[TASK] Add A New Quant Ball for FP32-MXFP Conversion

[shirohasuki]

Deliverables

• Add an MXFP ball RTL implementation in the prototype lib (under the arch path).
• A Pull Request (PR) containing a test written in C for this operation and a README to introduce your design.
• Report the performance results in this issue.

Task Description

• MXFP is a lower-precision floating-point representation designed to reduce data size and simplify computations in the following process. Using MXFP can improve throughput and hardware efficiency in bandwidth-sensitive workloads, while still maintaining acceptable numerical quality for many ML scenarios.
• You can learn this format and its variants, starting from this paper, "With Shared Microexponents, A Little Shifting Goes a Long Way".
• As we envisage, an FP32 matrix will be loaded into the banks, and then a your customised MXFP instruction will read the data from one bank into the ball you are to implement, before outputting it to another bank.
• You can refer to the previous Pull Request (Completed the development of ReluBall and further improved the operation manual #6) for the detailed implementation.

[Aurore-pixed]

I'd like to take this task. Thanks

[shirohasuki]

Assignee: @zdy2916353903-crypto

[shirohasuki]

I'd like to take this task. Thanks

Thank you for your interest!

[Aurore-pixed]

Implemented a V1 MXFP ball prototype and opened a PR.

Current result:

• Verilator simulation passes
• software golden model matches hardware output
• full 128-bit packed output block verified in the current test

Current implementation assumptions:

• InputNum = 16
• inputWidth = 32
• bankWidth = 128
• 4 input reads per block
• 1 packed output write per block

[shirohasuki]

Implemented a V1 MXFP ball prototype and opened a PR.

Current result:

• Verilator simulation passes
• software golden model matches hardware output
• full 128-bit packed output block verified in the current test

Current implementation assumptions:

• InputNum = 16
• inputWidth = 32
• bankWidth = 128
• 4 input reads per block
• 1 packed output write per block

Nice Progress! Let me deliver a quick review.