Question about CUDABuffer_<T>

[lanfeiying]

Hi,
I am new to CUDA programming and learning this open-source project.
Recently I was reading the source code of this project. I have a question about it. I will appreciate it if you can read this issue and answer my question.

In libvis/cuda/, the basic data structures are defined in the files under this directory.

For /applications/badslam/src/badslam/cuda_depth_processing.cu, the function ComputeNormalsCUDA has a argument CUDABuffer_<T>* out_normals. I can understand this step: the pointer of out_normals is passed into this function.

When this function launches the kernel function ComputeNormalsCUDAKernel(), it pass *out_normals as an object into the kernel function. However, the kernel function receives an object as input in its argument list, instead of a reference. My question is: how could the modifications on *out_normals in the kernel function affect the out_normals?

In other words, in the kernel function, it computes the normals and modifies the corresponding element of out_normals by using out_normals(y, x) = xxx, but it is not the reference of the object. It is a copy of the object.

Is it because when the object out_normals is passed into the kernel function, the copy of *out_normals copied its data head address, instead of allocating new memory space in GPU? So that when the copy of *out_normals use the operator out_normals(y, x), it modified the corresponding data with respect to the original data head pointer address?

Thank you very much for your time to read this question!

[puzzlepaint]

Is it because when the object out_normals is passed into the kernel function, the copy of *out_normals copied its data head address, instead of allocating new memory space in GPU? So that when the copy of *out_normals use the operator out_normals(y, x), it modified the corresponding data with respect to the original data head pointer address?

Yes, that is exactly how it works.

[puzzlepaint]

To elaborate a bit, there are two classes CUDABuffer and CUDABuffer_ that behave quite differently. The former one, CUDABuffer, is a bit higher-level and 'owns' the GPU buffer that it allocates. It is meant to be used from CPU code. For historical reasons (working with old CUDA versions), I could not pass such higher-level objects to CUDA code, since there had been incompatibilities. So I needed a light-weight helper class to access the same functionality from CUDA code. This is what CUDABuffer_ does. It is a simple wrapper over a few attributes of the buffer, and it can be retrieved with CUDABuffer::ToCUDA(). Copying the CUDABuffer_ object merely creates a copy of the pointer to the buffer, not a copy of the buffer itself.

[lanfeiying]

Thank you very much for your answer!
I have another question about applications/badslam/src/badslam/cuda_depth_processing.cu.
For the CUDA kernel function ComputeMinMaxDepthCUDAKernel(), it is a template function. The template arguments are block_width and block_height.

In ComputeMinMaxDepthCUDA(), this function lauch the above CUDA kernel function.
However, I see it passes the template argument block_width and block_height without pre-definition.

I searched these two variables in whole project, and it only appears here. These two variables are not defined, how does c++ compiler recognize the two variables?

[puzzlepaint]

block_width and block_height are defined by the macro CUDA_AUTO_TUNE_2D_TEMPLATED(), which calls CUDA_AUTO_TUNE_2D_BORDER_TEMPLATED(), in libvis/src/libvis/cuda/cuda_auto_tuner.h.