autotune

Function

Traverses the search space, tests different parameter combinations, and displays the running time of each combination and the optimal combination.

Function Prototype

def autotune(configs: List[Dict], warmup: int = 300, repeat: int = 1, device_ids = [0]):

Parameters

Parameter	Input/Output	Description
configs	Input	Search space definition. Data type: list[dict]. This parameter is required.
warmup	Input	Preheating time before performance collection. Longer preheating times typically lead to more stable operator performance. Unit: µs. This parameter is optional. The default value is 1000. The value is an integer ranging from 1 to 100000.
repeat	Input	Number of repeat times. The average running duration of multiple repeats is used as the operator duration. This parameter is optional. The default value is 1. The value is an integer ranging from 1 to 10000.
device_ids	Input	Device ID list. Currently, only the single-device mode is supported. If multiple device IDs are entered, only the first device ID takes effect. This parameter is optional. The default value is [0].

Return Value

None.

Example

@mskpp.autotune(configs=[
    {'L1TileShape': 'MatmulShape<64, 64, 64>', 'L0TileShape': 'MatmulShape<128, 256, 64>'},
    {'L1TileShape': 'MatmulShape<64, 64, 128>', 'L0TileShape': 'MatmulShape<128, 256, 64>'},
    {'L1TileShape': 'MatmulShape<64, 128, 128>', 'L0TileShape': 'MatmulShape<128, 256, 64>'},
    {'L1TileShape': 'MatmulShape<64, 128, 128>', 'L0TileShape': 'MatmulShape<64, 256, 64>'},
    {'L1TileShape': 'MatmulShape<128, 128, 128>', 'L0TileShape': 'MatmulShape<128, 256, 64>'},
], warmup=500, repeat=10, device_ids=[0])
def basic_matmul(problem_shape, a, layout_a, b, layout_b, c, layout_c):
    kernel = get_kernel()
    blockdim = 20
    return kernel[blockdim](problem_shape, a, layout_a, b, layout_b, c, layout_c)

Parent topic: API List