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_c16(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_c16(
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 >= 16 * sizeof(uint16_t); c -= 16 * sizeof(uint16_t)) {
47 const __m256 vtl456789AB = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
48 const __m256 vtr456789AB = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
49 const __m256 vbl456789AB = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
50 const __m256 vbr456789AB = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3));
51 const __m256 vtlCDEFGHIJ = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i0 + 8)));
52 const __m256 vtrCDEFGHIJ = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i1 + 8)));
53 const __m256 vblCDEFGHIJ = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i2 + 8)));
54 const __m256 vbrCDEFGHIJ = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i3 + 8)));
55 i0 += 16;
56 i1 += 16;
57 i2 += 16;
58 i3 += 16;
59
60 const __m256 vtd456789AB = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_sub_ps(vtr456789AB, vtl456789AB), _MM_FROUND_NO_EXC));
61 const __m256 vbd456789AB = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_sub_ps(vbr456789AB, vbl456789AB), _MM_FROUND_NO_EXC));
62 const __m256 vtdCDEFGHIJ = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_sub_ps(vtrCDEFGHIJ, vtlCDEFGHIJ), _MM_FROUND_NO_EXC));
63 const __m256 vbdCDEFGHIJ = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_sub_ps(vbrCDEFGHIJ, vblCDEFGHIJ), _MM_FROUND_NO_EXC));
64
65 const __m256 vt456789AB = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vtd456789AB, valphah, vtl456789AB), _MM_FROUND_NO_EXC));
66 const __m256 vb456789AB = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vbd456789AB, valphah, vbl456789AB), _MM_FROUND_NO_EXC));
67 const __m256 vtCDEFGHIJ = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vtdCDEFGHIJ, valphah, vtlCDEFGHIJ), _MM_FROUND_NO_EXC));
68 const __m256 vbCDEFGHIJ = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vbdCDEFGHIJ, valphah, vblCDEFGHIJ), _MM_FROUND_NO_EXC));
69
70 const __m256 vd456789AB = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_sub_ps(vb456789AB, vt456789AB), _MM_FROUND_NO_EXC));
71 const __m256 vdCDEFGHIJ = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_sub_ps(vbCDEFGHIJ, vtCDEFGHIJ), _MM_FROUND_NO_EXC));
72
73 const __m128i vo456789AB = _mm256_cvtps_ph(_mm256_fmadd_ps(vd456789AB, valphav, vt456789AB), _MM_FROUND_NO_EXC);
74 const __m128i voCDEFGHIJ = _mm256_cvtps_ph(_mm256_fmadd_ps(vdCDEFGHIJ, valphav, vtCDEFGHIJ), _MM_FROUND_NO_EXC);
75
76 _mm_storeu_si128((__m128i*) o, vo456789AB);
77 _mm_storeu_si128((__m128i*) (o + 8), voCDEFGHIJ);
78 o += 16;
79 }
80 for (; c >= 8 * sizeof(uint16_t); c -= 8 * sizeof(uint16_t)) {
81 const __m256 vtl = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
82 i0 += 8;
83 const __m256 vtr = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
84 i1 += 8;
85 const __m256 vbl = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
86 i2 += 8;
87 const __m256 vbr = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3));
88 i3 += 8;
89
90 const __m256 vtd = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_sub_ps(vtr, vtl), _MM_FROUND_NO_EXC));
91 const __m256 vbd = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_sub_ps(vbr, vbl), _MM_FROUND_NO_EXC));
92
93 const __m256 vt = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vtd, valphah, vtl), _MM_FROUND_NO_EXC));
94 const __m256 vb = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vbd, valphah, vbl), _MM_FROUND_NO_EXC));
95
96 const __m256 vd = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_sub_ps(vb, vt), _MM_FROUND_NO_EXC));
97
98 const __m128i vo = _mm256_cvtps_ph(_mm256_fmadd_ps(vd, valphav, vt), _MM_FROUND_NO_EXC);
99
100 _mm_storeu_si128((__m128i*) o, vo);
101 o += 8;
102 }
103 if XNN_UNLIKELY(c != 0) {
104 const __m256 vtl = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0));
105 i0 += 8;
106 const __m256 vtr = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1));
107 i1 += 8;
108 const __m256 vbl = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2));
109 i2 += 8;
110 const __m256 vbr = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i3));
111 i3 += 8;
112
113 const __m256 vtd = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_sub_ps(vtr, vtl), _MM_FROUND_NO_EXC));
114 const __m256 vbd = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_sub_ps(vbr, vbl), _MM_FROUND_NO_EXC));
115
116 const __m256 vt = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vtd, valphah, vtl), _MM_FROUND_NO_EXC));
117 const __m256 vb = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vbd, valphah, vbl), _MM_FROUND_NO_EXC));
118
119 const __m256 vd = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_sub_ps(vb, vt), _MM_FROUND_NO_EXC));
120
121 __m128i vo = _mm256_cvtps_ph(_mm256_fmadd_ps(vd, valphav, vt), _MM_FROUND_NO_EXC);
122 if (c & (4 * sizeof(uint16_t))) {
123 _mm_storel_epi64((__m128i*) o, vo);
124 vo = _mm_unpackhi_epi64(vo, vo);
125 o += 4;
126 }
127 if (c & (2 * sizeof(uint16_t))) {
128 _mm_storeu_si32(o, vo);
129 vo = _mm_srli_epi64(vo, 32);
130 o += 2;
131 }
132 if (c & (1 * sizeof(uint16_t))) {
133 *o = (uint16_t) _mm_extract_epi16(vo, 0);
134 o += 1;
135 }
136 }
137
138 o = (uint16_t*) ((uintptr_t) o + output_increment);
139 } while (--output_pixels != 0);
140 }
141