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 <xnnpack/dwconv.h>
13 #include <xnnpack/math.h>
14 #include <xnnpack/unaligned.h>
15
16
xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_fmagic(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_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_fmagic(
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_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29 assert(channels != 0);
30 assert(output_width != 0);
31
32 const float vscale = params->fp32_scalar_fmagic.scale;
33 const float voutput_min_less_zero_point = params->fp32_scalar_fmagic.output_min_less_zero_point;
34 const float voutput_max_less_zero_point = params->fp32_scalar_fmagic.output_max_less_zero_point;
35 const float vmagic_bias = params->fp32_scalar_fmagic.magic_bias;
36 const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar_fmagic.magic_bias_less_output_zero_point;
37 do {
38 const int8_t* i0 = input[0];
39 assert(i0 != NULL);
40 if XNN_UNPREDICTABLE(i0 != zero) {
41 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
42 }
43 const int8_t* i1 = input[1];
44 assert(i1 != NULL);
45 if XNN_UNPREDICTABLE(i1 != zero) {
46 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
47 }
48 const int8_t* i2 = input[2];
49 assert(i2 != NULL);
50 if XNN_UNPREDICTABLE(i2 != zero) {
51 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
52 }
53 const int8_t* i3 = input[3];
54 assert(i3 != NULL);
55 if XNN_UNPREDICTABLE(i3 != zero) {
56 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
57 }
58 const int8_t* i4 = input[4];
59 assert(i4 != NULL);
60 if XNN_UNPREDICTABLE(i4 != zero) {
61 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
62 }
63 const int8_t* i5 = input[5];
64 assert(i5 != NULL);
65 if XNN_UNPREDICTABLE(i5 != zero) {
66 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
67 }
68 const int8_t* i6 = input[6];
69 assert(i6 != NULL);
70 if XNN_UNPREDICTABLE(i6 != zero) {
71 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
72 }
73 const int8_t* i7 = input[7];
74 assert(i7 != NULL);
75 if XNN_UNPREDICTABLE(i7 != zero) {
76 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
77 }
78 const int8_t* i8 = input[8];
79 assert(i8 != NULL);
80 if XNN_UNPREDICTABLE(i8 != zero) {
81 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
82 }
83 input = (const int8_t**) ((uintptr_t) input + input_stride);
84
85 size_t c = channels;
86 const void* w = weights;
87 do {
88 int32_t vacc = unaligned_load_s32(w);
89
90 const int32_t vi0 = (int32_t) *i0++;
91 const int32_t vk0 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[0];
92 vacc += vi0 * vk0;
93 const int32_t vi1 = (int32_t) *i1++;
94 const int32_t vk1 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[1];
95 vacc += vi1 * vk1;
96 const int32_t vi2 = (int32_t) *i2++;
97 const int32_t vk2 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[2];
98 vacc += vi2 * vk2;
99 const int32_t vi3 = (int32_t) *i3++;
100 const int32_t vk3 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[3];
101 vacc += vi3 * vk3;
102 const int32_t vi4 = (int32_t) *i4++;
103 const int32_t vk4 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[4];
104 vacc += vi4 * vk4;
105 const int32_t vi5 = (int32_t) *i5++;
106 const int32_t vk5 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[5];
107 vacc += vi5 * vk5;
108 const int32_t vi6 = (int32_t) *i6++;
109 const int32_t vk6 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[6];
110 vacc += vi6 * vk6;
111 const int32_t vi7 = (int32_t) *i7++;
112 const int32_t vk7 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[7];
113 vacc += vi7 * vk7;
114 const int32_t vi8 = (int32_t) *i8++;
115 const int32_t vk8 = ((const int8_t*) ((uintptr_t) w + sizeof(int32_t)))[8];
116 vacc += vi8 * vk8;
117
118 w = (const void*) ((uintptr_t) w + sizeof(int32_t) + 9 * sizeof(int8_t));
119
120 float vfpacc = (float) vacc * vscale;
121
122 vfpacc = math_max_f32(vfpacc, voutput_min_less_zero_point);
123 vfpacc = math_min_f32(vfpacc, voutput_max_less_zero_point);
124 vfpacc += vmagic_bias;
125 int32_t vout = (int32_t) float_as_uint32(vfpacc) - vmagic_bias_less_output_zero_point;
126
127 *output++ = (int8_t) vout;
128 } while (--c != 0);
129
130 output = (int8_t*) ((uintptr_t) output + output_increment);
131 } while (--output_width != 0);
132 }
133