Nvidia Xavier devices: exception thrown during kernel execution on device Xavier #349
Comments
|
Since you have these packages Also, did you not see additional output from the device? It should have printed something (if anything, a recommendation to run under This also probably has nothing to do with running on a Xavier. |
Sorry, that wasn't a verbatim output sequence. I just added the top 4 lines as a MWE, then pasted the error below. I'll re-run this on the device and report more |
|
Yeah, this is happening on a GeForce GTX 1650 system too, with JLL-based CUDA. I've extracted the code which serves as a MWE. By the way, this code block has been unchanged recently, except for a switch from function keepdetections(input::CuArray) # THREADS:BLOCKS CAN BE OPTIMIZED WITH BETTER KERNEL
rows, cols = size(input)
bools = CUDA.zeros(Int32, cols)
@cuda blocks=cols threads=rows kern_genbools(input, bools)
idxs = cumsum(bools)
n = count(bools)
output = CuArray{Float32, 2}(undef, rows, n)
@cuda blocks=cols threads=rows kern_keepdetections(input, output, bools, idxs)
return output
end
function kern_genbools(input::CuDeviceArray, output::CuDeviceArray)
col = (blockIdx().x-1) * blockDim().x + threadIdx().x
cols = gridDim().x
if col < cols && input[5, col] > Float32(0)
@inbounds output[col] = Int32(1)
end
return
end
@inline function kern_keepdetections(input::CuDeviceArray, output::CuDeviceArray,
bools::CuDeviceArray, idxs::CuDeviceArray)
col = blockIdx().x
row = threadIdx().x
if bools[col] == Int32(1)
idx = idxs[col]
@inbounds output[row, idx] = input[row, col]
end
return
endAny idea why this has broken? |
|
Seems to be a problem with count |
|
Fixed with this. But I'm not sure why this started erroring. |
|
Yeah, it's probably this type error you were seeing: |
Describe the bug
Using ObjectDetector.jl (uses CUDA via Flux) on v1.5.0 I hit this kernel issue NVidia Xavier devices
Note that this is one of the NVIDIA jetpack-based devices, that uses a modified CUDA
Manifest.toml
Expected behavior
A clear and concise description of what you expected to happen.
Version info
Details on Julia:
Details on CUDA:
Additional context
The text was updated successfully, but these errors were encountered: