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_c2(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_c2(
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, 2) * sizeof(int8_t);
38
39 const int32_t vinit_bias = params->fp32_scalar_lrintf.init_bias;
40 int32_t* b = buffer;
41 for (ptrdiff_t c = (ptrdiff_t) channels; c > 0; c -= 2) {
42 const int32_t vi0x0 = (int32_t) i0[0];
43 const int32_t vi0x1 = (int32_t) i0[1];
44 i0 += 2;
45
46 int32_t vacc0 = vi0x0 + vinit_bias;
47 const int32_t vi1x0 = (int32_t) i1[0];
48 int32_t vacc1 = vi0x1 + vinit_bias;
49 const int32_t vi1x1 = (int32_t) i1[1];
50 i1 += 2;
51
52 vacc0 += vi1x0;
53 const int32_t vi2x0 = (int32_t) i2[0];
54 vacc1 += vi1x1;
55 const int32_t vi2x1 = (int32_t) i2[1];
56 i2 += 2;
57 vacc0 += vi2x0;
58 const int32_t vi3x0 = (int32_t) i3[0];
59 vacc1 += vi2x1;
60 const int32_t vi3x1 = (int32_t) i3[1];
61 i3 += 2;
62 vacc0 += vi3x0;
63 const int32_t vi4x0 = (int32_t) i4[0];
64 vacc1 += vi3x1;
65 const int32_t vi4x1 = (int32_t) i4[1];
66 i4 += 2;
67 vacc0 += vi4x0;
68 const int32_t vi5x0 = (int32_t) i5[0];
69 vacc1 += vi4x1;
70 const int32_t vi5x1 = (int32_t) i5[1];
71 i5 += 2;
72 vacc0 += vi5x0;
73 const int32_t vi6x0 = (int32_t) i6[0];
74 vacc1 += vi5x1;
75 const int32_t vi6x1 = (int32_t) i6[1];
76 i6 += 2;
77
78 vacc0 += vi6x0;
79 vacc1 += vi6x1;
80
81 b[0] = vacc0;
82 b[1] = vacc1;
83 b += 2;
84 }
85
86 for (rows -= 7; rows > 7; rows -= 7) {
87 i0 = (const int8_t*) ((uintptr_t) i0 + input_increment);
88 i1 = (const int8_t*) ((uintptr_t) i1 + input_increment);
89 i2 = (const int8_t*) ((uintptr_t) i2 + input_increment);
90 i3 = (const int8_t*) ((uintptr_t) i3 + input_increment);
91 i4 = (const int8_t*) ((uintptr_t) i4 + input_increment);
92 i5 = (const int8_t*) ((uintptr_t) i5 + input_increment);
93 i6 = (const int8_t*) ((uintptr_t) i6 + input_increment);
94
95 int32_t* b = buffer;
96 for (ptrdiff_t c = (ptrdiff_t) channels; c > 0; c -= 2) {
97 int32_t vacc0 = b[0];
98 const int32_t vi0x0 = (int32_t) i0[0];
99 int32_t vacc1 = b[1];
100 const int32_t vi0x1 = (int32_t) i0[1];
101 i0 += 2;
102
103 vacc0 += vi0x0;
104 const int32_t vi1x0 = (int32_t) i1[0];
105 vacc1 += vi0x1;
106 const int32_t vi1x1 = (int32_t) i1[1];
107 i1 += 2;
108 vacc0 += vi1x0;
109 const int32_t vi2x0 = (int32_t) i2[0];
110 vacc1 += vi1x1;
111 const int32_t vi2x1 = (int32_t) i2[1];
112 i2 += 2;
113 vacc0 += vi2x0;
114 const int32_t vi3x0 = (int32_t) i3[0];
115 vacc1 += vi2x1;
116 const int32_t vi3x1 = (int32_t) i3[1];
117 i3 += 2;
118 vacc0 += vi3x0;
119 const int32_t vi4x0 = (int32_t) i4[0];
120 vacc1 += vi3x1;
121 const int32_t vi4x1 = (int32_t) i4[1];
122 i4 += 2;
123 vacc0 += vi4x0;
124 const int32_t vi5x0 = (int32_t) i5[0];
125 vacc1 += vi4x1;
126 const int32_t vi5x1 = (int32_t) i5[1];
127 i5 += 2;
128 vacc0 += vi5x0;
129 const int32_t vi6x0 = (int32_t) i6[0];
130 vacc1 += vi5x1;
131 const int32_t vi6x1 = (int32_t) i6[1];
132 i6 += 2;
133
134 vacc0 += vi6x0;
135 vacc1 += vi6x1;
136
137 b[0] = vacc0;
138 b[1] = vacc1;
139 b += 2;
140 }
141 }
142
143 i0 = (const int8_t*) ((uintptr_t) i0 + input_increment);
144 i1 = (const int8_t*) ((uintptr_t) i1 + input_increment);
145 if XNN_UNPREDICTABLE(rows < 2) {
146 i1 = zero;
147 }
148 i2 = (const int8_t*) ((uintptr_t) i2 + input_increment);
149 if XNN_UNPREDICTABLE(rows <= 2) {
150 i2 = zero;
151 }
152 i3 = (const int8_t*) ((uintptr_t) i3 + input_increment);
153 if XNN_UNPREDICTABLE(rows < 4) {
154 i3 = zero;
155 }
156 i4 = (const int8_t*) ((uintptr_t) i4 + input_increment);
157 if XNN_UNPREDICTABLE(rows <= 4) {
158 i4 = zero;
159 }
160 i5 = (const int8_t*) ((uintptr_t) i5 + input_increment);
161 if XNN_UNPREDICTABLE(rows < 6) {
162 i5 = zero;
163 }
164 i6 = (const int8_t*) ((uintptr_t) i6 + input_increment);
165 if XNN_UNPREDICTABLE(rows <= 6) {
166 i6 = zero;
167 }
168
169 const float vscale = params->fp32_scalar_lrintf.scale;
170 const float voutput_min_less_zero_point = params->fp32_scalar_lrintf.output_min_less_zero_point;
171 const float voutput_max_less_zero_point = params->fp32_scalar_lrintf.output_max_less_zero_point;
172 const int32_t voutput_zero_point = params->fp32_scalar_lrintf.output_zero_point;
173 for (; channels >= 2; channels -= 2) {
174 int32_t vacc0 = buffer[0];
175 const int32_t vi0x0 = (int32_t) i0[0];
176 int32_t vacc1 = buffer[1];
177 const int32_t vi0x1 = (int32_t) i0[1];
178 buffer += 2;
179 i0 += 2;
180
181 vacc0 += vi0x0;
182 const int32_t vi1x0 = (int32_t) i1[0];
183 vacc1 += vi0x1;
184 const int32_t vi1x1 = (int32_t) i1[1];
185 i1 += 2;
186 vacc0 += vi1x0;
187 const int32_t vi2x0 = (int32_t) i2[0];
188 vacc1 += vi1x1;
189 const int32_t vi2x1 = (int32_t) i2[1];
190 i2 += 2;
191 vacc0 += vi2x0;
192 const int32_t vi3x0 = (int32_t) i3[0];
193 vacc1 += vi2x1;
194 const int32_t vi3x1 = (int32_t) i3[1];
195 i3 += 2;
196 vacc0 += vi3x0;
197 const int32_t vi4x0 = (int32_t) i4[0];
198 vacc1 += vi3x1;
199 const int32_t vi4x1 = (int32_t) i4[1];
200 i4 += 2;
201 vacc0 += vi4x0;
202 const int32_t vi5x0 = (int32_t) i5[0];
203 vacc1 += vi4x1;
204 const int32_t vi5x1 = (int32_t) i5[1];
205 i5 += 2;
206 vacc0 += vi5x0;
207 const int32_t vi6x0 = (int32_t) i6[0];
208 vacc1 += vi5x1;
209 const int32_t vi6x1 = (int32_t) i6[1];
210 i6 += 2;
211
212 vacc0 += vi6x0;
213 vacc1 += vi6x1;
214
215 float vfpacc0 = (float) vacc0 * vscale;
216 float vfpacc1 = (float) vacc1 * vscale;
217
218 vfpacc0 = math_max_f32(vfpacc0, voutput_min_less_zero_point);
219 vfpacc1 = math_max_f32(vfpacc1, voutput_min_less_zero_point);
220
221 vfpacc0 = math_min_f32(vfpacc0, voutput_max_less_zero_point);
222 vfpacc1 = math_min_f32(vfpacc1, voutput_max_less_zero_point);
223
224 const int32_t vrndacc0 = (int32_t) lrintf(vfpacc0);
225 const int32_t vrndacc1 = (int32_t) lrintf(vfpacc1);
226
227 int32_t vout0 = vrndacc0 + voutput_zero_point;
228 int32_t vout1 = vrndacc1 + voutput_zero_point;
229
230 output[0] = (int8_t) vout0;
231 output[1] = (int8_t) vout1;
232 output += 2;
233 }
234 if XNN_UNLIKELY(channels != 0) {
235 int32_t vacc = *buffer;
236 const int32_t vi0 = (int32_t) *i0;
237 const int32_t vi1 = (int32_t) *i1;
238
239 vacc += vi0;
240 const int32_t vi2 = (int32_t) *i2;
241 vacc += vi1;
242 const int32_t vi3 = (int32_t) *i3;
243 vacc += vi2;
244 const int32_t vi4 = (int32_t) *i4;
245 vacc += vi3;
246 const int32_t vi5 = (int32_t) *i5;
247 vacc += vi4;
248 const int32_t vi6 = (int32_t) *i6;
249
250 vacc += vi5;
251 vacc += vi6;
252
253 float vfpacc = (float) vacc * vscale;
254 vfpacc = math_max_f32(vfpacc, voutput_min_less_zero_point);
255 vfpacc = math_min_f32(vfpacc, voutput_max_less_zero_point);
256 const int32_t vrndacc = (int32_t) lrintf(vfpacc);
257 int32_t vout = vrndacc + voutput_zero_point;
258
259 *output = (int8_t) vout;
260 }
261 }
262