Merge pull request tensorflow#3269 from davidzchen/cuda-configure

Add cuda_configure repository rule to autodetect cuda.

Author: rmlarsen

.gitignore

@@ -9,13 +9,6 @@ node_modules
 /bazel-testlogs
 /bazel-tf
 /tensorflow/contrib/cmake/build
-/third_party/gpus/cuda/bin
-/third_party/gpus/cuda/cuda.config
-/third_party/gpus/cuda/extras
-/third_party/gpus/cuda/include
-/third_party/gpus/cuda/lib
-/third_party/gpus/cuda/lib64
-/third_party/gpus/cuda/nvvm
 /third_party/py/numpy/numpy_include
 /tools/bazel.rc
 /tools/python_bin_path.sh
@@ -24,4 +17,5 @@ node_modules
 /pip_test
 /_python_build
 *.pyc
-__pycache__
+__pycache__
+*.swp

configure

@@ -80,6 +80,7 @@ while [ "$TF_NEED_CUDA" == "" ]; do
   esac
 done
 
+export TF_NEED_CUDA
 if [ "$TF_NEED_CUDA" == "0" ]; then
   echo "Configuration finished"
   exit
@@ -97,6 +98,7 @@ while true; do
     fi
   fi
   if [ -e "$GCC_HOST_COMPILER_PATH" ]; then
+    export CC=$GCC_HOST_COMPILER_PATH
     break
   fi
   echo "Invalid gcc path. ${GCC_HOST_COMPILER_PATH} cannot be found" 1>&2
@@ -107,7 +109,6 @@ while true; do
   # Retry
 done
 
-
 # Find out where the CUDA toolkit is installed
 OSNAME=`uname -s`
 
@@ -140,6 +141,8 @@ while true; do
   fi
 
   if [ -e "${CUDA_TOOLKIT_PATH}/${CUDA_RT_LIB_PATH}" ]; then
+    export CUDA_TOOLKIT_PATH
+    export CUDA_VERSION=$TF_CUDA_VERSION
     break
   fi
   echo "Invalid path to CUDA $TF_CUDA_VERSION toolkit. ${CUDA_TOOLKIT_PATH}/${CUDA_RT_LIB_PATH} cannot be found"
@@ -200,13 +203,16 @@ while true; do
   fi
 
   if [ -e "$CUDNN_INSTALL_PATH/${CUDA_DNN_LIB_ALT_PATH}" -o -e "$CUDNN_INSTALL_PATH/${CUDA_DNN_LIB_PATH}" ]; then
+    export CUDNN_VERSION=$TF_CUDNN_VERSION
+    export CUDNN_INSTALL_PATH
     break
   fi
 
   if [ "$OSNAME" == "Linux" ]; then
     CUDNN_PATH_FROM_LDCONFIG="$(ldconfig -p | sed -n 's/.*libcudnn.so .* => \(.*\)/\1/p')"
     if [ -e "${CUDNN_PATH_FROM_LDCONFIG}${TF_CUDNN_EXT}" ]; then
-      CUDNN_INSTALL_PATH="$(dirname ${CUDNN_PATH_FROM_LDCONFIG})"
+      export CUDNN_VERSION=$TF_CUDNN_VERSION
+      export CUDNN_INSTALL_PATH="$(dirname ${CUDNN_PATH_FROM_LDCONFIG})"
       break
     fi
   fi
@@ -225,42 +231,11 @@ while true; do
   CUDNN_INSTALL_PATH=""
 done
 
-cat > third_party/gpus/cuda/cuda.config <<EOF
-# CUDA_TOOLKIT_PATH refers to the CUDA toolkit.
-CUDA_TOOLKIT_PATH="$CUDA_TOOLKIT_PATH"
-# CUDNN_INSTALL_PATH refers to the cuDNN toolkit. The cuDNN header and library
-# files can be either in this directory, or under include/ and lib64/
-# directories separately.
-CUDNN_INSTALL_PATH="$CUDNN_INSTALL_PATH"
-
-# The Cuda SDK version that should be used in this build (empty to use libcudart.so symlink)
-TF_CUDA_VERSION=$TF_CUDA_VERSION
-
-# The Cudnn version that should be used in this build
-TF_CUDNN_VERSION=$TF_CUDNN_VERSION
-EOF
-
-# Configure the gcc host compiler to use
-export WARNING=$DO_NOT_SUBMIT_WARNING
-perl -pi -e "s,CPU_COMPILER = \('.*'\),# \$ENV{WARNING}\nCPU_COMPILER = ('$GCC_HOST_COMPILER_PATH'),s" third_party/gpus/crosstool/clang/bin/crosstool_wrapper_driver_is_not_gcc
-perl -pi -e "s,GCC_HOST_COMPILER_PATH = \('.*'\),# \$ENV{WARNING}\nGCC_HOST_COMPILER_PATH = ('$GCC_HOST_COMPILER_PATH'),s" third_party/gpus/crosstool/clang/bin/crosstool_wrapper_driver_is_not_gcc
-
-# Configure the platform name.
-perl -pi -e "s,PLATFORM = \".*\",PLATFORM = \"$OSNAME\",s" third_party/gpus/cuda/platform.bzl
-
-# Configure the Cuda toolkit version to work with.
-perl -pi -e "s,(GetCudaVersion.*return )\"[0-9\.]*\",\1\"$TF_CUDA_VERSION\",s" tensorflow/stream_executor/dso_loader.cc
-perl -pi -e "s,CUDA_VERSION = \"[0-9\.]*\",CUDA_VERSION = \"$TF_CUDA_VERSION\",s" third_party/gpus/cuda/platform.bzl
-
-# Configure the Cudnn version to work with.
-perl -pi -e "s,(GetCudnnVersion.*return )\"[0-9\.]*\",\1\"$TF_CUDNN_VERSION\",s" tensorflow/stream_executor/dso_loader.cc
-perl -pi -e "s,CUDNN_VERSION = \"[0-9\.]*\",CUDNN_VERSION = \"$TF_CUDNN_VERSION\",s" third_party/gpus/cuda/platform.bzl
-
-
 # Configure the compute capabilities that TensorFlow builds for.
 # Since Cuda toolkit is not backward-compatible, this is not guaranteed to work.
 while true; do
   fromuser=""
+  default_cuda_compute_capabilities="3.5,5.2"
   if [ -z "$TF_CUDA_COMPUTE_CAPABILITIES" ]; then
 cat << EOF
 Please specify a list of comma-separated Cuda compute capabilities you want to build with.
@@ -270,6 +245,9 @@ EOF
     read -p "[Default is: \"3.5,5.2\"]: " TF_CUDA_COMPUTE_CAPABILITIES
     fromuser=1
   fi
+  if [ -z "$TF_CUDA_COMPUTE_CAPABILITIES" ]; then
+    TF_CUDA_COMPUTE_CAPABILITIES=$default_cuda_compute_capabilities
+  fi
   # Check whether all capabilities from the input is valid
   COMPUTE_CAPABILITIES=${TF_CUDA_COMPUTE_CAPABILITIES//,/ }
   ALL_VALID=1
@@ -285,34 +263,13 @@ EOF
       exit 1
     fi
   else
+    export CUDA_COMPUTE_CAPABILITIES=$TF_CUDA_COMPUTE_CAPABILITIES
     break
   fi
   TF_CUDA_COMPUTE_CAPABILITIES=""
 done
 
-if [ ! -z "$TF_CUDA_COMPUTE_CAPABILITIES" ]; then
-  export WARNING=$DO_NOT_SUBMIT_WARNING
-  function CudaGenCodeOpts() {
-    OUTPUT=""
-    for CAPABILITY in $@; do
-      OUTPUT=${OUTPUT}"   \"${CAPABILITY}\",     "
-    done
-    echo $OUTPUT
-  }
-  export CUDA_GEN_CODES_OPTS=$(CudaGenCodeOpts ${TF_CUDA_COMPUTE_CAPABILITIES//,/ })
-  perl -pi -0 -e 's,\n( *)([^\n]*supported_cuda_compute_capabilities\s*=\s*\[).*?(\]),\n\1# $ENV{WARNING}\n\1\2$ENV{CUDA_GEN_CODES_OPTS}\3,s' third_party/gpus/crosstool/clang/bin/crosstool_wrapper_driver_is_not_gcc
-  function CudaVersionOpts() {
-    OUTPUT=""
-    for CAPABILITY in $@; do
-      OUTPUT=$OUTPUT"CudaVersion(\"${CAPABILITY}\"), "
-    done
-    echo $OUTPUT
-  }
-  export CUDA_VERSION_OPTS=$(CudaVersionOpts ${TF_CUDA_COMPUTE_CAPABILITIES//,/ })
-  perl -pi -0 -e 's,\n( *)([^\n]*supported_cuda_compute_capabilities\s*=\s*\{).*?(\}),\n\1// $ENV{WARNING}\n\1\2$ENV{CUDA_VERSION_OPTS}\3,s' tensorflow/core/common_runtime/gpu/gpu_device.cc
-fi
-
-# Invoke the cuda_config.sh and set up the TensorFlow's canonical view of the Cuda libraries
-(cd third_party/gpus/cuda; ./cuda_config.sh;) || exit -1
+bazel clean --expunge
+bazel fetch //...
 
 echo "Configuration finished"

tensorflow/core/common_runtime/gpu/gpu_device.cc

@@ -785,10 +785,9 @@ struct CudaVersion {
   int minor_part = -1;
 };
 
-// "configure" uses the specific name to substitute the following string.
-// If you change it, make sure you modify "configure" as well.
 std::vector<CudaVersion> supported_cuda_compute_capabilities = {
-    CudaVersion("3.5"), CudaVersion("5.2")};
+  TF_CUDA_CAPABILITIES,
+};
 
 std::vector<CudaVersion> GetSupportedCudaComputeCapabilities() {
   auto cuda_caps = supported_cuda_compute_capabilities;

tensorflow/core/kernels/lrn_op.cc

@@ -31,7 +31,7 @@ limitations under the License.
 #endif
 
 #if GOOGLE_CUDA
-#include "third_party/gpus/cuda/include/cuda.h"
+#include "cuda/include/cuda.h"
 #include "tensorflow/core/platform/stream_executor.h"
 #include "tensorflow/core/util/stream_executor_util.h"
 #endif  // GOOGLE_CUDA

tensorflow/core/kernels/matmul_op.cc

@@ -25,7 +25,7 @@ limitations under the License.
 #include "tensorflow/core/kernels/fill_functor.h"
 
 #if GOOGLE_CUDA
-#include "third_party/gpus/cuda/include/cuda.h"
+#include "cuda/include/cuda.h"
 #include "tensorflow/core/platform/stream_executor.h"
 #endif  // GOOGLE_CUDA
 

tensorflow/core/platform/default/build_config/BUILD

@@ -9,7 +9,7 @@ exports_files(["LICENSE"])
 
 load("//tensorflow:tensorflow.bzl", "tf_copts")
 load("//tensorflow:tensorflow.bzl", "tf_cuda_library")
-load("//third_party/gpus/cuda:platform.bzl", "cuda_library_path")
+load("@local_config_cuda//cuda:platform.bzl", "cuda_library_path")
 
 cc_library(
     name = "gtest",
@@ -32,7 +32,7 @@ tf_cuda_library(
     deps = [
         "//tensorflow/stream_executor",
     ] + select({
-        "//third_party/gpus/cuda:darwin": ["IOKit"],
+        "@local_config_cuda//cuda:darwin": ["IOKit"],
         "//conditions:default": [],
     }),
 )
@@ -91,20 +91,20 @@ filegroup(
 cc_library(
     name = "cuda",
     data = [
-        "//third_party/gpus/cuda:{}".format(cuda_library_path("cudart")),
+        "@local_config_cuda//cuda:{}".format(cuda_library_path("cudart")),
     ],
     linkopts = select({
-        "//third_party/gpus/cuda:darwin": [
-            "-Wl,-rpath,third_party/gpus/cuda/lib",
-            "-Wl,-rpath,third_party/gpus/cuda/extras/CUPTI/lib",
+        "@local_config_cuda//cuda:darwin": [
+            "-Wl,-rpath,../local_config_cuda/cuda/lib",
+            "-Wl,-rpath,../local_config_cuda/cuda/extras/CUPTI/lib",
         ],
         "//conditions:default": [
-            "-Wl,-rpath,third_party/gpus/cuda/lib64",
-            "-Wl,-rpath,third_party/gpus/cuda/extras/CUPTI/lib64",
+            "-Wl,-rpath,../local_config_cuda/cuda/lib64",
+            "-Wl,-rpath,../local_config_cuda/cuda/extras/CUPTI/lib64",
         ],
     }),
     deps = [
-        "//third_party/gpus/cuda:cudart",
+        "@local_config_cuda//cuda:cudart",
     ],
 )
 

tensorflow/core/platform/default/gpu/BUILD

@@ -15,9 +15,9 @@ tf_cuda_library(
     copts = tf_copts(),
     cuda_deps = [
         "//tensorflow/core:stream_executor",
-        "//third_party/gpus/cuda:cuda_headers",
-        "//third_party/gpus/cuda:cupti_headers",
+        "@local_config_cuda//cuda:cuda_headers",
+        "@local_config_cuda//cuda:cupti_headers",
     ],
-    data = ["//third_party/gpus/cuda:cupti_dsos"],
+    data = ["@local_config_cuda//cuda:cupti_dsos"],
     visibility = ["//visibility:public"],
 )

tensorflow/core/platform/default/gpu/cupti_wrapper.h

@@ -21,7 +21,7 @@ limitations under the License.
 #include <stddef.h>
 #include <stdint.h>
 
-#include "third_party/gpus/cuda/extras/CUPTI/include/cupti.h"
+#include "cuda/extras/CUPTI/include/cupti.h"
 
 namespace perftools {
 namespace gputools {

tensorflow/core/util/port.cc

@@ -16,7 +16,7 @@ limitations under the License.
 #include "tensorflow/core/util/port.h"
 
 #if GOOGLE_CUDA
-#include "third_party/gpus/cuda/include/cuda.h"
+#include "cuda/include/cuda.h"
 #endif
 
 namespace tensorflow {

tensorflow/stream_executor/BUILD

@@ -27,17 +27,18 @@ cc_library(
     ]),
     data = [
         "//tensorflow/core:cuda",
-        "//third_party/gpus/cuda:cublas",
-        "//third_party/gpus/cuda:cudnn",
-        "//third_party/gpus/cuda:cufft",
+        "@local_config_cuda//cuda:cublas",
+        "@local_config_cuda//cuda:cudnn",
+        "@local_config_cuda//cuda:cufft",
+        "@local_config_cuda//cuda:curand",
     ],
     linkopts = [
         "-ldl",
     ],
     visibility = ["//visibility:public"],
     deps = [
         "//tensorflow/core:lib",
-        "//third_party/gpus/cuda:cuda_headers",
+        "@local_config_cuda//cuda:cuda_headers",
     ],
     alwayslink = 1,
 )