Porting parallel stuff to unix.

src/nvcore/Utils.h

@@ -25,6 +25,12 @@ namespace nv
     inline uint32 asUnsigned(int32 x) { return (uint32) x; }
     inline uint64 asUnsigned(int64 x) { return (uint64) x; }
 
+    template <typename T> inline uint32 toU32(T x) {
+        nvDebugCheck(x <= UINT32_MAX);
+        nvDebugCheck(x >= 0);
+        return (uint32) x;
+    }
+
     /*
     template <typename T> inline int8 toI8(T x) { 
         nvDebugCheck(x <= INT8_MAX);

src/nvthread/Atomic.h

@@ -61,7 +61,7 @@ namespace nv {
         *ptr = value;   // on x86, stores are Release
         nvCompilerWriteBarrier();
 #else
-#error "Not implemented"
+#error "Atomics not implemented."
 #endif
     }
 
@@ -84,6 +84,25 @@ namespace nv {
 
         return (uint32)_InterlockedDecrement((long *)value);
     }
+#elif NV_CC_GNUC
+    // Many alternative implementations at:
+    // http://www.memoryhole.net/kyle/2007/05/atomic_incrementing.html
+
+    inline uint32 atomicIncrement(uint32 * value)
+    {
+        nvDebugCheck((intptr_t(value) & 3) == 0);
+
+        return __sync_fetch_and_add(value, 1);
+    }
+
+    inline uint32 atomicDecrement(uint32 * value)
+    {
+        nvDebugCheck((intptr_t(value) & 3) == 0);
+
+        return __sync_fetch_and_sub(value, 1);
+    }
+#else
+#error "Atomics not implemented."
 #endif
 
 

src/nvthread/CMakeLists.txt

@@ -1,11 +1,13 @@
 PROJECT(nvthreads)
 
 SET(THREADS_SRCS
-	nvthreads.h
+	nvthread.h nvthread.cpp
+	Atomic.h
+	Event.h Event.cpp
 	Mutex.h Mutex.cpp
-	SpinWaiter.h SpinWaiter.cpp
+	ParallelFor.h ParallelFor.cpp
 	Thread.h Thread.cpp
-	ThreadLocalStorage.h ThreadLocalStorage.cpp)
+	ThreadPool.h ThreadPool.cpp)
 
 INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
 

src/nvthread/Event.cpp

@@ -48,5 +48,6 @@ void Event::wait() {
 }
 
 #elif NV_OS_UNIX
-    // @@ 
+    // @@ TODO
+#pragma NV_MESSAGE("Implement event using pthreads!")
 #endif	

src/nvthread/Mutex.cpp

@@ -86,4 +86,4 @@ void Mutex::unlock()
     nvDebugCheck(result == 0);
 }
 
-#endif // NV_OS
+#endif // NV_OS_UNIX

src/nvthread/Mutex.h

@@ -25,7 +25,7 @@ namespace nv
     private:
         struct Private;
         AutoPtr<Private> m;
-	};
+};
 
 
     // Templated lock that can be used with any mutex.

src/nvthread/Thread.cpp

@@ -23,6 +23,7 @@ struct Thread::Private
     void * arg;
 };
 
+
 #if NV_OS_WIN32
 
 unsigned long __stdcall threadFunc(void * arg) {
@@ -32,11 +33,13 @@ unsigned long __stdcall threadFunc(void * arg) {
 }
 
 #elif NV_OS_UNIX
+
 extern "C" void * threadFunc(void * arg) {
     Thread * thread = (Thread *)arg;
     thread->func(thread->arg);
     pthread_exit(0);
 }
+
 #endif
 
 
@@ -134,3 +137,4 @@ bool Thread::isRunning () const
     }
 //#endif
 }
+

src/nvthread/Thread.h

@@ -35,7 +35,7 @@ namespace nv
 	struct Private;
 	AutoPtr<Private> p;
 
-    public:
+    public: // @@ Why public? Also in private?!
         ThreadFunc * func;
         void * arg;
 

src/nvthread/ThreadPool.cpp

@@ -4,6 +4,8 @@
 #include "Mutex.h"
 #include "Thread.h"
 
+#include "nvcore/Utils.h"
+
 // Most of the time it's not necessary to protect the thread pool, but if it doesn't add a significant overhead, then it'd be safer to do it.
 #define PROTECT_THREAD_POOL 1
 
@@ -47,7 +49,7 @@ AutoPtr<ThreadPool> s_pool;
 
 
 /*static*/ void ThreadPool::workerFunc(void * arg) {
-    uint i = (uint)arg;
+    uint i = toU32((uintptr_t)arg); // This is OK, because workerCount should always be <<< 2^32
 
     while(true) 
     {

src/nvthread/nvthread.cpp

@@ -1,11 +1,16 @@
+// This code is in the public domain -- Ignacio Casta<74>o <castano@gmail.com>
 
 #include "nvthread.h"
 
 #include "Thread.h"
 
-#define WIN32_LEAN_AND_MEAN
-#define VC_EXTRALEAN
-#include <windows.h>
+#if NV_OS_WIN32
+    #include "Win32.h"
+#elif NV_OS_UNIX
+    #include <sys/types.h>
+    #include <sys/sysctl.h>
+#endif
+
 
 using namespace nv;
 

src/nvtt/tests/testsuite.cpp

@@ -269,6 +269,10 @@ struct MyOutputHandler : public nvtt::OutputHandler
         m_ptr = m_data;
     }
 
+    virtual void endImage()
+    {
+    }
+
     virtual bool writeData(const void * data, int size)
     {
         memcpy(m_ptr, data, size);