xref: /aosp_15_r20/external/XNNPACK/src/f16-vrnd/gen/vrndu-f16c-x16.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f16-vrnd/f16c.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2022 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 <immintrin.h>
13 
14 #include <xnnpack/common.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16 #include <xnnpack/math.h>
17 #include <xnnpack/vunary.h>
18 
19 
xnn_f16_vrndu_ukernel__f16c_x16(size_t n,const void * input,void * output,const union xnn_f16_rnd_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_f16_vrndu_ukernel__f16c_x16(
21     size_t n,
22     const void* input,
23     void* output,
24     const union xnn_f16_rnd_params params[restrict XNN_MIN_ELEMENTS(1)])
25 {
26   assert(n != 0);
27   assert(n % sizeof(uint16_t) == 0);
28 
29   const uint16_t* i = (const uint16_t*) input;
30   uint16_t* o = (uint16_t*) output;
31   for (; n >= 16 * sizeof(uint16_t); n -= 16 * sizeof(uint16_t)) {
32     __m256 vacc0 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
33     __m256 vacc1 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i + 8)));
34     i += 16;
35 
36     vacc0 = _mm256_round_ps(vacc0, _MM_FROUND_TO_POS_INF | _MM_FROUND_NO_EXC);
37     vacc1 = _mm256_round_ps(vacc1, _MM_FROUND_TO_POS_INF | _MM_FROUND_NO_EXC);
38 
39     _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vacc0, _MM_FROUND_NO_EXC));
40     _mm_storeu_si128((__m128i*) (o + 8), _mm256_cvtps_ph(vacc1, _MM_FROUND_NO_EXC));
41     o += 16;
42   }
43   for (; n >= 8 * sizeof(uint16_t); n -= 8 * sizeof(uint16_t)) {
44     __m256 vacc = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
45     i += 8;
46 
47     vacc = _mm256_round_ps(vacc, _MM_FROUND_TO_POS_INF | _MM_FROUND_NO_EXC);
48 
49     _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vacc, _MM_FROUND_NO_EXC));
50     o += 8;
51   }
52   if XNN_UNLIKELY(n != 0) {
53     assert(n >= 1 * sizeof(uint16_t));
54     assert(n <= 7 * sizeof(uint16_t));
55     __m256 vacc = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
56     vacc = _mm256_round_ps(vacc, _MM_FROUND_TO_POS_INF | _MM_FROUND_NO_EXC);
57     __m128i vh = _mm256_cvtps_ph(vacc, _MM_FROUND_NO_EXC);
58     if (n & (4 * sizeof(uint16_t))) {
59       _mm_storel_epi64((__m128i*) o, vh);
60       vh = _mm_unpackhi_epi64(vh, vh);
61       o += 4;
62     }
63     if (n & (2 * sizeof(uint16_t))) {
64       _mm_storeu_si32(o, vh);
65       vh = _mm_srli_epi64(vh, 32);
66       o += 2;
67     }
68     if (n & (1 * sizeof(uint16_t))) {
69       *o = (uint16_t) _mm_extract_epi16(vh, 0);
70     }
71   }
72 }
73