Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-9p2c-minmax-fp32-scalar-imagic.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_fp32_ukernel_9p2c__scalar_imagic(
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 float vscale = params->fp32_scalar.scale;
36 const float vmagic_bias = 12582912.0f;
37 const int32_t output_min_less_zero_point = (int32_t) params->fp32_scalar.output_min - (int32_t) params->fp32_scalar.output_zero_point;
38 const int32_t output_max_less_zero_point = (int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point;
39 const int32_t vmagic_min = (int32_t) float_as_uint32(12582912.0f + output_min_less_zero_point);
40 const int32_t vmagic_max = (int32_t) float_as_uint32(12582912.0f + output_max_less_zero_point);
41 const int32_t vmagic_bias_less_zero_point = INT32_C(0x4B400000) - (int32_t) params->fp32_scalar.output_zero_point;
42 do {
43 const int8_t* i0 = input[0];
44 assert(i0 != NULL);
45 if XNN_UNPREDICTABLE(i0 != zero) {
46 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
47 }
48 const int8_t* i1 = input[1];
49 assert(i1 != NULL);
50 if XNN_UNPREDICTABLE(i1 != zero) {
51 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
52 }
53 const int8_t* i2 = input[2];
54 assert(i2 != NULL);
55 if XNN_UNPREDICTABLE(i2 != zero) {
56 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
57 }
58 const int8_t* i3 = input[3];
59 assert(i3 != NULL);
60 if XNN_UNPREDICTABLE(i3 != zero) {
61 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
62 }
63 const int8_t* i4 = input[4];
64 assert(i4 != NULL);
65 if XNN_UNPREDICTABLE(i4 != zero) {
66 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
67 }
68 const int8_t* i5 = input[5];
69 assert(i5 != NULL);
70 if XNN_UNPREDICTABLE(i5 != zero) {
71 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
72 }
73 const int8_t* i6 = input[6];
74 assert(i6 != NULL);
75 if XNN_UNPREDICTABLE(i6 != zero) {
76 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
77 }
78 const int8_t* i7 = input[7];
79 assert(i7 != NULL);
80 if XNN_UNPREDICTABLE(i7 != zero) {
81 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
82 }
83 const int8_t* i8 = input[8];
84 assert(i8 != NULL);
85 if XNN_UNPREDICTABLE(i8 != zero) {
86 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
87 }
88 input = (const int8_t**) ((uintptr_t) input + input_stride);
89
90 size_t c = channels;
91 const void* w = weights;
92 for (; c >= 2; c -= 2) {
93 int32_t vacc0 = unaligned_indexed_load_s32(w, 0);
94 int32_t vacc1 = unaligned_indexed_load_s32(w, 1);
95
96
97 const int32_t vi0x0 = (int32_t) i0[0];
98 const int32_t vi0x1 = (int32_t) i0[1];
99 i0 += 2;
100
101 const int32_t vk0x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[0];
102 const int32_t vk0x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[1];
103
104 vacc0 += vi0x0 * vk0x0;
105 vacc1 += vi0x1 * vk0x1;
106
107 const int32_t vi1x0 = (int32_t) i1[0];
108 const int32_t vi1x1 = (int32_t) i1[1];
109 i1 += 2;
110
111 const int32_t vk1x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[2];
112 const int32_t vk1x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[3];
113
114 vacc0 += vi1x0 * vk1x0;
115 vacc1 += vi1x1 * vk1x1;
116
117 const int32_t vi2x0 = (int32_t) i2[0];
118 const int32_t vi2x1 = (int32_t) i2[1];
119 i2 += 2;
120
121 const int32_t vk2x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[4];
122 const int32_t vk2x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[5];
123
124 vacc0 += vi2x0 * vk2x0;
125 vacc1 += vi2x1 * vk2x1;
126
127 const int32_t vi3x0 = (int32_t) i3[0];
128 const int32_t vi3x1 = (int32_t) i3[1];
129 i3 += 2;
130
131 const int32_t vk3x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[6];
132 const int32_t vk3x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[7];
133
134 vacc0 += vi3x0 * vk3x0;
135 vacc1 += vi3x1 * vk3x1;
136
137 const int32_t vi4x0 = (int32_t) i4[0];
138 const int32_t vi4x1 = (int32_t) i4[1];
139 i4 += 2;
140
141 const int32_t vk4x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[8];
142 const int32_t vk4x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[9];
143
144 vacc0 += vi4x0 * vk4x0;
145 vacc1 += vi4x1 * vk4x1;
146
147 const int32_t vi5x0 = (int32_t) i5[0];
148 const int32_t vi5x1 = (int32_t) i5[1];
149 i5 += 2;
150
151 const int32_t vk5x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[10];
152 const int32_t vk5x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[11];
153
154 vacc0 += vi5x0 * vk5x0;
155 vacc1 += vi5x1 * vk5x1;
156
157 const int32_t vi6x0 = (int32_t) i6[0];
158 const int32_t vi6x1 = (int32_t) i6[1];
159 i6 += 2;
160
161 const int32_t vk6x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[12];
162 const int32_t vk6x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[13];
163
164 vacc0 += vi6x0 * vk6x0;
165 vacc1 += vi6x1 * vk6x1;
166
167 const int32_t vi7x0 = (int32_t) i7[0];
168 const int32_t vi7x1 = (int32_t) i7[1];
169 i7 += 2;
170
171 const int32_t vk7x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[14];
172 const int32_t vk7x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[15];
173
174 vacc0 += vi7x0 * vk7x0;
175 vacc1 += vi7x1 * vk7x1;
176
177 const int32_t vi8x0 = (int32_t) i8[0];
178 const int32_t vi8x1 = (int32_t) i8[1];
179 i8 += 2;
180
181 const int32_t vk8x0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[16];
182 const int32_t vk8x1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[17];
183
184 vacc0 += vi8x0 * vk8x0;
185 vacc1 += vi8x1 * vk8x1;
186
187 w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t) + 18 * sizeof(int8_t));
188
189 float vfpacc0 = (float) vacc0;
190 float vfpacc1 = (float) vacc1;
191
192 vfpacc0 *= vscale;
193 vfpacc1 *= vscale;
194
195 vfpacc0 += vmagic_bias;
196 vfpacc1 += vmagic_bias;
197
198 int32_t vout0 = (int32_t) float_as_uint32(vfpacc0);
199 int32_t vout1 = (int32_t) float_as_uint32(vfpacc1);
200
201 vout0 = math_max_s32(vout0, vmagic_min);
202 vout1 = math_max_s32(vout1, vmagic_min);
203
204 vout0 = math_min_s32(vout0, vmagic_max);
205 vout1 = math_min_s32(vout1, vmagic_max);
206
207 vout0 -= vmagic_bias_less_zero_point;
208 vout1 -= vmagic_bias_less_zero_point;
209
210 output[0] = (int8_t) vout0;
211 output[1] = (int8_t) vout1;
212 output += 2;
213 }
214 if XNN_UNLIKELY(c != 0) {
215 int32_t vacc = unaligned_load_s32(w);
216
217 const int32_t vi0 = (int32_t) *i0;
218 const int32_t vk0 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[0];
219 vacc += vi0 * vk0;
220 const int32_t vi1 = (int32_t) *i1;
221 const int32_t vk1 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[2];
222 vacc += vi1 * vk1;
223 const int32_t vi2 = (int32_t) *i2;
224 const int32_t vk2 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[4];
225 vacc += vi2 * vk2;
226 const int32_t vi3 = (int32_t) *i3;
227 const int32_t vk3 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[6];
228 vacc += vi3 * vk3;
229 const int32_t vi4 = (int32_t) *i4;
230 const int32_t vk4 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[8];
231 vacc += vi4 * vk4;
232 const int32_t vi5 = (int32_t) *i5;
233 const int32_t vk5 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[10];
234 vacc += vi5 * vk5;
235 const int32_t vi6 = (int32_t) *i6;
236 const int32_t vk6 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[12];
237 vacc += vi6 * vk6;
238 const int32_t vi7 = (int32_t) *i7;
239 const int32_t vk7 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[14];
240 vacc += vi7 * vk7;
241 const int32_t vi8 = (int32_t) *i8;
242 const int32_t vk8 = (int32_t) ((const int8_t*) ((uintptr_t) w + 2 * sizeof(int32_t)))[16];
243 vacc += vi8 * vk8;
244
245 float vfpacc = (float) vacc * vscale;
246
247 vfpacc += vmagic_bias;
248 int32_t vout = (int32_t) float_as_uint32(vfpacc);
249 vout = math_max_s32(vout, vmagic_min);
250 vout = math_min_s32(vout, vmagic_max);
251 vout -= vmagic_bias_less_zero_point;
252
253 *output++ = (int8_t) vout;
254 }
255
256 output = (int8_t*) ((uintptr_t) output + output_increment);
257 } while (--output_width != 0);
258 }