Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-9p2c-minmax-rndnu-scalar.c

raw
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 <stddef.h>
12 #include <stdint.h>
13
14 #include "xnnpack/common.h"
15 #include "xnnpack/dwconv.h"
16 #include "xnnpack/math.h"
17 #include "xnnpack/microparams.h"
18 #include "xnnpack/unaligned.h"
19
20 void xnn_qs8_dwconv_minmax_rndnu_ukernel_9p2c__scalar(
21 size_t channels,
22 size_t output_width,
23 const int8_t** input,
24 const void* weights,
25 int8_t* output,
26 intptr_t input_stride,
27 size_t output_increment,
28 size_t input_offset,
29 const int8_t* zero,
30 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
31 {
32 assert(channels != 0);
33 assert(output_width != 0);
34
35 const int32_t vmultiplier = params->rndnu_scalar.multiplier;
36 const int64_t vrounding = params->rndnu_scalar.rounding;
37 const uint32_t vshift = params->rndnu_scalar.shift;
38 const int32_t voutput_min_less_zero_point = (int32_t) params->rndnu_scalar.output_min - (int32_t) params->rndnu_scalar.output_zero_point;
39 const int32_t voutput_max_less_zero_point = (int32_t) params->rndnu_scalar.output_max - (int32_t) params->rndnu_scalar.output_zero_point;
40 const int32_t voutput_zero_point = params->rndnu_scalar.output_zero_point;
41 do {
42 const int8_t* i0 = input[0];
43 assert(i0 != NULL);
44 if XNN_UNPREDICTABLE(i0 != zero) {
45 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
46 }
47 const int8_t* i1 = input[1];
48 assert(i1 != NULL);
49 if XNN_UNPREDICTABLE(i1 != zero) {
50 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
51 }
52 const int8_t* i2 = input[2];
53 assert(i2 != NULL);
54 if XNN_UNPREDICTABLE(i2 != zero) {
55 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
56 }
57 const int8_t* i3 = input[3];
58 assert(i3 != NULL);
59 if XNN_UNPREDICTABLE(i3 != zero) {
60 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
61 }
62 const int8_t* i4 = input[4];
63 assert(i4 != NULL);
64 if XNN_UNPREDICTABLE(i4 != zero) {
65 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
66 }
67 const int8_t* i5 = input[5];
68 assert(i5 != NULL);
69 if XNN_UNPREDICTABLE(i5 != zero) {
70 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
71 }
72 const int8_t* i6 = input[6];
73 assert(i6 != NULL);
74 if XNN_UNPREDICTABLE(i6 != zero) {
75 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
76 }
77 const int8_t* i7 = input[7];
78 assert(i7 != NULL);
79 if XNN_UNPREDICTABLE(i7 != zero) {
80 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
81 }
82 const int8_t* i8 = input[8];
83 assert(i8 != NULL);
84 if XNN_UNPREDICTABLE(i8 != zero) {
85 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
86 }
87 input = (const int8_t**) ((uintptr_t) input + input_stride);
88
89 size_t c = channels;
90 const void* w = weights;
91 for (; c >= 2; c -= 2) {
92 int32_t vacc0 = unaligned_indexed_load_s32(w, 0);
93 int32_t vacc1 = unaligned_indexed_load_s32(w, 1);
94
95
96 const int32_t vi0x0 = (int32_t) i0[0];
97 const int32_t vi0x1 = (int32_t) i0[1];
98 i0 += 2;
99
100 const int32_t vk0x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[0];
101 const int32_t vk0x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[1];
102
103 vacc0 += vi0x0 * vk0x0;
104 vacc1 += vi0x1 * vk0x1;
105
106 const int32_t vi1x0 = (int32_t) i1[0];
107 const int32_t vi1x1 = (int32_t) i1[1];
108 i1 += 2;
109
110 const int32_t vk1x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[2];
111 const int32_t vk1x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[3];
112
113 vacc0 += vi1x0 * vk1x0;
114 vacc1 += vi1x1 * vk1x1;
115
116 const int32_t vi2x0 = (int32_t) i2[0];
117 const int32_t vi2x1 = (int32_t) i2[1];
118 i2 += 2;
119
120 const int32_t vk2x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[4];
121 const int32_t vk2x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[5];
122
123 vacc0 += vi2x0 * vk2x0;
124 vacc1 += vi2x1 * vk2x1;
125
126 const int32_t vi3x0 = (int32_t) i3[0];
127 const int32_t vi3x1 = (int32_t) i3[1];
128 i3 += 2;
129
130 const int32_t vk3x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[6];
131 const int32_t vk3x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[7];
132
133 vacc0 += vi3x0 * vk3x0;
134 vacc1 += vi3x1 * vk3x1;
135
136 const int32_t vi4x0 = (int32_t) i4[0];
137 const int32_t vi4x1 = (int32_t) i4[1];
138 i4 += 2;
139
140 const int32_t vk4x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[8];
141 const int32_t vk4x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[9];
142
143 vacc0 += vi4x0 * vk4x0;
144 vacc1 += vi4x1 * vk4x1;
145
146 const int32_t vi5x0 = (int32_t) i5[0];
147 const int32_t vi5x1 = (int32_t) i5[1];
148 i5 += 2;
149
150 const int32_t vk5x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[10];
151 const int32_t vk5x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[11];
152
153 vacc0 += vi5x0 * vk5x0;
154 vacc1 += vi5x1 * vk5x1;
155
156 const int32_t vi6x0 = (int32_t) i6[0];
157 const int32_t vi6x1 = (int32_t) i6[1];
158 i6 += 2;
159
160 const int32_t vk6x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[12];
161 const int32_t vk6x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[13];
162
163 vacc0 += vi6x0 * vk6x0;
164 vacc1 += vi6x1 * vk6x1;
165
166 const int32_t vi7x0 = (int32_t) i7[0];
167 const int32_t vi7x1 = (int32_t) i7[1];
168 i7 += 2;
169
170 const int32_t vk7x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[14];
171 const int32_t vk7x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[15];
172
173 vacc0 += vi7x0 * vk7x0;
174 vacc1 += vi7x1 * vk7x1;
175
176 const int32_t vi8x0 = (int32_t) i8[0];
177 const int32_t vi8x1 = (int32_t) i8[1];
178 i8 += 2;
179
180 const int32_t vk8x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[16];
181 const int32_t vk8x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[17];
182
183 vacc0 += vi8x0 * vk8x0;
184 vacc1 += vi8x1 * vk8x1;
185
186 w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t) + 18 * sizeof(int8_t));
187
188 const int64_t vextacc0 = math_mulext_s32(vacc0, vmultiplier) + vrounding;
189 const int64_t vextacc1 = math_mulext_s32(vacc1, vmultiplier) + vrounding;
190
191 int32_t vout0 = (int32_t) math_asr_s64(vextacc0, vshift);
192 int32_t vout1 = (int32_t) math_asr_s64(vextacc1, vshift);
193
194 vout0 = math_max_s32(vout0, voutput_min_less_zero_point);
195 vout1 = math_max_s32(vout1, voutput_min_less_zero_point);
196
197 vout0 = math_min_s32(vout0, voutput_max_less_zero_point);
198 vout1 = math_min_s32(vout1, voutput_max_less_zero_point);
199
200 vout0 += voutput_zero_point;
201 vout1 += voutput_zero_point;
202
203 output[0] = (int8_t) vout0;
204 output[1] = (int8_t) vout1;
205 output += 2;
206 }
207 if XNN_UNLIKELY(c != 0) {
208 int32_t vacc = unaligned_load_s32(w);
209
210 const int32_t vi0 = (int32_t) *i0;
211 const int32_t vk0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[0];
212 vacc += vi0 * vk0;
213 const int32_t vi1 = (int32_t) *i1;
214 const int32_t vk1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[2];
215 vacc += vi1 * vk1;
216 const int32_t vi2 = (int32_t) *i2;
217 const int32_t vk2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[4];
218 vacc += vi2 * vk2;
219 const int32_t vi3 = (int32_t) *i3;
220 const int32_t vk3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[6];
221 vacc += vi3 * vk3;
222 const int32_t vi4 = (int32_t) *i4;
223 const int32_t vk4 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[8];
224 vacc += vi4 * vk4;
225 const int32_t vi5 = (int32_t) *i5;
226 const int32_t vk5 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[10];
227 vacc += vi5 * vk5;
228 const int32_t vi6 = (int32_t) *i6;
229 const int32_t vk6 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[12];
230 vacc += vi6 * vk6;
231 const int32_t vi7 = (int32_t) *i7;
232 const int32_t vk7 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[14];
233 vacc += vi7 * vk7;
234 const int32_t vi8 = (int32_t) *i8;
235 const int32_t vk8 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[16];
236 vacc += vi8 * vk8;
237
238 const int64_t vextacc = math_mulext_s32(vacc, vmultiplier) + vrounding;
239 int32_t vout = (int32_t) math_asr_s64(vextacc, vshift);
240 vout = math_max_s32(vout, voutput_min_less_zero_point);
241 vout = math_min_s32(vout, voutput_max_less_zero_point);
242 vout += voutput_zero_point;
243
244 *output++ = (int8_t) vout;
245 }
246
247 output = (int8_t*) ((uintptr_t) output + output_increment);
248 } while (--output_width != 0);
249 }