CUTLASS 3.6.0 (#1850)

* v3.6

* update changelog

* update readme

* fix typo

* fixing typos

* hopper gemm with weight prefetch

---------

Co-authored-by: yuzhai <[email protected]>
Co-authored-by: Haicheng Wu <[email protected]>

Author: 3 people

CHANGELOG.md

@@ -1,5 +1,25 @@
 # NVIDIA CUTLASS Changelog
 
+## [3.6.0](https://github.com/NVIDIA/cutlass/releases/tag/v3.6.0) (2024-10-03)
+
+- [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).
+  Various improvements and fixed from the community and CUTLASS team. Thanks to everyone who submitted PRs!
+
 ## [3.5.1](https://github.com/NVIDIA/cutlass/releases/tag/v3.5.1) (2024-07-25)
 
 - [Minimal SM90 WGMMA + TMA GEMM example in 100 lines of code](./examples/cute/tutorial/wgmma_sm90.cu)

CMakeLists.txt

@@ -134,7 +134,6 @@ set(CUTLASS_ENABLE_PERFORMANCE ${CUTLASS_ENABLE_PROFILER} CACHE BOOL "Enable CUT
 set(CUTLASS_ENABLE_TESTS ${CUTLASS_ENABLE_TESTS_INIT} CACHE BOOL "Enable CUTLASS Tests")
 set(CUTLASS_ENABLE_GTEST_UNIT_TESTS ${CUTLASS_ENABLE_TESTS} CACHE BOOL "Enable CUTLASS GTest-based Unit Tests")
 set(CUTLASS_USE_SYSTEM_GOOGLETEST OFF CACHE BOOL "Use system/external installation of GTest")
-
 set(CUTLASS_USE_PACKED_TUPLE ON CACHE BOOL "If ON, make cute::tuple be new standard-layout tuple type; if OFF, use the original cute::tuple implementation that is _not_ standard-layout.")
 if (CUTLASS_USE_PACKED_TUPLE)
   list(APPEND CUTLASS_CUDA_NVCC_FLAGS -DCUTE_USE_PACKED_TUPLE=1)
@@ -234,7 +233,6 @@ set(CUTLASS_NVCC_EMBED_CUBIN ON CACHE BOOL "Embed compiled CUDA kernel binaries
 set(CUTLASS_NVCC_EMBED_PTX ON CACHE BOOL "Embed compiled PTX into executables.")
 set(CUTLASS_NVCC_KEEP OFF CACHE BOOL "Keep intermediate files generated by NVCC.")
 set(CUTLASS_ENABLE_F16C OFF CACHE BOOL "Enable F16C x86 extensions in host code.")
-set(CUTLASS_ENABLE_DIRECT_CUDA_DRIVER_CALL OFF CACHE BOOL "Enable CUTLASS to directly call driver API.")
 
 ################################################################################
 #
@@ -271,6 +269,7 @@ set(CUTLASS_LIBRARY_OPERATIONS "all" CACHE STRING "Comma-delimited list of opera
 set(CUTLASS_LIBRARY_KERNELS ${CUTLASS_LIBRARY_KERNELS_INIT} CACHE STRING "Comma-delimited list of kernel name filters. If unspecified, only the largest tile size is enabled. If the string 'all' is specified, all kernels are enabled.")
 set(CUTLASS_LIBRARY_IGNORE_KERNELS "" CACHE STRING "Comma-delimited list of kernels to exclude from build. This option ONLY takes effect if CUTLASS_LIBRARY_KERNELS is set.")
 set(CUTLASS_LIBRARY_EXCLUDE_KERNELS "" CACHE STRING "Comma-delimited list of kernels to exclude from build. This option always takes effect, whether or not CUTLASS_LIBRARY_KERNELS is set. It also can exclude kernels from the filter file (see KERNEL_FILTER_FILE).")
+set(CUTLASS_LIBRARY_INSTANTIATION_LEVEL "" CACHE STRING "Instantiation level for SM90 kernels. Set to `max` and make sure CUTLASS_LIBRARY_KERNELS is non-empty to stamp all possible kernel configurations.")
 
 ################################################################################
 
@@ -318,6 +317,8 @@ if(CUTLASS_ENABLE_SM90_EXTENDED_MMA_SHAPES)
   list(APPEND CUTLASS_CUDA_NVCC_FLAGS -DCUTE_SM90_EXTENDED_MMA_SHAPES_ENABLED)
 endif()
 
+set(CUTLASS_SKIP_REDUCTION_INIT OFF CACHE BOOL "Disable init reduction workspace")
+
 #
 # NOTE: running with asan and CUDA requires the following environment variable:
 #
@@ -345,6 +346,10 @@ if(CUTLASS_NVCC_EMBED_PTX)
   list(APPEND CUTLASS_CUDA_CLANG_FLAGS --cuda-include-ptx=all)
 endif()
 
+if (CUTLASS_SKIP_REDUCTION_INIT)
+  list(APPEND CUTLASS_CUDA_NVCC_FLAGS -DCUTLASS_SKIP_REDUCTION_INIT=1)
+endif()
+
 if (CUTLASS_ENABLE_TENSOR_CORE_MMA)
   list(APPEND CUTLASS_CUDA_FLAGS -DCUTLASS_ENABLE_TENSOR_CORE_MMA=1)
 endif()
@@ -354,6 +359,18 @@ if (CUTLASS_PROFILER_DISABLE_REFERENCE)
   list(APPEND CUTLASS_CUDA_NVCC_FLAGS -DCUTLASS_PROFILER_DISABLE_REFERENCE=1)
 endif()
 
+if (CUTLASS_ENABLE_GDC_FOR_SM90)
+  message(STATUS "Grid Dependency Control (GDC) is enabled for SM90 kernels (required for programmatic dependent launches).")
+  list(APPEND CUTLASS_CUDA_NVCC_FLAGS -DCUTLASS_ENABLE_GDC_FOR_SM90=1)
+endif()
+
+set(CUTLASS_ENABLE_SYNCLOG OFF CACHE BOOL "Enable synchronization event logging for race condition debugging. WARNING: This redefines __syncthreads() and __syncwarp() in all downstream code!")
+
+if (CUTLASS_ENABLE_SYNCLOG)
+  set(CMAKE_CUDA_SEPARABLE_COMPILATION ON)
+  string(APPEND CMAKE_CXX_FLAGS " -DCUTLASS_ENABLE_SYNCLOG=1")
+  string(APPEND CMAKE_CUDA_FLAGS " -DCUTLASS_ENABLE_SYNCLOG=1")
+endif()
 
 
 
@@ -880,12 +897,27 @@ function(cutlass_add_executable_tests NAME TARGET)
 
   set(TEST_GROUP_NAME ${NAME})
 
+  # To run the tests from an install package with tests enabled, we need to generate test files
+  # that don't rely on the current directory structure in build.
+
+  set(TEST_NAME c${NAME})
+  set(TEST_GEN_DIR ${CMAKE_CURRENT_BINARY_DIR}/ctest/${TEST_NAME})
+  file(MAKE_DIRECTORY ${TEST_GEN_DIR})
+
+  set(TEST_EXE_PATH $<TARGET_FILE:${TARGET}>)
+  set(TEST_USE_EXTENDED_FORMAT ON)
+  configure_file("${CUTLASS_CTEST_TEMPLATE_FILE}" "${TEST_GEN_DIR}/CTestTestfile.${TEST_NAME}.cmake" @ONLY)
+
+  set(TEST_EXE_PATH $<TARGET_FILE_NAME:${TARGET}>)
+  set(TEST_USE_EXTENDED_FORMAT OFF) # ctest does not support extended add_test format.
+  configure_file("${CUTLASS_CTEST_TEMPLATE_FILE}" "${TEST_GEN_DIR}/CTestTestfile.${TEST_NAME}.install.cmake.in" @ONLY)
+
   foreach(CMD_OPTIONS_VAR IN LISTS __TEST_COMMAND_OPTIONS)
 
     if (CMD_COUNT GREATER 1)
-      string(TOLOWER "${NAME}_${CMD_OPTIONS_VAR}" TEST_NAME)
+      string(TOLOWER "${NAME}_${CMD_OPTIONS_VAR}" TESTCASE_NAME)
     else()
-      string(TOLOWER "${NAME}" TEST_NAME)
+      string(TOLOWER "${NAME}" TESTCASE_NAME)
     endif()
 
     # The following rigmarole is needed to deal with spaces and possible quotes in
@@ -899,42 +931,28 @@ function(cutlass_add_executable_tests NAME TARGET)
     separate_arguments(TEST_COMMAND_OPTIONS)
 
     add_custom_target(
-      ${TEST_NAME}
+      ${TESTCASE_NAME}
       COMMAND
       ${CUTLASS_TEST_EXECUTION_ENVIRONMENT} $<TARGET_FILE:${TARGET}> ${TEST_COMMAND_OPTIONS}
       DEPENDS
       ${TARGET}
       )
 
     if (CMD_COUNT GREATER 1)
-      add_dependencies(${NAME} ${TEST_NAME})
+      add_dependencies(${NAME} ${TESTCASE_NAME})
     endif()
 
     foreach(DEPENDEE ${__DEPENDEES})
-      add_dependencies(${DEPENDEE} ${TEST_NAME})
+      add_dependencies(${DEPENDEE} ${TESTCASE_NAME})
     endforeach()
 
-    set(TEST_NAME c${TEST_NAME})
+    set(TESTCASE_NAME c${TESTCASE_NAME})
     string(CONFIGURE "${_INLINE_PER_TEST_CODE_TEMPLATE}" _TEST_CODE @ONLY)
-    string(APPEND _INLINE_PER_TEST_CODE "${_TEST_CODE}")
+    file(APPEND "${TEST_GEN_DIR}/CTestTestfile.${TEST_NAME}.cmake" "${_TEST_CODE}")
+    file(APPEND "${TEST_GEN_DIR}/CTestTestfile.${TEST_NAME}.install.cmake.in" "${_TEST_CODE}")
 
   endforeach()
 
-  # To run the tests from an install package with tests enabled, we need to generate test files
-  # that don't rely on the current directory structure in build.
-
-  set(TEST_NAME c${NAME})
-  set(TEST_GEN_DIR ${CMAKE_CURRENT_BINARY_DIR}/ctest/${TEST_NAME})
-  file(MAKE_DIRECTORY ${TEST_GEN_DIR})
-
-  set(TEST_EXE_PATH $<TARGET_FILE:${TARGET}>)
-  set(TEST_USE_EXTENDED_FORMAT ON)
-  configure_file("${CUTLASS_CTEST_TEMPLATE_FILE}" "${TEST_GEN_DIR}/CTestTestfile.${TEST_NAME}.cmake" @ONLY)
-
-  set(TEST_EXE_PATH $<TARGET_FILE_NAME:${TARGET}>)
-  set(TEST_USE_EXTENDED_FORMAT OFF) # ctest does not support extended add_test format.
-  configure_file("${CUTLASS_CTEST_TEMPLATE_FILE}" "${TEST_GEN_DIR}/CTestTestfile.${TEST_NAME}.install.cmake.in" @ONLY)
-
   # The following line imports the tests for immediate run via `make test`.
 
   include(${TEST_GEN_DIR}/CTestTestfile.${TEST_NAME}.cmake)

PUBLICATIONS.md

@@ -26,6 +26,8 @@
 
 - ["Mixed Precision Post Training Quantization of Neural Networks with Sensitivity Guided Search"](https://arxiv.org/abs/2302.01382). Clemens JS Schaefer, Elfie Guo, Caitlin Stanton, Xiaofan Zhang, Tom Jablin, Navid Lambert-Shirzad, Jian Li, Chiachen Chou, Siddharth Joshi, Yu Emma Wang. _arXiv_, Feburary 2023.
 
+- ["Dynamic N:M Fine-Grained Structured Sparse Attention Mechanism"](https://dl.acm.org/doi/abs/10.1145/3572848.3577500). Zhaodong Chen, Zheng Qu, Yuying Quan, Liu Liu, Yufei Ding, Yuan Xie. _Proceedings of the 28th ACM SIGPLAN Annual Symposium on Principles and Practice of Parallel Programming_, Feburary 2023.
+
 - ["Stream-K: Work-centric Parallel Decomposition for Dense Matrix-Matrix Multiplication on the GPU"](https://arxiv.org/abs/2301.03598). Muhammad Osama, Duane Merrill, Cris Cecka, Michael Garland, John D. Owens. _arXiv_, January 2023.
 
 ## 2022