This won't get inlined automatically by the compiler, meaning +1 on latency of the inner product for every output pixel due to the extra function call. It's the function_instantiate pragma that we should remove. That pragma is a leftover from the wild '10s of Vivado HLS where tricks like that were actual optimizations (note that there will only ever be one instance of this function since the weights array is fixed, so it is useless)

If function_instantiate is removed, there is then a clash with #pragma HLS PIPELINE .... If this one is removed, the issue is solved, but again, not sure about the impact on resulting rtl.

Looking at the surrounding code there's already a inline recursive pragma before each call to this function. Does that get ignored? Is inlining itself a problem (that messes with the pipeline of the depthwise_conv_1d_buffer_cl that ultimately calls this) or is the issue the existence of additional pipeline directive in the depthwise_product (which should be ignored with a warning, but is promoted to an error in Vitis)? I think we should test if this works better inlined or not, and if not, prevent inlining by removing those pragmas (and perhaps even keeping the one in depthwise_product but with off). Just not with function_instantiate :-)

To be honest, I have no clue about all this. I just have the impression from the error message that the multiple pragma are simple mutually exclusive with Vitis (obviously no issue with Vivado).

I am willing to dig a bit in all this and do some tests, but not sure when I will have the time to do it..