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_x16(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_x16(
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 >= 16 * sizeof(uint16_t); n -= 16 * sizeof(uint16_t)) {
37 const uint16x8_t vh0 = vld1q_u16(i); i += 8;
38 const uint16x8_t vh1 = vld1q_u16(i); i += 8;
39
40 const uint16x8_t vsign0 = vandq_u16(vh0, vsign_mask);
41 const uint16x8_t vsign1 = vandq_u16(vh1, vsign_mask);
42
43 const uint16x8_t vnonsign0 = veorq_u16(vh0, vsign0);
44 const uint16x8_t vnonsign1 = veorq_u16(vh1, vsign1);
45
46 const uint16x8x2_t vprenorm0 = vzipq_u16(vshlq_n_u16(vnonsign0, 13), vsraq_n_u16(vexp_offset, vnonsign0, 3));
47 const uint16x8x2_t vprenorm1 = vzipq_u16(vshlq_n_u16(vnonsign1, 13), vsraq_n_u16(vexp_offset, vnonsign1, 3));
48
49 const float32x4_t vnorm0 = vmulq_f32(vreinterpretq_f32_u16(vprenorm0.val[0]), vexp_scale);
50 const float32x4_t vnorm1 = vmulq_f32(vreinterpretq_f32_u16(vprenorm0.val[1]), vexp_scale);
51 const float32x4_t vnorm2 = vmulq_f32(vreinterpretq_f32_u16(vprenorm1.val[0]), vexp_scale);
52 const float32x4_t vnorm3 = vmulq_f32(vreinterpretq_f32_u16(vprenorm1.val[1]), vexp_scale);
53
54 const float32x4_t vdenorm0 = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_low_u16(vnonsign0))), vreinterpretq_f32_u32(vmagic_bias));
55 const float32x4_t vdenorm1 = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_high_u16(vnonsign0))), vreinterpretq_f32_u32(vmagic_bias));
56 const float32x4_t vdenorm2 = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_low_u16(vnonsign1))), vreinterpretq_f32_u32(vmagic_bias));
57 const float32x4_t vdenorm3 = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_high_u16(vnonsign1))), vreinterpretq_f32_u32(vmagic_bias));
58
59 const uint16x8_t vmask0 = vcgtq_u16(vnonsign0, vdenorm_cutoff);
60 const uint16x8_t vmask1 = vcgtq_u16(vnonsign1, vdenorm_cutoff);
61
62 const uint32x4_t vxmask0 = vreinterpretq_u32_s32(vmovl_s16(vreinterpret_s16_u16(vget_low_u16(vmask0))));
63 const uint32x4_t vf0 = vorrq_u32(vshll_n_u16(vget_low_u16(vsign0), 16),
64 vreinterpretq_u32_f32(vbslq_f32(vxmask0, vnorm0, vdenorm0)));
65 const uint32x4_t vxmask2 = vreinterpretq_u32_s32(vmovl_s16(vreinterpret_s16_u16(vget_low_u16(vmask1))));
66 const uint32x4_t vf2 = vorrq_u32(vshll_n_u16(vget_low_u16(vsign1), 16),
67 vreinterpretq_u32_f32(vbslq_f32(vxmask2, vnorm2, vdenorm2)));
68
69 const uint32x4_t vxmask1 = vreinterpretq_u32_s32(vmovl_s16(vreinterpret_s16_u16(vget_high_u16(vmask0))));
70 const uint32x4_t vf1 = vorrq_u32(vshll_n_u16(vget_high_u16(vsign0), 16),
71 vreinterpretq_u32_f32(vbslq_f32(vxmask1, vnorm1, vdenorm1)));
72 const uint32x4_t vxmask3 = vreinterpretq_u32_s32(vmovl_s16(vreinterpret_s16_u16(vget_high_u16(vmask1))));
73 const uint32x4_t vf3 = vorrq_u32(vshll_n_u16(vget_high_u16(vsign1), 16),
74 vreinterpretq_u32_f32(vbslq_f32(vxmask3, vnorm3, vdenorm3)));
75
76 vst1q_f32(output, vreinterpretq_f32_u32(vf0)); output += 4;
77 vst1q_f32(output, vreinterpretq_f32_u32(vf1)); output += 4;
78 vst1q_f32(output, vreinterpretq_f32_u32(vf2)); output += 4;
79 vst1q_f32(output, vreinterpretq_f32_u32(vf3)); output += 4;
80 }
81 for (; n >= 8 * sizeof(uint16_t); n -= 8 * sizeof(uint16_t)) {
82 const uint16x8_t vh = vld1q_u16(i); i += 8;
83
84 const uint16x8_t vsign = vandq_u16(vh, vsign_mask);
85
86 const uint16x8_t vnonsign = veorq_u16(vh, vsign);
87
88 const uint16x8x2_t vprenorm = vzipq_u16(vshlq_n_u16(vnonsign, 13), vsraq_n_u16(vexp_offset, vnonsign, 3));
89 const float32x4_t vnorm_lo = vmulq_f32(vreinterpretq_f32_u16(vprenorm.val[0]), vexp_scale);
90 const float32x4_t vnorm_hi = vmulq_f32(vreinterpretq_f32_u16(vprenorm.val[1]), vexp_scale);
91
92 const float32x4_t vdenorm_lo = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_low_u16(vnonsign))), vreinterpretq_f32_u32(vmagic_bias));
93 const float32x4_t vdenorm_hi = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_high_u16(vnonsign))), vreinterpretq_f32_u32(vmagic_bias));
94
95 const uint16x8_t vmask = vcgtq_u16(vnonsign, vdenorm_cutoff);
96
97 const uint32x4_t vxmask_lo = vreinterpretq_u32_s32(vmovl_s16(vreinterpret_s16_u16(vget_low_u16(vmask))));
98 const uint32x4_t vf_lo = vorrq_u32(vshll_n_u16(vget_low_u16(vsign), 16),
99 vreinterpretq_u32_f32(vbslq_f32(vxmask_lo, vnorm_lo, vdenorm_lo)));
100
101 const uint32x4_t vxmask_hi = vreinterpretq_u32_s32(vmovl_s16(vreinterpret_s16_u16(vget_high_u16(vmask))));
102 const uint32x4_t vf_hi = vorrq_u32(vshll_n_u16(vget_high_u16(vsign), 16),
103 vreinterpretq_u32_f32(vbslq_f32(vxmask_hi, vnorm_hi, vdenorm_hi)));
104
105 vst1q_f32(output, vreinterpretq_f32_u32(vf_lo)); output += 4;
106 vst1q_f32(output, vreinterpretq_f32_u32(vf_hi)); output += 4;
107 }
108 if XNN_UNPREDICTABLE(n != 0) {
109 const uint16x8_t vh = vld1q_u16(i); i += 8;
110
111 const uint16x8_t vsign = vandq_u16(vh, vsign_mask);
112
113 const uint16x8_t vnonsign = veorq_u16(vh, vsign);
114
115 const uint16x8x2_t vprenorm = vzipq_u16(vshlq_n_u16(vnonsign, 13), vsraq_n_u16(vexp_offset, vnonsign, 3));
116 const float32x4_t vnorm_lo = vmulq_f32(vreinterpretq_f32_u16(vprenorm.val[0]), vexp_scale);
117 const float32x4_t vnorm_hi = vmulq_f32(vreinterpretq_f32_u16(vprenorm.val[1]), vexp_scale);
118
119 const float32x4_t vdenorm_lo = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_low_u16(vnonsign))), vreinterpretq_f32_u32(vmagic_bias));
120 const float32x4_t vdenorm_hi = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_high_u16(vnonsign))), vreinterpretq_f32_u32(vmagic_bias));
121
122 const uint16x8_t vmask = vcgtq_u16(vnonsign, vdenorm_cutoff);
123
124 const uint32x4_t vxmask_lo = vreinterpretq_u32_s32(vmovl_s16(vreinterpret_s16_u16(vget_low_u16(vmask))));
125 uint32x4_t vf = vorrq_u32(vshll_n_u16(vget_low_u16(vsign), 16),
126 vreinterpretq_u32_f32(vbslq_f32(vxmask_lo, vnorm_lo, vdenorm_lo)));
127
128 if (n & (4 * sizeof(uint16_t))) {
129 vst1q_f32(output, vreinterpretq_f32_u32(vf)); output += 4;
130
131 const uint32x4_t vxmask_hi = vreinterpretq_u32_s32(vmovl_s16(vreinterpret_s16_u16(vget_high_u16(vmask))));
132 vf = vorrq_u32(vshll_n_u16(vget_high_u16(vsign), 16),
133 vreinterpretq_u32_f32(vbslq_f32(vxmask_hi, vnorm_hi, vdenorm_hi)));
134 }
135 uint32x2_t vf_lo = vget_low_u32(vf);
136 if (n & (2 * sizeof(uint16_t))) {
137 vst1_f32(output, vreinterpret_f32_u32(vf_lo)); output += 2;
138 vf_lo = vget_high_u32(vf);
139 }
140 if (n & (1 * sizeof(uint16_t))) {
141 vst1_lane_f32(output, vreinterpret_f32_u32(vf_lo), 0);
142 }
143 }
144 }
145