Hi, probably more for 'driver support' forum, but I would like here as well.
I have a problem with openCL program, clBuildProgram does segfault / crash with newer drivers.
Please help me improve the long time bad driver reputation AMD always had. The program was working fine with older drivers and older hw (5850).
I believe all 7xxx cards are affected with all 13.xx drivers.
Graphics Card
MSI R7850-1GD5/OC
and
GV-R795WF3-3GD
I believe all 7xxx cards are affected.
overclocking with/without - the same.
AMD Catalyst Driver Version, and Driver History
tested 12.2-12.10 works with tweaking,
13.4 and 13.6beta crashes
Operating System
Linux Fedora Core 16x64, 18x64, 19x64
Ubuntu 10.04 LTS and 12.04 LTS
Issue Details
OpenCL program using loops unrolling,sha256, ripemd160, elliptic curves, crashes / segfaults when calling clBuildProgram.
It can be tweaked so it compiles fine with 13.xx ati drivers, but then it produces bad results.
Motherboard or System Make & Model
ASUSTeK Computer INC. P8Z68-V LX
and
Gigabyte Technology Co., Ltd. Z77-D3H
Power Supply
ST-P0720PBA Seventeam Cilense 720w
and
Seasonic SS-500ET Active PFC T3
Display Device(s) and Connection(s) Used
HP ZR2740w
Applications and Games
opencl program - sha256, ripemd160, elliptic curves
CPU Details
Intel i5
and Intel i3
no overclocking
Motherboard BIOS Version
stock versions
System Memory Type & Amount
Kingston HyperX 4GB DDR 3 -
2x
and
1x
Additional Hardware
nothing special
Additional Details
no viruses in Linux,hw works well,memtest and Prime95 too.
---------------
Please help, there is another issue with new drivers, they even provide incorrect results,when offending crashing part is removed from code
I can assist with narrowing down the issue.
Thanks a lot.
best regards
.m.
You are posting on the right forum. Please provide us your kernel (attach it here as a thread). We will check out.
If required, we will raise a bug report internally.
Hi, please find enclosed description and crashing kernel.
Please do not hesitate to ask for clarification / amendments.
m.
zipped here :
sorry for checking this late, but the link provided is not working for us. Please use the forum itself for attaching the kernel instead of third party file hosting services.
Sorry, do not know, how to insert attachment, please find offending file here :
https://github.com/samr7/vanitygen/blob/master/calc_addrs.cl
The problem is now(latest drivers) only misbehaviour (not crash),it calculates incorrect GPU sum compared to CPU :
https://github.com/samr7/vanitygen/issues/19#issuecomment-22161310
but may be solved by downgrading AMDAPP SDK to 2.7 as somebody mentioned , will try ...
seems to be working well now with 13.8beta drivers and 3.10.5-201.fc19.x86_64, but I had to :
* install AMDAPP sdk 2.7
* manually overwrite all libs (x68,x64) with *ocl* name with those from app sdk
* ldconfig
* reboot
* run the app as a root
* disable 'quirks' (with '-S', probably just AMD_BFI_INT does not work with 7xxx cards)
tried with::
./oclvanitygen -d0 -v -v -S -i 1Lostxx ### wait about a minute
mulisak wrote:
* manually overwrite all libs (x68,x64) with *ocl* name with those from app sdk
* ldconfig
* reboot
* run the app as a root
* disable 'quirks' (with '-S', probably just AMD_BFI_INT does not work with 7xxx cards)
tried with::
./oclvanitygen -d0 -v -v -S -i 1Lostxx ### wait about a minute
What did you overwrite all libs with?
I guess you can give more information, as to what happens when you have latest APP SDK and catalyst 13.8beta. Also try using Kernel Analyzer for checking kernel compilations for different GPUs.
You can figure out if the compilation issue is specific to a GPU, or common for all generations. Also try modifying your kernel, and try make it compile, this can help you pinpoint the location of code that may be the culprit.
Hi, tested again with amd-driver-13.8-beta2 and amd sdk 2.8.1
when used under normal user -> coredump
[sq@private vanitygen.ori]$ ./oclvanitygen -S -i -v -v -d0 1Lostxx
Prefix difficulty: 941151070 1Lostxx
Difficulty: 941151070
Setting of real/effective user Id to 0/0 failed
Setting of real/effective user Id to 0/0 failed
Segmentation fault (core dumped)
when run under root -> incorrect GPU result compared to CPU.
I am sorry for my English, I do not know how to describe better. POST-BEFORE:: I unpacked AMD.APP.SDK.2.7 , took all libs (x86 && x64) that had *ocl* in their name, and overwrote those in system directories.
I am sorry I can not spend my time debugging your drivers anymore, I have to do paid work to pay my invoices.
It was working before with drivers 12.10 but is not working properly with 13.xx drivers - and is reported as such.
So please forward to developers.
Thanks a lot.
Best regards.
m.
how can i get it ?
I checked the discussion in github.
The patch was to disable AMD_BFI_INT
The "cl" code you have mentioned uses this define to utilize "cl_amd_meda_ops" function "amd_bytealign"
I wrote 2 kernels -- one which does bytealign() using amd_bytealign and other which does bytealign manually via software .
They both return the correct result...This is on Pitcairn device - 78xx device.
Here are the 2 kernels that I tested. They both give same results.
I checked it for 100000 iterations with various random numbers as "src0, src1 and src2"
I tested this on "uint" and not on vectors -- because I saw your code uses only scalar......
Can you run on this on your failing setup?
/*
Testing amd_bytealign
Build-in Function
uintn amd_bytealign (uintn src0, uintn src1, uintn src2)
Description
dst.s0 = (uint) (((((long)src0.s0) << 32) | (long)src1.s0) >> ((src2.s0 & 3)*8))
similar operation applied to other components of the vectors
*/
#pragma OPENCL EXTENSION cl_amd_media_ops : enable
__kernel void balignHW(uint src0, uint src1, uint src2, __global uint *result)
{
result[0] = amd_bytealign(src0, src1, src2);
}
__kernel void balignSW(uint src0, uint src1, uint src2, __global uint *result)
{
result[0] = (uint) (((((long)src0) << 32) | (long)src1) >> ((src2 & 3)*8));
}
Unfortunately your reply does not help make the program work under 13.xxx drivers.
I know the patch for Tahiti, I wrote it.
Program was working with older hardware and 12.10 drivers, it is not working with 13.xx amd drivers anymore (but it did work after using older amd sdk 2.7 and overwriting ocl* libraries - see above)
tested your program :
$ time ./a.out |grep -vE "(HW|SW|PASS|Corr|src0)"
Advanced Micro Devices, Inc.
Tahiti
real 0m9.954s
user 0m5.147s
sys 0m6.366s
==============
installed : amd-catalyst-13.11-beta1-linux-x86.x86_64.run
and latest AMD APP SDK
# uname -a
Linux private.rionet.cz 3.11.3-201.fc19.x86_64 #1 SMP Thu Oct 3 00:47:03 UTC 2013 x86_64 x86_64 x86_64 GNU/Linux
===but here CPU and GPU results differ====
D=vanitygen.worked.with.amd.12.10
git clone https://github.com/samr7/vanitygen.git $D
cd $D
wget https://gist.github.com/mnmnm/6954063/raw/e6546b40c82990106fe7dcf4f48554704b759cd7/tahiti.patch
patch -p1 < tahiti.patch
make
./oclvanitygen -v -i -d0 -k 1Lostx
Prefix difficulty: 32453485 1Lostx
Difficulty: 32453485
Device: Tahiti
Vendor: Advanced Micro Devices, Inc. (1002)
Driver: 1311.2 (VM)
Profile: FULL_PROFILE
Version: OpenCL 1.2 AMD-APP (1311.2)
Max compute units: 28
Max workgroup size: 256
Global memory: 2386558976
Max allocation: 1073741824
OpenCL compiler flags: -DDEEP_PREPROC_UNROLL -DVERY_EXPENSIVE_BRANCHES -DDEEP_VLIW
Loading kernel binary ce86061ed5322c51fec9fb904d9b205f.oclbin
Grid size: 3584x2048
Modular inverse: 7168 threads, 1024 ops each
Using OpenCL prefix matcher
[19.21 Mkey/s][total 22020096][Prob 49.3%][50% in 0.0s] Match idx: 12
CPU hash: a2dfd461c7c66f833277b65ae05616517241ee68
GPU hash: d947a4d0f2967514e326ff22924fd8c8b3235465
Found delta: 5898094 Start delta: 22020097
Match idx: 8
CPU hash: a1d0c4621d7627a9420415c5d6f4a8c38d4065ea
GPU hash: d932c65afdf4f267367904eea07ff9e00730ab77
So please forward to developers.
Thanks a lot.
Best regards.
m.
Hi Mulisak,
Thanks for coming back on this.
Unless I have a petite test-case (like what I wrote), I cannot raise a report
None has the time to look at vanitygen source code, understand what it does and then find out what is going on..
And what if, at the end of it --- it was a bug in your program....Well, Nobody can rule out this possibility.
So Practically, this is not going to work.
If you can isolate the bug to a particular fragment of OpenCL code,I will definitely file a case for you.
Best,
Bruhaspati
Hi,sorry,I am out of free time, but tested latest amd linux driver 13.11b6 and latest AMD R9 280x and AMD APP SDK 2.9
problem persists, different CPU and GPU hash results
$ ./oclvanitygen -v -i -d0 -k 1Lostx
Prefix difficulty: 32453485 1Lostx
Difficulty: 32453485
Device: Tahiti
Vendor: Advanced Micro Devices, Inc. (1002)
Driver: 1348.4 (VM)
Profile: FULL_PROFILE
Version: OpenCL 1.2 AMD-APP (1348.4)
Max compute units: 32
Max workgroup size: 256
Global memory: 2989490176
Max allocation: 1073741824
OpenCL compiler flags: -DDEEP_PREPROC_UNROLL -DVERY_EXPENSIVE_BRANCHES -DDEEP_VLIW
Compiling kernel, can take minutes...done!
Build log:
Grid size: 4096x2048
Modular inverse: 8192 threads, 1024 ops each
Using OpenCL prefix matcher
Match idx: 10
CPU hash: 93b02a4c6c1236e56a8b6113bbcfa454b1894016
GPU hash: d9332273aa7b03dca30303537a8ee223da015074
Found delta: 8075058 Start delta: 1
GPU idle: 1.18%
Match idx: 8
CPU hash: cb0a79aed222c28b22dd00ddac078bda2543f500
GPU hash: d932c66131afdf57fa7aa8ceb1afd80e5c3ee6d0
Found delta: 7358441 Start delta: 33554433
[19.71 Mkey/s][total 58720256][Prob 83.6%][90% in 0.8s]
Reviving the thread. Do you still see this problem?
--Prasad