Improving OpenCL On CPUs, Building Linux

[phoronix]

Phoronix: Improving OpenCL On CPUs, Building Linux

Back in April there was an LLVM European Conference in London where several interesting technical discussions happened. Among the topics covered were auto-vectorization with LLVM, building Linux with LLVM, and using LLVM to improve the performance of OpenCL on CPUs...

http://www.phoronix.com/vr.php?view=MTEyMTU

[uid313]

Why run OpenCL on the CPU?
Isn't OpenCL meant to be run on the GPU?

[deanjo]

Why run OpenCL on the CPU?
Isn't OpenCL meant to be run on the GPU?

openCL is meant to run on many devices including cpu's and DSP's.

[smitty3268]

Why run OpenCL on the CPU?
Isn't OpenCL meant to be run on the GPU?

The point is to have a fallback in cases where the GPU on the target platform doesn't support OpenCL.

That way developers can be sure that OpenCL will run everywhere - otherwise they may not want to go to the trouble of using it at all, if they have to provide their own alternative fallback.

[d.a.a.]

The point is to have a fallback in cases where the GPU on the target platform doesn't support OpenCL.

That way developers can be sure that OpenCL will run everywhere - otherwise they may not want to go to the trouble of using it at all, if they have to provide their own alternative fallback.

Not necessarily. Firstly, there are types of workloads that run faster on the CPU than on the GPU; mostly for algorithms (1) having low degree of parallelism/decomposition; (2) with very divergent work-flows or irregular data access patterns; and (3) that would need many data transfers between the host (CPU) and the GPU devices (via the relatively slow PCI-E interface). Finally, OpenCL is designed to take advantage of *all* available compute devices, which potentially includes GPUs, DSPs, FPGAs, and obviously CPUs.

[popper]

Why run OpenCL on the CPU?
Isn't OpenCL meant to be run on the GPU?

Open Computing Language (OpenCL) API is designed and expressly written to be run on ANY and all Co-processor(s) available to the system at the same time , as in a Heterogeneous computing systems that use a variety of different types of computational units including GPU CPU and FPGA's,

http://www.altera.com/education/webc...cl/player.html

http://www.marketwatch.com/story/alt...ers-2012-04-10
"April 10, 2012, 8:00 a.m. EDTAltera's OpenCL for FPGAs Program Delivers Dramatic Reductions in Development Times for Early Customers"


but AMD or at least Bridgman doesn't consider FPGA's as a valid option for placing them on their generic GPU and/or motherboard PCB's (i cant link the thread here as Michael has screwed up the phoronix database and you cant go back that far anymore apparently).

i do wonder though if that Negative stance has now changed with AMD joining ARM (and their Midgard Mali T-604 full OpenCL abilities etc) as founders of and the formation of the Heterogeneous System Architecture (HSA) Foundation along side Imagination, MediaTek and Texas Instruments.

plus with AMD officially to integrate ARM core into their APUs, and AMD also creating their new embedded IC business unit appointed Arun Iyengar, former head of Altera Corp.'s Military, Industrial, Computing business division, to run the new Embedded Solutions Group all carefully managed to be setup and to get worldwide PR coverage at the same time, it all points (plus other positive cash flow options not stated here) to FPGA's Vendors being encouraged and have their products integrated in to the Heterogeneous System Architecture (HSA) Foundation products and spec.


OC when i say FPGA i dont mean the full developer $1,999.00 a seat for a single board but rather the cheap as chips FPGA OpenCL end product and installed binaries to perform that selection of OpenCL/Video Encode/Decode etc tasks far faster if you don't want to go the full yard and make available binaries for real http://opencores.org/newsletter,2011,11,#n5

http://opencores.org/

[Qaridarium]

Open Computing Language (OpenCL) API is designed and expressly written to be run on ANY and all Co-processor(s) available to the system at the same time , as in a Heterogeneous computing systems that use a variety of different types of computational units including GPU CPU and FPGA's,

http://www.altera.com/education/webc...cl/player.html

http://www.marketwatch.com/story/alt...ers-2012-04-10
"April 10, 2012, 8:00 a.m. EDTAltera's OpenCL for FPGAs Program Delivers Dramatic Reductions in Development Times for Early Customers"


but AMD or at least Bridgman doesn't consider FPGA's as a valid option for placing them on their generic GPU and/or motherboard PCB's (i cant link the thread here as Michael has screwed up the phoronix database and you cant go back that far anymore apparently).

i do wonder though if that Negative stance has now changed with AMD joining ARM (and their Midgard Mali T-604 full OpenCL abilities etc) as founders of and the formation of the Heterogeneous System Architecture (HSA) Foundation along side Imagination, MediaTek and Texas Instruments.

plus with AMD officially to integrate ARM core into their APUs, and AMD also creating their new embedded IC business unit appointed Arun Iyengar, former head of Altera Corp.'s Military, Industrial, Computing business division, to run the new Embedded Solutions Group all carefully managed to be setup and to get worldwide PR coverage at the same time, it all points (plus other positive cash flow options not stated here) to FPGA's Vendors being encouraged and have their products integrated in to the Heterogeneous System Architecture (HSA) Foundation products and spec.


OC when i say FPGA i dont mean the full developer $1,999.00 a seat for a single board but rather the cheap as chips FPGA OpenCL end product and installed binaries to perform that selection of OpenCL/Video Encode/Decode etc tasks far faster if you don't want to go the full yard and make available binaries for real http://opencores.org/newsletter,2011,11,#n5

http://opencores.org/

you are right FPGA is the future and i think future CPUs will do have a Vector-SIMD unit and a FPGA part just to make sure you can make software run like hellfire speed.

[popper]

you are right FPGA is the future and i think future CPUs will do have a Vector-SIMD unit and a FPGA part just to make sure you can make software run like hellfire speed.

Qaridarium PLEASE don't use this thread to bash bridgman if he says anything you want to make a bad AMD point about then start another thread or add those comments to your threads and link him there so he can follow if he likes rather than here.

i Really want to see where if anywhere current and near future FPGA products might finally become popular and come into the generic Co-Processor usage everywhere, OpenCL/Encode/Decode for 2k/4k being top of my wish list progress OC.

[Qaridarium]

Qaridarium PLEASE don't use this thread to bash bridgman if he says anything you want to make a bad AMD point about then start another thread or add those comments to your threads and link him there so he can follow if he likes rather than here.

i Really want to see where if anywhere current and near future FPGA products might finally become popular and come into the generic Co-Processor usage everywhere, OpenCL/Encode/Decode for 2k/4k being top of my wish list progress OC.

LOL why do you spam me with crazy Bridgman stuff here?

i just admit that you are right about FPGA its a great technique similar to Vector-SIMD units.

IF a CPU like Loongson include a complete FPGA AND a Vector-SIMD unit AND a MIBS64bit cpu and this CPU becomes popular then you will see software what beats the socks of of these "Intel" Jokes

your really best hope for FPGA in general use is a CPU like Loongson ad this as a Co-processor.

but AMD or at least Bridgman doesn't consider FPGA's as a valid option for placing them on their generic GPU

this Bridgman insanity Quotes really become popular on Phoronix.

EDIT: try to buy a hd6950 on ebay right now just to make sure bridgman get the support he deserve to replace my old hd4770 the one who becomes useless after all work only go in hd5000+ openCL and Shader compiler and stuff.

[smitty3268]

you are right FPGA is the future and i think future CPUs will do have a Vector-SIMD unit and a FPGA part just to make sure you can make software run like hellfire speed.

I may not be up to speed on this story, but why in the world would AMD want an FPGA on their boards?

Wouldn't they just hardcode whatever functionality you are talking about? The point behind a FPGA is that it can be reprogrammed on the fly - that is it's defining characteristic, and it's what makes it so expensive. I can't imagine AMD would want to pay for that feature, they'd just put a chip on the board that would run OpenCL without being fully-programmable.