/
encode.sse.cpp
304 lines (238 loc) · 14.1 KB
/
encode.sse.cpp
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
namespace base64 {
namespace sse {
#define packed_dword(x) _mm_set1_epi32(x)
#define packed_byte(x) _mm_set1_epi8(char(x))
template <typename LOOKUP_FN>
void encode(LOOKUP_FN lookup, const uint8_t* input, size_t bytes, uint8_t* output) {
uint8_t* out = output;
const __m128i shuf = _mm_set_epi8(
10, 11, 9, 10,
7, 8, 6, 7,
4, 5, 3, 4,
1, 2, 0, 1
);
for (size_t i = 0; i < bytes; i += 4*3) {
// input = [xxxx|DDDC|CCBB|BAAA]
__m128i in = _mm_loadu_si128(reinterpret_cast<const __m128i*>(input + i));
// bytes from groups A, B and C are needed in separate 32-bit lanes
// in = [DDDD|CCCC|BBBB|AAAA]
//
// an input triplet has layout
// [????????|ccdddddd|bbbbcccc|aaaaaabb]
// byte 3 byte 2 byte 1 byte 0 -- byte 3 comes from the next triplet
//
// shuffling changes the order of bytes: 1, 0, 2, 1
// [bbbbcccc|ccdddddd|aaaaaabb|bbbbcccc]
// ^^^^ ^^^^^^^^ ^^^^^^^^ ^^^^
// processed bits
in = _mm_shuffle_epi8(in, shuf);
// unpacking
// t0 = [0000cccc|cc000000|aaaaaa00|00000000]
const __m128i t0 = _mm_and_si128(in, _mm_set1_epi32(0x0fc0fc00));
// t1 = [00000000|00cccccc|00000000|00aaaaaa]
// (c * (1 << 10), a * (1 << 6)) >> 16 (note: an unsigned multiplication)
const __m128i t1 = _mm_mulhi_epu16(t0, _mm_set1_epi32(0x04000040));
// t2 = [00000000|00dddddd|000000bb|bbbb0000]
const __m128i t2 = _mm_and_si128(in, _mm_set1_epi32(0x003f03f0));
// t3 = [00dddddd|00000000|00bbbbbb|00000000](
// (d * (1 << 8), b * (1 << 4))
const __m128i t3 = _mm_mullo_epi16(t2, _mm_set1_epi32(0x01000010));
// res = [00dddddd|00cccccc|00bbbbbb|00aaaaaa] = t1 | t3
const __m128i indices = _mm_or_si128(t1, t3);
const auto result = lookup(indices);
_mm_storeu_si128(reinterpret_cast<__m128i*>(out), result);
out += 16;
}
}
template <typename LOOKUP_FN>
void encode_unrolled(LOOKUP_FN lookup, const uint8_t* input, size_t bytes, uint8_t* output) {
uint8_t* out = output;
const __m128i shuf = _mm_set_epi8(
10, 11, 9, 10,
7, 8, 6, 7,
4, 5, 3, 4,
1, 2, 0, 1
);
for (size_t i = 0; i < bytes; i += 4*3 * 4) {
// unrolled improved version
__m128i in0 = _mm_loadu_si128(reinterpret_cast<const __m128i*>(input + i + 4*3 * 0));
__m128i in1 = _mm_loadu_si128(reinterpret_cast<const __m128i*>(input + i + 4*3 * 1));
__m128i in2 = _mm_loadu_si128(reinterpret_cast<const __m128i*>(input + i + 4*3 * 2));
__m128i in3 = _mm_loadu_si128(reinterpret_cast<const __m128i*>(input + i + 4*3 * 3));
in0 = _mm_shuffle_epi8(in0, shuf);
in1 = _mm_shuffle_epi8(in1, shuf);
in2 = _mm_shuffle_epi8(in2, shuf);
in3 = _mm_shuffle_epi8(in3, shuf);
const __m128i t0_0 = _mm_and_si128(in0, _mm_set1_epi32(0x0fc0fc00));
const __m128i t0_1 = _mm_and_si128(in1, _mm_set1_epi32(0x0fc0fc00));
const __m128i t0_2 = _mm_and_si128(in2, _mm_set1_epi32(0x0fc0fc00));
const __m128i t0_3 = _mm_and_si128(in3, _mm_set1_epi32(0x0fc0fc00));
const __m128i t1_0 = _mm_mulhi_epu16(t0_0, _mm_set1_epi32(0x04000040));
const __m128i t1_1 = _mm_mulhi_epu16(t0_1, _mm_set1_epi32(0x04000040));
const __m128i t1_2 = _mm_mulhi_epu16(t0_2, _mm_set1_epi32(0x04000040));
const __m128i t1_3 = _mm_mulhi_epu16(t0_3, _mm_set1_epi32(0x04000040));
const __m128i t2_0 = _mm_and_si128(in0, _mm_set1_epi32(0x003f03f0));
const __m128i t2_1 = _mm_and_si128(in1, _mm_set1_epi32(0x003f03f0));
const __m128i t2_2 = _mm_and_si128(in2, _mm_set1_epi32(0x003f03f0));
const __m128i t2_3 = _mm_and_si128(in3, _mm_set1_epi32(0x003f03f0));
const __m128i t3_0 = _mm_mullo_epi16(t2_0, _mm_set1_epi32(0x01000010));
const __m128i t3_1 = _mm_mullo_epi16(t2_1, _mm_set1_epi32(0x01000010));
const __m128i t3_2 = _mm_mullo_epi16(t2_2, _mm_set1_epi32(0x01000010));
const __m128i t3_3 = _mm_mullo_epi16(t2_3, _mm_set1_epi32(0x01000010));
const __m128i input0 = _mm_or_si128(t1_0, t3_0);
const __m128i input1 = _mm_or_si128(t1_1, t3_1);
const __m128i input2 = _mm_or_si128(t1_2, t3_2);
const __m128i input3 = _mm_or_si128(t1_3, t3_3);
{
const auto result = lookup(input0);
_mm_storeu_si128(reinterpret_cast<__m128i*>(out), result);
out += 16;
}
{
const auto result = lookup(input1);
_mm_storeu_si128(reinterpret_cast<__m128i*>(out), result);
out += 16;
}
{
const auto result = lookup(input2);
_mm_storeu_si128(reinterpret_cast<__m128i*>(out), result);
out += 16;
}
{
const auto result = lookup(input3);
_mm_storeu_si128(reinterpret_cast<__m128i*>(out), result);
out += 16;
}
}
}
void encode_full_unrolled(const uint8_t* input, size_t bytes, uint8_t* output) {
uint8_t* out = output;
const __m128i shuf = _mm_set_epi8(
10, 11, 9, 10,
7, 8, 6, 7,
4, 5, 3, 4,
1, 2, 0, 1
);
for (size_t i = 0; i < bytes; i += 4*3 * 4) {
// the same code as in encode_unrolled (no, no macros!)
__m128i in0 = _mm_loadu_si128(reinterpret_cast<const __m128i*>(input + i + 4*3 * 0));
__m128i in1 = _mm_loadu_si128(reinterpret_cast<const __m128i*>(input + i + 4*3 * 1));
__m128i in2 = _mm_loadu_si128(reinterpret_cast<const __m128i*>(input + i + 4*3 * 2));
__m128i in3 = _mm_loadu_si128(reinterpret_cast<const __m128i*>(input + i + 4*3 * 3));
in0 = _mm_shuffle_epi8(in0, shuf);
in1 = _mm_shuffle_epi8(in1, shuf);
in2 = _mm_shuffle_epi8(in2, shuf);
in3 = _mm_shuffle_epi8(in3, shuf);
const __m128i t0_0 = _mm_and_si128(in0, _mm_set1_epi32(0x0fc0fc00));
const __m128i t0_1 = _mm_and_si128(in1, _mm_set1_epi32(0x0fc0fc00));
const __m128i t0_2 = _mm_and_si128(in2, _mm_set1_epi32(0x0fc0fc00));
const __m128i t0_3 = _mm_and_si128(in3, _mm_set1_epi32(0x0fc0fc00));
const __m128i t1_0 = _mm_mulhi_epu16(t0_0, _mm_set1_epi32(0x04000040));
const __m128i t1_1 = _mm_mulhi_epu16(t0_1, _mm_set1_epi32(0x04000040));
const __m128i t1_2 = _mm_mulhi_epu16(t0_2, _mm_set1_epi32(0x04000040));
const __m128i t1_3 = _mm_mulhi_epu16(t0_3, _mm_set1_epi32(0x04000040));
const __m128i t2_0 = _mm_and_si128(in0, _mm_set1_epi32(0x003f03f0));
const __m128i t2_1 = _mm_and_si128(in1, _mm_set1_epi32(0x003f03f0));
const __m128i t2_2 = _mm_and_si128(in2, _mm_set1_epi32(0x003f03f0));
const __m128i t2_3 = _mm_and_si128(in3, _mm_set1_epi32(0x003f03f0));
const __m128i t3_0 = _mm_mullo_epi16(t2_0, _mm_set1_epi32(0x01000010));
const __m128i t3_1 = _mm_mullo_epi16(t2_1, _mm_set1_epi32(0x01000010));
const __m128i t3_2 = _mm_mullo_epi16(t2_2, _mm_set1_epi32(0x01000010));
const __m128i t3_3 = _mm_mullo_epi16(t2_3, _mm_set1_epi32(0x01000010));
const __m128i input0 = _mm_or_si128(t1_0, t3_0);
const __m128i input1 = _mm_or_si128(t1_1, t3_1);
const __m128i input2 = _mm_or_si128(t1_2, t3_2);
const __m128i input3 = _mm_or_si128(t1_3, t3_3);
// unrolled lookup_version2 from lookup.sse.cpp
__m128i result_0 = packed_byte(65);
__m128i result_1 = packed_byte(65);
__m128i result_2 = packed_byte(65);
__m128i result_3 = packed_byte(65);
const __m128i ge_26_0 = _mm_cmpgt_epi8(input0, packed_byte(25));
result_0 = _mm_add_epi8(result_0, _mm_and_si128(ge_26_0, packed_byte( 6)));
const __m128i ge_26_1 = _mm_cmpgt_epi8(input1, packed_byte(25));
result_1 = _mm_add_epi8(result_1, _mm_and_si128(ge_26_1, packed_byte( 6)));
const __m128i ge_26_2 = _mm_cmpgt_epi8(input2, packed_byte(25));
result_2 = _mm_add_epi8(result_2, _mm_and_si128(ge_26_2, packed_byte( 6)));
const __m128i ge_26_3 = _mm_cmpgt_epi8(input3, packed_byte(25));
result_3 = _mm_add_epi8(result_3, _mm_and_si128(ge_26_3, packed_byte( 6)));
const __m128i ge_52_0 = _mm_cmpgt_epi8(input0, packed_byte(51));
result_0 = _mm_sub_epi8(result_0, _mm_and_si128(ge_52_0, packed_byte( 75)));
const __m128i ge_52_1 = _mm_cmpgt_epi8(input1, packed_byte(51));
result_1 = _mm_sub_epi8(result_1, _mm_and_si128(ge_52_1, packed_byte( 75)));
const __m128i ge_52_2 = _mm_cmpgt_epi8(input2, packed_byte(51));
result_2 = _mm_sub_epi8(result_2, _mm_and_si128(ge_52_2, packed_byte( 75)));
const __m128i ge_52_3 = _mm_cmpgt_epi8(input3, packed_byte(51));
result_3 = _mm_sub_epi8(result_3, _mm_and_si128(ge_52_3, packed_byte( 75)));
const __m128i eq_62_0 = _mm_cmpeq_epi8(input0, packed_byte(62));
result_0 = _mm_add_epi8(result_0, _mm_and_si128(eq_62_0, packed_byte(241)));
const __m128i eq_62_1 = _mm_cmpeq_epi8(input1, packed_byte(62));
result_1 = _mm_add_epi8(result_1, _mm_and_si128(eq_62_1, packed_byte(241)));
const __m128i eq_62_2 = _mm_cmpeq_epi8(input2, packed_byte(62));
result_2 = _mm_add_epi8(result_2, _mm_and_si128(eq_62_2, packed_byte(241)));
const __m128i eq_62_3 = _mm_cmpeq_epi8(input3, packed_byte(62));
result_3 = _mm_add_epi8(result_3, _mm_and_si128(eq_62_3, packed_byte(241)));
const __m128i eq_63_0 = _mm_cmpeq_epi8(input0, packed_byte(63));
result_0 = _mm_sub_epi8(result_0, _mm_and_si128(eq_63_0, packed_byte( 12)));
const __m128i eq_63_1 = _mm_cmpeq_epi8(input1, packed_byte(63));
result_1 = _mm_sub_epi8(result_1, _mm_and_si128(eq_63_1, packed_byte( 12)));
const __m128i eq_63_2 = _mm_cmpeq_epi8(input2, packed_byte(63));
result_2 = _mm_sub_epi8(result_2, _mm_and_si128(eq_63_2, packed_byte( 12)));
const __m128i eq_63_3 = _mm_cmpeq_epi8(input3, packed_byte(63));
result_3 = _mm_sub_epi8(result_3, _mm_and_si128(eq_63_3, packed_byte( 12)));
result_0 = _mm_add_epi8(result_0, input0);
result_1 = _mm_add_epi8(result_1, input1);
result_2 = _mm_add_epi8(result_2, input2);
result_3 = _mm_add_epi8(result_3, input3);
_mm_storeu_si128(reinterpret_cast<__m128i*>(out), result_0);
out += 16;
_mm_storeu_si128(reinterpret_cast<__m128i*>(out), result_1);
out += 16;
_mm_storeu_si128(reinterpret_cast<__m128i*>(out), result_2);
out += 16;
_mm_storeu_si128(reinterpret_cast<__m128i*>(out), result_3);
out += 16;
}
}
#if defined(HAVE_BMI2_INSTRUCTIONS)
__m128i _mm_bswap_epi32(const __m128i v) {
return _mm_shuffle_epi8(v, _mm_setr_epi8(
7, 6, 5, 4,
3, 2, 1, 0,
15, 14, 13, 12,
11, 10, 9, 8
));
}
template <typename LOOKUP_FN>
void encode_bmi2(LOOKUP_FN lookup, const uint8_t* input, size_t bytes, uint8_t* output) {
uint8_t* out = output;
uint64_t lo = *reinterpret_cast<const uint64_t*>(input + 0);
uint64_t hi = *reinterpret_cast<const uint64_t*>(input + 0 + 6);
uint64_t t0 = __builtin_bswap64(lo) >> 16;
uint64_t t1 = __builtin_bswap64(hi) >> 16;
uint64_t expanded_lo = _pdep_u64(t0, 0x3f3f3f3f3f3f3f3flu);
uint64_t expanded_hi = _pdep_u64(t1, 0x3f3f3f3f3f3f3f3flu);
for (size_t i = 0; i < bytes; i += 2*6) {
#if 1
__m128i indices;
indices = _mm_insert_epi64(indices, expanded_lo, 0);
indices = _mm_insert_epi64(indices, expanded_hi, 1);
#else
const __m128i indices = _mm_set_epi64x(expanded_hi, expanded_lo);
#endif
lo = *reinterpret_cast<const uint64_t*>(input + i + 12);
hi = *reinterpret_cast<const uint64_t*>(input + i + 12 + 6);
uint64_t t0 = __builtin_bswap64(lo) >> 16;
uint64_t t1 = __builtin_bswap64(hi) >> 16;
expanded_lo = _pdep_u64(t0, 0x3f3f3f3f3f3f3f3flu);
expanded_hi = _pdep_u64(t1, 0x3f3f3f3f3f3f3f3flu);
const auto result = lookup(_mm_bswap_epi32(indices));
_mm_storeu_si128(reinterpret_cast<__m128i*>(out), result);
out += 16;
}
}
#endif
#undef packed_dword
#undef packed_byte
} // namespace sse
} // namespace base64