Update README.md

This commit is contained in:
catid 2017-05-20 22:48:45 -07:00 committed by GitHub
parent d867f3a7da
commit b970dedf9c
1 changed files with 0 additions and 36 deletions

View File

@ -7,39 +7,3 @@ This is an attempt at implementing a fast version of the algorithm described her
IEEE Trans. on Information Theory, pp. 6284-6299, November, 2016. IEEE Trans. on Information Theory, pp. 6284-6299, November, 2016.
~~~ ~~~
Available here: [http://ct.ee.ntust.edu.tw/it2016-2.pdf](http://ct.ee.ntust.edu.tw/it2016-2.pdf) Available here: [http://ct.ee.ntust.edu.tw/it2016-2.pdf](http://ct.ee.ntust.edu.tw/it2016-2.pdf)
I wasn't able to figure out how to speed up the critical piece of code:
~~~
// z = x + y (mod Q)
static inline GFSymbol AddModQ(GFSymbol a, GFSymbol b)
{
const unsigned sum = (unsigned)a + b;
// Partial reduction step, allowing for Q to be returned
return static_cast<GFSymbol>(sum + (sum >> kGFBits));
}
// vx[] += vy[] * z
static void muladd_mem(GFSymbol * vx, const GFSymbol * vy, GFSymbol z, unsigned symbolCount)
{
for (unsigned i = 0; i < symbolCount; ++i)
{
const GFSymbol a = vy[i];
if (a == 0)
continue;
const GFSymbol sum = static_cast<GFSymbol>(AddModQ(GFLog[a], z));
vx[i] ^= GFExp[sum];
}
}
~~~
This routine basically takes all the runtime. Since it involves normal addition via `AddModQ()`, it is pretty slow and is hard to vectorize.
Math over GF(2^^16) is a lot faster because multiplies can be decomposed linearly into separate operations that can be done via vector instructions.
So unfortunately it looks like while this field does have an FFT, the field does not permit fast math.
Maybe someone else can figure out how to speed this up?