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_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)19 void xnn_u8_ibilinear_ukernel__sse41_c16(
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 >= 16 * sizeof(uint8_t); c -= 16 * sizeof(uint8_t)) {
51 const __m128i vtl01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i0));
52 const __m128i vtr01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i1));
53 const __m128i vbl01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i2));
54 const __m128i vbr01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i3));
55 const __m128i vtl89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i0 + 8)));
56 const __m128i vtr89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i1 + 8)));
57 const __m128i vbl89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i2 + 8)));
58 const __m128i vbr89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i3 + 8)));
59 i0 += 16;
60 i1 += 16;
61 i2 += 16;
62 i3 += 16;
63
64
65 const __m128i vdr01234567 = _mm_sub_epi16(vbr01234567, vtr01234567);
66 const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah);
67 const __m128i vdl01234567 = _mm_sub_epi16(vbl01234567, vtl01234567);
68 const __m128i vt4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vtr01234567, vtl01234567), valphah);
69 const __m128i vdr89ABCDEF = _mm_sub_epi16(vbr89ABCDEF, vtr89ABCDEF);
70 const __m128i vt89AB = _mm_madd_epi16(_mm_unpacklo_epi16(vtr89ABCDEF, vtl89ABCDEF), valphah);
71 const __m128i vdl89ABCDEF = _mm_sub_epi16(vbl89ABCDEF, vtl89ABCDEF);
72 const __m128i vtCDEF = _mm_madd_epi16(_mm_unpackhi_epi16(vtr89ABCDEF, vtl89ABCDEF), valphah);
73
74 const __m128i vd0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vdr01234567, vdl01234567), valphah);
75 const __m128i vd4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vdr01234567, vdl01234567), valphah);
76 const __m128i vd89AB = _mm_madd_epi16(_mm_unpacklo_epi16(vdr89ABCDEF, vdl89ABCDEF), valphah);
77 const __m128i vdCDEF = _mm_madd_epi16(_mm_unpackhi_epi16(vdr89ABCDEF, vdl89ABCDEF), valphah);
78
79 __m128i vacc0123 = _mm_mullo_epi32(vd0123, valphav);
80 __m128i vacc4567 = _mm_mullo_epi32(vd4567, valphav);
81 __m128i vacc89AB = _mm_mullo_epi32(vd89AB, valphav);
82 __m128i vaccCDEF = _mm_mullo_epi32(vdCDEF, valphav);
83
84 vacc0123 = _mm_add_epi32(_mm_slli_epi32(vt0123, 11), vacc0123);
85 vacc4567 = _mm_add_epi32(_mm_slli_epi32(vt4567, 11), vacc4567);
86 vacc89AB = _mm_add_epi32(_mm_slli_epi32(vt89AB, 11), vacc89AB);
87 vaccCDEF = _mm_add_epi32(_mm_slli_epi32(vtCDEF, 11), vaccCDEF);
88
89 vacc0123 = _mm_srli_epi32(_mm_add_epi16(vacc0123, vrounding), 22);
90 vacc4567 = _mm_srli_epi32(_mm_add_epi16(vacc4567, vrounding), 22);
91 vacc89AB = _mm_srli_epi32(_mm_add_epi16(vacc89AB, vrounding), 22);
92 vaccCDEF = _mm_srli_epi32(_mm_add_epi16(vaccCDEF, vrounding), 22);
93
94 const __m128i vacc01234567 = _mm_packs_epi32(vacc0123, vacc4567);
95 const __m128i vacc89ABCDEF = _mm_packs_epi32(vacc89AB, vaccCDEF);
96
97 const __m128i vo0123456789ABCDEF = _mm_packus_epi16(vacc01234567, vacc89ABCDEF);
98
99 _mm_storeu_si128((__m128i*) output, vo0123456789ABCDEF);
100 output += 16;
101 }
102 for (; c >= 8 * sizeof(uint8_t); c -= 8 * sizeof(uint8_t)) {
103 const __m128i vtl01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i0));
104 i0 += 8;
105 const __m128i vtr01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i1));
106 i1 += 8;
107 const __m128i vbl01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i2));
108 i2 += 8;
109 const __m128i vbr01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i3));
110 i3 += 8;
111
112
113 const __m128i vdr01234567 = _mm_sub_epi16(vbr01234567, vtr01234567);
114 const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah);
115 const __m128i vdl01234567 = _mm_sub_epi16(vbl01234567, vtl01234567);
116 const __m128i vt4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vtr01234567, vtl01234567), valphah);
117
118 const __m128i vd0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vdr01234567, vdl01234567), valphah);
119 const __m128i vd4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vdr01234567, vdl01234567), valphah);
120
121 __m128i vacc0123 = _mm_mullo_epi32(vd0123, valphav);
122 __m128i vacc4567 = _mm_mullo_epi32(vd4567, valphav);
123
124 vacc0123 = _mm_add_epi32(_mm_slli_epi32(vt0123, 11), vacc0123);
125 vacc4567 = _mm_add_epi32(_mm_slli_epi32(vt4567, 11), vacc4567);
126
127 vacc0123 = _mm_srli_epi32(_mm_add_epi16(vacc0123, vrounding), 22);
128 vacc4567 = _mm_srli_epi32(_mm_add_epi16(vacc4567, vrounding), 22);
129
130 const __m128i vacc01234567 = _mm_packs_epi32(vacc0123, vacc4567);
131
132 const __m128i vo01234567 = _mm_packus_epi16(vacc01234567, vacc01234567);
133
134 _mm_storel_epi64((__m128i*) output, vo01234567);
135 output += 8;
136 }
137 if XNN_UNLIKELY(c != 0) {
138 const __m128i vtl01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i0));
139 const __m128i vtr01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i1));
140 const __m128i vbl01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i2));
141 const __m128i vbr01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i3));
142
143
144 const __m128i vdr01234567 = _mm_sub_epi16(vbr01234567, vtr01234567);
145 const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah);
146 const __m128i vdl01234567 = _mm_sub_epi16(vbl01234567, vtl01234567);
147 const __m128i vt4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vtr01234567, vtl01234567), valphah);
148
149 const __m128i vd0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vdr01234567, vdl01234567), valphah);
150 const __m128i vd4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vdr01234567, vdl01234567), valphah);
151
152 __m128i vacc0123 = _mm_mullo_epi32(vd0123, valphav);
153 __m128i vacc4567 = _mm_mullo_epi32(vd4567, valphav);
154
155 vacc0123 = _mm_add_epi32(_mm_slli_epi32(vt0123, 11), vacc0123);
156 vacc4567 = _mm_add_epi32(_mm_slli_epi32(vt4567, 11), vacc4567);
157
158 vacc0123 = _mm_srli_epi32(_mm_add_epi16(vacc0123, vrounding), 22);
159 vacc4567 = _mm_srli_epi32(_mm_add_epi16(vacc4567, vrounding), 22);
160
161 const __m128i vacc01234567 = _mm_packs_epi32(vacc0123, vacc4567);
162
163 __m128i vo01234567 = _mm_packus_epi16(vacc01234567, vacc01234567);
164
165 if (c & (4 * sizeof(uint8_t))) {
166 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vo01234567));
167 output += 4;
168 vo01234567 = _mm_srli_epi64(vo01234567, 32);
169 }
170 if (c & (2 * sizeof(uint8_t))) {
171 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vo01234567, 0));
172 output += 2;
173 vo01234567 = _mm_srli_epi32(vo01234567, 16);
174 }
175 if (c & (1 * sizeof(uint8_t))) {
176 *output++ = (uint8_t) _mm_extract_epi8(vo01234567, 0);
177 }
178 }
179
180 output = (uint8_t*) ((uintptr_t) output + output_increment);
181 } while (--output_pixels != 0);
182 }
183