conjugateGradientMultiDeviceCG.cu

/* Copyright (c) 2022, 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.
 */

/*
 * This sample implements a conjugate gradient solver on multiple GPU using
 * Unified Memory optimized prefetching and usage hints.
 *
 */

// includes, system
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <map>
#include <iostream>
#include <set>
#include <utility>

#include <cuda_runtime.h>

// Utilities and system includes
#include <helper_cuda.h>  // helper function CUDA error checking and initialization
#include <helper_functions.h>  // helper for shared functions common to CUDA Samples

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

namespace cg = cooperative_groups;

const char *sSDKname = "conjugateGradientMultiDeviceCG";

#define ENABLE_CPU_DEBUG_CODE 0
#define THREADS_PER_BLOCK 512

__device__ double grid_dot_result = 0.0;

/* genTridiag: generate a random tridiagonal symmetric matrix */
void genTridiag(int *I, int *J, float *val, int N, int nz) {
  I[0] = 0, J[0] = 0, J[1] = 1;
  val[0] = (float)rand() / RAND_MAX + 10.0f;
  val[1] = (float)rand() / RAND_MAX;
  int start;

  for (int i = 1; i < N; i++) {
    if (i > 1) {
      I[i] = I[i - 1] + 3;
    } else {
      I[1] = 2;
    }

    start = (i - 1) * 3 + 2;
    J[start] = i - 1;
    J[start + 1] = i;

    if (i < N - 1) {
      J[start + 2] = i + 1;
    }

    val[start] = val[start - 1];
    val[start + 1] = (float)rand() / RAND_MAX + 10.0f;

    if (i < N - 1) {
      val[start + 2] = (float)rand() / RAND_MAX;
    }
  }

  I[N] = nz;
}

// I - contains location of the given non-zero element in the row of the matrix
// J - contains location of the given non-zero element in the column of the
// matrix val - contains values of the given non-zero elements of the matrix
// inputVecX - input vector to be multiplied
// outputVecY - resultant vector
void cpuSpMV(int *I, int *J, float *val, int nnz, int num_rows, float alpha,
             float *inputVecX, float *outputVecY) {
  for (int i = 0; i < num_rows; i++) {
    int num_elems_this_row = I[i + 1] - I[i];

    float output = 0.0;
    for (int j = 0; j < num_elems_this_row; j++) {
      output += alpha * val[I[i] + j] * inputVecX[J[I[i] + j]];
    }
    outputVecY[i] = output;
  }

  return;
}

float dotProduct(float *vecA, float *vecB, int size) {
  float result = 0.0;

  for (int i = 0; i < size; i++) {
    result = result + (vecA[i] * vecB[i]);
  }

  return result;
}

void scaleVector(float *vec, float alpha, int size) {
  for (int i = 0; i < size; i++) {
    vec[i] = alpha * vec[i];
  }
}

void saxpy(float *x, float *y, float a, int size) {
  for (int i = 0; i < size; i++) {
    y[i] = a * x[i] + y[i];
  }
}

void cpuConjugateGrad(int *I, int *J, float *val, float *x, float *Ax, float *p,
                      float *r, int nnz, int N, float tol) {
  int max_iter = 10000;

  float alpha = 1.0;
  float alpham1 = -1.0;
  float r0 = 0.0, b, a, na;

  cpuSpMV(I, J, val, nnz, N, alpha, x, Ax);
  saxpy(Ax, r, alpham1, N);

  float r1 = dotProduct(r, r, N);

  int k = 1;

  while (r1 > tol * tol && k <= max_iter) {
    if (k > 1) {
      b = r1 / r0;
      scaleVector(p, b, N);

      saxpy(r, p, alpha, N);
    } else {
      for (int i = 0; i < N; i++) p[i] = r[i];
    }

    cpuSpMV(I, J, val, nnz, N, alpha, p, Ax);

    float dot = dotProduct(p, Ax, N);
    a = r1 / dot;

    saxpy(p, x, a, N);
    na = -a;
    saxpy(Ax, r, na, N);

    r0 = r1;
    r1 = dotProduct(r, r, N);

    printf("\nCPU code iteration = %3d, residual = %e\n", k, sqrt(r1));
    k++;
  }
}

// Data filled on CPU needed for MultiGPU operations.
struct MultiDeviceData {
  unsigned char *hostMemoryArrivedList;
  unsigned int numDevices;
  unsigned int deviceRank;
};

// Class used for coordination of multiple devices.
class PeerGroup {
  const MultiDeviceData &data;
  const cg::grid_group &grid;

  __device__ unsigned char load_arrived(unsigned char *arrived) const {
#if __CUDA_ARCH__ < 700
    return *(volatile unsigned char *)arrived;
#else
    unsigned int result;
    asm volatile("ld.acquire.sys.global.u8 %0, [%1];"
                 : "=r"(result)
                 : "l"(arrived)
                 : "memory");
    return result;
#endif
  }

  __device__ void store_arrived(unsigned char *arrived,
                                unsigned char val) const {
#if __CUDA_ARCH__ < 700
    *(volatile unsigned char *)arrived = val;
#else
    unsigned int reg_val = val;
    asm volatile(
        "st.release.sys.global.u8 [%1], %0;" ::"r"(reg_val) "l"(arrived)
        : "memory");

    // Avoids compiler warnings from unused variable val.
    (void)(reg_val = reg_val);
#endif
  }

 public:
  __device__ PeerGroup(const MultiDeviceData &data, const cg::grid_group &grid)
      : data(data), grid(grid){};

  __device__ unsigned int size() const { return data.numDevices * grid.size(); }

  __device__ unsigned int thread_rank() const {
    return data.deviceRank * grid.size() + grid.thread_rank();
  }

  __device__ void sync() const {
    grid.sync();

    // One thread from each grid participates in the sync.
    if (grid.thread_rank() == 0) {
      if (data.deviceRank == 0) {
        // Leader grid waits for others to join and then releases them.
        // Other GPUs can arrive in any order, so the leader have to wait for
        // all others.
        for (int i = 0; i < data.numDevices - 1; i++) {
          while (load_arrived(&data.hostMemoryArrivedList[i]) == 0)
            ;
        }
        for (int i = 0; i < data.numDevices - 1; i++) {
          store_arrived(&data.hostMemoryArrivedList[i], 0);
        }
        __threadfence_system();
      } else {
        // Other grids note their arrival and wait to be released.
        store_arrived(&data.hostMemoryArrivedList[data.deviceRank - 1], 1);
        while (load_arrived(&data.hostMemoryArrivedList[data.deviceRank - 1]) ==
               1)
          ;
      }
    }

    grid.sync();
  }
};

__device__ void gpuSpMV(int *I, int *J, float *val, int nnz, int num_rows,
                        float alpha, float *inputVecX, float *outputVecY,
                        const PeerGroup &peer_group) {
  for (int i = peer_group.thread_rank(); i < num_rows; i += peer_group.size()) {
    int row_elem = I[i];
    int next_row_elem = I[i + 1];
    int num_elems_this_row = next_row_elem - row_elem;

    float output = 0.0;
    for (int j = 0; j < num_elems_this_row; j++) {
      output += alpha * val[row_elem + j] * inputVecX[J[row_elem + j]];
    }

    outputVecY[i] = output;
  }
}

__device__ void gpuSaxpy(float *x, float *y, float a, int size,
                         const PeerGroup &peer_group) {
  for (int i = peer_group.thread_rank(); i < size; i += peer_group.size()) {
    y[i] = a * x[i] + y[i];
  }
}

__device__ void gpuDotProduct(float *vecA, float *vecB, int size,
                              const cg::thread_block &cta,
                              const PeerGroup &peer_group) {
  extern __shared__ double tmp[];

  double temp_sum = 0.0;

  for (int i = peer_group.thread_rank(); i < size; i += peer_group.size()) {
    temp_sum += (double)(vecA[i] * vecB[i]);
  }

  cg::thread_block_tile<32> tile32 = cg::tiled_partition<32>(cta);

  temp_sum = cg::reduce(tile32, temp_sum, cg::plus<double>());

  if (tile32.thread_rank() == 0) {
    tmp[tile32.meta_group_rank()] = temp_sum;
  }

  cg::sync(cta);

  if (tile32.meta_group_rank() == 0) {
    temp_sum = tile32.thread_rank() < tile32.meta_group_size()
                   ? tmp[tile32.thread_rank()]
                   : 0.0;
    temp_sum = cg::reduce(tile32, temp_sum, cg::plus<double>());

    if (tile32.thread_rank() == 0) {
      atomicAdd(&grid_dot_result, temp_sum);
    }
  }
}

__device__ void gpuCopyVector(float *srcA, float *destB, int size,
                              const PeerGroup &peer_group) {
  for (int i = peer_group.thread_rank(); i < size; i += peer_group.size()) {
    destB[i] = srcA[i];
  }
}

__device__ void gpuScaleVectorAndSaxpy(float *x, float *y, float a, float scale,
                                       int size, const PeerGroup &peer_group) {
  for (int i = peer_group.thread_rank(); i < size; i += peer_group.size()) {
    y[i] = a * x[i] + scale * y[i];
  }
}

extern "C" __global__ void multiGpuConjugateGradient(
    int *I, int *J, float *val, float *x, float *Ax, float *p, float *r,
    double *dot_result, int nnz, int N, float tol,
    MultiDeviceData multi_device_data) {
  cg::thread_block cta = cg::this_thread_block();
  cg::grid_group grid = cg::this_grid();
  PeerGroup peer_group(multi_device_data, grid);

  const int max_iter = 10000;

  float alpha = 1.0;
  float alpham1 = -1.0;
  float r0 = 0.0, r1, b, a, na;

  for (int i = peer_group.thread_rank(); i < N; i += peer_group.size()) {
    r[i] = 1.0;
    x[i] = 0.0;
  }

  cg::sync(grid);

  gpuSpMV(I, J, val, nnz, N, alpha, x, Ax, peer_group);

  cg::sync(grid);

  gpuSaxpy(Ax, r, alpham1, N, peer_group);

  cg::sync(grid);

  gpuDotProduct(r, r, N, cta, peer_group);

  cg::sync(grid);

  if (grid.thread_rank() == 0) {
    atomicAdd_system(dot_result, grid_dot_result);
    grid_dot_result = 0.0;
  }
  peer_group.sync();

  r1 = *dot_result;

  int k = 1;
  while (r1 > tol * tol && k <= max_iter) {
    if (k > 1) {
      b = r1 / r0;
      gpuScaleVectorAndSaxpy(r, p, alpha, b, N, peer_group);
    } else {
      gpuCopyVector(r, p, N, peer_group);
    }

    peer_group.sync();

    gpuSpMV(I, J, val, nnz, N, alpha, p, Ax, peer_group);

    if (peer_group.thread_rank() == 0) {
      *dot_result = 0.0;
    }
    peer_group.sync();

    gpuDotProduct(p, Ax, N, cta, peer_group);

    cg::sync(grid);

    if (grid.thread_rank() == 0) {
      atomicAdd_system(dot_result, grid_dot_result);
      grid_dot_result = 0.0;
    }
    peer_group.sync();

    a = r1 / *dot_result;

    gpuSaxpy(p, x, a, N, peer_group);

    na = -a;

    gpuSaxpy(Ax, r, na, N, peer_group);

    r0 = r1;

    peer_group.sync();

    if (peer_group.thread_rank() == 0) {
      *dot_result = 0.0;
    }

    peer_group.sync();

    gpuDotProduct(r, r, N, cta, peer_group);

    cg::sync(grid);

    if (grid.thread_rank() == 0) {
      atomicAdd_system(dot_result, grid_dot_result);
      grid_dot_result = 0.0;
    }
    peer_group.sync();

    r1 = *dot_result;
    k++;
  }
}

// Map of device version to device number
std::multimap<std::pair<int, int>, int> getIdenticalGPUs() {
  int numGpus = 0;
  checkCudaErrors(cudaGetDeviceCount(&numGpus));

  std::multimap<std::pair<int, int>, int> identicalGpus;

  for (int i = 0; i < numGpus; i++) {
    cudaDeviceProp deviceProp;
    checkCudaErrors(cudaGetDeviceProperties(&deviceProp, i));

    // Filter unsupported devices
    if (deviceProp.cooperativeLaunch && deviceProp.concurrentManagedAccess) {
      identicalGpus.emplace(std::make_pair(deviceProp.major, deviceProp.minor),
                            i);
    }
    printf("GPU Device %d: \"%s\" with compute capability %d.%d\n", i,
           deviceProp.name, deviceProp.major, deviceProp.minor);
  }

  return identicalGpus;
}

int main(int argc, char **argv) {
  constexpr size_t kNumGpusRequired = 2;
  int N = 0, nz = 0, *I = NULL, *J = NULL;
  float *val = NULL;
  const float tol = 1e-5f;
  float *x;
  float rhs = 1.0;
  float r1;
  float *r, *p, *Ax;

  printf("Starting [%s]...\n", sSDKname);
  auto gpusByArch = getIdenticalGPUs();

  auto it = gpusByArch.begin();
  auto end = gpusByArch.end();

  auto bestFit = std::make_pair(it, it);
  // use std::distance to find the largest number of GPUs amongst architectures
  auto distance = [](decltype(bestFit) p) {
    return std::distance(p.first, p.second);
  };

  // Read each unique key/pair element in order
  for (; it != end; it = gpusByArch.upper_bound(it->first)) {
    // first and second are iterators bounded within the architecture group
    auto testFit = gpusByArch.equal_range(it->first);
    // Always use devices with highest architecture version or whichever has the
    // most devices available
    if (distance(bestFit) <= distance(testFit)) bestFit = testFit;
  }

  if (distance(bestFit) < kNumGpusRequired) {
    printf(
        "No two or more GPUs with same architecture capable of "
        "concurrentManagedAccess found. "
        "\nWaiving the sample\n");
    exit(EXIT_WAIVED);
  }

  std::set<int> bestFitDeviceIds;

  // Check & select peer-to-peer access capable GPU devices as enabling p2p
  // access between participating GPUs gives better performance.
  for (auto itr = bestFit.first; itr != bestFit.second; itr++) {
    int deviceId = itr->second;
    checkCudaErrors(cudaSetDevice(deviceId));

    std::for_each(
        itr, bestFit.second,
        [&deviceId, &bestFitDeviceIds,
         &kNumGpusRequired](decltype(*itr) mapPair) {
          if (deviceId != mapPair.second) {
            int access = 0;
            checkCudaErrors(
                cudaDeviceCanAccessPeer(&access, deviceId, mapPair.second));
            printf("Device=%d %s Access Peer Device=%d\n", deviceId,
                   access ? "CAN" : "CANNOT", mapPair.second);
            if (access && bestFitDeviceIds.size() < kNumGpusRequired) {
              bestFitDeviceIds.emplace(deviceId);
              bestFitDeviceIds.emplace(mapPair.second);
            } else {
              printf("Ignoring device %i (max devices exceeded)\n",
                     mapPair.second);
            }
          }
        });

    if (bestFitDeviceIds.size() >= kNumGpusRequired) {
      printf("Selected p2p capable devices - ");
      for (auto devicesItr = bestFitDeviceIds.begin();
           devicesItr != bestFitDeviceIds.end(); devicesItr++) {
        printf("deviceId = %d  ", *devicesItr);
      }
      printf("\n");
      break;
    }
  }

  // if bestFitDeviceIds.size() == 0 it means the GPUs in system are not p2p
  // capable, hence we add it without p2p capability check.
  if (!bestFitDeviceIds.size()) {
    printf("Devices involved are not p2p capable.. selecting %zu of them\n",
           kNumGpusRequired);
    std::for_each(bestFit.first, bestFit.second,
                  [&bestFitDeviceIds,
                   &kNumGpusRequired](decltype(*bestFit.first) mapPair) {
                    if (bestFitDeviceIds.size() < kNumGpusRequired) {
                      bestFitDeviceIds.emplace(mapPair.second);
                    } else {
                      printf("Ignoring device %i (max devices exceeded)\n",
                             mapPair.second);
                    }
                    // Insert the sequence into the deviceIds set
                  });
  } else {
    // perform cudaDeviceEnablePeerAccess in both directions for all
    // participating devices.
    for (auto p1_itr = bestFitDeviceIds.begin();
         p1_itr != bestFitDeviceIds.end(); p1_itr++) {
      checkCudaErrors(cudaSetDevice(*p1_itr));
      for (auto p2_itr = bestFitDeviceIds.begin();
           p2_itr != bestFitDeviceIds.end(); p2_itr++) {
        if (*p1_itr != *p2_itr) {
          checkCudaErrors(cudaDeviceEnablePeerAccess(*p2_itr, 0));
          checkCudaErrors(cudaSetDevice(*p1_itr));
        }
      }
    }
  }

  /* Generate a random tridiagonal symmetric matrix in CSR format */
  N = 10485760 * 2;
  nz = (N - 2) * 3 + 4;

  checkCudaErrors(cudaMallocManaged((void **)&I, sizeof(int) * (N + 1)));
  checkCudaErrors(cudaMallocManaged((void **)&J, sizeof(int) * nz));
  checkCudaErrors(cudaMallocManaged((void **)&val, sizeof(float) * nz));

  float *val_cpu = (float *)malloc(sizeof(float) * nz);

  genTridiag(I, J, val_cpu, N, nz);

  memcpy(val, val_cpu, sizeof(float) * nz);
  checkCudaErrors(
      cudaMemAdvise(I, sizeof(int) * (N + 1), cudaMemAdviseSetReadMostly, 0));
  checkCudaErrors(
      cudaMemAdvise(J, sizeof(int) * nz, cudaMemAdviseSetReadMostly, 0));
  checkCudaErrors(
      cudaMemAdvise(val, sizeof(float) * nz, cudaMemAdviseSetReadMostly, 0));

  checkCudaErrors(cudaMallocManaged((void **)&x, sizeof(float) * N));

  double *dot_result;
  checkCudaErrors(cudaMallocManaged((void **)&dot_result, sizeof(double)));

  checkCudaErrors(cudaMemset(dot_result, 0, sizeof(double)));

  // temp memory for ConjugateGradient
  checkCudaErrors(cudaMallocManaged((void **)&r, N * sizeof(float)));
  checkCudaErrors(cudaMallocManaged((void **)&p, N * sizeof(float)));
  checkCudaErrors(cudaMallocManaged((void **)&Ax, N * sizeof(float)));

  std::cout << "\nRunning on GPUs = " << kNumGpusRequired << std::endl;
  cudaStream_t nStreams[kNumGpusRequired];

  int sMemSize = sizeof(double) * ((THREADS_PER_BLOCK / 32) + 1);
  int numBlocksPerSm = INT_MAX;
  int numThreads = THREADS_PER_BLOCK;
  int numSms = INT_MAX;
  auto deviceId = bestFitDeviceIds.begin();

  // set numSms & numBlocksPerSm to be lowest of 2 devices
  while (deviceId != bestFitDeviceIds.end()) {
    cudaDeviceProp deviceProp;
    checkCudaErrors(cudaSetDevice(*deviceId));
    checkCudaErrors(cudaGetDeviceProperties(&deviceProp, *deviceId));

    int numBlocksPerSm_current = 0;
    checkCudaErrors(cudaOccupancyMaxActiveBlocksPerMultiprocessor(
        &numBlocksPerSm_current, multiGpuConjugateGradient, numThreads,
        sMemSize));

    if (numBlocksPerSm > numBlocksPerSm_current) {
      numBlocksPerSm = numBlocksPerSm_current;
    }
    if (numSms > deviceProp.multiProcessorCount) {
      numSms = deviceProp.multiProcessorCount;
    }
    deviceId++;
  }

  if (!numBlocksPerSm) {
    printf(
        "Max active blocks per SM is returned as 0.\n Hence, Waiving the "
        "sample\n");
    exit(EXIT_WAIVED);
  }

  int device_count = 0;
  int totalThreadsPerGPU = numSms * numBlocksPerSm * THREADS_PER_BLOCK;
  deviceId = bestFitDeviceIds.begin();
  while (deviceId != bestFitDeviceIds.end()) {
    checkCudaErrors(cudaSetDevice(*deviceId));
    checkCudaErrors(cudaStreamCreate(&nStreams[device_count]));

    int perGPUIter = N / (totalThreadsPerGPU * kNumGpusRequired);
    int offset_Ax = device_count * totalThreadsPerGPU;
    int offset_r = device_count * totalThreadsPerGPU;
    int offset_p = device_count * totalThreadsPerGPU;
    int offset_x = device_count * totalThreadsPerGPU;

    checkCudaErrors(cudaMemPrefetchAsync(I, sizeof(int) * N, *deviceId,
                                         nStreams[device_count]));
    checkCudaErrors(cudaMemPrefetchAsync(val, sizeof(float) * nz, *deviceId,
                                         nStreams[device_count]));
    checkCudaErrors(cudaMemPrefetchAsync(J, sizeof(float) * nz, *deviceId,
                                         nStreams[device_count]));

    if (offset_Ax <= N) {
      for (int i = 0; i < perGPUIter; i++) {
        cudaMemAdvise(Ax + offset_Ax, sizeof(float) * totalThreadsPerGPU,
                      cudaMemAdviseSetPreferredLocation, *deviceId);
        cudaMemAdvise(r + offset_r, sizeof(float) * totalThreadsPerGPU,
                      cudaMemAdviseSetPreferredLocation, *deviceId);
        cudaMemAdvise(x + offset_x, sizeof(float) * totalThreadsPerGPU,
                      cudaMemAdviseSetPreferredLocation, *deviceId);
        cudaMemAdvise(p + offset_p, sizeof(float) * totalThreadsPerGPU,
                      cudaMemAdviseSetPreferredLocation, *deviceId);

        cudaMemAdvise(Ax + offset_Ax, sizeof(float) * totalThreadsPerGPU,
                      cudaMemAdviseSetAccessedBy, *deviceId);
        cudaMemAdvise(r + offset_r, sizeof(float) * totalThreadsPerGPU,
                      cudaMemAdviseSetAccessedBy, *deviceId);
        cudaMemAdvise(p + offset_p, sizeof(float) * totalThreadsPerGPU,
                      cudaMemAdviseSetAccessedBy, *deviceId);
        cudaMemAdvise(x + offset_x, sizeof(float) * totalThreadsPerGPU,
                      cudaMemAdviseSetAccessedBy, *deviceId);

        offset_Ax += totalThreadsPerGPU * kNumGpusRequired;
        offset_r += totalThreadsPerGPU * kNumGpusRequired;
        offset_p += totalThreadsPerGPU * kNumGpusRequired;
        offset_x += totalThreadsPerGPU * kNumGpusRequired;

        if (offset_Ax >= N) {
          break;
        }
      }
    }

    device_count++;
    deviceId++;
  }

#if ENABLE_CPU_DEBUG_CODE
  float *Ax_cpu = (float *)malloc(sizeof(float) * N);
  float *r_cpu = (float *)malloc(sizeof(float) * N);
  float *p_cpu = (float *)malloc(sizeof(float) * N);
  float *x_cpu = (float *)malloc(sizeof(float) * N);

  for (int i = 0; i < N; i++) {
    r_cpu[i] = 1.0;
    Ax_cpu[i] = x_cpu[i] = 0.0;
  }
#endif

  printf("Total threads per GPU = %d numBlocksPerSm  = %d\n",
         numSms * numBlocksPerSm * THREADS_PER_BLOCK, numBlocksPerSm);
  dim3 dimGrid(numSms * numBlocksPerSm, 1, 1),
      dimBlock(THREADS_PER_BLOCK, 1, 1);

  // Structure used for cross-grid synchronization.
  MultiDeviceData multi_device_data;
  checkCudaErrors(cudaHostAlloc(
      &multi_device_data.hostMemoryArrivedList,
      (kNumGpusRequired - 1) * sizeof(*multi_device_data.hostMemoryArrivedList),
      cudaHostAllocPortable));
  memset(multi_device_data.hostMemoryArrivedList, 0,
         (kNumGpusRequired - 1) *
             sizeof(*multi_device_data.hostMemoryArrivedList));
  multi_device_data.numDevices = kNumGpusRequired;
  multi_device_data.deviceRank = 0;

  void *kernelArgs[] = {
      (void *)&I,  (void *)&J, (void *)&val, (void *)&x,
      (void *)&Ax, (void *)&p, (void *)&r,   (void *)&dot_result,
      (void *)&nz, (void *)&N, (void *)&tol, (void *)&multi_device_data,
  };

  printf("Launching kernel\n");

  deviceId = bestFitDeviceIds.begin();
  device_count = 0;
  while (deviceId != bestFitDeviceIds.end()) {
    checkCudaErrors(cudaSetDevice(*deviceId));
    checkCudaErrors(cudaLaunchCooperativeKernel(
        (void *)multiGpuConjugateGradient, dimGrid, dimBlock, kernelArgs,
        sMemSize, nStreams[device_count++]));
    multi_device_data.deviceRank++;
    deviceId++;
  }

  checkCudaErrors(cudaMemPrefetchAsync(x, sizeof(float) * N, cudaCpuDeviceId));
  checkCudaErrors(
      cudaMemPrefetchAsync(dot_result, sizeof(double), cudaCpuDeviceId));

  deviceId = bestFitDeviceIds.begin();
  device_count = 0;
  while (deviceId != bestFitDeviceIds.end()) {
    checkCudaErrors(cudaSetDevice(*deviceId));
    checkCudaErrors(cudaStreamSynchronize(nStreams[device_count++]));
    deviceId++;
  }

  r1 = (float)*dot_result;

  printf("GPU Final, residual = %e \n  ", sqrt(r1));

#if ENABLE_CPU_DEBUG_CODE
  cpuConjugateGrad(I, J, val, x_cpu, Ax_cpu, p_cpu, r_cpu, nz, N, tol);
#endif

  float rsum, diff, err = 0.0;

  for (int i = 0; i < N; i++) {
    rsum = 0.0;

    for (int j = I[i]; j < I[i + 1]; j++) {
      rsum += val_cpu[j] * x[J[j]];
    }

    diff = fabs(rsum - rhs);

    if (diff > err) {
      err = diff;
    }
  }

  checkCudaErrors(cudaFreeHost(multi_device_data.hostMemoryArrivedList));
  checkCudaErrors(cudaFree(I));
  checkCudaErrors(cudaFree(J));
  checkCudaErrors(cudaFree(val));
  checkCudaErrors(cudaFree(x));
  checkCudaErrors(cudaFree(r));
  checkCudaErrors(cudaFree(p));
  checkCudaErrors(cudaFree(Ax));
  checkCudaErrors(cudaFree(dot_result));
  free(val_cpu);

#if ENABLE_CPU_DEBUG_CODE
  free(Ax_cpu);
  free(r_cpu);
  free(p_cpu);
  free(x_cpu);
#endif

  printf("Test Summary:  Error amount = %f \n", err);
  fprintf(stdout, "&&&& conjugateGradientMultiDeviceCG %s\n",
          (sqrt(r1) < tol) ? "PASSED" : "FAILED");
  exit((sqrt(r1) < tol) ? EXIT_SUCCESS : EXIT_FAILURE);
}