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 <fp16.h>
13
14 #include <xnnpack/gavgpool.h>
15 #include <xnnpack/math.h>
16
17
xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__scalar_fmagic_c2(size_t rows,size_t channels,const int8_t * input,size_t input_stride,const int8_t * zero,int8_t * output,const union xnn_qs8_avgpool_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__scalar_fmagic_c2(
19 size_t rows,
20 size_t channels,
21 const int8_t* input,
22 size_t input_stride,
23 const int8_t* zero,
24 int8_t* output,
25 const union xnn_qs8_avgpool_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
26 {
27 assert(rows != 0);
28 assert(rows <= 7);
29 assert(channels != 0);
30
31 const int8_t* i0 = input;
32 const int8_t* i1 = (const int8_t*) ((uintptr_t) i0 + input_stride);
33 if XNN_UNPREDICTABLE(rows < 2) {
34 i1 = zero;
35 }
36 const int8_t* i2 = (const int8_t*) ((uintptr_t) i1 + input_stride);
37 if XNN_UNPREDICTABLE(rows <= 2) {
38 i2 = zero;
39 }
40 const int8_t* i3 = (const int8_t*) ((uintptr_t) i2 + input_stride);
41 if XNN_UNPREDICTABLE(rows < 4) {
42 i3 = zero;
43 }
44 const int8_t* i4 = (const int8_t*) ((uintptr_t) i3 + input_stride);
45 if XNN_UNPREDICTABLE(rows <= 4) {
46 i4 = zero;
47 }
48 const int8_t* i5 = (const int8_t*) ((uintptr_t) i4 + input_stride);
49 if XNN_UNPREDICTABLE(rows < 6) {
50 i5 = zero;
51 }
52 const int8_t* i6 = (const int8_t*) ((uintptr_t) i5 + input_stride);
53 if XNN_UNPREDICTABLE(rows <= 6) {
54 i6 = zero;
55 }
56
57 const int32_t vinit_bias = params->fp32_scalar_fmagic.init_bias;
58 const float vscale = params->fp32_scalar_fmagic.scale;
59 const float voutput_min_less_zero_point = params->fp32_scalar_fmagic.output_min_less_zero_point;
60 const float voutput_max_less_zero_point = params->fp32_scalar_fmagic.output_max_less_zero_point;
61 const float vmagic_bias = params->fp32_scalar_fmagic.magic_bias;
62 const int32_t vmagic_bias_less_output_zero_point = params->fp32_scalar_fmagic.magic_bias_less_output_zero_point;
63 for (; channels >= 2; channels -= 2) {
64 const int32_t vi0x0 = (int32_t) i0[0];
65 const int32_t vi0x1 = (int32_t) i0[1];
66 i0 += 2;
67
68 int32_t vacc0 = vi0x0 + vinit_bias;
69 const int32_t vi1x0 = (int32_t) i1[0];
70 int32_t vacc1 = vi0x1 + vinit_bias;
71 const int32_t vi1x1 = (int32_t) i1[1];
72 i1 += 2;
73
74 vacc0 += vi1x0;
75 const int32_t vi2x0 = (int32_t) i2[0];
76 vacc1 += vi1x1;
77 const int32_t vi2x1 = (int32_t) i2[1];
78 i2 += 2;
79 vacc0 += vi2x0;
80 const int32_t vi3x0 = (int32_t) i3[0];
81 vacc1 += vi2x1;
82 const int32_t vi3x1 = (int32_t) i3[1];
83 i3 += 2;
84 vacc0 += vi3x0;
85 const int32_t vi4x0 = (int32_t) i4[0];
86 vacc1 += vi3x1;
87 const int32_t vi4x1 = (int32_t) i4[1];
88 i4 += 2;
89 vacc0 += vi4x0;
90 const int32_t vi5x0 = (int32_t) i5[0];
91 vacc1 += vi4x1;
92 const int32_t vi5x1 = (int32_t) i5[1];
93 i5 += 2;
94 vacc0 += vi5x0;
95 const int32_t vi6x0 = (int32_t) i6[0];
96 vacc1 += vi5x1;
97 const int32_t vi6x1 = (int32_t) i6[1];
98 i6 += 2;
99
100 vacc0 += vi6x0;
101 vacc1 += vi6x1;
102
103 float vfpacc0 = (float) vacc0 * vscale;
104 float vfpacc1 = (float) vacc1 * vscale;
105
106 vfpacc0 = math_max_f32(vfpacc0, voutput_min_less_zero_point);
107 vfpacc1 = math_max_f32(vfpacc1, voutput_min_less_zero_point);
108
109 vfpacc0 = math_min_f32(vfpacc0, voutput_max_less_zero_point);
110 vfpacc1 = math_min_f32(vfpacc1, voutput_max_less_zero_point);
111
112 vfpacc0 += vmagic_bias;
113 vfpacc1 += vmagic_bias;
114
115 int32_t vout0 = (int32_t) fp32_to_bits(vfpacc0) - vmagic_bias_less_output_zero_point;
116 int32_t vout1 = (int32_t) fp32_to_bits(vfpacc1) - vmagic_bias_less_output_zero_point;
117
118 output[0] = (int8_t) vout0;
119 output[1] = (int8_t) vout1;
120 output += 2;
121 }
122 if XNN_UNLIKELY(channels != 0) {
123 int32_t vacc = vinit_bias;
124 const int32_t vi0 = (int32_t) *i0;
125 const int32_t vi1 = (int32_t) *i1;
126
127 vacc += vi0;
128 const int32_t vi2 = (int32_t) *i2;
129 vacc += vi1;
130 const int32_t vi3 = (int32_t) *i3;
131 vacc += vi2;
132 const int32_t vi4 = (int32_t) *i4;
133 vacc += vi3;
134 const int32_t vi5 = (int32_t) *i5;
135 vacc += vi4;
136 const int32_t vi6 = (int32_t) *i6;
137
138 vacc += vi5;
139 vacc += vi6;
140
141 float vfpacc = (float) vacc * vscale;
142 vfpacc = math_max_f32(vfpacc, voutput_min_less_zero_point);
143 vfpacc = math_min_f32(vfpacc, voutput_max_less_zero_point);
144 vfpacc += vmagic_bias;
145 int32_t vout = (int32_t) fp32_to_bits(vfpacc) - vmagic_bias_less_output_zero_point;
146
147 *output = (int8_t) vout;
148 }
149 }
150