Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-6f6m7l8c8s8r-minmax-fp32-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_fp32_ukernel_6f6m7l8c8s8r__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 > 6);
36
37 const float32x4_t vscale = vld1q_dup_f32(&params->fp32_neon.scale);
38 const float32x4_t vmagic_bias = vld1q_dup_f32(&params->fp32_neon.magic_bias);
39 const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(&params->fp32_neon.magic_bias_less_output_zero_point);
40 const int8x8_t voutput_min = vld1_dup_s8(&params->fp32_neon.output_min);
41 const int8x8_t voutput_max = vld1_dup_s8(&params->fp32_neon.output_max);
42
43 do {
44 const void* w = weights;
45
46 // First pass to process 6 inputs.
47 {
48 int32_t* b = buffer;
49 const int8_t* i0 = input[0];
50 assert(i0 != NULL);
51 if XNN_UNPREDICTABLE(i0 != zero) {
52 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
53 }
54 const int8_t* i1 = input[1];
55 assert(i1 != NULL);
56 if XNN_UNPREDICTABLE(i1 != zero) {
57 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
58 }
59 const int8_t* i2 = input[2];
60 assert(i2 != NULL);
61 if XNN_UNPREDICTABLE(i2 != zero) {
62 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
63 }
64 const int8_t* i3 = input[3];
65 assert(i3 != NULL);
66 if XNN_UNPREDICTABLE(i3 != zero) {
67 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
68 }
69 const int8_t* i4 = input[4];
70 assert(i4 != NULL);
71 if XNN_UNPREDICTABLE(i4 != zero) {
72 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
73 }
74 const int8_t* i5 = input[5];
75 assert(i5 != NULL);
76 if XNN_UNPREDICTABLE(i5 != zero) {
77 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
78 }
79 input += 6;
80
81 size_t c = round_up_po2(channels, 8);
82
83 for (; c >= 8; c -= 8) {
84 int32x4_t vacc0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
85 int32x4_t vacc4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
86
87 const int16x8_t vi0x01234567 = vmovl_s8(vld1_s8(i0)); i0 += 8;
88 const int16x8_t vk0x01234567 = 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
93 const int16x8_t vi1x01234567 = vmovl_s8(vld1_s8(i1)); i1 += 8;
94 const int16x8_t vk1x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
95
96 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
97 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
98
99 const int16x8_t vi2x01234567 = vmovl_s8(vld1_s8(i2)); i2 += 8;
100 const int16x8_t vk2x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
101
102 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
103 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
104
105 const int16x8_t vi3x01234567 = vmovl_s8(vld1_s8(i3)); i3 += 8;
106 const int16x8_t vk3x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
107
108 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
109 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
110
111 const int16x8_t vi4x01234567 = vmovl_s8(vld1_s8(i4)); i4 += 8;
112 const int16x8_t vk4x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
113
114 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
115 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
116
117 const int16x8_t vi5x01234567 = vmovl_s8(vld1_s8(i5)); i5 += 8;
118 const int16x8_t vk5x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
119
120 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi5x01234567), vget_low_s16(vk5x01234567));
121 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi5x01234567), vget_high_s16(vk5x01234567));
122
123 vst1q_s32(b, vacc0123); b += 4;
124 vst1q_s32(b, vacc4567); b += 4;
125 }
126
127 assert(c == 0);
128 }
129
130 // Middle pass to process 6 inputs in each iteration.
131 for (size_t ks = kernel_size - 6; ks > 7; ks -= 6) {
132 int32_t* b = buffer;
133
134 const int8_t* i0 = input[0];
135 assert(i0 != NULL);
136 if XNN_UNPREDICTABLE(i0 != zero) {
137 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
138 }
139 const int8_t* i1 = input[1];
140 assert(i1 != NULL);
141 if XNN_UNPREDICTABLE(i1 != zero) {
142 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
143 }
144 const int8_t* i2 = input[2];
145 assert(i2 != NULL);
146 if XNN_UNPREDICTABLE(i2 != zero) {
147 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
148 }
149 const int8_t* i3 = input[3];
150 assert(i3 != NULL);
151 if XNN_UNPREDICTABLE(i3 != zero) {
152 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
153 }
154 const int8_t* i4 = input[4];
155 assert(i4 != NULL);
156 if XNN_UNPREDICTABLE(i4 != zero) {
157 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
158 }
159 const int8_t* i5 = input[5];
160 assert(i5 != NULL);
161 if XNN_UNPREDICTABLE(i5 != zero) {
162 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
163 }
164 input += 6;
165
166 size_t c = round_up_po2(channels, 8);
167
168 for (; c >= 8; c -= 8) {
169 int32x4_t vacc0123 = vld1q_s32(b);
170 int32x4_t vacc4567 = vld1q_s32(b + 4);
171
172 const int16x8_t vi0x01234567 = vmovl_s8(vld1_s8(i0)); i0 += 8;
173 const int16x8_t vk0x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
174
175 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
176 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
177
178 const int16x8_t vi1x01234567 = vmovl_s8(vld1_s8(i1)); i1 += 8;
179 const int16x8_t vk1x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
180
181 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
182 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
183
184 const int16x8_t vi2x01234567 = vmovl_s8(vld1_s8(i2)); i2 += 8;
185 const int16x8_t vk2x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
186
187 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
188 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
189
190 const int16x8_t vi3x01234567 = vmovl_s8(vld1_s8(i3)); i3 += 8;
191 const int16x8_t vk3x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
192
193 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
194 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
195
196 const int16x8_t vi4x01234567 = vmovl_s8(vld1_s8(i4)); i4 += 8;
197 const int16x8_t vk4x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
198
199 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
200 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
201
202 const int16x8_t vi5x01234567 = vmovl_s8(vld1_s8(i5)); i5 += 8;
203 const int16x8_t vk5x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
204
205 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi5x01234567), vget_low_s16(vk5x01234567));
206 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi5x01234567), vget_high_s16(vk5x01234567));
207
208 vst1q_s32(b, vacc0123); b += 4;
209 vst1q_s32(b, vacc4567); b += 4;
210 }
211
212 assert(c == 0);
213 }
214
215 // Last pass to process up to 7 inputs.
216 {
217 const int32_t* b = buffer;
218
219 const int8_t* i0 = input[0];
220 assert(i0 != NULL);
221 if XNN_UNPREDICTABLE(i0 != zero) {
222 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
223 }
224 const int8_t* i1 = input[1];
225 assert(i1 != NULL);
226 if XNN_UNPREDICTABLE(i1 != zero) {
227 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
228 }
229 const int8_t* i2 = input[2];
230 assert(i2 != NULL);
231 if XNN_UNPREDICTABLE(i2 != zero) {
232 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
233 }
234 const int8_t* i3 = input[3];
235 assert(i3 != NULL);
236 if XNN_UNPREDICTABLE(i3 != zero) {
237 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
238 }
239 const int8_t* i4 = input[4];
240 assert(i4 != NULL);
241 if XNN_UNPREDICTABLE(i4 != zero) {
242 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
243 }
244 const int8_t* i5 = input[5];
245 assert(i5 != NULL);
246 if XNN_UNPREDICTABLE(i5 != zero) {
247 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
248 }
249 const int8_t* i6 = input[6];
250 assert(i6 != NULL);
251 if XNN_UNPREDICTABLE(i6 != zero) {
252 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
253 }
254
255 size_t c = channels;
256 for (; c >= 8; c -= 8) {
257 int32x4_t vacc0123 = vld1q_s32(b); b += 4;
258 int32x4_t vacc4567 = vld1q_s32(b); b += 4;
259
260 const int16x8_t vi0x01234567 = vmovl_s8(vld1_s8(i0)); i0 += 8;
261 const int16x8_t vk0x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
262
263 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
264 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
265
266 const int16x8_t vi1x01234567 = vmovl_s8(vld1_s8(i1)); i1 += 8;
267 const int16x8_t vk1x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
268
269 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
270 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
271
272 const int16x8_t vi2x01234567 = vmovl_s8(vld1_s8(i2)); i2 += 8;
273 const int16x8_t vk2x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
274
275 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
276 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
277
278 const int16x8_t vi3x01234567 = vmovl_s8(vld1_s8(i3)); i3 += 8;
279 const int16x8_t vk3x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
280
281 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
282 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
283
284 const int16x8_t vi4x01234567 = vmovl_s8(vld1_s8(i4)); i4 += 8;
285 const int16x8_t vk4x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
286
287 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
288 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
289
290 const int16x8_t vi5x01234567 = vmovl_s8(vld1_s8(i5)); i5 += 8;
291 const int16x8_t vk5x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
292
293 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi5x01234567), vget_low_s16(vk5x01234567));
294 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi5x01234567), vget_high_s16(vk5x01234567));
295
296 const int16x8_t vi6x01234567 = vmovl_s8(vld1_s8(i6)); i6 += 8;
297 const int16x8_t vk6x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
298
299 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi6x01234567), vget_low_s16(vk6x01234567));
300 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi6x01234567), vget_high_s16(vk6x01234567));
301
302 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
303 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
304
305 vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
306 vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
307
308 vacc0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
309 vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
310
311 vacc0123 = vqsubq_s32(vacc0123, vmagic_bias_less_output_zero_point);
312 vacc4567 = vqsubq_s32(vacc4567, vmagic_bias_less_output_zero_point);
313
314 #if XNN_ARCH_ARM64
315 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
316
317
318 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
319 #else // !XNN_ARCH_ARM64
320 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
321
322
323 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
324 #endif // !XNN_ARCH_ARM64
325
326 vout01234567 = vmax_s8(vout01234567, voutput_min);
327
328 vout01234567 = vmin_s8(vout01234567, voutput_max);
329
330 vst1_s8(output, vout01234567); output += 8;
331 }
332
333 if XNN_UNLIKELY(c != 0) {
334 {
335 int32x4_t vacc0123 = vld1q_s32(b); b += 4;
336 int32x4_t vacc4567 = vld1q_s32(b); b += 4;
337
338 const int16x8_t vi0x01234567 = vmovl_s8(vld1_s8(i0));
339 const int16x8_t vk0x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
340
341 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
342 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
343
344 const int16x8_t vi1x01234567 = vmovl_s8(vld1_s8(i1));
345 const int16x8_t vk1x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
346
347 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
348 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
349
350 const int16x8_t vi2x01234567 = vmovl_s8(vld1_s8(i2));
351 const int16x8_t vk2x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
352
353 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
354 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
355
356 const int16x8_t vi3x01234567 = vmovl_s8(vld1_s8(i3));
357 const int16x8_t vk3x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
358
359 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
360 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
361
362 const int16x8_t vi4x01234567 = vmovl_s8(vld1_s8(i4));
363 const int16x8_t vk4x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
364
365 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
366 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
367
368 const int16x8_t vi5x01234567 = vmovl_s8(vld1_s8(i5));
369 const int16x8_t vk5x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
370
371 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi5x01234567), vget_low_s16(vk5x01234567));
372 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi5x01234567), vget_high_s16(vk5x01234567));
373
374 const int16x8_t vi6x01234567 = vmovl_s8(vld1_s8(i6));
375 const int16x8_t vk6x01234567 = vmovl_s8(vld1_s8(w)); w = (const int8_t*) w + 8;
376
377 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi6x01234567), vget_low_s16(vk6x01234567));
378 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi6x01234567), vget_high_s16(vk6x01234567));
379
380 float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
381 float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
382
383 vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
384 vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
385
386 vacc0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
387 vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
388
389 vacc0123 = vqsubq_s32(vacc0123, vmagic_bias_less_output_zero_point);
390 vacc4567 = vqsubq_s32(vacc4567, vmagic_bias_less_output_zero_point);
391
392 #if XNN_ARCH_ARM64
393 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
394 #else
395 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
396 #endif
397
398 int8x8_t vout01234567 = vqmovn_s16(vacc01234567);
399 vout01234567 = vmax_s8(vout01234567, voutput_min);
400 vout01234567 = vmin_s8(vout01234567, voutput_max);
401
402 if (c & 4) {
403 vst1_lane_u32((void*) output, vreinterpret_u32_s8(vout01234567), 0); output += 4;
404 vout01234567 = vext_s8(vout01234567, vout01234567, 4);
405 }
406 if (c & 2) {
407 vst1_lane_u16((void*) output, vreinterpret_u16_s8(vout01234567), 0); output += 2;
408 vout01234567 = vext_s8(vout01234567, vout01234567, 2);
409 }
410 if (c & 1) {
411 vst1_lane_s8(output, vout01234567, 0); output += 1;
412 }
413 }
414 }
415 }
416
417 input = (const int8_t**) ((uintptr_t) input + input_stride);
418 output = (int8_t*) ((uintptr_t) output + output_increment);
419 } while (--output_width != 0);
420 }