This is an automated email from the ASF dual-hosted git repository.
lunderberg pushed a commit to branch unity
in repository https://gitbox.apache.org/repos/asf/tvm.git
The following commit(s) were added to refs/heads/unity by this push:
new 163c7ac436 [Unity] Cutlass kernel compatibility with cmake 3.18+
(#16302)
163c7ac436 is described below
commit 163c7ac43610c958b481e6f39ed97bcc309f8af9
Author: Eric Lunderberg <[email protected]>
AuthorDate: Tue Jan 2 15:26:47 2024 -0600
[Unity] Cutlass kernel compatibility with cmake 3.18+ (#16302)
The updates to the cutlass kernels made in TVM PR #16244 require
symbols provided in cuda 7.5+. While the cuda architecture is
specified by setting `NVCC_FLAGS` in the `CMakeLists.txt` for each
kernel, cmake 3.18+ also sets it based on the
`CMAKE_CUDA_ARCHITECTURES` value. If not set, cmake will explicitly
pass the compute capability as nvidia's default of 5.2, *EVEN IF* it
has already been specified in `NVCC_FLAGS`. Because the kernels
cannot compile with compute capability of 5.2, this causes
compilation errors.
By setting `CMAKE_CUDA_ARCHITECTURES` to `OFF`, cmake does not add
5.2 as a target architecture.
See https://cmake.org/cmake/help/latest/policy/CMP0104.html for
details on CMake's policy for CUDA architecture flags.
See https://cmake.org/cmake/help/latest/policy/CMP0104.html for the
default CUDA architecture for each version of CUDA.
---
cmake/modules/CUDA.cmake | 22 ++++++++++++++++++++++
1 file changed, 22 insertions(+)
diff --git a/cmake/modules/CUDA.cmake b/cmake/modules/CUDA.cmake
index ce561c66a6..03a79326c8 100644
--- a/cmake/modules/CUDA.cmake
+++ b/cmake/modules/CUDA.cmake
@@ -38,6 +38,28 @@ if(USE_CUDA)
list(APPEND TVM_RUNTIME_LINKER_LIBS ${CUDA_CUDA_LIBRARY})
list(APPEND TVM_RUNTIME_LINKER_LIBS ${CUDA_NVRTC_LIBRARY})
+ # Compatibility with cmake 3.18+
+ #
+ # The updates to the cutlass kernels made in TVM PR#16244 require
+ # symbols provided in cuda 7.5+. While the cuda architecture is
+ # specified by setting `NVCC_FLAGS` in the `CMakeLists.txt` for each
+ # kernel, cmake 3.18+ also sets it based on the
+ # `CMAKE_CUDA_ARCHITECTURES` value. If not set, cmake will explicitly
+ # pass the compute capability as nvidia's default of 5.2, *EVEN IF* it
+ # has already been specified in `NVCC_FLAGS`. Because the kernels
+ # cannot compile with compute capability of 5.2, this causes
+ # compilation errors.
+ #
+ # By setting `CMAKE_CUDA_ARCHITECTURES` to `OFF`, cmake does not add
+ # 5.2 as a target architecture.
+ #
+ # See https://cmake.org/cmake/help/latest/policy/CMP0104.html for
+ # details on CMake's policy for CUDA architecture flags.
+ #
+ # See https://cmake.org/cmake/help/latest/policy/CMP0104.html for the
+ # default CUDA architecture for each version of CUDA.
+ set(CMAKE_CUDA_ARCHITECTURES OFF)
+
if(USE_CUDNN)
message(STATUS "Build with cuDNN support")
include_directories(SYSTEM ${CUDA_CUDNN_INCLUDE_DIRS})
