1 // Auto-generated file. Do not edit!
2 // Template: src/f32-prelu/wasmsimd-minmax.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <wasm_simd128.h>
13
14 #include <xnnpack/math.h>
15 #include <xnnpack/prelu.h>
16
17
xnn_f32_prelu_ukernel__wasmsimd_minmax_4x16(size_t rows,size_t channels,const float * restrict input,size_t input_stride,const float * restrict weights,float * restrict output,size_t output_stride)18 void xnn_f32_prelu_ukernel__wasmsimd_minmax_4x16(
19 size_t rows,
20 size_t channels,
21 const float*restrict input,
22 size_t input_stride,
23 const float*restrict weights,
24 float*restrict output,
25 size_t output_stride) XNN_OOB_READS
26 {
27 assert(rows != 0);
28 assert(channels != 0);
29 assert(channels % sizeof(float) == 0);
30
31 const float* i0 = input;
32 float* o0 = output;
33 const float* i1 = (const float*) ((uintptr_t) i0 + input_stride);
34 float* o1 = (float*) ((uintptr_t) o0 + output_stride);
35 const float* i2 = (const float*) ((uintptr_t) i1 + input_stride);
36 float* o2 = (float*) ((uintptr_t) o1 + output_stride);
37 const float* i3 = (const float*) ((uintptr_t) i2 + input_stride);
38 float* o3 = (float*) ((uintptr_t) o2 + output_stride);
39
40 const size_t input_increment = input_stride * 4 - channels;
41 const size_t output_increment = output_stride * 4 - channels;
42
43 const v128_t vzero = wasm_i32x4_const_splat(0);
44 do {
45 if XNN_UNPREDICTABLE(rows < 2) {
46 i1 = i0;
47 o1 = o0;
48 }
49 if XNN_UNPREDICTABLE(rows <= 2) {
50 i2 = i1;
51 o2 = o1;
52 }
53 if XNN_UNPREDICTABLE(rows < 4) {
54 i3 = i2;
55 o3 = o2;
56 }
57
58 const float* w = weights;
59 size_t c = channels;
60 for (; c >= 16 * sizeof(float); c -= 16 * sizeof(float)) {
61 const v128_t vw0123 = wasm_v128_load(w);
62 const v128_t vw4567 = wasm_v128_load(w + 4);
63 const v128_t vw89AB = wasm_v128_load(w + 8);
64 const v128_t vwCDEF = wasm_v128_load(w + 12);
65 w += 16;
66
67 v128_t vi0x0123 = wasm_v128_load(i0);
68 v128_t vi0x4567 = wasm_v128_load(i0 + 4);
69 v128_t vi0x89AB = wasm_v128_load(i0 + 8);
70 v128_t vi0xCDEF = wasm_v128_load(i0 + 12);
71 i0 += 16;
72 v128_t vi1x0123 = wasm_v128_load(i1);
73 v128_t vi1x4567 = wasm_v128_load(i1 + 4);
74 v128_t vi1x89AB = wasm_v128_load(i1 + 8);
75 v128_t vi1xCDEF = wasm_v128_load(i1 + 12);
76 i1 += 16;
77 v128_t vi2x0123 = wasm_v128_load(i2);
78 v128_t vi2x4567 = wasm_v128_load(i2 + 4);
79 v128_t vi2x89AB = wasm_v128_load(i2 + 8);
80 v128_t vi2xCDEF = wasm_v128_load(i2 + 12);
81 i2 += 16;
82 v128_t vi3x0123 = wasm_v128_load(i3);
83 v128_t vi3x4567 = wasm_v128_load(i3 + 4);
84 v128_t vi3x89AB = wasm_v128_load(i3 + 8);
85 v128_t vi3xCDEF = wasm_v128_load(i3 + 12);
86 i3 += 16;
87
88 v128_t vacc0x0123 = wasm_i32x4_max(vi0x0123, vzero);
89 vi0x0123 = wasm_i32x4_min(vi0x0123, vzero);
90 v128_t vacc0x4567 = wasm_i32x4_max(vi0x4567, vzero);
91 vi0x4567 = wasm_i32x4_min(vi0x4567, vzero);
92 v128_t vacc0x89AB = wasm_i32x4_max(vi0x89AB, vzero);
93 vi0x89AB = wasm_i32x4_min(vi0x89AB, vzero);
94 v128_t vacc0xCDEF = wasm_i32x4_max(vi0xCDEF, vzero);
95 vi0xCDEF = wasm_i32x4_min(vi0xCDEF, vzero);
96 v128_t vacc1x0123 = wasm_i32x4_max(vi1x0123, vzero);
97 vi1x0123 = wasm_i32x4_min(vi1x0123, vzero);
98 v128_t vacc1x4567 = wasm_i32x4_max(vi1x4567, vzero);
99 vi1x4567 = wasm_i32x4_min(vi1x4567, vzero);
100 v128_t vacc1x89AB = wasm_i32x4_max(vi1x89AB, vzero);
101 vi1x89AB = wasm_i32x4_min(vi1x89AB, vzero);
102 v128_t vacc1xCDEF = wasm_i32x4_max(vi1xCDEF, vzero);
103 vi1xCDEF = wasm_i32x4_min(vi1xCDEF, vzero);
104 v128_t vacc2x0123 = wasm_i32x4_max(vi2x0123, vzero);
105 vi2x0123 = wasm_i32x4_min(vi2x0123, vzero);
106 v128_t vacc2x4567 = wasm_i32x4_max(vi2x4567, vzero);
107 vi2x4567 = wasm_i32x4_min(vi2x4567, vzero);
108 v128_t vacc2x89AB = wasm_i32x4_max(vi2x89AB, vzero);
109 vi2x89AB = wasm_i32x4_min(vi2x89AB, vzero);
110 v128_t vacc2xCDEF = wasm_i32x4_max(vi2xCDEF, vzero);
111 vi2xCDEF = wasm_i32x4_min(vi2xCDEF, vzero);
112 v128_t vacc3x0123 = wasm_i32x4_max(vi3x0123, vzero);
113 vi3x0123 = wasm_i32x4_min(vi3x0123, vzero);
114 v128_t vacc3x4567 = wasm_i32x4_max(vi3x4567, vzero);
115 vi3x4567 = wasm_i32x4_min(vi3x4567, vzero);
116 v128_t vacc3x89AB = wasm_i32x4_max(vi3x89AB, vzero);
117 vi3x89AB = wasm_i32x4_min(vi3x89AB, vzero);
118 v128_t vacc3xCDEF = wasm_i32x4_max(vi3xCDEF, vzero);
119 vi3xCDEF = wasm_i32x4_min(vi3xCDEF, vzero);
120
121 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(vi0x0123, vw0123));
122 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(vi0x4567, vw4567));
123 vacc0x89AB = wasm_f32x4_add(vacc0x89AB, wasm_f32x4_mul(vi0x89AB, vw89AB));
124 vacc0xCDEF = wasm_f32x4_add(vacc0xCDEF, wasm_f32x4_mul(vi0xCDEF, vwCDEF));
125 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(vi1x0123, vw0123));
126 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(vi1x4567, vw4567));
127 vacc1x89AB = wasm_f32x4_add(vacc1x89AB, wasm_f32x4_mul(vi1x89AB, vw89AB));
128 vacc1xCDEF = wasm_f32x4_add(vacc1xCDEF, wasm_f32x4_mul(vi1xCDEF, vwCDEF));
129 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(vi2x0123, vw0123));
130 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(vi2x4567, vw4567));
131 vacc2x89AB = wasm_f32x4_add(vacc2x89AB, wasm_f32x4_mul(vi2x89AB, vw89AB));
132 vacc2xCDEF = wasm_f32x4_add(vacc2xCDEF, wasm_f32x4_mul(vi2xCDEF, vwCDEF));
133 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(vi3x0123, vw0123));
134 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(vi3x4567, vw4567));
135 vacc3x89AB = wasm_f32x4_add(vacc3x89AB, wasm_f32x4_mul(vi3x89AB, vw89AB));
136 vacc3xCDEF = wasm_f32x4_add(vacc3xCDEF, wasm_f32x4_mul(vi3xCDEF, vwCDEF));
137
138 wasm_v128_store(o0, vacc0x0123);
139 wasm_v128_store(o0 + 4, vacc0x4567);
140 wasm_v128_store(o0 + 8, vacc0x89AB);
141 wasm_v128_store(o0 + 12, vacc0xCDEF);
142 o0 += 16;
143 wasm_v128_store(o1, vacc1x0123);
144 wasm_v128_store(o1 + 4, vacc1x4567);
145 wasm_v128_store(o1 + 8, vacc1x89AB);
146 wasm_v128_store(o1 + 12, vacc1xCDEF);
147 o1 += 16;
148 wasm_v128_store(o2, vacc2x0123);
149 wasm_v128_store(o2 + 4, vacc2x4567);
150 wasm_v128_store(o2 + 8, vacc2x89AB);
151 wasm_v128_store(o2 + 12, vacc2xCDEF);
152 o2 += 16;
153 wasm_v128_store(o3, vacc3x0123);
154 wasm_v128_store(o3 + 4, vacc3x4567);
155 wasm_v128_store(o3 + 8, vacc3x89AB);
156 wasm_v128_store(o3 + 12, vacc3xCDEF);
157 o3 += 16;
158 }
159 for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
160 const v128_t vw0123 = wasm_v128_load(w);
161 w += 4;
162
163 v128_t vi0x0123 = wasm_v128_load(i0);
164 i0 += 4;
165 v128_t vi1x0123 = wasm_v128_load(i1);
166 i1 += 4;
167 v128_t vi2x0123 = wasm_v128_load(i2);
168 i2 += 4;
169 v128_t vi3x0123 = wasm_v128_load(i3);
170 i3 += 4;
171
172 v128_t vacc0x0123 = wasm_i32x4_max(vi0x0123, vzero);
173 vi0x0123 = wasm_i32x4_min(vi0x0123, vzero);
174 v128_t vacc1x0123 = wasm_i32x4_max(vi1x0123, vzero);
175 vi1x0123 = wasm_i32x4_min(vi1x0123, vzero);
176 v128_t vacc2x0123 = wasm_i32x4_max(vi2x0123, vzero);
177 vi2x0123 = wasm_i32x4_min(vi2x0123, vzero);
178 v128_t vacc3x0123 = wasm_i32x4_max(vi3x0123, vzero);
179 vi3x0123 = wasm_i32x4_min(vi3x0123, vzero);
180
181 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(vi0x0123, vw0123));
182 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(vi1x0123, vw0123));
183 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(vi2x0123, vw0123));
184 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(vi3x0123, vw0123));
185
186 wasm_v128_store(o0, vacc0x0123);
187 o0 += 4;
188 wasm_v128_store(o1, vacc1x0123);
189 o1 += 4;
190 wasm_v128_store(o2, vacc2x0123);
191 o2 += 4;
192 wasm_v128_store(o3, vacc3x0123);
193 o3 += 4;
194 }
195 if XNN_UNLIKELY(c != 0) {
196 const v128_t vw0123 = wasm_v128_load(w);
197 w = (const float*) ((uintptr_t) w + c);
198
199 v128_t vi0x0123 = wasm_v128_load(i0);
200 i0 = (const float*) ((uintptr_t) i0 + c);
201 v128_t vi1x0123 = wasm_v128_load(i1);
202 i1 = (const float*) ((uintptr_t) i1 + c);
203 v128_t vi2x0123 = wasm_v128_load(i2);
204 i2 = (const float*) ((uintptr_t) i2 + c);
205 v128_t vi3x0123 = wasm_v128_load(i3);
206 i3 = (const float*) ((uintptr_t) i3 + c);
207
208 v128_t vacc0x0123 = wasm_i32x4_max(vi0x0123, vzero);
209 vi0x0123 = wasm_i32x4_min(vi0x0123, vzero);
210 v128_t vacc1x0123 = wasm_i32x4_max(vi1x0123, vzero);
211 vi1x0123 = wasm_i32x4_min(vi1x0123, vzero);
212 v128_t vacc2x0123 = wasm_i32x4_max(vi2x0123, vzero);
213 vi2x0123 = wasm_i32x4_min(vi2x0123, vzero);
214 v128_t vacc3x0123 = wasm_i32x4_max(vi3x0123, vzero);
215 vi3x0123 = wasm_i32x4_min(vi3x0123, vzero);
216
217 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(vi0x0123, vw0123));
218 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(vi1x0123, vw0123));
219 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(vi2x0123, vw0123));
220 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(vi3x0123, vw0123));
221
222 if (c & (2 * sizeof(float))) {
223 *((double*) o0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
224 *((double*) o1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
225 *((double*) o2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
226 *((double*) o3) = wasm_f64x2_extract_lane(vacc3x0123, 0);
227
228 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
229 vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
230 vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
231 vacc3x0123 = wasm_v32x4_shuffle(vacc3x0123, vacc3x0123, 2, 3, 2, 3);
232
233 o0 += 2;
234 o1 += 2;
235 o2 += 2;
236 o3 += 2;
237 }
238 if (c & (1 * sizeof(float))) {
239 *o0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
240 *o1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
241 *o2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
242 *o3 = wasm_f32x4_extract_lane(vacc3x0123, 0);
243
244 o0 += 1;
245 o1 += 1;
246 o2 += 1;
247 o3 += 1;
248 }
249 }
250 i0 = (const float*) ((uintptr_t) i0 + input_increment);
251 o0 = (float*) ((uintptr_t) o0 + output_increment);
252 i1 = (const float*) ((uintptr_t) i1 + input_increment);
253 o1 = (float*) ((uintptr_t) o1 + output_increment);
254 i2 = (const float*) ((uintptr_t) i2 + input_increment);
255 o2 = (float*) ((uintptr_t) o2 + output_increment);
256 i3 = (const float*) ((uintptr_t) i3 + input_increment);
257 o3 = (float*) ((uintptr_t) o3 + output_increment);
258 rows = doz(rows, 4);
259 } while (rows != 0);
260 }
261