20 Replies Latest reply on Aug 18, 2010 1:17 PM by genaganna

    Catalyst 10.4 for some 5xxx GPUs is missing ?

    Raistmer
      I have at least one report that Catalyst 10.4 doesn't recognise HD 5830 .
      Cause Cat 10.3 just got restarted after few seconds of work when running my app, it means no support for 5830 GPUs again. Looks like same situation with 5970 GPU.
      User reports driver restart after ~20 seconds of work.
      App proved to work flawless on HD4870 and 5770 GPUs under SDK 2.1 and Cat 10.4.
      What can be done in this situation? Any possibility to get next Catalyst driver upate that will work correctly on mentioned GPUs?
        • Catalyst 10.4 for some 5xxx GPUs is missing ?
          michael.chu

          Raistmer, are you seeing this under Windows or Linux? (and if Windows, which version?)

            • Catalyst 10.4 for some 5xxx GPUs is missing ?
              freighter

               

              Originally posted by: michael.chu Raistmer, are you seeing this under Windows or Linux? (and if Windows, which version?)

               

              Hi,

              porting Raistmers app to 64bit Linux (using openSuse 11, GPU HD5670), there occurs a repeatable error. But instead of an openCL error message the application stops with a SIGSEGV. After searching for other reasons for some days i've run out of options other than to report this issue as a bug in the SDKv2.1 or driver Cat 10.4

              On an identical computer (see below, except GPU HD4670) the application produces wrong results, but is not erroring out.

              Identified this command as source of error: clEnqueueNDRangeKernel(...) in bool Science::find_single_pulse_range_cl(...), using identical kernels file like Windows app.

              GDB (backtrace) and Valgrind (both point to problems within /usr/lib64/libaticaldd.so)

              outcome from GDB:

              Program received signal SIGSEGV, Segmentation fault.
              [Switching to Thread 0x7ffff7e41950 (LWP 4469)]
              0x00007ffff45d2424 in ?? () from /usr/lib64/libaticaldd.so

              first example of thousands from Valgrind:

              ==4483== Conditional jump or move depends on uninitialised value(s)
              ==4483==    at 0x84BA337: (within /usr/lib64/libaticaldd.so)
              ==4483==    by 0x84B9BF7: (within /usr/lib64/libaticaldd.so)
              ==4483==    by 0x84AEF94: (within /usr/lib64/libaticaldd.so)
              ==4483==    by 0x849EDDE: (within /usr/lib64/libaticaldd.so)
              ==4483==    by 0x83B4FEE: (within /usr/lib64/libaticaldd.so)
              ==4483==    by 0x84F0D8E: (within /usr/lib64/libaticaldd.so)
              ==4483==    by 0x84E9F40: (within /usr/lib64/libaticaldd.so)
              ==4483==    by 0x84FAC54: (within /usr/lib64/libaticaldd.so)
              ==4483==    by 0x660B562: (within ~/development/ati-stream-sdk-v2.1-lnx64/lib/x86_64/libatiocl64.so)
              ==4483==    by 0x65ECA19: (within ~/development/ati-stream-sdk-v2.1-lnx64/lib/x86_64/libatiocl64.so)
              ==4483==    by 0x66169C9: (within ~/development/ati-stream-sdk-v2.1-lnx64/lib/x86_64/libatiocl64.so)
              ==4483==    by 0x65DBDB4: clGetPlatformInfo (in ~/development/ati-stream-sdk-v2.1-lnx64/lib/x86_64/libatiocl64.so)

               

              Hardware:

                              Desktop1                                         Desktop2

              CPU:        Intel T7200                                      Intel T7200

              GPU:        HD5670, 1GB                                   HD4670, 1GB

              RAM:          2GB                                                    2GB

              MB:                                    ASUS N4LVM-DH

              OS:                                    openSuse11, 64bit

               

              Urs

            • Catalyst 10.4 for some 5xxx GPUs is missing ?
              Raistmer
              First of all, it's not my own observation, I recived report from beta-tester of my app.
              AFAIK it was Windows7 x64.
              I asked GPU owner to describe situation by himself, hope he could provide more detailed description.
                • Catalyst 10.4 for some 5xxx GPUs is missing ?
                  arkayn

                   

                  Originally posted by: Raistmer First of all, it's not my own observation, I recived report from beta-tester of my app. AFAIK it was Windows7 x64. I asked GPU owner to describe situation by himself, hope he could provide more detailed description.


                   

                  Actually it is XP Pro-64, there is also a thread on the game side of the forums.

                  http://forums.amd.com/game/messageview.cfm?catid=279&threadid=132505&enterthread=y

                   

                  When I removed all the of the drivers and SDK and attempted to reinstall 10.4, there is no option for driver in installer. I tried both the preview and the final release. Hoping that the 10.5 preview will be coming out soon.

                  • Catalyst 10.4 for some 5xxx GPUs is missing ?
                    rosmo01

                    Hi,

                    I represent the 5970 behaviour.

                    Setup:

                    Windows 7 x64, 2x 5970, Cat 10.4. All GPUs reported as cat 10.4 by the app. 

                    The application never utilizes any of the GPUs, as the driver restarts. This goes on indefinitely.

                    Morten

                     

                    • Catalyst 10.4 for some 5xxx GPUs is missing ?
                      Claggy

                       

                      Originally posted by: Raistmer First of all, it's not my own observation, I recived report from beta-tester of my app. AFAIK it was Windows7 x64. I asked GPU owner to describe situation by himself, hope he could provide more detailed description.


                      My setup is an HD5770 and a Nvidia 9800GTX+, with Cat 10.4, SDK2.1_x86_64 and FW197.57, and Windows 7 64bit Premium, in this config Raistmer's OpenCL app crashes, even his Brook+ version crashes (which worked fine with Cat 10.2, SDK2.01_x86_64 and 197.25), and also GPU-Z crashes,

                      it isn't until i physically remove the 9800GTX+ and it's drivers, that i can get Raistmer's OpenCL app to function correctly, GPU-Z then also works.

                      If install Cat 10.2 and SDK2.1_x86_64 (with the 9800GTX+ fitted), Raistmer's Brook+ app now works, GPU-Z works, but the OpenCL app causes an ATI driver restart and doesn't progress.

                      Claggy

                        • Catalyst 10.4 for some 5xxx GPUs is missing ?
                          omkaranathan

                          freighter,

                          Could you post the source code?(A compilable testcase)

                          • Catalyst 10.4 for some 5xxx GPUs is missing ?
                            Claggy

                            Cat 10.5 didn't fix my problems, when Nvidia GPU also installed with 257.15 drivers, SDK2.1 installed, Brook+ and ATI OpenCL apps still crash,

                            reverted back to Cat 10.2, Brook+ apps now work again.

                             

                            Claggy

                             

                            Edit: HelloCl and Clinfo both Crash with Cat 10.5 & 257.15,  run with Cat 10.2 O.K, but with:

                            Error : Bytes mismatch!
                            Error : glSharing mismatch!
                            Error : images mismatch!
                            Failed!

                              • Catalyst 10.4 for some 5xxx GPUs is missing ?
                                omkaranathan

                                freighter,

                                 

                                The application is sending an invalid memory object.

                                You are doing typecast of data array to a memory object

                                  gpu_power = (cl_mem)&power;

                                 



                                  • Catalyst 10.4 for some 5xxx GPUs is missing ?
                                    freighter

                                     

                                    Originally posted by: omkaranathan freighter,

                                     

                                    The application is sending an invalid memory object.

                                    You are doing typecast of data array to a memory object

                                      gpu_power = (cl_mem)&power;



                                    Yes, i know that and did it this way to get a possibly partly matching backtrace.

                                    A version that reads the data array in the correct way, using a buffer, is also in test_case.cpp, uncommented, but that way there is no error, no backtrace.

                                    This is the closest that i can get to match a backtrace to the real problem. Like i wrote before, i was not able to reproduce the exact problem with such a short testcase.

                                    omkaranathan, please have a look into test_case.h . Try to combine the two backtraces there. Both mem addresses match upto a certain address.

                                    Sorry, if it is a bit more complicated to get to the problem.

                                      • Catalyst 10.4 for some 5xxx GPUs is missing ?
                                        genaganna

                                         

                                        Originally posted by: freighter
                                        Originally posted by: omkaranathan freighter,

                                         

                                        The application is sending an invalid memory object.

                                        You are doing typecast of data array to a memory object

                                          gpu_power = (cl_mem)&power;

                                         

                                        Yes, i know that and did it this way to get a possibly partly matching backtrace.

                                         

                                        A version that reads the data array in the correct way, using a buffer, is also in test_case.cpp, uncommented, but that way there is no error, no backtrace.

                                         

                                        This is the closest that i can get to match a backtrace to the real problem. Like i wrote before, i was not able to reproduce the exact problem with such a short testcase.

                                         

                                        omkaranathan, please have a look into test_case.h . Try to combine the two backtraces there. Both mem addresses match upto a certain address.

                                         

                                        Sorry, if it is a bit more complicated to get to the problem.

                                         



                                        As you are doing illegal casting, This testcase is not sufficient to reproduce your problem.  Please send your code to streamdeveloper@amd.com

                                    • Catalyst 10.4 for some 5xxx GPUs is missing ?
                                      Claggy

                                       

                                      Originally posted by: Claggy Cat 10.5 didn't fix my problems, when Nvidia GPU also installed with 257.15 drivers, SDK2.1 installed, Brook+ and ATI OpenCL apps still crash,

                                      reverted back to Cat 10.2, Brook+ apps now work again.

                                       

                                      Claggy

                                       

                                      Edit: HelloCl and Clinfo both Crash with Cat 10.5 & 257.15,  run with Cat 10.2 O.K, but with:

                                      Error : Bytes mismatch! Error : glSharing mismatch! Error : images mismatch! Failed!

                                       

                                      Still got this Issue, with Nvidia GPU fitted in 1st slot, HD5770 fitted in 2nd, Cat 10.6 or Cat 10.7 installed with latest Nvidia drivers,

                                      SDK_2.1 installed with Cat 10.6, SDK_2.2 installed with Cat 10.7,

                                      Brook+ and OpenCL apps still crash, GPU-Z crashes when SDK installed, O.K when not installed, at least latest version of Boinc now ignores ATI GPU when it detects SIGSEGV in ATI GPU detection,

                                      Is there any timescale for a fix?

                                      Claggy