How to perform GPU memory copy.

[npkhgl]

Hi, I got some troubles about GPU memory copy.
rgb24.PlanePtr().Export(surface_tensor.data_ptr(), w * 3, gpuID) # this works fine for me
However, when I tried to use pycuda to perform the GPU memory copy, it failed. The codes I used:

import pycuda.autoinit
import pycuda.driver as cudrv
texbufs_memcpy = cudrv.Memcpy2D()
# texbufs_memcpy.width_in_bytes = w * 3
texbufs_memcpy.width = w * 3
# texbufs_memcpy.height = h
texbufs_memcpy.set_src_device(rgb24.PlanePtr().GpuMem())
texbufs_memcpy.set_dst_device(surface_tensor.data_ptr())
texbufs_memcpy(aligned=True)

Could you give me some help? Thanks a lot.

[rarzumanyan]

Hi @npkhgl

Thanks for bringing this up.
I'm planning to add one more unit test for Surface class which will test (and illustrate) how use pycuda for memcpy.
Will update you a bit later as I test this on my machine.

[rarzumanyan]

@npkhgl

Please check out issue_297 branch in which I've added test_PySurface.py file.
It shows how to use pycuda API for memcpy.

[npkhgl]

@rarzumanyan
Thanks for your effort to write an example code. I read the codes in test_PySurface.py, but I still have problems to use pycuda API for memcpy. I found that memcpy between pytorch tensors is correct, however memcpy between pytorch tensor and surface_rgb is failed. Here are my codes, Please correct me if something wrong.

import PyNvCodec as nvc 
import torch 
import pycuda.driver as cuda
from pycuda import autoinit

w = 1280 
h = 720
gpuID = 0
surface_rgb = nvc.Surface.Make(nvc.PixelFormat.RGB, w, h, 0)
sp = surface_rgb.PlanePtr()
tensor = torch.zeros((h, w, 3), dtype=torch.uint8).cuda() + 1
sp.Import(tensor.data_ptr(), w*3, 0)
check_tensor = torch.zeros((h, w, 3), dtype=torch.uint8).cuda() 
test_tensor = torch.zeros((h, w, 3), dtype=torch.uint8).cuda() 

# something wrong when memcpy between tensor and surface_rgb
memcpy = cuda.Memcpy2D()
memcpy.width_in_bytes = w * 3
memcpy.height = h
memcpy.src_pitch = w * 3
memcpy.set_src_device(sp.GpuMem())
memcpy.set_dst_device(check_tensor.data_ptr())
memcpy(aligned=True)
print(check_tensor.equal(tensor))   # False

# memcpy performs correctly between tensors
memcpy.set_src_device(tensor.data_ptr())
memcpy.set_dst_device(test_tensor.data_ptr())
memcpy(aligned=True)
print(test_tensor.equal(tensor))   # True

# Export performs correctly
check_tensor[:,:,:] = 0
sp.Export(check_tensor.data_ptr(), w*3, 0)
print(check_tensor.equal(tensor))  # True

[rarzumanyan]

@npkhgl

At least one thing I can see is missing destination pitch.
VPF uses pitched allocation, it's actually illustrated in test_PySurface.py.
You just need to swap source and destination members in Memcpy2D class instance:

VideoProcessingFramework/Tests/test_PySurface.py

Lines 124 to 132 in d870bfb

	memcpy_2d = cuda.Memcpy2D()
	memcpy_2d.width_in_bytes = src_plane.Width() * src_plane.ElemSize()
	memcpy_2d.src_pitch = src_plane.Pitch()
	memcpy_2d.dst_pitch = self.nvDec.Width()
	memcpy_2d.width = src_plane.Width()
	memcpy_2d.height = src_plane.Height()
	memcpy_2d.set_src_device(src_plane.GpuMem())
	memcpy_2d.set_dst_device(dst_plane)
	memcpy_2d(self.cuda_str)