1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-neon-mul16.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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/dwconv.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16
17
xnn_qs8_dwconv_minmax_fp32_ukernel_up24x25__neonv8_mul16(size_t channels,size_t output_width,const int8_t ** input,const void * weights,int8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const int8_t * zero,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_dwconv_minmax_fp32_ukernel_up24x25__neonv8_mul16(
19 size_t channels,
20 size_t output_width,
21 const int8_t** input,
22 const void* weights,
23 int8_t* output,
24 size_t input_stride,
25 size_t output_increment,
26 size_t input_offset,
27 const int8_t* zero,
28 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(channels != 0);
31 assert(output_width != 0);
32
33 const float32x4_t vscale = vld1q_dup_f32(¶ms->fp32_neonv8.scale);
34 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->fp32_neonv8.output_zero_point);
35 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->fp32_neonv8.output_min);
36 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->fp32_neonv8.output_max);
37 do {
38 const int8_t* i0 = input[0];
39 assert(i0 != NULL);
40 if XNN_UNPREDICTABLE(i0 != zero) {
41 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
42 }
43 const int8_t* i1 = input[1];
44 assert(i1 != NULL);
45 if XNN_UNPREDICTABLE(i1 != zero) {
46 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
47 }
48 const int8_t* i2 = input[2];
49 assert(i2 != NULL);
50 if XNN_UNPREDICTABLE(i2 != zero) {
51 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
52 }
53 const int8_t* i3 = input[3];
54 assert(i3 != NULL);
55 if XNN_UNPREDICTABLE(i3 != zero) {
56 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
57 }
58 const int8_t* i4 = input[4];
59 assert(i4 != NULL);
60 if XNN_UNPREDICTABLE(i4 != zero) {
61 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
62 }
63 const int8_t* i5 = input[5];
64 assert(i5 != NULL);
65 if XNN_UNPREDICTABLE(i5 != zero) {
66 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
67 }
68 const int8_t* i6 = input[6];
69 assert(i6 != NULL);
70 if XNN_UNPREDICTABLE(i6 != zero) {
71 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
72 }
73 const int8_t* i7 = input[7];
74 assert(i7 != NULL);
75 if XNN_UNPREDICTABLE(i7 != zero) {
76 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
77 }
78 const int8_t* i8 = input[8];
79 assert(i8 != NULL);
80 if XNN_UNPREDICTABLE(i8 != zero) {
81 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
82 }
83 const int8_t* i9 = input[9];
84 assert(i9 != NULL);
85 if XNN_UNPREDICTABLE(i9 != zero) {
86 i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
87 }
88 const int8_t* i10 = input[10];
89 assert(i10 != NULL);
90 if XNN_UNPREDICTABLE(i10 != zero) {
91 i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
92 }
93 const int8_t* i11 = input[11];
94 assert(i11 != NULL);
95 if XNN_UNPREDICTABLE(i11 != zero) {
96 i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
97 }
98 const int8_t* i12 = input[12];
99 assert(i12 != NULL);
100 if XNN_UNPREDICTABLE(i12 != zero) {
101 i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
102 }
103 const int8_t* i13 = input[13];
104 assert(i13 != NULL);
105 if XNN_UNPREDICTABLE(i13 != zero) {
106 i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
107 }
108 const int8_t* i14 = input[14];
109 assert(i14 != NULL);
110 if XNN_UNPREDICTABLE(i14 != zero) {
111 i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
112 }
113 const int8_t* i15 = input[15];
114 assert(i15 != NULL);
115 if XNN_UNPREDICTABLE(i15 != zero) {
116 i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
117 }
118 const int8_t* i16 = input[16];
119 assert(i16 != NULL);
120 if XNN_UNPREDICTABLE(i16 != zero) {
121 i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
122 }
123 const int8_t* i17 = input[17];
124 assert(i17 != NULL);
125 if XNN_UNPREDICTABLE(i17 != zero) {
126 i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
127 }
128 const int8_t* i18 = input[18];
129 assert(i18 != NULL);
130 if XNN_UNPREDICTABLE(i18 != zero) {
131 i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
132 }
133 const int8_t* i19 = input[19];
134 assert(i19 != NULL);
135 if XNN_UNPREDICTABLE(i19 != zero) {
136 i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
137 }
138 const int8_t* i20 = input[20];
139 assert(i20 != NULL);
140 if XNN_UNPREDICTABLE(i20 != zero) {
141 i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
142 }
143 const int8_t* i21 = input[21];
144 assert(i21 != NULL);
145 if XNN_UNPREDICTABLE(i21 != zero) {
146 i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
147 }
148 const int8_t* i22 = input[22];
149 assert(i22 != NULL);
150 if XNN_UNPREDICTABLE(i22 != zero) {
151 i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
152 }
153 const int8_t* i23 = input[23];
154 assert(i23 != NULL);
155 if XNN_UNPREDICTABLE(i23 != zero) {
156 i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
157 }
158 const int8_t* i24 = input[24];
159 assert(i24 != NULL);
160 if XNN_UNPREDICTABLE(i24 != zero) {
161 i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
162 }
163 input = (const int8_t**) ((uintptr_t) input + input_stride);
164
165 size_t c = channels;
166 const void* w = weights;
167 for (; c >= 24; c -= 24) {
168 int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
169 int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
170 int32x4_t vacc89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
171 int32x4_t vaccCDEF = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
172 int32x4_t vaccGHIJ = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
173 int32x4_t vaccKLMN = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
174
175
176 const int16x8_t vi0x01234567 = vmovl_s8(vld1_s8(i0)); i0 += 8;
177 const int16x8_t vk0x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
178 const int16x8_t vi0x89ABCDEF = vmovl_s8(vld1_s8(i0)); i0 += 8;
179 const int16x8_t vk0x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
180 const int16x8_t vi0xGHIJKLMN = vmovl_s8(vld1_s8(i0)); i0 += 8;
181 const int16x8_t vk0xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
182
183 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
184 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
185 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi0x89ABCDEF), vget_low_s16(vk0x89ABCDEF));
186 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi0x89ABCDEF), vget_high_s16(vk0x89ABCDEF));
187 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi0xGHIJKLMN), vget_low_s16(vk0xGHIJKLMN));
188 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi0xGHIJKLMN), vget_high_s16(vk0xGHIJKLMN));
189
190 const int16x8_t vi1x01234567 = vmovl_s8(vld1_s8(i1)); i1 += 8;
191 const int16x8_t vk1x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
192 const int16x8_t vi1x89ABCDEF = vmovl_s8(vld1_s8(i1)); i1 += 8;
193 const int16x8_t vk1x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
194 const int16x8_t vi1xGHIJKLMN = vmovl_s8(vld1_s8(i1)); i1 += 8;
195 const int16x8_t vk1xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
196
197 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
198 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
199 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi1x89ABCDEF), vget_low_s16(vk1x89ABCDEF));
200 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi1x89ABCDEF), vget_high_s16(vk1x89ABCDEF));
201 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi1xGHIJKLMN), vget_low_s16(vk1xGHIJKLMN));
202 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi1xGHIJKLMN), vget_high_s16(vk1xGHIJKLMN));
203
204 const int16x8_t vi2x01234567 = vmovl_s8(vld1_s8(i2)); i2 += 8;
205 const int16x8_t vk2x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
206 const int16x8_t vi2x89ABCDEF = vmovl_s8(vld1_s8(i2)); i2 += 8;
207 const int16x8_t vk2x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
208 const int16x8_t vi2xGHIJKLMN = vmovl_s8(vld1_s8(i2)); i2 += 8;
209 const int16x8_t vk2xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
210
211 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
212 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
213 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi2x89ABCDEF), vget_low_s16(vk2x89ABCDEF));
214 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi2x89ABCDEF), vget_high_s16(vk2x89ABCDEF));
215 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi2xGHIJKLMN), vget_low_s16(vk2xGHIJKLMN));
216 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi2xGHIJKLMN), vget_high_s16(vk2xGHIJKLMN));
217
218 const int16x8_t vi3x01234567 = vmovl_s8(vld1_s8(i3)); i3 += 8;
219 const int16x8_t vk3x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
220 const int16x8_t vi3x89ABCDEF = vmovl_s8(vld1_s8(i3)); i3 += 8;
221 const int16x8_t vk3x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
222 const int16x8_t vi3xGHIJKLMN = vmovl_s8(vld1_s8(i3)); i3 += 8;
223 const int16x8_t vk3xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
224
225 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
226 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
227 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi3x89ABCDEF), vget_low_s16(vk3x89ABCDEF));
228 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi3x89ABCDEF), vget_high_s16(vk3x89ABCDEF));
229 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi3xGHIJKLMN), vget_low_s16(vk3xGHIJKLMN));
230 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi3xGHIJKLMN), vget_high_s16(vk3xGHIJKLMN));
231
232 const int16x8_t vi4x01234567 = vmovl_s8(vld1_s8(i4)); i4 += 8;
233 const int16x8_t vk4x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
234 const int16x8_t vi4x89ABCDEF = vmovl_s8(vld1_s8(i4)); i4 += 8;
235 const int16x8_t vk4x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
236 const int16x8_t vi4xGHIJKLMN = vmovl_s8(vld1_s8(i4)); i4 += 8;
237 const int16x8_t vk4xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
238
239 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
240 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
241 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi4x89ABCDEF), vget_low_s16(vk4x89ABCDEF));
242 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi4x89ABCDEF), vget_high_s16(vk4x89ABCDEF));
243 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi4xGHIJKLMN), vget_low_s16(vk4xGHIJKLMN));
244 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi4xGHIJKLMN), vget_high_s16(vk4xGHIJKLMN));
245
246 const int16x8_t vi5x01234567 = vmovl_s8(vld1_s8(i5)); i5 += 8;
247 const int16x8_t vk5x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
248 const int16x8_t vi5x89ABCDEF = vmovl_s8(vld1_s8(i5)); i5 += 8;
249 const int16x8_t vk5x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
250 const int16x8_t vi5xGHIJKLMN = vmovl_s8(vld1_s8(i5)); i5 += 8;
251 const int16x8_t vk5xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
252
253 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi5x01234567), vget_low_s16(vk5x01234567));
254 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi5x01234567), vget_high_s16(vk5x01234567));
255 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi5x89ABCDEF), vget_low_s16(vk5x89ABCDEF));
256 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi5x89ABCDEF), vget_high_s16(vk5x89ABCDEF));
257 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi5xGHIJKLMN), vget_low_s16(vk5xGHIJKLMN));
258 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi5xGHIJKLMN), vget_high_s16(vk5xGHIJKLMN));
259
260 const int16x8_t vi6x01234567 = vmovl_s8(vld1_s8(i6)); i6 += 8;
261 const int16x8_t vk6x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
262 const int16x8_t vi6x89ABCDEF = vmovl_s8(vld1_s8(i6)); i6 += 8;
263 const int16x8_t vk6x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
264 const int16x8_t vi6xGHIJKLMN = vmovl_s8(vld1_s8(i6)); i6 += 8;
265 const int16x8_t vk6xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
266
267 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi6x01234567), vget_low_s16(vk6x01234567));
268 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi6x01234567), vget_high_s16(vk6x01234567));
269 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi6x89ABCDEF), vget_low_s16(vk6x89ABCDEF));
270 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi6x89ABCDEF), vget_high_s16(vk6x89ABCDEF));
271 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi6xGHIJKLMN), vget_low_s16(vk6xGHIJKLMN));
272 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi6xGHIJKLMN), vget_high_s16(vk6xGHIJKLMN));
273
274 const int16x8_t vi7x01234567 = vmovl_s8(vld1_s8(i7)); i7 += 8;
275 const int16x8_t vk7x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
276 const int16x8_t vi7x89ABCDEF = vmovl_s8(vld1_s8(i7)); i7 += 8;
277 const int16x8_t vk7x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
278 const int16x8_t vi7xGHIJKLMN = vmovl_s8(vld1_s8(i7)); i7 += 8;
279 const int16x8_t vk7xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
280
281 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi7x01234567), vget_low_s16(vk7x01234567));
282 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi7x01234567), vget_high_s16(vk7x01234567));
283 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi7x89ABCDEF), vget_low_s16(vk7x89ABCDEF));
284 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi7x89ABCDEF), vget_high_s16(vk7x89ABCDEF));
285 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi7xGHIJKLMN), vget_low_s16(vk7xGHIJKLMN));
286 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi7xGHIJKLMN), vget_high_s16(vk7xGHIJKLMN));
287
288 const int16x8_t vi8x01234567 = vmovl_s8(vld1_s8(i8)); i8 += 8;
289 const int16x8_t vk8x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
290 const int16x8_t vi8x89ABCDEF = vmovl_s8(vld1_s8(i8)); i8 += 8;
291 const int16x8_t vk8x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
292 const int16x8_t vi8xGHIJKLMN = vmovl_s8(vld1_s8(i8)); i8 += 8;
293 const int16x8_t vk8xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
294
295 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi8x01234567), vget_low_s16(vk8x01234567));
296 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi8x01234567), vget_high_s16(vk8x01234567));
297 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi8x89ABCDEF), vget_low_s16(vk8x89ABCDEF));
298 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi8x89ABCDEF), vget_high_s16(vk8x89ABCDEF));
299 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi8xGHIJKLMN), vget_low_s16(vk8xGHIJKLMN));
300 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi8xGHIJKLMN), vget_high_s16(vk8xGHIJKLMN));
301
302 const int16x8_t vi9x01234567 = vmovl_s8(vld1_s8(i9)); i9 += 8;
303 const int16x8_t vk9x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
304 const int16x8_t vi9x89ABCDEF = vmovl_s8(vld1_s8(i9)); i9 += 8;
305 const int16x8_t vk9x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
306 const int16x8_t vi9xGHIJKLMN = vmovl_s8(vld1_s8(i9)); i9 += 8;
307 const int16x8_t vk9xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
308
309 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi9x01234567), vget_low_s16(vk9x01234567));
310 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi9x01234567), vget_high_s16(vk9x01234567));
311 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi9x89ABCDEF), vget_low_s16(vk9x89ABCDEF));
312 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi9x89ABCDEF), vget_high_s16(vk9x89ABCDEF));
313 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi9xGHIJKLMN), vget_low_s16(vk9xGHIJKLMN));
314 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi9xGHIJKLMN), vget_high_s16(vk9xGHIJKLMN));
315
316 const int16x8_t vi10x01234567 = vmovl_s8(vld1_s8(i10)); i10 += 8;
317 const int16x8_t vk10x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
318 const int16x8_t vi10x89ABCDEF = vmovl_s8(vld1_s8(i10)); i10 += 8;
319 const int16x8_t vk10x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
320 const int16x8_t vi10xGHIJKLMN = vmovl_s8(vld1_s8(i10)); i10 += 8;
321 const int16x8_t vk10xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
322
323 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi10x01234567), vget_low_s16(vk10x01234567));
324 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi10x01234567), vget_high_s16(vk10x01234567));
325 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi10x89ABCDEF), vget_low_s16(vk10x89ABCDEF));
326 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi10x89ABCDEF), vget_high_s16(vk10x89ABCDEF));
327 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi10xGHIJKLMN), vget_low_s16(vk10xGHIJKLMN));
328 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi10xGHIJKLMN), vget_high_s16(vk10xGHIJKLMN));
329
330 const int16x8_t vi11x01234567 = vmovl_s8(vld1_s8(i11)); i11 += 8;
331 const int16x8_t vk11x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
332 const int16x8_t vi11x89ABCDEF = vmovl_s8(vld1_s8(i11)); i11 += 8;
333 const int16x8_t vk11x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
334 const int16x8_t vi11xGHIJKLMN = vmovl_s8(vld1_s8(i11)); i11 += 8;
335 const int16x8_t vk11xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
336
337 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi11x01234567), vget_low_s16(vk11x01234567));
338 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi11x01234567), vget_high_s16(vk11x01234567));
339 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi11x89ABCDEF), vget_low_s16(vk11x89ABCDEF));
340 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi11x89ABCDEF), vget_high_s16(vk11x89ABCDEF));
341 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi11xGHIJKLMN), vget_low_s16(vk11xGHIJKLMN));
342 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi11xGHIJKLMN), vget_high_s16(vk11xGHIJKLMN));
343
344 const int16x8_t vi12x01234567 = vmovl_s8(vld1_s8(i12)); i12 += 8;
345 const int16x8_t vk12x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
346 const int16x8_t vi12x89ABCDEF = vmovl_s8(vld1_s8(i12)); i12 += 8;
347 const int16x8_t vk12x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
348 const int16x8_t vi12xGHIJKLMN = vmovl_s8(vld1_s8(i12)); i12 += 8;
349 const int16x8_t vk12xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
350
351 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi12x01234567), vget_low_s16(vk12x01234567));
352 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi12x01234567), vget_high_s16(vk12x01234567));
353 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi12x89ABCDEF), vget_low_s16(vk12x89ABCDEF));
354 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi12x89ABCDEF), vget_high_s16(vk12x89ABCDEF));
355 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi12xGHIJKLMN), vget_low_s16(vk12xGHIJKLMN));
356 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi12xGHIJKLMN), vget_high_s16(vk12xGHIJKLMN));
357
358 const int16x8_t vi13x01234567 = vmovl_s8(vld1_s8(i13)); i13 += 8;
359 const int16x8_t vk13x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
360 const int16x8_t vi13x89ABCDEF = vmovl_s8(vld1_s8(i13)); i13 += 8;
361 const int16x8_t vk13x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
362 const int16x8_t vi13xGHIJKLMN = vmovl_s8(vld1_s8(i13)); i13 += 8;
363 const int16x8_t vk13xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
364
365 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi13x01234567), vget_low_s16(vk13x01234567));
366 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi13x01234567), vget_high_s16(vk13x01234567));
367 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi13x89ABCDEF), vget_low_s16(vk13x89ABCDEF));
368 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi13x89ABCDEF), vget_high_s16(vk13x89ABCDEF));
369 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi13xGHIJKLMN), vget_low_s16(vk13xGHIJKLMN));
370 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi13xGHIJKLMN), vget_high_s16(vk13xGHIJKLMN));
371
372 const int16x8_t vi14x01234567 = vmovl_s8(vld1_s8(i14)); i14 += 8;
373 const int16x8_t vk14x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
374 const int16x8_t vi14x89ABCDEF = vmovl_s8(vld1_s8(i14)); i14 += 8;
375 const int16x8_t vk14x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
376 const int16x8_t vi14xGHIJKLMN = vmovl_s8(vld1_s8(i14)); i14 += 8;
377 const int16x8_t vk14xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
378
379 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi14x01234567), vget_low_s16(vk14x01234567));
380 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi14x01234567), vget_high_s16(vk14x01234567));
381 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi14x89ABCDEF), vget_low_s16(vk14x89ABCDEF));
382 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi14x89ABCDEF), vget_high_s16(vk14x89ABCDEF));
383 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi14xGHIJKLMN), vget_low_s16(vk14xGHIJKLMN));
384 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi14xGHIJKLMN), vget_high_s16(vk14xGHIJKLMN));
385
386 const int16x8_t vi15x01234567 = vmovl_s8(vld1_s8(i15)); i15 += 8;
387 const int16x8_t vk15x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
388 const int16x8_t vi15x89ABCDEF = vmovl_s8(vld1_s8(i15)); i15 += 8;
389 const int16x8_t vk15x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
390 const int16x8_t vi15xGHIJKLMN = vmovl_s8(vld1_s8(i15)); i15 += 8;
391 const int16x8_t vk15xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
392
393 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi15x01234567), vget_low_s16(vk15x01234567));
394 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi15x01234567), vget_high_s16(vk15x01234567));
395 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi15x89ABCDEF), vget_low_s16(vk15x89ABCDEF));
396 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi15x89ABCDEF), vget_high_s16(vk15x89ABCDEF));
397 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi15xGHIJKLMN), vget_low_s16(vk15xGHIJKLMN));
398 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi15xGHIJKLMN), vget_high_s16(vk15xGHIJKLMN));
399
400 const int16x8_t vi16x01234567 = vmovl_s8(vld1_s8(i16)); i16 += 8;
401 const int16x8_t vk16x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
402 const int16x8_t vi16x89ABCDEF = vmovl_s8(vld1_s8(i16)); i16 += 8;
403 const int16x8_t vk16x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
404 const int16x8_t vi16xGHIJKLMN = vmovl_s8(vld1_s8(i16)); i16 += 8;
405 const int16x8_t vk16xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
406
407 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi16x01234567), vget_low_s16(vk16x01234567));
408 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi16x01234567), vget_high_s16(vk16x01234567));
409 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi16x89ABCDEF), vget_low_s16(vk16x89ABCDEF));
410 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi16x89ABCDEF), vget_high_s16(vk16x89ABCDEF));
411 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi16xGHIJKLMN), vget_low_s16(vk16xGHIJKLMN));
412 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi16xGHIJKLMN), vget_high_s16(vk16xGHIJKLMN));
413
414 const int16x8_t vi17x01234567 = vmovl_s8(vld1_s8(i17)); i17 += 8;
415 const int16x8_t vk17x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
416 const int16x8_t vi17x89ABCDEF = vmovl_s8(vld1_s8(i17)); i17 += 8;
417 const int16x8_t vk17x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
418 const int16x8_t vi17xGHIJKLMN = vmovl_s8(vld1_s8(i17)); i17 += 8;
419 const int16x8_t vk17xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
420
421 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi17x01234567), vget_low_s16(vk17x01234567));
422 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi17x01234567), vget_high_s16(vk17x01234567));
423 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi17x89ABCDEF), vget_low_s16(vk17x89ABCDEF));
424 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi17x89ABCDEF), vget_high_s16(vk17x89ABCDEF));
425 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi17xGHIJKLMN), vget_low_s16(vk17xGHIJKLMN));
426 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi17xGHIJKLMN), vget_high_s16(vk17xGHIJKLMN));
427
428 const int16x8_t vi18x01234567 = vmovl_s8(vld1_s8(i18)); i18 += 8;
429 const int16x8_t vk18x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
430 const int16x8_t vi18x89ABCDEF = vmovl_s8(vld1_s8(i18)); i18 += 8;
431 const int16x8_t vk18x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
432 const int16x8_t vi18xGHIJKLMN = vmovl_s8(vld1_s8(i18)); i18 += 8;
433 const int16x8_t vk18xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
434
435 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi18x01234567), vget_low_s16(vk18x01234567));
436 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi18x01234567), vget_high_s16(vk18x01234567));
437 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi18x89ABCDEF), vget_low_s16(vk18x89ABCDEF));
438 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi18x89ABCDEF), vget_high_s16(vk18x89ABCDEF));
439 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi18xGHIJKLMN), vget_low_s16(vk18xGHIJKLMN));
440 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi18xGHIJKLMN), vget_high_s16(vk18xGHIJKLMN));
441
442 const int16x8_t vi19x01234567 = vmovl_s8(vld1_s8(i19)); i19 += 8;
443 const int16x8_t vk19x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
444 const int16x8_t vi19x89ABCDEF = vmovl_s8(vld1_s8(i19)); i19 += 8;
445 const int16x8_t vk19x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
446 const int16x8_t vi19xGHIJKLMN = vmovl_s8(vld1_s8(i19)); i19 += 8;
447 const int16x8_t vk19xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
448
449 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi19x01234567), vget_low_s16(vk19x01234567));
450 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi19x01234567), vget_high_s16(vk19x01234567));
451 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi19x89ABCDEF), vget_low_s16(vk19x89ABCDEF));
452 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi19x89ABCDEF), vget_high_s16(vk19x89ABCDEF));
453 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi19xGHIJKLMN), vget_low_s16(vk19xGHIJKLMN));
454 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi19xGHIJKLMN), vget_high_s16(vk19xGHIJKLMN));
455
456 const int16x8_t vi20x01234567 = vmovl_s8(vld1_s8(i20)); i20 += 8;
457 const int16x8_t vk20x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
458 const int16x8_t vi20x89ABCDEF = vmovl_s8(vld1_s8(i20)); i20 += 8;
459 const int16x8_t vk20x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
460 const int16x8_t vi20xGHIJKLMN = vmovl_s8(vld1_s8(i20)); i20 += 8;
461 const int16x8_t vk20xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
462
463 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi20x01234567), vget_low_s16(vk20x01234567));
464 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi20x01234567), vget_high_s16(vk20x01234567));
465 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi20x89ABCDEF), vget_low_s16(vk20x89ABCDEF));
466 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi20x89ABCDEF), vget_high_s16(vk20x89ABCDEF));
467 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi20xGHIJKLMN), vget_low_s16(vk20xGHIJKLMN));
468 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi20xGHIJKLMN), vget_high_s16(vk20xGHIJKLMN));
469
470 const int16x8_t vi21x01234567 = vmovl_s8(vld1_s8(i21)); i21 += 8;
471 const int16x8_t vk21x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
472 const int16x8_t vi21x89ABCDEF = vmovl_s8(vld1_s8(i21)); i21 += 8;
473 const int16x8_t vk21x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
474 const int16x8_t vi21xGHIJKLMN = vmovl_s8(vld1_s8(i21)); i21 += 8;
475 const int16x8_t vk21xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
476
477 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi21x01234567), vget_low_s16(vk21x01234567));
478 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi21x01234567), vget_high_s16(vk21x01234567));
479 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi21x89ABCDEF), vget_low_s16(vk21x89ABCDEF));
480 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi21x89ABCDEF), vget_high_s16(vk21x89ABCDEF));
481 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi21xGHIJKLMN), vget_low_s16(vk21xGHIJKLMN));
482 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi21xGHIJKLMN), vget_high_s16(vk21xGHIJKLMN));
483
484 const int16x8_t vi22x01234567 = vmovl_s8(vld1_s8(i22)); i22 += 8;
485 const int16x8_t vk22x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
486 const int16x8_t vi22x89ABCDEF = vmovl_s8(vld1_s8(i22)); i22 += 8;
487 const int16x8_t vk22x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
488 const int16x8_t vi22xGHIJKLMN = vmovl_s8(vld1_s8(i22)); i22 += 8;
489 const int16x8_t vk22xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
490
491 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi22x01234567), vget_low_s16(vk22x01234567));
492 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi22x01234567), vget_high_s16(vk22x01234567));
493 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi22x89ABCDEF), vget_low_s16(vk22x89ABCDEF));
494 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi22x89ABCDEF), vget_high_s16(vk22x89ABCDEF));
495 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi22xGHIJKLMN), vget_low_s16(vk22xGHIJKLMN));
496 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi22xGHIJKLMN), vget_high_s16(vk22xGHIJKLMN));
497
498 const int16x8_t vi23x01234567 = vmovl_s8(vld1_s8(i23)); i23 += 8;
499 const int16x8_t vk23x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
500 const int16x8_t vi23x89ABCDEF = vmovl_s8(vld1_s8(i23)); i23 += 8;
501 const int16x8_t vk23x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
502 const int16x8_t vi23xGHIJKLMN = vmovl_s8(vld1_s8(i23)); i23 += 8;
503 const int16x8_t vk23xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
504
505 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi23x01234567), vget_low_s16(vk23x01234567));
506 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi23x01234567), vget_high_s16(vk23x01234567));
507 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi23x89ABCDEF), vget_low_s16(vk23x89ABCDEF));
508 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi23x89ABCDEF), vget_high_s16(vk23x89ABCDEF));
509 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi23xGHIJKLMN), vget_low_s16(vk23xGHIJKLMN));
510 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi23xGHIJKLMN), vget_high_s16(vk23xGHIJKLMN));
511
512 const int16x8_t vi24x01234567 = vmovl_s8(vld1_s8(i24)); i24 += 8;
513 const int16x8_t vk24x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
514 const int16x8_t vi24x89ABCDEF = vmovl_s8(vld1_s8(i24)); i24 += 8;
515 const int16x8_t vk24x89ABCDEF = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
516 const int16x8_t vi24xGHIJKLMN = vmovl_s8(vld1_s8(i24)); i24 += 8;
517 const int16x8_t vk24xGHIJKLMN = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8);
518
519 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi24x01234567), vget_low_s16(vk24x01234567));
520 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi24x01234567), vget_high_s16(vk24x01234567));
521 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi24x89ABCDEF), vget_low_s16(vk24x89ABCDEF));
522 vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi24x89ABCDEF), vget_high_s16(vk24x89ABCDEF));
523 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi24xGHIJKLMN), vget_low_s16(vk24xGHIJKLMN));
524 vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi24xGHIJKLMN), vget_high_s16(vk24xGHIJKLMN));
525
526 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
527 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
528 float32x4_t vfpacc89AB = vcvtq_f32_s32(vacc89AB);
529 float32x4_t vfpaccCDEF = vcvtq_f32_s32(vaccCDEF);
530 float32x4_t vfpaccGHIJ = vcvtq_f32_s32(vaccGHIJ);
531 float32x4_t vfpaccKLMN = vcvtq_f32_s32(vaccKLMN);
532
533 vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
534 vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
535 vfpacc89AB = vmulq_f32(vfpacc89AB, vscale);
536 vfpaccCDEF = vmulq_f32(vfpaccCDEF, vscale);
537 vfpaccGHIJ = vmulq_f32(vfpaccGHIJ, vscale);
538 vfpaccKLMN = vmulq_f32(vfpaccKLMN, vscale);
539
540 vacc0123 = vcvtnq_s32_f32(vfpacc0123);
541 vacc4567 = vcvtnq_s32_f32(vfpacc4567);
542 vacc89AB = vcvtnq_s32_f32(vfpacc89AB);
543 vaccCDEF = vcvtnq_s32_f32(vfpaccCDEF);
544 vaccGHIJ = vcvtnq_s32_f32(vfpaccGHIJ);
545 vaccKLMN = vcvtnq_s32_f32(vfpaccKLMN);
546
547 #if XNN_ARCH_ARM64
548 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
549 int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
550 int16x8_t vaccGHIJKLMN = vqmovn_high_s32(vqmovn_s32(vaccGHIJ), vaccKLMN);
551
552 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
553 vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
554 vaccGHIJKLMN = vqaddq_s16(vaccGHIJKLMN, voutput_zero_point);
555
556 int8x16_t vout0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc01234567), vacc89ABCDEF);
557 int8x8_t voutGHIJKLMN = vqmovn_s16(vaccGHIJKLMN);
558 #else // !XNN_ARCH_ARM64
559 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
560 int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
561 int16x8_t vaccGHIJKLMN = vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN));
562
563 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
564 vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
565 vaccGHIJKLMN = vqaddq_s16(vaccGHIJKLMN, voutput_zero_point);
566
567 int8x16_t vout0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc01234567), vqmovn_s16(vacc89ABCDEF));
568 int8x8_t voutGHIJKLMN = vqmovn_s16(vaccGHIJKLMN);
569 #endif // !XNN_ARCH_ARM64
570
571 vout0123456789ABCDEF = vmaxq_s8(vout0123456789ABCDEF, voutput_min);
572 voutGHIJKLMN = vmax_s8(voutGHIJKLMN, vget_low_s8(voutput_min));
573
574 vout0123456789ABCDEF = vminq_s8(vout0123456789ABCDEF, voutput_max);
575 voutGHIJKLMN = vmin_s8(voutGHIJKLMN, vget_low_s8(voutput_max));
576
577 vst1q_s8(output, vout0123456789ABCDEF); output += 16;
578 vst1_s8(output, voutGHIJKLMN); output += 8;
579 }
580 if XNN_UNLIKELY(c != 0) {
581 const int8_t* k = (const int8_t*) ((const int32_t*) w + 24);
582 do {
583 int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
584 int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
585
586 const int16x8_t vi0x01234567 = vmovl_s8(vld1_s8(i0)); i0 += 8;
587 const int16x8_t vk0x01234567 = vmovl_s8(vld1_s8(k)); k += 8;
588
589 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
590 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
591 const int16x8_t vi1x01234567 = vmovl_s8(vld1_s8(i1)); i1 += 8;
592 const int16x8_t vk1x01234567 = vmovl_s8(vld1_s8((const void*) (k + 16)));
593
594 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
595 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
596 const int16x8_t vi2x01234567 = vmovl_s8(vld1_s8(i2)); i2 += 8;
597 const int16x8_t vk2x01234567 = vmovl_s8(vld1_s8((const void*) (k + 40)));
598
599 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
600 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
601 const int16x8_t vi3x01234567 = vmovl_s8(vld1_s8(i3)); i3 += 8;
602 const int16x8_t vk3x01234567 = vmovl_s8(vld1_s8((const void*) (k + 64)));
603
604 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
605 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
606 const int16x8_t vi4x01234567 = vmovl_s8(vld1_s8(i4)); i4 += 8;
607 const int16x8_t vk4x01234567 = vmovl_s8(vld1_s8((const void*) (k + 88)));
608
609 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
610 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
611 const int16x8_t vi5x01234567 = vmovl_s8(vld1_s8(i5)); i5 += 8;
612 const int16x8_t vk5x01234567 = vmovl_s8(vld1_s8((const void*) (k + 112)));
613
614 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi5x01234567), vget_low_s16(vk5x01234567));
615 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi5x01234567), vget_high_s16(vk5x01234567));
616 const int16x8_t vi6x01234567 = vmovl_s8(vld1_s8(i6)); i6 += 8;
617 const int16x8_t vk6x01234567 = vmovl_s8(vld1_s8((const void*) (k + 136)));
618
619 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi6x01234567), vget_low_s16(vk6x01234567));
620 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi6x01234567), vget_high_s16(vk6x01234567));
621 const int16x8_t vi7x01234567 = vmovl_s8(vld1_s8(i7)); i7 += 8;
622 const int16x8_t vk7x01234567 = vmovl_s8(vld1_s8((const void*) (k + 160)));
623
624 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi7x01234567), vget_low_s16(vk7x01234567));
625 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi7x01234567), vget_high_s16(vk7x01234567));
626 const int16x8_t vi8x01234567 = vmovl_s8(vld1_s8(i8)); i8 += 8;
627 const int16x8_t vk8x01234567 = vmovl_s8(vld1_s8((const void*) (k + 184)));
628
629 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi8x01234567), vget_low_s16(vk8x01234567));
630 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi8x01234567), vget_high_s16(vk8x01234567));
631 const int16x8_t vi9x01234567 = vmovl_s8(vld1_s8(i9)); i9 += 8;
632 const int16x8_t vk9x01234567 = vmovl_s8(vld1_s8((const void*) (k + 208)));
633
634 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi9x01234567), vget_low_s16(vk9x01234567));
635 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi9x01234567), vget_high_s16(vk9x01234567));
636 const int16x8_t vi10x01234567 = vmovl_s8(vld1_s8(i10)); i10 += 8;
637 const int16x8_t vk10x01234567 = vmovl_s8(vld1_s8((const void*) (k + 232)));
638
639 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi10x01234567), vget_low_s16(vk10x01234567));
640 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi10x01234567), vget_high_s16(vk10x01234567));
641 const int16x8_t vi11x01234567 = vmovl_s8(vld1_s8(i11)); i11 += 8;
642 const int16x8_t vk11x01234567 = vmovl_s8(vld1_s8((const void*) (k + 256)));
643
644 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi11x01234567), vget_low_s16(vk11x01234567));
645 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi11x01234567), vget_high_s16(vk11x01234567));
646 const int16x8_t vi12x01234567 = vmovl_s8(vld1_s8(i12)); i12 += 8;
647 const int16x8_t vk12x01234567 = vmovl_s8(vld1_s8((const void*) (k + 280)));
648
649 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi12x01234567), vget_low_s16(vk12x01234567));
650 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi12x01234567), vget_high_s16(vk12x01234567));
651 const int16x8_t vi13x01234567 = vmovl_s8(vld1_s8(i13)); i13 += 8;
652 const int16x8_t vk13x01234567 = vmovl_s8(vld1_s8((const void*) (k + 304)));
653
654 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi13x01234567), vget_low_s16(vk13x01234567));
655 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi13x01234567), vget_high_s16(vk13x01234567));
656 const int16x8_t vi14x01234567 = vmovl_s8(vld1_s8(i14)); i14 += 8;
657 const int16x8_t vk14x01234567 = vmovl_s8(vld1_s8((const void*) (k + 328)));
658
659 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi14x01234567), vget_low_s16(vk14x01234567));
660 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi14x01234567), vget_high_s16(vk14x01234567));
661 const int16x8_t vi15x01234567 = vmovl_s8(vld1_s8(i15)); i15 += 8;
662 const int16x8_t vk15x01234567 = vmovl_s8(vld1_s8((const void*) (k + 352)));
663
664 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi15x01234567), vget_low_s16(vk15x01234567));
665 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi15x01234567), vget_high_s16(vk15x01234567));
666 const int16x8_t vi16x01234567 = vmovl_s8(vld1_s8(i16)); i16 += 8;
667 const int16x8_t vk16x01234567 = vmovl_s8(vld1_s8((const void*) (k + 376)));
668
669 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi16x01234567), vget_low_s16(vk16x01234567));
670 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi16x01234567), vget_high_s16(vk16x01234567));
671 const int16x8_t vi17x01234567 = vmovl_s8(vld1_s8(i17)); i17 += 8;
672 const int16x8_t vk17x01234567 = vmovl_s8(vld1_s8((const void*) (k + 400)));
673
674 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi17x01234567), vget_low_s16(vk17x01234567));
675 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi17x01234567), vget_high_s16(vk17x01234567));
676 const int16x8_t vi18x01234567 = vmovl_s8(vld1_s8(i18)); i18 += 8;
677 const int16x8_t vk18x01234567 = vmovl_s8(vld1_s8((const void*) (k + 424)));
678
679 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi18x01234567), vget_low_s16(vk18x01234567));
680 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi18x01234567), vget_high_s16(vk18x01234567));
681 const int16x8_t vi19x01234567 = vmovl_s8(vld1_s8(i19)); i19 += 8;
682 const int16x8_t vk19x01234567 = vmovl_s8(vld1_s8((const void*) (k + 448)));
683
684 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi19x01234567), vget_low_s16(vk19x01234567));
685 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi19x01234567), vget_high_s16(vk19x01234567));
686 const int16x8_t vi20x01234567 = vmovl_s8(vld1_s8(i20)); i20 += 8;
687 const int16x8_t vk20x01234567 = vmovl_s8(vld1_s8((const void*) (k + 472)));
688
689 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi20x01234567), vget_low_s16(vk20x01234567));
690 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi20x01234567), vget_high_s16(vk20x01234567));
691 const int16x8_t vi21x01234567 = vmovl_s8(vld1_s8(i21)); i21 += 8;
692 const int16x8_t vk21x01234567 = vmovl_s8(vld1_s8((const void*) (k + 496)));
693
694 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi21x01234567), vget_low_s16(vk21x01234567));
695 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi21x01234567), vget_high_s16(vk21x01234567));
696 const int16x8_t vi22x01234567 = vmovl_s8(vld1_s8(i22)); i22 += 8;
697 const int16x8_t vk22x01234567 = vmovl_s8(vld1_s8((const void*) (k + 520)));
698
699 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi22x01234567), vget_low_s16(vk22x01234567));
700 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi22x01234567), vget_high_s16(vk22x01234567));
701 const int16x8_t vi23x01234567 = vmovl_s8(vld1_s8(i23)); i23 += 8;
702 const int16x8_t vk23x01234567 = vmovl_s8(vld1_s8((const void*) (k + 544)));
703
704 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi23x01234567), vget_low_s16(vk23x01234567));
705 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi23x01234567), vget_high_s16(vk23x01234567));
706 const int16x8_t vi24x01234567 = vmovl_s8(vld1_s8(i24)); i24 += 8;
707 const int16x8_t vk24x01234567 = vmovl_s8(vld1_s8((const void*) (k + 568)));
708
709 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi24x01234567), vget_low_s16(vk24x01234567));
710 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi24x01234567), vget_high_s16(vk24x01234567));
711
712 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
713 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
714
715 vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
716 vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
717
718 vacc0123 = vcvtnq_s32_f32(vfpacc0123);
719 vacc4567 = vcvtnq_s32_f32(vfpacc4567);
720
721 #if XNN_ARCH_ARM64
722 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
723 #else
724 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
725 #endif
726 vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
727
728 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
729 vout01234567 = vmax_s8(vout01234567, vget_low_s8(voutput_min));
730 vout01234567 = vmin_s8(vout01234567, vget_low_s8(voutput_max));
731
732 if XNN_LIKELY(c >= 8) {
733 vst1_s8(output, vout01234567); output += 8;
734 c -= 8;
735 } else {
736 if (c & 4) {
737 vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
738 vout01234567 = vext_s8(vout01234567, vout01234567, 4);
739 }
740 if (c & 2) {
741 vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
742 vout01234567 = vext_s8(vout01234567, vout01234567, 2);
743 }
744 if (c & 1) {
745 vst1_lane_s8(output, vout01234567, 0); output += 1;
746 }
747 c = 0;
748 }
749 } while (c != 0);
750 }
751
752 output = (int8_t*) ((uintptr_t) output + output_increment);
753 } while (--output_width != 0);
754 }
755