1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/neon-mull-addw-dup.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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/common.h>
15 #include <xnnpack/gemm.h>
16
17
xnn_qs8_gemm_minmax_rndnu_ukernel_3x8__neon_mull_addw_dup(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_3x8__neon_mull_addw_dup(
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 const int8_t* a0 = a;
40 int8_t* c0 = c;
41 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
42 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
43 if XNN_UNPREDICTABLE(mr < 2) {
44 a1 = a0;
45 c1 = c0;
46 }
47 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
48 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
49 if XNN_UNPREDICTABLE(mr <= 2) {
50 a2 = a1;
51 c2 = c1;
52 }
53
54 do {
55 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
56 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
57 int32x4_t vacc1x0123 = vacc0x0123;
58 int32x4_t vacc1x4567 = vacc0x4567;
59 int32x4_t vacc2x0123 = vacc0x0123;
60 int32x4_t vacc2x4567 = vacc0x4567;
61
62 size_t k = kc;
63 while (k >= 8 * sizeof(int8_t)) {
64 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
65 const int8x8_t va1 = vld1_s8(a1); a1 += 8;
66 const int8x8_t va2 = vld1_s8(a2); a2 += 8;
67
68 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
69
70 const int16x8_t vprod0x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va0, 0));
71 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c0));
72 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c0));
73 const int16x8_t vprod1x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va1, 0));
74 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c0));
75 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c0));
76 const int16x8_t vprod2x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va2, 0));
77 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c0));
78 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c0));
79 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
80
81 const int16x8_t vprod0x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va0, 1));
82 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c1));
83 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c1));
84 const int16x8_t vprod1x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va1, 1));
85 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c1));
86 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c1));
87 const int16x8_t vprod2x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va2, 1));
88 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c1));
89 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c1));
90 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
91
92 const int16x8_t vprod0x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va0, 2));
93 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c2));
94 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c2));
95 const int16x8_t vprod1x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va1, 2));
96 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c2));
97 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c2));
98 const int16x8_t vprod2x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va2, 2));
99 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c2));
100 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c2));
101 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
102
103 const int16x8_t vprod0x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va0, 3));
104 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c3));
105 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c3));
106 const int16x8_t vprod1x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va1, 3));
107 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c3));
108 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c3));
109 const int16x8_t vprod2x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va2, 3));
110 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c3));
111 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c3));
112 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
113
114 const int16x8_t vprod0x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va0, 4));
115 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c4));
116 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c4));
117 const int16x8_t vprod1x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va1, 4));
118 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c4));
119 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c4));
120 const int16x8_t vprod2x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va2, 4));
121 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c4));
122 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c4));
123 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
124
125 const int16x8_t vprod0x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va0, 5));
126 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c5));
127 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c5));
128 const int16x8_t vprod1x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va1, 5));
129 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c5));
130 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c5));
131 const int16x8_t vprod2x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va2, 5));
132 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c5));
133 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c5));
134 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
135
136 const int16x8_t vprod0x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va0, 6));
137 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c6));
138 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c6));
139 const int16x8_t vprod1x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va1, 6));
140 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c6));
141 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c6));
142 const int16x8_t vprod2x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va2, 6));
143 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c6));
144 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c6));
145 const int8x8_t vb01234567c7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
146
147 const int16x8_t vprod0x01234567c7 = vmull_s8(vb01234567c7, vdup_lane_s8(va0, 7));
148 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c7));
149 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c7));
150 const int16x8_t vprod1x01234567c7 = vmull_s8(vb01234567c7, vdup_lane_s8(va1, 7));
151 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c7));
152 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c7));
153 const int16x8_t vprod2x01234567c7 = vmull_s8(vb01234567c7, vdup_lane_s8(va2, 7));
154 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c7));
155 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c7));
156
157 k -= 8 * sizeof(int8_t);
158 }
159 if XNN_UNLIKELY(k != 0) {
160 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
161 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
162 const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
163
164 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
165
166 const int16x8_t vprod0x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va0, 0));
167 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c0));
168 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c0));
169 const int16x8_t vprod1x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va1, 0));
170 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c0));
171 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c0));
172 const int16x8_t vprod2x01234567c0 = vmull_s8(vb01234567c0, vdup_lane_s8(va2, 0));
173 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c0));
174 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c0));
175
176 if (k >= 2 * sizeof(int8_t)) {
177 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
178
179 const int16x8_t vprod0x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va0, 1));
180 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c1));
181 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c1));
182 const int16x8_t vprod1x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va1, 1));
183 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c1));
184 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c1));
185 const int16x8_t vprod2x01234567c1 = vmull_s8(vb01234567c1, vdup_lane_s8(va2, 1));
186 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c1));
187 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c1));
188
189 if (k > 2 * sizeof(int8_t)) {
190 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
191
192 const int16x8_t vprod0x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va0, 2));
193 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c2));
194 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c2));
195 const int16x8_t vprod1x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va1, 2));
196 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c2));
197 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c2));
198 const int16x8_t vprod2x01234567c2 = vmull_s8(vb01234567c2, vdup_lane_s8(va2, 2));
199 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c2));
200 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c2));
201
202 if (k >= 4 * sizeof(int8_t)) {
203 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
204
205 const int16x8_t vprod0x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va0, 3));
206 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c3));
207 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c3));
208 const int16x8_t vprod1x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va1, 3));
209 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c3));
210 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c3));
211 const int16x8_t vprod2x01234567c3 = vmull_s8(vb01234567c3, vdup_lane_s8(va2, 3));
212 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c3));
213 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c3));
214
215 if (k > 4 * sizeof(int8_t)) {
216 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
217
218 const int16x8_t vprod0x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va0, 4));
219 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c4));
220 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c4));
221 const int16x8_t vprod1x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va1, 4));
222 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c4));
223 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c4));
224 const int16x8_t vprod2x01234567c4 = vmull_s8(vb01234567c4, vdup_lane_s8(va2, 4));
225 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c4));
226 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c4));
227
228 if (k >= 6 * sizeof(int8_t)) {
229 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
230
231 const int16x8_t vprod0x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va0, 5));
232 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c5));
233 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c5));
234 const int16x8_t vprod1x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va1, 5));
235 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c5));
236 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c5));
237 const int16x8_t vprod2x01234567c5 = vmull_s8(vb01234567c5, vdup_lane_s8(va2, 5));
238 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c5));
239 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c5));
240
241 if (k > 6 * sizeof(int8_t)) {
242 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
243
244 const int16x8_t vprod0x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va0, 6));
245 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c6));
246 vacc0x4567 = vaddw_s16(vacc0x4567, vget_high_s16(vprod0x01234567c6));
247 const int16x8_t vprod1x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va1, 6));
248 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c6));
249 vacc1x4567 = vaddw_s16(vacc1x4567, vget_high_s16(vprod1x01234567c6));
250 const int16x8_t vprod2x01234567c6 = vmull_s8(vb01234567c6, vdup_lane_s8(va2, 6));
251 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c6));
252 vacc2x4567 = vaddw_s16(vacc2x4567, vget_high_s16(vprod2x01234567c6));
253 }
254 }
255 }
256 }
257 }
258 }
259 }
260
261 // Post-accumulation work
262 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
263 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
264 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
265
266 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
267 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
268 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
269 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
270 vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
271 vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
272
273 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
274 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
275 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
276 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
277 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
278 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
279
280 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
281 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
282 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
283 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
284 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
285 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
286
287 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
288 #if XNN_ARCH_ARM64
289 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
290 const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
291 const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
292
293 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
294 int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
295 #else
296 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
297 const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
298 const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
299
300 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
301 int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
302 #endif
303 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
304 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
305
306 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
307 vout2x01234567 = vmax_s8(vout2x01234567, vget_low_s8(voutput_min));
308
309 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
310 vout2x01234567 = vmin_s8(vout2x01234567, vget_low_s8(voutput_max));
311
312 if (nc >= 8) {
313 // Main case where there the 8 columns fit in the destination.
314 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
315 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
316 vst1_s8(c2 + 0, vout2x01234567);
317
318 // Advance to the next 8 columns.
319 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
320 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
321 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
322
323 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
324 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
325 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
326
327 nc -= 8;
328 } else {
329 // Final case where not all of the 8 columns fit in the destination.
330 if (nc & 4) {
331 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
332 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
333 vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
334 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
335 vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
336 }
337 if (nc & 2) {
338 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
339 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
340 vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
341 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
342 vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
343 }
344 if (nc & 1) {
345 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
346 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
347 vst1_lane_s8(c2, vout2x01234567, 0);
348 }
349
350 nc = 0;
351 }
352 } while (nc != 0);
353 }
354