source: trunk/lib/bitblock256.hpp @ 3447

Last change on this file since 3447 was 3447, checked in by cameron, 4 years ago

Fixes to leave carry-out variables in carry-in form.

File size: 12.5 KB
Line 
1#ifndef BITBLOCK256_HPP_
2#define BITBLOCK256_HPP_
3
4/*=============================================================================
5    bitblock256 - Specific 256 bit implementations.
6
7    Copyright (C) 2011, Robert D. Cameron, Kenneth S. Herdy, Hua Huang and Nigel Medforth.
8    Licensed to the public under the Open Software License 3.0.
9    Licensed to International Characters Inc.
10       under the Academic Free License version 3.0.
11
12=============================================================================*/
13
14#include "idisa128.hpp"
15#include "idisa256.hpp"
16#include "builtins.hpp"
17
18union ubitblock {
19        bitblock256_t _256;
20        bitblock128_t _128[sizeof(bitblock256_t)/sizeof(bitblock256_t)];
21        uint64_t _64[sizeof(bitblock256_t)/sizeof(uint64_t)];
22        uint32_t _32[sizeof(bitblock256_t)/sizeof(uint32_t)];
23        uint16_t _16[sizeof(bitblock256_t)/sizeof(uint16_t)];
24        uint8_t _8[sizeof(bitblock256_t)/sizeof(uint8_t)];
25};
26
27/* The type used to store a carry bit. */
28typedef bitblock256_t carry_t;
29
30static IDISA_ALWAYS_INLINE void add_ci_co(bitblock256_t x, bitblock256_t y, carry_t carry_in, carry_t & carry_out, bitblock256_t & sum);
31static IDISA_ALWAYS_INLINE void sub_bi_bo(bitblock256_t x, bitblock256_t y, carry_t borrow_in, carry_t & borrow_out, bitblock256_t & difference);
32static IDISA_ALWAYS_INLINE void adv_ci_co(bitblock256_t cursor, bitblock256_t carry_in, bitblock256_t & carry_out, bitblock256_t & rslt);
33
34
35
36
37static IDISA_ALWAYS_INLINE bitblock256_t carry2bitblock(carry_t carry);
38static IDISA_ALWAYS_INLINE carry_t bitblock2carry(bitblock256_t carry);
39
40static IDISA_ALWAYS_INLINE void adc(bitblock256_t x, bitblock256_t y, carry_t & carry, bitblock256_t & sum);
41static IDISA_ALWAYS_INLINE void sbb(bitblock256_t x, bitblock256_t y, carry_t & borrow, bitblock256_t & difference);
42static IDISA_ALWAYS_INLINE void advance_with_carry(bitblock256_t cursor, carry_t & carry, bitblock256_t & rslt);
43
44static IDISA_ALWAYS_INLINE void adc(bitblock256_t x, bitblock256_t y, carry_t carry_in, carry_t & carry_out, bitblock256_t & sum);
45static IDISA_ALWAYS_INLINE void sbb(bitblock256_t x, bitblock256_t y, carry_t borrow_in, carry_t & borrow_out, bitblock256_t & difference);
46static IDISA_ALWAYS_INLINE void advance_with_carry(bitblock256_t cursor, carry_t carry_in, carry_t & carry_out, bitblock256_t & rslt);
47
48static IDISA_ALWAYS_INLINE bitblock256_t convert (uint64_t s);
49static IDISA_ALWAYS_INLINE bitblock128_t convert_128 (uint64_t s);
50static IDISA_ALWAYS_INLINE uint64_t convert (bitblock256_t v);
51
52static IDISA_ALWAYS_INLINE bitblock256_t carry2bitblock(carry_t carry) {  return carry;}
53static IDISA_ALWAYS_INLINE carry_t bitblock2carry(bitblock256_t carry) {  return carry;}
54
55static IDISA_ALWAYS_INLINE void add_ci_co(bitblock256_t x, bitblock256_t y, carry_t carry_in, carry_t & carry_out, bitblock256_t & sum) {
56  bitblock256_t all_ones = simd256<1>::constant<1>();
57  bitblock256_t gen = simd_and(x, y);
58  bitblock256_t prop = simd_xor(x, y);
59  bitblock256_t partial_sum = simd256<64>::add(x, y);
60  bitblock256_t carry = simd_or(gen, simd_andc(prop, partial_sum));
61  bitblock256_t bubble = simd256<64>::eq(partial_sum, all_ones);
62  uint64_t carry_mask = hsimd256<64>::signmask(carry) * 2 + convert(carry_in);
63  uint64_t bubble_mask = hsimd256<64>::signmask(bubble);
64  uint64_t carry_scan_thru_bubbles = (carry_mask + bubble_mask) &~ bubble_mask;
65  uint64_t increments = carry_scan_thru_bubbles | (carry_scan_thru_bubbles - carry_mask);
66  carry_out = convert(increments >> 4);
67  uint64_t spread = 0x0000200040008001 * increments & 0x0001000100010001;
68  sum = simd256<64>::add(partial_sum, _mm256_cvtepu16_epi64(avx_select_lo128(convert(spread))));
69}
70
71static IDISA_ALWAYS_INLINE void sub_bi_bo(bitblock256_t x, bitblock256_t y, carry_t borrow_in, carry_t & borrow_out, bitblock256_t & difference){
72  bitblock256_t gen = simd_andc(y, x);
73  bitblock256_t prop = simd_not(simd_xor(x, y));
74  bitblock256_t partial_diff = simd256<64>::sub(x, y);
75  bitblock256_t borrow = simd_or(gen, simd_and(prop, partial_diff));
76  bitblock256_t bubble = simd256<64>::eq(partial_diff, simd<1>::constant<0>());
77  uint64_t borrow_mask = hsimd256<64>::signmask(borrow) * 2 + convert(borrow_in);
78  uint64_t bubble_mask = hsimd256<64>::signmask(bubble);
79  uint64_t borrow_scan_thru_bubbles = (borrow_mask + bubble_mask) &~ bubble_mask;
80  uint64_t decrements = borrow_scan_thru_bubbles | (borrow_scan_thru_bubbles - borrow_mask);
81  borrow_out = convert(decrements >> 4);
82  uint64_t spread = 0x0000200040008001 * decrements & 0x0001000100010001;
83  difference = simd256<64>::sub(partial_diff, _mm256_cvtepu16_epi64(avx_select_lo128(convert(spread))));
84}
85
86static IDISA_ALWAYS_INLINE void adv_ci_co(bitblock256_t cursor, bitblock256_t carry_in, bitblock256_t & carry_out, bitblock256_t & rslt){
87        bitblock256_t shift_out = simd256<64>::srli<63>(cursor);
88        bitblock256_t low_bits = simd_or(mvmd256<64>::slli<1>(shift_out), carry_in);
89        carry_out = mvmd256<64>::srli<3>(shift_out);
90        rslt = simd_or(simd256<64>::add(cursor, cursor), low_bits);
91}
92
93IDISA_ALWAYS_INLINE bitblock256_t convert(uint64_t s)
94{
95  ubitblock b = {b._256 = simd256<128>::constant<0>()}; // = {0};
96  b._64[0] = s;
97  return b._256;
98}
99
100IDISA_ALWAYS_INLINE bitblock128_t convert_128(uint64_t s)
101{
102  ubitblock b = {b._256 = simd256<128>::constant<0>()}; // = {0};
103  b._64[0] = s;
104  return b._128[0];
105}
106
107IDISA_ALWAYS_INLINE uint64_t convert(bitblock256_t v)
108{
109  return (uint64_t) mvmd256<64>::extract<0>(v);
110}
111
112// The code below is not used.
113
114#ifdef AVX
115#define avx_select_lo128(x) \
116        ((__m128i) _mm256_castps256_ps128(x))
117
118#define avx_select_hi128(x) \
119        ((__m128i)(_mm256_extractf128_ps(x, 1)))
120
121#define avx_general_combine256(x, y) \
122   (_mm256_insertf128_ps(_mm256_castps128_ps256((__m128) y), (__m128) x, 1))
123
124IDISA_ALWAYS_INLINE void adc128(bitblock128_t x, bitblock128_t y, bitblock128_t carry_in, bitblock128_t & carry_out, bitblock128_t & sum)
125{
126        bitblock128_t gen = simd_and(x, y);
127        bitblock128_t prop = simd_or(x, y);
128        bitblock128_t partial = simd128<64>::add(simd128<64>::add(x, y), carry_in);
129        bitblock128_t c1 = simd128<128>::slli<64>(simd128<64>::srli<63>(simd_or(gen, simd_andc(prop, partial))));
130        sum = simd128<64>::add(c1, partial);
131        carry_out = simd128<128>::srli<127>(simd_or(gen, simd_andc(prop, sum)));
132}
133
134#ifndef ADCMAGIC
135IDISA_ALWAYS_INLINE void adc(bitblock256_t x, bitblock256_t y, carry_t carry_in, carry_t & carry_out, bitblock256_t & sum)
136{
137//  Really Slow!
138//        bitblock256_t gen = simd_and(x, y);
139//        bitblock256_t prop = simd_or(x, y);
140//        sum = simd256<256>::add(simd256<256>::add(x, y), carry2bitblock(carry_in));
141//        carry_out = bitblock2carry(simd256<256>::srli<255>(simd_or(gen, simd_andc(prop, sum))));
142
143  bitblock128_t x0 = avx_select_lo128(x);
144  bitblock128_t x1 = avx_select_hi128(x);
145  bitblock128_t y0 = avx_select_lo128(y);
146  bitblock128_t y1 = avx_select_hi128(y);
147  bitblock128_t c0 = avx_select_lo128(carry2bitblock(carry_in));
148  bitblock128_t s0, s1, c1, c2;
149  adc128(x0, y0, c0, c1, s0);
150  adc128(x1, y1, c1, c2, s1);
151  sum = avx_general_combine256(s1, s0);
152  carry_out = _mm256_castps128_ps256((__m128) c2);
153}
154#endif
155
156#ifdef ADCMAGIC
157
158#ifdef AVX2
159static inline void adc(bitblock256_t x, bitblock256_t y, carry_t carry_in, carry_t & carry_out, bitblock256_t & sum) {
160        bitblock256_t all_ones = simd256<1>::constant<1>();
161        bitblock256_t gen = simd_and(x, y);
162        bitblock256_t prop = simd_xor(x, y);
163  bitblock256_t partial_sum = simd256<64>::add(x, y);
164        bitblock256_t carry = simd_or(gen, simd_andc(prop, partial_sum));
165  bitblock256_t bubble = simd256<64>::eq(partial_sum, all_ones);
166  uint64_t carry_mask = hsimd256<64>::signmask(carry) * 2 + convert(carry_in);
167  uint64_t bubble_mask = hsimd256<64>::signmask(bubble);
168        uint64_t carry_scan_thru_bubbles = (carry_mask + bubble_mask) &~ bubble_mask;
169  uint64_t increments = carry_scan_thru_bubbles | (carry_scan_thru_bubbles - carry_mask);
170  carry_out = convert(increments >> 4);
171  uint64_t spread = 0x0000200040008001 * increments & 0x0001000100010001;
172  sum = simd256<64>::add(partial_sum, _mm256_cvtepu8_epi64(convert_128(spread)));
173}
174#else
175static inline void adc(bitblock256_t x, bitblock256_t y, carry_t carry_in, carry_t & carry_out, bitblock256_t & sum) {
176  bitblock128_t all_ones = simd128<1>::constant<1>();
177  //bitblock256_t gen = simd_and(x, y);
178  //bitblock256_t prop = simd_xor(x, y);
179  bitblock128_t x0 = avx_select_lo128(x);
180  bitblock128_t x1 = avx_select_hi128(x);
181  bitblock128_t y0 = avx_select_lo128(y);
182  bitblock128_t y1 = avx_select_hi128(y);
183        bitblock128_t sum0 = simd128<64>::add(x0, y0);
184        bitblock128_t sum1 = simd128<64>::add(x1, y1);
185  //bitblock256_t icarry = simd_or(gen, simd_andc(prop, avx_general_combine256(sum1, sum0)));
186        bitblock128_t icarry0 = simd_or(simd_and(x0, y0), simd_andc(simd_or(x0, y0), sum0));
187        bitblock128_t icarry1 = simd_or(simd_and(x1, y1), simd_andc(simd_or(x1, y1), sum1));
188        // A carry may bubble through a field if it is all ones.
189        bitblock128_t bubble0 = simd128<64>::eq(sum0, all_ones);
190        bitblock128_t bubble1 = simd128<64>::eq(sum1, all_ones);
191        //bitblock128_t bubble = hsimd128<64>::packss(bubble1, bubble0);
192        bitblock256_t bubble = avx_general_combine256(bubble1, bubble0);
193        //uint64_t carry_mask = _mm256_movemask_pd((__m256d) icarry) * 2 + convert(carry_in);
194        uint64_t carry_mask = hsimd128<64>::signmask(icarry1) * 8 + hsimd128<64>::signmask(icarry0) * 2 + convert(carry_in);
195        uint64_t bubble_mask = _mm256_movemask_pd((__m256d) bubble);
196        //uint64_t bubble_mask = hsimd128<64>::signmask(bubble1) * 4 + hsimd128<64>::signmask(bubble0);
197  //uint64_t bubble_mask = hsimd128<32>::signmask(bubble);
198  uint64_t carry_scan_thru_bubbles = (carry_mask + bubble_mask) &~ bubble_mask;
199        uint64_t increments = carry_scan_thru_bubbles | (carry_scan_thru_bubbles - carry_mask);
200        carry_out = convert(increments >> 4);
201        uint64_t spread = 0x0000200040008001 * increments & 0x0001000100010001;
202        bitblock128_t inc_32 = _mm_cvtepu16_epi32(_mm_cvtsi64_si128(spread));
203        bitblock128_t inc_64_0 = esimd128<32>::mergel(simd128<1>::constant<0>(), inc_32);
204        bitblock128_t inc_64_1 = esimd128<32>::mergeh(simd128<1>::constant<0>(), inc_32);
205        sum = avx_general_combine256(simd128<64>::add(sum1, inc_64_1), simd128<64>::add(sum0, inc_64_0));
206}
207#endif
208
209#endif
210
211IDISA_ALWAYS_INLINE void sbb128(bitblock128_t x, bitblock128_t y, bitblock128_t borrow_in, bitblock128_t & borrow_out, bitblock128_t & difference)
212{
213        bitblock128_t gen = simd_andc(y, x);
214        bitblock128_t prop = simd_not(simd_xor(x, y));
215        bitblock128_t partial = simd128<64>::sub(simd128<64>::sub(x, y), borrow_in);
216        bitblock128_t b1 = simd128<128>::slli<64>(simd128<64>::srli<63>(simd_or(gen, simd_and(prop, partial))));
217        difference = simd128<64>::sub(partial, b1);
218        borrow_out = simd128<128>::srli<127>(simd_or(gen, simd_and(prop, difference)));
219}
220
221
222IDISA_ALWAYS_INLINE void sbb(bitblock256_t x, bitblock256_t y, carry_t borrow_in, carry_t & borrow_out, bitblock256_t & difference)
223{
224//        bitblock256_t gen = simd_andc(y, x);
225//        bitblock256_t prop = simd_not(simd_xor(x, y));
226//        difference = simd256<256>::sub(simd256<256>::sub(x, y), carry2bitblock(borrow_in));
227//        borrow_out = bitblock2carry(simd256<256>::srli<255>(simd_or(gen, simd_and(prop, difference))));
228  bitblock128_t x0 = avx_select_lo128(x);
229  bitblock128_t x1 = avx_select_hi128(x);
230  bitblock128_t y0 = avx_select_lo128(y);
231  bitblock128_t y1 = avx_select_hi128(y);
232  bitblock128_t b0 = avx_select_lo128(carry2bitblock(borrow_in));
233  bitblock128_t d0, d1, b1, b2;
234  sbb128(x0, y0, b0, b1, d0);
235  sbb128(x1, y1, b1, b2, d1);
236  difference = avx_general_combine256(d1, d0);
237  borrow_out = _mm256_castps128_ps256((__m128) b2);
238}
239
240IDISA_ALWAYS_INLINE void advance_with_carry128(bitblock128_t cursor, bitblock128_t carry_in, bitblock128_t & carry_out, bitblock128_t & rslt)
241{
242bitblock128_t shift_out = simd128<64>::srli<63>(cursor);
243bitblock128_t low_bits = esimd128<64>::mergel(shift_out, carry_in);
244carry_out = simd128<128>::srli<64>(shift_out);
245rslt = simd_or(simd128<64>::add(cursor, cursor), low_bits);
246}
247
248IDISA_ALWAYS_INLINE void advance_with_carry(bitblock256_t cursor, carry_t carry_in, carry_t & carry_out, bitblock256_t & rslt)
249{
250  bitblock128_t cursor0 = avx_select_lo128(cursor);
251  bitblock128_t cursor1 = avx_select_hi128(cursor);
252  bitblock128_t  carry0 = avx_select_lo128(carry_in);
253  bitblock128_t  carry1, carry2, rslt0, rslt1;
254  advance_with_carry128(cursor0, carry0, carry1, rslt0);
255  advance_with_carry128(cursor1, carry1, carry2, rslt1);
256  rslt = avx_general_combine256(rslt1, rslt0);
257  carry_out = _mm256_castps128_ps256((__m128)carry2);
258}
259#endif
260
261#endif // BITBLOCK256_HPP_
Note: See TracBrowser for help on using the repository browser.