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_c1(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_c1(
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 do {
62 int32_t vacc = vinit_bias;
63 const int32_t vi0 = (int32_t) *i0++;
64 const int32_t vi1 = (int32_t) *i1++;
65
66 vacc += vi0;
67 const int32_t vi2 = (int32_t) *i2++;
68 vacc += vi1;
69 const int32_t vi3 = (int32_t) *i3++;
70 vacc += vi2;
71 const int32_t vi4 = (int32_t) *i4++;
72 vacc += vi3;
73 const int32_t vi5 = (int32_t) *i5++;
74 vacc += vi4;
75 const int32_t vi6 = (int32_t) *i6++;
76
77 vacc += vi5;
78 vacc += vi6;
79
80 float vfpacc = (float) vacc * vscale;
81 vfpacc = math_max_f32(vfpacc, voutput_min_less_zero_point);
82 vfpacc = math_min_f32(vfpacc, voutput_max_less_zero_point);
83 vfpacc += vmagic_bias;
84 int32_t vout = (int32_t) float_as_uint32(vfpacc) - vmagic_bias_less_output_zero_point;
85
86 *output++ = (uint8_t) vout;
87 } while (--channels != 0);
88 }
89