Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-6x8-minmax-neon-mlal-lane.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
18
19 void xnn_qd8_f32_qc8w_igemm_minmax_ukernel_6x8__neon_mlal_lane(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 size_t ks,
24 const int8_t** restrict a,
25 const void* restrict w,
26 float* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 size_t a_offset,
30 const int8_t* zero,
31 const int8_t* zero_data,
32 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
33 const struct xnn_qd8_quantization_params quantization_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 float* c0 = c;
47 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 c1 = c0;
50 }
51 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 c2 = c1;
54 }
55 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr < 4) {
57 c3 = c2;
58 }
59 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
60 if XNN_UNPREDICTABLE(mr <= 4) {
61 c4 = c3;
62 }
63 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
64 if XNN_UNPREDICTABLE(mr != 6) {
65 c5 = c4;
66 }
67
68 do {
69 int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
70 int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
71 const int32x4_t vinput_zero_point = vld1q_dup_s32(&quantization_params->zero_point);
72 int32x4_t vacc0x0123 = vmulq_s32(vksum0123, vinput_zero_point);
73 int32x4_t vacc0x4567 = vmulq_s32(vksum4567, vinput_zero_point);
74 int32x4_t vacc1x0123 = vacc0x0123;
75 int32x4_t vacc1x4567 = vacc0x4567;
76 int32x4_t vacc2x0123 = vacc0x0123;
77 int32x4_t vacc2x4567 = vacc0x4567;
78 int32x4_t vacc3x0123 = vacc0x0123;
79 int32x4_t vacc3x4567 = vacc0x4567;
80 int32x4_t vacc4x0123 = vacc0x0123;
81 int32x4_t vacc4x4567 = vacc0x4567;
82 int32x4_t vacc5x0123 = vacc0x0123;
83 int32x4_t vacc5x4567 = vacc0x4567;
84
85 size_t p = ks;
86 do {
87 const int8_t* restrict a0 = a[0];
88 if XNN_UNPREDICTABLE(a0 != zero) {
89 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
90 } else {
91 a0 = zero_data;
92 }
93 const int8_t* restrict a1 = a[1];
94 if XNN_UNPREDICTABLE(a1 != zero) {
95 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
96 } else {
97 a1 = zero_data;
98 }
99 const int8_t* restrict a2 = a[2];
100 if XNN_UNPREDICTABLE(a2 != zero) {
101 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
102 } else {
103 a2 = zero_data;
104 }
105 const int8_t* restrict a3 = a[3];
106 if XNN_UNPREDICTABLE(a3 != zero) {
107 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
108 } else {
109 a3 = zero_data;
110 }
111 const int8_t* restrict a4 = a[4];
112 if XNN_UNPREDICTABLE(a4 != zero) {
113 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
114 } else {
115 a4 = zero_data;
116 }
117 const int8_t* restrict a5 = a[5];
118 if XNN_UNPREDICTABLE(a5 != zero) {
119 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
120 } else {
121 a5 = zero_data;
122 }
123 a += 6;
124
125 size_t k = kc;
126 while (k >= 8 * sizeof(int8_t)) {
127 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
128 const int16x8_t vxa0 = vmovl_s8(va0);
129 const int8x8_t va1 = vld1_s8(a1); a1 += 8;
130 const int16x8_t vxa1 = vmovl_s8(va1);
131 const int8x8_t va2 = vld1_s8(a2); a2 += 8;
132 const int16x8_t vxa2 = vmovl_s8(va2);
133 const int8x8_t va3 = vld1_s8(a3); a3 += 8;
134 const int16x8_t vxa3 = vmovl_s8(va3);
135 const int8x8_t va4 = vld1_s8(a4); a4 += 8;
136 const int16x8_t vxa4 = vmovl_s8(va4);
137 const int8x8_t va5 = vld1_s8(a5); a5 += 8;
138 const int16x8_t vxa5 = vmovl_s8(va5);
139
140 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
141 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
142
143 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
144 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
145 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
146 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
147 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
148 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
149 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
150 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
151 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa4), 0);
152 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa4), 0);
153 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa5), 0);
154 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa5), 0);
155 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
156 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
157
158 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
159 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
160 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
161 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
162 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
163 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
164 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
165 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
166 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa4), 1);
167 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa4), 1);
168 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa5), 1);
169 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa5), 1);
170 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
171 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
172
173 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
174 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
175 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
176 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
177 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
178 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
179 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
180 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
181 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa4), 2);
182 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa4), 2);
183 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa5), 2);
184 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa5), 2);
185 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
186 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
187
188 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
189 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
190 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
191 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
192 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
193 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
194 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
195 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
196 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa4), 3);
197 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa4), 3);
198 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa5), 3);
199 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa5), 3);
200
201
202 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
203 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
204
205 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
206 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
207 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
208 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
209 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
210 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
211 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
212 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
213 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa4), 0);
214 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa4), 0);
215 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa5), 0);
216 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa5), 0);
217 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
218 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
219
220 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
221 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
222 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
223 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
224 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
225 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
226 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
227 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
228 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa4), 1);
229 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa4), 1);
230 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa5), 1);
231 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa5), 1);
232 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
233 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
234
235 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
236 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
237 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
238 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
239 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
240 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
241 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
242 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
243 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa4), 2);
244 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa4), 2);
245 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa5), 2);
246 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa5), 2);
247 const int8x8_t vb01234567c7 = vld1_s8(w); w = (const int8_t*) w + 8;
248 const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7);
249
250 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
251 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
252 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
253 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
254 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
255 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
256 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa3), 3);
257 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa3), 3);
258 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa4), 3);
259 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa4), 3);
260 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa5), 3);
261 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa5), 3);
262
263 k -= 8 * sizeof(int8_t);
264 }
265 if XNN_UNLIKELY(k != 0) {
266 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
267 const int16x8_t vxa0 = vmovl_s8(va0);
268 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
269 const int16x8_t vxa1 = vmovl_s8(va1);
270 const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
271 const int16x8_t vxa2 = vmovl_s8(va2);
272 const int8x8_t va3 = vld1_s8(a3); a3 = (const int8_t*) ((uintptr_t) a3 + k);
273 const int16x8_t vxa3 = vmovl_s8(va3);
274 const int8x8_t va4 = vld1_s8(a4); a4 = (const int8_t*) ((uintptr_t) a4 + k);
275 const int16x8_t vxa4 = vmovl_s8(va4);
276 const int8x8_t va5 = vld1_s8(a5); a5 = (const int8_t*) ((uintptr_t) a5 + k);
277 const int16x8_t vxa5 = vmovl_s8(va5);
278
279 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
280 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
281
282 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
283 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
284 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
285 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
286 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
287 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
288 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
289 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
290 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa4), 0);
291 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa4), 0);
292 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa5), 0);
293 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa5), 0);
294
295 if (k >= 2 * sizeof(int8_t)) {
296 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
297 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
298
299 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
300 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
301 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
302 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
303 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
304 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
305 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
306 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
307 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa4), 1);
308 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa4), 1);
309 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa5), 1);
310 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa5), 1);
311
312 if (k > 2 * sizeof(int8_t)) {
313 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
314 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
315
316 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
317 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
318 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
319 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
320 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
321 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
322 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
323 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
324 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa4), 2);
325 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa4), 2);
326 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa5), 2);
327 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa5), 2);
328
329 if (k >= 4 * sizeof(int8_t)) {
330 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
331 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
332
333 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
334 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
335 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
336 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
337 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
338 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
339 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
340 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
341 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa4), 3);
342 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa4), 3);
343 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa5), 3);
344 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa5), 3);
345
346 if (k > 4 * sizeof(int8_t)) {
347 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
348 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
349
350 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
351 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
352 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
353 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
354 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
355 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
356 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
357 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
358 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa4), 0);
359 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa4), 0);
360 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa5), 0);
361 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa5), 0);
362
363 if (k >= 6 * sizeof(int8_t)) {
364 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
365 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
366
367 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
368 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
369 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
370 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
371 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
372 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
373 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
374 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
375 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa4), 1);
376 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa4), 1);
377 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa5), 1);
378 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa5), 1);
379
380 if (k > 6 * sizeof(int8_t)) {
381 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
382 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
383
384 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
385 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
386 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
387 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
388 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
389 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
390 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
391 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
392 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa4), 2);
393 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa4), 2);
394 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa5), 2);
395 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa5), 2);
396 }
397 }
398 }
399 }
400 }
401 }
402 }
403 p -= 6 * sizeof(void*);
404 } while (p != 0);
405
406 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
407 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
408 float32x4_t vout1x0123 = vcvtq_f32_s32(vacc1x0123);
409 float32x4_t vout1x4567 = vcvtq_f32_s32(vacc1x4567);
410 float32x4_t vout2x0123 = vcvtq_f32_s32(vacc2x0123);
411 float32x4_t vout2x4567 = vcvtq_f32_s32(vacc2x4567);
412 float32x4_t vout3x0123 = vcvtq_f32_s32(vacc3x0123);
413 float32x4_t vout3x4567 = vcvtq_f32_s32(vacc3x4567);
414 float32x4_t vout4x0123 = vcvtq_f32_s32(vacc4x0123);
415 float32x4_t vout4x4567 = vcvtq_f32_s32(vacc4x4567);
416 float32x4_t vout5x0123 = vcvtq_f32_s32(vacc5x0123);
417 float32x4_t vout5x4567 = vcvtq_f32_s32(vacc5x4567);
418
419 const float32x4_t vinput_scale = vld1q_dup_f32(&quantization_params->inv_scale);
420 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale);
421 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale);
422 vout1x0123 = vmulq_f32(vout1x0123, vinput_scale);
423 vout1x4567 = vmulq_f32(vout1x4567, vinput_scale);
424 vout2x0123 = vmulq_f32(vout2x0123, vinput_scale);
425 vout2x4567 = vmulq_f32(vout2x4567, vinput_scale);
426 vout3x0123 = vmulq_f32(vout3x0123, vinput_scale);
427 vout3x4567 = vmulq_f32(vout3x4567, vinput_scale);
428 vout4x0123 = vmulq_f32(vout4x0123, vinput_scale);
429 vout4x4567 = vmulq_f32(vout4x4567, vinput_scale);
430 vout5x0123 = vmulq_f32(vout5x0123, vinput_scale);
431 vout5x4567 = vmulq_f32(vout5x4567, vinput_scale);
432
433 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
434 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
435
436 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
437 #if XNN_ARCH_ARM64
438 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
439 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
440 vout2x0123 = vfmaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
441 vout3x0123 = vfmaq_f32(vbias0123, vout3x0123, vfilter_output_scale0123);
442 vout4x0123 = vfmaq_f32(vbias0123, vout4x0123, vfilter_output_scale0123);
443 vout5x0123 = vfmaq_f32(vbias0123, vout5x0123, vfilter_output_scale0123);
444 #else
445 vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
446 vout1x0123 = vmlaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
447 vout2x0123 = vmlaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
448 vout3x0123 = vmlaq_f32(vbias0123, vout3x0123, vfilter_output_scale0123);
449 vout4x0123 = vmlaq_f32(vbias0123, vout4x0123, vfilter_output_scale0123);
450 vout5x0123 = vmlaq_f32(vbias0123, vout5x0123, vfilter_output_scale0123);
451 #endif
452 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
453 #if XNN_ARCH_ARM64
454 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
455 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
456 vout2x4567 = vfmaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
457 vout3x4567 = vfmaq_f32(vbias4567, vout3x4567, vfilter_output_scale4567);
458 vout4x4567 = vfmaq_f32(vbias4567, vout4x4567, vfilter_output_scale4567);
459 vout5x4567 = vfmaq_f32(vbias4567, vout5x4567, vfilter_output_scale4567);
460 #else
461 vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
462 vout1x4567 = vmlaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
463 vout2x4567 = vmlaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
464 vout3x4567 = vmlaq_f32(vbias4567, vout3x4567, vfilter_output_scale4567);
465 vout4x4567 = vmlaq_f32(vbias4567, vout4x4567, vfilter_output_scale4567);
466 vout5x4567 = vmlaq_f32(vbias4567, vout5x4567, vfilter_output_scale4567);
467 #endif
468
469 const float32x4_t voutput_min = vld1q_dup_f32(&params->scalar.min);
470 vout0x0123 = vmaxq_f32(vout0x0123, voutput_min);
471 vout0x4567 = vmaxq_f32(vout0x4567, voutput_min);
472 vout1x0123 = vmaxq_f32(vout1x0123, voutput_min);
473 vout1x4567 = vmaxq_f32(vout1x4567, voutput_min);
474 vout2x0123 = vmaxq_f32(vout2x0123, voutput_min);
475 vout2x4567 = vmaxq_f32(vout2x4567, voutput_min);
476 vout3x0123 = vmaxq_f32(vout3x0123, voutput_min);
477 vout3x4567 = vmaxq_f32(vout3x4567, voutput_min);
478 vout4x0123 = vmaxq_f32(vout4x0123, voutput_min);
479 vout4x4567 = vmaxq_f32(vout4x4567, voutput_min);
480 vout5x0123 = vmaxq_f32(vout5x0123, voutput_min);
481 vout5x4567 = vmaxq_f32(vout5x4567, voutput_min);
482
483 const float32x4_t voutput_max = vld1q_dup_f32(&params->scalar.max);
484 vout0x0123 = vminq_f32(vout0x0123, voutput_max);
485 vout0x4567 = vminq_f32(vout0x4567, voutput_max);
486 vout1x0123 = vminq_f32(vout1x0123, voutput_max);
487 vout1x4567 = vminq_f32(vout1x4567, voutput_max);
488 vout2x0123 = vminq_f32(vout2x0123, voutput_max);
489 vout2x4567 = vminq_f32(vout2x4567, voutput_max);
490 vout3x0123 = vminq_f32(vout3x0123, voutput_max);
491 vout3x4567 = vminq_f32(vout3x4567, voutput_max);
492 vout4x0123 = vminq_f32(vout4x0123, voutput_max);
493 vout4x4567 = vminq_f32(vout4x4567, voutput_max);
494 vout5x0123 = vminq_f32(vout5x0123, voutput_max);
495 vout5x4567 = vminq_f32(vout5x4567, voutput_max);
496
497 if XNN_LIKELY(nc >= 8) {
498 vst1q_f32(&c5[0], vout5x0123);
499 vst1q_f32(&c5[4], vout5x4567);
500 vst1q_f32(&c4[0], vout4x0123);
501 vst1q_f32(&c4[4], vout4x4567);
502 vst1q_f32(&c3[0], vout3x0123);
503 vst1q_f32(&c3[4], vout3x4567);
504 vst1q_f32(&c2[0], vout2x0123);
505 vst1q_f32(&c2[4], vout2x4567);
506 vst1q_f32(&c1[0], vout1x0123);
507 vst1q_f32(&c1[4], vout1x4567);
508 vst1q_f32(&c0[0], vout0x0123);
509 vst1q_f32(&c0[4], vout0x4567);
510
511 a = (const int8_t**restrict) ((uintptr_t) a - ks);
512
513 c0 = (float*) ((uintptr_t) c0 + cn_stride);
514 c1 = (float*) ((uintptr_t) c1 + cn_stride);
515 c2 = (float*) ((uintptr_t) c2 + cn_stride);
516 c3 = (float*) ((uintptr_t) c3 + cn_stride);
517 c4 = (float*) ((uintptr_t) c4 + cn_stride);
518 c5 = (float*) ((uintptr_t) c5 + cn_stride);
519
520 nc -= 8;
521 } else {
522 if (nc & 4) {
523 vst1q_f32(c5, vout5x0123); c5 += 4;
524 vout5x0123 = vout5x4567;
525 vst1q_f32(c4, vout4x0123); c4 += 4;
526 vout4x0123 = vout4x4567;
527 vst1q_f32(c3, vout3x0123); c3 += 4;
528 vout3x0123 = vout3x4567;
529 vst1q_f32(c2, vout2x0123); c2 += 4;
530 vout2x0123 = vout2x4567;
531 vst1q_f32(c1, vout1x0123); c1 += 4;
532 vout1x0123 = vout1x4567;
533 vst1q_f32(c0, vout0x0123); c0 += 4;
534 vout0x0123 = vout0x4567;
535 }
536 float32x2_t vout5x01 = vget_low_f32(vout5x0123);
537 float32x2_t vout4x01 = vget_low_f32(vout4x0123);
538 float32x2_t vout3x01 = vget_low_f32(vout3x0123);
539 float32x2_t vout2x01 = vget_low_f32(vout2x0123);
540 float32x2_t vout1x01 = vget_low_f32(vout1x0123);
541 float32x2_t vout0x01 = vget_low_f32(vout0x0123);
542 if (nc & 2) {
543 vst1_f32(c5, vout5x01); c5 += 2;
544 vst1_f32(c4, vout4x01); c4 += 2;
545 vst1_f32(c3, vout3x01); c3 += 2;
546 vst1_f32(c2, vout2x01); c2 += 2;
547 vst1_f32(c1, vout1x01); c1 += 2;
548 vst1_f32(c0, vout0x01); c0 += 2;
549 vout5x01 = vget_high_f32(vout5x0123);
550 vout4x01 = vget_high_f32(vout4x0123);
551 vout3x01 = vget_high_f32(vout3x0123);
552 vout2x01 = vget_high_f32(vout2x0123);
553 vout1x01 = vget_high_f32(vout1x0123);
554 vout0x01 = vget_high_f32(vout0x0123);
555 }
556 if (nc & 1) {
557 vst1_lane_f32(c5, vout5x01, 0);
558 vst1_lane_f32(c4, vout4x01, 0);
559 vst1_lane_f32(c3, vout3x01, 0);
560 vst1_lane_f32(c2, vout2x01, 0);
561 vst1_lane_f32(c1, vout1x01, 0);
562 vst1_lane_f32(c0, vout0x01, 0);
563 }
564 nc = 0;
565 }
566 } while (nc != 0);
567 }