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_qc8_dwconv_minmax_fp32_ukernel_up1x25__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)])17 void xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic(
18 size_t channels,
19 size_t output_width,
20 const int8_t** input,
21 const void* weights,
22 int8_t* output,
23 size_t input_stride,
24 size_t output_increment,
25 size_t input_offset,
26 const int8_t* zero,
27 const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29 assert(channels != 0);
30 assert(output_width != 0);
31
32 const float vmagic_bias = params->fp32_scalar_imagic.magic_bias;
33 const int32_t vmagic_min = params->fp32_scalar_imagic.magic_min;
34 const int32_t vmagic_max = params->fp32_scalar_imagic.magic_max;
35 const int32_t vmagic_bias_less_zero_point = params->fp32_scalar_imagic.magic_bias_less_zero_point;
36 do {
37 const int8_t* i0 = input[0];
38 assert(i0 != NULL);
39 if XNN_UNPREDICTABLE(i0 != zero) {
40 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
41 }
42 const int8_t* i1 = input[1];
43 assert(i1 != NULL);
44 if XNN_UNPREDICTABLE(i1 != zero) {
45 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
46 }
47 const int8_t* i2 = input[2];
48 assert(i2 != NULL);
49 if XNN_UNPREDICTABLE(i2 != zero) {
50 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
51 }
52 const int8_t* i3 = input[3];
53 assert(i3 != NULL);
54 if XNN_UNPREDICTABLE(i3 != zero) {
55 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
56 }
57 const int8_t* i4 = input[4];
58 assert(i4 != NULL);
59 if XNN_UNPREDICTABLE(i4 != zero) {
60 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
61 }
62 const int8_t* i5 = input[5];
63 assert(i5 != NULL);
64 if XNN_UNPREDICTABLE(i5 != zero) {
65 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
66 }
67 const int8_t* i6 = input[6];
68 assert(i6 != NULL);
69 if XNN_UNPREDICTABLE(i6 != zero) {
70 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
71 }
72 const int8_t* i7 = input[7];
73 assert(i7 != NULL);
74 if XNN_UNPREDICTABLE(i7 != zero) {
75 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
76 }
77 const int8_t* i8 = input[8];
78 assert(i8 != NULL);
79 if XNN_UNPREDICTABLE(i8 != zero) {
80 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
81 }
82 const int8_t* i9 = input[9];
83 assert(i9 != NULL);
84 if XNN_UNPREDICTABLE(i9 != zero) {
85 i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
86 }
87 const int8_t* i10 = input[10];
88 assert(i10 != NULL);
89 if XNN_UNPREDICTABLE(i10 != zero) {
90 i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
91 }
92 const int8_t* i11 = input[11];
93 assert(i11 != NULL);
94 if XNN_UNPREDICTABLE(i11 != zero) {
95 i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
96 }
97 const int8_t* i12 = input[12];
98 assert(i12 != NULL);
99 if XNN_UNPREDICTABLE(i12 != zero) {
100 i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
101 }
102 const int8_t* i13 = input[13];
103 assert(i13 != NULL);
104 if XNN_UNPREDICTABLE(i13 != zero) {
105 i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
106 }
107 const int8_t* i14 = input[14];
108 assert(i14 != NULL);
109 if XNN_UNPREDICTABLE(i14 != zero) {
110 i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
111 }
112 const int8_t* i15 = input[15];
113 assert(i15 != NULL);
114 if XNN_UNPREDICTABLE(i15 != zero) {
115 i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
116 }
117 const int8_t* i16 = input[16];
118 assert(i16 != NULL);
119 if XNN_UNPREDICTABLE(i16 != zero) {
120 i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
121 }
122 const int8_t* i17 = input[17];
123 assert(i17 != NULL);
124 if XNN_UNPREDICTABLE(i17 != zero) {
125 i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
126 }
127 const int8_t* i18 = input[18];
128 assert(i18 != NULL);
129 if XNN_UNPREDICTABLE(i18 != zero) {
130 i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
131 }
132 const int8_t* i19 = input[19];
133 assert(i19 != NULL);
134 if XNN_UNPREDICTABLE(i19 != zero) {
135 i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
136 }
137 const int8_t* i20 = input[20];
138 assert(i20 != NULL);
139 if XNN_UNPREDICTABLE(i20 != zero) {
140 i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
141 }
142 const int8_t* i21 = input[21];
143 assert(i21 != NULL);
144 if XNN_UNPREDICTABLE(i21 != zero) {
145 i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
146 }
147 const int8_t* i22 = input[22];
148 assert(i22 != NULL);
149 if XNN_UNPREDICTABLE(i22 != zero) {
150 i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
151 }
152 const int8_t* i23 = input[23];
153 assert(i23 != NULL);
154 if XNN_UNPREDICTABLE(i23 != zero) {
155 i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
156 }
157 const int8_t* i24 = input[24];
158 assert(i24 != NULL);
159 if XNN_UNPREDICTABLE(i24 != zero) {
160 i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
161 }
162 input = (const int8_t**) ((uintptr_t) input + input_stride);
163
164 size_t c = channels;
165 const void* w = weights;
166 do {
167 int32_t vacc = unaligned_load_s32(w);
168
169 const int32_t vi0 = (int32_t) *i0++;
170 const int32_t vk0 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[0];
171 vacc += vi0 * vk0;
172 const int32_t vi1 = (int32_t) *i1++;
173 const int32_t vk1 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[1];
174 vacc += vi1 * vk1;
175 const int32_t vi2 = (int32_t) *i2++;
176 const int32_t vk2 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[2];
177 vacc += vi2 * vk2;
178 const int32_t vi3 = (int32_t) *i3++;
179 const int32_t vk3 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[3];
180 vacc += vi3 * vk3;
181 const int32_t vi4 = (int32_t) *i4++;
182 const int32_t vk4 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[4];
183 vacc += vi4 * vk4;
184 const int32_t vi5 = (int32_t) *i5++;
185 const int32_t vk5 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[5];
186 vacc += vi5 * vk5;
187 const int32_t vi6 = (int32_t) *i6++;
188 const int32_t vk6 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[6];
189 vacc += vi6 * vk6;
190 const int32_t vi7 = (int32_t) *i7++;
191 const int32_t vk7 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[7];
192 vacc += vi7 * vk7;
193 const int32_t vi8 = (int32_t) *i8++;
194 const int32_t vk8 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[8];
195 vacc += vi8 * vk8;
196 const int32_t vi9 = (int32_t) *i9++;
197 const int32_t vk9 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[9];
198 vacc += vi9 * vk9;
199 const int32_t vi10 = (int32_t) *i10++;
200 const int32_t vk10 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[10];
201 vacc += vi10 * vk10;
202 const int32_t vi11 = (int32_t) *i11++;
203 const int32_t vk11 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[11];
204 vacc += vi11 * vk11;
205 const int32_t vi12 = (int32_t) *i12++;
206 const int32_t vk12 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[12];
207 vacc += vi12 * vk12;
208 const int32_t vi13 = (int32_t) *i13++;
209 const int32_t vk13 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[13];
210 vacc += vi13 * vk13;
211 const int32_t vi14 = (int32_t) *i14++;
212 const int32_t vk14 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[14];
213 vacc += vi14 * vk14;
214 const int32_t vi15 = (int32_t) *i15++;
215 const int32_t vk15 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[15];
216 vacc += vi15 * vk15;
217 const int32_t vi16 = (int32_t) *i16++;
218 const int32_t vk16 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[16];
219 vacc += vi16 * vk16;
220 const int32_t vi17 = (int32_t) *i17++;
221 const int32_t vk17 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[17];
222 vacc += vi17 * vk17;
223 const int32_t vi18 = (int32_t) *i18++;
224 const int32_t vk18 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[18];
225 vacc += vi18 * vk18;
226 const int32_t vi19 = (int32_t) *i19++;
227 const int32_t vk19 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[19];
228 vacc += vi19 * vk19;
229 const int32_t vi20 = (int32_t) *i20++;
230 const int32_t vk20 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[20];
231 vacc += vi20 * vk20;
232 const int32_t vi21 = (int32_t) *i21++;
233 const int32_t vk21 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[21];
234 vacc += vi21 * vk21;
235 const int32_t vi22 = (int32_t) *i22++;
236 const int32_t vk22 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[22];
237 vacc += vi22 * vk22;
238 const int32_t vi23 = (int32_t) *i23++;
239 const int32_t vk23 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[23];
240 vacc += vi23 * vk23;
241 const int32_t vi24 = (int32_t) *i24++;
242 const int32_t vk24 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[24];
243 vacc += vi24 * vk24;
244
245 w = (const void*) ((uintptr_t) w + sizeof(int32_t) + 25 * sizeof(int8_t));
246
247 const float vscale = unaligned_load_f32(w);
248 w = (const void*) ((const float*) w + 1);
249 float vfpacc = (float) vacc * vscale;
250
251 vfpacc += vmagic_bias;
252 int32_t vout = (int32_t) float_as_uint32(vfpacc);
253 vout = math_max_s32(vout, vmagic_min);
254 vout = math_min_s32(vout, vmagic_max);
255 vout -= vmagic_bias_less_zero_point;
256
257 *output++ = (int8_t) vout;
258 } while (--c != 0);
259
260 output = (int8_t*) ((uintptr_t) output + output_increment);
261 } while (--output_width != 0);
262 }
263