flashattention fwd add (80, 96) instance (#3415)

* add hdim (96,96) instance

* change to (80,96)

* format py

* remove 96 in optdim

* when N=6 change to llvm_amdgcn_raw_buffer_load_i32x3
This commit is contained in:
ltqin
2025-12-18 01:16:11 +08:00
committed by GitHub
parent fe3d52d9b0
commit 92653168c2
6 changed files with 127 additions and 6 deletions

View File

@@ -40,7 +40,16 @@ DTYPE_BITS = {
"bf8": 8,
}
K0_MAX_SUBMAX_MAP = {32: 32, 48: 48, 64: 64, 96: 128, 128: 128, 192: 192, 256: 256}
K0_MAX_SUBMAX_MAP = {
32: 32,
48: 48,
64: 64,
80: 96,
96: 128,
128: 128,
192: 192,
256: 256,
}
FMHA_FWD_KERNEL_HEADER = """// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.\n
@@ -930,6 +939,7 @@ class KernelComponentFactoryGfx9(CompatibilityRuleFactoryGfx9):
( 64, 64) : [FmhaFwdTileSize( 16, 32, 64, 64, 32, 64, 1, 1, 1, 1, 1, 1, 16, 16, 32, 16, 16, 32, -1),
FmhaFwdTileSize( 32, 32, 64, 64, 32, 64, 1, 1, 1, 1, 1, 1, 32, 32, 16, 32, 32, 16, -1),
FmhaFwdTileSize(128, 64, 32, 64, 32, 64, 4, 1, 1, 4, 1, 1, 32, 32, 16, 32, 32, 16, -1)],
( 80, 96) : [FmhaFwdTileSize(128, 128, 16, 96, 32, 80, 4, 1, 1, 4, 1, 1, 32, 32, 16, 32, 32, 16, -1)],
( 96, 128) : [FmhaFwdTileSize(128, 128, 32, 128, 32, 96, 4, 1, 1, 4, 1, 1, 32, 32, 16, 32, 32, 16, -1)],
(128, 128) : [FmhaFwdTileSize( 16, 32, 64, 128, 32, 128, 1, 1, 1, 1, 1, 1, 16, 16, 32, 16, 16, 32, -1),
FmhaFwdTileSize( 32, 32, 128, 128, 32, 128, 1, 1, 1, 1, 1, 1, 32, 32, 16, 32, 32, 16, -1),