1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gavgpool/multipass-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/gavgpool.h>
14 #include <xnnpack/math.h>
15
16
xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_lrintf_c1(size_t rows,size_t channels,const int8_t * input,size_t input_stride,const int8_t * zero,int32_t * buffer,int8_t * output,const union xnn_qs8_avgpool_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_lrintf_c1(
18 size_t rows,
19 size_t channels,
20 const int8_t* input,
21 size_t input_stride,
22 const int8_t* zero,
23 int32_t* buffer,
24 int8_t* output,
25 const union xnn_qs8_avgpool_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
26 {
27 assert(rows > 7);
28 assert(channels != 0);
29
30 const int8_t* i0 = input;
31 const int8_t* i1 = (const int8_t*) ((uintptr_t) i0 + input_stride);
32 const int8_t* i2 = (const int8_t*) ((uintptr_t) i1 + input_stride);
33 const int8_t* i3 = (const int8_t*) ((uintptr_t) i2 + input_stride);
34 const int8_t* i4 = (const int8_t*) ((uintptr_t) i3 + input_stride);
35 const int8_t* i5 = (const int8_t*) ((uintptr_t) i4 + input_stride);
36 const int8_t* i6 = (const int8_t*) ((uintptr_t) i5 + input_stride);
37 const size_t input_increment = 7 * input_stride - round_up_po2(channels, 1) * sizeof(int8_t);
38
39 const int32_t vinit_bias = params->fp32_scalar_lrintf.init_bias;
40 int32_t* b = buffer;
41 size_t c = channels;
42 do {
43 int32_t vacc = vinit_bias;
44 const int32_t vi0 = (int32_t) *i0++;
45 const int32_t vi1 = (int32_t) *i1++;
46
47 vacc += vi0;
48 const int32_t vi2 = (int32_t) *i2++;
49 vacc += vi1;
50 const int32_t vi3 = (int32_t) *i3++;
51 vacc += vi2;
52 const int32_t vi4 = (int32_t) *i4++;
53 vacc += vi3;
54 const int32_t vi5 = (int32_t) *i5++;
55 vacc += vi4;
56 const int32_t vi6 = (int32_t) *i6++;
57
58 vacc += vi5;
59 vacc += vi6;
60
61 *b++ = vacc;
62 } while (--c != 0);
63
64 for (rows -= 7; rows > 7; rows -= 7) {
65 i0 = (const int8_t*) ((uintptr_t) i0 + input_increment);
66 i1 = (const int8_t*) ((uintptr_t) i1 + input_increment);
67 i2 = (const int8_t*) ((uintptr_t) i2 + input_increment);
68 i3 = (const int8_t*) ((uintptr_t) i3 + input_increment);
69 i4 = (const int8_t*) ((uintptr_t) i4 + input_increment);
70 i5 = (const int8_t*) ((uintptr_t) i5 + input_increment);
71 i6 = (const int8_t*) ((uintptr_t) i6 + input_increment);
72
73 int32_t* b = buffer;
74 size_t c = channels;
75 do {
76 int32_t vacc = *b;
77 const int32_t vi0 = (int32_t) *i0++;
78 const int32_t vi1 = (int32_t) *i1++;
79
80 vacc += vi0;
81 const int32_t vi2 = (int32_t) *i2++;
82 vacc += vi1;
83 const int32_t vi3 = (int32_t) *i3++;
84 vacc += vi2;
85 const int32_t vi4 = (int32_t) *i4++;
86 vacc += vi3;
87 const int32_t vi5 = (int32_t) *i5++;
88 vacc += vi4;
89 const int32_t vi6 = (int32_t) *i6++;
90
91 vacc += vi5;
92 vacc += vi6;
93
94 *b++ = vacc;
95 } while (--c != 0);
96 }
97
98 i0 = (const int8_t*) ((uintptr_t) i0 + input_increment);
99 i1 = (const int8_t*) ((uintptr_t) i1 + input_increment);
100 if XNN_UNPREDICTABLE(rows < 2) {
101 i1 = zero;
102 }
103 i2 = (const int8_t*) ((uintptr_t) i2 + input_increment);
104 if XNN_UNPREDICTABLE(rows <= 2) {
105 i2 = zero;
106 }
107 i3 = (const int8_t*) ((uintptr_t) i3 + input_increment);
108 if XNN_UNPREDICTABLE(rows < 4) {
109 i3 = zero;
110 }
111 i4 = (const int8_t*) ((uintptr_t) i4 + input_increment);
112 if XNN_UNPREDICTABLE(rows <= 4) {
113 i4 = zero;
114 }
115 i5 = (const int8_t*) ((uintptr_t) i5 + input_increment);
116 if XNN_UNPREDICTABLE(rows < 6) {
117 i5 = zero;
118 }
119 i6 = (const int8_t*) ((uintptr_t) i6 + input_increment);
120 if XNN_UNPREDICTABLE(rows <= 6) {
121 i6 = zero;
122 }
123
124 const float vscale = params->fp32_scalar_lrintf.scale;
125 const float voutput_min_less_zero_point = params->fp32_scalar_lrintf.output_min_less_zero_point;
126 const float voutput_max_less_zero_point = params->fp32_scalar_lrintf.output_max_less_zero_point;
127 const int32_t voutput_zero_point = params->fp32_scalar_lrintf.output_zero_point;
128 do {
129 int32_t vacc = *buffer++;
130 const int32_t vi0 = (int32_t) *i0++;
131 const int32_t vi1 = (int32_t) *i1++;
132
133 vacc += vi0;
134 const int32_t vi2 = (int32_t) *i2++;
135 vacc += vi1;
136 const int32_t vi3 = (int32_t) *i3++;
137 vacc += vi2;
138 const int32_t vi4 = (int32_t) *i4++;
139 vacc += vi3;
140 const int32_t vi5 = (int32_t) *i5++;
141 vacc += vi4;
142 const int32_t vi6 = (int32_t) *i6++;
143
144 vacc += vi5;
145 vacc += vi6;
146
147 float vfpacc = (float) vacc * vscale;
148 vfpacc = math_max_f32(vfpacc, voutput_min_less_zero_point);
149 vfpacc = math_min_f32(vfpacc, voutput_max_less_zero_point);
150 const int32_t vrndacc = (int32_t) lrintf(vfpacc);
151 int32_t vout = vrndacc + voutput_zero_point;
152
153 *output++ = (int8_t) vout;
154 } while (--channels != 0);
155 }
156