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 #include <assert.h>
11
12 #include <arm_neon.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/igemm.h>
16
17
xnn_qu8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane(size_t mr,size_t nc,size_t kc,size_t ks,const uint8_t ** restrict a,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const uint8_t** restrict a,
24 const void* restrict w,
25 uint8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const uint8_t* zero,
30 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 6);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (6 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(uint8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 uint8_t* c0 = c;
44 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 c1 = c0;
47 }
48 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
49 if XNN_UNPREDICTABLE(mr <= 2) {
50 c2 = c1;
51 }
52 uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
53 if XNN_UNPREDICTABLE(mr < 4) {
54 c3 = c2;
55 }
56 uint8_t* c4 = (uint8_t*) ((uintptr_t) c3 + cm_stride);
57 if XNN_UNPREDICTABLE(mr <= 4) {
58 c4 = c3;
59 }
60 uint8_t* c5 = (uint8_t*) ((uintptr_t) c4 + cm_stride);
61 if XNN_UNPREDICTABLE(mr != 6) {
62 c5 = c4;
63 }
64
65 const uint8x8_t vb_zero_point = vld1_dup_u8(¶ms->rndnu_neon.kernel_zero_point[0]);
66 do {
67 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
68 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
69 int32x4_t vacc1x0123 = vacc0x0123;
70 int32x4_t vacc1x4567 = vacc0x4567;
71 int32x4_t vacc2x0123 = vacc0x0123;
72 int32x4_t vacc2x4567 = vacc0x4567;
73 int32x4_t vacc3x0123 = vacc0x0123;
74 int32x4_t vacc3x4567 = vacc0x4567;
75 int32x4_t vacc4x0123 = vacc0x0123;
76 int32x4_t vacc4x4567 = vacc0x4567;
77 int32x4_t vacc5x0123 = vacc0x0123;
78 int32x4_t vacc5x4567 = vacc0x4567;
79
80 size_t p = ks;
81 do {
82 const uint8_t* restrict a0 = a[0];
83 if XNN_UNPREDICTABLE(a0 != zero) {
84 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
85 }
86 const uint8_t* restrict a1 = a[1];
87 if XNN_UNPREDICTABLE(a1 != zero) {
88 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
89 }
90 const uint8_t* restrict a2 = a[2];
91 if XNN_UNPREDICTABLE(a2 != zero) {
92 a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
93 }
94 const uint8_t* restrict a3 = a[3];
95 if XNN_UNPREDICTABLE(a3 != zero) {
96 a3 = (const uint8_t*) ((uintptr_t) a3 + a_offset);
97 }
98 const uint8_t* restrict a4 = a[4];
99 if XNN_UNPREDICTABLE(a4 != zero) {
100 a4 = (const uint8_t*) ((uintptr_t) a4 + a_offset);
101 }
102 const uint8_t* restrict a5 = a[5];
103 if XNN_UNPREDICTABLE(a5 != zero) {
104 a5 = (const uint8_t*) ((uintptr_t) a5 + a_offset);
105 }
106 a += 6;
107
108 size_t k = kc;
109 while (k >= 8 * sizeof(uint8_t)) {
110 const uint8x8_t va0 = vld1_u8(a0); a0 += 8;
111 const int16x8_t vxa0 = vreinterpretq_s16_u16(vmovl_u8(va0));
112 const uint8x8_t va1 = vld1_u8(a1); a1 += 8;
113 const int16x8_t vxa1 = vreinterpretq_s16_u16(vmovl_u8(va1));
114 const uint8x8_t va2 = vld1_u8(a2); a2 += 8;
115 const int16x8_t vxa2 = vreinterpretq_s16_u16(vmovl_u8(va2));
116 const uint8x8_t va3 = vld1_u8(a3); a3 += 8;
117 const int16x8_t vxa3 = vreinterpretq_s16_u16(vmovl_u8(va3));
118 const uint8x8_t va4 = vld1_u8(a4); a4 += 8;
119 const int16x8_t vxa4 = vreinterpretq_s16_u16(vmovl_u8(va4));
120 const uint8x8_t va5 = vld1_u8(a5); a5 += 8;
121 const int16x8_t vxa5 = vreinterpretq_s16_u16(vmovl_u8(va5));
122
123 const uint8x8_t vb01234567c0 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
124 const int16x8_t vxb01234567c0 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c0, vb_zero_point));
125
126 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
127 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
128 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
129 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
130 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
131 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
132 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
133 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
134 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa4), 0);
135 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa4), 0);
136 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa5), 0);
137 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa5), 0);
138 const uint8x8_t vb01234567c1 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
139 const int16x8_t vxb01234567c1 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c1, vb_zero_point));
140
141 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
142 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
143 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
144 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
145 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
146 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
147 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
148 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
149 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa4), 1);
150 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa4), 1);
151 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa5), 1);
152 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa5), 1);
153 const uint8x8_t vb01234567c2 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
154 const int16x8_t vxb01234567c2 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c2, vb_zero_point));
155
156 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
157 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
158 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
159 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
160 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
161 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
162 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
163 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
164 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa4), 2);
165 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa4), 2);
166 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa5), 2);
167 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa5), 2);
168 const uint8x8_t vb01234567c3 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
169 const int16x8_t vxb01234567c3 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c3, vb_zero_point));
170
171 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
172 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
173 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
174 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
175 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
176 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
177 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
178 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
179 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa4), 3);
180 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa4), 3);
181 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa5), 3);
182 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa5), 3);
183
184
185 const uint8x8_t vb01234567c4 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
186 const int16x8_t vxb01234567c4 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c4, vb_zero_point));
187
188 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
189 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
190 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
191 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
192 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
193 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
194 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
195 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
196 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa4), 0);
197 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa4), 0);
198 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa5), 0);
199 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa5), 0);
200 const uint8x8_t vb01234567c5 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
201 const int16x8_t vxb01234567c5 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c5, vb_zero_point));
202
203 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
204 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
205 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
206 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
207 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
208 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
209 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
210 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
211 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa4), 1);
212 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa4), 1);
213 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa5), 1);
214 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa5), 1);
215 const uint8x8_t vb01234567c6 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
216 const int16x8_t vxb01234567c6 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c6, vb_zero_point));
217
218 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
219 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
220 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
221 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
222 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
223 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
224 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
225 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
226 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa4), 2);
227 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa4), 2);
228 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa5), 2);
229 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa5), 2);
230 const uint8x8_t vb01234567c7 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
231 const int16x8_t vxb01234567c7 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c7, vb_zero_point));
232
233 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
234 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
235 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
236 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
237 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
238 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
239 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa3), 3);
240 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa3), 3);
241 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa4), 3);
242 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa4), 3);
243 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa5), 3);
244 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa5), 3);
245
246 k -= 8 * sizeof(uint8_t);
247 }
248 if XNN_UNLIKELY(k != 0) {
249 const uint8x8_t va0 = vld1_u8(a0); a0 = (const uint8_t*) ((uintptr_t) a0 + k);
250 const int16x8_t vxa0 = vreinterpretq_s16_u16(vmovl_u8(va0));
251 const uint8x8_t va1 = vld1_u8(a1); a1 = (const uint8_t*) ((uintptr_t) a1 + k);
252 const int16x8_t vxa1 = vreinterpretq_s16_u16(vmovl_u8(va1));
253 const uint8x8_t va2 = vld1_u8(a2); a2 = (const uint8_t*) ((uintptr_t) a2 + k);
254 const int16x8_t vxa2 = vreinterpretq_s16_u16(vmovl_u8(va2));
255 const uint8x8_t va3 = vld1_u8(a3); a3 = (const uint8_t*) ((uintptr_t) a3 + k);
256 const int16x8_t vxa3 = vreinterpretq_s16_u16(vmovl_u8(va3));
257 const uint8x8_t va4 = vld1_u8(a4); a4 = (const uint8_t*) ((uintptr_t) a4 + k);
258 const int16x8_t vxa4 = vreinterpretq_s16_u16(vmovl_u8(va4));
259 const uint8x8_t va5 = vld1_u8(a5); a5 = (const uint8_t*) ((uintptr_t) a5 + k);
260 const int16x8_t vxa5 = vreinterpretq_s16_u16(vmovl_u8(va5));
261
262 const uint8x8_t vb01234567c0 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
263 const int16x8_t vxb01234567c0 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c0, vb_zero_point));
264
265 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
266 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
267 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
268 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
269 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
270 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
271 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
272 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
273 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa4), 0);
274 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa4), 0);
275 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa5), 0);
276 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa5), 0);
277
278 if (k >= 2 * sizeof(uint8_t)) {
279 const uint8x8_t vb01234567c1 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
280 const int16x8_t vxb01234567c1 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c1, vb_zero_point));
281
282 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
283 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
284 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
285 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
286 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
287 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
288 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
289 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
290 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa4), 1);
291 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa4), 1);
292 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa5), 1);
293 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa5), 1);
294
295 if (k > 2 * sizeof(uint8_t)) {
296 const uint8x8_t vb01234567c2 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
297 const int16x8_t vxb01234567c2 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c2, vb_zero_point));
298
299 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
300 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
301 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
302 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
303 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
304 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
305 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
306 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
307 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa4), 2);
308 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa4), 2);
309 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa5), 2);
310 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa5), 2);
311
312 if (k >= 4 * sizeof(uint8_t)) {
313 const uint8x8_t vb01234567c3 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
314 const int16x8_t vxb01234567c3 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c3, vb_zero_point));
315
316 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
317 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
318 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
319 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
320 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
321 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
322 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
323 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
324 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa4), 3);
325 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa4), 3);
326 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa5), 3);
327 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa5), 3);
328
329 if (k > 4 * sizeof(uint8_t)) {
330 const uint8x8_t vb01234567c4 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
331 const int16x8_t vxb01234567c4 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c4, vb_zero_point));
332
333 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
334 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
335 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
336 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
337 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
338 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
339 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
340 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
341 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa4), 0);
342 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa4), 0);
343 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa5), 0);
344 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa5), 0);
345
346 if (k >= 6 * sizeof(uint8_t)) {
347 const uint8x8_t vb01234567c5 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
348 const int16x8_t vxb01234567c5 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c5, vb_zero_point));
349
350 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
351 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
352 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
353 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
354 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
355 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
356 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
357 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
358 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa4), 1);
359 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa4), 1);
360 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa5), 1);
361 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa5), 1);
362
363 if (k > 6 * sizeof(uint8_t)) {
364 const uint8x8_t vb01234567c6 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
365 const int16x8_t vxb01234567c6 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c6, vb_zero_point));
366
367 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
368 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
369 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
370 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
371 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
372 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
373 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
374 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
375 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa4), 2);
376 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa4), 2);
377 vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa5), 2);
378 vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa5), 2);
379 }
380 }
381 }
382 }
383 }
384 }
385 }
386 p -= 6 * sizeof(void*);
387 } while (p != 0);
388
389 // Post-accumulation work
390 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
391 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
392 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
393
394 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
395 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
396 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
397 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
398 vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
399 vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
400 vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
401 vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
402 vacc4x0123 = vqshlq_s32(vacc4x0123, vright_pre_shift);
403 vacc4x4567 = vqshlq_s32(vacc4x4567, vright_pre_shift);
404 vacc5x0123 = vqshlq_s32(vacc5x0123, vright_pre_shift);
405 vacc5x4567 = vqshlq_s32(vacc5x4567, vright_pre_shift);
406
407 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
408 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
409 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
410 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
411 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
412 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
413 vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
414 vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
415 vacc4x0123 = vqdmulhq_s32(vacc4x0123, vmultiplier);
416 vacc4x4567 = vqdmulhq_s32(vacc4x4567, vmultiplier);
417 vacc5x0123 = vqdmulhq_s32(vacc5x0123, vmultiplier);
418 vacc5x4567 = vqdmulhq_s32(vacc5x4567, vmultiplier);
419
420 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
421 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
422 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
423 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
424 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
425 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
426 vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
427 vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
428 vacc4x0123 = vrshlq_s32(vacc4x0123, vright_post_shift);
429 vacc4x4567 = vrshlq_s32(vacc4x4567, vright_post_shift);
430 vacc5x0123 = vrshlq_s32(vacc5x0123, vright_post_shift);
431 vacc5x4567 = vrshlq_s32(vacc5x4567, vright_post_shift);
432
433 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
434 #if XNN_ARCH_ARM64
435 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
436 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
437 int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
438 int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
439 int16x8_t vacc4x01234567 = vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567);
440 int16x8_t vacc5x01234567 = vqmovn_high_s32(vqmovn_s32(vacc5x0123), vacc5x4567);
441
442 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
443 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
444 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
445 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
446 vacc4x01234567 = vqaddq_s16(vacc4x01234567, voutput_zero_point);
447 vacc5x01234567 = vqaddq_s16(vacc5x01234567, voutput_zero_point);
448
449 uint8x16_t vout0x01234567_1x01234567 = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc1x01234567);
450 uint8x16_t vout2x01234567_3x01234567 = vqmovun_high_s16(vqmovun_s16(vacc2x01234567), vacc3x01234567);
451 uint8x16_t vout4x01234567_5x01234567 = vqmovun_high_s16(vqmovun_s16(vacc4x01234567), vacc5x01234567);
452 #else
453 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
454 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
455 int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
456 int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
457 int16x8_t vacc4x01234567 = vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x4567));
458 int16x8_t vacc5x01234567 = vcombine_s16(vqmovn_s32(vacc5x0123), vqmovn_s32(vacc5x4567));
459
460 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
461 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
462 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
463 vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
464 vacc4x01234567 = vqaddq_s16(vacc4x01234567, voutput_zero_point);
465 vacc5x01234567 = vqaddq_s16(vacc5x01234567, voutput_zero_point);
466
467 uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc1x01234567));
468 uint8x16_t vout2x01234567_3x01234567 = vcombine_u8(vqmovun_s16(vacc2x01234567), vqmovun_s16(vacc3x01234567));
469 uint8x16_t vout4x01234567_5x01234567 = vcombine_u8(vqmovun_s16(vacc4x01234567), vqmovun_s16(vacc5x01234567));
470 #endif
471
472 const uint8x16_t voutput_min = vld1q_dup_u8(¶ms->rndnu_neon.output_min);
473 vout0x01234567_1x01234567 = vmaxq_u8(vout0x01234567_1x01234567, voutput_min);
474 vout2x01234567_3x01234567 = vmaxq_u8(vout2x01234567_3x01234567, voutput_min);
475 vout4x01234567_5x01234567 = vmaxq_u8(vout4x01234567_5x01234567, voutput_min);
476
477 const uint8x16_t voutput_max = vld1q_dup_u8(¶ms->rndnu_neon.output_max);
478 vout0x01234567_1x01234567 = vminq_u8(vout0x01234567_1x01234567, voutput_max);
479 vout2x01234567_3x01234567 = vminq_u8(vout2x01234567_3x01234567, voutput_max);
480 vout4x01234567_5x01234567 = vminq_u8(vout4x01234567_5x01234567, voutput_max);
481
482 if (nc >= 8) {
483 vst1_u8(c5 + 0, vget_high_u8(vout4x01234567_5x01234567));
484 vst1_u8(c4 + 0, vget_low_u8(vout4x01234567_5x01234567));
485 vst1_u8(c3 + 0, vget_high_u8(vout2x01234567_3x01234567));
486 vst1_u8(c2 + 0, vget_low_u8(vout2x01234567_3x01234567));
487 vst1_u8(c1 + 0, vget_high_u8(vout0x01234567_1x01234567));
488 vst1_u8(c0 + 0, vget_low_u8(vout0x01234567_1x01234567));
489
490 c5 = (uint8_t*) ((uintptr_t) c5 + cn_stride);
491 c4 = (uint8_t*) ((uintptr_t) c4 + cn_stride);
492 c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
493 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
494 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
495 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
496
497 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
498
499 nc -= 8;
500 } else {
501 if (nc & 4) {
502 vst1q_lane_u32((void*) c5, vreinterpretq_u32_u8(vout4x01234567_5x01234567), 2); c5 += 4;
503 vst1q_lane_u32((void*) c4, vreinterpretq_u32_u8(vout4x01234567_5x01234567), 0); c4 += 4;
504 vst1q_lane_u32((void*) c3, vreinterpretq_u32_u8(vout2x01234567_3x01234567), 2); c3 += 4;
505 vst1q_lane_u32((void*) c2, vreinterpretq_u32_u8(vout2x01234567_3x01234567), 0); c2 += 4;
506 vst1q_lane_u32((void*) c1, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
507 vst1q_lane_u32((void*) c0, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
508 vout4x01234567_5x01234567 = vextq_u8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 4);
509 vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
510 vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
511 }
512 if (nc & 2) {
513 vst1q_lane_u16((void*) c5, vreinterpretq_u16_u8(vout4x01234567_5x01234567), 4); c5 += 2;
514 vst1q_lane_u16((void*) c4, vreinterpretq_u16_u8(vout4x01234567_5x01234567), 0); c4 += 2;
515 vst1q_lane_u16((void*) c3, vreinterpretq_u16_u8(vout2x01234567_3x01234567), 4); c3 += 2;
516 vst1q_lane_u16((void*) c2, vreinterpretq_u16_u8(vout2x01234567_3x01234567), 0); c2 += 2;
517 vst1q_lane_u16((void*) c1, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
518 vst1q_lane_u16((void*) c0, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
519 vout4x01234567_5x01234567 = vextq_u8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 2);
520 vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
521 vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
522 }
523 if (nc & 1) {
524 vst1q_lane_u8(c5, vout4x01234567_5x01234567, 8);
525 vst1q_lane_u8(c4, vout4x01234567_5x01234567, 0);
526 vst1q_lane_u8(c3, vout2x01234567_3x01234567, 8);
527 vst1q_lane_u8(c2, vout2x01234567_3x01234567, 0);
528 vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
529 vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
530 }
531
532 nc = 0;
533 }
534 } while (nc != 0);
535 }
536