• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f16-f32-vcvt/wasmsimd-int16.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_int16_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_int16_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_int16.sign_mask);
30   const v128_t vexp_offset = wasm_v128_load64_splat(params->wasmsimd_int16.exp_offset);
31   const v128_t vexp_scale = wasm_v128_load64_splat(params->wasmsimd_int16.exp_scale);
32   const v128_t vmagic_mask = wasm_v128_load64_splat(params->wasmsimd_int16.magic_mask);
33   const v128_t vmagic_bias = wasm_v128_load64_splat(params->wasmsimd_int16.magic_bias);
34   const v128_t vdenorm_cutoff = wasm_v128_load64_splat(params->wasmsimd_int16.denorm_cutoff);
35 
36   const uint16_t* i = (const uint16_t*) input;
37   for (; n >= 8 * sizeof(uint16_t); n -= 8 * sizeof(uint16_t)) {
38     const v128_t vh = wasm_v128_load(i);
39     i += 8;
40 
41     const v128_t vsign = wasm_v128_and(vh, vsign_mask);
42 
43     const v128_t vnonsign = wasm_v128_xor(vh, vsign);
44 
45     const v128_t vprenorm_lo = wasm_i16x8_shl(vnonsign, 13);
46     const v128_t vprenorm_hi = wasm_i16x8_add(wasm_u16x8_shr(vnonsign, 3), vexp_offset);
47 
48     const v128_t vnorm_lo = wasm_f32x4_mul(wasm_v16x8_shuffle(vprenorm_lo, vprenorm_hi, 0,  8, 1,  9, 2, 10, 3, 11), vexp_scale);
49     const v128_t vnorm_hi = wasm_f32x4_mul(wasm_v16x8_shuffle(vprenorm_lo, vprenorm_hi, 4, 12, 5, 13, 6, 14, 7, 15), vexp_scale);
50 
51     const v128_t vdenorm_lo = wasm_f32x4_sub(wasm_v16x8_shuffle(vnonsign, vmagic_mask, 0,  8, 1,  9, 2, 10, 3, 11), vmagic_bias);
52     const v128_t vdenorm_hi = wasm_f32x4_sub(wasm_v16x8_shuffle(vnonsign, vmagic_mask, 4, 12, 5, 13, 6, 14, 7, 15), vmagic_bias);
53 
54     const v128_t vmask = wasm_i16x8_gt(vnonsign, vdenorm_cutoff);
55     const v128_t vzero = wasm_i16x8_const_splat(0);
56 
57     const v128_t vxmask_lo = wasm_i32x4_extend_low_i16x8(vmask);
58     const v128_t vf_lo = wasm_v128_or(wasm_v16x8_shuffle(vzero, vsign, 0,  8, 1,  9, 2, 10, 3, 11),
59       wasm_v128_bitselect(vnorm_lo, vdenorm_lo, vxmask_lo));
60 
61     const v128_t vxmask_hi = wasm_i32x4_extend_high_i16x8(vmask);
62     const v128_t vf_hi = wasm_v128_or(wasm_v16x8_shuffle(vzero, vsign, 4, 12, 5, 13, 6, 14, 7, 15),
63       wasm_v128_bitselect(vnorm_hi, vdenorm_hi, vxmask_hi));
64 
65     wasm_v128_store(output, vf_lo);
66     wasm_v128_store(output + 4, vf_hi);
67     output += 8;
68   }
69   if XNN_UNLIKELY(n != 0) {
70     assert(n >= 1 * sizeof(uint16_t));
71     assert(n <= 7 * sizeof(uint16_t));
72     const v128_t vh = wasm_v128_load(i);
73 
74     const v128_t vsign = wasm_v128_and(vh, vsign_mask);
75 
76     const v128_t vnonsign = wasm_v128_xor(vh, vsign);
77 
78     const v128_t vprenorm_lo = wasm_i16x8_shl(vnonsign, 13);
79     const v128_t vprenorm_hi = wasm_i16x8_add(wasm_u16x8_shr(vnonsign, 3), vexp_offset);
80 
81     const v128_t vnorm_lo = wasm_f32x4_mul(wasm_v16x8_shuffle(vprenorm_lo, vprenorm_hi, 0,  8, 1,  9, 2, 10, 3, 11), vexp_scale);
82     const v128_t vnorm_hi = wasm_f32x4_mul(wasm_v16x8_shuffle(vprenorm_lo, vprenorm_hi, 4, 12, 5, 13, 6, 14, 7, 15), vexp_scale);
83 
84     const v128_t vdenorm_lo = wasm_f32x4_sub(wasm_v16x8_shuffle(vnonsign, vmagic_mask, 0,  8, 1,  9, 2, 10, 3, 11), vmagic_bias);
85     const v128_t vdenorm_hi = wasm_f32x4_sub(wasm_v16x8_shuffle(vnonsign, vmagic_mask, 4, 12, 5, 13, 6, 14, 7, 15), vmagic_bias);
86 
87     const v128_t vmask = wasm_i16x8_gt(vnonsign, vdenorm_cutoff);
88     const v128_t vzero = wasm_i16x8_const_splat(0);
89 
90     const v128_t vxmask_lo = wasm_i32x4_extend_low_i16x8(vmask);
91     v128_t vf = wasm_v128_or(wasm_v16x8_shuffle(vzero, vsign, 0, 8, 1, 9, 2, 10, 3, 11),
92       wasm_v128_bitselect(vnorm_lo, vdenorm_lo, vxmask_lo));
93 
94     if (n & (4 * sizeof(uint16_t))) {
95       wasm_v128_store(output, vf);
96       output += 4;
97 
98       const v128_t vxmask_hi = wasm_i32x4_extend_high_i16x8(vmask);
99       vf = wasm_v128_or(wasm_v16x8_shuffle(vzero, vsign, 4, 12, 5, 13, 6, 14, 7, 15),
100         wasm_v128_bitselect(vnorm_hi, vdenorm_hi, vxmask_hi));
101     }
102     if (n & (2 * sizeof(uint16_t))) {
103       *((double*) output) = wasm_f64x2_extract_lane(vf, 0);
104       output += 2;
105 
106       vf = wasm_v64x2_shuffle(vf, vf, 1, 1);
107     }
108     if (n & (1 * sizeof(uint16_t))) {
109       *((float*) output) = wasm_f32x4_extract_lane(vf, 0);
110     }
111   }
112 }
113