1 // Auto-generated file. Do not edit!
2 // Template: src/f16-f32-vcvt/wasmsimd-int32.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 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/vcvt.h>
16
17
xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x8(size_t n,const void * input,float * output,const union xnn_f16_f32_cvt_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f16_f32_vcvt_ukernel__wasmsimd_int32_x8(
19 size_t n,
20 const void* input,
21 float* output,
22 const union xnn_f16_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23 {
24 assert(n != 0);
25 assert(n % sizeof(uint16_t) == 0);
26 assert(input != NULL);
27 assert(output != NULL);
28
29 const v128_t vsign_mask = wasm_v128_load64_splat(params->wasmsimd_int32.sign_mask);
30 const v128_t vexp_offset = wasm_v128_load64_splat(params->wasmsimd_int32.exp_offset);
31 const v128_t vexp_scale = wasm_v128_load64_splat(params->wasmsimd_int32.exp_scale);
32 const v128_t vmagic_bias = wasm_v128_load64_splat(params->wasmsimd_int32.magic_bias);
33 const v128_t vdenorm_cutoff = wasm_v128_load64_splat(params->wasmsimd_int32.denorm_cutoff);
34
35 const uint16_t* i = (const uint16_t*) input;
36 for (; n >= 8 * sizeof(uint16_t); n -= 8 * sizeof(uint16_t)) {
37 const v128_t vh = wasm_v128_load(i);
38 i += 8;
39
40 const v128_t vzero = wasm_i16x8_const_splat(0);
41 const v128_t vw_lo = wasm_v16x8_shuffle(vzero, vh, 0, 8, 1, 9, 2, 10, 3, 11);
42 const v128_t vw_hi = wasm_v16x8_shuffle(vzero, vh, 4, 12, 5, 13, 6, 14, 7, 15);
43
44 const v128_t vsign_lo = wasm_v128_and(vw_lo, vsign_mask);
45 const v128_t vsign_hi = wasm_v128_and(vw_hi, vsign_mask);
46
47 const v128_t vnonsign_lo = wasm_v128_xor(vw_lo, vsign_lo);
48 const v128_t vnonsign_hi = wasm_v128_xor(vw_hi, vsign_hi);
49
50 const v128_t vnorm_lo = wasm_f32x4_mul(wasm_i32x4_add(wasm_u32x4_shr(vnonsign_lo, 3), vexp_offset), vexp_scale);
51 const v128_t vnorm_hi = wasm_f32x4_mul(wasm_i32x4_add(wasm_u32x4_shr(vnonsign_hi, 3), vexp_offset), vexp_scale);
52
53 const v128_t vdenorm_lo = wasm_f32x4_sub(wasm_v128_or(wasm_u32x4_shr(vnonsign_lo, 16), vmagic_bias), vmagic_bias);
54 const v128_t vdenorm_hi = wasm_f32x4_sub(wasm_v128_or(wasm_u32x4_shr(vnonsign_hi, 16), vmagic_bias), vmagic_bias);
55
56 const v128_t vxmask_lo = wasm_i32x4_gt(vnonsign_lo, vdenorm_cutoff);
57 const v128_t vxmask_hi = wasm_i32x4_gt(vnonsign_hi, vdenorm_cutoff);
58
59 const v128_t vf_lo = wasm_v128_or(vsign_lo, wasm_v128_bitselect(vnorm_lo, vdenorm_lo, vxmask_lo));
60 const v128_t vf_hi = wasm_v128_or(vsign_hi, wasm_v128_bitselect(vnorm_hi, vdenorm_hi, vxmask_hi));
61
62 wasm_v128_store(output, vf_lo);
63 wasm_v128_store(output + 4, vf_hi);
64 output += 8;
65 }
66 if XNN_UNLIKELY(n != 0) {
67 assert(n >= 1 * sizeof(uint16_t));
68 assert(n <= 7 * sizeof(uint16_t));
69 const v128_t vh = wasm_v128_load(i);
70
71 const v128_t vzero = wasm_i16x8_const_splat(0);
72 const v128_t vw_lo = wasm_v16x8_shuffle(vzero, vh, 0, 8, 1, 9, 2, 10, 3, 11);
73 const v128_t vw_hi = wasm_v16x8_shuffle(vzero, vh, 4, 12, 5, 13, 6, 14, 7, 15);
74
75 const v128_t vsign_lo = wasm_v128_and(vw_lo, vsign_mask);
76 const v128_t vsign_hi = wasm_v128_and(vw_hi, vsign_mask);
77
78 const v128_t vnonsign_lo = wasm_v128_xor(vw_lo, vsign_lo);
79 const v128_t vnonsign_hi = wasm_v128_xor(vw_hi, vsign_hi);
80
81 const v128_t vnorm_lo = wasm_f32x4_mul(wasm_i32x4_add(wasm_u32x4_shr(vnonsign_lo, 3), vexp_offset), vexp_scale);
82 const v128_t vnorm_hi = wasm_f32x4_mul(wasm_i32x4_add(wasm_u32x4_shr(vnonsign_hi, 3), vexp_offset), vexp_scale);
83
84 const v128_t vdenorm_lo = wasm_f32x4_sub(wasm_v128_or(wasm_u32x4_shr(vnonsign_lo, 16), vmagic_bias), vmagic_bias);
85 const v128_t vdenorm_hi = wasm_f32x4_sub(wasm_v128_or(wasm_u32x4_shr(vnonsign_hi, 16), vmagic_bias), vmagic_bias);
86
87 const v128_t vxmask_lo = wasm_i32x4_gt(vnonsign_lo, vdenorm_cutoff);
88 v128_t vf = wasm_v128_or(vsign_lo, wasm_v128_bitselect(vnorm_lo, vdenorm_lo, vxmask_lo));
89
90 if (n & (4 * sizeof(uint16_t))) {
91 wasm_v128_store(output, vf);
92 output += 4;
93
94 const v128_t vxmask_hi = wasm_i32x4_gt(vnonsign_hi, vdenorm_cutoff);
95 vf = wasm_v128_or(vsign_hi, wasm_v128_bitselect(vnorm_hi, vdenorm_hi, vxmask_hi));
96 }
97 if (n & (2 * sizeof(uint16_t))) {
98 *((double*) output) = wasm_f64x2_extract_lane(vf, 0);
99 output += 2;
100
101 vf = wasm_v64x2_shuffle(vf, vf, 1, 1);
102 }
103 if (n & (1 * sizeof(uint16_t))) {
104 *((float*) output) = wasm_f32x4_extract_lane(vf, 0);
105 }
106 }
107 }
108