1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c4-neondot.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/igemm.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16 #include <xnnpack/math.h>
17
18
xnn_qc8_igemm_minmax_fp32_ukernel_4x16c4__neondot(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qc8_igemm_minmax_fp32_ukernel_4x16c4__neondot(
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 int8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 size_t a_offset,
30 const int8_t* zero,
31 const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(mr != 0);
34 assert(mr <= 4);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(ks != 0);
38 assert(ks % (4 * sizeof(void*)) == 0);
39 assert(a_offset % sizeof(int8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 kc = round_up_po2(kc, 4 * sizeof(int8_t));
45 int8_t* c0 = c;
46 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr < 2) {
48 c1 = c0;
49 }
50 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 c2 = c1;
53 }
54 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
55 if XNN_UNPREDICTABLE(mr != 4) {
56 c3 = c2;
57 }
58
59 do {
60 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
61 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
62 int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
63 int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
64 int32x4_t vacc1x0123 = vacc0x0123;
65 int32x4_t vacc1x4567 = vacc0x4567;
66 int32x4_t vacc1x89AB = vacc0x89AB;
67 int32x4_t vacc1xCDEF = vacc0xCDEF;
68 int32x4_t vacc2x0123 = vacc0x0123;
69 int32x4_t vacc2x4567 = vacc0x4567;
70 int32x4_t vacc2x89AB = vacc0x89AB;
71 int32x4_t vacc2xCDEF = vacc0xCDEF;
72 int32x4_t vacc3x0123 = vacc0x0123;
73 int32x4_t vacc3x4567 = vacc0x4567;
74 int32x4_t vacc3x89AB = vacc0x89AB;
75 int32x4_t vacc3xCDEF = vacc0xCDEF;
76
77 size_t p = ks;
78 do {
79 const int8_t* restrict a0 = a[0];
80 if XNN_UNPREDICTABLE(a0 != zero) {
81 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
82 }
83 const int8_t* restrict a1 = a[1];
84 if XNN_UNPREDICTABLE(a1 != zero) {
85 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
86 }
87 const int8_t* restrict a2 = a[2];
88 if XNN_UNPREDICTABLE(a2 != zero) {
89 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
90 }
91 const int8_t* restrict a3 = a[3];
92 if XNN_UNPREDICTABLE(a3 != zero) {
93 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
94 }
95 a += 4;
96
97 // Inner accumulation loop along the 16 columns.
98 size_t k = kc;
99 // 2x partial unrolled loop to load 8 bytes at a time.
100 while (k >= 8 * sizeof(int8_t)) {
101 // Load a 4x8 block of activations.
102 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
103 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 8;
104 const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 8;
105 const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 8;
106
107 // Load a 8x16 block of weights.
108 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
109 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
110 const int8x16_t vb0123x89AB = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
111 const int8x16_t vb0123xCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
112 const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
113 const int8x16_t vb4567x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
114 const int8x16_t vb4567x89AB = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
115 const int8x16_t vb4567xCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
116
117 // Multiply-accumulate: 4x8 * 8x16 --> 4x16.
118 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
119 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
120 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
121 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
122 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
123 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
124 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb0123x89AB, va1x01234567, 0);
125 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
126 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
127 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
128 vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb0123x89AB, va2x01234567, 0);
129 vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
130 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
131 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
132 vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb0123x89AB, va3x01234567, 0);
133 vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
134 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
135 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
136 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1);
137 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
138 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1);
139 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1);
140 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb4567x89AB, va1x01234567, 1);
141 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb4567xCDEF, va1x01234567, 1);
142 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1);
143 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb4567x4567, va2x01234567, 1);
144 vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb4567x89AB, va2x01234567, 1);
145 vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb4567xCDEF, va2x01234567, 1);
146 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1);
147 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb4567x4567, va3x01234567, 1);
148 vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb4567x89AB, va3x01234567, 1);
149 vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb4567xCDEF, va3x01234567, 1);
150
151 k -= 8 * sizeof(int8_t);
152 }
153 // Handle up to 4 final positions of `k`
154 if XNN_UNLIKELY(k != 0) {
155 // Load a 4x4 block of activations.
156 const int8x8_t va0x01234567 = vld1_s8(a0);
157 const int8x8_t va1x01234567 = vld1_s8(a1);
158 const int8x8_t va2x01234567 = vld1_s8(a2);
159 const int8x8_t va3x01234567 = vld1_s8(a3);
160
161 // Load a 4x16 block of weights.
162 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
163 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
164 const int8x16_t vb0123x89AB = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
165 const int8x16_t vb0123xCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
166
167 // Multiply-accumulate: 4x4 * 4x16 --> 4x16.
168 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
169 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
170 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
171 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
172 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
173 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
174 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb0123x89AB, va1x01234567, 0);
175 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
176 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
177 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
178 vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb0123x89AB, va2x01234567, 0);
179 vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
180 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
181 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
182 vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb0123x89AB, va3x01234567, 0);
183 vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
184 }
185 p -= 4 * sizeof(void*);
186 } while (p != 0);
187
188 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
189 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
190 float32x4_t vfpacc0x89AB = vcvtq_f32_s32(vacc0x89AB);
191 float32x4_t vfpacc0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
192 float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
193 float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
194 float32x4_t vfpacc1x89AB = vcvtq_f32_s32(vacc1x89AB);
195 float32x4_t vfpacc1xCDEF = vcvtq_f32_s32(vacc1xCDEF);
196 float32x4_t vfpacc2x0123 = vcvtq_f32_s32(vacc2x0123);
197 float32x4_t vfpacc2x4567 = vcvtq_f32_s32(vacc2x4567);
198 float32x4_t vfpacc2x89AB = vcvtq_f32_s32(vacc2x89AB);
199 float32x4_t vfpacc2xCDEF = vcvtq_f32_s32(vacc2xCDEF);
200 float32x4_t vfpacc3x0123 = vcvtq_f32_s32(vacc3x0123);
201 float32x4_t vfpacc3x4567 = vcvtq_f32_s32(vacc3x4567);
202 float32x4_t vfpacc3x89AB = vcvtq_f32_s32(vacc3x89AB);
203 float32x4_t vfpacc3xCDEF = vcvtq_f32_s32(vacc3xCDEF);
204
205 const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
206 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
207 vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale0123);
208 vfpacc2x0123 = vmulq_f32(vfpacc2x0123, vscale0123);
209 vfpacc3x0123 = vmulq_f32(vfpacc3x0123, vscale0123);
210 const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
211 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
212 vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale4567);
213 vfpacc2x4567 = vmulq_f32(vfpacc2x4567, vscale4567);
214 vfpacc3x4567 = vmulq_f32(vfpacc3x4567, vscale4567);
215 const float32x4_t vscale89AB = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
216 vfpacc0x89AB = vmulq_f32(vfpacc0x89AB, vscale89AB);
217 vfpacc1x89AB = vmulq_f32(vfpacc1x89AB, vscale89AB);
218 vfpacc2x89AB = vmulq_f32(vfpacc2x89AB, vscale89AB);
219 vfpacc3x89AB = vmulq_f32(vfpacc3x89AB, vscale89AB);
220 const float32x4_t vscaleCDEF = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
221 vfpacc0xCDEF = vmulq_f32(vfpacc0xCDEF, vscaleCDEF);
222 vfpacc1xCDEF = vmulq_f32(vfpacc1xCDEF, vscaleCDEF);
223 vfpacc2xCDEF = vmulq_f32(vfpacc2xCDEF, vscaleCDEF);
224 vfpacc3xCDEF = vmulq_f32(vfpacc3xCDEF, vscaleCDEF);
225
226 vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
227 vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
228 vacc0x89AB = vcvtnq_s32_f32(vfpacc0x89AB);
229 vacc0xCDEF = vcvtnq_s32_f32(vfpacc0xCDEF);
230 vacc1x0123 = vcvtnq_s32_f32(vfpacc1x0123);
231 vacc1x4567 = vcvtnq_s32_f32(vfpacc1x4567);
232 vacc1x89AB = vcvtnq_s32_f32(vfpacc1x89AB);
233 vacc1xCDEF = vcvtnq_s32_f32(vfpacc1xCDEF);
234 vacc2x0123 = vcvtnq_s32_f32(vfpacc2x0123);
235 vacc2x4567 = vcvtnq_s32_f32(vfpacc2x4567);
236 vacc2x89AB = vcvtnq_s32_f32(vfpacc2x89AB);
237 vacc2xCDEF = vcvtnq_s32_f32(vfpacc2xCDEF);
238 vacc3x0123 = vcvtnq_s32_f32(vfpacc3x0123);
239 vacc3x4567 = vcvtnq_s32_f32(vfpacc3x4567);
240 vacc3x89AB = vcvtnq_s32_f32(vfpacc3x89AB);
241 vacc3xCDEF = vcvtnq_s32_f32(vfpacc3xCDEF);
242
243 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->fp32_neonv8.output_zero_point);
244 #if XNN_ARCH_ARM64
245 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
246 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
247 const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
248 const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point);
249 const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
250 const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF), voutput_zero_point);
251 const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point);
252 const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF), voutput_zero_point);
253
254 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
255 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
256 int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
257 int8x16_t vout3x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc3x01234567), vacc3x89ABCDEF);
258 #else
259 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
260 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
261 const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
262 const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point);
263 const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
264 const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF)), voutput_zero_point);
265 const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point);
266 const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF)), voutput_zero_point);
267
268 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
269 int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
270 int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
271 int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCDEF));
272 #endif
273 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->fp32_neonv8.output_min);
274 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->fp32_neonv8.output_max);
275
276 vout3x0123456789ABCDEF = vmaxq_s8(vout3x0123456789ABCDEF, voutput_min);
277 vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
278 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
279 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
280
281 vout3x0123456789ABCDEF = vminq_s8(vout3x0123456789ABCDEF, voutput_max);
282 vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
283 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
284 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
285
286 if (nc >= 16) {
287 vst1q_s8(c3 + 0, vout3x0123456789ABCDEF);
288 vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
289 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
290 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
291
292 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
293 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
294 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
295 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
296
297 a = (const int8_t**restrict) ((uintptr_t) a - ks);
298
299 nc -= 16;
300 } else {
301 int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8(vout3x0123456789ABCDEF));
302 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
303 if (nc & 8) {
304 vst1_s8(c3, vget_high_s8(vout2x01234567_3x01234567)); c3 += 8;
305 vst1_s8(c2, vget_low_s8(vout2x01234567_3x01234567)); c2 += 8;
306 vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
307 vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
308 vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0123456789ABCDEF));
309 vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
310 }
311 if (nc & 4) {
312 vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
313 vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
314 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
315 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
316 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
317 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
318 }
319 if (nc & 2) {
320 vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
321 vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
322 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
323 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
324 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
325 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
326 }
327 if (nc & 1) {
328 vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
329 vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
330 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
331 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
332 }
333
334 nc = 0;
335 }
336 } while (nc != 0);
337 }
338