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 #include <xnnpack/unaligned.h>
15
16
xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_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)])17 void xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_fmagic(
18 size_t channels,
19 size_t output_width,
20 const uint8_t** input,
21 const void* weights,
22 uint8_t* output,
23 size_t input_stride,
24 size_t output_increment,
25 size_t input_offset,
26 const uint8_t* zero,
27 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29 assert(channels != 0);
30 assert(output_width != 0);
31
32 const float vscale = params->fp32_scalar_fmagic.scale;
33 const float voutput_min_less_zero_point = params->fp32_scalar_fmagic.output_min_less_zero_point;
34 const float voutput_max_less_zero_point = params->fp32_scalar_fmagic.output_max_less_zero_point;
35 const float vmagic_bias = params->fp32_scalar_fmagic.magic_bias;
36 const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar_fmagic.magic_bias_less_output_zero_point;
37 const int32_t vkernel_zero_point = params->fp32_scalar_fmagic.kernel_zero_point;
38 do {
39 const uint8_t* i0 = input[0];
40 assert(i0 != NULL);
41 if XNN_UNPREDICTABLE(i0 != zero) {
42 i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
43 }
44 const uint8_t* i1 = input[1];
45 assert(i1 != NULL);
46 if XNN_UNPREDICTABLE(i1 != zero) {
47 i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
48 }
49 const uint8_t* i2 = input[2];
50 assert(i2 != NULL);
51 if XNN_UNPREDICTABLE(i2 != zero) {
52 i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
53 }
54 const uint8_t* i3 = input[3];
55 assert(i3 != NULL);
56 if XNN_UNPREDICTABLE(i3 != zero) {
57 i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
58 }
59 const uint8_t* i4 = input[4];
60 assert(i4 != NULL);
61 if XNN_UNPREDICTABLE(i4 != zero) {
62 i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
63 }
64 const uint8_t* i5 = input[5];
65 assert(i5 != NULL);
66 if XNN_UNPREDICTABLE(i5 != zero) {
67 i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
68 }
69 const uint8_t* i6 = input[6];
70 assert(i6 != NULL);
71 if XNN_UNPREDICTABLE(i6 != zero) {
72 i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
73 }
74 const uint8_t* i7 = input[7];
75 assert(i7 != NULL);
76 if XNN_UNPREDICTABLE(i7 != zero) {
77 i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
78 }
79 const uint8_t* i8 = input[8];
80 assert(i8 != NULL);
81 if XNN_UNPREDICTABLE(i8 != zero) {
82 i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
83 }
84 const uint8_t* i9 = input[9];
85 assert(i9 != NULL);
86 if XNN_UNPREDICTABLE(i9 != zero) {
87 i9 = (const uint8_t*) ((uintptr_t) i9 + input_offset);
88 }
89 const uint8_t* i10 = input[10];
90 assert(i10 != NULL);
91 if XNN_UNPREDICTABLE(i10 != zero) {
92 i10 = (const uint8_t*) ((uintptr_t) i10 + input_offset);
93 }
94 const uint8_t* i11 = input[11];
95 assert(i11 != NULL);
96 if XNN_UNPREDICTABLE(i11 != zero) {
97 i11 = (const uint8_t*) ((uintptr_t) i11 + input_offset);
98 }
99 const uint8_t* i12 = input[12];
100 assert(i12 != NULL);
101 if XNN_UNPREDICTABLE(i12 != zero) {
102 i12 = (const uint8_t*) ((uintptr_t) i12 + input_offset);
103 }
104 const uint8_t* i13 = input[13];
105 assert(i13 != NULL);
106 if XNN_UNPREDICTABLE(i13 != zero) {
107 i13 = (const uint8_t*) ((uintptr_t) i13 + input_offset);
108 }
109 const uint8_t* i14 = input[14];
110 assert(i14 != NULL);
111 if XNN_UNPREDICTABLE(i14 != zero) {
112 i14 = (const uint8_t*) ((uintptr_t) i14 + input_offset);
113 }
114 const uint8_t* i15 = input[15];
115 assert(i15 != NULL);
116 if XNN_UNPREDICTABLE(i15 != zero) {
117 i15 = (const uint8_t*) ((uintptr_t) i15 + input_offset);
118 }
119 const uint8_t* i16 = input[16];
120 assert(i16 != NULL);
121 if XNN_UNPREDICTABLE(i16 != zero) {
122 i16 = (const uint8_t*) ((uintptr_t) i16 + input_offset);
123 }
124 const uint8_t* i17 = input[17];
125 assert(i17 != NULL);
126 if XNN_UNPREDICTABLE(i17 != zero) {
127 i17 = (const uint8_t*) ((uintptr_t) i17 + input_offset);
128 }
129 const uint8_t* i18 = input[18];
130 assert(i18 != NULL);
131 if XNN_UNPREDICTABLE(i18 != zero) {
132 i18 = (const uint8_t*) ((uintptr_t) i18 + input_offset);
133 }
134 const uint8_t* i19 = input[19];
135 assert(i19 != NULL);
136 if XNN_UNPREDICTABLE(i19 != zero) {
137 i19 = (const uint8_t*) ((uintptr_t) i19 + input_offset);
138 }
139 const uint8_t* i20 = input[20];
140 assert(i20 != NULL);
141 if XNN_UNPREDICTABLE(i20 != zero) {
142 i20 = (const uint8_t*) ((uintptr_t) i20 + input_offset);
143 }
144 const uint8_t* i21 = input[21];
145 assert(i21 != NULL);
146 if XNN_UNPREDICTABLE(i21 != zero) {
147 i21 = (const uint8_t*) ((uintptr_t) i21 + input_offset);
148 }
149 const uint8_t* i22 = input[22];
150 assert(i22 != NULL);
151 if XNN_UNPREDICTABLE(i22 != zero) {
152 i22 = (const uint8_t*) ((uintptr_t) i22 + input_offset);
153 }
154 const uint8_t* i23 = input[23];
155 assert(i23 != NULL);
156 if XNN_UNPREDICTABLE(i23 != zero) {
157 i23 = (const uint8_t*) ((uintptr_t) i23 + input_offset);
158 }
159 const uint8_t* i24 = input[24];
160 assert(i24 != NULL);
161 if XNN_UNPREDICTABLE(i24 != zero) {
162 i24 = (const uint8_t*) ((uintptr_t) i24 + input_offset);
163 }
164 input = (const uint8_t**) ((uintptr_t) input + input_stride);
165
166 size_t c = channels;
167 const void* w = weights;
168 do {
169 int32_t vacc = unaligned_load_s32(w);
170
171 const int32_t vi0 = (int32_t) (uint32_t) *i0++;
172 const int32_t vk0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[0] - vkernel_zero_point;
173 vacc += vi0 * vk0;
174 const int32_t vi1 = (int32_t) (uint32_t) *i1++;
175 const int32_t vk1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[1] - vkernel_zero_point;
176 vacc += vi1 * vk1;
177 const int32_t vi2 = (int32_t) (uint32_t) *i2++;
178 const int32_t vk2 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[2] - vkernel_zero_point;
179 vacc += vi2 * vk2;
180 const int32_t vi3 = (int32_t) (uint32_t) *i3++;
181 const int32_t vk3 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[3] - vkernel_zero_point;
182 vacc += vi3 * vk3;
183 const int32_t vi4 = (int32_t) (uint32_t) *i4++;
184 const int32_t vk4 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[4] - vkernel_zero_point;
185 vacc += vi4 * vk4;
186 const int32_t vi5 = (int32_t) (uint32_t) *i5++;
187 const int32_t vk5 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[5] - vkernel_zero_point;
188 vacc += vi5 * vk5;
189 const int32_t vi6 = (int32_t) (uint32_t) *i6++;
190 const int32_t vk6 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[6] - vkernel_zero_point;
191 vacc += vi6 * vk6;
192 const int32_t vi7 = (int32_t) (uint32_t) *i7++;
193 const int32_t vk7 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[7] - vkernel_zero_point;
194 vacc += vi7 * vk7;
195 const int32_t vi8 = (int32_t) (uint32_t) *i8++;
196 const int32_t vk8 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[8] - vkernel_zero_point;
197 vacc += vi8 * vk8;
198 const int32_t vi9 = (int32_t) (uint32_t) *i9++;
199 const int32_t vk9 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[9] - vkernel_zero_point;
200 vacc += vi9 * vk9;
201 const int32_t vi10 = (int32_t) (uint32_t) *i10++;
202 const int32_t vk10 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[10] - vkernel_zero_point;
203 vacc += vi10 * vk10;
204 const int32_t vi11 = (int32_t) (uint32_t) *i11++;
205 const int32_t vk11 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[11] - vkernel_zero_point;
206 vacc += vi11 * vk11;
207 const int32_t vi12 = (int32_t) (uint32_t) *i12++;
208 const int32_t vk12 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[12] - vkernel_zero_point;
209 vacc += vi12 * vk12;
210 const int32_t vi13 = (int32_t) (uint32_t) *i13++;
211 const int32_t vk13 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[13] - vkernel_zero_point;
212 vacc += vi13 * vk13;
213 const int32_t vi14 = (int32_t) (uint32_t) *i14++;
214 const int32_t vk14 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[14] - vkernel_zero_point;
215 vacc += vi14 * vk14;
216 const int32_t vi15 = (int32_t) (uint32_t) *i15++;
217 const int32_t vk15 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[15] - vkernel_zero_point;
218 vacc += vi15 * vk15;
219 const int32_t vi16 = (int32_t) (uint32_t) *i16++;
220 const int32_t vk16 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[16] - vkernel_zero_point;
221 vacc += vi16 * vk16;
222 const int32_t vi17 = (int32_t) (uint32_t) *i17++;
223 const int32_t vk17 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[17] - vkernel_zero_point;
224 vacc += vi17 * vk17;
225 const int32_t vi18 = (int32_t) (uint32_t) *i18++;
226 const int32_t vk18 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[18] - vkernel_zero_point;
227 vacc += vi18 * vk18;
228 const int32_t vi19 = (int32_t) (uint32_t) *i19++;
229 const int32_t vk19 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[19] - vkernel_zero_point;
230 vacc += vi19 * vk19;
231 const int32_t vi20 = (int32_t) (uint32_t) *i20++;
232 const int32_t vk20 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[20] - vkernel_zero_point;
233 vacc += vi20 * vk20;
234 const int32_t vi21 = (int32_t) (uint32_t) *i21++;
235 const int32_t vk21 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[21] - vkernel_zero_point;
236 vacc += vi21 * vk21;
237 const int32_t vi22 = (int32_t) (uint32_t) *i22++;
238 const int32_t vk22 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[22] - vkernel_zero_point;
239 vacc += vi22 * vk22;
240 const int32_t vi23 = (int32_t) (uint32_t) *i23++;
241 const int32_t vk23 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[23] - vkernel_zero_point;
242 vacc += vi23 * vk23;
243 const int32_t vi24 = (int32_t) (uint32_t) *i24++;
244 const int32_t vk24 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[24] - vkernel_zero_point;
245 vacc += vi24 * vk24;
246
247 w = (const void*) ((uintptr_t) w + sizeof(int32_t) + 25 * sizeof(uint8_t));
248
249 float vfpacc = (float) vacc * vscale;
250
251 vfpacc = math_max_f32(vfpacc, voutput_min_less_zero_point);
252 vfpacc = math_min_f32(vfpacc, voutput_max_less_zero_point);
253 vfpacc += vmagic_bias;
254 int32_t vout = (int32_t) float_as_uint32(vfpacc) - vmagic_bias_less_output_zero_point;
255
256 *output++ = (uint8_t) vout;
257 } while (--c != 0);
258
259 output = (uint8_t*) ((uintptr_t) output + output_increment);
260 } while (--output_width != 0);
261 }
262