How can I validated a mapping found by Timeloop in ZigZag?

[morainer]

How can I validated a mapping found by Timeloop in ZigZag? I can only set the spatial mapping and temporal mapping respectively. But I find that I can not allocate the memory level at the exact loop of the mapping as I wish. So I can not get the result that Zigzag can obtain energy that are approximate to the Timeloop.

[asyms]

Hi,

I don't understand your question completely. Could you provide a MWE?

The ZigZag spatial and temporal mapping representations work on a memory-level basis, so if you hardcode the loops in each level, that should achieve what you're after?

Kind regards, Arne

[morainer]

Hi,

Sorry, maybe my expression is wrong.

When we obtain a mapping in TimeLoop, I want to transform this timeloop mapping into Zigzag way, so that I can run the same mapping both in TimeLoop and Zigzag and then get the results of them. However, when I divide the timeloop mapping into spatial and temporal mappings and input them into zigzag and run it in debug mode, I observe that the loops on each memory level do not align with the timeloop mapping. So the energy results of timeloop and zigzag are quite different.

Therefore, I cannot reproduce the result that Zigzag can obtain energy that are approximate to the Timeloop, which mentioned in the verification paper of zigzag. As shown in the figure below.

I wonder that if there are other ways to solve such problem？

[asyms]

I understand what you're trying to achieve, but can you elaborate on "I observe the loops on each memory level do not align with the timeloop mapping"? A MWE will help me pinpoint the difference.

[morainer]

tlzz.zip Okay, here is an example that I try converting a Timeloop mapping into a Zigzag mapping, but I'm not sure if the conversion is correct. If it is wrong, could you please help me point out the error. Thank you very much.

[JiacongSun]

Hi morainer,

I've noticed several errors in your configurations:

• The processing element (PE) counts differ between the two frameworks: 168 PEs for Timeloop and 96 PEs for Zigzag.
• The PE array dimension definitions are different: 2D (14x12) for timeloop, 4D (3x4x4x2) for zigzag.
• You've used an incorrect bandwidth value for memory instances in Zigzag. Timeloop's bandwidth unit is per operand, whereas Zigzag's is per bit. To rectify this, you should multiply the value used for Timeloop by the operand bit precision (32 in your case) for Zigzag.
• The temporal mapping defined for both frameworks are different. It seems you may have misinterpreted some output dimensions from Timeloop. For instance, in Timeloop, the dimension Q represents the input/output kernel size. Therefore, for Q in [0:4) in Timeloop's mapping output should be interpreted as ('OX', 4)instead of ('FY', 4).

Please review these discrepancies and adjust your settings accordingly. Let me know if you need further assistance.

Best regards, Jiacong

[asyms]

Closing due to inactivity