mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-19 22:39:03 +00:00

Files

Po Yen Chen 40d2ed0f2a [CK_TILE] Share partition index across threads and specify offset in load_tile()/async_load_tile()/load_tile_transpose() (#2905 )

* Allow sharing partition index across threads

* Fix typo PartitoinIndex -> PartitionIndex

* Remove C++20 'requires' usages

* Add missing template arguments

* Fix load_tile() overload ambiguity issue

* Use SFINAE to exclude invalid arguments

* Add additional offset parameter to the async_load_tile()

* Remove async_load_tile() default argument to avoid ambiguity

* Extract tile_window coordinate compute logic as method

* Use warp-shared LDS base address in tile_window::async_load()

* Add constraint to tile_window::load() templates

* Fix wrong type traits is_class_v<> usages

* Add missing constraint to async_load_tile()

* Add missing tile_window::load() overload

* Add more constraint to avoid load_tile() call ambiguity

* Rename ParitionIndex as ReplacementPartitionIndex

* Update pre_computed_warp_coords_ in move_extended()

* Fix inconsistency between template parameters and documentation

* Allow specifying pre-computed parition index

* Add type straits is_sequence<> & is_tile_distribution<>

* Add type straits is_tensor_view<>

* Add type constraints to make_tile_window() templates

* Allow passing partition_index to set_tile_if()

* Allow specifying partition_index to store_tile()

* Add missing template parameter of replace_bottom_tensor_view()

* Allow passing partition_index to Default2DEpilogue

* Make get_partition_index() public

* Add _with_offset() postfix to avoid resolution error

* Remove ReplacementPartitionIndex template param

* Add missing comments

* Add load_tile_transpose_with_offset() overload

2025-11-12 10:26:14 +08:00

algorithm

[CK_Tile] Merge multiple convolution groups into a single GEMM batch (#2986 )

2025-10-29 16:49:28 +02:00

arch

[CK_BUILDER] Add grouped conv fwd ck tile traits (#3183 )

2025-11-11 13:55:33 -08:00

container

[CK_TILE] Share partition index across threads and specify offset in load_tile()/async_load_tile()/load_tile_transpose() (#2905 )

2025-11-12 10:26:14 +08:00

numeric

[CK_TILE] Support WMMA (gfx12) in FMHA (#2528 )

2025-10-29 13:31:08 -07:00

tensor

[CK_TILE] Share partition index across threads and specify offset in load_tile()/async_load_tile()/load_tile_transpose() (#2905 )

2025-11-12 10:26:14 +08:00

utility

Fix compilation errors with clang22. (#3164 )

2025-11-05 15:42:22 -08:00

config.hpp

[CK_TILE] Support WMMA (gfx12) in FMHA (#2528 )

2025-10-29 13:31:08 -07:00

README.md

introducing ck_tile! (#1216 )

2024-04-15 19:27:12 -05:00

README.md

ck_tile/core

ck_tile/core contains every basic functions and structures to create a GPU kernel using ck_tile. User should only include ck_tile/core.hpp this single header to use all the functionality. Everything is under ck_tile namespace. The coding style under this folder should be similar to std (snake_case for structure/function, Camel for template types...)

algorithm/
    coordinate transform and some other reusable algorithm
arch/
    contains some basic device building block like mma, buffer addressing, etc...
container/
    contains basic container data structure, array/sequence/tuple/...
numeric/
    data type, and data type related math
tensor/
    tensor descriptors and tile level API
utility/
    other utility function for both host/device