Doc updates for 3.2.2

CHANGELOG.md

@@ -1,5 +1,8 @@
 # NVIDIA CUTLASS Changelog
 
+## [3.2.2](https://github.com/NVIDIA/cutlass/releases/tag/v3.2.2) (2023-10-25)
+* Fixes illegal memory access issue [1138](https://github.com/NVIDIA/cutlass/issues/1138) hit by FlashAttention tests in PyTorch.
+
 ## [3.2.1](https://github.com/NVIDIA/cutlass/releases/tag/v3.2.1) (2023-09-22)
 * Python support SM90 Epilogue Visitor Tree (EVT) on top of the C++ support released in 3.2.0.
 * SM80 EVT support in C++ and Python.

CMakeLists.txt

@@ -40,7 +40,7 @@ endif()
 message(STATUS "CMake Version: ${CMAKE_VERSION}")
 set(IMPLICIT_CMAKE_CXX_STANDARD OFF CACHE BOOL "Do not explicitly specify -std=c++11 if set")
 
-project(CUTLASS VERSION 3.2.1 LANGUAGES CXX)
+project(CUTLASS VERSION 3.2.2 LANGUAGES CXX)
 include(${CMAKE_CURRENT_SOURCE_DIR}/CUDA.cmake)
 
 if (CUDA_VERSION VERSION_LESS 11.3)

README.md

@@ -62,6 +62,9 @@ CUTLASS 3.2.1 is an update to CUTLASS adding:
 - SM90 rasterization direction support in the CUTLASS profiler.
 - Improvement for CUTLASS profiler build times.
 
+CUTLASS 3.2.2 is a minor update to CUTLASS adding:
+- Bug fix for illegal memory access issue hit by Flash Attention tests in PyTorch. See [1138](https://github.com/NVIDIA/cutlass/issues/1138) for details.
+
 Minimum requirements:
 
 - Architecture: Volta

include/cute/pointer.hpp

@@ -93,8 +93,6 @@ struct device_ptr
 {
   using value_type = T;
   
-  static const uint32_t ElementsPerStoredItem = sizeof(T) * 8 / sizeof_bits_v<T>;
-
   CUTE_HOST_DEVICE constexpr
   device_ptr(T* ptr) : ptr_(ptr) {}
 
@@ -113,7 +111,7 @@ struct device_ptr
 
   template <class Index>
   CUTE_HOST_DEVICE constexpr
-  DerivedType operator+(Index const& i) const { return {ptr_ + i / ElementsPerStoredItem}; }
+  DerivedType operator+(Index const& i) const { return {ptr_ + i}; }
 
   CUTE_HOST_DEVICE constexpr friend
   ptrdiff_t operator-(device_ptr<T,DerivedType> const& a,