• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-vbinary/vopc-wasmsimd.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/vbinary.h>
16 
17 
xnn_f32_vmulc_minmax_ukernel__wasmsimd_x86_x16(size_t n,const float * a,const float * b,float * y,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_vmulc_minmax_ukernel__wasmsimd_x86_x16(
19     size_t n,
20     const float* a,
21     const float* b,
22     float* y,
23     const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24 {
25   assert(n != 0);
26   assert(n % sizeof(float) == 0);
27   assert(a != NULL);
28   assert(b != NULL);
29   assert(y != NULL);
30 
31   const v128_t vy_min = wasm_v128_load64_splat(params->wasmsimd.min);
32   const v128_t vy_max = wasm_v128_load64_splat(params->wasmsimd.max);
33   const v128_t vb = wasm_v128_load32_splat(b);
34   for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
35     const v128_t va0123 = wasm_v128_load(a);
36     const v128_t va4567 = wasm_v128_load(a + 4);
37     const v128_t va89AB = wasm_v128_load(a + 8);
38     const v128_t vaCDEF = wasm_v128_load(a + 12);
39     a += 16;
40 
41     v128_t vy0123 = wasm_f32x4_mul(va0123, vb);
42     v128_t vy4567 = wasm_f32x4_mul(va4567, vb);
43     v128_t vy89AB = wasm_f32x4_mul(va89AB, vb);
44     v128_t vyCDEF = wasm_f32x4_mul(vaCDEF, vb);
45 
46 
47     vy0123 = wasm_f32x4_pmax(vy_min, vy0123);
48     vy4567 = wasm_f32x4_pmax(vy_min, vy4567);
49     vy89AB = wasm_f32x4_pmax(vy_min, vy89AB);
50     vyCDEF = wasm_f32x4_pmax(vy_min, vyCDEF);
51 
52     vy0123 = wasm_f32x4_pmin(vy_max, vy0123);
53     vy4567 = wasm_f32x4_pmin(vy_max, vy4567);
54     vy89AB = wasm_f32x4_pmin(vy_max, vy89AB);
55     vyCDEF = wasm_f32x4_pmin(vy_max, vyCDEF);
56 
57     wasm_v128_store(y, vy0123);
58     wasm_v128_store(y + 4, vy4567);
59     wasm_v128_store(y + 8, vy89AB);
60     wasm_v128_store(y + 12, vyCDEF);
61     y += 16;
62   }
63   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
64     const v128_t va = wasm_v128_load(a);
65     a += 4;
66 
67     v128_t vy = wasm_f32x4_mul(va, vb);
68 
69     vy = wasm_f32x4_pmax(vy_min, vy);
70     vy = wasm_f32x4_pmin(vy_max, vy);
71 
72     wasm_v128_store(y, vy);
73     y += 4;
74   }
75   if XNN_UNLIKELY(n != 0) {
76     const v128_t va = wasm_v128_load(a);
77 
78     v128_t vy = wasm_f32x4_mul(va, vb);
79 
80     vy = wasm_f32x4_pmax(vy_min, vy);
81     vy = wasm_f32x4_pmin(vy_max, vy);
82 
83     if (n & (2 * sizeof(float))) {
84       *((double*) y) = wasm_f64x2_extract_lane(vy, 0);
85       vy = wasm_v32x4_shuffle(vy, vy, 2, 3, 2, 3);
86       y += 2;
87     }
88     if (n & (1 * sizeof(float))) {
89       *y = wasm_f32x4_extract_lane(vy, 0);
90     }
91   }
92 }
93