Feature request (RGB(A) to planar and back)

[Nikolai Vorontsov]

Hello Ihar,

could you please add SIMD accelearted functions to transform pixel format data (RGB, RGBA) to a plane-by-plane format (RRRRR.., GGGGG.., BBBBB.., etc)?

What I found that some of functions in your library is 6x times slower in RGB mode than in single channel. So, I assume that in triple planar format I can get 2x times speed up (not taking into account convertion to/from planar type).

Thanks in advance,
Nikolai

[Yermalayeu Ihar]

Yes, I can do it. I will implement this functionality when I have enough free time.

[Yermalayeu Ihar]

I've implemented these functions (DeinterleaveBgr, DeinterleaveBgra, InterleaveBgr, InterleaveBgra).

[Nikolai Vorontsov]

Hello Ihar,

thank you. I'll check it in a new SIMD release.

Nikolai

[Nikolai Vorontsov]

Hello Ihar,

thank you for implementing (de)interleaving functions, they are quite useful in some cases.

Is it possible to use your library somehow to implement erode/dilate functions? Something similar to Blur3x3...

Nikolai

[Yermalayeu Ihar]

Hello, Nikolai.

Yes. I can implement erode/dilate algorithms. But there are many variations of these algorithms and possible parameters (positive and negative values in the mask, use or don't use diagonal values for erode/dilate and so on). Also it would be good if the algorithms were suitable for common using. So if you give me a basic implementation of the algorithms I will add them to the library and perform platform specific optimizations.

Even better if you join to the project and add base implementation. If you agree I will add you to the list of developers of Simd LIbrary.

Sincerely, Ihar.

[Nikolai Vorontsov]

Hello Ihar,

well right now I'm using implementation taken here http://blog.ostermiller.org/dilate-and-erode

From a first glance it's possible to use SIMD to implement Manhattan oracle and a single step erode/dilate function.

Give me some time to play a bit with it, before you make it yourself :-) If I get some results, I'll toss it to you to improve SIMD lib.

Nikolai

[Nikolai Vorontsov]

Ah, got it - you also need a plain c/c++ code. OK, let me brush up a bit what I have now and then I share it with you
N.

[Yermalayeu Ihar]

Ok. I have looked the algorithms at your link. It looks like it is possible to implement with using of SIMD. I will wait your decision.

[Nikolai Vorontsov]

Hello Ihar,

sorry for delay, as usual when deadline approaches - less time for other tasks.

Is everything OK with the latest release? I've downloaded (twice) simd.3.3.48.844.zip file, but was not able to decompress it :-(
The simd.3.3.47.809.zip was downloaded and opened OK.

[Yermalayeu Ihar]

Thank you for bug report. Could you download the file again - I've just updated it?

[Nikolai Vorontsov]

it works now, thanks!

[Nikolai Vorontsov]

Ihar, I'm trying to compile it with VS2013 and it has some issues. Not 100% sure if VS2015 handles type matching better, but what I see in a new neural code is:

1>....\src\Simd\SimdAvx2Neural.cpp(111): error C2872: 'F' : ambiguous symbol
1> could be 'SIMD2\src\Simd/SimdConst.h(216) : const size_t Simd::Avx::F'
1> or 'SIMD2\src\Simd/SimdConst.h(82) : const size_t Simd::Sse::F'
1> ....\src\Simd\SimdAvx2Neural.cpp(133) : see reference to function template instantiation 'void Simd::Avx2::NeuralRoughSigmoid2<true>(const float ,size_t,const float ,float *)' being compiled

As I see NeuralRoughSigmoid2 is in Avx2 namespace and F is in Avx and Sse namespaces.
So, when it tries to compile:
namespace Avx2 ....
template <bool align=""> SIMD_INLINE void NeuralRoughSigmoid2(const float * src, size_t size, const float * slope, float * dst)
{
size_t partialAlignedSize = Simd::AlignLo(size, F);

it doesn't match well between Avx::F and Sse::F.

Does it work for you in VS2015?

[Yermalayeu Ihar]

It fine works in VS2015. I've fixed this bug for VS2013.

[Nikolai Vorontsov]

just out of curiosity - how it decides what value (4 or 8) to select???

[Yermalayeu Ihar]

See file SimdConst.h :

namespace Sse2
{
    using namespace Sse;
}

...

namespace Avx2
{
    using namespace Avx;
}

It fine works for VS2015 and GCC, but VS2013 can't resolve this situation.