Interpreter has difficulty with signed numbers

[susan-garry]

When I run a simple example with negative numbers through a 2d-matrix multiplier, I get two different outputs depending on if I use verilator or the interpreter (although the result of the multiplication is the same).

This is the example:

{
  "out": {
    "data": [[0, 0], [0, 0]],
    "format": {
      "numeric_type": "bitnum",
      "is_signed": true,
      "width": 32
    }
  },

  "A": {
    "data": [[1, 0], [-1, 1]],
    "format": {
      "numeric_type": "bitnum",
      "is_signed": true,
      "width": 32
    }
  },

  "B": {
    "data": [[1, 0], [1, 1]],
    "format": {
      "numeric_type": "bitnum",
      "is_signed": true,
      "width": 32
    }
  }
}

These are the final values in memory when I run it through verilator:

"memories": {
    "A": [
      [
        1,
        0
      ],
      [
        -1,
        1
      ]
    ],
    "B": [
      [
        1,
        0
      ],
      [
        1,
        1
      ]
    ],
    "out": [
      [
        1,
        0
      ],
      [
        0,
        1
      ]
    ]
  }

And these are the final values when I run it through the interpreter:

"memories": {
    "main": {
      "A": [
        [
          1,
          0
        ],
        [
          4294967295,
          1
        ]
      ],
      "B": [
        [
          1,
          0
        ],
        [
          1,
          1
        ]
      ],
      "out": [
        [
          1,
          0
        ],
        [
          0,
          1
        ]
      ]
    }
  }

Since is_signed is true for all arrays, it seems like the interpreter may not be handling signed numbers.

[EclecticGriffin]

If you don't pass the --raw flag, the interpreter will just output an unsigned int interpretation of all numbers. With that flag fud should do some fun data formatting. See https://github.com/cucapra/calyx/blob/master/fud/fud/stages/interpreter.py#L162

There are also some examples in the runt test suite for the interpreter.

I should perhaps add that flag to the default config for fud.

Anyway, if running with that flag doesn't fix things then there is likely an error in fud's post-interpretation data processing code

[sampsyo]

Interesting point! Yeah, maybe that should indeed be the default… if for no other reason than "Verilator and interpreter modes match" is a pretty desirable property.

[susan-garry]

Hmm. When I try something like fud e --raw --to interpreter-out -s verilog.data calyx-tutorial/2d-matmul1.json calyx-tutorial/2d-matmul.futil, I get this error message: unrecognized arguments: --raw. I also can't find any documentation for this option. Where might I find an example of how to use the --raw flag?

[sampsyo]

Good question—in fact, --raw is an argument to the actual interpreter executable. So if you were to run that directly (not through fud), you would use --raw. To pass this through fud, you'll need to use the interpreter.flags config option. So something like fud e -s interpreter.flags '--raw' ... or similar.

[EclecticGriffin]

@susan-garry, is this issue still happening? If so, could you send me the program & data so I can investigate