runs on GPU but need further tuning

examples/portable/inplaceTraspose/Makefile_ptx

@@ -1,12 +1,12 @@
-PROG=radixSort
+PROG=inplaceTranspose
-ISPC_SRC=radixSort.ispc
+ISPC_SRC=inplaceTranspose.ispc
 
-CU_SRC=radixSort.cu
+#CU_SRC=inplaceTranspose.cu
 # NVCC_FLAGS=-Xptxas=-O1
-CXX_SRC=radixSort.cpp  radixSort.cpp
+CXX_SRC=inplaceTranspose.cpp
-PTXCC_REGMAX=64
+PTXCC_REGMAX=32
 
-LLVM_GPU=1
+# LLVM_GPU=1
 NVVM_GPU=1
 
 include ../common_ptx.mk

examples/portable/inplaceTraspose/inplaceTranspose.cpp

@@ -43,38 +43,44 @@ int main(int argc, char * argv[])
       m*n*sizeof(int)*2/1e6);
 
 
-  std::vector< std::pair<int,int> > A(m*n);
+  std::pair<int,int> *A     = new std::pair<int,int>[m*n];
+  std::pair<int,int> *Acopy = new std::pair<int,int>[m*n];
+
+  for (int j = 0; j < n; j++)
+    for (int i = 0; i < m; i++)
+      A[j*m+i] = std::make_pair(i,j);
+
+  if (verbose)
+  {
+    fprintf(stderr, "Original: \n");
+    for (int j = 0; j < n; j++)
+    {
+      for (int i = 0; i < m; i++)
+      {
+        fprintf(stderr, "(%2d,%2d) ", A[j*m+i].first, A[j*m+i].second);
+      }
+      fprintf(stderr, "\n");
+    }
+    fprintf(stderr, "\n");
+    for (int i = 0; i < m*n; i++)
+      fprintf(stderr, "(%2d,%2d) ", A[i].first, A[i].second);
+    fprintf(stderr, "\n");
+    fprintf(stderr, "\n");
+  }
+
+
+  for (int j = 0; j < n; j++)
+    for (int i = 0; i < m; i++)
+      assert(A[j*m+i].first == i && A[j*m+i].second == j);
+
+  ispcSetMallocHeapLimit(1024ull*1024*1024*8);
+  ispcMemcpy(&Acopy[0], &A[0], sizeof(T)*m*n);
 
   int nrep = 10;
   double dt = 1e10;
   for (int r = 0; r < nrep; r++)
   {
-    for (int j = 0; j < n; j++)
+    ispcMemcpy(&A[0], &Acopy[0], sizeof(T)*m*n);
-      for (int i = 0; i < m; i++)
-        A[j*m+i] = std::make_pair(i,j);
-
-    if (r == 0 && verbose)
-    {
-      fprintf(stderr, "Original: \n");
-      for (int j = 0; j < n; j++)
-      {
-        for (int i = 0; i < m; i++)
-        {
-          fprintf(stderr, "(%2d,%2d) ", A[j*m+i].first, A[j*m+i].second);
-        }
-        fprintf(stderr, "\n");
-      }
-      fprintf(stderr, "\n");
-      for (int i = 0; i < m*n; i++)
-        fprintf(stderr, "(%2d,%2d) ", A[i].first, A[i].second);
-      fprintf(stderr, "\n");
-      fprintf(stderr, "\n");
-    }
-
-    for (int j = 0; j < n; j++)
-      for (int i = 0; i < m; i++)
-        assert(A[j*m+i].first == i && A[j*m+i].second == j);
-
     reset_and_start_timer();
     ispc::transpose((T*)&A[0], n, m);
     const double t1 = rtc();

examples/portable/inplaceTraspose/inplaceTranspose.ispc

@@ -31,6 +31,7 @@ int __sj(const int i, const uniform int j, const uniform int m, const uniform in
 }
 
 
+#if 0
 static inline 
 void transpose_serial(uniform T A[], const uniform int m, const uniform int n)
 {
@@ -83,6 +84,7 @@ void transpose_serial(uniform T A[], const uniform int m, const uniform int n)
   delete joverb;
   delete tmp;
 }
+#endif
 
 static uniform   T * uniform tmpAll = NULL;
 static uniform int * uniform joverb = NULL;
@@ -90,9 +92,9 @@ static uniform int * uniform iovera = NULL;
 static uniform int a,b,c;
 
 static 
-void transpose_init(const uniform int m, const uniform int n, const uniform int taskCount)
+void transpose_init(const uniform int m, const uniform int n, const uniform int nTask)
 {
-  const uniform int tmpSize = max(m,n) * programCount * taskCount;
+  const uniform int tmpSize = max(m,n) * programCount * nTask;
   tmpAll = uniform new uniform T  [tmpSize];
   joverb = uniform new uniform int[n];
   iovera = uniform new uniform int[m];
@@ -178,7 +180,7 @@ void transpose(uniform T A[], const uniform int m, const uniform int n)
 #if 0
   transpose_serial(A, m, n);
 #else
-  const uniform int nTask = num_cores()*4;
+  const uniform int nTask = 32*8; //num_cores()*4;
   transpose_init(m,n,nTask);
 
   launch [nTask] transpose_step1(A, m, n);
@@ -192,5 +194,6 @@ void transpose(uniform T A[], const uniform int m, const uniform int n)
 
   transpose_finalize();
 #endif
+  sync;
 }