a kh@sFddddddddddddddd Zd d Zd d ZdS)cKs d|dS)Ng2CwidthkwargsrrC/var/www/auris/lib/python3.9/site-packages/triton/profiler/specs.pyrcKs d|dS)NgbBrrrrrrrrcKs |d|ddd|dS)Nrgx:AigmBrrrnum_sms clock_raterrrrrscKs|d|dd|dS)Ni@@@g.Arrr rrrr r)80Z89Z90Z100cKs d|dS)Ng//bBrrrrrrrrcKs d|dS)Ng?y"Crrrrrrrr)Zgfx90aZgfx942)ZCUDAZHIPcCsJ|tvrtd||t|vr0td|t||}||||dS)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: )r r )flops_by_device ValueError)Z device_typearchrr r Z flops_funcrrr max_flopss   rcCsd||ddS)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. r rr)Z bus_widthZmemory_clock_raterrrmax_bps-s rN)rrrrrrrs