1 // Auto-generated file. Do not edit!
2 // Template: src/f16-f32-vcvt/neon-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 <arm_neon.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/vcvt.h>
16
17
xnn_f16_f32_vcvt_ukernel__neon_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__neon_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 uint32x4_t vsign_mask = vmovq_n_u32(0x80000000);
30 const uint32x4_t vexp_offset = vmovq_n_u32(0x70000000);
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 uint32x4_t vdenorm_cutoff = vmovq_n_u32(0x04000000);
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 uint32x4_t vw_lo = vshll_n_u16(vget_low_u16(vh), 16);
40 const uint32x4_t vw_hi = vshll_n_u16(vget_high_u16(vh), 16);
41
42 const uint32x4_t vsign_lo = vandq_u32(vw_lo, vsign_mask);
43 const uint32x4_t vsign_hi = vandq_u32(vw_hi, vsign_mask);
44
45 const uint32x4_t vnonsign_lo = veorq_u32(vw_lo, vsign_lo);
46 const uint32x4_t vnonsign_hi = veorq_u32(vw_hi, vsign_hi);
47
48 const float32x4_t vnorm_lo = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign_lo, 3)), vexp_scale);
49 const float32x4_t vnorm_hi = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign_hi, 3)), vexp_scale);
50
51 const float32x4_t vdenorm_lo = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign_lo, 16)), vreinterpretq_f32_u32(vmagic_bias));
52 const float32x4_t vdenorm_hi = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign_hi, 16)), vreinterpretq_f32_u32(vmagic_bias));
53
54 const uint32x4_t vxmask_lo = vcgtq_u32(vnonsign_lo, vdenorm_cutoff);
55 const uint32x4_t vf_lo = vorrq_u32(vsign_lo, vreinterpretq_u32_f32(vbslq_f32(vxmask_lo, vnorm_lo, vdenorm_lo)));
56
57 const uint32x4_t vxmask_hi = vcgtq_u32(vnonsign_hi, vdenorm_cutoff);
58 const uint32x4_t vf_hi = vorrq_u32(vsign_hi, 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 uint32x4_t vw_lo = vshll_n_u16(vget_low_u16(vh), 16);
67 const uint32x4_t vw_hi = vshll_n_u16(vget_high_u16(vh), 16);
68
69 const uint32x4_t vsign_lo = vandq_u32(vw_lo, vsign_mask);
70 const uint32x4_t vsign_hi = vandq_u32(vw_hi, vsign_mask);
71
72 const uint32x4_t vnonsign_lo = veorq_u32(vw_lo, vsign_lo);
73 const uint32x4_t vnonsign_hi = veorq_u32(vw_hi, vsign_hi);
74
75 const float32x4_t vnorm_lo = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign_lo, 3)), vexp_scale);
76 const float32x4_t vnorm_hi = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign_hi, 3)), vexp_scale);
77
78 const float32x4_t vdenorm_lo = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign_lo, 16)), vreinterpretq_f32_u32(vmagic_bias));
79 const float32x4_t vdenorm_hi = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign_hi, 16)), vreinterpretq_f32_u32(vmagic_bias));
80
81 const uint32x4_t vxmask_lo = vcgtq_u32(vnonsign_lo, vdenorm_cutoff);
82 uint32x4_t vf = vorrq_u32(vsign_lo, vreinterpretq_u32_f32(vbslq_f32(vxmask_lo, vnorm_lo, vdenorm_lo)));
83
84 if (n & (4 * sizeof(uint16_t))) {
85 vst1q_f32(output, vreinterpretq_f32_u32(vf)); output += 4;
86
87 const uint32x4_t vxmask_hi = vcgtq_u32(vnonsign_hi, vdenorm_cutoff);
88 vf = vorrq_u32(vsign_hi, 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