Adding Mi300 to 2.x

[coleramos425]

Overview

This pull request adds Mi300 support from our mi300 branch into the 2.x redesign. I've tested these changes on all of the Mi300 varients that we'll be supporting.

From now on, our "source of truth" for supported SoC's is the __supported_archs property of the main Omniperf() class. This value is passed to all sub-classes that need this arch -> GPU name mapping. https://github.com/AMDResearch/omniperf/blob/131429087b1fe7ec6997c554c7e2f9069d8a057b/src/omniperf_base.py#L50-L57

Major Changes

• New class definitions for MI300A_A0, MI300X_A0, and MI300A_A1 + MI300X_A1
• Support for rocprofv2
• Standardization of rocprofiler csv headers to maintain full backward compatibility. This way standard headers can be used throughout Omniperf https://github.com/AMDResearch/omniperf/blob/131429087b1fe7ec6997c554c7e2f9069d8a057b/src/utils/utils.py#L246-L271
• New MachineSpecs() properties with improved naming conventions and additional fields for Mi300 support (i.e., compute_partion and memory_partition) https://github.com/AMDResearch/omniperf/blob/131429087b1fe7ec6997c554c7e2f9069d8a057b/src/utils/specs.py#L38-L65
• Options to override detected machine specs via --specs-correction. At any time, peek current machine specs via --specs
• New Memory Chart added to CLI view
• Redesigned L2 Cache per channel implementation improves maintainability via dynamic detection of num. of channels in SoC, i.e. https://github.com/AMDResearch/omniperf/blob/df6471db54f3f2cd97dfa69983a14f15957ac43d/src/omniperf_soc/analysis_configs/gfx940/1800_L2_cache_per_channel.yaml#L136-L151
• cli_style property which can define infographic style for CLI analysis view
• Improved --list-metrics implementation to highlight both individual metrics in addition to sections that can be toggled, i.e.

Notes

• You'll see we withheld the plotext submodule implemented in the mi300 branch. This will not be merged upstream until the PyPi module is updated. Current ETA is ~ March https://github.com/piccolomo/plotext/issues/196#issuecomment-1890988471
• Since we're withholding plotext module, some CLI elements are stuck in table view (i.e. l2 cache per channel, mem chart, etc.). Pending the availability of the updated module, we may pivot to other existing Python utilities.
• Roofline analysis in Mi300 has been temporarily disabled. We still need to verify this feature via independent tests.

Special thanks to @feizheng10 for his contributions in this PR.