get_noc_addr_from_bank_id needed for first eltwise PyKernel

[vtangTT]

This PR implements the last TTKernel op needed by eltwise kernel example

get_noc_addr_from_bank_id is templated in dataflow_api.h on the metal side

template <bool DRAM>
get_noc_addr_from_bank_id(uint32_t bank_id, uint32_t bank_address_offset, uint8_t noc = noc_index) {

Further investigation shows that every kernel written sets DRAM to true.

Similar API like get_noc_addr is templated the same way and this template is not exposed in the tablegen for mlir (would love to know how to expose it), so I will also not expose it.

[vroubtsovTT]

need to deal with the template arg

[vtangTT]

I added a unit test for the ttkernel -> emitc conversion. I'm not sure if this belongs in this PR and if so, should I add it for the other ttkernel ops that were recently added?

[vroubtsovTT]

Thank you for testing your changes! I propose some small tweaks. This existing test is a little hard to read because it abuses the [[C:.*]] to capture some substrings into C that is never used. It is also not necessary to match the end of the line. I suggest we make things more readable with this version of your change:

    %bank_id = arith.constant 1 : i32
    %addr_offset = arith.constant 262400 : i32
    %noc_addr = "ttkernel.get_noc_addr_from_bank_id"(%bank_id, %addr_offset) : (i32, i32) -> !ttkernel.noc_addr
    // CHECK: = emitc.call_opaque "get_noc_addr_from_bank_id"({{.*}}) {template_args = [#emitc.opaque<"true">]}

This makes the call more readable and we are not matching anything we don't need.

(We also have test/ttmlir/Translate/TTKernel/ttkernel_{noc, tensix}.mlir that test cpp translation, but I want to clean those up via a separate ticket.)

[vtangTT]

Sounds good, ty for the pointer!

[vroubtsovTT]

cleaner version of ttkernel.mlir test, per conversation

[vroubtsovTT]

Looking good.