CUTLASS 3.6.0 (#1850)

* v3.6

* update changelog

* update readme

* fix typo

* fixing typos

* hopper gemm with weight prefetch

---------

Co-authored-by: yuzhai <yuzhai@nvidia.com>
Co-authored-by: Haicheng Wu <haichengw@nvidia.com>

README.md

@@ -1,8 +1,8 @@
 ![ALT](./media/images/gemm-hierarchy-with-epilogue-no-labels.png "Complete CUDA GEMM decomposition")
 
-# CUTLASS 3.5.1
+# CUTLASS 3.6.0
 
-_CUTLASS 3.5.1 - July 2024_
+_CUTLASS 3.6.0 - October 2024_
 
 CUTLASS is a collection of CUDA C++ template abstractions for implementing
 high-performance matrix-matrix multiplication (GEMM) and related computations at all levels 
@@ -42,48 +42,26 @@ and improves code composability and readability. More documentation specific to
 In addition to GEMMs, CUTLASS implements high-performance convolution via the implicit GEMM algorithm. Implicit GEMM is the formulation of a convolution operation as a GEMM thereby taking advantage of CUTLASS's modular GEMM pipeline. This allows CUTLASS to build convolutions by reusing highly-optimized GEMM components.
 
 
-# What's New in CUTLASS 3.5
+# What's New in CUTLASS 3.6
 
-CUTLASS 3.5.1 is an update to CUTLASS adding:
+CUTLASS 3.6.0 is an update to CUTLASS adding:
 
-- [Minimal SM90 WGMMA + TMA GEMM example in 100 lines of code](./examples/cute/tutorial/wgmma_sm90.cu).
-- [Exposure of L2 `cache_hint`s in TMA copy atoms](./include/cute/arch/copy_sm90_tma.hpp#L48)
-- Exposure of raster order and tile swizzle extent in [CUTLASS library profiler](./media/docs/profiler.md#GEMM), and
-[example 48](./examples/48_hopper_warp_specialized_gemm/48_hopper_warp_specialized_gemm.cu).
-- [TMA store based and EVT supported epilogues](./include/cutlass/epilogue/collective/sm90_epilogue_array_tma_warpspecialized.hpp) for [Hopper pointer array batched kernels](./test/unit/gemm/device/sm90_gemm_f16_f16_f16_tensor_op_f32_ptr_array.cu).
-- A new [`GemmSparseUniversal` API for CUTLASS 2.x Ampere kernels](./include/cutlass/gemm/device/gemm_sparse_universal.h) to enable serial and parallel split-k for sparse tensor cores and new tiny tile sizes to better support LLM inference.
-- [CUDA host adapter](./include/cutlass/cuda_host_adapter.hpp) extensions to support TMA descriptor construction driver APIs.
-- Inclusion of more [Hopper fprop, dgrad, and wgrad convolution kernels in CUTLASS library and profiler](./python/cutlass_library/generator.py).
-- Support for residual add (beta != 0) in convolution kernels.
-- A new convolution [epilogue](./examples/16_ampere_tensorop_conv2dfprop/ampere_tensorop_conv2dfprop.cu#L269) for CUTLASS 2.x to support non-packed NHWC output.
-- A refactor of [include files throughout CUTLASS core directories](./include/cutlass/gemm/collective/collective_mma_decl.hpp) to reduce circular dependencies and [tests to guard against them](./test/self_contained_includes/CMakeLists.txt).
-- [A guide for setting up VSCode to work well with CUTLASS](./media/docs/ide_setup.md) and [expanded code style guide](./media/docs/programming_guidelines.md).
-- Better support for MSVC as a host compiler.
-- Many performance optimizations, improvements, and bug fixes including fixes for FlashAttention-2.
-- Optimal code generation with CUDA toolkit versions 12.4 and 12.5u1.
-- NOTICE:
-  + Upcoming CUTLASS 3.6 release will include a breaking refactor to the CUTLASS 3.x convolution `kernel::ConvUniversal` API to bring it in line with `gemm::GemmUniversal`. After this, the 3.x convolution API will no longer be considered as a beta API.
-  + Upcoming CUTLASS 3.6 release will include a breaking refactor to the Hopper TMA pointer array batched epilogue in order to support grouped GEMMs.
-
-CUTLASS 3.5.0 is an update to CUTLASS adding:
-
-- Implicit GEMM Convolutions targeting Hopper SM90A via WGMMA + [TMA im2col](./include/cute/atom/copy_traits_sm90_im2col.hpp).
-  + Native implementation in CUTLASS 3.x using CuTe, mirroring the [same design hierarchy as that of GEMMs](./media/docs/gemm_api_3x.md).
-  + Support for 1D, 2D, and 3D convolutions in a [rank-agnostic fashion](./include/cutlass/conv/convnd_problem_shape.hpp).
-  + Support for [Fprop](./test/unit/conv/device_3x/fprop/sm90_conv3d_fprop_implicit_gemm_s8_s8_s32_tensorop_s32.cu), [Dgrad](./test/unit/conv/device_3x/dgrad/sm90_conv2d_dgrad_implicit_gemm_f16_f16_f32_tensorop_f16.cu), and [Wgrad](./test/unit/conv/device_3x/wgrad/sm90_conv1d_wgrad_implicit_gemm_f16_f16_f32_tensorop_f16.cu) algorithms.
-  + [CUTLASS profiler support](./python/cutlass_library/conv3x_emitter.py) for 2D and 3D convolutions implemented via the 3.x API.
-  + NOTE: this is a beta release. Further updates to CUTLASS will include major performance improvements, feature enablement, and possible breaking changes to the API until 3.7 release. Your feedback is welcome on the design!
-- Support for [Ada (SM89) FP8 tensor cores via the 2.x API](./examples/58_ada_fp8_gemm/ada_fp8_gemm.cu). Requires CUDA 12.4 or newer.
-- [Ampere gather/scatter convolution example](./examples/59_ampere_gather_scatter_gemm/README.md) in CuTe and CUTLASS 3.x.
-  + Showcasing how custom kernels can be written and optimized using CUTLASS 3.x and CuTe and the general strategy for implementing convolutions as specializations of GETTs.
-  + Implementation of a coarse grained sparse gather/scatter kernel achieving peak performance on Ampere class tensor cores.
-- 32x and 16x tile sizes are added to CUTLASS 2.x to improve the performance of narrow-tall and wide-short matrices.
-- Updates to CuTe documentation for [`cute::Tensor<>`](./media/docs/cute/03_tensor.md), [MMA atoms](./media/docs/cute/0t_mma_atom.md), and an overhauled [CuTe GEMM tutorial series](./examples/cute/tutorial).
-- Extensions to CuTe to support [L2 prefetching](./include/cute/algorithm/prefetch.hpp) and [TMA store+reductions](./include/cute/arch/copy_sm90_tma.hpp#L1337).
-- Remove C++11 requirement on a few CUTLASS 2.x API header files. All CUTLASS files now require C++17.
-- Fixes to greatly reduce build warnings.
-- Updates and bugfixes from the community (thanks!)
-- CUTLASS 3.5.1 is a minor update to CUTLASS containing small bug fixes and improvements, including fixes for FlashAttention-2 builds.
+- [Hopper structured sparse GEMM](./examples/62_hopper_sparse_gemm/62_hopper_sparse_gemm.cu).
+  + [FP16](./test/unit/gemm/device/sm90_sparse_gemm_f16_f16_f32_tensor_op_f32.cu)
+  + [FP8](./test/unit/gemm/device/sm90_sparse_gemm_f8_f8_f32_tensor_op_f32.cu)
+  + [INT8](./test/unit/gemm/device/sm90_sparse_gemm_s8_s8_s32_tensor_op_s32.cu)
+  + [TF32](./test/unit/gemm/device/sm90_sparse_gemm_tf32_tf32_f32_tensor_op_f32.cu)
+- A refactor to the CUTLASS 3.x convolution `kernel::ConvUniversal` [API](./include/cutlass/conv/kernel/sm90_implicit_gemm_tma_warpspecialized.hpp) to bring it in line with `gemm::GemmUniversal`. Now the 3.x convolution API is no longer considered as a beta API.
+- [An improved mixed input GEMM](./examples/55_hopper_mixed_dtype_gemm/README.md) and a [lookup table implementation](./examples/55_hopper_mixed_dtype_gemm/55_hopper_int4_fp8_gemm.cu) for `INT4`x`FP8` scale-only mode.
+- [EVT nodes for Top-K selection and softmax](./include/cutlass/epilogue/fusion/sm90_visitor_topk_softmax.hpp) and [GEMM example using those](./examples/61_hopper_gemm_with_topk_and_softmax/61_hopper_gemm_with_topk_and_softmax.cu).
+- [Programmatic Dependent Launch](./include/cutlass/arch/grid_dependency_control.h) (PDL) that leverages a new Hopper feature to speedup two back-to-back kernels, and its corresponding [documentations](./media/docs/dependent_kernel_launch.md).
+- [A new debugging tool, synclog](./include/cutlass/arch/synclog.hpp), for dumping out all synchronization events from within a kernel to a file. Please see [synclog documentation](./media/docs/utilities.md#debugging-asynchronous-kernels-with-cutlasss-built-in-synclog-tool) for details.
+- A new TMA-enabled [epilogue](./include/cutlass/epilogue/collective/sm90_epilogue_array_tma_warpspecialized.hpp) for grouped GEMM that brings significant performance improvement, as well as its EVT support.
+- A SIMT-enabled pointer-array [epilogue](./include/cutlass/epilogue/collective/sm70_epilogue_vectorized_array.hpp).
+- A new [Ping-Pong kernel schedule for Grouped GEMM](./include/cutlass/gemm/kernel/sm90_gemm_array_tma_warpspecialized_pingpong.hpp) and some other optimizations.
+- [A new instantiation strategy for CUTLASS profiler kernels](./python/cutlass_library/sm90_shapes.py) along with [improved documentation for instantiation level in CUTLASS profiler](./media/docs/profiler.md#instantiating-more-kernels-with-hopper).
+- A new hardware support for comparisons and computations of [`cutlass::bfloat16_t`](./include/cutlass/bfloat16.h)
+- Fixed use of isnan on Windows for [`half_t`](./test/unit/core/functional.cu).
 
 Minimum requirements:
 
@@ -163,7 +141,7 @@ CUTLASS runs successfully on the following NVIDIA GPUs, and it is expected to be
 
 In general, PTX code generated for one target architecture can be run on future architectures (i.e., it is forward compatible).  However, CUDA 12.0 introduced the concept of "architecture-accelerated features" whose PTX does not have forward compatibility guarantees. Several Hopper PTX instructions fall under this category of architecture-accelerated features, and thus require a `sm_90a` target architecture (note the "a" appended). For more details on this and other architecture-accelerated instructions, please refer to the [CUDA Documentation](https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#feature-availability).
 
-The target architecture information is passed on to CUTLASS via the cmake flag `CUTLASS_NVCC_ARCHS`. In order to maximize performance on Hopper GH100, users are required to build CUTLASS with `90a` as the target architecture. If a user accidentally builds a kernel which uses SM90a features (e.g. Hopper Tensor Core Instructions), using the SM90 target (note the lack of "a"), with either CTK 12 or 11.8, the kernel is expected to fail with a runtime error.
+The target architecture information is passed on to CUTLASS via the cmake flag `CUTLASS_NVCC_ARCHS`. In order to maximize performance on Hopper GH100, users are required to build CUTLASS with `90a` as the target architecture. If a user accidentally builds a kernel which uses SM90a features (e.g. Hopper Tensor Core Instructions), using the SM90 target (note the lack of "a"), with either CUDA Toolkit 12 or 11.8, the kernel is expected to fail with a runtime error.
 
 ```
 cmake .. -DCUTLASS_NVCC_ARCHS="90a" 
@@ -191,6 +169,8 @@ CUTLASS is described in the following documents and the accompanying
 - [Tile Iterators](./media/docs/tile_iterator_concept.md) - describes C++ concepts for iterating over tiles of matrices in memory
 - [CUTLASS Profiler](./media/docs/profiler.md) - command-line driven profiling application
 - [CUTLASS Utilities](./media/docs/utilities.md) - additional templates used to facilate rapid development
+- [Dependent kernel launch](./media/docs/dependent_kernel_launch.md) - describes a new feature in Hopper which allows overlapping dependent 
+kernels in the same stream, and how it is used in CUTLASS.
 
 # Resources
 We have also described the structure of an efficient GEMM in our talk at the