cancel
Showing results for 
Search instead for 
Did you mean: 

Archives Discussions

gbilotta
Adept III

Re: Blender Cycles(Opencl on AMD GPUS)


P.S. Inferior cuda have unified memory, where's Opencl 2.0 with unified memory on AMD?


Oh please, CUDA's “unified memory” is 90% marketing and 10% software (the driver part that does the actual copying between host and device without exposing the need to do so to the developer). NVIDIA hardware is simply not capable to do truly unified memory, as opposed to AMD HSA-enabled APUs.

Regarding the OpenCL 2.0 SVM, you can already play around with them in the latest AMD drivers, even though they are still 1.2 formally (cl_amd_svm extension).

0 Likes
sdar
Journeyman III

Re: Blender Cycles(Opencl on AMD GPUS)

As far as I know Opencl 2.0 is only working on HSA capable APUs.

HSA is only working on APUs, AMD APUs performance are far behind my needs so no HSA nor Opencl 2.0 for high performance hardware, nice.

90% marketing 10% software seems to be better than 80% marketing 20% promises.

0 Likes
VincentSC
Adept II

Re: Blender Cycles(Opencl on AMD GPUS)

Nvidia doesn't have unified memory indeed. See "CUDA 6 Unified Memory explained" for more info. Nvidia put this in the world, as it is afraid of embedded GPUs, as it should. See also my post "7 things NVIDIA doesn't want you to know" on the disputable company ethics of Nvidia. Exactly this lying is the reason I prefer OpenCL and companies like AMD - I don't want to buy from companies with low company ethics.

Current drivers of AMD are not aware of the dependency-tree of all the parts in a kernel-file. for instance a large .cl-file with 20 independent kernels compiles much slower than having them split over 20 files and compiling them one by one. So there is plenty of room for improvements.

0 Likes
gbilotta
Adept III

Re: Blender Cycles(Opencl on AMD GPUS)


sdar wrote:



As far as I know Opencl 2.0 is only working on HSA capable APUs.



HSA is only working on APUs, AMD APUs performance are far behind my needs so no HSA nor Opencl 2.0 for high performance hardware, nice.



90% marketing 10% software seems to be better than 80% marketing 20% promises.


OpenCL 2.0 is not only for HSA-capable APUs. The specification is very wide and it encompasses a LOT of features, of which SVM is only a small part. And SVM itself is not limited to HSA-capable APUs, as it can be supported by any device that is capable of reading/writing host memory using the same address space as the CPU. This includes any NVIDIA devices with UVA (i.e. since Fermi).

0 Likes
mdriftmeyer
Adept II

Re: Blender Cycles(Opencl on AMD GPUS)


Reading all the emotional commentary about choosing Nvidia over AMD for Cycles truly misses the boat on the lack of work invested, by Blender, with respect to implementing OpenCL within Blender proper.



There are dozens of areas Blender could architect to leverage OpenCL on either Nvidia or most importantly, AMD GPGPUs and CPUs.



And yet it is AMD who gets slammed for ``dropping' the ball, while other projects manage to implement AMD's OpenCL stack well within their projects.



I am waiting for the next round of AMD GPGPUs to buy and later will be buying a Mac Pro, so the FirePro investment is assured.



I personally have zero interest in CUDA and complete interest in OpenCL, OpenGL, Metal and eventually Mantle.



It seems with all the heavy lifting that has been committed to LLVM/Clang, and a branch of OpenMP for LLVM-Clang 3.5 now branched, you would hope Blender spends some cycles working on making LLVM a first class citizen within it's project, thus making the likes of AMD, Apple and others that much more interested in pitching in knowing their investment in the LLVM community is assured inside a project like Blender.



Perhaps Blender will have some future GSoC projects centered around these areas, instead of complaining that AMD hasn't done the heavy lifting for them.



After all, the OpenCL quality from Nvidia in Blender is abysmal, but then again Nvidia has zero interest in pushing a superior solution [OpenCL] when it's heavily invested in an inferior proprietary solution [CUDA].


There is an entire community of developers, free of charge, at Blender's disposal, who are extremely helpful in all matters LLVM/Clang, and yet Blender devs are silent on the LLVM/Clang Dev Lists. Hell, you can even talk to AMD devs on there, including Google, SONY, Apple, IBM, and more about OpenCL.

Blender devs: Silent.

Blender is a beautiful project, but whining that AMD isn't doing the heavy lifting in working around Cycles architecture, instead of Cycles working with groups of experts to evolve its architecture speaks very poorly of Blender.

0 Likes
sdar
Journeyman III

Re: Blender Cycles(Opencl on AMD GPUS)

LLVM/Clang != AMD Opencl Compiler.

LLVM/clang is not going to help Cycles render to work on AMD GPUs and probably is going to make it run slower on CPUs too.

LLVM is faster compiling but creates a slightly slower code than gcc, also blender devs are not having problems with llvm are having problems with AMD OPENCL COMPILER.


Blender devs: Silent.


Ton Roosendaal twitter:


We are in contact with too many AMD people already. It's their turn. JUST FIX THE COMPILER!!!


0 Likes
Ziple
Adept I

Re: Blender Cycles(Opencl on AMD GPUS)

AMD OpenCL compiler uses LLVM.

0 Likes
sdar
Journeyman III

Re: Blender Cycles(Opencl on AMD GPUS)

llvm don't choke to death when compiling large kernels, so it's better to say that it's based on llvm but is not working like llvm anymore.

Just test Intel Opencl compiler that is based on llvm and you'll compile almost any large kernel within seconds.

Ex. Luxrender all materials and textures enabled

Intel OpenCL CPU 4.5 Seconds

AMD OpenCL CPU >15 minutes

AMD OpenCL GPU  N/A


Source


Anyway using a different compiler for the rest of blender is not going to help cycles run on AMD GPUs cause cycles for GPU Opencl uses the compiler included in your drivers and it compiles itself the first time you use it.

0 Likes
cusa123
Adept I

Re: Blender Cycles(Opencl on AMD GPUS)

That's why many times Sdar to compile the kernel gpu must be charged first cpu. That last ah me sometimes if fee directly to gpu gpu often the program hangs and Blender.

0 Likes
cusa123
Adept I

Re: Blender Cycles(Opencl on AMD GPUS)

Sorry, but this is what we expect since June ?. Because I do not see any improvement in either performance or anything that is visible at first glance.

0 Likes