* Add separate mask checking for scope [aligned_physical_seqlen_k_start, physical_seqlen_k_end) in pagedkv pipeline
* i_nhead_ conversion type to prevent overflow
---------
Co-authored-by: ltqin <letaoqin@amd.com>
* add prefetching physical block id for pagedkv
* start add pagedkv prefill
* rename pipeline
* add kernel for pagedkv
* add an init version pagedkv prefill
* fix redefine issue
* add struct BlockFmhaFwdPagedKVPipelineProblem and fmha_fwd_pagedkv_args
* generate dispatch code
* add body generating code
* comipling pass
* remove dropout from pagedkv
* set lse to false in generating code
* start changing qr kernel to pagedkv
* init version of kernerl with pagedkv
* change names of file that are generated
* chang host validation for pagedkv prefill
* using iglp to change blockgemm
* add kernel files to op head file
* show parameters
* rewrite print parameter fun
* add fwd
* remove default parameter of GridSize
* format
* fix nhead issue and add seqlen_k_ptr to batch mode
* format code
* remove no-longer used code
* format
* fix some comments
---------
Co-authored-by: ltqin <letaoqin@amd.com>
Co-authored-by: Po Yen Chen <PoYen.Chen@amd.com>
* Do not use warpSize as compile time constant as it is removed
* Update tile_image_to_column_shape.hpp
update warpSize usage.
* clean-up all use of warpSize, make sure code builds
* fix
---------
Co-authored-by: Illia Silin <98187287+illsilin@users.noreply.github.com>
Co-authored-by: illsilin <Illia.Silin@amd.com>
Co-authored-by: Bartlomiej Kocot <barkocot@amd.com>
* Fix fmha fwd precision issue on MI3XX series
For fmha fwd fp16 cases, we found that using
impl::cast_tile_pk_fp16_fp32 for casting P would lead to precision
issues, since it uses __builtin_amdgcn_cvt_pkrtz, which is round to zero.
For examaple, fixing K,V to be all 1, and Q is random, which outputs are
expected to be all 1. But we found that it would have some incorrect
outputs 0.9995, which are smaller than the atol 0.001. (1 - 0.9995 =
0.0005 < 0.001) Thus, ck do not report this error.
* Add option to switch rtn/rtz for fmha fwd
* Write soft-sign in inline asm
* Change tile idx computation
* Add macro to turn off soft-sign asm opt
* Use simple for loop to avoid register spill
* Only do block id transform for masking cases
* Re-implement qr_ks_vs_async pipeline by using kLoadOnce
* Remove last block_sync_lds() in the loop
* Tiny adjustment in qr_ks_vs_async pipeline for better performance
* Rename MakeQDramTileDistribution to MakeQRegTileDistribution for QLoadOnce pipeline
* Use LDS as intermediary stop when loading Q from global memory for qr_ks_vs_async pipeline
* Use un-rolled gemm for Gemm-0
* Use k0_loops small tile load/store to replace the big tile load/store for K
* Remove the commented lines in qx_ks_vs_custom_policy.hpp
* Tune the prefetching of V in qr_ks_vs_async pipeline
* Move the codes for storing the first v_lds tile some later
* Let BlockDropout reuse LDS with V
* Switch to separate code blocks according to iteration index
* Interleave code blocks for better performance
* Move clear_tile(s_acc) for better interleaving
* Move code interleaving
* Use MakeQDramTileDistribution for q_dram_window
* Roll-back to load Q directly from global memory instead of using LDS as intermediary stop
* Let V reuse the LDS of K
* Use array of tiles to represent Q in vgprs
* Use QLoadOnce == false for qr_ks_vs_async pipeline
* Special treatment for hdim-96 to save vgprs in qr_ks_vs_async pipeline
* Define statically indexed array k_lds_windows[] to reduce the using of get_slice_tile()
* Move the definition of v_tiles out from the loop
* Define statically indexed array v_lds_windows[] to reduce using of get_slice_tile()
* Remove using KLoadOnce in qx_ks_vs_custom_policy
* Remove un-used get_slice_tile() call
* Move the code line of clear_tile(s_acc)
* Tune the lines of codes to make them more tidy
* Re-arrange the codes before the main-loop
* Add comments
* Unify the alignment to be 8 for Q/K/V Lds decriptors
* Tuning to K pre-loading
* Tune K Lds and V Lds reuse for kPreloadWholeNextIterationK == false
* Adjust the pipeline codes
* Use NumPrefetchV to separate from NumVLdsBuffers
* Tune the location of a scheduler barrier code line
* Prefetch first v_tile at earlier time for both kPreloadNextWholeIterationK true/false paths
* Adjust the using of kPadSeqLenQ and kPadSeqLenK in the kernel
* Use __builtin_amdgcn_sched_barrier(0x7f) in the pipeline
* Move the location for store_tile() of first v_tile
* Rename the qr_ks_vs_async pipeline to qr_ks_vs_whole_k_prefetch pipeline
* Re-add NumPrefetchK as template for BlockFmhaPipelineQXKSVSCustomPolicy<>
* Try to fix old bugs in qx_ks_vs_custom_policy
* Remove K_LDS_LOAD_USE_OFFSET_TRANSFORM code-path to make qr_ks_vs_async and qx_ks_vs_custom_policy simpler
* Fix in MakeKDramTileDistribution() in qx_ks_vs_custom_policy
* Update to LdsBufferSequence and introduce NumKVLdsBuffers for max(NumPrefetchK, NumPrefetchV)
* Tiny Fix (#1888)
* Ck tile/paged attention workaround (#1894)
* Correction in GetRangeAlongX()
* Work-around to solve the failures in test_paged_attention_ck in xformers
* Tiny code adjustment in the qr_ks_vs_whole_k_prefetch pipeline
* Remove one call of move_tile_window for q_dram_window
* Refine the codes in GetNumPrefetchV()/GetNumKLdsBuffers()
* Tiny fix in qr_ks_vs_whole_k_prefetch pipeline
* Adjust the location of codes for storing the first V tile to LDS
* Tiny fix and add comments
* Change GetSmemKPackK size to improve performance
* Move the codes related to K-Lds to the pipeline default policy due to some override on the generic custom_policy
* Update MakeKDramTileDistribution() and MakeKLdsDescriptor() to completely remove bank conflicts for K-Lds access
* Adjustment in intermediate iteration codes for tiny performance improvement
* Reduce the number of VLds buffers to 2 for whole_k_prefetch situtation
* Use IsFirstKLdsBufferOverlapLastVLdsBuffer() to avoid potential Lds issue
* Adjust the code location for calling IsFirstKLdsBufferOverlapLastVLdsBuffer()
* Remove useless AsyncopyV
* Rename MakeQDramTileDistribution to MakeQRegTileDistribution when LDS is not used
* Keep qx_ks_vs_custom_policy work for other pipelines and move whole_k_prefetch specific codes to whole_k_prefetch default policy
* Recover the qr_ks_vs_async pipeline
* Recover qr_ks_vs_async in fmha.hpp and tiny fix in qr_ks_vs pipeline
* Revert "Try to fix old bugs in qx_ks_vs_custom_policy"
This reverts commit 39b82ca194.
* Tiny fix with regard to whole_k_prefetch pipeline compiling
* Update kPadSeqLenK setting in fmha_fwd_kernel
* Use q_element_func and k_element_func
* Use single q_tile rather than multiple sliced q_tiles
* Codes refine according to the comments
* Re-format one file
* Mark qr_ks_vs_whole_k_prefetch as QLoadOnec == true
* Update for fmha_fwd qs_ks_vs pipeline
* Remove _builtin_amdgcn_sched_barrier(0)
* Move p_compute to p converting earlier for trying to increase vgprs re-using
* Enable GetQKBlockGemm to use WarpGemm-16x16x16 for QLoadOnce==false situation
* Re-add __builtin_amdgcn_sched_barrier(0)
---------
Co-authored-by: Po Yen Chen <PoYen.Chen@amd.com>
* Update license year
* Add initial code to override decode problem
* Fix splitkv traits/args overriding error
* Reshape and transpose lse for decode
* Remove debug code
* Prettify example code
* Use better function name
* Add kMergeNumHeadGroupsSeqLenQ flag
Kernel user can use this switch to turn on/off optimization for
some problem sizes
* Add missing flag declarations
* Default turn off kMergeNumHeadGroupsSeqLenQ in codegen
* Group similar statements together
* Remove assumption of seqlen_q=1
* Remove kMergeNumHeadGroupsSeqLenQ from splitkv combine kernel
* Support kMergeNumHeadGroupsSeqLenQ=true in fmha splitkv kernel
* Run kMergeNumHeadGroupsSeqLenQ=true kernels when need
* Fix group mode block skip logics
* Undo changes of normal fwd kernel
* Update in GridSize() and using GridSize() for splitkv kernel (#1799)
---------
Co-authored-by: Qianfeng <qianfeng.zhang@amd.com>
* Add check for zero values
* Add static assertions
* Remove invalid option '-e' in smoke_test.sh
* Use correct path of smoke_test.sh
* Avoid zero-sized shared memory array
* Add warning comment
* Replace expr by integer_divide_ceil() call
* Use more readable constant names
* Write down assumption as static assertion
* Add more diagnostic error messages
* Fix wrong BlockWarps when using default pipeline policy
* Add more static assertions for A LDS desc
* Allow using vector size < 8 for data type fp16/bf16
* Align vector size between DRAM dist & LDS desc
* Remove no-longer used func decl
* Fix wrong displayed piepline name
* Undo policy template changes for tile_example_gemm_basic
* Add missing space and make error message stands out
* Unify print precision
* Add missing include directive <iomanip>
* Replace constant 64 by get_warp_size() call
* Replace constant 128 by named variable: BankLength
* Add kAMBlock/kBNBlock attributes
* Allow usig different A/B warp dist for multiple blocks
* Add helper function to get warp dist encodings
* Add 4x64x4 fp16 warp gemm attribute impl
* Complete the A/B warp dist encoding logic
* Fix wrong thread mapping for C matrix
* Use smaller vector size for small tile
* Add static assert to block unsupported warp gemm impl
* Extract common code out as helper method
* Add 4x64x16 fp16 warp gemm type alias
* Add comment to warning developers
* Undo WarpGemmAtrributeMfma<> changes
* Use more clear static assertion error message
* Add trivial wrapper to get warp dstr encodings
* Only transpose warp gemm result if it's square
* Fix compilation error
* Support multi-block warp gemm (on N direction)
* Remove duplicated code
* Fix output encoding of warp gemm
* Fix wrong shape of WarpGemmAtrributeMfmaIterateK<>
* Remove unused code
* Fix wrong shape of WarpGemmAttributeMfmaImplF16F16F32M4N64K4
* Add type config for bf16_t
* Add 4x64x16 bf16 warp gemm
* Update WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution
* Add 64x4x4 fp16/bf16 warp gemm impl
* Add 64x4x16 fp16/bf16 warp gemm
* Add static assertion for better error diagnostic
* Get Q dram dstr directly form block gemm
* Add missing header: fused_moe.hpp
* Allow specifying different warp-gemm for gemm0 & gemm1
* Store P matrix into LDS before gemm1
* Fix inconsistant kernel name
* Remove constraint on gemm0 & gemm1 block warps
* Remove unsupported vector size from checking list
* Allow using 4x64x16 warp gemm for gemm0
* Finish policy customization
* Finish pipeline modification
F#
* Use block warps in codegen
* Fix wrong rank of m_lds_window origin
* Use better distributed tensor
* Make P-store earlier
* Remove duplicated experssions
* Remove unnecessary tile window
* Create new files for new splitkv pipeline
* Separate old/new pipeline codegen logic
* Sync changes form develop
* Undo gemm kernel/pipeline changes
* Undo gemm example changes
* Remove blank lines
* Fix typo
* Use new warp gemm interface
* Fix link error
* Fix wrong pipeline tag
* Fix more link error
* Avoid unnecessary padding
* Always use vector load for K
* Padding on fastest dimension when necessary
* Force padding Q on hdim_q
* Set high dimension padding flag to false
* Re-format headers
* Use warps=<1, 4, 1> for both gemm0 & gemm1
* Fix complilation errors
* Remove m/l shuffle logics
* Ignore duplicate data when write lse_acc
* Use gemm0 block warps as lds tile width
* Remove hard-coded numbers
* Fix wrong distribution width
* Remove unnecessary code
* Add s_barrier before writing to LDS
* Store Q into LDS before gemm0
* Fix wrong Q tile size
* Use simple Q lds descriptor for debuging
* Use more realistic Q lds descriptor
* Add comment & use better variable name
* Make Q lds space not overlapped with others
* Remove unnecessary block_tile_reduce_sync() call
* Move Q load statements
* Move block_sync_lds() right before use
* Re-order instructions
* Remove necessary lambda expression
* Use 8 threads on kMaxSplits direction while doing reduction
* Tiny correction for using 8 threads on kMaxSplits direction for combine kernel
* Padding num_split direction of o_acc tile window to 4x
* Update splitkv combine pipeline design
* Add kN1 back to splitkv combine pipeline problem
* Fix compilation errors
* Add missing template parameter
* Fix wrong splitkv combine kernel name
* Fix wrong origin
* Fix wrong LDS descriptor shape
* Fix sync & reduction logics
* Remove unnecessary static assertions
* Extract tile size computation logics
* Make sure we can reuse padding flags in combine kernels
* Rename variables
* Use OaccDataType in BlockFmhaSplitKVCombinePipelineTileSizes<>
* Remove unnecessary static assertion
* Fix function name typo
* Add constraint on kN1 template parameter
* Hide K tile loading latency in earlier iteration
* Fix wrong splitkv kernel name
* Use s_shuffling to replace p_shuffling which removes the needs of cross-warp reduction
* Rename pipeline
* Fix wrong pipeline name attribute
* Add GetAlignmentQ() for NWarpSShuffle pipeline
* Separate Q tile into dram tile & register tile concepts
* Remove non-squre warp gemm transpose c type alias
* Fallback tile size changes for fmha fwd splitkv
* Remove redundant change
* Refine naming for the S tile
* Use better naming of the S tile dstr (read from lds)
* Share Q lds with K lds
* Tiny change
* Fix with using static_for for passing CI checking
---------
Co-authored-by: Qianfeng Zhang <Qianfeng.Zhang@amd.com>
* Change in fwd-splitkv kernel to support num_splits=1 case
* Update in codegen fwd-splitkv to make num_splits > 1 cases pass
* Specify instance traits in dispatch
* Fix link error for fp8 kernels
---------
Co-authored-by: Po Yen Chen <PoYen.Chen@amd.com>
* Add ceil_to_qualified_tile_length()
* Rename kK0BlockLength to kQKHeaddim
* Add kSubQKHeaddim concept to support headdim96
* Fix in math.hpp to avoid using __half interfaces
* Add LdsBufferSequence instance for headdim96
* Update in fmha_fwd/fmha_fwd_splitkv codegen to support hd96 testing
* Disable hd96 instance generation in codegen fmha_fwd and fmha_fwd_splitkv to save compiling time
* Reformat one file
* Fix text alignment in fmha_fwd_splitkv.py
---------
Co-authored-by: Po Yen Chen <PoYen.Chen@amd.com>
* Use pre-defined constants for readability
* Use vector write for o_acc tensor
* Remove no-longer used policy method
* Deprecate no-longer used policy/pipeline
* Specify gemm0/gemm1 block warps separately in codegen
* Fix wrong ps_idx creation logic
* Add single-warp block gemm
* Supoprt single-warp gemm0
* Make MakeCBlockTile() as static method
* Use MakeCBlockTile() to get underlying tile distribution
* Use kNumGemm1Warps to compute # threads for gemm1
* Put normal case in the if clause
* Refine fmha splitkv block mapping
* Refine & fix the lse_acc/o_acc layout
* Fix wrong LDS size for K tile
* Use kK0=64 for hdim=128,256 fmha splitkv kernels
* Use kK1=64 for hdim=32,64,128 fmha splitkv kernels
* Undo kK0/kK1 changes
* Use more reasonable GetAlignmentV() computation
* Using store_tile() in fmha splitkv kernel epilogue
* Add kQKHeaddimForGemmN and kVHeaddimForGemmN in order to support headdim 96
* Remove the using of MakeKRegBlockDescriptor and MakeVRegBlockDescriptor
* Fix in bwd_piple_default_policy
* Remove kQKHeaddim and rename kQKHeaddimForGemmN to kQKHeaddim in the bwd kernel and pipelines
* Replace kVHeaddimForGemmN by kVHeaddim and kDoDvHeaddim
* Update to hd96 tile settings
* Add smoke test scripts for fmha-bwd hd96
* Revert "Add smoke test scripts for fmha-bwd hd96"
This reverts commit 7ca7e1a93d.
* Remove hd96 tile settings in fmha_bwd codegen to save compiling
* Fix lost code line in bwd_pipeline_default_policy
* Merge kDoDvHeaddim/kPadHeadDimDoDv to kVHeaddim/kPadHeadDimV and remove TileFmhaBwdTraits
* Rename KRegSliceBlockDescriptor/VRegSliceBlockDescriptor to KRegBlockDescriptor/VRegBlockDescriptor
* tiny adjustments
---------
Co-authored-by: Po Yen Chen <PoYen.Chen@amd.com>
Co-authored-by: danyao12 <Dan.Yao@amd.com>
* ake the cshuffle compilable
* modify Mhe reference on gpu and cpu. Correaccess of cshuffle
* fix the cpu reference code
* Complete the in tile shuffle logic
* restructure the kernel template input
* change the naming pattern of ck_tile gemm pipeline
* Re-format files using remod.py
* Solve the fmha conflict with gemm
* Comment Addressed from Carlus
---------
Co-authored-by: Po Yen, Chen <PoYen.Chen@amd.com>
* Simplify the codes in splitkv_combine pipeline
* Always set kPadSeqLenK=true for fmha splitkv kernels
* Change in Oacc Alignment and TileDistribution to be more adaptable to tile sizes
---------
Co-authored-by: Po Yen Chen <PoYen.Chen@amd.com>
* Use same layout for o_acc and o tensor
* Use better param names in partitioner
* Remove redundant kargs 'max_seqlen_q'
* Use better param names in splitkv kernel
* Add comment for additional kernel arguments
* Sync empty loop early return logics between pipelines
* Pass more arguments to cmake in scripts
* Align backslashes
* Fix wrong o_acc tensor view strides
* Change o_acc layout if o_perm=0
* Handle whole row masked via attn_bias
* Use use vector width = 1 for o_acc
* Use more even split sizes
