1 // Auto-generated file. Do not edit!
2 // Template: src/f16-ibilinear/fma3.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/ibilinear.h>
16 #include <xnnpack/intrinsics-polyfill.h>
17
18
xnn_f16_ibilinear_ukernel__fma3_c8(size_t output_pixels,size_t channels,const void ** restrict input,size_t input_offset,const void * restrict weights,void * restrict output,size_t output_increment)19 void xnn_f16_ibilinear_ukernel__fma3_c8(
20 size_t output_pixels,
21 size_t channels,
22 const void**restrict input,
23 size_t input_offset,
24 const void*restrict weights,
25 void*restrict output,
26 size_t output_increment) XNN_OOB_READS
27 {
28 assert(output_pixels != 0);
29 assert(channels != 0);
30 assert(channels % sizeof(uint16_t) == 0);
31
32 uint16_t* o = (uint16_t*) output;
33 do {
34 const uint16_t* i0 = (const uint16_t*) ((uintptr_t) input[0] + input_offset);
35 const uint16_t* i1 = (const uint16_t*) ((uintptr_t) input[1] + input_offset);
36 const uint16_t* i2 = (const uint16_t*) ((uintptr_t) input[2] + input_offset);
37 const uint16_t* i3 = (const uint16_t*) ((uintptr_t) input[3] + input_offset);
38 input += 4;
39
40 const __m256 valphahv = _mm256_cvtph_ps(_mm_castps_si128(_mm_broadcast_ss(weights)));
41 const __m256 valphah = _mm256_permute_ps(valphahv, _MM_SHUFFLE(2, 0, 2, 0));
42 const __m256 valphav = _mm256_permute_ps(valphahv, _MM_SHUFFLE(3, 1, 3, 1));
43 weights = (const uint16_t*) weights + 2;
44
45 size_t c = channels;
46 for (; c >= 8 * sizeof(uint16_t); c -= 8 * sizeof(uint16_t)) {
47 const __m256 vtl = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
48 i0 += 8;
49 const __m256 vtr = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
50 i1 += 8;
51 const __m256 vbl = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
52 i2 += 8;
53 const __m256 vbr = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3));
54 i3 += 8;
55
56 const __m256 vtd = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_sub_ps(vtr, vtl), _MM_FROUND_NO_EXC));
57 const __m256 vbd = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_sub_ps(vbr, vbl), _MM_FROUND_NO_EXC));
58
59 const __m256 vt = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vtd, valphah, vtl), _MM_FROUND_NO_EXC));
60 const __m256 vb = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vbd, valphah, vbl), _MM_FROUND_NO_EXC));
61
62 const __m256 vd = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_sub_ps(vb, vt), _MM_FROUND_NO_EXC));
63
64 const __m128i vo = _mm256_cvtps_ph(_mm256_fmadd_ps(vd, valphav, vt), _MM_FROUND_NO_EXC);
65
66 _mm_storeu_si128((__m128i*) o, vo);
67 o += 8;
68 }
69 if XNN_UNLIKELY(c != 0) {
70 const __m256 vtl = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
71 i0 += 8;
72 const __m256 vtr = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
73 i1 += 8;
74 const __m256 vbl = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
75 i2 += 8;
76 const __m256 vbr = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3));
77 i3 += 8;
78
79 const __m256 vtd = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_sub_ps(vtr, vtl), _MM_FROUND_NO_EXC));
80 const __m256 vbd = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_sub_ps(vbr, vbl), _MM_FROUND_NO_EXC));
81
82 const __m256 vt = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vtd, valphah, vtl), _MM_FROUND_NO_EXC));
83 const __m256 vb = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vbd, valphah, vbl), _MM_FROUND_NO_EXC));
84
85 const __m256 vd = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_sub_ps(vb, vt), _MM_FROUND_NO_EXC));
86
87 __m128i vo = _mm256_cvtps_ph(_mm256_fmadd_ps(vd, valphav, vt), _MM_FROUND_NO_EXC);
88 if (c & (4 * sizeof(uint16_t))) {
89 _mm_storel_epi64((__m128i*) o, vo);
90 vo = _mm_unpackhi_epi64(vo, vo);
91 o += 4;
92 }
93 if (c & (2 * sizeof(uint16_t))) {
94 _mm_storeu_si32(o, vo);
95 vo = _mm_srli_epi64(vo, 32);
96 o += 2;
97 }
98 if (c & (1 * sizeof(uint16_t))) {
99 *o = (uint16_t) _mm_extract_epi16(vo, 0);
100 o += 1;
101 }
102 }
103
104 o = (uint16_t*) ((uintptr_t) o + output_increment);
105 } while (--output_pixels != 0);
106 }
107