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_up1x9__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_up1x9__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 input = (const uint8_t**) ((uintptr_t) input + input_stride);
86
87 size_t c = channels;
88 const void* w = weights;
89 do {
90 int32_t vacc = *((const int32_t*) w);
91
92 const int32_t vi0 = (int32_t) (uint32_t) *i0++;
93 const int32_t vk0 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[0] - vkernel_zero_point;
94 vacc += vi0 * vk0;
95 const int32_t vi1 = (int32_t) (uint32_t) *i1++;
96 const int32_t vk1 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[1] - vkernel_zero_point;
97 vacc += vi1 * vk1;
98 const int32_t vi2 = (int32_t) (uint32_t) *i2++;
99 const int32_t vk2 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[2] - vkernel_zero_point;
100 vacc += vi2 * vk2;
101 const int32_t vi3 = (int32_t) (uint32_t) *i3++;
102 const int32_t vk3 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[3] - vkernel_zero_point;
103 vacc += vi3 * vk3;
104 const int32_t vi4 = (int32_t) (uint32_t) *i4++;
105 const int32_t vk4 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[4] - vkernel_zero_point;
106 vacc += vi4 * vk4;
107 const int32_t vi5 = (int32_t) (uint32_t) *i5++;
108 const int32_t vk5 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[5] - vkernel_zero_point;
109 vacc += vi5 * vk5;
110 const int32_t vi6 = (int32_t) (uint32_t) *i6++;
111 const int32_t vk6 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[6] - vkernel_zero_point;
112 vacc += vi6 * vk6;
113 const int32_t vi7 = (int32_t) (uint32_t) *i7++;
114 const int32_t vk7 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[7] - vkernel_zero_point;
115 vacc += vi7 * vk7;
116 const int32_t vi8 = (int32_t) (uint32_t) *i8++;
117 const int32_t vk8 = (int32_t) (uint32_t) ((const uint8_t*) ((uintptr_t) w + sizeof(int32_t)))[8] - vkernel_zero_point;
118 vacc += vi8 * vk8;
119
120 w = (const void*) ((uintptr_t) w + sizeof(int32_t) + 9 * sizeof(uint8_t));
121
122 float vfpacc = (float) vacc * vscale;
123
124 vfpacc += vmagic_bias;
125 int32_t vout = (int32_t) fp32_to_bits(vfpacc);
126 vout = math_max_s32(vout, vmagic_min);
127 vout = math_min_s32(vout, vmagic_max);
128 vout -= vmagic_bias_less_zero_point;
129
130 *output++ = (uint8_t) vout;
131 } while (--c != 0);
132
133 output = (uint8_t*) ((uintptr_t) output + output_increment);
134 } while (--output_width != 0);
135 }
136