I tried your patch and the bug is fixed.
Thanks.
Yan Wang
> LGTM, thanks.
>
>> -Original Message-
>> From: Gong, Zhigang
>> Sent: Friday, October 24, 2014 11:41
>> To: beignet@lists.freedesktop.org
>> Cc: yan.w...@linux.intel.com; Yang, Rong R; Gong, Zhigang
>> Subject: [PATCH] GBE: handl
LGTM, thanks.
> -Original Message-
> From: Gong, Zhigang
> Sent: Friday, October 24, 2014 11:41
> To: beignet@lists.freedesktop.org
> Cc: yan.w...@linux.intel.com; Yang, Rong R; Gong, Zhigang
> Subject: [PATCH] GBE: handle dead loop BBs in liveness analysis.
>
> Considering the following
Considering the following CFG, our previous liveness analysis will only
back traverse from the exit point BB, thus the BB 6 and 7 will not be handled.
4---
| |
| |
5 --|---> 10 ---> ret
| |
| |
6<--
|
|<--
7 |
| |
---
Although the CFG l
Hi,
For IVB, I just checked the 3.18-rc1, it has the following patch:
commit c9224faa59c3071ecfa2d4b24592f4eb61e57069
Author: Brad Volkin
Date: Tue Jun 17 14:10:34 2014 -0700
drm/i915: Add some L3 registers to the parser whitelist
Beignet needs these in order to program the L3 cache c
---
src/CMakeLists.txt | 5 ++---
src/git_sha1.sh| 4 ++--
2 files changed, 4 insertions(+), 5 deletions(-)
diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
index 0d22589..9e65856 100644
--- a/src/CMakeLists.txt
+++ b/src/CMakeLists.txt
@@ -110,11 +110,10 @@ SET(CMAKE_C_FLAGS "-DHAS_OCLIc
`gpgpu->binded_buf[i]->offset != 0' failed.
Segmentation fault (core dumped)
Sometimes this kind of assert may have relationship with multi-thread
bug.
On 四, 2014-10-23 at 06:12 -0700, yan.w...@linux.intel.com wrote:
> Update:
> Under root account, it will crash.
> Under user account(e.g. yanwa
Ok. So this is likely a BYT T platform only issue, it may trigger
a GPU hang at OCL_ImageProc/Filter2D.Mat/256. You can use dmesg to
check whether there was a hang.
On Thu, Oct 23, 2014 at 06:04:47AM -0700, yan.w...@linux.intel.com wrote:
> In my baytail T platform, it is still exist.
>
> > Befo
Update:
Under root account, it will crash.
Under user account(e.g. yanwang), it will pass.
> In my baytail T platform, it is still exist.
>
>> Before leaving office, I have built the latest beignet. I will try it
>> again tomorrow.
>> Thanks.
>>
>> Yan Wang
>>
>>> Can't reproduce it here. Did you
In my baytail T platform, it is still exist.
> Before leaving office, I have built the latest beignet. I will try it
> again tomorrow.
> Thanks.
>
> Yan Wang
>
>> Can't reproduce it here. Did you use the latest git master?
>>
>> The latest git version is
>> commit 33a043e903f9d61445d3fb7972becf83c
Before leaving office, I have built the latest beignet. I will try it
again tomorrow.
Thanks.
Yan Wang
> Can't reproduce it here. Did you use the latest git master?
>
> The latest git version is
> commit 33a043e903f9d61445d3fb7972becf83c2842276
> Author: Junyan He
> Date: Thu Oct 23 13:09:00 2
Hi,
As you maybe know, any application which uses beignet OpenCL
implementation crashes on Ivy Bridge hardware when using vanilla
3.17.1 kernel.
I guess it's due to batchbuffer security and patch to disable
batchbuffer security is required, but guys, it fails since 3.16, and
3.17 was released quit
And please also give the platform id in the report, such as
device_name "Intel(R) HD Graphics IvyBridge GT2"
device_opencl_c_version "OpenCL C 1.2 beignet 0.9 (git-33a043e)"
The above info should be helpful.
Thanks,
Zhigang Gong.
On Thu, Oct 23, 2014 at 05:52:43PM +0800, Zhigang Gong wrote:
> Can
Can't reproduce it here. Did you use the latest git master?
The latest git version is
commit 33a043e903f9d61445d3fb7972becf83c2842276
Author: Junyan He
Date: Thu Oct 23 13:09:00 2014 +0800
Add the disasm support for Gen8
Signed-off-by: Junyan He
Reviewed-by: Zhigang Gong
Re
Hi, All,
I found one possible bug for review.
if run the following:
./opencv_test_imgproc --gtest_filter=OCL_ImageProc/Filter2D.Mat*.
OCL_ImageProc/Filter2D.Mat/256 failed and continue. But the whole test
flow will crash in OCL_ImageProc/Filter2D.Mat/257:
[ FAILED ] OCL_ImageProc/Filter2
Thanks.
On Thu, Oct 23, 2014 at 02:59:14AM -0700, yan.w...@linux.intel.com wrote:
> Done.
>
> > Sure. I will try again and file it.
> >
> >> I just tried on my IVB machine, and it could be reproduced easily, just
> >> need
> >> to unset the environment variable.
> >> And this is a platform indepe
Done.
> Sure. I will try again and file it.
>
>> I just tried on my IVB machine, and it could be reproduced easily, just
>> need
>> to unset the environment variable.
>> And this is a platform independent issue. The backend is just a
>> compiler.
>> The compiler could complain warnings or
>> error
Sure. I will try again and file it.
> I just tried on my IVB machine, and it could be reproduced easily, just
> need
> to unset the environment variable.
> And this is a platform independent issue. The backend is just a compiler.
> The compiler could complain warnings or
> errors but should never
I just tried on my IVB machine, and it could be reproduced easily, just need
to unset the environment variable.
And this is a platform independent issue. The backend is just a compiler.
The compiler could complain warnings or
errors but should never crash. If you met a crash in gcc or clang interna
I am not sure whether it is platform issue. So far I have only Baytail T
platform no others. The assertion will not be hit now. Still need me
submit bug? I could do it.
It may be the bug of OpenCL implementation of opecv hough lines for fail
case. Need more investigation.
> Yan,
>
> No matter you
I confirmed it is fixed.
Thanks.
> Yan,
>
> How about this one?
>
> On Thu, Oct 23, 2014 at 12:29:34AM -0700, yan.w...@linux.intel.com wrote:
>> Sorry, I should update beignet not opencv.
>> I will try.
>>
>> Yan Wang
>>
>> > You commit seems not to be merged yet.
>> > If merged, I will try.
>> >
Yan,
How about this one?
On Thu, Oct 23, 2014 at 12:29:34AM -0700, yan.w...@linux.intel.com wrote:
> Sorry, I should update beignet not opencv.
> I will try.
>
> Yan Wang
>
> > You commit seems not to be merged yet.
> > If merged, I will try.
> > Thanks.
> >
> > Yan Wang
> >
> >> Did you try la
Yan,
No matter you set or not set the test data. It should never crash beignet
in such an assertion. That's the point. Anyway, we will look into it.
On Thu, Oct 23, 2014 at 01:40:28AM -0700, yan.w...@linux.intel.com wrote:
> My previous mail may misunderstand you. Sorry.
> After I cloned opencv_e
My previous mail may misunderstand you. Sorry.
After I cloned opencv_extra git and set path rightly, this assertion will
not be hit But 1 case of 8 will failed because dst size should be 1x60 and
is 1x61 in fact. I am trying to debug it. The other 7 cases passed.
> Yan,
>
> It's still an valid bu
Yan,
It's still an valid bug. Beignet should never hit such an assertion
if everything works well. Could you open a bug on fd.o for beignet module?
You may need to specify the git-hash of the beignet and the opencv you are
using. And you need to specify that don't set the opencv test data environm
After cloned related test data, it will not crashed.
Sorry.
> Hi, All,
> This case will crash as the following:
> [ RUN ] OCL_Imgproc/HoughLines.RealImage/0
> ASSERTION FAILED: defIt != bbIt->second->end() && defIt->second != NULL
> at file /home/yanwang/beignet/backend/src/ir/value.cpp
Thanks for your comment.
Actually, no need to add such a comment, as it is not special case.
We haven't add all the other instructions into this list because all of them
haven't use CMP/SEL/... in it.
Unaligned store is just one of those instructions, and you can check the
gen_insn_selection.cpp,
Hi, All,
This case will crash as the following:
[ RUN ] OCL_Imgproc/HoughLines.RealImage/0
ASSERTION FAILED: defIt != bbIt->second->end() && defIt->second != NULL
at file /home/yanwang/beignet/backend/src/ir/value.cpp, function
gbe::ir::LiveOutSet::RegDefSet& gbe::ir::LiveOutSet::getDefS
LGTM.
Add comments for why unaligned store instruction not excluded before push,
please.
Thanks.
Luo Xionghu
Best Regards
-Original Message-
From: Beignet [mailto:beignet-boun...@lists.freedesktop.org] On Behalf Of
Zhigang Gong
Sent: Thursday, October 23, 2014 2:26 PM
To: beignet@lists.
LGTM, thanks.
-Original Message-
From: Beignet [mailto:beignet-boun...@lists.freedesktop.org] On Behalf Of
Zhenyu Wang
Sent: Thursday, October 23, 2014 3:19 PM
To: beignet@lists.freedesktop.org
Subject: [Beignet] [PATCH v2 5/5] Fix AUX buffer for page alignment
Apply ALIGN() for aux buff
LGTM, thanks.
-Original Message-
From: Beignet [mailto:beignet-boun...@lists.freedesktop.org] On Behalf Of
Zhenyu Wang
Sent: Thursday, October 23, 2014 3:19 PM
To: beignet@lists.freedesktop.org
Subject: [Beignet] [PATCH v2 1/5] Make use of write enable flag for mem bo map
Use drm/intel o
Sorry, I should update beignet not opencv.
I will try.
Yan Wang
> You commit seems not to be merged yet.
> If merged, I will try.
> Thanks.
>
> Yan Wang
>
>> Did you try latest git master?
>> It should has been fixed by commit:
>> commit 5417a7e2afb3f2043dbb8abbaebca5115b87ab70
>> Author: Zhigang
You commit seems not to be merged yet.
If merged, I will try.
Thanks.
Yan Wang
> Did you try latest git master?
> It should has been fixed by commit:
> commit 5417a7e2afb3f2043dbb8abbaebca5115b87ab70
> Author: Zhigang Gong
> Date: Tue Oct 21 21:03:46 2014 +0800
>
> GBE: add basic PHINode s
Did you try latest git master?
It should has been fixed by commit:
commit 5417a7e2afb3f2043dbb8abbaebca5115b87ab70
Author: Zhigang Gong
Date: Tue Oct 21 21:03:46 2014 +0800
GBE: add basic PHINode support in legalize pass.
Lack of the incomplete PHINode support currently.
After this
Fix libdrm warning about unmap the unmapped buffer, as when
batch terminate we will try to do unmap again, but batch has
already been unmapped.
Signed-off-by: Zhenyu Wang
---
src/intel/intel_batchbuffer.c | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/src/intel/intel_batch
Instead of mmap, use pread/write interface for bo read/write with
optimized operations.
Result on one mem bandwidth benchmark for buffer enqueue read/write on HSW.
Host to device copy - workgroup_size=512
old code with mmap new code with pread/pwrite
WG=512 SZ=
Almost all 64bit related instructions and unaligned load
instruction should be complex instruction. We need to exclude
them from simple block.
Signed-off-by: Zhigang Gong
---
backend/src/backend/gen_insn_selection.cpp | 28
1 file changed, 28 insertions(+)
diff --gi
On recent kernel with full PPGTT support, we can possibly bind buffer
offset with 0, but intel_gpgpu_check_binded_buf_address() always thinks
it's invalid, which is not true. So simply remove the check.
Signed-off-by: Zhenyu Wang
---
src/intel/intel_gpgpu.c | 9 -
1 file changed, 9 delet
Apply ALIGN() for aux buffer size from beginning has no effect.
Move to the end of all state offsets set for page alignment.
v2: Update comments
Signed-off-by: Zhenyu Wang
---
src/intel/intel_gpgpu.c | 9 ++---
1 file changed, 6 insertions(+), 3 deletions(-)
diff --git a/src/intel/intel_gp
Use drm/intel optimization for mem bo mapping in case of read or write.
So we could be possibly waiting less.
This also adds 'map_flags' check in clEnqueueMapBuffer/clEnqueueMapImage
for actual read or write mapping.
But currently leave clMapBufferIntel untouched which might break ABI/API.
v2: F
Hi, All,
Today I tried the latest opencv code (3.0 master) and found morph perf
case will crash when test CV_8UC1 type only.
The case number is the following:
opencv_perf_imgproc
--gtest_filter=OCL_MorphologyExFixture_MorphologyEx.MorphologyEx/4.
Its operation type is MORPHY_GRADIENT.
I d
40 matches
Mail list logo
