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_x24(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_x24(
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 >= 24 * sizeof(uint16_t); n -= 24 * sizeof(uint16_t)) {
37 const uint16x8_t vh0 = vld1q_u16(i); i += 8;
38 const uint16x8_t vh1 = vld1q_u16(i); i += 8;
39 const uint16x8_t vh2 = vld1q_u16(i); i += 8;
40
41 const uint32x4_t vw0 = vshll_n_u16(vget_low_u16(vh0), 16);
42 const uint32x4_t vw1 = vshll_n_u16(vget_high_u16(vh0), 16);
43 const uint32x4_t vw2 = vshll_n_u16(vget_low_u16(vh1), 16);
44 const uint32x4_t vw3 = vshll_n_u16(vget_high_u16(vh1), 16);
45 const uint32x4_t vw4 = vshll_n_u16(vget_low_u16(vh2), 16);
46 const uint32x4_t vw5 = vshll_n_u16(vget_high_u16(vh2), 16);
47
48 const uint32x4_t vsign0 = vandq_u32(vw0, vsign_mask);
49 const uint32x4_t vsign1 = vandq_u32(vw1, vsign_mask);
50 const uint32x4_t vsign2 = vandq_u32(vw2, vsign_mask);
51 const uint32x4_t vsign3 = vandq_u32(vw3, vsign_mask);
52 const uint32x4_t vsign4 = vandq_u32(vw4, vsign_mask);
53 const uint32x4_t vsign5 = vandq_u32(vw5, vsign_mask);
54
55 const uint32x4_t vnonsign0 = veorq_u32(vw0, vsign0);
56 const uint32x4_t vnonsign1 = veorq_u32(vw1, vsign1);
57 const uint32x4_t vnonsign2 = veorq_u32(vw2, vsign2);
58 const uint32x4_t vnonsign3 = veorq_u32(vw3, vsign3);
59 const uint32x4_t vnonsign4 = veorq_u32(vw4, vsign4);
60 const uint32x4_t vnonsign5 = veorq_u32(vw5, vsign5);
61
62 const float32x4_t vnorm0 = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign0, 3)), vexp_scale);
63 const float32x4_t vnorm1 = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign1, 3)), vexp_scale);
64 const float32x4_t vnorm2 = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign2, 3)), vexp_scale);
65 const float32x4_t vnorm3 = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign3, 3)), vexp_scale);
66 const float32x4_t vnorm4 = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign4, 3)), vexp_scale);
67 const float32x4_t vnorm5 = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign5, 3)), vexp_scale);
68
69 const float32x4_t vdenorm0 = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign0, 16)), vreinterpretq_f32_u32(vmagic_bias));
70 const float32x4_t vdenorm1 = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign1, 16)), vreinterpretq_f32_u32(vmagic_bias));
71 const float32x4_t vdenorm2 = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign2, 16)), vreinterpretq_f32_u32(vmagic_bias));
72 const float32x4_t vdenorm3 = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign3, 16)), vreinterpretq_f32_u32(vmagic_bias));
73 const float32x4_t vdenorm4 = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign4, 16)), vreinterpretq_f32_u32(vmagic_bias));
74 const float32x4_t vdenorm5 = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign5, 16)), vreinterpretq_f32_u32(vmagic_bias));
75
76 const uint32x4_t vxmask0 = vcgtq_u32(vnonsign0, vdenorm_cutoff);
77 const uint32x4_t vxmask1 = vcgtq_u32(vnonsign1, vdenorm_cutoff);
78 const uint32x4_t vxmask2 = vcgtq_u32(vnonsign2, vdenorm_cutoff);
79 const uint32x4_t vxmask3 = vcgtq_u32(vnonsign3, vdenorm_cutoff);
80 const uint32x4_t vxmask4 = vcgtq_u32(vnonsign4, vdenorm_cutoff);
81 const uint32x4_t vxmask5 = vcgtq_u32(vnonsign5, vdenorm_cutoff);
82
83 const uint32x4_t vf0 = vorrq_u32(vsign0, vreinterpretq_u32_f32(vbslq_f32(vxmask0, vnorm0, vdenorm0)));
84 const uint32x4_t vf1 = vorrq_u32(vsign1, vreinterpretq_u32_f32(vbslq_f32(vxmask1, vnorm1, vdenorm1)));
85 const uint32x4_t vf2 = vorrq_u32(vsign2, vreinterpretq_u32_f32(vbslq_f32(vxmask2, vnorm2, vdenorm2)));
86 const uint32x4_t vf3 = vorrq_u32(vsign3, vreinterpretq_u32_f32(vbslq_f32(vxmask3, vnorm3, vdenorm3)));
87 const uint32x4_t vf4 = vorrq_u32(vsign4, vreinterpretq_u32_f32(vbslq_f32(vxmask4, vnorm4, vdenorm4)));
88 const uint32x4_t vf5 = vorrq_u32(vsign5, vreinterpretq_u32_f32(vbslq_f32(vxmask5, vnorm5, vdenorm5)));
89
90 vst1q_f32(output, vreinterpretq_f32_u32(vf0)); output += 4;
91 vst1q_f32(output, vreinterpretq_f32_u32(vf1)); output += 4;
92 vst1q_f32(output, vreinterpretq_f32_u32(vf2)); output += 4;
93 vst1q_f32(output, vreinterpretq_f32_u32(vf3)); output += 4;
94 vst1q_f32(output, vreinterpretq_f32_u32(vf4)); output += 4;
95 vst1q_f32(output, vreinterpretq_f32_u32(vf5)); output += 4;
96 }
97 for (; n >= 8 * sizeof(uint16_t); n -= 8 * sizeof(uint16_t)) {
98 const uint16x8_t vh = vld1q_u16(i); i += 8;
99
100 const uint32x4_t vw_lo = vshll_n_u16(vget_low_u16(vh), 16);
101 const uint32x4_t vw_hi = vshll_n_u16(vget_high_u16(vh), 16);
102
103 const uint32x4_t vsign_lo = vandq_u32(vw_lo, vsign_mask);
104 const uint32x4_t vsign_hi = vandq_u32(vw_hi, vsign_mask);
105
106 const uint32x4_t vnonsign_lo = veorq_u32(vw_lo, vsign_lo);
107 const uint32x4_t vnonsign_hi = veorq_u32(vw_hi, vsign_hi);
108
109 const float32x4_t vnorm_lo = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign_lo, 3)), vexp_scale);
110 const float32x4_t vnorm_hi = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign_hi, 3)), vexp_scale);
111
112 const float32x4_t vdenorm_lo = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign_lo, 16)), vreinterpretq_f32_u32(vmagic_bias));
113 const float32x4_t vdenorm_hi = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign_hi, 16)), vreinterpretq_f32_u32(vmagic_bias));
114
115 const uint32x4_t vxmask_lo = vcgtq_u32(vnonsign_lo, vdenorm_cutoff);
116 const uint32x4_t vf_lo = vorrq_u32(vsign_lo, vreinterpretq_u32_f32(vbslq_f32(vxmask_lo, vnorm_lo, vdenorm_lo)));
117
118 const uint32x4_t vxmask_hi = vcgtq_u32(vnonsign_hi, vdenorm_cutoff);
119 const uint32x4_t vf_hi = vorrq_u32(vsign_hi, vreinterpretq_u32_f32(vbslq_f32(vxmask_hi, vnorm_hi, vdenorm_hi)));
120
121 vst1q_f32(output, vreinterpretq_f32_u32(vf_lo)); output += 4;
122 vst1q_f32(output, vreinterpretq_f32_u32(vf_hi)); output += 4;
123 }
124 if XNN_UNPREDICTABLE(n != 0) {
125 const uint16x8_t vh = vld1q_u16(i); i += 8;
126
127 const uint32x4_t vw_lo = vshll_n_u16(vget_low_u16(vh), 16);
128 const uint32x4_t vw_hi = vshll_n_u16(vget_high_u16(vh), 16);
129
130 const uint32x4_t vsign_lo = vandq_u32(vw_lo, vsign_mask);
131 const uint32x4_t vsign_hi = vandq_u32(vw_hi, vsign_mask);
132
133 const uint32x4_t vnonsign_lo = veorq_u32(vw_lo, vsign_lo);
134 const uint32x4_t vnonsign_hi = veorq_u32(vw_hi, vsign_hi);
135
136 const float32x4_t vnorm_lo = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign_lo, 3)), vexp_scale);
137 const float32x4_t vnorm_hi = vmulq_f32(vreinterpretq_f32_u32(vsraq_n_u32(vexp_offset, vnonsign_hi, 3)), vexp_scale);
138
139 const float32x4_t vdenorm_lo = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign_lo, 16)), vreinterpretq_f32_u32(vmagic_bias));
140 const float32x4_t vdenorm_hi = vsubq_f32(vreinterpretq_f32_u32(vsriq_n_u32(vmagic_bias, vnonsign_hi, 16)), vreinterpretq_f32_u32(vmagic_bias));
141
142 const uint32x4_t vxmask_lo = vcgtq_u32(vnonsign_lo, vdenorm_cutoff);
143 uint32x4_t vf = vorrq_u32(vsign_lo, vreinterpretq_u32_f32(vbslq_f32(vxmask_lo, vnorm_lo, vdenorm_lo)));
144
145 if (n & (4 * sizeof(uint16_t))) {
146 vst1q_f32(output, vreinterpretq_f32_u32(vf)); output += 4;
147
148 const uint32x4_t vxmask_hi = vcgtq_u32(vnonsign_hi, vdenorm_cutoff);
149 vf = vorrq_u32(vsign_hi, vreinterpretq_u32_f32(vbslq_f32(vxmask_hi, vnorm_hi, vdenorm_hi)));
150 }
151 uint32x2_t vf_lo = vget_low_u32(vf);
152 if (n & (2 * sizeof(uint16_t))) {
153 vst1_f32(output, vreinterpret_f32_u32(vf_lo)); output += 2;
154 vf_lo = vget_high_u32(vf);
155 }
156 if (n & (1 * sizeof(uint16_t))) {
157 vst1_lane_f32(output, vreinterpret_f32_u32(vf_lo), 0);
158 }
159 }
160 }
161