Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-6f6m7l2c1s1r-minmax-fp32-scalar-imagic.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/multipass-scalar.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 #include <assert.h>
11 #include <stddef.h>
12 #include <stdint.h>
13
14 #include "xnnpack/common.h"
15 #include "xnnpack/dwconv.h"
16 #include "xnnpack/math.h"
17 #include "xnnpack/microparams.h"
18 #include "xnnpack/unaligned.h"
19
20
21 void xnn_qs8_dwconv_minmax_fp32_ukernel_6f6m7l2c1s1r__scalar_imagic(
22 size_t channels,
23 size_t output_width,
24 const int8_t** input,
25 const void* weights,
26 int8_t* output,
27 intptr_t input_stride,
28 size_t output_increment,
29 size_t input_offset,
30 const int8_t* zero,
31 size_t kernel_size,
32 int32_t* buffer,
33 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
34 {
35 assert(channels != 0);
36 assert(output_width != 0);
37 assert(kernel_size > 6);
38
39 const float vscale = params->fp32_scalar.scale;
40 const float vmagic_bias = 12582912.0f;
41 const int32_t output_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point;
42 const int32_t output_max_less_zero_point = (int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point;
43 const int32_t vmagic_min = (int32_t) float_as_uint32(12582912.0f + output_min_less_zero_point);
44 const int32_t vmagic_max = (int32_t) float_as_uint32(12582912.0f + output_max_less_zero_point);
45 const int32_t vmagic_bias_less_zero_point = INT32_C(0x4B400000) - (int32_t) params->fp32_scalar.output_zero_point;
46 do {
47 const void* w = weights;
48
49 // First pass to process 6 inputs.
50 {
51 int32_t* b = buffer;
52 const int8_t* i0 = input[0];
53 assert(i0 != NULL);
54 if XNN_UNPREDICTABLE(i0 != zero) {
55 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
56 }
57 const int8_t* i1 = input[1];
58 assert(i1 != NULL);
59 if XNN_UNPREDICTABLE(i1 != zero) {
60 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
61 }
62 const int8_t* i2 = input[2];
63 assert(i2 != NULL);
64 if XNN_UNPREDICTABLE(i2 != zero) {
65 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
66 }
67 const int8_t* i3 = input[3];
68 assert(i3 != NULL);
69 if XNN_UNPREDICTABLE(i3 != zero) {
70 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
71 }
72 const int8_t* i4 = input[4];
73 assert(i4 != NULL);
74 if XNN_UNPREDICTABLE(i4 != zero) {
75 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
76 }
77 const int8_t* i5 = input[5];
78 assert(i5 != NULL);
79 if XNN_UNPREDICTABLE(i5 != zero) {
80 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
81 }
82 input += 6;
83
84 size_t c = channels;
85 for (; c >= 2; c -= 2) {
86 int32_t vacc0 = unaligned_indexed_load_s32(w, 0);
87 int32_t vacc1 = unaligned_indexed_load_s32(w, 1);
88
89 const int32_t vi0x0 = (int32_t) i0[0];
90 const int32_t vi0x1 = (int32_t) i0[1];
91 i0 += 2;
92
93 const int32_t vk0x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[0];
94 const int32_t vk0x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[1];
95
96 vacc0 += vi0x0 * vk0x0;
97 vacc1 += vi0x1 * vk0x1;
98
99 const int32_t vi1x0 = (int32_t) i1[0];
100 const int32_t vi1x1 = (int32_t) i1[1];
101 i1 += 2;
102
103 const int32_t vk1x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[2];
104 const int32_t vk1x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[3];
105
106 vacc0 += vi1x0 * vk1x0;
107 vacc1 += vi1x1 * vk1x1;
108
109 const int32_t vi2x0 = (int32_t) i2[0];
110 const int32_t vi2x1 = (int32_t) i2[1];
111 i2 += 2;
112
113 const int32_t vk2x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[4];
114 const int32_t vk2x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[5];
115
116 vacc0 += vi2x0 * vk2x0;
117 vacc1 += vi2x1 * vk2x1;
118
119 const int32_t vi3x0 = (int32_t) i3[0];
120 const int32_t vi3x1 = (int32_t) i3[1];
121 i3 += 2;
122
123 const int32_t vk3x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[6];
124 const int32_t vk3x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[7];
125
126 vacc0 += vi3x0 * vk3x0;
127 vacc1 += vi3x1 * vk3x1;
128
129 const int32_t vi4x0 = (int32_t) i4[0];
130 const int32_t vi4x1 = (int32_t) i4[1];
131 i4 += 2;
132
133 const int32_t vk4x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[8];
134 const int32_t vk4x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[9];
135
136 vacc0 += vi4x0 * vk4x0;
137 vacc1 += vi4x1 * vk4x1;
138
139 const int32_t vi5x0 = (int32_t) i5[0];
140 const int32_t vi5x1 = (int32_t) i5[1];
141 i5 += 2;
142
143 const int32_t vk5x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[10];
144 const int32_t vk5x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[11];
145
146 vacc0 += vi5x0 * vk5x0;
147 vacc1 += vi5x1 * vk5x1;
148
149 w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t) + 12 * sizeof(int8_t));
150 b[0] = vacc0;
151 b[1] = vacc1;
152 b += 2;
153 }
154 if XNN_UNLIKELY(c != 0) {
155 int32_t vacc = unaligned_load_s32(w);
156
157 const int32_t vi0 = (int32_t) *i0;
158 const int32_t vk0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 1 * sizeof(int32_t)))[0];
159 vacc += vi0 * vk0;
160 const int32_t vi1 = (int32_t) *i1;
161 const int32_t vk1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 1 * sizeof(int32_t)))[1];
162 vacc += vi1 * vk1;
163 const int32_t vi2 = (int32_t) *i2;
164 const int32_t vk2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 1 * sizeof(int32_t)))[2];
165 vacc += vi2 * vk2;
166 const int32_t vi3 = (int32_t) *i3;
167 const int32_t vk3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 1 * sizeof(int32_t)))[3];
168 vacc += vi3 * vk3;
169 const int32_t vi4 = (int32_t) *i4;
170 const int32_t vk4 = (int32_t) ((const int8_t*) ((uintptr_t) w + 1 * sizeof(int32_t)))[4];
171 vacc += vi4 * vk4;
172 const int32_t vi5 = (int32_t) *i5;
173 const int32_t vk5 = (int32_t) ((const int8_t*) ((uintptr_t) w + 1 * sizeof(int32_t)))[5];
174 vacc += vi5 * vk5;
175
176 w = (const void*) ((uintptr_t) w + 1 * sizeof(int32_t) + 6 * sizeof(int8_t));
177 *b++ = vacc;
178 }
179 }
180
181 // Middle pass to process 6 inputs in each iteration.
182 for (size_t ks = kernel_size - 6; ks > 7; ks -= 6) {
183 int32_t* b = buffer;
184 const int8_t* i0 = input[0];
185 assert(i0 != NULL);
186 if XNN_UNPREDICTABLE(i0 != zero) {
187 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
188 }
189 const int8_t* i1 = input[1];
190 assert(i1 != NULL);
191 if XNN_UNPREDICTABLE(i1 != zero) {
192 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
193 }
194 const int8_t* i2 = input[2];
195 assert(i2 != NULL);
196 if XNN_UNPREDICTABLE(i2 != zero) {
197 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
198 }
199 const int8_t* i3 = input[3];
200 assert(i3 != NULL);
201 if XNN_UNPREDICTABLE(i3 != zero) {
202 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
203 }
204 const int8_t* i4 = input[4];
205 assert(i4 != NULL);
206 if XNN_UNPREDICTABLE(i4 != zero) {
207 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
208 }
209 const int8_t* i5 = input[5];
210 assert(i5 != NULL);
211 if XNN_UNPREDICTABLE(i5 != zero) {
212 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
213 }
214 input += 6;
215
216 size_t c = channels;
217 for (; c >= 2; c -= 2) {
218 int32_t vacc0 = b[0];
219 int32_t vacc1 = b[1];
220
221 const int32_t vi0x0 = (int32_t) i0[0];
222 const int32_t vi0x1 = (int32_t) i0[1];
223 i0 += 2;
224
225 const int32_t vk0x0 = (int32_t) ((const int8_t*) w)[0];
226 const int32_t vk0x1 = (int32_t) ((const int8_t*) w)[1];
227
228 vacc0 += vi0x0 * vk0x0;
229 vacc1 += vi0x1 * vk0x1;
230
231 const int32_t vi1x0 = (int32_t) i1[0];
232 const int32_t vi1x1 = (int32_t) i1[1];
233 i1 += 2;
234
235 const int32_t vk1x0 = (int32_t) ((const int8_t*) w)[2];
236 const int32_t vk1x1 = (int32_t) ((const int8_t*) w)[3];
237
238 vacc0 += vi1x0 * vk1x0;
239 vacc1 += vi1x1 * vk1x1;
240
241 const int32_t vi2x0 = (int32_t) i2[0];
242 const int32_t vi2x1 = (int32_t) i2[1];
243 i2 += 2;
244
245 const int32_t vk2x0 = (int32_t) ((const int8_t*) w)[4];
246 const int32_t vk2x1 = (int32_t) ((const int8_t*) w)[5];
247
248 vacc0 += vi2x0 * vk2x0;
249 vacc1 += vi2x1 * vk2x1;
250
251 const int32_t vi3x0 = (int32_t) i3[0];
252 const int32_t vi3x1 = (int32_t) i3[1];
253 i3 += 2;
254
255 const int32_t vk3x0 = (int32_t) ((const int8_t*) w)[6];
256 const int32_t vk3x1 = (int32_t) ((const int8_t*) w)[7];
257
258 vacc0 += vi3x0 * vk3x0;
259 vacc1 += vi3x1 * vk3x1;
260
261 const int32_t vi4x0 = (int32_t) i4[0];
262 const int32_t vi4x1 = (int32_t) i4[1];
263 i4 += 2;
264
265 const int32_t vk4x0 = (int32_t) ((const int8_t*) w)[8];
266 const int32_t vk4x1 = (int32_t) ((const int8_t*) w)[9];
267
268 vacc0 += vi4x0 * vk4x0;
269 vacc1 += vi4x1 * vk4x1;
270
271 const int32_t vi5x0 = (int32_t) i5[0];
272 const int32_t vi5x1 = (int32_t) i5[1];
273 i5 += 2;
274
275 const int32_t vk5x0 = (int32_t) ((const int8_t*) w)[10];
276 const int32_t vk5x1 = (int32_t) ((const int8_t*) w)[11];
277
278 vacc0 += vi5x0 * vk5x0;
279 vacc1 += vi5x1 * vk5x1;
280
281 w = (const void*) ((uintptr_t) w + 12 * sizeof(int8_t));
282 b[0] = vacc0;
283 b[1] = vacc1;
284 b += 2;
285 }
286 if XNN_UNLIKELY(c != 0) {
287 int32_t vacc = b[0];
288
289 const int32_t vi0 = (int32_t) *i0;
290 const int32_t vk0 = (int32_t) ((const int8_t*) w)[0];
291 vacc += vi0 * vk0;
292 const int32_t vi1 = (int32_t) *i1;
293 const int32_t vk1 = (int32_t) ((const int8_t*) w)[1];
294 vacc += vi1 * vk1;
295 const int32_t vi2 = (int32_t) *i2;
296 const int32_t vk2 = (int32_t) ((const int8_t*) w)[2];
297 vacc += vi2 * vk2;
298 const int32_t vi3 = (int32_t) *i3;
299 const int32_t vk3 = (int32_t) ((const int8_t*) w)[3];
300 vacc += vi3 * vk3;
301 const int32_t vi4 = (int32_t) *i4;
302 const int32_t vk4 = (int32_t) ((const int8_t*) w)[4];
303 vacc += vi4 * vk4;
304 const int32_t vi5 = (int32_t) *i5;
305 const int32_t vk5 = (int32_t) ((const int8_t*) w)[5];
306 vacc += vi5 * vk5;
307 w = (const void*) ((uintptr_t) w + 6 * sizeof(int8_t));
308 *b++ = vacc;
309 }
310 }
311
312 // Last pass to process up to 7 inputs.
313 {
314 const int32_t* b = buffer;
315 const int8_t* i0 = input[0];
316 assert(i0 != NULL);
317 if XNN_UNPREDICTABLE(i0 != zero) {
318 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
319 }
320 const int8_t* i1 = input[1];
321 assert(i1 != NULL);
322 if XNN_UNPREDICTABLE(i1 != zero) {
323 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
324 }
325 const int8_t* i2 = input[2];
326 assert(i2 != NULL);
327 if XNN_UNPREDICTABLE(i2 != zero) {
328 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
329 }
330 const int8_t* i3 = input[3];
331 assert(i3 != NULL);
332 if XNN_UNPREDICTABLE(i3 != zero) {
333 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
334 }
335 const int8_t* i4 = input[4];
336 assert(i4 != NULL);
337 if XNN_UNPREDICTABLE(i4 != zero) {
338 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
339 }
340 const int8_t* i5 = input[5];
341 assert(i5 != NULL);
342 if XNN_UNPREDICTABLE(i5 != zero) {
343 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
344 }
345 const int8_t* i6 = input[6];
346 assert(i6 != NULL);
347 if XNN_UNPREDICTABLE(i6 != zero) {
348 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
349 }
350
351 size_t c = channels;
352 for (; c >= 2; c -= 2) {
353 int32_t vacc0 = b[0];
354 int32_t vacc1 = b[1];
355 b += 2;
356
357 const int32_t vi0x0 = (int32_t) i0[0];
358 const int32_t vi0x1 = (int32_t) i0[1];
359 i0 += 2;
360
361 const int32_t vk0x0 = (int32_t) ((const int8_t*) w)[0];
362 const int32_t vk0x1 = (int32_t) ((const int8_t*) w)[1];
363
364 vacc0 += vi0x0 * vk0x0;
365 vacc1 += vi0x1 * vk0x1;
366
367 const int32_t vi1x0 = (int32_t) i1[0];
368 const int32_t vi1x1 = (int32_t) i1[1];
369 i1 += 2;
370
371 const int32_t vk1x0 = (int32_t) ((const int8_t*) w)[2];
372 const int32_t vk1x1 = (int32_t) ((const int8_t*) w)[3];
373
374 vacc0 += vi1x0 * vk1x0;
375 vacc1 += vi1x1 * vk1x1;
376
377 const int32_t vi2x0 = (int32_t) i2[0];
378 const int32_t vi2x1 = (int32_t) i2[1];
379 i2 += 2;
380
381 const int32_t vk2x0 = (int32_t) ((const int8_t*) w)[4];
382 const int32_t vk2x1 = (int32_t) ((const int8_t*) w)[5];
383
384 vacc0 += vi2x0 * vk2x0;
385 vacc1 += vi2x1 * vk2x1;
386
387 const int32_t vi3x0 = (int32_t) i3[0];
388 const int32_t vi3x1 = (int32_t) i3[1];
389 i3 += 2;
390
391 const int32_t vk3x0 = (int32_t) ((const int8_t*) w)[6];
392 const int32_t vk3x1 = (int32_t) ((const int8_t*) w)[7];
393
394 vacc0 += vi3x0 * vk3x0;
395 vacc1 += vi3x1 * vk3x1;
396
397 const int32_t vi4x0 = (int32_t) i4[0];
398 const int32_t vi4x1 = (int32_t) i4[1];
399 i4 += 2;
400
401 const int32_t vk4x0 = (int32_t) ((const int8_t*) w)[8];
402 const int32_t vk4x1 = (int32_t) ((const int8_t*) w)[9];
403
404 vacc0 += vi4x0 * vk4x0;
405 vacc1 += vi4x1 * vk4x1;
406
407 const int32_t vi5x0 = (int32_t) i5[0];
408 const int32_t vi5x1 = (int32_t) i5[1];
409 i5 += 2;
410
411 const int32_t vk5x0 = (int32_t) ((const int8_t*) w)[10];
412 const int32_t vk5x1 = (int32_t) ((const int8_t*) w)[11];
413
414 vacc0 += vi5x0 * vk5x0;
415 vacc1 += vi5x1 * vk5x1;
416
417 const int32_t vi6x0 = (int32_t) i6[0];
418 const int32_t vi6x1 = (int32_t) i6[1];
419 i6 += 2;
420
421 const int32_t vk6x0 = (int32_t) ((const int8_t*) w)[12];
422 const int32_t vk6x1 = (int32_t) ((const int8_t*) w)[13];
423
424 vacc0 += vi6x0 * vk6x0;
425 vacc1 += vi6x1 * vk6x1;
426
427 w = (const void*) ((uintptr_t) w + 14 * sizeof(int8_t));
428
429 float vfpacc0 = (float) vacc0;
430 float vfpacc1 = (float) vacc1;
431
432 vfpacc0 *= vscale;
433 vfpacc1 *= vscale;
434
435 vfpacc0 += vmagic_bias;
436 vfpacc1 += vmagic_bias;
437
438 int32_t vout0 = (int32_t) float_as_uint32(vfpacc0);
439 int32_t vout1 = (int32_t) float_as_uint32(vfpacc1);
440
441 vout0 = math_max_s32(vout0, vmagic_min);
442 vout1 = math_max_s32(vout1, vmagic_min);
443
444 vout0 = math_min_s32(vout0, vmagic_max);
445 vout1 = math_min_s32(vout1, vmagic_max);
446
447 vout0 -= vmagic_bias_less_zero_point;
448 vout1 -= vmagic_bias_less_zero_point;
449
450 output[0] = (int8_t) vout0;
451 output[1] = (int8_t) vout1;
452 output += 2;
453 }
454 if XNN_UNLIKELY(c != 0) {
455 int32_t vacc = b[0];
456
457 const int32_t vi0 = (int32_t) *i0;
458 const int32_t vk0 = (int32_t) ((const int8_t*) w)[0];
459 vacc += vi0 * vk0;
460 const int32_t vi1 = (int32_t) *i1;
461 const int32_t vk1 = (int32_t) ((const int8_t*) w)[1];
462 vacc += vi1 * vk1;
463 const int32_t vi2 = (int32_t) *i2;
464 const int32_t vk2 = (int32_t) ((const int8_t*) w)[2];
465 vacc += vi2 * vk2;
466 const int32_t vi3 = (int32_t) *i3;
467 const int32_t vk3 = (int32_t) ((const int8_t*) w)[3];
468 vacc += vi3 * vk3;
469 const int32_t vi4 = (int32_t) *i4;
470 const int32_t vk4 = (int32_t) ((const int8_t*) w)[4];
471 vacc += vi4 * vk4;
472 const int32_t vi5 = (int32_t) *i5;
473 const int32_t vk5 = (int32_t) ((const int8_t*) w)[5];
474 vacc += vi5 * vk5;
475 const int32_t vi6 = (int32_t) *i6;
476 const int32_t vk6 = (int32_t) ((const int8_t*) w)[6];
477 vacc += vi6 * vk6;
478
479 float vfpacc = (float) vacc * vscale;
480
481 vfpacc += vmagic_bias;
482 int32_t vout = (int32_t) float_as_uint32(vfpacc);
483 vout = math_max_s32(vout, vmagic_min);
484 vout = math_min_s32(vout, vmagic_max);
485 vout -= vmagic_bias_less_zero_point;
486
487 *output++ = (int8_t) vout;
488 }
489 }
490
491 input = (const int8_t**) ((uintptr_t) input + input_stride);
492 output = (int8_t*) ((uintptr_t) output + output_increment);
493 } while (--output_width != 0);
494 }