Debian Codegraph @

src/qs8-dwconv/gen/qs8-dwconv-9p1c-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_9p1c__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 do {
93 int32_t vacc = unaligned_load_s32(w);
94
95 const int32_t vi0 = (int32_t) *i0++;
96 const int32_t vk0 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[0];
97 vacc += vi0 * vk0;
98 const int32_t vi1 = (int32_t) *i1++;
99 const int32_t vk1 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[1];
100 vacc += vi1 * vk1;
101 const int32_t vi2 = (int32_t) *i2++;
102 const int32_t vk2 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[2];
103 vacc += vi2 * vk2;
104 const int32_t vi3 = (int32_t) *i3++;
105 const int32_t vk3 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[3];
106 vacc += vi3 * vk3;
107 const int32_t vi4 = (int32_t) *i4++;
108 const int32_t vk4 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[4];
109 vacc += vi4 * vk4;
110 const int32_t vi5 = (int32_t) *i5++;
111 const int32_t vk5 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[5];
112 vacc += vi5 * vk5;
113 const int32_t vi6 = (int32_t) *i6++;
114 const int32_t vk6 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[6];
115 vacc += vi6 * vk6;
116 const int32_t vi7 = (int32_t) *i7++;
117 const int32_t vk7 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[7];
118 vacc += vi7 * vk7;
119 const int32_t vi8 = (int32_t) *i8++;
120 const int32_t vk8 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[8];
121 vacc += vi8 * vk8;
122
123 w = (const void*) ((uintptr_t) w + sizeof(int32_t) + 9 * sizeof(int8_t));
124
125 float vfpacc = (float) vacc * vscale;
126
127 vfpacc += vmagic_bias;
128 int32_t vout = (int32_t) float_as_uint32(vfpacc);
129 vout = math_max_s32(vout, vmagic_min);
130 vout = math_min_s32(vout, vmagic_max);
131 vout -= vmagic_bias_less_zero_point;
132
133 *output++ = (int8_t) vout;
134 } while (--c != 0);
135
136 output = (int8_t*) ((uintptr_t) output + output_increment);
137 } while (--output_width != 0);
138 }