• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 <fp16.h>
13 
14 #include <xnnpack/dwconv.h>
15 #include <xnnpack/math.h>
16 
17 
xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic(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)])18 void xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic(
19     size_t channels,
20     size_t output_width,
21     const uint8_t** input,
22     const void* weights,
23     uint8_t* output,
24     size_t input_stride,
25     size_t output_increment,
26     size_t input_offset,
27     const uint8_t* zero,
28     const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
29 {
30   assert(channels != 0);
31   assert(output_width != 0);
32 
33   const float vscale = params->fp32_scalar_imagic.scale;
34   const float vmagic_bias = params->fp32_scalar_imagic.magic_bias;
35   const int32_t vmagic_min = params->fp32_scalar_imagic.magic_min;
36   const int32_t vmagic_max = params->fp32_scalar_imagic.magic_max;
37   const int32_t vmagic_bias_less_zero_point = params->fp32_scalar_imagic.magic_bias_less_zero_point;
38   const int32_t vkernel_zero_point = params->fp32_scalar_imagic.kernel_zero_point;
39   do {
40     const uint8_t* i0 = input[0];
41     assert(i0 != NULL);
42     if XNN_UNPREDICTABLE(i0 != zero) {
43       i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
44     }
45     const uint8_t* i1 = input[1];
46     assert(i1 != NULL);
47     if XNN_UNPREDICTABLE(i1 != zero) {
48       i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
49     }
50     const uint8_t* i2 = input[2];
51     assert(i2 != NULL);
52     if XNN_UNPREDICTABLE(i2 != zero) {
53       i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
54     }
55     const uint8_t* i3 = input[3];
56     assert(i3 != NULL);
57     if XNN_UNPREDICTABLE(i3 != zero) {
58       i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
59     }
60     const uint8_t* i4 = input[4];
61     assert(i4 != NULL);
62     if XNN_UNPREDICTABLE(i4 != zero) {
63       i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
64     }
65     const uint8_t* i5 = input[5];
66     assert(i5 != NULL);
67     if XNN_UNPREDICTABLE(i5 != zero) {
68       i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
69     }
70     const uint8_t* i6 = input[6];
71     assert(i6 != NULL);
72     if XNN_UNPREDICTABLE(i6 != zero) {
73       i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
74     }
75     const uint8_t* i7 = input[7];
76     assert(i7 != NULL);
77     if XNN_UNPREDICTABLE(i7 != zero) {
78       i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
79     }
80     const uint8_t* i8 = input[8];
81     assert(i8 != NULL);
82     if XNN_UNPREDICTABLE(i8 != zero) {
83       i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
84     }
85     const uint8_t* i9 = input[9];
86     assert(i9 != NULL);
87     if XNN_UNPREDICTABLE(i9 != zero) {
88       i9 = (const uint8_t*) ((uintptr_t) i9 + input_offset);
89     }
90     const uint8_t* i10 = input[10];
91     assert(i10 != NULL);
92     if XNN_UNPREDICTABLE(i10 != zero) {
93       i10 = (const uint8_t*) ((uintptr_t) i10 + input_offset);
94     }
95     const uint8_t* i11 = input[11];
96     assert(i11 != NULL);
97     if XNN_UNPREDICTABLE(i11 != zero) {
98       i11 = (const uint8_t*) ((uintptr_t) i11 + input_offset);
99     }
100     const uint8_t* i12 = input[12];
101     assert(i12 != NULL);
102     if XNN_UNPREDICTABLE(i12 != zero) {
103       i12 = (const uint8_t*) ((uintptr_t) i12 + input_offset);
104     }
105     const uint8_t* i13 = input[13];
106     assert(i13 != NULL);
107     if XNN_UNPREDICTABLE(i13 != zero) {
108       i13 = (const uint8_t*) ((uintptr_t) i13 + input_offset);
109     }
110     const uint8_t* i14 = input[14];
111     assert(i14 != NULL);
112     if XNN_UNPREDICTABLE(i14 != zero) {
113       i14 = (const uint8_t*) ((uintptr_t) i14 + input_offset);
114     }
115     const uint8_t* i15 = input[15];
116     assert(i15 != NULL);
117     if XNN_UNPREDICTABLE(i15 != zero) {
118       i15 = (const uint8_t*) ((uintptr_t) i15 + input_offset);
119     }
120     const uint8_t* i16 = input[16];
121     assert(i16 != NULL);
122     if XNN_UNPREDICTABLE(i16 != zero) {
123       i16 = (const uint8_t*) ((uintptr_t) i16 + input_offset);
124     }
125     const uint8_t* i17 = input[17];
126     assert(i17 != NULL);
127     if XNN_UNPREDICTABLE(i17 != zero) {
128       i17 = (const uint8_t*) ((uintptr_t) i17 + input_offset);
129     }
130     const uint8_t* i18 = input[18];
131     assert(i18 != NULL);
132     if XNN_UNPREDICTABLE(i18 != zero) {
133       i18 = (const uint8_t*) ((uintptr_t) i18 + input_offset);
134     }
135     const uint8_t* i19 = input[19];
136     assert(i19 != NULL);
137     if XNN_UNPREDICTABLE(i19 != zero) {
138       i19 = (const uint8_t*) ((uintptr_t) i19 + input_offset);
139     }
140     const uint8_t* i20 = input[20];
141     assert(i20 != NULL);
142     if XNN_UNPREDICTABLE(i20 != zero) {
143       i20 = (const uint8_t*) ((uintptr_t) i20 + input_offset);
144     }
145     const uint8_t* i21 = input[21];
146     assert(i21 != NULL);
147     if XNN_UNPREDICTABLE(i21 != zero) {
148       i21 = (const uint8_t*) ((uintptr_t) i21 + input_offset);
149     }
150     const uint8_t* i22 = input[22];
151     assert(i22 != NULL);
152     if XNN_UNPREDICTABLE(i22 != zero) {
153       i22 = (const uint8_t*) ((uintptr_t) i22 + input_offset);
154     }
155     const uint8_t* i23 = input[23];
156     assert(i23 != NULL);
157     if XNN_UNPREDICTABLE(i23 != zero) {
158       i23 = (const uint8_t*) ((uintptr_t) i23 + input_offset);
159     }
160     const uint8_t* i24 = input[24];
161     assert(i24 != NULL);
162     if XNN_UNPREDICTABLE(i24 != zero) {
163       i24 = (const uint8_t*) ((uintptr_t) i24 + input_offset);
164     }
165     input = (const uint8_t**) ((uintptr_t) input + input_stride);
166 
167     size_t c = channels;
168     const void* w = weights;
169     do {
170       int32_t vacc = *((const int32_t*) w);
171 
172       const int32_t vi0 = (int32_t) (uint32_t) *i0++;
173       const int32_t vk0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[0] - vkernel_zero_point;
174       vacc += vi0 * vk0;
175       const int32_t vi1 = (int32_t) (uint32_t) *i1++;
176       const int32_t vk1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[1] - vkernel_zero_point;
177       vacc += vi1 * vk1;
178       const int32_t vi2 = (int32_t) (uint32_t) *i2++;
179       const int32_t vk2 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[2] - vkernel_zero_point;
180       vacc += vi2 * vk2;
181       const int32_t vi3 = (int32_t) (uint32_t) *i3++;
182       const int32_t vk3 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[3] - vkernel_zero_point;
183       vacc += vi3 * vk3;
184       const int32_t vi4 = (int32_t) (uint32_t) *i4++;
185       const int32_t vk4 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[4] - vkernel_zero_point;
186       vacc += vi4 * vk4;
187       const int32_t vi5 = (int32_t) (uint32_t) *i5++;
188       const int32_t vk5 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[5] - vkernel_zero_point;
189       vacc += vi5 * vk5;
190       const int32_t vi6 = (int32_t) (uint32_t) *i6++;
191       const int32_t vk6 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[6] - vkernel_zero_point;
192       vacc += vi6 * vk6;
193       const int32_t vi7 = (int32_t) (uint32_t) *i7++;
194       const int32_t vk7 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[7] - vkernel_zero_point;
195       vacc += vi7 * vk7;
196       const int32_t vi8 = (int32_t) (uint32_t) *i8++;
197       const int32_t vk8 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[8] - vkernel_zero_point;
198       vacc += vi8 * vk8;
199       const int32_t vi9 = (int32_t) (uint32_t) *i9++;
200       const int32_t vk9 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[9] - vkernel_zero_point;
201       vacc += vi9 * vk9;
202       const int32_t vi10 = (int32_t) (uint32_t) *i10++;
203       const int32_t vk10 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[10] - vkernel_zero_point;
204       vacc += vi10 * vk10;
205       const int32_t vi11 = (int32_t) (uint32_t) *i11++;
206       const int32_t vk11 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[11] - vkernel_zero_point;
207       vacc += vi11 * vk11;
208       const int32_t vi12 = (int32_t) (uint32_t) *i12++;
209       const int32_t vk12 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[12] - vkernel_zero_point;
210       vacc += vi12 * vk12;
211       const int32_t vi13 = (int32_t) (uint32_t) *i13++;
212       const int32_t vk13 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[13] - vkernel_zero_point;
213       vacc += vi13 * vk13;
214       const int32_t vi14 = (int32_t) (uint32_t) *i14++;
215       const int32_t vk14 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[14] - vkernel_zero_point;
216       vacc += vi14 * vk14;
217       const int32_t vi15 = (int32_t) (uint32_t) *i15++;
218       const int32_t vk15 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[15] - vkernel_zero_point;
219       vacc += vi15 * vk15;
220       const int32_t vi16 = (int32_t) (uint32_t) *i16++;
221       const int32_t vk16 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[16] - vkernel_zero_point;
222       vacc += vi16 * vk16;
223       const int32_t vi17 = (int32_t) (uint32_t) *i17++;
224       const int32_t vk17 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[17] - vkernel_zero_point;
225       vacc += vi17 * vk17;
226       const int32_t vi18 = (int32_t) (uint32_t) *i18++;
227       const int32_t vk18 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[18] - vkernel_zero_point;
228       vacc += vi18 * vk18;
229       const int32_t vi19 = (int32_t) (uint32_t) *i19++;
230       const int32_t vk19 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[19] - vkernel_zero_point;
231       vacc += vi19 * vk19;
232       const int32_t vi20 = (int32_t) (uint32_t) *i20++;
233       const int32_t vk20 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[20] - vkernel_zero_point;
234       vacc += vi20 * vk20;
235       const int32_t vi21 = (int32_t) (uint32_t) *i21++;
236       const int32_t vk21 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[21] - vkernel_zero_point;
237       vacc += vi21 * vk21;
238       const int32_t vi22 = (int32_t) (uint32_t) *i22++;
239       const int32_t vk22 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[22] - vkernel_zero_point;
240       vacc += vi22 * vk22;
241       const int32_t vi23 = (int32_t) (uint32_t) *i23++;
242       const int32_t vk23 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[23] - vkernel_zero_point;
243       vacc += vi23 * vk23;
244       const int32_t vi24 = (int32_t) (uint32_t) *i24++;
245       const int32_t vk24 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[24] - vkernel_zero_point;
246       vacc += vi24 * vk24;
247 
248       w = (const void*) ((uintptr_t) w + sizeof(int32_t) + 25 * sizeof(uint8_t));
249 
250       float vfpacc = (float) vacc * vscale;
251 
252       vfpacc += vmagic_bias;
253       int32_t vout = (int32_t) fp32_to_bits(vfpacc);
254       vout = math_max_s32(vout, vmagic_min);
255       vout = math_min_s32(vout, vmagic_max);
256       vout -= vmagic_bias_less_zero_point;
257 
258       *output++ = (uint8_t) vout;
259     } while (--c != 0);
260 
261     output = (uint8_t*) ((uintptr_t) output + output_increment);
262   } while (--output_width != 0);
263 }
264