From dbabd5e399b132654e75a3ca9589f35eca09f6cb Mon Sep 17 00:00:00 2001
From: Pavel Krajcevski <pavel@cs.unc.edu>
Date: Wed, 6 Mar 2013 20:58:01 -0500
Subject: [PATCH] Add command line option to choose atomics path for
 compression.

---
 CLTool/src/clwin32.cpp |  9 +++++++++
 Core/src/TexComp.cpp   | 15 ++++++++++++---
 2 files changed, 21 insertions(+), 3 deletions(-)

diff --git a/CLTool/src/clwin32.cpp b/CLTool/src/clwin32.cpp
index 7185746..5cb6663 100644
--- a/CLTool/src/clwin32.cpp
+++ b/CLTool/src/clwin32.cpp
@@ -59,6 +59,7 @@ void PrintUsage() {
   fprintf(stderr, "\t-n <num>\tCompress the image num times and give the average time and PSNR. Default: 1\n");
   fprintf(stderr, "\t-simd\t\tUse SIMD compression path\n");
   fprintf(stderr, "\t-t <num>\tCompress the image using <num> threads. Default: 1\n");
+  fprintf(stderr, "\t-a \tCompress the image using synchronization via atomic operations. Default: Off\n");
   fprintf(stderr, "\t-j <num>\tUse <num> blocks for each work item in a worker queue threading model. Default: (Blocks / Threads)\n");
 }
 
@@ -91,6 +92,7 @@ int _tmain(int argc, _TCHAR* argv[])
   int numCompressions = 1;
   bool bUseSIMD = false;
   bool bSaveLog = false;
+  bool bUseAtomics = false;
   
   bool knowArg = false;
   do {
@@ -162,6 +164,12 @@ int _tmain(int argc, _TCHAR* argv[])
       continue;
     }
 
+    if(strcmp(argv[fileArg], "-a") == 0) {
+      fileArg++;
+      bUseAtomics = true;
+      knowArg = true;
+      continue;
+    }
   } while(knowArg && fileArg < argc);
 
   if(numThreads > 1 && bSaveLog) {
@@ -194,6 +202,7 @@ int _tmain(int argc, _TCHAR* argv[])
   
   SCompressionSettings settings;
   settings.bUseSIMD = bUseSIMD;
+  settings.bUseAtomics = bUseAtomics;
   settings.iNumThreads = numThreads;
   settings.iQuality = quality;
   settings.iNumCompressions = numCompressions;
diff --git a/Core/src/TexComp.cpp b/Core/src/TexComp.cpp
index 80d19f9..9434a40 100644
--- a/Core/src/TexComp.cpp
+++ b/Core/src/TexComp.cpp
@@ -184,13 +184,15 @@ class AtomicThreadUnit : public TCCallable {
     m_OutBuf(outBuf),
     m_Height(height),
     m_Width(width),
-    m_Barrier(barrier)
+    m_Barrier(barrier),
+    m_NumCompressions(nCompressions),
+    m_CmpFnc(f)
   { }
 
   virtual ~AtomicThreadUnit() { }
   virtual void operator()() {
     m_Barrier->Wait();
-    for(int i = 0; i < m_NumCompressions; i++)
+    for(uint32 i = 0; i < m_NumCompressions; i++)
       (*m_CmpFnc)(m_InBuf, m_OutBuf, m_Width, m_Height);
   }
 };
@@ -374,7 +376,14 @@ bool CompressImageData(
     double cmpMSTime = 0.0;
 
     if(settings.iNumThreads > 1) {
-      if(settings.iJobSize > 0)
+      if(settings.bUseAtomics) {
+        //!KLUDGE!
+        unsigned int height = 4;
+        unsigned int width = dataSz / 16;
+
+        cmpMSTime = CompressImageWithAtomics(data, width, height, settings, cmpData);
+      }
+      else if(settings.iJobSize > 0)
         cmpMSTime = CompressImageWithWorkerQueue(data, dataSz, settings, cmpData);
       else
         cmpMSTime = CompressImageWithThreads(data, dataSz, settings, cmpData);