xref: /aosp_15_r20/external/XNNPACK/src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-vlrelu/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/common.h>
15 #include <xnnpack/vunary.h>
16 
17 
xnn_f32_vlrelu_ukernel__wasmsimd_bitselect_x8(size_t n,const float * x,float * y,const union xnn_f32_lrelu_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_vlrelu_ukernel__wasmsimd_bitselect_x8(
19     size_t n,
20     const float* x,
21     float* y,
22     const union xnn_f32_lrelu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23 {
24   assert(n != 0);
25   assert(n % sizeof(float) == 0);
26 
27   const v128_t vslope = wasm_v128_load64_splat(params->wasmsimd.slope);
28   for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
29     const v128_t vx0123 = wasm_v128_load(x);
30     const v128_t vx4567 = wasm_v128_load(x + 4);
31     x += 8;
32 
33     v128_t vacc0123 = wasm_f32x4_mul(vx0123, vslope);
34     const v128_t vmask0123 = wasm_i32x4_shr(vx0123, 31);
35     v128_t vacc4567 = wasm_f32x4_mul(vx4567, vslope);
36     const v128_t vmask4567 = wasm_i32x4_shr(vx4567, 31);
37 
38     vacc0123 = wasm_v128_bitselect(vacc0123, vx0123, vmask0123);
39     vacc4567 = wasm_v128_bitselect(vacc4567, vx4567, vmask4567);
40 
41     wasm_v128_store(y, vacc0123);
42     wasm_v128_store(y + 4, vacc4567);
43     y += 8;
44   }
45   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
46     const v128_t vx = wasm_v128_load(x);
47     x += 4;
48     v128_t vacc = wasm_f32x4_mul(vx, vslope);
49     const v128_t vmask = wasm_i32x4_shr(vx, 31);
50     vacc = wasm_v128_bitselect(vacc, vx, vmask);
51     wasm_v128_store(y, vacc);
52     y += 4;
53   }
54   if XNN_UNLIKELY(n != 0) {
55     const v128_t vx = wasm_v128_load(x);
56     v128_t vacc = wasm_f32x4_mul(vx, vslope);
57     const v128_t vmask = wasm_i32x4_shr(vx, 31);
58     vacc = wasm_v128_bitselect(vacc, vx, vmask);
59 
60     if (n & (2 * sizeof(float))) {
61       *((double*) y) = wasm_f64x2_extract_lane(vacc, 0);
62       vacc = wasm_v32x4_shuffle(vacc, vacc, 2, 3, 2, 3);
63       y += 2;
64     }
65     if (n & (1 * sizeof(float))) {
66       *y = wasm_f32x4_extract_lane(vacc, 0);
67     }
68   }
69 }
70