-
Notifications
You must be signed in to change notification settings - Fork 0
/
decode_ans.cpp
240 lines (196 loc) · 9.39 KB
/
decode_ans.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
#include <cstdint>
#include "immintrin.h"
#include "/home/christoph/intel/iaca/include/iacaMarks.h"
__m512i _mm512_mulhi_epi32(__m512i a, __m512i b) {
auto a_lo = _mm512_srli_epi64(a, 32);
auto b_lo = _mm512_srli_epi64(b, 32);
auto c_lo = _mm512_mul_epi32(a_lo, b_lo);
auto c_hi = _mm512_mul_epi32(a , b );
//return _mm512_mask_blend_epi32(0x5555, c_lo, _mm512_srli_epi64(c_hi, 32));
return _mm512_mask_alignr_epi32(c_lo, 0x5555, c_hi, c_hi, 1);
}
__m512 div22_by_lt32(__m512 a, __m512i b) {
// reciprocals: ceiled to next float
const __m512 t1 = _mm512_set_ps(
0.032258067280, 0.033333335072, 0.034482762218, 0.035714287311, 0.037037037313, 0.038461539894, 0.040000002831, 0.041666667908,
0.043478261679, 0.045454546809, 0.047619048506, 0.050000000745, 0.052631579340, 0.055555555969, 0.058823529631, 0.062500000000
);
const __m512 t2 = _mm512_set_ps(
0.066666670144, 0.071428574622, 0.076923079789, 0.083333335817, 0.090909093618, 0.100000001490, 0.111111111939, 0.125000000000,
0.142857149243, 0.166666671634, 0.200000002980, 0.250000000000, 0.333333343267, 0.500000000000, 1.000000000000, 1.000000000000
);
// vpermi2ps to destroy b
auto rcp = _mm512_permutex2var_ps(t2, b, t1);
// divide
auto c = _mm512_fmadd_round_ps(a, rcp, _mm512_set1_ps(1<<23), _MM_FROUND_TO_ZERO | _MM_FROUND_NO_EXC);
return _mm512_sub_ps(c, _mm512_set1_ps(1<<23)); // floor c
}
__m512i div22_by_lt32(__m512i a, __m512i b) {
// reciprocals: ceiled to next float
const __m512i t1 = _mm512_set_epi32(
0x08421085, 0x08888889, 0x08d3dcb1, 0x0924924a, 0x097b425f, 0x09d89d8a, 0x0a3d70a4, 0x0aaaaaab,
0x0b21642d, 0x0ba2e8bb, 0x0c30c30d, 0x0ccccccd, 0x0d79435f, 0x0e38e38f, 0x0f0f0f10, 0x10000000
);
const __m512i t2 = _mm512_set_epi32(
0x11111112, 0x12492493, 0x13b13b14, 0x15555556, 0x1745d175, 0x1999999a, 0x1c71c71d, 0x20000000,
0x24924925, 0x2aaaaaab, 0x33333334, 0x40000000, 0x55555556, 0x80000000, 0xFFFFFFFF, 0xFFFFFFFF
);
// vpermi2ps to destroy b
auto rcp = _mm512_permutex2var_epi32(t2, b, t1);
// divide
return _mm512_mulhi_epi32(a, rcp); // floor c
}
template<class T>
inline T* _addr(T* base, std::size_t byte_step) {
return reinterpret_cast<T*>(
reinterpret_cast<std::uint8_t*>(base) + byte_step
);
}
// constants
constexpr std::size_t m512i_size = sizeof(__m512i);
constexpr std::size_t small_bits = 23;
constexpr std::size_t small_shft = 8;
constexpr std::size_t large_bits = 52;
constexpr std::size_t large_shft = 32;
constexpr std::size_t small_smin = 1 << (small_bits - small_shft);
constexpr std::size_t large_smin = 1 << (large_bits - large_shft);
inline __m512i decode_ans(
__m512& state_sml0, // current state, the rolling is done in the calling loop
__m512i base_int32, // base to decode with
std::uint8_t*& stream_ptr // where to read bits
) {
// check if there are small codings
auto small_mask = _mm512_cmpgt_epi32_mask(_mm512_set1_epi32(32), base_int32);
__m512i digit_int32;
if (!_mm512_kortestc(small_mask, small_mask)) {
// @todo Add fallback for large bases
_mm512_storeu_si512(reinterpret_cast<__m512i*>(stream_ptr), base_int32); // just to keep the branch
}
// convert to float (should this use the mask?)
auto base_float = _mm512_cvtepi32_ps(base_int32);
// decode the symbol
auto state_sml1 = div22_by_lt32(state_sml0, base_int32);
auto digit = _mm512_fnmadd_ps(state_sml1, base_float, state_sml0); // digit = s0 % b = s0 - s1 * b
// test for underflow NOTE: we could compare to c in div22_by_lt32 in order to skip latency of the sub
auto uflow_mask = _mm512_cmp_ps_mask(_mm512_set1_ps(small_smin), state_sml1, _CMP_GT_OQ);
// load bytes
auto stream_in = _mm_loadu_si128(reinterpret_cast<__m128i*>(stream_ptr));
auto stream_int = _mm512_maskz_expand_epi32(uflow_mask, _mm512_cvtepi8_epi32(stream_in));
auto stream_flt = _mm512_cvtepi32_ps(stream_int);
stream_ptr += _mm_popcnt_u64(uflow_mask);
// rescale and update state
state_sml0 = _mm512_mask_fmadd_ps(state_sml1, uflow_mask, _mm512_set1_ps(256), stream_flt);
// convert and return
return _mm512_mask_cvtps_epi32(digit_int32, small_mask, digit);
}
inline __m512i decode_ans_div(
__m512& state_sml0, // current state, the rolling is done in the calling loop
__m512i base_int32, // base to decode with
std::uint8_t*& stream_ptr // where to read bits
) {
// check if there are small codings
auto small_mask = _mm512_cmpgt_epi32_mask(_mm512_set1_epi32(32), base_int32);
__m512i digit_int32;
if (!_mm512_kortestc(small_mask, small_mask)) {
// @todo Add fallback for large bases
_mm512_storeu_si512(reinterpret_cast<__m512i*>(stream_ptr), base_int32); // just to keep the branch
}
// convert to float (should this use the mask?)
auto base_float = _mm512_cvtepi32_ps(base_int32);
// decode the symbol
auto state_sml1 = _mm512_roundscale_ps(_mm512_div_ps(state_sml0, base_float), 1);
auto digit = _mm512_fnmadd_ps(state_sml1, base_float, state_sml0); // digit = s0 % b = s0 - s1 * b
// test for underflow NOTE: we could compare to c in div22_by_lt32 in order to skip latency of the sub
auto uflow_mask = _mm512_cmp_ps_mask(_mm512_set1_ps(small_smin), state_sml1, _CMP_GT_OQ);
// load bytes
auto stream_in = _mm_loadu_si128(reinterpret_cast<__m128i*>(stream_ptr));
auto stream_int = _mm512_maskz_expand_epi32(uflow_mask, _mm512_cvtepi8_epi32(stream_in));
auto stream_flt = _mm512_cvtepi32_ps(stream_int);
stream_ptr += _mm_popcnt_u64(uflow_mask);
// rescale and update state
state_sml0 = _mm512_mask_fmadd_ps(state_sml1, uflow_mask, _mm512_set1_ps(256), stream_flt);
// convert and return
return _mm512_mask_cvtps_epi32(digit_int32, small_mask, digit);
}
inline __m256i decode_ans_big(
__m512d& state_lrg0, // current state, the rolling is done in the calling loop
__mmask8 mask,
__m256i base_int32, // base to decode with
std::uint8_t*& stream_ptr // where to read bits
) {
// convert to float (should this use the mask?)
auto base_float = _mm512_cvtepi32_pd(base_int32);
// decode the symbol
auto state_lrg1 = _mm512_roundscale_pd(_mm512_div_pd(state_lrg0, base_float), 1);
auto digit = _mm512_fnmadd_pd(state_lrg1, base_float, state_lrg0); // digit = s0 % b = s0 - s1 * b
// test for underflow
auto uflow_mask = _mm512_mask_cmp_pd_mask(mask, _mm512_set1_pd(small_smin), state_lrg1, _CMP_GT_OQ);
// load bytes
auto stream_int = _mm256_maskz_expandloadu_epi32(uflow_mask, stream_ptr);
auto stream_flt = _mm512_cvtepi32_pd(stream_int);
stream_ptr = _addr(stream_ptr, _mm_popcnt_u64(uflow_mask) * sizeof(uint32_t));
// rescale and update state
state_lrg0 = _mm512_mask_fmadd_pd(state_lrg1, uflow_mask, _mm512_set1_pd(UINT64_C(1) << large_shft), stream_flt);
// convert and return
return _mm512_cvtpd_epi32(digit);
}
inline __m512i decode_ans_int(
__m512i& state_sml0, // current state, the rolling is done in the calling loop
__m512i base_int32, // base to decode with
std::uint8_t*& stream_ptr // where to read bits
) {
// check if there are small codings
auto small_mask = _mm512_cmp_epi32_mask(_mm512_set1_epi32(31), base_int32, 2);
__m512i digit_int32;
if (!_mm512_kortestc(small_mask, small_mask)) {
// @todo change parameter layout and code this correctly
// decode hi
//decode_ans_big();
// decode lo
//decode_ans_big();
}
// decode the symbol
auto state_sml1 = div22_by_lt32(state_sml0, base_int32);
auto digit = _mm512_sub_epi32(state_sml0, _mm512_mullo_epi32(state_sml1, base_int32)); // digit = s0 % b = s0 - s1 * b
// test for underflow NOTE: we could compare to c in div22_by_lt32 in order to skip latency of the sub
auto uflow_mask = _mm512_mask_cmpgt_epi32_mask(small_mask, _mm512_set1_epi32(small_smin), state_sml1);
// load bytes
auto stream_in = _mm_loadu_si128(reinterpret_cast<__m128i*>(stream_ptr));
auto stream_int = _mm512_maskz_expand_epi32(uflow_mask, _mm512_cvtepi8_epi32(stream_in));
stream_ptr += _mm_popcnt_u64(uflow_mask);
// shift left and add
state_sml0 = _mm512_mask_slli_epi32(state_sml1, uflow_mask, state_sml1, 8);
state_sml0 = _mm512_add_epi32(state_sml0, stream_int);
// add and return
return digit;
}
void minimal_loop(__m512i* base, std::uint8_t* stream_ptr) {
auto state_sml0 = _mm512_set1_ps(small_smin); // actually loaded from stream_ptr
auto state_sml1 = _mm512_set1_ps(small_smin); // actually loaded from stream_ptr
auto state_sml2 = _mm512_set1_ps(small_smin); // actually loaded from stream_ptr
for (__m512i* be = base + 1024; base < be; base++) {
//IACA_START
*base = decode_ans(state_sml0, *base, stream_ptr);
//*base = decode_ans_div(state_sml0, *base, stream_ptr);
// auto tmp = state_sml0;
// state_sml0 = state_sml1;
// state_sml1 = state_sml2;
// state_sml2 = tmp;
}
//IACA_END
}
void minimal_loop_int(__m512i* base, std::uint8_t* stream_ptr) {
auto state_sml0 = _mm512_set1_epi32(small_smin); // actually loaded from stream_ptr
auto state_sml1 = _mm512_set1_epi32(small_smin); // actually loaded from stream_ptr
auto state_sml2 = _mm512_set1_epi32(small_smin); // actually loaded from stream_ptr
for (__m512i* be = base + 1024; base < be; base++) {
// IACA_START
*base = decode_ans_int(state_sml0, *base, stream_ptr);
// auto tmp = state_sml0;
// state_sml0 = state_sml1;
// state_sml1 = state_sml2;
// state_sml2 = tmp;
}
// IACA_END
}
int main() {}