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_3x16c4__neon_mlal_ld1r(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_3x16c4__neon_mlal_ld1r(
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 <= 3);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (3 * 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 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 c2 = c1;
52 }
53
54 do {
55 int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
56 int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
57 int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
58 int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
59 int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
60 int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
61 int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
62 int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
63 int32x4_t vacc1x01 = vacc0x01;
64 int32x4_t vacc1x23 = vacc0x23;
65 int32x4_t vacc1x45 = vacc0x45;
66 int32x4_t vacc1x67 = vacc0x67;
67 int32x4_t vacc1x89 = vacc0x89;
68 int32x4_t vacc1xAB = vacc0xAB;
69 int32x4_t vacc1xCD = vacc0xCD;
70 int32x4_t vacc1xEF = vacc0xEF;
71 int32x4_t vacc2x01 = vacc0x01;
72 int32x4_t vacc2x23 = vacc0x23;
73 int32x4_t vacc2x45 = vacc0x45;
74 int32x4_t vacc2x67 = vacc0x67;
75 int32x4_t vacc2x89 = vacc0x89;
76 int32x4_t vacc2xAB = vacc0xAB;
77 int32x4_t vacc2xCD = vacc0xCD;
78 int32x4_t vacc2xEF = vacc0xEF;
79
80 size_t p = ks;
81 do {
82 const int8_t* restrict a0 = a[0];
83 if XNN_UNPREDICTABLE(a0 != zero) {
84 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
85 }
86 const int8_t* restrict a1 = a[1];
87 if XNN_UNPREDICTABLE(a1 != zero) {
88 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
89 }
90 const int8_t* restrict a2 = a[2];
91 if XNN_UNPREDICTABLE(a2 != zero) {
92 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
93 }
94 a += 3;
95
96 size_t k = kc;
97
98 while (k >= 16 * sizeof(int8_t)) {
99 const int32x2_t va00x0 = vld1_dup_s32((const void*)a0);
100 const int32x2_t va01x0 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
101 const int32x2_t va00x1 = vld1_dup_s32((const void*)a0);
102 const int32x2_t va01x1 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
103 const int32x2_t va10x0 = vld1_dup_s32((const void*)a1);
104 const int32x2_t va11x0 = vld1_dup_s32((const void*)(a1 + 4)); a1 += 8;
105 const int32x2_t va10x1 = vld1_dup_s32((const void*)a1);
106 const int32x2_t va11x1 = vld1_dup_s32((const void*)(a1 + 4)); a1 += 8;
107 const int32x2_t va20x0 = vld1_dup_s32((const void*)a2);
108 const int32x2_t va21x0 = vld1_dup_s32((const void*)(a2 + 4)); a2 += 8;
109 const int32x2_t va20x1 = vld1_dup_s32((const void*)a2);
110 const int32x2_t va21x1 = vld1_dup_s32((const void*)(a2 + 4)); a2 += 8;
111
112 const int8x8_t vb01c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
113 const int8x8_t vb23c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
114 const int8x8_t vb45c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
115 const int8x8_t vb67c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
116 const int8x8_t vb89c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
117 const int8x8_t vbABc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
118 const int8x8_t vbCDc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
119 const int8x8_t vbEFc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
120 const int8x8_t vb01c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
121 const int8x8_t vb23c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
122 const int8x8_t vb45c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
123 const int8x8_t vb67c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
124 const int8x8_t vb89c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
125 const int8x8_t vbABc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
126 const int8x8_t vbCDc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
127 const int8x8_t vbEFc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
128
129 const int8x8_t va0c0x0 = vreinterpret_s8_s32(va00x0);
130 const int8x8_t va0c0x1 = vreinterpret_s8_s32(va00x1);
131 const int8x8_t va1c0x0 = vreinterpret_s8_s32(va10x0);
132 const int8x8_t va1c0x1 = vreinterpret_s8_s32(va10x1);
133 const int8x8_t va2c0x0 = vreinterpret_s8_s32(va20x0);
134 const int8x8_t va2c0x1 = vreinterpret_s8_s32(va20x1);
135
136 int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0c0x0);
137 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0);
138 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2c0x0);
139 const int8x8_t vb01c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
140 vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0c0x1);
141 vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1);
142 vprod2x01c0 = vmlal_s8(vprod2x01c0, vb01c0x1, va2c0x1);
143 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
144 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
145 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
146 int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0c0x0);
147 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0);
148 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0);
149 const int8x8_t vb23c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
150 vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0c0x1);
151 vprod1x23c0 = vmlal_s8(vprod1x23c0, vb23c0x1, va1c0x1);
152 vprod2x23c0 = vmlal_s8(vprod2x23c0, vb23c0x1, va2c0x1);
153 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
154 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
155 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
156 int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0c0x0);
157 int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1c0x0);
158 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2c0x0);
159 const int8x8_t vb45c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
160 vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0c0x1);
161 vprod1x45c0 = vmlal_s8(vprod1x45c0, vb45c0x1, va1c0x1);
162 vprod2x45c0 = vmlal_s8(vprod2x45c0, vb45c0x1, va2c0x1);
163 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
164 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
165 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
166 int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0c0x0);
167 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0);
168 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0);
169 const int8x8_t vb67c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
170 vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0c0x1);
171 vprod1x67c0 = vmlal_s8(vprod1x67c0, vb67c0x1, va1c0x1);
172 vprod2x67c0 = vmlal_s8(vprod2x67c0, vb67c0x1, va2c0x1);
173 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
174 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
175 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
176 int16x8_t vprod0x89c0 = vmull_s8(vb89c0x0, va0c0x0);
177 int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1c0x0);
178 int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2c0x0);
179 const int8x8_t vb89c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
180 vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0c0x1);
181 vprod1x89c0 = vmlal_s8(vprod1x89c0, vb89c0x1, va1c0x1);
182 vprod2x89c0 = vmlal_s8(vprod2x89c0, vb89c0x1, va2c0x1);
183 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
184 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
185 vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c0);
186 int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0c0x0);
187 int16x8_t vprod1xABc0 = vmull_s8(vbABc0x0, va1c0x0);
188 int16x8_t vprod2xABc0 = vmull_s8(vbABc0x0, va2c0x0);
189 const int8x8_t vbABc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
190 vprod0xABc0 = vmlal_s8(vprod0xABc0, vbABc0x1, va0c0x1);
191 vprod1xABc0 = vmlal_s8(vprod1xABc0, vbABc0x1, va1c0x1);
192 vprod2xABc0 = vmlal_s8(vprod2xABc0, vbABc0x1, va2c0x1);
193 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
194 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
195 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0);
196 int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0c0x0);
197 int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0x0, va1c0x0);
198 int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2c0x0);
199 const int8x8_t vbCDc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
200 vprod0xCDc0 = vmlal_s8(vprod0xCDc0, vbCDc0x1, va0c0x1);
201 vprod1xCDc0 = vmlal_s8(vprod1xCDc0, vbCDc0x1, va1c0x1);
202 vprod2xCDc0 = vmlal_s8(vprod2xCDc0, vbCDc0x1, va2c0x1);
203 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
204 vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
205 vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc0);
206 int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0x0, va0c0x0);
207 int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0x0, va1c0x0);
208 int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0x0, va2c0x0);
209 const int8x8_t vbEFc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
210 vprod0xEFc0 = vmlal_s8(vprod0xEFc0, vbEFc0x1, va0c0x1);
211 vprod1xEFc0 = vmlal_s8(vprod1xEFc0, vbEFc0x1, va1c0x1);
212 vprod2xEFc0 = vmlal_s8(vprod2xEFc0, vbEFc0x1, va2c0x1);
213 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
214 vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
215 vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc0);
216 const int8x8_t va0c1x0 = vreinterpret_s8_s32(va01x0);
217 const int8x8_t va0c1x1 = vreinterpret_s8_s32(va01x1);
218 const int8x8_t va1c1x0 = vreinterpret_s8_s32(va11x0);
219 const int8x8_t va1c1x1 = vreinterpret_s8_s32(va11x1);
220 const int8x8_t va2c1x0 = vreinterpret_s8_s32(va21x0);
221 const int8x8_t va2c1x1 = vreinterpret_s8_s32(va21x1);
222
223 int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0c1x0);
224 int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1c1x0);
225 int16x8_t vprod2x01c1 = vmull_s8(vb01c1x0, va2c1x0);
226 const int8x8_t vb01c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
227 vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0c1x1);
228 vprod1x01c1 = vmlal_s8(vprod1x01c1, vb01c1x1, va1c1x1);
229 vprod2x01c1 = vmlal_s8(vprod2x01c1, vb01c1x1, va2c1x1);
230 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
231 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
232 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
233 int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0c1x0);
234 int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1c1x0);
235 int16x8_t vprod2x23c1 = vmull_s8(vb23c1x0, va2c1x0);
236 const int8x8_t vb23c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
237 vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0c1x1);
238 vprod1x23c1 = vmlal_s8(vprod1x23c1, vb23c1x1, va1c1x1);
239 vprod2x23c1 = vmlal_s8(vprod2x23c1, vb23c1x1, va2c1x1);
240 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
241 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
242 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
243 int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0c1x0);
244 int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1c1x0);
245 int16x8_t vprod2x45c1 = vmull_s8(vb45c1x0, va2c1x0);
246 const int8x8_t vb45c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
247 vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0c1x1);
248 vprod1x45c1 = vmlal_s8(vprod1x45c1, vb45c1x1, va1c1x1);
249 vprod2x45c1 = vmlal_s8(vprod2x45c1, vb45c1x1, va2c1x1);
250 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
251 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
252 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
253 int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0c1x0);
254 int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1c1x0);
255 int16x8_t vprod2x67c1 = vmull_s8(vb67c1x0, va2c1x0);
256 const int8x8_t vb67c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
257 vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0c1x1);
258 vprod1x67c1 = vmlal_s8(vprod1x67c1, vb67c1x1, va1c1x1);
259 vprod2x67c1 = vmlal_s8(vprod2x67c1, vb67c1x1, va2c1x1);
260 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
261 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
262 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
263 int16x8_t vprod0x89c1 = vmull_s8(vb89c1x0, va0c1x0);
264 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1c1x0);
265 int16x8_t vprod2x89c1 = vmull_s8(vb89c1x0, va2c1x0);
266 const int8x8_t vb89c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
267 vprod0x89c1 = vmlal_s8(vprod0x89c1, vb89c1x1, va0c1x1);
268 vprod1x89c1 = vmlal_s8(vprod1x89c1, vb89c1x1, va1c1x1);
269 vprod2x89c1 = vmlal_s8(vprod2x89c1, vb89c1x1, va2c1x1);
270 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
271 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
272 vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c1);
273 int16x8_t vprod0xABc1 = vmull_s8(vbABc1x0, va0c1x0);
274 int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1c1x0);
275 int16x8_t vprod2xABc1 = vmull_s8(vbABc1x0, va2c1x0);
276 const int8x8_t vbABc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
277 vprod0xABc1 = vmlal_s8(vprod0xABc1, vbABc1x1, va0c1x1);
278 vprod1xABc1 = vmlal_s8(vprod1xABc1, vbABc1x1, va1c1x1);
279 vprod2xABc1 = vmlal_s8(vprod2xABc1, vbABc1x1, va2c1x1);
280 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
281 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
282 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc1);
283 int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1x0, va0c1x0);
284 int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1c1x0);
285 int16x8_t vprod2xCDc1 = vmull_s8(vbCDc1x0, va2c1x0);
286 const int8x8_t vbCDc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
287 vprod0xCDc1 = vmlal_s8(vprod0xCDc1, vbCDc1x1, va0c1x1);
288 vprod1xCDc1 = vmlal_s8(vprod1xCDc1, vbCDc1x1, va1c1x1);
289 vprod2xCDc1 = vmlal_s8(vprod2xCDc1, vbCDc1x1, va2c1x1);
290 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
291 vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
292 vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc1);
293 int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1x0, va0c1x0);
294 int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1x0, va1c1x0);
295 int16x8_t vprod2xEFc1 = vmull_s8(vbEFc1x0, va2c1x0);
296 const int8x8_t vbEFc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
297 vprod0xEFc1 = vmlal_s8(vprod0xEFc1, vbEFc1x1, va0c1x1);
298 vprod1xEFc1 = vmlal_s8(vprod1xEFc1, vbEFc1x1, va1c1x1);
299 vprod2xEFc1 = vmlal_s8(vprod2xEFc1, vbEFc1x1, va2c1x1);
300 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
301 vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
302 vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc1);
303
304 k -= 16 * sizeof(int8_t);
305 }
306
307 if (k >= 8 * sizeof(int8_t)) {
308 const int32x2_t va00 = vld1_dup_s32((const void*)a0);
309 const int32x2_t va01 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
310 const int32x2_t va10 = vld1_dup_s32((const void*)a1);
311 const int32x2_t va11 = vld1_dup_s32((const void*)(a1 + 4)); a1 += 8;
312 const int32x2_t va20 = vld1_dup_s32((const void*)a2);
313 const int32x2_t va21 = vld1_dup_s32((const void*)(a2 + 4)); a2 += 8;
314
315 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
316 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
317 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
318 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
319 const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
320 const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
321 const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
322 const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
323 const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
324 const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
325 const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
326 const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
327 const int8x8_t vb89c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
328 const int8x8_t vbABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
329 const int8x8_t vbCDc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
330 const int8x8_t vbEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
331
332 const int8x8_t va0c0 = vreinterpret_s8_s32(va00);
333 const int8x8_t va1c0 = vreinterpret_s8_s32(va10);
334 const int8x8_t va2c0 = vreinterpret_s8_s32(va20);
335
336 const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
337 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
338 const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
339 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
340 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
341 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
342 const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
343 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
344 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0);
345 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
346 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
347 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
348 const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
349 const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
350 const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
351 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
352 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
353 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
354 const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
355 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
356 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
357 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
358 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
359 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
360 const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
361 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0);
362 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0);
363 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
364 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
365 vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c0);
366 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
367 const int16x8_t vprod1xABc0 = vmull_s8(vbABc0, va1c0);
368 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0);
369 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
370 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
371 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0);
372 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
373 const int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0, va1c0);
374 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0);
375 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
376 vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
377 vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc0);
378 const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
379 const int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0, va1c0);
380 const int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0, va2c0);
381 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
382 vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
383 vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc0);
384 const int8x8_t va0c1 = vreinterpret_s8_s32(va01);
385 const int8x8_t va1c1 = vreinterpret_s8_s32(va11);
386 const int8x8_t va2c1 = vreinterpret_s8_s32(va21);
387
388 const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
389 const int16x8_t vprod1x01c1 = vmull_s8(vb01c1, va1c1);
390 const int16x8_t vprod2x01c1 = vmull_s8(vb01c1, va2c1);
391 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
392 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
393 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
394 const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
395 const int16x8_t vprod1x23c1 = vmull_s8(vb23c1, va1c1);
396 const int16x8_t vprod2x23c1 = vmull_s8(vb23c1, va2c1);
397 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
398 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
399 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
400 const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
401 const int16x8_t vprod1x45c1 = vmull_s8(vb45c1, va1c1);
402 const int16x8_t vprod2x45c1 = vmull_s8(vb45c1, va2c1);
403 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
404 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
405 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
406 const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
407 const int16x8_t vprod1x67c1 = vmull_s8(vb67c1, va1c1);
408 const int16x8_t vprod2x67c1 = vmull_s8(vb67c1, va2c1);
409 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
410 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
411 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
412 const int16x8_t vprod0x89c1 = vmull_s8(vb89c1, va0c1);
413 const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1);
414 const int16x8_t vprod2x89c1 = vmull_s8(vb89c1, va2c1);
415 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
416 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
417 vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c1);
418 const int16x8_t vprod0xABc1 = vmull_s8(vbABc1, va0c1);
419 const int16x8_t vprod1xABc1 = vmull_s8(vbABc1, va1c1);
420 const int16x8_t vprod2xABc1 = vmull_s8(vbABc1, va2c1);
421 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
422 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
423 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc1);
424 const int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1, va0c1);
425 const int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1, va1c1);
426 const int16x8_t vprod2xCDc1 = vmull_s8(vbCDc1, va2c1);
427 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
428 vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
429 vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc1);
430 const int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1, va0c1);
431 const int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1, va1c1);
432 const int16x8_t vprod2xEFc1 = vmull_s8(vbEFc1, va2c1);
433 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
434 vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
435 vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc1);
436
437 k -= 8 * sizeof(int8_t);
438 }
439
440 if XNN_UNLIKELY(k != 0) {
441 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
442 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
443 const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
444
445 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
446 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
447 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
448 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
449 const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
450 const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
451 const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
452 const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
453
454 const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
455 const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
456 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
457 const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
458 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
459 const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
460 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
461 const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
462 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
463 const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
464 vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
465 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
466 vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
467 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
468 vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
469 const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
470 vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
471 const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
472 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
473 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
474 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
475 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
476 const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
477 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
478 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
479 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
480 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0);
481 vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
482 const int16x8_t vprod1xABc0 = vmull_s8(vbABc0, va1c0);
483 vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
484 const int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0, va1c0);
485 vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
486 const int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0, va1c0);
487 vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
488 const int8x8_t va2c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2), 0));
489 const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
490 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
491 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0);
492 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
493 const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
494 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
495 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
496 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
497 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0);
498 vacc2x89 = vpadalq_s16(vacc2x89, vprod2x89c0);
499 const int16x8_t vprod2xABc0 = vmull_s8(vbABc0, va2c0);
500 vacc2xAB = vpadalq_s16(vacc2xAB, vprod2xABc0);
501 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0);
502 vacc2xCD = vpadalq_s16(vacc2xCD, vprod2xCDc0);
503 const int16x8_t vprod2xEFc0 = vmull_s8(vbEFc0, va2c0);
504 vacc2xEF = vpadalq_s16(vacc2xEF, vprod2xEFc0);
505 }
506 p -= 3 * sizeof(void*);
507 } while (p != 0);
508
509 #if XNN_ARCH_ARM64
510 int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
511 int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
512 int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB);
513 int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF);
514 int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
515 int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
516 int32x4_t vacc1x89AB = vpaddq_s32(vacc1x89, vacc1xAB);
517 int32x4_t vacc1xCDEF = vpaddq_s32(vacc1xCD, vacc1xEF);
518 int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23);
519 int32x4_t vacc2x4567 = vpaddq_s32(vacc2x45, vacc2x67);
520 int32x4_t vacc2x89AB = vpaddq_s32(vacc2x89, vacc2xAB);
521 int32x4_t vacc2xCDEF = vpaddq_s32(vacc2xCD, vacc2xEF);
522 #else
523 const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
524 const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
525 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
526 const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
527 const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
528 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
529 const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89));
530 const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB));
531 int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB);
532 const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD));
533 const int32x2_t vsum0xEF = vpadd_s32(vget_low_s32(vacc0xEF), vget_high_s32(vacc0xEF));
534 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF);
535 const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
536 const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
537 int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
538 const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
539 const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
540 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
541 const int32x2_t vsum1x89 = vpadd_s32(vget_low_s32(vacc1x89), vget_high_s32(vacc1x89));
542 const int32x2_t vsum1xAB = vpadd_s32(vget_low_s32(vacc1xAB), vget_high_s32(vacc1xAB));
543 int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB);
544 const int32x2_t vsum1xCD = vpadd_s32(vget_low_s32(vacc1xCD), vget_high_s32(vacc1xCD));
545 const int32x2_t vsum1xEF = vpadd_s32(vget_low_s32(vacc1xEF), vget_high_s32(vacc1xEF));
546 int32x4_t vacc1xCDEF = vcombine_s32(vsum1xCD, vsum1xEF);
547 const int32x2_t vsum2x01 = vpadd_s32(vget_low_s32(vacc2x01), vget_high_s32(vacc2x01));
548 const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23));
549 int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23);
550 const int32x2_t vsum2x45 = vpadd_s32(vget_low_s32(vacc2x45), vget_high_s32(vacc2x45));
551 const int32x2_t vsum2x67 = vpadd_s32(vget_low_s32(vacc2x67), vget_high_s32(vacc2x67));
552 int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67);
553 const int32x2_t vsum2x89 = vpadd_s32(vget_low_s32(vacc2x89), vget_high_s32(vacc2x89));
554 const int32x2_t vsum2xAB = vpadd_s32(vget_low_s32(vacc2xAB), vget_high_s32(vacc2xAB));
555 int32x4_t vacc2x89AB = vcombine_s32(vsum2x89, vsum2xAB);
556 const int32x2_t vsum2xCD = vpadd_s32(vget_low_s32(vacc2xCD), vget_high_s32(vacc2xCD));
557 const int32x2_t vsum2xEF = vpadd_s32(vget_low_s32(vacc2xEF), vget_high_s32(vacc2xEF));
558 int32x4_t vacc2xCDEF = vcombine_s32(vsum2xCD, vsum2xEF);
559 #endif
560
561 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
562 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
563 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
564
565 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
566 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
567 vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
568 vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
569 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
570 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
571 vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
572 vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
573 vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
574 vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
575 vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
576 vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
577
578 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
579 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
580 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
581 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
582 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
583 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
584 vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
585 vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
586 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
587 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
588 vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
589 vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
590
591 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
592 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
593 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
594 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
595 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
596 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
597 vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
598 vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
599 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
600 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
601 vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
602 vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
603
604 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
605 #if XNN_ARCH_ARM64
606 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
607 int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
608 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
609 int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
610 int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
611 int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
612
613 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
614 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
615 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
616 vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
617 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
618 vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
619
620 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
621 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
622 int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
623 #else
624 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
625 int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
626 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
627 int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
628 int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
629 int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
630
631 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
632 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
633 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
634 vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
635 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
636 vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
637
638 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
639 int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
640 int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
641 #endif
642
643 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
644 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
645 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
646 vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
647
648 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
649 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
650 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
651 vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
652
653 if (nc >= 16) {
654 vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
655 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
656 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
657
658 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
659 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
660 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
661
662 a = (const int8_t**restrict) ((uintptr_t) a - ks);
663
664 nc -= 16;
665 } else {
666 int8x8_t vout2x01234567 = vget_low_s8(vout2x0123456789ABCDEF);
667 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
668 if (nc & 8) {
669 vst1_s8(c2, vout2x01234567); c2 += 8;
670 vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
671 vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
672 vout2x01234567 = vget_high_s8(vout2x0123456789ABCDEF);
673 vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
674 }
675 if (nc & 4) {
676 vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
677 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
678 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
679 vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
680 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
681 }
682 if (nc & 2) {
683 vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
684 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
685 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
686 vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
687 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
688 }
689 if (nc & 1) {
690 vst1_lane_s8(c2, vout2x01234567, 0);
691 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
692 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
693 }
694
695 nc = 0;
696 }
697 } while (nc != 0);
698 }
699