xref: /aosp_15_r20/external/XNNPACK/src/f16-f32-vcvt/gen/vcvt-neonfp16-x16.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f16-f32-vcvt/neonfp16.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__neonfp16_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__neonfp16_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 uint16_t* i = (const uint16_t*) input;
30   for (; n >= 16 * sizeof(uint16_t); n -= 16 * sizeof(uint16_t)) {
31     const float16x8_t vh0 = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
32     const float16x8_t vh1 = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
33 
34     const float32x4_t vf0 = vcvt_f32_f16(vget_low_f16(vh0));
35     const float32x4_t vf1 = vcvt_f32_f16(vget_high_f16(vh0));
36     const float32x4_t vf2 = vcvt_f32_f16(vget_low_f16(vh1));
37     const float32x4_t vf3 = vcvt_f32_f16(vget_high_f16(vh1));
38 
39     vst1q_f32(output, vf0); output += 4;
40     vst1q_f32(output, vf1); output += 4;
41     vst1q_f32(output, vf2); output += 4;
42     vst1q_f32(output, vf3); output += 4;
43   }
44   for (; n >= 8 * sizeof(uint16_t); n -= 8 * sizeof(uint16_t)) {
45     const float16x8_t vh = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
46 
47     const float32x4_t vf_lo = vcvt_f32_f16(vget_low_f16(vh));
48     const float32x4_t vf_hi = vcvt_f32_f16(vget_high_f16(vh));
49 
50     vst1q_f32(output, vf_lo); output += 4;
51     vst1q_f32(output, vf_hi); output += 4;
52   }
53   if XNN_UNLIKELY(n != 0) {
54     assert(n >= 1 * sizeof(uint16_t));
55     assert(n <= 7 * sizeof(uint16_t));
56     const float16x8_t vh = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
57 
58     float32x4_t vf = vcvt_f32_f16(vget_low_f16(vh));
59     if (n & (4 * sizeof(uint16_t))) {
60       vst1q_f32(output, vf); output += 4;
61       vf = vcvt_f32_f16(vget_high_f16(vh));
62     }
63     float32x2_t vf_lo = vget_low_f32(vf);
64     if (n & (2 * sizeof(uint16_t))) {
65       vst1_f32(output, vf_lo); output += 2;
66       vf_lo = vget_high_f32(vf);
67     }
68     if (n & (1 * sizeof(uint16_t))) {
69       vst1_lane_f32(output, vf_lo, 0);
70     }
71   }
72 }
73