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 #include <xnnpack/unaligned.h>
17
18
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)19 void xnn_u8_ibilinear_ukernel__sse41_c8(
20 size_t output_pixels,
21 size_t channels,
22 const uint8_t**restrict input,
23 size_t input_offset,
24 const int16_t*restrict weights,
25 uint8_t*restrict output,
26 size_t output_increment) XNN_OOB_READS
27 {
28 assert(output_pixels != 0);
29 assert(channels != 0);
30
31 do {
32 const uint8_t* i0 = (const uint8_t*) ((uintptr_t) input[0] + input_offset);
33 const uint8_t* i1 = (const uint8_t*) ((uintptr_t) input[1] + input_offset);
34 const uint8_t* i2 = (const uint8_t*) ((uintptr_t) input[2] + input_offset);
35 const uint8_t* i3 = (const uint8_t*) ((uintptr_t) input[3] + input_offset);
36 input += 4;
37
38 const __m128i valpha = _mm_cvtsi32_si128(*((const int*) weights));
39 weights += 2;
40 __m128i valphah = _mm_shufflelo_epi16(valpha, _MM_SHUFFLE(0, 0, 0, 0));
41 valphah = _mm_unpacklo_epi64(valphah, valphah);
42 __m128i valphav = _mm_srli_epi32(valpha, 16);
43 valphav = _mm_shuffle_epi32(valphav, _MM_SHUFFLE(0, 0, 0, 0));
44
45 valphah = _mm_blend_epi16(valphah, _mm_sub_epi16(_mm_set1_epi32(0x08000000), valphah), 0xAA);
46
47 const __m128i vrounding = _mm_set1_epi32(0x00200000);
48
49 size_t c = channels;
50 for (; c >= 8 * sizeof(uint8_t); c -= 8 * sizeof(uint8_t)) {
51 const __m128i vtl01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i0));
52 i0 += 8;
53 const __m128i vtr01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i1));
54 i1 += 8;
55 const __m128i vbl01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i2));
56 i2 += 8;
57 const __m128i vbr01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i3));
58 i3 += 8;
59
60
61 const __m128i vdr01234567 = _mm_sub_epi16(vbr01234567, vtr01234567);
62 const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah);
63 const __m128i vdl01234567 = _mm_sub_epi16(vbl01234567, vtl01234567);
64 const __m128i vt4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vtr01234567, vtl01234567), valphah);
65
66 const __m128i vd0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vdr01234567, vdl01234567), valphah);
67 const __m128i vd4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vdr01234567, vdl01234567), valphah);
68
69 __m128i vacc0123 = _mm_mullo_epi32(vd0123, valphav);
70 __m128i vacc4567 = _mm_mullo_epi32(vd4567, valphav);
71
72 vacc0123 = _mm_add_epi32(_mm_slli_epi32(vt0123, 11), vacc0123);
73 vacc4567 = _mm_add_epi32(_mm_slli_epi32(vt4567, 11), vacc4567);
74
75 vacc0123 = _mm_srli_epi32(_mm_add_epi16(vacc0123, vrounding), 22);
76 vacc4567 = _mm_srli_epi32(_mm_add_epi16(vacc4567, vrounding), 22);
77
78 const __m128i vacc01234567 = _mm_packs_epi32(vacc0123, vacc4567);
79
80 const __m128i vo01234567 = _mm_packus_epi16(vacc01234567, vacc01234567);
81
82 _mm_storel_epi64((__m128i*) output, vo01234567);
83 output += 8;
84 }
85 if XNN_UNLIKELY(c != 0) {
86 const __m128i vtl01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i0));
87 const __m128i vtr01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i1));
88 const __m128i vbl01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i2));
89 const __m128i vbr01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i3));
90
91
92 const __m128i vdr01234567 = _mm_sub_epi16(vbr01234567, vtr01234567);
93 const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah);
94 const __m128i vdl01234567 = _mm_sub_epi16(vbl01234567, vtl01234567);
95 const __m128i vt4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vtr01234567, vtl01234567), valphah);
96
97 const __m128i vd0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vdr01234567, vdl01234567), valphah);
98 const __m128i vd4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vdr01234567, vdl01234567), valphah);
99
100 __m128i vacc0123 = _mm_mullo_epi32(vd0123, valphav);
101 __m128i vacc4567 = _mm_mullo_epi32(vd4567, valphav);
102
103 vacc0123 = _mm_add_epi32(_mm_slli_epi32(vt0123, 11), vacc0123);
104 vacc4567 = _mm_add_epi32(_mm_slli_epi32(vt4567, 11), vacc4567);
105
106 vacc0123 = _mm_srli_epi32(_mm_add_epi16(vacc0123, vrounding), 22);
107 vacc4567 = _mm_srli_epi32(_mm_add_epi16(vacc4567, vrounding), 22);
108
109 const __m128i vacc01234567 = _mm_packs_epi32(vacc0123, vacc4567);
110
111 __m128i vo01234567 = _mm_packus_epi16(vacc01234567, vacc01234567);
112
113 if (c & (4 * sizeof(uint8_t))) {
114 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vo01234567));
115 output += 4;
116 vo01234567 = _mm_srli_epi64(vo01234567, 32);
117 }
118 if (c & (2 * sizeof(uint8_t))) {
119 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vo01234567, 0));
120 output += 2;
121 vo01234567 = _mm_srli_epi32(vo01234567, 16);
122 }
123 if (c & (1 * sizeof(uint8_t))) {
124 *output++ = (uint8_t) _mm_extract_epi8(vo01234567, 0);
125 }
126 }
127
128 output = (uint8_t*) ((uintptr_t) output + output_increment);
129 } while (--output_pixels != 0);
130 }
131