* fix cppcheck errors, first pass
* fix format
* fix returned value in examples
* add macro definitions for cppcheck
* fix the profile_gemm logic
* update the gemm profiler logic
* add more difinitions to cppcheck, fix couple more errors
* replace runtime error with message in device function
* fix a couple of int4 issues
* no return for fill function
* fix errors in data_types.hpp
* fix format
* fix few remaining errors
* fix errors in data_types.hpp
* fix last couple of errors in datat_types.hpp
This PR introduces support for double buffering in LDS into GEMM kernels that use direct load instructions.
Direct loads now use inline asm instead of intrinsics. Usage of intrinsics results in compiler adding additional waitcnt instructions what breaks possible load/compute overlap in case of double buffering.
Usage of inline asm results in the need to use sched_barrier in order to make sure that compiler cannot incorrectly reschedule instructions since it does not know the data dependencies between global->LDS and LDS->registers.
* 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
* Introduce LocalBlockToCTileMap.
* Change the signature of CalculateBottomIndex() function which now does
not accept any argument. The B2C map which is already passed as an
argument to the kernel Run function is calculating block's local id
already outside at kernel entry point __global__ function.
The LocalB2C map stores as members local block ID.
* Use LocalBlockToCTile map in device ops.
* First draft of tile loop work distribution.
* Fix typo.
* Simplify kernel arguments.
Calculate descriptors & B2C maps on the device.
* Use looping kernel.
* Fix B2C constructor.
* Fix Navi21 errors.
* Calculate tile start/end in device kernel.
* Change Run API to accept user provided workspace buffer.
* Add new line at EOF.
* Move Gemm KernelArguments to device op interface.
* Remove unused code.
* Update API.
* Launch grid size which is min of occupancy vs tile count
* Get back to use constant memory for gemm descriptors.
* Remove unused code.
* Add default virtual method implementation.
* Update comments to conform with doxygen style.
* Fix doc style and unused parameters.
* Add thread cluster lengths to kernel name.
* Remove old splitk impl and replace it with tile looping one.
* Modify instances.
* set KPerBlock to 64
* maximize wherever possible vector load size.
* Fix instances cluster lengths.
* Change comment style.
* Use 128b store where possible in instances.
* Update test cases, since KPerBlock has doubled.
* Update output stream operator for Sequence.
* Add pipeline version to GroupedGEMM device op type string.
* Fix pipeline version type logging.
* Fix input tensors type after merge.
* Fix compiler error.
* Fix output stream operator for Pipeline version.
* Store using 128b.
* Set of instances with kpb 32/64
* Limit number of instances
* Remove commented out instances.
* Fix function name.
* Limit the number of instances.
Add pipline version to the regular instances
* Change thr cluster layout for reading B tensor.
* disabled failed instances
---------
Co-authored-by: Adam Osewski <aosewski@amd.com>
Co-authored-by: zjing14 <zhangjing14@gmail.com>
Co-authored-by: Jing Zhang <jizha@amd.com>
* workaround nan problem by changing output to fp16
* enable f8/bf8 gemm tests on MI200
* workaround f16 to f8 conversion
---------
Co-authored-by: Jing Zhang <jizha@amd.com>
* Add column to image kernel
* Minor fixes for dtypes and client examples
* Disable tests for disabled dtypes
* Disable add instances functions for disabled data types
* Minor stylistic fixes
* Revert "Disable add instances functions for disabled data types"
This reverts commit 728b869563.
* Instances reduction
* Add comments in device_column_to_image_impl
* Update changelog and Copyrights
* Improve changelog
* 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
* Add maxpool instances
* Rename index pool to max pool.
* Add maxpool bwd bf16 instances
* Add avg pool bwd instances
* Rename avgpool and maxpool to avg_pool3d and max_pool
* Add bf16 pool fwd instances
* Add max pool bwd to ckProfiler
* Add avg pool3d bwd to ckProfiler
* Add avg pool bwd test
* Fix bug of reference pool fwd (dilation)
* Fix bug of max pool bwd (dilation and initZero)
* Support bf16 compute data type
* Force compute type be f32. Because atomicAdd only support f32
* Add max pool bwd test
* Rename folder
* Rename pool
* Add max pool bwd client example
* Add avg pool bwd client example
* Add missing workspace
* clang format
* Rename macro
* remove useless header
* remove useless layout
* Add s_nops after v_dot to avoid hazard
* Fix builtin for inner_produxt fp16
* Skip inline version to builtin
* Add comments regarding isa
* Fix comment regarding s_nop
* 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
* Add basic fp8 definitions and prn-generator
* Format
* Add fp8<->fp32 type_convert
* Format
* Split type_convert and cast_to/from_f8
* Format
* Minor fix
* Minor fix
* Move fp8 utils to a separate header
* Add elementwise ops
* Add fp8_convert_sr
* Format
* Add element op
* Eliminate magic numbers
* Split f8_convert_sr in host and device
* Format
* Add some constexpr
* Add a datatype test
* Format
* Another format
* Add fp8<->fp16 tests
* Update type_converts
* Format
* Add fp16 casting functions
* Format
* Use seed as a runtime arg
* Use element location for PRNG
* Format
* Add fp8<->fp16 to PassThrough element op
* Clean up
* Merge host and device implementations
* Add comments on rounding modes
* Remove leftover code
* Put type_converts into a separate header
* Put random number gen to a separate header
* Rearrange f8_utils' namespaces
* Refactor type_convert.hpp
* Move f8_t definition
* enable gfx941/942 targets
* fix clang format
* fix the cmake logic for multiple targets
* fix cmake syntax for looping over targets
* add gfx941/942 support for gemm_xdl instances
* Add overloaded version of __builtin_amdgcn_readfirstlane()
* Remove 'static' specifiers
* Remove more 'static' specifier
* Replace unsigne char by std::byte
* Add 'const' specifier to never changing variable
* Add 'inline' specifier to funcion definition
* Fix wrong boundar calculation logic
* Rename type trait
* Remove std:: qualifier from standard types
* Replace 'size_t' by 'unsigned'
* Use type alias to hint usage
* Replace static_for<> by ordinary 'for' loop
* Rename readfirstlane() to amd_wave_read_first_lane()
* Rename file readfirstlance.hpp as amd_wave_read_first_lane.hpp
* Reorder statements
* Add TypeConvert class and start refactoring
* Refactor TypeConvert as a struct
* Get back to template functions type_convert
* Add a type_convert_bf16_rtn, set rtz as default
* Clean up
* Add UnaryConvertPrecision struct for high-precision workloads
* Format
* Update type_convert to UnaryConvert on threadwise level
* Update UnaryConvertPrecision
* Format
* Fix chmod
* Add a flag to pick converion method
* Format
* Remove the added flag
* Merge elementwise op with type conversion
* Move type_convert to elemwise op, update the op
* Update type_convert_precision -> bf16_convert_rtn
* Clean up
* Update comments
* Update the CK_WORKAROUND_DENORM_FIX flag handling
* Update the unneeded op to work but warn user
* Remove the message
* Use a PassThrough instead of ConvertBF16RTN to calcaulate reference
* Format
* Add missing include
* Rename to proper naming
* Add example of groupnorm + swish
* Extract duplicate code in example
* Add groupnorm + swish instances
* Ractor instance generation, split into multiple cpp file
* Add external api and client example
* Refine profiler message
* Use ck math version of exp
* Refine problem size in example
* Add host version of exp
* Add type_convert implementations for bf16
* Add the fix for conv_fwd
* Add the fix for conv_bwd_data
* Add the fix for conv_bwd_weight
* Format
* Format
* Another format
* Add a macro to use workaround on MI200 only
* Format
---------
Co-authored-by: Rosty Geyyer <rosty.geyyer@amd.com>
Co-authored-by: zjing14 <zhangjing14@gmail.com>
