Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-5f5m5l16c8s8r-minmax-rndnu-neon-mul16.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/multipass-neon-mul16.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2023 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
11 #include <assert.h>
12
13 #include <arm_neon.h>
14
15 #include "xnnpack/dwconv.h"
16 #include "xnnpack/math.h"
17
18
19 void xnn_qs8_dwconv_minmax_rndnu_ukernel_5f5m5l16c8s8r__neon_mul16(
20 size_t channels,
21 size_t output_width,
22 const int8_t** input,
23 const void* weights,
24 int8_t* output,
25 intptr_t input_stride,
26 size_t output_increment,
27 size_t input_offset,
28 const int8_t* zero,
29 size_t kernel_size,
30 int32_t* buffer,
31 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(channels != 0);
34 assert(output_width != 0);
35 assert(kernel_size > 5);
36
37 const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
38 const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
39 const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
40 const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
41 const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
42 const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
43
44 do {
45 const void* w = weights;
46
47 // First pass to process 5 inputs.
48 {
49 int32_t* b = buffer;
50 const int8_t* i0 = input[0];
51 assert(i0 != NULL);
52 if XNN_UNPREDICTABLE(i0 != zero) {
53 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
54 }
55 const int8_t* i1 = input[1];
56 assert(i1 != NULL);
57 if XNN_UNPREDICTABLE(i1 != zero) {
58 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
59 }
60 const int8_t* i2 = input[2];
61 assert(i2 != NULL);
62 if XNN_UNPREDICTABLE(i2 != zero) {
63 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
64 }
65 const int8_t* i3 = input[3];
66 assert(i3 != NULL);
67 if XNN_UNPREDICTABLE(i3 != zero) {
68 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
69 }
70 const int8_t* i4 = input[4];
71 assert(i4 != NULL);
72 if XNN_UNPREDICTABLE(i4 != zero) {
73 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
74 }
75 input += 5;
76
77 size_t c = round_up_po2(channels, 8);
78
79 for (; c >= 16; c -= 16) {
80 int32x4_t vacc0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
81 int32x4_t vacc4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
82 int32x4_t vacc89AB = vld1q_s32(w); w = (const int32_t*) w + 4;
83 int32x4_t vaccCDEF = vld1q_s32(w); w = (const int32_t*) w + 4;
84
85 const int16x8_t vi0x01234567 = vmovl_s8(vld1_s8(i0)); i0 += 8;
86 const int16x8_t vk0x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
87 const int16x8_t vi0x89ABCDEF = vmovl_s8(vld1_s8(i0)); i0 += 8;
88 const int16x8_t vk0x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
89
90 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
91 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
92 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi0x89ABCDEF), vget_low_s16(vk0x89ABCDEF));
93 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi0x89ABCDEF), vget_high_s16(vk0x89ABCDEF));
94
95 const int16x8_t vi1x01234567 = vmovl_s8(vld1_s8(i1)); i1 += 8;
96 const int16x8_t vk1x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
97 const int16x8_t vi1x89ABCDEF = vmovl_s8(vld1_s8(i1)); i1 += 8;
98 const int16x8_t vk1x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
99
100 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
101 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
102 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi1x89ABCDEF), vget_low_s16(vk1x89ABCDEF));
103 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi1x89ABCDEF), vget_high_s16(vk1x89ABCDEF));
104
105 const int16x8_t vi2x01234567 = vmovl_s8(vld1_s8(i2)); i2 += 8;
106 const int16x8_t vk2x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
107 const int16x8_t vi2x89ABCDEF = vmovl_s8(vld1_s8(i2)); i2 += 8;
108 const int16x8_t vk2x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
109
110 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
111 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
112 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi2x89ABCDEF), vget_low_s16(vk2x89ABCDEF));
113 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi2x89ABCDEF), vget_high_s16(vk2x89ABCDEF));
114
115 const int16x8_t vi3x01234567 = vmovl_s8(vld1_s8(i3)); i3 += 8;
116 const int16x8_t vk3x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
117 const int16x8_t vi3x89ABCDEF = vmovl_s8(vld1_s8(i3)); i3 += 8;
118 const int16x8_t vk3x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
119
120 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
121 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
122 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi3x89ABCDEF), vget_low_s16(vk3x89ABCDEF));
123 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi3x89ABCDEF), vget_high_s16(vk3x89ABCDEF));
124
125 const int16x8_t vi4x01234567 = vmovl_s8(vld1_s8(i4)); i4 += 8;
126 const int16x8_t vk4x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
127 const int16x8_t vi4x89ABCDEF = vmovl_s8(vld1_s8(i4)); i4 += 8;
128 const int16x8_t vk4x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
129
130 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
131 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
132 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi4x89ABCDEF), vget_low_s16(vk4x89ABCDEF));
133 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi4x89ABCDEF), vget_high_s16(vk4x89ABCDEF));
134
135 vst1q_s32(b, vacc0123); b += 4;
136 vst1q_s32(b, vacc4567); b += 4;
137 vst1q_s32(b, vacc89AB); b += 4;
138 vst1q_s32(b, vaccCDEF); b += 4;
139 }
140
141 if XNN_UNLIKELY(c != 0) {
142 do {
143 int32x4_t vacc0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
144 int32x4_t vacc4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
145
146 const int16x8_t vi0x01234567 = vmovl_s8(vld1_s8(i0)); i0 += 8;
147 const int16x8_t vk0x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
148
149 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
150 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
151
152 const int16x8_t vi1x01234567 = vmovl_s8(vld1_s8(i1)); i1 += 8;
153 const int16x8_t vk1x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
154
155 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
156 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
157
158 const int16x8_t vi2x01234567 = vmovl_s8(vld1_s8(i2)); i2 += 8;
159 const int16x8_t vk2x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
160
161 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
162 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
163
164 const int16x8_t vi3x01234567 = vmovl_s8(vld1_s8(i3)); i3 += 8;
165 const int16x8_t vk3x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
166
167 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
168 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
169
170 const int16x8_t vi4x01234567 = vmovl_s8(vld1_s8(i4)); i4 += 8;
171 const int16x8_t vk4x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
172
173 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
174 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
175
176 vst1q_s32(b, vacc0123); b += 4;
177 vst1q_s32(b, vacc4567); b += 4;
178 c -= 8;
179 } while (c != 0);
180 }
181 }
182
183 // Middle pass to process 5 inputs in each iteration.
184 for (size_t ks = kernel_size - 5; ks > 5; ks -= 5) {
185 int32_t* b = buffer;
186
187 const int8_t* i0 = input[0];
188 assert(i0 != NULL);
189 if XNN_UNPREDICTABLE(i0 != zero) {
190 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
191 }
192 const int8_t* i1 = input[1];
193 assert(i1 != NULL);
194 if XNN_UNPREDICTABLE(i1 != zero) {
195 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
196 }
197 const int8_t* i2 = input[2];
198 assert(i2 != NULL);
199 if XNN_UNPREDICTABLE(i2 != zero) {
200 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
201 }
202 const int8_t* i3 = input[3];
203 assert(i3 != NULL);
204 if XNN_UNPREDICTABLE(i3 != zero) {
205 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
206 }
207 const int8_t* i4 = input[4];
208 assert(i4 != NULL);
209 if XNN_UNPREDICTABLE(i4 != zero) {
210 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
211 }
212 input += 5;
213
214 size_t c = round_up_po2(channels, 8);
215
216 for (; c >= 16; c -= 16) {
217 int32x4_t vacc0123 = vld1q_s32(b);
218 int32x4_t vacc4567 = vld1q_s32(b + 4);
219 int32x4_t vacc89AB = vld1q_s32(b + 8);
220 int32x4_t vaccCDEF = vld1q_s32(b + 12);
221
222 const int16x8_t vi0x01234567 = vmovl_s8(vld1_s8(i0)); i0 += 8;
223 const int16x8_t vk0x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
224 const int16x8_t vi0x89ABCDEF = vmovl_s8(vld1_s8(i0)); i0 += 8;
225 const int16x8_t vk0x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
226
227 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
228 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
229 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi0x89ABCDEF), vget_low_s16(vk0x89ABCDEF));
230 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi0x89ABCDEF), vget_high_s16(vk0x89ABCDEF));
231
232 const int16x8_t vi1x01234567 = vmovl_s8(vld1_s8(i1)); i1 += 8;
233 const int16x8_t vk1x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
234 const int16x8_t vi1x89ABCDEF = vmovl_s8(vld1_s8(i1)); i1 += 8;
235 const int16x8_t vk1x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
236
237 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
238 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
239 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi1x89ABCDEF), vget_low_s16(vk1x89ABCDEF));
240 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi1x89ABCDEF), vget_high_s16(vk1x89ABCDEF));
241
242 const int16x8_t vi2x01234567 = vmovl_s8(vld1_s8(i2)); i2 += 8;
243 const int16x8_t vk2x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
244 const int16x8_t vi2x89ABCDEF = vmovl_s8(vld1_s8(i2)); i2 += 8;
245 const int16x8_t vk2x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
246
247 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
248 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
249 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi2x89ABCDEF), vget_low_s16(vk2x89ABCDEF));
250 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi2x89ABCDEF), vget_high_s16(vk2x89ABCDEF));
251
252 const int16x8_t vi3x01234567 = vmovl_s8(vld1_s8(i3)); i3 += 8;
253 const int16x8_t vk3x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
254 const int16x8_t vi3x89ABCDEF = vmovl_s8(vld1_s8(i3)); i3 += 8;
255 const int16x8_t vk3x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
256
257 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
258 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
259 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi3x89ABCDEF), vget_low_s16(vk3x89ABCDEF));
260 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi3x89ABCDEF), vget_high_s16(vk3x89ABCDEF));
261
262 const int16x8_t vi4x01234567 = vmovl_s8(vld1_s8(i4)); i4 += 8;
263 const int16x8_t vk4x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
264 const int16x8_t vi4x89ABCDEF = vmovl_s8(vld1_s8(i4)); i4 += 8;
265 const int16x8_t vk4x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
266
267 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
268 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
269 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi4x89ABCDEF), vget_low_s16(vk4x89ABCDEF));
270 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi4x89ABCDEF), vget_high_s16(vk4x89ABCDEF));
271
272 vst1q_s32(b, vacc0123); b += 4;
273 vst1q_s32(b, vacc4567); b += 4;
274 vst1q_s32(b, vacc89AB); b += 4;
275 vst1q_s32(b, vaccCDEF); b += 4;
276 }
277
278 if XNN_UNLIKELY(c != 0) {
279 do {
280 int32x4_t vacc0123 = vld1q_s32(b);
281 int32x4_t vacc4567 = vld1q_s32(b + 4);
282
283 const int16x8_t vi0x01234567 = vmovl_s8(vld1_s8(i0)); i0 += 8;
284 const int16x8_t vk0x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
285
286 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
287 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
288
289 const int16x8_t vi1x01234567 = vmovl_s8(vld1_s8(i1)); i1 += 8;
290 const int16x8_t vk1x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
291
292 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
293 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
294
295 const int16x8_t vi2x01234567 = vmovl_s8(vld1_s8(i2)); i2 += 8;
296 const int16x8_t vk2x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
297
298 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
299 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
300
301 const int16x8_t vi3x01234567 = vmovl_s8(vld1_s8(i3)); i3 += 8;
302 const int16x8_t vk3x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
303
304 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
305 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
306
307 const int16x8_t vi4x01234567 = vmovl_s8(vld1_s8(i4)); i4 += 8;
308 const int16x8_t vk4x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
309
310 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
311 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
312
313 vst1q_s32(b, vacc0123); b += 4;
314 vst1q_s32(b, vacc4567); b += 4;
315 c -= 8;
316 } while (c != 0);
317 }
318 }
319
320 // Last pass to process up to 5 inputs.
321 {
322 const int32_t* b = buffer;
323
324 const int8_t* i0 = input[0];
325 assert(i0 != NULL);
326 if XNN_UNPREDICTABLE(i0 != zero) {
327 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
328 }
329 const int8_t* i1 = input[1];
330 assert(i1 != NULL);
331 if XNN_UNPREDICTABLE(i1 != zero) {
332 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
333 }
334 const int8_t* i2 = input[2];
335 assert(i2 != NULL);
336 if XNN_UNPREDICTABLE(i2 != zero) {
337 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
338 }
339 const int8_t* i3 = input[3];
340 assert(i3 != NULL);
341 if XNN_UNPREDICTABLE(i3 != zero) {
342 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
343 }
344 const int8_t* i4 = input[4];
345 assert(i4 != NULL);
346 if XNN_UNPREDICTABLE(i4 != zero) {
347 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
348 }
349
350 size_t c = channels;
351 for (; c >= 16; c -= 16) {
352 int32x4_t vacc0123 = vld1q_s32(b); b += 4;
353 int32x4_t vacc4567 = vld1q_s32(b); b += 4;
354 int32x4_t vacc89AB = vld1q_s32(b); b += 4;
355 int32x4_t vaccCDEF = vld1q_s32(b); b += 4;
356
357 const int16x8_t vi0x01234567 = vmovl_s8(vld1_s8(i0)); i0 += 8;
358 const int16x8_t vk0x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
359 const int16x8_t vi0x89ABCDEF = vmovl_s8(vld1_s8(i0)); i0 += 8;
360 const int16x8_t vk0x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
361
362 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
363 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
364 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi0x89ABCDEF), vget_low_s16(vk0x89ABCDEF));
365 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi0x89ABCDEF), vget_high_s16(vk0x89ABCDEF));
366
367 const int16x8_t vi1x01234567 = vmovl_s8(vld1_s8(i1)); i1 += 8;
368 const int16x8_t vk1x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
369 const int16x8_t vi1x89ABCDEF = vmovl_s8(vld1_s8(i1)); i1 += 8;
370 const int16x8_t vk1x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
371
372 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
373 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
374 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi1x89ABCDEF), vget_low_s16(vk1x89ABCDEF));
375 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi1x89ABCDEF), vget_high_s16(vk1x89ABCDEF));
376
377 const int16x8_t vi2x01234567 = vmovl_s8(vld1_s8(i2)); i2 += 8;
378 const int16x8_t vk2x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
379 const int16x8_t vi2x89ABCDEF = vmovl_s8(vld1_s8(i2)); i2 += 8;
380 const int16x8_t vk2x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
381
382 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
383 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
384 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi2x89ABCDEF), vget_low_s16(vk2x89ABCDEF));
385 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi2x89ABCDEF), vget_high_s16(vk2x89ABCDEF));
386
387 const int16x8_t vi3x01234567 = vmovl_s8(vld1_s8(i3)); i3 += 8;
388 const int16x8_t vk3x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
389 const int16x8_t vi3x89ABCDEF = vmovl_s8(vld1_s8(i3)); i3 += 8;
390 const int16x8_t vk3x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
391
392 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
393 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
394 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi3x89ABCDEF), vget_low_s16(vk3x89ABCDEF));
395 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi3x89ABCDEF), vget_high_s16(vk3x89ABCDEF));
396
397 const int16x8_t vi4x01234567 = vmovl_s8(vld1_s8(i4)); i4 += 8;
398 const int16x8_t vk4x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
399 const int16x8_t vi4x89ABCDEF = vmovl_s8(vld1_s8(i4)); i4 += 8;
400 const int16x8_t vk4x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
401
402 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
403 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
404 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi4x89ABCDEF), vget_low_s16(vk4x89ABCDEF));
405 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi4x89ABCDEF), vget_high_s16(vk4x89ABCDEF));
406
407 vacc0123 = vqshlq_s32(vacc0123, vright_pre_shift);
408 vacc4567 = vqshlq_s32(vacc4567, vright_pre_shift);
409 vacc89AB = vqshlq_s32(vacc89AB, vright_pre_shift);
410 vaccCDEF = vqshlq_s32(vaccCDEF, vright_pre_shift);
411
412 vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
413 vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
414 vacc89AB = vqdmulhq_s32(vacc89AB, vmultiplier);
415 vaccCDEF = vqdmulhq_s32(vaccCDEF, vmultiplier);
416
417 vacc0123 = vrshlq_s32(vacc0123, vright_post_shift);
418 vacc4567 = vrshlq_s32(vacc4567, vright_post_shift);
419 vacc89AB = vrshlq_s32(vacc89AB, vright_post_shift);
420 vaccCDEF = vrshlq_s32(vaccCDEF, vright_post_shift);
421
422 #if XNN_ARCH_ARM64
423 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
424 int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
425
426 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
427 vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
428
429 int8x16_t vout0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc01234567), vacc89ABCDEF);
430 #else // !XNN_ARCH_ARM64
431 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
432 int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
433
434 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
435 vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
436
437 int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF));
438 #endif // !XNN_ARCH_ARM64
439
440 vout0123456789ABCDEF = vmaxq_s8(vout0123456789ABCDEF, voutput_min);
441
442 vout0123456789ABCDEF = vminq_s8(vout0123456789ABCDEF, voutput_max);
443
444 vst1q_s8(output, vout0123456789ABCDEF); output += 16;
445 }
446
447 if XNN_UNLIKELY(c != 0) {
448 do {
449 int32x4_t vacc0123 = vld1q_s32(b); b += 4;
450 int32x4_t vacc4567 = vld1q_s32(b); b += 4;
451
452 const int16x8_t vi0x01234567 = vmovl_s8(vld1_s8(i0)); i0 += 8;
453 const int16x8_t vk0x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
454
455 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
456 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
457
458 const int16x8_t vi1x01234567 = vmovl_s8(vld1_s8(i1)); i1 += 8;
459 const int16x8_t vk1x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
460
461 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
462 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
463
464 const int16x8_t vi2x01234567 = vmovl_s8(vld1_s8(i2)); i2 += 8;
465 const int16x8_t vk2x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
466
467 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
468 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
469
470 const int16x8_t vi3x01234567 = vmovl_s8(vld1_s8(i3)); i3 += 8;
471 const int16x8_t vk3x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
472
473 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
474 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
475
476 const int16x8_t vi4x01234567 = vmovl_s8(vld1_s8(i4)); i4 += 8;
477 const int16x8_t vk4x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
478
479 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
480 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
481
482 vacc0123 = vqshlq_s32(vacc0123, vright_pre_shift);
483 vacc4567 = vqshlq_s32(vacc4567, vright_pre_shift);
484
485 vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
486 vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
487
488 vacc0123 = vrshlq_s32(vacc0123, vright_post_shift);
489 vacc4567 = vrshlq_s32(vacc4567, vright_post_shift);
490
491 #if XNN_ARCH_ARM64
492 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
493 #else
494 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
495 #endif
496 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
497
498 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
499 vout01234567 = vmax_s8(vout01234567, vget_low_s8(voutput_min));
500 vout01234567 = vmin_s8(vout01234567, vget_low_s8(voutput_max));
501
502 if XNN_LIKELY(c >= 8) {
503 vst1_s8(output, vout01234567); output += 8;
504 c -= 8;
505 } else {
506 if (c & 4) {
507 vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
508 vout01234567 = vext_s8(vout01234567, vout01234567, 4);
509 }
510 if (c & 2) {
511 vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
512 vout01234567 = vext_s8(vout01234567, vout01234567, 2);
513 }
514 if (c & 1) {
515 vst1_lane_s8(output, vout01234567, 0); output += 1;
516 }
517 c = 0;
518 }
519 } while (c != 0);
520 }
521 }
522
523 input = (const int8_t**) ((uintptr_t) input + input_stride);
524 output = (int8_t*) ((uintptr_t) output + output_increment);
525 } while (--output_width != 0);
526 }