1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/c16-neon-mlal.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 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/gemm.h>
15 #include <xnnpack/math.h>
16
17
xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c16__neon_mlal(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c16__neon_mlal(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const int8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(mr != 0);
31 assert(mr <= 3);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(int8_t) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38
39 kc = round_up_po2(kc, 16 * sizeof(int8_t));
40 const int8_t* a0 = a;
41 int8_t* c0 = c;
42 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
43 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
44 if XNN_UNPREDICTABLE(mr < 2) {
45 a1 = a0;
46 c1 = c0;
47 }
48 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
49 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 a2 = a1;
52 c2 = c1;
53 }
54
55 do {
56 int32x4_t vacc0x0 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
57 int32x4_t vacc0x1 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
58 int32x4_t vacc0x2 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
59 int32x4_t vacc0x3 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
60 int32x4_t vacc0x4 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
61 int32x4_t vacc0x5 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
62 int32x4_t vacc0x6 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
63 int32x4_t vacc0x7 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
64 int32x4_t vacc1x0 = vacc0x0;
65 int32x4_t vacc1x1 = vacc0x1;
66 int32x4_t vacc1x2 = vacc0x2;
67 int32x4_t vacc1x3 = vacc0x3;
68 int32x4_t vacc1x4 = vacc0x4;
69 int32x4_t vacc1x5 = vacc0x5;
70 int32x4_t vacc1x6 = vacc0x6;
71 int32x4_t vacc1x7 = vacc0x7;
72 int32x4_t vacc2x0 = vacc0x0;
73 int32x4_t vacc2x1 = vacc0x1;
74 int32x4_t vacc2x2 = vacc0x2;
75 int32x4_t vacc2x3 = vacc0x3;
76 int32x4_t vacc2x4 = vacc0x4;
77 int32x4_t vacc2x5 = vacc0x5;
78 int32x4_t vacc2x6 = vacc0x6;
79 int32x4_t vacc2x7 = vacc0x7;
80
81 // KC loop of 16
82 size_t k = kc;
83 while (k != 0) {
84 const int8x16_t va0 = vld1q_s8(a0); a0 += 16;
85 const int8x16_t va1 = vld1q_s8(a1); a1 += 16;
86 const int8x16_t va2 = vld1q_s8(a2); a2 += 16;
87
88 const int8x16_t vb0 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
89 const int8x16_t vb1 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
90 const int8x16_t vb2 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
91 const int8x16_t vb3 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
92 const int8x16_t vb4 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
93 const int8x16_t vb5 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
94 const int8x16_t vb6 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
95 const int8x16_t vb7 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
96
97 int16x8_t vprod0x0 = vmull_s8(vget_low_s8(vb0), vget_low_s8(va0));
98 int16x8_t vprod1x0 = vmull_s8(vget_low_s8(vb0), vget_low_s8(va1));
99 int16x8_t vprod2x0 = vmull_s8(vget_low_s8(vb0), vget_low_s8(va2));
100 vprod0x0 = vmlal_s8(vprod0x0, vget_high_s8(vb0), vget_high_s8(va0));
101 vprod1x0 = vmlal_s8(vprod1x0, vget_high_s8(vb0), vget_high_s8(va1));
102 vprod2x0 = vmlal_s8(vprod2x0, vget_high_s8(vb0), vget_high_s8(va2));
103 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
104 vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
105 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0);
106 int16x8_t vprod0x1 = vmull_s8(vget_low_s8(vb1), vget_low_s8(va0));
107 int16x8_t vprod1x1 = vmull_s8(vget_low_s8(vb1), vget_low_s8(va1));
108 int16x8_t vprod2x1 = vmull_s8(vget_low_s8(vb1), vget_low_s8(va2));
109 vprod0x1 = vmlal_s8(vprod0x1, vget_high_s8(vb1), vget_high_s8(va0));
110 vprod1x1 = vmlal_s8(vprod1x1, vget_high_s8(vb1), vget_high_s8(va1));
111 vprod2x1 = vmlal_s8(vprod2x1, vget_high_s8(vb1), vget_high_s8(va2));
112 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
113 vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
114 vacc2x1 = vpadalq_s16(vacc2x1, vprod2x1);
115 int16x8_t vprod0x2 = vmull_s8(vget_low_s8(vb2), vget_low_s8(va0));
116 int16x8_t vprod1x2 = vmull_s8(vget_low_s8(vb2), vget_low_s8(va1));
117 int16x8_t vprod2x2 = vmull_s8(vget_low_s8(vb2), vget_low_s8(va2));
118 vprod0x2 = vmlal_s8(vprod0x2, vget_high_s8(vb2), vget_high_s8(va0));
119 vprod1x2 = vmlal_s8(vprod1x2, vget_high_s8(vb2), vget_high_s8(va1));
120 vprod2x2 = vmlal_s8(vprod2x2, vget_high_s8(vb2), vget_high_s8(va2));
121 vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
122 vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
123 vacc2x2 = vpadalq_s16(vacc2x2, vprod2x2);
124 int16x8_t vprod0x3 = vmull_s8(vget_low_s8(vb3), vget_low_s8(va0));
125 int16x8_t vprod1x3 = vmull_s8(vget_low_s8(vb3), vget_low_s8(va1));
126 int16x8_t vprod2x3 = vmull_s8(vget_low_s8(vb3), vget_low_s8(va2));
127 vprod0x3 = vmlal_s8(vprod0x3, vget_high_s8(vb3), vget_high_s8(va0));
128 vprod1x3 = vmlal_s8(vprod1x3, vget_high_s8(vb3), vget_high_s8(va1));
129 vprod2x3 = vmlal_s8(vprod2x3, vget_high_s8(vb3), vget_high_s8(va2));
130 vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
131 vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
132 vacc2x3 = vpadalq_s16(vacc2x3, vprod2x3);
133 int16x8_t vprod0x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va0));
134 int16x8_t vprod1x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va1));
135 int16x8_t vprod2x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va2));
136 vprod0x4 = vmlal_s8(vprod0x4, vget_high_s8(vb4), vget_high_s8(va0));
137 vprod1x4 = vmlal_s8(vprod1x4, vget_high_s8(vb4), vget_high_s8(va1));
138 vprod2x4 = vmlal_s8(vprod2x4, vget_high_s8(vb4), vget_high_s8(va2));
139 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
140 vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
141 vacc2x4 = vpadalq_s16(vacc2x4, vprod2x4);
142 int16x8_t vprod0x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va0));
143 int16x8_t vprod1x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va1));
144 int16x8_t vprod2x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va2));
145 vprod0x5 = vmlal_s8(vprod0x5, vget_high_s8(vb5), vget_high_s8(va0));
146 vprod1x5 = vmlal_s8(vprod1x5, vget_high_s8(vb5), vget_high_s8(va1));
147 vprod2x5 = vmlal_s8(vprod2x5, vget_high_s8(vb5), vget_high_s8(va2));
148 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
149 vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
150 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5);
151 int16x8_t vprod0x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va0));
152 int16x8_t vprod1x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va1));
153 int16x8_t vprod2x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va2));
154 vprod0x6 = vmlal_s8(vprod0x6, vget_high_s8(vb6), vget_high_s8(va0));
155 vprod1x6 = vmlal_s8(vprod1x6, vget_high_s8(vb6), vget_high_s8(va1));
156 vprod2x6 = vmlal_s8(vprod2x6, vget_high_s8(vb6), vget_high_s8(va2));
157 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
158 vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
159 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6);
160 int16x8_t vprod0x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va0));
161 int16x8_t vprod1x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va1));
162 int16x8_t vprod2x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va2));
163 vprod0x7 = vmlal_s8(vprod0x7, vget_high_s8(vb7), vget_high_s8(va0));
164 vprod1x7 = vmlal_s8(vprod1x7, vget_high_s8(vb7), vget_high_s8(va1));
165 vprod2x7 = vmlal_s8(vprod2x7, vget_high_s8(vb7), vget_high_s8(va2));
166 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
167 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
168 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7);
169
170 k -= 16 * sizeof(int8_t);
171 }
172
173 #if XNN_ARCH_ARM64
174 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1);
175 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3);
176 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5);
177 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7);
178 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1);
179 const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3);
180 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5);
181 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7);
182 const int32x4_t vsum2x01 = vpaddq_s32(vacc2x0, vacc2x1);
183 const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3);
184 const int32x4_t vsum2x45 = vpaddq_s32(vacc2x4, vacc2x5);
185 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7);
186 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23);
187 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67);
188 int32x4_t vacc1x0123 = vpaddq_s32(vsum1x01, vsum1x23);
189 int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67);
190 int32x4_t vacc2x0123 = vpaddq_s32(vsum2x01, vsum2x23);
191 int32x4_t vacc2x4567 = vpaddq_s32(vsum2x45, vsum2x67);
192 #else
193 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0));
194 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1));
195 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2));
196 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3));
197 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1);
198 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3);
199 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 );
200 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4));
201 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5));
202 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6));
203 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7));
204 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5);
205 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7);
206 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 );
207 const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0));
208 const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1));
209 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2));
210 const int32x2_t vpsum1x3 = vadd_s32(vget_low_s32(vacc1x3), vget_high_s32(vacc1x3));
211 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1);
212 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3);
213 int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23 );
214 const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4));
215 const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(vacc1x5));
216 const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6));
217 const int32x2_t vpsum1x7 = vadd_s32(vget_low_s32(vacc1x7), vget_high_s32(vacc1x7));
218 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5);
219 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7);
220 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 );
221 const int32x2_t vpsum2x0 = vadd_s32(vget_low_s32(vacc2x0), vget_high_s32(vacc2x0));
222 const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1));
223 const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2));
224 const int32x2_t vpsum2x3 = vadd_s32(vget_low_s32(vacc2x3), vget_high_s32(vacc2x3));
225 const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1);
226 const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3);
227 int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23 );
228 const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4));
229 const int32x2_t vpsum2x5 = vadd_s32(vget_low_s32(vacc2x5), vget_high_s32(vacc2x5));
230 const int32x2_t vpsum2x6 = vadd_s32(vget_low_s32(vacc2x6), vget_high_s32(vacc2x6));
231 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7));
232 const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5);
233 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7);
234 int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67 );
235 #endif
236
237 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
238 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
239 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
240
241 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
242 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
243 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
244 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
245 vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
246 vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
247
248 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
249 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
250 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
251 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
252 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
253 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
254
255 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
256 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
257 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
258 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
259 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
260 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
261
262 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
263 #if XNN_ARCH_ARM64
264 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
265 const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
266 const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
267 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
268 int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
269 #else
270 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
271 const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
272 const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
273
274 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
275 int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
276 #endif
277 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
278 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
279
280 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
281 vout2x01234567 = vmax_s8(vout2x01234567, vget_low_s8(voutput_min));
282
283 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
284 vout2x01234567 = vmin_s8(vout2x01234567, vget_low_s8(voutput_max));
285
286 if (nc >= 8) {
287 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
288 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
289 vst1_s8(c2 + 0, vout2x01234567);
290
291 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
292 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
293 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
294
295 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
296 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
297 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
298
299 nc -= 8;
300 } else {
301 // Final case where not all of the 8 columns fit in the destination.
302 if (nc & 4) {
303 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
304 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
305 vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
306 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
307 vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
308 }
309 if (nc & 2) {
310 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
311 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
312 vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
313 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
314 vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
315 }
316 if (nc & 1) {
317 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
318 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
319 vst1_lane_s8(c2, vout2x01234567, 0);
320 }
321
322 nc = 0;
323 }
324 } while (nc != 0);
325 }
326