500 ms kernel execution latency on ZynqMP

[apinas]

Hello, I have created a dummy kernel to see the XRT latency overhead. It just takes an argument as input but does not write back to memory. Basically:

void basic_kernel(int input_val) {
        int temp;
        temp = input_val + 1;
    }

In my host program I have the following code:

auto krnl = xrt::kernel(device, uuid, "basic_kernel");
xrt::run run(krnl);
run.set_arg(0, dummy_data);
auto t_start = std::chrono::steady_clock::now();
run.start();
run.wait();
auto t_end = std::chrono::steady_clock::now();
auto elapsed_ms = std::chrono::duration_cast<std::chrono::milliseconds>(t_end - t_start).count();

I run this in a loop, and I get ~500ms of elapsed time. I am using the following versions:

Version              : 2.16.0
Branch               : 2023.2
Hash                 : 2865a62b6a417dea523d2d5646154aa94a2cbc28
Hash Date            : 2023-10-06 21:08:57
ZOCL                 : 2.16.0, 2865a62b6a417dea523d2d5646154aa94a2cbc28

[apinas]

Ok so turns out that the driver/runtime is not setting the MER register in the AXI Interrupt Controller, therefore interrupts are not raised when the kernels stop executing and the must be some kind of polling fallback to check this, thats why it takes ~500ms.

If I manually enable the MER register, it works as expected with lower execution time.

Why it does not enable the MER, while for example the IER register is enabled? I dont know. If someone know how to trace this down, help is welcomed