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_up1x9__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_up1x9__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 do {
86 int32_t vacc = *((const int32_t*) w);
87
88 const int32_t vi0 = (int32_t) *i0++;
89 const int32_t vk0 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[0];
90 vacc += vi0 * vk0;
91 const int32_t vi1 = (int32_t) *i1++;
92 const int32_t vk1 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[1];
93 vacc += vi1 * vk1;
94 const int32_t vi2 = (int32_t) *i2++;
95 const int32_t vk2 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[2];
96 vacc += vi2 * vk2;
97 const int32_t vi3 = (int32_t) *i3++;
98 const int32_t vk3 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[3];
99 vacc += vi3 * vk3;
100 const int32_t vi4 = (int32_t) *i4++;
101 const int32_t vk4 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[4];
102 vacc += vi4 * vk4;
103 const int32_t vi5 = (int32_t) *i5++;
104 const int32_t vk5 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[5];
105 vacc += vi5 * vk5;
106 const int32_t vi6 = (int32_t) *i6++;
107 const int32_t vk6 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[6];
108 vacc += vi6 * vk6;
109 const int32_t vi7 = (int32_t) *i7++;
110 const int32_t vk7 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[7];
111 vacc += vi7 * vk7;
112 const int32_t vi8 = (int32_t) *i8++;
113 const int32_t vk8 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[8];
114 vacc += vi8 * vk8;
115
116 w = (const void*) ((uintptr_t) w + sizeof(int32_t) + 9 * sizeof(int8_t));
117
118 typedef XNN_UNALIGNED float unaligned_float;
119 const float vscale = *((const unaligned_float*) w);
120 w = (const void*) ((const float*) w + 1);
121 float vfpacc = (float) vacc * vscale;
122
123 vfpacc = math_max_f32(vfpacc, voutput_min_less_zero_point);
124 vfpacc = math_min_f32(vfpacc, voutput_max_less_zero_point);
125 const int32_t vrndacc = (int32_t) lrintf(vfpacc);
126 int32_t vout = vrndacc + voutput_zero_point;
127
128 *output++ = (int8_t) vout;
129 } while (--c != 0);
130
131 output = (int8_t*) ((uintptr_t) output + output_increment);
132 } while (--output_width != 0);
133 }
134