init src 0.9.2

This commit is contained in:
2026-01-09 15:09:53 +08:00
parent 0eb2c0a4b3
commit 41d98d4359
1438 changed files with 417605 additions and 683 deletions

View File

@@ -0,0 +1,244 @@
{
"1536_7168": {
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"14": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"64": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 8,
"num_ldmatrixes": 1
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
},
"2048": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
}
}
}

View File

@@ -0,0 +1,244 @@
{
"1536_7168": {
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"14": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 0
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 0
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 8,
"num_ldmatrixes": 1
},
"256": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 8,
"num_ldmatrixes": 1
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
}
}
}

View File

@@ -0,0 +1,244 @@
{
"3072_1536": {
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"14": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"64": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 8,
"num_ldmatrixes": 1
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 8,
"num_ldmatrixes": 1
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8,
"num_ldmatrixes": 1
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"2048": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
}
}
}

View File

@@ -0,0 +1,244 @@
{
"3072_1536": {
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"14": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 0
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"256": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8,
"num_ldmatrixes": 1
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
}
}
}

View File

@@ -0,0 +1,244 @@
{
"4096_512": {
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"14": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"64": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8,
"num_ldmatrixes": 1
},
"128": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8,
"num_ldmatrixes": 1
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8,
"num_ldmatrixes": 1
},
"512": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
}
}
}

View File

@@ -0,0 +1,244 @@
{
"4096_512": {
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"14": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 0
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 0
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"256": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8,
"num_ldmatrixes": 1
},
"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": 8,
"num_ldmatrixes": 1
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
}
}
}

View File

@@ -0,0 +1,244 @@
{
"4608_7168": {
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 2,
"num_ldmatrixes": 1
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 2,
"num_ldmatrixes": 1
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 2,
"num_ldmatrixes": 1
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 2,
"num_ldmatrixes": 1
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 2,
"num_ldmatrixes": 1
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 2,
"num_ldmatrixes": 1
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 2,
"num_ldmatrixes": 1
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 2,
"num_ldmatrixes": 1
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 2,
"num_ldmatrixes": 1
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 2,
"num_ldmatrixes": 1
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 2,
"num_ldmatrixes": 1
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 2,
"num_ldmatrixes": 1
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 2,
"num_ldmatrixes": 1
},
"14": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 2,
"num_ldmatrixes": 1
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 2,
"num_ldmatrixes": 1
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 2,
"num_ldmatrixes": 1
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"64": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
},
"128": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
},
"256": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
},
"512": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
},
"1024": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
},
"2048": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
}
}
}

View File

@@ -0,0 +1,244 @@
{
"4608_7168": {
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"14": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"64": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 8,
"num_ldmatrixes": 1
},
"128": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
}
}
}

View File

@@ -0,0 +1,244 @@
{
"512_7168": {
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"14": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 8,
"num_ldmatrixes": 1
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 8,
"num_ldmatrixes": 1
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 8,
"num_ldmatrixes": 1
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 8,
"num_ldmatrixes": 1
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
},
"2048": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
}
}
}

View File

@@ -0,0 +1,244 @@
{
"512_7168": {
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"14": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"128": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 0
},
"256": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 8,
"num_ldmatrixes": 1
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 8,
"num_ldmatrixes": 1
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 8,
"num_ldmatrixes": 1
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 8,
"num_ldmatrixes": 1
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
}
}
}

View File

@@ -0,0 +1,244 @@
{
"576_7168": {
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"14": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 8,
"num_ldmatrixes": 1
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 8,
"num_ldmatrixes": 1
},
"256": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 8,
"num_ldmatrixes": 1
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 8,
"num_ldmatrixes": 1
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
},
"2048": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
}
}
}

View File

@@ -0,0 +1,244 @@
{
"576_7168": {
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"14": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"64": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"128": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"256": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 0,
"num_warps": 8,
"num_ldmatrixes": 1
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 0,
"num_warps": 8,
"num_ldmatrixes": 1
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
}
}
}

View File

@@ -0,0 +1,244 @@
{
"7168_2048": {
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"14": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"64": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 8,
"num_ldmatrixes": 1
},
"128": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8,
"num_ldmatrixes": 1
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
},
"512": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"1024": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
},
"2048": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
}
}
}

View File

@@ -0,0 +1,244 @@
{
"7168_2048": {
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"14": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"64": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"128": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
}
}
}

View File

@@ -0,0 +1,244 @@
{
"7168_2304": {
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"14": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"64": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 8,
"num_ldmatrixes": 1
},
"128": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8,
"num_ldmatrixes": 1
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
},
"512": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"1024": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
},
"2048": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
}
}
}

View File

@@ -0,0 +1,244 @@
{
"7168_2304": {
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"14": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 8,
"SPLIT_K": 8,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 0
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 4,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 0
},
"64": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 2,
"num_stages": 0,
"num_warps": 8,
"num_ldmatrixes": 1
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8,
"num_ldmatrixes": 1
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 8,
"num_ldmatrixes": 1
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 8,
"num_ldmatrixes": 1
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
}
}
}

View File

@@ -0,0 +1,244 @@
{
"7168_256": {
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"13": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"14": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"64": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"128": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"512": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"1024": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"2048": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"4096": {
"BLOCK_SIZE_M": 256,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
}
}
}

View File

@@ -0,0 +1,244 @@
{
"7168_256": {
"1": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 0
},
"2": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 0
},
"3": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"4": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 0
},
"5": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"6": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"7": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 8,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"9": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"10": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"11": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"12": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"13": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 0
},
"14": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 0
},
"15": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"16": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 0
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 0
},
"64": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"128": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 0,
"num_warps": 4,
"num_ldmatrixes": 1
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 4,
"SPLIT_K": 1,
"num_stages": 1,
"num_warps": 4,
"num_ldmatrixes": 1
}
}
}