ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-06-30 03:37:38 +00:00
Code Issues Packages Projects Releases Wiki Activity
2,187 Commits 776 Branches 40 Tags
4a264eb9ed95dc4d0191fa5c1efc5fb5fc20645d
Commit Graph

680 Commits

Author SHA1 Message Date
MHYang
4a264eb9ed Fix register spilling and K0 tile size issues 2025-07-28 14:54:51 -04:00
YC Lin
eb737b8f82 [GEMM] Fix num_loop issues 2025-07-28 14:54:51 -04:00
Clement Lin
f71b2c7e55 Add generate.py for codegen 2025-07-28 14:54:51 -04:00
YC Lin
9e020272c4 [GEMM] Remove redundant GetBlockGemm 2025-07-28 14:54:51 -04:00
YC Lin
4e6a792b82 [GEMM] Implement local prefetch and refactor block gemm pipeline 2025-07-28 14:54:51 -04:00
Clement Lin
cbc660acc7 Refactor flash_attention_fwd_traits_ for codegen 2025-07-28 14:54:51 -04:00
YC Lin
f30071289c [GEMM] Merge universal_block_gemm into block_gemm 2025-07-28 14:54:51 -04:00
mhYang
26b73c0ed1 Fix flash attention 1 tile case 2025-07-28 14:54:51 -04:00
Clement Lin
e614acfdd8 Refactor FlashAttnArgs usage for codegen 2025-07-28 14:54:51 -04:00
Clement Lin
0cc5130818 Add codegen test example 2025-07-28 14:54:51 -04:00
YC Lin
5b1c397806 [GEMM] Refactor GetStaticLdsSize and remove GetSmemSize 2025-07-28 14:54:51 -04:00
Clement Lin
2499b8d401 Fix indentation 2025-07-28 14:54:51 -04:00
Clement Lin
d98fb3e0b5 Remove unused code 2025-07-28 14:54:51 -04:00
YC Lin
ae275aa105 [GEMM] Refactor block gemm, pipeline, and policy of instruction schedule opt 2025-07-28 14:54:51 -04:00
YC Lin
6113ca8062 [Add] Add build option for generating assembly 2025-07-28 14:54:51 -04:00
YC Lin
97a960042b [GEMM] Refactor block gemm and pipeline policy of instruction schedule 2025-07-28 14:54:51 -04:00
Clement Lin
8785e6599e Add flash_attention_fwd toy_example 2025-07-28 14:54:51 -04:00
mhYang
a949b82c9f Update tile size and use slc 2025-07-28 14:54:51 -04:00
mhYang
9158612a9f Fix add flops calculation 2025-07-28 14:54:51 -04:00
ClementLinCF
88a4c7414f Create README.md 2025-07-28 14:54:51 -04:00
mhYang
ac972bfd11 Use mfma 16x16x32 2025-07-28 14:54:51 -04:00
mhYang
5326d403e4 Fix KERNEL_D config 2025-07-28 14:54:51 -04:00
YC Lin
fe319b97ae [GEMM] Add pragma message for different MFMA options 2025-07-28 14:54:51 -04:00
YC Lin
76751567b5 [GEMM] Fix print typos 2025-07-28 14:54:51 -04:00
Clement Lin
4c526ab140 Fix indentation typo 2025-07-28 14:54:51 -04:00
Clement Lin
5b10e9f3dd [GEMM] Fix MFMA condition checks 2025-07-28 14:54:51 -04:00
Clement Lin
a95665a6af [GEMM] Add new macor options check 2025-07-28 14:54:51 -04:00
Clement Lin
1099762267 [GEMM] Add macros for multiple optimization options 2025-07-28 14:54:51 -04:00
YC Lin
890a159877 [GEMM] default MFMA config 2025-07-28 14:54:51 -04:00
YC Lin
8d75ae7c96 git push test 2025-07-28 14:54:51 -04:00
root
a36d246cc0 [GEMM] fix MFMA configurations 2025-07-28 14:54:51 -04:00
mhYang
15e6f36f66 Adjust mfma schedule order 2025-07-28 14:54:51 -04:00
Clement Lin
e9f7c9bf42 [GEMM] Replace const auto with constexpr index_t 2025-07-28 14:54:51 -04:00
Clement Lin
cef77c1dcb [GEMM] Update cache-aware wg schedule 2025-07-28 14:54:51 -04:00
bobofang
127e742e96 Add MFMA M16N16K16 and M16N16K32 methods 
these two methods are default off
 2025-07-28 14:54:51 -04:00
YC Lin
e866f814f9 [GEMM] remove a_col_major/b_row_majro case 2025-07-28 14:54:51 -04:00
root
bf69235cfb [GEMM] modify if-else locations 2025-07-28 14:54:51 -04:00
mhYang
ba8b5112c4 Fix AccDataType and CDataType 
1. Fix AccDataType and CDataType
2. Remove indent
3. Align merge_transform for tutorial
 2025-07-28 14:54:51 -04:00
mhYang
d6fd468603 Fix build error 2025-07-28 14:54:51 -04:00
root
b3986c32a6 [GEMM] disable/enable instruction scheduling 2025-07-28 14:54:51 -04:00
mhYang
42f2e21865 Fix missing message 2025-07-28 14:54:51 -04:00
mhYang
38ce4dd8c3 Fix xor transform dim. 2025-07-28 14:54:51 -04:00
Clement Lin
b03668fe8a [GEMM] Add cache-aware WG schedule and adjust block tile 
113 -> 121.7 TFops
 2025-07-28 14:54:51 -04:00
mhYang
39ca852330 Add LDS bank conlict solutions 2025-07-28 14:54:51 -04:00
bobofang
22147ace51 Fix add accuracy issue 
2673 GB/s -> 3271 GB/s
Perf: 0.0512898 ms, 3271.06 GB/s
 2025-07-28 14:54:51 -04:00
root
d7d9fdaf1b [GEMM] use mfma k8 warp gemm 2025-07-28 14:54:51 -04:00
root
1b8d7cd1b9 [GEMM] disable/enable prefetch 2025-07-28 14:54:50 -04:00
Clement Lin
6a2036015e [CK TILE] Toy example - basic gemm 2025-07-28 14:54:50 -04:00
Clement Lin
077056b32d Adjust block shape 
2673 GB/s -> 3647 GB/s
 2025-07-28 14:54:50 -04:00
Clement Lin
2ff691f3f2 Utilize vectorized memory access 
1998.24 GB/s -> 2673 GB/s
 2025-07-28 14:54:50 -04:00