1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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/igemm.h>
15 #include <xnnpack/math.h>
16
17
xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const int8_t** restrict a,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 2);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (2 * 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 kc = round_up_po2(kc, 8 * sizeof(int8_t));
44 int8_t* c0 = c;
45 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr != 2) {
47 c1 = c0;
48 }
49
50 do {
51 int32x4_t vacc0x0 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
52 int32x4_t vacc0x1 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
53 int32x4_t vacc0x2 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
54 int32x4_t vacc0x3 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
55 int32x4_t vacc0x4 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
56 int32x4_t vacc0x5 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
57 int32x4_t vacc0x6 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
58 int32x4_t vacc0x7 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
59 int32x4_t vacc0x8 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
60 int32x4_t vacc0x9 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
61 int32x4_t vacc0x10 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
62 int32x4_t vacc0x11 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
63 int32x4_t vacc0x12 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
64 int32x4_t vacc0x13 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
65 int32x4_t vacc0x14 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
66 int32x4_t vacc0x15 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
67 int32x4_t vacc1x0 = vacc0x0;
68 int32x4_t vacc1x1 = vacc0x1;
69 int32x4_t vacc1x2 = vacc0x2;
70 int32x4_t vacc1x3 = vacc0x3;
71 int32x4_t vacc1x4 = vacc0x4;
72 int32x4_t vacc1x5 = vacc0x5;
73 int32x4_t vacc1x6 = vacc0x6;
74 int32x4_t vacc1x7 = vacc0x7;
75 int32x4_t vacc1x8 = vacc0x8;
76 int32x4_t vacc1x9 = vacc0x9;
77 int32x4_t vacc1x10 = vacc0x10;
78 int32x4_t vacc1x11 = vacc0x11;
79 int32x4_t vacc1x12 = vacc0x12;
80 int32x4_t vacc1x13 = vacc0x13;
81 int32x4_t vacc1x14 = vacc0x14;
82 int32x4_t vacc1x15 = vacc0x15;
83
84 size_t p = ks;
85 do {
86 const int8_t* restrict a0 = a[0];
87 if XNN_UNPREDICTABLE(a0 != zero) {
88 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
89 }
90 const int8_t* restrict a1 = a[1];
91 if XNN_UNPREDICTABLE(a1 != zero) {
92 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
93 }
94 a += 2;
95
96 size_t k = kc;
97
98 // Handle 8 bytes at a time using MUL.
99 while (k != 0) {
100 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
101 const int8x8_t va1 = vld1_s8(a1); a1 += 8;
102
103 const int8x8_t vb0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
104 const int16x8_t vprod0x0 = vmull_s8(vb0, va0);
105 const int16x8_t vprod1x0 = vmull_s8(vb0, va1);
106 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
107 vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
108 const int8x8_t vb1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
109 const int16x8_t vprod0x1 = vmull_s8(vb1, va0);
110 const int16x8_t vprod1x1 = vmull_s8(vb1, va1);
111 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
112 vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
113 const int8x8_t vb2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
114 const int16x8_t vprod0x2 = vmull_s8(vb2, va0);
115 const int16x8_t vprod1x2 = vmull_s8(vb2, va1);
116 vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
117 vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
118 const int8x8_t vb3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
119 const int16x8_t vprod0x3 = vmull_s8(vb3, va0);
120 const int16x8_t vprod1x3 = vmull_s8(vb3, va1);
121 vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
122 vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
123 const int8x8_t vb4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
124 const int16x8_t vprod0x4 = vmull_s8(vb4, va0);
125 const int16x8_t vprod1x4 = vmull_s8(vb4, va1);
126 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
127 vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
128 const int8x8_t vb5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
129 const int16x8_t vprod0x5 = vmull_s8(vb5, va0);
130 const int16x8_t vprod1x5 = vmull_s8(vb5, va1);
131 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
132 vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
133 const int8x8_t vb6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
134 const int16x8_t vprod0x6 = vmull_s8(vb6, va0);
135 const int16x8_t vprod1x6 = vmull_s8(vb6, va1);
136 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
137 vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
138 const int8x8_t vb7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
139 const int16x8_t vprod0x7 = vmull_s8(vb7, va0);
140 const int16x8_t vprod1x7 = vmull_s8(vb7, va1);
141 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
142 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
143 const int8x8_t vb8 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
144 const int16x8_t vprod0x8 = vmull_s8(vb8, va0);
145 const int16x8_t vprod1x8 = vmull_s8(vb8, va1);
146 vacc0x8 = vpadalq_s16(vacc0x8, vprod0x8);
147 vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8);
148 const int8x8_t vb9 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
149 const int16x8_t vprod0x9 = vmull_s8(vb9, va0);
150 const int16x8_t vprod1x9 = vmull_s8(vb9, va1);
151 vacc0x9 = vpadalq_s16(vacc0x9, vprod0x9);
152 vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9);
153 const int8x8_t vb10 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
154 const int16x8_t vprod0x10 = vmull_s8(vb10, va0);
155 const int16x8_t vprod1x10 = vmull_s8(vb10, va1);
156 vacc0x10 = vpadalq_s16(vacc0x10, vprod0x10);
157 vacc1x10 = vpadalq_s16(vacc1x10, vprod1x10);
158 const int8x8_t vb11 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
159 const int16x8_t vprod0x11 = vmull_s8(vb11, va0);
160 const int16x8_t vprod1x11 = vmull_s8(vb11, va1);
161 vacc0x11 = vpadalq_s16(vacc0x11, vprod0x11);
162 vacc1x11 = vpadalq_s16(vacc1x11, vprod1x11);
163 const int8x8_t vb12 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
164 const int16x8_t vprod0x12 = vmull_s8(vb12, va0);
165 const int16x8_t vprod1x12 = vmull_s8(vb12, va1);
166 vacc0x12 = vpadalq_s16(vacc0x12, vprod0x12);
167 vacc1x12 = vpadalq_s16(vacc1x12, vprod1x12);
168 const int8x8_t vb13 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
169 const int16x8_t vprod0x13 = vmull_s8(vb13, va0);
170 const int16x8_t vprod1x13 = vmull_s8(vb13, va1);
171 vacc0x13 = vpadalq_s16(vacc0x13, vprod0x13);
172 vacc1x13 = vpadalq_s16(vacc1x13, vprod1x13);
173 const int8x8_t vb14 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
174 const int16x8_t vprod0x14 = vmull_s8(vb14, va0);
175 const int16x8_t vprod1x14 = vmull_s8(vb14, va1);
176 vacc0x14 = vpadalq_s16(vacc0x14, vprod0x14);
177 vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14);
178 const int8x8_t vb15 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
179 const int16x8_t vprod0x15 = vmull_s8(vb15, va0);
180 const int16x8_t vprod1x15 = vmull_s8(vb15, va1);
181 vacc0x15 = vpadalq_s16(vacc0x15, vprod0x15);
182 vacc1x15 = vpadalq_s16(vacc1x15, vprod1x15);
183
184 k -= 8 * sizeof(int8_t);
185 }
186
187 p -= 2 * sizeof(void*);
188 } while (p != 0);
189
190 #if XNN_ARCH_ARM64
191 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1);
192 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3);
193 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5);
194 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7);
195 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9);
196 const int32x4_t vsum0xAB = vpaddq_s32(vacc0x10, vacc0x11);
197 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13);
198 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15);
199 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1);
200 const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3);
201 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5);
202 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7);
203 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9);
204 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11);
205 const int32x4_t vsum1xCD = vpaddq_s32(vacc1x12, vacc1x13);
206 const int32x4_t vsum1xEF = vpaddq_s32(vacc1x14, vacc1x15);
207
208 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23);
209 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67);
210 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB);
211 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF);
212 int32x4_t vacc1x0123 = vpaddq_s32(vsum1x01, vsum1x23);
213 int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67);
214 int32x4_t vacc1x89AB = vpaddq_s32(vsum1x89, vsum1xAB);
215 int32x4_t vacc1xCDEF = vpaddq_s32(vsum1xCD, vsum1xEF);
216 #else
217 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0));
218 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1));
219 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2));
220 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3));
221 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1);
222 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3);
223 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 );
224 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4));
225 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5));
226 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6));
227 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7));
228 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5);
229 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7);
230 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 );
231 const int32x2_t vpsum0x8 = vadd_s32(vget_low_s32(vacc0x8), vget_high_s32(vacc0x8));
232 const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9));
233 const int32x2_t vpsum0xA = vadd_s32(vget_low_s32(vacc0x10), vget_high_s32(vacc0x10));
234 const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11));
235 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9);
236 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB);
237 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB );
238 const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12));
239 const int32x2_t vpsum0xD = vadd_s32(vget_low_s32(vacc0x13), vget_high_s32(vacc0x13));
240 const int32x2_t vpsum0xE = vadd_s32(vget_low_s32(vacc0x14), vget_high_s32(vacc0x14));
241 const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(vacc0x15));
242 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD);
243 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF);
244 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF );
245 const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0));
246 const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1));
247 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2));
248 const int32x2_t vpsum1x3 = vadd_s32(vget_low_s32(vacc1x3), vget_high_s32(vacc1x3));
249 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1);
250 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3);
251 int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23 );
252 const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4));
253 const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(vacc1x5));
254 const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6));
255 const int32x2_t vpsum1x7 = vadd_s32(vget_low_s32(vacc1x7), vget_high_s32(vacc1x7));
256 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5);
257 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7);
258 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 );
259 const int32x2_t vpsum1x8 = vadd_s32(vget_low_s32(vacc1x8), vget_high_s32(vacc1x8));
260 const int32x2_t vpsum1x9 = vadd_s32(vget_low_s32(vacc1x9), vget_high_s32(vacc1x9));
261 const int32x2_t vpsum1xA = vadd_s32(vget_low_s32(vacc1x10), vget_high_s32(vacc1x10));
262 const int32x2_t vpsum1xB = vadd_s32(vget_low_s32(vacc1x11), vget_high_s32(vacc1x11));
263 const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9);
264 const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB);
265 int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB );
266 const int32x2_t vpsum1xC = vadd_s32(vget_low_s32(vacc1x12), vget_high_s32(vacc1x12));
267 const int32x2_t vpsum1xD = vadd_s32(vget_low_s32(vacc1x13), vget_high_s32(vacc1x13));
268 const int32x2_t vpsum1xE = vadd_s32(vget_low_s32(vacc1x14), vget_high_s32(vacc1x14));
269 const int32x2_t vpsum1xF = vadd_s32(vget_low_s32(vacc1x15), vget_high_s32(vacc1x15));
270 const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD);
271 const int32x2_t vsum1xEF = vpadd_s32(vpsum1xE, vpsum1xF);
272 int32x4_t vacc1xCDEF = vcombine_s32(vsum1xCD, vsum1xEF );
273 #endif
274
275 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
276 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
277 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
278
279 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
280 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
281 vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
282 vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
283 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
284 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
285 vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
286 vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
287
288 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
289 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
290 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
291 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
292 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
293 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
294 vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
295 vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
296
297 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
298 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
299 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
300 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
301 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
302 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
303 vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
304 vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
305
306 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
307 #if XNN_ARCH_ARM64
308 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
309 int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
310 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
311 int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
312
313 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
314 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
315 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
316 vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
317
318 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
319 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
320 #else
321 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
322 int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
323 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
324 int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
325
326 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
327 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
328 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
329 vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
330
331 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
332 int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
333 #endif
334
335 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
336 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
337 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
338
339 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
340 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
341 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
342
343 if (nc >= 16) {
344 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
345 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
346
347 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
348 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
349
350 a = (const int8_t**restrict) ((uintptr_t) a - ks);
351
352 nc -= 16;
353 } else {
354 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
355 if (nc & 8) {
356 vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
357 vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
358 vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
359 }
360 if (nc & 4) {
361 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
362 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
363 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
364 }
365 if (nc & 2) {
366 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
367 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
368 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
369 }
370 if (nc & 1) {
371 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
372 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
373 }
374
375 nc = 0;
376 }
377 } while (nc != 0);
378 }
379