[AIE2] Fix memory access cycle for TM instructions

[katerynamuts]

Exploring the architecture design, it was found that the instructions that store to and load from the tile memory (TM) map access the TM at cycle 4 instead of cycle 5. This PR changes MemoryCycle in AIE2Schedule.td file from 5 to 4 for such instructions.

The fix changes some tests. After checking the changed tests with @martien-de-jong, it became clear that changing the memory cycle changes the first and last memory cycle for the TM instructions from 5 to 4 which are used in AIEBaseInstrInfo::getConservativeMemoryLatency used in MaxLatencyFinder. As a result, in the tests an additional nop is added because of the worst-case analysis used for inter-block scheduling.

[gbossu]

As discussed offline, that extra might hurt QoR. I'd be happy to measure that impact.

If the results aren't good, we could extend

  int getMinFirstMemoryCycle() const;
  int getMaxFirstMemoryCycle() const;
  int getMinLastMemoryCycle() const;
  int getMaxLastMemoryCycle() const;

into e.g.

  int getMinFirstMemoryCycle(unsigned AddrSpace = 0) const;
  int getMaxFirstMemoryCycle(unsigned AddrSpace = 0) const;
  int getMinLastMemoryCycle(unsigned AddrSpace = 0) const;
  int getMaxLastMemoryCycle(unsigned AddrSpace = 0) const;

Taking the example of AIE2:

• getMinFirstMemoryCycle(AIE2::AddressSpaces::none) should return 4
• getMinFirstMemoryCycle(AIE2::AddressSpaces::DM) should return 5
• getMinFirstMemoryCycle(AIE2::AddressSpaces::TM) should return 4

But then I guess we'll have the same problem because many load/stores will have the default none address space, even though they actually are in DM. That would be quite a big change if we also need to infer the AddressSpaces::DM address space.

[krishnamtibrewala]

Why do you think it will be a big change if we need to infer none to DM ?

Note : access to tile memory is always annotated at intrinsic level so we will always infer that

[krishnamtibrewala]

Hi @katerynamuts did you find this info in the architecture file that you had shared with me a month back ?

[katerynamuts]

Hi @katerynamuts did you find this info in the architecture file that you had shared with me a month back ?

@krishnamtibrewala yes