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_qu8_gavgpool_minmax_fp32_ukernel_7x__scalar_imagic_c4(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)])18 void xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__scalar_imagic_c4(
19 size_t rows,
20 size_t channels,
21 const uint8_t* input,
22 size_t input_stride,
23 const uint8_t* zero,
24 uint8_t* output,
25 const union xnn_qu8_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 uint8_t* i0 = input;
32 const uint8_t* i1 = (const uint8_t*) ((uintptr_t) i0 + input_stride);
33 if XNN_UNPREDICTABLE(rows < 2) {
34 i1 = zero;
35 }
36 const uint8_t* i2 = (const uint8_t*) ((uintptr_t) i1 + input_stride);
37 if XNN_UNPREDICTABLE(rows <= 2) {
38 i2 = zero;
39 }
40 const uint8_t* i3 = (const uint8_t*) ((uintptr_t) i2 + input_stride);
41 if XNN_UNPREDICTABLE(rows < 4) {
42 i3 = zero;
43 }
44 const uint8_t* i4 = (const uint8_t*) ((uintptr_t) i3 + input_stride);
45 if XNN_UNPREDICTABLE(rows <= 4) {
46 i4 = zero;
47 }
48 const uint8_t* i5 = (const uint8_t*) ((uintptr_t) i4 + input_stride);
49 if XNN_UNPREDICTABLE(rows < 6) {
50 i5 = zero;
51 }
52 const uint8_t* i6 = (const uint8_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_imagic.init_bias;
58 const float vscale = params->fp32_scalar_imagic.scale;
59 const float vmagic_bias = params->fp32_scalar_imagic.magic_bias;
60 const int32_t vmagic_min = params->fp32_scalar_imagic.magic_min;
61 const int32_t vmagic_max = params->fp32_scalar_imagic.magic_max;
62 const int32_t vmagic_bias_less_zero_point = params->fp32_scalar_imagic.magic_bias_less_zero_point;
63 for (; channels >= 4; channels -= 4) {
64 const int32_t vi0x0 = (int32_t) i0[0];
65 const int32_t vi0x1 = (int32_t) i0[1];
66 const int32_t vi0x2 = (int32_t) i0[2];
67 const int32_t vi0x3 = (int32_t) i0[3];
68 i0 += 4;
69
70 int32_t vacc0 = vi0x0 + vinit_bias;
71 const int32_t vi1x0 = (int32_t) i1[0];
72 int32_t vacc1 = vi0x1 + vinit_bias;
73 const int32_t vi1x1 = (int32_t) i1[1];
74 int32_t vacc2 = vi0x2 + vinit_bias;
75 const int32_t vi1x2 = (int32_t) i1[2];
76 int32_t vacc3 = vi0x3 + vinit_bias;
77 const int32_t vi1x3 = (int32_t) i1[3];
78 i1 += 4;
79
80 vacc0 += vi1x0;
81 const int32_t vi2x0 = (int32_t) i2[0];
82 vacc1 += vi1x1;
83 const int32_t vi2x1 = (int32_t) i2[1];
84 vacc2 += vi1x2;
85 const int32_t vi2x2 = (int32_t) i2[2];
86 vacc3 += vi1x3;
87 const int32_t vi2x3 = (int32_t) i2[3];
88 i2 += 4;
89 vacc0 += vi2x0;
90 const int32_t vi3x0 = (int32_t) i3[0];
91 vacc1 += vi2x1;
92 const int32_t vi3x1 = (int32_t) i3[1];
93 vacc2 += vi2x2;
94 const int32_t vi3x2 = (int32_t) i3[2];
95 vacc3 += vi2x3;
96 const int32_t vi3x3 = (int32_t) i3[3];
97 i3 += 4;
98 vacc0 += vi3x0;
99 const int32_t vi4x0 = (int32_t) i4[0];
100 vacc1 += vi3x1;
101 const int32_t vi4x1 = (int32_t) i4[1];
102 vacc2 += vi3x2;
103 const int32_t vi4x2 = (int32_t) i4[2];
104 vacc3 += vi3x3;
105 const int32_t vi4x3 = (int32_t) i4[3];
106 i4 += 4;
107 vacc0 += vi4x0;
108 const int32_t vi5x0 = (int32_t) i5[0];
109 vacc1 += vi4x1;
110 const int32_t vi5x1 = (int32_t) i5[1];
111 vacc2 += vi4x2;
112 const int32_t vi5x2 = (int32_t) i5[2];
113 vacc3 += vi4x3;
114 const int32_t vi5x3 = (int32_t) i5[3];
115 i5 += 4;
116 vacc0 += vi5x0;
117 const int32_t vi6x0 = (int32_t) i6[0];
118 vacc1 += vi5x1;
119 const int32_t vi6x1 = (int32_t) i6[1];
120 vacc2 += vi5x2;
121 const int32_t vi6x2 = (int32_t) i6[2];
122 vacc3 += vi5x3;
123 const int32_t vi6x3 = (int32_t) i6[3];
124 i6 += 4;
125
126 vacc0 += vi6x0;
127 vacc1 += vi6x1;
128 vacc2 += vi6x2;
129 vacc3 += vi6x3;
130
131 float vfpacc0 = (float) vacc0 * vscale;
132 float vfpacc1 = (float) vacc1 * vscale;
133 float vfpacc2 = (float) vacc2 * vscale;
134 float vfpacc3 = (float) vacc3 * vscale;
135
136 vfpacc0 += vmagic_bias;
137 vfpacc1 += vmagic_bias;
138 vfpacc2 += vmagic_bias;
139 vfpacc3 += vmagic_bias;
140
141 int32_t vout0 = (int32_t) fp32_to_bits(vfpacc0);
142 int32_t vout1 = (int32_t) fp32_to_bits(vfpacc1);
143 int32_t vout2 = (int32_t) fp32_to_bits(vfpacc2);
144 int32_t vout3 = (int32_t) fp32_to_bits(vfpacc3);
145
146 vout0 = math_max_s32(vout0, vmagic_min);
147 vout1 = math_max_s32(vout1, vmagic_min);
148 vout2 = math_max_s32(vout2, vmagic_min);
149 vout3 = math_max_s32(vout3, vmagic_min);
150
151 vout0 = math_min_s32(vout0, vmagic_max);
152 vout1 = math_min_s32(vout1, vmagic_max);
153 vout2 = math_min_s32(vout2, vmagic_max);
154 vout3 = math_min_s32(vout3, vmagic_max);
155
156 vout0 -= vmagic_bias_less_zero_point;
157 vout1 -= vmagic_bias_less_zero_point;
158 vout2 -= vmagic_bias_less_zero_point;
159 vout3 -= vmagic_bias_less_zero_point;
160
161 output[0] = (uint8_t) vout0;
162 output[1] = (uint8_t) vout1;
163 output[2] = (uint8_t) vout2;
164 output[3] = (uint8_t) vout3;
165 output += 4;
166 }
167 if XNN_UNLIKELY(channels != 0) {
168 do {
169 int32_t vacc = vinit_bias;
170 const int32_t vi0 = (int32_t) *i0++;
171 const int32_t vi1 = (int32_t) *i1++;
172
173 vacc += vi0;
174 const int32_t vi2 = (int32_t) *i2++;
175 vacc += vi1;
176 const int32_t vi3 = (int32_t) *i3++;
177 vacc += vi2;
178 const int32_t vi4 = (int32_t) *i4++;
179 vacc += vi3;
180 const int32_t vi5 = (int32_t) *i5++;
181 vacc += vi4;
182 const int32_t vi6 = (int32_t) *i6++;
183
184 vacc += vi5;
185 vacc += vi6;
186
187 float vfpacc = (float) vacc * vscale;
188 vfpacc += vmagic_bias;
189 int32_t vout = (int32_t) fp32_to_bits(vfpacc);
190 vout = math_max_s32(vout, vmagic_min);
191 vout = math_min_s32(vout, vmagic_max);
192 vout -= vmagic_bias_less_zero_point;
193
194 *output++ = (uint8_t) vout;
195 } while (--channels != 0);
196 }
197 }
198