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_qs8_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_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_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_qs8_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_imagic.scale;
33 const float vmagic_bias = params->fp32_scalar_imagic.magic_bias;
34 const int32_t vmagic_min = params->fp32_scalar_imagic.magic_min;
35 const int32_t vmagic_max = params->fp32_scalar_imagic.magic_max;
36 const int32_t vmagic_bias_less_zero_point = params->fp32_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 const int8_t* i9 = input[9];
84 assert(i9 != NULL);
85 if XNN_UNPREDICTABLE(i9 != zero) {
86 i9 = (const int8_t*) ((uintptr_t) i9 + input_offset);
87 }
88 const int8_t* i10 = input[10];
89 assert(i10 != NULL);
90 if XNN_UNPREDICTABLE(i10 != zero) {
91 i10 = (const int8_t*) ((uintptr_t) i10 + input_offset);
92 }
93 const int8_t* i11 = input[11];
94 assert(i11 != NULL);
95 if XNN_UNPREDICTABLE(i11 != zero) {
96 i11 = (const int8_t*) ((uintptr_t) i11 + input_offset);
97 }
98 const int8_t* i12 = input[12];
99 assert(i12 != NULL);
100 if XNN_UNPREDICTABLE(i12 != zero) {
101 i12 = (const int8_t*) ((uintptr_t) i12 + input_offset);
102 }
103 const int8_t* i13 = input[13];
104 assert(i13 != NULL);
105 if XNN_UNPREDICTABLE(i13 != zero) {
106 i13 = (const int8_t*) ((uintptr_t) i13 + input_offset);
107 }
108 const int8_t* i14 = input[14];
109 assert(i14 != NULL);
110 if XNN_UNPREDICTABLE(i14 != zero) {
111 i14 = (const int8_t*) ((uintptr_t) i14 + input_offset);
112 }
113 const int8_t* i15 = input[15];
114 assert(i15 != NULL);
115 if XNN_UNPREDICTABLE(i15 != zero) {
116 i15 = (const int8_t*) ((uintptr_t) i15 + input_offset);
117 }
118 const int8_t* i16 = input[16];
119 assert(i16 != NULL);
120 if XNN_UNPREDICTABLE(i16 != zero) {
121 i16 = (const int8_t*) ((uintptr_t) i16 + input_offset);
122 }
123 const int8_t* i17 = input[17];
124 assert(i17 != NULL);
125 if XNN_UNPREDICTABLE(i17 != zero) {
126 i17 = (const int8_t*) ((uintptr_t) i17 + input_offset);
127 }
128 const int8_t* i18 = input[18];
129 assert(i18 != NULL);
130 if XNN_UNPREDICTABLE(i18 != zero) {
131 i18 = (const int8_t*) ((uintptr_t) i18 + input_offset);
132 }
133 const int8_t* i19 = input[19];
134 assert(i19 != NULL);
135 if XNN_UNPREDICTABLE(i19 != zero) {
136 i19 = (const int8_t*) ((uintptr_t) i19 + input_offset);
137 }
138 const int8_t* i20 = input[20];
139 assert(i20 != NULL);
140 if XNN_UNPREDICTABLE(i20 != zero) {
141 i20 = (const int8_t*) ((uintptr_t) i20 + input_offset);
142 }
143 const int8_t* i21 = input[21];
144 assert(i21 != NULL);
145 if XNN_UNPREDICTABLE(i21 != zero) {
146 i21 = (const int8_t*) ((uintptr_t) i21 + input_offset);
147 }
148 const int8_t* i22 = input[22];
149 assert(i22 != NULL);
150 if XNN_UNPREDICTABLE(i22 != zero) {
151 i22 = (const int8_t*) ((uintptr_t) i22 + input_offset);
152 }
153 const int8_t* i23 = input[23];
154 assert(i23 != NULL);
155 if XNN_UNPREDICTABLE(i23 != zero) {
156 i23 = (const int8_t*) ((uintptr_t) i23 + input_offset);
157 }
158 const int8_t* i24 = input[24];
159 assert(i24 != NULL);
160 if XNN_UNPREDICTABLE(i24 != zero) {
161 i24 = (const int8_t*) ((uintptr_t) i24 + input_offset);
162 }
163 input = (const int8_t**) ((uintptr_t) input + input_stride);
164
165 size_t c = channels;
166 const void* w = weights;
167 do {
168 int32_t vacc = unaligned_load_s32(w);
169
170 const int32_t vi0 = (int32_t) *i0++;
171 const int32_t vk0 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[0];
172 vacc += vi0 * vk0;
173 const int32_t vi1 = (int32_t) *i1++;
174 const int32_t vk1 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[1];
175 vacc += vi1 * vk1;
176 const int32_t vi2 = (int32_t) *i2++;
177 const int32_t vk2 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[2];
178 vacc += vi2 * vk2;
179 const int32_t vi3 = (int32_t) *i3++;
180 const int32_t vk3 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[3];
181 vacc += vi3 * vk3;
182 const int32_t vi4 = (int32_t) *i4++;
183 const int32_t vk4 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[4];
184 vacc += vi4 * vk4;
185 const int32_t vi5 = (int32_t) *i5++;
186 const int32_t vk5 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[5];
187 vacc += vi5 * vk5;
188 const int32_t vi6 = (int32_t) *i6++;
189 const int32_t vk6 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[6];
190 vacc += vi6 * vk6;
191 const int32_t vi7 = (int32_t) *i7++;
192 const int32_t vk7 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[7];
193 vacc += vi7 * vk7;
194 const int32_t vi8 = (int32_t) *i8++;
195 const int32_t vk8 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[8];
196 vacc += vi8 * vk8;
197 const int32_t vi9 = (int32_t) *i9++;
198 const int32_t vk9 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[9];
199 vacc += vi9 * vk9;
200 const int32_t vi10 = (int32_t) *i10++;
201 const int32_t vk10 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[10];
202 vacc += vi10 * vk10;
203 const int32_t vi11 = (int32_t) *i11++;
204 const int32_t vk11 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[11];
205 vacc += vi11 * vk11;
206 const int32_t vi12 = (int32_t) *i12++;
207 const int32_t vk12 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[12];
208 vacc += vi12 * vk12;
209 const int32_t vi13 = (int32_t) *i13++;
210 const int32_t vk13 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[13];
211 vacc += vi13 * vk13;
212 const int32_t vi14 = (int32_t) *i14++;
213 const int32_t vk14 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[14];
214 vacc += vi14 * vk14;
215 const int32_t vi15 = (int32_t) *i15++;
216 const int32_t vk15 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[15];
217 vacc += vi15 * vk15;
218 const int32_t vi16 = (int32_t) *i16++;
219 const int32_t vk16 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[16];
220 vacc += vi16 * vk16;
221 const int32_t vi17 = (int32_t) *i17++;
222 const int32_t vk17 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[17];
223 vacc += vi17 * vk17;
224 const int32_t vi18 = (int32_t) *i18++;
225 const int32_t vk18 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[18];
226 vacc += vi18 * vk18;
227 const int32_t vi19 = (int32_t) *i19++;
228 const int32_t vk19 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[19];
229 vacc += vi19 * vk19;
230 const int32_t vi20 = (int32_t) *i20++;
231 const int32_t vk20 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[20];
232 vacc += vi20 * vk20;
233 const int32_t vi21 = (int32_t) *i21++;
234 const int32_t vk21 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[21];
235 vacc += vi21 * vk21;
236 const int32_t vi22 = (int32_t) *i22++;
237 const int32_t vk22 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[22];
238 vacc += vi22 * vk22;
239 const int32_t vi23 = (int32_t) *i23++;
240 const int32_t vk23 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[23];
241 vacc += vi23 * vk23;
242 const int32_t vi24 = (int32_t) *i24++;
243 const int32_t vk24 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[24];
244 vacc += vi24 * vk24;
245
246 w = (const void*) ((uintptr_t) w + sizeof(int32_t) + 25 * sizeof(int8_t));
247
248 float vfpacc = (float) vacc * vscale;
249
250 vfpacc += vmagic_bias;
251 int32_t vout = (int32_t) float_as_uint32(vfpacc);
252 vout = math_max_s32(vout, vmagic_min);
253 vout = math_min_s32(vout, vmagic_max);
254 vout -= vmagic_bias_less_zero_point;
255
256 *output++ = (int8_t) vout;
257 } while (--c != 0);
258
259 output = (int8_t*) ((uintptr_t) output + output_increment);
260 } while (--output_width != 0);
261 }
262