• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Copyright 2021 Google LLC
2 //
3 // This source code is licensed under the BSD-style license found in the
4 // LICENSE file in the root directory of this source tree.
5 
6 #include <assert.h>
7 
8 #include <arm_neon.h>
9 
10 #include <xnnpack/maxpool.h>
11 
12 
xnn_s8_maxpool_minmax_ukernel_9p8x__neon_c16(size_t output_pixels,size_t kernel_elements,size_t channels,const int8_t ** input,size_t input_offset,int8_t * output,size_t input_increment,size_t output_increment,const union xnn_s8_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])13 void xnn_s8_maxpool_minmax_ukernel_9p8x__neon_c16(
14     size_t output_pixels,
15     size_t kernel_elements,
16     size_t channels,
17     const int8_t** input,
18     size_t input_offset,
19     int8_t* output,
20     size_t input_increment,
21     size_t output_increment,
22     const union xnn_s8_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23 {
24   assert(output_pixels != 0);
25   assert(kernel_elements != 0);
26   assert(channels != 0);
27 
28   const int8x16_t voutput_max = vld1q_dup_s8(&params->neon.max);
29   const int8x16_t voutput_min = vld1q_dup_s8(&params->neon.min);
30   do {
31     int8_t* o = output;
32     {
33       const int8_t* i0 = *input++;
34       const int8_t* i1 = *input++;
35       const int8_t* i2 = *input++;
36       const int8_t* i3 = *input++;
37       const int8_t* i4 = *input++;
38       const int8_t* i5 = *input++;
39       const int8_t* i6 = *input++;
40       const int8_t* i7 = *input++;
41       const int8_t* i8 = *input++;
42       i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
43       i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
44       i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
45       i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
46       i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
47       i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
48       i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
49       i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
50       i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
51       if (kernel_elements < 2) {
52         i1 = i0;
53       }
54       if (kernel_elements <= 2) {
55         i2 = i0;
56       }
57       if (kernel_elements < 4) {
58         i3 = i0;
59       }
60       if (kernel_elements <= 4) {
61         i4 = i0;
62       }
63       if (kernel_elements < 6) {
64         i5 = i0;
65       }
66       if (kernel_elements <= 6) {
67         i6 = i0;
68       }
69       if (kernel_elements < 8) {
70         i7 = i0;
71       }
72       if (kernel_elements <= 8) {
73         i8 = i0;
74       }
75 
76       size_t c = channels;
77       for (; c >= 16; c -= 16) {
78         const int8x16_t vi0 = vld1q_s8(i0); i0 += 16;
79         const int8x16_t vi1 = vld1q_s8(i1); i1 += 16;
80         const int8x16_t vi2 = vld1q_s8(i2); i2 += 16;
81         const int8x16_t vi3 = vld1q_s8(i3); i3 += 16;
82         const int8x16_t vi4 = vld1q_s8(i4); i4 += 16;
83         const int8x16_t vi5 = vld1q_s8(i5); i5 += 16;
84         const int8x16_t vi6 = vld1q_s8(i6); i6 += 16;
85         const int8x16_t vi7 = vld1q_s8(i7); i7 += 16;
86         const int8x16_t vi8 = vld1q_s8(i8); i8 += 16;
87 
88         const int8x16_t vmax018 = vmaxq_s8(vmaxq_s8(vi0, vi1), vi8);
89         const int8x16_t vmax23 = vmaxq_s8(vi2, vi3);
90         const int8x16_t vmax45 = vmaxq_s8(vi4, vi5);
91         const int8x16_t vmax67 = vmaxq_s8(vi6, vi7);
92 
93         const int8x16_t vmax2345 = vmaxq_s8(vmax23, vmax45);
94         const int8x16_t vmax01678 = vmaxq_s8(vmax018, vmax67);
95         int8x16_t vout = vmaxq_s8(vmax2345, vmax01678);
96         vout = vmaxq_s8(vout, voutput_min);
97         vout = vminq_s8(vout, voutput_max);
98 
99         vst1q_s8(o, vout); o += 16;
100       }
101       if (c != 0) {
102         const int8x16_t vi0 = vld1q_s8(i0);
103         const int8x16_t vi1 = vld1q_s8(i1);
104         const int8x16_t vi2 = vld1q_s8(i2);
105         const int8x16_t vi3 = vld1q_s8(i3);
106         const int8x16_t vi4 = vld1q_s8(i4);
107         const int8x16_t vi5 = vld1q_s8(i5);
108         const int8x16_t vi6 = vld1q_s8(i6);
109         const int8x16_t vi7 = vld1q_s8(i7);
110         const int8x16_t vi8 = vld1q_s8(i8);
111 
112         const int8x16_t vmax018 = vmaxq_s8(vmaxq_s8(vi0, vi1), vi8);
113         const int8x16_t vmax23 = vmaxq_s8(vi2, vi3);
114         const int8x16_t vmax45 = vmaxq_s8(vi4, vi5);
115         const int8x16_t vmax67 = vmaxq_s8(vi6, vi7);
116 
117         const int8x16_t vmax2345 = vmaxq_s8(vmax23, vmax45);
118         const int8x16_t vmax01678 = vmaxq_s8(vmax018, vmax67);
119         int8x16_t vout = vmaxq_s8(vmax2345, vmax01678);
120         vout = vmaxq_s8(vout, voutput_min);
121         vout = vminq_s8(vout, voutput_max);
122 
123         int8x8_t vout_lo = vget_low_s8(vout);
124         if (c & 8) {
125           vst1_s8(o, vout_lo); o += 8;
126           vout_lo = vget_high_s8(vout);
127         }
128         if (c & 4) {
129           vst1_lane_u32((void*) o, vreinterpret_u32_s8(vout_lo), 0); o += 4;
130           vout_lo = vext_s8(vout_lo, vout_lo, 4);
131         }
132         if (c & 2) {
133           vst1_lane_u16((void*) o, vreinterpret_u16_s8(vout_lo), 0); o += 2;
134           vout_lo = vext_s8(vout_lo, vout_lo, 2);
135         }
136         if (c & 1) {
137           vst1_lane_s8(o, vout_lo, 0); o += 1;
138         }
139       }
140     }
141 
142     for (ptrdiff_t k = (ptrdiff_t) kernel_elements - 9; k > 0; k -= 8) {
143       const int8_t* i0 = *input++;
144       const int8_t* i1 = *input++;
145       const int8_t* i2 = *input++;
146       const int8_t* i3 = *input++;
147       const int8_t* i4 = *input++;
148       const int8_t* i5 = *input++;
149       const int8_t* i6 = *input++;
150       const int8_t* i7 = *input++;
151       i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
152       i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
153       i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
154       i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
155       i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
156       i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
157       i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
158       i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
159       if (k < 2) {
160         i1 = i0;
161       }
162       if (k <= 2) {
163         i2 = i0;
164       }
165       if (k < 4) {
166         i3 = i0;
167       }
168       if (k <= 4) {
169         i4 = i0;
170       }
171       if (k < 6) {
172         i5 = i0;
173       }
174       if (k <= 6) {
175         i6 = i0;
176       }
177       if (k < 8) {
178         i7 = i0;
179       }
180 
181       o = output;
182       size_t c = channels;
183       for (; c >= 16; c -= 16) {
184         const int8x16_t vi0 = vld1q_s8(i0); i0 += 16;
185         const int8x16_t vi1 = vld1q_s8(i1); i1 += 16;
186         const int8x16_t vi2 = vld1q_s8(i2); i2 += 16;
187         const int8x16_t vi3 = vld1q_s8(i3); i3 += 16;
188         const int8x16_t vi4 = vld1q_s8(i4); i4 += 16;
189         const int8x16_t vi5 = vld1q_s8(i5); i5 += 16;
190         const int8x16_t vi6 = vld1q_s8(i6); i6 += 16;
191         const int8x16_t vi7 = vld1q_s8(i7); i7 += 16;
192         const int8x16_t vo = vld1q_s8(o);
193 
194         const int8x16_t vmax01 = vmaxq_s8(vmaxq_s8(vi0, vi1), vo);
195         const int8x16_t vmax23 = vmaxq_s8(vi2, vi3);
196         const int8x16_t vmax45 = vmaxq_s8(vi4, vi5);
197         const int8x16_t vmax67 = vmaxq_s8(vi6, vi7);
198 
199         const int8x16_t vmax2345 = vmaxq_s8(vmax23, vmax45);
200         const int8x16_t vmax0167 = vmaxq_s8(vmax01, vmax67);
201         int8x16_t vout = vmaxq_s8(vmax2345, vmax0167);
202         vout = vmaxq_s8(vout, voutput_min);
203         vout = vminq_s8(vout, voutput_max);
204 
205         vst1q_s8(o, vout); o += 16;
206       }
207       if (c != 0) {
208         const int8x16_t vi0 = vld1q_s8(i0);
209         const int8x16_t vi1 = vld1q_s8(i1);
210         const int8x16_t vi2 = vld1q_s8(i2);
211         const int8x16_t vi3 = vld1q_s8(i3);
212         const int8x16_t vi4 = vld1q_s8(i4);
213         const int8x16_t vi5 = vld1q_s8(i5);
214         const int8x16_t vi6 = vld1q_s8(i6);
215         const int8x16_t vi7 = vld1q_s8(i7);
216         const int8x16_t vo = vld1q_s8(o);
217 
218         const int8x16_t vmax01 = vmaxq_s8(vmaxq_s8(vi0, vi1), vo);
219         const int8x16_t vmax23 = vmaxq_s8(vi2, vi3);
220         const int8x16_t vmax45 = vmaxq_s8(vi4, vi5);
221         const int8x16_t vmax67 = vmaxq_s8(vi6, vi7);
222 
223         const int8x16_t vmax2345 = vmaxq_s8(vmax23, vmax45);
224         const int8x16_t vmax0167 = vmaxq_s8(vmax01, vmax67);
225         int8x16_t vout = vmaxq_s8(vmax2345, vmax0167);
226         vout = vmaxq_s8(vout, voutput_min);
227         vout = vminq_s8(vout, voutput_max);
228 
229         int8x8_t vout_lo = vget_low_s8(vout);
230         if (c & 8) {
231           vst1_s8(o, vout_lo); o += 8;
232           vout_lo = vget_high_s8(vout);
233         }
234         if (c & 4) {
235           vst1_lane_u32((void*) o, vreinterpret_u32_s8(vout_lo), 0); o += 4;
236           vout_lo = vext_s8(vout_lo, vout_lo, 4);
237         }
238         if (c & 2) {
239           vst1_lane_u16((void*) o, vreinterpret_u16_s8(vout_lo), 0); o += 2;
240           vout_lo = vext_s8(vout_lo, vout_lo, 2);
241         }
242         if (c & 1) {
243           vst1_lane_s8(o, vout_lo, 0); o += 1;
244         }
245       }
246     }
247     input = (const int8_t**) ((uintptr_t) input + input_increment);
248     output = (int8_t*) ((uintptr_t) o + output_increment);
249   } while (--output_pixels != 0);
250 }
251