1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/c4-neon-mull-dup.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_2x16c4__neon_mull_dup(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_2x16c4__neon_mull_dup(
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 <= 2);
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, 4 * 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
49 do {
50 int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
51 int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
52 int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
53 int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
54 int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
55 int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
56 int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
57 int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
58 int32x4_t vacc1x01 = vacc0x01;
59 int32x4_t vacc1x23 = vacc0x23;
60 int32x4_t vacc1x45 = vacc0x45;
61 int32x4_t vacc1x67 = vacc0x67;
62 int32x4_t vacc1x89 = vacc0x89;
63 int32x4_t vacc1xAB = vacc0xAB;
64 int32x4_t vacc1xCD = vacc0xCD;
65 int32x4_t vacc1xEF = vacc0xEF;
66
67 size_t k = kc;
68
69
70 while (k >= 8 * sizeof(int8_t)) {
71 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
72 const int8x8_t va1 = vld1_s8(a1); a1 += 8;
73
74 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
75 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
76 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
77 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
78 const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
79 const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
80 const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
81 const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
82 const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83 const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
84 const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
85 const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
86 const int8x8_t vb89c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
87 const int8x8_t vbABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
88 const int8x8_t vbCDc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
89 const int8x8_t vbEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
90
91 const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
92 const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
93
94 const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
95 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
96 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
97 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
98 const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
99 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
100 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
101 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
102 const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
103 const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
104 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
105 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
106 const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
107 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
108 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
109 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
110 const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
111 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0);
112 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
113 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
114 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
115 const int16x8_t vprod1xABc0 = vmull_s8(vbABc0, va1c0);
116 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
117 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
118 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
119 const int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0, va1c0);
120 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
121 vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
122 const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
123 const int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0, va1c0);
124 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
125 vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
126 const int8x8_t va0c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 1));
127 const int8x8_t va1c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 1));
128
129 const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
130 const int16x8_t vprod1x01c1 = vmull_s8(vb01c1, va1c1);
131 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
132 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
133 const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
134 const int16x8_t vprod1x23c1 = vmull_s8(vb23c1, va1c1);
135 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
136 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
137 const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
138 const int16x8_t vprod1x45c1 = vmull_s8(vb45c1, va1c1);
139 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
140 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
141 const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
142 const int16x8_t vprod1x67c1 = vmull_s8(vb67c1, va1c1);
143 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
144 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
145 const int16x8_t vprod0x89c1 = vmull_s8(vb89c1, va0c1);
146 const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1);
147 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
148 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
149 const int16x8_t vprod0xABc1 = vmull_s8(vbABc1, va0c1);
150 const int16x8_t vprod1xABc1 = vmull_s8(vbABc1, va1c1);
151 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
152 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
153 const int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1, va0c1);
154 const int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1, va1c1);
155 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
156 vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
157 const int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1, va0c1);
158 const int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1, va1c1);
159 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
160 vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
161
162 k -= 8 * sizeof(int8_t);
163 }
164
165 if XNN_UNLIKELY(k != 0) {
166 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
167 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
168
169 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
170 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
171 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
172 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
173 const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
174 const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
175 const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
176 const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
177
178 const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
179 const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
180 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
181 const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
182 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
183 const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
184 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
185 const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
186 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
187 const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
188 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
189 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
190 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
191 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
192 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
193 const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
194 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
195 const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
196 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
197 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
198 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
199 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
200 const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
201 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
202 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
203 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
204 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0);
205 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
206 const int16x8_t vprod1xABc0 = vmull_s8(vbABc0, va1c0);
207 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
208 const int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0, va1c0);
209 vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
210 const int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0, va1c0);
211 vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
212 }
213
214 #if XNN_ARCH_ARM64
215 int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
216 int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
217 int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB);
218 int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF);
219 int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
220 int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
221 int32x4_t vacc1x89AB = vpaddq_s32(vacc1x89, vacc1xAB);
222 int32x4_t vacc1xCDEF = vpaddq_s32(vacc1xCD, vacc1xEF);
223 #else
224 const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
225 const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
226 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
227 const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
228 const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
229 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
230 const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89));
231 const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB));
232 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB);
233 const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD));
234 const int32x2_t vsum0xEF = vpadd_s32(vget_low_s32(vacc0xEF), vget_high_s32(vacc0xEF));
235 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF);
236 const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
237 const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
238 int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
239 const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
240 const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
241 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
242 const int32x2_t vsum1x89 = vpadd_s32(vget_low_s32(vacc1x89), vget_high_s32(vacc1x89));
243 const int32x2_t vsum1xAB = vpadd_s32(vget_low_s32(vacc1xAB), vget_high_s32(vacc1xAB));
244 int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB);
245 const int32x2_t vsum1xCD = vpadd_s32(vget_low_s32(vacc1xCD), vget_high_s32(vacc1xCD));
246 const int32x2_t vsum1xEF = vpadd_s32(vget_low_s32(vacc1xEF), vget_high_s32(vacc1xEF));
247 int32x4_t vacc1xCDEF = vcombine_s32(vsum1xCD, vsum1xEF);
248 #endif
249
250 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
251 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
252 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
253
254 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
255 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
256 vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
257 vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
258 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
259 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
260 vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
261 vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
262
263 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
264 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
265 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
266 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
267 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
268 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
269 vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
270 vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
271
272 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
273 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
274 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
275 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
276 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
277 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
278 vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
279 vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
280
281 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
282 #if XNN_ARCH_ARM64
283 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
284 int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
285 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
286 int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
287
288 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
289 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
290 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
291 vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
292
293 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
294 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
295 #else
296 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
297 int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
298 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
299 int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
300
301 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
302 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
303 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
304 vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
305
306 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
307 int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
308 #endif
309
310 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
311 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
312 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
313
314 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
315 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
316 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
317
318 if (nc >= 16) {
319 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
320 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
321
322 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
323 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
324
325 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
326 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
327
328 nc -= 16;
329 } else {
330 // Final case where not all of the 16 columns fit in the destination.
331 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
332 if (nc & 8) {
333 vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
334 vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
335 vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
336 }
337 if (nc & 4) {
338 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
339 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
340 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
341 }
342 if (nc & 2) {
343 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
344 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
345 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
346 }
347 if (nc & 1) {
348 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
349 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
350 }
351
352 nc = 0;
353 }
354 } while (nc != 0);
355 }
356