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_4x8(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_4x8(
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 >= 8 * sizeof(float); c -= 8 * sizeof(float)) {
61 const v128_t vw0123 = wasm_v128_load(w);
62 const v128_t vw4567 = wasm_v128_load(w + 4);
63 w += 8;
64
65 v128_t vi0x0123 = wasm_v128_load(i0);
66 v128_t vi0x4567 = wasm_v128_load(i0 + 4);
67 i0 += 8;
68 v128_t vi1x0123 = wasm_v128_load(i1);
69 v128_t vi1x4567 = wasm_v128_load(i1 + 4);
70 i1 += 8;
71 v128_t vi2x0123 = wasm_v128_load(i2);
72 v128_t vi2x4567 = wasm_v128_load(i2 + 4);
73 i2 += 8;
74 v128_t vi3x0123 = wasm_v128_load(i3);
75 v128_t vi3x4567 = wasm_v128_load(i3 + 4);
76 i3 += 8;
77
78 v128_t vacc0x0123 = wasm_i32x4_max(vi0x0123, vzero);
79 vi0x0123 = wasm_i32x4_min(vi0x0123, vzero);
80 v128_t vacc0x4567 = wasm_i32x4_max(vi0x4567, vzero);
81 vi0x4567 = wasm_i32x4_min(vi0x4567, vzero);
82 v128_t vacc1x0123 = wasm_i32x4_max(vi1x0123, vzero);
83 vi1x0123 = wasm_i32x4_min(vi1x0123, vzero);
84 v128_t vacc1x4567 = wasm_i32x4_max(vi1x4567, vzero);
85 vi1x4567 = wasm_i32x4_min(vi1x4567, vzero);
86 v128_t vacc2x0123 = wasm_i32x4_max(vi2x0123, vzero);
87 vi2x0123 = wasm_i32x4_min(vi2x0123, vzero);
88 v128_t vacc2x4567 = wasm_i32x4_max(vi2x4567, vzero);
89 vi2x4567 = wasm_i32x4_min(vi2x4567, vzero);
90 v128_t vacc3x0123 = wasm_i32x4_max(vi3x0123, vzero);
91 vi3x0123 = wasm_i32x4_min(vi3x0123, vzero);
92 v128_t vacc3x4567 = wasm_i32x4_max(vi3x4567, vzero);
93 vi3x4567 = wasm_i32x4_min(vi3x4567, vzero);
94
95 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(vi0x0123, vw0123));
96 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(vi0x4567, vw4567));
97 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(vi1x0123, vw0123));
98 vacc1x4567 = wasm_f32x4_add(vacc1x4567, wasm_f32x4_mul(vi1x4567, vw4567));
99 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(vi2x0123, vw0123));
100 vacc2x4567 = wasm_f32x4_add(vacc2x4567, wasm_f32x4_mul(vi2x4567, vw4567));
101 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(vi3x0123, vw0123));
102 vacc3x4567 = wasm_f32x4_add(vacc3x4567, wasm_f32x4_mul(vi3x4567, vw4567));
103
104 wasm_v128_store(o0, vacc0x0123);
105 wasm_v128_store(o0 + 4, vacc0x4567);
106 o0 += 8;
107 wasm_v128_store(o1, vacc1x0123);
108 wasm_v128_store(o1 + 4, vacc1x4567);
109 o1 += 8;
110 wasm_v128_store(o2, vacc2x0123);
111 wasm_v128_store(o2 + 4, vacc2x4567);
112 o2 += 8;
113 wasm_v128_store(o3, vacc3x0123);
114 wasm_v128_store(o3 + 4, vacc3x4567);
115 o3 += 8;
116 }
117 for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
118 const v128_t vw0123 = wasm_v128_load(w);
119 w += 4;
120
121 v128_t vi0x0123 = wasm_v128_load(i0);
122 i0 += 4;
123 v128_t vi1x0123 = wasm_v128_load(i1);
124 i1 += 4;
125 v128_t vi2x0123 = wasm_v128_load(i2);
126 i2 += 4;
127 v128_t vi3x0123 = wasm_v128_load(i3);
128 i3 += 4;
129
130 v128_t vacc0x0123 = wasm_i32x4_max(vi0x0123, vzero);
131 vi0x0123 = wasm_i32x4_min(vi0x0123, vzero);
132 v128_t vacc1x0123 = wasm_i32x4_max(vi1x0123, vzero);
133 vi1x0123 = wasm_i32x4_min(vi1x0123, vzero);
134 v128_t vacc2x0123 = wasm_i32x4_max(vi2x0123, vzero);
135 vi2x0123 = wasm_i32x4_min(vi2x0123, vzero);
136 v128_t vacc3x0123 = wasm_i32x4_max(vi3x0123, vzero);
137 vi3x0123 = wasm_i32x4_min(vi3x0123, vzero);
138
139 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(vi0x0123, vw0123));
140 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(vi1x0123, vw0123));
141 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(vi2x0123, vw0123));
142 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(vi3x0123, vw0123));
143
144 wasm_v128_store(o0, vacc0x0123);
145 o0 += 4;
146 wasm_v128_store(o1, vacc1x0123);
147 o1 += 4;
148 wasm_v128_store(o2, vacc2x0123);
149 o2 += 4;
150 wasm_v128_store(o3, vacc3x0123);
151 o3 += 4;
152 }
153 if XNN_UNLIKELY(c != 0) {
154 const v128_t vw0123 = wasm_v128_load(w);
155 w = (const float*) ((uintptr_t) w + c);
156
157 v128_t vi0x0123 = wasm_v128_load(i0);
158 i0 = (const float*) ((uintptr_t) i0 + c);
159 v128_t vi1x0123 = wasm_v128_load(i1);
160 i1 = (const float*) ((uintptr_t) i1 + c);
161 v128_t vi2x0123 = wasm_v128_load(i2);
162 i2 = (const float*) ((uintptr_t) i2 + c);
163 v128_t vi3x0123 = wasm_v128_load(i3);
164 i3 = (const float*) ((uintptr_t) i3 + c);
165
166 v128_t vacc0x0123 = wasm_i32x4_max(vi0x0123, vzero);
167 vi0x0123 = wasm_i32x4_min(vi0x0123, vzero);
168 v128_t vacc1x0123 = wasm_i32x4_max(vi1x0123, vzero);
169 vi1x0123 = wasm_i32x4_min(vi1x0123, vzero);
170 v128_t vacc2x0123 = wasm_i32x4_max(vi2x0123, vzero);
171 vi2x0123 = wasm_i32x4_min(vi2x0123, vzero);
172 v128_t vacc3x0123 = wasm_i32x4_max(vi3x0123, vzero);
173 vi3x0123 = wasm_i32x4_min(vi3x0123, vzero);
174
175 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(vi0x0123, vw0123));
176 vacc1x0123 = wasm_f32x4_add(vacc1x0123, wasm_f32x4_mul(vi1x0123, vw0123));
177 vacc2x0123 = wasm_f32x4_add(vacc2x0123, wasm_f32x4_mul(vi2x0123, vw0123));
178 vacc3x0123 = wasm_f32x4_add(vacc3x0123, wasm_f32x4_mul(vi3x0123, vw0123));
179
180 if (c & (2 * sizeof(float))) {
181 *((double*) o0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
182 *((double*) o1) = wasm_f64x2_extract_lane(vacc1x0123, 0);
183 *((double*) o2) = wasm_f64x2_extract_lane(vacc2x0123, 0);
184 *((double*) o3) = wasm_f64x2_extract_lane(vacc3x0123, 0);
185
186 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
187 vacc1x0123 = wasm_v32x4_shuffle(vacc1x0123, vacc1x0123, 2, 3, 2, 3);
188 vacc2x0123 = wasm_v32x4_shuffle(vacc2x0123, vacc2x0123, 2, 3, 2, 3);
189 vacc3x0123 = wasm_v32x4_shuffle(vacc3x0123, vacc3x0123, 2, 3, 2, 3);
190
191 o0 += 2;
192 o1 += 2;
193 o2 += 2;
194 o3 += 2;
195 }
196 if (c & (1 * sizeof(float))) {
197 *o0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
198 *o1 = wasm_f32x4_extract_lane(vacc1x0123, 0);
199 *o2 = wasm_f32x4_extract_lane(vacc2x0123, 0);
200 *o3 = wasm_f32x4_extract_lane(vacc3x0123, 0);
201
202 o0 += 1;
203 o1 += 1;
204 o2 += 1;
205 o3 += 1;
206 }
207 }
208 i0 = (const float*) ((uintptr_t) i0 + input_increment);
209 o0 = (float*) ((uintptr_t) o0 + output_increment);
210 i1 = (const float*) ((uintptr_t) i1 + input_increment);
211 o1 = (float*) ((uintptr_t) o1 + output_increment);
212 i2 = (const float*) ((uintptr_t) i2 + input_increment);
213 o2 = (float*) ((uintptr_t) o2 + output_increment);
214 i3 = (const float*) ((uintptr_t) i3 + input_increment);
215 o3 = (float*) ((uintptr_t) o3 + output_increment);
216 rows = doz(rows, 4);
217 } while (rows != 0);
218 }
219