/
sse2.cpp
323 lines (273 loc) · 11.1 KB
/
sse2.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
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
// SSE2 implementation according to http://0x80.pl/articles/sse-itoa.html
// Modifications: (1) fix incorrect digits (2) accept all ranges (3) write to user provided buffer.
#if defined(i386) || defined(__amd64) || defined(_M_IX86) || defined(_M_X64)
#include <cassert>
#include <emmintrin.h>
#include <stdint.h>
#include "digitslut.h"
#include "test.h"
#ifdef _MSC_VER
#include "intrin.h"
#endif
#ifdef _MSC_VER
#define ALIGN_PRE __declspec(align(16))
#define ALIGN_SUF
#else
#define ALIGN_PRE
#define ALIGN_SUF __attribute__ ((aligned(16)))
#endif
static const uint32_t kDiv10000 = 0xd1b71759;
ALIGN_PRE static const uint32_t kDiv10000Vector[4] ALIGN_SUF = { kDiv10000, kDiv10000, kDiv10000, kDiv10000 };
ALIGN_PRE static const uint32_t k10000Vector[4] ALIGN_SUF = { 10000, 10000, 10000, 10000 };
ALIGN_PRE static const uint16_t kDivPowersVector[8] ALIGN_SUF = { 8389, 5243, 13108, 32768, 8389, 5243, 13108, 32768 }; // 10^3, 10^2, 10^1, 10^0
ALIGN_PRE static const uint16_t kShiftPowersVector[8] ALIGN_SUF = {
1 << (16 - (23 + 2 - 16)),
1 << (16 - (19 + 2 - 16)),
1 << (16 - 1 - 2),
1 << (15),
1 << (16 - (23 + 2 - 16)),
1 << (16 - (19 + 2 - 16)),
1 << (16 - 1 - 2),
1 << (15)
};
ALIGN_PRE static const uint16_t k10Vector[8] ALIGN_SUF = { 10, 10, 10, 10, 10, 10, 10, 10 };
ALIGN_PRE static const char kAsciiZero[16] ALIGN_SUF = { '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0' };
inline __m128i Convert8DigitsSSE2(uint32_t value) {
assert(value <= 99999999);
// abcd, efgh = abcdefgh divmod 10000
const __m128i abcdefgh = _mm_cvtsi32_si128(value);
const __m128i abcd = _mm_srli_epi64(_mm_mul_epu32(abcdefgh, reinterpret_cast<const __m128i*>(kDiv10000Vector)[0]), 45);
const __m128i efgh = _mm_sub_epi32(abcdefgh, _mm_mul_epu32(abcd, reinterpret_cast<const __m128i*>(k10000Vector)[0]));
// v1 = [ abcd, efgh, 0, 0, 0, 0, 0, 0 ]
const __m128i v1 = _mm_unpacklo_epi16(abcd, efgh);
// v1a = v1 * 4 = [ abcd * 4, efgh * 4, 0, 0, 0, 0, 0, 0 ]
const __m128i v1a = _mm_slli_epi64(v1, 2);
// v2 = [ abcd * 4, abcd * 4, abcd * 4, abcd * 4, efgh * 4, efgh * 4, efgh * 4, efgh * 4 ]
const __m128i v2a = _mm_unpacklo_epi16(v1a, v1a);
const __m128i v2 = _mm_unpacklo_epi32(v2a, v2a);
// v4 = v2 div 10^3, 10^2, 10^1, 10^0 = [ a, ab, abc, abcd, e, ef, efg, efgh ]
const __m128i v3 = _mm_mulhi_epu16(v2, reinterpret_cast<const __m128i*>(kDivPowersVector)[0]);
const __m128i v4 = _mm_mulhi_epu16(v3, reinterpret_cast<const __m128i*>(kShiftPowersVector)[0]);
// v5 = v4 * 10 = [ a0, ab0, abc0, abcd0, e0, ef0, efg0, efgh0 ]
const __m128i v5 = _mm_mullo_epi16(v4, reinterpret_cast<const __m128i*>(k10Vector)[0]);
// v6 = v5 << 16 = [ 0, a0, ab0, abc0, 0, e0, ef0, efg0 ]
const __m128i v6 = _mm_slli_epi64(v5, 16);
// v7 = v4 - v6 = { a, b, c, d, e, f, g, h }
const __m128i v7 = _mm_sub_epi16(v4, v6);
return v7;
}
inline __m128i ShiftDigits_SSE2(__m128i a, unsigned digit) {
assert(digit <= 8);
switch (digit) {
case 0: return a;
case 1: return _mm_srli_si128(a, 1);
case 2: return _mm_srli_si128(a, 2);
case 3: return _mm_srli_si128(a, 3);
case 4: return _mm_srli_si128(a, 4);
case 5: return _mm_srli_si128(a, 5);
case 6: return _mm_srli_si128(a, 6);
case 7: return _mm_srli_si128(a, 7);
case 8: return _mm_srli_si128(a, 8);
}
return a; // should not execute here.
}
inline void u32toa_sse2(uint32_t value, char* buffer) {
if (value < 10000) {
const uint32_t d1 = (value / 100) << 1;
const uint32_t d2 = (value % 100) << 1;
if (value >= 1000)
*buffer++ = gDigitsLut[d1];
if (value >= 100)
*buffer++ = gDigitsLut[d1 + 1];
if (value >= 10)
*buffer++ = gDigitsLut[d2];
*buffer++ = gDigitsLut[d2 + 1];
*buffer++ = '\0';
}
else if (value < 100000000) {
// Experiment shows that this case SSE2 is slower
#if 0
const __m128i a = Convert8DigitsSSE2(value);
// Convert to bytes, add '0'
const __m128i va = _mm_add_epi8(_mm_packus_epi16(a, _mm_setzero_si128()), reinterpret_cast<const __m128i*>(kAsciiZero)[0]);
// Count number of digit
const unsigned mask = _mm_movemask_epi8(_mm_cmpeq_epi8(va, reinterpret_cast<const __m128i*>(kAsciiZero)[0]));
unsigned long digit;
#ifdef _MSC_VER
_BitScanForward(&digit, ~mask | 0x8000);
#else
digit = __builtin_ctz(~mask | 0x8000);
#endif
// Shift digits to the beginning
__m128i result = ShiftDigits_SSE2(va, digit);
//__m128i result = _mm_srl_epi64(va, _mm_cvtsi32_si128(digit * 8));
_mm_storel_epi64(reinterpret_cast<__m128i*>(buffer), result);
buffer[8 - digit] = '\0';
#else
// value = bbbbcccc
const uint32_t b = value / 10000;
const uint32_t c = value % 10000;
const uint32_t d1 = (b / 100) << 1;
const uint32_t d2 = (b % 100) << 1;
const uint32_t d3 = (c / 100) << 1;
const uint32_t d4 = (c % 100) << 1;
if (value >= 10000000)
*buffer++ = gDigitsLut[d1];
if (value >= 1000000)
*buffer++ = gDigitsLut[d1 + 1];
if (value >= 100000)
*buffer++ = gDigitsLut[d2];
*buffer++ = gDigitsLut[d2 + 1];
*buffer++ = gDigitsLut[d3];
*buffer++ = gDigitsLut[d3 + 1];
*buffer++ = gDigitsLut[d4];
*buffer++ = gDigitsLut[d4 + 1];
*buffer++ = '\0';
#endif
}
else {
// value = aabbbbbbbb in decimal
const uint32_t a = value / 100000000; // 1 to 42
value %= 100000000;
if (a >= 10) {
const unsigned i = a << 1;
*buffer++ = gDigitsLut[i];
*buffer++ = gDigitsLut[i + 1];
}
else
*buffer++ = '0' + static_cast<char>(a);
const __m128i b = Convert8DigitsSSE2(value);
const __m128i ba = _mm_add_epi8(_mm_packus_epi16(_mm_setzero_si128(), b), reinterpret_cast<const __m128i*>(kAsciiZero)[0]);
const __m128i result = _mm_srli_si128(ba, 8);
_mm_storel_epi64(reinterpret_cast<__m128i*>(buffer), result);
buffer[8] = '\0';
}
}
void i32toa_sse2(int32_t value, char* buffer) {
uint32_t u = static_cast<uint32_t>(value);
if (value < 0) {
*buffer++ = '-';
u = ~u + 1;
}
u32toa_sse2(u, buffer);
}
inline void u64toa_sse2(uint64_t value, char* buffer) {
if (value < 100000000) {
uint32_t v = static_cast<uint32_t>(value);
if (v < 10000) {
const uint32_t d1 = (v / 100) << 1;
const uint32_t d2 = (v % 100) << 1;
if (v >= 1000)
*buffer++ = gDigitsLut[d1];
if (v >= 100)
*buffer++ = gDigitsLut[d1 + 1];
if (v >= 10)
*buffer++ = gDigitsLut[d2];
*buffer++ = gDigitsLut[d2 + 1];
*buffer++ = '\0';
}
else {
// Experiment shows that this case SSE2 is slower
#if 0
const __m128i a = Convert8DigitsSSE2(v);
// Convert to bytes, add '0'
const __m128i va = _mm_add_epi8(_mm_packus_epi16(a, _mm_setzero_si128()), reinterpret_cast<const __m128i*>(kAsciiZero)[0]);
// Count number of digit
const unsigned mask = _mm_movemask_epi8(_mm_cmpeq_epi8(va, reinterpret_cast<const __m128i*>(kAsciiZero)[0]));
unsigned long digit;
#ifdef _MSC_VER
_BitScanForward(&digit, ~mask | 0x8000);
#else
digit = __builtin_ctz(~mask | 0x8000);
#endif
// Shift digits to the beginning
__m128i result = ShiftDigits_SSE2(va, digit);
_mm_storel_epi64(reinterpret_cast<__m128i*>(buffer), result);
buffer[8 - digit] = '\0';
#else
// value = bbbbcccc
const uint32_t b = v / 10000;
const uint32_t c = v % 10000;
const uint32_t d1 = (b / 100) << 1;
const uint32_t d2 = (b % 100) << 1;
const uint32_t d3 = (c / 100) << 1;
const uint32_t d4 = (c % 100) << 1;
if (value >= 10000000)
*buffer++ = gDigitsLut[d1];
if (value >= 1000000)
*buffer++ = gDigitsLut[d1 + 1];
if (value >= 100000)
*buffer++ = gDigitsLut[d2];
*buffer++ = gDigitsLut[d2 + 1];
*buffer++ = gDigitsLut[d3];
*buffer++ = gDigitsLut[d3 + 1];
*buffer++ = gDigitsLut[d4];
*buffer++ = gDigitsLut[d4 + 1];
*buffer++ = '\0';
#endif
}
}
else if (value < 10000000000000000) {
const uint32_t v0 = static_cast<uint32_t>(value / 100000000);
const uint32_t v1 = static_cast<uint32_t>(value % 100000000);
const __m128i a0 = Convert8DigitsSSE2(v0);
const __m128i a1 = Convert8DigitsSSE2(v1);
// Convert to bytes, add '0'
const __m128i va = _mm_add_epi8(_mm_packus_epi16(a0, a1), reinterpret_cast<const __m128i*>(kAsciiZero)[0]);
// Count number of digit
const unsigned mask = _mm_movemask_epi8(_mm_cmpeq_epi8(va, reinterpret_cast<const __m128i*>(kAsciiZero)[0]));
#ifdef _MSC_VER
unsigned long digit;
_BitScanForward(&digit, ~mask | 0x8000);
#else
unsigned digit = __builtin_ctz(~mask | 0x8000);
#endif
// Shift digits to the beginning
__m128i result = ShiftDigits_SSE2(va, digit);
_mm_storeu_si128(reinterpret_cast<__m128i*>(buffer), result);
buffer[16 - digit] = '\0';
}
else {
const uint32_t a = static_cast<uint32_t>(value / 10000000000000000); // 1 to 1844
value %= 10000000000000000;
if (a < 10)
*buffer++ = '0' + static_cast<char>(a);
else if (a < 100) {
const uint32_t i = a << 1;
*buffer++ = gDigitsLut[i];
*buffer++ = gDigitsLut[i + 1];
}
else if (a < 1000) {
*buffer++ = '0' + static_cast<char>(a / 100);
const uint32_t i = (a % 100) << 1;
*buffer++ = gDigitsLut[i];
*buffer++ = gDigitsLut[i + 1];
}
else {
const uint32_t i = (a / 100) << 1;
const uint32_t j = (a % 100) << 1;
*buffer++ = gDigitsLut[i];
*buffer++ = gDigitsLut[i + 1];
*buffer++ = gDigitsLut[j];
*buffer++ = gDigitsLut[j + 1];
}
const uint32_t v0 = static_cast<uint32_t>(value / 100000000);
const uint32_t v1 = static_cast<uint32_t>(value % 100000000);
const __m128i a0 = Convert8DigitsSSE2(v0);
const __m128i a1 = Convert8DigitsSSE2(v1);
// Convert to bytes, add '0'
const __m128i va = _mm_add_epi8(_mm_packus_epi16(a0, a1), reinterpret_cast<const __m128i*>(kAsciiZero)[0]);
_mm_storeu_si128(reinterpret_cast<__m128i*>(buffer), va);
buffer[16] = '\0';
}
}
void i64toa_sse2(int64_t value, char* buffer) {
uint64_t u = static_cast<uint64_t>(value);
if (value < 0) {
*buffer++ = '-';
u = ~u + 1;
}
u64toa_sse2(u, buffer);
}
REGISTER_TEST(sse2);
#endif