mirror of
https://github.com/NVIDIA/cuda-samples.git
synced 2026-10-25 22:28:17 +08:00
Apply consistent code formatting across the repo. Add clang-format and pre-commit hooks.
This commit is contained in:
Rob Armstrong
@@ -48,28 +48,28 @@
|
||||
|
||||
COMMAND LINE ARGUMENTS
|
||||
|
||||
"--shmoo": Test performance for 1 to 32M elements with each of the
|
||||
"--shmoo": Test performance for 1 to 32M elements with each of the
|
||||
7 different kernels
|
||||
"--n=<N>": Specify the number of elements to reduce (default 1048576)
|
||||
"--threads=<N>": Specify the number of threads per block (default 128)
|
||||
"--maxblocks=<N>": Specify the maximum number of thread blocks to launch
|
||||
"--maxblocks=<N>": Specify the maximum number of thread blocks to launch
|
||||
(kernel 6 only, default 64)
|
||||
"--cpufinal": Read back the per-block results and do final sum of block
|
||||
"--cpufinal": Read back the per-block results and do final sum of block
|
||||
sums on CPU (default false)
|
||||
"--cputhresh=<N>": The threshold of number of blocks sums below which to
|
||||
"--cputhresh=<N>": The threshold of number of blocks sums below which to
|
||||
perform a CPU final reduction (default 1)
|
||||
"--multipass": Use a multipass reduction instead of a single-pass reduction
|
||||
*/
|
||||
|
||||
// includes, system
|
||||
#include <stdlib.h>
|
||||
#include <stdio.h>
|
||||
#include <string.h>
|
||||
#include <math.h>
|
||||
#include <stdio.h>
|
||||
#include <stdlib.h>
|
||||
#include <string.h>
|
||||
|
||||
// includes, project
|
||||
#include <helper_functions.h>
|
||||
#include <helper_cuda.h>
|
||||
#include <helper_functions.h>
|
||||
|
||||
#define VERSION_MAJOR (CUDART_VERSION / 1000)
|
||||
#define VERSION_MINOR (CUDART_VERSION % 100) / 10
|
||||
@@ -86,51 +86,52 @@ const char *sSDKsample = "threadFenceReduction";
|
||||
// declaration, forward
|
||||
bool runTest(int argc, char **argv);
|
||||
|
||||
extern "C" {
|
||||
void reduce(int size, int threads, int blocks, float *d_idata, float *d_odata);
|
||||
void reduceSinglePass(int size, int threads, int blocks, float *d_idata,
|
||||
float *d_odata);
|
||||
extern "C"
|
||||
{
|
||||
void reduce(int size, int threads, int blocks, float *d_idata, float *d_odata);
|
||||
void reduceSinglePass(int size, int threads, int blocks, float *d_idata, float *d_odata);
|
||||
}
|
||||
|
||||
#if CUDART_VERSION < 2020
|
||||
void reduce(int size, int threads, int blocks, float *d_idata, float *d_odata) {
|
||||
printf("reduce(), compiler not supported, aborting tests\n");
|
||||
void reduce(int size, int threads, int blocks, float *d_idata, float *d_odata)
|
||||
{
|
||||
printf("reduce(), compiler not supported, aborting tests\n");
|
||||
}
|
||||
|
||||
void reduceSinglePass(int size, int threads, int blocks, float *d_idata,
|
||||
float *d_odata) {
|
||||
printf("reduceSinglePass(), compiler not supported, aborting tests\n");
|
||||
void reduceSinglePass(int size, int threads, int blocks, float *d_idata, float *d_odata)
|
||||
{
|
||||
printf("reduceSinglePass(), compiler not supported, aborting tests\n");
|
||||
}
|
||||
#endif
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
// Program main
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
int main(int argc, char **argv) {
|
||||
cudaDeviceProp deviceProp;
|
||||
deviceProp.major = 0;
|
||||
deviceProp.minor = 0;
|
||||
int dev;
|
||||
int main(int argc, char **argv)
|
||||
{
|
||||
cudaDeviceProp deviceProp;
|
||||
deviceProp.major = 0;
|
||||
deviceProp.minor = 0;
|
||||
int dev;
|
||||
|
||||
printf("%s Starting...\n\n", sSDKsample);
|
||||
printf("%s Starting...\n\n", sSDKsample);
|
||||
|
||||
dev = findCudaDevice(argc, (const char **)argv);
|
||||
dev = findCudaDevice(argc, (const char **)argv);
|
||||
|
||||
checkCudaErrors(cudaGetDeviceProperties(&deviceProp, dev));
|
||||
checkCudaErrors(cudaGetDeviceProperties(&deviceProp, dev));
|
||||
|
||||
printf("GPU Device supports SM %d.%d compute capability\n\n",
|
||||
deviceProp.major, deviceProp.minor);
|
||||
printf("GPU Device supports SM %d.%d compute capability\n\n", deviceProp.major, deviceProp.minor);
|
||||
|
||||
bool bTestResult = false;
|
||||
bool bTestResult = false;
|
||||
|
||||
#if CUDART_VERSION >= 2020
|
||||
bTestResult = runTest(argc, argv);
|
||||
bTestResult = runTest(argc, argv);
|
||||
#else
|
||||
print_NVCC_min_spec(sSDKsample, "2.2", "Version 185");
|
||||
exit(EXIT_SUCCESS);
|
||||
print_NVCC_min_spec(sSDKsample, "2.2", "Version 185");
|
||||
exit(EXIT_SUCCESS);
|
||||
#endif
|
||||
|
||||
exit(bTestResult ? EXIT_SUCCESS : EXIT_FAILURE);
|
||||
exit(bTestResult ? EXIT_SUCCESS : EXIT_FAILURE);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
@@ -141,137 +142,147 @@ int main(int argc, char **argv) {
|
||||
//! @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;
|
||||
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;
|
||||
}
|
||||
for (int i = 1; i < size; i++) {
|
||||
T y = data[i] - c;
|
||||
T t = sum + y;
|
||||
c = (t - sum) - y;
|
||||
sum = t;
|
||||
}
|
||||
|
||||
return sum;
|
||||
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;
|
||||
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));
|
||||
}
|
||||
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);
|
||||
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, bool multiPass,
|
||||
bool cpuFinalReduction, int cpuFinalThreshold,
|
||||
StopWatchInterface *timer, float *h_odata, float *d_idata,
|
||||
float *d_odata) {
|
||||
float gpu_result = 0;
|
||||
bool bNeedReadback = true;
|
||||
cudaError_t error;
|
||||
float benchmarkReduce(int n,
|
||||
int numThreads,
|
||||
int numBlocks,
|
||||
int maxThreads,
|
||||
int maxBlocks,
|
||||
int testIterations,
|
||||
bool multiPass,
|
||||
bool cpuFinalReduction,
|
||||
int cpuFinalThreshold,
|
||||
StopWatchInterface *timer,
|
||||
float *h_odata,
|
||||
float *d_idata,
|
||||
float *d_odata)
|
||||
{
|
||||
float gpu_result = 0;
|
||||
bool bNeedReadback = true;
|
||||
cudaError_t error;
|
||||
|
||||
for (int i = 0; i < testIterations; ++i) {
|
||||
gpu_result = 0;
|
||||
unsigned int retCnt = 0;
|
||||
error = setRetirementCount(retCnt);
|
||||
checkCudaErrors(error);
|
||||
|
||||
cudaDeviceSynchronize();
|
||||
sdkStartTimer(&timer);
|
||||
|
||||
if (multiPass) {
|
||||
// execute the kernel
|
||||
reduce(n, numThreads, numBlocks, d_idata, d_odata);
|
||||
|
||||
// check if kernel execution generated an error
|
||||
getLastCudaError("Kernel execution failed");
|
||||
|
||||
if (cpuFinalReduction) {
|
||||
// sum partial sums from each block on CPU
|
||||
// copy result from device to host
|
||||
error = cudaMemcpy(h_odata, d_odata, numBlocks * sizeof(float),
|
||||
cudaMemcpyDeviceToHost);
|
||||
for (int i = 0; i < testIterations; ++i) {
|
||||
gpu_result = 0;
|
||||
unsigned int retCnt = 0;
|
||||
error = setRetirementCount(retCnt);
|
||||
checkCudaErrors(error);
|
||||
|
||||
for (int i = 0; i < numBlocks; i++) {
|
||||
gpu_result += h_odata[i];
|
||||
cudaDeviceSynchronize();
|
||||
sdkStartTimer(&timer);
|
||||
|
||||
if (multiPass) {
|
||||
// execute the kernel
|
||||
reduce(n, numThreads, numBlocks, d_idata, d_odata);
|
||||
|
||||
// check if kernel execution generated an error
|
||||
getLastCudaError("Kernel execution failed");
|
||||
|
||||
if (cpuFinalReduction) {
|
||||
// sum partial sums from each block on CPU
|
||||
// copy result from device to host
|
||||
error = cudaMemcpy(h_odata, d_odata, numBlocks * sizeof(float), cudaMemcpyDeviceToHost);
|
||||
checkCudaErrors(error);
|
||||
|
||||
for (int i = 0; i < numBlocks; i++) {
|
||||
gpu_result += h_odata[i];
|
||||
}
|
||||
|
||||
bNeedReadback = false;
|
||||
}
|
||||
else {
|
||||
// sum partial block sums on GPU
|
||||
int s = numBlocks;
|
||||
|
||||
while (s > cpuFinalThreshold) {
|
||||
int threads = 0, blocks = 0;
|
||||
getNumBlocksAndThreads(s, maxBlocks, maxThreads, blocks, threads);
|
||||
|
||||
reduce(s, threads, blocks, d_odata, d_odata);
|
||||
|
||||
s = s / (threads * 2);
|
||||
}
|
||||
|
||||
if (s > 1) {
|
||||
// copy result from device to host
|
||||
error = cudaMemcpy(h_odata, d_odata, s * sizeof(float), cudaMemcpyDeviceToHost);
|
||||
checkCudaErrors(error);
|
||||
|
||||
for (int i = 0; i < s; i++) {
|
||||
gpu_result += h_odata[i];
|
||||
}
|
||||
|
||||
bNeedReadback = false;
|
||||
}
|
||||
}
|
||||
}
|
||||
else {
|
||||
getLastCudaError("Kernel execution failed");
|
||||
|
||||
// execute the kernel
|
||||
reduceSinglePass(n, numThreads, numBlocks, d_idata, d_odata);
|
||||
|
||||
// check if kernel execution generated an error
|
||||
getLastCudaError("Kernel execution failed");
|
||||
}
|
||||
|
||||
bNeedReadback = false;
|
||||
} else {
|
||||
// sum partial block sums on GPU
|
||||
int s = numBlocks;
|
||||
|
||||
while (s > cpuFinalThreshold) {
|
||||
int threads = 0, blocks = 0;
|
||||
getNumBlocksAndThreads(s, maxBlocks, maxThreads, blocks, threads);
|
||||
|
||||
reduce(s, threads, blocks, d_odata, d_odata);
|
||||
|
||||
s = s / (threads * 2);
|
||||
}
|
||||
|
||||
if (s > 1) {
|
||||
// copy result from device to host
|
||||
error = cudaMemcpy(h_odata, d_odata, s * sizeof(float),
|
||||
cudaMemcpyDeviceToHost);
|
||||
checkCudaErrors(error);
|
||||
|
||||
for (int i = 0; i < s; i++) {
|
||||
gpu_result += h_odata[i];
|
||||
}
|
||||
|
||||
bNeedReadback = false;
|
||||
}
|
||||
}
|
||||
} else {
|
||||
getLastCudaError("Kernel execution failed");
|
||||
|
||||
// execute the kernel
|
||||
reduceSinglePass(n, numThreads, numBlocks, d_idata, d_odata);
|
||||
|
||||
// check if kernel execution generated an error
|
||||
getLastCudaError("Kernel execution failed");
|
||||
cudaDeviceSynchronize();
|
||||
sdkStopTimer(&timer);
|
||||
}
|
||||
|
||||
cudaDeviceSynchronize();
|
||||
sdkStopTimer(&timer);
|
||||
}
|
||||
if (bNeedReadback) {
|
||||
// copy final sum from device to host
|
||||
error = cudaMemcpy(&gpu_result, d_odata, sizeof(float), cudaMemcpyDeviceToHost);
|
||||
checkCudaErrors(error);
|
||||
}
|
||||
|
||||
if (bNeedReadback) {
|
||||
// copy final sum from device to host
|
||||
error =
|
||||
cudaMemcpy(&gpu_result, d_odata, sizeof(float), cudaMemcpyDeviceToHost);
|
||||
checkCudaErrors(error);
|
||||
}
|
||||
|
||||
return gpu_result;
|
||||
return gpu_result;
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
@@ -280,186 +291,203 @@ float benchmarkReduce(int n, int numThreads, int numBlocks, int maxThreads,
|
||||
// for generating a "shmoo" plot showing the performance for each kernel
|
||||
// variation over a wide range of input sizes.
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
void shmoo(int minN, int maxN, int maxThreads, int maxBlocks) {
|
||||
// create random input data on CPU
|
||||
unsigned int bytes = maxN * sizeof(float);
|
||||
|
||||
float *h_idata = (float *)malloc(bytes);
|
||||
|
||||
for (int i = 0; i < maxN; i++) {
|
||||
// Keep the numbers small so we don't get truncation error in the sum
|
||||
h_idata[i] = (rand() & 0xFF) / (float)RAND_MAX;
|
||||
}
|
||||
|
||||
int maxNumBlocks = min(65535, maxN / maxThreads);
|
||||
|
||||
// allocate mem for the result on host side
|
||||
float *h_odata = (float *)malloc(maxNumBlocks * 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, maxNumBlocks * sizeof(float)));
|
||||
|
||||
// copy data directly to device memory
|
||||
checkCudaErrors(cudaMemcpy(d_idata, h_idata, bytes, cudaMemcpyHostToDevice));
|
||||
checkCudaErrors(cudaMemcpy(d_odata, h_idata, maxNumBlocks * sizeof(float),
|
||||
cudaMemcpyHostToDevice));
|
||||
|
||||
// warm-up
|
||||
reduce(maxN, maxThreads, maxNumBlocks, d_idata, d_odata);
|
||||
int testIterations = 100;
|
||||
|
||||
StopWatchInterface *timer = NULL;
|
||||
sdkCreateTimer(&timer);
|
||||
|
||||
// print headers
|
||||
printf("N, %d blocks one pass, %d blocks multipass\n", maxBlocks, maxBlocks);
|
||||
|
||||
for (int i = minN; i <= maxN; i *= 2) {
|
||||
printf("%d, ", i);
|
||||
|
||||
for (int multiPass = 0; multiPass <= 1; multiPass++) {
|
||||
sdkResetTimer(&timer);
|
||||
int numBlocks = 0;
|
||||
int numThreads = 0;
|
||||
getNumBlocksAndThreads(i, maxBlocks, maxThreads, numBlocks, numThreads);
|
||||
|
||||
benchmarkReduce(i, numThreads, numBlocks, maxThreads, maxBlocks,
|
||||
testIterations, multiPass == 1, false, 1, timer, h_odata,
|
||||
d_idata, d_odata);
|
||||
|
||||
float reduceTime = sdkGetAverageTimerValue(&timer);
|
||||
printf("%f%s", reduceTime, multiPass == 0 ? ", " : "\n");
|
||||
}
|
||||
}
|
||||
|
||||
printf("\n");
|
||||
|
||||
// cleanup
|
||||
sdkDeleteTimer(&timer);
|
||||
free(h_idata);
|
||||
free(h_odata);
|
||||
|
||||
cudaFree(d_idata);
|
||||
cudaFree(d_odata);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
// The main function which runs the reduction test.
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
bool runTest(int argc, char **argv) {
|
||||
int size = 1 << 20; // number of elements to reduce
|
||||
int maxThreads = 128; // number of threads per block
|
||||
int maxBlocks = 64;
|
||||
bool cpuFinalReduction = false;
|
||||
int cpuFinalThreshold = 1;
|
||||
bool multipass = false;
|
||||
bool bTestResult = false;
|
||||
|
||||
if (checkCmdLineFlag(argc, (const char **)argv, "n")) {
|
||||
size = getCmdLineArgumentInt(argc, (const char **)argv, "n");
|
||||
}
|
||||
|
||||
if (checkCmdLineFlag(argc, (const char **)argv, "threads")) {
|
||||
maxThreads = getCmdLineArgumentInt(argc, (const char **)argv, "threads");
|
||||
}
|
||||
|
||||
if (checkCmdLineFlag(argc, (const char **)argv, "maxblocks")) {
|
||||
maxBlocks = getCmdLineArgumentInt(argc, (const char **)argv, "maxblocks");
|
||||
}
|
||||
|
||||
printf("%d elements\n", size);
|
||||
printf("%d threads (max)\n", maxThreads);
|
||||
|
||||
cpuFinalReduction = checkCmdLineFlag(argc, (const char **)argv, "cpufinal");
|
||||
multipass = checkCmdLineFlag(argc, (const char **)argv, "multipass");
|
||||
|
||||
if (checkCmdLineFlag(argc, (const char **)argv, "cputhresh")) {
|
||||
cpuFinalThreshold =
|
||||
getCmdLineArgumentInt(argc, (const char **)argv, "cputhresh");
|
||||
}
|
||||
|
||||
bool runShmoo = checkCmdLineFlag(argc, (const char **)argv, "shmoo");
|
||||
|
||||
if (runShmoo) {
|
||||
shmoo(1, 33554432, maxThreads, maxBlocks);
|
||||
} else {
|
||||
void shmoo(int minN, int maxN, int maxThreads, int maxBlocks)
|
||||
{
|
||||
// create random input data on CPU
|
||||
unsigned int bytes = size * sizeof(float);
|
||||
unsigned int bytes = maxN * 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;
|
||||
for (int i = 0; i < maxN; i++) {
|
||||
// Keep the numbers small so we don't get truncation error in the sum
|
||||
h_idata[i] = (rand() & 0xFF) / (float)RAND_MAX;
|
||||
}
|
||||
|
||||
int numBlocks = 0;
|
||||
int numThreads = 0;
|
||||
getNumBlocksAndThreads(size, maxBlocks, maxThreads, numBlocks, numThreads);
|
||||
|
||||
if (numBlocks == 1) {
|
||||
cpuFinalThreshold = 1;
|
||||
}
|
||||
int maxNumBlocks = min(65535, maxN / maxThreads);
|
||||
|
||||
// allocate mem for the result on host side
|
||||
float *h_odata = (float *)malloc(numBlocks * sizeof(float));
|
||||
|
||||
printf("%d blocks\n", numBlocks);
|
||||
float *h_odata = (float *)malloc(maxNumBlocks * 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)));
|
||||
checkCudaErrors(cudaMalloc((void **)&d_odata, maxNumBlocks * 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));
|
||||
checkCudaErrors(cudaMemcpy(d_idata, h_idata, bytes, cudaMemcpyHostToDevice));
|
||||
checkCudaErrors(cudaMemcpy(d_odata, h_idata, maxNumBlocks * sizeof(float), cudaMemcpyHostToDevice));
|
||||
|
||||
// warm-up
|
||||
reduce(size, numThreads, numBlocks, d_idata, d_odata);
|
||||
reduce(maxN, maxThreads, maxNumBlocks, d_idata, d_odata);
|
||||
int testIterations = 100;
|
||||
|
||||
StopWatchInterface *timer = 0;
|
||||
StopWatchInterface *timer = NULL;
|
||||
sdkCreateTimer(&timer);
|
||||
|
||||
float gpu_result = 0;
|
||||
// print headers
|
||||
printf("N, %d blocks one pass, %d blocks multipass\n", maxBlocks, maxBlocks);
|
||||
|
||||
gpu_result =
|
||||
benchmarkReduce(size, numThreads, numBlocks, maxThreads, maxBlocks,
|
||||
testIterations, multipass, cpuFinalReduction,
|
||||
cpuFinalThreshold, timer, h_odata, d_idata, d_odata);
|
||||
for (int i = minN; i <= maxN; i *= 2) {
|
||||
printf("%d, ", i);
|
||||
|
||||
float reduceTime = sdkGetAverageTimerValue(&timer);
|
||||
printf("Average time: %f ms\n", reduceTime);
|
||||
printf("Bandwidth: %f GB/s\n\n",
|
||||
(size * sizeof(int)) / (reduceTime * 1.0e6));
|
||||
for (int multiPass = 0; multiPass <= 1; multiPass++) {
|
||||
sdkResetTimer(&timer);
|
||||
int numBlocks = 0;
|
||||
int numThreads = 0;
|
||||
getNumBlocksAndThreads(i, maxBlocks, maxThreads, numBlocks, numThreads);
|
||||
|
||||
// compute reference solution
|
||||
float cpu_result = reduceCPU<float>(h_idata, size);
|
||||
benchmarkReduce(i,
|
||||
numThreads,
|
||||
numBlocks,
|
||||
maxThreads,
|
||||
maxBlocks,
|
||||
testIterations,
|
||||
multiPass == 1,
|
||||
false,
|
||||
1,
|
||||
timer,
|
||||
h_odata,
|
||||
d_idata,
|
||||
d_odata);
|
||||
|
||||
printf("GPU result = %0.12f\n", gpu_result);
|
||||
printf("CPU result = %0.12f\n", cpu_result);
|
||||
float reduceTime = sdkGetAverageTimerValue(&timer);
|
||||
printf("%f%s", reduceTime, multiPass == 0 ? ", " : "\n");
|
||||
}
|
||||
}
|
||||
|
||||
double threshold = 1e-8 * size;
|
||||
double diff = abs((double)gpu_result - (double)cpu_result);
|
||||
bTestResult = (diff < threshold);
|
||||
printf("\n");
|
||||
|
||||
// cleanup
|
||||
sdkDeleteTimer(&timer);
|
||||
|
||||
free(h_idata);
|
||||
free(h_odata);
|
||||
|
||||
cudaFree(d_idata);
|
||||
cudaFree(d_odata);
|
||||
}
|
||||
|
||||
return bTestResult;
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
// The main function which runs the reduction test.
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
bool runTest(int argc, char **argv)
|
||||
{
|
||||
int size = 1 << 20; // number of elements to reduce
|
||||
int maxThreads = 128; // number of threads per block
|
||||
int maxBlocks = 64;
|
||||
bool cpuFinalReduction = false;
|
||||
int cpuFinalThreshold = 1;
|
||||
bool multipass = false;
|
||||
bool bTestResult = false;
|
||||
|
||||
if (checkCmdLineFlag(argc, (const char **)argv, "n")) {
|
||||
size = getCmdLineArgumentInt(argc, (const char **)argv, "n");
|
||||
}
|
||||
|
||||
if (checkCmdLineFlag(argc, (const char **)argv, "threads")) {
|
||||
maxThreads = getCmdLineArgumentInt(argc, (const char **)argv, "threads");
|
||||
}
|
||||
|
||||
if (checkCmdLineFlag(argc, (const char **)argv, "maxblocks")) {
|
||||
maxBlocks = getCmdLineArgumentInt(argc, (const char **)argv, "maxblocks");
|
||||
}
|
||||
|
||||
printf("%d elements\n", size);
|
||||
printf("%d threads (max)\n", maxThreads);
|
||||
|
||||
cpuFinalReduction = checkCmdLineFlag(argc, (const char **)argv, "cpufinal");
|
||||
multipass = checkCmdLineFlag(argc, (const char **)argv, "multipass");
|
||||
|
||||
if (checkCmdLineFlag(argc, (const char **)argv, "cputhresh")) {
|
||||
cpuFinalThreshold = getCmdLineArgumentInt(argc, (const char **)argv, "cputhresh");
|
||||
}
|
||||
|
||||
bool runShmoo = checkCmdLineFlag(argc, (const char **)argv, "shmoo");
|
||||
|
||||
if (runShmoo) {
|
||||
shmoo(1, 33554432, maxThreads, maxBlocks);
|
||||
}
|
||||
else {
|
||||
// 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;
|
||||
}
|
||||
|
||||
int numBlocks = 0;
|
||||
int numThreads = 0;
|
||||
getNumBlocksAndThreads(size, maxBlocks, maxThreads, numBlocks, numThreads);
|
||||
|
||||
if (numBlocks == 1) {
|
||||
cpuFinalThreshold = 1;
|
||||
}
|
||||
|
||||
// allocate mem for the result on host side
|
||||
float *h_odata = (float *)malloc(numBlocks * sizeof(float));
|
||||
|
||||
printf("%d blocks\n", numBlocks);
|
||||
|
||||
// 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));
|
||||
|
||||
// warm-up
|
||||
reduce(size, numThreads, numBlocks, d_idata, d_odata);
|
||||
int testIterations = 100;
|
||||
|
||||
StopWatchInterface *timer = 0;
|
||||
sdkCreateTimer(&timer);
|
||||
|
||||
float gpu_result = 0;
|
||||
|
||||
gpu_result = benchmarkReduce(size,
|
||||
numThreads,
|
||||
numBlocks,
|
||||
maxThreads,
|
||||
maxBlocks,
|
||||
testIterations,
|
||||
multipass,
|
||||
cpuFinalReduction,
|
||||
cpuFinalThreshold,
|
||||
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);
|
||||
bTestResult = (diff < threshold);
|
||||
|
||||
// cleanup
|
||||
sdkDeleteTimer(&timer);
|
||||
|
||||
free(h_idata);
|
||||
free(h_odata);
|
||||
cudaFree(d_idata);
|
||||
cudaFree(d_odata);
|
||||
}
|
||||
|
||||
return bTestResult;
|
||||
}
|
||||
|
||||
@@ -32,8 +32,8 @@
|
||||
#ifndef _REDUCE_KERNEL_H_
|
||||
#define _REDUCE_KERNEL_H_
|
||||
|
||||
#include <cuda_runtime_api.h>
|
||||
#include <cooperative_groups.h>
|
||||
#include <cuda_runtime_api.h>
|
||||
|
||||
namespace cg = cooperative_groups;
|
||||
|
||||
@@ -51,85 +51,87 @@ namespace cg = cooperative_groups;
|
||||
*/
|
||||
|
||||
template <unsigned int blockSize>
|
||||
__device__ void reduceBlock(volatile float *sdata, float mySum,
|
||||
const unsigned int tid, cg::thread_block cta) {
|
||||
cg::thread_block_tile<32> tile32 = cg::tiled_partition<32>(cta);
|
||||
sdata[tid] = mySum;
|
||||
cg::sync(tile32);
|
||||
|
||||
const int VEC = 32;
|
||||
const int vid = tid & (VEC - 1);
|
||||
|
||||
float beta = mySum;
|
||||
float temp;
|
||||
|
||||
for (int i = VEC / 2; i > 0; i >>= 1) {
|
||||
if (vid < i) {
|
||||
temp = sdata[tid + i];
|
||||
beta += temp;
|
||||
sdata[tid] = beta;
|
||||
}
|
||||
__device__ void reduceBlock(volatile float *sdata, float mySum, const unsigned int tid, cg::thread_block cta)
|
||||
{
|
||||
cg::thread_block_tile<32> tile32 = cg::tiled_partition<32>(cta);
|
||||
sdata[tid] = mySum;
|
||||
cg::sync(tile32);
|
||||
}
|
||||
cg::sync(cta);
|
||||
|
||||
if (cta.thread_rank() == 0) {
|
||||
beta = 0;
|
||||
for (int i = 0; i < blockDim.x; i += VEC) {
|
||||
beta += sdata[i];
|
||||
const int VEC = 32;
|
||||
const int vid = tid & (VEC - 1);
|
||||
|
||||
float beta = mySum;
|
||||
float temp;
|
||||
|
||||
for (int i = VEC / 2; i > 0; i >>= 1) {
|
||||
if (vid < i) {
|
||||
temp = sdata[tid + i];
|
||||
beta += temp;
|
||||
sdata[tid] = beta;
|
||||
}
|
||||
cg::sync(tile32);
|
||||
}
|
||||
sdata[0] = beta;
|
||||
}
|
||||
cg::sync(cta);
|
||||
cg::sync(cta);
|
||||
|
||||
if (cta.thread_rank() == 0) {
|
||||
beta = 0;
|
||||
for (int i = 0; i < blockDim.x; i += VEC) {
|
||||
beta += sdata[i];
|
||||
}
|
||||
sdata[0] = beta;
|
||||
}
|
||||
cg::sync(cta);
|
||||
}
|
||||
|
||||
template <unsigned int blockSize, bool nIsPow2>
|
||||
__device__ void reduceBlocks(const float *g_idata, float *g_odata,
|
||||
unsigned int n, cg::thread_block cta) {
|
||||
extern __shared__ float sdata[];
|
||||
__device__ void reduceBlocks(const float *g_idata, float *g_odata, unsigned int n, cg::thread_block cta)
|
||||
{
|
||||
extern __shared__ float sdata[];
|
||||
|
||||
// perform first level of reduction,
|
||||
// reading from global memory, writing to shared memory
|
||||
unsigned int tid = threadIdx.x;
|
||||
unsigned int i = blockIdx.x * (blockSize * 2) + threadIdx.x;
|
||||
unsigned int gridSize = blockSize * 2 * gridDim.x;
|
||||
float mySum = 0;
|
||||
// perform first level of reduction,
|
||||
// reading from global memory, writing to shared memory
|
||||
unsigned int tid = threadIdx.x;
|
||||
unsigned int i = blockIdx.x * (blockSize * 2) + threadIdx.x;
|
||||
unsigned int gridSize = blockSize * 2 * gridDim.x;
|
||||
float mySum = 0;
|
||||
|
||||
// we reduce multiple elements per thread. The number is determined by the
|
||||
// number of active thread blocks (via gridDim). More blocks will result
|
||||
// in a larger gridSize and therefore fewer elements per thread
|
||||
while (i < n) {
|
||||
mySum += g_idata[i];
|
||||
// we reduce multiple elements per thread. The number is determined by the
|
||||
// number of active thread blocks (via gridDim). More blocks will result
|
||||
// in a larger gridSize and therefore fewer elements per thread
|
||||
while (i < n) {
|
||||
mySum += g_idata[i];
|
||||
|
||||
// ensure we don't read out of bounds -- this is optimized away for powerOf2
|
||||
// sized arrays
|
||||
if (nIsPow2 || i + blockSize < n) mySum += g_idata[i + blockSize];
|
||||
// ensure we don't read out of bounds -- this is optimized away for powerOf2
|
||||
// sized arrays
|
||||
if (nIsPow2 || i + blockSize < n)
|
||||
mySum += g_idata[i + blockSize];
|
||||
|
||||
i += gridSize;
|
||||
}
|
||||
i += gridSize;
|
||||
}
|
||||
|
||||
// do reduction in shared mem
|
||||
reduceBlock<blockSize>(sdata, mySum, tid, cta);
|
||||
// do reduction in shared mem
|
||||
reduceBlock<blockSize>(sdata, mySum, tid, cta);
|
||||
|
||||
// write result for this block to global mem
|
||||
if (tid == 0) g_odata[blockIdx.x] = sdata[0];
|
||||
// write result for this block to global mem
|
||||
if (tid == 0)
|
||||
g_odata[blockIdx.x] = sdata[0];
|
||||
}
|
||||
|
||||
template <unsigned int blockSize, bool nIsPow2>
|
||||
__global__ void reduceMultiPass(const float *g_idata, float *g_odata,
|
||||
unsigned int n) {
|
||||
// Handle to thread block group
|
||||
cg::thread_block cta = cg::this_thread_block();
|
||||
reduceBlocks<blockSize, nIsPow2>(g_idata, g_odata, n, cta);
|
||||
__global__ void reduceMultiPass(const float *g_idata, float *g_odata, unsigned int n)
|
||||
{
|
||||
// Handle to thread block group
|
||||
cg::thread_block cta = cg::this_thread_block();
|
||||
reduceBlocks<blockSize, nIsPow2>(g_idata, g_odata, n, cta);
|
||||
}
|
||||
|
||||
// Global variable used by reduceSinglePass to count how many blocks have
|
||||
// finished
|
||||
__device__ unsigned int retirementCount = 0;
|
||||
|
||||
cudaError_t setRetirementCount(int retCnt) {
|
||||
return cudaMemcpyToSymbol(retirementCount, &retCnt, sizeof(unsigned int), 0,
|
||||
cudaMemcpyHostToDevice);
|
||||
cudaError_t setRetirementCount(int retCnt)
|
||||
{
|
||||
return cudaMemcpyToSymbol(retirementCount, &retCnt, sizeof(unsigned int), 0, cudaMemcpyHostToDevice);
|
||||
}
|
||||
|
||||
// This reduction kernel reduces an arbitrary size array in a single kernel
|
||||
@@ -148,59 +150,59 @@ cudaError_t setRetirementCount(int retCnt) {
|
||||
// For more details on the reduction algorithm (notably the multi-pass
|
||||
// approach), see the "reduction" sample in the CUDA SDK.
|
||||
template <unsigned int blockSize, bool nIsPow2>
|
||||
__global__ void reduceSinglePass(const float *g_idata, float *g_odata,
|
||||
unsigned int n) {
|
||||
// Handle to thread block group
|
||||
cg::thread_block cta = cg::this_thread_block();
|
||||
//
|
||||
// PHASE 1: Process all inputs assigned to this block
|
||||
//
|
||||
__global__ void reduceSinglePass(const float *g_idata, float *g_odata, unsigned int n)
|
||||
{
|
||||
// Handle to thread block group
|
||||
cg::thread_block cta = cg::this_thread_block();
|
||||
//
|
||||
// PHASE 1: Process all inputs assigned to this block
|
||||
//
|
||||
|
||||
reduceBlocks<blockSize, nIsPow2>(g_idata, g_odata, n, cta);
|
||||
reduceBlocks<blockSize, nIsPow2>(g_idata, g_odata, n, cta);
|
||||
|
||||
//
|
||||
// PHASE 2: Last block finished will process all partial sums
|
||||
//
|
||||
//
|
||||
// PHASE 2: Last block finished will process all partial sums
|
||||
//
|
||||
|
||||
if (gridDim.x > 1) {
|
||||
const unsigned int tid = threadIdx.x;
|
||||
__shared__ bool amLast;
|
||||
extern float __shared__ smem[];
|
||||
if (gridDim.x > 1) {
|
||||
const unsigned int tid = threadIdx.x;
|
||||
__shared__ bool amLast;
|
||||
extern float __shared__ smem[];
|
||||
|
||||
// wait until all outstanding memory instructions in this thread are
|
||||
// finished
|
||||
__threadfence();
|
||||
// wait until all outstanding memory instructions in this thread are
|
||||
// finished
|
||||
__threadfence();
|
||||
|
||||
// Thread 0 takes a ticket
|
||||
if (tid == 0) {
|
||||
unsigned int ticket = atomicInc(&retirementCount, gridDim.x);
|
||||
// If the ticket ID is equal to the number of blocks, we are the last
|
||||
// block!
|
||||
amLast = (ticket == gridDim.x - 1);
|
||||
// Thread 0 takes a ticket
|
||||
if (tid == 0) {
|
||||
unsigned int ticket = atomicInc(&retirementCount, gridDim.x);
|
||||
// If the ticket ID is equal to the number of blocks, we are the last
|
||||
// block!
|
||||
amLast = (ticket == gridDim.x - 1);
|
||||
}
|
||||
|
||||
cg::sync(cta);
|
||||
|
||||
// The last block sums the results of all other blocks
|
||||
if (amLast) {
|
||||
int i = tid;
|
||||
float mySum = 0;
|
||||
|
||||
while (i < gridDim.x) {
|
||||
mySum += g_odata[i];
|
||||
i += blockSize;
|
||||
}
|
||||
|
||||
reduceBlock<blockSize>(smem, mySum, tid, cta);
|
||||
|
||||
if (tid == 0) {
|
||||
g_odata[0] = smem[0];
|
||||
|
||||
// reset retirement count so that next run succeeds
|
||||
retirementCount = 0;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
cg::sync(cta);
|
||||
|
||||
// The last block sums the results of all other blocks
|
||||
if (amLast) {
|
||||
int i = tid;
|
||||
float mySum = 0;
|
||||
|
||||
while (i < gridDim.x) {
|
||||
mySum += g_odata[i];
|
||||
i += blockSize;
|
||||
}
|
||||
|
||||
reduceBlock<blockSize>(smem, mySum, tid, cta);
|
||||
|
||||
if (tid == 0) {
|
||||
g_odata[0] = smem[0];
|
||||
|
||||
// reset retirement count so that next run succeeds
|
||||
retirementCount = 0;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
bool isPow2(unsigned int x) { return ((x & (x - 1)) == 0); }
|
||||
@@ -208,233 +210,194 @@ bool isPow2(unsigned int x) { return ((x & (x - 1)) == 0); }
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
// Wrapper function for kernel launch
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
extern "C" void reduce(int size, int threads, int blocks, float *d_idata,
|
||||
float *d_odata) {
|
||||
dim3 dimBlock(threads, 1, 1);
|
||||
dim3 dimGrid(blocks, 1, 1);
|
||||
int smemSize =
|
||||
(threads <= 32) ? 2 * threads * sizeof(float) : threads * sizeof(float);
|
||||
extern "C" void reduce(int size, int threads, int blocks, float *d_idata, float *d_odata)
|
||||
{
|
||||
dim3 dimBlock(threads, 1, 1);
|
||||
dim3 dimGrid(blocks, 1, 1);
|
||||
int smemSize = (threads <= 32) ? 2 * threads * sizeof(float) : threads * sizeof(float);
|
||||
|
||||
// choose which of the optimized versions of reduction to launch
|
||||
if (isPow2(size)) {
|
||||
switch (threads) {
|
||||
case 512:
|
||||
reduceMultiPass<512, true>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
// choose which of the optimized versions of reduction to launch
|
||||
if (isPow2(size)) {
|
||||
switch (threads) {
|
||||
case 512:
|
||||
reduceMultiPass<512, true><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
|
||||
case 256:
|
||||
reduceMultiPass<256, true>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
case 256:
|
||||
reduceMultiPass<256, true><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
|
||||
case 128:
|
||||
reduceMultiPass<128, true>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
case 128:
|
||||
reduceMultiPass<128, true><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
|
||||
case 64:
|
||||
reduceMultiPass<64, true>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
case 64:
|
||||
reduceMultiPass<64, true><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
|
||||
case 32:
|
||||
reduceMultiPass<32, true>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
case 32:
|
||||
reduceMultiPass<32, true><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
|
||||
case 16:
|
||||
reduceMultiPass<16, true>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
case 16:
|
||||
reduceMultiPass<16, true><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
|
||||
case 8:
|
||||
reduceMultiPass<8, true>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
case 8:
|
||||
reduceMultiPass<8, true><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
|
||||
case 4:
|
||||
reduceMultiPass<4, true>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
case 4:
|
||||
reduceMultiPass<4, true><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
|
||||
case 2:
|
||||
reduceMultiPass<2, true>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
case 2:
|
||||
reduceMultiPass<2, true><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
|
||||
case 1:
|
||||
reduceMultiPass<1, true>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
case 1:
|
||||
reduceMultiPass<1, true><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
}
|
||||
}
|
||||
} else {
|
||||
switch (threads) {
|
||||
case 512:
|
||||
reduceMultiPass<512, false>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
else {
|
||||
switch (threads) {
|
||||
case 512:
|
||||
reduceMultiPass<512, false><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
|
||||
case 256:
|
||||
reduceMultiPass<256, false>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
case 256:
|
||||
reduceMultiPass<256, false><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
|
||||
case 128:
|
||||
reduceMultiPass<128, false>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
case 128:
|
||||
reduceMultiPass<128, false><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
|
||||
case 64:
|
||||
reduceMultiPass<64, false>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
case 64:
|
||||
reduceMultiPass<64, false><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
|
||||
case 32:
|
||||
reduceMultiPass<32, false>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
case 32:
|
||||
reduceMultiPass<32, false><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
|
||||
case 16:
|
||||
reduceMultiPass<16, false>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
case 16:
|
||||
reduceMultiPass<16, false><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
|
||||
case 8:
|
||||
reduceMultiPass<8, false>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
case 8:
|
||||
reduceMultiPass<8, false><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
|
||||
case 4:
|
||||
reduceMultiPass<4, false>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
case 4:
|
||||
reduceMultiPass<4, false><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
|
||||
case 2:
|
||||
reduceMultiPass<2, false>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
case 2:
|
||||
reduceMultiPass<2, false><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
|
||||
case 1:
|
||||
reduceMultiPass<1, false>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
case 1:
|
||||
reduceMultiPass<1, false><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
extern "C" void reduceSinglePass(int size, int threads, int blocks,
|
||||
float *d_idata, float *d_odata) {
|
||||
dim3 dimBlock(threads, 1, 1);
|
||||
dim3 dimGrid(blocks, 1, 1);
|
||||
int smemSize = threads * sizeof(float);
|
||||
extern "C" void reduceSinglePass(int size, int threads, int blocks, float *d_idata, float *d_odata)
|
||||
{
|
||||
dim3 dimBlock(threads, 1, 1);
|
||||
dim3 dimGrid(blocks, 1, 1);
|
||||
int smemSize = threads * sizeof(float);
|
||||
|
||||
// choose which of the optimized versions of reduction to launch
|
||||
if (isPow2(size)) {
|
||||
switch (threads) {
|
||||
case 512:
|
||||
reduceSinglePass<512, true>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
// choose which of the optimized versions of reduction to launch
|
||||
if (isPow2(size)) {
|
||||
switch (threads) {
|
||||
case 512:
|
||||
reduceSinglePass<512, true><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
|
||||
case 256:
|
||||
reduceSinglePass<256, true>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
case 256:
|
||||
reduceSinglePass<256, true><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
|
||||
case 128:
|
||||
reduceSinglePass<128, true>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
case 128:
|
||||
reduceSinglePass<128, true><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
|
||||
case 64:
|
||||
reduceSinglePass<64, true>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
case 64:
|
||||
reduceSinglePass<64, true><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
|
||||
case 32:
|
||||
reduceSinglePass<32, true>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
case 32:
|
||||
reduceSinglePass<32, true><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
|
||||
case 16:
|
||||
reduceSinglePass<16, true>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
case 16:
|
||||
reduceSinglePass<16, true><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
|
||||
case 8:
|
||||
reduceSinglePass<8, true>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
case 8:
|
||||
reduceSinglePass<8, true><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
|
||||
case 4:
|
||||
reduceSinglePass<4, true>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
case 4:
|
||||
reduceSinglePass<4, true><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
|
||||
case 2:
|
||||
reduceSinglePass<2, true>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
case 2:
|
||||
reduceSinglePass<2, true><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
|
||||
case 1:
|
||||
reduceSinglePass<1, true>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
case 1:
|
||||
reduceSinglePass<1, true><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
}
|
||||
}
|
||||
} else {
|
||||
switch (threads) {
|
||||
case 512:
|
||||
reduceSinglePass<512, false>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
else {
|
||||
switch (threads) {
|
||||
case 512:
|
||||
reduceSinglePass<512, false><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
|
||||
case 256:
|
||||
reduceSinglePass<256, false>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
case 256:
|
||||
reduceSinglePass<256, false><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
|
||||
case 128:
|
||||
reduceSinglePass<128, false>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
case 128:
|
||||
reduceSinglePass<128, false><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
|
||||
case 64:
|
||||
reduceSinglePass<64, false>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
case 64:
|
||||
reduceSinglePass<64, false><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
|
||||
case 32:
|
||||
reduceSinglePass<32, false>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
case 32:
|
||||
reduceSinglePass<32, false><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
|
||||
case 16:
|
||||
reduceSinglePass<16, false>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
case 16:
|
||||
reduceSinglePass<16, false><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
|
||||
case 8:
|
||||
reduceSinglePass<8, false>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
case 8:
|
||||
reduceSinglePass<8, false><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
|
||||
case 4:
|
||||
reduceSinglePass<4, false>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
case 4:
|
||||
reduceSinglePass<4, false><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
|
||||
case 2:
|
||||
reduceSinglePass<2, false>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
case 2:
|
||||
reduceSinglePass<2, false><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
|
||||
case 1:
|
||||
reduceSinglePass<1, false>
|
||||
<<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
case 1:
|
||||
reduceSinglePass<1, false><<<dimGrid, dimBlock, smemSize>>>(d_idata, d_odata, size);
|
||||
break;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
#endif // #ifndef _REDUCE_KERNEL_H_
|
||||
#endif // #ifndef _REDUCE_KERNEL_H_
|
||||
|
||||
Reference in New Issue
Block a user