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_1x16(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_1x16(
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
34 const size_t input_increment = input_stride * 1 - channels;
35 const size_t output_increment = output_stride * 1 - channels;
36
37 const v128_t vzero = wasm_i32x4_const_splat(0);
38 do {
39
40 const float* w = weights;
41 size_t c = channels;
42 for (; c >= 16 * sizeof(float); c -= 16 * sizeof(float)) {
43 const v128_t vw0123 = wasm_v128_load(w);
44 const v128_t vw4567 = wasm_v128_load(w + 4);
45 const v128_t vw89AB = wasm_v128_load(w + 8);
46 const v128_t vwCDEF = wasm_v128_load(w + 12);
47 w += 16;
48
49 v128_t vi0x0123 = wasm_v128_load(i0);
50 v128_t vi0x4567 = wasm_v128_load(i0 + 4);
51 v128_t vi0x89AB = wasm_v128_load(i0 + 8);
52 v128_t vi0xCDEF = wasm_v128_load(i0 + 12);
53 i0 += 16;
54
55 v128_t vacc0x0123 = wasm_i32x4_max(vi0x0123, vzero);
56 vi0x0123 = wasm_i32x4_min(vi0x0123, vzero);
57 v128_t vacc0x4567 = wasm_i32x4_max(vi0x4567, vzero);
58 vi0x4567 = wasm_i32x4_min(vi0x4567, vzero);
59 v128_t vacc0x89AB = wasm_i32x4_max(vi0x89AB, vzero);
60 vi0x89AB = wasm_i32x4_min(vi0x89AB, vzero);
61 v128_t vacc0xCDEF = wasm_i32x4_max(vi0xCDEF, vzero);
62 vi0xCDEF = wasm_i32x4_min(vi0xCDEF, vzero);
63
64 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(vi0x0123, vw0123));
65 vacc0x4567 = wasm_f32x4_add(vacc0x4567, wasm_f32x4_mul(vi0x4567, vw4567));
66 vacc0x89AB = wasm_f32x4_add(vacc0x89AB, wasm_f32x4_mul(vi0x89AB, vw89AB));
67 vacc0xCDEF = wasm_f32x4_add(vacc0xCDEF, wasm_f32x4_mul(vi0xCDEF, vwCDEF));
68
69 wasm_v128_store(o0, vacc0x0123);
70 wasm_v128_store(o0 + 4, vacc0x4567);
71 wasm_v128_store(o0 + 8, vacc0x89AB);
72 wasm_v128_store(o0 + 12, vacc0xCDEF);
73 o0 += 16;
74 }
75 for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
76 const v128_t vw0123 = wasm_v128_load(w);
77 w += 4;
78
79 v128_t vi0x0123 = wasm_v128_load(i0);
80 i0 += 4;
81
82 v128_t vacc0x0123 = wasm_i32x4_max(vi0x0123, vzero);
83 vi0x0123 = wasm_i32x4_min(vi0x0123, vzero);
84
85 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(vi0x0123, vw0123));
86
87 wasm_v128_store(o0, vacc0x0123);
88 o0 += 4;
89 }
90 if XNN_UNLIKELY(c != 0) {
91 const v128_t vw0123 = wasm_v128_load(w);
92 w = (const float*) ((uintptr_t) w + c);
93
94 v128_t vi0x0123 = wasm_v128_load(i0);
95 i0 = (const float*) ((uintptr_t) i0 + c);
96
97 v128_t vacc0x0123 = wasm_i32x4_max(vi0x0123, vzero);
98 vi0x0123 = wasm_i32x4_min(vi0x0123, vzero);
99
100 vacc0x0123 = wasm_f32x4_add(vacc0x0123, wasm_f32x4_mul(vi0x0123, vw0123));
101
102 if (c & (2 * sizeof(float))) {
103 *((double*) o0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
104
105 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
106
107 o0 += 2;
108 }
109 if (c & (1 * sizeof(float))) {
110 *o0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
111
112 o0 += 1;
113 }
114 }
115 i0 = (const float*) ((uintptr_t) i0 + input_increment);
116 o0 = (float*) ((uintptr_t) o0 + output_increment);
117 rows = doz(rows, 1);
118 } while (rows != 0);
119 }
120