gFluidSurface fails on GTX 970 on Win10..

[oscarbg]

testing new SDK 1.1 reléase on Windows 10 GTX 970 and 391.35 driver..
all samples work ok except: gFluidSurface.exe
it's due to 4gb GPU memory not enough?
thanks..
error log:

FLUID SYSTEM, CUDA ERROR:
Launch status: unspecified launch failure
Kernel status: no error
Caller: FluidSystem::TransferFromCUDA
Call: cuMemcpyDtoH
Args: FPOS
Error. Application will exit.

full log below:

gFluidSurface.exe
Reading TGA: arial.tga
Creating OptiX context.
Setting Ray Generation program.
Loading: optix_trace_primary.ptx, trace_primary
Loading: optix_trace_primary.ptx, exception
Setting Miss program.
Loading: optix_trace_miss.ptx, miss
Creating random number buffer.
Creating Mesh intersection programs.
Loading: optix_mesh_intersect.ptx, mesh_intersect
Loading: optix_mesh_intersect.ptx, mesh_bounds
Creating Volume intersection programs.
Loading: optix_vol_intersect.ptx, vol_intersect
Loading: optix_vol_intersect.ptx, vol_levelset
Loading: optix_vol_intersect.ptx, vol_deep
Loading: optix_vol_intersect.ptx, vol_bounds
Creating OptiX Main Group and BVH.
PERF_INIT: No CPU markers.
PERF_INIT: Disabling GPU markers. Did not find nvToolsExt64_1.dll.
PERF_INIT: Disabling perf. No CPU or GPU markers.
Starting GVDB Voxels. ver 1.11
Creating Allocator..
Creating Scene..
Creating CUDPP..
Starting Fluid System.
Add Particles. Density: 400.000000, Spacing: 1.393426, PDist: 0.008008
CUDA Config:
Pnts: 1500000, t:2930x512=1500160, Size:1500160
Grid: 1203052, t:2350x512=1203200, bufGrid:0, Res: 134x67x134
Creating screen buffer. 1024 x 768
Loading: optix_trace_surface.ptx, trace_surface
Loading: optix_trace_surface.ptx, trace_shadow
Adding GVDB Volume to OptiX graph.
Validating OptiX.
Update GVDB Volume.
Rebuild Optix.. Done.
Running..
OptiX buffer size: 1024, 768
FLUID SYSTEM, CUDA ERROR:
Launch status: unspecified launch failure
Kernel status: no error
Caller: FluidSystem::TransferFromCUDA
Call: cuMemcpyDtoH
Args: FPOS
Error. Application will exit.

[davidxmiller]

I see the exact same output on my NVIDIA Quadro M1200

[TuotuoLi]

I am having the same problem here.

[YuxingQiu]

I have the same problem with you, and even the error log is the same. I'm testing on Windows 10 GTX 1070 Ti and the driver 411.63. My GPU memory is 8gb, while the program still crashed. T^T

[geekzzz]

I build and run it on 1080Ti, win10+vs2015+cuda 9.1+optix 4.0.0. and I have the same problem here.

[mrvanr]

Built on w10 with a Quadro M4000 using vs2015 and cuda 9.0 and getting the same error. The other samples work normally.

[kukuhbr]

Same problem with:
GTX1060, win10+vs2015+cuda9.0+optix 4.0.0+driver 385.54

[shenberg]

Same problem with 1080Ti on ubuntu 18.04.

I did some comment-out debugging and noticed that the crash is caused by CountingSortFullCUDA - specifically, if I comment-out the call to the CUDA kernel doing the copying sort, I won't get a crash on TransferFromCUDA() (of course, the fluid simulation is not really running any more). I suspect a memory corruption but the code inside the kernel hasn't changed between 1.0 and 1.1 so I'm unsure as to the difference between the two versions.

One thing that I thought made a difference was that the number of threads per block was increased from 192 to 512 but changing it back did not make a difference.

[shenberg]

Found and fixed the bug. PR: #68