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_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__sse2_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_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 >= 16 * sizeof(uint8_t); c -= 16 * sizeof(uint8_t)) {
52 __m128i vtl01234567 = _mm_loadl_epi64((const __m128i*) i0);
53 __m128i vtr01234567 = _mm_loadl_epi64((const __m128i*) i1);
54 __m128i vbl01234567 = _mm_loadl_epi64((const __m128i*) i2);
55 __m128i vbr01234567 = _mm_loadl_epi64((const __m128i*) i3);
56 __m128i vtl89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8));
57 __m128i vtr89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8));
58 __m128i vbl89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8));
59 __m128i vbr89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8));
60 i0 += 16;
61 i1 += 16;
62 i2 += 16;
63 i3 += 16;
64
65 __m128i vzero = _mm_setzero_si128();
66 vtl01234567 = _mm_unpacklo_epi8(vtl01234567, vzero);
67 vtr01234567 = _mm_unpacklo_epi8(vtr01234567, vzero);
68 vbl01234567 = _mm_unpacklo_epi8(vbl01234567, vzero);
69 vbr01234567 = _mm_unpacklo_epi8(vbr01234567, vzero);
70 vtl89ABCDEF = _mm_unpacklo_epi8(vtl89ABCDEF, vzero);
71 vtr89ABCDEF = _mm_unpacklo_epi8(vtr89ABCDEF, vzero);
72 vbl89ABCDEF = _mm_unpacklo_epi8(vbl89ABCDEF, vzero);
73 vbr89ABCDEF = _mm_unpacklo_epi8(vbr89ABCDEF, vzero);
74
75 const __m128i vdr01234567 = _mm_sub_epi16(vbr01234567, vtr01234567);
76 const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah);
77 const __m128i vdl01234567 = _mm_sub_epi16(vbl01234567, vtl01234567);
78 const __m128i vt4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vtr01234567, vtl01234567), valphah);
79 const __m128i vdr89ABCDEF = _mm_sub_epi16(vbr89ABCDEF, vtr89ABCDEF);
80 const __m128i vt89AB = _mm_madd_epi16(_mm_unpacklo_epi16(vtr89ABCDEF, vtl89ABCDEF), valphah);
81 const __m128i vdl89ABCDEF = _mm_sub_epi16(vbl89ABCDEF, vtl89ABCDEF);
82 const __m128i vtCDEF = _mm_madd_epi16(_mm_unpackhi_epi16(vtr89ABCDEF, vtl89ABCDEF), valphah);
83
84 const __m128i vd0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vdr01234567, vdl01234567), valphah);
85 const __m128i vd4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vdr01234567, vdl01234567), valphah);
86 const __m128i vd89AB = _mm_madd_epi16(_mm_unpacklo_epi16(vdr89ABCDEF, vdl89ABCDEF), valphah);
87 const __m128i vdCDEF = _mm_madd_epi16(_mm_unpackhi_epi16(vdr89ABCDEF, vdl89ABCDEF), valphah);
88
89 __m128i vacc0123 = _mm_slli_epi32(_mm_mulhi_epu16(vd0123, valphav), 16);
90 __m128i vacc4567 = _mm_slli_epi32(_mm_mulhi_epu16(vd4567, valphav), 16);
91 __m128i vacc89AB = _mm_slli_epi32(_mm_mulhi_epu16(vd89AB, valphav), 16);
92 __m128i vaccCDEF = _mm_slli_epi32(_mm_mulhi_epu16(vdCDEF, valphav), 16);
93
94 vacc0123 = _mm_add_epi16(_mm_mullo_epi16(vd0123, valphav), vacc0123);
95 vacc4567 = _mm_add_epi16(_mm_mullo_epi16(vd4567, valphav), vacc4567);
96 vacc89AB = _mm_add_epi16(_mm_mullo_epi16(vd89AB, valphav), vacc89AB);
97 vaccCDEF = _mm_add_epi16(_mm_mullo_epi16(vdCDEF, valphav), vaccCDEF);
98
99 vacc0123 = _mm_add_epi32(_mm_slli_epi32(vt0123, 11), vacc0123);
100 vacc4567 = _mm_add_epi32(_mm_slli_epi32(vt4567, 11), vacc4567);
101 vacc89AB = _mm_add_epi32(_mm_slli_epi32(vt89AB, 11), vacc89AB);
102 vaccCDEF = _mm_add_epi32(_mm_slli_epi32(vtCDEF, 11), vaccCDEF);
103
104 vacc0123 = _mm_srli_epi32(_mm_add_epi16(vacc0123, vrounding), 22);
105 vacc4567 = _mm_srli_epi32(_mm_add_epi16(vacc4567, vrounding), 22);
106 vacc89AB = _mm_srli_epi32(_mm_add_epi16(vacc89AB, vrounding), 22);
107 vaccCDEF = _mm_srli_epi32(_mm_add_epi16(vaccCDEF, vrounding), 22);
108
109 const __m128i vacc01234567 = _mm_packs_epi32(vacc0123, vacc4567);
110 const __m128i vacc89ABCDEF = _mm_packs_epi32(vacc89AB, vaccCDEF);
111
112 const __m128i vo0123456789ABCDEF = _mm_packus_epi16(vacc01234567, vacc89ABCDEF);
113
114 _mm_storeu_si128((__m128i*) output, vo0123456789ABCDEF);
115 output += 16;
116 }
117 for (; c >= 8 * sizeof(uint8_t); c -= 8 * sizeof(uint8_t)) {
118 __m128i vtl01234567 = _mm_loadl_epi64((const __m128i*) i0);
119 i0 += 8;
120 __m128i vtr01234567 = _mm_loadl_epi64((const __m128i*) i1);
121 i1 += 8;
122 __m128i vbl01234567 = _mm_loadl_epi64((const __m128i*) i2);
123 i2 += 8;
124 __m128i vbr01234567 = _mm_loadl_epi64((const __m128i*) i3);
125 i3 += 8;
126
127 __m128i vzero = _mm_setzero_si128();
128 vtl01234567 = _mm_unpacklo_epi8(vtl01234567, vzero);
129 vtr01234567 = _mm_unpacklo_epi8(vtr01234567, vzero);
130 vbl01234567 = _mm_unpacklo_epi8(vbl01234567, vzero);
131 vbr01234567 = _mm_unpacklo_epi8(vbr01234567, vzero);
132
133 const __m128i vdr01234567 = _mm_sub_epi16(vbr01234567, vtr01234567);
134 const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah);
135 const __m128i vdl01234567 = _mm_sub_epi16(vbl01234567, vtl01234567);
136 const __m128i vt4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vtr01234567, vtl01234567), valphah);
137
138 const __m128i vd0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vdr01234567, vdl01234567), valphah);
139 const __m128i vd4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vdr01234567, vdl01234567), valphah);
140
141 __m128i vacc0123 = _mm_slli_epi32(_mm_mulhi_epu16(vd0123, valphav), 16);
142 __m128i vacc4567 = _mm_slli_epi32(_mm_mulhi_epu16(vd4567, valphav), 16);
143
144 vacc0123 = _mm_add_epi16(_mm_mullo_epi16(vd0123, valphav), vacc0123);
145 vacc4567 = _mm_add_epi16(_mm_mullo_epi16(vd4567, valphav), vacc4567);
146
147 vacc0123 = _mm_add_epi32(_mm_slli_epi32(vt0123, 11), vacc0123);
148 vacc4567 = _mm_add_epi32(_mm_slli_epi32(vt4567, 11), vacc4567);
149
150 vacc0123 = _mm_srli_epi32(_mm_add_epi16(vacc0123, vrounding), 22);
151 vacc4567 = _mm_srli_epi32(_mm_add_epi16(vacc4567, vrounding), 22);
152
153 const __m128i vacc01234567 = _mm_packs_epi32(vacc0123, vacc4567);
154
155 const __m128i vo01234567 = _mm_packus_epi16(vacc01234567, vacc01234567);
156
157 _mm_storel_epi64((__m128i*) output, vo01234567);
158 output += 8;
159 }
160 if XNN_UNLIKELY(c != 0) {
161 __m128i vtl01234567 = _mm_loadl_epi64((const __m128i*) i0);
162 __m128i vtr01234567 = _mm_loadl_epi64((const __m128i*) i1);
163 __m128i vbl01234567 = _mm_loadl_epi64((const __m128i*) i2);
164 __m128i vbr01234567 = _mm_loadl_epi64((const __m128i*) i3);
165
166 __m128i vzero = _mm_setzero_si128();
167 vtl01234567 = _mm_unpacklo_epi8(vtl01234567, vzero);
168 vtr01234567 = _mm_unpacklo_epi8(vtr01234567, vzero);
169 vbl01234567 = _mm_unpacklo_epi8(vbl01234567, vzero);
170 vbr01234567 = _mm_unpacklo_epi8(vbr01234567, vzero);
171
172 const __m128i vdr01234567 = _mm_sub_epi16(vbr01234567, vtr01234567);
173 const __m128i vt0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vtr01234567, vtl01234567), valphah);
174 const __m128i vdl01234567 = _mm_sub_epi16(vbl01234567, vtl01234567);
175 const __m128i vt4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vtr01234567, vtl01234567), valphah);
176
177 const __m128i vd0123 = _mm_madd_epi16(_mm_unpacklo_epi16(vdr01234567, vdl01234567), valphah);
178 const __m128i vd4567 = _mm_madd_epi16(_mm_unpackhi_epi16(vdr01234567, vdl01234567), valphah);
179
180 __m128i vacc0123 = _mm_slli_epi32(_mm_mulhi_epu16(vd0123, valphav), 16);
181 __m128i vacc4567 = _mm_slli_epi32(_mm_mulhi_epu16(vd4567, valphav), 16);
182
183 vacc0123 = _mm_add_epi16(_mm_mullo_epi16(vd0123, valphav), vacc0123);
184 vacc4567 = _mm_add_epi16(_mm_mullo_epi16(vd4567, valphav), vacc4567);
185
186 vacc0123 = _mm_add_epi32(_mm_slli_epi32(vt0123, 11), vacc0123);
187 vacc4567 = _mm_add_epi32(_mm_slli_epi32(vt4567, 11), vacc4567);
188
189 vacc0123 = _mm_srli_epi32(_mm_add_epi16(vacc0123, vrounding), 22);
190 vacc4567 = _mm_srli_epi32(_mm_add_epi16(vacc4567, vrounding), 22);
191
192 const __m128i vacc01234567 = _mm_packs_epi32(vacc0123, vacc4567);
193
194 __m128i vo01234567 = _mm_packus_epi16(vacc01234567, vacc01234567);
195
196 if (c & (4 * sizeof(uint8_t))) {
197 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vo01234567));
198 output += 4;
199 vo01234567 = _mm_srli_epi64(vo01234567, 32);
200 }
201 uint32_t vo0123 = (uint32_t) _mm_cvtsi128_si32(vo01234567);
202 if (c & (2 * sizeof(uint8_t))) {
203 unaligned_store_u16(output, (uint16_t) vo0123);
204 output += 2;
205 vo0123 >>= 16;
206 }
207 if (c & (1 * sizeof(uint8_t))) {
208 *output++ = (uint8_t) vo0123;
209 }
210 }
211
212 output = (uint8_t*) ((uintptr_t) output + output_increment);
213 } while (--output_pixels != 0);
214 }
215