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