嗨,假设你有两个不同的独立64位二进制矩阵A
和T
(T
是它自己的转置版本,使用矩阵的转置版本允许在乘法运算T
的行而不是二进制算术超酷的列)并且你想要将这些矩阵相乘,唯一的事情是矩阵乘法结果被截断为64位,如果在某个特定矩阵单元格中产生的值大于1
,则生成的矩阵单元格将包含1
,否则0
A T
00000001 01111101
01010100 01100101
10010111 00010100
10110000 00011000 <-- This matrix is transposed
11000100 00111110
10000011 10101111
11110101 11000100
10100000 01100010
二元和传统乘法结果:
Binary Traditional
11000100 11000100
11111111 32212121
11111111 32213421
11111111 21112211
11101111 22101231
11001111 11001311
11111111 54213432
11001111 11001211
如何以最有效的方式以上述方式将这些矩阵相乘?
我试图利用二进制and
(即&
运算符)而不是在单独的位上执行乘法,在这种情况下,我必须为乘法准备数据:
ulong u;
u = T & 0xFF;
u = (u << 00) + (u << 08) + (u << 16) + (u << 24)
+ (u << 32) + (u << 40) + (u << 48) + (u << 56);
现在通过执行二进制and
超过两个整数A
和u
它将屈服于以下:
A u R C
00000001 01111101 00000001 1
01010100 01111101 01010100 3
10010111 01111101 00010101 3
10110000 01111101 00110000 2
11000100 01111101 01000100 2
10000011 01111101 00000001 1
11110101 01111101 01110101 5
10100000 01111101 00100000 1
在上面的例子中,R
包含A
位与u
位的乘法结果,并且为了获得最终值,我们必须sum
连续的所有位。请注意,列C
包含的值等于上面生成的Traditional
矩阵乘法的第一列中的值。问题是在这个步骤中我必须操作一个单独的位,我认为这是次优的方法,我通过http://graphics.stanford.edu/~seander/bithacks.html读取并行寻找方法,但没有运气,如果有人知道如何将位于R
列中的值“展平”并“合并”为生成的64位矩阵,如果你给我几行,我将不胜感激,
谢谢,
非常感谢David Eisenstat,最终算法将如下所示:
var A = ...;
var T = ...; // T == transpose(t), t is original matrix, algorithm works with transposed matrix
var D = 0x8040201008040201UL;
U = A & T; U |= U >> 1; U |= U >> 2; U |= U >> 4; U &= 0x0101010101010101UL; U = (U << 8) - U; r |= (U & D); T = (T << 8) | (T >> 56); D = (D << 8) | (D >> 56);
U = A & T; U |= U >> 1; U |= U >> 2; U |= U >> 4; U &= 0x0101010101010101UL; U = (U << 8) - U; r |= (U & D); T = (T << 8) | (T >> 56); D = (D << 8) | (D >> 56);
U = A & T; U |= U >> 1; U |= U >> 2; U |= U >> 4; U &= 0x0101010101010101UL; U = (U << 8) - U; r |= (U & D); T = (T << 8) | (T >> 56); D = (D << 8) | (D >> 56);
U = A & T; U |= U >> 1; U |= U >> 2; U |= U >> 4; U &= 0x0101010101010101UL; U = (U << 8) - U; r |= (U & D); T = (T << 8) | (T >> 56); D = (D << 8) | (D >> 56);
U = A & T; U |= U >> 1; U |= U >> 2; U |= U >> 4; U &= 0x0101010101010101UL; U = (U << 8) - U; r |= (U & D); T = (T << 8) | (T >> 56); D = (D << 8) | (D >> 56);
U = A & T; U |= U >> 1; U |= U >> 2; U |= U >> 4; U &= 0x0101010101010101UL; U = (U << 8) - U; r |= (U & D); T = (T << 8) | (T >> 56); D = (D << 8) | (D >> 56);
U = A & T; U |= U >> 1; U |= U >> 2; U |= U >> 4; U &= 0x0101010101010101UL; U = (U << 8) - U; r |= (U & D); T = (T << 8) | (T >> 56); D = (D << 8) | (D >> 56);
U = A & T; U |= U >> 1; U |= U >> 2; U |= U >> 4; U &= 0x0101010101010101UL; U = (U << 8) - U; r |= (U & D);
以下代码:
public static void Main (string[] args){
ulong U;
var Random = new Xor128 ();
var timer = DateTime.Now;
var A = Random.As<IUniformRandom<UInt64>>().Evaluate();
var T = Random.As<IUniformRandom<UInt64>>().Evaluate();
var steps = 10000000;
for (var i = 0; i < steps; i++) {
ulong r = 0;
var d = 0x8040201008040201UL;
U = A & T; U |= U >> 1; U |= U >> 2; U |= U >> 4; U &= 0x0101010101010101UL; U = (U << 8) - U; r |= (U & d); T = (T << 8) | (T >> 56); d = (d << 8) | (d >> 56);
U = A & T; U |= U >> 1; U |= U >> 2; U |= U >> 4; U &= 0x0101010101010101UL; U = (U << 8) - U; r |= (U & d); T = (T << 8) | (T >> 56); d = (d << 8) | (d >> 56);
U = A & T; U |= U >> 1; U |= U >> 2; U |= U >> 4; U &= 0x0101010101010101UL; U = (U << 8) - U; r |= (U & d); T = (T << 8) | (T >> 56); d = (d << 8) | (d >> 56);
U = A & T; U |= U >> 1; U |= U >> 2; U |= U >> 4; U &= 0x0101010101010101UL; U = (U << 8) - U; r |= (U & d); T = (T << 8) | (T >> 56); d = (d << 8) | (d >> 56);
U = A & T; U |= U >> 1; U |= U >> 2; U |= U >> 4; U &= 0x0101010101010101UL; U = (U << 8) - U; r |= (U & d); T = (T << 8) | (T >> 56); d = (d << 8) | (d >> 56);
U = A & T; U |= U >> 1; U |= U >> 2; U |= U >> 4; U &= 0x0101010101010101UL; U = (U << 8) - U; r |= (U & d); T = (T << 8) | (T >> 56); d = (d << 8) | (d >> 56);
U = A & T; U |= U >> 1; U |= U >> 2; U |= U >> 4; U &= 0x0101010101010101UL; U = (U << 8) - U; r |= (U & d); T = (T << 8) | (T >> 56); d = (d << 8) | (d >> 56);
U = A & T; U |= U >> 1; U |= U >> 2; U |= U >> 4; U &= 0x0101010101010101UL; U = (U << 8) - U; r |= (U & d);
}
Console.WriteLine (DateTime.Now - timer);
var m1 = new Int32[8,8];
var m2 = new Int32[8,8];
var m3 = new Int32[8,8];
for (int row = 0; row < 8; row++) {
for (int col = 0; col < 8; col++) {
m1 [row, col] = Random.As<IUniformRandom<Int32>> ().Evaluate(0, 1);
m2 [row, col] = Random.As<IUniformRandom<Int32>> ().Evaluate(0, 1);
m3 [row, col] = Random.As<IUniformRandom<Int32>> ().Evaluate(0, 1);
}
}
timer = DateTime.Now;
for (int i = 0; i < steps; i++) {
for (int row = 0; row < 8; row++) {
for (int col = 0; col < 8; col++) {
var sum = 0;
for (int temp = 0; temp < 8; temp++) {
sum += m1 [row, temp] * m2 [temp, row];
}
m3 [row, col] = sum;
}
}
}
Console.WriteLine (DateTime.Now - timer);
}
向我展示以下结果:
00:00:02.4035870
00:00:57.5147150
感谢大家,Mac OS X / Mono的性能提升了23倍
我不确定最有效率,但这里有一些尝试。以下指令序列计算乘积A * T'的主对角线。将T和D旋转8位并重复7次迭代。
// uint64_t A, T;
uint64_t D = UINT64_C(0x8040201008040201);
uint64_t P = A & T;
// test whether each byte is nonzero
P |= P >> 1;
P |= P >> 2;
P |= P >> 4;
P &= UINT64_C(0x0101010101010101);
// fill each nonzero byte with ones
P *= 255; // or P = (P << 8) - P;
// leave only the current diagonal
P &= D;
如果您正在寻找一种并行执行密集矩阵乘法的方法,请将结果矩阵划分为块并并行计算每个块。
http://en.wikipedia.org/wiki/Block_matrix#Block_matrix_multiplication
目前尚不清楚您使用的是哪种数据结构,哪种语言(是的,我知道您说的是“任何语言”),以及您要优化的内容(速度?记忆?)等等所有这些都可能对您产生深远的影响解。
一些例子:
|
)而不是+
。有些语言可能会懒惰地对此进行评估,在遇到的第一个“1”处停止。顺便说一下,我假设你有很多矩阵需要处理,否则我会使用直接的,可读的代码。我的猜测是,即使有很多矩阵,性能的提升也可以忽略不计。
如果你允许C / C ++的更低层次结构,那么SSE / AVX机器指令和内在编译器功能允许编写更快的代码(根据我制作的一些基准测试4x)。您需要使用非标准向量变量(至少由GCC,ICC,CLang支持):
using epu = uint8_t __attribute__((vector_size(16)));
我正在使用像
class BMat8 {
[...]
private:
uint64_t _data;
};
那么,以下代码应该做你想要的
static constexpr epu rothigh { 0, 1, 2, 3, 4, 5, 6, 7,15, 8, 9,10,11,12,13,14};
static constexpr epu rot2 { 6, 7, 0, 1, 2, 3, 4, 5,14,15, 8, 9,10,11,12,13};
inline BMat8 operator*(BMat8 const& tr) const {
epu x = _mm_set_epi64x(_data, _data);
epu y = _mm_shuffle_epi8(_mm_set_epi64x(tr._data, tr._data), rothigh);
epu data {};
epu diag = {0x01,0x02,0x04,0x08,0x10,0x20,0x40,0x80,
0x80,0x01,0x02,0x04,0x08,0x10,0x20,0x40};
for (int i = 0; i < 4; ++i) {
data |= ((x & y) != epu {}) & diag;
y = _mm_shuffle_epi8(y, rot2);
diag = _mm_shuffle_epi8(diag, rot2);
}
return BMat8(_mm_extract_epi64(data, 0) | _mm_extract_epi64(data, 1));
}
特别是,使用128位寄存器,我可以一次完成两次迭代。
使用我在此描述的解决方案,可以在x86-64上非常有效地实现严格布尔代数的解决方案:
https://stackoverflow.com/a/55307540/11147804
唯一的区别是来自转置矩阵的数据也需要通过列提取并在每个64位产品之前重新打包到行。幸运的是,使用BMI2指令进行并行位提取是很容易的,可以在GCC上使用内部_pext_u64访问:
uint64_t mul8x8T (uint64_t A, uint64_t B) {
const uint64_t COL = 0x0101010101010101;
uint64_t C = 0;
for (int i=0; i<8; ++i) {
uint64_t p = COL & (A>>i); // select column
uint64_t r = torow( COL & (B>>i) );
C |= (p*r); // use ^ for GF(2) instead
}
return C;
}
uint64_t torow (uint64_t c) {
const uint64_t ROW = 0x00000000000000FF; // mask of the first row
const uint64_t COL = 0x0101010101010101; // mask of the first column
// select bits of c in positions marked by COL,
// and pack them consecutively
// last 'and' is included for clarity and is not
// really necessary
return _pext_u64(c, COL) & ROW;
}
在不支持该特定指令的处理器中,一种可能的解决方案是调整典型的位技巧以进行打包,例如在使用64位乘法的字节的经典位顺序反转中使用:
https://graphics.stanford.edu/~seander/bithacks.html#ReverseByteWith64BitsDiv
使用掩码和带有一些常量的整数乘法导致包含打包结果的四字作为位子串,然后可以使用位移和掩码提取。
我们的想法是将乘法步骤视为并行位移,其中输入中的每个位移位不同的量,在常量中指定。只要两个数字的步幅不在结果中的某个位置上发生碰撞,即只要来自乘法的每个部分和更新结果中的不同位位置,这总是可能的。这避免了任何潜在的进位,这使得逐位和相当于位并行OR(或XOR)。
uint64_t torow (uint64_t c) {
const uint64_t ROW = 0x00000000000000FF; // select 8 lowest consecutive bits to get the first row
const uint64_t COL = 0x0101010101010101; // select every 8th bit to get the first column
const uint64_t DIA = 0x8040201008040201; // select every 8+1 bit to obtain a diagonal
c *= ROW; // "copies" first column to the rest
c &= DIA; // only use diagonal bits or else there will be position collisions and unexpected carries
c *= COL; // "scatters" every bit to all rows after itself; the last row will now contain the packed bits
return c >> 56; // move last row to first & discard the rest
}
此功能还有其他可能的替代实现,使用更多低强度的操作,其中最快的操作将取决于目标体系结构。