#2004: Added shlo to ttir conversion pass for all_gather and updated shlo to ttir conversion test cases for all gather

[tapspatel]

Meshes we will support

• 1x2
• 1x8
• 2x4
• 1x32
• 8x4

Multi-device tests consolidation

test/ttmlir/Conversion/StableHLOToTTIR/ccl_ops.mlir

• all_reduce tests for all meshes
• all_gather tests for all meshes

[tapspatel]

One issue I am trying to get around is preventing all the single chip test cases from running on multi chip systems. LIT gives us the ability to do boolean expressions in its REQUIRES: directive

OR

// REQUIRES: feature1 || feature2

AND

// REQUIRES: feature1, feature2

So I was thinking we can set something like this

// REQUIRES: num-chips-1 || num-chips-2

// REQUIRES: num-chips-32

[tapspatel]

and in test/lit.cfg.py , we already do this for multi-chip systems

def set_system_desc_features(system_desc):
    config.available_features.add(system_desc["chip_descs"][0]["arch"])
    if len(system_desc["chip_desc_indices"]) > 1:
        config.available_features.add("multi-chip")
    config.available_features.add(
        "multi-chip-x" + str(len(system_desc["chip_desc_indices"]))
    )

[tapspatel]

we got some tg machine failures related to huge pages. Fixing the machine

[nsmithtt]

Looks good!