source: trunk/lib/bitblock256.hpp @ 3370

Last change on this file since 3370 was 2719, checked in by cameron, 7 years ago

Some copyright and description fixes

File size: 8.6 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        bitblock256_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 bitblock256_t carry2bitblock(carry_t carry);
31static IDISA_ALWAYS_INLINE carry_t bitblock2carry(bitblock256_t carry);
32
33static IDISA_ALWAYS_INLINE void adc(bitblock256_t x, bitblock256_t y, carry_t & carry, bitblock256_t & sum);
34static IDISA_ALWAYS_INLINE void sbb(bitblock256_t x, bitblock256_t y, carry_t & borrow, bitblock256_t & difference);
35static IDISA_ALWAYS_INLINE void advance_with_carry(bitblock256_t cursor, carry_t & carry, bitblock256_t & rslt);
36
37static IDISA_ALWAYS_INLINE void adc(bitblock256_t x, bitblock256_t y, carry_t carry_in, carry_t & carry_out, bitblock256_t & sum);
38static IDISA_ALWAYS_INLINE void sbb(bitblock256_t x, bitblock256_t y, carry_t borrow_in, carry_t & borrow_out, bitblock256_t & difference);
39static IDISA_ALWAYS_INLINE void advance_with_carry(bitblock256_t cursor, carry_t carry_in, carry_t & carry_out, bitblock256_t & rslt);
40
41static IDISA_ALWAYS_INLINE bitblock256_t convert (uint64_t s);
42static IDISA_ALWAYS_INLINE uint64_t convert (bitblock256_t v);
43
44static IDISA_ALWAYS_INLINE bitblock256_t carry2bitblock(carry_t carry) {  return carry;}
45static IDISA_ALWAYS_INLINE carry_t bitblock2carry(bitblock256_t carry) {  return carry;}
46
47
48
49#define avx_select_lo128(x) \
50        ((__m128i) _mm256_castps256_ps128(x))
51
52#define avx_select_hi128(x) \
53        ((__m128i)(_mm256_extractf128_ps(x, 1)))
54
55#define avx_general_combine256(x, y) \
56   (_mm256_insertf128_ps(_mm256_castps128_ps256((__m128) y), (__m128) x, 1))
57
58IDISA_ALWAYS_INLINE void adc128(bitblock128_t x, bitblock128_t y, bitblock128_t carry_in, bitblock128_t & carry_out, bitblock128_t & sum)
59{
60        bitblock128_t gen = simd_and(x, y);
61        bitblock128_t prop = simd_or(x, y);
62        bitblock128_t partial = simd128<64>::add(simd128<64>::add(x, y), carry_in);
63        bitblock128_t c1 = simd128<128>::slli<64>(simd128<64>::srli<63>(simd_or(gen, simd_andc(prop, partial))));
64        sum = simd128<64>::add(c1, partial);
65        carry_out = simd128<128>::srli<127>(simd_or(gen, simd_andc(prop, sum)));
66}
67
68#ifndef ADCMAGIC
69IDISA_ALWAYS_INLINE void adc(bitblock256_t x, bitblock256_t y, carry_t carry_in, carry_t & carry_out, bitblock256_t & sum)
70{
71//  Really Slow!
72//        bitblock256_t gen = simd_and(x, y);
73//        bitblock256_t prop = simd_or(x, y);
74//        sum = simd256<256>::add(simd256<256>::add(x, y), carry2bitblock(carry_in));
75//        carry_out = bitblock2carry(simd256<256>::srli<255>(simd_or(gen, simd_andc(prop, sum))));
76
77  bitblock128_t x0 = avx_select_lo128(x);
78  bitblock128_t x1 = avx_select_hi128(x);
79  bitblock128_t y0 = avx_select_lo128(y);
80  bitblock128_t y1 = avx_select_hi128(y);
81  bitblock128_t c0 = avx_select_lo128(carry2bitblock(carry_in));
82  bitblock128_t s0, s1, c1, c2;
83  adc128(x0, y0, c0, c1, s0);
84  adc128(x1, y1, c1, c2, s1);
85  sum = avx_general_combine256(s1, s0);
86  carry_out = _mm256_castps128_ps256((__m128) c2);
87}
88#endif
89
90#ifdef ADCMAGIC
91static inline void adc(bitblock256_t x, bitblock256_t y, carry_t carry_in, carry_t & carry_out, bitblock256_t & sum) {
92        bitblock128_t all_ones = simd128<1>::constant<1>();
93        //bitblock256_t gen = simd_and(x, y);
94        //bitblock256_t prop = simd_xor(x, y);
95        bitblock128_t x0 = avx_select_lo128(x);
96        bitblock128_t x1 = avx_select_hi128(x);
97        bitblock128_t y0 = avx_select_lo128(y);
98        bitblock128_t y1 = avx_select_hi128(y);
99        bitblock128_t sum0 = simd128<64>::add(x0, y0);
100        bitblock128_t sum1 = simd128<64>::add(x1, y1);
101        //bitblock256_t icarry = simd_or(gen, simd_andc(prop, avx_general_combine256(sum1, sum0)));
102        bitblock128_t icarry0 = simd_or(simd_and(x0, y0), simd_andc(simd_or(x0, y0), sum0));
103        bitblock128_t icarry1 = simd_or(simd_and(x1, y1), simd_andc(simd_or(x1, y1), sum1));
104        // A carry may bubble through a field if it is all ones.
105        bitblock128_t bubble0 = simd128<64>::eq(sum0, all_ones);
106        bitblock128_t bubble1 = simd128<64>::eq(sum1, all_ones);
107        //bitblock128_t bubble = hsimd128<64>::packss(bubble1, bubble0);
108        bitblock256_t bubble = avx_general_combine256(bubble1, bubble0);
109        //uint64_t carry_mask = _mm256_movemask_pd((__m256d) icarry) * 2 + convert(carry_in);
110        uint64_t carry_mask = hsimd128<64>::signmask(icarry1) * 8 + hsimd128<64>::signmask(icarry0) * 2 + convert(carry_in);
111        uint64_t bubble_mask = _mm256_movemask_pd((__m256d) bubble);
112        //uint64_t bubble_mask = hsimd128<64>::signmask(bubble1) * 4 + hsimd128<64>::signmask(bubble0);
113        //uint64_t bubble_mask = hsimd128<32>::signmask(bubble);
114        uint64_t carry_scan_thru_bubbles = (carry_mask + bubble_mask) &~ bubble_mask;
115        uint64_t increments = carry_scan_thru_bubbles | (carry_scan_thru_bubbles - carry_mask);
116        carry_out = convert(increments >> 4);
117        uint64_t spread = 0x0000200040008001 * increments & 0x0001000100010001;
118        bitblock128_t inc_32 = _mm_cvtepu16_epi32(_mm_cvtsi64_si128(spread));
119        bitblock128_t inc_64_0 = esimd128<32>::mergel(simd128<1>::constant<0>(), inc_32);
120        bitblock128_t inc_64_1 = esimd128<32>::mergeh(simd128<1>::constant<0>(), inc_32);
121        sum = avx_general_combine256(simd128<64>::add(sum1, inc_64_1), simd128<64>::add(sum0, inc_64_0));
122}
123#endif
124
125
126
127
128
129IDISA_ALWAYS_INLINE void sbb128(bitblock128_t x, bitblock128_t y, bitblock128_t borrow_in, bitblock128_t & borrow_out, bitblock128_t & difference)
130{
131        bitblock128_t gen = simd_andc(y, x);
132        bitblock128_t prop = simd_not(simd_xor(x, y));
133        bitblock128_t partial = simd128<64>::sub(simd128<64>::sub(x, y), borrow_in);
134        bitblock128_t b1 = simd128<128>::slli<64>(simd128<64>::srli<63>(simd_or(gen, simd_and(prop, partial))));
135        difference = simd128<64>::sub(partial, b1);
136        borrow_out = simd128<128>::srli<127>(simd_or(gen, simd_and(prop, difference)));
137}
138
139
140IDISA_ALWAYS_INLINE void sbb(bitblock256_t x, bitblock256_t y, carry_t borrow_in, carry_t & borrow_out, bitblock256_t & difference)
141{
142//        bitblock256_t gen = simd_andc(y, x);
143//        bitblock256_t prop = simd_not(simd_xor(x, y));
144//        difference = simd256<256>::sub(simd256<256>::sub(x, y), carry2bitblock(borrow_in));
145//        borrow_out = bitblock2carry(simd256<256>::srli<255>(simd_or(gen, simd_and(prop, difference))));
146  bitblock128_t x0 = avx_select_lo128(x);
147  bitblock128_t x1 = avx_select_hi128(x);
148  bitblock128_t y0 = avx_select_lo128(y);
149  bitblock128_t y1 = avx_select_hi128(y);
150  bitblock128_t b0 = avx_select_lo128(carry2bitblock(borrow_in));
151  bitblock128_t d0, d1, b1, b2;
152  sbb128(x0, y0, b0, b1, d0);
153  sbb128(x1, y1, b1, b2, d1);
154  difference = avx_general_combine256(d1, d0);
155  borrow_out = _mm256_castps128_ps256((__m128) b2);
156}
157
158IDISA_ALWAYS_INLINE void advance_with_carry128(bitblock128_t cursor, bitblock128_t carry_in, bitblock128_t & carry_out, bitblock128_t & rslt)
159{
160bitblock128_t shift_out = simd128<64>::srli<63>(cursor);
161bitblock128_t low_bits = esimd128<64>::mergel(shift_out, carry_in);
162carry_out = simd128<128>::srli<64>(shift_out);
163rslt = simd_or(simd128<64>::add(cursor, cursor), low_bits);
164}
165
166IDISA_ALWAYS_INLINE void advance_with_carry(bitblock256_t cursor, carry_t carry_in, carry_t & carry_out, bitblock256_t & rslt)
167{
168  bitblock128_t cursor0 = avx_select_lo128(cursor);
169  bitblock128_t cursor1 = avx_select_hi128(cursor);
170  bitblock128_t  carry0 = avx_select_lo128(carry_in);
171  bitblock128_t  carry1, carry2, rslt0, rslt1;
172  advance_with_carry128(cursor0, carry0, carry1, rslt0);
173  advance_with_carry128(cursor1, carry1, carry2, rslt1);
174  rslt = avx_general_combine256(rslt1, rslt0);
175  carry_out = _mm256_castps128_ps256((__m128)carry2);
176}
177
178IDISA_ALWAYS_INLINE bitblock256_t convert(uint64_t s)
179{
180        ubitblock b = {b._256 = simd256<128>::constant<0>()}; // = {0};
181        b._64[0] = s;
182        return b._256;
183}
184
185IDISA_ALWAYS_INLINE uint64_t convert (bitblock256_t v)
186{
187        return (uint64_t) mvmd256<64>::extract<0>(v);
188}
189
190
191#endif // BITBLOCK256_HPP_
Note: See TracBrowser for help on using the repository browser.