tsadakane

> R2C and C2R may work. Worked. ([a39e6b2](https://github.com/CERN/TIGRE/pull/423/commits/a39e6b28679a77d2d7e2db885e924e099c8dc352)) ## Results of python test code ### Conditions 0: NVIDIA GeForce GTX 1060 6GB {'name': 'NVIDIA GeForce GTX 1060 6GB', 'devices': [0]}...

@AnderBiguri Do you think we should keep the option to use/not to use GPU for FBP filtering for `filtering` and `FDK` function?

@AnderBiguri Thank you for your reply. I think it would be better to remove the old code if possible. Let me know the result and I can remove the option...

> In your time comments, you are saying that most of the time (7s) is on the FFT maths, and the rest is memory and context management? Sorry for unclear...

I have no idea so far. * Commenting out all but the last line of `filtering` function in `filtering.py`, which is called by `FDK` before Atb is called, the time...

I wrote multi-gpu version and cudaMemcpyAsync version, but as expected, the both were no effect.

Just removing padding from `filtering` function in `filtering.py` reduced 3 sec., so I moved the padding into cuda function. This reduces the amount of memory to transfer. For effective strided...

## About aa51ec6 "CPU" uses CPU for convolution. In "GPU 1", the convolution is processed in a GPU, even if `gpuids` includes multiple GPUs. (=[a39e6b2](https://github.com/CERN/TIGRE/pull/423/commits/a39e6b28679a77d2d7e2db885e924e099c8dc352)) In "GPU 2", the padding...

The table below is the summary of the results above. The result of python is acceptable for me, but the result of matlab is hard to understand. At first, I...

I ran the python test program above after adding `gpuids.devices.pop(-1)` to use only one GPU. Python | | 2x GPU [s], | 1x GPU [s] | |---|---|---| | CPU |...