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_qu8_dwconv_minmax_fp32_ukernel_up4x9__wasm_fmagic(size_t channels,size_t output_width,const uint8_t ** input,const void * weights,uint8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])16 void xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__wasm_fmagic(
17 size_t channels,
18 size_t output_width,
19 const uint8_t** input,
20 const void* weights,
21 uint8_t* output,
22 size_t input_stride,
23 size_t output_increment,
24 size_t input_offset,
25 const uint8_t* zero,
26 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
27 {
28 assert(channels != 0);
29 assert(output_width != 0);
30
31 const float vscale = params->fp32_scalar_fmagic.scale;
32 const float voutput_min_less_zero_point = params->fp32_scalar_fmagic.output_min_less_zero_point;
33 const float voutput_max_less_zero_point = params->fp32_scalar_fmagic.output_max_less_zero_point;
34 const float vmagic_bias = params->fp32_scalar_fmagic.magic_bias;
35 const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar_fmagic.magic_bias_less_output_zero_point;
36 const int32_t vkernel_zero_point = params->fp32_scalar_fmagic.kernel_zero_point;
37 do {
38 const uint8_t* i0 = input[0];
39 assert(i0 != NULL);
40 if XNN_UNPREDICTABLE(i0 != zero) {
41 i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
42 }
43 const uint8_t* i1 = input[1];
44 assert(i1 != NULL);
45 if XNN_UNPREDICTABLE(i1 != zero) {
46 i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
47 }
48 const uint8_t* i2 = input[2];
49 assert(i2 != NULL);
50 if XNN_UNPREDICTABLE(i2 != zero) {
51 i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
52 }
53 const uint8_t* i3 = input[3];
54 assert(i3 != NULL);
55 if XNN_UNPREDICTABLE(i3 != zero) {
56 i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
57 }
58 const uint8_t* i4 = input[4];
59 assert(i4 != NULL);
60 if XNN_UNPREDICTABLE(i4 != zero) {
61 i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
62 }
63 const uint8_t* i5 = input[5];
64 assert(i5 != NULL);
65 if XNN_UNPREDICTABLE(i5 != zero) {
66 i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
67 }
68 const uint8_t* i6 = input[6];
69 assert(i6 != NULL);
70 if XNN_UNPREDICTABLE(i6 != zero) {
71 i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
72 }
73 const uint8_t* i7 = input[7];
74 assert(i7 != NULL);
75 if XNN_UNPREDICTABLE(i7 != zero) {
76 i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
77 }
78 const uint8_t* i8 = input[8];
79 assert(i8 != NULL);
80 if XNN_UNPREDICTABLE(i8 != zero) {
81 i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
82 }
83 input = (const uint8_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) (uint32_t) i0[0];
95 const int32_t vi0x1 = (int32_t) (uint32_t) i0[1];
96 const int32_t vi0x2 = (int32_t) (uint32_t) i0[2];
97 const int32_t vi0x3 = (int32_t) (uint32_t) i0[3];
98 i0 += 4;
99
100 const int32_t vk0x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[0] - vkernel_zero_point;
101 const int32_t vk0x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[1] - vkernel_zero_point;
102 const int32_t vk0x2 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[2] - vkernel_zero_point;
103 const int32_t vk0x3 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[3] - vkernel_zero_point;
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) (uint32_t) i1[0];
111 const int32_t vi1x1 = (int32_t) (uint32_t) i1[1];
112 const int32_t vi1x2 = (int32_t) (uint32_t) i1[2];
113 const int32_t vi1x3 = (int32_t) (uint32_t) i1[3];
114 i1 += 4;
115
116 const int32_t vk1x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[4] - vkernel_zero_point;
117 const int32_t vk1x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[5] - vkernel_zero_point;
118 const int32_t vk1x2 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[6] - vkernel_zero_point;
119 const int32_t vk1x3 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[7] - vkernel_zero_point;
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) (uint32_t) i2[0];
127 const int32_t vi2x1 = (int32_t) (uint32_t) i2[1];
128 const int32_t vi2x2 = (int32_t) (uint32_t) i2[2];
129 const int32_t vi2x3 = (int32_t) (uint32_t) i2[3];
130 i2 += 4;
131
132 const int32_t vk2x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[8] - vkernel_zero_point;
133 const int32_t vk2x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[9] - vkernel_zero_point;
134 const int32_t vk2x2 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[10] - vkernel_zero_point;
135 const int32_t vk2x3 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[11] - vkernel_zero_point;
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) (uint32_t) i3[0];
143 const int32_t vi3x1 = (int32_t) (uint32_t) i3[1];
144 const int32_t vi3x2 = (int32_t) (uint32_t) i3[2];
145 const int32_t vi3x3 = (int32_t) (uint32_t) i3[3];
146 i3 += 4;
147
148 const int32_t vk3x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[12] - vkernel_zero_point;
149 const int32_t vk3x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[13] - vkernel_zero_point;
150 const int32_t vk3x2 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[14] - vkernel_zero_point;
151 const int32_t vk3x3 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[15] - vkernel_zero_point;
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) (uint32_t) i4[0];
159 const int32_t vi4x1 = (int32_t) (uint32_t) i4[1];
160 const int32_t vi4x2 = (int32_t) (uint32_t) i4[2];
161 const int32_t vi4x3 = (int32_t) (uint32_t) i4[3];
162 i4 += 4;
163
164 const int32_t vk4x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[16] - vkernel_zero_point;
165 const int32_t vk4x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[17] - vkernel_zero_point;
166 const int32_t vk4x2 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[18] - vkernel_zero_point;
167 const int32_t vk4x3 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[19] - vkernel_zero_point;
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) (uint32_t) i5[0];
175 const int32_t vi5x1 = (int32_t) (uint32_t) i5[1];
176 const int32_t vi5x2 = (int32_t) (uint32_t) i5[2];
177 const int32_t vi5x3 = (int32_t) (uint32_t) i5[3];
178 i5 += 4;
179
180 const int32_t vk5x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[20] - vkernel_zero_point;
181 const int32_t vk5x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[21] - vkernel_zero_point;
182 const int32_t vk5x2 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[22] - vkernel_zero_point;
183 const int32_t vk5x3 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[23] - vkernel_zero_point;
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) (uint32_t) i6[0];
191 const int32_t vi6x1 = (int32_t) (uint32_t) i6[1];
192 const int32_t vi6x2 = (int32_t) (uint32_t) i6[2];
193 const int32_t vi6x3 = (int32_t) (uint32_t) i6[3];
194 i6 += 4;
195
196 const int32_t vk6x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[24] - vkernel_zero_point;
197 const int32_t vk6x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[25] - vkernel_zero_point;
198 const int32_t vk6x2 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[26] - vkernel_zero_point;
199 const int32_t vk6x3 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[27] - vkernel_zero_point;
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) (uint32_t) i7[0];
207 const int32_t vi7x1 = (int32_t) (uint32_t) i7[1];
208 const int32_t vi7x2 = (int32_t) (uint32_t) i7[2];
209 const int32_t vi7x3 = (int32_t) (uint32_t) i7[3];
210 i7 += 4;
211
212 const int32_t vk7x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[28] - vkernel_zero_point;
213 const int32_t vk7x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[29] - vkernel_zero_point;
214 const int32_t vk7x2 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[30] - vkernel_zero_point;
215 const int32_t vk7x3 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[31] - vkernel_zero_point;
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) (uint32_t) i8[0];
223 const int32_t vi8x1 = (int32_t) (uint32_t) i8[1];
224 const int32_t vi8x2 = (int32_t) (uint32_t) i8[2];
225 const int32_t vi8x3 = (int32_t) (uint32_t) i8[3];
226 i8 += 4;
227
228 const int32_t vk8x0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[32] - vkernel_zero_point;
229 const int32_t vk8x1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[33] - vkernel_zero_point;
230 const int32_t vk8x2 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[34] - vkernel_zero_point;
231 const int32_t vk8x3 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t)))[35] - vkernel_zero_point;
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(uint8_t));
239
240 float vfpacc0 = (float) vacc0;
241 float vfpacc1 = (float) vacc1;
242 float vfpacc2 = (float) vacc2;
243 float vfpacc3 = (float) vacc3;
244
245 vfpacc0 *= vscale;
246 vfpacc1 *= vscale;
247 vfpacc2 *= vscale;
248 vfpacc3 *= vscale;
249
250 vfpacc0 = __builtin_wasm_max_f32(vfpacc0, voutput_min_less_zero_point);
251 vfpacc1 = __builtin_wasm_max_f32(vfpacc1, voutput_min_less_zero_point);
252 vfpacc2 = __builtin_wasm_max_f32(vfpacc2, voutput_min_less_zero_point);
253 vfpacc3 = __builtin_wasm_max_f32(vfpacc3, voutput_min_less_zero_point);
254
255 vfpacc0 = __builtin_wasm_min_f32(vfpacc0, voutput_max_less_zero_point);
256 vfpacc1 = __builtin_wasm_min_f32(vfpacc1, voutput_max_less_zero_point);
257 vfpacc2 = __builtin_wasm_min_f32(vfpacc2, voutput_max_less_zero_point);
258 vfpacc3 = __builtin_wasm_min_f32(vfpacc3, voutput_max_less_zero_point);
259
260 vfpacc0 += vmagic_bias;
261 vfpacc1 += vmagic_bias;
262 vfpacc2 += vmagic_bias;
263 vfpacc3 += vmagic_bias;
264
265 int32_t vout0 = (int32_t) float_as_uint32(vfpacc0) - vmagic_bias_less_output_zero_point;
266 int32_t vout1 = (int32_t) float_as_uint32(vfpacc1) - vmagic_bias_less_output_zero_point;
267 int32_t vout2 = (int32_t) float_as_uint32(vfpacc2) - vmagic_bias_less_output_zero_point;
268 int32_t vout3 = (int32_t) float_as_uint32(vfpacc3) - vmagic_bias_less_output_zero_point;
269
270 output[0] = (uint8_t) vout0;
271 output[1] = (uint8_t) vout1;
272 output[2] = (uint8_t) vout2;
273 output[3] = (uint8_t) vout3;
274 output += 4;
275 }
276 if XNN_UNLIKELY(c != 0) {
277 const uint8_t* k = (const uint8_t*) ((uintptr_t) w + 4 * sizeof(int32_t));
278 do {
279 int32_t vacc = *((const int32_t*) w);
280 w = (const void*) ((uintptr_t) w + sizeof(int32_t));
281
282 const int32_t vi0 = (int32_t) (uint32_t) *i0++;
283 const int32_t vk0 = (int32_t) (uint32_t) k[0] - vkernel_zero_point;
284 vacc += vi0 * vk0;
285 const int32_t vi1 = (int32_t) (uint32_t) *i1++;
286 const int32_t vk1 = (int32_t) (uint32_t) k[4] - vkernel_zero_point;
287 vacc += vi1 * vk1;
288 const int32_t vi2 = (int32_t) (uint32_t) *i2++;
289 const int32_t vk2 = (int32_t) (uint32_t) k[8] - vkernel_zero_point;
290 vacc += vi2 * vk2;
291 const int32_t vi3 = (int32_t) (uint32_t) *i3++;
292 const int32_t vk3 = (int32_t) (uint32_t) k[12] - vkernel_zero_point;
293 vacc += vi3 * vk3;
294 const int32_t vi4 = (int32_t) (uint32_t) *i4++;
295 const int32_t vk4 = (int32_t) (uint32_t) k[16] - vkernel_zero_point;
296 vacc += vi4 * vk4;
297 const int32_t vi5 = (int32_t) (uint32_t) *i5++;
298 const int32_t vk5 = (int32_t) (uint32_t) k[20] - vkernel_zero_point;
299 vacc += vi5 * vk5;
300 const int32_t vi6 = (int32_t) (uint32_t) *i6++;
301 const int32_t vk6 = (int32_t) (uint32_t) k[24] - vkernel_zero_point;
302 vacc += vi6 * vk6;
303 const int32_t vi7 = (int32_t) (uint32_t) *i7++;
304 const int32_t vk7 = (int32_t) (uint32_t) k[28] - vkernel_zero_point;
305 vacc += vi7 * vk7;
306 const int32_t vi8 = (int32_t) (uint32_t) *i8++;
307 const int32_t vk8 = (int32_t) (uint32_t) k[32] - vkernel_zero_point;
308 vacc += vi8 * vk8;
309 k += 1;
310
311 float vfpacc = (float) vacc * vscale;
312
313 vfpacc = __builtin_wasm_max_f32(vfpacc, voutput_min_less_zero_point);
314 vfpacc = __builtin_wasm_min_f32(vfpacc, voutput_max_less_zero_point);
315 vfpacc += vmagic_bias;
316 int32_t vout = (int32_t) float_as_uint32(vfpacc) - vmagic_bias_less_output_zero_point;
317
318 *output++ = (uint8_t) vout;
319 } while (--c != 0);
320 }
321
322 output = (uint8_t*) ((uintptr_t) output + output_increment);
323 } while (--output_width != 0);
324 }
325