• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Copyright 2020 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 <wasm_simd128.h>
9 
10 #include <xnnpack/maxpool.h>
11 
12 
xnn_f32_maxpool_minmax_ukernel_9p8x__wasmsimd_x86_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__wasmsimd_x86_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_OOB_READS
23 {
24   assert(output_pixels != 0);
25   assert(kernel_elements != 0);
26   assert(channels != 0);
27 
28   const v128_t voutput_max = wasm_v128_load64_splat(params->wasmsimd.max);
29   const v128_t voutput_min = wasm_v128_load64_splat(params->wasmsimd.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 v128_t vi0 = wasm_v128_load(i0);
79         i0 += 4;
80         const v128_t vi1 = wasm_v128_load(i1);
81         i1 += 4;
82         const v128_t vi2 = wasm_v128_load(i2);
83         i2 += 4;
84         const v128_t vi3 = wasm_v128_load(i3);
85         i3 += 4;
86         const v128_t vi4 = wasm_v128_load(i4);
87         i4 += 4;
88         const v128_t vi5 = wasm_v128_load(i5);
89         i5 += 4;
90         const v128_t vi6 = wasm_v128_load(i6);
91         i6 += 4;
92         const v128_t vi7 = wasm_v128_load(i7);
93         i7 += 4;
94         const v128_t vi8 = wasm_v128_load(i8);
95         i8 += 4;
96 
97         const v128_t vmax01 = wasm_f32x4_pmax(vi1, vi0);
98         const v128_t vmax23 = wasm_f32x4_pmax(vi3, vi2);
99         const v128_t vmax45 = wasm_f32x4_pmax(vi5, vi4);
100         const v128_t vmax018 = wasm_f32x4_pmax(vi8, vmax01);
101         const v128_t vmax67 = wasm_f32x4_pmax(vi7, vi6);
102 
103         const v128_t vmax2345 = wasm_f32x4_pmax(vmax45, vmax23);
104         const v128_t vmax01678 = wasm_f32x4_pmax(vmax67, vmax018);
105         const v128_t vmax = wasm_f32x4_pmax(vmax2345, vmax01678);
106 
107         v128_t vout = wasm_f32x4_pmax(voutput_min, vmax);
108         vout = wasm_f32x4_pmin(voutput_max, vout);
109 
110         wasm_v128_store(o, vout);
111         o += 4;
112       }
113       if (c != 0) {
114         const v128_t vi0 = wasm_v128_load(i0);
115         i0 += 4;
116         const v128_t vi1 = wasm_v128_load(i1);
117         i1 += 4;
118         const v128_t vi2 = wasm_v128_load(i2);
119         i2 += 4;
120         const v128_t vi3 = wasm_v128_load(i3);
121         i3 += 4;
122         const v128_t vi4 = wasm_v128_load(i4);
123         i4 += 4;
124         const v128_t vi5 = wasm_v128_load(i5);
125         i5 += 4;
126         const v128_t vi6 = wasm_v128_load(i6);
127         i6 += 4;
128         const v128_t vi7 = wasm_v128_load(i7);
129         i7 += 4;
130         const v128_t vi8 = wasm_v128_load(i8);
131         i8 += 4;
132 
133         const v128_t vmax01 = wasm_f32x4_pmax(vi1, vi0);
134         const v128_t vmax23 = wasm_f32x4_pmax(vi3, vi2);
135         const v128_t vmax45 = wasm_f32x4_pmax(vi5, vi4);
136         const v128_t vmax018 = wasm_f32x4_pmax(vi8, vmax01);
137         const v128_t vmax67 = wasm_f32x4_pmax(vi7, vi6);
138 
139         const v128_t vmax2345 = wasm_f32x4_pmax(vmax45, vmax23);
140         const v128_t vmax01678 = wasm_f32x4_pmax(vmax67, vmax018);
141         const v128_t vmax = wasm_f32x4_pmax(vmax2345, vmax01678);
142 
143         v128_t vout = wasm_f32x4_pmax(voutput_min, vmax);
144         vout = wasm_f32x4_pmin(voutput_max, vout);
145 
146         if (c & 2) {
147           *((double*) o) = wasm_f64x2_extract_lane(vout, 0);
148           vout = wasm_v32x4_shuffle(vout, vout, 2, 3, 2, 3);
149           o += 2;
150         }
151         if (c & 1) {
152           *o++ = wasm_f32x4_extract_lane(vout, 0);
153         }
154       }
155     }
156 
157     for (ptrdiff_t k = (ptrdiff_t) kernel_elements - 9; k > 0; k -= 8) {
158       const float* i0 = *input++;
159       const float* i1 = *input++;
160       const float* i2 = *input++;
161       const float* i3 = *input++;
162       const float* i4 = *input++;
163       const float* i5 = *input++;
164       const float* i6 = *input++;
165       const float* i7 = *input++;
166       i0 = (const float*) ((uintptr_t) i0 + input_offset);
167       i1 = (const float*) ((uintptr_t) i1 + input_offset);
168       i2 = (const float*) ((uintptr_t) i2 + input_offset);
169       i3 = (const float*) ((uintptr_t) i3 + input_offset);
170       i4 = (const float*) ((uintptr_t) i4 + input_offset);
171       i5 = (const float*) ((uintptr_t) i5 + input_offset);
172       i6 = (const float*) ((uintptr_t) i6 + input_offset);
173       i7 = (const float*) ((uintptr_t) i7 + input_offset);
174       if (k < 2) {
175         i1 = i0;
176       }
177       if (k <= 2) {
178         i2 = i0;
179       }
180       if (k < 4) {
181         i3 = i0;
182       }
183       if (k <= 4) {
184         i4 = i0;
185       }
186       if (k < 6) {
187         i5 = i0;
188       }
189       if (k <= 6) {
190         i6 = i0;
191       }
192       if (k < 8) {
193         i7 = i0;
194       }
195 
196       o = output;
197       size_t c = channels;
198       for (; c >= 4; c -= 4) {
199         const v128_t vi0 = wasm_v128_load(i0);
200         i0 += 4;
201         const v128_t vi1 = wasm_v128_load(i1);
202         i1 += 4;
203         const v128_t vi2 = wasm_v128_load(i2);
204         i2 += 4;
205         const v128_t vi3 = wasm_v128_load(i3);
206         i3 += 4;
207         const v128_t vi4 = wasm_v128_load(i4);
208         i4 += 4;
209         const v128_t vi5 = wasm_v128_load(i5);
210         i5 += 4;
211         const v128_t vi6 = wasm_v128_load(i6);
212         i6 += 4;
213         const v128_t vi7 = wasm_v128_load(i7);
214         i7 += 4;
215         const v128_t vo = wasm_v128_load(o);
216 
217         const v128_t vmax01 = wasm_f32x4_pmax(vi1, vi0);
218         const v128_t vmax23 = wasm_f32x4_pmax(vi3, vi2);
219         const v128_t vmax45 = wasm_f32x4_pmax(vi5, vi4);
220         const v128_t vmax01o = wasm_f32x4_pmax(vo, vmax01);
221         const v128_t vmax67 = wasm_f32x4_pmax(vi7, vi6);
222 
223         const v128_t vmax2345 = wasm_f32x4_pmax(vmax45, vmax23);
224         const v128_t vmax0167 = wasm_f32x4_pmax(vmax67, vmax01o);
225         const v128_t vmax = wasm_f32x4_pmax(vmax2345, vmax0167);
226 
227         v128_t vout = wasm_f32x4_pmax(voutput_min, vmax);
228         vout = wasm_f32x4_pmin(voutput_max, vout);
229 
230         wasm_v128_store(o, vout);
231         o += 4;
232       }
233       if (c != 0) {
234         const v128_t vi0 = wasm_v128_load(i0);
235         const v128_t vi1 = wasm_v128_load(i1);
236         const v128_t vi2 = wasm_v128_load(i2);
237         const v128_t vi3 = wasm_v128_load(i3);
238         const v128_t vi4 = wasm_v128_load(i4);
239         const v128_t vi5 = wasm_v128_load(i5);
240         const v128_t vi6 = wasm_v128_load(i6);
241         const v128_t vi7 = wasm_v128_load(i7);
242         const v128_t vo = wasm_v128_load(o);
243 
244         const v128_t vmax01 = wasm_f32x4_pmax(vi1, vi0);
245         const v128_t vmax23 = wasm_f32x4_pmax(vi3, vi2);
246         const v128_t vmax45 = wasm_f32x4_pmax(vi5, vi4);
247         const v128_t vmax01o = wasm_f32x4_pmax(vo, vmax01);
248         const v128_t vmax67 = wasm_f32x4_pmax(vi7, vi6);
249 
250         const v128_t vmax2345 = wasm_f32x4_pmax(vmax45, vmax23);
251         const v128_t vmax0167 = wasm_f32x4_pmax(vmax67, vmax01o);
252         const v128_t vmax = wasm_f32x4_pmax(vmax2345, vmax0167);
253 
254         v128_t vout = wasm_f32x4_pmax(voutput_min, vmax);
255         vout = wasm_f32x4_pmin(voutput_max, vout);
256 
257         if (c & 2) {
258           *((double*) o) = wasm_f64x2_extract_lane(vout, 0);
259           vout = wasm_v32x4_shuffle(vout, vout, 2, 3, 2, 3);
260           o += 2;
261         }
262         if (c & 1) {
263           *o++ = wasm_f32x4_extract_lane(vout, 0);
264         }
265       }
266     }
267     input = (const float**) ((uintptr_t) input + input_increment);
268     output = (float*) ((uintptr_t) o + output_increment);
269   } while (--output_pixels != 0);
270 }
271