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 #include <math.h>
12
13 #include <xnnpack/dwconv.h>
14 #include <xnnpack/math.h>
15
16
xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf(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_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf(
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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29 assert(channels != 0);
30 assert(output_width != 0);
31
32 const float voutput_min_less_zero_point = params->scalar_lrintf.output_min_less_zero_point;
33 const float voutput_max_less_zero_point = params->scalar_lrintf.output_max_less_zero_point;
34 const int32_t voutput_zero_point = params->scalar_lrintf.output_zero_point;
35 do {
36 const int8_t* i0 = input[0];
37 assert(i0 != NULL);
38 if XNN_UNPREDICTABLE(i0 != zero) {
39 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
40 }
41 const int8_t* i1 = input[1];
42 assert(i1 != NULL);
43 if XNN_UNPREDICTABLE(i1 != zero) {
44 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
45 }
46 const int8_t* i2 = input[2];
47 assert(i2 != NULL);
48 if XNN_UNPREDICTABLE(i2 != zero) {
49 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
50 }
51 const int8_t* i3 = input[3];
52 assert(i3 != NULL);
53 if XNN_UNPREDICTABLE(i3 != zero) {
54 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
55 }
56 const int8_t* i4 = input[4];
57 assert(i4 != NULL);
58 if XNN_UNPREDICTABLE(i4 != zero) {
59 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
60 }
61 const int8_t* i5 = input[5];
62 assert(i5 != NULL);
63 if XNN_UNPREDICTABLE(i5 != zero) {
64 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
65 }
66 const int8_t* i6 = input[6];
67 assert(i6 != NULL);
68 if XNN_UNPREDICTABLE(i6 != zero) {
69 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
70 }
71 const int8_t* i7 = input[7];
72 assert(i7 != NULL);
73 if XNN_UNPREDICTABLE(i7 != zero) {
74 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
75 }
76 const int8_t* i8 = input[8];
77 assert(i8 != NULL);
78 if XNN_UNPREDICTABLE(i8 != zero) {
79 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
80 }
81 input = (const int8_t**) ((uintptr_t) input + input_stride);
82
83 size_t c = channels;
84 const void* w = weights;
85 for (; c >= 2; c -= 2) {
86 int32_t vacc0 = ((const int32_t*) w)[0];
87 int32_t vacc1 = ((const int32_t*) w)[1];
88
89
90 const int32_t vi0x0 = (int32_t) i0[0];
91 const int32_t vi0x1 = (int32_t) i0[1];
92 i0 += 2;
93
94 const int32_t vk0x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[0];
95 const int32_t vk0x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[1];
96
97 vacc0 += vi0x0 * vk0x0;
98 vacc1 += vi0x1 * vk0x1;
99
100 const int32_t vi1x0 = (int32_t) i1[0];
101 const int32_t vi1x1 = (int32_t) i1[1];
102 i1 += 2;
103
104 const int32_t vk1x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[2];
105 const int32_t vk1x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[3];
106
107 vacc0 += vi1x0 * vk1x0;
108 vacc1 += vi1x1 * vk1x1;
109
110 const int32_t vi2x0 = (int32_t) i2[0];
111 const int32_t vi2x1 = (int32_t) i2[1];
112 i2 += 2;
113
114 const int32_t vk2x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[4];
115 const int32_t vk2x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[5];
116
117 vacc0 += vi2x0 * vk2x0;
118 vacc1 += vi2x1 * vk2x1;
119
120 const int32_t vi3x0 = (int32_t) i3[0];
121 const int32_t vi3x1 = (int32_t) i3[1];
122 i3 += 2;
123
124 const int32_t vk3x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[6];
125 const int32_t vk3x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[7];
126
127 vacc0 += vi3x0 * vk3x0;
128 vacc1 += vi3x1 * vk3x1;
129
130 const int32_t vi4x0 = (int32_t) i4[0];
131 const int32_t vi4x1 = (int32_t) i4[1];
132 i4 += 2;
133
134 const int32_t vk4x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[8];
135 const int32_t vk4x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[9];
136
137 vacc0 += vi4x0 * vk4x0;
138 vacc1 += vi4x1 * vk4x1;
139
140 const int32_t vi5x0 = (int32_t) i5[0];
141 const int32_t vi5x1 = (int32_t) i5[1];
142 i5 += 2;
143
144 const int32_t vk5x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[10];
145 const int32_t vk5x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[11];
146
147 vacc0 += vi5x0 * vk5x0;
148 vacc1 += vi5x1 * vk5x1;
149
150 const int32_t vi6x0 = (int32_t) i6[0];
151 const int32_t vi6x1 = (int32_t) i6[1];
152 i6 += 2;
153
154 const int32_t vk6x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[12];
155 const int32_t vk6x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[13];
156
157 vacc0 += vi6x0 * vk6x0;
158 vacc1 += vi6x1 * vk6x1;
159
160 const int32_t vi7x0 = (int32_t) i7[0];
161 const int32_t vi7x1 = (int32_t) i7[1];
162 i7 += 2;
163
164 const int32_t vk7x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[14];
165 const int32_t vk7x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[15];
166
167 vacc0 += vi7x0 * vk7x0;
168 vacc1 += vi7x1 * vk7x1;
169
170 const int32_t vi8x0 = (int32_t) i8[0];
171 const int32_t vi8x1 = (int32_t) i8[1];
172 i8 += 2;
173
174 const int32_t vk8x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[16];
175 const int32_t vk8x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[17];
176
177 vacc0 += vi8x0 * vk8x0;
178 vacc1 += vi8x1 * vk8x1;
179
180 w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t) + 18 * sizeof(int8_t));
181
182 float vfpacc0 = (float) vacc0;
183 float vfpacc1 = (float) vacc1;
184
185 typedef XNN_UNALIGNED float unaligned_float;
186 const float vscale0 = ((const unaligned_float*) w)[0];
187 const float vscale1 = ((const unaligned_float*) w)[1];
188 w = (const void*) ((const float*) w + 2);
189
190 vfpacc0 *= vscale0;
191 vfpacc1 *= vscale1;
192
193 vfpacc0 = math_max_f32(vfpacc0, voutput_min_less_zero_point);
194 vfpacc1 = math_max_f32(vfpacc1, voutput_min_less_zero_point);
195
196 vfpacc0 = math_min_f32(vfpacc0, voutput_max_less_zero_point);
197 vfpacc1 = math_min_f32(vfpacc1, voutput_max_less_zero_point);
198
199 const int32_t vrndacc0 = (int32_t) lrintf(vfpacc0);
200 const int32_t vrndacc1 = (int32_t) lrintf(vfpacc1);
201
202 int32_t vout0 = (int32_t) vrndacc0 + voutput_zero_point;
203 int32_t vout1 = (int32_t) vrndacc1 + voutput_zero_point;
204
205 output[0] = (int8_t) vout0;
206 output[1] = (int8_t) vout1;
207 output += 2;
208 }
209 if XNN_UNLIKELY(c != 0) {
210 int32_t vacc = *((const int32_t*) w);
211
212 const int32_t vi0 = (int32_t) *i0;
213 const int32_t vk0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[0];
214 vacc += vi0 * vk0;
215 const int32_t vi1 = (int32_t) *i1;
216 const int32_t vk1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[2];
217 vacc += vi1 * vk1;
218 const int32_t vi2 = (int32_t) *i2;
219 const int32_t vk2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[4];
220 vacc += vi2 * vk2;
221 const int32_t vi3 = (int32_t) *i3;
222 const int32_t vk3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[6];
223 vacc += vi3 * vk3;
224 const int32_t vi4 = (int32_t) *i4;
225 const int32_t vk4 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[8];
226 vacc += vi4 * vk4;
227 const int32_t vi5 = (int32_t) *i5;
228 const int32_t vk5 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[10];
229 vacc += vi5 * vk5;
230 const int32_t vi6 = (int32_t) *i6;
231 const int32_t vk6 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[12];
232 vacc += vi6 * vk6;
233 const int32_t vi7 = (int32_t) *i7;
234 const int32_t vk7 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[14];
235 vacc += vi7 * vk7;
236 const int32_t vi8 = (int32_t) *i8;
237 const int32_t vk8 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[16];
238 vacc += vi8 * vk8;
239
240 typedef XNN_UNALIGNED float unaligned_float;
241 const float vscale = *((const unaligned_float*) ((uintptr_t) w + 2 * sizeof(int32_t) + 18 * sizeof(int8_t)));
242 float vfpacc = (float) vacc * vscale;
243
244 vfpacc = math_max_f32(vfpacc, voutput_min_less_zero_point);
245 vfpacc = math_min_f32(vfpacc, voutput_max_less_zero_point);
246 const int32_t vrndacc = (int32_t) lrintf(vfpacc);
247 int32_t vout = vrndacc + voutput_zero_point;
248
249 *output++ = (int8_t) vout;
250 }
251
252 output = (int8_t*) ((uintptr_t) output + output_increment);
253 } while (--output_width != 0);
254 }
255