source: trunk/lib/bitblock256.hpp @ 3519

Last change on this file since 3519 was 3519, checked in by cameron, 6 years ago

Updates for carry_t = uint64_t support

File size: 13.5 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. */
28#ifndef CARRY64
29typedef bitblock256_t carry_t;
30#endif
31#ifdef CARRY64
32typedef uint64_t carry_t;
33#endif
34
35static IDISA_ALWAYS_INLINE void add_ci_co(bitblock256_t x, bitblock256_t y, carry_t carry_in, carry_t & carry_out, bitblock256_t & sum);
36static IDISA_ALWAYS_INLINE void sub_bi_bo(bitblock256_t x, bitblock256_t y, carry_t borrow_in, carry_t & borrow_out, bitblock256_t & difference);
37static IDISA_ALWAYS_INLINE void adv_ci_co(bitblock256_t cursor, carry_t carry_in, carry_t & carry_out, bitblock256_t & rslt);
38
39
40
41
42static IDISA_ALWAYS_INLINE bitblock256_t carry2bitblock(carry_t carry);
43static IDISA_ALWAYS_INLINE carry_t bitblock2carry(bitblock256_t carry);
44static IDISA_ALWAYS_INLINE carry_t carryout2carry(bitblock256_t carryout);
45
46
47static IDISA_ALWAYS_INLINE void adc(bitblock256_t x, bitblock256_t y, carry_t & carry, bitblock256_t & sum);
48static IDISA_ALWAYS_INLINE void sbb(bitblock256_t x, bitblock256_t y, carry_t & borrow, bitblock256_t & difference);
49static IDISA_ALWAYS_INLINE void advance_with_carry(bitblock256_t cursor, carry_t & carry, bitblock256_t & rslt);
50
51static IDISA_ALWAYS_INLINE void adc(bitblock256_t x, bitblock256_t y, carry_t carry_in, carry_t & carry_out, bitblock256_t & sum);
52static IDISA_ALWAYS_INLINE void sbb(bitblock256_t x, bitblock256_t y, carry_t borrow_in, carry_t & borrow_out, bitblock256_t & difference);
53static IDISA_ALWAYS_INLINE void advance_with_carry(bitblock256_t cursor, carry_t carry_in, carry_t & carry_out, bitblock256_t & rslt);
54
55static IDISA_ALWAYS_INLINE bitblock256_t convert (uint64_t s);
56static IDISA_ALWAYS_INLINE bitblock128_t convert_128 (uint64_t s);
57static IDISA_ALWAYS_INLINE uint64_t convert (bitblock256_t v);
58
59#ifndef CARRY64
60static IDISA_ALWAYS_INLINE bitblock256_t carry2bitblock(carry_t carry) {  return carry;}
61static IDISA_ALWAYS_INLINE carry_t bitblock2carry(bitblock256_t carry) {  return carry;}
62static IDISA_ALWAYS_INLINE uint64_t carry2uint64(carry_t carry) {  return convert(carry);}
63static IDISA_ALWAYS_INLINE carry_t uint2carry(uint64_t carry) {  return convert(carry);}
64#endif
65#ifdef CARRY64
66static IDISA_ALWAYS_INLINE bitblock256_t carry2bitblock(carry_t carry) {  return convert(carry);}
67static IDISA_ALWAYS_INLINE carry_t bitblock2carry(bitblock256_t carry) {  return convert(carry);}
68static IDISA_ALWAYS_INLINE uint64_t carry2uint64(carry_t carry) {  return carry;}
69static IDISA_ALWAYS_INLINE carry_t uint2carry(uint64_t carry) {  return carry;}
70#endif
71
72static IDISA_ALWAYS_INLINE carry_t carryout2carry(carry_t carryout) {
73  return carryout;
74}
75
76static IDISA_ALWAYS_INLINE void add_ci_co(bitblock256_t x, bitblock256_t y, carry_t carry_in, carry_t & carry_out, bitblock256_t & sum) {
77  bitblock256_t all_ones = simd256<1>::constant<1>();
78  bitblock256_t gen = simd_and(x, y);
79  bitblock256_t prop = simd_xor(x, y);
80  bitblock256_t partial_sum = simd256<64>::add(x, y);
81  bitblock256_t carry = simd_or(gen, simd_andc(prop, partial_sum));
82  bitblock256_t bubble = simd256<64>::eq(partial_sum, all_ones);
83  uint64_t carry_mask = hsimd256<64>::signmask(carry) * 2 + carry2uint64(carry_in);
84  uint64_t bubble_mask = hsimd256<64>::signmask(bubble);
85  uint64_t carry_scan_thru_bubbles = (carry_mask + bubble_mask) &~ bubble_mask;
86  uint64_t increments = carry_scan_thru_bubbles | (carry_scan_thru_bubbles - carry_mask);
87  carry_out = uint2carry(increments >> 4);
88  uint64_t spread = 0x0000200040008001 * increments & 0x0001000100010001;
89  sum = simd256<64>::add(partial_sum, _mm256_cvtepu16_epi64(avx_select_lo128(convert(spread))));
90}
91
92static IDISA_ALWAYS_INLINE void sub_bi_bo(bitblock256_t x, bitblock256_t y, carry_t borrow_in, carry_t & borrow_out, bitblock256_t & difference){
93  bitblock256_t gen = simd_andc(y, x);
94  bitblock256_t prop = simd_not(simd_xor(x, y));
95  bitblock256_t partial_diff = simd256<64>::sub(x, y);
96  bitblock256_t borrow = simd_or(gen, simd_and(prop, partial_diff));
97  bitblock256_t bubble = simd256<64>::eq(partial_diff, simd<1>::constant<0>());
98  uint64_t borrow_mask = hsimd256<64>::signmask(borrow) * 2 + carry2uint64(borrow_in);
99  uint64_t bubble_mask = hsimd256<64>::signmask(bubble);
100  uint64_t borrow_scan_thru_bubbles = (borrow_mask + bubble_mask) &~ bubble_mask;
101  uint64_t decrements = borrow_scan_thru_bubbles | (borrow_scan_thru_bubbles - borrow_mask);
102  borrow_out = uint2carry(decrements >> 4);
103  uint64_t spread = 0x0000200040008001 * decrements & 0x0001000100010001;
104  difference = simd256<64>::sub(partial_diff, _mm256_cvtepu16_epi64(avx_select_lo128(convert(spread))));
105}
106
107static IDISA_ALWAYS_INLINE void adv_ci_co(bitblock256_t cursor, carry_t carry_in, carry_t & carry_out, bitblock256_t & rslt){
108        bitblock256_t shift_out = simd256<64>::srli<63>(cursor);
109        bitblock256_t low_bits = simd_or(mvmd256<64>::slli<1>(shift_out), carry2bitblock(carry_in));
110        carry_out = bitblock2carry(mvmd256<64>::srli<3>(shift_out));
111        rslt = simd_or(simd256<64>::add(cursor, cursor), low_bits);
112}
113
114IDISA_ALWAYS_INLINE bitblock256_t convert(uint64_t s)
115{ 
116  return _mm256_castsi128_si256(_mm_cvtsi64_si128(s));
117  // ubitblock b = {b._256 = simd256<128>::constant<0>()}; // = {0};
118  // b._64[0] = s;
119  // return b._256;
120}
121
122IDISA_ALWAYS_INLINE bitblock128_t convert_128(uint64_t s)
123{
124  ubitblock b = {b._256 = simd256<128>::constant<0>()}; // = {0};
125  b._64[0] = s;
126  return b._128[0];
127}
128
129IDISA_ALWAYS_INLINE uint64_t convert(bitblock256_t v)
130{
131  return (uint64_t) _mm_cvtsi128_si64(avx_select_lo128(v));
132  // return (uint64_t) mvmd256<64>::extract<0>(v);
133}
134
135// The code below is not used.
136
137#ifdef AVX
138#define avx_select_lo128(x) \
139        ((__m128i) _mm256_castps256_ps128(x))
140
141#define avx_select_hi128(x) \
142        ((__m128i)(_mm256_extractf128_ps(x, 1)))
143
144#define avx_general_combine256(x, y) \
145   (_mm256_insertf128_ps(_mm256_castps128_ps256((__m128) y), (__m128) x, 1))
146
147IDISA_ALWAYS_INLINE void adc128(bitblock128_t x, bitblock128_t y, bitblock128_t carry_in, bitblock128_t & carry_out, bitblock128_t & sum)
148{
149        bitblock128_t gen = simd_and(x, y);
150        bitblock128_t prop = simd_or(x, y);
151        bitblock128_t partial = simd128<64>::add(simd128<64>::add(x, y), carry_in);
152        bitblock128_t c1 = simd128<128>::slli<64>(simd128<64>::srli<63>(simd_or(gen, simd_andc(prop, partial))));
153        sum = simd128<64>::add(c1, partial);
154        carry_out = simd128<128>::srli<127>(simd_or(gen, simd_andc(prop, sum)));
155}
156
157#ifndef ADCMAGIC
158IDISA_ALWAYS_INLINE void adc(bitblock256_t x, bitblock256_t y, carry_t carry_in, carry_t & carry_out, bitblock256_t & sum)
159{
160//  Really Slow!
161//        bitblock256_t gen = simd_and(x, y);
162//        bitblock256_t prop = simd_or(x, y);
163//        sum = simd256<256>::add(simd256<256>::add(x, y), carry2bitblock(carry_in));
164//        carry_out = bitblock2carry(simd256<256>::srli<255>(simd_or(gen, simd_andc(prop, sum))));
165
166  bitblock128_t x0 = avx_select_lo128(x);
167  bitblock128_t x1 = avx_select_hi128(x);
168  bitblock128_t y0 = avx_select_lo128(y);
169  bitblock128_t y1 = avx_select_hi128(y);
170  bitblock128_t c0 = avx_select_lo128(carry2bitblock(carry_in));
171  bitblock128_t s0, s1, c1, c2;
172  adc128(x0, y0, c0, c1, s0);
173  adc128(x1, y1, c1, c2, s1);
174  sum = avx_general_combine256(s1, s0);
175  carry_out = _mm256_castps128_ps256((__m128) c2);
176}
177#endif
178
179#ifdef ADCMAGIC
180
181#ifdef AVX2
182static inline void adc(bitblock256_t x, bitblock256_t y, carry_t carry_in, carry_t & carry_out, bitblock256_t & sum) {
183        bitblock256_t all_ones = simd256<1>::constant<1>();
184        bitblock256_t gen = simd_and(x, y);
185        bitblock256_t prop = simd_xor(x, y);
186  bitblock256_t partial_sum = simd256<64>::add(x, y);
187        bitblock256_t carry = simd_or(gen, simd_andc(prop, partial_sum));
188  bitblock256_t bubble = simd256<64>::eq(partial_sum, all_ones);
189  uint64_t carry_mask = hsimd256<64>::signmask(carry) * 2 + convert(carry_in);
190  uint64_t bubble_mask = hsimd256<64>::signmask(bubble);
191        uint64_t carry_scan_thru_bubbles = (carry_mask + bubble_mask) &~ bubble_mask;
192  uint64_t increments = carry_scan_thru_bubbles | (carry_scan_thru_bubbles - carry_mask);
193  carry_out = convert(increments >> 4);
194  uint64_t spread = 0x0000200040008001 * increments & 0x0001000100010001;
195  sum = simd256<64>::add(partial_sum, _mm256_cvtepu8_epi64(convert_128(spread)));
196}
197#else
198static inline void adc(bitblock256_t x, bitblock256_t y, carry_t carry_in, carry_t & carry_out, bitblock256_t & sum) {
199  bitblock128_t all_ones = simd128<1>::constant<1>();
200  //bitblock256_t gen = simd_and(x, y);
201  //bitblock256_t prop = simd_xor(x, y);
202  bitblock128_t x0 = avx_select_lo128(x);
203  bitblock128_t x1 = avx_select_hi128(x);
204  bitblock128_t y0 = avx_select_lo128(y);
205  bitblock128_t y1 = avx_select_hi128(y);
206        bitblock128_t sum0 = simd128<64>::add(x0, y0);
207        bitblock128_t sum1 = simd128<64>::add(x1, y1);
208  //bitblock256_t icarry = simd_or(gen, simd_andc(prop, avx_general_combine256(sum1, sum0)));
209        bitblock128_t icarry0 = simd_or(simd_and(x0, y0), simd_andc(simd_or(x0, y0), sum0));
210        bitblock128_t icarry1 = simd_or(simd_and(x1, y1), simd_andc(simd_or(x1, y1), sum1));
211        // A carry may bubble through a field if it is all ones.
212        bitblock128_t bubble0 = simd128<64>::eq(sum0, all_ones);
213        bitblock128_t bubble1 = simd128<64>::eq(sum1, all_ones);
214        //bitblock128_t bubble = hsimd128<64>::packss(bubble1, bubble0);
215        bitblock256_t bubble = avx_general_combine256(bubble1, bubble0);
216        //uint64_t carry_mask = _mm256_movemask_pd((__m256d) icarry) * 2 + convert(carry_in);
217        uint64_t carry_mask = hsimd128<64>::signmask(icarry1) * 8 + hsimd128<64>::signmask(icarry0) * 2 + convert(carry_in);
218        uint64_t bubble_mask = _mm256_movemask_pd((__m256d) bubble);
219        //uint64_t bubble_mask = hsimd128<64>::signmask(bubble1) * 4 + hsimd128<64>::signmask(bubble0);
220  //uint64_t bubble_mask = hsimd128<32>::signmask(bubble);
221  uint64_t carry_scan_thru_bubbles = (carry_mask + bubble_mask) &~ bubble_mask;
222        uint64_t increments = carry_scan_thru_bubbles | (carry_scan_thru_bubbles - carry_mask);
223        carry_out = convert(increments >> 4);
224        uint64_t spread = 0x0000200040008001 * increments & 0x0001000100010001;
225        bitblock128_t inc_32 = _mm_cvtepu16_epi32(_mm_cvtsi64_si128(spread));
226        bitblock128_t inc_64_0 = esimd128<32>::mergel(simd128<1>::constant<0>(), inc_32);
227        bitblock128_t inc_64_1 = esimd128<32>::mergeh(simd128<1>::constant<0>(), inc_32);
228        sum = avx_general_combine256(simd128<64>::add(sum1, inc_64_1), simd128<64>::add(sum0, inc_64_0));
229}
230#endif
231
232#endif
233
234IDISA_ALWAYS_INLINE void sbb128(bitblock128_t x, bitblock128_t y, bitblock128_t borrow_in, bitblock128_t & borrow_out, bitblock128_t & difference)
235{
236        bitblock128_t gen = simd_andc(y, x);
237        bitblock128_t prop = simd_not(simd_xor(x, y));
238        bitblock128_t partial = simd128<64>::sub(simd128<64>::sub(x, y), borrow_in);
239        bitblock128_t b1 = simd128<128>::slli<64>(simd128<64>::srli<63>(simd_or(gen, simd_and(prop, partial))));
240        difference = simd128<64>::sub(partial, b1);
241        borrow_out = simd128<128>::srli<127>(simd_or(gen, simd_and(prop, difference)));
242}
243
244
245IDISA_ALWAYS_INLINE void sbb(bitblock256_t x, bitblock256_t y, carry_t borrow_in, carry_t & borrow_out, bitblock256_t & difference)
246{
247//        bitblock256_t gen = simd_andc(y, x);
248//        bitblock256_t prop = simd_not(simd_xor(x, y));
249//        difference = simd256<256>::sub(simd256<256>::sub(x, y), carry2bitblock(borrow_in));
250//        borrow_out = bitblock2carry(simd256<256>::srli<255>(simd_or(gen, simd_and(prop, difference))));
251  bitblock128_t x0 = avx_select_lo128(x);
252  bitblock128_t x1 = avx_select_hi128(x);
253  bitblock128_t y0 = avx_select_lo128(y);
254  bitblock128_t y1 = avx_select_hi128(y);
255  bitblock128_t b0 = avx_select_lo128(carry2bitblock(borrow_in));
256  bitblock128_t d0, d1, b1, b2;
257  sbb128(x0, y0, b0, b1, d0);
258  sbb128(x1, y1, b1, b2, d1);
259  difference = avx_general_combine256(d1, d0);
260  borrow_out = _mm256_castps128_ps256((__m128) b2);
261}
262
263IDISA_ALWAYS_INLINE void advance_with_carry128(bitblock128_t cursor, bitblock128_t carry_in, bitblock128_t & carry_out, bitblock128_t & rslt)
264{
265bitblock128_t shift_out = simd128<64>::srli<63>(cursor);
266bitblock128_t low_bits = esimd128<64>::mergel(shift_out, carry_in);
267carry_out = simd128<128>::srli<64>(shift_out);
268rslt = simd_or(simd128<64>::add(cursor, cursor), low_bits);
269}
270
271IDISA_ALWAYS_INLINE void advance_with_carry(bitblock256_t cursor, carry_t carry_in, carry_t & carry_out, bitblock256_t & rslt)
272{
273  bitblock128_t cursor0 = avx_select_lo128(cursor);
274  bitblock128_t cursor1 = avx_select_hi128(cursor);
275  bitblock128_t  carry0 = avx_select_lo128(carry_in);
276  bitblock128_t  carry1, carry2, rslt0, rslt1;
277  advance_with_carry128(cursor0, carry0, carry1, rslt0);
278  advance_with_carry128(cursor1, carry1, carry2, rslt1);
279  rslt = avx_general_combine256(rslt1, rslt0);
280  carry_out = _mm256_castps128_ps256((__m128)carry2);
281}
282#endif
283
284#endif // BITBLOCK256_HPP_
Note: See TracBrowser for help on using the repository browser.