dee185d80c
* Change splitk_batch_offset parameter to k_size in UniversalGemmKernel::MakeGemmTensorViews function Prior to this change, the splitk_batch_offset parameter of MakeGemmTensorViews had type SplitKBatchOffset. But, the only member variable of the SplitKBatchOffset class used in the MakeGemmTensorViews function was splitted_k (an int32_t). The splitted_k value was used as part of defining the dimensions of the tensor view. That said, for Stream K, we do not need to use the SplitKBatchOffset class since we are not using Split K. Thus, this commit changes the splitk_batch_offset parameter to a int32_t called k_size. This will avoid the constraint of requiring a caller of MakeGemmTensorViews to use the SplitKBatchOffset class while still providing the same functionality. Calls to UniversalGemmKernel::MakeGemmTensorViews have been updated accordingly. * StreamK Kernel RunGemm Implementation Stream K cannot simply use UniversalGemmKernel's RunGemm for the following reasons: 1. The UniversalGemmKernel::RunGemm function computes num_loop based on a static function of the TilePartitioner. That said, for Stream K, num_loop must be computed using a member function (namely GetCurrentIterLength from PR #2708). 2. The UniversalGemmKernel::RunGemm function requires the use of a SplitKBatchOffset object which is not used for Stream K since we are not using Split K. Thus, this change adds a RunGemm function in the StreamKKernel class. * initial implementation for operator() for StreamKKernel: adding stream-k algorithm and calls to RunGemm * Fix indexing and offset issues for StreamK These changes do the following: - Ensure offsets along the M and N dimensions are multiplied by MPerblock or NPerBlock, respectively. This ensures tile window origins are at the correct locations. - Fix bug in the tile partitioner's GetTileIdxWithOffset. Now, we apply divmod to the given references to ensure correct values are available to the caller. - Added documentation in the Stream-K operator() * Initial gtests for Stream-K These changes add an initial gtest suite for the CK Tile Stream-K kernel. Currently, due to bugs in the StreamKTilePartitioner (which will be handled in a future PR), there are validation issues for certain cases which may differ on different architectures. Thus, we opted to run cases that are only fully data-parallel (skipping others). A guard was added to Stream-K's IsSupportedArgument method to ensure that callers are aware of this constraint. Additionally, to ensure testing reproducibility, options for setting the number of CUs and occupancy were added to MakeKernelArgs. * Use GemmPipeline operator() variant that takes hot loop and tail num In Stream-K, the num_loop value varies per WG and per iteration of a Stream-K loop. So instead, we use the version of the GemmPipeline's operator() function that takes in has_hot_loop and tail_num. This is similar to what is done in Grouped GEMM. * changes from review: comments, move readfirstlane, remove ifndef * Switch direction of C tensor traversal & add padding guard Prior to this change, WGs travelled backwards through their assigned macro tiles in the C tensor. For instance, if WG0 is responsible for C tiles 0 and 1, it would first visit tile 1 then tile 0. This means that the iter_end decrements in each iteration of the stream-K while loop. Since we are working with unsigned integers, the subtraction operation may not be safe. Thus, this change makes is such that WGs travel forward so that their iter_start is incremented and their iter_end remains fixed. Additionally, we added a guard against WGs that are neither sk_blocks nor dp_blocks to ensure such WGs do not participate in the GEMM. Together, these changes make is such that the algorithm is correct when sk_blocks is greater than zero. * Disable StreamK_M256_N256_K256_SKBlocks12 test case This instance involves >=3 WGs contributing to each macro tile in C. Due to the use of atomics, this is resulting in precision errors. These errors will not persist once the reduction strategy is implemented. We will re-enable this test then. --------- Co-authored-by: Astha Rai <astha.rai713@gmail.com>
Composable Kernel Tile
concept
ck_tile provides a programming model with templated abstractions to enable users to implement performance-critical kernels for machine learning workloads. introduces following basic concepts to help users building your own operator
- tensor coordinate transformation, this is the core concept of layout/index transform abstraction in both compiler time and run time.
- tile-based programming model, including tile-level api and the concept of distributed tensor.
ck_tile is independently from the old ck, located under /include/ck_tile. You don't need to include anything from old CK, ck_tile has similiar (indeed almost the same) implementations for users to build operators. We will have a transition period to pull everything from old ck into ck_tile, stay tuned.
component
ck_tile is splitted into several componenets including core, host, ops/gemm, ops/fmha... each component you only need to include a single header (e.g #include "ck_tile/core.hpp", #include "ck_tile/ops/fmha.hpp") then you are able to use the function/structure inside (different from old ck)
[core]
ck_tile/core contains all the basic data structure and function to build the kernel, you can only include this header and build your own operators that utilizing all the basic building blocks introduced in ck.
core/container
- array, store runtime variables with fixed length (tensor index, register buffer, etc...)
- tuple, same as std::tuple, hold different type of data, and one of the solution to achieve multiple buffer.
- sequence, compile time integer sequence used to build various internal structures, or to describe tile size
- other convenient structure build on top of above 3
core/numeric
- gpu data type like
fp16_t,bf16_t,fp8_t... and the conversion between each other - constexpr integer similiar to std::integral_constant to be used as compile time integer.
- math functions and numeric utilities
core/algorithm
- coordinate transformation system, used to build tensor transform and compile time indexing. This is the core idea introduced in old
ckto describe how a tensor is build by several basic transform primitives likemerge/unmerge/embedetc... and how we indexing into a ND tensor that finally mapped to 1D memory offset.
core/tensor
- tensor descriptor, to describe how a ND tensor
- distributed tensor, describe the storage of this tensor, and the distribution of how a collection of threads collaborately work for this tensor.
- tile level API, including
load_tile,store_tile,shuffle_tile,slice_tile, etc...
[host]
ck_tile/host contains all the host side utilities to launch a kernel, create the device buffer, and some reference implementations. This can be used to create examples (like that under ck_tile example folder) and simple executable to invoke this kernel, so if you only need ck_tile to build your own device library then it's OK to not include this. Based on this, it is recommended to include the specific header you needed under this folder to avoid including unwanted headers (e.g, only include ck_tile/host/kernel_launch.hpp), unless you are writing a host executable.
[ops/gemm, ops/fmha, ops/reduce...]
our implementation of different device operators.
- warp, warp tile level operator
- block, block tile level operator
- pipeline, pipeline that can achieve a customized tile level mainloop (or epilogue). By switching different pipeline to the kernel template you can have different kind of pipeline optimizations.
- kernel, template interface for users to instantiate a particular kernel
[ops/epilogue]
epilogue part of our kernel. We may extend this epilogue part to let users to build their own cutomized epilogues.
[ref]
reference implementation of cpu or gpu. This folder is supposed to include a specific header on demand.
examples
currently we put all ck_tile related example under /example/ck_tile folder. Please check each example's subfolder.