aditya.kukekar Posted March 4, 2014 Share Posted March 4, 2014 Hey every one..im trying to use open cl functionality in the pyro solver but im getting this error OpenCL Exception: Could not create OpenCL device of type (HOUDINI_OCL_DEVICETYPE): CPU on platform vendor (HOUDINI_OCL_VENDOR): Intel® Corporation (-6) OpenCL Exception: clFinish (-36) OpenCL Exception: Failed to create compute grid. (-4) OpenCL Exception: clCreateProgramWithSource (-34) OpenCL Exception: Failed to create compute grid. (-4) OpenCL Exception: clCreateProgramWithSource (-34) OpenCL Exception: Failed to create compute grid. (-4) OpenCL Exception: clCreateProgramWithSource (-34) OpenCL Exception: clCreateProgramWithSource (-34) OpenCL Exception: clCreateProgramWithSource (-34) OpenCL Exception: clCreateProgramWithSource (-34) OpenCL Exception: Failed to create compute grid. (-4) OpenCL Exception: clFinish (-36) im using dell precision mobile workstation with k4000m quadro nvidia gpu on win 7 64bit v can anyone help me pls Quote Link to comment Share on other sites More sharing options...
aditya.kukekar Posted May 28, 2016 Author Share Posted May 28, 2016 Solved 2 Quote Link to comment Share on other sites More sharing options...
Cockx Posted July 12, 2016 Share Posted July 12, 2016 Hi, How did you fix this problem? I have a simular problem... Thanks! Quote Link to comment Share on other sites More sharing options...
vfxhound Posted January 3, 2017 Share Posted January 3, 2017 Hello, I get the same errors trying to use openCL on my Geforce GTX 1080. Can anyone please tell me how to solve this? Quote Link to comment Share on other sites More sharing options...
Guest tar Posted January 3, 2017 Share Posted January 3, 2017 1 hour ago, vfxhound said: Hello, I get the same errors trying to use openCL on my Geforce GTX 1080. Can anyone please tell me how to solve this? Can you post your opencl info from the About Houdini menu, and, the error you are receiving. Also the driver version. Quote Link to comment Share on other sites More sharing options...
vfxhound Posted January 3, 2017 Share Posted January 3, 2017 Hi Marty, thanks a lot for taking the time to look at my problem. Here's the opencl info: OpenCL Platform NVIDIA CUDA Platform Vendor NVIDIA Corporation Platform Version OpenCL 1.2 CUDA 8.0.0 OpenCL Device GeForce GTX 1080 OpenCL Type GPU Device Version OpenCL 1.2 CUDA Frequency 1835 MHz Compute Units 20 Device Address Bits 64 Global Memory 8192 MB Max Allocation 2048 MB Global Cache 320 KB Max Constant Args 9 Max Constant Size 64 KB Local Mem Size 48 KB 2D Image Support 16384x32768 3D Image Support 16384x16384x16384 OpenCL Platform Extensions: cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_khr_fp64 cl_khr_byte_addressable_store cl_khr_icd cl_khr_gl_sharing cl_nv_compiler_options cl_nv_device_attribute_query cl_nv_pragma_unroll cl_nv_d3d10_sharing cl_khr_d3d10_sharing cl_nv_d3d11_sharing cl_nv_copy_opts OpenCL Device Extensions: cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_khr_fp64 cl_khr_byte_addressable_store cl_khr_icd cl_khr_gl_sharing cl_nv_compiler_options cl_nv_device_attribute_query cl_nv_pragma_unroll cl_nv_d3d10_sharing cl_khr_d3d10_sharing cl_nv_d3d11_sharing cl_nv_copy_opts ====================================== This is my driver version: 21.21.13.7633 (from the device manager, right?) ====================================== This is the error I'm receiving: OpenCL Exception: clEnqueueNDRangeKernel (-4) OpenCL Exception: clEnqueueNDRangeKernel (-4) OpenCL Exception: clEnqueueNDRangeKernel (-4) OpenCL Exception: clEnqueueNDRangeKernel (-4) OpenCL Context error: CL_MEM_OBJECT_ALLOCATION_FAILURE error executing CL_COMMAND_NDRANGE_KERNEL on GeForce GTX 1080 (Device 0). ======================================= These errors only happen when I increase the sim resolution too high. I'm on Windows 10 64x and using Houdini 15.5 Thanks again Marty for your time! Quote Link to comment Share on other sites More sharing options...
dimovfx Posted January 3, 2017 Share Posted January 3, 2017 You should monitor your VRAM. Probably 8GB is not enough for the sim. Quote Link to comment Share on other sites More sharing options...
vfxhound Posted January 3, 2017 Share Posted January 3, 2017 41 minutes ago, sasho78 said: You should monitor your VRAM. Probably 8GB is not enough for the sim. It spikes to 7002MB when the error occurs Quote Link to comment Share on other sites More sharing options...
Atom Posted January 3, 2017 Share Posted January 3, 2017 (edited) I just get out right crashes with no message when using OpenCL with FLIP. OpenCL is kind of a pipe dream IMHO. It gives you the illusion that you get some acceleration but under heavy load it never holds up. It can help speed up the dev of a sim in the early stages but whenever I try to crank out a final version of a sim at a higher resolution, with OpenCL enabled, I get failure. I don't think it is my hardware setup either. I can render under full load for several days with Redshift using 100% of my GPU without crashes. When I try an overnight export with OpenCL enabled I only get a few hundred frames before the system hangs or crashes. Houdini only uses about 30% load on the GPU so I don't think heat or power are the culprit, just an old fashion memory leak bug that has not been squashed. Edited January 3, 2017 by Atom 1 Quote Link to comment Share on other sites More sharing options...
dimovfx Posted January 3, 2017 Share Posted January 3, 2017 Not sure where do you monitor your VRAM but still sounds like it's reaching the limit. Try with a little bigger voxels and see if it works (bet it does). For me OpenCL is faster, but when I reach the final look I turn off OpenCL because VRAM is not limitless and I don't want my few hours sim to fail because of that. 1 Quote Link to comment Share on other sites More sharing options...
vfxhound Posted January 3, 2017 Share Posted January 3, 2017 25 minutes ago, Atom said: I just get out right crashes with no message when using OpenCL with FLIP. OpenCL is kind of a pipe dream IMHO. It gives you the illusion that you get some acceleration but under heavy load it never holds up. It can help speed up the dev of a sim in the early stages but whenever I try to crank out a final version of a sim at a higher resolution, with OpenCL enabled, I get failure. I don't think it is my hardware setup either. I can render under full load for several days with Redshift using 100% of my GPU without crashes. When I try an overnight export with OpenCL enabled I only get a few hundred frames before the system hangs or crashes. Houdini only uses about 30% load on the GPU so I don't think heat or power are the culprit, just an old fashion memory leak bug that has not been squashed. I guess like you said it's best to use it at the early stages to develop the sim and iterate faster. But yeah, higher res seems to always crash/run out of memory. Quote Link to comment Share on other sites More sharing options...
vfxhound Posted January 3, 2017 Share Posted January 3, 2017 (edited) 48 minutes ago, sasho78 said: Not sure where do you monitor your VRAM but still sounds like it's reaching the limit. Try with a little bigger voxels and see if it works (bet it does). For me OpenCL is faster, but when I reach the final look I turn off OpenCL because VRAM is not limitless and I don't want my few hours sim to fail because of that. I monitor it using GPU-Z. And yeah seems to be running out of memory. That doesn't happen in lower res sims. Thanks for your help! Edited January 3, 2017 by vfxhound Quote Link to comment Share on other sites More sharing options...
vfxhound Posted January 3, 2017 Share Posted January 3, 2017 In all fairness though, the smoke container had 107M voxels before it crashed It's a pretty decent limit Quote Link to comment Share on other sites More sharing options...
Guest tar Posted January 3, 2017 Share Posted January 3, 2017 3 hours ago, Atom said: I don't think it is my hardware setup either. I can render under full load for several days with Redshift using 100% of my GPU without crashes. When I try an overnight export with OpenCL enabled I only get a few hundred frames before the system hangs or crashes. Houdini only uses about 30% load on the GPU so I don't think heat or power are the culprit, just an old fashion memory leak bug that has not been squashed. To test the memory leak hypothesis - you can run OpenCL in CPU mode to see if it still fails. The Redshift comparison is irrelevant. Quote Link to comment Share on other sites More sharing options...
Atom Posted January 4, 2017 Share Posted January 4, 2017 (edited) The Redshift comparison is to indicate that my system can run under full GPU load for days without crashing. OpenCL uses the GPU just like Redshift. I just conducted the same test with an overnight export of my waterfall with OpenCL, on the FLIP Solver, turned off. No crash and all my frames were exported. The night before it failed with OpenCL turned on. So that sums it up, there is a bug in Houdini's 15.5.673 OpenCL acceleration code. It does not work with large data sets. Edited January 4, 2017 by Atom Quote Link to comment Share on other sites More sharing options...
Guest tar Posted January 4, 2017 Share Posted January 4, 2017 6 hours ago, Atom said: So that sums it up, there is a bug in Houdini's 15.5.673 OpenCL acceleration code. It does not work with large data sets. Until you test with OpenCL CPU you are simply confirming that you are running out of GPU Ram. We have run tests using OpenCL CPU up to 64GB. What is a large dataset in your opinion? Quote Link to comment Share on other sites More sharing options...
symek Posted January 5, 2017 Share Posted January 5, 2017 10 hours ago, Atom said: The Redshift comparison is to indicate that my system can run under full GPU load for days without crashing. OpenCL uses the GPU just like Redshift. I just conducted the same test with an overnight export of my waterfall with OpenCL, on the FLIP Solver, turned off. No crash and all my frames were exported. The night before it failed with OpenCL turned on. So that sums it up, there is a bug in Houdini's 15.5.673 OpenCL acceleration code. It does not work with large data sets. Redshift uses CUDA, not OpenCL, and it can offload assets from GPU memory on demand (out-of-core rendering), meanwhile Pyro OpenCL can't. This comparison is meaningless. It's not a bug, but limitation. Quote Link to comment Share on other sites More sharing options...
Recommended Posts
Join the conversation
You can post now and register later. If you have an account, sign in now to post with your account.
Note: Your post will require moderator approval before it will be visible.