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_3x16c8__neon_mull(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_3x16c8__neon_mull(
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 <= 3);
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
55 do {
56 int32x4_t vacc0x0 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
57 int32x4_t vacc0x1 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
58 int32x4_t vacc0x2 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
59 int32x4_t vacc0x3 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
60 int32x4_t vacc0x4 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
61 int32x4_t vacc0x5 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
62 int32x4_t vacc0x6 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
63 int32x4_t vacc0x7 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
64 int32x4_t vacc0x8 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
65 int32x4_t vacc0x9 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
66 int32x4_t vacc0x10 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
67 int32x4_t vacc0x11 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
68 int32x4_t vacc0x12 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
69 int32x4_t vacc0x13 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
70 int32x4_t vacc0x14 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
71 int32x4_t vacc0x15 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
72 int32x4_t vacc1x0 = vacc0x0;
73 int32x4_t vacc1x1 = vacc0x1;
74 int32x4_t vacc1x2 = vacc0x2;
75 int32x4_t vacc1x3 = vacc0x3;
76 int32x4_t vacc1x4 = vacc0x4;
77 int32x4_t vacc1x5 = vacc0x5;
78 int32x4_t vacc1x6 = vacc0x6;
79 int32x4_t vacc1x7 = vacc0x7;
80 int32x4_t vacc1x8 = vacc0x8;
81 int32x4_t vacc1x9 = vacc0x9;
82 int32x4_t vacc1x10 = vacc0x10;
83 int32x4_t vacc1x11 = vacc0x11;
84 int32x4_t vacc1x12 = vacc0x12;
85 int32x4_t vacc1x13 = vacc0x13;
86 int32x4_t vacc1x14 = vacc0x14;
87 int32x4_t vacc1x15 = vacc0x15;
88 int32x4_t vacc2x0 = vacc0x0;
89 int32x4_t vacc2x1 = vacc0x1;
90 int32x4_t vacc2x2 = vacc0x2;
91 int32x4_t vacc2x3 = vacc0x3;
92 int32x4_t vacc2x4 = vacc0x4;
93 int32x4_t vacc2x5 = vacc0x5;
94 int32x4_t vacc2x6 = vacc0x6;
95 int32x4_t vacc2x7 = vacc0x7;
96 int32x4_t vacc2x8 = vacc0x8;
97 int32x4_t vacc2x9 = vacc0x9;
98 int32x4_t vacc2x10 = vacc0x10;
99 int32x4_t vacc2x11 = vacc0x11;
100 int32x4_t vacc2x12 = vacc0x12;
101 int32x4_t vacc2x13 = vacc0x13;
102 int32x4_t vacc2x14 = vacc0x14;
103 int32x4_t vacc2x15 = vacc0x15;
104
105 size_t k = kc;
106
107 // Handle 8 bytes at a time using MUL.
108 while (k != 0) {
109 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
110 const int8x8_t va1 = vld1_s8(a1); a1 += 8;
111 const int8x8_t va2 = vld1_s8(a2); a2 += 8;
112
113 const int8x8_t vb0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
114 const int16x8_t vprod0x0 = vmull_s8(vb0, va0);
115 const int16x8_t vprod1x0 = vmull_s8(vb0, va1);
116 const int16x8_t vprod2x0 = vmull_s8(vb0, va2);
117 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
118 vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
119 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0);
120 const int8x8_t vb1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
121 const int16x8_t vprod0x1 = vmull_s8(vb1, va0);
122 const int16x8_t vprod1x1 = vmull_s8(vb1, va1);
123 const int16x8_t vprod2x1 = vmull_s8(vb1, va2);
124 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
125 vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
126 vacc2x1 = vpadalq_s16(vacc2x1, vprod2x1);
127 const int8x8_t vb2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
128 const int16x8_t vprod0x2 = vmull_s8(vb2, va0);
129 const int16x8_t vprod1x2 = vmull_s8(vb2, va1);
130 const int16x8_t vprod2x2 = vmull_s8(vb2, va2);
131 vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
132 vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
133 vacc2x2 = vpadalq_s16(vacc2x2, vprod2x2);
134 const int8x8_t vb3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
135 const int16x8_t vprod0x3 = vmull_s8(vb3, va0);
136 const int16x8_t vprod1x3 = vmull_s8(vb3, va1);
137 const int16x8_t vprod2x3 = vmull_s8(vb3, va2);
138 vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
139 vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
140 vacc2x3 = vpadalq_s16(vacc2x3, vprod2x3);
141 const int8x8_t vb4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
142 const int16x8_t vprod0x4 = vmull_s8(vb4, va0);
143 const int16x8_t vprod1x4 = vmull_s8(vb4, va1);
144 const int16x8_t vprod2x4 = vmull_s8(vb4, va2);
145 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
146 vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
147 vacc2x4 = vpadalq_s16(vacc2x4, vprod2x4);
148 const int8x8_t vb5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
149 const int16x8_t vprod0x5 = vmull_s8(vb5, va0);
150 const int16x8_t vprod1x5 = vmull_s8(vb5, va1);
151 const int16x8_t vprod2x5 = vmull_s8(vb5, va2);
152 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
153 vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
154 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5);
155 const int8x8_t vb6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
156 const int16x8_t vprod0x6 = vmull_s8(vb6, va0);
157 const int16x8_t vprod1x6 = vmull_s8(vb6, va1);
158 const int16x8_t vprod2x6 = vmull_s8(vb6, va2);
159 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
160 vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
161 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6);
162 const int8x8_t vb7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
163 const int16x8_t vprod0x7 = vmull_s8(vb7, va0);
164 const int16x8_t vprod1x7 = vmull_s8(vb7, va1);
165 const int16x8_t vprod2x7 = vmull_s8(vb7, va2);
166 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
167 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
168 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7);
169 const int8x8_t vb8 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
170 const int16x8_t vprod0x8 = vmull_s8(vb8, va0);
171 const int16x8_t vprod1x8 = vmull_s8(vb8, va1);
172 const int16x8_t vprod2x8 = vmull_s8(vb8, va2);
173 vacc0x8 = vpadalq_s16(vacc0x8, vprod0x8);
174 vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8);
175 vacc2x8 = vpadalq_s16(vacc2x8, vprod2x8);
176 const int8x8_t vb9 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
177 const int16x8_t vprod0x9 = vmull_s8(vb9, va0);
178 const int16x8_t vprod1x9 = vmull_s8(vb9, va1);
179 const int16x8_t vprod2x9 = vmull_s8(vb9, va2);
180 vacc0x9 = vpadalq_s16(vacc0x9, vprod0x9);
181 vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9);
182 vacc2x9 = vpadalq_s16(vacc2x9, vprod2x9);
183 const int8x8_t vb10 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
184 const int16x8_t vprod0x10 = vmull_s8(vb10, va0);
185 const int16x8_t vprod1x10 = vmull_s8(vb10, va1);
186 const int16x8_t vprod2x10 = vmull_s8(vb10, va2);
187 vacc0x10 = vpadalq_s16(vacc0x10, vprod0x10);
188 vacc1x10 = vpadalq_s16(vacc1x10, vprod1x10);
189 vacc2x10 = vpadalq_s16(vacc2x10, vprod2x10);
190 const int8x8_t vb11 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
191 const int16x8_t vprod0x11 = vmull_s8(vb11, va0);
192 const int16x8_t vprod1x11 = vmull_s8(vb11, va1);
193 const int16x8_t vprod2x11 = vmull_s8(vb11, va2);
194 vacc0x11 = vpadalq_s16(vacc0x11, vprod0x11);
195 vacc1x11 = vpadalq_s16(vacc1x11, vprod1x11);
196 vacc2x11 = vpadalq_s16(vacc2x11, vprod2x11);
197 const int8x8_t vb12 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
198 const int16x8_t vprod0x12 = vmull_s8(vb12, va0);
199 const int16x8_t vprod1x12 = vmull_s8(vb12, va1);
200 const int16x8_t vprod2x12 = vmull_s8(vb12, va2);
201 vacc0x12 = vpadalq_s16(vacc0x12, vprod0x12);
202 vacc1x12 = vpadalq_s16(vacc1x12, vprod1x12);
203 vacc2x12 = vpadalq_s16(vacc2x12, vprod2x12);
204 const int8x8_t vb13 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
205 const int16x8_t vprod0x13 = vmull_s8(vb13, va0);
206 const int16x8_t vprod1x13 = vmull_s8(vb13, va1);
207 const int16x8_t vprod2x13 = vmull_s8(vb13, va2);
208 vacc0x13 = vpadalq_s16(vacc0x13, vprod0x13);
209 vacc1x13 = vpadalq_s16(vacc1x13, vprod1x13);
210 vacc2x13 = vpadalq_s16(vacc2x13, vprod2x13);
211 const int8x8_t vb14 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
212 const int16x8_t vprod0x14 = vmull_s8(vb14, va0);
213 const int16x8_t vprod1x14 = vmull_s8(vb14, va1);
214 const int16x8_t vprod2x14 = vmull_s8(vb14, va2);
215 vacc0x14 = vpadalq_s16(vacc0x14, vprod0x14);
216 vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14);
217 vacc2x14 = vpadalq_s16(vacc2x14, vprod2x14);
218 const int8x8_t vb15 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
219 const int16x8_t vprod0x15 = vmull_s8(vb15, va0);
220 const int16x8_t vprod1x15 = vmull_s8(vb15, va1);
221 const int16x8_t vprod2x15 = vmull_s8(vb15, va2);
222 vacc0x15 = vpadalq_s16(vacc0x15, vprod0x15);
223 vacc1x15 = vpadalq_s16(vacc1x15, vprod1x15);
224 vacc2x15 = vpadalq_s16(vacc2x15, vprod2x15);
225
226 k -= 8 * sizeof(int8_t);
227 }
228
229 #if XNN_ARCH_ARM64
230 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1);
231 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3);
232 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5);
233 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7);
234 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9);
235 const int32x4_t vsum0xAB = vpaddq_s32(vacc0x10, vacc0x11);
236 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13);
237 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15);
238 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1);
239 const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3);
240 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5);
241 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7);
242 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9);
243 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11);
244 const int32x4_t vsum1xCD = vpaddq_s32(vacc1x12, vacc1x13);
245 const int32x4_t vsum1xEF = vpaddq_s32(vacc1x14, vacc1x15);
246 const int32x4_t vsum2x01 = vpaddq_s32(vacc2x0, vacc2x1);
247 const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3);
248 const int32x4_t vsum2x45 = vpaddq_s32(vacc2x4, vacc2x5);
249 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7);
250 const int32x4_t vsum2x89 = vpaddq_s32(vacc2x8, vacc2x9);
251 const int32x4_t vsum2xAB = vpaddq_s32(vacc2x10, vacc2x11);
252 const int32x4_t vsum2xCD = vpaddq_s32(vacc2x12, vacc2x13);
253 const int32x4_t vsum2xEF = vpaddq_s32(vacc2x14, vacc2x15);
254
255 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23);
256 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67);
257 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB);
258 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF);
259 int32x4_t vacc1x0123 = vpaddq_s32(vsum1x01, vsum1x23);
260 int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67);
261 int32x4_t vacc1x89AB = vpaddq_s32(vsum1x89, vsum1xAB);
262 int32x4_t vacc1xCDEF = vpaddq_s32(vsum1xCD, vsum1xEF);
263 int32x4_t vacc2x0123 = vpaddq_s32(vsum2x01, vsum2x23);
264 int32x4_t vacc2x4567 = vpaddq_s32(vsum2x45, vsum2x67);
265 int32x4_t vacc2x89AB = vpaddq_s32(vsum2x89, vsum2xAB);
266 int32x4_t vacc2xCDEF = vpaddq_s32(vsum2xCD, vsum2xEF);
267 #else
268 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0));
269 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1));
270 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2));
271 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3));
272 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1);
273 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3);
274 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 );
275 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4));
276 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5));
277 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6));
278 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7));
279 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5);
280 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7);
281 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 );
282 const int32x2_t vpsum0x8 = vadd_s32(vget_low_s32(vacc0x8), vget_high_s32(vacc0x8));
283 const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9));
284 const int32x2_t vpsum0xA = vadd_s32(vget_low_s32(vacc0x10), vget_high_s32(vacc0x10));
285 const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11));
286 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9);
287 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB);
288 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB );
289 const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12));
290 const int32x2_t vpsum0xD = vadd_s32(vget_low_s32(vacc0x13), vget_high_s32(vacc0x13));
291 const int32x2_t vpsum0xE = vadd_s32(vget_low_s32(vacc0x14), vget_high_s32(vacc0x14));
292 const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(vacc0x15));
293 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD);
294 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF);
295 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF );
296 const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0));
297 const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1));
298 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2));
299 const int32x2_t vpsum1x3 = vadd_s32(vget_low_s32(vacc1x3), vget_high_s32(vacc1x3));
300 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1);
301 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3);
302 int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23 );
303 const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4));
304 const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(vacc1x5));
305 const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6));
306 const int32x2_t vpsum1x7 = vadd_s32(vget_low_s32(vacc1x7), vget_high_s32(vacc1x7));
307 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5);
308 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7);
309 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 );
310 const int32x2_t vpsum1x8 = vadd_s32(vget_low_s32(vacc1x8), vget_high_s32(vacc1x8));
311 const int32x2_t vpsum1x9 = vadd_s32(vget_low_s32(vacc1x9), vget_high_s32(vacc1x9));
312 const int32x2_t vpsum1xA = vadd_s32(vget_low_s32(vacc1x10), vget_high_s32(vacc1x10));
313 const int32x2_t vpsum1xB = vadd_s32(vget_low_s32(vacc1x11), vget_high_s32(vacc1x11));
314 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9);
315 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB);
316 int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB );
317 const int32x2_t vpsum1xC = vadd_s32(vget_low_s32(vacc1x12), vget_high_s32(vacc1x12));
318 const int32x2_t vpsum1xD = vadd_s32(vget_low_s32(vacc1x13), vget_high_s32(vacc1x13));
319 const int32x2_t vpsum1xE = vadd_s32(vget_low_s32(vacc1x14), vget_high_s32(vacc1x14));
320 const int32x2_t vpsum1xF = vadd_s32(vget_low_s32(vacc1x15), vget_high_s32(vacc1x15));
321 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD);
322 const int32x2_t vsum1xEF = vpadd_s32(vpsum1xE, vpsum1xF);
323 int32x4_t vacc1xCDEF = vcombine_s32(vsum1xCD, vsum1xEF );
324 const int32x2_t vpsum2x0 = vadd_s32(vget_low_s32(vacc2x0), vget_high_s32(vacc2x0));
325 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1));
326 const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2));
327 const int32x2_t vpsum2x3 = vadd_s32(vget_low_s32(vacc2x3), vget_high_s32(vacc2x3));
328 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1);
329 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3);
330 int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23 );
331 const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4));
332 const int32x2_t vpsum2x5 = vadd_s32(vget_low_s32(vacc2x5), vget_high_s32(vacc2x5));
333 const int32x2_t vpsum2x6 = vadd_s32(vget_low_s32(vacc2x6), vget_high_s32(vacc2x6));
334 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7));
335 const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5);
336 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7);
337 int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67 );
338 const int32x2_t vpsum2x8 = vadd_s32(vget_low_s32(vacc2x8), vget_high_s32(vacc2x8));
339 const int32x2_t vpsum2x9 = vadd_s32(vget_low_s32(vacc2x9), vget_high_s32(vacc2x9));
340 const int32x2_t vpsum2xA = vadd_s32(vget_low_s32(vacc2x10), vget_high_s32(vacc2x10));
341 const int32x2_t vpsum2xB = vadd_s32(vget_low_s32(vacc2x11), vget_high_s32(vacc2x11));
342 const int32x2_t vsum2x89 = vpadd_s32(vpsum2x8, vpsum2x9);
343 const int32x2_t vsum2xAB = vpadd_s32(vpsum2xA, vpsum2xB);
344 int32x4_t vacc2x89AB = vcombine_s32(vsum2x89, vsum2xAB );
345 const int32x2_t vpsum2xC = vadd_s32(vget_low_s32(vacc2x12), vget_high_s32(vacc2x12));
346 const int32x2_t vpsum2xD = vadd_s32(vget_low_s32(vacc2x13), vget_high_s32(vacc2x13));
347 const int32x2_t vpsum2xE = vadd_s32(vget_low_s32(vacc2x14), vget_high_s32(vacc2x14));
348 const int32x2_t vpsum2xF = vadd_s32(vget_low_s32(vacc2x15), vget_high_s32(vacc2x15));
349 const int32x2_t vsum2xCD = vpadd_s32(vpsum2xC, vpsum2xD);
350 const int32x2_t vsum2xEF = vpadd_s32(vpsum2xE, vpsum2xF);
351 int32x4_t vacc2xCDEF = vcombine_s32(vsum2xCD, vsum2xEF );
352 #endif
353
354 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
355 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
356 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
357
358 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
359 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
360 vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
361 vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
362 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
363 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
364 vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
365 vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
366 vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
367 vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
368 vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
369 vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
370
371 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
372 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
373 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
374 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
375 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
376 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
377 vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
378 vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
379 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
380 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
381 vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
382 vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
383
384 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
385 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
386 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
387 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
388 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
389 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
390 vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
391 vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
392 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
393 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
394 vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
395 vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
396
397 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
398 #if XNN_ARCH_ARM64
399 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
400 int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
401 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
402 int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
403 int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
404 int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
405
406 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
407 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
408 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
409 vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
410 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
411 vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
412
413 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
414 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
415 int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
416 #else
417 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
418 int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
419 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
420 int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
421 int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
422 int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
423
424 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
425 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
426 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
427 vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
428 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
429 vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
430
431 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
432 int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
433 int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
434 #endif
435
436 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
437 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
438 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
439 vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
440
441 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
442 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
443 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
444 vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
445
446 if (nc >= 16) {
447 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
448 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
449 vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
450
451 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
452 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
453 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
454
455 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
456 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
457 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
458
459 nc -= 16;
460 } else {
461 // Final case where not all of the 16 columns fit in the destination.
462 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
463 int8x8_t vout2x01234567 = vget_low_s8(vout2x0123456789ABCDEF);
464 if (nc & 8) {
465 vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
466 vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
467 vst1_s8(c2, vout2x01234567); c2 += 8;
468 vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
469 vout2x01234567 = vget_high_s8(vout2x0123456789ABCDEF);
470 }
471 if (nc & 4) {
472 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
473 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
474 vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
475 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
476 vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
477 }
478 if (nc & 2) {
479 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
480 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
481 vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
482 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
483 vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
484 }
485 if (nc & 1) {
486 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
487 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
488 vst1_lane_s8(c2, vout2x01234567, 0);
489 }
490
491 nc = 0;
492 }
493 } while (nc != 0);
494 }
495