add and update samples with CUDA 11.3 support

Samples/binaryPartitionCG/README.md

@@ -24,7 +24,7 @@ x86_64, ppc64le, armv7l
 
 ## Prerequisites
 
-Download and install the [CUDA Toolkit 11.2](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.
+Download and install the [CUDA Toolkit 11.3](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.
 
 ## Build and Run
 

Samples/binaryPartitionCG/binaryPartitionCG.cu

@@ -31,14 +31,16 @@
  * 1.) Each thread loads a value from random array.
  * 2.) then checks if it is odd or even.
  * 3.) create binary partition group based on the above predicate
- * 4.) we count the number of odd/even in the group based on size of the binary groups
+ * 4.) we count the number of odd/even in the group based on size of the binary
+       groups
  * 5.) write it global counter of odd.
- * 6.) sum the values loaded by individual threads(using reduce) and write it to global 
- *     even & odd elements sum.
+ * 6.) sum the values loaded by individual threads(using reduce) and write it to
+       global even & odd elements sum.
  *
- * **NOTE** : binary_partition results in splitting warp into divergent thread groups
-              this is not good from performance perspective, but in cases where warp 
-              divergence is inevitable one can use binary_partition group.
+ * **NOTE** :
+ *    binary_partition results in splitting warp into divergent thread groups
+ *    this is not good from performance perspective, but in cases where warp
+ *    divergence is inevitable one can use binary_partition group.
 */
 
 #include <stdio.h>
@@ -48,108 +50,110 @@
 
 namespace cg = cooperative_groups;
 
-void initOddEvenArr(int *inputArr, unsigned int size)
-{
-    for (int i=0; i < size; i++)
-    {
-        inputArr[i] = rand() % 50;
-    }
+void initOddEvenArr(int *inputArr, unsigned int size) {
+  for (int i = 0; i < size; i++) {
+    inputArr[i] = rand() % 50;
+  }
 }
 
-
 /**
  * CUDA kernel device code
- * 
+ *
  * Creates cooperative groups and performs odd/even counting & summation.
  */
-__global__ void oddEvenCountAndSumCG(int *inputArr, int *numOfOdds, int *sumOfOddAndEvens, unsigned int size)
-{
-    cg::thread_block cta = cg::this_thread_block();
-    cg::grid_group grid = cg::this_grid();
-    cg::thread_block_tile<32> tile32 = cg::tiled_partition<32>(cta);
+__global__ void oddEvenCountAndSumCG(int *inputArr, int *numOfOdds,
+                                     int *sumOfOddAndEvens, unsigned int size) {
+  cg::thread_block cta = cg::this_thread_block();
+  cg::grid_group grid = cg::this_grid();
+  cg::thread_block_tile<32> tile32 = cg::tiled_partition<32>(cta);
 
-    for (int i = grid.thread_rank(); i < size; i += grid.size())
+  for (int i = grid.thread_rank(); i < size; i += grid.size()) {
+    int elem = inputArr[i];
+    auto subTile = cg::binary_partition(tile32, elem & 1);
+    if (elem & 1)  // Odd numbers group
     {
-        int elem = inputArr[i];
-        auto subTile = cg::binary_partition(tile32, elem & 1);
-        if (elem & 1) // Odd numbers group
-        {
-            int oddGroupSum = cg::reduce(subTile, elem, cg::plus<int>());
+      int oddGroupSum = cg::reduce(subTile, elem, cg::plus<int>());
 
-            if (subTile.thread_rank() == 0)
-            {
-                // Add number of odds present in this group of Odds.
-                atomicAdd(numOfOdds, subTile.size());
+      if (subTile.thread_rank() == 0) {
+        // Add number of odds present in this group of Odds.
+        atomicAdd(numOfOdds, subTile.size());
 
-                // Add local reduction of odds present in this group of Odds.
-                atomicAdd(&sumOfOddAndEvens[0], oddGroupSum);
+        // Add local reduction of odds present in this group of Odds.
+        atomicAdd(&sumOfOddAndEvens[0], oddGroupSum);
+      }
+    } else  // Even numbers group
+    {
+      int evenGroupSum = cg::reduce(subTile, elem, cg::plus<int>());
 
-            }
-        }
-        else // Even numbers group
-        {
-            int evenGroupSum = cg::reduce(subTile, elem, cg::plus<int>());
-
-            if (subTile.thread_rank() == 0)
-            {
-                // Add local reduction of even present in this group of evens.
-                atomicAdd(&sumOfOddAndEvens[1], evenGroupSum);
-            }
-        }
-        // reconverge warp so for next loop iteration we ensure convergence of 
-        // above diverged threads to perform coalesced loads of inputArr.
-        cg::sync(tile32);
+      if (subTile.thread_rank() == 0) {
+        // Add local reduction of even present in this group of evens.
+        atomicAdd(&sumOfOddAndEvens[1], evenGroupSum);
+      }
     }
+    // reconverge warp so for next loop iteration we ensure convergence of
+    // above diverged threads to perform coalesced loads of inputArr.
+    cg::sync(tile32);
+  }
 }
 
-
 /**
  * Host main routine
  */
-int main(int argc, const char **argv)
-{
-    int deviceId = findCudaDevice(argc, argv);
-    int *h_inputArr, *d_inputArr;
-    int *h_numOfOdds, *d_numOfOdds;
-    int *h_sumOfOddEvenElems, *d_sumOfOddEvenElems;
-    unsigned int arrSize = 1024 * 100;
+int main(int argc, const char **argv) {
+  int deviceId = findCudaDevice(argc, argv);
+  int *h_inputArr, *d_inputArr;
+  int *h_numOfOdds, *d_numOfOdds;
+  int *h_sumOfOddEvenElems, *d_sumOfOddEvenElems;
+  unsigned int arrSize = 1024 * 100;
 
-    h_inputArr = new int[arrSize];
-    h_numOfOdds = new int[1];
-    h_sumOfOddEvenElems = new int[2];
-    initOddEvenArr(h_inputArr, arrSize);
-   
-    cudaStream_t stream;
-    checkCudaErrors(cudaStreamCreateWithFlags(&stream, cudaStreamNonBlocking));
-    checkCudaErrors(cudaMalloc(&d_inputArr, sizeof(int)*arrSize));
-    checkCudaErrors(cudaMalloc(&d_numOfOdds, sizeof(int)));
-    checkCudaErrors(cudaMalloc(&d_sumOfOddEvenElems, sizeof(int)*2));
+  checkCudaErrors(cudaMallocHost(&h_inputArr, sizeof(int) * arrSize));
+  checkCudaErrors(cudaMallocHost(&h_numOfOdds, sizeof(int)));
+  checkCudaErrors(cudaMallocHost(&h_sumOfOddEvenElems, sizeof(int) * 2));
+  initOddEvenArr(h_inputArr, arrSize);
 
-    checkCudaErrors(cudaMemcpyAsync(d_inputArr, h_inputArr, sizeof(int)*arrSize, cudaMemcpyHostToDevice, stream));
-    checkCudaErrors(cudaMemsetAsync(d_numOfOdds, 0, sizeof(int), stream));
-    checkCudaErrors(cudaMemsetAsync(d_sumOfOddEvenElems, 0, 2*sizeof(int), stream));
+  cudaStream_t stream;
+  checkCudaErrors(cudaStreamCreateWithFlags(&stream, cudaStreamNonBlocking));
+  checkCudaErrors(cudaMalloc(&d_inputArr, sizeof(int) * arrSize));
+  checkCudaErrors(cudaMalloc(&d_numOfOdds, sizeof(int)));
+  checkCudaErrors(cudaMalloc(&d_sumOfOddEvenElems, sizeof(int) * 2));
 
-    //Launch the kernel
-    int threadsPerBlock=1024;
-    int blocksPerGrid = arrSize / threadsPerBlock;
+  checkCudaErrors(cudaMemcpyAsync(d_inputArr, h_inputArr, sizeof(int) * arrSize,
+                                  cudaMemcpyHostToDevice, stream));
+  checkCudaErrors(cudaMemsetAsync(d_numOfOdds, 0, sizeof(int), stream));
+  checkCudaErrors(
+      cudaMemsetAsync(d_sumOfOddEvenElems, 0, 2 * sizeof(int), stream));
 
-    printf("\nLaunching %d blocks with %d threads...\n\n",blocksPerGrid, threadsPerBlock);
+  // Launch the kernel
+  int threadsPerBlock = 0;
+  int blocksPerGrid = 0;
+  checkCudaErrors(cudaOccupancyMaxPotentialBlockSize(
+      &blocksPerGrid, &threadsPerBlock, oddEvenCountAndSumCG, 0, 0));
 
-    oddEvenCountAndSumCG<<<blocksPerGrid, threadsPerBlock, 0, stream>>>(d_inputArr, d_numOfOdds, d_sumOfOddEvenElems, arrSize);
+  printf("\nLaunching %d blocks with %d threads...\n\n", blocksPerGrid,
+         threadsPerBlock);
 
-    checkCudaErrors(cudaMemcpyAsync(h_numOfOdds, d_numOfOdds, sizeof(int), cudaMemcpyDeviceToHost, stream));
-    checkCudaErrors(cudaMemcpyAsync(h_sumOfOddEvenElems, d_sumOfOddEvenElems, 2*sizeof(int), cudaMemcpyDeviceToHost, stream));
-    
-    printf("Array size = %d Num of Odds = %d Sum of Odds = %d Sum of Evens %d\n", arrSize, h_numOfOdds[0], h_sumOfOddEvenElems[0], h_sumOfOddEvenElems[1]);
-    printf("\n...Done.\n\n");
+  oddEvenCountAndSumCG<<<blocksPerGrid, threadsPerBlock, 0, stream>>>(
+      d_inputArr, d_numOfOdds, d_sumOfOddEvenElems, arrSize);
 
-    delete[] h_inputArr;
-    delete[] h_numOfOdds;
-    delete[] h_sumOfOddEvenElems;
+  checkCudaErrors(cudaMemcpyAsync(h_numOfOdds, d_numOfOdds, sizeof(int),
+                                  cudaMemcpyDeviceToHost, stream));
+  checkCudaErrors(cudaMemcpyAsync(h_sumOfOddEvenElems, d_sumOfOddEvenElems,
+                                  2 * sizeof(int), cudaMemcpyDeviceToHost,
+                                  stream));
+  checkCudaErrors(cudaStreamSynchronize(stream));
 
-    checkCudaErrors(cudaFree(d_inputArr));
-    checkCudaErrors(cudaFree(d_numOfOdds));
-    checkCudaErrors(cudaFree(d_sumOfOddEvenElems));
+  printf("Array size = %d Num of Odds = %d Sum of Odds = %d Sum of Evens %d\n",
+         arrSize, h_numOfOdds[0], h_sumOfOddEvenElems[0],
+         h_sumOfOddEvenElems[1]);
+  printf("\n...Done.\n\n");
 
-    return EXIT_SUCCESS;
+  checkCudaErrors(cudaFreeHost(h_inputArr));
+  checkCudaErrors(cudaFreeHost(h_numOfOdds));
+  checkCudaErrors(cudaFreeHost(h_sumOfOddEvenElems));
+
+  checkCudaErrors(cudaFree(d_inputArr));
+  checkCudaErrors(cudaFree(d_numOfOdds));
+  checkCudaErrors(cudaFree(d_sumOfOddEvenElems));
+
+  return EXIT_SUCCESS;
 }

Samples/binaryPartitionCG/binaryPartitionCG_vs2017.vcxproj

@@ -38,7 +38,7 @@
   </PropertyGroup>
   <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
   <ImportGroup Label="ExtensionSettings">
-    <Import Project="$(CUDAPropsPath)\CUDA 11.2.props" />
+    <Import Project="$(CUDAPropsPath)\CUDA 11.3.props" />
   </ImportGroup>
   <ImportGroup Label="PropertySheets">
     <Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
@@ -108,6 +108,6 @@
   </ItemGroup>
   <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
   <ImportGroup Label="ExtensionTargets">
-    <Import Project="$(CUDAPropsPath)\CUDA 11.2.targets" />
+    <Import Project="$(CUDAPropsPath)\CUDA 11.3.targets" />
   </ImportGroup>
 </Project>

Samples/binaryPartitionCG/binaryPartitionCG_vs2019.vcxproj

@@ -34,7 +34,7 @@
   </PropertyGroup>
   <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
   <ImportGroup Label="ExtensionSettings">
-    <Import Project="$(CUDAPropsPath)\CUDA 11.2.props" />
+    <Import Project="$(CUDAPropsPath)\CUDA 11.3.props" />
   </ImportGroup>
   <ImportGroup Label="PropertySheets">
     <Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
@@ -104,6 +104,6 @@
   </ItemGroup>
   <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
   <ImportGroup Label="ExtensionTargets">
-    <Import Project="$(CUDAPropsPath)\CUDA 11.2.targets" />
+    <Import Project="$(CUDAPropsPath)\CUDA 11.3.targets" />
   </ImportGroup>
 </Project>