GitLab

* start fixing 16bit data packing

* adding StaticTensor

* adding StaticTensor

* adding StaticTensor

* add missing constexpr

* adding static tensor

* adding static tensor

* adding transpose

* add inline asm for transpose 2x2 of half_t

* add general transpose_vectors(), but have unnecessary register initialization using v_mov

* fix unnecessary register initialization in transpose_vector by using more pass-by-reference

* add hardcoded logic for NHWC wrw

* improve asm for v_pack

* make ThreadwiseTensorSliceTransfer_v3r2 support any tensor

* tweak

* reorganize file

* add DeviceGemmXdl

* update script

* fix naming issue

* fix comment

* output HostTensorDescriptor

* rename

* padded GEMM for fwd v4r4r4 nhwc

* refactor

* refactor

* refactor

* adding ckProfiler

* adding ckProfiler

* refactor

* fix tuning parameter bug

* add more gemm instances

* add more fp16 GEMM instances

* fix profiler driver

* fix bug in tuning parameter

* add fp32 gemm instances

* small fix

* refactor

* rename

* refactor gemm profiler; adding DeviceConv and conv profiler

* refactor

* fix

* add conv profiler

* refactor

* adding more GEMM and Conv instance

* Create README.md

Add build instruction for ckProfiler

* Create README.md

Add Readme for gemm_xdl example

* Update README.md

Remove build instruction from top most folder

* Update README.md

* clean up

* add parameters

* tweak gemm

* tweak

* update conv

* update script

* adding bwd 1x1

* update script

* adding 1x1 bwd

* debugging bwd 1x1 failure

* update script

* update script

* test

* test v100

* clean up

* add f32/i32 atomicAdd support into dynamicBuffer, and enable it in v1r3

* fixed

* fixed

* update comment
Co-authored-by: Chao Liu <chao.liu2@amd.com>

* Squashed 'src/composable_kernel/' content from commit f6edda61

git-subtree-dir: src/composable_kernel
git-subtree-split: f6edda61

* add solver ConvIgemmFwdV6r1DlopsNchwKcyxNkhw; rename static ck source files

* Squashed 'src/composable_kernel/' changes from f6edda61..5781adf5

5781adf5 Update develop (#5) (#6)
97e6d514 Merge pull request #4 from ROCmSoftwarePlatform/separate_online_compile
7b1ec41e refactor
49c33aae refactor
54b3e73d rename

git-subtree-dir: src/composable_kernel
git-subtree-split: 5781adf5



* fix

* refactor

* remove online compilation from CK

* refactor

* fix

* add ctest

* add c-style pointer cast

* vector/scalar pointer cast use c-style pointer cast instead of reinterpret_cast

* fix clang warning suppression

* tidy

* suppress cppcheck

* fix enum issue

* revert chagnes to hip build

* fix kernel filename

* update CK build script

* rename

* rename

* make innner product compatiable on gfx900

* Update src/include/miopen/solver/ck_utility_common.hpp
Co-authored-by: JD <Jehandad.Khan@amd.com>

* compiler parameter use stream

* use int instead of index_t in kernel wrapper

* DynamicBuffer, StaticBuffer, amd_buffer_load support customized value for invalid element

* refactor

* refactor

* change cmakelist

* change ck common utility

* fix
Co-authored-by: JD <Jehandad.Khan@amd.com>

* update online kernel wrapper bundle all descriptors in a tuple

* change __CONSTANT__ to CONSTANT

* rename

* adding tuning

* added IsValidCompileParameter

* reorginze

* adding tunable for fp16 and int8

* fix kernel compile warning and bug fixes

* suppress warning about cast CONSTANT (address space 4) pointer

* fix building issue

* change olc cmake

* adding online compile to fwd-v4r5r2

* update scripts

* remane fwd-v4r5r2 to fwd-v6r1

* clean up

* update default launch bounds

* add threadwise copy the copy a tensor in one copy, added kpack to DL GEMM

* add kpack into fwd v4r5 nchw fp32

* create files for xdlops

* working on blockwise_gemm_xdlops

* add KReduction

* add m/n repeats

* add 2x2 pipeline

* added 128x128 wavegemm

* use StaticBuffer of vector_type

* break vector type to blk_size

* add kpack into xldops_gemm and blockwise_gemm

* abroadcast only

* add fp32 mfma instructions

* adding fp16 mfma

* pack half4_t

* rename kperwave to kpack

* add 32x32x8fp16

* add fp16 mfma

* clean code

* clean code

* V4r4 xdlops kpack (#35)

* add kpack with incorrect results

* bug fix for make_dynamic_naive_tensor_descriptor_aligned_v2

* add 1x1 kernel

* add gridwise_gemm_v2 - single_buffer

* enabled dwordx4 for fp16
Co-authored-by: Chao Liu <chao.liu2@amd.com>

* refactor fwd-v4r4-xdlops

* add v4r4-nhwc-xdlop

* improve some perf of nhwc and nchw by tuning parameters, and change scheuduling in gridwise-gemm loop

* tweak scheduling in gridwise gemm

* add v4r3 with a single output copy

* init commit: output with slice win

* adding sliceWin

* add multiple repeats pattern

* starting adding bwd-v4r1-xdlops

* use tuple as SrcBuffer

* adding bwd-data v4r1 nhwc xdlops

* fix bug in make_dynamic_naive_tensor_descriptor_aligned_v2()

* fix bug in host bwd-data conv

* initial implementation of bwd-data v4r1 nhwc xdlops

* add launch bound flags

* enable launch bound

* add m/nrepeat=4

* tweak bwd-data v4r1 nhwc xdlops

* added bwd-data v4r1 nhwc xlops with output A and weight B

* add fwd-v4r4 nhwc xdlops, A input, B weight, C output
Co-authored-by: Chao Liu <chao.liu2@amd.com>

* Add online-compiling facility

* Synchronize from fwd-v4r5 and implement host interfaces to call conv-fwd v4r4/v4r5 using on-line compiling method

* Tiny adjustment to time reporting

* Use object assignment to replace explicit bytes copying in the first kernel of v4r4/v4r5

* Use single thread to assign descriptor object to device memory

* Adjust to the workload assignment of the two kernels of v4r4 (experimental)

* Revert "Adjust to the workload assignment of the two kernels of v4r4 (experimental)"

This reverts commit eb38461456bb0c82b6c0d32cdd616e181907e20c.

* Update to make constexpr for generating descriptor types in kernel 2 of dynamic conv-fwd v4r4

* Update to dynamic conv-fwd v4r4 online-compiling

* Update to dynamic conv-fwd v4r5 online-compiling (result not accurate)

* Tiny update to driver/CMakeLists.txt

* clang-format

* Tiny comments change

* Add env OLC_DUMP_SAVE_TMP_DIR to support saving of temperary dir

* Fwd v4r5 olc perf (#39)

* added hip-clang flags that fix perf issue of online compilation

* fix bug for olc fwd-v4r5-nchw

* Move constexpr and type reference statements out of the function body in conv-fwd v4r4/v4r5 kernel wrapper

* Remove printing in hip_build_utils.cpp

* Update to root CMakeLists.txt

* Revert "Move constexpr and type reference statements out of the function body in conv-fwd v4r4/v4r5 kernel wrapper"

This reverts commit 3d2c5d8ecdd8298b72d127110500ed5b38d9835c.
Co-authored-by: Chao Liu <chao.liu2@amd.com>
Co-authored-by: Chao Liu <lc.roy86@gmail.com>
Co-authored-by: root <root@dc-smc-18.amd.com>

* pass-by-void-pointer for gridwise_dynamic_gemm_v1r2

* use pass-by-value by default

* experimenting magic number division

* overhauling fwd-v4r4 to clearly reflect transformation graph

* added fwd-v4r5

* bug fix for make_dynamic_naive_tensor_descriptor_aligned_v2

* bug fix and added sanity-check in transform_dynamic_tensor_descriptor

* added conv_driver_v2

* Use DynamicBuffer to hold raw pointer (to global and LDS memory)

* add workaround for compiler issue (inefficient ISA) of ds_write for int8x4, int8x8, int8x16

* replacing array with tuple and vector for tensor data

* overhaul vector_type, make int8x4_t real vector instead of aliasing from int32_t

* initial implementation for magic number division and DynamicMerge_v2_magic_division that uses it

* turn off DynamicMerge_v2_magic_division that use magic number division by default

* use address_space(4) in kernel signature to fix performance issue when passing tensor descriptor from host to kernel by (void) pointers

* remove passing by pointer* option (only use pass by value or void*)

* support dynamic tensor descriptor

* use buffer load OOB feature for padding case

* add navi support

* add int8x4 inference kernel
Co-authored-by: Chao Liu <chao@ixt-rack-81.local.lan>
Co-authored-by: Jing Zhang <jizhan@amd.com>

* Use buffer load built-in OOB check. buffer size is limited to 2GB.
* buffer APIs use combined wave and thread offset
* use uint32_t for addr shift in buffer addressing

* tuning para,

* testing on v100

* add fp16

* remove deprecated tensor descriptor

* sync with miopen

* update build script
Co-authored-by: Jing Zhang <jizhan@amd.com>

* update for miopen integration: cosmetic refactor

* update for MIOpen integration

* Added bwd data v3r1: breaking down compute into a series of load balanced GEMM, and launch in a single kernel
* Added bwd data v4r1: like v3r1, but launch GEMMs in multiple kernels
* Tweaked v1r1  and v1r2 (atomic) on AMD GPU

* enabled atomic add in tensor copy
* added gridwise GEMM
* added backward data conv using GEMM + atomic
* added backward data conv using GEMM, no atomic

Refactor, so can bring multi-index transformation and padding support into MIOpen