Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-9p4c-minmax-fp32-scalar-imagic.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-scalar.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11 #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
19 void xnn_qs8_dwconv_minmax_fp32_ukernel_9p4c__scalar_imagic(
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 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
30 {
31 assert(channels != 0);
32 assert(output_width != 0);
33
34 const float vscale = params->fp32_scalar.scale;
35 const float vmagic_bias = 12582912.0f;
36 const int32_t output_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point;
37 const int32_t output_max_less_zero_point = (int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point;
38 const int32_t vmagic_min = (int32_t) float_as_uint32(12582912.0f + output_min_less_zero_point);
39 const int32_t vmagic_max = (int32_t) float_as_uint32(12582912.0f + output_max_less_zero_point);
40 const int32_t vmagic_bias_less_zero_point = INT32_C(0x4B400000) - (int32_t) params->fp32_scalar.output_zero_point;
41 do {
42 const int8_t* i0 = input[0];
43 assert(i0 != NULL);
44 if XNN_UNPREDICTABLE(i0 != zero) {
45 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
46 }
47 const int8_t* i1 = input[1];
48 assert(i1 != NULL);
49 if XNN_UNPREDICTABLE(i1 != zero) {
50 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
51 }
52 const int8_t* i2 = input[2];
53 assert(i2 != NULL);
54 if XNN_UNPREDICTABLE(i2 != zero) {
55 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
56 }
57 const int8_t* i3 = input[3];
58 assert(i3 != NULL);
59 if XNN_UNPREDICTABLE(i3 != zero) {
60 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
61 }
62 const int8_t* i4 = input[4];
63 assert(i4 != NULL);
64 if XNN_UNPREDICTABLE(i4 != zero) {
65 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
66 }
67 const int8_t* i5 = input[5];
68 assert(i5 != NULL);
69 if XNN_UNPREDICTABLE(i5 != zero) {
70 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
71 }
72 const int8_t* i6 = input[6];
73 assert(i6 != NULL);
74 if XNN_UNPREDICTABLE(i6 != zero) {
75 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
76 }
77 const int8_t* i7 = input[7];
78 assert(i7 != NULL);
79 if XNN_UNPREDICTABLE(i7 != zero) {
80 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
81 }
82 const int8_t* i8 = input[8];
83 assert(i8 != NULL);
84 if XNN_UNPREDICTABLE(i8 != zero) {
85 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
86 }
87 input = (const int8_t**) ((uintptr_t) input + input_stride);
88
89 size_t c = channels;
90 const void* w = weights;
91 for (; c >= 4; c -= 4) {
92 int32_t vacc0 = ((const int32_t*) w)[0];
93 int32_t vacc1 = ((const int32_t*) w)[1];
94 int32_t vacc2 = ((const int32_t*) w)[2];
95 int32_t vacc3 = ((const int32_t*) w)[3];
96
97
98 const int32_t vi0x0 = (int32_t) i0[0];
99 const int32_t vi0x1 = (int32_t) i0[1];
100 const int32_t vi0x2 = (int32_t) i0[2];
101 const int32_t vi0x3 = (int32_t) i0[3];
102 i0 += 4;
103
104 const int32_t vk0x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[0];
105 const int32_t vk0x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[1];
106 const int32_t vk0x2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[2];
107 const int32_t vk0x3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[3];
108
109 vacc0 += vi0x0 * vk0x0;
110 vacc1 += vi0x1 * vk0x1;
111 vacc2 += vi0x2 * vk0x2;
112 vacc3 += vi0x3 * vk0x3;
113
114 const int32_t vi1x0 = (int32_t) i1[0];
115 const int32_t vi1x1 = (int32_t) i1[1];
116 const int32_t vi1x2 = (int32_t) i1[2];
117 const int32_t vi1x3 = (int32_t) i1[3];
118 i1 += 4;
119
120 const int32_t vk1x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[4];
121 const int32_t vk1x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[5];
122 const int32_t vk1x2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[6];
123 const int32_t vk1x3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[7];
124
125 vacc0 += vi1x0 * vk1x0;
126 vacc1 += vi1x1 * vk1x1;
127 vacc2 += vi1x2 * vk1x2;
128 vacc3 += vi1x3 * vk1x3;
129
130 const int32_t vi2x0 = (int32_t) i2[0];
131 const int32_t vi2x1 = (int32_t) i2[1];
132 const int32_t vi2x2 = (int32_t) i2[2];
133 const int32_t vi2x3 = (int32_t) i2[3];
134 i2 += 4;
135
136 const int32_t vk2x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[8];
137 const int32_t vk2x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[9];
138 const int32_t vk2x2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[10];
139 const int32_t vk2x3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[11];
140
141 vacc0 += vi2x0 * vk2x0;
142 vacc1 += vi2x1 * vk2x1;
143 vacc2 += vi2x2 * vk2x2;
144 vacc3 += vi2x3 * vk2x3;
145
146 const int32_t vi3x0 = (int32_t) i3[0];
147 const int32_t vi3x1 = (int32_t) i3[1];
148 const int32_t vi3x2 = (int32_t) i3[2];
149 const int32_t vi3x3 = (int32_t) i3[3];
150 i3 += 4;
151
152 const int32_t vk3x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[12];
153 const int32_t vk3x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[13];
154 const int32_t vk3x2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[14];
155 const int32_t vk3x3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[15];
156
157 vacc0 += vi3x0 * vk3x0;
158 vacc1 += vi3x1 * vk3x1;
159 vacc2 += vi3x2 * vk3x2;
160 vacc3 += vi3x3 * vk3x3;
161
162 const int32_t vi4x0 = (int32_t) i4[0];
163 const int32_t vi4x1 = (int32_t) i4[1];
164 const int32_t vi4x2 = (int32_t) i4[2];
165 const int32_t vi4x3 = (int32_t) i4[3];
166 i4 += 4;
167
168 const int32_t vk4x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[16];
169 const int32_t vk4x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[17];
170 const int32_t vk4x2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[18];
171 const int32_t vk4x3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[19];
172
173 vacc0 += vi4x0 * vk4x0;
174 vacc1 += vi4x1 * vk4x1;
175 vacc2 += vi4x2 * vk4x2;
176 vacc3 += vi4x3 * vk4x3;
177
178 const int32_t vi5x0 = (int32_t) i5[0];
179 const int32_t vi5x1 = (int32_t) i5[1];
180 const int32_t vi5x2 = (int32_t) i5[2];
181 const int32_t vi5x3 = (int32_t) i5[3];
182 i5 += 4;
183
184 const int32_t vk5x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[20];
185 const int32_t vk5x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[21];
186 const int32_t vk5x2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[22];
187 const int32_t vk5x3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[23];
188
189 vacc0 += vi5x0 * vk5x0;
190 vacc1 += vi5x1 * vk5x1;
191 vacc2 += vi5x2 * vk5x2;
192 vacc3 += vi5x3 * vk5x3;
193
194 const int32_t vi6x0 = (int32_t) i6[0];
195 const int32_t vi6x1 = (int32_t) i6[1];
196 const int32_t vi6x2 = (int32_t) i6[2];
197 const int32_t vi6x3 = (int32_t) i6[3];
198 i6 += 4;
199
200 const int32_t vk6x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[24];
201 const int32_t vk6x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[25];
202 const int32_t vk6x2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[26];
203 const int32_t vk6x3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[27];
204
205 vacc0 += vi6x0 * vk6x0;
206 vacc1 += vi6x1 * vk6x1;
207 vacc2 += vi6x2 * vk6x2;
208 vacc3 += vi6x3 * vk6x3;
209
210 const int32_t vi7x0 = (int32_t) i7[0];
211 const int32_t vi7x1 = (int32_t) i7[1];
212 const int32_t vi7x2 = (int32_t) i7[2];
213 const int32_t vi7x3 = (int32_t) i7[3];
214 i7 += 4;
215
216 const int32_t vk7x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[28];
217 const int32_t vk7x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[29];
218 const int32_t vk7x2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[30];
219 const int32_t vk7x3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[31];
220
221 vacc0 += vi7x0 * vk7x0;
222 vacc1 += vi7x1 * vk7x1;
223 vacc2 += vi7x2 * vk7x2;
224 vacc3 += vi7x3 * vk7x3;
225
226 const int32_t vi8x0 = (int32_t) i8[0];
227 const int32_t vi8x1 = (int32_t) i8[1];
228 const int32_t vi8x2 = (int32_t) i8[2];
229 const int32_t vi8x3 = (int32_t) i8[3];
230 i8 += 4;
231
232 const int32_t vk8x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[32];
233 const int32_t vk8x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[33];
234 const int32_t vk8x2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[34];
235 const int32_t vk8x3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[35];
236
237 vacc0 += vi8x0 * vk8x0;
238 vacc1 += vi8x1 * vk8x1;
239 vacc2 += vi8x2 * vk8x2;
240 vacc3 += vi8x3 * vk8x3;
241
242 w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t) + 36 * sizeof(int8_t));
243
244 float vfpacc0 = (float) vacc0;
245 float vfpacc1 = (float) vacc1;
246 float vfpacc2 = (float) vacc2;
247 float vfpacc3 = (float) vacc3;
248
249 vfpacc0 *= vscale;
250 vfpacc1 *= vscale;
251 vfpacc2 *= vscale;
252 vfpacc3 *= vscale;
253
254 vfpacc0 += vmagic_bias;
255 vfpacc1 += vmagic_bias;
256 vfpacc2 += vmagic_bias;
257 vfpacc3 += vmagic_bias;
258
259 int32_t vout0 = (int32_t) float_as_uint32(vfpacc0);
260 int32_t vout1 = (int32_t) float_as_uint32(vfpacc1);
261 int32_t vout2 = (int32_t) float_as_uint32(vfpacc2);
262 int32_t vout3 = (int32_t) float_as_uint32(vfpacc3);
263
264 vout0 = math_max_s32(vout0, vmagic_min);
265 vout1 = math_max_s32(vout1, vmagic_min);
266 vout2 = math_max_s32(vout2, vmagic_min);
267 vout3 = math_max_s32(vout3, vmagic_min);
268
269 vout0 = math_min_s32(vout0, vmagic_max);
270 vout1 = math_min_s32(vout1, vmagic_max);
271 vout2 = math_min_s32(vout2, vmagic_max);
272 vout3 = math_min_s32(vout3, vmagic_max);
273
274 vout0 -= vmagic_bias_less_zero_point;
275 vout1 -= vmagic_bias_less_zero_point;
276 vout2 -= vmagic_bias_less_zero_point;
277 vout3 -= vmagic_bias_less_zero_point;
278
279 output[0] = (int8_t) vout0;
280 output[1] = (int8_t) vout1;
281 output[2] = (int8_t) vout2;
282 output[3] = (int8_t) vout3;
283 output += 4;
284 }
285 if XNN_UNLIKELY(c != 0) {
286 const int8_t* k = (const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t));
287 do {
288 int32_t vacc = *((const int32_t*) w);
289 w = (const int32_t*) w + 1;
290
291 const int32_t vi0 = (int32_t) *i0++;
292 const int32_t vk0 = (int32_t) k[0];
293 vacc += vi0 * vk0;
294 const int32_t vi1 = (int32_t) *i1++;
295 const int32_t vk1 = (int32_t) k[4];
296 vacc += vi1 * vk1;
297 const int32_t vi2 = (int32_t) *i2++;
298 const int32_t vk2 = (int32_t) k[8];
299 vacc += vi2 * vk2;
300 const int32_t vi3 = (int32_t) *i3++;
301 const int32_t vk3 = (int32_t) k[12];
302 vacc += vi3 * vk3;
303 const int32_t vi4 = (int32_t) *i4++;
304 const int32_t vk4 = (int32_t) k[16];
305 vacc += vi4 * vk4;
306 const int32_t vi5 = (int32_t) *i5++;
307 const int32_t vk5 = (int32_t) k[20];
308 vacc += vi5 * vk5;
309 const int32_t vi6 = (int32_t) *i6++;
310 const int32_t vk6 = (int32_t) k[24];
311 vacc += vi6 * vk6;
312 const int32_t vi7 = (int32_t) *i7++;
313 const int32_t vk7 = (int32_t) k[28];
314 vacc += vi7 * vk7;
315 const int32_t vi8 = (int32_t) *i8++;
316 const int32_t vk8 = (int32_t) k[32];
317 vacc += vi8 * vk8;
318 k += 1;
319
320 float vfpacc = (float) vacc * vscale;
321
322 vfpacc += vmagic_bias;
323 int32_t vout = (int32_t) float_as_uint32(vfpacc);
324 vout = math_max_s32(vout, vmagic_min);
325 vout = math_min_s32(vout, vmagic_max);
326 vout -= vmagic_bias_less_zero_point;
327
328 *output++ = (int8_t) vout;
329 } while (--c != 0);
330 }
331
332 output = (int8_t*) ((uintptr_t) output + output_increment);
333 } while (--output_width != 0);
334 }