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 <emmintrin.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/ibilinear.h>
16 #include <xnnpack/unaligned.h>
17
18
xnn_u8_ibilinear_ukernel__sse2_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__sse2_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_shufflelo_epi16(valpha, _MM_SHUFFLE(1, 1, 1, 1));
43 valphav = _mm_unpacklo_epi64(valphav, valphav);
44
45 valphah = _mm_xor_si128(valphah, _mm_set1_epi32(0xFFFF0000));
46 valphah = _mm_add_epi16(valphah, _mm_set1_epi32(0x08010000));
47
48 const __m128i vrounding = _mm_set1_epi32(0x00200000);
49
50 size_t c = channels;
51 for (; c >= 8 * sizeof(uint8_t); c -= 8 * sizeof(uint8_t)) {
52 __m128i vtl01234567 = _mm_loadl_epi64((const __m128i*) i0);
53 i0 += 8;
54 __m128i vtr01234567 = _mm_loadl_epi64((const __m128i*) i1);
55 i1 += 8;
56 __m128i vbl01234567 = _mm_loadl_epi64((const __m128i*) i2);
57 i2 += 8;
58 __m128i vbr01234567 = _mm_loadl_epi64((const __m128i*) i3);
59 i3 += 8;
60
61 __m128i vzero = _mm_setzero_si128();
62 vtl01234567 = _mm_unpacklo_epi8(vtl01234567, vzero);
63 vtr01234567 = _mm_unpacklo_epi8(vtr01234567, vzero);
64 vbl01234567 = _mm_unpacklo_epi8(vbl01234567, vzero);
65 vbr01234567 = _mm_unpacklo_epi8(vbr01234567, vzero);
66
67 const __m128i vdr01234567 = _mm_sub_epi16(vbr01234567, vtr01234567);
68 const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah);
69 const __m128i vdl01234567 = _mm_sub_epi16(vbl01234567, vtl01234567);
70 const __m128i vt4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vtr01234567, vtl01234567), valphah);
71
72 const __m128i vd0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vdr01234567, vdl01234567), valphah);
73 const __m128i vd4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vdr01234567, vdl01234567), valphah);
74
75 __m128i vacc0123 = _mm_slli_epi32(_mm_mulhi_epu16(vd0123, valphav), 16);
76 __m128i vacc4567 = _mm_slli_epi32(_mm_mulhi_epu16(vd4567, valphav), 16);
77
78 vacc0123 = _mm_add_epi16(_mm_mullo_epi16(vd0123, valphav), vacc0123);
79 vacc4567 = _mm_add_epi16(_mm_mullo_epi16(vd4567, valphav), vacc4567);
80
81 vacc0123 = _mm_add_epi32(_mm_slli_epi32(vt0123, 11), vacc0123);
82 vacc4567 = _mm_add_epi32(_mm_slli_epi32(vt4567, 11), vacc4567);
83
84 vacc0123 = _mm_srli_epi32(_mm_add_epi16(vacc0123, vrounding), 22);
85 vacc4567 = _mm_srli_epi32(_mm_add_epi16(vacc4567, vrounding), 22);
86
87 const __m128i vacc01234567 = _mm_packs_epi32(vacc0123, vacc4567);
88
89 const __m128i vo01234567 = _mm_packus_epi16(vacc01234567, vacc01234567);
90
91 _mm_storel_epi64((__m128i*) output, vo01234567);
92 output += 8;
93 }
94 if XNN_UNLIKELY(c != 0) {
95 __m128i vtl01234567 = _mm_loadl_epi64((const __m128i*) i0);
96 __m128i vtr01234567 = _mm_loadl_epi64((const __m128i*) i1);
97 __m128i vbl01234567 = _mm_loadl_epi64((const __m128i*) i2);
98 __m128i vbr01234567 = _mm_loadl_epi64((const __m128i*) i3);
99
100 __m128i vzero = _mm_setzero_si128();
101 vtl01234567 = _mm_unpacklo_epi8(vtl01234567, vzero);
102 vtr01234567 = _mm_unpacklo_epi8(vtr01234567, vzero);
103 vbl01234567 = _mm_unpacklo_epi8(vbl01234567, vzero);
104 vbr01234567 = _mm_unpacklo_epi8(vbr01234567, vzero);
105
106 const __m128i vdr01234567 = _mm_sub_epi16(vbr01234567, vtr01234567);
107 const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah);
108 const __m128i vdl01234567 = _mm_sub_epi16(vbl01234567, vtl01234567);
109 const __m128i vt4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vtr01234567, vtl01234567), valphah);
110
111 const __m128i vd0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vdr01234567, vdl01234567), valphah);
112 const __m128i vd4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vdr01234567, vdl01234567), valphah);
113
114 __m128i vacc0123 = _mm_slli_epi32(_mm_mulhi_epu16(vd0123, valphav), 16);
115 __m128i vacc4567 = _mm_slli_epi32(_mm_mulhi_epu16(vd4567, valphav), 16);
116
117 vacc0123 = _mm_add_epi16(_mm_mullo_epi16(vd0123, valphav), vacc0123);
118 vacc4567 = _mm_add_epi16(_mm_mullo_epi16(vd4567, valphav), vacc4567);
119
120 vacc0123 = _mm_add_epi32(_mm_slli_epi32(vt0123, 11), vacc0123);
121 vacc4567 = _mm_add_epi32(_mm_slli_epi32(vt4567, 11), vacc4567);
122
123 vacc0123 = _mm_srli_epi32(_mm_add_epi16(vacc0123, vrounding), 22);
124 vacc4567 = _mm_srli_epi32(_mm_add_epi16(vacc4567, vrounding), 22);
125
126 const __m128i vacc01234567 = _mm_packs_epi32(vacc0123, vacc4567);
127
128 __m128i vo01234567 = _mm_packus_epi16(vacc01234567, vacc01234567);
129
130 if (c & (4 * sizeof(uint8_t))) {
131 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vo01234567));
132 output += 4;
133 vo01234567 = _mm_srli_epi64(vo01234567, 32);
134 }
135 uint32_t vo0123 = (uint32_t) _mm_cvtsi128_si32(vo01234567);
136 if (c & (2 * sizeof(uint8_t))) {
137 unaligned_store_u16(output, (uint16_t) vo0123);
138 output += 2;
139 vo0123 >>= 16;
140 }
141 if (c & (1 * sizeof(uint8_t))) {
142 *output++ = (uint8_t) vo0123;
143 }
144 }
145
146 output = (uint8_t*) ((uintptr_t) output + output_increment);
147 } while (--output_pixels != 0);
148 }
149