* fix async copytest bug
* Add block_sync_lds_direct_load utility
* fix the s_waitcnt_imm calculation
* Improve s_waitcnt_imm calculation
* fix vmcnt shift
* add input validation and bug fix
* remove unnecessary output
* move test_copy into test
* change bit width check
* refactor macros into constexpr functions
which still get inlined
* wrap s_waitcnt api
* parameterize test
* cleanup
* cleanup fp8 stub
* add fp8 test cases; todo which input parameters are valid?
* replace n for fp8 in test cases
* add large shapes; fp8 fails again
* change input init
* test sync/async
* time the test
* clang-format test
* use float instead of bfloat to cover a 4-byte type
* fix logic - arg sections should be 'or'd
* make block_sync_lds_direct_load interface similar to old ck
* fix a few comment typos
* name common shapes
* revert the example to original logic of not waiting lds
* clang-format
---------
Co-authored-by: Max Podkorytov <4273004+tenpercent@users.noreply.github.com>
Co-authored-by: Thomas Ning <Thomas.Ning@amd.com>
* ck_tile kernel for gemm with groupwise quantized A or B tensor.
This change introduces new pipelines with Intrawave scheduler and block gemm primitives that loads the scale tensor to registers to perform dequantization post MFMA on C tensor in registers.
Scale tensor data, AQ/BQ is spliced across threads in registers and not stored in LDS.
Current support is for the following combinations, but it should be fairly straightforward to extend support to more formats.
1. fp8, fp8 -> f32
2. bf8, bf8 -> f32
3. i4, fp8 -> f32
4. i4, bf8 -> f32
Group size can go down to as low as K length of underlying WarpGemm primitive.
For Gemm problems with quantized B tensor, this change also introduces preliminary support for flatmm pipeline which loads B tensor directly into registers.
* [Block Scale Gemm] Only run gemm quant examples on __gfx94__
- Only run gemm quant examples on __gfx94__ for usage of
`v_cvt_pk_fp8_f32`
- Format the code
* [Block Scale Gemm] Remove Bquant Gemm BlockScale
This cleanup is in preparation for future development of bquant. By
isolating Aquant-related code, we can streamline the codebase and make
it easier to add and maintain bquant functionality in subsequent
updates.
* [Block Scale Gemm] Format code with clang-format-12
The latest clang-format (v19) in ROCm 7.0 generate different result than
clang-format-12 which is used in CK CI.
Format code with clang-format-12 for consistency.
* [Block Scale Gemm] Split the k direction loop
- Split the k direction loop in block_universal_gemm_as_quant_bs_cr.hpp
to make the logic clearer.
- Disable C transposition.
* [Block Scale Gemm] Move block scale gemm example to 38_block_scale_gemm
* [Block Scale Gemm] Update copyright
* test
* Add TailHandler
* Move TileDistributionEncodingPatternAQ
* Refactor
* refactor
* fix bug
* fix bug
* help solve the PR comment
* Format the code
* [Block Scale Gemm] Add unit tests
* [Block Scale Gemm] Add support to 16x16x32 MFMA
- Add support to 16x16x32 MFMA
- Fix a bug when exchange data crossing lanes
---------
Co-authored-by: Vijay Krishnamoorthy <vjkrish@meta.com>
Co-authored-by: Cong MA <congma13@ctr2-alola-ctrl-01.amd.com>
Co-authored-by: ThomasNing <thomas.ning@amd.com>
* Use read_tr in universal gemm
* Enable all instances back
* Revert example37 changes
* Resolve comments
* resolve comments 2
* Fix assertion msg
* fix the gemm basic
* change index_t to bool for preshuffle variable
* Solve the comment
---------
Co-authored-by: Thomas Ning <Thomas.Ning@amd.com>
Co-authored-by: Po Yen Chen <PoYen.Chen@amd.com>
Co-authored-by: Max Podkorytov <4273004+tenpercent@users.noreply.github.com>
Co-authored-by: AviralGoelAMD <aviral.goel@amd.com>
Without this DataType = unknown -
``` sh
Run Flatmm kernel with DataType = unknown M =1280 N =16384 K =1024 StrideA =1024 StrideB =1024 StrideC =16384 : 0.228837 ms, 187.687 TFlops, 341.374 GB/s,
```
after this change
```sh
Run Flatmm kernel with DataType = bf16 M =1280 N =16384 K =1024 StrideA =1024 StrideB =1024 StrideC =16384 : 0.227029 ms, 189.181 TFlops, 344.092 GB/s,
```
* fix bug in loops that need use local tokens to compute
* support extra chain local_token
* update
* update
* refine some main
* update
* support dispatch_policy
* fix 15 example
* Initial commit
* Adding new tile partitioner to flatmm
* intermediate changes
* debugging kernels
* Updating flatmm example to universal gemm example
* updated flatmm kernel to run via gemmKernel
* update universal gemm to incorporate flatmm
* debug
* Fix flatmm call
* Fixing other kernels and tests for API changes
* clang formatted
* fixing gemm tests
* added test for flatmm and simplify kernel arguments
* adding flatmm test
* fix test for flatmm
* simplify gemm kernel with flatmm
* remove flatmm related files
* addressing review comments and code clean up
* resolving empty file
* resolving empty file
* clang formatted
* addressing review comments
* enable persistent kernel for flatmm
* reverted the removed files for flatmm
* reverted the removed files for flatmm
* changed flatmm to weightPReshuffle; removed the _1 added in teh faltmm example
* some more renames
* clang formatted
* Wrap tile size mapping as class method
* Warp pipeline generating as class method
* Add constraint as kernel dispatching criteria
* Support mutltiple tile size for a (hdim, hdim_v) combination
* Use smaller tile size if CU utilization is low
* Use integar as the key of the tile size map
* Fix type error
* Simply override parent class method return value
* Add attribute to eliminate warnging
* Allow using environment variables to turn on/off custom factory
* Unify param naming style
* Add missing HIP runtime include directive
* Fix os.environ.get() usage
* add for async load builtin
* add async load api
* fix some compiling errors
* fix a compiling error
* fix some compiling errors
* add a pipeline which copies from v4
* add a new pipeline for async load
* fix some compiling errors
* add async load tests
* fix some issues in async load
* fix
* fix async inline assembly
* fix async inline assembly
* add ignore header file
* comment some not gfx950 codes
* comment some not gfx950 codes
* fix a error
* update async load apis
* fix lds descriptor
* fix a compiling error
* fix some compiling errors
* fix a descriptor issue
* update lds descriptor
* change async pipeline's tile distribution pattern from thread to warp
* fix clang format
* update async policy
* fix a CRTP issue
* fix a typo error
* change lds layout
* fix some sync issues
* improve codes
* delete the async test
* fix a commented format issue
* avoid compiling device functions when compile host
* make gemm run
* add the copy kernel support
* finish the feature
* Address comment
* add the support for buffer_builtin
* solved the merging problem
* Comment Addressed
---------
Co-authored-by: joye <joye@amd.com>
Co-authored-by: joyeamd <John.Ye@amd.com>
* add prefetching physical block id for pagedkv
* start add pagedkv prefill
* rename pipeline
* add kernel for pagedkv
* add an init version pagedkv prefill
* fix redefine issue
* add struct BlockFmhaFwdPagedKVPipelineProblem and fmha_fwd_pagedkv_args
* generate dispatch code
* add body generating code
* comipling pass
* remove dropout from pagedkv
* set lse to false in generating code
* start changing qr kernel to pagedkv
* init version of kernerl with pagedkv
* change names of file that are generated
* chang host validation for pagedkv prefill
* using iglp to change blockgemm
* add kernel files to op head file
* show parameters
* rewrite print parameter fun
* add fwd
* remove default parameter of GridSize
* format
* fix nhead issue and add seqlen_k_ptr to batch mode
* format code
* remove no-longer used code
* format
* fix some comments
---------
Co-authored-by: ltqin <letaoqin@amd.com>
Co-authored-by: Po Yen Chen <PoYen.Chen@amd.com>
* updates to support int8 in 03_gemm example
* added comments, using aliases, helper functions
* test(gemm_universal): add test cases for int8 gemm pipeline
* fix(test_gemm): fix for failing test unit test for int8
* test(ck_tile): add int8 unit test for gemm universal
* refactor(gemm_universal): GPU reference verification for GEMM code improved
* style(gemm_universal): removed extra comments and did clang format
* merging recent changes to universal gemm to tile_engine
* ck tile engine integration work
* feat(tile_engine): add int8 support to tile engine ops/gemm
* feat(tile_engine): added 32 32 16 mfma instances to tile engine for int8
* style: Format code with clang-format-12
* refactor(tile_engine): address review comments
* style: removed unhelpful comments & unused variables.
* build: tile engine uses default config
* feat: add int8 support for CK_TILE GEMM
* style: added trailing commas to codegen_utils.py
* refactor: tile engine
* refactor: formatting and code review
* refactor: code formatting for python files
* fix: suppress build warning
* add support for gfx950
* refactor:KWarpTile size in gemms util
* Fix the branch and wrap up the k warp tile
* Add bf8 integration
* refactor: clang format and rebase
---------
Co-authored-by: zjli2013 <leezhengjiang@gmail.com>
Co-authored-by: AviralGoelAMD <aviral.goel@amd.com>
Co-authored-by: Khushbu Agarwal <khuagarw@amd.com>
* [CK_TILE] Refine fp8 in flatmm
1. Replace USING_MFMA_16x16x32 & USING_MFMA_16x16x32 with constexpr
2. Add an additional const check to avoid build error in HotLoopScheduler
3. Refine shuffleb to support both tile 32x32 and 16x16
4. Support command option -init
5. Move Gemm warp defintion to a separate struct
* fix clang format
* fix clang format
* keep default bhavior unchanged (warp tile = 16x16)
* fix tile engine build error
* fix a typo in codegen_utils.py
* address review comments
* address review comments
---------
Co-authored-by: Thomas Ning <Thomas.Ning@amd.com>
* Avoid passing indices (std::vector) by value to host tensor's operator()
Each access requires 2 allocations and copies of the vector.
* Remove 1 unneeded vector copy from the slowest part of fmha_bwd's verification
* Compute ds_hp_host_ref in parallel
This sequntial ForEach is the slowest part of validation and it benefits
from parallel computation.
* Do not use ForEach for simple copy and conversion of large tensors
These tensors all have the same shape {nhead, real_seqlen_q, real_seqlen_k} and
can be copied/converted without complex computations of linear indices.
