1 // Auto-generated file. Do not edit!
2 // Template: src/s8-ibilinear/sse.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 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 <smmintrin.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/ibilinear.h>
16
17
xnn_u8_ibilinear_ukernel__sse41_c8(size_t output_pixels,size_t channels,const uint8_t ** restrict input,size_t input_offset,const int16_t * restrict weights,uint8_t * restrict output,size_t output_increment)18 void xnn_u8_ibilinear_ukernel__sse41_c8(
19 size_t output_pixels,
20 size_t channels,
21 const uint8_t**restrict input,
22 size_t input_offset,
23 const int16_t*restrict weights,
24 uint8_t*restrict output,
25 size_t output_increment) XNN_OOB_READS
26 {
27 assert(output_pixels != 0);
28 assert(channels != 0);
29
30 do {
31 const uint8_t* i0 = (const uint8_t*) ((uintptr_t) input[0] + input_offset);
32 const uint8_t* i1 = (const uint8_t*) ((uintptr_t) input[1] + input_offset);
33 const uint8_t* i2 = (const uint8_t*) ((uintptr_t) input[2] + input_offset);
34 const uint8_t* i3 = (const uint8_t*) ((uintptr_t) input[3] + input_offset);
35 input += 4;
36
37 const __m128i valpha = _mm_cvtsi32_si128(*((const int*) weights));
38 weights += 2;
39 __m128i valphah = _mm_shufflelo_epi16(valpha, _MM_SHUFFLE(0, 0, 0, 0));
40 valphah = _mm_unpacklo_epi64(valphah, valphah);
41 __m128i valphav = _mm_srli_epi32(valpha, 16);
42 valphav = _mm_shuffle_epi32(valphav, _MM_SHUFFLE(0, 0, 0, 0));
43
44 valphah = _mm_blend_epi16(valphah, _mm_sub_epi16(_mm_set1_epi32(0x08000000), valphah), 0xAA);
45
46 const __m128i vrounding = _mm_set1_epi32(0x00200000);
47
48 size_t c = channels;
49 for (; c >= 8 * sizeof(uint8_t); c -= 8 * sizeof(uint8_t)) {
50 const __m128i vtl01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i0));
51 i0 += 8;
52 const __m128i vtr01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i1));
53 i1 += 8;
54 const __m128i vbl01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i2));
55 i2 += 8;
56 const __m128i vbr01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i3));
57 i3 += 8;
58
59
60 const __m128i vdr01234567 = _mm_sub_epi16(vbr01234567, vtr01234567);
61 const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah);
62 const __m128i vdl01234567 = _mm_sub_epi16(vbl01234567, vtl01234567);
63 const __m128i vt4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vtr01234567, vtl01234567), valphah);
64
65 const __m128i vd0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vdr01234567, vdl01234567), valphah);
66 const __m128i vd4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vdr01234567, vdl01234567), valphah);
67
68 __m128i vacc0123 = _mm_mullo_epi32(vd0123, valphav);
69 __m128i vacc4567 = _mm_mullo_epi32(vd4567, valphav);
70
71 vacc0123 = _mm_add_epi32(_mm_slli_epi32(vt0123, 11), vacc0123);
72 vacc4567 = _mm_add_epi32(_mm_slli_epi32(vt4567, 11), vacc4567);
73
74 vacc0123 = _mm_srli_epi32(_mm_add_epi16(vacc0123, vrounding), 22);
75 vacc4567 = _mm_srli_epi32(_mm_add_epi16(vacc4567, vrounding), 22);
76
77 const __m128i vacc01234567 = _mm_packs_epi32(vacc0123, vacc4567);
78
79 const __m128i vo01234567 = _mm_packus_epi16(vacc01234567, vacc01234567);
80
81 _mm_storel_epi64((__m128i*) output, vo01234567);
82 output += 8;
83 }
84 if XNN_UNLIKELY(c != 0) {
85 const __m128i vtl01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i0));
86 const __m128i vtr01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i1));
87 const __m128i vbl01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i2));
88 const __m128i vbr01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i3));
89
90
91 const __m128i vdr01234567 = _mm_sub_epi16(vbr01234567, vtr01234567);
92 const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah);
93 const __m128i vdl01234567 = _mm_sub_epi16(vbl01234567, vtl01234567);
94 const __m128i vt4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vtr01234567, vtl01234567), valphah);
95
96 const __m128i vd0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vdr01234567, vdl01234567), valphah);
97 const __m128i vd4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vdr01234567, vdl01234567), valphah);
98
99 __m128i vacc0123 = _mm_mullo_epi32(vd0123, valphav);
100 __m128i vacc4567 = _mm_mullo_epi32(vd4567, valphav);
101
102 vacc0123 = _mm_add_epi32(_mm_slli_epi32(vt0123, 11), vacc0123);
103 vacc4567 = _mm_add_epi32(_mm_slli_epi32(vt4567, 11), vacc4567);
104
105 vacc0123 = _mm_srli_epi32(_mm_add_epi16(vacc0123, vrounding), 22);
106 vacc4567 = _mm_srli_epi32(_mm_add_epi16(vacc4567, vrounding), 22);
107
108 const __m128i vacc01234567 = _mm_packs_epi32(vacc0123, vacc4567);
109
110 __m128i vo01234567 = _mm_packus_epi16(vacc01234567, vacc01234567);
111
112 if (c & (4 * sizeof(uint8_t))) {
113 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vo01234567);
114 output += 4;
115 vo01234567 = _mm_srli_epi64(vo01234567, 32);
116 }
117 if (c & (2 * sizeof(uint8_t))) {
118 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vo01234567, 0);
119 output += 2;
120 vo01234567 = _mm_srli_epi32(vo01234567, 16);
121 }
122 if (c & (1 * sizeof(uint8_t))) {
123 *output++ = (uint8_t) _mm_extract_epi8(vo01234567, 0);
124 }
125 }
126
127 output = (uint8_t*) ((uintptr_t) output + output_increment);
128 } while (--output_pixels != 0);
129 }
130