In the pico sample program, how does pevb_dma_h2c2h_multi() kick the CUDA kernel and let the CPU know about it after it finishes?

[TatsuhiroOshima]

In the pico sample program, how does pevb_dma_h2c2h_multi() kick the CUDA kernel and let the CPU know about it after it finishes?

		ret = pevb_dma_h2c_single(pevb, src_addr, 0, len_chunk);
		if (ret)
			goto unlock;

		ret = pevb_dma_c2h_single(pevb, dst_addr, 0, len_chunk);
		if (ret)
			goto unlock;

We want to control the CPU driver from the CUDA kernel.
Do I need to implement my own FPGA functionality like mailbox for this?

[pateldipen1984-nv]

It just post the data and let CUDA handle it, but there is no notification from cuda driver if it had processed the data if that is what you are asking.

[pateldipen1984-nv]

The DMA of FPGA notifies fpga driver that DMA has completed through pevb_irq_handler and at that point control returns to the userspace. The GPU part I mentioned above is where there is no notification if GPU had processed/consumed the data in this sample code.