1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c16-neon-mlal.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <arm_neon.h>
13
14 #include <xnnpack/gemm.h>
15 #include <xnnpack/math.h>
16
17
xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const int8_t** restrict a,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 3);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (3 * 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, 16 * 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 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 c2 = c1;
52 }
53
54 do {
55 int32x4_t vacc0x0 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
56 int32x4_t vacc0x1 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
57 int32x4_t vacc0x2 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
58 int32x4_t vacc0x3 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
59 int32x4_t vacc0x4 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
60 int32x4_t vacc0x5 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
61 int32x4_t vacc0x6 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
62 int32x4_t vacc0x7 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
63 int32x4_t vacc1x0 = vacc0x0;
64 int32x4_t vacc1x1 = vacc0x1;
65 int32x4_t vacc1x2 = vacc0x2;
66 int32x4_t vacc1x3 = vacc0x3;
67 int32x4_t vacc1x4 = vacc0x4;
68 int32x4_t vacc1x5 = vacc0x5;
69 int32x4_t vacc1x6 = vacc0x6;
70 int32x4_t vacc1x7 = vacc0x7;
71 int32x4_t vacc2x0 = vacc0x0;
72 int32x4_t vacc2x1 = vacc0x1;
73 int32x4_t vacc2x2 = vacc0x2;
74 int32x4_t vacc2x3 = vacc0x3;
75 int32x4_t vacc2x4 = vacc0x4;
76 int32x4_t vacc2x5 = vacc0x5;
77 int32x4_t vacc2x6 = vacc0x6;
78 int32x4_t vacc2x7 = vacc0x7;
79
80 size_t p = ks;
81 do {
82 const int8_t* restrict a0 = a[0];
83 if XNN_UNPREDICTABLE(a0 != zero) {
84 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
85 }
86 const int8_t* restrict a1 = a[1];
87 if XNN_UNPREDICTABLE(a1 != zero) {
88 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
89 }
90 const int8_t* restrict a2 = a[2];
91 if XNN_UNPREDICTABLE(a2 != zero) {
92 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
93 }
94 a += 3;
95
96 // KC loop of 16 with up to 15 remainder
97 size_t k = kc;
98 while (k != 0) {
99 const int8x16_t va0 = vld1q_s8(a0); a0 += 16;
100 const int8x16_t va1 = vld1q_s8(a1); a1 += 16;
101 const int8x16_t va2 = vld1q_s8(a2); a2 += 16;
102
103 const int8x16_t vb0 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
104 const int8x16_t vb1 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
105 const int8x16_t vb2 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
106 const int8x16_t vb3 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
107 const int8x16_t vb4 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
108 const int8x16_t vb5 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
109 const int8x16_t vb6 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
110 const int8x16_t vb7 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
111
112 int16x8_t vprod0x0 = vmull_s8(vget_low_s8(vb0), vget_low_s8(va0));
113 int16x8_t vprod1x0 = vmull_s8(vget_low_s8(vb0), vget_low_s8(va1));
114 int16x8_t vprod2x0 = vmull_s8(vget_low_s8(vb0), vget_low_s8(va2));
115 vprod0x0 = vmlal_s8(vprod0x0, vget_high_s8(vb0), vget_high_s8(va0));
116 vprod1x0 = vmlal_s8(vprod1x0, vget_high_s8(vb0), vget_high_s8(va1));
117 vprod2x0 = vmlal_s8(vprod2x0, vget_high_s8(vb0), vget_high_s8(va2));
118 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
119 vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
120 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0);
121 int16x8_t vprod0x1 = vmull_s8(vget_low_s8(vb1), vget_low_s8(va0));
122 int16x8_t vprod1x1 = vmull_s8(vget_low_s8(vb1), vget_low_s8(va1));
123 int16x8_t vprod2x1 = vmull_s8(vget_low_s8(vb1), vget_low_s8(va2));
124 vprod0x1 = vmlal_s8(vprod0x1, vget_high_s8(vb1), vget_high_s8(va0));
125 vprod1x1 = vmlal_s8(vprod1x1, vget_high_s8(vb1), vget_high_s8(va1));
126 vprod2x1 = vmlal_s8(vprod2x1, vget_high_s8(vb1), vget_high_s8(va2));
127 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
128 vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
129 vacc2x1 = vpadalq_s16(vacc2x1, vprod2x1);
130 int16x8_t vprod0x2 = vmull_s8(vget_low_s8(vb2), vget_low_s8(va0));
131 int16x8_t vprod1x2 = vmull_s8(vget_low_s8(vb2), vget_low_s8(va1));
132 int16x8_t vprod2x2 = vmull_s8(vget_low_s8(vb2), vget_low_s8(va2));
133 vprod0x2 = vmlal_s8(vprod0x2, vget_high_s8(vb2), vget_high_s8(va0));
134 vprod1x2 = vmlal_s8(vprod1x2, vget_high_s8(vb2), vget_high_s8(va1));
135 vprod2x2 = vmlal_s8(vprod2x2, vget_high_s8(vb2), vget_high_s8(va2));
136 vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
137 vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
138 vacc2x2 = vpadalq_s16(vacc2x2, vprod2x2);
139 int16x8_t vprod0x3 = vmull_s8(vget_low_s8(vb3), vget_low_s8(va0));
140 int16x8_t vprod1x3 = vmull_s8(vget_low_s8(vb3), vget_low_s8(va1));
141 int16x8_t vprod2x3 = vmull_s8(vget_low_s8(vb3), vget_low_s8(va2));
142 vprod0x3 = vmlal_s8(vprod0x3, vget_high_s8(vb3), vget_high_s8(va0));
143 vprod1x3 = vmlal_s8(vprod1x3, vget_high_s8(vb3), vget_high_s8(va1));
144 vprod2x3 = vmlal_s8(vprod2x3, vget_high_s8(vb3), vget_high_s8(va2));
145 vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
146 vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
147 vacc2x3 = vpadalq_s16(vacc2x3, vprod2x3);
148 int16x8_t vprod0x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va0));
149 int16x8_t vprod1x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va1));
150 int16x8_t vprod2x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va2));
151 vprod0x4 = vmlal_s8(vprod0x4, vget_high_s8(vb4), vget_high_s8(va0));
152 vprod1x4 = vmlal_s8(vprod1x4, vget_high_s8(vb4), vget_high_s8(va1));
153 vprod2x4 = vmlal_s8(vprod2x4, vget_high_s8(vb4), vget_high_s8(va2));
154 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
155 vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
156 vacc2x4 = vpadalq_s16(vacc2x4, vprod2x4);
157 int16x8_t vprod0x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va0));
158 int16x8_t vprod1x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va1));
159 int16x8_t vprod2x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va2));
160 vprod0x5 = vmlal_s8(vprod0x5, vget_high_s8(vb5), vget_high_s8(va0));
161 vprod1x5 = vmlal_s8(vprod1x5, vget_high_s8(vb5), vget_high_s8(va1));
162 vprod2x5 = vmlal_s8(vprod2x5, vget_high_s8(vb5), vget_high_s8(va2));
163 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
164 vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
165 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5);
166 int16x8_t vprod0x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va0));
167 int16x8_t vprod1x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va1));
168 int16x8_t vprod2x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va2));
169 vprod0x6 = vmlal_s8(vprod0x6, vget_high_s8(vb6), vget_high_s8(va0));
170 vprod1x6 = vmlal_s8(vprod1x6, vget_high_s8(vb6), vget_high_s8(va1));
171 vprod2x6 = vmlal_s8(vprod2x6, vget_high_s8(vb6), vget_high_s8(va2));
172 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
173 vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
174 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6);
175 int16x8_t vprod0x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va0));
176 int16x8_t vprod1x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va1));
177 int16x8_t vprod2x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va2));
178 vprod0x7 = vmlal_s8(vprod0x7, vget_high_s8(vb7), vget_high_s8(va0));
179 vprod1x7 = vmlal_s8(vprod1x7, vget_high_s8(vb7), vget_high_s8(va1));
180 vprod2x7 = vmlal_s8(vprod2x7, vget_high_s8(vb7), vget_high_s8(va2));
181 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
182 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
183 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7);
184
185 k -= 16 * sizeof(int8_t);
186 }
187
188 p -= 3 * sizeof(void*);
189 } while (p != 0);
190
191 #if XNN_ARCH_ARM64
192 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1);
193 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3);
194 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5);
195 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7);
196 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1);
197 const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3);
198 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5);
199 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7);
200 const int32x4_t vsum2x01 = vpaddq_s32(vacc2x0, vacc2x1);
201 const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3);
202 const int32x4_t vsum2x45 = vpaddq_s32(vacc2x4, vacc2x5);
203 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7);
204 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23);
205 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67);
206 int32x4_t vacc1x0123 = vpaddq_s32(vsum1x01, vsum1x23);
207 int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67);
208 int32x4_t vacc2x0123 = vpaddq_s32(vsum2x01, vsum2x23);
209 int32x4_t vacc2x4567 = vpaddq_s32(vsum2x45, vsum2x67);
210 #else
211 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0));
212 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1));
213 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2));
214 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3));
215 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1);
216 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3);
217 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 );
218 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4));
219 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5));
220 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6));
221 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7));
222 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5);
223 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7);
224 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 );
225 const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0));
226 const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1));
227 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2));
228 const int32x2_t vpsum1x3 = vadd_s32(vget_low_s32(vacc1x3), vget_high_s32(vacc1x3));
229 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1);
230 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3);
231 int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23 );
232 const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4));
233 const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(vacc1x5));
234 const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6));
235 const int32x2_t vpsum1x7 = vadd_s32(vget_low_s32(vacc1x7), vget_high_s32(vacc1x7));
236 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5);
237 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7);
238 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 );
239 const int32x2_t vpsum2x0 = vadd_s32(vget_low_s32(vacc2x0), vget_high_s32(vacc2x0));
240 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1));
241 const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2));
242 const int32x2_t vpsum2x3 = vadd_s32(vget_low_s32(vacc2x3), vget_high_s32(vacc2x3));
243 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1);
244 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3);
245 int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23 );
246 const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4));
247 const int32x2_t vpsum2x5 = vadd_s32(vget_low_s32(vacc2x5), vget_high_s32(vacc2x5));
248 const int32x2_t vpsum2x6 = vadd_s32(vget_low_s32(vacc2x6), vget_high_s32(vacc2x6));
249 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7));
250 const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5);
251 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7);
252 int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67 );
253 #endif
254
255 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
256 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
257 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
258
259 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
260 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
261 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
262 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
263 vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
264 vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
265
266 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
267 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
268 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
269 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
270 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
271 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
272
273 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
274 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
275 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
276 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
277 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
278 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
279
280 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
281 #if XNN_ARCH_ARM64
282 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
283 const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
284 const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
285 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
286 int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
287 #else
288 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
289 const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
290 const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
291
292 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
293 int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
294 #endif
295 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
296 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
297
298 vout2x01234567 = vmax_s8(vout2x01234567, vget_low_s8(voutput_min));
299 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
300
301 vout2x01234567 = vmin_s8(vout2x01234567, vget_low_s8(voutput_max));
302 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
303
304 if (nc >= 8) {
305 vst1_s8(c2 + 0, vout2x01234567);
306 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
307 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
308
309 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
310 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
311 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
312
313 a = (const int8_t**restrict) ((uintptr_t) a - ks);
314
315 nc -= 8;
316 } else {
317 if (nc & 4) {
318 vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
319 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
320 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
321 vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
322 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
323 }
324 if (nc & 2) {
325 vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
326 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
327 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
328 vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
329 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
330 }
331 if (nc & 1) {
332 vst1_lane_s8(c2, vout2x01234567, 0);
333 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
334 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
335 }
336
337 nc = 0;
338 }
339 } while (nc != 0);
340 }
341