Initial commit

CMakeLists.txt

@@ -0,0 +1,45 @@
+cmake_minimum_required(VERSION 3.1)
+
+project(external_sort)
+
+set(CMAKE_CXX_STANDARD 11)
+
+option(GPU_CUDA_SUPPORT "CUDA support." OFF)
+
+find_package(OpenMP)
+if (OpenMP_CXX_FOUND)
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${OpenMP_CXX_FLAGS}")
+    set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} ${OpenMP_EXE_LINKER_FLAGS}")
+else()
+    message(WARNING "OpenMP not found!")
+endif()
+
+add_subdirectory(libs)
+
+convertIntoHeader(src/cl/merge_sort.cl src/cl/merge_sort_cl.h merge_sort_kernel)
+
+set(SOURCES
+        src/cl/merge_sort_cl.h
+        src/io_utils/buffer_reader.cpp
+        src/io_utils/buffer_reader.h
+        src/io_utils/buffer_writer.cpp
+        src/io_utils/buffer_writer.h
+        src/io_utils/file_reader.cpp
+        src/io_utils/file_reader.h
+        src/io_utils/file_writer.cpp
+        src/io_utils/file_writer.h
+)
+
+add_executable(input_generator src/main_generator.cpp ${SOURCES})
+target_link_libraries(input_generator libclew libgpu libutils)
+
+if (GPU_CUDA_SUPPORT)
+    find_package(CUDA REQUIRED)
+    add_definitions(-DCUDA_SUPPORT)
+    set(SOURCES ${SOURCES} src/cu/merge_sort.cu)
+    cuda_add_executable(${PROJECT_NAME} src/main_sorter.cpp ${SOURCES})
+else()
+    add_executable(${PROJECT_NAME} src/main_sorter.cpp ${SOURCES})
+endif()
+
+target_link_libraries(${PROJECT_NAME} libclew libgpu libutils)

README.md

@@ -0,0 +1,34 @@
+# Generating input data
+
+```
+./input_generator 1000000000 input.data
+```
+
+```
+Saving 1000000000 random floats (3814 MB) to file input.data
+Done in 9.94176 seconds (383.704 MB/s)
+```
+
+# Sorting
+
+```
+./external_sort input.data output.data
+```
+
+```
+Values number: 1000000000 (3814 MB)
+Pass #0: sorting part by part in core...
+    In core parts number: 60
+    Limit for values in core: 16777216 (64 MB)
+    IO: 51.1294 MB/s
+    Finished in 149.217 s (9% reading + 82% sorting + 10% writing)
+Pass #1: merging groups of 16 parts...
+    Input parts: 60 with 16777216 values (64 MB) in each
+    IO: 629.504 MB/s
+    Finished in 12.1199 s
+Pass #2: merging groups of 16 parts...
+    Input parts: 4 with 268435456 values (1024 MB) in each
+    IO: 343.152 MB/s
+    Finished in 22.2333 s
+Finished in 183.571 s
+```

libs/CMakeLists.txt

@@ -0,0 +1,3 @@
+add_subdirectory(clew)
+add_subdirectory(gpu)
+add_subdirectory(utils)