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_vsqrdiff_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_vsqrdiff_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_sub_ps(va01234567, vb01234567), _MM_FROUND_NO_EXC));
46 __m256 vy456789AB = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_sub_ps(va456789AB, vb456789AB), _MM_FROUND_NO_EXC));
47
48 vy01234567 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vy01234567, vy01234567), _MM_FROUND_NO_EXC));
49 vy456789AB = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vy456789AB, vy456789AB), _MM_FROUND_NO_EXC));
50
51
52 _mm_storeu_si128((__m128i*) y, _mm256_cvtps_ph(vy01234567, _MM_FROUND_NO_EXC));
53 _mm_storeu_si128((__m128i*) (y + 8), _mm256_cvtps_ph(vy456789AB, _MM_FROUND_NO_EXC));
54 y += 16;
55 }
56 for (; n >= 8 * sizeof(uint16_t); n -= 8 * sizeof(uint16_t)) {
57 const __m256 va = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) a));
58 const __m256 vb = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) b));
59 a += 8;
60 b += 8;
61
62 __m256 vy = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_sub_ps(va, vb), _MM_FROUND_NO_EXC));
63 vy = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vy, vy), _MM_FROUND_NO_EXC));
64
65
66 _mm_storeu_si128((__m128i*) y, _mm256_cvtps_ph(vy, _MM_FROUND_NO_EXC));
67 y += 8;
68 }
69 if XNN_UNLIKELY(n != 0) {
70 const __m256 va = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) a));
71 const __m256 vb = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) b));
72
73 __m256 vy = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_sub_ps(va, vb), _MM_FROUND_NO_EXC));
74 vy = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vy, vy), _MM_FROUND_NO_EXC));
75
76
77 __m128i vh = _mm256_cvtps_ph(vy, _MM_FROUND_NO_EXC);
78 if (n & (4 * sizeof(uint16_t))) {
79 _mm_storel_epi64((__m128i*) y, vh);
80 vh = _mm_unpackhi_epi64(vh, vh);
81 y += 4;
82 }
83 if (n & (2 * sizeof(uint16_t))) {
84 _mm_storeu_si32(y, vh);
85 vh = _mm_srli_epi64(vh, 32);
86 y += 2;
87 }
88 if (n & (1 * sizeof(uint16_t))) {
89 *y = (uint16_t) _mm_extract_epi16(vh, 0);
90 }
91 }
92 }
93