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_3x8c4__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_3x8c4__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 vacc1x01 = vacc0x01;
60 int32x4_t vacc1x23 = vacc0x23;
61 int32x4_t vacc1x45 = vacc0x45;
62 int32x4_t vacc1x67 = vacc0x67;
63 int32x4_t vacc2x01 = vacc0x01;
64 int32x4_t vacc2x23 = vacc0x23;
65 int32x4_t vacc2x45 = vacc0x45;
66 int32x4_t vacc2x67 = vacc0x67;
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 const int8_t* restrict a2 = a[2];
79 if XNN_UNPREDICTABLE(a2 != zero) {
80 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
81 }
82 a += 3;
83
84 size_t k = kc;
85
86 while (k >= 16 * sizeof(int8_t)) {
87 const int32x2_t va00x0 = vld1_dup_s32((const void*)a0);
88 const int32x2_t va01x0 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
89 const int32x2_t va00x1 = vld1_dup_s32((const void*)a0);
90 const int32x2_t va01x1 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
91 const int32x2_t va10x0 = vld1_dup_s32((const void*)a1);
92 const int32x2_t va11x0 = vld1_dup_s32((const void*)(a1 + 4)); a1 += 8;
93 const int32x2_t va10x1 = vld1_dup_s32((const void*)a1);
94 const int32x2_t va11x1 = vld1_dup_s32((const void*)(a1 + 4)); a1 += 8;
95 const int32x2_t va20x0 = vld1_dup_s32((const void*)a2);
96 const int32x2_t va21x0 = vld1_dup_s32((const void*)(a2 + 4)); a2 += 8;
97 const int32x2_t va20x1 = vld1_dup_s32((const void*)a2);
98 const int32x2_t va21x1 = vld1_dup_s32((const void*)(a2 + 4)); a2 += 8;
99
100 const int8x8_t vb01c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
101 const int8x8_t vb23c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
102 const int8x8_t vb45c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
103 const int8x8_t vb67c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
104 const int8x8_t vb01c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
105 const int8x8_t vb23c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
106 const int8x8_t vb45c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
107 const int8x8_t vb67c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
108
109 const int8x8_t va0c0x0 = vreinterpret_s8_s32(va00x0);
110 const int8x8_t va0c0x1 = vreinterpret_s8_s32(va00x1);
111 const int8x8_t va1c0x0 = vreinterpret_s8_s32(va10x0);
112 const int8x8_t va1c0x1 = vreinterpret_s8_s32(va10x1);
113 const int8x8_t va2c0x0 = vreinterpret_s8_s32(va20x0);
114 const int8x8_t va2c0x1 = vreinterpret_s8_s32(va20x1);
115
116 int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0c0x0);
117 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0);
118 int16x8_t vprod2x01c0 = vmull_s8(vb01c0x0, va2c0x0);
119 const int8x8_t vb01c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
120 vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0c0x1);
121 vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1);
122 vprod2x01c0 = vmlal_s8(vprod2x01c0, vb01c0x1, va2c0x1);
123 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
124 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
125 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
126 int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0c0x0);
127 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0);
128 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0);
129 const int8x8_t vb23c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
130 vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0c0x1);
131 vprod1x23c0 = vmlal_s8(vprod1x23c0, vb23c0x1, va1c0x1);
132 vprod2x23c0 = vmlal_s8(vprod2x23c0, vb23c0x1, va2c0x1);
133 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
134 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
135 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
136 int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0c0x0);
137 int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1c0x0);
138 int16x8_t vprod2x45c0 = vmull_s8(vb45c0x0, va2c0x0);
139 const int8x8_t vb45c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
140 vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0c0x1);
141 vprod1x45c0 = vmlal_s8(vprod1x45c0, vb45c0x1, va1c0x1);
142 vprod2x45c0 = vmlal_s8(vprod2x45c0, vb45c0x1, va2c0x1);
143 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
144 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
145 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
146 int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0c0x0);
147 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0);
148 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0);
149 const int8x8_t vb67c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
150 vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0c0x1);
151 vprod1x67c0 = vmlal_s8(vprod1x67c0, vb67c0x1, va1c0x1);
152 vprod2x67c0 = vmlal_s8(vprod2x67c0, vb67c0x1, va2c0x1);
153 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
154 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
155 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
156 const int8x8_t va0c1x0 = vreinterpret_s8_s32(va01x0);
157 const int8x8_t va0c1x1 = vreinterpret_s8_s32(va01x1);
158 const int8x8_t va1c1x0 = vreinterpret_s8_s32(va11x0);
159 const int8x8_t va1c1x1 = vreinterpret_s8_s32(va11x1);
160 const int8x8_t va2c1x0 = vreinterpret_s8_s32(va21x0);
161 const int8x8_t va2c1x1 = vreinterpret_s8_s32(va21x1);
162
163 int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0c1x0);
164 int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1c1x0);
165 int16x8_t vprod2x01c1 = vmull_s8(vb01c1x0, va2c1x0);
166 const int8x8_t vb01c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
167 vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0c1x1);
168 vprod1x01c1 = vmlal_s8(vprod1x01c1, vb01c1x1, va1c1x1);
169 vprod2x01c1 = vmlal_s8(vprod2x01c1, vb01c1x1, va2c1x1);
170 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
171 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
172 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
173 int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0c1x0);
174 int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1c1x0);
175 int16x8_t vprod2x23c1 = vmull_s8(vb23c1x0, va2c1x0);
176 const int8x8_t vb23c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
177 vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0c1x1);
178 vprod1x23c1 = vmlal_s8(vprod1x23c1, vb23c1x1, va1c1x1);
179 vprod2x23c1 = vmlal_s8(vprod2x23c1, vb23c1x1, va2c1x1);
180 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
181 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
182 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
183 int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0c1x0);
184 int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1c1x0);
185 int16x8_t vprod2x45c1 = vmull_s8(vb45c1x0, va2c1x0);
186 const int8x8_t vb45c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
187 vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0c1x1);
188 vprod1x45c1 = vmlal_s8(vprod1x45c1, vb45c1x1, va1c1x1);
189 vprod2x45c1 = vmlal_s8(vprod2x45c1, vb45c1x1, va2c1x1);
190 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
191 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
192 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
193 int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0c1x0);
194 int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1c1x0);
195 int16x8_t vprod2x67c1 = vmull_s8(vb67c1x0, va2c1x0);
196 const int8x8_t vb67c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
197 vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0c1x1);
198 vprod1x67c1 = vmlal_s8(vprod1x67c1, vb67c1x1, va1c1x1);
199 vprod2x67c1 = vmlal_s8(vprod2x67c1, vb67c1x1, va2c1x1);
200 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
201 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
202 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
203
204 k -= 16 * sizeof(int8_t);
205 }
206
207 if (k >= 8 * sizeof(int8_t)) {
208 const int32x2_t va00 = vld1_dup_s32((const void*)a0);
209 const int32x2_t va01 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
210 const int32x2_t va10 = vld1_dup_s32((const void*)a1);
211 const int32x2_t va11 = vld1_dup_s32((const void*)(a1 + 4)); a1 += 8;
212 const int32x2_t va20 = vld1_dup_s32((const void*)a2);
213 const int32x2_t va21 = vld1_dup_s32((const void*)(a2 + 4)); a2 += 8;
214
215 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
216 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
217 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
218 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
219 const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
220 const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
221 const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
222 const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
223
224 const int8x8_t va0c0 = vreinterpret_s8_s32(va00);
225 const int8x8_t va1c0 = vreinterpret_s8_s32(va10);
226 const int8x8_t va2c0 = vreinterpret_s8_s32(va20);
227
228 const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
229 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
230 const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
231 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
232 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
233 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
234 const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
235 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
236 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0);
237 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
238 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
239 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
240 const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
241 const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
242 const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
243 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
244 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
245 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
246 const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
247 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
248 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
249 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
250 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
251 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
252 const int8x8_t va0c1 = vreinterpret_s8_s32(va01);
253 const int8x8_t va1c1 = vreinterpret_s8_s32(va11);
254 const int8x8_t va2c1 = vreinterpret_s8_s32(va21);
255
256 const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
257 const int16x8_t vprod1x01c1 = vmull_s8(vb01c1, va1c1);
258 const int16x8_t vprod2x01c1 = vmull_s8(vb01c1, va2c1);
259 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
260 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
261 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
262 const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
263 const int16x8_t vprod1x23c1 = vmull_s8(vb23c1, va1c1);
264 const int16x8_t vprod2x23c1 = vmull_s8(vb23c1, va2c1);
265 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
266 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
267 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
268 const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
269 const int16x8_t vprod1x45c1 = vmull_s8(vb45c1, va1c1);
270 const int16x8_t vprod2x45c1 = vmull_s8(vb45c1, va2c1);
271 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
272 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
273 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
274 const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
275 const int16x8_t vprod1x67c1 = vmull_s8(vb67c1, va1c1);
276 const int16x8_t vprod2x67c1 = vmull_s8(vb67c1, va2c1);
277 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
278 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
279 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
280
281 k -= 8 * sizeof(int8_t);
282 }
283
284 if XNN_UNLIKELY(k != 0) {
285 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
286 const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
287 const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
288
289 const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
290 const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
291 const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
292 const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
293
294 const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
295 const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
296 vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
297 const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
298 vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
299 const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
300 vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
301 const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
302 vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
303 const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
304 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
305 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
306 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
307 vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
308 const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
309 vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
310 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
311 vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
312 const int8x8_t va2c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2), 0));
313 const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
314 vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
315 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0);
316 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
317 const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
318 vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
319 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
320 vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
321 }
322 p -= 3 * sizeof(void*);
323 } while (p != 0);
324
325 #if XNN_ARCH_ARM64
326 int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
327 int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
328 int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
329 int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
330 int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23);
331 int32x4_t vacc2x4567 = vpaddq_s32(vacc2x45, vacc2x67);
332 #else
333 const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
334 const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
335 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
336 const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
337 const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
338 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
339 const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
340 const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
341 int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
342 const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
343 const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
344 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
345 const int32x2_t vsum2x01 = vpadd_s32(vget_low_s32(vacc2x01), vget_high_s32(vacc2x01));
346 const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23));
347 int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23);
348 const int32x2_t vsum2x45 = vpadd_s32(vget_low_s32(vacc2x45), vget_high_s32(vacc2x45));
349 const int32x2_t vsum2x67 = vpadd_s32(vget_low_s32(vacc2x67), vget_high_s32(vacc2x67));
350 int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67);
351 #endif
352
353 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
354 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
355 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
356
357 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
358 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
359 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
360 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
361 vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
362 vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
363
364 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
365 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
366 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
367 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
368 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
369 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
370
371 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
372 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
373 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
374 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
375 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
376 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
377
378 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
379 #if XNN_ARCH_ARM64
380 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
381 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
382 int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
383
384 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
385 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
386 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
387
388 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
389 int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
390 #else
391 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
392 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
393 int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
394
395 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
396 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
397 vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
398
399 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
400 int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
401 #endif
402
403 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
404 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
405 vout2x01234567 = vmax_s8(vout2x01234567, vget_low_s8(voutput_min));
406
407 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
408 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
409 vout2x01234567 = vmin_s8(vout2x01234567, vget_low_s8(voutput_max));
410
411 if (nc >= 8) {
412 vst1_s8(c2 + 0, vout2x01234567);
413 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
414 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
415
416 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
417 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
418 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
419
420 a = (const int8_t**restrict) ((uintptr_t) a - ks);
421
422 nc -= 8;
423 } else {
424 if (nc & 4) {
425 vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
426 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
427 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
428 vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
429 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
430 }
431 if (nc & 2) {
432 vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
433 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
434 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
435 vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
436 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
437 }
438 if (nc & 1) {
439 vst1_lane_s8(c2, vout2x01234567, 0);
440 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
441 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
442 }
443
444 nc = 0;
445 }
446 } while (nc != 0);
447 }
448