1 // Copyright 2019 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_f32_maxpool_minmax_ukernel_9p8x__neon_c4(size_t output_pixels,size_t kernel_elements,size_t channels,const float ** input,size_t input_offset,float * output,size_t input_increment,size_t output_increment,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])13 void xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4(
14 size_t output_pixels,
15 size_t kernel_elements,
16 size_t channels,
17 const float** input,
18 size_t input_offset,
19 float* output,
20 size_t input_increment,
21 size_t output_increment,
22 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
23 {
24 assert(output_pixels != 0);
25 assert(kernel_elements != 0);
26 assert(channels != 0);
27
28 const float32x4_t voutput_max = vld1q_dup_f32(¶ms->scalar.max);
29 const float32x4_t voutput_min = vld1q_dup_f32(¶ms->scalar.min);
30 do {
31 float* o = output;
32 {
33 const float* i0 = *input++;
34 const float* i1 = *input++;
35 const float* i2 = *input++;
36 const float* i3 = *input++;
37 const float* i4 = *input++;
38 const float* i5 = *input++;
39 const float* i6 = *input++;
40 const float* i7 = *input++;
41 const float* i8 = *input++;
42 i0 = (const float*) ((uintptr_t) i0 + input_offset);
43 i1 = (const float*) ((uintptr_t) i1 + input_offset);
44 i2 = (const float*) ((uintptr_t) i2 + input_offset);
45 i3 = (const float*) ((uintptr_t) i3 + input_offset);
46 i4 = (const float*) ((uintptr_t) i4 + input_offset);
47 i5 = (const float*) ((uintptr_t) i5 + input_offset);
48 i6 = (const float*) ((uintptr_t) i6 + input_offset);
49 i7 = (const float*) ((uintptr_t) i7 + input_offset);
50 i8 = (const float*) ((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 >= 4; c -= 4) {
78 const float32x4_t vi0 = vld1q_f32(i0);
79 i0 += 4;
80 const float32x4_t vi1 = vld1q_f32(i1);
81 i1 += 4;
82 const float32x4_t vi2 = vld1q_f32(i2);
83 i2 += 4;
84 const float32x4_t vi3 = vld1q_f32(i3);
85 i3 += 4;
86 const float32x4_t vi4 = vld1q_f32(i4);
87 i4 += 4;
88 const float32x4_t vi5 = vld1q_f32(i5);
89 i5 += 4;
90 const float32x4_t vi6 = vld1q_f32(i6);
91 i6 += 4;
92 const float32x4_t vi7 = vld1q_f32(i7);
93 i7 += 4;
94 const float32x4_t vi8 = vld1q_f32(i8);
95 i8 += 4;
96
97 const float32x4_t vmax018 = vmaxq_f32(vmaxq_f32(vi0, vi1), vi8);
98 const float32x4_t vmax23 = vmaxq_f32(vi2, vi3);
99 const float32x4_t vmax45 = vmaxq_f32(vi4, vi5);
100 const float32x4_t vmax67 = vmaxq_f32(vi6, vi7);
101
102 const float32x4_t vmax2345 = vmaxq_f32(vmax23, vmax45);
103 const float32x4_t vmax01678 = vmaxq_f32(vmax018, vmax67);
104 const float32x4_t vmax = vmaxq_f32(vmax2345, vmax01678);
105 const float32x4_t vout = vmaxq_f32(vminq_f32(vmax, voutput_max), voutput_min);
106
107 vst1q_f32(o, vout);
108 o += 4;
109 }
110 if (c != 0) {
111 const float32x4_t vi0 = vld1q_f32(i0);
112 i0 += 4;
113 const float32x4_t vi1 = vld1q_f32(i1);
114 i1 += 4;
115 const float32x4_t vi2 = vld1q_f32(i2);
116 i2 += 4;
117 const float32x4_t vi3 = vld1q_f32(i3);
118 i3 += 4;
119 const float32x4_t vi4 = vld1q_f32(i4);
120 i4 += 4;
121 const float32x4_t vi5 = vld1q_f32(i5);
122 i5 += 4;
123 const float32x4_t vi6 = vld1q_f32(i6);
124 i6 += 4;
125 const float32x4_t vi7 = vld1q_f32(i7);
126 i7 += 4;
127 const float32x4_t vi8 = vld1q_f32(i8);
128 i8 += 4;
129
130 const float32x4_t vmax018 = vmaxq_f32(vmaxq_f32(vi0, vi1), vi8);
131 const float32x4_t vmax23 = vmaxq_f32(vi2, vi3);
132 const float32x4_t vmax45 = vmaxq_f32(vi4, vi5);
133 const float32x4_t vmax67 = vmaxq_f32(vi6, vi7);
134
135 const float32x4_t vmax2345 = vmaxq_f32(vmax23, vmax45);
136 const float32x4_t vmax01678 = vmaxq_f32(vmax018, vmax67);
137 const float32x4_t vmax = vmaxq_f32(vmax2345, vmax01678);
138 float32x4_t vout = vmaxq_f32(vminq_f32(vmax, voutput_max), voutput_min);
139
140 float32x2_t vout_lo = vget_low_f32(vout);
141 if (c & 2) {
142 vst1_f32(o, vout_lo);
143 vout_lo = vget_high_f32(vout);
144 o += 2;
145 }
146 if (c & 1) {
147 vst1_lane_f32(o, vout_lo, 0);
148 o += 1;
149 }
150 }
151 }
152
153 for (ptrdiff_t k = (ptrdiff_t) kernel_elements - 9; k > 0; k -= 8) {
154 const float* i0 = *input++;
155 const float* i1 = *input++;
156 const float* i2 = *input++;
157 const float* i3 = *input++;
158 const float* i4 = *input++;
159 const float* i5 = *input++;
160 const float* i6 = *input++;
161 const float* i7 = *input++;
162 i0 = (const float*) ((uintptr_t) i0 + input_offset);
163 i1 = (const float*) ((uintptr_t) i1 + input_offset);
164 i2 = (const float*) ((uintptr_t) i2 + input_offset);
165 i3 = (const float*) ((uintptr_t) i3 + input_offset);
166 i4 = (const float*) ((uintptr_t) i4 + input_offset);
167 i5 = (const float*) ((uintptr_t) i5 + input_offset);
168 i6 = (const float*) ((uintptr_t) i6 + input_offset);
169 i7 = (const float*) ((uintptr_t) i7 + input_offset);
170 if (k < 2) {
171 i1 = i0;
172 }
173 if (k <= 2) {
174 i2 = i0;
175 }
176 if (k < 4) {
177 i3 = i0;
178 }
179 if (k <= 4) {
180 i4 = i0;
181 }
182 if (k < 6) {
183 i5 = i0;
184 }
185 if (k <= 6) {
186 i6 = i0;
187 }
188 if (k < 8) {
189 i7 = i0;
190 }
191
192 o = output;
193 size_t c = channels;
194 for (; c >= 4; c -= 4) {
195 const float32x4_t vi0 = vld1q_f32(i0);
196 i0 += 4;
197 const float32x4_t vi1 = vld1q_f32(i1);
198 i1 += 4;
199 const float32x4_t vi2 = vld1q_f32(i2);
200 i2 += 4;
201 const float32x4_t vi3 = vld1q_f32(i3);
202 i3 += 4;
203 const float32x4_t vi4 = vld1q_f32(i4);
204 i4 += 4;
205 const float32x4_t vi5 = vld1q_f32(i5);
206 i5 += 4;
207 const float32x4_t vi6 = vld1q_f32(i6);
208 i6 += 4;
209 const float32x4_t vi7 = vld1q_f32(i7);
210 i7 += 4;
211 const float32x4_t vo = vld1q_f32(o);
212
213 const float32x4_t vmax01 = vmaxq_f32(vmaxq_f32(vi0, vi1), vo);
214 const float32x4_t vmax23 = vmaxq_f32(vi2, vi3);
215 const float32x4_t vmax45 = vmaxq_f32(vi4, vi5);
216 const float32x4_t vmax67 = vmaxq_f32(vi6, vi7);
217
218 const float32x4_t vmax2345 = vmaxq_f32(vmax23, vmax45);
219 const float32x4_t vmax0167 = vmaxq_f32(vmax01, vmax67);
220 const float32x4_t vmax = vmaxq_f32(vmax2345, vmax0167);
221 const float32x4_t vout = vmaxq_f32(vminq_f32(vmax, voutput_max), voutput_min);
222
223 vst1q_f32(o, vout);
224 o += 4;
225 }
226 if (c != 0) {
227 const float32x4_t vi0 = vld1q_f32(i0);
228 const float32x4_t vi1 = vld1q_f32(i1);
229 const float32x4_t vi2 = vld1q_f32(i2);
230 const float32x4_t vi3 = vld1q_f32(i3);
231 const float32x4_t vi4 = vld1q_f32(i4);
232 const float32x4_t vi5 = vld1q_f32(i5);
233 const float32x4_t vi6 = vld1q_f32(i6);
234 const float32x4_t vi7 = vld1q_f32(i7);
235 const float32x4_t vo = vld1q_f32(o);
236
237 const float32x4_t vmax01 = vmaxq_f32(vmaxq_f32(vi0, vi1), vo);
238 const float32x4_t vmax23 = vmaxq_f32(vi2, vi3);
239 const float32x4_t vmax45 = vmaxq_f32(vi4, vi5);
240 const float32x4_t vmax67 = vmaxq_f32(vi6, vi7);
241
242 const float32x4_t vmax2345 = vmaxq_f32(vmax23, vmax45);
243 const float32x4_t vmax0167 = vmaxq_f32(vmax01, vmax67);
244 const float32x4_t vmax = vmaxq_f32(vmax2345, vmax0167);
245 float32x4_t vout = vmaxq_f32(vminq_f32(vmax, voutput_max), voutput_min);
246
247 float32x2_t vout_lo = vget_low_f32(vout);
248 if (c & 2) {
249 vst1_f32(o, vout_lo);
250 vout_lo = vget_high_f32(vout);
251 o += 2;
252 }
253 if (c & 1) {
254 vst1_lane_f32(o, vout_lo, 0);
255 o += 1;
256 }
257 }
258 }
259 input = (const float**) ((uintptr_t) input + input_increment);
260 output = (float*) ((uintptr_t) o + output_increment);
261 } while (--output_pixels != 0);
262 }
263