Category: Reverse
Solved: 2/490
The function f performs matrix multiplication in $M_{256}(\mathbb{F}{7^3})$. With given $A, B \in M{256}(\mathbb{F}_{7^3})$ and power $k\in\mathbb{N}$, f calculates $A\cdot B^k$.
The buffer_dict is used to handle the different compute capability of Nvidia GPUs. On google colab, user might get Nvidia K80s, T4s, P4s or P100s [1]. We cannot select the type of GPUs, so we need to include versions of cubin.
To findout the content of the cubin, nvdisasm provided by Nvidia is recommanded to get the assembly code of the cubin. You could generate control flow graph with nvdisasm -bbcfg a.cubin | dot -obbcfg.png -Tpng. Also, you can find the instrcution tables of Nvidia architectures here [2]. These tools could help to recover the semantics of the program.
It is a parallel matrix multiplication algorithm on GPU (.L7). The dependent blocks was loaded into shared memory first with LDS instrcution. Then, finite field arithmetic was parallel calculated.
To obtain the original $A$(flag), you need to calculate $B^{-1}$. You could get the inverse matrix $B^{-1}$ in sage. I tried to calculate $B^{-k}$ with sage but it seems too slow with CPU version, you could speedup with exponentiating by squaring or use GPU version on colab.