• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Copyright (c) Facebook, Inc. and its affiliates.
2 // All rights reserved.
3 //
4 // Copyright 2019 Google LLC
5 //
6 // This source code is licensed under the BSD-style license found in the
7 // LICENSE file in the root directory of this source tree.
8 
9 #include <assert.h>
10 
11 #include <xnnpack/avgpool.h>
12 #include <xnnpack/math.h>
13 
14 
xnn_qu8_avgpool_minmax_ukernel_9p8x__scalar_c1(size_t output_pixels,size_t kernel_elements,size_t channels,const uint8_t ** input,size_t input_offset,const uint8_t * zero,int32_t * buffer,uint8_t * output,size_t input_increment,size_t output_increment,const union xnn_qu8_avgpool_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])15 void xnn_qu8_avgpool_minmax_ukernel_9p8x__scalar_c1(
16     size_t output_pixels,
17     size_t kernel_elements,
18     size_t channels,
19     const uint8_t** input,
20     size_t input_offset,
21     const uint8_t* zero,
22     int32_t* buffer,
23     uint8_t* output,
24     size_t input_increment,
25     size_t output_increment,
26     const union xnn_qu8_avgpool_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
27 {
28   assert(output_pixels != 0);
29   assert(kernel_elements > 9);
30   assert(channels != 0);
31 
32   const int32_t vbias = params->scalar.bias;
33   const int32_t vmultiplier = params->scalar.multiplier;
34   const int64_t vrounding = params->scalar.rounding;
35   const uint32_t vshift = params->scalar.right_shift;
36   const int32_t voutput_min = params->scalar.output_min_less_zero_point;
37   const int32_t voutput_max = params->scalar.output_max_less_zero_point;
38   const int32_t voutput_zero_point = params->scalar.output_zero_point;
39   do {
40     // First pass.
41     {
42       const uint8_t* i0 = *input++;
43       assert(i0 != NULL);
44       if XNN_UNPREDICTABLE(i0 != zero) {
45         i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
46       }
47       const uint8_t* i1 = *input++;
48       assert(i1 != NULL);
49       if XNN_UNPREDICTABLE(i1 != zero) {
50         i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
51       }
52       const uint8_t* i2 = *input++;
53       assert(i2 != NULL);
54       if XNN_UNPREDICTABLE(i2 != zero) {
55         i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
56       }
57       const uint8_t* i3 = *input++;
58       assert(i3 != NULL);
59       if XNN_UNPREDICTABLE(i3 != zero) {
60         i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
61       }
62       const uint8_t* i4 = *input++;
63       assert(i4 != NULL);
64       if XNN_UNPREDICTABLE(i4 != zero) {
65         i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
66       }
67       const uint8_t* i5 = *input++;
68       assert(i5 != NULL);
69       if XNN_UNPREDICTABLE(i5 != zero) {
70         i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
71       }
72       const uint8_t* i6 = *input++;
73       assert(i6 != NULL);
74       if XNN_UNPREDICTABLE(i6 != zero) {
75         i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
76       }
77       const uint8_t* i7 = *input++;
78       assert(i7 != NULL);
79       if XNN_UNPREDICTABLE(i7 != zero) {
80         i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
81       }
82       const uint8_t* i8 = *input++;
83       assert(i8 != NULL);
84       if XNN_UNPREDICTABLE(i8 != zero) {
85         i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
86       }
87 
88       int32_t* b = buffer;
89       size_t c = channels;
90       do {
91         const uint32_t vi0 = (uint32_t) *i0++;
92         const uint32_t vi1 = (uint32_t) *i1++;
93         const uint32_t vi2 = (uint32_t) *i2++;
94         const uint32_t vi3 = (uint32_t) *i3++;
95         const uint32_t vi4 = (uint32_t) *i4++;
96         const uint32_t vi5 = (uint32_t) *i5++;
97         const uint32_t vi6 = (uint32_t) *i6++;
98         const uint32_t vi7 = (uint32_t) *i7++;
99         const uint32_t vi8 = (uint32_t) *i8++;
100 
101         const uint32_t vsum01 = vi0 + vi1;
102         const uint32_t vsum23 = vi2 + vi3;
103         const uint32_t vsum45 = vi4 + vi5;
104         const uint32_t vsum67 = vi6 + vi7;
105         const uint32_t vsum018 = vsum01 + vi8;
106         const uint32_t vsum2345 = vsum23 + vsum45;
107         const uint32_t vsum01678 = vsum018 + vsum67;
108         int32_t vacc = vbias + (int32_t) vsum2345;
109         vacc += (int32_t) vsum01678;
110         *b++ = vacc;
111       } while (--c != 0);
112     }
113 
114     size_t k = kernel_elements;
115     // Intermediate passes.
116     for (k -= 9; k > 8; k -= 8) {
117       const uint8_t* i0 = *input++;
118       assert(i0 != NULL);
119       if XNN_UNPREDICTABLE(i0 != zero) {
120         i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
121       }
122       const uint8_t* i1 = *input++;
123       assert(i1 != NULL);
124       if XNN_UNPREDICTABLE(i1 != zero) {
125         i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
126       }
127       const uint8_t* i2 = *input++;
128       assert(i2 != NULL);
129       if XNN_UNPREDICTABLE(i2 != zero) {
130         i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
131       }
132       const uint8_t* i3 = *input++;
133       assert(i3 != NULL);
134       if XNN_UNPREDICTABLE(i3 != zero) {
135         i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
136       }
137       const uint8_t* i4 = *input++;
138       assert(i4 != NULL);
139       if XNN_UNPREDICTABLE(i4 != zero) {
140         i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
141       }
142       const uint8_t* i5 = *input++;
143       assert(i5 != NULL);
144       if XNN_UNPREDICTABLE(i5 != zero) {
145         i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
146       }
147       const uint8_t* i6 = *input++;
148       assert(i6 != NULL);
149       if XNN_UNPREDICTABLE(i6 != zero) {
150         i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
151       }
152       const uint8_t* i7 = *input++;
153       assert(i7 != NULL);
154       if XNN_UNPREDICTABLE(i7 != zero) {
155         i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
156       }
157 
158       int32_t* b = buffer;
159       size_t c = channels;
160       do {
161         int32_t vacc = *b;
162 
163         const uint32_t vi0 = (uint32_t) *i0++;
164         const uint32_t vi1 = (uint32_t) *i1++;
165         const uint32_t vi2 = (uint32_t) *i2++;
166         const uint32_t vi3 = (uint32_t) *i3++;
167         const uint32_t vi4 = (uint32_t) *i4++;
168         const uint32_t vi5 = (uint32_t) *i5++;
169         const uint32_t vi6 = (uint32_t) *i6++;
170         const uint32_t vi7 = (uint32_t) *i7++;
171 
172         const uint32_t vsum01 = vi0 + vi1;
173         const uint32_t vsum23 = vi2 + vi3;
174         const uint32_t vsum45 = vi4 + vi5;
175         const uint32_t vsum67 = vi6 + vi7;
176         const uint32_t vsum0123 = vsum01 + vsum23;
177         const uint32_t vsum4567 = vsum45 + vsum67;
178         vacc += (int32_t) vsum0123;
179         vacc += (int32_t) vsum4567;
180 
181         *b++ = vacc;
182       } while (--c != 0);
183     }
184 
185     // Last pass.
186     {
187       const uint8_t* i0 = input[0];
188       assert(i0 != NULL);
189       const uint8_t* i1 = input[1];
190       const uint8_t* i2 = input[2];
191       const uint8_t* i3 = input[3];
192       const uint8_t* i4 = input[4];
193       const uint8_t* i5 = input[5];
194       const uint8_t* i6 = input[6];
195       const uint8_t* i7 = input[7];
196       input = (const uint8_t**) ((uintptr_t) input + input_increment);
197       if (k < 2) {
198         i1 = zero;
199       }
200       assert(i1 != NULL);
201       if (k <= 2) {
202         i2 = zero;
203       }
204       assert(i2 != NULL);
205       if (k < 4) {
206         i3 = zero;
207       }
208       assert(i3 != NULL);
209       if (k <= 4) {
210         i4 = zero;
211       }
212       assert(i4 != NULL);
213       if (k < 6) {
214         i5 = zero;
215       }
216       assert(i5 != NULL);
217       if (k <= 6) {
218         i6 = zero;
219       }
220       assert(i6 != NULL);
221       if (k < 8) {
222         i7 = zero;
223       }
224       assert(i7 != NULL);
225       if XNN_UNPREDICTABLE(i0 != zero) {
226         i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
227       }
228       if XNN_UNPREDICTABLE(i1 != zero) {
229         i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
230       }
231       if XNN_UNPREDICTABLE(i2 != zero) {
232         i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
233       }
234       if XNN_UNPREDICTABLE(i3 != zero) {
235         i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
236       }
237       if XNN_UNPREDICTABLE(i4 != zero) {
238         i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
239       }
240       if XNN_UNPREDICTABLE(i5 != zero) {
241         i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
242       }
243       if XNN_UNPREDICTABLE(i6 != zero) {
244         i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
245       }
246       if XNN_UNPREDICTABLE(i7 != zero) {
247         i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
248       }
249 
250       size_t c = channels;
251       int32_t* b = buffer;
252       do {
253         int32_t vacc = *b++;
254 
255         const uint32_t vi0 = (uint32_t) *i0++;
256         const uint32_t vi1 = (uint32_t) *i1++;
257         const uint32_t vi2 = (uint32_t) *i2++;
258         const uint32_t vi3 = (uint32_t) *i3++;
259         const uint32_t vi4 = (uint32_t) *i4++;
260         const uint32_t vi5 = (uint32_t) *i5++;
261         const uint32_t vi6 = (uint32_t) *i6++;
262         const uint32_t vi7 = (uint32_t) *i7++;
263 
264         const uint32_t vsum01 = vi0 + vi1;
265         const uint32_t vsum23 = vi2 + vi3;
266         const uint32_t vsum45 = vi4 + vi5;
267         const uint32_t vsum67 = vi6 + vi7;
268         const uint32_t vsum0123 = vsum01 + vsum23;
269         const uint32_t vsum4567 = vsum45 + vsum67;
270         vacc += (int32_t) vsum0123;
271         vacc += (int32_t) vsum4567;
272 
273         const int64_t vproduct = (int64_t) vacc * (int64_t) vmultiplier;
274         const int64_t vadjusted_product = vproduct - (int64_t) (vacc < 0);
275         int32_t vout = (int32_t) math_asr_s64(vadjusted_product + vrounding, vshift);
276         vout = vout < voutput_min ? voutput_min : vout;
277         vout = vout > voutput_max ? voutput_max : vout;
278         vout += voutput_zero_point;
279 
280         *output++ = (uint8_t) vout;
281       } while (--c != 0);
282     }
283     output = (uint8_t*) ((uintptr_t) output + output_increment);
284   } while (--output_pixels != 0);
285 }
286