* Introduces the new partitioner to implement the reduction StreamK kernel
* Add more doc text to functions
* Add persistent-dp option to streamk example
* Update example/ck_tile/40_streamk_gemm/README.md
This change replaces pipeline macros like CK_TILE_PIPELINE_COMPUTE_V3,
CK_TILE_PIPELINE_MEMORY, etc in the CK Tile examples with a common enum
called GemmPipeline to reduce code duplication.
* Adding a ds permute fallback for the gfx908 and older for row_newbcast:7 instruction
* Better macro for selecting ROW_NEWBCAST
* clang-format the update
---------
Co-authored-by: Illia Silin <98187287+illsilin@users.noreply.github.com>
* Refactor quant group size to be configurable for M/N/K, not just K
* add some asserts for configurations not implemented
* start setting of group size for N dimension
* enable 2d for reference quant gemm
* WIP: trying to figure out tile dstr and/or indexing for scale matrix
* WIP
* Fix handling of n dim blocks in tile windows etc
* remove commented code and enable all tests again
* fix formatting
* Add more specialized tile distributions
* Enable NWarps replication for bquant tile dstr
* fix formatting
* fix format
* Fix some issues from the merge
* fix formatting
* one more fix to tile dstr, and revert debug initialization
* Remove commented code
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
* simplify conditions that are needed for tile distributions
* only enable the working group sizes in tests
* fix formatting
* Update tile distribution for 2D bquant
* add some documentation and 2d block scale example
* fix formatting
* Add in Changlog and restructure the quant 2d example
* fix CMake
* support the change for blockscale 2d
* fix the test file
---------
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
Co-authored-by: Cong Ma <congma13@amd.com>
Co-authored-by: ThomasNing <thomas.ning@amd.com>
* Refactor split-image implementation: simplify code and remove redundant variables
* Add padding debug output to split-image implementation
- Added debug prints for padding calculations in transform_conv_fwd_to_gemm.hpp
- Verified padding works correctly with all tests passing
* Fix sign comparison warning after rebase with origin/develop
- Cast blockIdX from unsigned to signed index_t for comparisons
- Integrated with new GetOutputTileIndex logic from upstream
- Updated to use amd_wave_read_first_lane instead of __builtin_amdgcn_readfirstlane
* Fix Split-N with groups bug and clean up unused parameters
- Fixed batch stride calculation to include G dimension for grouped convolutions
- When moving between batches in NHWGC/NWGC/NDHWGC layouts, need to account for all groups
- Removed unused multi-split parameters (we only support 2-way split)
- All tests now pass: G=1 with Split-N, G>1 with Split-N, G>1 without Split-N
* Implement recursive queue-based split-image detection and calculation
- Add LaunchKernelWithSplitIfNeeded() helper method in transform_conv_fwd_to_gemm.hpp
- Implement recursive binary splitting algorithm (10GB→5GB+5GB→...)
- Correctly handle odd dimensions (61→30+31)
- Calculate proper offsets for each split piece
- Update invoker to use split-image helper
Note: Split detection and calculation work correctly but kernel launching
for individual pieces requires kernel modification to handle different
spatial dimensions (unlike Split-N which uses blockIdx.z).
* WIP: Split-Image investigation - found architecture mismatch
- Split-N modifies N_ directly in transformer constructor
- Split-Image needs different approach due to varying dimensions
- Added split calculation logic for 1D and 2D convolutions
- Still facing memory issues when creating piece transformers
Key finding: Split-N uses blockIdx.z for parallel execution,
while Split-Image needs sequential execution of non-uniform pieces.
* Add 1D split-image implementation for grouped convolution (N=1 working)
Implements split-image for 1D convolution to handle large tensors that
exceed memory thresholds. This is a critical milestone with N=1 fully
working and tested.
Key Changes:
- Invoker: Add split-image logic that splits W dimension in half
- Transformer: Add SplitConvProblem helper for recursive splitting
- Calculate offsets for LEFT and RIGHT pieces
- Launch two kernels sequentially (LEFT then RIGHT)
Implementation Details:
- Binary split: divides W dimension by 2
- LEFT piece: W=0 to W/2, keeps left padding, removes right padding
- RIGHT piece: W/2 to W, removes left padding, keeps right padding
- Offset calculation accounts for stride, dilation, and padding
- Physical memory offset (no padding in memory)
Test Results (N=1):
✅ 94/94 tests passing
- Comprehensive tests: 36/36 (channels, padding, stride, dilation, filters, groups)
- Edge case tests: 31/31 (odd dimensions, extreme parameters, boundaries)
- Stress tests: 27/27 (maximum dimensions, up to 91.4 TFlops)
Known Limitations:
- Only works with N=1 (single batch)
- N>1 fails when split-image triggers (offset calculation issue with Split-N)
- Root cause: Split-N modifies N in transformer, but offset calculated in invoker
- Solution planned: Move offset calculation to transformer (next phase)
Files Modified:
- grouped_convolution_forward_invoker.hpp: Add split-image logic
- transform_conv_fwd_to_gemm.hpp: Add SplitConvProblem helper
This commit represents a stable, tested 1D split-image implementation
for N=1 cases. It's an important milestone before extending to N>1
and multi-dimensional splits.
* Add basic split-image implementation for 1D/2D/3D grouped convolution
This is a working baseline implementation that splits large spatial
dimensions to handle memory constraints.
Implementation:
- 1D: W-split for NWGC layout (36/36 tests passing)
- 2D: H-split for NHWGC layout (20/20 tests passing)
- 3D: D-split for NDHWGC layout (verified working)
Features:
- Binary split of outermost spatial dimension
- Sequential LEFT/RIGHT kernel launches
- Proper padding adjustment at split boundaries
- Offset calculation for pointer arithmetic
- Debug output for verification
Threshold: 100KB (configurable in transformer)
Known limitations:
- No safety checks for edge cases (to be added)
- Offset calculated before Split-N (incompatible with N>1, to be fixed)
- No recursive splitting for very large tensors
Next steps:
- Add safety checks (is_possible_to_split_*)
- Move offset calculation to transformer (after Split-N)
- Test with N>1 + split-image combination
* Refactor split-image to unified structure for 1D/2D/3D
Unified the three separate dimension-specific blocks into a single
common implementation with dimension-specific stride calculations.
Benefits:
- Reduced code from 636 → 348 lines (45% reduction)
- Eliminated code duplication
- Easier to maintain and extend
- Single source of truth for split logic
Implementation:
- Common: Binary split, offset calc, padding adjustment, kernel launch
- Dimension-specific: Stride calculation only
- 1D: stride = G * C
- 2D: stride = W_in * G * C
- 3D: stride = H_in * W_in * G * C
Test results (all passing):
- 1D: 36/36 tests ✅
- 2D: 20/20 tests ✅
- 3D: 28/28 tests ✅
- Total: 84/84 (100%)
All test scenarios verified:
- Varying channels, padding, stride, dilation
- Filter sizes (1x1 pointwise to 7x7)
- Multiple groups (G=1,2,4)
- Odd dimensions
- Complex combinations
* Add safety checks for split-image in all dimensions
Added is_possible_to_split safety checks to prevent crashes when
splitting is not feasible.
Safety checks verify:
1. Output dimension > 1 (can't split single element)
2. RIGHT piece starts after left padding
3. LEFT piece ends within input bounds
If checks fail, falls back to normal kernel launch.
Verified for all dimensions:
- 1D (W-split): Wo=1 case triggers fallback
- 2D (H-split): Ho=1 case triggers fallback
- 3D (D-split): Do=1 case triggers fallback
Original 84 tests still pass - they use normal configurations
that naturally satisfy safety conditions.
Safety checks protect against pathological edge cases with:
- Very small spatial dimensions
- Extreme stride/dilation combinations
- Invalid padding configurations
* Fix Split-N + Split-Image compatibility issue
Fixed critical bug where Split-N and Split-Image working together
caused ~50% incorrect results due to wrong batch stride calculation.
Problem:
- Batch stride was calculated using MODIFIED spatial dimensions
(e.g., W=50000 after split) instead of ORIGINAL dimensions (W=100000)
- Spatial offset was applied globally in invoker, not per-batch in kernel
- Each batch (blockIdx.z) got wrong memory offset
Solution:
1. Store spatial offset in kargs (don't apply to pointer in invoker)
2. Copy correct batch_stride from temp_kargs to left/right kargs
3. Apply formula in operator(): ptr = base + (batch × stride) + spatial_offset
Changes:
- grouped_convolution_forward_kernel.hpp:
* Added spatial_offset_in/out fields to KernelArgs
* Apply batch + spatial offset in operator()
- grouped_convolution_forward_invoker.hpp:
* Keep base pointer, store spatial offset in kargs
* Copy batch_stride from temp_kargs (has original dimensions)
- transform_conv_fwd_to_gemm.hpp:
* Add debug output for split-image calculation
Results:
- N=1 tests: 84/84 passing (100%)
- N>1 tests: Now all passing (previously ~50% errors)
- Tested: 1D, 2D, 3D with N=1,2,4,8,16,20
* Implement unified threshold for Split-N and Split-Image
This commit consolidates threshold management for both Split-N and
Split-Image operations into a single source of truth, eliminating
code duplication and fixing offset calculation issues.
Key Changes:
============
1. Transformer (transform_conv_fwd_to_gemm.hpp):
- Moved TwoGB constant to public section for unified access
- CalculateSplitImage() now takes no parameters
- Uses internal threshold: TwoGB / sizeof(CDataType)
- Calculates offsets using N_ (after Split-N) for correctness
2. Kernel (grouped_convolution_forward_kernel.hpp):
- GetSplitImageInfo() simplified to take no parameters
- Forwards to transformer's CalculateSplitImage()
- Clean interface with unified threshold internally
3. Invoker (grouped_convolution_forward_invoker.hpp):
- Removed redundant threshold calculation
- Simplified to call kargs.GetSplitImageInfo() with no params
- Clean early-return pattern (no unnecessary else blocks)
- Removed duplicate/dead code paths
Benefits:
=========
- Single source of truth: TwoGB defined once in transformer
- No parameter passing for threshold between components
- Correct offset calculation using N_ (post-Split-N)
- Cleaner code with no duplication
- All tests passing: 1D/2D/3D with various N values
Testing:
========
- Split-Image only (N=1, large spatial): PASS
- Split-N only (N>1, small spatial): PASS
- Both splits active (N>1, large spatial): PASS
- No splits (N=1, small spatial): PASS
- CPU verification correct for all scenarios
* Comment out outdated split-image code (SplitConvProblem/LaunchKernelWithSplitIfNeeded)
The old recursive queue-based implementation has been replaced by the
new CalculateSplitImage() method which is simpler and correctly handles
Split-N + Split-Image interaction.
Changes:
- Wrapped lines 381-1078 in #if 0...#endif
- Old methods: SplitConvProblem() and LaunchKernelWithSplitIfNeeded()
- Preserved for reference but disabled from compilation
- No functional changes - all tests still pass
The new implementation (CalculateSplitImage at line ~2163) provides:
- Correct offset calculation using N_ (after Split-N)
- Simpler binary split logic
- Better integration with unified threshold approach
* Implement recursive split-image with depth limit (MAX_DEPTH=10)
Changes:
- Add depth tracking to SplitPiece struct
- Implement two stopping conditions:
1. Piece size below threshold (optimal case)
2. Depth >= MAX_DEPTH (prevents infinite recursion)
- Remove MAX_PIECES limit in favor of depth-based control
- Support up to 2^10 = 1024 pieces with depth 10
This allows handling extreme tensor sizes while ensuring termination.
Pieces larger than threshold will still launch correctly if depth limit reached.
Tested with H=100 (4 levels), H=2000 (6 levels), H=4000 (9 levels) - all pass CPU verification.
* Summary of recursive split-image implementation:
- Recursive queue-based splitting with depth limit (MAX_DEPTH=10, up to 1024 pieces)
- Two stopping conditions: size below threshold OR max depth reached
- Cumulative offset tracking through all recursion levels
- LEFT piece inherits parent offset, RIGHT accumulates (parent + local)
- Per-batch spatial offset application in kernel operator()
- Batch stride uses original dimensions (before split)
- Works with Split-N: split-N first, then recursive split-image
- Handles odd dimensions, padding, stride, dilation correctly
- All 1D/2D/3D tests pass with CPU verification
* Add comment explaining MAX_DEPTH capacity for 2GB threshold
* Refactor: move recursive split-image logic to transformer
- Move LaunchWithRecursiveSplit() from invoker to transform_conv_fwd_to_gemm.hpp
- Simplify invoker from ~250 lines to ~140 lines (removed 110 lines of inline logic)
- Encapsulate SplitPiece struct and BFS splitting algorithm in transformer
- Remove unused includes (queue, vector) from invoker
- Add documentation comment for AreDescriptorsSmallerThan2GB()
- Improve code organization and reusability
- No performance overhead (static template function, compiler inlines)
- All tests passing with 2GB production threshold
* Apply clang-format-18 formatting
- Format invoker and transformer files with clang-format-18
- Fix brace placement and alignment
- No functional changes
* Fix clang-format-18 issues in forward kernel
- Remove extra blank lines
- Fix line wrapping for template calls
- Consolidate GetSplitImageInfo() to single line
* Update include/ck_tile/ops/grouped_convolution/utils/transform_conv_fwd_to_gemm.hpp
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
* Update include/ck_tile/ops/grouped_convolution/utils/transform_conv_fwd_to_gemm.hpp
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
* Update include/ck_tile/ops/grouped_convolution/kernel/grouped_convolution_forward_kernel.hpp
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
* Update include/ck_tile/ops/grouped_convolution/kernel/grouped_convolution_forward_kernel.hpp
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
* Split-Image implementation with temporary fixed divider
- Implemented spatial dimension splitting (Split-Image) for large tensors
- Added piece-based coordinate transformation for 1D/2D/3D convolutions
- Integrated Split-N (batch splitting) with automatic threshold detection
- Fixed M dimension calculation to include batch: M = N × spatial_size
- Added spatial offset support in kernel arguments
- Verified 20/20 test cases passing for Split-Image alone
- Known issue: Split-N + Split-Image combination needs coordinate fix
Implementation Details:
- Split factors: 4 (1D), 4×4 (2D), 4×4×4 (3D) - temporary fixed values
- Batch strides properly calculated for NWGC/NHWGC/NDHWGC layouts
- Piece descriptors track spatial boundaries and block ranges
- No performance overhead for N=1 cases
* Fix 1D split-image padding issue with per-piece dimensions
- Store actual size per piece to handle non-uniform splits
- Remove dead code from transform utils
* Fix 2D/3D split-image with independent split factors per dimension
Problem: Single split factor caused non-uniform pieces when dimensions
didn't divide evenly. Result: 18/25 (72%) 2D padding combinations failed.
Solution: Independent split factor selection for W, H, D dimensions.
Each dimension gets optimal factor based on its own size.
Test Results:
- 1D: 42/42 pass (100%)
- 2D: 25/25 pass (100%)
- Total: 67/67 combinations verified
* Remove unused split-image struct fields
Cleanup of split-image implementation:
- Removed unused piece_d, piece_h, piece_w fields from SplitImageInfo struct
- These fields were declared but never used in the kernel
- Per-piece dimensions are already stored in pieces[] array
- Reduces struct size and improves code clarity
Tested: 1D/2D/3D convolutions with split-image, padding, stride all pass
* Refactor split-image invoker code for improved readability
- Extract piece calculation logic into calculate_piece lambda helper
- Extract kernel args population into populate_split_image_kargs lambda
- Use aggregate initialization for cleaner struct population
- Reduce nesting depth and improve maintainability
- Fix outdated comment about split-image implementation status
* Refactor split-image code and remove debug prints
- Extract GPU kernel helper lambdas for better readability
- Remove all split-image debug print statements
- Set memory threshold to 2GB for production
- All tests pass with CPU verification
* Add split-image safety constraints and refactor to utils
- Add MAX_TOTAL_PIECES=64 limit to prevent segfault
- Move calculate_spatial_piece to library utils
- Add layout validation (NWGC, NHWGC, NDHWGC only)
- Fix hierarchical splitting to respect piece limits
- Add proper documentation and formatting
* Change split-image from runtime to compile-time branching
Response to @bartekxk review comment:
Convert 'if(kargs.num_spatial_pieces > 1)' to 'if constexpr(EnableSplitImage)'
Changes:
- Add EnableSplitImage template parameter to kernel
- Change runtime if to compile-time if constexpr
- Update invoker to instantiate kernel variants with true/false
Benefits:
- Eliminates runtime branching in GPU kernel
- Dead code elimination (each variant is smaller)
- Better compiler optimization
Files modified: 2
Lines changed: 20 total (6 in kernel, 14 in invoker)
Tests: 27/27 passed (100%)
Performance: No regression
* Add split-image example as separate binary
- Create grouped_convolution_forward_split_image example
- Add grouped_convolution_forward_split_image_invoker.hpp
- Update CMakeLists.txt to build split_image binary
* Replace linear search with binary search in find_piece_id
- Change O(n) to O(log n) for finding piece ownership
- Matches reference implementation in large_tensor_cshuffle
* Simplify split-image code and fix integer overflow
- Extract lambda functions to static helper methods
- Pre-calculate constants in invoker
- Fix integer overflow in tensor size calculation for large tensors
* Trigger CI rerun - fix merge conflicts
* Fix merge conflict markers
* Fix clang-format: remove space before {}
* Fix clang-format: comment wrapping and Swish constructor
* Rename split_image to large_tensor for clarity
- Renamed grouped_convolution_forward_split_image.cpp -> grouped_convolution_forward_large_tensor.cpp
- Renamed grouped_convolution_forward_split_image_invoker.hpp -> grouped_convolution_forward_large_tensor_invoker.hpp
- Updated CMakeLists.txt target name: tile_example_grouped_conv_fwd_split_image -> tile_example_grouped_conv_fwd_large_tensor
- Updated comments to refer to 'large tensor' instead of 'split-image'
* Update comments and include in large_tensor example
- Updated header comments to use 'large tensor' terminology
- Fixed include path to use large_tensor_invoker.hpp
* Remove test code, restore 2GB threshold
* Update include/ck_tile/ops/grouped_convolution/utils/transform_conv_fwd_to_gemm.hpp
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
* Fix build errors after develop merge and complete rename to large_tensor
This commit addresses compilation errors from the develop merge and
completes the rename from split_image to large_tensor.
Changes:
1. Fix CDEElementWise typo in grouped_convolution_forward_invoker.hpp
2. Fix template parameter order in large_tensor_invoker.hpp
- TransformConvFwdToGemm signature changed in develop
- NumGroupsToMerge and SplitN parameters swapped positions
3. Fix missing template parameter in GroupedConvFwdHostArgs
4. Fix EpiloguePipeline scope in kernel (merge conflict)
5. Update binary name references in test scripts
* Restore 2GB threshold for split-image
Changed threshold from 100MB (testing) back to 2GB for production use.
* Fix const-correctness in ds_ptr cast
* Update include/ck_tile/ops/grouped_convolution/kernel/grouped_convolution_forward_kernel.hpp
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
* Apply clang-format-18
* update c++ 18 format
* Apply clang-format-18 to transform_conv_fwd_to_gemm.hpp
---------
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
* add tensorwise quant in grouped gemm
* fix example issue
* update test cases
* format codes
* clang format
* use GTEST_FAIL
* add bquant to grouped_gemm
* add tensorwise quant in grouped gemm
* fix example issue
* update test cases
* format codes
* clang format
* use GTEST_FAIL
* fix a bug in test_grouped_gemm_util
* skip test when use wmma on grouped_quant kernel
* change cmake
* fix a bug in test_grouped_gemm_util
* skip test when use wmma on grouped_quant kernel
* change cmake
* tests(quant_grouped_gemm): add unit tests to cover bquant in grouped_gemm
* Update test/ck_tile/grouped_gemm_quant/test_grouped_gemm_util_quant.hpp
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
* Update example/ck_tile/17_grouped_gemm/quant_grouped_gemm.hpp
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
* feat: add bf8 support
* chore: remove unnecessary decltype usage
* chore: add default quant_mode to function signature as fallback
* fix: pass correct runtime pipeline params in grouped_gemm bquant kernel
Calculate has_hot_loop, num_loop, and tail_number on device side for each
GEMM problem instead of using default values. This fixes incorrect results
when different problems in the group have different K dimensions.
* chore: set default quant mode in function signature
* test: add additional test cases to cover edge case of no hotloop
* change code based on comments
* WIP: bquant preshuffle b compiles but gives numerical error
* feat(grouped_gemm_quant): bquant with preshuffleB support added to grouped_gemm example & kernel
* refactor: refactor code after merge commit
* chore: remove print statements
* test(grouped_gemm): split test cases by quant mode to reduce compilation time and add bquant-preshuffleB mode test cases
---------
Co-authored-by: kyle-256 <Kyle.Zhao@amd.com>
Co-authored-by: ThomasNing <thomas.ning@amd.com>
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
Fixes LDS bank conflicts on gfx950 for universal gemm v3 pipeline
Replaces hardcoded LDS layer calculations with dynamic computation using the new architecture helpers
Adds architecture-specific helper function get_n_lds_banks()
Changes function attributes from CK_TILE_HOST_DEVICE to CK_TILE_DEVICE in universal gemm policy
BlockFmhaPipelineQRKSVS reuses LDS for K and dropout so there must be
block_sync_lds between loading k_lds_window by gemm_0 and storing
dropout randval.
* HasHotLoop is a constexpr
* Remove an unused function
* Remove some unused include statements
* Add implementation and tests for fp8 x bf8 weight preshuffle GEMM
* Add implementation and tests for fp8 x bf8 in CK Tile basic and universal GEMMs
* Remove two barrier calls that HotLoopScheduler already calls
* No need to suppress a variable that hasn't been declared
* Replace six arg_parser arguments with constexpr literals
* Simplify run_gemm_test_prec_type
* The strides don't need to be passed via arg_parser as we use their default values
* The layouts don't need to be passed as arguments twice
* Pass M N and K as regular arguments, not using the argument parser
* We can now remove the argument parser
* Add a common file for precision types to be used in testing
* Convert basic and universal GEMM tests to use gtest
* Make GemmConfig a test parameter, and form test cases as the cartesian product GemmConfigs x PrecTypes
* Add GemmConfigComputeV4 to the GEMM configs to run the universal tests on
* Added a changelog entry
* Add missing copyright statements
* ifndef-define-endif is not needed with pragma once
* Fix a comment
* Add F8 x BF8 tests for CompV4 in test_gemm_pipeline_kernel_types.hpp
* Disable the unreliable test MoeSortingCase4
---------
Co-authored-by: Adam Osewski <19374865+aosewski@users.noreply.github.com>
* Improve the grouped conv kernel name generation in CK Tile.
* Fix building CShuffle epilogue tests.
---------
Co-authored-by: Bartłomiej Kocot <barkocot@amd.com>
* 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.
* Fix compilation of the grouped conv examples.
* Fix grouped conv bwd weight example output in CK Tile.
* Add number of groups to merge to ck tile grouped gemm example.
* Initial set of tests for TransformConvBwdWeightToGemm.
* Added unit tests for TransformConvBwdWeightToGemm conv groups are merged.
* WIP: Tensor transformations.
* Add unit tests for coordinate transforms.
* Fully working conv group merging for TransformConvBwdWeightToGemm.
* WIP: Merged conv groups offset calculation.
* Adde unit tests for tensor view.
* WIP: Merged conv groups epilogue.
* Enable running multiple conv groups per batch.
* Add tests for tile_distribution_encoding.
* Change example to match optimally depthwise convolution with merged groups.
* Add more tests for tensor view.
* Integration test for reading diagonal blocks from grouped distributed tensor.
* Improved integration test.
* Improve test for accessing diagonal blocks.
* Added integration test for cshuffle epilogue LDS tile distribution.
* Add more logging.
* Increase the max number of reported errors.
* WIP: merged conv groups GEMM epilogue changes.
* LDS to global memory copy.
* Fix tile window size for c block.
* Integration test for CShuffle epilogue.
* Improved CShuffle test.
* WIP: Separate epilogue for merged conv groups.
* Tile example parameters changes to match depthwise conv.
* Offset fixes.
* Epilogue fixes.
* Working baseline for depthwise covolution with merged conv groups.
* Fix build.
* Initial unit tests for tensor descriptor.
* Add one more unit test for tensor view.
* WIP: LDS to global mem transfer using CK tile tensor descriptor and tile distribution encoding.
* Fully functional LDS to global mem transfer using tensor descriptor and tile distribution encoding.
* Add more comments, disable debug code.
* Remove debug and other dead code.
* Code clean-up for bwd tensor transformations.
* Enable running multiple GEMM batches of merged conv groups.
* Add compile check for assumed row-mjor layout.
* Fix strides in 1D conv to gemm transformation.
* WIP: Simplify conv to gemm transformations and handle K > 1 and C > 1 cases.
* Fix case k > 1 and c=1.
* Remove debug code.
* Make MPerGroup and NPerGroup template parameters.
* Add additional check for non-supported c > 1 case.
* WIP: Put back the generic tensor descriptors for convolutions.
* Fix tensor descriptors.
* Remove the obsolete template parameters.
* Add more instances.
* Fix bugs in merged conv groups tensor descriptors.
* Fix tensor descriptors for merged conv groups when K > 1.
* Remove debug output.
* Remove dead code.
* Fix merge conflicts.
* Code clean-up.
* Remove unused code.
* Run clang-formatting.
* Remove debug prints and obsolete tests.
* Check that number of convolution groups is multiple of merged groups.
* Fix build after removing obsolete functionality.
* Remove obsolete enumeration.
* Fix new unit projects.
* Remove unnecessary includes.
* Fix passing the number of merged groups.
* Remove unrelated tests.
* Fix IsSupportedArgument for bwd weight conv kernel.
* Fix clang formatting.
* Fix the bwd weight conv to gemm mapping for num merged groups > 1.
* GEMM config for conv group merging.
* Fix clang-formatting.
* Remove obsolete comment.
* Fix typos in comment strings.
* Increase the max number of reported errors when testing against reference implementation.
* Rename gemm_config to conv_config.
* Rename GemmConfig to ConvConfig and move NumGroupsToMerge into ConvConfig.
* Change num_groups_to_merge to a boolean flag in the ck tile grouped conv example.
* Run clang-format.
* Add number of merged groups into kernel name string.
* Remove group merging flag from CK Tile grouped conv example.
* Add indexing support to pooling operator
- Add IndexDataType template parameter to pooling problem and kernel
definitions
- Enable pooling kernel to output indices of selected elements during
max/absmax pooling
- Add overloaded operators for Max and AbsMax that track when values
change using bool changed parameter
- Support optional index buffer allocation and management in device
memory
- Modify BlockReduce2d classes to handle index tensors alongside value
tensors
- Add separate shared memory allocation for index data in cross-warp
reductions
- Create validate_pool_indices function to verify index correctness
- Modify pool3d.cpp example to demonstrate index output functionality
- Add tests for index output
* fixes
* Refactor BlockReduce2D functions to get rid auxiliary private types.
* comment resolutions and some changes to block_reduce2d
- index reference implementation improved
- reduce_operator.hpp cleanedup
- updated the block_reduce2d.hpp to have index calculation for
BlockReduce2dLinearCrossWarpSync as well
* conditionally used variable declaration improvement
- the conditionally used vairbales are used only when indexing is
enabled. To inform the compiler that they may be unused and declare them
with least size possible. This may allow it to be optimized compared to
the previous declarations
* comment resolutions
* lexical ordering of the indicies
- introduced accumulate methods that handle the intermediate steps if
needed to order the indexes
* add reduce_operator_accumulate.hpp to core.hpp
---------
Co-authored-by: Adam Osewski <Adam.Osewski@amd.com>
* [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>
* update test cases
* format codes
* use GTEST_FAIL
* add bquant to grouped_gemm
* fix a bug in test_grouped_gemm_util
* skip test when use wmma on grouped_quant kernel
* add tensorwise quant in grouped gemm
* fix example issue
* update test cases
* format codes
* fix a bug in test_grouped_gemm_util
* tests(quant_grouped_gemm): add unit tests to cover bquant in grouped_gemm
* Update test/ck_tile/grouped_gemm_quant/test_grouped_gemm_util_quant.hpp
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
* Update example/ck_tile/17_grouped_gemm/quant_grouped_gemm.hpp
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
* feat: add bf8 support
* chore: remove unnecessary decltype usage
* chore: add default quant_mode to function signature as fallback
* fix: pass correct runtime pipeline params in grouped_gemm bquant kernel
Calculate has_hot_loop, num_loop, and tail_number on device side for each
GEMM problem instead of using default values. This fixes incorrect results
when different problems in the group have different K dimensions.
* chore: set default quant mode in function signature
* test: add additional test cases to cover edge case of no hotloop
* chore: clang formatting
---------
Co-authored-by: kyle-256 <Kyle.Zhao@amd.com>
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
* Implement argument passing to element-wise functions for fwd convolution
* Add files for fwd + bias + clamp example
* Implement Bias
* Implement Clamp
* Elementwise function composition
* Composition unit test
* Implement fwd + bias + clamp example
* Simplify argument passing and composition
* elfunc -> bias_and_clamp
* Rename function to specify example
* Move element-wise function instantiation to kernel
* Make bias a runtime tensor
* No ugly namespace aliasing
* Initialize element-wise function on host
* Remove function initialization helper, simplify Compose initialization
* Remove unintended LSP compatibility patch
* Clean up includes and unused code
* Switch names in cshuffle epilogue
* Move CDElementwise to conv traits
* Re-add required include
* Initialize bias in same way as other tensors
* Better type specification for ds pointer
* Disable 1D convolution
* Add warning for non-group-constant bias
* Persistent Stream-K Kernel Implementation
This change implements an operator() function in the
reboot::StreamKKernel class that is enabled when the Persistent flag is
set to true. In this case, the data-parallel portion and the Stream-K
portion of the kernel are fully persistent.
The changes were made in the reboot namespace. A future PR will remove
the old Stream-K kernel class and remove the reboot namespace.
* Unit Tests for Persistent Stream-K Kernel
This change contains the inital test suite for the Persitent Stream-K
Kernel. The files contain "reboot" in the name; a future PR will remove
tests for the old Stream-K Kernel and remove the "reboot" naming.
A future commit will add tests for the non-persistent kernel.
Also added estimate_num_wgs_per_tile to the StreamKTilePartitionerBase
class. This allows us to estimate the number of accumulations done per
macro tile in C to use during validation when computing relative and
absolute tolerance.
* Adding implementation for the Non-Persistent Stream-K kernel
This code is adding the operator() function for the Non-Persistent Stream-K
kernel. Persistency of the kernel is determined through a template argument.
The Non-Persistent kernel will allocate additional workgroups for the data
parallel section, leading to a different structure for processing the data
parallel and Stream-K sections.
There has been an addition to the TilePartitioner to get access to the whether
Persistent has been set to true or false in the StreamKKernel.
* Adding in the tests for the Non-Persistent Stream-K kernel
* Refactor Stream-K Reboot Unit Tests
This commit makes the following changes:
- Update test cases to determine M, N, and K based on the number of CUs.
This ensures that each test case is one of Edge Case, SK Only, DP
Only, or DP + 2 Tile SK regardless of the architecture.
- Since the DP + 2 Tile SK test case takes long to run, this change
moves this case into a separate .inc file and labels it as an extended
test.
- Since the extended test takes > 30 seconds to run, this test is added
to the list of regression tests.
* Fix spelling errors in comments for test cases
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
* Changes based on review
Removed const volatile for typenames
Set up alias for is_tuple_t
Naming changes for clarity: GemmCommon -> BaseGemm
Moved std::enable_if_t out of template parameters and changed to a return type for operator()
Added constructor for StreamKKernelArgs to clarify UniversalGemm inheritance
---------
Co-authored-by: Emily Martins <emily.martins@amd.com>
Co-authored-by: Christopher Millette <63608002+cgmillette@users.noreply.github.com>
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
* Add name member to unary elementwise ops.
* Update elementwise_op_name to check for name attribute.
* Require that the layout is derived from BaseTensorLayout struct.
* add tensorwise quant in grouped gemm
* fix example issue
* update test cases
* format codes
* clang format
* use GTEST_FAIL
* fix a bug in test_grouped_gemm_util
* skip test when use wmma on grouped_quant kernel
* change cmake
* change code based on comments
---------
Co-authored-by: ThomasNing <thomas.ning@amd.com>
* Introduce tree reduction for BlockReduce2dCrossWarpSync
* Rename original impl to BlockReduce2dLinearCrossWarpSync
* Replace warp_size with get_warp_size()
---------
Co-authored-by: Illia Silin <98187287+illsilin@users.noreply.github.com>
Prior to this change, the number of accumulations passed into
calculate_rtol_atol was 1. That said, in most cases, this is not correct
when there are multiple workgroups contributing to the same macro tile
in C.
This change ensures uses the function estimate_num_wgs_per_tile, which
was extracted into a common file and generalized, to estimate the number
of workgroups per macro tile. This estimate is passed into
calculate_rtol_atol to ensure we get a better relative and absolute
tolerance.
The following changes were made
- Renamed iter to iter_start
- Renamed tile_iter to tile_iter_start
- Moved documentation from member variables to getters
- Removed double underscore from extra_iters_before_me variable
- Defined parent header in impl file
- Removed unused inlcudes
There are 2 derived structs based on whether Stream-K is persistent or not.
If it's persistent that means that both the data parallel and Stream-K sections
are data parallel. If it's non-persistent that means that only the
Stream-K section is persistent, while the data parallel section will have
separate workgroups allocated for it. Both structs will have a template
argument for Persistent.
The 2 derived classes will inherit common variables and functions from the
Stream-K TilePartitioner base class. There are additional variables for the
differing data parallel sections that will be added to each derived class,
that are in charge of the indexing/bookkeeping for the data parallel sections.
The only additional function that will differ between the 2 structs is GridSize(),
as the non-persistent will allocate extra workgroups for data parallel.
Unit tests for the derived structs are included.
To better align with the original Stream-K paper, this change implements
a new Stream-K tile partitioner base class. This class will handle the
Stream-K setup that is common to both a persistent and non-persistent DP
section. A later change will implement derived classes to handle the
differences between persistent and non-persistent DP.
This change also includes unit tests for the base tile partitioner.
