[GPU-Plugin] Support for building to specific GPU architectures (#2390)

* Support for builing gpu-plugins to specific GPU architectures
1. Option GPU_COMPUTE_VER exposed from both Makefile and CMakeLists.txt
2. updater_gpu documentation updated accordingly

* Re-introduced GPU_COMPUTE_VER option in the cmake flow.
This seems to fix the compile-time, rdc=true and copy-constructor related
errors seen and discussed in PR #2390.

CMakeLists.txt

@@ -3,6 +3,8 @@ project (xgboost)
 find_package(OpenMP)
 
 option(PLUGIN_UPDATER_GPU "Build GPU accelerated tree construction plugin")
+set(GPU_COMPUTE_VER 50;52;60;61 CACHE STRING
+    "Space separated list of compute versions to be built against")
 if(PLUGIN_UPDATER_GPU)
   find_package(CUDA REQUIRED)
 endif()
@@ -107,35 +109,29 @@ if(PLUGIN_UPDATER_GPU)
     set(NCCL_DIRECTORY ${PROJECT_SOURCE_DIR}/nccl)
     include_directories(${NCCL_DIRECTORY}/src)
     set(LINK_LIBRARIES ${LINK_LIBRARIES} ${CUDA_LIBRARIES})
-
+    #Find cub
-	#Find cub
+    set(CUB_DIRECTORY ${PROJECT_SOURCE_DIR}/cub/)
-	set(CUB_DIRECTORY ${PROJECT_SOURCE_DIR}/cub/)
+    include_directories(${CUB_DIRECTORY})
-	include_directories(${CUB_DIRECTORY})
-
     #Find googletest
     set(GTEST_DIRECTORY "${CACHE_PREFIX}" CACHE PATH "Googletest directory")
     include_directories(${GTEST_DIRECTORY}/include)
-
+    #gencode flags
-    # plugin
+    set(GENCODE_FLAGS "")
-    set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS};-lineinfo;--expt-extended-lambda")
+    foreach(ver ${GPU_COMPUTE_VER})
+      set(GENCODE_FLAGS "${GENCODE_FLAGS}-gencode arch=compute_${ver},code=sm_${ver};")
+    endforeach()
+    set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS};--expt-extended-lambda;${GENCODE_FLAGS};-lineinfo;")
+    if(NOT MSVC)
+      set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS};-Xcompiler -fPIC")
+    endif()
     set(CUDA_SOURCES
       plugin/updater_gpu/src/updater_gpu.cu
       plugin/updater_gpu/src/gpu_hist_builder.cu
       )
-    include(${PROJECT_SOURCE_DIR}/cmake/Utils.cmake)
-    include(${PROJECT_SOURCE_DIR}/cmake/Cuda.cmake)
     # use below for forcing specific arch
-    #cuda_compile(CUDA_OBJS ${CUDA_SOURCES} ${CUDA_NVCC_FLAGS} -arch=compute_52)
+    cuda_compile(CUDA_OBJS ${CUDA_SOURCES} ${CUDA_NVCC_FLAGS})
-    # use below for auto-detect, but gpu_grow currently doesn't work with 61
-    xgboost_cuda_compile(CUDA_OBJS ${CUDA_SOURCES} ${CUDA_NVCC_FLAGS})
-    if(MSVC)
-    else()
-      cuda_add_library(updater_gpu STATIC ${CUDA_SOURCES})
-      set(LINK_LIBRARIES ${LINK_LIBRARIES} updater_gpu)
-    endif()
 else()
     set(CUDA_OBJS "")
-    set(updater_gpu "")
 endif()
 
 add_library(objxgboost OBJECT ${SOURCES})

Makefile

@@ -93,12 +93,12 @@ endif
 CFLAGS += $(OPENMP_FLAGS)
 
 # for using GPUs
-COMPUTE ?= 60 35
+GPU_COMPUTE_VER ?= 50 52 60 61
 NVCC = nvcc
 INCLUDES = -Iinclude -I$(DMLC_CORE)/include -I$(RABIT)/include
 INCLUDES += -I$(CUB_PATH)
 INCLUDES += -I$(GTEST_PATH)/include
-CODE = $(foreach ver,$(COMPUTE),-gencode arch=compute_$(ver),code=sm_$(ver))
+CODE = $(foreach ver,$(GPU_COMPUTE_VER),-gencode arch=compute_$(ver),code=sm_$(ver))
 NVCC_FLAGS = --std=c++11 $(CODE) $(INCLUDES) -lineinfo --expt-extended-lambda
 NVCC_FLAGS += -Xcompiler=$(OPENMP_FLAGS) -Xcompiler=-fPIC
 ifeq ($(PLUGIN_UPDATER_GPU),ON)

plugin/updater_gpu/README.md

@@ -90,12 +90,27 @@ Visual studio community 2015, supported by cuda toolkit (http://docs.nvidia.com/
 
 ### For Developers!
 
+In case you want to build only for a specific GPU(s), for eg. GP100 and GP102,
+whose compute capability are 60 and 61 respectively:
+```bash
+$ cmake .. -DPLUGIN_UPDATER_GPU=ON -DGPU_COMPUTE_VER="60;61"
+```
+By default, the versions will include support for all GPUs in Maxwell and Pascal architectures.
+
 ### Using make
 Now, it also supports the usual 'make' flow to build gpu-enabled tree construction plugins. It's currently only tested on Linux. From the xgboost directory
 ```bash
 # make sure CUDA SDK bin directory is in the 'PATH' env variable
 $ make PLUGIN_UPDATER_GPU=ON
 ```
+
+Similar to cmake, if you want to build only for a specific GPU(s):
+```bash
+$ make PLUGIN_UPDATER_GPU=ON GPU_COMPUTE_VER="60 61"
+```
+
+### For Developers!
+
 Now, some of the code-base inside gpu plugins have googletest unit-tests inside 'tests/'.
 They can be enabled run along with other unit-tests inside '<xgboostRoot>/tests/cpp' using:
 ```bash
@@ -131,19 +146,11 @@ $ make PLUGIN_UPDATER_GPU=ON GTEST_PATH=${CACHE_PREFIX} test
 [Mitchell, Rory, and Eibe Frank. Accelerating the XGBoost algorithm using GPU computing. No. e2911v1. PeerJ Preprints, 2017.](https://peerj.com/preprints/2911/)
 
 ## Author
-<<<<<<< HEAD
-Rory Mitchell,
-Jonathan C. McKinney,
-Shankara Rao Thejaswi Nanditale,
-Vinay Deshpande,
-and the rest of the H2O.ai and NVIDIA team.
-=======
 Rory Mitchell
 Jonathan C. McKinney
 Shankara Rao Thejaswi Nanditale
 Vinay Deshpande
 ... and the rest of the H2O.ai and NVIDIA team.
->>>>>>> d2fbbdf4a39fa1f0af5cbd59a7912cf5caade34e
 
 Please report bugs to the xgboost/issues page.