• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f16-vbinary/vopc-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/vbinary.h>
16 
17 
xnn_f16_vmaxc_ukernel__f16c_x16(size_t n,const void * restrict a_ptr,const void * restrict b_ptr,void * restrict y_ptr,const union xnn_f16_default_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f16_vmaxc_ukernel__f16c_x16(
19     size_t n,
20     const void* restrict a_ptr,
21     const void* restrict b_ptr,
22     void* restrict y_ptr,
23     const union xnn_f16_default_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24 {
25   assert(n != 0);
26   assert(n % sizeof(uint16_t) == 0);
27   assert(a_ptr != NULL);
28   assert(b_ptr != NULL);
29   assert(y_ptr != NULL);
30 
31   const uint16_t* a = (const uint16_t*) a_ptr;
32   const uint16_t* b = (const uint16_t*) b_ptr;
33   uint16_t* y = (uint16_t*) y_ptr;
34 
35 
36   const __m256 vb = _mm256_cvtph_ps(_mm_set1_epi16((short) *b));
37   for (; n >= 16 * sizeof(uint16_t); n -= 16 * sizeof(uint16_t)) {
38     const __m256 va01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) a));
39     const __m256 va456789AB = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (a + 8)));
40     a += 16;
41 
42     __m256 vy01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_max_ps(va01234567, vb), _MM_FROUND_NO_EXC));
43     __m256 vy456789AB = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_max_ps(va456789AB, vb), _MM_FROUND_NO_EXC));
44 
45 
46 
47     _mm_storeu_si128((__m128i*) y, _mm256_cvtps_ph(vy01234567, _MM_FROUND_NO_EXC));
48     _mm_storeu_si128((__m128i*) (y + 8), _mm256_cvtps_ph(vy456789AB, _MM_FROUND_NO_EXC));
49     y += 16;
50   }
51   for (; n >= 8 * sizeof(uint16_t); n -= 8 * sizeof(uint16_t)) {
52     const __m256 va = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) a));
53     a += 8;
54 
55     __m256 vy = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_max_ps(va, vb), _MM_FROUND_NO_EXC));
56 
57 
58     _mm_storeu_si128((__m128i*) y, _mm256_cvtps_ph(vy, _MM_FROUND_NO_EXC));
59     y += 8;
60   }
61   if XNN_UNLIKELY(n != 0) {
62     const __m256 va = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) a));
63 
64     __m256 vy = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_max_ps(va, vb), _MM_FROUND_NO_EXC));
65 
66 
67     __m128i vh = _mm256_cvtps_ph(vy, _MM_FROUND_NO_EXC);
68     if (n & (4 * sizeof(uint16_t))) {
69       _mm_storel_epi64((__m128i*) y, vh);
70       vh = _mm_unpackhi_epi64(vh, vh);
71       y += 4;
72     }
73     if (n & (2 * sizeof(uint16_t))) {
74       *((uint32_t*) y) = (uint32_t) _mm_cvtsi128_si32(vh);
75       vh = _mm_srli_epi64(vh, 32);
76       y += 2;
77     }
78     if (n & (1 * sizeof(uint16_t))) {
79       *y = (uint16_t) _mm_extract_epi16(vh, 0);
80     }
81   }
82 }
83