+1

examples_ptx/radixSort/radixSort.ispc

@@ -19,30 +19,37 @@ void countPass(
   uniform int * uniform     counts = countsAll + blockIdx*NUMDIGITS;
   const uniform int           nloc = min(numElements - blockIdx*blockDim, blockDim);
 
-  foreach (digit = 0 ... NUMDIGITS)
-    counts[digit] = 0;
 
-#if 1
+#if 0
+  uniform int shcounts[NUMDIGITS];
+#else
+#define shcounts counts
+#endif
+
+  foreach (digit = 0 ... NUMDIGITS)
+    shcounts[digit] = 0;
+
   foreach (i = 0 ... nloc)
   {
     const int key = mask & ((unsigned int)keys[i] >> bit);
     uniform int skey;
     if (reduce_equal(key, &skey) == true)
-      counts[skey] += reduce_add(1);
+      shcounts[skey] += reduce_add(1);
     else
     {
 #ifdef __NVPTX__
-      atomic_add_global(&counts[key], 1);
+      atomic_add_global(&shcounts[key], 1);
 #else
-      atomic_add_local(&counts[key], 1);
+      atomic_add_local(&shcounts[key], 1);
 #endif
     }
   }
-#else
-#endif
 
   foreach (digit = 0 ... NUMDIGITS)
-    atomic_add_global(&countsGlobal[digit], counts[digit]);
+  {
+    counts[digit] = shcounts[digit];
+    atomic_add_global(&countsGlobal[digit], shcounts[digit]);
+  }
 }
 
 task
@@ -51,22 +58,23 @@ void sortPass(
     uniform int sorted[],
     uniform int bit,
     uniform int numElements,
-    uniform int digitOffsetsAll[],
+    uniform int digitOffsetsAll[])
-    uniform int sharedCounts[])
 {
   const uniform int  blockIdx = taskIndex;
   const uniform int numBlocks = taskCount;
 
   const uniform int  blockDim = (numElements + numBlocks - 1) / numBlocks;
 
-
+  uniform int localCounts[NUMDIGITS];
-  uniform int * uniform localCounts = sharedCounts + blockIdx*NUMDIGITS;
 
   const uniform int keyIndex = blockIdx * blockDim;
   uniform int * uniform keys = keysAll + keyIndex;
   uniform int * uniform digitOffsets = digitOffsetsAll + blockIdx*NUMDIGITS;
   const uniform int nloc = min(numElements - keyIndex, blockDim);
 
+  const int unitScan = exclusive_scan_add(1);
+
+
   const uniform int mask = (1 << NUMBITS) - 1;
   foreach (i = 0 ... NUMDIGITS)
     localCounts[i] = 0;
@@ -75,10 +83,18 @@ void sortPass(
   {
     const int key = mask & ((unsigned int)keys[i] >> bit);
     int rel;
-    foreach_active(iv)
+    if (reduce_equal(key) == true)
     {
-      rel = localCounts[key];
+      rel = localCounts[key] + unitScan;
-      localCounts[key]++;
+      localCounts[key] = rel+1;
+    }
+    else
+    {
+#ifdef __NVPTX__
+      rel = atomic_add_global(&localCounts[key],1);
+#else
+      rel = atomic_add_local(&localCounts[key],1);
+#endif
     }
     const int scatter = rel + digitOffsets[key];
     sorted [scatter] = keys[i];
@@ -177,14 +193,12 @@ inline void radixExclusiveScan(
 
 static uniform int * uniform memoryPool = NULL;
 static uniform int numBlocks;
-static uniform int nSharedCounts;
 static uniform int nCountsGlobal;
 static uniform int nExcScan;
 static uniform int nCountsBlock;
 static uniform int nPartialSum;
 static uniform int nPrefixSum;
 
-static uniform int * uniform sharedCounts;
 static uniform int * uniform countsGlobal;
 static uniform int * uniform excScan;
 static uniform int * uniform counts;
@@ -198,7 +212,6 @@ export void radixSort_alloc(const uniform int n)
 {
   assert(memoryPool == NULL);
   numBlocks     = num_cores()*4;
-  nSharedCounts = NUMDIGITS*numBlocks;
   nCountsGlobal = NUMDIGITS;
   nExcScan      = NUMDIGITS*numBlocks;
   nCountsBlock  = NUMDIGITS*numBlocks;
@@ -207,7 +220,6 @@ export void radixSort_alloc(const uniform int n)
 
 
   const uniform int nalloc = 
-    nSharedCounts +
     nCountsGlobal +
     nExcScan +
     nCountsBlock + 
@@ -216,8 +228,7 @@ export void radixSort_alloc(const uniform int n)
 
   memoryPool = uniform new uniform int[nalloc];
 
-  sharedCounts = memoryPool;
+  countsGlobal = memoryPool;
-  countsGlobal = sharedCounts + nSharedCounts;
   excScan      = countsGlobal + nCountsGlobal;
   counts       = excScan      + nExcScan;
   partialSum   = counts       + nCountsBlock;
@@ -292,8 +303,7 @@ export void radixSort(
           bufKeys, 
           bit, 
           numElements,
-          excScan,
+          excScan);
-          sharedCounts);
     sync;
 
     uniform int * uniform tmp = keys;