1 // Auto-generated file. Do not edit!
2 // Template: src/f32-prelu/wasmsimd-bitselect.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_bitselect_1x8(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_bitselect_1x8(
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_DISABLE_TSAN
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_splat(0);
38 do {
39 const float* w = weights;
40 size_t c = channels;
41 for (; c >= 8 * sizeof(float); c -= 8 * sizeof(float)) {
42 const v128_t vw0123 = wasm_v128_load(w);
43 const v128_t vw4567 = wasm_v128_load(w + 4);
44 w += 8;
45
46 const v128_t vi0x0123 = wasm_v128_load(i0);
47 const v128_t vi0x4567 = wasm_v128_load(i0 + 4);
48 i0 += 8;
49
50 v128_t vacc0x0123 = wasm_f32x4_mul(vi0x0123, vw0123);
51 const v128_t vmask0x0123 = wasm_i32x4_lt(vi0x0123, vzero);
52 v128_t vacc0x4567 = wasm_f32x4_mul(vi0x4567, vw4567);
53 const v128_t vmask0x4567 = wasm_i32x4_lt(vi0x4567, vzero);
54
55 vacc0x0123 = wasm_v128_bitselect(vacc0x0123, vi0x0123, vmask0x0123);
56 vacc0x4567 = wasm_v128_bitselect(vacc0x4567, vi0x4567, vmask0x4567);
57
58 wasm_v128_store(o0, vacc0x0123);
59 wasm_v128_store(o0 + 4, vacc0x4567);
60 o0 += 8;
61 }
62 for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
63 const v128_t vw0123 = wasm_v128_load(w);
64 w += 4;
65
66 const v128_t vi0x0123 = wasm_v128_load(i0);
67 i0 += 4;
68
69 v128_t vacc0x0123 = wasm_f32x4_mul(vi0x0123, vw0123);
70 const v128_t vmask0x0123 = wasm_i32x4_lt(vi0x0123, vzero);
71
72 vacc0x0123 = wasm_v128_bitselect(vacc0x0123, vi0x0123, vmask0x0123);
73
74 wasm_v128_store(o0, vacc0x0123);
75 o0 += 4;
76 }
77 if XNN_UNLIKELY(c != 0) {
78 const v128_t vw0123 = wasm_v128_load(w);
79 w = (const float*) ((uintptr_t) w + c);
80
81 const v128_t vi0x0123 = wasm_v128_load(i0);
82 i0 = (const float*) ((uintptr_t) i0 + c);
83
84 v128_t vacc0x0123 = wasm_f32x4_mul(vi0x0123, vw0123);
85 const v128_t vmask0x0123 = wasm_i32x4_lt(vi0x0123, vzero);
86
87 vacc0x0123 = wasm_v128_bitselect(vacc0x0123, vi0x0123, vmask0x0123);
88
89 if (c & (2 * sizeof(float))) {
90 *((double*) o0) = wasm_f64x2_extract_lane(vacc0x0123, 0);
91
92 vacc0x0123 = wasm_v32x4_shuffle(vacc0x0123, vacc0x0123, 2, 3, 2, 3);
93
94 o0 += 2;
95 }
96 if (c & (1 * sizeof(float))) {
97 *o0 = wasm_f32x4_extract_lane(vacc0x0123, 0);
98
99 o0 += 1;
100 }
101 }
102 i0 = (const float*) ((uintptr_t) i0 + input_increment);
103 o0 = (float*) ((uintptr_t) o0 + output_increment);
104 rows = doz(rows, 1);
105 } while (rows != 0);
106 }
107