1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vbinary/vop-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_vmul_minmax_ukernel__wasmsimd_x86_x8(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_vmul_minmax_ukernel__wasmsimd_x86_x8(
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_DISABLE_TSAN
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_v32x4_load_splat(¶ms->scalar.min);
32 const v128_t vy_max = wasm_v32x4_load_splat(¶ms->scalar.max);
33
34 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
35 const v128_t va0123 = wasm_v128_load(a);
36 const v128_t va4567 = wasm_v128_load(a + 4);
37 a += 8;
38
39 const v128_t vb0123 = wasm_v128_load(b);
40 const v128_t vb4567 = wasm_v128_load(b + 4);
41 b += 8;
42
43 v128_t vy0123 = wasm_f32x4_mul(va0123, vb0123);
44 v128_t vy4567 = wasm_f32x4_mul(va4567, vb4567);
45
46
47 const v128_t vltmask0123 = wasm_f32x4_lt(vy0123, vy_min);
48 const v128_t vltmask4567 = wasm_f32x4_lt(vy4567, vy_min);
49
50 const v128_t vngtmask0123 = wasm_f32x4_le(vy0123, vy_max);
51 vy0123 = wasm_v128_bitselect(vy_min, vy0123, vltmask0123);
52 const v128_t vngtmask4567 = wasm_f32x4_le(vy4567, vy_max);
53 vy4567 = wasm_v128_bitselect(vy_min, vy4567, vltmask4567);
54
55 vy0123 = wasm_v128_bitselect(vy0123, vy_max, vngtmask0123);
56 vy4567 = wasm_v128_bitselect(vy4567, vy_max, vngtmask4567);
57
58 wasm_v128_store(y, vy0123);
59 wasm_v128_store(y + 4, vy4567);
60 y += 8;
61 }
62 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
63 const v128_t va = wasm_v128_load(a);
64 a += 4;
65
66 const v128_t vb = wasm_v128_load(b);
67 b += 4;
68
69 v128_t vy = wasm_f32x4_mul(va, vb);
70
71 const v128_t vltmask = wasm_f32x4_lt(vy, vy_min);
72 const v128_t vngtmask = wasm_f32x4_le(vy, vy_max);
73 vy = wasm_v128_bitselect(vy_min, vy, vltmask);
74 vy = wasm_v128_bitselect(vy, vy_max, vngtmask);
75
76 wasm_v128_store(y, vy);
77 y += 4;
78 }
79 if XNN_UNLIKELY(n != 0) {
80 const v128_t va = wasm_v128_load(a);
81 const v128_t vb = wasm_v128_load(b);
82
83 v128_t vy = wasm_f32x4_mul(va, vb);
84
85 const v128_t vltmask = wasm_f32x4_lt(vy, vy_min);
86 const v128_t vngtmask = wasm_f32x4_le(vy, vy_max);
87 vy = wasm_v128_bitselect(vy_min, vy, vltmask);
88 vy = wasm_v128_bitselect(vy, vy_max, vngtmask);
89
90 if (n & (2 * sizeof(float))) {
91 *((double*) y) = wasm_f64x2_extract_lane(vy, 0);
92 vy = wasm_v32x4_shuffle(vy, vy, 2, 3, 2, 3);
93 y += 2;
94 }
95 if (n & (1 * sizeof(float))) {
96 *y = wasm_f32x4_extract_lane(vy, 0);
97 }
98 }
99 }
100