1242 Commits

Author	SHA1	Message	Date
Qianfeng Zhang	660604635d	Clarifying the using of CK_TILE_HOST and CK_TILE_HOST_DEVICE trying to save compiling time	2026-06-23 09:27:59 +00:00
Qianfeng Zhang	3565656005	Change in updating max_uih_seqlen in the example	2026-06-23 09:27:59 +00:00
Qianfeng Zhang	639332b5de	Use supplement_array_by_last_element(num_targets, ) in example	2026-06-23 09:27:59 +00:00
Qianfeng Zhang	19a887cffa	Use supplement_array_by_last_element() in example to simplify the codes	2026-06-23 09:27:59 +00:00
Qianfeng Zhang	9537b958ab	Update to README.md	2026-06-23 09:27:59 +00:00
Qianfeng Zhang	2ae759c5b3	Add scripts for testing the using of separate sequence lengths for k/v	2026-06-23 09:27:59 +00:00
Qianfeng Zhang	a926db031a	Support separate sequence lengths for q and kv	2026-06-23 09:27:59 +00:00
Qianfeng Zhang	bcefc197d8	Use separate pipelines for using or not-using softmax situations	2026-06-23 09:27:59 +00:00
Qianfeng Zhang	4b3901e989	Implementation of hstu attention pipeline using trload for v on mi350	2026-06-23 09:27:59 +00:00
Qianfeng Zhang	5db875df2c	Fix in the comments	2026-06-23 09:27:59 +00:00
Qianfeng Zhang	563229a9e4	Update to gemm_0's CBlockDistribution encoding so that it is compatible with gemm_1's ABlockDistribution encoding	2026-06-23 09:27:59 +00:00
Qianfeng Zhang	4e9d4f7487	Using separate tile settings for no-softmax and with-softmax hstu attention situations	2026-06-23 09:27:59 +00:00
Qianfeng Zhang	91afdfee40	Update to benchmark scripts to consider for using softmax	2026-06-23 09:27:59 +00:00
Qianfeng Zhang	be3b27edd3	Add support of softmax in hstu attention	2026-06-23 09:27:59 +00:00
Qianfeng Zhang	ec4f174ac4	Add template parameter to gemm_0 MakeCBlockTile() for the need of defining PcompBlockTileType	2026-06-23 09:27:59 +00:00
Qianfeng Zhang	da50eea674	Move scaling by attn_scale to inside the main-loop	2026-06-23 09:27:59 +00:00
Qianfeng Zhang	3639fb8e38	Let IsTokenPairInsideMask() return bool type	2026-06-23 09:27:59 +00:00
Qianfeng Zhang	a71c996049	Add instances to consider for adding softmax support	2026-06-23 09:27:59 +00:00
Qianfeng Zhang	0abb52004a	Remove K0 from tile setting since it is not used	2026-06-23 09:27:58 +00:00
Qianfeng Zhang	de198549ad	Change to pipeline so that it is easier to add support of using softmax	2026-06-23 09:27:58 +00:00
Qianfeng Zhang	7e79736df7	Remove using IGLP method for instruction scheduling for kUseLocal true path	2026-06-23 09:27:58 +00:00
Qianfeng Zhang	08f50c2c51	Fix in GetQKBlockGemm()	2026-06-23 09:27:58 +00:00
Qianfeng Zhang	5199dbd027	Simplify the warp_gemm definitions in GetQKBlockGemm and GetKVBlockGemm	2026-06-23 09:27:58 +00:00
Qianfeng Zhang	b0710e8871	Remove useless constant statement in the kernel	2026-06-23 09:27:58 +00:00
Qianfeng Zhang	2c861d541f	Remove un-necessary HSTU_CHECK() callings	2026-06-23 09:27:58 +00:00
Qianfeng Zhang	bc5616f1dc	Add HSTU_CHECK() and use it in example codes	2026-06-23 09:27:58 +00:00
Qianfeng Zhang	f7f90c539e	Smalle update in reference hstu attention	2026-06-23 09:27:58 +00:00
Qianfeng Zhang	9c4d76d96b	Detach HstuBlockMask from pipeline definition and construct the HstuBlockMask type in the kernel according to window_size	2026-06-23 09:27:58 +00:00
Qianfeng Zhang	8313b34543	Unify the license statements on all the source files	2026-06-23 09:27:58 +00:00
Qianfeng Zhang	2668bb3aee	Remove using MakeKargsImpl() to simplify the hstu kernel	2026-06-23 09:27:58 +00:00
Qianfeng Zhang	a8c62920bf	Clarify the using of kSubQKHeaddim and kQKHeaddim so that less regular hdim (eg. 96, 160) can be efficiently supported	2026-06-23 09:27:58 +00:00
Qianfeng Zhang	7a948eee1d	Use set_slice_tilie() to replace direct thread_buffer assignment	2026-06-23 09:27:58 +00:00
Qianfeng Zhang	260ef2fdf2	Use xor transform to implement Q/K Lds descriptor for kKpack == 8 cases	2026-06-23 09:27:58 +00:00
Qianfeng Zhang	56dab1298f	Remove selectable VLayout for simplifying the codes since hdim is always fatest dimension	2026-06-23 09:27:58 +00:00
Qianfeng Zhang	e1ebd780c1	Using separate settings for gfx942 and gfx950	2026-06-23 09:27:58 +00:00
Qianfeng Zhang	40091814c6	Tiny change in pipeline BlockGemm definition to adapt to the latest merging with develop branch	2026-06-23 09:27:58 +00:00
Qianfeng Zhang	ee06d0b4fc	Add norm_dist parameter for hstu example to select either normal or uniform distribution to initialize data	2026-06-23 09:27:58 +00:00
Qianfeng Zhang	da5db1773d	Tiny fix in HstuBlockMaskWithLocal::GetTileRangeAlongX()	2026-06-23 09:27:58 +00:00
Qianfeng Zhang	446b62ad82	Adjust the atol and rtol and fix the check_err() using in example_hstu_attention.cpp	2026-06-23 09:27:58 +00:00
Qianfeng Zhang	dde0729405	Add simple handling for max_atten_seqlen bigger than max_uih_len situations	2026-06-23 09:27:58 +00:00
Qianfeng Zhang	c649b9b049	Tiny fix and comments in HstuBlockMaskWithLocal::IsFullTimeInsideMask()	2026-06-23 09:27:58 +00:00
Qianfeng Zhang	de47bfe752	Update HstuBlockMaskWithLocal::GetTileRangeAlongX, add comments and test cases for causal == false	2026-06-23 09:27:58 +00:00
Qianfeng Zhang	4ad55eab4d	Update to support min_full_attn_seqlen be bigger than max_uih_len	2026-06-23 09:27:58 +00:00
Qianfeng Zhang	0b54f1f43d	Add attn_scale MakeKargs() parameter support and update in example, reference codes	2026-06-23 09:20:58 +00:00
Qianfeng Zhang	6364b641b8	Replace the integer max_seqlen by float scale_p as kernel/pipeline parameter	2026-06-23 09:20:58 +00:00
Qianfeng Zhang	6784c0be5a	Use __builtin_amdgcn_sched_barrier(0x1) to prevent the compiler from unexpected codes arrangement	2026-06-23 09:20:58 +00:00
Qianfeng Zhang	492c724b14	Fix added case in test_hstu_attention.sh	2026-06-23 09:20:58 +00:00
Qianfeng Zhang	d3ed6ac473	Update in GetTileRangeAlongX to consider for non-causal+local_size>0 situation and add test case to test_hstu_attention.sh	2026-06-23 09:20:58 +00:00
Qianfeng Zhang	7d698c2b78	Add three scripts for verification of jagged causal cases	2026-06-23 09:20:58 +00:00
Qianfeng Zhang	14c955aade	Fix in GetTileRangeAlongX() and IsFullTileInsideMask() of HstuBlockMaskWithLocal	2026-06-23 09:20:58 +00:00