I allocate memory using dma_alloc_coherent and receive video frames via PCIe/DMA.
How can I pass this memory to GPU/Cuda without cudaMemcpy?
YES, but you will need to map it with EGL interface.
We have an example to demonstrate this.
Please check /usr/src/tegra_multimedia_api/samples/12_camera_v4l2_cuda for more information.