* 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
* Refactor the design of DeviceGemmMultipleDMultipleR_Xdl_CShuffle
* Add 'DeviceGroupedConvFwdMultipleDMultipleR' interface
* Add DeviceGroupedConvFwdMultipleDMultipleR_Xdl_CShuffle
* Remove 'GridwiseConvFwdMultipleDMultipleR_xdl_cshuffle'
* Add 'TransformConvFwdToGemm<>' utility class (from Chao)
* Use 'TransformConvFwdToGemm<>' to shorten code
* Fix ill-formed method declaration
* Re-implement MakeRGridDescriptor_M() function
* Change problem description
* Use macro to define layout types
* Define K-reduced output tensor layout types
* Let user to decide R output tensor layout
* Rename variables
* Add padding to the reduced output tensor if necessary
* Extract common code as helper method
* Remove debug message
* Add missing include directive
* Add partial fp16 Conv + Reduction example
* Add example verification code for 2D Conv problem
* Use type alias to simplify code
* Share code across different-dimension Conv problems
* Rename file/functions from run_conv_fwd* to run_convnd_fwd*
* Make example code more verbose
* Add code to support 1D & 3D Conv + Reduction on host
* Add more examples for data type: bf16, fp32
* Add example for int8
* Add custom target to group examples
* Use more general custom target name
* Change the description in error message
* Disable testing for example other than fp32
* Add examplel for int4 (just copy from int8)
* Fix wrong data type
* Use larger data type for intermediate tensors
* Finish int4 example
* Undefine macro PP_DEFINE_LAYOUT_TYPE() after use
* Use named variables to replace magic numbers
* Remove debug messages
* Use same A/B data type for host Conv in int4 example
* Add check for the 'RLayout' type argument
* Group same-dim-layouts together in 'LayoutSetting<>'
* Add 'final' specifier to utility classes
* Use different initialization method for examples
* Remove macro PP_DEFINE_LAYOUT_TYPE()
* Fix code-comment mismatch
* Use more reasonable initialization value for all data types
* Default use init_method=1 for all examples
* Remove never-used code
* Remove confusing out-of-date comments
* clean
Co-authored-by: Chao Liu <chao.liu2@amd.com>
Co-authored-by: Chao Liu <lc.roy86@gmail.com>
* add padding algo for bmm+scale+softmax+bmm. Version for verification
* remove verification code
* remove comments
* add padded bmm scale softmax bmm example
* format
* refactor
* add comments for usages of padding bmm+scale+softmax+bmm
Co-authored-by: Chao Liu <lc.roy86@gmail.com>
* comment on specialization for TensorSpecialization::Packed
* gemm_softmax_gemm with output permutation
* scaling
* refactor MatrixPadder; rename to GemmPadder
* remove old sanity check
* restore original gemm_softmax_gemm
* revise comment in gemm_softmax_gemm example
* use GetElementSpaceSize()
* remove extra header
* typo
* remove archaic DeviceOpPtr
* add examples into grouped/batched_gemm
* adding splitK examples
* fixed splitK
* add bfp16 int8 example into splitK
* formatting
* use static_cast
* added common for batched_gemm
* add commons for examples of splitK/batched/grouped_gemm
* return true
* adjust splitK check tol
* update example
Co-authored-by: Chao Liu <lc.roy86@gmail.com>
* GemmPadder and GemmGemmPadder
* proper padding using GemmGemmPadder
* test gemm_gemm padding
* properly check size K in IsSupportedArgument()
* properly check size requirement given SrcScalarPerVector in IsSupportedArgument()
* comment
* format
* Implement multiple-reduction in one kernel (kernels, device ops, examples)
* Add generic elementwise kernel and device interface
* Add generator for normal-distributed data initialization
* Add host refer implementation of batchnorm-forward and batchnorm-infer
* Add examples for implementing batchnorm-forward and batchnorm-infer using generic kernels
* Remove un-needed including in batchnorm example
* Renaming generic_elementwise to elementiwise in kernel and device classes/functions
* Change in gemm_layernorm examples to use DeviceElementwise instead of Device5AryElementwise
* Change in exampe 19_binary_elementwise to use DeviceElementwise instead of DeviceBinaryElementwise
* Change in device_cgemm_4gemm_xdl_cshuffle.hpp to use kernel_elementwise instead of kernel_binary_elementwise
* Add DeviceElementwiseBase and use it in device_normalize_instance.cpp
* Removing and renaming files
* Update to synchronize gemm_layernorm client example to the generic element-wise device op API
* Update to synchronize with the latest headers directory and HostTensorDescriptor interface renaming
* Merge two static member functions in device_elementwise.hpp
* Remove unary_elementwise_1d kernel and device
* 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>
* Implement layernorm kernel and deviceOp
* verify gpu kernel with host code
* 1. Separate gamma aand beta from affine
2. Check if argument is valid
* clean
* Sync the naming
* Support sweep once mode if we can put k dimension data inside one block
* [What] Get length from upper length.
[Why] if we get length directly, we may get length after padding.
* We only use one block in K dimension.
Hence, we can simplify the indexing of global R/W.
* Use 1d descriptor for gamma and beta
* Add accElementwiseOp
* Extract layernorm host code
* Support different YVectorDim in GridwiseLayernorm
* Rename XSrcVectorDim to XYSrcVectorDim. Because we use same parameter in deviceOp
* Gamma and beta can share the VGPR.
* Add test for fp32 and fp16
* Fix bug of concurrency and add test case which may fail orignally
* Propagate NaN for layernorm
Co-authored-by: Chao Liu <chao.liu2@amd.com>
* dump lds content in appropriate precision type
* add squared add reduction op; allows sq sum
* initial stub from regular gemm impl
* layernorm example code & host verification
* initial layernorm implementation
* tidy up
* make C0 precision type consistent with C
* clang-tidy and additional comments
* tighten up example code
* account for extra flops/bytes from normalization
* clang-format
* c0 bias/beta/gamma now have its own precision type
* AccElemOp for gemm outputs prior to feeding to layernorm
* update workgroup mapping
* rename kernel template param to reflect its dual use
* use LDS mem pool for reduction workspace
* change cshuffle precision type to f16; clean up
* clang-format
* correct naming
* explicit cast
* fully implemented gemm + bias + activation + add + norm
* activation in correct order
* reflect reduction API's recent change
* amend
* clean up; add comment
* keep up with recent changes in reduction API
* format
* resolve merge conflicts
Co-authored-by: Chao Liu <chao.liu2@amd.com>
* use 'sweep once' softmax kernel where applicable
* threadwise copy's dst buffer can specify invalid element value
* add int8 in/out float compute softmax support
give a bit of leeway for int absolute tolerance as there's a single data point of all test cases showing off-by-1 error
* format
* softmax inherits DeviceNormalization
* softmax profiler stub
* tighten up reference softmax interface
* example prints tensor dimension
* add fp32 to softmax profiler
* rename header
* hook with ckProfiler
* format
* resolve merge conflict
* resolve merge conflicts
* update normalization profiler help string
* resolve conflict
* typo
* remove residual
* softmax profiler: address feedback
* test for mixed precision input/output
* fully qualify ck::math::isnan
* add comment for device normalization interface
* revise wording
* constness for alpha/beta scaler pointer
* Extract base class for elementwise
* Refactor interface of DeviceGemmReduce. Do not use tuple in interface
* [What] Rename d into reduce in gemm + reduction related code
[Why] Prepare to add d term for add
* Unify base class of gemm + reduce and gemm + bias + add + reduce
* 1. Rename gemm_bias_add_reduce for external api
2. Refine cmake
* Add normalize device operation
* [What] Reorder the argument
[Why] Because d0 is also the input of c.
* Add type string
* Add example of gemm_bias_add_layernorm via external api
* Refactor example code
* clang-format
* Fix compile error
* clang-format
* Add external api for gemm_add_add_layernorm and normalize
* Add client example
* clang-format
* 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
* Remove template from Reducton operation classes and add template to their operator() and GetIdentityValue() interfaces
* Change to unary elementwise operators and the reduce_unary_operator (class for mapping) and dependent variations in all host layers
* Remove the data type template parameter from reduce_binary_operator (class for mapping) and dependent variations in host layers
* Add InMemoryDataOperatonSupportedOnDataType to check the matching between data type and InMemoryDataOperation
* Use struct-scope operator template instantiation for binary and unary element-wise operations
* Change a few more elementwise operations to use template for operator()
* Tiny correction in Normalize operator
* Add static_assert to check the data type appliability for some reduction accumulator and element-wise operatons
* Correction in some examples with regard to using ReduceAccDataType
* Use static_assert for UnaryDivide
* Update to merged codes to use Element-wise operations and Reduction Accumulator operations correctly
* Tiny fix with regard to SetWorkSpacePointer()
* Copy "gemm reduce" to "gemm bias add reduce"
* Implement gemm bias add reduction
* Fix compiler error due to merge from develop
* Add tensor operation for gemm + bias + add + reduce
* Add gemm_bais_add_reduce to ckProfiler
* Add c1 functor
* Refine type
* Use reduceAccDataType instead of explicitly float
* Change to use check_err()
* Do relu in float32 instead of bhalf_t. Because bhalf_t is unsigned
* Refactor relu. using type_trait instead of overloading
* Rename DxsReduceAccElementwiseOperation to DxsReduceAccElementwiseOperation
* Fix denominator
* Refine nameing
* Fix denominator in host
* Remove useless include header
* Use AccDataType
* Fix static_cast order
* Refine type
* [What] Remove tuple type in the base class
[Why] External api depend on base class. if base class has relationship with type, we will need many class for different type
* add GetWorkSpaceSize to base arg and make an example on convnd_bwd_weight
* add bwd weight for bf16: init
* remove redundant compute
* use datatype and split k to check whether a workspace is used
* remove unused computation for work space size
* add some code for bfp16
* add device/grid unary op
* add unary type convert to bwd-weight example
* support bf16 splitk kernel for convnd bwd weight
* 1. remove comments. 2. add checkvalidity. 3. add gridsize computation
* add workspace size check
* fix format
* change function name
* Use the unified naming for math functions on host and HIP kernel
* Corresponding change/simplification in reduction host/profiler/examples due to unified math functions renaming
* Renaming GetReductionZeroVal() to GetIdentityValue()
* Tiny renaming in profile_reduce_impl.hpp
* More renaming in profile_reduce_impl.hpp
* Replace zeroVal by identiyVal
* Remove ck_ prefix in the naming of ck::math provided functions
* moved gemm_descs_args into const buff
* use CK_CONSTANT_ADDRESS_SPACE instead of global constant
* clean
* moved hipMemAlloc outside of deviceOp
* add SetWorkSpacePointer
* fix ignore
