Commit Graph

5 Commits

Author SHA1 Message Date
Jeff Huang
518d24e662 Add sequence padding and variable length support in fmha (#2932)
* * [CK_TILE] Add sequence padding and variable length support in fmha (and v3)

 - Group Mode Padding: Introduces the `-s_qpad` argument to support
   physically padded layouts. Kernels now use padded start pointers
   (`seqstart_padded_*_ptr`) for memory addressing.

 - Batch Mode Variable Length: Adds `-q_eff_lens` and `-kv_eff_lens`
   arguments for efficient processing of variable-length sequences by
   passing cumulative effective lengths (`cu_seqlen_*_ptr`) to the kernel.

 - FMHA examples: Support padding and variable length both in
   group and batch mode. Dispatcher is updated as well (dispatch to
   kPadSeqLenK enabled pipeline).

 - New padding test cases: Add padding test cases to `smoke_test_fwd.sh` and
   `test_fmha_fwd.inc`, and add benchmarks to `benchmark_fwd.sh` and
   `benchmark_fwd_v3.sh` as well. These test cases and benchmarks that
   specifically validate/benchmark the new padding and variable-length
   functionalities in both group and batch modes.

* [CK_TILE] Fix build error in fmha unit tests

* [CK_TILE] add mqa, gqa to sequence padding unit tests

* [CI_TILE] Reduce the number of padding seqlen unit tests in FMHA to avoid timeouts in CI

* [CK_TILE] remove unnecessary MageKArgs overload in FmhaFwdV3Kernel and FmhaFwdKernel
2025-09-26 12:36:27 +08:00
Illia Silin
b765fe78f3 Revert "[CK_TILE] Add sequence padding and variable length support in fmha (a…" (#2883)
This reverts commit 86dd59cd01.
2025-09-19 08:15:02 -07:00
Jeff Huang
86dd59cd01 [CK_TILE] Add sequence padding and variable length support in fmha (a… (#2851)
* [CK_TILE] Add sequence padding and variable length support in fmha (and v3)

 - Group Mode Padding: Introduces the `-s_qpad` argument to support
   physically padded layouts. Kernels now use padded start pointers
   (`seqstart_padded_*_ptr`) for memory addressing.

 - Batch Mode Variable Length: Adds `-q_eff_lens` and `-kv_eff_lens`
   arguments for efficient processing of variable-length sequences by
   passing cumulative effective lengths (`cu_seqlen_*_ptr`) to the kernel.

 - FMHA examples: Support padding and variable length both in
   group and batch mode. Dispatcher is updated as well (dispatch to
   kPadSeqLenK enabled pipeline).

 - New padding test cases: Add padding test cases to `smoke_test_fwd.sh`,
   and add benchmarks to `benchmark_fwd.sh` and `benchmark_fwd_v3.sh` as well.
   These test cases and benchmarks that specifically validate/benchmark the
   new padding and variable-length functionalities in both group and batch modes.

* [CK_TILE] Fix build error in fmha unit tests

---------

Co-authored-by: Po Yen Chen <PoYen.Chen@amd.com>
Co-authored-by: Yi DING <yi.ding@amd.com>
2025-09-19 17:36:49 +08:00
Po Yen Chen
7fbc9d6c97 [CK_TILE] FMHA FAv3 scheduling fine-tuning for performance (#2833)
* Re-mapping thread block indices for causal=True kernels

* Use more intuitive remap_opt value

* Fallback to origin remapping if seqlen_q >= 64K

* Use GenericAttentionMask to reduce mask computation

* Avoid unnecessary boundary check for IsMasking=false case

* Fix wrong kernel entry specifier

* Add s_nop to prevent delay wave0-3

* Refine scheduling

* Remove unnecessary sched_group_barrier()

* Move sched_group_barrier() call to scheduler

* Replace inline asm s_setprio with intrinsics

* Rephrase comments

* Expend some o_acc rescaling insts to avoid SIMD idle

* Fix block idx special mapping logic

* Tune block index mapping for causal=False cases

* Tune block index mapping for causal=True cases

* Fix wrong vmcnt()

* Remove parameter name

* Use boolean option for turn on/off causal mask

* Update benchmark_fwd_v3.sh option usages

* Add option if compiler support it
2025-09-16 11:32:38 +08:00
Po Yen Chen
d876e87fe4 [CK_TILE] Add FAv3 fwd pipeline (#2731)
* Add FAv3 fwd pipeline

* Unpack v_pk_mul to hide v_mov

* Avoid compiler moving l compute across phase

* Sync sched_group_barrier() setting for masking cases
2025-09-01 09:16:45 +08:00