1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c2-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
xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__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)])17 void xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 size_t ks,
22 const int8_t** restrict a,
23 const void* restrict w,
24 int8_t* restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 size_t a_offset,
28 const int8_t* zero,
29 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 1);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(ks != 0);
36 assert(ks % (1 * sizeof(void*)) == 0);
37 assert(a_offset % sizeof(int8_t) == 0);
38 assert(a != NULL);
39 assert(w != NULL);
40 assert(c != NULL);
41
42 kc = round_up_po2(kc, 2 * sizeof(int8_t));
43 int8_t* c0 = c;
44
45 do {
46 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
47 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
48 int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
49 int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
50
51 size_t p = ks;
52 do {
53 const int8_t* restrict a0 = a[0];
54 if XNN_UNPREDICTABLE(a0 != zero) {
55 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
56 }
57 a += 1;
58
59 size_t k = kc;
60
61 while (k >= 16 * sizeof(int8_t)) {
62 const int16x4_t va00x0 = vld1_dup_s16((const void*)a0);
63 const int16x4_t va01x0 = vld1_dup_s16((const void*)(a0 + 2));
64 const int16x4_t va02x0 = vld1_dup_s16((const void*)(a0 + 4));
65 const int16x4_t va03x0 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
66 const int16x4_t va00x1 = vld1_dup_s16((const void*)a0);
67 const int16x4_t va01x1 = vld1_dup_s16((const void*)(a0 + 2));
68 const int16x4_t va02x1 = vld1_dup_s16((const void*)(a0 + 4));
69 const int16x4_t va03x1 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
70
71 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
72 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
73 const int8x8_t vb89ABc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
74 const int8x8_t vbCDEFc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
75 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
76 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
77 const int8x8_t vb89ABc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
78 const int8x8_t vbCDEFc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
79 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
80 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
81 const int8x8_t vb89ABc2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
82 const int8x8_t vbCDEFc2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
84 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
85 const int8x8_t vb89ABc3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
86 const int8x8_t vbCDEFc3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
87
88 const int8x8_t va0c0x0 = vreinterpret_s8_s16(va00x0);
89 const int8x8_t va0c0x1 = vreinterpret_s8_s16(va00x1);
90
91 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0c0x0);
92 const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
93 vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0c0x1);
94 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
95 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0c0x0);
96 const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
97 vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0c0x1);
98 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
99 int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, va0c0x0);
100 const int8x8_t vb89ABc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
101 vprod0x89ABc0 = vmlal_s8(vprod0x89ABc0, vb89ABc0x1, va0c0x1);
102 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
103 int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0x0, va0c0x0);
104 const int8x8_t vbCDEFc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
105 vprod0xCDEFc0 = vmlal_s8(vprod0xCDEFc0, vbCDEFc0x1, va0c0x1);
106 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
107 const int8x8_t va0c1x0 = vreinterpret_s8_s16(va01x0);
108 const int8x8_t va0c1x1 = vreinterpret_s8_s16(va01x1);
109
110 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0c1x0);
111 const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
112 vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0c1x1);
113 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
114 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0);
115 const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
116 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1);
117 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
118 int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0c1x0);
119 const int8x8_t vb89ABc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
120 vprod0x89ABc1 = vmlal_s8(vprod0x89ABc1, vb89ABc1x1, va0c1x1);
121 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
122 int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0c1x0);
123 const int8x8_t vbCDEFc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
124 vprod0xCDEFc1 = vmlal_s8(vprod0xCDEFc1, vbCDEFc1x1, va0c1x1);
125 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
126 const int8x8_t va0c2x0 = vreinterpret_s8_s16(va02x0);
127 const int8x8_t va0c2x1 = vreinterpret_s8_s16(va02x1);
128
129 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0c2x0);
130 const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
131 vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0c2x1);
132 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
133 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0c2x0);
134 const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
135 vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0c2x1);
136 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
137 int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2x0, va0c2x0);
138 const int8x8_t vb89ABc2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
139 vprod0x89ABc2 = vmlal_s8(vprod0x89ABc2, vb89ABc2x1, va0c2x1);
140 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
141 int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0c2x0);
142 const int8x8_t vbCDEFc2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
143 vprod0xCDEFc2 = vmlal_s8(vprod0xCDEFc2, vbCDEFc2x1, va0c2x1);
144 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
145 const int8x8_t va0c3x0 = vreinterpret_s8_s16(va03x0);
146 const int8x8_t va0c3x1 = vreinterpret_s8_s16(va03x1);
147
148 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0c3x0);
149 const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
150 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1);
151 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
152 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0c3x0);
153 const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
154 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1);
155 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
156 int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3x0, va0c3x0);
157 const int8x8_t vb89ABc3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
158 vprod0x89ABc3 = vmlal_s8(vprod0x89ABc3, vb89ABc3x1, va0c3x1);
159 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
160 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0c3x0);
161 const int8x8_t vbCDEFc3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
162 vprod0xCDEFc3 = vmlal_s8(vprod0xCDEFc3, vbCDEFc3x1, va0c3x1);
163 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
164
165 k -= 16 * sizeof(int8_t);
166 }
167
168 if (k >= 8 * sizeof(int8_t)) {
169 const int16x4_t va00 = vld1_dup_s16((const void*)a0);
170 const int16x4_t va01 = vld1_dup_s16((const void*)(a0 + 2));
171 const int16x4_t va02 = vld1_dup_s16((const void*)(a0 + 4));
172 const int16x4_t va03 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
173
174 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
175 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
176 const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
177 const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
178 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
179 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
180 const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
181 const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
182 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
183 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
184 const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
185 const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
186 const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
187 const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
188 const int8x8_t vb89ABc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
189 const int8x8_t vbCDEFc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
190
191 const int8x8_t va0c0 = vreinterpret_s8_s16(va00);
192
193 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
194 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
195 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
196 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
197 const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
198 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
199 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
200 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
201 const int8x8_t va0c1 = vreinterpret_s8_s16(va01);
202
203 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
204 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
205 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
206 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
207 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
208 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
209 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
210 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
211 const int8x8_t va0c2 = vreinterpret_s8_s16(va02);
212
213 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
214 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
215 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
216 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
217 const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
218 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
219 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
220 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
221 const int8x8_t va0c3 = vreinterpret_s8_s16(va03);
222
223 const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
224 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
225 const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
226 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
227 const int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3, va0c3);
228 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
229 const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3);
230 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
231
232 k -= 8 * sizeof(int8_t);
233 }
234
235 if XNN_UNLIKELY(k != 0) {
236 const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
237
238 const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
239 const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
240 const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
241 const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
242
243 const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
244 const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
245 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
246 const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
247 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
248 const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
249 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
250 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
251 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
252
253 if (k > 2 * sizeof(int8_t)) {
254 const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
255 const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
256 const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
257 const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
258
259 const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
260 const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
261 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
262 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
263 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
264 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
265 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
266 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
267 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
268
269 if (k > 4 * sizeof(int8_t)) {
270 const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
271 const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
272 const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
273 const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
274
275 const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
276 const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
277 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
278 const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
279 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
280 const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
281 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
282 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
283 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
284 }
285 }
286 }
287 p -= 1 * sizeof(void*);
288 } while (p != 0);
289
290 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
291 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
292 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
293
294 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
295 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
296 vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
297 vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
298
299 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
300 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
301 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
302 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
303
304 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
305 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
306 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
307 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
308
309 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
310 #if XNN_ARCH_ARM64
311 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
312 int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
313
314 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
315 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
316
317 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
318 #else
319 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
320 int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
321
322 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
323 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
324
325 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
326 #endif
327
328 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
329 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
330
331 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
332 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
333
334 if (nc >= 16) {
335 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
336
337 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
338
339 a = (const int8_t**restrict) ((uintptr_t) a - ks);
340
341 nc -= 16;
342 } else {
343 int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF);
344 if (nc & 8) {
345 vst1_s8(c0, vout0x01234567); c0 += 8;
346 vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF);
347 }
348 if (nc & 4) {
349 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
350 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
351 }
352 if (nc & 2) {
353 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
354 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
355 }
356 if (nc & 1) {
357 vst1_lane_s8(c0, vout0x01234567, 0);
358 }
359
360 nc = 0;
361 }
362 } while (nc != 0);
363 }
364