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_qs8_gavgpool_minmax_fp32_ukernel_7x__scalar_fmagic_c4(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)])16 void xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__scalar_fmagic_c4(
17 size_t rows,
18 size_t channels,
19 const int8_t* input,
20 size_t input_stride,
21 const int8_t* zero,
22 int8_t* output,
23 const union xnn_qs8_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 int8_t* i0 = input;
30 const int8_t* i1 = (const int8_t*) ((uintptr_t) i0 + input_stride);
31 if XNN_UNPREDICTABLE(rows < 2) {
32 i1 = zero;
33 }
34 const int8_t* i2 = (const int8_t*) ((uintptr_t) i1 + input_stride);
35 if XNN_UNPREDICTABLE(rows <= 2) {
36 i2 = zero;
37 }
38 const int8_t* i3 = (const int8_t*) ((uintptr_t) i2 + input_stride);
39 if XNN_UNPREDICTABLE(rows < 4) {
40 i3 = zero;
41 }
42 const int8_t* i4 = (const int8_t*) ((uintptr_t) i3 + input_stride);
43 if XNN_UNPREDICTABLE(rows <= 4) {
44 i4 = zero;
45 }
46 const int8_t* i5 = (const int8_t*) ((uintptr_t) i4 + input_stride);
47 if XNN_UNPREDICTABLE(rows < 6) {
48 i5 = zero;
49 }
50 const int8_t* i6 = (const int8_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 >= 4; channels -= 4) {
62 const int32_t vi0x0 = (int32_t) i0[0];
63 const int32_t vi0x1 = (int32_t) i0[1];
64 const int32_t vi0x2 = (int32_t) i0[2];
65 const int32_t vi0x3 = (int32_t) i0[3];
66 i0 += 4;
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 int32_t vacc2 = vi0x2 + vinit_bias;
73 const int32_t vi1x2 = (int32_t) i1[2];
74 int32_t vacc3 = vi0x3 + vinit_bias;
75 const int32_t vi1x3 = (int32_t) i1[3];
76 i1 += 4;
77
78 vacc0 += vi1x0;
79 const int32_t vi2x0 = (int32_t) i2[0];
80 vacc1 += vi1x1;
81 const int32_t vi2x1 = (int32_t) i2[1];
82 vacc2 += vi1x2;
83 const int32_t vi2x2 = (int32_t) i2[2];
84 vacc3 += vi1x3;
85 const int32_t vi2x3 = (int32_t) i2[3];
86 i2 += 4;
87 vacc0 += vi2x0;
88 const int32_t vi3x0 = (int32_t) i3[0];
89 vacc1 += vi2x1;
90 const int32_t vi3x1 = (int32_t) i3[1];
91 vacc2 += vi2x2;
92 const int32_t vi3x2 = (int32_t) i3[2];
93 vacc3 += vi2x3;
94 const int32_t vi3x3 = (int32_t) i3[3];
95 i3 += 4;
96 vacc0 += vi3x0;
97 const int32_t vi4x0 = (int32_t) i4[0];
98 vacc1 += vi3x1;
99 const int32_t vi4x1 = (int32_t) i4[1];
100 vacc2 += vi3x2;
101 const int32_t vi4x2 = (int32_t) i4[2];
102 vacc3 += vi3x3;
103 const int32_t vi4x3 = (int32_t) i4[3];
104 i4 += 4;
105 vacc0 += vi4x0;
106 const int32_t vi5x0 = (int32_t) i5[0];
107 vacc1 += vi4x1;
108 const int32_t vi5x1 = (int32_t) i5[1];
109 vacc2 += vi4x2;
110 const int32_t vi5x2 = (int32_t) i5[2];
111 vacc3 += vi4x3;
112 const int32_t vi5x3 = (int32_t) i5[3];
113 i5 += 4;
114 vacc0 += vi5x0;
115 const int32_t vi6x0 = (int32_t) i6[0];
116 vacc1 += vi5x1;
117 const int32_t vi6x1 = (int32_t) i6[1];
118 vacc2 += vi5x2;
119 const int32_t vi6x2 = (int32_t) i6[2];
120 vacc3 += vi5x3;
121 const int32_t vi6x3 = (int32_t) i6[3];
122 i6 += 4;
123
124 vacc0 += vi6x0;
125 vacc1 += vi6x1;
126 vacc2 += vi6x2;
127 vacc3 += vi6x3;
128
129 float vfpacc0 = (float) vacc0 * vscale;
130 float vfpacc1 = (float) vacc1 * vscale;
131 float vfpacc2 = (float) vacc2 * vscale;
132 float vfpacc3 = (float) vacc3 * vscale;
133
134 vfpacc0 = math_max_f32(vfpacc0, voutput_min_less_zero_point);
135 vfpacc1 = math_max_f32(vfpacc1, voutput_min_less_zero_point);
136 vfpacc2 = math_max_f32(vfpacc2, voutput_min_less_zero_point);
137 vfpacc3 = math_max_f32(vfpacc3, voutput_min_less_zero_point);
138
139 vfpacc0 = math_min_f32(vfpacc0, voutput_max_less_zero_point);
140 vfpacc1 = math_min_f32(vfpacc1, voutput_max_less_zero_point);
141 vfpacc2 = math_min_f32(vfpacc2, voutput_max_less_zero_point);
142 vfpacc3 = math_min_f32(vfpacc3, voutput_max_less_zero_point);
143
144 vfpacc0 += vmagic_bias;
145 vfpacc1 += vmagic_bias;
146 vfpacc2 += vmagic_bias;
147 vfpacc3 += vmagic_bias;
148
149 int32_t vout0 = (int32_t) float_as_uint32(vfpacc0) - vmagic_bias_less_output_zero_point;
150 int32_t vout1 = (int32_t) float_as_uint32(vfpacc1) - vmagic_bias_less_output_zero_point;
151 int32_t vout2 = (int32_t) float_as_uint32(vfpacc2) - vmagic_bias_less_output_zero_point;
152 int32_t vout3 = (int32_t) float_as_uint32(vfpacc3) - vmagic_bias_less_output_zero_point;
153
154 output[0] = (int8_t) vout0;
155 output[1] = (int8_t) vout1;
156 output[2] = (int8_t) vout2;
157 output[3] = (int8_t) vout3;
158 output += 4;
159 }
160 if XNN_UNLIKELY(channels != 0) {
161 do {
162 int32_t vacc = vinit_bias;
163 const int32_t vi0 = (int32_t) *i0++;
164 const int32_t vi1 = (int32_t) *i1++;
165
166 vacc += vi0;
167 const int32_t vi2 = (int32_t) *i2++;
168 vacc += vi1;
169 const int32_t vi3 = (int32_t) *i3++;
170 vacc += vi2;
171 const int32_t vi4 = (int32_t) *i4++;
172 vacc += vi3;
173 const int32_t vi5 = (int32_t) *i5++;
174 vacc += vi4;
175 const int32_t vi6 = (int32_t) *i6++;
176
177 vacc += vi5;
178 vacc += vi6;
179
180 float vfpacc = (float) vacc * vscale;
181 vfpacc = math_max_f32(vfpacc, voutput_min_less_zero_point);
182 vfpacc = math_min_f32(vfpacc, voutput_max_less_zero_point);
183 vfpacc += vmagic_bias;
184 int32_t vout = (int32_t) float_as_uint32(vfpacc) - vmagic_bias_less_output_zero_point;
185
186 *output++ = (int8_t) vout;
187 } while (--channels != 0);
188 }
189 }
190