On Monday, 12 December 2016 at 22:34:32 UTC, Johan Engelen wrote:
On Monday, 12 December 2016 at 21:58:23 UTC, Relja
Ljubobratovic wrote:
Hey guys,
We have just published another post on "Writing efficient
numerical code in D", to Mir's Blog[1].
Nice :)
Thanks! :)
Some comments:
- You can make the numbers in the table more readable by adding
commas?
For example, 1624687 --> 1,624,687. Probably even better is to
only show the most significant digits. 1624687 us --> 1.625 s
for example.
Great suggestion - I was focusing on the speedup, so I didn't pay
attention to run time, but obviously other people will.
- Add a sentence about the benchmarking set up. " The numbers
shown are the best times of 10 trials, with LDC commandline
options `ldc2 -O3 -release`" Something like that. Or add it to
the configuration table.
Thanks, will do. I agree there's an explanation missing on how
benchmark is configured.
- For the "convolution" example, could you elaborate on why the
Mir-based implementation is so much faster than the C-like
implementation? From the big table, I see that the speedup for
DCV's convolution algorithms is not as large as 10x.
DCV's convolution is parallelized, which I believe, could be the
reason. Ilya is probably the best man to elaborate the speedup vs
C implementation.
- The "threshold" algorithm has a 19x speed up?! (I would
remove "threshold" from the average improvement calculation.
The numbers are already awesome, no need to oversell it!)
I agree. I was thrilled when I saw the comparison at first, but
to be honest, I'm not sure what's the main reason. Here's the
commit diff if anyone is interested to explain:
https://github.com/libmir/dcv/pull/58/commits/0c744a7bb7308f07bd45d041c6e88e906c905fcf
- Overall, does this mean that DCV became more than twice as
fast? How does it stack up to other vision libs after this
massive speed boost?
I'm not sure. But I'm afraid we're not even close to opencv,
since its base image processing (and other) algorithms have been
intensively (manually) optimized with SIMD, but also have almost
100% coverage with OpenCL (and CUDA from 3.x). Anyhow, I'd like
not to bother with comparisons like this until we are satisfied
with DCV - there's much work yet to be done, IMHO.
cheers,
Johan
Thanks for the feedback Johan!
