1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/c2-neon-mull-shuffle.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <arm_neon.h>
13
14 #include <xnnpack/gemm.h>
15 #include <xnnpack/math.h>
16
17
xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2s4__neon_mlal(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2s4__neon_mlal(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const int8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(mr != 0);
31 assert(mr <= 2);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(int8_t) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38
39 const int8_t* a0 = a;
40 int8_t* c0 = c;
41 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
42 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
43 if XNN_UNPREDICTABLE(mr != 2) {
44 a1 = a0;
45 c1 = c0;
46 }
47
48 kc = round_up_po2(kc, 8 * sizeof(int8_t));
49 do {
50 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
51 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
52 int32x4_t vacc0x89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
53 int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
54 int32x4_t vacc1x0123 = vacc0x0123;
55 int32x4_t vacc1x4567 = vacc0x4567;
56 int32x4_t vacc1x89AB = vacc0x89AB;
57 int32x4_t vacc1xCDEF = vacc0xCDEF;
58
59 size_t k = kc;
60 while (k >= 16 * sizeof(int8_t)) {
61 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
62 int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
63 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
64 int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
65
66 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
67 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
68 const int8x8_t vb89ABc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
69 const int8x8_t vbCDEFc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
70 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
71 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
72 const int8x8_t vb89ABc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
73 const int8x8_t vbCDEFc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
74 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
75 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
76 const int8x8_t vb89ABc2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
77 const int8x8_t vbCDEFc2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
78 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
79 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
80 const int8x8_t vb89ABc3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
81 const int8x8_t vbCDEFc3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
82
83 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
84 int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1x0);
85 const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
86 vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0x1);
87 vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, va1x1);
88 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
89 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
90 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
91 int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1x0);
92 const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
93 vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0x1);
94 vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, va1x1);
95 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
96 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
97 int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, va0x0);
98 int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1x0);
99 const int8x8_t vb89ABc0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
100 vprod0x89ABc0 = vmlal_s8(vprod0x89ABc0, vb89ABc0x1, va0x1);
101 vprod1x89ABc0 = vmlal_s8(vprod1x89ABc0, vb89ABc0x1, va1x1);
102 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
103 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
104 int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0x0, va0x0);
105 int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0x0, va1x0);
106 const int8x8_t vbCDEFc0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
107 vprod0xCDEFc0 = vmlal_s8(vprod0xCDEFc0, vbCDEFc0x1, va0x1);
108 vprod1xCDEFc0 = vmlal_s8(vprod1xCDEFc0, vbCDEFc0x1, va1x1);
109 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
110 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
111 va0x0 = vext_s8(va0x0, va0x0, 2);
112 va0x1 = vext_s8(va0x1, va0x1, 2);
113 va1x0 = vext_s8(va1x0, va1x0, 2);
114 va1x1 = vext_s8(va1x1, va1x1, 2);
115 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
116 int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1x0);
117 const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
118 vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0x1);
119 vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, va1x1);
120 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
121 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
122 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
123 int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1x0);
124 const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
125 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0x1);
126 vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, va1x1);
127 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
128 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
129 int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0x0);
130 int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1x0, va1x0);
131 const int8x8_t vb89ABc1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
132 vprod0x89ABc1 = vmlal_s8(vprod0x89ABc1, vb89ABc1x1, va0x1);
133 vprod1x89ABc1 = vmlal_s8(vprod1x89ABc1, vb89ABc1x1, va1x1);
134 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
135 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
136 int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0x0);
137 int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1x0);
138 const int8x8_t vbCDEFc1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
139 vprod0xCDEFc1 = vmlal_s8(vprod0xCDEFc1, vbCDEFc1x1, va0x1);
140 vprod1xCDEFc1 = vmlal_s8(vprod1xCDEFc1, vbCDEFc1x1, va1x1);
141 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
142 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
143 va0x0 = vext_s8(va0x0, va0x0, 2);
144 va0x1 = vext_s8(va0x1, va0x1, 2);
145 va1x0 = vext_s8(va1x0, va1x0, 2);
146 va1x1 = vext_s8(va1x1, va1x1, 2);
147 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
148 int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1x0);
149 const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
150 vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0x1);
151 vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, va1x1);
152 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
153 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
154 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
155 int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1x0);
156 const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
157 vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0x1);
158 vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, va1x1);
159 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
160 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
161 int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2x0, va0x0);
162 int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2x0, va1x0);
163 const int8x8_t vb89ABc2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
164 vprod0x89ABc2 = vmlal_s8(vprod0x89ABc2, vb89ABc2x1, va0x1);
165 vprod1x89ABc2 = vmlal_s8(vprod1x89ABc2, vb89ABc2x1, va1x1);
166 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
167 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
168 int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0x0);
169 int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2x0, va1x0);
170 const int8x8_t vbCDEFc2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
171 vprod0xCDEFc2 = vmlal_s8(vprod0xCDEFc2, vbCDEFc2x1, va0x1);
172 vprod1xCDEFc2 = vmlal_s8(vprod1xCDEFc2, vbCDEFc2x1, va1x1);
173 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
174 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
175 va0x0 = vext_s8(va0x0, va0x0, 2);
176 va0x1 = vext_s8(va0x1, va0x1, 2);
177 va1x0 = vext_s8(va1x0, va1x0, 2);
178 va1x1 = vext_s8(va1x1, va1x1, 2);
179 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
180 int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1x0);
181 const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
182 vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0x1);
183 vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1x1);
184 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
185 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
186 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
187 int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1x0);
188 const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
189 vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0x1);
190 vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1x1);
191 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
192 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
193 int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3x0, va0x0);
194 int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3x0, va1x0);
195 const int8x8_t vb89ABc3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
196 vprod0x89ABc3 = vmlal_s8(vprod0x89ABc3, vb89ABc3x1, va0x1);
197 vprod1x89ABc3 = vmlal_s8(vprod1x89ABc3, vb89ABc3x1, va1x1);
198 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
199 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
200 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0x0);
201 int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3x0, va1x0);
202 const int8x8_t vbCDEFc3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
203 vprod0xCDEFc3 = vmlal_s8(vprod0xCDEFc3, vbCDEFc3x1, va0x1);
204 vprod1xCDEFc3 = vmlal_s8(vprod1xCDEFc3, vbCDEFc3x1, va1x1);
205 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
206 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
207
208 k -= 16 * sizeof(int8_t);
209 }
210 if (k != 0) {
211 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
212 int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
213
214 const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
215 const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
216 const int8x8_t vb89ABc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
217 const int8x8_t vbCDEFc0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
218 const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
219 const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
220 const int8x8_t vb89ABc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
221 const int8x8_t vbCDEFc1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
222 const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
223 const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
224 const int8x8_t vb89ABc2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
225 const int8x8_t vbCDEFc2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
226 const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
227 const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
228 const int8x8_t vb89ABc3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
229 const int8x8_t vbCDEFc3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
230
231 int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
232 int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1x0);
233 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
234 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
235 int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
236 int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1x0);
237 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
238 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
239 int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, va0x0);
240 int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1x0);
241 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
242 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
243 int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0x0, va0x0);
244 int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0x0, va1x0);
245 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
246 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
247 va0x0 = vext_s8(va0x0, va0x0, 2);
248 va1x0 = vext_s8(va1x0, va1x0, 2);
249 int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
250 int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1x0);
251 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
252 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
253 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
254 int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1x0);
255 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
256 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
257 int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0x0);
258 int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1x0, va1x0);
259 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
260 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
261 int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0x0);
262 int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1x0);
263 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
264 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
265 va0x0 = vext_s8(va0x0, va0x0, 2);
266 va1x0 = vext_s8(va1x0, va1x0, 2);
267 int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
268 int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1x0);
269 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
270 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
271 int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
272 int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1x0);
273 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
274 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
275 int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2x0, va0x0);
276 int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2x0, va1x0);
277 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
278 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
279 int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0x0);
280 int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2x0, va1x0);
281 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
282 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
283 va0x0 = vext_s8(va0x0, va0x0, 2);
284 va1x0 = vext_s8(va1x0, va1x0, 2);
285 int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
286 int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1x0);
287 vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
288 vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
289 int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
290 int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1x0);
291 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
292 vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
293 int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3x0, va0x0);
294 int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3x0, va1x0);
295 vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
296 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
297 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0x0);
298 int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3x0, va1x0);
299 vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
300 vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
301
302 }
303
304 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
305 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
306 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
307
308 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
309 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
310 vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
311 vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
312 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
313 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
314 vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
315 vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
316
317 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
318 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
319 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
320 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
321 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
322 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
323 vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
324 vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
325
326 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
327 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
328 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
329 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
330 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
331 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
332 vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
333 vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
334
335 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
336 #if XNN_ARCH_ARM64
337 int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
338 int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
339 int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
340 int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
341
342 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
343 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
344 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
345 vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
346
347 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
348 int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
349 #else
350 int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
351 int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
352 int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
353 int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
354
355 vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
356 vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
357 vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
358 vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
359
360 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
361 int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
362 #endif
363
364 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
365 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
366 vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
367
368 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
369 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
370 vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
371
372 if (nc >= 16) {
373 vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
374 vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
375
376 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
377 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
378
379 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
380 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
381
382 nc -= 16;
383 } else {
384 // Final case where not all of the 16 columns fit in the destination.
385 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
386 if (nc & 8) {
387 vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
388 vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
389 vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
390 }
391 if (nc & 4) {
392 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
393 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
394 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
395 }
396 if (nc & 2) {
397 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
398 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
399 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
400 }
401 if (nc & 1) {
402 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
403 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
404 }
405
406 nc = 0;
407 }
408 } while (nc != 0);
409 }
410