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
12 #include <fp16.h>
13
14 #include <xnnpack/dwconv.h>
15 #include <xnnpack/math.h>
16
17
xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic(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_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic(
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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
29 {
30 assert(channels != 0);
31 assert(output_width != 0);
32
33 const float vmagic_bias = params->scalar_imagic.magic_bias;
34 const int32_t vmagic_min = params->scalar_imagic.magic_min;
35 const int32_t vmagic_max = params->scalar_imagic.magic_max;
36 const int32_t vmagic_bias_less_zero_point = params->scalar_imagic.magic_bias_less_zero_point;
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 input = (const int8_t**) ((uintptr_t) input + input_stride);
84
85 size_t c = channels;
86 const void* w = weights;
87 for (; c >= 4; c -= 4) {
88 int32_t vacc0 = ((const int32_t*) w)[0];
89 int32_t vacc1 = ((const int32_t*) w)[1];
90 int32_t vacc2 = ((const int32_t*) w)[2];
91 int32_t vacc3 = ((const int32_t*) w)[3];
92
93
94 const int32_t vi0x0 = (int32_t) i0[0];
95 const int32_t vi0x1 = (int32_t) i0[1];
96 const int32_t vi0x2 = (int32_t) i0[2];
97 const int32_t vi0x3 = (int32_t) i0[3];
98 i0 += 4;
99
100 const int32_t vk0x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[0];
101 const int32_t vk0x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[1];
102 const int32_t vk0x2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[2];
103 const int32_t vk0x3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[3];
104
105 vacc0 += vi0x0 * vk0x0;
106 vacc1 += vi0x1 * vk0x1;
107 vacc2 += vi0x2 * vk0x2;
108 vacc3 += vi0x3 * vk0x3;
109
110 const int32_t vi1x0 = (int32_t) i1[0];
111 const int32_t vi1x1 = (int32_t) i1[1];
112 const int32_t vi1x2 = (int32_t) i1[2];
113 const int32_t vi1x3 = (int32_t) i1[3];
114 i1 += 4;
115
116 const int32_t vk1x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[4];
117 const int32_t vk1x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[5];
118 const int32_t vk1x2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[6];
119 const int32_t vk1x3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[7];
120
121 vacc0 += vi1x0 * vk1x0;
122 vacc1 += vi1x1 * vk1x1;
123 vacc2 += vi1x2 * vk1x2;
124 vacc3 += vi1x3 * vk1x3;
125
126 const int32_t vi2x0 = (int32_t) i2[0];
127 const int32_t vi2x1 = (int32_t) i2[1];
128 const int32_t vi2x2 = (int32_t) i2[2];
129 const int32_t vi2x3 = (int32_t) i2[3];
130 i2 += 4;
131
132 const int32_t vk2x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[8];
133 const int32_t vk2x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[9];
134 const int32_t vk2x2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[10];
135 const int32_t vk2x3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[11];
136
137 vacc0 += vi2x0 * vk2x0;
138 vacc1 += vi2x1 * vk2x1;
139 vacc2 += vi2x2 * vk2x2;
140 vacc3 += vi2x3 * vk2x3;
141
142 const int32_t vi3x0 = (int32_t) i3[0];
143 const int32_t vi3x1 = (int32_t) i3[1];
144 const int32_t vi3x2 = (int32_t) i3[2];
145 const int32_t vi3x3 = (int32_t) i3[3];
146 i3 += 4;
147
148 const int32_t vk3x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[12];
149 const int32_t vk3x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[13];
150 const int32_t vk3x2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[14];
151 const int32_t vk3x3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[15];
152
153 vacc0 += vi3x0 * vk3x0;
154 vacc1 += vi3x1 * vk3x1;
155 vacc2 += vi3x2 * vk3x2;
156 vacc3 += vi3x3 * vk3x3;
157
158 const int32_t vi4x0 = (int32_t) i4[0];
159 const int32_t vi4x1 = (int32_t) i4[1];
160 const int32_t vi4x2 = (int32_t) i4[2];
161 const int32_t vi4x3 = (int32_t) i4[3];
162 i4 += 4;
163
164 const int32_t vk4x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[16];
165 const int32_t vk4x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[17];
166 const int32_t vk4x2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[18];
167 const int32_t vk4x3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[19];
168
169 vacc0 += vi4x0 * vk4x0;
170 vacc1 += vi4x1 * vk4x1;
171 vacc2 += vi4x2 * vk4x2;
172 vacc3 += vi4x3 * vk4x3;
173
174 const int32_t vi5x0 = (int32_t) i5[0];
175 const int32_t vi5x1 = (int32_t) i5[1];
176 const int32_t vi5x2 = (int32_t) i5[2];
177 const int32_t vi5x3 = (int32_t) i5[3];
178 i5 += 4;
179
180 const int32_t vk5x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[20];
181 const int32_t vk5x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[21];
182 const int32_t vk5x2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[22];
183 const int32_t vk5x3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[23];
184
185 vacc0 += vi5x0 * vk5x0;
186 vacc1 += vi5x1 * vk5x1;
187 vacc2 += vi5x2 * vk5x2;
188 vacc3 += vi5x3 * vk5x3;
189
190 const int32_t vi6x0 = (int32_t) i6[0];
191 const int32_t vi6x1 = (int32_t) i6[1];
192 const int32_t vi6x2 = (int32_t) i6[2];
193 const int32_t vi6x3 = (int32_t) i6[3];
194 i6 += 4;
195
196 const int32_t vk6x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[24];
197 const int32_t vk6x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[25];
198 const int32_t vk6x2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[26];
199 const int32_t vk6x3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[27];
200
201 vacc0 += vi6x0 * vk6x0;
202 vacc1 += vi6x1 * vk6x1;
203 vacc2 += vi6x2 * vk6x2;
204 vacc3 += vi6x3 * vk6x3;
205
206 const int32_t vi7x0 = (int32_t) i7[0];
207 const int32_t vi7x1 = (int32_t) i7[1];
208 const int32_t vi7x2 = (int32_t) i7[2];
209 const int32_t vi7x3 = (int32_t) i7[3];
210 i7 += 4;
211
212 const int32_t vk7x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[28];
213 const int32_t vk7x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[29];
214 const int32_t vk7x2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[30];
215 const int32_t vk7x3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[31];
216
217 vacc0 += vi7x0 * vk7x0;
218 vacc1 += vi7x1 * vk7x1;
219 vacc2 += vi7x2 * vk7x2;
220 vacc3 += vi7x3 * vk7x3;
221
222 const int32_t vi8x0 = (int32_t) i8[0];
223 const int32_t vi8x1 = (int32_t) i8[1];
224 const int32_t vi8x2 = (int32_t) i8[2];
225 const int32_t vi8x3 = (int32_t) i8[3];
226 i8 += 4;
227
228 const int32_t vk8x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[32];
229 const int32_t vk8x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[33];
230 const int32_t vk8x2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[34];
231 const int32_t vk8x3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[35];
232
233 vacc0 += vi8x0 * vk8x0;
234 vacc1 += vi8x1 * vk8x1;
235 vacc2 += vi8x2 * vk8x2;
236 vacc3 += vi8x3 * vk8x3;
237
238 w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t) + 36 * sizeof(int8_t));
239
240 float vfpacc0 = (float) vacc0;
241 float vfpacc1 = (float) vacc1;
242 float vfpacc2 = (float) vacc2;
243 float vfpacc3 = (float) vacc3;
244
245 const float vscale0 = ((const float*) w)[0];
246 const float vscale1 = ((const float*) w)[1];
247 const float vscale2 = ((const float*) w)[2];
248 const float vscale3 = ((const float*) w)[3];
249 w = (const void*) ((const float*) w + 4);
250
251 vfpacc0 *= vscale0;
252 vfpacc1 *= vscale1;
253 vfpacc2 *= vscale2;
254 vfpacc3 *= vscale3;
255
256 vfpacc0 += vmagic_bias;
257 vfpacc1 += vmagic_bias;
258 vfpacc2 += vmagic_bias;
259 vfpacc3 += vmagic_bias;
260
261 int32_t vout0 = (int32_t) fp32_to_bits(vfpacc0);
262 int32_t vout1 = (int32_t) fp32_to_bits(vfpacc1);
263 int32_t vout2 = (int32_t) fp32_to_bits(vfpacc2);
264 int32_t vout3 = (int32_t) fp32_to_bits(vfpacc3);
265
266 vout0 = math_max_s32(vout0, vmagic_min);
267 vout1 = math_max_s32(vout1, vmagic_min);
268 vout2 = math_max_s32(vout2, vmagic_min);
269 vout3 = math_max_s32(vout3, vmagic_min);
270
271 vout0 = math_min_s32(vout0, vmagic_max);
272 vout1 = math_min_s32(vout1, vmagic_max);
273 vout2 = math_min_s32(vout2, vmagic_max);
274 vout3 = math_min_s32(vout3, vmagic_max);
275
276 vout0 -= vmagic_bias_less_zero_point;
277 vout1 -= vmagic_bias_less_zero_point;
278 vout2 -= vmagic_bias_less_zero_point;
279 vout3 -= vmagic_bias_less_zero_point;
280
281 output[0] = (int8_t) vout0;
282 output[1] = (int8_t) vout1;
283 output[2] = (int8_t) vout2;
284 output[3] = (int8_t) vout3;
285 output += 4;
286 }
287 if XNN_UNLIKELY(c != 0) {
288 const int8_t* k = (const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t));
289 do {
290 int32_t vacc = *((const int32_t*) w);
291 w = (const void*) ((uintptr_t) w + sizeof(int32_t));
292
293 const int32_t vi0 = (int32_t) *i0++;
294 const int32_t vk0 = (int32_t) k[0];
295 vacc += vi0 * vk0;
296 const int32_t vi1 = (int32_t) *i1++;
297 const int32_t vk1 = (int32_t) k[4];
298 vacc += vi1 * vk1;
299 const int32_t vi2 = (int32_t) *i2++;
300 const int32_t vk2 = (int32_t) k[8];
301 vacc += vi2 * vk2;
302 const int32_t vi3 = (int32_t) *i3++;
303 const int32_t vk3 = (int32_t) k[12];
304 vacc += vi3 * vk3;
305 const int32_t vi4 = (int32_t) *i4++;
306 const int32_t vk4 = (int32_t) k[16];
307 vacc += vi4 * vk4;
308 const int32_t vi5 = (int32_t) *i5++;
309 const int32_t vk5 = (int32_t) k[20];
310 vacc += vi5 * vk5;
311 const int32_t vi6 = (int32_t) *i6++;
312 const int32_t vk6 = (int32_t) k[24];
313 vacc += vi6 * vk6;
314 const int32_t vi7 = (int32_t) *i7++;
315 const int32_t vk7 = (int32_t) k[28];
316 vacc += vi7 * vk7;
317 const int32_t vi8 = (int32_t) *i8++;
318 const int32_t vk8 = (int32_t) k[32];
319 vacc += vi8 * vk8;
320 k += 1;
321
322 const float vscale = *((const float*) ((uintptr_t) w + 3 * sizeof(int32_t) + 36 * sizeof(int8_t)));
323 float vfpacc = (float) vacc * vscale;
324
325 vfpacc += vmagic_bias;
326 int32_t vout = (int32_t) fp32_to_bits(vfpacc);
327 vout = math_max_s32(vout, vmagic_min);
328 vout = math_min_s32(vout, vmagic_max);
329 vout -= vmagic_bias_less_zero_point;
330
331 *output++ = (int8_t) vout;
332 } while (--c != 0);
333 }
334
335 output = (int8_t*) ((uintptr_t) output + output_increment);
336 } while (--output_width != 0);
337 }
338