• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f16-vbinary/vop-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_vdiv_minmax_ukernel__f16c_x16(size_t n,const void * restrict a_ptr,const void * restrict b_ptr,void * restrict y_ptr,const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f16_vdiv_minmax_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_minmax_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   const __m256 vy_min = _mm256_load_ps(params->avx.min);
36   const __m256 vy_max = _mm256_load_ps(params->avx.max);
37 
38   for (; n >= 16 * sizeof(uint16_t); n -= 16 * sizeof(uint16_t)) {
39     const __m256 va01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) a));
40     const __m256 vb01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) b));
41     const __m256 va456789AB = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (a + 8)));
42     const __m256 vb456789AB = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (b + 8)));
43     a += 16;
44     b += 16;
45 
46     __m256 vy01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_div_ps(va01234567, vb01234567), _MM_FROUND_NO_EXC));
47     __m256 vy456789AB = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_div_ps(va456789AB, vb456789AB), _MM_FROUND_NO_EXC));
48 
49 
50     vy01234567 = _mm256_max_ps(vy01234567, vy_min);
51     vy456789AB = _mm256_max_ps(vy456789AB, vy_min);
52 
53     vy01234567 = _mm256_min_ps(vy01234567, vy_max);
54     vy456789AB = _mm256_min_ps(vy456789AB, vy_max);
55 
56     _mm_storeu_si128((__m128i*) y, _mm256_cvtps_ph(vy01234567, _MM_FROUND_NO_EXC));
57     _mm_storeu_si128((__m128i*) (y + 8), _mm256_cvtps_ph(vy456789AB, _MM_FROUND_NO_EXC));
58     y += 16;
59   }
60   for (; n >= 8 * sizeof(uint16_t); n -= 8 * sizeof(uint16_t)) {
61     const __m256 va = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) a));
62     const __m256 vb = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) b));
63     a += 8;
64     b += 8;
65 
66     __m256 vy = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_div_ps(va, vb), _MM_FROUND_NO_EXC));
67 
68     vy = _mm256_max_ps(vy, vy_min);
69     vy = _mm256_min_ps(vy, vy_max);
70 
71     _mm_storeu_si128((__m128i*) y, _mm256_cvtps_ph(vy, _MM_FROUND_NO_EXC));
72     y += 8;
73   }
74   if XNN_UNLIKELY(n != 0) {
75     const __m256 va = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) a));
76     const __m256 vb = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) b));
77 
78     __m256 vy = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_div_ps(va, vb), _MM_FROUND_NO_EXC));
79 
80     vy = _mm256_max_ps(vy, vy_min);
81     vy = _mm256_min_ps(vy, vy_max);
82 
83     __m128i vh = _mm256_cvtps_ph(vy, _MM_FROUND_NO_EXC);
84     if (n & (4 * sizeof(uint16_t))) {
85       _mm_storel_epi64((__m128i*) y, vh);
86       vh = _mm_unpackhi_epi64(vh, vh);
87       y += 4;
88     }
89     if (n & (2 * sizeof(uint16_t))) {
90       *((uint32_t*) y) = (uint32_t) _mm_cvtsi128_si32(vh);
91       vh = _mm_srli_epi64(vh, 32);
92       y += 2;
93     }
94     if (n & (1 * sizeof(uint16_t))) {
95       *y = (uint16_t) _mm_extract_epi16(vh, 0);
96     }
97   }
98 }
99