Index
> Windows > for tomasz, please replace the FractalExplorer64 example |
| Author |
|
|
tthsqe 30 Jan 2014, 04:13
The new version contains a deep version that goes up to 4096 bits of precision, though I doubt anyone has the patience to zoom that far. I was comparing two methods of squaring and it seems like 'mulx' does not give a huge benefit over the plain 'mul'. Does anyone see something wrong with the 'mulx' functions? They are in 'Sqrx.inc'. Here are the results (in iterations per second) for various precision levels:
Code: bits mul mulx mulx/mul 128 293M 302M 1.03 192 190M 201M 1.05 256 128M 144M 1.12 384 77M 87M 1.13 |
|||
30 Jan 2014, 04:13 |
|
|
Tomasz Grysztar 12 Feb 2014, 12:50
The DeepExplorer silently quits on my AVX machine, what are its requirements?
Also this new version does not have an FPU code path at all - I liked in the previous version that it demonstrates the usage of so many different instruction sets. |
|||
|
12 Feb 2014, 12:50 |
|
|
tthsqe 13 Feb 2014, 01:36
oops - the problem with the deep explorer is that I forgot to make a code switch for mul/mulx. Now it should detect if BMI2 is present and use the correct path.
Also, I added an fpu path to the shallow explorer. The reason I don't like the old version is that my register reloader introduced too many complications and made the code cryptic; this new version is very clear, has good performance, and can be followed easily by another person. |
|||
|
13 Feb 2014, 01:36 |
|
|
Tomasz Grysztar 13 Feb 2014, 16:37
Thank you for all the improvements, I have replaced the example on the official page.
|
|||
|
13 Feb 2014, 16:37 |
|
< Last Thread | Next Thread > |
Forum Rules:
|