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_c16(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_c16(
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 >= 16 * sizeof(uint8_t); c -= 16 * sizeof(uint8_t)) {
50 const __m128i vtl01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i0));
51 const __m128i vtr01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i1));
52 const __m128i vbl01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i2));
53 const __m128i vbr01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i3));
54 const __m128i vtl89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i0 + 8)));
55 const __m128i vtr89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i1 + 8)));
56 const __m128i vbl89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i2 + 8)));
57 const __m128i vbr89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i3 + 8)));
58 i0 += 16;
59 i1 += 16;
60 i2 += 16;
61 i3 += 16;
62
63
64 const __m128i vdr01234567 = _mm_sub_epi16(vbr01234567, vtr01234567);
65 const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah);
66 const __m128i vdl01234567 = _mm_sub_epi16(vbl01234567, vtl01234567);
67 const __m128i vt4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vtr01234567, vtl01234567), valphah);
68 const __m128i vdr89ABCDEF = _mm_sub_epi16(vbr89ABCDEF, vtr89ABCDEF);
69 const __m128i vt89AB = _mm_madd_epi16(_mm_unpacklo_epi16(vtr89ABCDEF, vtl89ABCDEF), valphah);
70 const __m128i vdl89ABCDEF = _mm_sub_epi16(vbl89ABCDEF, vtl89ABCDEF);
71 const __m128i vtCDEF = _mm_madd_epi16(_mm_unpackhi_epi16(vtr89ABCDEF, vtl89ABCDEF), valphah);
72
73 const __m128i vd0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vdr01234567, vdl01234567), valphah);
74 const __m128i vd4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vdr01234567, vdl01234567), valphah);
75 const __m128i vd89AB = _mm_madd_epi16(_mm_unpacklo_epi16(vdr89ABCDEF, vdl89ABCDEF), valphah);
76 const __m128i vdCDEF = _mm_madd_epi16(_mm_unpackhi_epi16(vdr89ABCDEF, vdl89ABCDEF), valphah);
77
78 __m128i vacc0123 = _mm_mullo_epi32(vd0123, valphav);
79 __m128i vacc4567 = _mm_mullo_epi32(vd4567, valphav);
80 __m128i vacc89AB = _mm_mullo_epi32(vd89AB, valphav);
81 __m128i vaccCDEF = _mm_mullo_epi32(vdCDEF, valphav);
82
83 vacc0123 = _mm_add_epi32(_mm_slli_epi32(vt0123, 11), vacc0123);
84 vacc4567 = _mm_add_epi32(_mm_slli_epi32(vt4567, 11), vacc4567);
85 vacc89AB = _mm_add_epi32(_mm_slli_epi32(vt89AB, 11), vacc89AB);
86 vaccCDEF = _mm_add_epi32(_mm_slli_epi32(vtCDEF, 11), vaccCDEF);
87
88 vacc0123 = _mm_srli_epi32(_mm_add_epi16(vacc0123, vrounding), 22);
89 vacc4567 = _mm_srli_epi32(_mm_add_epi16(vacc4567, vrounding), 22);
90 vacc89AB = _mm_srli_epi32(_mm_add_epi16(vacc89AB, vrounding), 22);
91 vaccCDEF = _mm_srli_epi32(_mm_add_epi16(vaccCDEF, vrounding), 22);
92
93 const __m128i vacc01234567 = _mm_packs_epi32(vacc0123, vacc4567);
94 const __m128i vacc89ABCDEF = _mm_packs_epi32(vacc89AB, vaccCDEF);
95
96 const __m128i vo0123456789ABCDEF = _mm_packus_epi16(vacc01234567, vacc89ABCDEF);
97
98 _mm_storeu_si128((__m128i*) output, vo0123456789ABCDEF);
99 output += 16;
100 }
101 for (; c >= 8 * sizeof(uint8_t); c -= 8 * sizeof(uint8_t)) {
102 const __m128i vtl01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i0));
103 i0 += 8;
104 const __m128i vtr01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i1));
105 i1 += 8;
106 const __m128i vbl01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i2));
107 i2 += 8;
108 const __m128i vbr01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i3));
109 i3 += 8;
110
111
112 const __m128i vdr01234567 = _mm_sub_epi16(vbr01234567, vtr01234567);
113 const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah);
114 const __m128i vdl01234567 = _mm_sub_epi16(vbl01234567, vtl01234567);
115 const __m128i vt4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vtr01234567, vtl01234567), valphah);
116
117 const __m128i vd0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vdr01234567, vdl01234567), valphah);
118 const __m128i vd4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vdr01234567, vdl01234567), valphah);
119
120 __m128i vacc0123 = _mm_mullo_epi32(vd0123, valphav);
121 __m128i vacc4567 = _mm_mullo_epi32(vd4567, valphav);
122
123 vacc0123 = _mm_add_epi32(_mm_slli_epi32(vt0123, 11), vacc0123);
124 vacc4567 = _mm_add_epi32(_mm_slli_epi32(vt4567, 11), vacc4567);
125
126 vacc0123 = _mm_srli_epi32(_mm_add_epi16(vacc0123, vrounding), 22);
127 vacc4567 = _mm_srli_epi32(_mm_add_epi16(vacc4567, vrounding), 22);
128
129 const __m128i vacc01234567 = _mm_packs_epi32(vacc0123, vacc4567);
130
131 const __m128i vo01234567 = _mm_packus_epi16(vacc01234567, vacc01234567);
132
133 _mm_storel_epi64((__m128i*) output, vo01234567);
134 output += 8;
135 }
136 if XNN_UNLIKELY(c != 0) {
137 const __m128i vtl01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i0));
138 const __m128i vtr01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i1));
139 const __m128i vbl01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i2));
140 const __m128i vbr01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i3));
141
142
143 const __m128i vdr01234567 = _mm_sub_epi16(vbr01234567, vtr01234567);
144 const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah);
145 const __m128i vdl01234567 = _mm_sub_epi16(vbl01234567, vtl01234567);
146 const __m128i vt4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vtr01234567, vtl01234567), valphah);
147
148 const __m128i vd0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vdr01234567, vdl01234567), valphah);
149 const __m128i vd4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vdr01234567, vdl01234567), valphah);
150
151 __m128i vacc0123 = _mm_mullo_epi32(vd0123, valphav);
152 __m128i vacc4567 = _mm_mullo_epi32(vd4567, valphav);
153
154 vacc0123 = _mm_add_epi32(_mm_slli_epi32(vt0123, 11), vacc0123);
155 vacc4567 = _mm_add_epi32(_mm_slli_epi32(vt4567, 11), vacc4567);
156
157 vacc0123 = _mm_srli_epi32(_mm_add_epi16(vacc0123, vrounding), 22);
158 vacc4567 = _mm_srli_epi32(_mm_add_epi16(vacc4567, vrounding), 22);
159
160 const __m128i vacc01234567 = _mm_packs_epi32(vacc0123, vacc4567);
161
162 __m128i vo01234567 = _mm_packus_epi16(vacc01234567, vacc01234567);
163
164 if (c & (4 * sizeof(uint8_t))) {
165 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vo01234567);
166 output += 4;
167 vo01234567 = _mm_srli_epi64(vo01234567, 32);
168 }
169 if (c & (2 * sizeof(uint8_t))) {
170 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vo01234567, 0);
171 output += 2;
172 vo01234567 = _mm_srli_epi32(vo01234567, 16);
173 }
174 if (c & (1 * sizeof(uint8_t))) {
175 *output++ = (uint8_t) _mm_extract_epi8(vo01234567, 0);
176 }
177 }
178
179 output = (uint8_t*) ((uintptr_t) output + output_increment);
180 } while (--output_pixels != 0);
181 }
182