* Add basic support for direct loads from global to LDS
* Clean the code and comments
* Add support for fp16
* Add comments
* Add check for thread cluster lengths
* Align non-direct-load fp16 example
* Small fixes
* Extend IsSupported to check for supported GPU gens
* Build examples only on the supported HW
* Do not throw when instance not supported in 04 example
* Review: Apply review suggestions
* Review: small fix
* Review: small fix
* Redesign the DPP8 GEMM kernel to use warp-wise component
* Review: Improve error messages
* Review: Remove unnecessary empty lines
* Review: Fix M, N per thread names
* Review: Rename mfma_input_type to dpp_input_type
* Review: Fix tensor adaptor; remove unnecessary element
* Review: Remove calls to dpp_gemm's MakeCDescriptor
* Review: Add blockwise doc, change function names to include dimension names
* Review: Remove duplicated code; Move Block2CtileMap alias to the top of the file
* Review: Add __restrict__ keywords
* Review: Use MatrixPadder for padding A, B, C matrices
* Review: Remove hardcoded datatypes
* Review: Change names from FloatX to XDataType
* Review: Introduce AK0 and BK0 instead of a single K0
* Review: Remove construction of dpp_datatypes object
* Review: Rename DppInstrRunner to DppLanegroupGemm
* experiment with config file
* experiment with version.h config
* add more info to version.h
* minor updates
* minor updates
* fix case where DTYPE is not used
* large amount of files but minor changes
* remove white space
* minor changes to add more MACROs
* fix cmakedefine01
* fix issue with CK internal conflict
* fix define and define value
* fix clang-format
* fix formatting issue
* experiment with cmake
* clang format v12 to be consistent with miopen
* avoid clang-format for config file
* initial stream-k implementation with example
* fix unexpected change in err
* improve a little bit performance by reorganize pipeline.
* improve perf a little bit by swizzle block idx
* add profiler
* update example
* fix spelling
* shrink karg for streamk
* support dynamic buffer using memory coherence glc_slc bit from template
* control memory coherence while construct dynamic buffer
* update reduction for streamk(not ready yet)
* Add template parameter to make_dynamic_buffer to support amd_buffer coherence setting
* fix build issue
* fix several bug
* now result is correct, everything works (but has scratch)
* remove scratch by manually reset coordinate
* update device code
* fix a bug in final reduce
* fix something in example
* update async memset
* fix enum as camel case
* modify coherence enum name
* clean code and use atomic streamk by default
* remove unused var
* throw exception if have empty pointer
* fix format
* fix CI warning
* fix type in init
* modify CI error
* filter out on gfx10+
* restore changed example code
---------
Co-authored-by: Qianfeng Zhang <Qianfeng.Zhang@amd.com>
* Use dim 0 as faster dim for writing mean/var/count workspace in batchnorm multiblock method [performance]
* Add CountDataType as template parameter in blockwise_welford
* Add utility/get_shift.hpp
* Add BatchNorm multiblock single-kernel implementation
* Add smem inline assembly based implementation of gms_init/gms_barrier/gms_reset for gfx90a
* Renaming in device_batchnorm_forward_impl.hpp
* Tiny fix in the batchnorm_fwd profiler
* Revert "Add smem inline assembly based implementation of gms_init/gms_barrier/gms_reset for gfx90a"
This reverts commit d16d00919c.
* Use the old two-kernel batchnorm multiblock method for gfx1030
* Use the old two-kernel batchnorm multiblock method for gfx908
* use the single-kernel batchnorm multiblock method only for gfx90a
* Remove get_wave_id() from utility/get_id.hpp since it is not used
* Set true for testing running mean/variance and saving mean/invvariance in the examples
* Fix to copy-right words
* Remove un-needed including in utility/get_id.hpp
* Add comments to workgroup_synchronization.hpp
* Remove un-used codes in gridwise_multiblock_batchnorm_forward.hpp
* Renaming in the kernels
* Remove un-used kernel file
* Modify Doxygen config to pick up include directories recursively
* Add DeviceMem struct to API Reference guide
* Add classes that are used in Flash Attention kernel
* Add a reference and config for generating bibliography
Co-authored-by: Philip Maybank <Philip.Maybank@amd.com>
* wmma_op + unit test
* add arch limitation to wmma test
* change arch limitation
* Refactor + Add all type unit test(int4 compile failed)
* Add f32_16x16x16_bf16 unit test
* tempsave
* tempsave
* tempsave
* runtime bug, cannot find symbol
* workaround for incorrect HIP warpSize return value
* debugging
* tempsave
* Correctness OK, waiting for optimization
* Tidy up + format
* temp save
* temp save, reproduce the v_bfi_b32 issue
* add inline asm for wmmaop test
* tidy up
* clean some debug purpose code
* discard some codes
* clang format
* clang format
* compiler issue fixed + increase tile size
* Add gridwise gemm pipeline v1/v2 selector
* Pipeline selector working, test-wise add pipeline options to one instance
* Add gemm instances
* Add debug info to DeviceGemmXdl
* Add debug info to DeviceGemmXdl_CShuffle
* Add debug info to DeviceGemmXdl_CShuffle and instances to gemm_add_add_fastgelu
* Minor fix
* Add debug info to DeviceBatchedGemmXdl and instances to batched_gemm
* set up inter-wave configuration
* use defualt loop scheduling for supported gemm ops
for blanket-applying interwave scheduling for all supported gemm ops, define macro CK_EXPERIMENTAL_DEFAULT_TO_INTER_WAVE_SCHEDULING=1. this should be discouraged though as it is not covered by CI
* Add enum PipelineVersion
* Update instances
* Format
* Fix the merge conflict
* Add flags to disable added instances
* Test disable flag check
* Disable flag check
* Enable the instances
Co-authored-by: Anthony Chang <ac.chang@outlook.com>
* reopen masking att instance due to CI is upgraded
* re-enable instances previously failed on 9110
* enable ksize-kpadding pair validity test
* add non-masked attention+permute test; expose masking boolean to attention kernel handles
* disable bench
* fix test
* move files
* bulk rename batched_gemm_masking_scale_softmax_gemm_permute to batched_gemm_softmax_gemm_permute
* format
* amend rename
* disable bench in test
* add mask/no-mask test for non-permute attention kernels
* disable broken kernel instance
* example working
add non-permuted problem statement
evaluating whether overhead comes from permutation or the extra kernel arg
* interface for bias addition without implementing it
* test and profiler running
* tidy
* mask type determined by enum class
* unify example code
* move masking specialization to its own header
* align formats
* extract helper functions
* experiment merging dims for attn w/ permute; shows perf parity with attn wo/ permute
* add tensor specialization to template args
since tensor spec packed shows perf parity when permutation isn't needed
remove redundant template args
comment on 'packed' tensor specialization
* grouped attention with input/output permute example
* format
* clean up
* refactor acc0 tile visitor
Co-authored-by: shaojiewang <wsjmessi@163.com>
Co-authored-by: Chao Liu <chao.liu2@amd.com>
* modify comment
* trim unnecessary check
* add gemm spec in kernel name
* add TNTT gemm_gemm + atten kernel instances
* refactor attention padding to better fit in unit tests
This streamlines usage where "ResetNaNToMinusInf" is now hidden from user facing device op.
Also added compile-time conditionals that load OOB value as NaN only after padding is enabled
* add adhoc padding test for atten
* shrink input value range for attention kernel validation to avoid occasional error by 1e-3
Still unsure whether this kind of deterministic floating point accurary issue is expected
or not. May want to try exact same approach as the GPU kernel in the host reference
GEMM+Softmax+GEMM function to see if the accuracy discrepancy goes away. Until then,
shrink the input value range as it is less likely to produce errors of around ~1e-3.
* attention kernel proper granular padding for all 4 dims
* IsSupportedArgument checks
* test more padded cases
* block PadK specialization in attention kernels
* workaround clang crash for gfx908
(gfx908 only) workaround for compiler crash in fused kernels on mainline #9110; #10738 seems ok
error message was "fatal error: error in backend: Error while trying to spill VGPR0 from class
VGPR_32: Cannot scavenge register without an emergency spill slot!"
this fall back to less ideal way of handle NPadding in fused attention kernel
* comment out kernels giving wrong results on MI100; MI200 doesn't seem affected
* Add threadwise and blockwise welford
* Rename gridwise op, prepare to add welford version
* implement welford and integrate welford into layernorm
* Take care of tail loop
* Fix buf when ThreadSliceK > 1
* Fix bug of merging of two empty set
* Rename clip to clamp
* 1. Fix type of count
2. Remove useless static_assert
* Do not inherit Reduction::Argument
* [What] replace __syncthreads() with block_sync_lds()
[Why] __syncthreads might wait both lgkmcnt(0) and vmcnt(0)
* Add y stride
* Rename.
DeviceLayernorm -> DeviceLayernormImpl
DeviceNormalization2 -> DeviceLayernorm
* Move literal ""_uz & ""_zu into namespace 'literals'
* Move namespace 'literals' as 'ck::literals'
Co-authored-by: Po-Yen, Chen <PoYen.Chen@amd.com>
Co-authored-by: Chao Liu <chao.liu2@amd.com>
* initial stub for gemm_gemm_xdl_cshuffle
* set up example code
* compiles
* prevent integer overflow
* harmonize interface between ref_gemm and ref_batched_gemm
* batched_gemm_gemm
* fix example
* host tensor gen: diagonal pattern in lowest two-dimensions only
* make c descriptors containing only integral constants
* clean up
* add BlockwiseGemmXdlops_v2 while exploring an unified approach
* implement proper interface
* tidy up example
* fix compilation warnings
* coarsely controlled 2nd gemm padding
* remove rocm-cmake's hard requirement for certain revision
* clang-format
* resolve merge conflict
* fix compilation error on gfx10
* adds acc0 elementwise op to interface
* add gemm_gemm instances and tests
* avoid LDS data hazard
* fix build
Co-authored-by: Chao Liu <chao.liu2@amd.com>
* initial stub for gemm_gemm_xdl_cshuffle
* set up example code
* compiles
* prevent integer overflow
* harmonize interface between ref_gemm and ref_batched_gemm
* batched_gemm_gemm
* fix example
* host tensor gen: diagonal pattern in lowest two-dimensions only
* make c descriptors containing only integral constants
* clean up
* add BlockwiseGemmXdlops_v2 while exploring an unified approach
* implement proper interface
* tidy up example
* fix compilation warnings
* coarsely controlled 2nd gemm padding
* remove rocm-cmake's hard requirement for certain revision
* clang-format
* resolve merge conflict
* fix compilation error on gfx10
* adds acc0 elementwise op to interface
* attention host validation
* add blockwsie softmax v1
* iteratively update softmax+gemm
* transpose both gemm0 and gemm1 xdl output so as to avoid broadcasting softmax max/sum
* add init method for easier debugging
* do away with manual thread cluster calculation
* generalize blockwise softmax interface
* row-wise softmax sum & max
* format
* rename to DeviceBatchedGemmSoftmaxGemm
* add gemm_softmax_gemm instances and tests
* comment
Co-authored-by: ltqin <letao.qin@amd.com>
Co-authored-by: Chao Liu <chao.liu2@amd.com>
* initial stub for standalone softmax
* start device_softmax_mk_to_mk as a wrapper to device_reduce_mk_to_m
* host softmax validates
* compiles; to implement beta scaling
* use NaN trick to efficiently ignore OOB values during sum of exponentials
* freeload device_reduce's utility functions
* clean up interface
* adding prior value (beta scaling)
* remove restriction related to perf considerations
* apply clang-format
* clean; disable diagnostics
* resolve conflicts
* add exp wrapper
* honor HostTensorDesc interface; allow implicit cast from different vector<T> type
* test softmax for fp16/fp32
* update readme
* amend commit NaN trick
* remove redundant param added during development
* format
* replace ScalarDataType with AccDataType
* separate out test programs by precision type
* move softmax sample code to its own folder
* format
* keep up with recent changes in reduction API
* remove extra header
* start adding navi21 GEMM
* navi_gemm_km_kn_mn_fp32 compiles and passes one test.
* rename variables and functions in gridwise_gemm_dlops_v1r3
* add other 3 layouts; format instance
* adding more tuning parameters
add tuning parameters for other 3 layouts
* add gemm_dlops_f16
* tmp
* add dependence of DeviceGemm::IsSupportedArg() on arch
* minor changes
* minor changes
* minor changes
* minor changes
* minor changes
* minor changes
* minor changes
* push gemm_dlops into profiler
* minor changes
* if using xdl or dlops is moved into profiler_gemm_impl
* minor changes
* minor changes
* remove is_xdl from profile_gemm_impl
* make IsSupportedArg dependent on arch for other device_gemm
* minor changes
* minor changes
* fix a bug in f_generate_tensor_value
* add 64x64x64 for gemm_dlops_int8
* add 64x64x64 for gemm_dlops_int8
* comment out 3 layouts in gemm_dlops_int8; add 32x32x32 for gemm_dlops_int8; init A values to 1
* fix
* start fixing tuning parameters
* monir
* minor changes
* minor changes
* minor changes
* fixing
* adding example
* adding example
* adding example
* add gemm fp32 example
* clean up
* use 128x128x16 as MNK tile in navi21 gemm example
* bug fix
* fix test
* use new block c tile
* clean
* fix build
Co-authored-by: Chao Liu <chao.liu2@amd.com>
Co-authored-by: shaojiewang <wsjmessi@163.com>
* manual control of MAC cluster for improved 2-wave performance
ensure setprio's order; ensure inner loop size >= local read size
synchronize when single mac cluster
* format
* use value field from ck::integral_constant
* roll out inter-wave loop scheduler to c-shuffle gemm variants
will gradually roll out to other applicable device ops when occasional reg spill is resolved
* additional comments
* format
* fix mismatch between inter-wave pipeline and interwave blockwise gemm
* address review feedback
* amend
* Add ThreadwiseReduction functor as per-thread reduction api
* Using ThreadwiseReduce api and some change in using PartitionedBlockwiseReduction api to simply the kernels
* Add comments and remove useless declarations in the kernels
* Tiny updates
* Use thread cluster descriptor and explicit M_K 2d descriptor to simply Blockwise Reduction
* Change by replacing ReduceDims by NumReduceDims as Device Reduce interface template parameter
* Rename the folder name for the pool2d and reduce examples
* Update to reduction test scripts
* Add Readme for pool2d_fwd and reduce_blockwise examples
* Tiny fix in reduce profiler and tiny update in reduce testing scripts
* Tiny fix in testing script profile_reduce_no_index.sh
* Tiny change in script/profile_reduce_with_index.sh
* Renaming and refining in Reduction profiler/device layer/examples
* Renaming and refining in Reduction profiler/device layer/examples
* Renaming all NumReduceDims to NumReduceDim
