cuda-samples/Samples/reductionMultiBlockCG/reductionMultiBlockCG.cu

/* Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions
 * are met:
 *  * Redistributions of source code must retain the above copyright
 *    notice, this list of conditions and the following disclaimer.
 *  * Redistributions in binary form must reproduce the above copyright
 *    notice, this list of conditions and the following disclaimer in the
 *    documentation and/or other materials provided with the distribution.
 *  * Neither the name of NVIDIA CORPORATION nor the names of its
 *    contributors may be used to endorse or promote products derived
 *    from this software without specific prior written permission.
 *
 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
 * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT OWNER OR
 * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
 * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
 * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
 * OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 */

/*
  Parallel reduction

  This sample shows how to perform a reduction operation on an array of values
  to produce a single value in a single kernel (as opposed to two or more
  kernel calls as shown in the "reduction" CUDA Sample).  Single-pass
  reduction requires Cooperative Groups.

  Reductions are a very common computation in parallel algorithms.  Any time
  an array of values needs to be reduced to a single value using a binary
  associative operator, a reduction can be used.  Example applications include
  statistics computations such as mean and standard deviation, and image
  processing applications such as finding the total luminance of an
  image.

  This code performs sum reductions, but any associative operator such as
  min() or max() could also be used.

  It assumes the input size is a power of 2.

  COMMAND LINE ARGUMENTS

  "--n=<N>"         :Specify the number of elements to reduce (default 33554432)
  "--threads=<N>"   :Specify the number of threads per block (default 128)
  "--maxblocks=<N>" :Specify the maximum number of thread blocks to launch
 (kernel 6 only, default 64)
*/

// includes, system
#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <math.h>

// includes, project
#include <helper_functions.h>
#include <helper_cuda.h>

#include <cuda_runtime.h>

const char *sSDKsample = "reductionMultiBlockCG";

#include <cuda_runtime_api.h>
#include <cooperative_groups.h>
#include <cooperative_groups/reduce.h>

namespace cg = cooperative_groups;

/*
  Parallel sum reduction using shared memory
  - takes log(n) steps for n input elements
  - uses n/2 threads
  - only works for power-of-2 arrays

  This version adds multiple elements per thread sequentially. This reduces the
  overall cost of the algorithm while keeping the work complexity O(n) and the
  step complexity O(log n).
  (Brent's Theorem optimization)

  See the CUDA SDK "reduction" sample for more information.
*/

__device__ void reduceBlock(double *sdata, const cg::thread_block &cta) {
  const unsigned int tid = cta.thread_rank();
  cg::thread_block_tile<32> tile32 = cg::tiled_partition<32>(cta);

  sdata[tid] = cg::reduce(tile32, sdata[tid], cg::plus<double>());
  cg::sync(cta);

  double beta = 0.0;
  if (cta.thread_rank() == 0) {
    beta = 0;
    for (int i = 0; i < blockDim.x; i += tile32.size()) {
      beta += sdata[i];
    }
    sdata[0] = beta;
  }
  cg::sync(cta);
}

// This reduction kernel reduces an arbitrary size array in a single kernel
// invocation
//
// For more details on the reduction algorithm (notably the multi-pass
// approach), see the "reduction" sample in the CUDA SDK.
extern "C" __global__ void reduceSinglePassMultiBlockCG(const float *g_idata,
                                                        float *g_odata,
                                                        unsigned int n) {
  // Handle to thread block group
  cg::thread_block block = cg::this_thread_block();
  cg::grid_group grid = cg::this_grid();

  extern double __shared__ sdata[];

  // Stride over grid and add the values to a shared memory buffer
  sdata[block.thread_rank()] = 0;

  for (int i = grid.thread_rank(); i < n; i += grid.size()) {
    sdata[block.thread_rank()] += g_idata[i];
  }

  cg::sync(block);

  // Reduce each block (called once per block)
  reduceBlock(sdata, block);
  // Write out the result to global memory
  if (block.thread_rank() == 0) {
    g_odata[blockIdx.x] = sdata[0];
  }
  cg::sync(grid);

  if (grid.thread_rank() == 0) {
    for (int block = 1; block < gridDim.x; block++) {
      g_odata[0] += g_odata[block];
    }
  }
}

////////////////////////////////////////////////////////////////////////////////
// Wrapper function for kernel launch
////////////////////////////////////////////////////////////////////////////////
void call_reduceSinglePassMultiBlockCG(int size, int threads, int numBlocks,
                                       float *d_idata, float *d_odata) {
  int smemSize = threads * sizeof(double);
  void *kernelArgs[] = {
      (void *)&d_idata, (void *)&d_odata, (void *)&size,
  };

  dim3 dimBlock(threads, 1, 1);
  dim3 dimGrid(numBlocks, 1, 1);

  cudaLaunchCooperativeKernel((void *)reduceSinglePassMultiBlockCG, dimGrid,
                              dimBlock, kernelArgs, smemSize, NULL);
  // check if kernel execution generated an error
  getLastCudaError("Kernel execution failed");
}

////////////////////////////////////////////////////////////////////////////////
// declaration, forward
bool runTest(int argc, char **argv, int device);

////////////////////////////////////////////////////////////////////////////////
// Program main
////////////////////////////////////////////////////////////////////////////////
int main(int argc, char **argv) {
  cudaDeviceProp deviceProp = {0};
  int dev;

  printf("%s Starting...\n\n", sSDKsample);

  dev = findCudaDevice(argc, (const char **)argv);
  checkCudaErrors(cudaGetDeviceProperties(&deviceProp, dev));
  if (!deviceProp.cooperativeLaunch) {
    printf(
        "\nSelected GPU (%d) does not support Cooperative Kernel Launch, "
        "Waiving the run\n",
        dev);
    exit(EXIT_WAIVED);
  }

  bool bTestPassed = false;
  bTestPassed = runTest(argc, argv, dev);

  exit(bTestPassed ? EXIT_SUCCESS : EXIT_FAILURE);
}

////////////////////////////////////////////////////////////////////////////////
//! Compute sum reduction on CPU
//! We use Kahan summation for an accurate sum of large arrays.
//! http://en.wikipedia.org/wiki/Kahan_summation_algorithm
//!
//! @param data       pointer to input data
//! @param size       number of input data elements
////////////////////////////////////////////////////////////////////////////////
template <class T>
T reduceCPU(T *data, int size) {
  T sum = data[0];
  T c = (T)0.0;

  for (int i = 1; i < size; i++) {
    T y = data[i] - c;
    T t = sum + y;
    c = (t - sum) - y;
    sum = t;
  }

  return sum;
}

unsigned int nextPow2(unsigned int x) {
  --x;
  x |= x >> 1;
  x |= x >> 2;
  x |= x >> 4;
  x |= x >> 8;
  x |= x >> 16;
  return ++x;
}

////////////////////////////////////////////////////////////////////////////////
// Compute the number of threads and blocks to use for the reduction
// We set threads / block to the minimum of maxThreads and n/2.
////////////////////////////////////////////////////////////////////////////////
void getNumBlocksAndThreads(int n, int maxBlocks, int maxThreads, int &blocks,
                            int &threads) {
  if (n == 1) {
    threads = 1;
    blocks = 1;
  } else {
    threads = (n < maxThreads * 2) ? nextPow2(n / 2) : maxThreads;
    blocks = max(1, n / (threads * 2));
  }

  blocks = min(maxBlocks, blocks);
}

////////////////////////////////////////////////////////////////////////////////
// This function performs a reduction of the input data multiple times and
// measures the average reduction time.
////////////////////////////////////////////////////////////////////////////////
float benchmarkReduce(int n, int numThreads, int numBlocks, int maxThreads,
                      int maxBlocks, int testIterations,
                      StopWatchInterface *timer, float *h_odata, float *d_idata,
                      float *d_odata) {
  float gpu_result = 0;
  cudaError_t error;

  printf("\nLaunching %s kernel\n",
         "SinglePass Multi Block Cooperative Groups");
  for (int i = 0; i < testIterations; ++i) {
    gpu_result = 0;
    sdkStartTimer(&timer);
    call_reduceSinglePassMultiBlockCG(n, numThreads, numBlocks, d_idata,
                                      d_odata);
    cudaDeviceSynchronize();
    sdkStopTimer(&timer);
  }

  // copy final sum from device to host
  error =
      cudaMemcpy(&gpu_result, d_odata, sizeof(float), cudaMemcpyDeviceToHost);
  checkCudaErrors(error);

  return gpu_result;
}

////////////////////////////////////////////////////////////////////////////////
// The main function which runs the reduction test.
////////////////////////////////////////////////////////////////////////////////
bool runTest(int argc, char **argv, int device) {
  int size = 1 << 25;  // number of elements to reduce
  bool bTestPassed = false;

  if (checkCmdLineFlag(argc, (const char **)argv, "n")) {
    size = getCmdLineArgumentInt(argc, (const char **)argv, "n");
  }

  printf("%d elements\n", size);

  // Set the device to be used
  cudaDeviceProp prop = {0};
  checkCudaErrors(cudaSetDevice(device));
  checkCudaErrors(cudaGetDeviceProperties(&prop, device));

  // create random input data on CPU
  unsigned int bytes = size * sizeof(float);

  float *h_idata = (float *)malloc(bytes);

  for (int i = 0; i < size; i++) {
    // Keep the numbers small so we don't get truncation error in the sum
    h_idata[i] = (rand() & 0xFF) / (float)RAND_MAX;
  }

  // Determine the launch configuration (threads, blocks)
  int maxThreads = 0;
  int maxBlocks = 0;

  if (checkCmdLineFlag(argc, (const char **)argv, "threads")) {
    maxThreads = getCmdLineArgumentInt(argc, (const char **)argv, "threads");
  } else {
    maxThreads = prop.maxThreadsPerBlock;
  }

  if (checkCmdLineFlag(argc, (const char **)argv, "maxblocks")) {
    maxBlocks = getCmdLineArgumentInt(argc, (const char **)argv, "maxblocks");
  } else {
    maxBlocks = prop.multiProcessorCount *
                (prop.maxThreadsPerMultiProcessor / prop.maxThreadsPerBlock);
  }

  int numBlocks = 0;
  int numThreads = 0;
  getNumBlocksAndThreads(size, maxBlocks, maxThreads, numBlocks, numThreads);

  // We calculate the occupancy to know how many block can actually fit on the
  // GPU
  int numBlocksPerSm = 0;
  checkCudaErrors(cudaOccupancyMaxActiveBlocksPerMultiprocessor(
      &numBlocksPerSm, reduceSinglePassMultiBlockCG, numThreads,
      numThreads * sizeof(double)));

  int numSms = prop.multiProcessorCount;
  if (numBlocks > numBlocksPerSm * numSms) {
    numBlocks = numBlocksPerSm * numSms;
  }
  printf("numThreads: %d\n", numThreads);
  printf("numBlocks: %d\n", numBlocks);

  // allocate mem for the result on host side
  float *h_odata = (float *)malloc(numBlocks * sizeof(float));

  // allocate device memory and data
  float *d_idata = NULL;
  float *d_odata = NULL;

  checkCudaErrors(cudaMalloc((void **)&d_idata, bytes));
  checkCudaErrors(cudaMalloc((void **)&d_odata, numBlocks * sizeof(float)));

  // copy data directly to device memory
  checkCudaErrors(cudaMemcpy(d_idata, h_idata, bytes, cudaMemcpyHostToDevice));
  checkCudaErrors(cudaMemcpy(d_odata, h_idata, numBlocks * sizeof(float),
                             cudaMemcpyHostToDevice));

  int testIterations = 100;

  StopWatchInterface *timer = 0;
  sdkCreateTimer(&timer);

  float gpu_result = 0;

  gpu_result =
      benchmarkReduce(size, numThreads, numBlocks, maxThreads, maxBlocks,
                      testIterations, timer, h_odata, d_idata, d_odata);

  float reduceTime = sdkGetAverageTimerValue(&timer);
  printf("Average time: %f ms\n", reduceTime);
  printf("Bandwidth:    %f GB/s\n\n",
         (size * sizeof(int)) / (reduceTime * 1.0e6));

  // compute reference solution
  float cpu_result = reduceCPU<float>(h_idata, size);
  printf("GPU result = %0.12f\n", gpu_result);
  printf("CPU result = %0.12f\n", cpu_result);

  double threshold = 1e-8 * size;
  double diff = abs((double)gpu_result - (double)cpu_result);
  bTestPassed = (diff < threshold);

  // cleanup
  sdkDeleteTimer(&timer);

  free(h_idata);
  free(h_odata);
  cudaFree(d_idata);
  cudaFree(d_odata);

  return bTestPassed;
}