1 // Auto-generated file. Do not edit!
2 // Template: src/f16-f32-vcvt/neon-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 <arm_neon.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/vcvt.h>
16
17
xnn_f16_f32_vcvt_ukernel__neon_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__neon_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 uint16x8_t vsign_mask = vmovq_n_u16(0x8000);
30 const uint16x8_t vexp_offset = vmovq_n_u16(0x7000);
31 const float32x4_t vexp_scale = vld1q_dup_f32(¶ms->neon.exp_scale);
32 const uint32x4_t vmagic_bias = vmovq_n_u32(0x3F000000);
33 const uint16x8_t vdenorm_cutoff = vmovq_n_u16(0x0400);
34
35 const uint16_t* i = (const uint16_t*) input;
36 for (; n >= 8 * sizeof(uint16_t); n -= 8 * sizeof(uint16_t)) {
37 const uint16x8_t vh = vld1q_u16(i); i += 8;
38
39 const uint16x8_t vsign = vandq_u16(vh, vsign_mask);
40
41 const uint16x8_t vnonsign = veorq_u16(vh, vsign);
42
43 const uint16x8x2_t vprenorm = vzipq_u16(vshlq_n_u16(vnonsign, 13), vsraq_n_u16(vexp_offset, vnonsign, 3));
44 const float32x4_t vnorm_lo = vmulq_f32(vreinterpretq_f32_u16(vprenorm.val[0]), vexp_scale);
45 const float32x4_t vnorm_hi = vmulq_f32(vreinterpretq_f32_u16(vprenorm.val[1]), vexp_scale);
46
47 const float32x4_t vdenorm_lo = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_low_u16(vnonsign))), vreinterpretq_f32_u32(vmagic_bias));
48 const float32x4_t vdenorm_hi = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_high_u16(vnonsign))), vreinterpretq_f32_u32(vmagic_bias));
49
50 const uint16x8_t vmask = vcgtq_u16(vnonsign, vdenorm_cutoff);
51
52 const uint32x4_t vxmask_lo = vreinterpretq_u32_s32(vmovl_s16(vreinterpret_s16_u16(vget_low_u16(vmask))));
53 const uint32x4_t vf_lo = vorrq_u32(vshll_n_u16(vget_low_u16(vsign), 16),
54 vreinterpretq_u32_f32(vbslq_f32(vxmask_lo, vnorm_lo, vdenorm_lo)));
55
56 const uint32x4_t vxmask_hi = vreinterpretq_u32_s32(vmovl_s16(vreinterpret_s16_u16(vget_high_u16(vmask))));
57 const uint32x4_t vf_hi = vorrq_u32(vshll_n_u16(vget_high_u16(vsign), 16),
58 vreinterpretq_u32_f32(vbslq_f32(vxmask_hi, vnorm_hi, vdenorm_hi)));
59
60 vst1q_f32(output, vreinterpretq_f32_u32(vf_lo)); output += 4;
61 vst1q_f32(output, vreinterpretq_f32_u32(vf_hi)); output += 4;
62 }
63 if XNN_UNPREDICTABLE(n != 0) {
64 const uint16x8_t vh = vld1q_u16(i); i += 8;
65
66 const uint16x8_t vsign = vandq_u16(vh, vsign_mask);
67
68 const uint16x8_t vnonsign = veorq_u16(vh, vsign);
69
70 const uint16x8x2_t vprenorm = vzipq_u16(vshlq_n_u16(vnonsign, 13), vsraq_n_u16(vexp_offset, vnonsign, 3));
71 const float32x4_t vnorm_lo = vmulq_f32(vreinterpretq_f32_u16(vprenorm.val[0]), vexp_scale);
72 const float32x4_t vnorm_hi = vmulq_f32(vreinterpretq_f32_u16(vprenorm.val[1]), vexp_scale);
73
74 const float32x4_t vdenorm_lo = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_low_u16(vnonsign))), vreinterpretq_f32_u32(vmagic_bias));
75 const float32x4_t vdenorm_hi = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_high_u16(vnonsign))), vreinterpretq_f32_u32(vmagic_bias));
76
77 const uint16x8_t vmask = vcgtq_u16(vnonsign, vdenorm_cutoff);
78
79 const uint32x4_t vxmask_lo = vreinterpretq_u32_s32(vmovl_s16(vreinterpret_s16_u16(vget_low_u16(vmask))));
80 uint32x4_t vf = vorrq_u32(vshll_n_u16(vget_low_u16(vsign), 16),
81 vreinterpretq_u32_f32(vbslq_f32(vxmask_lo, vnorm_lo, vdenorm_lo)));
82
83 if (n & (4 * sizeof(uint16_t))) {
84 vst1q_f32(output, vreinterpretq_f32_u32(vf)); output += 4;
85
86 const uint32x4_t vxmask_hi = vreinterpretq_u32_s32(vmovl_s16(vreinterpret_s16_u16(vget_high_u16(vmask))));
87 vf = vorrq_u32(vshll_n_u16(vget_high_u16(vsign), 16),
88 vreinterpretq_u32_f32(vbslq_f32(vxmask_hi, vnorm_hi, vdenorm_hi)));
89 }
90 uint32x2_t vf_lo = vget_low_u32(vf);
91 if (n & (2 * sizeof(uint16_t))) {
92 vst1_f32(output, vreinterpret_f32_u32(vf_lo)); output += 2;
93 vf_lo = vget_high_u32(vf);
94 }
95 if (n & (1 * sizeof(uint16_t))) {
96 vst1_lane_f32(output, vreinterpret_f32_u32(vf_lo), 0);
97 }
98 }
99 }
100