Commits · 55d1b19c5fe89bacdecf7fc0ae8b10c9db889099 · gaoqiong / composable_kernel_ROCM · GitLab

11 Feb, 2025 3 commits
- Move the location for store_tile() of first v_tile · 55d1b19c
  Qianfeng Zhang authored Feb 11, 2025
  
  55d1b19c
- Use __builtin_amdgcn_sched_barrier(0x7f) in the pipeline · f881fa70
  Qianfeng Zhang authored Feb 11, 2025
  
  f881fa70
- Adjust the using of kPadSeqLenQ and kPadSeqLenK in the kernel · 660261a2
  Qianfeng Zhang authored Feb 11, 2025
  
  660261a2
10 Feb, 2025 2 commits
- Prefetch first v_tile at earlier time for both kPreloadNextWholeIterationK true/false paths · aa304c67
  Qianfeng Zhang authored Feb 10, 2025
  
  aa304c67
- Tune the location of a scheduler barrier code line · e472af36
  Qianfeng Zhang authored Feb 10, 2025
  
  e472af36
09 Feb, 2025 3 commits
- Use NumPrefetchV to separate from NumVLdsBuffers · d362410d
  Qianfeng Zhang authored Feb 09, 2025
  
  d362410d
- Adjust the pipeline codes · 2e612c02
  Qianfeng Zhang authored Feb 09, 2025
  
  2e612c02
- Tune K Lds and V Lds reuse for kPreloadWholeNextIterationK == false · a72e100e
  Qianfeng Zhang authored Feb 09, 2025
  
  a72e100e
08 Feb, 2025 2 commits
- Merge branch 'ck_tile/improve_async_pipeline' of... · 53766479
  Qianfeng Zhang authored Feb 08, 2025
```
Merge branch 'ck_tile/improve_async_pipeline' of https://github.com/ROCm/composable_kernel into ck_tile/improve_async_pipeline
```
  53766479
- Tuning to K pre-loading · a151fd6d
  Qianfeng Zhang authored Feb 08, 2025
  
  a151fd6d
06 Feb, 2025 1 commit
- Unify the alignment to be 8 for Q/K/V Lds decriptors · 02b6c6c2
  Qianfeng Zhang authored Feb 06, 2025
  
  02b6c6c2
04 Feb, 2025 4 commits
- Add comments · fb0f56b3
  Qianfeng Zhang authored Feb 04, 2025
  
  fb0f56b3
- Re-arrange the codes before the main-loop · d3b01d2b
  Qianfeng Zhang authored Feb 04, 2025
  
  d3b01d2b
- Tune the lines of codes to make them more tidy · d55852bc
  Qianfeng Zhang authored Feb 04, 2025
  
  d55852bc
- Move the code line of clear_tile(s_acc) · ee324424
  Qianfeng Zhang authored Feb 04, 2025
  
  ee324424
03 Feb, 2025 5 commits
- Remove un-used get_slice_tile() call · 1a6a3715
  Qianfeng Zhang authored Feb 03, 2025
  
  1a6a3715
- Remove using KLoadOnce in qx_ks_vs_custom_policy · c5206c90
  Qianfeng Zhang authored Feb 03, 2025
  
  c5206c90
- Define statically indexed array v_lds_windows[] to reduce using of get_slice_tile() · 87b206fb
  Qianfeng Zhang authored Feb 03, 2025
  
  87b206fb
- Move the definition of v_tiles out from the loop · cde3b677
  Qianfeng Zhang authored Feb 03, 2025
  
  cde3b677
- Define statically indexed array k_lds_windows[] to reduce the using of get_slice_tile() · 08598523
  Qianfeng Zhang authored Feb 03, 2025
  
  08598523
02 Feb, 2025 3 commits
- Special treatment for hdim-96 to save vgprs in qr_ks_vs_async pipeline · 97efebdb
  Qianfeng Zhang authored Feb 02, 2025
  
  97efebdb
- Use QLoadOnce == false for qr_ks_vs_async pipeline · a94ac4bb
  Qianfeng Zhang authored Feb 02, 2025
  
  a94ac4bb
- Use array of tiles to represent Q in vgprs · 475c0d2c
  Qianfeng Zhang authored Feb 02, 2025
  
  475c0d2c
31 Jan, 2025 3 commits
- Let V reuse the LDS of K · 119dd2ac
  Qianfeng Zhang authored Jan 31, 2025
  
  119dd2ac
- Roll-back to load Q directly from global memory instead of using LDS as intermediary stop · 512eeecb
  Qianfeng Zhang authored Jan 31, 2025
  
  512eeecb
- Use MakeQDramTileDistribution for q_dram_window · 3e411ef0
  Qianfeng Zhang authored Jan 31, 2025
  
  3e411ef0
30 Jan, 2025 3 commits
- Move code interleaving · ceea2722
  Qianfeng Zhang authored Jan 30, 2025
  
  ceea2722
- Move clear_tile(s_acc) for better interleaving · 76871a6f
  Qianfeng Zhang authored Jan 30, 2025
  
  76871a6f
- Interleave code blocks for better performance · a979d030
  Qianfeng Zhang authored Jan 30, 2025
  
  a979d030
29 Jan, 2025 1 commit
- Switch to separate code blocks according to iteration index · b1da29ba
  Qianfeng Zhang authored Jan 27, 2025
  
  b1da29ba
26 Jan, 2025 6 commits
- Let BlockDropout reuse LDS with V · 90e99a95
  Qianfeng Zhang authored Jan 26, 2025
  
  90e99a95
- Move the codes for storing the first v_lds tile some later · 281110cf
  Qianfeng Zhang authored Jan 26, 2025
  
  281110cf
- Tune the prefetching of V in qr_ks_vs_async pipeline · 5f4bfa4a
  Qianfeng Zhang authored Jan 26, 2025
  
  5f4bfa4a
- Remove the commented lines in qx_ks_vs_custom_policy.hpp · 45398bf4
  Qianfeng Zhang authored Jan 26, 2025
  
  45398bf4
- Use k0_loops small tile load/store to replace the big tile load/store for K · 916daf59
  Qianfeng Zhang authored Jan 26, 2025
  
  916daf59
- Use un-rolled gemm for Gemm-0 · 4776c8c0
  Qianfeng Zhang authored Jan 26, 2025
  
  4776c8c0
24 Jan, 2025 4 commits
- Use LDS as intermediary stop when loading Q from global memory for qr_ks_vs_async pipeline · 00fe0752
  Qianfeng Zhang authored Jan 24, 2025
  
  00fe0752
- Rename MakeQDramTileDistribution to MakeQRegTileDistribution for QLoadOnce pipeline · 34157f26
  Qianfeng Zhang authored Jan 24, 2025
  
  34157f26
- Tiny adjustment in qr_ks_vs_async pipeline for better performance · 80c84d08
  Qianfeng Zhang authored Jan 24, 2025
  
  80c84d08
- Remove last block_sync_lds() in the loop · 59e3cb05
  Qianfeng Zhang authored Jan 24, 2025
  
  59e3cb05