McPAT 1.3 Modified.

What's modified?

1. Initialization phase and computation phase are decoupled.

   • Initialization of a processor is done directly in the instantiation of it.

   • Area computing is left unchanged in initialization (a chip is just a chip, not a transformer)

   • Power computing is picked out into a new function, compute.

2. Consecutive simulation is enabled.

   • Modified McPAT can initialize for just once and keep computing new power stats upon new XML files.

3. 2025.2.10 Output Format is Modified for Hotspot

   All power breakdowns are summarized.

4. 2025.2.11 Interface for Custom Blocks.

   A very simple interface added. You can see two new files "custom_block.cc" & "custom_block.h", which contains a very simple definition. An example is instantiated as "CIM_SRAM" under "processor" at "processor.h" & "processor.cc".

How to use?

• The command of original McPAT is still supported.

  mcpat -infile <name> -print_level <0~5> -opt_for_clk <0/1>

• To enable consecutive simulation, use command -trace.

  So the command becomes:

  mcpat -trace -infile <name> -print_level <0~5> -opt_for_clk <0/1>

  After simulating upon the initial XML file, McPAT waits for you to input a new filename. This continues until you input "exit" and ends the program.

  If you input "repeat", McPAT computes upon the last XML interface.

• Cutsom Blocks

  Now, you can already see the power stat of "CIM_SRAM" whatever benchmark you're using (but the value is crazy, since the detailed computation is not specified).

  To write a more realistic computation for it, you have to read three parts:

  1. The simple definition in "custom_block.h".

  2. The instantiation of it in "processor.h".

  3. The calculation for it in "processor.cc". It's at the beginning of void Processor::compute(ParseXML *fresh_XML).