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