Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-6x8-minmax-fp32-neonv8-mlal-lane-prfm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/neon-mlal-lane.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
11 #include <assert.h>
12
13 #include <arm_neon.h>
14
15 #include "xnnpack/common.h"
16 #include "xnnpack/igemm.h"
17 #include "xnnpack/intrinsics-polyfill.h"
18 #include "xnnpack/prefetch.h"
19
20
21 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane_prfm(
22 size_t mr,
23 size_t nc,
24 size_t kc,
25 size_t ks,
26 const int8_t** restrict a,
27 const void* restrict w,
28 int8_t* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 size_t a_offset,
32 const int8_t* zero,
33 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
34 {
35 assert(mr != 0);
36 assert(mr <= 6);
37 assert(nc != 0);
38 assert(kc != 0);
39 assert(ks != 0);
40 assert(ks % (6 * sizeof(void*)) == 0);
41 assert(a_offset % sizeof(int8_t) == 0);
42 assert(a != NULL);
43 assert(w != NULL);
44 assert(c != NULL);
45
46 int8_t* c0 = c;
47 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 c1 = c0;
50 }
51 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 c2 = c1;
54 }
55 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr < 4) {
57 c3 = c2;
58 }
59 int8_t* c4 = (int8_t*) ((uintptr_t) c3 + cm_stride);
60 if XNN_UNPREDICTABLE(mr <= 4) {
61 c4 = c3;
62 }
63 int8_t* c5 = (int8_t*) ((uintptr_t) c4 + cm_stride);
64 if XNN_UNPREDICTABLE(mr != 6) {
65 c5 = c4;
66 }
67
68 do {
69 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
70 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
71 int32x4_t vacc1x0123 = vacc0x0123;
72 int32x4_t vacc1x4567 = vacc0x4567;
73 int32x4_t vacc2x0123 = vacc0x0123;
74 int32x4_t vacc2x4567 = vacc0x4567;
75 int32x4_t vacc3x0123 = vacc0x0123;
76 int32x4_t vacc3x4567 = vacc0x4567;
77 int32x4_t vacc4x0123 = vacc0x0123;
78 int32x4_t vacc4x4567 = vacc0x4567;
79 int32x4_t vacc5x0123 = vacc0x0123;
80 int32x4_t vacc5x4567 = vacc0x4567;
81
82 size_t p = ks;
83 do {
84 const int8_t* restrict a0 = a[0];
85 if XNN_UNPREDICTABLE(a0 != zero) {
86 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
87 }
88 const int8_t* restrict a1 = a[1];
89 if XNN_UNPREDICTABLE(a1 != zero) {
90 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
91 }
92 const int8_t* restrict a2 = a[2];
93 if XNN_UNPREDICTABLE(a2 != zero) {
94 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
95 }
96 const int8_t* restrict a3 = a[3];
97 if XNN_UNPREDICTABLE(a3 != zero) {
98 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
99 }
100 const int8_t* restrict a4 = a[4];
101 if XNN_UNPREDICTABLE(a4 != zero) {
102 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
103 }
104 const int8_t* restrict a5 = a[5];
105 if XNN_UNPREDICTABLE(a5 != zero) {
106 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
107 }
108 a += 6;
109
110 size_t k = kc;
111 while (k >= 8 * sizeof(int8_t)) {
112 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
113 const int16x8_t vxa0 = vmovl_s8(va0);
114 const int8x8_t va1 = vld1_s8(a1); a1 += 8;
115 const int16x8_t vxa1 = vmovl_s8(va1);
116 const int8x8_t va2 = vld1_s8(a2); a2 += 8;
117 const int16x8_t vxa2 = vmovl_s8(va2);
118 const int8x8_t va3 = vld1_s8(a3); a3 += 8;
119 const int16x8_t vxa3 = vmovl_s8(va3);
120 const int8x8_t va4 = vld1_s8(a4); a4 += 8;
121 const int16x8_t vxa4 = vmovl_s8(va4);
122 const int8x8_t va5 = vld1_s8(a5); a5 += 8;
123 const int16x8_t vxa5 = vmovl_s8(va5);
124
125 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
126 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
127
128 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
129 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
130 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
131 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
132 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
133 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
134 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
135 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
136 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa4), 0);
137 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa4), 0);
138 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa5), 0);
139 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa5), 0);
140 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
141 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
142
143 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
144 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
145 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
146 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
147 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
148 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
149 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
150 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
151 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa4), 1);
152 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa4), 1);
153 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa5), 1);
154 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa5), 1);
155 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
156 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
157
158 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
159 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
160 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
161 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
162 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
163 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
164 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
165 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
166 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa4), 2);
167 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa4), 2);
168 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa5), 2);
169 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa5), 2);
170 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
171 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
172
173 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
174 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
175 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
176 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
177 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
178 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
179 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
180 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
181 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa4), 3);
182 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa4), 3);
183 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa5), 3);
184 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa5), 3);
185
186 xnn_prefetch_to_l1((const int8_t*) w + 448);
187
188 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
189 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
190
191 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
192 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
193 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
194 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
195 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
196 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
197 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
198 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
199 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa4), 0);
200 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa4), 0);
201 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa5), 0);
202 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa5), 0);
203 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
204 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
205
206 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
207 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
208 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
209 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
210 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
211 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
212 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
213 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
214 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa4), 1);
215 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa4), 1);
216 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa5), 1);
217 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa5), 1);
218 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
219 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
220
221 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
222 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
223 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
224 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
225 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
226 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
227 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
228 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
229 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa4), 2);
230 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa4), 2);
231 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa5), 2);
232 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa5), 2);
233 const int8x8_t vb01234567c7 = vld1_s8(w); w = (const int8_t*) w + 8;
234 const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7);
235
236 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
237 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
238 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
239 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
240 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
241 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
242 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa3), 3);
243 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa3), 3);
244 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa4), 3);
245 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa4), 3);
246 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa5), 3);
247 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa5), 3);
248
249 k -= 8 * sizeof(int8_t);
250 }
251 if XNN_UNLIKELY(k != 0) {
252 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
253 const int16x8_t vxa0 = vmovl_s8(va0);
254 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
255 const int16x8_t vxa1 = vmovl_s8(va1);
256 const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
257 const int16x8_t vxa2 = vmovl_s8(va2);
258 const int8x8_t va3 = vld1_s8(a3); a3 = (const int8_t*) ((uintptr_t) a3 + k);
259 const int16x8_t vxa3 = vmovl_s8(va3);
260 const int8x8_t va4 = vld1_s8(a4); a4 = (const int8_t*) ((uintptr_t) a4 + k);
261 const int16x8_t vxa4 = vmovl_s8(va4);
262 const int8x8_t va5 = vld1_s8(a5); a5 = (const int8_t*) ((uintptr_t) a5 + k);
263 const int16x8_t vxa5 = vmovl_s8(va5);
264
265 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
266 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
267
268 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
269 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
270 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
271 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
272 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
273 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
274 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
275 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
276 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa4), 0);
277 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa4), 0);
278 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa5), 0);
279 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa5), 0);
280
281 if (k >= 2 * sizeof(int8_t)) {
282 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
283 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
284
285 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
286 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
287 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
288 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
289 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
290 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
291 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
292 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
293 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa4), 1);
294 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa4), 1);
295 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa5), 1);
296 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa5), 1);
297
298 if (k > 2 * sizeof(int8_t)) {
299 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
300 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
301
302 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
303 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
304 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
305 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
306 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
307 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
308 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
309 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
310 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa4), 2);
311 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa4), 2);
312 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa5), 2);
313 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa5), 2);
314
315 if (k >= 4 * sizeof(int8_t)) {
316 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
317 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
318
319 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
320 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
321 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
322 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
323 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
324 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
325 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
326 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
327 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa4), 3);
328 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa4), 3);
329 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa5), 3);
330 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa5), 3);
331
332 if (k > 4 * sizeof(int8_t)) {
333 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
334 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
335
336 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
337 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
338 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
339 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
340 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
341 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
342 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
343 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
344 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa4), 0);
345 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa4), 0);
346 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa5), 0);
347 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa5), 0);
348
349 if (k >= 6 * sizeof(int8_t)) {
350 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
351 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
352
353 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
354 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
355 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
356 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
357 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
358 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
359 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
360 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
361 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa4), 1);
362 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa4), 1);
363 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa5), 1);
364 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa5), 1);
365
366 if (k > 6 * sizeof(int8_t)) {
367 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
368 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
369
370 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
371 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
372 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
373 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
374 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
375 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
376 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
377 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
378 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa4), 2);
379 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa4), 2);
380 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa5), 2);
381 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa5), 2);
382 }
383 }
384 }
385 }
386 }
387 }
388 }
389 p -= 6 * sizeof(void*);
390 } while (p != 0);
391
392 // Post-accumulation work
393 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
394 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
395 float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
396 float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
397 float32x4_t vfpacc2x0123 = vcvtq_f32_s32(vacc2x0123);
398 float32x4_t vfpacc2x4567 = vcvtq_f32_s32(vacc2x4567);
399 float32x4_t vfpacc3x0123 = vcvtq_f32_s32(vacc3x0123);
400 float32x4_t vfpacc3x4567 = vcvtq_f32_s32(vacc3x4567);
401 float32x4_t vfpacc4x0123 = vcvtq_f32_s32(vacc4x0123);
402 float32x4_t vfpacc4x4567 = vcvtq_f32_s32(vacc4x4567);
403 float32x4_t vfpacc5x0123 = vcvtq_f32_s32(vacc5x0123);
404 float32x4_t vfpacc5x4567 = vcvtq_f32_s32(vacc5x4567);
405
406 const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const float*) w + 4;
407 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
408 vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale0123);
409 vfpacc2x0123 = vmulq_f32(vfpacc2x0123, vscale0123);
410 vfpacc3x0123 = vmulq_f32(vfpacc3x0123, vscale0123);
411 vfpacc4x0123 = vmulq_f32(vfpacc4x0123, vscale0123);
412 vfpacc5x0123 = vmulq_f32(vfpacc5x0123, vscale0123);
413 const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const float*) w + 4;
414 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
415 vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale4567);
416 vfpacc2x4567 = vmulq_f32(vfpacc2x4567, vscale4567);
417 vfpacc3x4567 = vmulq_f32(vfpacc3x4567, vscale4567);
418 vfpacc4x4567 = vmulq_f32(vfpacc4x4567, vscale4567);
419 vfpacc5x4567 = vmulq_f32(vfpacc5x4567, vscale4567);
420
421 vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
422 vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
423 vacc1x0123 = vcvtnq_s32_f32(vfpacc1x0123);
424 vacc1x4567 = vcvtnq_s32_f32(vfpacc1x4567);
425 vacc2x0123 = vcvtnq_s32_f32(vfpacc2x0123);
426 vacc2x4567 = vcvtnq_s32_f32(vfpacc2x4567);
427 vacc3x0123 = vcvtnq_s32_f32(vfpacc3x0123);
428 vacc3x4567 = vcvtnq_s32_f32(vfpacc3x4567);
429 vacc4x0123 = vcvtnq_s32_f32(vfpacc4x0123);
430 vacc4x4567 = vcvtnq_s32_f32(vfpacc4x4567);
431 vacc5x0123 = vcvtnq_s32_f32(vfpacc5x0123);
432 vacc5x4567 = vcvtnq_s32_f32(vfpacc5x4567);
433
434 const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->fp32_neonv8.output_zero_point);
435 #if XNN_ARCH_ARM64
436 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
437 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
438 int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
439 int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
440 int16x8_t vacc4x01234567 = vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567);
441 int16x8_t vacc5x01234567 = vqmovn_high_s32(vqmovn_s32(vacc5x0123), vacc5x4567);
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 vacc4x01234567 = vqaddq_s16(vacc4x01234567, voutput_zero_point);
448 vacc5x01234567 = vqaddq_s16(vacc5x01234567, voutput_zero_point);
449
450 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
451 int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567);
452 int8x16_t vout4x01234567_5x01234567 = vqmovn_high_s16(vqmovn_s16(vacc4x01234567), vacc5x01234567);
453 #else
454 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
455 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
456 int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
457 int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
458 int16x8_t vacc4x01234567 = vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x4567));
459 int16x8_t vacc5x01234567 = vcombine_s16(vqmovn_s32(vacc5x0123), vqmovn_s32(vacc5x4567));
460
461 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
462 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
463 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
464 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
465 vacc4x01234567 = vqaddq_s16(vacc4x01234567, voutput_zero_point);
466 vacc5x01234567 = vqaddq_s16(vacc5x01234567, voutput_zero_point);
467
468 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
469 int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567));
470 int8x16_t vout4x01234567_5x01234567 = vcombine_s8(vqmovn_s16(vacc4x01234567), vqmovn_s16(vacc5x01234567));
471 #endif
472
473 const int8x16_t voutput_min = vld1q_dup_s8(&params->fp32_neonv8.output_min);
474 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
475 vout2x01234567_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, voutput_min);
476 vout4x01234567_5x01234567 = vmaxq_s8(vout4x01234567_5x01234567, voutput_min);
477
478 const int8x16_t voutput_max = vld1q_dup_s8(&params->fp32_neonv8.output_max);
479 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
480 vout2x01234567_3x01234567 = vminq_s8(vout2x01234567_3x01234567, voutput_max);
481 vout4x01234567_5x01234567 = vminq_s8(vout4x01234567_5x01234567, voutput_max);
482
483 if (nc >= 8) {
484 vst1_s8(c5 + 0, vget_high_s8(vout4x01234567_5x01234567));
485 vst1_s8(c4 + 0, vget_low_s8(vout4x01234567_5x01234567));
486 vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567));
487 vst1_s8(c2 + 0, vget_low_s8(vout2x01234567_3x01234567));
488 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
489 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
490
491 c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);
492 c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
493 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
494 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
495 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
496 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
497
498 a = (const int8_t**restrict) ((uintptr_t) a - ks);
499
500 nc -= 8;
501 } else {
502 if (nc & 4) {
503 vst1q_lane_u32((void*) c5, vreinterpretq_u32_s8(vout4x01234567_5x01234567), 2); c5 += 4;
504 vst1q_lane_u32((void*) c4, vreinterpretq_u32_s8(vout4x01234567_5x01234567), 0); c4 += 4;
505 vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
506 vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
507 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
508 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
509 vout4x01234567_5x01234567 = vextq_s8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 4);
510 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
511 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
512 }
513 if (nc & 2) {
514 vst1q_lane_u16((void*) c5, vreinterpretq_u16_s8(vout4x01234567_5x01234567), 4); c5 += 2;
515 vst1q_lane_u16((void*) c4, vreinterpretq_u16_s8(vout4x01234567_5x01234567), 0); c4 += 2;
516 vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
517 vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
518 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
519 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
520 vout4x01234567_5x01234567 = vextq_s8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 2);
521 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
522 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
523 }
524 if (nc & 1) {
525 vst1q_lane_s8(c5, vout4x01234567_5x01234567, 8);
526 vst1q_lane_s8(c4, vout4x01234567_5x01234567, 0);
527 vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
528 vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
529 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
530 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
531 }
532
533 nc = 0;
534 }
535 } while (nc != 0);
536 }