Kmeans||

CMakeLists.txt

@@ -7,151 +7,143 @@ SET_DEFAULT_CONFIGURATION_RELEASE()
 
 FIND_PACKAGE(OpenMP)
 FIND_PACKAGE(BLAS REQUIRED)
-FIND_PACKAGE(SWIG REQUIRED)
-FIND_PACKAGE(PythonLibs REQUIRED) # SWIG
-
-INCLUDE(${SWIG_USE_FILE})
 
 #============= OPTIONS & SETTINGS
 OPTION(USE_CUDA  "Build with GPU acceleration" ON)
+OPTION(USE_SWIG "Use swig to generate language bindings." ON)
+OPTION(BUILD_TESTS "Build tests." ON)
+OPTION(USE_SYSTEM_GTEST "Use system google tests." ON)
 OPTION(DEV_BUILD  "Dev build" OFF)
 
 # Compiler flags
 SET(CMAKE_CXX_STANDARD 11)
 SET(CMAKE_CXX_STANDARD_REQUIRED ON)
 SET(CMAKE_POSITION_INDEPENDENT_CODE ON)
-SET(CMAKE_CXX_FLAGS  "${CMAKE_CXX_FLAGS} -w")
-
-# PythonLibs' PYTHON_INCLUDE_PATH doesn't take into account virtualenv etc.
-# Open to suggestions how to do this better.
-EXECUTE_PROCESS(COMMAND python -c "import numpy; print(numpy.get_include())"
-        OUTPUT_VARIABLE PYTHON_INCLUDE_PATH_CUST
-        OUTPUT_STRIP_TRAILING_WHITESPACE)
+SET(CMAKE_CXX_FLAGS  "${CMAKE_CXX_FLAGS}")
+SET(GPU_COMPUTE_VER "" CACHE STRING
+  "Semicolon separated list of compute versions to be built against,  e.g. -DGPU_COMPUTE_VER='35;61'")
 
 if(OpenMP_CXX_FOUND OR OPENMP_FOUND)
-        SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${OpenMP_CXX_FLAGS}")
+  SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${OpenMP_CXX_FLAGS}")
 endif()
 #============= OPTIONS & SETTINGS
 
 # TODO probably to be removed after POGS is out in favor of XGboost GLM
 ADD_DEFINITIONS(
-        -D_GITHASH_=0
-        -DH2O4GPU_DOUBLE
-        -DH2O4GPU_SINGLE
-)
+  -D_GITHASH_=0
+  -DH2O4GPU_DOUBLE
+  -DH2O4GPU_SINGLE)
 
 #============= BUILD COMMON CPU/GPU CODE
 FILE(GLOB_RECURSE COMMON_SOURCES
-        src/common/*.cpp
-        src/common/*.h
-        src/interface_c/*.cpp
-        src/interface_c/*.h
-        )
+  src/common/*.cpp
+  src/common/*.h
+  src/interface_c/*.cpp
+  src/interface_c/*.h)
 
 INCLUDE_DIRECTORIES(
-        src/include
-        src/cpu/include
-        # Here and not in target_include_directories b/c cmake < 3.7 which we use in Dockerfiles does not support it
-        src/gpu/include
-        ${PYTHON_INCLUDE_PATH}
-        ${PYTHON_INCLUDE_PATH_CUST}
-)
+  src/include
+  src/cpu/include
+  # Here and not in target_include_directories b/c cmake < 3.7 which we use in Dockerfiles does not support it
+  src/gpu/include)
 
 ADD_LIBRARY(commonh2o4gpu OBJECT ${COMMON_SOURCES})
 #============= BUILD COMMON CPU/GPU CODE
 
 #============= BUILD CPU LIBRARY
 FILE(GLOB_RECURSE CPU_SOURCES
-        src/cpu/*.cpp
-        src/cpu/*.h
-        )
+  src/cpu/*.cpp
+  src/cpu/*.h)
 
 ADD_LIBRARY(cpuh2o4gpu STATIC ${CPU_SOURCES} $<TARGET_OBJECTS:commonh2o4gpu>)
 TARGET_LINK_LIBRARIES(cpuh2o4gpu ${BLAS_LIBRARIES})
 #============= BUILD CPU LIBRARY
 
-#============= SWIG
-SET(CMAKE_SWIG_FLAGS -Werror)
-#============= SWIG
-
-#============= CPU SWIG
-SET_SOURCE_FILES_PROPERTIES(src/swig/ch2o4gpu_cpu.i PROPERTIES CPLUSPLUS ON)
-
-if (${CMAKE_VERSION} VERSION_LESS "3.8.0")
-        SWIG_ADD_MODULE(ch2o4gpu_cpu python src/swig/ch2o4gpu_cpu.i)
+if (USE_CUDA)
+  SET(HG_USE_CUDA 1)
 else()
-        SWIG_ADD_LIBRARY(ch2o4gpu_cpu LANGUAGE python SOURCES src/swig/ch2o4gpu_cpu.i)
-endif()
-
-SWIG_LINK_LIBRARIES(ch2o4gpu_cpu cpuh2o4gpu ${PYTHON_LIBRARIES})
-
-SET_TARGET_PROPERTIES(${SWIG_MODULE_ch2o4gpu_cpu_REAL_NAME} PROPERTIES
-        LINK_FLAGS ${OpenMP_CXX_FLAGS})
-#============= CPU SWIG
+  SET(HG_USE_CUDA 0)
+endif(USE_CUDA)
+ADD_SUBDIRECTORY(${CMAKE_CURRENT_LIST_DIR}/src/common)
 
 if(USE_CUDA)
-        FIND_PACKAGE(CUDA 8.0 REQUIRED)
-        FIND_PACKAGE(NVML REQUIRED)
-
-        #============= BUILD GPU LIBRARY
-        ADD_DEFINITIONS(
-                -DCUDA_MAJOR=${CUDA_VERSION_MAJOR}
-                -DHAVECUDA
-        )
-
-        if(DEV_BUILD)
-                MESSAGE(STATUS "Building DEVELOPER compute capability version.")
-                SET(GPU_COMPUTE_VER 61 CACHE STRING
-                        "Space separated list of compute versions to be built against")
-        else()
-                MESSAGE(STATUS "Building RELEASE compute capability version.")
-                SET(GPU_COMPUTE_VER 35;50;52;60;61 CACHE STRING
-                        "Space separated list of compute versions to be built against")
-        endif()
-
-        if(((CUDA_VERSION_MAJOR EQUAL 9) OR (CUDA_VERSION_MAJOR GREATER 9)) AND NOT DEV_BUILD)
-                MESSAGE(STATUS "CUDA 9.0 detected, adding Volta compute capability (7.0).")
-                SET(GPU_COMPUTE_VER "${GPU_COMPUTE_VER};70")
-        endif()
-
-        SET(GENCODE_FLAGS "")
-        FORMAT_GENCODE_FLAGS("${GPU_COMPUTE_VER}" GENCODE_FLAGS)
-        SET(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS};-Xcompiler -fPIC; -std=c++11;--expt-extended-lambda;--expt-relaxed-constexpr;${GENCODE_FLAGS};-lineinfo; -w;")
-
-        FILE(GLOB_RECURSE GPU_SOURCES
-                src/*.cu
-                src/*.cuh
-                src/common/*.cpp
-                src/common/*.h
-                )
-
-        CUDA_ADD_LIBRARY(gpuh2o4gpu ${GPU_SOURCES} $<TARGET_OBJECTS:commonh2o4gpu> STATIC)
-
-        if($ENV{USENVTX})
-                MESSAGE(STATUS "Building with NVTX support on.")
-                SET(NVTX_LIBRARY nvToolsExt)
-        endif()
-
-        TARGET_LINK_LIBRARIES(gpuh2o4gpu
-                ${CUDA_CUBLAS_LIBRARIES}
-                ${CUDA_cusolver_LIBRARY}
-                ${CUDA_cusparse_LIBRARY}
-                ${BLAS_LIBRARIES}
-                ${NVTX_LIBRARY}
-                ${NVML_LIBRARY})
-        #============= BUILD GPU LIBRARY
-
-        #============= GPU SWIG
-        SET_SOURCE_FILES_PROPERTIES(src/swig/ch2o4gpu_gpu.i PROPERTIES CPLUSPLUS ON)
-
-        if (${CMAKE_VERSION} VERSION_LESS "3.8.0")
-                SWIG_ADD_MODULE(ch2o4gpu_gpu python src/swig/ch2o4gpu_gpu.i)
-        else()
-                SWIG_ADD_LIBRARY(ch2o4gpu_gpu LANGUAGE python SOURCES src/swig/ch2o4gpu_gpu.i)
-        endif()
-        SWIG_LINK_LIBRARIES(ch2o4gpu_gpu gpuh2o4gpu ${PYTHON_LIBRARIES})
-
-        SET_TARGET_PROPERTIES(${SWIG_MODULE_ch2o4gpu_gpu_REAL_NAME} PROPERTIES
-                LINK_FLAGS ${OpenMP_CXX_FLAGS})
-        #============= GPU SWIG
+  FIND_PACKAGE(CUDA 8.0 REQUIRED)
+  FIND_PACKAGE(NVML REQUIRED)
+
+  #============= BUILD GPU LIBRARY
+  ADD_DEFINITIONS(
+    -DCUDA_MAJOR=${CUDA_VERSION_MAJOR}
+    -DHAVECUDA
+    )
+
+  if(DEV_BUILD AND NOT GPU_COMPUTE_VER)
+    MESSAGE(STATUS "Building DEVELOPER compute capability version.")
+    SET(GPU_COMPUTE_VER 61)
+  elseif(NOT GPU_COMPUTE_VER)
+    MESSAGE(STATUS "Building RELEASE compute capability version.")
+    SET(GPU_COMPUTE_VER 35;50;52;60;61)
+  endif()
+
+  if(((CUDA_VERSION_MAJOR EQUAL 9)
+	OR (CUDA_VERSION_MAJOR GREATER 9))
+      AND NOT DEV_BUILD
+      AND NOT GPU_COMPUTE_VER)
+    MESSAGE(STATUS "CUDA 9.0 detected, adding Volta compute capability (7.0).")
+    SET(GPU_COMPUTE_VER "${GPU_COMPUTE_VER};70")
+  endif()
+
+  SET(GENCODE_FLAGS "")
+  FORMAT_GENCODE_FLAGS("${GPU_COMPUTE_VER}" GENCODE_FLAGS)
+  MESSAGE("CUDA architecture flags ${GENCODE_FLAGS}")
+
+  SET(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS};-Xcompiler -fPIC; -std=c++11;--expt-extended-lambda;--expt-relaxed-constexpr;${GENCODE_FLAGS};-lineinfo;")
+
+  FILE(GLOB_RECURSE GPU_SOURCES
+    src/*.cu
+    src/*.cuh
+    src/gpu/matrix/*.cpp
+    src/common/*.cpp
+    src/common/*.h)
+
+  CUDA_ADD_LIBRARY(gpuh2o4gpu ${GPU_SOURCES} $<TARGET_OBJECTS:commonh2o4gpu> STATIC)
+
+  if($ENV{USENVTX})
+    MESSAGE(STATUS "Building with NVTX support on.")
+    SET(NVTX_LIBRARY nvToolsExt)
+  endif()
+
+  TARGET_LINK_LIBRARIES(gpuh2o4gpu
+    ${CUDA_CUBLAS_LIBRARIES}
+    ${CUDA_cusolver_LIBRARY}
+    ${CUDA_cusparse_LIBRARY}
+    ${BLAS_LIBRARIES}
+    ${NVTX_LIBRARY}
+    ${NVML_LIBRARY})
+  #============= BUILD GPU LIBRARY
 endif()
+
+if(USE_SWIG)
+  ADD_SUBDIRECTORY(${CMAKE_CURRENT_LIST_DIR}/src/swig)
+endif(USE_SWIG)
+
+#============= Tests
+if(BUILD_TESTS)
+  ENABLE_TESTING()
+  if (USE_SYSTEM_GTEST)
+    FIND_PACKAGE(GTest REQUIRED)
+  else ()
+    ADD_SUBDIRECTORY(${CMAKE_CURRENT_LIST_DIR}/tests/googletest)
+    SET(GTEST_LIBRARIES gtest gtest_main)
+  endif (USE_SYSTEM_GTEST)
+  if (USE_CUDA)
+    FILE(GLOB_RECURSE CUDA_TEST_SOURCES "tests/cpp/*.cu")
+    CUDA_COMPILE(CUDA_TEST_OBJS ${CUDA_TEST_SOURCES})
+  endif(USE_CUDA)
+
+  ADD_EXECUTABLE(test-h2o4gpu ${CUDA_TEST_OBJS})
+  TARGET_LINK_LIBRARIES(test-h2o4gpu gpuh2o4gpu)
+  TARGET_LINK_LIBRARIES(test-h2o4gpu ${GTEST_LIBRARIES})
+
+  ADD_TEST(TestH2O4GPU test-h2o4gpu)
+endif(BUILD_TESTS)
+#============= Tests

src/common/CMakeLists.txt

@@ -0,0 +1 @@
+configure_file(${CMAKE_CURRENT_SOURCE_DIR}/utils.h.in ${CMAKE_CURRENT_SOURCE_DIR}/utils.h)

src/common/utils.h.in

@@ -0,0 +1,54 @@
+/*!
+ * Copyright 2017-2018 H2O.ai, Inc.
+ * License   Apache License Version 2.0 (see LICENSE for details)
+ */
+#pragma once
+#include <vector>
+
+#include <iostream>
+#include <sstream>
+
+#include "cblas/cblas.h"
+
+#define USE_CUDA() @HG_USE_CUDA@
+
+template<typename T>
+void self_dot(std::vector<T> array_in, int n, int dim,
+              std::vector<T>& dots);
+
+void compute_distances(std::vector<double> data_in,
+                       std::vector<double> centroids_in,
+                       std::vector<double> &pairwise_distances,
+                       int n, int dim, int k);
+
+void compute_distances(std::vector<float> data_in,
+                       std::vector<float> centroids_in,
+                       std::vector<float> &pairwise_distances,
+                       int n, int dim, int k);
+
+// Matrix host dev
+#define HG_HOSTDEV       __host__   __device__
+#define HG_DEV           __device__
+#define HG_DEVINLINE     __device__ __forceinline__
+#define HG_HOSTDEVINLINE __host__   __device__      __forceinline__
+
+#define h2o4gpu_error(x) error(x, __FILE__, __LINE__);
+
+inline void error(const char* e, const char* file, int line)
+{
+  std::stringstream ss;
+  ss << e << " - " << file << "(" << line << ")";
+  //throw error_text;
+  std::cerr << ss.str() << std::endl;
+  exit(-1);
+}
+
+#define h2o4gpu_check(condition, msg) check(condition, msg, __FILE__, __LINE__);
+
+inline void check(bool val, const char* e, const char* file, int line)
+{
+  if (!val)
+  {
+    error(e, file, line);
+  }
+}

src/cpu/h2o4gpuglm.cpp

@@ -6,6 +6,7 @@
 #include <algorithm>
 #include <limits>
 #include <deque>
+#include <numeric>
 
 #include <functional>
 #include <cstring>

src/cpu/h2o4gpukmeans.cpp

@@ -10,7 +10,7 @@
 #include <algorithm>
 #include <vector>
 //#include "mkl.h"
-#include "cblas.h"
+#include "cblas/cblas.h"
 #include <atomic>
 #include <csignal>
 

src/gpu/h2o4gpuglm.cu

@@ -10,6 +10,7 @@
 #include <thrust/transform.h>
 
 #include <algorithm>
+#include <numeric>
 #include <limits>
 #include <deque>
 

src/gpu/kmeans/kmeans_general.h

@@ -4,6 +4,8 @@
  */
 #pragma once
 #include "../../common/logger.h"
+#include "../utils/utils.cuh"
+#include "stdio.h"
 #define MAX_NGPUS 16
 
 #define VERBOSE 0
@@ -12,8 +14,6 @@
 
 // TODO(pseudotensor): Avoid throw for python exception handling.  Need to avoid all exit's and return exit code all the way back.
 #define gpuErrchk(ans) { gpu_assert((ans), __FILE__, __LINE__); }
-#define safe_cuda(ans) throw_on_cuda_error((ans), __FILE__, __LINE__);
-#define safe_cublas(ans) throw_on_cublas_error((ans), __FILE__, __LINE__);
 
 #define CUDACHECK(cmd) do {                           \
     cudaError_t e = cmd;                              \