1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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_gemm_minmax_rndnu_ukernel_4x8c4s2__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_4x8c4s2__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 vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
63 int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
64 int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
65 int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const int32_t*) w + 2;
66 int32x4_t vacc1x01 = vacc0x01;
67 int32x4_t vacc1x23 = vacc0x23;
68 int32x4_t vacc1x45 = vacc0x45;
69 int32x4_t vacc1x67 = vacc0x67;
70 int32x4_t vacc2x01 = vacc0x01;
71 int32x4_t vacc2x23 = vacc0x23;
72 int32x4_t vacc2x45 = vacc0x45;
73 int32x4_t vacc2x67 = vacc0x67;
74 int32x4_t vacc3x01 = vacc0x01;
75 int32x4_t vacc3x23 = vacc0x23;
76 int32x4_t vacc3x45 = vacc0x45;
77 int32x4_t vacc3x67 = vacc0x67;
78
79 size_t k = kc;
80 while (k >= 16 * sizeof(int8_t)) {
81 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
82 int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
83 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
84 int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
85 int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
86 int8x8_t va2x1 = vld1_s8(a2); a2 += 8;
87 int8x8_t va3x0 = vld1_s8(a3); a3 += 8;
88 int8x8_t va3x1 = vld1_s8(a3); a3 += 8;
89
90 const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
91 const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
92 const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
93 const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
94 const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
95 const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
96 const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
97 const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
98
99 int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
100 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1x0);
101 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2x0);
102 int16x8_t vprod3x01c0 = vmull_s8(vb01c0x0, va3x0);
103 const int8x8_t vb01c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
104 vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0x1);
105 vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1x1);
106 vprod2x01c0 = vmlal_s8(vprod2x01c0, vb01c0x1, va2x1);
107 vprod3x01c0 = vmlal_s8(vprod3x01c0, vb01c0x1, va3x1);
108 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
109 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
110 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
111 vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0);
112 int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
113 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1x0);
114 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2x0);
115 int16x8_t vprod3x23c0 = vmull_s8(vb23c0x0, va3x0);
116 const int8x8_t vb23c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
117 vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0x1);
118 vprod1x23c0 = vmlal_s8(vprod1x23c0, vb23c0x1, va1x1);
119 vprod2x23c0 = vmlal_s8(vprod2x23c0, vb23c0x1, va2x1);
120 vprod3x23c0 = vmlal_s8(vprod3x23c0, vb23c0x1, va3x1);
121 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
122 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
123 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
124 vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0);
125 int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
126 int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1x0);
127 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2x0);
128 int16x8_t vprod3x45c0 = vmull_s8(vb45c0x0, va3x0);
129 const int8x8_t vb45c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
130 vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0x1);
131 vprod1x45c0 = vmlal_s8(vprod1x45c0, vb45c0x1, va1x1);
132 vprod2x45c0 = vmlal_s8(vprod2x45c0, vb45c0x1, va2x1);
133 vprod3x45c0 = vmlal_s8(vprod3x45c0, vb45c0x1, va3x1);
134 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
135 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
136 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
137 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0);
138 int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
139 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1x0);
140 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2x0);
141 int16x8_t vprod3x67c0 = vmull_s8(vb67c0x0, va3x0);
142 const int8x8_t vb67c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
143 vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0x1);
144 vprod1x67c0 = vmlal_s8(vprod1x67c0, vb67c0x1, va1x1);
145 vprod2x67c0 = vmlal_s8(vprod2x67c0, vb67c0x1, va2x1);
146 vprod3x67c0 = vmlal_s8(vprod3x67c0, vb67c0x1, va3x1);
147 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
148 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
149 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
150 vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0);
151 va0x0 = vext_s8(va0x0, va0x0, 4);
152 va0x1 = vext_s8(va0x1, va0x1, 4);
153 va1x0 = vext_s8(va1x0, va1x0, 4);
154 va1x1 = vext_s8(va1x1, va1x1, 4);
155 va2x0 = vext_s8(va2x0, va2x0, 4);
156 va2x1 = vext_s8(va2x1, va2x1, 4);
157 va3x0 = vext_s8(va3x0, va3x0, 4);
158 va3x1 = vext_s8(va3x1, va3x1, 4);
159 int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
160 int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1x0);
161 int16x8_t vprod2x01c1 = vmull_s8(vb01c1x0, va2x0);
162 int16x8_t vprod3x01c1 = vmull_s8(vb01c1x0, va3x0);
163 const int8x8_t vb01c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
164 vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0x1);
165 vprod1x01c1 = vmlal_s8(vprod1x01c1, vb01c1x1, va1x1);
166 vprod2x01c1 = vmlal_s8(vprod2x01c1, vb01c1x1, va2x1);
167 vprod3x01c1 = vmlal_s8(vprod3x01c1, vb01c1x1, va3x1);
168 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
169 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
170 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
171 vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c1);
172 int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
173 int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1x0);
174 int16x8_t vprod2x23c1 = vmull_s8(vb23c1x0, va2x0);
175 int16x8_t vprod3x23c1 = vmull_s8(vb23c1x0, va3x0);
176 const int8x8_t vb23c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
177 vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0x1);
178 vprod1x23c1 = vmlal_s8(vprod1x23c1, vb23c1x1, va1x1);
179 vprod2x23c1 = vmlal_s8(vprod2x23c1, vb23c1x1, va2x1);
180 vprod3x23c1 = vmlal_s8(vprod3x23c1, vb23c1x1, va3x1);
181 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
182 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
183 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
184 vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c1);
185 int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
186 int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1x0);
187 int16x8_t vprod2x45c1 = vmull_s8(vb45c1x0, va2x0);
188 int16x8_t vprod3x45c1 = vmull_s8(vb45c1x0, va3x0);
189 const int8x8_t vb45c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
190 vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0x1);
191 vprod1x45c1 = vmlal_s8(vprod1x45c1, vb45c1x1, va1x1);
192 vprod2x45c1 = vmlal_s8(vprod2x45c1, vb45c1x1, va2x1);
193 vprod3x45c1 = vmlal_s8(vprod3x45c1, vb45c1x1, va3x1);
194 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
195 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
196 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
197 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c1);
198 int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
199 int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1x0);
200 int16x8_t vprod2x67c1 = vmull_s8(vb67c1x0, va2x0);
201 int16x8_t vprod3x67c1 = vmull_s8(vb67c1x0, va3x0);
202 const int8x8_t vb67c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
203 vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0x1);
204 vprod1x67c1 = vmlal_s8(vprod1x67c1, vb67c1x1, va1x1);
205 vprod2x67c1 = vmlal_s8(vprod2x67c1, vb67c1x1, va2x1);
206 vprod3x67c1 = vmlal_s8(vprod3x67c1, vb67c1x1, va3x1);
207 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
208 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
209 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
210 vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c1);
211
212 k -= 16 * sizeof(int8_t);
213 }
214 if (k != 0) {
215 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
216 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
217 int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
218 int8x8_t va3x0 = vld1_s8(a3); a3 += 8;
219
220 const int8x8_t vb01c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
221 const int8x8_t vb23c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
222 const int8x8_t vb45c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
223 const int8x8_t vb67c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
224 const int8x8_t vb01c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
225 const int8x8_t vb23c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
226 const int8x8_t vb45c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
227 const int8x8_t vb67c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
228
229 int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0x0);
230 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1x0);
231 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2x0);
232 int16x8_t vprod3x01c0 = vmull_s8(vb01c0x0, va3x0);
233 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
234 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
235 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
236 vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0);
237 int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0x0);
238 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1x0);
239 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2x0);
240 int16x8_t vprod3x23c0 = vmull_s8(vb23c0x0, va3x0);
241 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
242 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
243 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
244 vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0);
245 int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0x0);
246 int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1x0);
247 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2x0);
248 int16x8_t vprod3x45c0 = vmull_s8(vb45c0x0, va3x0);
249 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
250 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
251 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
252 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0);
253 int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0x0);
254 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1x0);
255 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2x0);
256 int16x8_t vprod3x67c0 = vmull_s8(vb67c0x0, va3x0);
257 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
258 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
259 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
260 vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0);
261 va0x0 = vext_s8(va0x0, va0x0, 4);
262 va1x0 = vext_s8(va1x0, va1x0, 4);
263 va2x0 = vext_s8(va2x0, va2x0, 4);
264 va3x0 = vext_s8(va3x0, va3x0, 4);
265 int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0x0);
266 int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1x0);
267 int16x8_t vprod2x01c1 = vmull_s8(vb01c1x0, va2x0);
268 int16x8_t vprod3x01c1 = vmull_s8(vb01c1x0, va3x0);
269 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
270 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
271 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
272 vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c1);
273 int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0x0);
274 int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1x0);
275 int16x8_t vprod2x23c1 = vmull_s8(vb23c1x0, va2x0);
276 int16x8_t vprod3x23c1 = vmull_s8(vb23c1x0, va3x0);
277 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
278 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
279 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
280 vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c1);
281 int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0x0);
282 int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1x0);
283 int16x8_t vprod2x45c1 = vmull_s8(vb45c1x0, va2x0);
284 int16x8_t vprod3x45c1 = vmull_s8(vb45c1x0, va3x0);
285 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
286 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
287 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
288 vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c1);
289 int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0x0);
290 int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1x0);
291 int16x8_t vprod2x67c1 = vmull_s8(vb67c1x0, va2x0);
292 int16x8_t vprod3x67c1 = vmull_s8(vb67c1x0, va3x0);
293 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
294 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
295 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
296 vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c1);
297
298 }
299
300 #if XNN_ARCH_ARM64
301 int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
302 int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
303 int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
304 int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
305 int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23);
306 int32x4_t vacc2x4567 = vpaddq_s32(vacc2x45, vacc2x67);
307 int32x4_t vacc3x0123 = vpaddq_s32(vacc3x01, vacc3x23);
308 int32x4_t vacc3x4567 = vpaddq_s32(vacc3x45, vacc3x67);
309 #else
310 const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
311 const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
312 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
313 const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
314 const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
315 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
316 const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
317 const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
318 int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
319 const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
320 const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
321 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
322 const int32x2_t vsum2x01 = vpadd_s32(vget_low_s32(vacc2x01), vget_high_s32(vacc2x01));
323 const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23));
324 int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23);
325 const int32x2_t vsum2x45 = vpadd_s32(vget_low_s32(vacc2x45), vget_high_s32(vacc2x45));
326 const int32x2_t vsum2x67 = vpadd_s32(vget_low_s32(vacc2x67), vget_high_s32(vacc2x67));
327 int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67);
328 const int32x2_t vsum3x01 = vpadd_s32(vget_low_s32(vacc3x01), vget_high_s32(vacc3x01));
329 const int32x2_t vsum3x23 = vpadd_s32(vget_low_s32(vacc3x23), vget_high_s32(vacc3x23));
330 int32x4_t vacc3x0123 = vcombine_s32(vsum3x01, vsum3x23);
331 const int32x2_t vsum3x45 = vpadd_s32(vget_low_s32(vacc3x45), vget_high_s32(vacc3x45));
332 const int32x2_t vsum3x67 = vpadd_s32(vget_low_s32(vacc3x67), vget_high_s32(vacc3x67));
333 int32x4_t vacc3x4567 = vcombine_s32(vsum3x45, vsum3x67);
334 #endif
335
336 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
337 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
338 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
339
340 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
341 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
342 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
343 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
344 vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
345 vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
346 vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
347 vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
348
349 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
350 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
351 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
352 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
353 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
354 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
355 vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
356 vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
357
358 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
359 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
360 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
361 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
362 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
363 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
364 vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
365 vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
366
367 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
368 #if XNN_ARCH_ARM64
369 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
370 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
371 int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
372 int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
373
374 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
375 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
376 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
377 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
378
379 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
380 int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567);
381 #else
382 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
383 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
384 int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
385 int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
386
387 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
388 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
389 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
390 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
391
392 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
393 int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567));
394 #endif
395
396 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
397 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
398 vout2x01234567_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, voutput_min);
399
400 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
401 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
402 vout2x01234567_3x01234567 = vminq_s8(vout2x01234567_3x01234567, voutput_max);
403
404 if (nc >= 8) {
405 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
406 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
407 vst1_s8(c2 + 0, vget_low_s8(vout2x01234567_3x01234567));
408 vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567));
409
410 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
411 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
412 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
413 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
414
415 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
416 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
417 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
418 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
419
420 nc -= 8;
421 } else {
422 // Final case where not all of the 8 columns fit in the destination.
423 if (nc & 4) {
424 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
425 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
426 vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
427 vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
428 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
429 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
430 }
431 if (nc & 2) {
432 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
433 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
434 vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
435 vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
436 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
437 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
438 }
439 if (nc & 1) {
440 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
441 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
442 vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
443 vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
444 }
445
446 nc = 0;
447 }
448 } while (nc != 0);
449 }
450