1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup(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_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const int8_t** restrict a,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 2);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (2 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 4 * sizeof(int8_t));
44 int8_t* c0 = c;
45 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr != 2) {
47 c1 = c0;
48 }
49
50 do {
51 int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
52 int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
53 int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
54 int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
55 int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
56 int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
57 int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
58 int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
59 int32x4_t vacc1x01 = vacc0x01;
60 int32x4_t vacc1x23 = vacc0x23;
61 int32x4_t vacc1x45 = vacc0x45;
62 int32x4_t vacc1x67 = vacc0x67;
63 int32x4_t vacc1x89 = vacc0x89;
64 int32x4_t vacc1xAB = vacc0xAB;
65 int32x4_t vacc1xCD = vacc0xCD;
66 int32x4_t vacc1xEF = vacc0xEF;
67
68 size_t p = ks;
69 do {
70 const int8_t* restrict a0 = a[0];
71 if XNN_UNPREDICTABLE(a0 != zero) {
72 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
73 }
74 const int8_t* restrict a1 = a[1];
75 if XNN_UNPREDICTABLE(a1 != zero) {
76 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
77 }
78 a += 2;
79
80 size_t k = kc;
81
82
83 while (k >= 8 * sizeof(int8_t)) {
84 const int8x8_t va0 = vld1_s8(a0); a0 += 8;
85 const int8x8_t va1 = vld1_s8(a1); a1 += 8;
86
87 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
88 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
89 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
90 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
91 const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
92 const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
93 const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
94 const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
95 const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
96 const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
97 const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
98 const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
99 const int8x8_t vb89c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
100 const int8x8_t vbABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
101 const int8x8_t vbCDc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
102 const int8x8_t vbEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
103
104 const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
105 const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
106
107 const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
108 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
109 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
110 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
111 const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
112 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
113 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
114 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
115 const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
116 const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
117 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
118 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
119 const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
120 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
121 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
122 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
123 const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
124 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0);
125 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
126 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
127 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
128 const int16x8_t vprod1xABc0 = vmull_s8(vbABc0, va1c0);
129 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
130 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
131 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
132 const int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0, va1c0);
133 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
134 vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
135 const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
136 const int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0, va1c0);
137 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
138 vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
139 const int8x8_t va0c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 1));
140 const int8x8_t va1c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 1));
141
142 const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
143 const int16x8_t vprod1x01c1 = vmull_s8(vb01c1, va1c1);
144 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
145 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
146 const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
147 const int16x8_t vprod1x23c1 = vmull_s8(vb23c1, va1c1);
148 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
149 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
150 const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
151 const int16x8_t vprod1x45c1 = vmull_s8(vb45c1, va1c1);
152 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
153 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
154 const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
155 const int16x8_t vprod1x67c1 = vmull_s8(vb67c1, va1c1);
156 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
157 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
158 const int16x8_t vprod0x89c1 = vmull_s8(vb89c1, va0c1);
159 const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1);
160 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
161 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
162 const int16x8_t vprod0xABc1 = vmull_s8(vbABc1, va0c1);
163 const int16x8_t vprod1xABc1 = vmull_s8(vbABc1, va1c1);
164 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
165 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
166 const int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1, va0c1);
167 const int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1, va1c1);
168 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
169 vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
170 const int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1, va0c1);
171 const int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1, va1c1);
172 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
173 vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
174
175 k -= 8 * sizeof(int8_t);
176 }
177
178 if XNN_UNLIKELY(k != 0) {
179 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
180 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
181
182 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
183 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
184 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
185 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
186 const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
187 const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
188 const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
189 const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
190
191 const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
192 const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
193 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
194 const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
195 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
196 const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
197 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
198 const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
199 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
200 const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
201 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
202 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
203 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
204 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
205 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
206 const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
207 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
208 const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
209 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
210 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
211 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
212 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
213 const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
214 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
215 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
216 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
217 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0);
218 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
219 const int16x8_t vprod1xABc0 = vmull_s8(vbABc0, va1c0);
220 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
221 const int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0, va1c0);
222 vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
223 const int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0, va1c0);
224 vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
225 }
226 p -= 2 * sizeof(void*);
227 } while (p != 0);
228
229 #if XNN_ARCH_ARM64
230 int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
231 int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
232 int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB);
233 int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF);
234 int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
235 int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
236 int32x4_t vacc1x89AB = vpaddq_s32(vacc1x89, vacc1xAB);
237 int32x4_t vacc1xCDEF = vpaddq_s32(vacc1xCD, vacc1xEF);
238 #else
239 const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
240 const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
241 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
242 const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
243 const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
244 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
245 const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89));
246 const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB));
247 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB);
248 const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD));
249 const int32x2_t vsum0xEF = vpadd_s32(vget_low_s32(vacc0xEF), vget_high_s32(vacc0xEF));
250 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF);
251 const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
252 const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
253 int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
254 const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
255 const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
256 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
257 const int32x2_t vsum1x89 = vpadd_s32(vget_low_s32(vacc1x89), vget_high_s32(vacc1x89));
258 const int32x2_t vsum1xAB = vpadd_s32(vget_low_s32(vacc1xAB), vget_high_s32(vacc1xAB));
259 int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB);
260 const int32x2_t vsum1xCD = vpadd_s32(vget_low_s32(vacc1xCD), vget_high_s32(vacc1xCD));
261 const int32x2_t vsum1xEF = vpadd_s32(vget_low_s32(vacc1xEF), vget_high_s32(vacc1xEF));
262 int32x4_t vacc1xCDEF = vcombine_s32(vsum1xCD, vsum1xEF);
263 #endif
264
265 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
266 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
267 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
268
269 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
270 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
271 vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
272 vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
273 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
274 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
275 vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
276 vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
277
278 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
279 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
280 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
281 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
282 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
283 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
284 vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
285 vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
286
287 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
288 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
289 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
290 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
291 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
292 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
293 vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
294 vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
295
296 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
297 #if XNN_ARCH_ARM64
298 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
299 int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
300 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
301 int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
302
303 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
304 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
305 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
306 vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
307
308 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
309 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
310 #else
311 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
312 int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
313 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
314 int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
315
316 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
317 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
318 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
319 vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
320
321 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
322 int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
323 #endif
324
325 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
326 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
327 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
328
329 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
330 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
331 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
332
333 if (nc >= 16) {
334 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
335 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
336
337 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
338 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
339
340 a = (const int8_t**restrict) ((uintptr_t) a - ks);
341
342 nc -= 16;
343 } else {
344 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
345 if (nc & 8) {
346 vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
347 vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
348 vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
349 }
350 if (nc & 4) {
351 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
352 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
353 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
354 }
355 if (nc & 2) {
356 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
357 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
358 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
359 }
360 if (nc & 1) {
361 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
362 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
363 }
364
365 nc = 0;
366 }
367 } while (nc != 0);
368 }
369