fix a bug where Magma binaries were failing with on K80 GPUs. More co…

…ntext in pytorch/pytorch#29096
pytorch · Nov 4, 2019 · dd7edfd · dd7edfd
1 parent d5e3f7f
commit dd7edfd
Show file tree

Hide file tree

Showing 4 changed files with 38 additions and 14 deletions.
diff --git a/conda/README.md b/conda/README.md
@@ -41,14 +41,11 @@ export PYTORCH_BRANCH=master
 
 ```
 nvidia-docker run -it --ipc=host --rm -v $(pwd):/remote soumith/conda-cuda bash
-yum install -y yum-utils centos-release-scl
-yum-config-manager --enable rhel-server-rhscl-7-rpms
-yum install -y devtoolset-3-gcc devtoolset-3-gcc-c++ devtoolset-3-gcc-gfortran devtoolset-3-binutils
-export PATH=/opt/rh/devtoolset-3/root/usr/bin:$PATH
-export LD_LIBRARY_PATH=/opt/rh/devtoolset-3/root/usr/lib64:/opt/rh/devtoolset-3/root/usr/lib:$LD_LIBRARY_PATH
 git clone https://github.com/pytorch/builder
 cd builder/conda
 conda install -y conda-build
 . ./switch_cuda_version.sh 10.0
-conda build magma-cuda100-2.5.0
+conda build magma-cuda100-2.5.1
 ```
+
+Test magma builds on K80 machine, to check against the bug https://github.com/pytorch/pytorch/issues/29096
diff --git a/conda/magma-cuda100-2.5.1/cmakelists.patch b/conda/magma-cuda100-2.5.1/cmakelists.patch
@@ -1,5 +1,5 @@
---- CMakeLists.txt	2019-08-02 14:09:54.000000000 -0700
-+++ CMakeLists.txt	2019-08-02 18:00:11.818541038 -0700
+--- CMakeLists.txt	2019-11-04 15:04:50.452110612 -0800
++++ CMakeLists.txt	2019-11-04 15:05:39.889521970 -0800
 @@ -3,7 +3,7 @@
  # ----------------------------------------
  # to disable Fortran, set this to "off"
@@ -43,6 +43,15 @@
 
 
  # ----------------------------------------
+@@ -119,7 +121,7 @@
+     set( NV_SM    "" )
+     set( NV_COMP  "" )
+
+-    set(CUDA_SEPARABLE_COMPILATION ON)
++    set(CUDA_SEPARABLE_COMPILATION OFF)
+
+     # nvcc >= 6.5 supports -std=c++11, so propagate CXXFLAGS to NVCCFLAGS.
+     # Older nvcc didn't support -std=c++11, so previously we disabled propagation.
 @@ -259,11 +261,18 @@
          message( STATUS "    compile for CUDA arch 7.5 (Turing)" )
      endif()

diff --git a/conda/magma-cuda101-2.5.1/cmakelists.patch b/conda/magma-cuda101-2.5.1/cmakelists.patch
@@ -1,5 +1,5 @@
---- CMakeLists.txt	2019-08-02 14:09:54.000000000 -0700
-+++ CMakeLists.txt	2019-08-02 18:00:11.818541038 -0700
+--- CMakeLists.txt	2019-11-04 15:04:50.452110612 -0800
++++ CMakeLists.txt	2019-11-04 15:05:39.889521970 -0800
 @@ -3,7 +3,7 @@
  # ----------------------------------------
  # to disable Fortran, set this to "off"
@@ -43,13 +43,22 @@
 
 
  # ----------------------------------------
+@@ -119,7 +121,7 @@
+     set( NV_SM    "" )
+     set( NV_COMP  "" )
+
+-    set(CUDA_SEPARABLE_COMPILATION ON)
++    set(CUDA_SEPARABLE_COMPILATION OFF)
+
+     # nvcc >= 6.5 supports -std=c++11, so propagate CXXFLAGS to NVCCFLAGS.
+     # Older nvcc didn't support -std=c++11, so previously we disabled propagation.
 @@ -259,11 +261,18 @@
          message( STATUS "    compile for CUDA arch 7.5 (Turing)" )
      endif()
 
 +    if ( ${GPU_TARGET} MATCHES "All")
 +      set( MIN_ARCH 370)
-+      SET( NV_SM -gencode arch=compute_37,code=compute_37 -gencode arch=compute_37,code=sm_37 -gencode arch=compute_50,code=sm_50 -gencode arch=compute_60,code=sm_60 -gencode arch=compute_70,code=sm_70 )
++      SET( NV_SM -gencode arch=compute_37,code=sm_37 -gencode arch=compute_50,code=sm_50 -gencode arch=compute_60,code=sm_60 -gencode arch=compute_70,code=sm_70 )
 +      SET( NV_COMP "")
 +    endif()
 +    

diff --git a/conda/magma-cuda92-2.5.1/cmakelists.patch b/conda/magma-cuda92-2.5.1/cmakelists.patch
@@ -1,5 +1,5 @@
---- CMakeLists.txt	2019-08-02 14:09:54.000000000 -0700
-+++ CMakeLists.txt	2019-08-02 18:00:11.818541038 -0700
+--- CMakeLists.txt	2019-11-04 15:04:50.452110612 -0800
++++ CMakeLists.txt	2019-11-04 15:05:39.889521970 -0800
 @@ -3,7 +3,7 @@
  # ----------------------------------------
  # to disable Fortran, set this to "off"
@@ -43,13 +43,22 @@
 
 
  # ----------------------------------------
+@@ -119,7 +121,7 @@
+     set( NV_SM    "" )
+     set( NV_COMP  "" )
+
+-    set(CUDA_SEPARABLE_COMPILATION ON)
++    set(CUDA_SEPARABLE_COMPILATION OFF)
+
+     # nvcc >= 6.5 supports -std=c++11, so propagate CXXFLAGS to NVCCFLAGS.
+     # Older nvcc didn't support -std=c++11, so previously we disabled propagation.
 @@ -259,11 +261,18 @@
          message( STATUS "    compile for CUDA arch 7.5 (Turing)" )
      endif()
 
 +    if ( ${GPU_TARGET} MATCHES "All")
 +      set( MIN_ARCH 370)
-+      SET( NV_SM -gencode arch=compute_37,code=compute_37 -gencode arch=compute_37,code=sm_37 -gencode arch=compute_50,code=sm_50 -gencode arch=compute_60,code=sm_60 -gencode arch=compute_70,code=sm_70 )
++      SET( NV_SM -gencode arch=compute_37,code=sm_37 -gencode arch=compute_50,code=sm_50 -gencode arch=compute_60,code=sm_60 -gencode arch=compute_70,code=sm_70 )
 +      SET( NV_COMP "")
 +    endif()
 +