1 // Auto-generated file. Do not edit!
2 // Template: src/qu8-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/math.h>
16
17
xnn_qu8_igemm_minmax_rndnu_ukernel_6x8c4__neondot(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_6x8c4__neondot(
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 kc = round_up_po2(kc, 4 * sizeof(uint8_t));
44 uint8_t* c0 = c;
45 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr < 2) {
47 c1 = c0;
48 }
49 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 c2 = c1;
52 }
53 uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
54 if XNN_UNPREDICTABLE(mr < 4) {
55 c3 = c2;
56 }
57 uint8_t* c4 = (uint8_t*) ((uintptr_t) c3 + cm_stride);
58 if XNN_UNPREDICTABLE(mr <= 4) {
59 c4 = c3;
60 }
61 uint8_t* c5 = (uint8_t*) ((uintptr_t) c4 + cm_stride);
62 if XNN_UNPREDICTABLE(mr != 6) {
63 c5 = c4;
64 }
65
66 const uint8x8_t va_zero_point = vld1_dup_u8(¶ms->rndnu_neon.kernel_zero_point[0]);
67
68 do {
69 // Initialize accumulators with bias. 8 bias values are loaded from the
70 // weight matrix, at the start of the group of 8 columns.
71 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
72 uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
73 uint32x4_t vpacc1x0123 = vpacc0x0123;
74 uint32x4_t vpacc1x4567 = vpacc0x4567;
75 uint32x4_t vpacc2x0123 = vpacc0x0123;
76 uint32x4_t vpacc2x4567 = vpacc0x4567;
77 uint32x4_t vpacc3x0123 = vpacc0x0123;
78 uint32x4_t vpacc3x4567 = vpacc0x4567;
79 uint32x4_t vpacc4x0123 = vpacc0x0123;
80 uint32x4_t vpacc4x4567 = vpacc0x4567;
81 uint32x4_t vpacc5x0123 = vpacc0x0123;
82 uint32x4_t vpacc5x4567 = vpacc0x4567;
83 uint32x2_t vnacc0 = vmov_n_u32(0);
84 uint32x2_t vnacc1 = vmov_n_u32(0);
85 uint32x2_t vnacc2 = vmov_n_u32(0);
86 uint32x2_t vnacc3 = vmov_n_u32(0);
87 uint32x2_t vnacc4 = vmov_n_u32(0);
88 uint32x2_t vnacc5 = vmov_n_u32(0);
89
90 size_t p = ks;
91 do {
92 const uint8_t* restrict a0 = a[0];
93 if XNN_UNPREDICTABLE(a0 != zero) {
94 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
95 }
96 const uint8_t* restrict a1 = a[1];
97 if XNN_UNPREDICTABLE(a1 != zero) {
98 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
99 }
100 const uint8_t* restrict a2 = a[2];
101 if XNN_UNPREDICTABLE(a2 != zero) {
102 a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
103 }
104 const uint8_t* restrict a3 = a[3];
105 if XNN_UNPREDICTABLE(a3 != zero) {
106 a3 = (const uint8_t*) ((uintptr_t) a3 + a_offset);
107 }
108 const uint8_t* restrict a4 = a[4];
109 if XNN_UNPREDICTABLE(a4 != zero) {
110 a4 = (const uint8_t*) ((uintptr_t) a4 + a_offset);
111 }
112 const uint8_t* restrict a5 = a[5];
113 if XNN_UNPREDICTABLE(a5 != zero) {
114 a5 = (const uint8_t*) ((uintptr_t) a5 + a_offset);
115 }
116 a += 6;
117
118 // Inner accumulation loop along the 8 columns.
119 size_t k = kc;
120 // 2x partial unrolled loop to load 8 bytes at a time.
121 while (k >= 8 * sizeof(uint8_t)) {
122 // Load a 6x8 block of activations.
123 const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
124 const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 8;
125 const uint8x8_t va2x01234567 = vld1_u8(a2); a2 += 8;
126 const uint8x8_t va3x01234567 = vld1_u8(a3); a3 += 8;
127 const uint8x8_t va4x01234567 = vld1_u8(a4); a4 += 8;
128 const uint8x8_t va5x01234567 = vld1_u8(a5); a5 += 8;
129
130 // Load a 8x8 block of weights.
131 const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
132 const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
133 const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
134 const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
135
136 // Multiply-accumulate: 6x8 * 8x8 --> 6x8.
137 vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
138 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
139 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
140 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
141 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
142 vnacc1 = vdot_u32(vnacc1, va_zero_point, va1x01234567);
143 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
144 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
145 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1);
146 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb4567x4567, va1x01234567, 1);
147 vnacc2 = vdot_u32(vnacc2, va_zero_point, va2x01234567);
148 vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
149 vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
150 vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb4567x0123, va2x01234567, 1);
151 vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb4567x4567, va2x01234567, 1);
152 vnacc3 = vdot_u32(vnacc3, va_zero_point, va3x01234567);
153 vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb0123x0123, va3x01234567, 0);
154 vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb0123x4567, va3x01234567, 0);
155 vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb4567x0123, va3x01234567, 1);
156 vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb4567x4567, va3x01234567, 1);
157 vnacc4 = vdot_u32(vnacc4, va_zero_point, va4x01234567);
158 vpacc4x0123 = vdotq_lane_u32(vpacc4x0123, vb0123x0123, va4x01234567, 0);
159 vpacc4x4567 = vdotq_lane_u32(vpacc4x4567, vb0123x4567, va4x01234567, 0);
160 vpacc4x0123 = vdotq_lane_u32(vpacc4x0123, vb4567x0123, va4x01234567, 1);
161 vpacc4x4567 = vdotq_lane_u32(vpacc4x4567, vb4567x4567, va4x01234567, 1);
162 vnacc5 = vdot_u32(vnacc5, va_zero_point, va5x01234567);
163 vpacc5x0123 = vdotq_lane_u32(vpacc5x0123, vb0123x0123, va5x01234567, 0);
164 vpacc5x4567 = vdotq_lane_u32(vpacc5x4567, vb0123x4567, va5x01234567, 0);
165 vpacc5x0123 = vdotq_lane_u32(vpacc5x0123, vb4567x0123, va5x01234567, 1);
166 vpacc5x4567 = vdotq_lane_u32(vpacc5x4567, vb4567x4567, va5x01234567, 1);
167
168 k -= 8 * sizeof(uint8_t);
169 }
170 // Handle up to 4 final positions of `k`
171 if XNN_UNLIKELY(k != 0) {
172 // Load a 6x4 block of activations.
173 const uint8x8_t va0x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a0, vmov_n_u32(0), 0)); a0 += 4;
174 const uint8x8_t va1x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a1, vmov_n_u32(0), 0)); a1 += 4;
175 const uint8x8_t va2x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a2, vmov_n_u32(0), 0)); a2 += 4;
176 const uint8x8_t va3x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a3, vmov_n_u32(0), 0)); a3 += 4;
177 const uint8x8_t va4x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a4, vmov_n_u32(0), 0)); a4 += 4;
178 const uint8x8_t va5x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a5, vmov_n_u32(0), 0)); a5 += 4;
179
180 // Load a 4x8 block of weights.
181 const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
182 const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
183
184 // Multiply-accumulate: 6x4 * 4x8 --> 6x8.
185 vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
186 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
187 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
188 vnacc1 = vdot_u32(vnacc1, va_zero_point, va1x01234567);
189 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
190 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
191 vnacc2 = vdot_u32(vnacc2, va_zero_point, va2x01234567);
192 vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
193 vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
194 vnacc3 = vdot_u32(vnacc3, va_zero_point, va3x01234567);
195 vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb0123x0123, va3x01234567, 0);
196 vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb0123x4567, va3x01234567, 0);
197 vnacc4 = vdot_u32(vnacc4, va_zero_point, va4x01234567);
198 vpacc4x0123 = vdotq_lane_u32(vpacc4x0123, vb0123x0123, va4x01234567, 0);
199 vpacc4x4567 = vdotq_lane_u32(vpacc4x4567, vb0123x4567, va4x01234567, 0);
200 vnacc5 = vdot_u32(vnacc5, va_zero_point, va5x01234567);
201 vpacc5x0123 = vdotq_lane_u32(vpacc5x0123, vb0123x0123, va5x01234567, 0);
202 vpacc5x4567 = vdotq_lane_u32(vpacc5x4567, vb0123x4567, va5x01234567, 0);
203 }
204 p -= 6 * sizeof(void*);
205 } while (p != 0);
206
207 // Subtract zero point from accumulators.
208 vnacc0 = vpadd_u32(vnacc0, vnacc0);
209 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0);
210 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
211 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123));
212 vnacc1 = vpadd_u32(vnacc1, vnacc1);
213 const uint32x4_t vnacc1x0123 = vcombine_u32(vnacc1, vnacc1);
214 int32x4_t vacc1x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x0123, vnacc1x0123));
215 int32x4_t vacc1x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x4567, vnacc1x0123));
216 vnacc2 = vpadd_u32(vnacc2, vnacc2);
217 const uint32x4_t vnacc2x0123 = vcombine_u32(vnacc2, vnacc2);
218 int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123));
219 int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x0123));
220 vnacc3 = vpadd_u32(vnacc3, vnacc3);
221 const uint32x4_t vnacc3x0123 = vcombine_u32(vnacc3, vnacc3);
222 int32x4_t vacc3x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc3x0123, vnacc3x0123));
223 int32x4_t vacc3x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc3x4567, vnacc3x0123));
224 vnacc4 = vpadd_u32(vnacc4, vnacc4);
225 const uint32x4_t vnacc4x0123 = vcombine_u32(vnacc4, vnacc4);
226 int32x4_t vacc4x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc4x0123, vnacc4x0123));
227 int32x4_t vacc4x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc4x4567, vnacc4x0123));
228 vnacc5 = vpadd_u32(vnacc5, vnacc5);
229 const uint32x4_t vnacc5x0123 = vcombine_u32(vnacc5, vnacc5);
230 int32x4_t vacc5x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc5x0123, vnacc5x0123));
231 int32x4_t vacc5x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc5x4567, vnacc5x0123));
232
233 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
234 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
235 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
236
237 vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
238 vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
239 vacc1x0123 = vshlq_s32(vacc1x0123, vright_pre_shift);
240 vacc1x4567 = vshlq_s32(vacc1x4567, vright_pre_shift);
241 vacc2x0123 = vshlq_s32(vacc2x0123, vright_pre_shift);
242 vacc2x4567 = vshlq_s32(vacc2x4567, vright_pre_shift);
243 vacc3x0123 = vshlq_s32(vacc3x0123, vright_pre_shift);
244 vacc3x4567 = vshlq_s32(vacc3x4567, vright_pre_shift);
245 vacc4x0123 = vshlq_s32(vacc4x0123, vright_pre_shift);
246 vacc4x4567 = vshlq_s32(vacc4x4567, vright_pre_shift);
247 vacc5x0123 = vshlq_s32(vacc5x0123, vright_pre_shift);
248 vacc5x4567 = vshlq_s32(vacc5x4567, vright_pre_shift);
249
250 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
251 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
252 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
253 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
254 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
255 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
256 vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
257 vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
258 vacc4x0123 = vqdmulhq_s32(vacc4x0123, vmultiplier);
259 vacc4x4567 = vqdmulhq_s32(vacc4x4567, vmultiplier);
260 vacc5x0123 = vqdmulhq_s32(vacc5x0123, vmultiplier);
261 vacc5x4567 = vqdmulhq_s32(vacc5x4567, vmultiplier);
262
263 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
264 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
265 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
266 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
267 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
268 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
269 vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
270 vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
271 vacc4x0123 = vrshlq_s32(vacc4x0123, vright_post_shift);
272 vacc4x4567 = vrshlq_s32(vacc4x4567, vright_post_shift);
273 vacc5x0123 = vrshlq_s32(vacc5x0123, vright_post_shift);
274 vacc5x4567 = vrshlq_s32(vacc5x4567, vright_post_shift);
275
276 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
277 #if XNN_ARCH_ARM64
278 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
279 const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
280 const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
281 const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point);
282 const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), voutput_zero_point);
283 const int16x8_t vacc5x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc5x0123), vacc5x4567), voutput_zero_point);
284
285 uint8x16_t vout0x01234567_1x01234567 = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc1x01234567);
286 uint8x16_t vout2x01234567_3x01234567 = vqmovun_high_s16(vqmovun_s16(vacc2x01234567), vacc3x01234567);
287 uint8x16_t vout4x01234567_5x01234567 = vqmovun_high_s16(vqmovun_s16(vacc4x01234567), vacc5x01234567);
288 #else
289 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
290 const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
291 const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
292 const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point);
293 const int16x8_t vacc4x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x4567)), voutput_zero_point);
294 const int16x8_t vacc5x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc5x0123), vqmovn_s32(vacc5x4567)), voutput_zero_point);
295
296 uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc1x01234567));
297 uint8x16_t vout2x01234567_3x01234567 = vcombine_u8(vqmovun_s16(vacc2x01234567), vqmovun_s16(vacc3x01234567));
298 uint8x16_t vout4x01234567_5x01234567 = vcombine_u8(vqmovun_s16(vacc4x01234567), vqmovun_s16(vacc5x01234567));
299 #endif
300 const uint8x16_t voutput_min = vld1q_dup_u8(¶ms->rndnu_neon.output_min);
301 const uint8x16_t voutput_max = vld1q_dup_u8(¶ms->rndnu_neon.output_max);
302
303 vout0x01234567_1x01234567 = vmaxq_u8(vout0x01234567_1x01234567, voutput_min);
304 vout2x01234567_3x01234567 = vmaxq_u8(vout2x01234567_3x01234567, voutput_min);
305 vout4x01234567_5x01234567 = vmaxq_u8(vout4x01234567_5x01234567, voutput_min);
306
307 vout0x01234567_1x01234567 = vminq_u8(vout0x01234567_1x01234567, voutput_max);
308 vout2x01234567_3x01234567 = vminq_u8(vout2x01234567_3x01234567, voutput_max);
309 vout4x01234567_5x01234567 = vminq_u8(vout4x01234567_5x01234567, voutput_max);
310
311 if (nc >= 8) {
312 vst1_u8(c5 + 0, vget_high_u8(vout4x01234567_5x01234567));
313 vst1_u8(c4 + 0, vget_low_u8(vout4x01234567_5x01234567));
314 vst1_u8(c3 + 0, vget_high_u8(vout2x01234567_3x01234567));
315 vst1_u8(c2 + 0, vget_low_u8(vout2x01234567_3x01234567));
316 vst1_u8(c1 + 0, vget_high_u8(vout0x01234567_1x01234567));
317 vst1_u8(c0 + 0, vget_low_u8(vout0x01234567_1x01234567));
318
319 c5 = (uint8_t*) ((uintptr_t) c5 + cn_stride);
320 c4 = (uint8_t*) ((uintptr_t) c4 + cn_stride);
321 c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
322 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
323 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
324 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
325
326 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
327
328 nc -= 8;
329 } else {
330 if (nc & 4) {
331 vst1q_lane_u32((void*) c5, vreinterpretq_u32_u8(vout4x01234567_5x01234567), 2); c5 += 4;
332 vst1q_lane_u32((void*) c4, vreinterpretq_u32_u8(vout4x01234567_5x01234567), 0); c4 += 4;
333 vst1q_lane_u32((void*) c3, vreinterpretq_u32_u8(vout2x01234567_3x01234567), 2); c3 += 4;
334 vst1q_lane_u32((void*) c2, vreinterpretq_u32_u8(vout2x01234567_3x01234567), 0); c2 += 4;
335 vst1q_lane_u32((void*) c1, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
336 vst1q_lane_u32((void*) c0, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
337 vout4x01234567_5x01234567 = vextq_u8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 4);
338 vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
339 vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
340 }
341 if (nc & 2) {
342 vst1q_lane_u16((void*) c5, vreinterpretq_u16_u8(vout4x01234567_5x01234567), 4); c5 += 2;
343 vst1q_lane_u16((void*) c4, vreinterpretq_u16_u8(vout4x01234567_5x01234567), 0); c4 += 2;
344 vst1q_lane_u16((void*) c3, vreinterpretq_u16_u8(vout2x01234567_3x01234567), 4); c3 += 2;
345 vst1q_lane_u16((void*) c2, vreinterpretq_u16_u8(vout2x01234567_3x01234567), 0); c2 += 2;
346 vst1q_lane_u16((void*) c1, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
347 vst1q_lane_u16((void*) c0, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
348 vout4x01234567_5x01234567 = vextq_u8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 2);
349 vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
350 vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
351 }
352 if (nc & 1) {
353 vst1q_lane_u8(c5, vout4x01234567_5x01234567, 8);
354 vst1q_lane_u8(c4, vout4x01234567_5x01234567, 0);
355 vst1q_lane_u8(c3, vout2x01234567_3x01234567, 8);
356 vst1q_lane_u8(c2, vout2x01234567_3x01234567, 0);
357 vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
358 vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
359 }
360
361 nc = 0;
362 }
363 } while (nc != 0);
364 }
365