#32313: CRAQMM op

[ssundaramTT]

Ticket

tenstorrent/tt-metal#32313

Problem description

Upon investigating performance of UNPACK and MATH for a single-tile matmul, we observe that [1, 7168] x [7168, 32] is more performant with [32, 32] tile shape compared to with [1 ,32] tile shape, when we expect the opposite. The changes need to be integrated into the existing matmul test: models/demos/deepseek_v3_b1/micro_ops/matmul/kernels/matmul_compute.cpp

What's changed

• Adds new CRAQMM (Custom Replay And Queue Matrix Multiply) kernel optimized for K-dimension reduction
• Uses MOP replay buffers + hardware address counters to reduce RISC overhead

Type of change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Documentation update

Checklist

• All post commit CI passes
• Blackhole Post commit CI passes (if applicable)

[github-actions]

Thank you for your contribution! 🚀
If you want to run metal post-commit tests, you can add the metal-post-commit-tests label to this pull request.
📖 For more information, please refer to our CONTRIBUTING guide.

[github-actions]

🚀 tt-metal post-commit tests

Branch: lpremovic/deepseek
Test Configuration:

• Wormhole tests: ⚪ Disabled
• Blackhole tests: 🟢 Enabled

Test Results:

• Wormhole tests: ⚪ Not run
• C++ post-commit tests: 🎯 Passed - View run
• Blackhole tests: 🎯 Passed - View run
• Device perf regressions tests: 🎯 Passed - View run

🔗 Links

📊 Post-commit workflow: #20529484939

[github-actions]

🚀 tt-metal post-commit tests

Branch: lpremovic/deepseek
Test Configuration:

• Wormhole tests: ⚪ Disabled
• Blackhole tests: 🟢 Enabled

Test Results:

• Wormhole tests: ⚪ Not run
• C++ post-commit tests: 🎯 Passed - View run
• Blackhole tests: 🎯 Passed - View run
• Device perf regressions tests: 🎯 Passed - View run

🔗 Links

📊 Post-commit workflow: #20625927787