* Pass hdim to tile_example_fmha_fwd in fp8 tests
* Add WMMA support to fwd FMHA pipelines
* Tune tile sizes a bit for less spilling
fp16 256 is still quite slow
* Fix Q grad tile distribution for warp size = 32 and hdim >= 256
With AccDataType = float and warp size = 32, K0 becomes 0, K repeat is required to correcty distribute the tile.
* Use code based on BlockDropout in BlockDropoutBwd
* Fix split KV combine kernel for gfx12 (warp size 32) and make it more universal
* Fix LSE LDS tensor descriptors: kMaxSplits and kM0 were swapped, it worked on gfx9
because they both equal to 8 while on gfx12 they are 8 and 4;
* Fix Oacc LDS tensor descriptor: it was transposed even though its shape=[4 * kM0, kN1],
it worked on gfx9 because 4 * kM == kN1 == 32;
* Removing these hidden dependecies allows to support:
* any number of warps (power-of-2), not only 4;
* kN1 = 16, not only 32;
* any number of splits;
* Rename ids like o_acc_4 and Oacc4 to eliminate confusion: kNumWarps doesn't have to be 4 now
* Replace hard-coded kN1 in dispatch code with the requested tile size
* Add gfx12-specific tile sizes for split KV
* Pass GPU architecture to kernel generation scripts
This is still a temporary solution.
* Build and run FMHA CI tests for gfx12
* Fix issue after merging
* Fix bwd tile sizes
The current pipelines always read only one tile K and V tile, this
requires bk0 == bhdq and bk2 == bhdv (kK0 == kQKHeaddim and
kK2 == kVHeaddim).
* Use hardware f32->f8 on gfx12, remove v_perm
__builtin_amdgcn_perm is not needed because
__builtin_amdgcn_cvt_pk_fp8_f32 allows to specify which word (16 bit of
32-bit dword) is used to store results (two f8 values).
* Update changelog
* Add WMMA support to pagedkv
* Fix scripts after rebasing
* Support 16x16 (MFMA, WMMA) and 32x32 (MFMA) tiles in fwd and bwd BlockDropout
Add comments with dropout implementation details
Fix performance regression of fwd+dropout
* Remove some usage of type punning (reinterpret_cast with ref or ptr) in Philox;
* "scalarize" seed and offset, they may come either from kernel args or from device memory
(presumably loaded with vector loads).
These changes help the compiler to procude more optimal code and reduce register spilling.
Use WarpGemmDispatcher instead of explicit WarpGemmMfma... to get CWarpDstrEncoding
Use code based on BlockDropout in BlockDropoutBwd
Refactor BlockDropout (fwd)
Implement BlockDropout (fwd) for WMMA
Originally BlockDropout only supported 32x32 tiles (IsWG32 = true),
this version supports 16x16 tiles.
If MPerBlock > MWarp * 16, it can generate numbers for two 16x16 tiles, similarly
to BlockDropoutBwd.
Implement BlockDropoutBwd for WMMA
Remove MakeRandValLds* functions unused in BlockDropoutBwd
Remove unused Run overload from BlockDropoutBwd
* Fix regression with philox seed and offset when they exceed 32-bit int
__builtin_amdgcn_readfirstlane works with 32-bit values, seed and offset
are 64-bit so they get truncated.
* Fix names after cherry-picking
* Fix selection of a fallback tile based on bm0
The assumption that the largest bm0 == 128 is not always true for
current fp32 tiles.
* Do not use filters related to qr_async_trload
They disable tiles/pipelines which are valid for gfx12.
* Use different dstr encoding when C is transposed
* Do not call GetQKBlockGemm (and hence WarpGemmDispatcher) in host code
Some WarpGemmDispatcher instantiations are defined only
for specific archs and undefined on host.
Calculations related to sched barriers are moved from Pipeline's public
fields into pipeline's operator().
* Fix incorrect name WarpGemmMfmaFp8Fp8F32M32N32K16SwizzleBTransposedCDistribution
Correct name is WarpGemmMfmaFp8Fp8F32M32N32K32SwizzleBTransposedCDistribution
because it's 32x32x16 with IterateK = 2 so K = 32, also all tiles used
in codegen scripts are 32, 32, 32.
* Generalize usages of WarpGemmDispatcher for MFMA and WMMA
WarpGemmMfmaFp8Fp8F32M32N32K32SwizzleBTransposedCDistribution is still
used explicitly becaus of swizzle factor = 4.
* Mark has_load_tr as maybe_unused
There are no transpose loading for RDNA.
* Remove CK_TILE_USE_MFMA/WMMA from fmha-related code
* Detect BlockSize on host based on warp size of the current device
If kBlockSize == kNumWarps * get_warp_size(), the kernel is launched with
kBlockSize / 2 because on host get_warp_size() == 64 always.
* Fix calculation of grid size for combine kernel with warp size = 32
* Add missing includes and header
* Support multiple archs in one binary for fwd
* Support multiple archs in one binary for fwd_splitkv, fwd_appendkv, pagedkv_prefill
* Support multiple archs in one binary for bwd
* trload kernels are compiled only for gfx950;
* instances with padding are checked after instances without padding so
they can be used as fallbacks (similarly to fwd);
* Extract common code from register_traits
* Revert "Fix regression with philox seed and offset when they exceed 32-bit int"
To simplify merging , the proper fix is in develop already.
* Support new numerical d paddings in trait ordering checks
* Build fp32 tests only on gfx9
* Do not use hardcoded M0 = 64 for dot bwd kernel
* Use textwrap.indent from standard library
* Make fp8 pipelines on gfx12 consistent with gfx9
* Update tests for current pipelines
* Make ninja check more responsive in CI
ninja buffers output so this job looks hanging.
* Support fp8fp32 by limiting O vector size
The fp32 output type requires storing 8 * sizeof(float) = 32 bytes,
which is not implemented (here 8 is the number of C values per lane for
v_wmma_f32_16x16x16...).
* Remove unused cmake options
* Unify including amd_buffer_addressing.hpp/_builtins.hpp
* Temporarily use amd_buffer_addressing.hpp on >=gfx10
amd_buffer_addressing_builtins.hpp uses inline asm for loads/stores
which is not compatible with >=gfx10:
* 1 scalar for exec masks instead of 2,
* gfx12 uses different instruction names etc.
* Update asm in bf16 conversions to work with warp 32
* Do not generate splitkv/appendkv with vlayout=col for consistency with fwd
* Add arch tags to kernels/host funcs, compile for each arch separately
* Add kM0 to fmha_bwd_dot_do_o kernel name to match filename
* Add workaround for miscompilation of bwd with padded hdim
SWDEV-559729: v_wmma instructions can be incorrectly placed in divergent
branches used to store padded tensors (when some lanes are inactive due
to padding). Inline asm with dummy dependencies on VGPRs of the tensors
prevents the compiler doing this.
* Fix add_gtest_executable for absolute paths
Some tests (like gemm_tile_engine) pass absolute paths to source files.
In CI the branch name is a part of the root dir, and if the branch name
contains "wmma", "xdl" etc., files can be incorrectly excluded.
* Run only hdim 128 smoke tests for fp8fp32
There are no instances for hdim 64 and 256.
* Format py with ruff to simplify merging develop
* Fix incorrect var name
* Codegen for gfx9,gfx950 when --targets is not specified
Aiter and Pytorch require changes for passing their targets to the codegen scripts.
With this temporary solution the files are generated but not all of them
have to be really built (depending on the used --offload-arch=).
* Combine arch-related values into ArchTrait
This more centralized approach removes duplication of various formatting templates.
* Try a workaround for Jenkins error "groovyjarjarasm.asm.MethodTooLargeException: Method too large"
Some code is extracted into a function.
* [CK_TILE] fmha: Add query padding support to backward pass
Introduces support for query sequence padding (q_padding) in the FMHA backward pass kernels.
- Passing `seqlen_q_ptr` to the backward kernels to distinguish logical from physical sequence lengths.
- Updating `OGradDotO`, `ConvertQGrad`, and `DQDKDV` kernels to respect logical lengths and handle zero-length sequences.
- Aligning LSE indexing in the forward kernel with the padded layout for consistency.
- Adding a new GTest suite (`test_fmha_bwd_kernel_padding.cpp`) with comprehensive tests for various padding scenarios, including zero-length
sequences and deterministic mode.
* fix clang format
* Adapt fmha_bwd_runner.cpp to new q, kv sequence padding
Add backward q/kv sequence padding unit tests.
* [CK_TILE] fmha: Unify sequence length and padding handling
Refactor the handling of sequence lengths and padding in the
FMHA forward and backward kernels to provide a more unified and flexible
interface.
- Replaced `seqstart_padded_*_ptr` with a more robust system that uses
`seqstart_*_ptr` for physical sequence lengths and introduces
`seqlen_*_ptr` and `cu_seqlen_*_ptr` for logical (unpadded) lengths.
- Established a clear order of precedence for determining sequence
length: cumulative lengths (`cu_seqlen_*_ptr`) take priority,
followed by per-sequence lengths (`seqlen_*_ptr`), and finally
physical lengths derived from `seqstart_*_ptr`.
- Clarified the distinction between "group mode" and "batch mode" and
how sequence lengths are handled in each case.
- Renamed `cu_seqlen_kv_ptr` to `cu_seqlen_k_ptr` for consistency.
- Updated comments and documentation to reflect the new argument
structure and usage.
---------
Co-authored-by: illsilin_amdeng <Illia.Silin@amd.com>
* Support 16x16 (MFMA, WMMA) and 32x32 (MFMA) tiles in fwd and bwd BlockDropout
Add comments with dropout implementation details
Fix performance regression of fwd+dropout
* Remove some usage of type punning (reinterpret_cast with ref or ptr) in Philox;
* "scalarize" seed and offset, they may come either from kernel args or from device memory
(presumably loaded with vector loads).
These changes help the compiler to procude more optimal code and reduce register spilling.
Use WarpGemmDispatcher instead of explicit WarpGemmMfma... to get CWarpDstrEncoding
Use code based on BlockDropout in BlockDropoutBwd
Refactor BlockDropout (fwd)
Implement BlockDropout (fwd) for WMMA
Originally BlockDropout only supported 32x32 tiles (IsWG32 = true),
this version supports 16x16 tiles.
If MPerBlock > MWarp * 16, it can generate numbers for two 16x16 tiles, similarly
to BlockDropoutBwd.
Implement BlockDropoutBwd for WMMA
Remove MakeRandValLds* functions unused in BlockDropoutBwd
Remove unused Run overload from BlockDropoutBwd
* Fix regression with philox seed and offset when they exceed 32-bit int
__builtin_amdgcn_readfirstlane works with 32-bit values, seed and offset
are 64-bit so they get truncated.
* Add F32 MFMA warp gemms
* Support f32 in fwd FMHA
* Implement transpose_vectors for 4-byte types (float)
* Fix unexpected implicit f32->uint32 cast in buffer_store<4>
__builtin_amdgcn_raw_buffer_store_b32 expects unsigned int but float was passed (implicitly casted to uint).
mbuf_t types in other buffer_store<> are changed for consistency.
* Support F32 in bwd FMHA
hdim = 256 is disabled for now because it uses too much memory on gfx90a
* Support Headdim = 48 (divisible by 16) in fwd
* Add fp32-specific receipts (800 and 801)
* Tune fwd tiles
* Tune bwd tiles
* Use small tiles only for small seqlen_q
* Fix after rebasing
* Fix selection of a fallback tile based on bm0
The assumption that the largest bm0 == 128 is not always true for
current fp32 tiles.
* Remove constraints and adjust filtering for fp32
Custom constraints are no longer needed because now the smallest tile
is selected automtically based on seqlen_q.
Filters related to qr_async_trload disabled valid fp32 tiles.
* Add fp32 tests
* Make splitkv and appendkv compile for fp32 only
There are no instances yet, but API still must compile when only fp32 is
requested.
* Remove unimportant f32 instances
* Add test_ck_tile_fmha_*_fp32 to REGRESSION_TESTS
* Replace magic numbers with a constant, improve comments for dropout
* Update changelog
* Fix condition that dq_acc must be set to zero when mask is used
The change was introduced in #2799
* Replace warp_uniform with recently added amd_wave_read_first_lane
* Add hdim = 96 and 192 to fwd
* * [CK_TILE] Add sequence padding and variable length support in fmha (and v3)
- Group Mode Padding: Introduces the `-s_qpad` argument to support
physically padded layouts. Kernels now use padded start pointers
(`seqstart_padded_*_ptr`) for memory addressing.
- Batch Mode Variable Length: Adds `-q_eff_lens` and `-kv_eff_lens`
arguments for efficient processing of variable-length sequences by
passing cumulative effective lengths (`cu_seqlen_*_ptr`) to the kernel.
- FMHA examples: Support padding and variable length both in
group and batch mode. Dispatcher is updated as well (dispatch to
kPadSeqLenK enabled pipeline).
- New padding test cases: Add padding test cases to `smoke_test_fwd.sh` and
`test_fmha_fwd.inc`, and add benchmarks to `benchmark_fwd.sh` and
`benchmark_fwd_v3.sh` as well. These test cases and benchmarks that
specifically validate/benchmark the new padding and variable-length
functionalities in both group and batch modes.
* [CK_TILE] Fix build error in fmha unit tests
* [CK_TILE] add mqa, gqa to sequence padding unit tests
* [CI_TILE] Reduce the number of padding seqlen unit tests in FMHA to avoid timeouts in CI
* [CK_TILE] remove unnecessary MageKArgs overload in FmhaFwdV3Kernel and FmhaFwdKernel
* Have a workable version for SGPR
* have a workable version for atomic add
* Revert "have a workable version for atomic add"
This reverts commit 792377a590c26cfff9c8f545d9a9e8484a7422eb.
* substitute with the new sgpr read api
* update the CHANGELOG
* have a workable version for atomic add
* Revert "have a workable version for atomic add"
This reverts commit 792377a590c26cfff9c8f545d9a9e8484a7422eb.
* change to static for logic
* have a workable version for atomic add
* Revert "have a workable version for atomic add"
This reverts commit 792377a590c26cfff9c8f545d9a9e8484a7422eb.
* [CK_TILE] Add sequence padding and variable length support in fmha (and v3)
- Group Mode Padding: Introduces the `-s_qpad` argument to support
physically padded layouts. Kernels now use padded start pointers
(`seqstart_padded_*_ptr`) for memory addressing.
- Batch Mode Variable Length: Adds `-q_eff_lens` and `-kv_eff_lens`
arguments for efficient processing of variable-length sequences by
passing cumulative effective lengths (`cu_seqlen_*_ptr`) to the kernel.
- FMHA examples: Support padding and variable length both in
group and batch mode. Dispatcher is updated as well (dispatch to
kPadSeqLenK enabled pipeline).
- New padding test cases: Add padding test cases to `smoke_test_fwd.sh`,
and add benchmarks to `benchmark_fwd.sh` and `benchmark_fwd_v3.sh` as well.
These test cases and benchmarks that specifically validate/benchmark the
new padding and variable-length functionalities in both group and batch modes.
* [CK_TILE] Fix build error in fmha unit tests
---------
Co-authored-by: Po Yen Chen <PoYen.Chen@amd.com>
Co-authored-by: Yi DING <yi.ding@amd.com>
* change host using fp16 to check
* fp8 to fp8 compare
* rewrite input parameters
* add not squant
* remove some output code
* for scale = 1
* format
* saturates only for fp8
* add fp8bf16 data type
* add fp8bf16 data type
* fix test fp8 code
* add run_fp8bf16_tests
* change fmha fwd example parameter(adding fp8bf16)
* Support fp8bf16 for Aiter
* Support aiter fp8bf16 in c++
* fix comment about fp8 in readme.md
* add fp8fp32
* add fp8fp32 test
* remove range_q etc.
* format
* fix test parameters about squant and fmha example input fp8bf16 fp8fp32 data type
* add fp8bf16 to data_type function
* change colmajor to rowmajor in test_ck_tile_fmha_fwd_fp8
* format
* reset atol for fp8
* fix bug for atol
---------
Co-authored-by: rocking <ChunYu.Lai@amd.com>
Co-authored-by: asleepzzz <hanwen.chang@amd.com>
Downstream libraries aren't migrated to c++20 yet, so replace a use of c++20 concept with equivalent SFINAE logic. The template checks for both the existence and the truthiness of the static member variable.
* Remove some duplicate code in fmha_fwd_appendkv_kernel.hpp
* Simplify two templated operator calls by having the templated types deduced automatically
* Simplify two GemmPipeline calls
* Fix GemmPipelineAgBgCrCompV4::GetName
* Refactor use of ArgParser in CK tile GEMM examples
* Update args in README.md to match the implementation in create_args
* Remove some unnecessary include statements
* Rename two variables
* Factor out common code
* Factor out do_verify
* Add and use type aliases for memory operation integral constants
* In gemm_basic.cpp, use kPadM, kPadN, kPadK, and kBlockPerCu from GemmConfig
---------
Co-authored-by: Adam Osewski <19374865+aosewski@users.noreply.github.com>
* Add separate mask checking for scope [aligned_physical_seqlen_k_start, physical_seqlen_k_end) in pagedkv pipeline
* i_nhead_ conversion type to prevent overflow
---------
Co-authored-by: ltqin <letaoqin@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>
* Write soft-sign in inline asm
* Change tile idx computation
* Add macro to turn off soft-sign asm opt
* Use simple for loop to avoid register spill
* Only do block id transform for masking cases
* Re-implement qr_ks_vs_async pipeline by using kLoadOnce
* Remove last block_sync_lds() in the loop
* Tiny adjustment in qr_ks_vs_async pipeline for better performance
* Rename MakeQDramTileDistribution to MakeQRegTileDistribution for QLoadOnce pipeline
* Use LDS as intermediary stop when loading Q from global memory for qr_ks_vs_async pipeline
* Use un-rolled gemm for Gemm-0
* Use k0_loops small tile load/store to replace the big tile load/store for K
* Remove the commented lines in qx_ks_vs_custom_policy.hpp
* Tune the prefetching of V in qr_ks_vs_async pipeline
* Move the codes for storing the first v_lds tile some later
* Let BlockDropout reuse LDS with V
* Switch to separate code blocks according to iteration index
* Interleave code blocks for better performance
* Move clear_tile(s_acc) for better interleaving
* Move code interleaving
* Use MakeQDramTileDistribution for q_dram_window
* Roll-back to load Q directly from global memory instead of using LDS as intermediary stop
* Let V reuse the LDS of K
* Use array of tiles to represent Q in vgprs
* Use QLoadOnce == false for qr_ks_vs_async pipeline
* Special treatment for hdim-96 to save vgprs in qr_ks_vs_async pipeline
* Define statically indexed array k_lds_windows[] to reduce the using of get_slice_tile()
* Move the definition of v_tiles out from the loop
* Define statically indexed array v_lds_windows[] to reduce using of get_slice_tile()
* Remove using KLoadOnce in qx_ks_vs_custom_policy
* Remove un-used get_slice_tile() call
* Move the code line of clear_tile(s_acc)
* Tune the lines of codes to make them more tidy
* Re-arrange the codes before the main-loop
* Add comments
* Unify the alignment to be 8 for Q/K/V Lds decriptors
* Tuning to K pre-loading
* Tune K Lds and V Lds reuse for kPreloadWholeNextIterationK == false
* Adjust the pipeline codes
* Use NumPrefetchV to separate from NumVLdsBuffers
* Tune the location of a scheduler barrier code line
* Prefetch first v_tile at earlier time for both kPreloadNextWholeIterationK true/false paths
* Adjust the using of kPadSeqLenQ and kPadSeqLenK in the kernel
* Use __builtin_amdgcn_sched_barrier(0x7f) in the pipeline
* Move the location for store_tile() of first v_tile
* Rename the qr_ks_vs_async pipeline to qr_ks_vs_whole_k_prefetch pipeline
* Re-add NumPrefetchK as template for BlockFmhaPipelineQXKSVSCustomPolicy<>
* Try to fix old bugs in qx_ks_vs_custom_policy
* Remove K_LDS_LOAD_USE_OFFSET_TRANSFORM code-path to make qr_ks_vs_async and qx_ks_vs_custom_policy simpler
* Fix in MakeKDramTileDistribution() in qx_ks_vs_custom_policy
* Update to LdsBufferSequence and introduce NumKVLdsBuffers for max(NumPrefetchK, NumPrefetchV)
* Tiny Fix (#1888)
* Ck tile/paged attention workaround (#1894)
* Correction in GetRangeAlongX()
* Work-around to solve the failures in test_paged_attention_ck in xformers
* Tiny code adjustment in the qr_ks_vs_whole_k_prefetch pipeline
* Remove one call of move_tile_window for q_dram_window
* Refine the codes in GetNumPrefetchV()/GetNumKLdsBuffers()
* Tiny fix in qr_ks_vs_whole_k_prefetch pipeline
* Adjust the location of codes for storing the first V tile to LDS
* Tiny fix and add comments
* Change GetSmemKPackK size to improve performance
* Move the codes related to K-Lds to the pipeline default policy due to some override on the generic custom_policy
* Update MakeKDramTileDistribution() and MakeKLdsDescriptor() to completely remove bank conflicts for K-Lds access
* Adjustment in intermediate iteration codes for tiny performance improvement
* Reduce the number of VLds buffers to 2 for whole_k_prefetch situtation
* Use IsFirstKLdsBufferOverlapLastVLdsBuffer() to avoid potential Lds issue
* Adjust the code location for calling IsFirstKLdsBufferOverlapLastVLdsBuffer()
* Remove useless AsyncopyV
* Rename MakeQDramTileDistribution to MakeQRegTileDistribution when LDS is not used
* Keep qx_ks_vs_custom_policy work for other pipelines and move whole_k_prefetch specific codes to whole_k_prefetch default policy
* Recover the qr_ks_vs_async pipeline
* Recover qr_ks_vs_async in fmha.hpp and tiny fix in qr_ks_vs pipeline
* Revert "Try to fix old bugs in qx_ks_vs_custom_policy"
This reverts commit 39b82ca194.
* Tiny fix with regard to whole_k_prefetch pipeline compiling
* Update kPadSeqLenK setting in fmha_fwd_kernel
* Use q_element_func and k_element_func
* Use single q_tile rather than multiple sliced q_tiles
* Codes refine according to the comments
* Re-format one file
* Mark qr_ks_vs_whole_k_prefetch as QLoadOnec == true
* Update license year
* Add initial code to override decode problem
* Fix splitkv traits/args overriding error
* Reshape and transpose lse for decode
* Remove debug code
* Prettify example code
* Use better function name
* Add kMergeNumHeadGroupsSeqLenQ flag
Kernel user can use this switch to turn on/off optimization for
some problem sizes
* Add missing flag declarations
* Default turn off kMergeNumHeadGroupsSeqLenQ in codegen
* Group similar statements together
* Remove assumption of seqlen_q=1
* Remove kMergeNumHeadGroupsSeqLenQ from splitkv combine kernel
* Support kMergeNumHeadGroupsSeqLenQ=true in fmha splitkv kernel
* Run kMergeNumHeadGroupsSeqLenQ=true kernels when need
* Fix group mode block skip logics
* Undo changes of normal fwd kernel
* Update in GridSize() and using GridSize() for splitkv kernel (#1799)
---------
Co-authored-by: Qianfeng <qianfeng.zhang@amd.com>
* Remove using tile partitioner for fmha_fwd_kernel
* Remove using tile partitioner for fmha_fwd_splitkv and splitkv-combine kernels
* Remove using tile partitioner for fmha_fwd_appendkv kernel
* Unify the format of GetTileIndex
* Add check for zero values
* Add static assertions
* Remove invalid option '-e' in smoke_test.sh
* Use correct path of smoke_test.sh
* Avoid zero-sized shared memory array
* Add warning comment
* Replace expr by integer_divide_ceil() call
* Use more readable constant names
* Write down assumption as static assertion
* Add more diagnostic error messages
* Fix wrong BlockWarps when using default pipeline policy
* Add more static assertions for A LDS desc
* Allow using vector size < 8 for data type fp16/bf16
* Align vector size between DRAM dist & LDS desc
* Remove no-longer used func decl
* Fix wrong displayed piepline name
* Undo policy template changes for tile_example_gemm_basic
* Add missing space and make error message stands out
* Unify print precision
* Add missing include directive <iomanip>
* Replace constant 64 by get_warp_size() call
* Replace constant 128 by named variable: BankLength
* Add kAMBlock/kBNBlock attributes
* Allow usig different A/B warp dist for multiple blocks
* Add helper function to get warp dist encodings
* Add 4x64x4 fp16 warp gemm attribute impl
* Complete the A/B warp dist encoding logic
* Fix wrong thread mapping for C matrix
* Use smaller vector size for small tile
* Add static assert to block unsupported warp gemm impl
* Extract common code out as helper method
* Add 4x64x16 fp16 warp gemm type alias
* Add comment to warning developers
* Undo WarpGemmAtrributeMfma<> changes
* Use more clear static assertion error message
* Add trivial wrapper to get warp dstr encodings
* Only transpose warp gemm result if it's square
* Fix compilation error
* Support multi-block warp gemm (on N direction)
* Remove duplicated code
* Fix output encoding of warp gemm
* Fix wrong shape of WarpGemmAtrributeMfmaIterateK<>
* Remove unused code
* Fix wrong shape of WarpGemmAttributeMfmaImplF16F16F32M4N64K4
* Add type config for bf16_t
* Add 4x64x16 bf16 warp gemm
* Update WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution
* Add 64x4x4 fp16/bf16 warp gemm impl
* Add 64x4x16 fp16/bf16 warp gemm
* Add static assertion for better error diagnostic
* Get Q dram dstr directly form block gemm
* Add missing header: fused_moe.hpp
* Allow specifying different warp-gemm for gemm0 & gemm1
* Store P matrix into LDS before gemm1
* Fix inconsistant kernel name
* Remove constraint on gemm0 & gemm1 block warps
* Remove unsupported vector size from checking list
* Allow using 4x64x16 warp gemm for gemm0
* Finish policy customization
* Finish pipeline modification
F#
* Use block warps in codegen
* Fix wrong rank of m_lds_window origin
* Use better distributed tensor
* Make P-store earlier
* Remove duplicated experssions
* Remove unnecessary tile window
* Create new files for new splitkv pipeline
* Separate old/new pipeline codegen logic
* Sync changes form develop
* Undo gemm kernel/pipeline changes
* Undo gemm example changes
* Remove blank lines
* Fix typo
* Use new warp gemm interface
* Fix link error
* Fix wrong pipeline tag
* Fix more link error
* Avoid unnecessary padding
* Always use vector load for K
* Padding on fastest dimension when necessary
* Force padding Q on hdim_q
* Set high dimension padding flag to false
* Re-format headers
* Use warps=<1, 4, 1> for both gemm0 & gemm1
* Fix complilation errors
* Remove m/l shuffle logics
* Ignore duplicate data when write lse_acc
* Use gemm0 block warps as lds tile width
* Remove hard-coded numbers
* Fix wrong distribution width
* Remove unnecessary code
* Add s_barrier before writing to LDS
* Store Q into LDS before gemm0
* Fix wrong Q tile size
* Use simple Q lds descriptor for debuging
* Use more realistic Q lds descriptor
* Add comment & use better variable name
* Make Q lds space not overlapped with others
* Remove unnecessary block_tile_reduce_sync() call
* Move Q load statements
* Move block_sync_lds() right before use
* Re-order instructions
* Remove necessary lambda expression
* Use 8 threads on kMaxSplits direction while doing reduction
* Tiny correction for using 8 threads on kMaxSplits direction for combine kernel
* Padding num_split direction of o_acc tile window to 4x
* Update splitkv combine pipeline design
* Add kN1 back to splitkv combine pipeline problem
* Fix compilation errors
* Add missing template parameter
* Fix wrong splitkv combine kernel name
* Fix wrong origin
* Fix wrong LDS descriptor shape
* Fix sync & reduction logics
* Remove unnecessary static assertions
* Extract tile size computation logics
* Make sure we can reuse padding flags in combine kernels
* Rename variables
* Use OaccDataType in BlockFmhaSplitKVCombinePipelineTileSizes<>
* Remove unnecessary static assertion
* Fix function name typo
* Add constraint on kN1 template parameter
* Hide K tile loading latency in earlier iteration
* Fix wrong splitkv kernel name
* Use s_shuffling to replace p_shuffling which removes the needs of cross-warp reduction
* Rename pipeline
* Fix wrong pipeline name attribute
* Add GetAlignmentQ() for NWarpSShuffle pipeline
* Separate Q tile into dram tile & register tile concepts
* Remove non-squre warp gemm transpose c type alias
* Fallback tile size changes for fmha fwd splitkv
* Remove redundant change
* Refine naming for the S tile
* Use better naming of the S tile dstr (read from lds)
* Share Q lds with K lds
* Tiny change
* Fix with using static_for for passing CI checking
---------
Co-authored-by: Qianfeng Zhang <Qianfeng.Zhang@amd.com>
