cancel
Showing results for 
Search instead for 
Did you mean: 

Archives Discussions

Raistmer
Adept II

Catalyst 11.11 is broken too

Here the example of difference between 11.2 and 12.1preview elapsed times for my app:

11.2:
App Name Task name AR CPU time Elapsed

MB7_win_x86_SSE3_OpenCL_ATi_r390 PG0009_v7.wu 0.008955 34.975 129.685
MB7_win_x86_SSE3_OpenCL_ATi_r390 PG0009_v7_1.wu 0.008955 36.879 125.781
MB7_win_x86_SSE3_OpenCL_ATi_r390 PG0009_v7_2.wu 0.008955 32.511 251.097
MB7_win_x86_SSE3_OpenCL_ATi_r390 PG0009_v7_3.wu 0.008955 32.823 274.248
MB7_win_x86_SSE3_OpenCL_ATi_r390 PG0009_v7_4.wu 0.008955 35.085 124.304
MB7_win_x86_SSE3_OpenCL_ATi_r390 PG0009_v7_5.wu 0.008955 31.731 276.672
MB7_win_x86_SSE3_OpenCL_ATi_r390 PG0009_v7_6.wu 0.008955 31.762 291.919
MB7_win_x86_SSE3_OpenCL_ATi_r390 PG0009_v7_7.wu 0.008955 33.025 259.631
MB7_win_x86_SSE3_OpenCL_ATi_r390 PG0009_v7_8.wu 0.008955 32.76 256.995
MB7_win_x86_SSE3_OpenCL_ATi_r390 PG0009_v7.wu 0.008955 35.35 132.97
MB7_win_x86_SSE3_OpenCL_ATi_r390 PG0009_v7_1.wu 0.008955 37.596 127.986
MB7_win_x86_SSE3_OpenCL_ATi_r390 PG0009_v7_2.wu 0.008955 34.024 252.894
MB7_win_x86_SSE3_OpenCL_ATi_r390 PG0009_v7_3.wu 0.008955 36.707 125.958
MB7_win_x86_SSE3_OpenCL_ATi_r390 PG0009_v7_4.wu 0.008955 35.802 125.538
MB7_win_x86_SSE3_OpenCL_ATi_r390 PG0009_v7_5.wu 0.008955 35.319 124.983
MB7_win_x86_SSE3_OpenCL_ATi_r390 PG0009_v7_6.wu 0.008955 33.837 259.316
MB7_win_x86_SSE3_OpenCL_ATi_r390 PG0009_v7_7.wu 0.008955 35.802 126.343
MB7_win_x86_SSE3_OpenCL_ATi_r390 PG0009_v7_8.wu 0.008955 36.411 126.186

12.1p:
App Name Task name AR CPU time Elapsed

MB7_win_x86_SSE3_OpenCL_ATi_r390 PG0009_v7.wu 0.008955 29.734 470.622
MB7_win_x86_SSE3_OpenCL_ATi_r390 PG0009_v7_1.wu 0.008955 29.718 497.89
MB7_win_x86_SSE3_OpenCL_ATi_r390 PG0009_v7_2.wu 0.008955 29.812 471.589
MB7_win_x86_SSE3_OpenCL_ATi_r390 PG0009_v7_3.wu 0.008955 29.562 474.88
MB7_win_x86_SSE3_OpenCL_ATi_r390 PG0009_v7_4.wu 0.008955 34.632 116.392
MB7_win_x86_SSE3_OpenCL_ATi_r390 PG0009_v7_5.wu 0.008955 34.648 116.47
MB7_win_x86_SSE3_OpenCL_ATi_r390 PG0009_v7_6.wu 0.008955 30.763 388.191
MB7_win_x86_SSE3_OpenCL_ATi_r390 PG0009_v7_7.wu 0.008955 34.897 116.855
MB7_win_x86_SSE3_OpenCL_ATi_r390 PG0009_v7_8.wu 0.008955 34.866 116.674
MB7_win_x86_SSE3_OpenCL_ATi_r390 PG0009_v7.wu 0.008955 30.795 583.831
MB7_win_x86_SSE3_OpenCL_ATi_r390 PG0009_v7_1.wu 0.008955 30.311 576.561
MB7_win_x86_SSE3_OpenCL_ATi_r390 PG0009_v7_2.wu 0.008955 31.091 499.032
MB7_win_x86_SSE3_OpenCL_ATi_r390 PG0009_v7_3.wu 0.008955 32.745 173.332
MB7_win_x86_SSE3_OpenCL_ATi_r390 PG0009_v7_4.wu 0.008955 33.384 186.435
MB7_win_x86_SSE3_OpenCL_ATi_r390 PG0009_v7_5.wu 0.008955 35.506 117.972
MB7_win_x86_SSE3_OpenCL_ATi_r390 PG0009_v7_6.wu 0.008955 33.01 185.255
MB7_win_x86_SSE3_OpenCL_ATi_r390 PG0009_v7_7.wu 0.008955 35.677 116.281
MB7_win_x86_SSE3_OpenCL_ATi_r390 PG0009_v7_8.wu 0.008955 31.887 250.409

As one can see, sometimes 12.1p much slower.

This leads to 2 issues:

1) under both drivers elapsed time increased a lot sometimes
2) under new 12.1p driver this increae can be much bigger

Testing environment included few idle-priority CPU-intencive processes + GPU app running at increased (increased above normal process) priority.
0 Likes
_heinz
Journeyman III

Catalyst 11.11 is broken too

I bought a HD 4670 AGP to have a machine to test OCL programs, OS W7 x32. Latest driver http://support.amd.com/de/kbarticles/Pages/CatalystAGPHotfix.aspx  forces CPU always to 100%. This is a error in latest driver. CPU usage should be max 5%. AMD produce good hardware, but it is useless if driver fails.

I'm very disappointed from AMD company. I will revise my decision to bouhgt new AMD server's if driver support is so bad. This way you lost all my trust to AMD.

_heinz

 

 

0 Likes
nou
Exemplar

Catalyst 11.11 is broken too

you bouht the worst card for opencl. first it is 4xxx card which can hardly manage opencl. second it is AGP which have another set of problem of own.

0 Likes
Meteorhead
Challenger

Catalyst 11.11 is broken too

Indeed. There have been a few topics where people were asking what sort of card to buy for testing. Consulting some of these before buying would've saved some money and time. 4670 sure supports OpenCL and might seem an economically wise decision, however 4xxx cards only emulate some features, such as __local (or shared for CUDA) memory as it is. It resides in __global (or device) memory, therefore it hurts performance badly.

I don't know about AGP connector, bud naturally, it is badly deprecated.

Testing performance on 3 years old HW is somewhat self-controversial. If you don't want to spend a lot of money on a test machine, consider buying 6850, which is downclocked second largest chip of the presently newest generation, holds all features and shows decent performance. 7000 cards won't hit market until middle of January most likely.

It is if I'd test NV on GT220 and say that it is bad perf-wise. Naturally, cause it is 3 years old and it is the lowest of the class.

0 Likes
freighter
Journeyman III

Catalyst 11.11 is broken too

Originally posted by: freighter
Originally posted by: MicahVillmow Raistmer, It was in response to raistmer/freighter. XP support is being dropped, so if issues arise, they probably won't be fixed.


 

Will repeat the tests after newer Win7x64 has finished install on that XP host i'm currently running tests.

 

Finally was able to reproduce the issue, first seen on WinXPsp3, now on Win7x64sp1, Cat 11.12 :

Some bluescreen data (added TdrLevel=1 in windows registry) right from crash  :

Technical Info:

STOP: 0x0116 (0xFFFFFA8001B044E0, 0xFFFFF88002C078C8, 0x0000000000000000, 0x0000000000000001)

atikmpag.sys - Address FFFFF88002C078C8 base at FFFFF88002C00000, DateStamp 0x4ebb331c

AND additional after restarting the testmachine found in dump :

STOP: 0x0116 (0xFFFFFA8001B044E0, 0xFFFFF88002C078C8, 0x0000000000000000, 0x0000000000000001)

dxgmms1.sys - Address FFFFF88003F8FF07 base at FFFFF88003F80000, DateStamp 0x4ce799c1

 

Will retry with Cat 12.1preview also.

0 Likes
freighter
Journeyman III

Catalyst 11.11 is broken too

Originally posted by: Meteorhead Indeed. There have been a few topics where people were asking what sort of card to buy for testing. Consulting some of these before buying would've saved some money and time. 4670 sure supports OpenCL and might seem an economically wise decision, however 4xxx cards only emulate some features, such as __local (or shared for CUDA) memory as it is. It resides in __global (or device) memory, therefore it hurts performance badly.

 

I don't know about AGP connector, bud naturally, it is badly deprecated.

 

Testing performance on 3 years old HW is somewhat self-controversial. If you don't want to spend a lot of money on a test machine, consider buying 6850, which is downclocked second largest chip of the presently newest generation, holds all features and shows decent performance. 7000 cards won't hit market until middle of January most likely.

 

It is if I'd test NV on GT220 and say that it is bad perf-wise. Naturally, cause it is 3 years old and it is the lowest of the class.

 

Reasoning : If an OpenCL app works on one of these lower end HD4xxx GPUs it will run on all the newer ones, too. That seems to be true at least for Seti@home's OpenCL app.  This is not about performance or new features it is about providing compatibilty for a wider userbase.

0 Likes
MicahVillmow
Staff
Staff

Catalyst 11.11 is broken too

freighter,
If you want compatibility, don't target a specific device(HD4XXX), but target the language version. You can buy a higher end card but target only core OpenCL 1.0(without images), and it should work on the lower end cards. If it doesn't work, then it is a problem with the our software and should be reported as such.
0 Likes
freighter
Journeyman III

Catalyst 11.11 is broken too

Originally posted by: MicahVillmow freighter, If you want compatibility, don't target a specific device(HD4XXX), but target the language version. You can buy a higher end card but target only core OpenCL 1.0(without images), and it should work on the lower end cards. If it doesn't work, then it is a problem with the our software and should be reported as such.


MicahVillmov,

sorry if my issue is still a problem.

Where did i state that i target a "specific device(HD4xxx)" ? Only using a specific device for testing an OpenCL(1.0 without images) version of Seti@home that tries to follow AMDs recommendations for "HD4xxx GPUs with a max. workgroup size of 128" : Use not more than workgroup sizes of 32 if you want to be on the safe side. Workgroup sizes of 64 and higher can lead to incorrect results.

So, i made sure that AMDs recommendation is in use and works ok on at least one of the (former) supported platforms : openSuse 11.3, 64bit and Catalyst driver version 11.9 and older. Newer drivers lead to the same failure like i reported earlier in this thread for the windows version.

???

0 Likes
MicahVillmow
Staff
Staff

Catalyst 11.11 is broken too

freighter/Raistmer,
Does the crashing occur on anything outside of the HD4XXX series? We believe we have the performance issue understood and are working on a fix.
0 Likes
freighter
Journeyman III

Catalyst 11.11 is broken too

Originally posted by: MicahVillmow freighter/Raistmer, Does the crashing occur on anything outside of the HD4XXX series? We believe we have the performance issue understood and are working on a fix.


No, only  HD4xxx GPUs with max. workgroup size of 128 are showing the crashes.

0 Likes