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/intrinsics-polyfill.h>
16 #include <xnnpack/vbinary.h>
17
18
xnn_f16_vmin_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)])19 void xnn_f16_vmin_ukernel__f16c_x16(
20 size_t n,
21 const void* restrict a_ptr,
22 const void* restrict b_ptr,
23 void* restrict y_ptr,
24 const union xnn_f16_default_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
25 {
26 assert(n != 0);
27 assert(n % sizeof(uint16_t) == 0);
28 assert(a_ptr != NULL);
29 assert(b_ptr != NULL);
30 assert(y_ptr != NULL);
31
32 const uint16_t* a = (const uint16_t*) a_ptr;
33 const uint16_t* b = (const uint16_t*) b_ptr;
34 uint16_t* y = (uint16_t*) y_ptr;
35
36
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 vb01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) b));
40 const __m256 va456789AB = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (a + 8)));
41 const __m256 vb456789AB = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (b + 8)));
42 a += 16;
43 b += 16;
44
45 __m256 vy01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_min_ps(va01234567, vb01234567), _MM_FROUND_NO_EXC));
46 __m256 vy456789AB = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_min_ps(va456789AB, vb456789AB), _MM_FROUND_NO_EXC));
47
48
49
50 _mm_storeu_si128((__m128i*) y, _mm256_cvtps_ph(vy01234567, _MM_FROUND_NO_EXC));
51 _mm_storeu_si128((__m128i*) (y + 8), _mm256_cvtps_ph(vy456789AB, _MM_FROUND_NO_EXC));
52 y += 16;
53 }
54 for (; n >= 8 * sizeof(uint16_t); n -= 8 * sizeof(uint16_t)) {
55 const __m256 va = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) a));
56 const __m256 vb = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) b));
57 a += 8;
58 b += 8;
59
60 __m256 vy = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_min_ps(va, vb), _MM_FROUND_NO_EXC));
61
62
63 _mm_storeu_si128((__m128i*) y, _mm256_cvtps_ph(vy, _MM_FROUND_NO_EXC));
64 y += 8;
65 }
66 if XNN_UNLIKELY(n != 0) {
67 const __m256 va = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) a));
68 const __m256 vb = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) b));
69
70 __m256 vy = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_min_ps(va, vb), _MM_FROUND_NO_EXC));
71
72
73 __m128i vh = _mm256_cvtps_ph(vy, _MM_FROUND_NO_EXC);
74 if (n & (4 * sizeof(uint16_t))) {
75 _mm_storel_epi64((__m128i*) y, vh);
76 vh = _mm_unpackhi_epi64(vh, vh);
77 y += 4;
78 }
79 if (n & (2 * sizeof(uint16_t))) {
80 _mm_storeu_si32(y, vh);
81 vh = _mm_srli_epi64(vh, 32);
82 y += 2;
83 }
84 if (n & (1 * sizeof(uint16_t))) {
85 *y = (uint16_t) _mm_extract_epi16(vh, 0);
86 }
87 }
88 }
89