AMD's Graphics Core Next- a reason to accelerate towards OpenCL? - Page 2

Ethan (EO) · 2012-02-13, 19:54

Quote:

Originally Posted by msft

I find AMD-APP-SDK-v2.6-RC3-lnx64/samples/opencl/cl/app/FFT .
Can someone run this ?

From a quick glance at the kernel, I believe this is hard coded to 1K FFT length :\ Regardless, here is usage and output:

Code:

:~/AMD-APP-SDK-v2.6-RC3-lnx64/samples/opencl/bin/x86_64# ./FFT --help
Usage
-h, --help      Display this information
--device        Execute the openCL kernel on a device [cpu|gpu]
-q, --quiet     Quiet mode. Suppress all text output.
-e, --verify    Verify results against reference implementation.
-t, --timing    Print timing.
--dump  Dump binary image for all devices
--load  Load binary image and execute on device
--flags Specify compiler flags to build kernel
-p, --platformId        Select platformId to be used[0 to N-1 where N is number platforms available].
-v, --version   AMD APP SDK version string.
-d, --deviceId  Select deviceId to be used[0 to N-1 where N is number devices available].
-i, --iterations        Number of iterations to execute kernel

Code:

:~/AMD-APP-SDK-v2.6-RC3-lnx64/samples/opencl/bin/x86_64# ./FFT -t -i 1000
Platform 0 : Advanced Micro Devices, Inc.

Original Input Real
15.3732 201.81 51.9855 89.2322 92.572 34.4675 96.2478 66.3863 11.345 225.168


Original Input Img
0.0600514 0.788318 0.203068 0.348563 0.361609 0.134639 0.375968 0.259322 0.0443163 0.879562

Platform found : Advanced Micro Devices, Inc.

Selected Platform Vendor : Advanced Micro Devices, Inc.
Device 0 : Cypress Device ID is 0x1557a90
Executing kernel for 1000 iterations
-------------------------------------------

Output real
131643 -1085.95 -997.15 -1791.52 532.118 1659.74 -166.271 969.692 1189.76 -862.707


Output img
514.23 2289.84 936.489 -603.839 699.7 1018.18 1900.06 795.439 -1328.03 -293.334

Length                   Time(sec)                [Transfer+Kernel]Time(sec)
1024                     0.178297                 0.000297

msft · 2012-02-13, 22:16

Hi ,Ethan(EO)
34Mflops.
Only 1K FFT is problem.
Thank you,

jasong · 2012-02-19, 15:26

2 quick questions:

(1) Isn't the programming community trying to adapt CUDA to work with AMD GPUs, or possibly the reverse?

(2) BOINC is telling me to run GPU-Z to find out my adapter type. There's no obvious word adapter and I have no clue what an adapter is. I googled a bit and wonder if it means my bus interface. Am I right? Or even close to right?

jasonp · 2012-02-19, 16:03

"Graphics adapter" is another term for "graphics card".

The point of OpenCL was to have a single set of code that could be made to run on different graphics cards. AMD and Nvidia both have their own OpenCL toolkits to make code run on the own graphics cards. Right now the answer to your first question is that for trial factoring of large numbers the first high-performance code that ran on graphics cards needed the Nvidia tools, and other developers have taken that and rewrote it to use OpenCL. For actually running primality tests I don't think things have gotten to that point yet.

kladner · 2012-02-19, 16:09

Quote:

Originally Posted by jasong

2 quick questions:

(1) Isn't the programming community trying to adapt CUDA to work with AMD GPUs, or possibly the reverse?

(2) BOINC is telling me to run GPU-Z to find out my adapter type. There's no obvious word adapter and I have no clue what an adapter is. I googled a bit and wonder if it means my bus interface. Am I right? Or even close to right?

(1) I believe it is the reverse. nVidia can be programmed to run Open CL.

(2) I'm not sure just what aspect of the adapter they are seeking. In the screen shot attached, GF104 is the chip in use. GTX 460 is the particular implementation of that chip. The Subvendor, Gigabyte, built the actual card. Bus width is the size of the "pipe" through which the GPU communicates with the memory.

Of these, the first three are the most likely targets, and either "Name" or "GPU" most likely to be significant information.

Why are you being instructed to find this out? To determine Compute Capability?

jasong · 2012-02-19, 16:31

Thanks for the answer jasonp :thumb:

Quote:

Originally Posted by kladner

(1) I believe it is the reverse. nVidia can be programmed to run Open CL.
...
Why are you being instructed to find this out? To determine Compute Capability?

It's part of the instructions to find out if it is capable of running boinc. So yes(just reread your question) it's to determine compute capability.

Dubslow · 2012-02-19, 19:05

To expand on nVidia -> OpenCL, I actually mentioned this once in the mfakto thread, and Bdot attempted to compile mfakto using nVidia's GTX OpenCL compiler, but ran into a major buttload of errors. Link! This means that for now, we'll use CUDA for nVidia, and OpenCL for AMD. (CUDA is proprietary to nVidia, and thus is not adaptable. OpenCL is supposed to be universal, but like I said, in practice it's not.)

msft · 2012-03-05, 02:34

http://developer.amd.com/libraries/a...s/default.aspx
AMD Accelerated Parallel Processing Math Libraries (APPML)

I bought 7750,but linux driver not yet.

Bdot · 2012-03-07, 18:58

Quote:

Originally Posted by msft

http://developer.amd.com/libraries/a...s/default.aspx
AMD Accelerated Parallel Processing Math Libraries (APPML)

I bought 7750,but linux driver not yet.

They also ship faster FFT libs, maybe we'll see LL tests on AMD GPUs soon

You seem to be the first one around here who owns one of the HD 7xxx series. Do you see any chance to do a test run of mfakto on that machine? I'm just too curious how the new cores will perform ... I would send you a special testing version that can time the different kernels ...

But of course it requires some installed AMD driver.

2012-02-19, 16:03	#15
jasonp Tribal Bullet Oct 2004 DED₁₆ Posts	"Graphics adapter" is another term for "graphics card". The point of OpenCL was to have a single set of code that could be made to run on different graphics cards. AMD and Nvidia both have their own OpenCL toolkits to make code run on the own graphics cards. Right now the answer to your first question is that for trial factoring of large numbers the first high-performance code that ran on graphics cards needed the Nvidia tools, and other developers have taken that and rewrote it to use OpenCL. For actually running primality tests I don't think things have gotten to that point yet. Last fiddled with by jasonp on 2012-02-19 at 16:05

2012-02-19, 19:05	#18
Dubslow Basketry That Evening! "Bunslow the Bold" Jun 2011 40<A<43 -89<O<-88 3×29×83 Posts	To expand on nVidia -> OpenCL, I actually mentioned this once in the mfakto thread, and Bdot attempted to compile mfakto using nVidia's GTX OpenCL compiler, but ran into a major buttload of errors. Link! This means that for now, we'll use CUDA for nVidia, and OpenCL for AMD. (CUDA is proprietary to nVidia, and thus is not adaptable. OpenCL is supposed to be universal, but like I said, in practice it's not.) Last fiddled with by Dubslow on 2012-02-19 at 19:09

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
The crime of reason	jasonp	Soap Box	1	2009-12-20 15:09
Is there any good reason for why GMP does this?	ShiningArcanine	Software	2	2007-12-11 15:11
My fatalistic(sp?) reason for believing in God.	jasong	Soap Box	22	2006-05-04 19:26
Another reason for high attrition...	guest	PrimeNet	27	2004-03-17 13:50
eBay suspending users for no reason?	eepiccolo	Lounge	12	2003-05-05 19:25

2012-02-13, 22:16	#13
msft Jul 2009 Tokyo 610₁₀ Posts	Hi ,Ethan(EO) 34Mflops. Only 1K FFT is problem. Thank you,

2012-02-19, 15:26	#14
jasong "Jason Goatcher" Mar 2005 3·7·167 Posts	2 quick questions: (1) Isn't the programming community trying to adapt CUDA to work with AMD GPUs, or possibly the reverse? (2) BOINC is telling me to run GPU-Z to find out my adapter type. There's no obvious word adapter and I have no clue what an adapter is. I googled a bit and wonder if it means my bus interface. Am I right? Or even close to right?

2012-03-05, 02:34	#19
msft Jul 2009 Tokyo 2×5×61 Posts	http://developer.amd.com/libraries/a...s/default.aspx AMD Accelerated Parallel Processing Math Libraries (APPML) I bought 7750,but linux driver not yet.