:j6dddddddddZd Zd Zy ) c d|dz z S)Ng2Cwidthkwargss j/home/mcse/projects/srt_converter/srt-converter-venv/lib/python3.12/site-packages/triton/profiler/specs.pyr s%!) 4c d|dz z S)NgbBrrrs rr r s%!) <r c 0|dz |zdz dzdz|dz z S)Nrgx:AigmBrrrnum_sms clock_raters rr r s-w}z7QU_7`cg7gko6o 6r c *|dz|dz zdz|dz z S)Ni@@@g.Arrrs rr r s'goVYIY6Z]`6`ejmnen5or )808990100c d|dz z S)Ng//bBrrrs rr r s&EAI*>r c d|dz z S)Ng?y"Crrrs rr r s)uqy*Ar )gfx90agfx942)CUDAHIPc|tvrtd||t|vrtd|t||}||||S)a Calculate the maximum FLOPS for a given device type and width. Args: device_type (str): The type of device (e.g., "CUDA", "HIP"). arch (str): The architecture of the device (e.g., "80", "90"). width (int): The width in bits. num_sms (int): The number of streaming multiprocessors. clock_rate (float): The clock rate in GHz. Returns: float: The maximum FLOPS for the given device type and width. zUnsupported device type: zUnsupported architecture: )rr)flops_by_device ValueError) device_typearchrrr flops_funcs r max_flopsr#s]/)4[MBCC ?;//5dV<== -d3J eW DDr cd|z|zdzdz S)a Calculate the maximum bytes per second for a given bus width and memory clock rate. Args: bus_width (int): The bus width in bits. memory_clock_rate (float): The memory clock rate in GHz. Returns: float: The maximum bytes per second. rrr) bus_widthmemory_clock_rates rmax_bpsr(-s y=, ,s 2Q 66r N)rr#r(rr rr)s< 5<  p ?A &E2 7r