1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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/gemm.h>
15 #include <xnnpack/math.h>
16
17
xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__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_4x8c8__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 kc = round_up_po2(kc, 8 * sizeof(int8_t));
40 const int8_t* a0 = a;
41 int8_t* c0 = c;
42 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
43 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
44 if XNN_UNPREDICTABLE(mr < 2) {
45 a1 = a0;
46 c1 = c0;
47 }
48 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
49 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 a2 = a1;
52 c2 = c1;
53 }
54 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
55 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr != 4) {
57 a3 = a2;
58 c3 = c2;
59 }
60
61 do {
62 int32x4_t vacc0x0 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
63 int32x4_t vacc0x1 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
64 int32x4_t vacc0x2 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
65 int32x4_t vacc0x3 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
66 int32x4_t vacc0x4 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
67 int32x4_t vacc0x5 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
68 int32x4_t vacc0x6 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
69 int32x4_t vacc0x7 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
70 int32x4_t vacc1x0 = vacc0x0;
71 int32x4_t vacc1x1 = vacc0x1;
72 int32x4_t vacc1x2 = vacc0x2;
73 int32x4_t vacc1x3 = vacc0x3;
74 int32x4_t vacc1x4 = vacc0x4;
75 int32x4_t vacc1x5 = vacc0x5;
76 int32x4_t vacc1x6 = vacc0x6;
77 int32x4_t vacc1x7 = vacc0x7;
78 int32x4_t vacc2x0 = vacc0x0;
79 int32x4_t vacc2x1 = vacc0x1;
80 int32x4_t vacc2x2 = vacc0x2;
81 int32x4_t vacc2x3 = vacc0x3;
82 int32x4_t vacc2x4 = vacc0x4;
83 int32x4_t vacc2x5 = vacc0x5;
84 int32x4_t vacc2x6 = vacc0x6;
85 int32x4_t vacc2x7 = vacc0x7;
86 int32x4_t vacc3x0 = vacc0x0;
87 int32x4_t vacc3x1 = vacc0x1;
88 int32x4_t vacc3x2 = vacc0x2;
89 int32x4_t vacc3x3 = vacc0x3;
90 int32x4_t vacc3x4 = vacc0x4;
91 int32x4_t vacc3x5 = vacc0x5;
92 int32x4_t vacc3x6 = vacc0x6;
93 int32x4_t vacc3x7 = vacc0x7;
94
95 size_t k = kc;
96 // 2x partial unrolled loop to load 16 bytes at a time using MLA.
97 while (k >= 16 * sizeof(int8_t)) {
98 const int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
99 const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
100 const int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
101 const int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
102 const int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
103 const int8x8_t va2x1 = vld1_s8(a2); a2 += 8;
104 const int8x8_t va3x0 = vld1_s8(a3); a3 += 8;
105 const int8x8_t va3x1 = vld1_s8(a3); a3 += 8;
106
107 const int8x8_t vb0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
108 const int8x8_t vb1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
109 const int8x8_t vb2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
110 const int8x8_t vb3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
111 const int8x8_t vb4x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
112 const int8x8_t vb5x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
113 const int8x8_t vb6x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
114 const int8x8_t vb7x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
115
116 const int8x8_t vb0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
117 int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0);
118 int16x8_t vprod1x0 = vmull_s8(vb0x0, va1x0);
119 int16x8_t vprod2x0 = vmull_s8(vb0x0, va2x0);
120 int16x8_t vprod3x0 = vmull_s8(vb0x0, va3x0);
121 vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1);
122 vprod1x0 = vmlal_s8(vprod1x0, vb0x1, va1x1);
123 vprod2x0 = vmlal_s8(vprod2x0, vb0x1, va2x1);
124 vprod3x0 = vmlal_s8(vprod3x0, vb0x1, va3x1);
125 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
126 vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
127 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0);
128 vacc3x0 = vpadalq_s16(vacc3x0, vprod3x0);
129 const int8x8_t vb1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
130 int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0);
131 int16x8_t vprod1x1 = vmull_s8(vb1x0, va1x0);
132 int16x8_t vprod2x1 = vmull_s8(vb1x0, va2x0);
133 int16x8_t vprod3x1 = vmull_s8(vb1x0, va3x0);
134 vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1);
135 vprod1x1 = vmlal_s8(vprod1x1, vb1x1, va1x1);
136 vprod2x1 = vmlal_s8(vprod2x1, vb1x1, va2x1);
137 vprod3x1 = vmlal_s8(vprod3x1, vb1x1, va3x1);
138 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
139 vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
140 vacc2x1 = vpadalq_s16(vacc2x1, vprod2x1);
141 vacc3x1 = vpadalq_s16(vacc3x1, vprod3x1);
142 const int8x8_t vb2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
143 int16x8_t vprod0x2 = vmull_s8(vb2x0, va0x0);
144 int16x8_t vprod1x2 = vmull_s8(vb2x0, va1x0);
145 int16x8_t vprod2x2 = vmull_s8(vb2x0, va2x0);
146 int16x8_t vprod3x2 = vmull_s8(vb2x0, va3x0);
147 vprod0x2 = vmlal_s8(vprod0x2, vb2x1, va0x1);
148 vprod1x2 = vmlal_s8(vprod1x2, vb2x1, va1x1);
149 vprod2x2 = vmlal_s8(vprod2x2, vb2x1, va2x1);
150 vprod3x2 = vmlal_s8(vprod3x2, vb2x1, va3x1);
151 vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
152 vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
153 vacc2x2 = vpadalq_s16(vacc2x2, vprod2x2);
154 vacc3x2 = vpadalq_s16(vacc3x2, vprod3x2);
155 const int8x8_t vb3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
156 int16x8_t vprod0x3 = vmull_s8(vb3x0, va0x0);
157 int16x8_t vprod1x3 = vmull_s8(vb3x0, va1x0);
158 int16x8_t vprod2x3 = vmull_s8(vb3x0, va2x0);
159 int16x8_t vprod3x3 = vmull_s8(vb3x0, va3x0);
160 vprod0x3 = vmlal_s8(vprod0x3, vb3x1, va0x1);
161 vprod1x3 = vmlal_s8(vprod1x3, vb3x1, va1x1);
162 vprod2x3 = vmlal_s8(vprod2x3, vb3x1, va2x1);
163 vprod3x3 = vmlal_s8(vprod3x3, vb3x1, va3x1);
164 vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
165 vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
166 vacc2x3 = vpadalq_s16(vacc2x3, vprod2x3);
167 vacc3x3 = vpadalq_s16(vacc3x3, vprod3x3);
168 const int8x8_t vb4x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
169 int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0);
170 int16x8_t vprod1x4 = vmull_s8(vb4x0, va1x0);
171 int16x8_t vprod2x4 = vmull_s8(vb4x0, va2x0);
172 int16x8_t vprod3x4 = vmull_s8(vb4x0, va3x0);
173 vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1);
174 vprod1x4 = vmlal_s8(vprod1x4, vb4x1, va1x1);
175 vprod2x4 = vmlal_s8(vprod2x4, vb4x1, va2x1);
176 vprod3x4 = vmlal_s8(vprod3x4, vb4x1, va3x1);
177 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
178 vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
179 vacc2x4 = vpadalq_s16(vacc2x4, vprod2x4);
180 vacc3x4 = vpadalq_s16(vacc3x4, vprod3x4);
181 const int8x8_t vb5x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
182 int16x8_t vprod0x5 = vmull_s8(vb5x0, va0x0);
183 int16x8_t vprod1x5 = vmull_s8(vb5x0, va1x0);
184 int16x8_t vprod2x5 = vmull_s8(vb5x0, va2x0);
185 int16x8_t vprod3x5 = vmull_s8(vb5x0, va3x0);
186 vprod0x5 = vmlal_s8(vprod0x5, vb5x1, va0x1);
187 vprod1x5 = vmlal_s8(vprod1x5, vb5x1, va1x1);
188 vprod2x5 = vmlal_s8(vprod2x5, vb5x1, va2x1);
189 vprod3x5 = vmlal_s8(vprod3x5, vb5x1, va3x1);
190 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
191 vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
192 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5);
193 vacc3x5 = vpadalq_s16(vacc3x5, vprod3x5);
194 const int8x8_t vb6x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
195 int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0);
196 int16x8_t vprod1x6 = vmull_s8(vb6x0, va1x0);
197 int16x8_t vprod2x6 = vmull_s8(vb6x0, va2x0);
198 int16x8_t vprod3x6 = vmull_s8(vb6x0, va3x0);
199 vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1);
200 vprod1x6 = vmlal_s8(vprod1x6, vb6x1, va1x1);
201 vprod2x6 = vmlal_s8(vprod2x6, vb6x1, va2x1);
202 vprod3x6 = vmlal_s8(vprod3x6, vb6x1, va3x1);
203 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
204 vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
205 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6);
206 vacc3x6 = vpadalq_s16(vacc3x6, vprod3x6);
207 const int8x8_t vb7x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof( int8_t));
208 int16x8_t vprod0x7 = vmull_s8(vb7x0, va0x0);
209 int16x8_t vprod1x7 = vmull_s8(vb7x0, va1x0);
210 int16x8_t vprod2x7 = vmull_s8(vb7x0, va2x0);
211 int16x8_t vprod3x7 = vmull_s8(vb7x0, va3x0);
212 vprod0x7 = vmlal_s8(vprod0x7, vb7x1, va0x1);
213 vprod1x7 = vmlal_s8(vprod1x7, vb7x1, va1x1);
214 vprod2x7 = vmlal_s8(vprod2x7, vb7x1, va2x1);
215 vprod3x7 = vmlal_s8(vprod3x7, vb7x1, va3x1);
216 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
217 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
218 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7);
219 vacc3x7 = vpadalq_s16(vacc3x7, vprod3x7);
220
221 k -= 16 * sizeof(int8_t);
222 }
223
224 // Handle 8 bytes at a time using MUL.
225 if (k != 0) {
226 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
227 const int8x8_t va1 = vld1_s8(a1); a1 += 8;
228 const int8x8_t va2 = vld1_s8(a2); a2 += 8;
229 const int8x8_t va3 = vld1_s8(a3); a3 += 8;
230
231 const int8x8_t vb0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
232 const int16x8_t vprod0x0 = vmull_s8(vb0, va0);
233 const int16x8_t vprod1x0 = vmull_s8(vb0, va1);
234 const int16x8_t vprod2x0 = vmull_s8(vb0, va2);
235 const int16x8_t vprod3x0 = vmull_s8(vb0, va3);
236 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
237 vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
238 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0);
239 vacc3x0 = vpadalq_s16(vacc3x0, vprod3x0);
240 const int8x8_t vb1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
241 const int16x8_t vprod0x1 = vmull_s8(vb1, va0);
242 const int16x8_t vprod1x1 = vmull_s8(vb1, va1);
243 const int16x8_t vprod2x1 = vmull_s8(vb1, va2);
244 const int16x8_t vprod3x1 = vmull_s8(vb1, va3);
245 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
246 vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
247 vacc2x1 = vpadalq_s16(vacc2x1, vprod2x1);
248 vacc3x1 = vpadalq_s16(vacc3x1, vprod3x1);
249 const int8x8_t vb2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
250 const int16x8_t vprod0x2 = vmull_s8(vb2, va0);
251 const int16x8_t vprod1x2 = vmull_s8(vb2, va1);
252 const int16x8_t vprod2x2 = vmull_s8(vb2, va2);
253 const int16x8_t vprod3x2 = vmull_s8(vb2, va3);
254 vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
255 vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
256 vacc2x2 = vpadalq_s16(vacc2x2, vprod2x2);
257 vacc3x2 = vpadalq_s16(vacc3x2, vprod3x2);
258 const int8x8_t vb3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
259 const int16x8_t vprod0x3 = vmull_s8(vb3, va0);
260 const int16x8_t vprod1x3 = vmull_s8(vb3, va1);
261 const int16x8_t vprod2x3 = vmull_s8(vb3, va2);
262 const int16x8_t vprod3x3 = vmull_s8(vb3, va3);
263 vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
264 vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
265 vacc2x3 = vpadalq_s16(vacc2x3, vprod2x3);
266 vacc3x3 = vpadalq_s16(vacc3x3, vprod3x3);
267 const int8x8_t vb4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
268 const int16x8_t vprod0x4 = vmull_s8(vb4, va0);
269 const int16x8_t vprod1x4 = vmull_s8(vb4, va1);
270 const int16x8_t vprod2x4 = vmull_s8(vb4, va2);
271 const int16x8_t vprod3x4 = vmull_s8(vb4, va3);
272 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
273 vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
274 vacc2x4 = vpadalq_s16(vacc2x4, vprod2x4);
275 vacc3x4 = vpadalq_s16(vacc3x4, vprod3x4);
276 const int8x8_t vb5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
277 const int16x8_t vprod0x5 = vmull_s8(vb5, va0);
278 const int16x8_t vprod1x5 = vmull_s8(vb5, va1);
279 const int16x8_t vprod2x5 = vmull_s8(vb5, va2);
280 const int16x8_t vprod3x5 = vmull_s8(vb5, va3);
281 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
282 vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
283 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5);
284 vacc3x5 = vpadalq_s16(vacc3x5, vprod3x5);
285 const int8x8_t vb6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
286 const int16x8_t vprod0x6 = vmull_s8(vb6, va0);
287 const int16x8_t vprod1x6 = vmull_s8(vb6, va1);
288 const int16x8_t vprod2x6 = vmull_s8(vb6, va2);
289 const int16x8_t vprod3x6 = vmull_s8(vb6, va3);
290 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
291 vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
292 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6);
293 vacc3x6 = vpadalq_s16(vacc3x6, vprod3x6);
294 const int8x8_t vb7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
295 const int16x8_t vprod0x7 = vmull_s8(vb7, va0);
296 const int16x8_t vprod1x7 = vmull_s8(vb7, va1);
297 const int16x8_t vprod2x7 = vmull_s8(vb7, va2);
298 const int16x8_t vprod3x7 = vmull_s8(vb7, va3);
299 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
300 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
301 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7);
302 vacc3x7 = vpadalq_s16(vacc3x7, vprod3x7);
303
304 k -= 8 * sizeof(int8_t);
305 }
306
307 #if XNN_ARCH_ARM64
308 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1);
309 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3);
310 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5);
311 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7);
312 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1);
313 const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3);
314 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5);
315 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7);
316 const int32x4_t vsum2x01 = vpaddq_s32(vacc2x0, vacc2x1);
317 const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3);
318 const int32x4_t vsum2x45 = vpaddq_s32(vacc2x4, vacc2x5);
319 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7);
320 const int32x4_t vsum3x01 = vpaddq_s32(vacc3x0, vacc3x1);
321 const int32x4_t vsum3x23 = vpaddq_s32(vacc3x2, vacc3x3);
322 const int32x4_t vsum3x45 = vpaddq_s32(vacc3x4, vacc3x5);
323 const int32x4_t vsum3x67 = vpaddq_s32(vacc3x6, vacc3x7);
324
325 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23);
326 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67);
327 int32x4_t vacc1x0123 = vpaddq_s32(vsum1x01, vsum1x23);
328 int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67);
329 int32x4_t vacc2x0123 = vpaddq_s32(vsum2x01, vsum2x23);
330 int32x4_t vacc2x4567 = vpaddq_s32(vsum2x45, vsum2x67);
331 int32x4_t vacc3x0123 = vpaddq_s32(vsum3x01, vsum3x23);
332 int32x4_t vacc3x4567 = vpaddq_s32(vsum3x45, vsum3x67);
333 #else
334 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0));
335 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1));
336 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2));
337 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3));
338 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1);
339 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3);
340 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 );
341 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4));
342 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5));
343 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6));
344 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7));
345 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5);
346 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7);
347 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 );
348 const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0));
349 const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1));
350 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2));
351 const int32x2_t vpsum1x3 = vadd_s32(vget_low_s32(vacc1x3), vget_high_s32(vacc1x3));
352 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1);
353 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3);
354 int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23 );
355 const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4));
356 const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(vacc1x5));
357 const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6));
358 const int32x2_t vpsum1x7 = vadd_s32(vget_low_s32(vacc1x7), vget_high_s32(vacc1x7));
359 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5);
360 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7);
361 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 );
362 const int32x2_t vpsum2x0 = vadd_s32(vget_low_s32(vacc2x0), vget_high_s32(vacc2x0));
363 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1));
364 const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2));
365 const int32x2_t vpsum2x3 = vadd_s32(vget_low_s32(vacc2x3), vget_high_s32(vacc2x3));
366 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1);
367 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3);
368 int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23 );
369 const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4));
370 const int32x2_t vpsum2x5 = vadd_s32(vget_low_s32(vacc2x5), vget_high_s32(vacc2x5));
371 const int32x2_t vpsum2x6 = vadd_s32(vget_low_s32(vacc2x6), vget_high_s32(vacc2x6));
372 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7));
373 const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5);
374 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7);
375 int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67 );
376 const int32x2_t vpsum3x0 = vadd_s32(vget_low_s32(vacc3x0), vget_high_s32(vacc3x0));
377 const int32x2_t vpsum3x1 = vadd_s32(vget_low_s32(vacc3x1), vget_high_s32(vacc3x1));
378 const int32x2_t vpsum3x2 = vadd_s32(vget_low_s32(vacc3x2), vget_high_s32(vacc3x2));
379 const int32x2_t vpsum3x3 = vadd_s32(vget_low_s32(vacc3x3), vget_high_s32(vacc3x3));
380 const int32x2_t vsum3x01 = vpadd_s32(vpsum3x0, vpsum3x1);
381 const int32x2_t vsum3x23 = vpadd_s32(vpsum3x2, vpsum3x3);
382 int32x4_t vacc3x0123 = vcombine_s32(vsum3x01, vsum3x23 );
383 const int32x2_t vpsum3x4 = vadd_s32(vget_low_s32(vacc3x4), vget_high_s32(vacc3x4));
384 const int32x2_t vpsum3x5 = vadd_s32(vget_low_s32(vacc3x5), vget_high_s32(vacc3x5));
385 const int32x2_t vpsum3x6 = vadd_s32(vget_low_s32(vacc3x6), vget_high_s32(vacc3x6));
386 const int32x2_t vpsum3x7 = vadd_s32(vget_low_s32(vacc3x7), vget_high_s32(vacc3x7));
387 const int32x2_t vsum3x45 = vpadd_s32(vpsum3x4, vpsum3x5);
388 const int32x2_t vsum3x67 = vpadd_s32(vpsum3x6, vpsum3x7);
389 int32x4_t vacc3x4567 = vcombine_s32(vsum3x45, vsum3x67 );
390 #endif
391
392 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
393 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
394 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
395
396 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
397 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
398 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
399 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
400 vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
401 vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
402 vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
403 vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
404
405 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
406 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
407 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
408 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
409 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
410 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
411 vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
412 vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
413
414 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
415 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
416 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
417 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
418 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
419 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
420 vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
421 vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
422
423 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
424 #if XNN_ARCH_ARM64
425 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
426 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
427 int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
428 int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
429
430 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
431 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
432 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
433 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
434
435 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
436 int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567);
437 #else
438 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
439 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
440 int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
441 int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
442
443 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
444 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
445 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
446 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
447
448 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
449 int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567));
450 #endif
451
452 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
453 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
454 vout2x01234567_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, voutput_min);
455
456 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
457 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
458 vout2x01234567_3x01234567 = vminq_s8(vout2x01234567_3x01234567, voutput_max);
459
460 if (nc >= 8) {
461 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
462 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
463 vst1_s8(c2 + 0, vget_low_s8(vout2x01234567_3x01234567));
464 vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567));
465
466 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
467 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
468 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
469 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
470
471 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
472 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
473 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
474 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
475
476 nc -= 8;
477 } else {
478 // Final case where not all of the 8 columns fit in the destination.
479 if (nc & 4) {
480 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
481 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
482 vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
483 vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
484 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
485 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
486 }
487 if (nc & 2) {
488 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
489 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
490 vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
491 vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
492 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
493 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
494 }
495 if (nc & 1) {
496 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
497 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
498 vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
499 vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
500 }
501
502 nc = 0;
503 }
504 } while (nc != 0);
505 }
506