
211 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
212 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
213 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
214 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c1);
215 int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
216 int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1x0);
217 int16x8_t vprod2x67c1 = vmull_s8(vb67c1x0, va2x0);
218 int16x8_t vprod3x67c1 = vmull_s8(vb67c1x0, va3x0);
219 const int8x8_t vb67c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
220 vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0x1);
221 vprod1x67c1 = vmlal_s8(vprod1x67c1, vb67c1x1, va1x1);
222 vprod2x67c1 = vmlal_s8(vprod2x67c1, vb67c1x1, va2x1);
223 vprod3x67c1 = vmlal_s8(vprod3x67c1, vb67c1x1, va3x1);
224 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
225 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
226 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
227 vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c1);
228
229 k -= 16 * sizeof(int8_t);
230 }
231 if (k != 0) {
232 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
233 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
234 int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
235 int8x8_t va3x0 = vld1_s8(a3); a3 += 8;
236
237 const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
238 const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
239 const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
240 const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
241 const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
242 const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
243 const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
244 const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
245
246 int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
247 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1x0);
248 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2x0);
249 int16x8_t vprod3x01c0 = vmull_s8(vb01c0x0, va3x0);
250 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
251 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
252 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
253 vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0);
254 int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
255 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1x0);
256 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2x0);
257 int16x8_t vprod3x23c0 = vmull_s8(vb23c0x0, va3x0);
258 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
259 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
260 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
261 vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0);
262 int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
263 int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1x0);
264 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2x0);
265 int16x8_t vprod3x45c0 = vmull_s8(vb45c0x0, va3x0);
266 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
267 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
268 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
269 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0);
270 int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
271 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1x0);
272 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2x0);
273 int16x8_t vprod3x67c0 = vmull_s8(vb67c0x0, va3x0);
274 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
275 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
276 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
277 vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0);
278 va0x0 = vext_s8(va0x0, va0x0, 4);
279 va1x0 = vext_s8(va1x0, va1x0, 4);
280 va2x0 = vext_s8(va2x0, va2x0, 4);
281 va3x0 = vext_s8(va3x0, va3x0, 4);
282 int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
283 int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1x0);
284 int16x8_t vprod2x01c1 = vmull_s8(vb01c1x0, va2x0);
285 int16x8_t vprod3x01c1 = vmull_s8(vb01c1x0, va3x0);
286 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
287 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
288 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
289 vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c1);
290 int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
291 int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1x0);
292 int16x8_t vprod2x23c1 = vmull_s8(vb23c1x0, va2x0);
293 int16x8_t vprod3x23c1 = vmull_s8(vb23c1x0, va3x0);
294 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
295 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
296 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
297 vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c1);
298 int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
299 int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1x0);
300 int16x8_t vprod2x45c1 = vmull_s8(vb45c1x0, va2x0);
301 int16x8_t vprod3x45c1 = vmull_s8(vb45c1x0, va3x0);
302 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
303 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
304 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
305 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c1);
306 int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
307 int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1x0);
308 int16x8_t vprod2x67c1 = vmull_s8(vb67c1x0, va2x0);
309 int16x8_t vprod3x67c1 = vmull_s8(vb67c1x0, va3x0);
310 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
311 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
312 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
313 vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c1);
314
315 }
316
317 p -= 4 * sizeof(void*);
318 } while (p != 0);
319
320 #if XNN_ARCH_ARM64
321 int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
322 int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
323 int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
324 int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
325 int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23);
326 int32x4_t vacc2x4567 = vpaddq_s32(vacc2x45, vacc2x67);
327 int32x4_t vacc3x0123 = vpaddq_s32(vacc3x01, vacc3x23);
328 int32x4_t vacc3x4567 = vpaddq_s32(vacc3x45, vacc3x67);
329 #else
330 const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
331 const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
332 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
333 const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
334 const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
335 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
336 const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
337 const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
338 int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
339 const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
340 const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
341 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
342 const int32x2_t vsum2x01 = vpadd_s32(vget_low_s32(vacc2x01), vget_high_s32(vacc2x01));
343 const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23));
344 int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23);
345 const int32x2_t vsum2x45 = vpadd_s32(vget_low_s32(vacc2x45), vget_high_s32(vacc2x45));
346 const int32x2_t vsum2x67 = vpadd_s32(vget_low_s32(vacc2x67), vget_high_s32(vacc2x67));
347 int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67);
348 const int32x2_t vsum3x01 = vpadd_s32(vget_low_s32(vacc3x01), vget_high_s32(vacc3x01));
349 const int32x2_t vsum3x23 = vpadd_s32(vget_low_s32(vacc3x23), vget_high_s32(vacc3x23));
350 int32x4_t vacc3x0123 = vcombine_s32(vsum3x01, vsum3x23);
351 const int32x2_t vsum3x45 = vpadd_s32(vget_low_s32(vacc3x45), vget_high_s32(vacc3x45));
352 const int32x2_t vsum3x67 = vpadd_s32(vget_low_s32(vacc3x67), vget_high_s32(vacc3x67));
353 int32x4_t vacc3x4567 = vcombine_s32(vsum3x45, vsum3x67);
354 #endif
355
356 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
357 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
358 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
359
360 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
361 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
362 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
363 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
364 vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
365 vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
366 vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
367 vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
368
369 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
370 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
371 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
372 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
373 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
374 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
375 vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
376 vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
377
378 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
379 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
380 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
381 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
382 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
383 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
384 vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
385 vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
386
387 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
388 #if XNN_ARCH_ARM64
389 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
390 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
391 int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
392 int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
393
394 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
395 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
396 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
397 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
398
399 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
400 int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567);
401 #else
402 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
403 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
404 int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
405 int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
406
407 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
408 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
409 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
410 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
411
412 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
413 int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567));
414 #endif
415
416 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
417 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
418 vout2x01234567_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, voutput_min);
419
420 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
421 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
422 vout2x01234567_3x01234567 = vminq_s8(vout2x01234567_3x01234567, voutput_max);
423
424 if (nc >= 8) {
425 vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567));
426 vst1_s8(c2 + 0, vget_low_s8(vout2x01234567_3x01234567));
427 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
428 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
429
430 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
431 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
432 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
433 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
434
435 a = (const int8_t**restrict) ((uintptr_t) a - ks);
436
437 nc -= 8;
438 } else {
439 if (nc & 4) {
440 vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
441 vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
442 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
443 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
444 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
445 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
446 }
447 if (nc & 2) {
448 vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
449 vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
450 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
451 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
452 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
453 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
454 }
455 if (nc & 1) {
456 vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
457 vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
458 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
459 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
460 }
461
462 nc = 0;
463 }
464 } while (nc != 0);
465 }
466