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 <xnnpack/dwconv.h>
13 #include <xnnpack/math.h>
14
15
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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])16 void xnn_qc8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic(
17 size_t channels,
18 size_t output_width,
19 const int8_t** input,
20 const void* weights,
21 int8_t* output,
22 size_t input_stride,
23 size_t output_increment,
24 size_t input_offset,
25 const int8_t* zero,
26 const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
27 {
28 assert(channels != 0);
29 assert(output_width != 0);
30
31 const float vmagic_bias = params->fp32_scalar_imagic.magic_bias;
32 const int32_t vmagic_min = params->fp32_scalar_imagic.magic_min;
33 const int32_t vmagic_max = params->fp32_scalar_imagic.magic_max;
34 const int32_t vmagic_bias_less_zero_point = params->fp32_scalar_imagic.magic_bias_less_zero_point;
35 do {
36 const int8_t* i0 = input[0];
37 assert(i0 != NULL);
38 if XNN_UNPREDICTABLE(i0 != zero) {
39 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
40 }
41 const int8_t* i1 = input[1];
42 assert(i1 != NULL);
43 if XNN_UNPREDICTABLE(i1 != zero) {
44 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
45 }
46 const int8_t* i2 = input[2];
47 assert(i2 != NULL);
48 if XNN_UNPREDICTABLE(i2 != zero) {
49 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
50 }
51 const int8_t* i3 = input[3];
52 assert(i3 != NULL);
53 if XNN_UNPREDICTABLE(i3 != zero) {
54 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
55 }
56 const int8_t* i4 = input[4];
57 assert(i4 != NULL);
58 if XNN_UNPREDICTABLE(i4 != zero) {
59 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
60 }
61 const int8_t* i5 = input[5];
62 assert(i5 != NULL);
63 if XNN_UNPREDICTABLE(i5 != zero) {
64 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
65 }
66 const int8_t* i6 = input[6];
67 assert(i6 != NULL);
68 if XNN_UNPREDICTABLE(i6 != zero) {
69 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
70 }
71 const int8_t* i7 = input[7];
72 assert(i7 != NULL);
73 if XNN_UNPREDICTABLE(i7 != zero) {
74 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
75 }
76 const int8_t* i8 = input[8];
77 assert(i8 != NULL);
78 if XNN_UNPREDICTABLE(i8 != zero) {
79 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
80 }
81 input = (const int8_t**) ((uintptr_t) input + input_stride);
82
83 size_t c = channels;
84 const void* w = weights;
85 for (; c >= 4; c -= 4) {
86 int32_t vacc0 = ((const int32_t*) w)[0];
87 int32_t vacc1 = ((const int32_t*) w)[1];
88 int32_t vacc2 = ((const int32_t*) w)[2];
89 int32_t vacc3 = ((const int32_t*) w)[3];
90
91
92 const int32_t vi0x0 = (int32_t) i0[0];
93 const int32_t vi0x1 = (int32_t) i0[1];
94 const int32_t vi0x2 = (int32_t) i0[2];
95 const int32_t vi0x3 = (int32_t) i0[3];
96 i0 += 4;
97
98 const int32_t vk0x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[0];
99 const int32_t vk0x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[1];
100 const int32_t vk0x2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[2];
101 const int32_t vk0x3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[3];
102
103 vacc0 += vi0x0 * vk0x0;
104 vacc1 += vi0x1 * vk0x1;
105 vacc2 += vi0x2 * vk0x2;
106 vacc3 += vi0x3 * vk0x3;
107
108 const int32_t vi1x0 = (int32_t) i1[0];
109 const int32_t vi1x1 = (int32_t) i1[1];
110 const int32_t vi1x2 = (int32_t) i1[2];
111 const int32_t vi1x3 = (int32_t) i1[3];
112 i1 += 4;
113
114 const int32_t vk1x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[4];
115 const int32_t vk1x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[5];
116 const int32_t vk1x2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[6];
117 const int32_t vk1x3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[7];
118
119 vacc0 += vi1x0 * vk1x0;
120 vacc1 += vi1x1 * vk1x1;
121 vacc2 += vi1x2 * vk1x2;
122 vacc3 += vi1x3 * vk1x3;
123
124 const int32_t vi2x0 = (int32_t) i2[0];
125 const int32_t vi2x1 = (int32_t) i2[1];
126 const int32_t vi2x2 = (int32_t) i2[2];
127 const int32_t vi2x3 = (int32_t) i2[3];
128 i2 += 4;
129
130 const int32_t vk2x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[8];
131 const int32_t vk2x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[9];
132 const int32_t vk2x2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[10];
133 const int32_t vk2x3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[11];
134
135 vacc0 += vi2x0 * vk2x0;
136 vacc1 += vi2x1 * vk2x1;
137 vacc2 += vi2x2 * vk2x2;
138 vacc3 += vi2x3 * vk2x3;
139
140 const int32_t vi3x0 = (int32_t) i3[0];
141 const int32_t vi3x1 = (int32_t) i3[1];
142 const int32_t vi3x2 = (int32_t) i3[2];
143 const int32_t vi3x3 = (int32_t) i3[3];
144 i3 += 4;
145
146 const int32_t vk3x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[12];
147 const int32_t vk3x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[13];
148 const int32_t vk3x2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[14];
149 const int32_t vk3x3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[15];
150
151 vacc0 += vi3x0 * vk3x0;
152 vacc1 += vi3x1 * vk3x1;
153 vacc2 += vi3x2 * vk3x2;
154 vacc3 += vi3x3 * vk3x3;
155
156 const int32_t vi4x0 = (int32_t) i4[0];
157 const int32_t vi4x1 = (int32_t) i4[1];
158 const int32_t vi4x2 = (int32_t) i4[2];
159 const int32_t vi4x3 = (int32_t) i4[3];
160 i4 += 4;
161
162 const int32_t vk4x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[16];
163 const int32_t vk4x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[17];
164 const int32_t vk4x2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[18];
165 const int32_t vk4x3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[19];
166
167 vacc0 += vi4x0 * vk4x0;
168 vacc1 += vi4x1 * vk4x1;
169 vacc2 += vi4x2 * vk4x2;
170 vacc3 += vi4x3 * vk4x3;
171
172 const int32_t vi5x0 = (int32_t) i5[0];
173 const int32_t vi5x1 = (int32_t) i5[1];
174 const int32_t vi5x2 = (int32_t) i5[2];
175 const int32_t vi5x3 = (int32_t) i5[3];
176 i5 += 4;
177
178 const int32_t vk5x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[20];
179 const int32_t vk5x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[21];
180 const int32_t vk5x2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[22];
181 const int32_t vk5x3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[23];
182
183 vacc0 += vi5x0 * vk5x0;
184 vacc1 += vi5x1 * vk5x1;
185 vacc2 += vi5x2 * vk5x2;
186 vacc3 += vi5x3 * vk5x3;
187
188 const int32_t vi6x0 = (int32_t) i6[0];
189 const int32_t vi6x1 = (int32_t) i6[1];
190 const int32_t vi6x2 = (int32_t) i6[2];
191 const int32_t vi6x3 = (int32_t) i6[3];
192 i6 += 4;
193
194 const int32_t vk6x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[24];
195 const int32_t vk6x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[25];
196 const int32_t vk6x2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[26];
197 const int32_t vk6x3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[27];
198
199 vacc0 += vi6x0 * vk6x0;
200 vacc1 += vi6x1 * vk6x1;
201 vacc2 += vi6x2 * vk6x2;
202 vacc3 += vi6x3 * vk6x3;
203
204 const int32_t vi7x0 = (int32_t) i7[0];
205 const int32_t vi7x1 = (int32_t) i7[1];
206 const int32_t vi7x2 = (int32_t) i7[2];
207 const int32_t vi7x3 = (int32_t) i7[3];
208 i7 += 4;
209
210 const int32_t vk7x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[28];
211 const int32_t vk7x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[29];
212 const int32_t vk7x2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[30];
213 const int32_t vk7x3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[31];
214
215 vacc0 += vi7x0 * vk7x0;
216 vacc1 += vi7x1 * vk7x1;
217 vacc2 += vi7x2 * vk7x2;
218 vacc3 += vi7x3 * vk7x3;
219
220 const int32_t vi8x0 = (int32_t) i8[0];
221 const int32_t vi8x1 = (int32_t) i8[1];
222 const int32_t vi8x2 = (int32_t) i8[2];
223 const int32_t vi8x3 = (int32_t) i8[3];
224 i8 += 4;
225
226 const int32_t vk8x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[32];
227 const int32_t vk8x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[33];
228 const int32_t vk8x2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[34];
229 const int32_t vk8x3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[35];
230
231 vacc0 += vi8x0 * vk8x0;
232 vacc1 += vi8x1 * vk8x1;
233 vacc2 += vi8x2 * vk8x2;
234 vacc3 += vi8x3 * vk8x3;
235
236 w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t) + 36 * sizeof(int8_t));
237
238 float vfpacc0 = (float) vacc0;
239 float vfpacc1 = (float) vacc1;
240 float vfpacc2 = (float) vacc2;
241 float vfpacc3 = (float) vacc3;
242
243 const float vscale0 = ((const float*) w)[0];
244 const float vscale1 = ((const float*) w)[1];
245 const float vscale2 = ((const float*) w)[2];
246 const float vscale3 = ((const float*) w)[3];
247 w = (const void*) ((const float*) w + 4);
248
249 vfpacc0 *= vscale0;
250 vfpacc1 *= vscale1;
251 vfpacc2 *= vscale2;
252 vfpacc3 *= vscale3;
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 void*) ((uintptr_t) w + sizeof(int32_t));
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 const float vscale = *((const float*) ((uintptr_t) w + 3 * sizeof(int32_t) + 36 * sizeof(int8_t)));
321 float vfpacc = (float) vacc * vscale;
322
323 vfpacc += vmagic_bias;
324 int32_t vout = (int32_t) float_as_uint32(vfpacc);
325 vout = math_max_s32(vout, vmagic_min);
326 vout = math_min_s32(vout, vmagic_max);
327 vout -= vmagic_bias_less_zero_point;
328
329 *output++ = (int8_t) vout;
330 } while (--c != 0);
331 }
332
333 output = (int8_t*) ((uintptr_t) output + output_increment);
334 } while (--output_width != 0);
335 }
336