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 <wasm_simd128.h>
9
10 #include <xnnpack/maxpool.h>
11
12
xnn_s8_maxpool_minmax_ukernel_9p8x__wasmsimd_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__wasmsimd_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 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
31 do {
32 int8_t* o = output;
33 {
34 const int8_t* i0 = *input++;
35 const int8_t* i1 = *input++;
36 const int8_t* i2 = *input++;
37 const int8_t* i3 = *input++;
38 const int8_t* i4 = *input++;
39 const int8_t* i5 = *input++;
40 const int8_t* i6 = *input++;
41 const int8_t* i7 = *input++;
42 const int8_t* i8 = *input++;
43 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
44 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
45 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
46 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
47 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
48 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
49 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
50 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
51 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
52 if (kernel_elements < 2) {
53 i1 = i0;
54 }
55 if (kernel_elements <= 2) {
56 i2 = i0;
57 }
58 if (kernel_elements < 4) {
59 i3 = i0;
60 }
61 if (kernel_elements <= 4) {
62 i4 = i0;
63 }
64 if (kernel_elements < 6) {
65 i5 = i0;
66 }
67 if (kernel_elements <= 6) {
68 i6 = i0;
69 }
70 if (kernel_elements < 8) {
71 i7 = i0;
72 }
73 if (kernel_elements <= 8) {
74 i8 = i0;
75 }
76
77 size_t c = channels;
78 for (; c >= 16; c -= 16) {
79 const v128_t vi0 = wasm_v128_load(i0);
80 i0 += 16;
81 const v128_t vi1 = wasm_v128_load(i1);
82 i1 += 16;
83 const v128_t vi2 = wasm_v128_load(i2);
84 i2 += 16;
85 const v128_t vi3 = wasm_v128_load(i3);
86 i3 += 16;
87 const v128_t vi4 = wasm_v128_load(i4);
88 i4 += 16;
89 const v128_t vi5 = wasm_v128_load(i5);
90 i5 += 16;
91 const v128_t vi6 = wasm_v128_load(i6);
92 i6 += 16;
93 const v128_t vi7 = wasm_v128_load(i7);
94 i7 += 16;
95 const v128_t vi8 = wasm_v128_load(i8);
96 i8 += 16;
97
98 const v128_t vmax018 = wasm_i8x16_max(wasm_i8x16_max(vi0, vi1), vi8);
99 const v128_t vmax23 = wasm_i8x16_max(vi2, vi3);
100 const v128_t vmax45 = wasm_i8x16_max(vi4, vi5);
101 const v128_t vmax67 = wasm_i8x16_max(vi6, vi7);
102
103 const v128_t vmax2345 = wasm_i8x16_max(vmax23, vmax45);
104 const v128_t vmax01678 = wasm_i8x16_max(vmax018, vmax67);
105 v128_t vout = wasm_i8x16_max(vmax2345, vmax01678);
106 vout = wasm_i8x16_min(vout, voutput_max);
107 vout = wasm_i8x16_max(vout, voutput_min);
108
109 wasm_v128_store(o, vout); o += 16;
110 }
111 if (c != 0) {
112 const v128_t vi0 = wasm_v128_load(i0);
113 const v128_t vi1 = wasm_v128_load(i1);
114 const v128_t vi2 = wasm_v128_load(i2);
115 const v128_t vi3 = wasm_v128_load(i3);
116 const v128_t vi4 = wasm_v128_load(i4);
117 const v128_t vi5 = wasm_v128_load(i5);
118 const v128_t vi6 = wasm_v128_load(i6);
119 const v128_t vi7 = wasm_v128_load(i7);
120 const v128_t vi8 = wasm_v128_load(i8);
121
122 const v128_t vmax018 = wasm_i8x16_max(wasm_i8x16_max(vi0, vi1), vi8);
123 const v128_t vmax23 = wasm_i8x16_max(vi2, vi3);
124 const v128_t vmax45 = wasm_i8x16_max(vi4, vi5);
125 const v128_t vmax67 = wasm_i8x16_max(vi6, vi7);
126
127 const v128_t vmax2345 = wasm_i8x16_max(vmax23, vmax45);
128 const v128_t vmax01678 = wasm_i8x16_max(vmax018, vmax67);
129 v128_t vout = wasm_i8x16_max(vmax2345, vmax01678);
130 vout = wasm_i8x16_min(vout, voutput_max);
131 vout = wasm_i8x16_max(vout, voutput_min);
132
133 if (c & 8) {
134 *((double*) o) = wasm_f64x2_extract_lane(vout, 0);
135 vout = wasm_v64x2_shuffle(vout, vout, 1, 1);
136 o += 8;
137 }
138 if (c & 4) {
139 *((float*) o) = wasm_f32x4_extract_lane(vout, 0);
140 vout = wasm_u64x2_shr(vout, 32);
141 o += 4;
142 }
143 uint32_t vout_lo = wasm_i32x4_extract_lane(vout, 0);
144 if (c & 2) {
145 *((uint16_t*) o) = (uint16_t) vout_lo;
146 vout_lo >>= 16;
147 o += 2;
148 }
149 if (c & 1) {
150 *o = (int8_t) vout_lo;
151 o += 1;
152 }
153 }
154 }
155
156 for (ptrdiff_t k = (ptrdiff_t) kernel_elements - 9; k > 0; k -= 8) {
157 const int8_t* i0 = *input++;
158 const int8_t* i1 = *input++;
159 const int8_t* i2 = *input++;
160 const int8_t* i3 = *input++;
161 const int8_t* i4 = *input++;
162 const int8_t* i5 = *input++;
163 const int8_t* i6 = *input++;
164 const int8_t* i7 = *input++;
165 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
166 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
167 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
168 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
169 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
170 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
171 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
172 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
173 if (k < 2) {
174 i1 = i0;
175 }
176 if (k <= 2) {
177 i2 = i0;
178 }
179 if (k < 4) {
180 i3 = i0;
181 }
182 if (k <= 4) {
183 i4 = i0;
184 }
185 if (k < 6) {
186 i5 = i0;
187 }
188 if (k <= 6) {
189 i6 = i0;
190 }
191 if (k < 8) {
192 i7 = i0;
193 }
194
195 o = output;
196 size_t c = channels;
197 for (; c >= 16; c -= 16) {
198 const v128_t vi0 = wasm_v128_load(i0);
199 i0 += 16;
200 const v128_t vi1 = wasm_v128_load(i1);
201 i1 += 16;
202 const v128_t vi2 = wasm_v128_load(i2);
203 i2 += 16;
204 const v128_t vi3 = wasm_v128_load(i3);
205 i3 += 16;
206 const v128_t vi4 = wasm_v128_load(i4);
207 i4 += 16;
208 const v128_t vi5 = wasm_v128_load(i5);
209 i5 += 16;
210 const v128_t vi6 = wasm_v128_load(i6);
211 i6 += 16;
212 const v128_t vi7 = wasm_v128_load(i7);
213 i7 += 16;
214 const v128_t vo = wasm_v128_load(o);
215
216 const v128_t vmax01 = wasm_i8x16_max(wasm_i8x16_max(vi0, vi1), vo);
217 const v128_t vmax23 = wasm_i8x16_max(vi2, vi3);
218 const v128_t vmax45 = wasm_i8x16_max(vi4, vi5);
219 const v128_t vmax67 = wasm_i8x16_max(vi6, vi7);
220
221 const v128_t vmax2345 = wasm_i8x16_max(vmax23, vmax45);
222 const v128_t vmax0167 = wasm_i8x16_max(vmax01, vmax67);
223 v128_t vout = wasm_i8x16_max(vmax2345, vmax0167);
224 vout = wasm_i8x16_min(vout, voutput_max);
225 vout = wasm_i8x16_max(vout, voutput_min);
226
227 wasm_v128_store(o, vout);
228 o += 16;
229 }
230 if (c != 0) {
231 const v128_t vi0 = wasm_v128_load(i0);
232 const v128_t vi1 = wasm_v128_load(i1);
233 const v128_t vi2 = wasm_v128_load(i2);
234 const v128_t vi3 = wasm_v128_load(i3);
235 const v128_t vi4 = wasm_v128_load(i4);
236 const v128_t vi5 = wasm_v128_load(i5);
237 const v128_t vi6 = wasm_v128_load(i6);
238 const v128_t vi7 = wasm_v128_load(i7);
239 const v128_t vo = wasm_v128_load(o);
240
241 const v128_t vmax01 = wasm_i8x16_max(wasm_i8x16_max(vi0, vi1), vo);
242 const v128_t vmax23 = wasm_i8x16_max(vi2, vi3);
243 const v128_t vmax45 = wasm_i8x16_max(vi4, vi5);
244 const v128_t vmax67 = wasm_i8x16_max(vi6, vi7);
245
246 const v128_t vmax2345 = wasm_i8x16_max(vmax23, vmax45);
247 const v128_t vmax0167 = wasm_i8x16_max(vmax01, vmax67);
248 v128_t vout = wasm_i8x16_max(vmax2345, vmax0167);
249 vout = wasm_i8x16_min(vout, voutput_max);
250 vout = wasm_i8x16_max(vout, voutput_min);
251
252 if (c & 8) {
253 *((double*) o) = wasm_f64x2_extract_lane(vout, 0);
254 vout = wasm_v64x2_shuffle(vout, vout, 1, 1);
255 o += 8;
256 }
257 if (c & 4) {
258 *((float*) o) = wasm_f32x4_extract_lane(vout, 0);
259 vout = wasm_u64x2_shr(vout, 32);
260 o += 4;
261 }
262 uint32_t vout_lo = wasm_i32x4_extract_lane(vout, 0);
263 if (c & 2) {
264 *((uint16_t*) o) = (uint16_t) vout_lo;
265 vout_lo >>= 16;
266 o += 2;
267 }
268 if (c & 1) {
269 *o = (int8_t) vout_lo;
270 o += 1;
271 }
272 }
273 }
274 input = (const int8_t**) ((uintptr_t) input + input_increment);
275 output = (int8_t*) ((uintptr_t) o + output_increment);
276 } while (--output_pixels != 0);
277 }
278