1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gavgpool/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/gavgpool.h>
13 #include <xnnpack/math.h>
14
15
xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__scalar_fmagic_c2(size_t rows,size_t channels,const uint8_t * input,size_t input_stride,const uint8_t * zero,uint8_t * output,const union xnn_qu8_avgpool_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])16 void xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__scalar_fmagic_c2(
17 size_t rows,
18 size_t channels,
19 const uint8_t* input,
20 size_t input_stride,
21 const uint8_t* zero,
22 uint8_t* output,
23 const union xnn_qu8_avgpool_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
24 {
25 assert(rows != 0);
26 assert(rows <= 7);
27 assert(channels != 0);
28
29 const uint8_t* i0 = input;
30 const uint8_t* i1 = (const uint8_t*) ((uintptr_t) i0 + input_stride);
31 if XNN_UNPREDICTABLE(rows < 2) {
32 i1 = zero;
33 }
34 const uint8_t* i2 = (const uint8_t*) ((uintptr_t) i1 + input_stride);
35 if XNN_UNPREDICTABLE(rows <= 2) {
36 i2 = zero;
37 }
38 const uint8_t* i3 = (const uint8_t*) ((uintptr_t) i2 + input_stride);
39 if XNN_UNPREDICTABLE(rows < 4) {
40 i3 = zero;
41 }
42 const uint8_t* i4 = (const uint8_t*) ((uintptr_t) i3 + input_stride);
43 if XNN_UNPREDICTABLE(rows <= 4) {
44 i4 = zero;
45 }
46 const uint8_t* i5 = (const uint8_t*) ((uintptr_t) i4 + input_stride);
47 if XNN_UNPREDICTABLE(rows < 6) {
48 i5 = zero;
49 }
50 const uint8_t* i6 = (const uint8_t*) ((uintptr_t) i5 + input_stride);
51 if XNN_UNPREDICTABLE(rows <= 6) {
52 i6 = zero;
53 }
54
55 const int32_t vinit_bias = params->fp32_scalar_fmagic.init_bias;
56 const float vscale = params->fp32_scalar_fmagic.scale;
57 const float voutput_min_less_zero_point = params->fp32_scalar_fmagic.output_min_less_zero_point;
58 const float voutput_max_less_zero_point = params->fp32_scalar_fmagic.output_max_less_zero_point;
59 const float vmagic_bias = params->fp32_scalar_fmagic.magic_bias;
60 const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar_fmagic.magic_bias_less_output_zero_point;
61 for (; channels >= 2; channels -= 2) {
62 const int32_t vi0x0 = (int32_t) i0[0];
63 const int32_t vi0x1 = (int32_t) i0[1];
64 i0 += 2;
65
66 int32_t vacc0 = vi0x0 + vinit_bias;
67 const int32_t vi1x0 = (int32_t) i1[0];
68 int32_t vacc1 = vi0x1 + vinit_bias;
69 const int32_t vi1x1 = (int32_t) i1[1];
70 i1 += 2;
71
72 vacc0 += vi1x0;
73 const int32_t vi2x0 = (int32_t) i2[0];
74 vacc1 += vi1x1;
75 const int32_t vi2x1 = (int32_t) i2[1];
76 i2 += 2;
77 vacc0 += vi2x0;
78 const int32_t vi3x0 = (int32_t) i3[0];
79 vacc1 += vi2x1;
80 const int32_t vi3x1 = (int32_t) i3[1];
81 i3 += 2;
82 vacc0 += vi3x0;
83 const int32_t vi4x0 = (int32_t) i4[0];
84 vacc1 += vi3x1;
85 const int32_t vi4x1 = (int32_t) i4[1];
86 i4 += 2;
87 vacc0 += vi4x0;
88 const int32_t vi5x0 = (int32_t) i5[0];
89 vacc1 += vi4x1;
90 const int32_t vi5x1 = (int32_t) i5[1];
91 i5 += 2;
92 vacc0 += vi5x0;
93 const int32_t vi6x0 = (int32_t) i6[0];
94 vacc1 += vi5x1;
95 const int32_t vi6x1 = (int32_t) i6[1];
96 i6 += 2;
97
98 vacc0 += vi6x0;
99 vacc1 += vi6x1;
100
101 float vfpacc0 = (float) vacc0 * vscale;
102 float vfpacc1 = (float) vacc1 * vscale;
103
104 vfpacc0 = math_max_f32(vfpacc0, voutput_min_less_zero_point);
105 vfpacc1 = math_max_f32(vfpacc1, voutput_min_less_zero_point);
106
107 vfpacc0 = math_min_f32(vfpacc0, voutput_max_less_zero_point);
108 vfpacc1 = math_min_f32(vfpacc1, voutput_max_less_zero_point);
109
110 vfpacc0 += vmagic_bias;
111 vfpacc1 += vmagic_bias;
112
113 int32_t vout0 = (int32_t) float_as_uint32(vfpacc0) - vmagic_bias_less_output_zero_point;
114 int32_t vout1 = (int32_t) float_as_uint32(vfpacc1) - vmagic_bias_less_output_zero_point;
115
116 output[0] = (uint8_t) vout0;
117 output[1] = (uint8_t) vout1;
118 output += 2;
119 }
120 if XNN_UNLIKELY(channels != 0) {
121 int32_t vacc = vinit_bias;
122 const int32_t vi0 = (int32_t) *i0;
123 const int32_t vi1 = (int32_t) *i1;
124
125 vacc += vi0;
126 const int32_t vi2 = (int32_t) *i2;
127 vacc += vi1;
128 const int32_t vi3 = (int32_t) *i3;
129 vacc += vi2;
130 const int32_t vi4 = (int32_t) *i4;
131 vacc += vi3;
132 const int32_t vi5 = (int32_t) *i5;
133 vacc += vi4;
134 const int32_t vi6 = (int32_t) *i6;
135
136 vacc += vi5;
137 vacc += vi6;
138
139 float vfpacc = (float) vacc * vscale;
140 vfpacc = math_max_f32(vfpacc, voutput_min_less_zero_point);
141 vfpacc = math_min_f32(vfpacc, voutput_max_less_zero_point);
142 vfpacc += vmagic_bias;
143 int32_t vout = (int32_t) float_as_uint32(vfpacc) - vmagic_bias_less_output_zero_point;
144
145 *output = (uint8_t) vout;
146 }
147 }
148