Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-6x8-minmax-neon-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/prefetch.h"
18
19
20 void xnn_qd8_f32_qc8w_igemm_minmax_ukernel_6x8__neon_mlal_lane_prfm(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 size_t ks,
25 const int8_t** restrict a,
26 const void* restrict w,
27 float* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 size_t a_offset,
31 const int8_t* zero,
32 const int8_t* zero_data,
33 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
34 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
35 {
36 assert(mr != 0);
37 assert(mr <= 6);
38 assert(nc != 0);
39 assert(kc != 0);
40 assert(ks != 0);
41 assert(ks % (6 * sizeof(void*)) == 0);
42 assert(a_offset % sizeof(int8_t) == 0);
43 assert(a != NULL);
44 assert(w != NULL);
45 assert(c != NULL);
46
47 float* c0 = c;
48 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
49 if XNN_UNPREDICTABLE(mr < 2) {
50 c1 = c0;
51 }
52 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
53 if XNN_UNPREDICTABLE(mr <= 2) {
54 c2 = c1;
55 }
56 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
57 if XNN_UNPREDICTABLE(mr < 4) {
58 c3 = c2;
59 }
60 float* c4 = (float*) ((uintptr_t) c3 + cm_stride);
61 if XNN_UNPREDICTABLE(mr <= 4) {
62 c4 = c3;
63 }
64 float* c5 = (float*) ((uintptr_t) c4 + cm_stride);
65 if XNN_UNPREDICTABLE(mr != 6) {
66 c5 = c4;
67 }
68
69 do {
70 int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
71 int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
72 const int32x4_t vinput_zero_point = vld1q_dup_s32(&quantization_params->zero_point);
73 int32x4_t vacc0x0123 = vmulq_s32(vksum0123, vinput_zero_point);
74 int32x4_t vacc0x4567 = vmulq_s32(vksum4567, vinput_zero_point);
75 int32x4_t vacc1x0123 = vacc0x0123;
76 int32x4_t vacc1x4567 = vacc0x4567;
77 int32x4_t vacc2x0123 = vacc0x0123;
78 int32x4_t vacc2x4567 = vacc0x4567;
79 int32x4_t vacc3x0123 = vacc0x0123;
80 int32x4_t vacc3x4567 = vacc0x4567;
81 int32x4_t vacc4x0123 = vacc0x0123;
82 int32x4_t vacc4x4567 = vacc0x4567;
83 int32x4_t vacc5x0123 = vacc0x0123;
84 int32x4_t vacc5x4567 = vacc0x4567;
85
86 size_t p = ks;
87 do {
88 const int8_t* restrict a0 = a[0];
89 if XNN_UNPREDICTABLE(a0 != zero) {
90 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
91 } else {
92 a0 = zero_data;
93 }
94 const int8_t* restrict a1 = a[1];
95 if XNN_UNPREDICTABLE(a1 != zero) {
96 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
97 } else {
98 a1 = zero_data;
99 }
100 const int8_t* restrict a2 = a[2];
101 if XNN_UNPREDICTABLE(a2 != zero) {
102 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
103 } else {
104 a2 = zero_data;
105 }
106 const int8_t* restrict a3 = a[3];
107 if XNN_UNPREDICTABLE(a3 != zero) {
108 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
109 } else {
110 a3 = zero_data;
111 }
112 const int8_t* restrict a4 = a[4];
113 if XNN_UNPREDICTABLE(a4 != zero) {
114 a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
115 } else {
116 a4 = zero_data;
117 }
118 const int8_t* restrict a5 = a[5];
119 if XNN_UNPREDICTABLE(a5 != zero) {
120 a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
121 } else {
122 a5 = zero_data;
123 }
124 a += 6;
125
126 size_t k = kc;
127 while (k >= 8 * sizeof(int8_t)) {
128 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
129 const int16x8_t vxa0 = vmovl_s8(va0);
130 const int8x8_t va1 = vld1_s8(a1); a1 += 8;
131 const int16x8_t vxa1 = vmovl_s8(va1);
132 const int8x8_t va2 = vld1_s8(a2); a2 += 8;
133 const int16x8_t vxa2 = vmovl_s8(va2);
134 const int8x8_t va3 = vld1_s8(a3); a3 += 8;
135 const int16x8_t vxa3 = vmovl_s8(va3);
136 const int8x8_t va4 = vld1_s8(a4); a4 += 8;
137 const int16x8_t vxa4 = vmovl_s8(va4);
138 const int8x8_t va5 = vld1_s8(a5); a5 += 8;
139 const int16x8_t vxa5 = vmovl_s8(va5);
140
141 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
142 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
143
144 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
145 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
146 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
147 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
148 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
149 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
150 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
151 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
152 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa4), 0);
153 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa4), 0);
154 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa5), 0);
155 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa5), 0);
156 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
157 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
158
159 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
160 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
161 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
162 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
163 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
164 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
165 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
166 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
167 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa4), 1);
168 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa4), 1);
169 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa5), 1);
170 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa5), 1);
171 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
172 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
173
174 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
175 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
176 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
177 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
178 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
179 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
180 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
181 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
182 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa4), 2);
183 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa4), 2);
184 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa5), 2);
185 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa5), 2);
186 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
187 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
188
189 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
190 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
191 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
192 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
193 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
194 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
195 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
196 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
197 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa4), 3);
198 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa4), 3);
199 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa5), 3);
200 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa5), 3);
201
202 xnn_prefetch_to_l1((const int8_t*) w + 448);
203
204 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
205 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
206
207 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
208 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
209 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
210 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
211 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
212 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
213 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
214 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
215 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa4), 0);
216 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa4), 0);
217 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa5), 0);
218 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa5), 0);
219 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
220 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
221
222 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
223 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
224 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
225 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
226 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
227 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
228 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
229 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
230 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa4), 1);
231 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa4), 1);
232 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa5), 1);
233 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa5), 1);
234 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
235 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
236
237 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
238 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
239 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
240 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
241 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
242 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
243 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
244 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
245 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa4), 2);
246 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa4), 2);
247 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa5), 2);
248 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa5), 2);
249 const int8x8_t vb01234567c7 = vld1_s8(w); w = (const int8_t*) w + 8;
250 const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7);
251
252 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
253 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
254 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
255 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
256 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
257 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
258 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa3), 3);
259 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa3), 3);
260 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa4), 3);
261 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa4), 3);
262 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa5), 3);
263 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa5), 3);
264
265 k -= 8 * sizeof(int8_t);
266 }
267 if XNN_UNLIKELY(k != 0) {
268 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
269 const int16x8_t vxa0 = vmovl_s8(va0);
270 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
271 const int16x8_t vxa1 = vmovl_s8(va1);
272 const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
273 const int16x8_t vxa2 = vmovl_s8(va2);
274 const int8x8_t va3 = vld1_s8(a3); a3 = (const int8_t*) ((uintptr_t) a3 + k);
275 const int16x8_t vxa3 = vmovl_s8(va3);
276 const int8x8_t va4 = vld1_s8(a4); a4 = (const int8_t*) ((uintptr_t) a4 + k);
277 const int16x8_t vxa4 = vmovl_s8(va4);
278 const int8x8_t va5 = vld1_s8(a5); a5 = (const int8_t*) ((uintptr_t) a5 + k);
279 const int16x8_t vxa5 = vmovl_s8(va5);
280
281 const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8;
282 const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
283
284 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
285 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
286 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
287 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
288 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
289 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
290 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
291 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
292 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa4), 0);
293 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa4), 0);
294 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa5), 0);
295 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa5), 0);
296
297 if (k >= 2 * sizeof(int8_t)) {
298 const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8;
299 const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
300
301 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
302 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
303 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
304 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
305 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
306 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
307 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
308 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
309 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa4), 1);
310 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa4), 1);
311 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa5), 1);
312 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa5), 1);
313
314 if (k > 2 * sizeof(int8_t)) {
315 const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8;
316 const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
317
318 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
319 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
320 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
321 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
322 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
323 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
324 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
325 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
326 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa4), 2);
327 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa4), 2);
328 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa5), 2);
329 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa5), 2);
330
331 if (k >= 4 * sizeof(int8_t)) {
332 const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8;
333 const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
334
335 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
336 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
337 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
338 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
339 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
340 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
341 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
342 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
343 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa4), 3);
344 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa4), 3);
345 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa5), 3);
346 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa5), 3);
347
348 if (k > 4 * sizeof(int8_t)) {
349 const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8;
350 const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
351
352 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
353 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
354 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
355 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
356 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
357 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
358 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
359 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
360 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa4), 0);
361 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa4), 0);
362 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa5), 0);
363 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa5), 0);
364
365 if (k >= 6 * sizeof(int8_t)) {
366 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8;
367 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
368
369 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
370 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
371 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
372 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
373 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
374 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
375 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
376 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
377 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa4), 1);
378 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa4), 1);
379 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa5), 1);
380 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa5), 1);
381
382 if (k > 6 * sizeof(int8_t)) {
383 const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8;
384 const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
385
386 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
387 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
388 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
389 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
390 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
391 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
392 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
393 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
394 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa4), 2);
395 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa4), 2);
396 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa5), 2);
397 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa5), 2);
398 }
399 }
400 }
401 }
402 }
403 }
404 }
405 p -= 6 * sizeof(void*);
406 } while (p != 0);
407
408 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
409 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
410 float32x4_t vout1x0123 = vcvtq_f32_s32(vacc1x0123);
411 float32x4_t vout1x4567 = vcvtq_f32_s32(vacc1x4567);
412 float32x4_t vout2x0123 = vcvtq_f32_s32(vacc2x0123);
413 float32x4_t vout2x4567 = vcvtq_f32_s32(vacc2x4567);
414 float32x4_t vout3x0123 = vcvtq_f32_s32(vacc3x0123);
415 float32x4_t vout3x4567 = vcvtq_f32_s32(vacc3x4567);
416 float32x4_t vout4x0123 = vcvtq_f32_s32(vacc4x0123);
417 float32x4_t vout4x4567 = vcvtq_f32_s32(vacc4x4567);
418 float32x4_t vout5x0123 = vcvtq_f32_s32(vacc5x0123);
419 float32x4_t vout5x4567 = vcvtq_f32_s32(vacc5x4567);
420
421 const float32x4_t vinput_scale = vld1q_dup_f32(&quantization_params->inv_scale);
422 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale);
423 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale);
424 vout1x0123 = vmulq_f32(vout1x0123, vinput_scale);
425 vout1x4567 = vmulq_f32(vout1x4567, vinput_scale);
426 vout2x0123 = vmulq_f32(vout2x0123, vinput_scale);
427 vout2x4567 = vmulq_f32(vout2x4567, vinput_scale);
428 vout3x0123 = vmulq_f32(vout3x0123, vinput_scale);
429 vout3x4567 = vmulq_f32(vout3x4567, vinput_scale);
430 vout4x0123 = vmulq_f32(vout4x0123, vinput_scale);
431 vout4x4567 = vmulq_f32(vout4x4567, vinput_scale);
432 vout5x0123 = vmulq_f32(vout5x0123, vinput_scale);
433 vout5x4567 = vmulq_f32(vout5x4567, vinput_scale);
434
435 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
436 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
437
438 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
439 #if XNN_ARCH_ARM64
440 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
441 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
442 vout2x0123 = vfmaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
443 vout3x0123 = vfmaq_f32(vbias0123, vout3x0123, vfilter_output_scale0123);
444 vout4x0123 = vfmaq_f32(vbias0123, vout4x0123, vfilter_output_scale0123);
445 vout5x0123 = vfmaq_f32(vbias0123, vout5x0123, vfilter_output_scale0123);
446 #else
447 vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
448 vout1x0123 = vmlaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
449 vout2x0123 = vmlaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
450 vout3x0123 = vmlaq_f32(vbias0123, vout3x0123, vfilter_output_scale0123);
451 vout4x0123 = vmlaq_f32(vbias0123, vout4x0123, vfilter_output_scale0123);
452 vout5x0123 = vmlaq_f32(vbias0123, vout5x0123, vfilter_output_scale0123);
453 #endif
454 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
455 #if XNN_ARCH_ARM64
456 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
457 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
458 vout2x4567 = vfmaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
459 vout3x4567 = vfmaq_f32(vbias4567, vout3x4567, vfilter_output_scale4567);
460 vout4x4567 = vfmaq_f32(vbias4567, vout4x4567, vfilter_output_scale4567);
461 vout5x4567 = vfmaq_f32(vbias4567, vout5x4567, vfilter_output_scale4567);
462 #else
463 vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
464 vout1x4567 = vmlaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
465 vout2x4567 = vmlaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
466 vout3x4567 = vmlaq_f32(vbias4567, vout3x4567, vfilter_output_scale4567);
467 vout4x4567 = vmlaq_f32(vbias4567, vout4x4567, vfilter_output_scale4567);
468 vout5x4567 = vmlaq_f32(vbias4567, vout5x4567, vfilter_output_scale4567);
469 #endif
470
471 const float32x4_t voutput_min = vld1q_dup_f32(&params->scalar.min);
472 vout0x0123 = vmaxq_f32(vout0x0123, voutput_min);
473 vout0x4567 = vmaxq_f32(vout0x4567, voutput_min);
474 vout1x0123 = vmaxq_f32(vout1x0123, voutput_min);
475 vout1x4567 = vmaxq_f32(vout1x4567, voutput_min);
476 vout2x0123 = vmaxq_f32(vout2x0123, voutput_min);
477 vout2x4567 = vmaxq_f32(vout2x4567, voutput_min);
478 vout3x0123 = vmaxq_f32(vout3x0123, voutput_min);
479 vout3x4567 = vmaxq_f32(vout3x4567, voutput_min);
480 vout4x0123 = vmaxq_f32(vout4x0123, voutput_min);
481 vout4x4567 = vmaxq_f32(vout4x4567, voutput_min);
482 vout5x0123 = vmaxq_f32(vout5x0123, voutput_min);
483 vout5x4567 = vmaxq_f32(vout5x4567, voutput_min);
484
485 const float32x4_t voutput_max = vld1q_dup_f32(&params->scalar.max);
486 vout0x0123 = vminq_f32(vout0x0123, voutput_max);
487 vout0x4567 = vminq_f32(vout0x4567, voutput_max);
488 vout1x0123 = vminq_f32(vout1x0123, voutput_max);
489 vout1x4567 = vminq_f32(vout1x4567, voutput_max);
490 vout2x0123 = vminq_f32(vout2x0123, voutput_max);
491 vout2x4567 = vminq_f32(vout2x4567, voutput_max);
492 vout3x0123 = vminq_f32(vout3x0123, voutput_max);
493 vout3x4567 = vminq_f32(vout3x4567, voutput_max);
494 vout4x0123 = vminq_f32(vout4x0123, voutput_max);
495 vout4x4567 = vminq_f32(vout4x4567, voutput_max);
496 vout5x0123 = vminq_f32(vout5x0123, voutput_max);
497 vout5x4567 = vminq_f32(vout5x4567, voutput_max);
498
499 if XNN_LIKELY(nc >= 8) {
500 vst1q_f32(&c5[0], vout5x0123);
501 vst1q_f32(&c5[4], vout5x4567);
502 vst1q_f32(&c4[0], vout4x0123);
503 vst1q_f32(&c4[4], vout4x4567);
504 vst1q_f32(&c3[0], vout3x0123);
505 vst1q_f32(&c3[4], vout3x4567);
506 vst1q_f32(&c2[0], vout2x0123);
507 vst1q_f32(&c2[4], vout2x4567);
508 vst1q_f32(&c1[0], vout1x0123);
509 vst1q_f32(&c1[4], vout1x4567);
510 vst1q_f32(&c0[0], vout0x0123);
511 vst1q_f32(&c0[4], vout0x4567);
512
513 a = (const int8_t**restrict) ((uintptr_t) a - ks);
514
515 c0 = (float*) ((uintptr_t) c0 + cn_stride);
516 c1 = (float*) ((uintptr_t) c1 + cn_stride);
517 c2 = (float*) ((uintptr_t) c2 + cn_stride);
518 c3 = (float*) ((uintptr_t) c3 + cn_stride);
519 c4 = (float*) ((uintptr_t) c4 + cn_stride);
520 c5 = (float*) ((uintptr_t) c5 + cn_stride);
521
522 nc -= 8;
523 } else {
524 if (nc & 4) {
525 vst1q_f32(c5, vout5x0123); c5 += 4;
526 vout5x0123 = vout5x4567;
527 vst1q_f32(c4, vout4x0123); c4 += 4;
528 vout4x0123 = vout4x4567;
529 vst1q_f32(c3, vout3x0123); c3 += 4;
530 vout3x0123 = vout3x4567;
531 vst1q_f32(c2, vout2x0123); c2 += 4;
532 vout2x0123 = vout2x4567;
533 vst1q_f32(c1, vout1x0123); c1 += 4;
534 vout1x0123 = vout1x4567;
535 vst1q_f32(c0, vout0x0123); c0 += 4;
536 vout0x0123 = vout0x4567;
537 }
538 float32x2_t vout5x01 = vget_low_f32(vout5x0123);
539 float32x2_t vout4x01 = vget_low_f32(vout4x0123);
540 float32x2_t vout3x01 = vget_low_f32(vout3x0123);
541 float32x2_t vout2x01 = vget_low_f32(vout2x0123);
542 float32x2_t vout1x01 = vget_low_f32(vout1x0123);
543 float32x2_t vout0x01 = vget_low_f32(vout0x0123);
544 if (nc & 2) {
545 vst1_f32(c5, vout5x01); c5 += 2;
546 vst1_f32(c4, vout4x01); c4 += 2;
547 vst1_f32(c3, vout3x01); c3 += 2;
548 vst1_f32(c2, vout2x01); c2 += 2;
549 vst1_f32(c1, vout1x01); c1 += 2;
550 vst1_f32(c0, vout0x01); c0 += 2;
551 vout5x01 = vget_high_f32(vout5x0123);
552 vout4x01 = vget_high_f32(vout4x0123);
553 vout3x01 = vget_high_f32(vout3x0123);
554 vout2x01 = vget_high_f32(vout2x0123);
555 vout1x01 = vget_high_f32(vout1x0123);
556 vout0x01 = vget_high_f32(vout0x0123);
557 }
558 if (nc & 1) {
559 vst1_lane_f32(c5, vout5x01, 0);
560 vst1_lane_f32(c4, vout4x01, 0);
561 vst1_lane_f32(c3, vout3x01, 0);
562 vst1_lane_f32(c2, vout2x01, 0);
563 vst1_lane_f32(c1, vout1x01, 0);
564 vst1_lane_f32(c0, vout0x01, 0);
565 }
566 nc = 0;
567 }
568 } while (nc != 0);
569 }