[Martijn]

the port is a bit on hold as we first need to adjust our build system. No point in doing work that has to be redone again.
The goal will be to get rid of precompiled libs and compile them ourselves like we do on other platforms

[phate89]

IMO the best test would be to run ffmpeg 32-bit vs ffmpeg 64-bit, possibly compiled without HW acceleration. Not sure it has HW acceleration for HEVC though :-).

I'm actually surprised the gain is that big. Yes, there are gains to be had from a 64-bit arch. But that much? sounds weird.

Where is the port managed? Which github repo? I have quite a bit of experience with 32-bit to 64-bit conversions...

It's not only about 64 bit capability. 64 bit version has a lot more cpu optimizations and developers. It's the same with other newer codecs like vp9
If you want to do a benchmark of ffmpeg decoding (software) is really easy with ffmpeg binaries you can find in ffmpeg site:
ffmpeg -benchmark -i INPUT -f null out.null

[jpsdr]

I'm actually surprised the gain is that big. Yes, there are gains to be had from a 64-bit arch. But that much? sounds weird.

I'm not surprised. Having almost twice more of CPU registers may have a great increase speed in some asm optimised video processing function, especialy when on a loop sub-routine, you can remove memory access (for static data coef for exemple) and replace them by register access, because now you can store them in the registers because you have more. For exemple.
I've personnaly, with of course the exact same file and same parameters, a 40% speedup between the x64 and x86 version of x264.

It doesn't mean that x64 will always provide great speedup, but in some specific case, it can. And video decoding/encoding are processing which often fit the "specific case".

[phate89]

I'm actually surprised the gain is that big. Yes, there are gains to be had from a 64-bit arch. But that much? sounds weird.

I'm not surprised. Having almost twice more of CPU registers may have a great increase speed in some asm optimised video processing function, especialy when on a loop sub-routine, you can remove memory access (for static data coef for exemple) and replace them by register access, because now you can store them in the registers because you have more. For exemple.
I've personnaly, with of course the exact same file and same parameters, a 40% speedup between the x64 and x86 version of x264.

It doesn't mean that x64 will always provide great speedup, but in some specific case, it can. And video decoding/encoding are processing which often fit the "specific case".

I did several test with my haswell CPU (so with latest instructions to get all optimizations) and that's not true, at least for h264. I always get approximately 10% better results with 64 bit version on a ffmpeg build on windows. Not a lot of difference compared with the doubling of speed you get with hevc or vp9 (where devs stated they don't really care to optimize 32 bit version)

[jpsdr]

I'm just saying that such speed difference between x86 and x64 is possible. Maybe with decoding h264 with ffmpeg you have only 10%, but with encoding a video with x264 i've noticed around 40%, with the parameters i've used (according the parameters you're encoding with, difference can be smaller, i've apparently hit almost the max with the parameters i'm usied).

[phate89]

I'm just saying that such speed difference between x86 and x64 is possible. Maybe with decoding h264 with ffmpeg you have only 10%, but with encoding a video with x264 i've noticed around 40%, with the parameters i've used (according the parameters you're encoding with, difference can be smaller, i've apparently hit almost the max with the parameters i'm usied).

Sure with encoding there is a very consistent advantage but decoding is a lot simpler than encoding so the advantage is smaller and that's not the reason because hevc is twice as fast in x64

[voguemaster]

I'm actually surprised the gain is that big. Yes, there are gains to be had from a 64-bit arch. But that much? sounds weird.

I'm not surprised. Having almost twice more of CPU registers may have a great increase speed in some asm optimised video processing function, especialy when on a loop sub-routine, you can remove memory access (for static data coef for exemple) and replace them by register access, because now you can store them in the registers because you have more. For exemple.
I've personnaly, with of course the exact same file and same parameters, a 40% speedup between the x64 and x86 version of x264.

It doesn't mean that x64 will always provide great speedup, but in some specific case, it can. And video decoding/encoding are processing which often fit the "specific case".

40% is pretty steep. From what you said those micro-optimizations mean that there is a real CPU bottleneck in the processing of video. May I remind you that in most cases the real problem is not the CPU but memory access times? So yeah, if you optimize your asm code correctly you can gain 40% but i'm still skeptical about it just being CPU cycles. Maybe you can optimize a lot more when you work your data properly but still - i'm a bit surprised by those numbers.

[jpsdr]

May I remind you that in most cases the real problem is not the CPU but memory access times?

With almost twice registers avaible in x64, you can reduce memory access and replace them by register access, when of course it's possible.
Also, not all, of course, but a lot of x264 code is highly ultra Intel ASM optimized... Maybe it reaches the possible peak. I've this result on a i7@980.

[hentai23]

man i didnt even know that x64 bit kodi build was in progress , this is great news , since we over at the dsplayer thread could use this as madshi has released his x64 version of madvr already , all thats left is to have kodi be fully x64 based and im sure reclock for those who use it will follow

[aquiles2k]

Is this still on hold?

[hentai23]

any news on this?

[Paxxi]

No news, lack of time and Windows developers

[hentai23]

well heres to hoping some news will surface soonish , keep up the great work guys , thanks

[HeresJohnny]

I see some win64 builds amongst the nightlies. Any info on that?

[FXB78]

I see some win64 builds amongst the nightlies. Any info on that?

There's a thread about it somewhere, I believe they are looking to maybe get 18 released in 64-bit.