"lib/vscode:/vscode.git/clone" did not exist on "923e1ee729605a5c3abf2e43b535e9b2da90e771"
Commit 083b80ea authored by zhuwenwen's avatar zhuwenwen
Browse files

增加w8a8相关修改

parent 09428eec
{
"4096_4096": {
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"14": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"17": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"18": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 8
},
"19": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 8
},
"20": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 8
},
"21": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"22": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"23": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"25": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"26": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"27": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"28": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"29": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"30": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"31": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"256": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"2048": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"8192": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
}
}
{
"4096_4096": {
"20": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 8
},
"24": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 8
},
"28": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 8
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 8
},
"36": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"40": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"44": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"52": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"56": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"60": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"72": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"80": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"88": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"96": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"104": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"112": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"120": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"136": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"144": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"152": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"160": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"14": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"256": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"2048": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"8192": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
}
}
}
\ No newline at end of file
{
"4608_3584": {
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"14": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"20": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"24": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"28": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"36": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"40": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"44": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"52": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"56": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"60": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"72": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"80": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"88": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"96": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"104": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"112": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"120": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"136": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"144": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"152": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"160": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"256": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"8192": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
}
}
}
\ No newline at end of file
{
"5120_13824": {
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"14": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"20": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"24": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"28": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"36": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"40": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"44": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"48": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"52": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"56": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"60": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"64": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"72": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 2,
"num_warps": 8
},
"80": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"88": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"96": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"104": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"112": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"120": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"136": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"144": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"152": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"160": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"2048": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"8192": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
}
}
}
\ No newline at end of file
{
"5120_2560": {
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 2
},
"14": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 2
},
"20": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"24": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"28": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"36": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"40": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"44": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"52": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"56": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"60": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"72": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"80": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"88": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"96": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"104": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"112": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"120": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 16
},
"136": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"144": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"152": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"160": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"256": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"8192": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
}
}
}
\ No newline at end of file
{
"5120_5120": {
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"14": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"20": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"24": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"28": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"36": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"40": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"44": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"52": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"56": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"60": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"72": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"80": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"88": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"96": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"104": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"112": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"120": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"136": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"144": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"152": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"160": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"8192": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
}
}
}
\ No newline at end of file
{
"5120_6912": {
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 2
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 2
},
"14": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 2
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"20": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"24": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"28": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"36": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"40": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"44": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"48": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"52": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"56": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"60": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"72": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"80": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 16
},
"88": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"96": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 16
},
"104": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"112": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"120": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 16
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 16
},
"136": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"144": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 16
},
"152": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 16
},
"160": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"8192": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
}
}
}
\ No newline at end of file
{
"5120_8192": {
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"9": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"10": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"11": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"12": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"13": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"14": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"15": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"16": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 2
},
"20": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"24": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"28": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"36": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"40": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"44": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 1,
"num_warps": 16
},
"48": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 16
},
"52": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 16
},
"56": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 16
},
"60": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 16
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 16
},
"72": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 16
},
"80": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 16
},
"88": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 16
},
"96": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 16
},
"104": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 16
},
"112": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 16
},
"120": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 16
},
"128": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 16
},
"136": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 16
},
"144": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 16
},
"152": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 16
},
"160": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 16
},
"256": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"8192": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
}
}
}
\ No newline at end of file
{
"6144_4096": {
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"5": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"6": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"7": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"8": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"9": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"10": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"11": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"12": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"13": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"14": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"15": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"16": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"20": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"24": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"28": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"36": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"40": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"44": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"48": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"52": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"56": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"60": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"72": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"80": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"88": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"96": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"104": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"112": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"120": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"136": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"144": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"152": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"160": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"256": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"8192": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
}
}
}
\ No newline at end of file
{
"7168_8192": {
"1": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"2": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"3": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"4": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"6": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"14": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"15": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"16": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"20": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"24": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"28": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"32": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"36": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"40": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"44": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"48": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"52": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"56": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"60": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"64": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"72": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"80": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"88": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"96": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"104": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"112": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"120": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 2,
"num_warps": 8
},
"136": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 8
},
"144": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 2,
"num_warps": 8
},
"152": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 2,
"num_warps": 8
},
"160": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 2,
"num_warps": 8
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 8
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 2,
"num_warps": 8
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 1,
"num_warps": 8
},
"2048": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"8192": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
}
}
}
\ No newline at end of file
{
"7680_5120": {
"1": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"5": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"6": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"7": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"8": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"9": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"10": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"13": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"14": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"15": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"16": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"20": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"24": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"28": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"36": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"40": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"44": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"48": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"52": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"56": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"60": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"64": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"72": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"80": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"88": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"96": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"104": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"112": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"120": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"136": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"144": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"152": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"160": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"1024": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"2048": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"8192": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
}
}
}
\ No newline at end of file
{
"8192_1024": {
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"14": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"20": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"24": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"28": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"36": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"40": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"44": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"48": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"52": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"56": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"60": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"72": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"80": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"88": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"96": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"104": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"112": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"120": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"136": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"144": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"152": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"160": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"256": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"8192": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
}
}
}
\ No newline at end of file
{
"8192_14336": {
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"14": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"20": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"28": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"36": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"40": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"44": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"48": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"52": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"56": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"60": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"72": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"80": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"88": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"96": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"104": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"112": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"120": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"136": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"144": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"152": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"160": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"512": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"1024": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"2048": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"8192": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
}
}
}
\ No newline at end of file
{
"8192_2048": {
"1": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"3": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"14": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 16,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"17": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"20": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"24": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"28": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"36": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"40": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"44": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"48": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"52": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"56": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"60": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"72": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"80": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"88": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"96": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"104": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"112": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"120": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 16
},
"136": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"144": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"152": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"160": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"1024": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"2048": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"8192": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"16384": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
}
}
}
\ No newline at end of file
{
"8192_3584": {
"1": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 2
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"4": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 2
},
"5": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 2
},
"6": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 2
},
"7": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 2
},
"8": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 2
},
"9": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 2
},
"10": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 2
},
"11": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 2
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"14": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"15": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"20": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"24": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"28": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"36": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"40": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"44": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"48": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"52": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"56": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"60": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"72": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"80": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"88": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"96": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"104": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"112": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"120": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"136": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"144": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"152": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"160": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"256": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"8192": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
}
}
}
\ No newline at end of file
{
"8192_4096": {
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"14": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"20": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"28": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"36": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"40": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"44": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 1,
"num_warps": 4
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 1,
"num_warps": 4
},
"52": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"56": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"60": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"72": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"80": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"88": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"96": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"104": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"112": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"120": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"136": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"144": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"152": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"160": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 512,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"512": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"1024": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"2048": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"8192": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
}
}
}
\ No newline at end of file
{
"8192_7168": {
"1": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 2
},
"2": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"3": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8
},
"4": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 8
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"14": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 2,
"num_warps": 4
},
"17": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"20": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"28": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"36": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"40": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"44": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"48": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"52": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"56": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"60": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"64": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4
},
"72": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"80": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"88": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"96": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"104": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"112": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"120": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"136": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"144": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"152": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"160": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"512": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"1024": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 2,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"2048": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
},
"8192": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8
}
}
}
\ No newline at end of file
......@@ -238,27 +238,40 @@ def apply_int8_linear(
m=x_q.shape[0]
k=x_q.shape[1]
n=weight.shape[1]
if f"{m}_{n}_{k}" in W8A8_TRITONJSON.triton_json_dict[0]:
best_config=W8A8_TRITONJSON.triton_json_dict[0][f"{m}_{n}_{k}"]
#print("json files:",best_config)
elif f"1_{n}_{k}" in W8A8_TRITONJSON.triton_json_dict[0]:
if m<64:
m_= 32
elif m<128:
m_=64
elif m<256:
m_=128
elif m<512:
#print("m:{},n:{},k:{}".format(m,n,k))
if f"1_{n}_{k}" in W8A8_TRITONJSON.triton_json_dict[0]:
if m<=16:
m_=m
#best_config=W8A8_TRITONJSON.triton_json_dict[0][f"{m}_{n}_{k}"]
elif m<=64:
m_= (m + 3) & -4 #取值到最近的4的倍数
elif m<=160:
m_=(m + 7) & -8
elif m<200: #256
m_=160
elif m<480: #512
m_=256
elif m<1024:
elif m<960: #1024
m_=512
else:
elif m<2048:
m_=1024
elif m<4096:
m_=2048
elif m<6000:
m_=4096
else:
m_=8192
best_config=W8A8_TRITONJSON.triton_json_dict[0][f"{m_}_{n}_{k}"]
else:
best_config=None
print("config not found!")
# if best_config==None:
# print("m:{},n:{},k:{}".format(m,n,k))
# print("config not found!")
return ops.triton_scaled_mm(x_q,
weight,
......
......@@ -462,7 +462,7 @@ class FalconForCausalLM(nn.Module, SupportsPP):
self.use_gemm_pad = os.environ.get('GEMM_PAD') == '1'
self.use_fa_pad = os.environ.get('FA_PAD') == '1'
self.use_awq_pad = os.environ.get('AWQ_PAD') == '1'
self.w8a8_strategy=int(os.getenv('W8A8_SUPPORT_METHODS', '0'))
self.w8a8_strategy=int(os.getenv('W8A8_SUPPORT_METHODS', '1'))
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
return self.transformer.get_input_embeddings(input_ids)
......
......@@ -355,7 +355,7 @@ class LlamaModel(nn.Module):
self.use_gemm_pad = os.environ.get('GEMM_PAD') == '1'
self.use_fa_pad = os.environ.get('FA_PAD') == '1'
self.use_awq_pad = os.environ.get('AWQ_PAD') == '1'
self.w8a8_strategy=int(os.getenv('W8A8_SUPPORT_METHODS', '0'))
self.w8a8_strategy=int(os.getenv('W8A8_SUPPORT_METHODS', '1'))
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
return self.embed_tokens(input_ids)
......@@ -553,6 +553,10 @@ class LlamaModel(nn.Module):
if matches and "scale" not in layername:
weight_data =params_dict[layername]
n=weight_data.shape[0]
# k=weight_data.shape[1]
# #判断当前size是否在优化的范围内,假如存在则走triton,假如不存在则走rocblas
# json_file=self.tritonsingleton.get_w8a8json_name(n,k)
#rocblas和cutlass目前都需要weight做处理,但是triton不用
if self.w8a8_strategy!=1:
......
......@@ -913,12 +913,13 @@ class QWenBaseModel(nn.Module, SupportsPP, SupportsLoRA):
if quant_config is not None:
self.quant_method=quant_config.get_name()
self.quant_config=quant_config
self.tritonsingleton= W8a8GetCacheJSON()
self.use_llama_nn = os.environ.get('LLAMA_NN') == '1'
self.use_gemm_pad = os.environ.get('GEMM_PAD') == '1'
self.use_fa_pad = os.environ.get('FA_PAD') == '1'
self.use_awq_pad = os.environ.get('AWQ_PAD') == '1'
self.w8a8_strategy=int(os.getenv('W8A8_SUPPORT_METHODS', '0'))
self.w8a8_strategy=int(os.getenv('W8A8_SUPPORT_METHODS', '1'))
def _get_image_input_type(
self,
......@@ -1147,7 +1148,7 @@ class QWenBaseModel(nn.Module, SupportsPP, SupportsLoRA):
m=int(key.split('_')[0])
n=int(key.split('_')[1])
k=int(key.split('_')[2])
ops._int8_gemm_helper(m=m,n=n,k=k,per_token_act_quant=True,per_out_channel_weight_quant=True,use_bias=False,best_config=value)
ops.triton_int8_gemm_helper(m=m,n=n,k=k,per_token_act_quant=True,per_out_channel_weight_quant=True,use_bias=False,best_config=value)
return loaded_params
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment