source: trunk/lib/bitblock256.hpp @ 3833

Last change on this file since 3833 was 3833, checked in by ksherdy, 4 years ago

Added _bitblock to unions.

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 _bitblock;
20        bitblock256_t _256;
21        bitblock128_t _128[sizeof(bitblock256_t)/sizeof(bitblock256_t)];
22        uint64_t _64[sizeof(bitblock256_t)/sizeof(uint64_t)];
23        uint32_t _32[sizeof(bitblock256_t)/sizeof(uint32_t)];
24        uint16_t _16[sizeof(bitblock256_t)/sizeof(uint16_t)];
25        uint8_t _8[sizeof(bitblock256_t)/sizeof(uint8_t)];
26};
27
28/* The type used to store a carry bit. */
29#ifndef CARRY64
30typedef bitblock256_t carry_t;
31#endif
32#ifdef CARRY64
33typedef uint64_t carry_t;
34#endif
35
36static IDISA_ALWAYS_INLINE void add_ci_co(bitblock256_t x, bitblock256_t y, carry_t carry_in, carry_t & carry_out, bitblock256_t & sum);
37static IDISA_ALWAYS_INLINE void sub_bi_bo(bitblock256_t x, bitblock256_t y, carry_t borrow_in, carry_t & borrow_out, bitblock256_t & difference);
38static IDISA_ALWAYS_INLINE void adv_ci_co(bitblock256_t cursor, carry_t carry_in, carry_t & carry_out, bitblock256_t & rslt);
39
40
41
42
43static IDISA_ALWAYS_INLINE bitblock256_t carry2bitblock(carry_t carry);
44static IDISA_ALWAYS_INLINE carry_t bitblock2carry(bitblock256_t carry);
45static IDISA_ALWAYS_INLINE carry_t carryout2carry(bitblock256_t carryout);
46
47
48static IDISA_ALWAYS_INLINE void adc(bitblock256_t x, bitblock256_t y, carry_t & carry, bitblock256_t & sum);
49static IDISA_ALWAYS_INLINE void sbb(bitblock256_t x, bitblock256_t y, carry_t & borrow, bitblock256_t & difference);
50static IDISA_ALWAYS_INLINE void advance_with_carry(bitblock256_t cursor, carry_t & carry, bitblock256_t & rslt);
51
52static IDISA_ALWAYS_INLINE void adc(bitblock256_t x, bitblock256_t y, carry_t carry_in, carry_t & carry_out, bitblock256_t & sum);
53static IDISA_ALWAYS_INLINE void sbb(bitblock256_t x, bitblock256_t y, carry_t borrow_in, carry_t & borrow_out, bitblock256_t & difference);
54static IDISA_ALWAYS_INLINE void advance_with_carry(bitblock256_t cursor, carry_t carry_in, carry_t & carry_out, bitblock256_t & rslt);
55
56static IDISA_ALWAYS_INLINE bitblock256_t convert (uint64_t s);
57static IDISA_ALWAYS_INLINE bitblock128_t convert_128 (uint64_t s);
58static IDISA_ALWAYS_INLINE uint64_t convert (bitblock256_t v);
59
60#ifndef CARRY64
61static IDISA_ALWAYS_INLINE bitblock256_t carry2bitblock(carry_t carry) {  return carry;}
62static IDISA_ALWAYS_INLINE carry_t bitblock2carry(bitblock256_t carry) {  return carry;}
63static IDISA_ALWAYS_INLINE uint64_t carry2uint64(carry_t carry) {  return convert(carry);}
64static IDISA_ALWAYS_INLINE carry_t uint2carry(uint64_t carry) {  return convert(carry);}
65#endif
66#ifdef CARRY64
67static IDISA_ALWAYS_INLINE bitblock256_t carry2bitblock(carry_t carry) {  return convert(carry);}
68static IDISA_ALWAYS_INLINE carry_t bitblock2carry(bitblock256_t carry) {  return convert(carry);}
69static IDISA_ALWAYS_INLINE uint64_t carry2uint64(carry_t carry) {  return carry;}
70static IDISA_ALWAYS_INLINE carry_t uint2carry(uint64_t carry) {  return carry;}
71#endif
72
73static IDISA_ALWAYS_INLINE carry_t carryout2carry(carry_t carryout) {
74  return carryout;
75}
76
77static IDISA_ALWAYS_INLINE void add_ci_co(bitblock256_t x, bitblock256_t y, carry_t carry_in, carry_t & carry_out, bitblock256_t & sum) {
78  bitblock256_t all_ones = simd256<1>::constant<1>();
79  bitblock256_t gen = simd_and(x, y);
80  bitblock256_t prop = simd_xor(x, y);
81  bitblock256_t partial_sum = simd256<64>::add(x, y);
82  bitblock256_t carry = simd_or(gen, simd_andc(prop, partial_sum));
83  bitblock256_t bubble = simd256<64>::eq(partial_sum, all_ones);
84  uint64_t carry_mask = hsimd256<64>::signmask(carry) * 2 + carry2uint64(carry_in);
85  uint64_t bubble_mask = hsimd256<64>::signmask(bubble);
86  uint64_t carry_scan_thru_bubbles = (carry_mask + bubble_mask) &~ bubble_mask;
87  uint64_t increments = carry_scan_thru_bubbles | (carry_scan_thru_bubbles - carry_mask);
88  carry_out = uint2carry(increments >> 4);
89  uint64_t spread = 0x0000200040008001 * increments & 0x0001000100010001;
90  sum = simd256<64>::add(partial_sum, _mm256_cvtepu16_epi64(avx_select_lo128(convert(spread))));
91}
92
93static IDISA_ALWAYS_INLINE void sub_bi_bo(bitblock256_t x, bitblock256_t y, carry_t borrow_in, carry_t & borrow_out, bitblock256_t & difference){
94  bitblock256_t gen = simd_andc(y, x);
95  bitblock256_t prop = simd_not(simd_xor(x, y));
96  bitblock256_t partial_diff = simd256<64>::sub(x, y);
97  bitblock256_t borrow = simd_or(gen, simd_and(prop, partial_diff));
98  bitblock256_t bubble = simd256<64>::eq(partial_diff, simd<1>::constant<0>());
99  uint64_t borrow_mask = hsimd256<64>::signmask(borrow) * 2 + carry2uint64(borrow_in);
100  uint64_t bubble_mask = hsimd256<64>::signmask(bubble);
101  uint64_t borrow_scan_thru_bubbles = (borrow_mask + bubble_mask) &~ bubble_mask;
102  uint64_t decrements = borrow_scan_thru_bubbles | (borrow_scan_thru_bubbles - borrow_mask);
103  borrow_out = uint2carry(decrements >> 4);
104  uint64_t spread = 0x0000200040008001 * decrements & 0x0001000100010001;
105  difference = simd256<64>::sub(partial_diff, _mm256_cvtepu16_epi64(avx_select_lo128(convert(spread))));
106}
107
108static IDISA_ALWAYS_INLINE void adv_ci_co(bitblock256_t cursor, carry_t carry_in, carry_t & carry_out, bitblock256_t & rslt){
109        bitblock256_t shift_out = simd256<64>::srli<63>(cursor);
110        bitblock256_t low_bits = simd_or(mvmd256<64>::slli<1>(shift_out), carry2bitblock(carry_in));
111        carry_out = bitblock2carry(mvmd256<64>::srli<3>(shift_out));
112        rslt = simd_or(simd256<64>::add(cursor, cursor), low_bits);
113}
114
115IDISA_ALWAYS_INLINE bitblock256_t convert(uint64_t s)
116{ 
117  return _mm256_castsi128_si256(_mm_cvtsi64_si128(s));
118  // ubitblock b = {b._256 = simd256<128>::constant<0>()}; // = {0};
119  // b._64[0] = s;
120  // return b._256;
121}
122
123IDISA_ALWAYS_INLINE bitblock128_t convert_128(uint64_t s)
124{
125  ubitblock b = {b._256 = simd256<128>::constant<0>()}; // = {0};
126  b._64[0] = s;
127  return b._128[0];
128}
129
130IDISA_ALWAYS_INLINE uint64_t convert(bitblock256_t v)
131{
132  return (uint64_t) _mm_cvtsi128_si64(avx_select_lo128(v));
133  // return (uint64_t) mvmd256<64>::extract<0>(v);
134}
135
136// The code below is not used.
137
138#ifdef AVX
139#define avx_select_lo128(x) \
140        ((__m128i) _mm256_castps256_ps128(x))
141
142#define avx_select_hi128(x) \
143        ((__m128i)(_mm256_extractf128_ps(x, 1)))
144
145#define avx_general_combine256(x, y) \
146   (_mm256_insertf128_ps(_mm256_castps128_ps256((__m128) y), (__m128) x, 1))
147
148IDISA_ALWAYS_INLINE void adc128(bitblock128_t x, bitblock128_t y, bitblock128_t carry_in, bitblock128_t & carry_out, bitblock128_t & sum)
149{
150        bitblock128_t gen = simd_and(x, y);
151        bitblock128_t prop = simd_or(x, y);
152        bitblock128_t partial = simd128<64>::add(simd128<64>::add(x, y), carry_in);
153        bitblock128_t c1 = simd128<128>::slli<64>(simd128<64>::srli<63>(simd_or(gen, simd_andc(prop, partial))));
154        sum = simd128<64>::add(c1, partial);
155        carry_out = simd128<128>::srli<127>(simd_or(gen, simd_andc(prop, sum)));
156}
157
158#ifndef ADCMAGIC
159IDISA_ALWAYS_INLINE void adc(bitblock256_t x, bitblock256_t y, carry_t carry_in, carry_t & carry_out, bitblock256_t & sum)
160{
161//  Really Slow!
162//        bitblock256_t gen = simd_and(x, y);
163//        bitblock256_t prop = simd_or(x, y);
164//        sum = simd256<256>::add(simd256<256>::add(x, y), carry2bitblock(carry_in));
165//        carry_out = bitblock2carry(simd256<256>::srli<255>(simd_or(gen, simd_andc(prop, sum))));
166
167  bitblock128_t x0 = avx_select_lo128(x);
168  bitblock128_t x1 = avx_select_hi128(x);
169  bitblock128_t y0 = avx_select_lo128(y);
170  bitblock128_t y1 = avx_select_hi128(y);
171  bitblock128_t c0 = avx_select_lo128(carry2bitblock(carry_in));
172  bitblock128_t s0, s1, c1, c2;
173  adc128(x0, y0, c0, c1, s0);
174  adc128(x1, y1, c1, c2, s1);
175  sum = avx_general_combine256(s1, s0);
176  carry_out = _mm256_castps128_ps256((__m128) c2);
177}
178#endif
179
180#ifdef ADCMAGIC
181
182#ifdef AVX2
183static inline void adc(bitblock256_t x, bitblock256_t y, carry_t carry_in, carry_t & carry_out, bitblock256_t & sum) {
184        bitblock256_t all_ones = simd256<1>::constant<1>();
185        bitblock256_t gen = simd_and(x, y);
186        bitblock256_t prop = simd_xor(x, y);
187  bitblock256_t partial_sum = simd256<64>::add(x, y);
188        bitblock256_t carry = simd_or(gen, simd_andc(prop, partial_sum));
189  bitblock256_t bubble = simd256<64>::eq(partial_sum, all_ones);
190  uint64_t carry_mask = hsimd256<64>::signmask(carry) * 2 + convert(carry_in);
191  uint64_t bubble_mask = hsimd256<64>::signmask(bubble);
192        uint64_t carry_scan_thru_bubbles = (carry_mask + bubble_mask) &~ bubble_mask;
193  uint64_t increments = carry_scan_thru_bubbles | (carry_scan_thru_bubbles - carry_mask);
194  carry_out = convert(increments >> 4);
195  uint64_t spread = 0x0000200040008001 * increments & 0x0001000100010001;
196  sum = simd256<64>::add(partial_sum, _mm256_cvtepu8_epi64(convert_128(spread)));
197}
198#else
199static inline void adc(bitblock256_t x, bitblock256_t y, carry_t carry_in, carry_t & carry_out, bitblock256_t & sum) {
200  bitblock128_t all_ones = simd128<1>::constant<1>();
201  //bitblock256_t gen = simd_and(x, y);
202  //bitblock256_t prop = simd_xor(x, y);
203  bitblock128_t x0 = avx_select_lo128(x);
204  bitblock128_t x1 = avx_select_hi128(x);
205  bitblock128_t y0 = avx_select_lo128(y);
206  bitblock128_t y1 = avx_select_hi128(y);
207        bitblock128_t sum0 = simd128<64>::add(x0, y0);
208        bitblock128_t sum1 = simd128<64>::add(x1, y1);
209  //bitblock256_t icarry = simd_or(gen, simd_andc(prop, avx_general_combine256(sum1, sum0)));
210        bitblock128_t icarry0 = simd_or(simd_and(x0, y0), simd_andc(simd_or(x0, y0), sum0));
211        bitblock128_t icarry1 = simd_or(simd_and(x1, y1), simd_andc(simd_or(x1, y1), sum1));
212        // A carry may bubble through a field if it is all ones.
213        bitblock128_t bubble0 = simd128<64>::eq(sum0, all_ones);
214        bitblock128_t bubble1 = simd128<64>::eq(sum1, all_ones);
215        //bitblock128_t bubble = hsimd128<64>::packss(bubble1, bubble0);
216        bitblock256_t bubble = avx_general_combine256(bubble1, bubble0);
217        //uint64_t carry_mask = _mm256_movemask_pd((__m256d) icarry) * 2 + convert(carry_in);
218        uint64_t carry_mask = hsimd128<64>::signmask(icarry1) * 8 + hsimd128<64>::signmask(icarry0) * 2 + convert(carry_in);
219        uint64_t bubble_mask = _mm256_movemask_pd((__m256d) bubble);
220        //uint64_t bubble_mask = hsimd128<64>::signmask(bubble1) * 4 + hsimd128<64>::signmask(bubble0);
221  //uint64_t bubble_mask = hsimd128<32>::signmask(bubble);
222  uint64_t carry_scan_thru_bubbles = (carry_mask + bubble_mask) &~ bubble_mask;
223        uint64_t increments = carry_scan_thru_bubbles | (carry_scan_thru_bubbles - carry_mask);
224        carry_out = convert(increments >> 4);
225        uint64_t spread = 0x0000200040008001 * increments & 0x0001000100010001;
226        bitblock128_t inc_32 = _mm_cvtepu16_epi32(_mm_cvtsi64_si128(spread));
227        bitblock128_t inc_64_0 = esimd128<32>::mergel(simd128<1>::constant<0>(), inc_32);
228        bitblock128_t inc_64_1 = esimd128<32>::mergeh(simd128<1>::constant<0>(), inc_32);
229        sum = avx_general_combine256(simd128<64>::add(sum1, inc_64_1), simd128<64>::add(sum0, inc_64_0));
230}
231#endif
232
233#endif
234
235IDISA_ALWAYS_INLINE void sbb128(bitblock128_t x, bitblock128_t y, bitblock128_t borrow_in, bitblock128_t & borrow_out, bitblock128_t & difference)
236{
237        bitblock128_t gen = simd_andc(y, x);
238        bitblock128_t prop = simd_not(simd_xor(x, y));
239        bitblock128_t partial = simd128<64>::sub(simd128<64>::sub(x, y), borrow_in);
240        bitblock128_t b1 = simd128<128>::slli<64>(simd128<64>::srli<63>(simd_or(gen, simd_and(prop, partial))));
241        difference = simd128<64>::sub(partial, b1);
242        borrow_out = simd128<128>::srli<127>(simd_or(gen, simd_and(prop, difference)));
243}
244
245
246IDISA_ALWAYS_INLINE void sbb(bitblock256_t x, bitblock256_t y, carry_t borrow_in, carry_t & borrow_out, bitblock256_t & difference)
247{
248//        bitblock256_t gen = simd_andc(y, x);
249//        bitblock256_t prop = simd_not(simd_xor(x, y));
250//        difference = simd256<256>::sub(simd256<256>::sub(x, y), carry2bitblock(borrow_in));
251//        borrow_out = bitblock2carry(simd256<256>::srli<255>(simd_or(gen, simd_and(prop, difference))));
252  bitblock128_t x0 = avx_select_lo128(x);
253  bitblock128_t x1 = avx_select_hi128(x);
254  bitblock128_t y0 = avx_select_lo128(y);
255  bitblock128_t y1 = avx_select_hi128(y);
256  bitblock128_t b0 = avx_select_lo128(carry2bitblock(borrow_in));
257  bitblock128_t d0, d1, b1, b2;
258  sbb128(x0, y0, b0, b1, d0);
259  sbb128(x1, y1, b1, b2, d1);
260  difference = avx_general_combine256(d1, d0);
261  borrow_out = _mm256_castps128_ps256((__m128) b2);
262}
263
264IDISA_ALWAYS_INLINE void advance_with_carry128(bitblock128_t cursor, bitblock128_t carry_in, bitblock128_t & carry_out, bitblock128_t & rslt)
265{
266bitblock128_t shift_out = simd128<64>::srli<63>(cursor);
267bitblock128_t low_bits = esimd128<64>::mergel(shift_out, carry_in);
268carry_out = simd128<128>::srli<64>(shift_out);
269rslt = simd_or(simd128<64>::add(cursor, cursor), low_bits);
270}
271
272IDISA_ALWAYS_INLINE void advance_with_carry(bitblock256_t cursor, carry_t carry_in, carry_t & carry_out, bitblock256_t & rslt)
273{
274  bitblock128_t cursor0 = avx_select_lo128(cursor);
275  bitblock128_t cursor1 = avx_select_hi128(cursor);
276  bitblock128_t  carry0 = avx_select_lo128(carry_in);
277  bitblock128_t  carry1, carry2, rslt0, rslt1;
278  advance_with_carry128(cursor0, carry0, carry1, rslt0);
279  advance_with_carry128(cursor1, carry1, carry2, rslt1);
280  rslt = avx_general_combine256(rslt1, rslt0);
281  carry_out = _mm256_castps128_ps256((__m128)carry2);
282}
283#endif
284
285#endif // BITBLOCK256_HPP_
Note: See TracBrowser for help on using the repository browser.