making vectorforth independent from TBB

Author: janm31415

.gitmodules

@@ -0,0 +1,3 @@
+[submodule "jtk"]
+	path = jtk
+	url = https://github.com/janm31415/jtk

CMakeLists.txt

@@ -9,6 +9,10 @@ set(CMAKE_RUNTIME_OUTPUT_DIRECTORY "${CMAKE_CURRENT_BINARY_DIR}/bin")
 set(CMAKE_CXX_STANDARD 17)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
 
+set(JTK_THREADING "std" CACHE STRING  "Choose your threading library: tbb is Intel's Threading Building Blocks library, ppl is Windows' concurrency library, std uses std::thread, and none does not use threading.")
+ 
+set_property(CACHE JTK_THREADING PROPERTY STRINGS tbb ppl std none)
+
 set_property(GLOBAL PROPERTY USE_FOLDERS ON)
 
 set(AVX512 OFF CACHE BOOL "Set ON for AVX-512, set OFF for AVX2.")
@@ -21,6 +25,7 @@ message(STATUS "SINCOS_METHOD='${SINCOS_METHOD}'")
 
 add_subdirectory(asm)
 add_subdirectory(asm.tests)
+add_subdirectory(jtk)
 add_subdirectory(sf)
 add_subdirectory(shaderforth)
 add_subdirectory(vectorforth)

jtk

@@ -10,6 +10,12 @@ if (UNIX)
   set(OPENGL_LIBRARIES "/usr/lib/x86_64-linux-gnu/libOpenGL.so.0.0.0")
 endif (UNIX)
 
+if (APPLE)
+  set(TBB_INCLUDE_DIR "")
+  set(TBB_LIBRARIES "")
+  set(OPENGL_LIBRARIES "/System/Library/Frameworks/OpenGL.framework/OpenGL")
+endif (APPLE)
+
 set(HDRS
 window.h
     )
@@ -50,17 +56,30 @@ if (${SINCOS_METHOD} STREQUAL "approximate")
 add_definitions(-DSINCOSAPPROX)
 endif (${SINCOS_METHOD} STREQUAL "approximate")
 
+if (${JTK_THREADING} STREQUAL "tbb")
+add_definitions(-D_ENABLE_TBB)
+endif (${JTK_THREADING} STREQUAL "tbb")
+
+if (${JTK_THREADING} STREQUAL "ppl")
+add_definitions(-D_ENABLE_PPL)
+endif (${JTK_THREADING} STREQUAL "ppl")
+
+if (${JTK_THREADING} STREQUAL "std")
+add_definitions(-D_ENABLE_THREADS)
+endif (${JTK_THREADING} STREQUAL "std")
+
 add_executable(sf ${HDRS} ${SRCS})
 source_group("Header Files" FILES ${hdrs})
 source_group("Source Files" FILES ${srcs})
-
+    
+if (WIN32)	
  target_include_directories(sf
     PRIVATE
     ${CMAKE_CURRENT_SOURCE_DIR}/../
+    ${CMAKE_CURRENT_SOURCE_DIR}/../jtk/
     ${TBB_INCLUDE_DIR}
     )	
-    
-if (WIN32)	
+
 target_link_libraries(sf
     PRIVATE
     asm
@@ -70,11 +89,28 @@ target_link_libraries(sf
 endif (WIN32)    
 
 if (UNIX)	
+
+set(XLIBINCLUDE "")
+set(XLIBLIBRARY "X11.so")
+
+if (APPLE)
+set(XLIBINCLUDE "/usr/X11/include/")
+set(XLIBLIBRARY "/usr/X11/lib/libX11.dylib")
+endif (APPLE)
+
+target_link_libraries(sf
+    PRIVATE
+    asm
+    vectorforth
+    ${TBB_LIBRARIES}    
+    ${XLIBINCLUDE}
+    )
+
 target_link_libraries(sf
     PRIVATE
     asm
     vectorforth
-    X11.so
+    ${XLIBLIBRARY}
     ${TBB_LIBRARIES}    
     )
 endif (UNIX)

sf/CMakeLists.txt

@@ -16,8 +16,7 @@
 #include <vectorforth/stdlib.h>
 #include <vectorforth/sincos_table.h>
 
-#include <tbb/parallel_for.h>
-#include <tbb/enumerable_thread_specific.h>
+#include <jtk/concurrency.h>
 
 #include "window.h"
 
@@ -282,7 +281,7 @@ int main(int argc, char** argv)
   int frame = 0;
   auto last_tic = std::chrono::high_resolution_clock::now();
   float time = 0.f;
-  tbb::enumerable_thread_specific< VF::context > local_context;
+  jtk::combinable< VF::context > local_context;
 
   while (!l.quit)
     {
@@ -318,7 +317,7 @@ int main(int argc, char** argv)
 #ifdef SINGLE
     for (int y = 0; y < h; ++y)
 #else
-    tbb::parallel_for((int)0, h, [&](int y)    
+    jtk::parallel_for((int)0, h, [&](int y)    
 #endif
       {
 
@@ -463,10 +462,11 @@ int main(int argc, char** argv)
 
   _mm_free(image);
 
-  for (auto& ctxt : local_context)
+  local_context.combine_each([](VF::context& ctxt)
     {
     VF::destroy_context(ctxt);
-    }
+    });
+
   ASM::free_assembled_function((void*)fun, fun_size);
 
   printf("\n");

sf/main.cpp

@@ -14,6 +14,14 @@ if (UNIX)
   set(OPENGL_LIBRARIES "/usr/lib/x86_64-linux-gnu/libOpenGL.so.0.0.0")
 endif (UNIX)
 
+if (APPLE)
+  set(SDL_INCLUDE_DIR "/Library/Frameworks/SDL2.framework/Headers/")
+  set(SDL_LIBRARIES "/Library/Frameworks/SDL2.framework/SDL2")
+  set(TBB_INCLUDE_DIR "")
+  set(TBB_LIBRARIES "")
+  set(OPENGL_LIBRARIES "/System/Library/Frameworks/OpenGL.framework/OpenGL")
+endif (APPLE)
+
 set(IMGUI
 ${CMAKE_CURRENT_SOURCE_DIR}/../thirdparty/imgui_sdl/imgui.h
 ${CMAKE_CURRENT_SOURCE_DIR}/../thirdparty/imgui_sdl/imgui_impl_opengl2.h
@@ -62,10 +70,10 @@ set(CMAKE_CXX_FLAGS_RELEASE "/W4 /MP /GF /O2 /Ob2 /Oi /Ot /MD /Zi")
 endif (WIN32)
 
 if (UNIX)
-set(CMAKE_C_FLAGS_DEBUG "-mavx2 -mfma")
-set(CMAKE_CXX_FLAGS_DEBUG "-mavx2 -mfma")
-set(CMAKE_C_FLAGS_RELEASE "-mavx2 -mfma")
-set(CMAKE_CXX_FLAGS_RELEASE "-mavx2 -mfma")
+set(CMAKE_C_FLAGS_DEBUG "${CMAKE_C_FLAGS_DEBUG} -mavx2 -mfma -pthread")
+set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} -mavx2 -mfma -pthread")
+set(CMAKE_C_FLAGS_RELEASE "${CMAKE_C_FLAGS_RELEASE} -mavx2 -mfma -pthread")
+set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -mavx2 -mfma -pthread")
 endif (UNIX)
 
 # general build definitions
@@ -84,6 +92,18 @@ if (${SINCOS_METHOD} STREQUAL "approximate")
 add_definitions(-DSINCOSAPPROX)
 endif (${SINCOS_METHOD} STREQUAL "approximate")
 
+if (${JTK_THREADING} STREQUAL "tbb")
+add_definitions(-D_ENABLE_TBB)
+endif (${JTK_THREADING} STREQUAL "tbb")
+
+if (${JTK_THREADING} STREQUAL "ppl")
+add_definitions(-D_ENABLE_PPL)
+endif (${JTK_THREADING} STREQUAL "ppl")
+
+if (${JTK_THREADING} STREQUAL "std")
+add_definitions(-D_ENABLE_THREADS)
+endif (${JTK_THREADING} STREQUAL "std")
+
 if (WIN32)
 add_executable(shaderforth WIN32 ${HDRS} ${SRCS} ${IMGUI} ${JSON})
 endif (WIN32)
@@ -100,6 +120,7 @@ source_group("ThirdParty/json" FILES ${JSON})
  target_include_directories(shaderforth
     PRIVATE
     ${CMAKE_CURRENT_SOURCE_DIR}/../
+    ${CMAKE_CURRENT_SOURCE_DIR}/../jtk/
     ${CMAKE_CURRENT_SOURCE_DIR}/../thirdparty/imgui_sdl/
     ${CMAKE_CURRENT_SOURCE_DIR}/../thirdparty/json/
     ${SDL_INCLUDE_DIR}

shaderforth/CMakeLists.txt

@@ -5,8 +5,7 @@
 
 #include <sstream>
 
-#include <tbb/parallel_for.h>
-#include <tbb/enumerable_thread_specific.h>
+#include <jtk/concurrency.h>
 
 #include <vectorforth/context.h>
 #include <vectorforth/compiler.h>
@@ -98,10 +97,10 @@ shader_program::shader_program(int w, int h) : _w(w), _h(h), _fun_size(0), _fun(
 
 shader_program::~shader_program()
   {  
-  for (auto& ctxt : local_context)
+  local_context.combine_each([](VF::context& ctxt)
     {
     VF::destroy_context(ctxt);
-    }
+    });
   if (_fun)
     ASM::free_assembled_function((void*)_fun, _fun_size);
   }
@@ -233,7 +232,7 @@ void shader_program::run(image<uint32_t>& im)
   __m256 global_time_val = _mm256_set1_ps(_input.global_time);
 #endif
 
-  tbb::parallel_for((int)0, _h, [&](int y)
+  jtk::parallel_for((int)0, _h, [&](int y)
     {
 
     uint32_t* p_im = im.data() + y * _w;

shaderforth/shader_program.cpp

@@ -6,7 +6,11 @@
 #include <string>
 #include <stdint.h>
 
-#include <tbb/enumerable_thread_specific.h>
+#ifdef _WIN32
+#include <windows.h>
+#endif
+
+#include <jtk/concurrency.h>
 
 #include <vectorforth/context.h>
 
@@ -33,5 +37,5 @@ class shader_program
     typedef void(*fun_ptr)(void*);
     fun_ptr _fun;
 
-    tbb::enumerable_thread_specific< VF::context > local_context;
+    jtk::combinable< VF::context > local_context;
   };

shaderforth/shader_program.h

@@ -26,10 +26,10 @@ set(CMAKE_CXX_FLAGS_RELEASE "/W4 /MP /GF /O2 /Ob2 /Oi /Ot /MD /Zi /DNDEBUG")
 endif (WIN32)
 
 if (UNIX)
-set(CMAKE_C_FLAGS_DEBUG "-mavx2 -mfma")
-set(CMAKE_CXX_FLAGS_DEBUG "-mavx2 -mfma")
-set(CMAKE_C_FLAGS_RELEASE "-mavx2 -mfma")
-set(CMAKE_CXX_FLAGS_RELEASE "-mavx2 -mfma")
+set(CMAKE_C_FLAGS_DEBUG "${CMAKE_C_FLAGS_DEBUG} -mavx2 -mfma")
+set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} -mavx2 -mfma")
+set(CMAKE_C_FLAGS_RELEASE "${CMAKE_C_FLAGS_RELEASE} -mavx2 -mfma")
+set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -mavx2 -mfma")
 endif (UNIX)
 
 # general build definitions

vectorforth.tests/CMakeLists.txt

@@ -48,10 +48,10 @@ set(CMAKE_CXX_FLAGS_RELEASE "/W4 /MP /GF /O2 /Ob2 /Oi /Ot /MD /Zi /DNDEBUG /arch
 endif (WIN32)
 
 if (UNIX)
-set(CMAKE_C_FLAGS_DEBUG "-mavx2 -mfma")
-set(CMAKE_CXX_FLAGS_DEBUG "-mavx2 -mfma")
-set(CMAKE_C_FLAGS_RELEASE "-mavx2 -mfma")
-set(CMAKE_CXX_FLAGS_RELEASE "-mavx2 -mfma")
+set(CMAKE_C_FLAGS_DEBUG "${CMAKE_C_FLAGS_DEBUG} -mavx2 -mfma")
+set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} -mavx2 -mfma")
+set(CMAKE_C_FLAGS_RELEASE "${CMAKE_C_FLAGS_RELEASE} -mavx2 -mfma")
+set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -mavx2 -mfma")
 endif (UNIX)
 
 # general build definitions