Support matrix AB ReadNoSnp in OpenLLC

[zykucas]

OpenLLC Matrix A/B Non-Coherent Read Support (ReadNoSnp)

Summary

This PR adds OpenLLC-side support for Matrix A/B non-coherent reads issued as CHI ReadNoSnp.

Together with the matching HBL2 change, Matrix A/B Get requests can bypass the coherent read path and use ReadNoSnp, reducing snoop/ownership overhead for matrix/GEMM workloads.

Changes

• Added matrixAB metadata to OpenLLC self directory entries.
• Added ReadNoSnp handling in the LLC MainPipe.
• Added self/client directory update behavior for ReadNoSnp.
• Added Matrix A/B replacement handling for cached clean ReadNoSnp lines.
• Added memory read/refill support for ReadNoSnp.
• Connected downstream memory CompData into the RefillUnit.
• Updated ResponseUnit and MemUnit handling for ReadNoSnp / WriteNoSnpFull interactions.
• Updated RequestArb read classification to include ReadNoSnp.
• Added a TestTopMatrix switch:
  • --matrix-ab-nc 1: enable Matrix A/B ReadNoSnp
  • --matrix-ab-nc 0: disable it for comparison

Validation

Basic Regression Tests

Test	Result	Notes
openLLC-test-matrix-trace	PASS	line_trace-64-1024-64-matrix.txt
openLLC-test-l2l3	PASS	finalized at cycle 540000

GEMM Trace Sweep

For openLLC-test-matrix-trace:

The following full GEMM line traces were tested with Matrix A/B ReadNoSnp enabled and disabled.

GEMM Trace	ReadNoSnp on	ReadNoSnp off	Saved cycles	Cycle reduction	Speedup
line_trace-64-256-64-matrix.txt	16,312	17,324	1,012	5.84%	1.062x
line_trace-64-512-64-matrix.txt	27,580	30,618	3,038	9.92%	1.110x
line_trace-64-1024-64-matrix.txt	51,914	59,066	7,152	12.11%	1.138x
line_trace-64-2048-64-matrix.txt	99,820	115,962	16,142	13.92%	1.162x
line_trace-64-4096-64-matrix.txt	195,898	229,754	33,856	14.74%	1.173x
line_trace-64-8192-64-matrix.txt	388,290	457,338	69,048	15.10%	1.178x
line_trace-64-16384-64-matrix.txt	773,822	912,506	138,684	15.20%	1.179x

Overall GEMM sweep result:

• Arithmetic mean speedup: 1.143x
• Arithmetic mean cycle reduction: 12.40%
• Weighted speedup by total cycles: 1.173x
• Weighted cycle reduction by total cycles: 14.76%

The benefit increases with the GEMM trace size and stabilizes around 1.18x for the larger traces.

[Ivyfeather]

This work has been migrated to OpenXiangShan/XSAICache#5. Closing this PR in favor of the consolidated implementation there.