* add space_filling_curve
* cleanup and move space_filling_curve into test
* WIP: start refactoring threadwise_transfer_v1r3
* threadwise_copy works but needs further refactoring
* add some comments
* add SpaceFillingCurve::GetIndices()
* minor changes
* removed GetIndices; refactored GetDstCoordinateResetStep
* add DynamicBuffer::Transfer, but Add is not tested
* rebased agaist develop
* threadwise_copy_v6r1/v6r2/v6r3 using space-filling curve start to work
* minor changes
* refactored threadcopy v3r1, v2; removed old implementations
* clang-format
* cleanup
* fix a typo in v6r3
* format
Co-authored-by: Chao Liu <chao.liu2@amd.com>
* start conv2d bwd api
* kernel running
* add bwd reference
* change to no shuffle
* fix bwd reference
* pass verification
* add Filter1x1Stride1Pad0 and start testing
* change some tuning parameter
* fix test error
* add fp16 tuning parameter
* add bf16 tuning parameter
* add int8 tuning parameters
* change fp32 tuning parameter
* add bwd to profiler
* fix bug for bwd profiler
* fix ckProfiler bug
* change conv2d_bwd_xdl to fp16
* fix bug in comments
* fix precompile id
* fix enum conv name
* chage _bwd_ to _bwd_data_
* change conv2d_bwd example id
* bwd to bwd data
* fix prehead
* fix MakeDefaultBlock2CTileMap ,import form merge develop
* format bwd instance
* bwd to bwd data
* change name bwd to bwd data
* change name bwd to bwd data in example
* formate code
* change conv2d bwd data id in example
* rewrite readme for example
* fix CalculateMagicNumbers about div zero
* add workaround CK_WORKAROUND_SWDEV_325164
* change test_conf2d_bwd_data show info
* format
* fix bug for workaround:CK_WORKAROUND_SWDEV_325164
* formate tuning parameters
* formate tuning parameters again
* formate tuning parameters 3
* formate tuning parameters 4
* remove add function template
* format
* update comment
Co-authored-by: ltqin <letaoqin@amd.com>
Co-authored-by: Chao Liu <chao.liu2@amd.com>
* add docker file and make default target buildable
* add Jenkinsfile
* remove empty env block
* fix package stage
* remove render group from docker run
* clean up Jenkins file
* add cppcheck as dev dependency
* update cmake file
* Add profiler build stage
* add hip_version config file for reduction operator
* correct jenkins var name
* Build release instead of debug
* Update test CMakeLists.txt
reorg test dir
add test stage
* reduce compile threads to prevent compiler crash
* add optional debug stage, update second test
* remove old test target
* fix tests to return proper results and self review
* Fix package name and make test run without args
* change Dockerfile to ues rocm4.3.1
* remove parallelism from build
* Lower paralellism
Co-authored-by: Chao Liu <chao.liu2@amd.com>
* add gitignore
* host tensor: allow generating sequentially increasing value in a given dimension
* gridwise gemm v3r1: allow distinct K0/K1 values for A/B block descriptor
- remove dangling header include
- modify example gemm_xdl accordingly
- infer KPack value from M/NPerXdl
- device conv2d fwd: update parameters accordingly for the underlying gridwise gemm v3r1
(API for conv2d fwd stays the same for now until we decide to expose individual K0s for activation and weight)
* add LDS data dump utility
* profiler: reflect API change for distinct K0/K1 for A/B matrices
* profiler: add conflict-free LDS write FP16 kernel instances
* fix accidental perf regression
* address feedback; cosmetic changes
* clang-format for new files
* format
Co-authored-by: Chao Liu <chao.liu2@amd.com>
* init for splitk f16
* a working prototype
* debug
* perf debug
* update example
* instances for mk kn
* add instances for all layers
* clean
* clean
* add tuning
* format
* add mn_padding into irregular tile
* clean
Co-authored-by: Chao Liu <chao.liu2@amd.com>
* add space_filling_curve
* cleanup and move space_filling_curve into test
* add functions for backward and forward step; hard coded results in unit test
* minor changes
* Convolution ND
* Code unification across dimensions for generating tensor descriptors.
* Example
* Instances
* Move convnd f32 instance file to comply with repo structure.
* Conv 1D tensor layouts.
* Formatting and use ReferenceConv
* Reference ConvFwd supporting 1D and 2D convolution.
* Debug printing TensorLayout name.
* Conv fwd 1D instance f32
* Refactor conv ND example.
Needed to support various conv dimensio.
Needed to support various conv dimensions
* Rename conv nd example director to prevent conflicts.
* Refactor some common utility to single file.
Plus some tests.
* Refactor GetHostTensorDescriptor + UT.
* Add 1D test case.
* Test reference convolution 1d/2d
* Remove some leftovers.
* Fix convolution example error for 1D
* Refactor test check errors utility function.
* Test Conv2D Fwd XDL
* More UT for 1D case.
* Parameterize input & weight initializers.
* Rename example to prevent conflicts.
* Split convnd instance into separate files for 1d/2d
* Address review comments.
* Fix data type for flops/gbytes calculations.
* Assign example number 11.
Co-authored-by: Adam Osewski <aosewski@amd.com>
Co-authored-by: Chao Liu <chao.liu2@amd.com>
* conv3d compiles but has memory error
* conv3d works
* fix performance issue by using __builtin_amdgc_readfirstlane
* change MakeBlock2CTileMap to MakeDefaultBlock2CTileMap; change c_blockid_to* to cblockid_to*
* clang-format
* remove CK_EXPERIMENTAL_PASS_TENSOR_DECRIPTOR_BY_*; moved wrapper into DeviceConv3d
* format
* remove useless marc
* add comment
Co-authored-by: Chao Liu <chao.liu2@amd.com>
* [What] Refactor verification of gemm alpha_beta, move to reference operation
[Why] Sync with other verification
* Profile mk_nk for gemm bias 2d
* Support bias 2d with mn * kn in profiler
* Support bias 2d with km*kn and km*nk in profiler
* Support fp32 bias 2d in profiler
* format
* format
Co-authored-by: rocking <chunylai@amd.com>
Co-authored-by: Chao Liu <chao.liu2@amd.com>
* [What] Add 2d version of bias, prepare to implement alpha / beta scaling
* Add alpha / beta functor
* Refine parameter of example
* [What] Use real type instead of template
[Why] Prevent implicit cast
* Rename parameter for general operator
* Remove redundant comment
* Fix compile error
Co-authored-by: rocking <chunylai@amd.com>
Co-authored-by: Chao Liu <chao.liu2@amd.com>
- device_gemm_xdl_c_shuffle function signature matches split-k
- retire host_driver since it is no longer maintained
- linter error (unused variable)
Co-authored-by: Chao Liu <chao.liu2@amd.com>
* tweak conv for odd C
* update script
* clean up elementwise op
* fix build
* clean up
* added example for gemm+bias+relu+add
* added example for gemm+bias+relu
* add profiler for gemm_s_shuffle; re-org files
* add profiler
* fix build
* clean up
* clean up
* clean up
* fix build
* Do not hardcode the function parameter, use template instead.
* [What] Remove AThreadTransferSrcResetCoordinateAfterRun and BThreadTransferSrcResetCoordinateAfterRun in host API
[Why] "C_Shuffle" version is supposed to be similar to the vanilla one
* Fix typo
Let DeviceGemmXdl_C_Shuffle use kernel_gemm_xdlops_v3r1
* [What]
1. Add DeviceGemmXdl_C_Shuffle
2. Revise example of gemm_xdl
[Why] Prepare to add shuffle version of D = alpha * (A * B) + beta * C
[How] Imitate DeviceGemmXdl and device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp
* fix relu
* clean up
* clean up
* adding 1x1 conv
* adding 1x1 conv
* added 1x1 conv
* refactor
* refactor
* refactor
* added profiler for conv+bias+relu+add
* clean up
* adding conv+bias+relu
* adding conv+bias+relu
* added conv+bias+relu
* Update README.md
* update cpu verification
* adding c shuffle
* update static_tensor for dealing with invalid element
* adding c shuffle
* debugging
* fix bug
* convert to fp16 before shuffle
* shuffle more than one M/NRepeat
* clean up
* remove coordinate step hack from GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
* clean up
* remove coordinate step hack from all gridwise gemm xdl
* clean up coordinate step hack
* clean up coordinate step hack
* ThreadwiseTensorSliceTransfer_v3r2 support pointwise op on both src and dst
* adding output shuffle in conv+bias+relu+add
* update
* added conv+bias+relu+add with c shuffle
* added conv+bias+relu+add with c shuffle
* fix forward_sweep bugs in threadwise copy
* clean up
* refactor
* clean up
* clean up
* added conv_c_shuffle+bias_relu
* clean up
* added conv+bias+relu+atomic_add
* clean up
* clean up
* clean up
* clean up
* clean up
* clean up
* misc fixes; add 1x1 specialization
* clean up
* delete unused device op
* clean up
* add support for odd C value
* Bug in BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2()
* Bug in ThreadwiseTensorSliceTransfer_v1r3 logic for calculating "forward_sweep"
* add add new algorithm from v4r4r2
* program once issue
* add split k functiion
* redefine code
* add a matrix unmerge
* add b matrix unmerge k0
* trans a and b to gridegemm
* nhwc init
* no hacks and vector load
* add hacks
* modify some parameter
* fix tuning prometer for fp32
* fix tuning prometer for fp16
* start change gridwise k split
* init ok
* revome a b matrix k0mk1 desc in grid
* carewrite lculate gridsize
* add kbatch to CalculateBottomIndex
* remove some unused funtion
* add clear data function before call kernel
* out hacks
* in hacks
* rename device convolution file and function name
* modify kBatch value
* fix some tuning code
* start from v4r4 nhwc
* nhwc atomic is able to run
* just for fp32
* enable nchw atomic
* tweak
* tweak
* re-arrange gridwise gemm hot loop for wrw
* add wrw v4r5
* v4r4r5 fp16
* v4r4r4 fp16
* v4r4r2 fp16
* V4R4R4XDLNHWC fp16
* V4R4R2XDLATOMICNCHW fp16
* adjust for fp16
* input gridsize
* change kbatch to gridsize
* testing wrw
* clean up
* k_batch to gridsize
* fix bug
* wrw v4r4r4 kbatch change to gride size
* wrw v4r4r2 kbatch change to gride size
* after merge , change gridwise gemm v2r4
* change MakeCBlockClusterAdaptor
* other method use new gridwise gemm
* clean up
* chapad method nge to make_right_pad_transform
* kbatch out from transform function
* clean up and fix bug
* fix bug
* using function type reduce template parameters
* using auto replace define fuction type
* clean up
Co-authored-by: ltqin <letaoqin@amd.com>
Co-authored-by: Chao Liu <chao.liu2@amd.com>
Co-authored-by: Jing Zhang <jizhan@amd.com>
* Tiny fix in using data type template parameters in blockwise and direct_threadwise kernel
* Fix with regard to implementing GetZeroVal() in both kernel and host
* Avoid convert to compType from dstDataType before writting the output value
* Add half_t support to NumericLimits and make constexpr GetZeroVal() of binary operator
* Add CONSTANT decorator for descriptor read buffer
* Use get_thread_local_1d_id() for thread local Id
* Rename GetZeroVal() to GetReductionZeroVal() in the kernels
* Remove constexpr from initialized zeroVal and tiny fix in reduction_operator.hpp
* Occasional tiny simplification and update in the kernel files
* Update to re-order tensor dimensions on the host, split second_call kernel wrapper files and simplify reduce_all kernel wrappers
* Update to remove OpenCL tidy checking failures
* Update for better readability
* Remove unused codes and not-needed template parameters in the kernel wrappers
Co-authored-by: Chao Liu <chao.liu2@amd.com>
