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
@@ -53,257 +53,237 @@ static const char *sSDKsample = "[simpleVoteIntrinsics]\0";
|
||||
#include "simpleVote_kernel.cuh"
|
||||
|
||||
// Generate the test pattern for Tests 1 and 2
|
||||
void genVoteTestPattern(unsigned int *VOTE_PATTERN, int size) {
|
||||
// For testing VOTE.Any (all of these threads will return 0)
|
||||
for (int i = 0; i < size / 4; i++) {
|
||||
VOTE_PATTERN[i] = 0x00000000;
|
||||
}
|
||||
|
||||
// For testing VOTE.Any (1/2 these threads will return 1)
|
||||
for (int i = 2 * size / 8; i < 4 * size / 8; i++) {
|
||||
VOTE_PATTERN[i] = (i & 0x01) ? i : 0;
|
||||
}
|
||||
|
||||
// For testing VOTE.all (1/2 of these threads will return 0)
|
||||
for (int i = 2 * size / 4; i < 3 * size / 4; i++) {
|
||||
VOTE_PATTERN[i] = (i & 0x01) ? 0 : i;
|
||||
}
|
||||
|
||||
// For testing VOTE.all (all of these threads will return 1)
|
||||
for (int i = 3 * size / 4; i < 4 * size / 4; i++) {
|
||||
VOTE_PATTERN[i] = 0xffffffff;
|
||||
}
|
||||
}
|
||||
|
||||
int checkErrors1(unsigned int *h_result, int start, int end, int warp_size,
|
||||
const char *voteType) {
|
||||
int i, sum = 0;
|
||||
|
||||
for (sum = 0, i = start; i < end; i++) {
|
||||
sum += h_result[i];
|
||||
}
|
||||
|
||||
if (sum > 0) {
|
||||
printf("\t<%s>[%d - %d] = ", voteType, start, end - 1);
|
||||
|
||||
for (i = start; i < end; i++) {
|
||||
printf("%d", h_result[i]);
|
||||
void genVoteTestPattern(unsigned int *VOTE_PATTERN, int size)
|
||||
{
|
||||
// For testing VOTE.Any (all of these threads will return 0)
|
||||
for (int i = 0; i < size / 4; i++) {
|
||||
VOTE_PATTERN[i] = 0x00000000;
|
||||
}
|
||||
|
||||
printf("%d values FAILED\n", sum);
|
||||
}
|
||||
|
||||
return (sum > 0);
|
||||
}
|
||||
|
||||
int checkErrors2(unsigned int *h_result, int start, int end, int warp_size,
|
||||
const char *voteType) {
|
||||
int i, sum = 0;
|
||||
|
||||
for (sum = 0, i = start; i < end; i++) {
|
||||
sum += h_result[i];
|
||||
}
|
||||
|
||||
if (sum != warp_size) {
|
||||
printf("\t<%s>[%d - %d] = ", voteType, start, end - 1);
|
||||
|
||||
for (i = start; i < end; i++) {
|
||||
printf("%d", h_result[i]);
|
||||
// For testing VOTE.Any (1/2 these threads will return 1)
|
||||
for (int i = 2 * size / 8; i < 4 * size / 8; i++) {
|
||||
VOTE_PATTERN[i] = (i & 0x01) ? i : 0;
|
||||
}
|
||||
|
||||
printf(" - FAILED\n");
|
||||
}
|
||||
// For testing VOTE.all (1/2 of these threads will return 0)
|
||||
for (int i = 2 * size / 4; i < 3 * size / 4; i++) {
|
||||
VOTE_PATTERN[i] = (i & 0x01) ? 0 : i;
|
||||
}
|
||||
|
||||
return (sum != warp_size);
|
||||
// For testing VOTE.all (all of these threads will return 1)
|
||||
for (int i = 3 * size / 4; i < 4 * size / 4; i++) {
|
||||
VOTE_PATTERN[i] = 0xffffffff;
|
||||
}
|
||||
}
|
||||
|
||||
int checkErrors1(unsigned int *h_result, int start, int end, int warp_size, const char *voteType)
|
||||
{
|
||||
int i, sum = 0;
|
||||
|
||||
for (sum = 0, i = start; i < end; i++) {
|
||||
sum += h_result[i];
|
||||
}
|
||||
|
||||
if (sum > 0) {
|
||||
printf("\t<%s>[%d - %d] = ", voteType, start, end - 1);
|
||||
|
||||
for (i = start; i < end; i++) {
|
||||
printf("%d", h_result[i]);
|
||||
}
|
||||
|
||||
printf("%d values FAILED\n", sum);
|
||||
}
|
||||
|
||||
return (sum > 0);
|
||||
}
|
||||
|
||||
int checkErrors2(unsigned int *h_result, int start, int end, int warp_size, const char *voteType)
|
||||
{
|
||||
int i, sum = 0;
|
||||
|
||||
for (sum = 0, i = start; i < end; i++) {
|
||||
sum += h_result[i];
|
||||
}
|
||||
|
||||
if (sum != warp_size) {
|
||||
printf("\t<%s>[%d - %d] = ", voteType, start, end - 1);
|
||||
|
||||
for (i = start; i < end; i++) {
|
||||
printf("%d", h_result[i]);
|
||||
}
|
||||
|
||||
printf(" - FAILED\n");
|
||||
}
|
||||
|
||||
return (sum != warp_size);
|
||||
}
|
||||
|
||||
// Verification code for Kernel #1
|
||||
int checkResultsVoteAnyKernel1(unsigned int *h_result, int size,
|
||||
int warp_size) {
|
||||
int error_count = 0;
|
||||
int checkResultsVoteAnyKernel1(unsigned int *h_result, int size, int warp_size)
|
||||
{
|
||||
int error_count = 0;
|
||||
|
||||
error_count += checkErrors1(h_result, 0, VOTE_DATA_GROUP * warp_size / 4,
|
||||
warp_size, "Vote.Any");
|
||||
error_count +=
|
||||
checkErrors2(h_result, VOTE_DATA_GROUP * warp_size / 4,
|
||||
2 * VOTE_DATA_GROUP * warp_size / 4, warp_size, "Vote.Any");
|
||||
error_count +=
|
||||
checkErrors2(h_result, 2 * VOTE_DATA_GROUP * warp_size / 4,
|
||||
3 * VOTE_DATA_GROUP * warp_size / 4, warp_size, "Vote.Any");
|
||||
error_count +=
|
||||
checkErrors2(h_result, 3 * VOTE_DATA_GROUP * warp_size / 4,
|
||||
4 * VOTE_DATA_GROUP * warp_size / 4, warp_size, "Vote.Any");
|
||||
error_count += checkErrors1(h_result, 0, VOTE_DATA_GROUP * warp_size / 4, warp_size, "Vote.Any");
|
||||
error_count += checkErrors2(
|
||||
h_result, VOTE_DATA_GROUP * warp_size / 4, 2 * VOTE_DATA_GROUP * warp_size / 4, warp_size, "Vote.Any");
|
||||
error_count += checkErrors2(
|
||||
h_result, 2 * VOTE_DATA_GROUP * warp_size / 4, 3 * VOTE_DATA_GROUP * warp_size / 4, warp_size, "Vote.Any");
|
||||
error_count += checkErrors2(
|
||||
h_result, 3 * VOTE_DATA_GROUP * warp_size / 4, 4 * VOTE_DATA_GROUP * warp_size / 4, warp_size, "Vote.Any");
|
||||
|
||||
printf((error_count == 0) ? "\tOK\n" : "\tERROR\n");
|
||||
return error_count;
|
||||
printf((error_count == 0) ? "\tOK\n" : "\tERROR\n");
|
||||
return error_count;
|
||||
}
|
||||
|
||||
// Verification code for Kernel #2
|
||||
int checkResultsVoteAllKernel2(unsigned int *h_result, int size,
|
||||
int warp_size) {
|
||||
int error_count = 0;
|
||||
int checkResultsVoteAllKernel2(unsigned int *h_result, int size, int warp_size)
|
||||
{
|
||||
int error_count = 0;
|
||||
|
||||
error_count += checkErrors1(h_result, 0, VOTE_DATA_GROUP * warp_size / 4,
|
||||
warp_size, "Vote.All");
|
||||
error_count +=
|
||||
checkErrors1(h_result, VOTE_DATA_GROUP * warp_size / 4,
|
||||
2 * VOTE_DATA_GROUP * warp_size / 4, warp_size, "Vote.All");
|
||||
error_count +=
|
||||
checkErrors1(h_result, 2 * VOTE_DATA_GROUP * warp_size / 4,
|
||||
3 * VOTE_DATA_GROUP * warp_size / 4, warp_size, "Vote.All");
|
||||
error_count +=
|
||||
checkErrors2(h_result, 3 * VOTE_DATA_GROUP * warp_size / 4,
|
||||
4 * VOTE_DATA_GROUP * warp_size / 4, warp_size, "Vote.All");
|
||||
error_count += checkErrors1(h_result, 0, VOTE_DATA_GROUP * warp_size / 4, warp_size, "Vote.All");
|
||||
error_count += checkErrors1(
|
||||
h_result, VOTE_DATA_GROUP * warp_size / 4, 2 * VOTE_DATA_GROUP * warp_size / 4, warp_size, "Vote.All");
|
||||
error_count += checkErrors1(
|
||||
h_result, 2 * VOTE_DATA_GROUP * warp_size / 4, 3 * VOTE_DATA_GROUP * warp_size / 4, warp_size, "Vote.All");
|
||||
error_count += checkErrors2(
|
||||
h_result, 3 * VOTE_DATA_GROUP * warp_size / 4, 4 * VOTE_DATA_GROUP * warp_size / 4, warp_size, "Vote.All");
|
||||
|
||||
printf((error_count == 0) ? "\tOK\n" : "\tERROR\n");
|
||||
return error_count;
|
||||
printf((error_count == 0) ? "\tOK\n" : "\tERROR\n");
|
||||
return error_count;
|
||||
}
|
||||
|
||||
// Verification code for Kernel #3
|
||||
int checkResultsVoteAnyKernel3(bool *hinfo, int size) {
|
||||
int i, error_count = 0;
|
||||
int checkResultsVoteAnyKernel3(bool *hinfo, int size)
|
||||
{
|
||||
int i, error_count = 0;
|
||||
|
||||
for (i = 0; i < size * 3; i++) {
|
||||
switch (i % 3) {
|
||||
case 0:
|
||||
for (i = 0; i < size * 3; i++) {
|
||||
switch (i % 3) {
|
||||
case 0:
|
||||
|
||||
// First warp should be all zeros.
|
||||
if (hinfo[i] != (i >= size * 1)) {
|
||||
error_count++;
|
||||
// First warp should be all zeros.
|
||||
if (hinfo[i] != (i >= size * 1)) {
|
||||
error_count++;
|
||||
}
|
||||
|
||||
break;
|
||||
|
||||
case 1:
|
||||
|
||||
// First warp and half of second should be all zeros.
|
||||
if (hinfo[i] != (i >= size * 3 / 2)) {
|
||||
error_count++;
|
||||
}
|
||||
|
||||
break;
|
||||
|
||||
case 2:
|
||||
|
||||
// First two warps should be all zeros.
|
||||
if (hinfo[i] != (i >= size * 2)) {
|
||||
error_count++;
|
||||
}
|
||||
|
||||
break;
|
||||
}
|
||||
|
||||
break;
|
||||
|
||||
case 1:
|
||||
|
||||
// First warp and half of second should be all zeros.
|
||||
if (hinfo[i] != (i >= size * 3 / 2)) {
|
||||
error_count++;
|
||||
}
|
||||
|
||||
break;
|
||||
|
||||
case 2:
|
||||
|
||||
// First two warps should be all zeros.
|
||||
if (hinfo[i] != (i >= size * 2)) {
|
||||
error_count++;
|
||||
}
|
||||
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
printf((error_count == 0) ? "\tOK\n" : "\tERROR\n");
|
||||
return error_count;
|
||||
printf((error_count == 0) ? "\tOK\n" : "\tERROR\n");
|
||||
return error_count;
|
||||
}
|
||||
|
||||
int main(int argc, char **argv) {
|
||||
unsigned int *h_input, *h_result;
|
||||
unsigned int *d_input, *d_result;
|
||||
int main(int argc, char **argv)
|
||||
{
|
||||
unsigned int *h_input, *h_result;
|
||||
unsigned int *d_input, *d_result;
|
||||
|
||||
bool *dinfo = NULL, *hinfo = NULL;
|
||||
int error_count[3] = {0, 0, 0};
|
||||
bool *dinfo = NULL, *hinfo = NULL;
|
||||
int error_count[3] = {0, 0, 0};
|
||||
|
||||
cudaDeviceProp deviceProp;
|
||||
int devID, warp_size = 32;
|
||||
cudaDeviceProp deviceProp;
|
||||
int devID, warp_size = 32;
|
||||
|
||||
printf("%s\n", sSDKsample);
|
||||
printf("%s\n", sSDKsample);
|
||||
|
||||
// This will pick the best possible CUDA capable device
|
||||
devID = findCudaDevice(argc, (const char **)argv);
|
||||
// This will pick the best possible CUDA capable device
|
||||
devID = findCudaDevice(argc, (const char **)argv);
|
||||
|
||||
checkCudaErrors(cudaGetDeviceProperties(&deviceProp, devID));
|
||||
checkCudaErrors(cudaGetDeviceProperties(&deviceProp, devID));
|
||||
|
||||
// Statistics about the GPU device
|
||||
printf(
|
||||
"> GPU device has %d Multi-Processors, SM %d.%d compute capabilities\n\n",
|
||||
deviceProp.multiProcessorCount, deviceProp.major, deviceProp.minor);
|
||||
// Statistics about the GPU device
|
||||
printf("> GPU device has %d Multi-Processors, SM %d.%d compute capabilities\n\n",
|
||||
deviceProp.multiProcessorCount,
|
||||
deviceProp.major,
|
||||
deviceProp.minor);
|
||||
|
||||
h_input = (unsigned int *)malloc(VOTE_DATA_GROUP * warp_size *
|
||||
sizeof(unsigned int));
|
||||
h_result = (unsigned int *)malloc(VOTE_DATA_GROUP * warp_size *
|
||||
sizeof(unsigned int));
|
||||
checkCudaErrors(
|
||||
cudaMalloc(reinterpret_cast<void **>(&d_input),
|
||||
VOTE_DATA_GROUP * warp_size * sizeof(unsigned int)));
|
||||
checkCudaErrors(
|
||||
cudaMalloc(reinterpret_cast<void **>(&d_result),
|
||||
VOTE_DATA_GROUP * warp_size * sizeof(unsigned int)));
|
||||
genVoteTestPattern(h_input, VOTE_DATA_GROUP * warp_size);
|
||||
checkCudaErrors(cudaMemcpy(d_input, h_input,
|
||||
VOTE_DATA_GROUP * warp_size * sizeof(unsigned int),
|
||||
cudaMemcpyHostToDevice));
|
||||
h_input = (unsigned int *)malloc(VOTE_DATA_GROUP * warp_size * sizeof(unsigned int));
|
||||
h_result = (unsigned int *)malloc(VOTE_DATA_GROUP * warp_size * sizeof(unsigned int));
|
||||
checkCudaErrors(
|
||||
cudaMalloc(reinterpret_cast<void **>(&d_input), VOTE_DATA_GROUP * warp_size * sizeof(unsigned int)));
|
||||
checkCudaErrors(
|
||||
cudaMalloc(reinterpret_cast<void **>(&d_result), VOTE_DATA_GROUP * warp_size * sizeof(unsigned int)));
|
||||
genVoteTestPattern(h_input, VOTE_DATA_GROUP * warp_size);
|
||||
checkCudaErrors(
|
||||
cudaMemcpy(d_input, h_input, VOTE_DATA_GROUP * warp_size * sizeof(unsigned int), cudaMemcpyHostToDevice));
|
||||
|
||||
// Start of Vote Any Test Kernel #1
|
||||
printf("[VOTE Kernel Test 1/3]\n");
|
||||
printf("\tRunning <<Vote.Any>> kernel1 ...\n");
|
||||
{
|
||||
checkCudaErrors(cudaDeviceSynchronize());
|
||||
dim3 gridBlock(1, 1);
|
||||
dim3 threadBlock(VOTE_DATA_GROUP * warp_size, 1);
|
||||
VoteAnyKernel1<<<gridBlock, threadBlock>>>(d_input, d_result,
|
||||
VOTE_DATA_GROUP * warp_size);
|
||||
getLastCudaError("VoteAnyKernel() execution failed\n");
|
||||
checkCudaErrors(cudaDeviceSynchronize());
|
||||
}
|
||||
checkCudaErrors(cudaMemcpy(h_result, d_result,
|
||||
VOTE_DATA_GROUP * warp_size * sizeof(unsigned int),
|
||||
cudaMemcpyDeviceToHost));
|
||||
error_count[0] += checkResultsVoteAnyKernel1(
|
||||
h_result, VOTE_DATA_GROUP * warp_size, warp_size);
|
||||
// Start of Vote Any Test Kernel #1
|
||||
printf("[VOTE Kernel Test 1/3]\n");
|
||||
printf("\tRunning <<Vote.Any>> kernel1 ...\n");
|
||||
{
|
||||
checkCudaErrors(cudaDeviceSynchronize());
|
||||
dim3 gridBlock(1, 1);
|
||||
dim3 threadBlock(VOTE_DATA_GROUP * warp_size, 1);
|
||||
VoteAnyKernel1<<<gridBlock, threadBlock>>>(d_input, d_result, VOTE_DATA_GROUP * warp_size);
|
||||
getLastCudaError("VoteAnyKernel() execution failed\n");
|
||||
checkCudaErrors(cudaDeviceSynchronize());
|
||||
}
|
||||
checkCudaErrors(
|
||||
cudaMemcpy(h_result, d_result, VOTE_DATA_GROUP * warp_size * sizeof(unsigned int), cudaMemcpyDeviceToHost));
|
||||
error_count[0] += checkResultsVoteAnyKernel1(h_result, VOTE_DATA_GROUP * warp_size, warp_size);
|
||||
|
||||
// Start of Vote All Test Kernel #2
|
||||
printf("\n[VOTE Kernel Test 2/3]\n");
|
||||
printf("\tRunning <<Vote.All>> kernel2 ...\n");
|
||||
{
|
||||
checkCudaErrors(cudaDeviceSynchronize());
|
||||
dim3 gridBlock(1, 1);
|
||||
dim3 threadBlock(VOTE_DATA_GROUP * warp_size, 1);
|
||||
VoteAllKernel2<<<gridBlock, threadBlock>>>(d_input, d_result,
|
||||
VOTE_DATA_GROUP * warp_size);
|
||||
getLastCudaError("VoteAllKernel() execution failed\n");
|
||||
checkCudaErrors(cudaDeviceSynchronize());
|
||||
}
|
||||
checkCudaErrors(cudaMemcpy(h_result, d_result,
|
||||
VOTE_DATA_GROUP * warp_size * sizeof(unsigned int),
|
||||
cudaMemcpyDeviceToHost));
|
||||
error_count[1] += checkResultsVoteAllKernel2(
|
||||
h_result, VOTE_DATA_GROUP * warp_size, warp_size);
|
||||
// Start of Vote All Test Kernel #2
|
||||
printf("\n[VOTE Kernel Test 2/3]\n");
|
||||
printf("\tRunning <<Vote.All>> kernel2 ...\n");
|
||||
{
|
||||
checkCudaErrors(cudaDeviceSynchronize());
|
||||
dim3 gridBlock(1, 1);
|
||||
dim3 threadBlock(VOTE_DATA_GROUP * warp_size, 1);
|
||||
VoteAllKernel2<<<gridBlock, threadBlock>>>(d_input, d_result, VOTE_DATA_GROUP * warp_size);
|
||||
getLastCudaError("VoteAllKernel() execution failed\n");
|
||||
checkCudaErrors(cudaDeviceSynchronize());
|
||||
}
|
||||
checkCudaErrors(
|
||||
cudaMemcpy(h_result, d_result, VOTE_DATA_GROUP * warp_size * sizeof(unsigned int), cudaMemcpyDeviceToHost));
|
||||
error_count[1] += checkResultsVoteAllKernel2(h_result, VOTE_DATA_GROUP * warp_size, warp_size);
|
||||
|
||||
// Second Vote Kernel Test #3 (both Any/All)
|
||||
hinfo = reinterpret_cast<bool *>(calloc(warp_size * 3 * 3, sizeof(bool)));
|
||||
cudaMalloc(reinterpret_cast<void **>(&dinfo),
|
||||
warp_size * 3 * 3 * sizeof(bool));
|
||||
cudaMemcpy(dinfo, hinfo, warp_size * 3 * 3 * sizeof(bool),
|
||||
cudaMemcpyHostToDevice);
|
||||
// Second Vote Kernel Test #3 (both Any/All)
|
||||
hinfo = reinterpret_cast<bool *>(calloc(warp_size * 3 * 3, sizeof(bool)));
|
||||
cudaMalloc(reinterpret_cast<void **>(&dinfo), warp_size * 3 * 3 * sizeof(bool));
|
||||
cudaMemcpy(dinfo, hinfo, warp_size * 3 * 3 * sizeof(bool), cudaMemcpyHostToDevice);
|
||||
|
||||
printf("\n[VOTE Kernel Test 3/3]\n");
|
||||
printf("\tRunning <<Vote.Any>> kernel3 ...\n");
|
||||
{
|
||||
checkCudaErrors(cudaDeviceSynchronize());
|
||||
VoteAnyKernel3<<<1, warp_size * 3>>>(dinfo, warp_size);
|
||||
checkCudaErrors(cudaDeviceSynchronize());
|
||||
}
|
||||
printf("\n[VOTE Kernel Test 3/3]\n");
|
||||
printf("\tRunning <<Vote.Any>> kernel3 ...\n");
|
||||
{
|
||||
checkCudaErrors(cudaDeviceSynchronize());
|
||||
VoteAnyKernel3<<<1, warp_size * 3>>>(dinfo, warp_size);
|
||||
checkCudaErrors(cudaDeviceSynchronize());
|
||||
}
|
||||
|
||||
cudaMemcpy(hinfo, dinfo, warp_size * 3 * 3 * sizeof(bool),
|
||||
cudaMemcpyDeviceToHost);
|
||||
cudaMemcpy(hinfo, dinfo, warp_size * 3 * 3 * sizeof(bool), cudaMemcpyDeviceToHost);
|
||||
|
||||
error_count[2] = checkResultsVoteAnyKernel3(hinfo, warp_size * 3);
|
||||
error_count[2] = checkResultsVoteAnyKernel3(hinfo, warp_size * 3);
|
||||
|
||||
// Now free these resources for Test #1,2
|
||||
checkCudaErrors(cudaFree(d_input));
|
||||
checkCudaErrors(cudaFree(d_result));
|
||||
free(h_input);
|
||||
free(h_result);
|
||||
// Now free these resources for Test #1,2
|
||||
checkCudaErrors(cudaFree(d_input));
|
||||
checkCudaErrors(cudaFree(d_result));
|
||||
free(h_input);
|
||||
free(h_result);
|
||||
|
||||
// Free resources from Test #3
|
||||
free(hinfo);
|
||||
cudaFree(dinfo);
|
||||
// Free resources from Test #3
|
||||
free(hinfo);
|
||||
cudaFree(dinfo);
|
||||
|
||||
printf("\tShutting down...\n");
|
||||
printf("\tShutting down...\n");
|
||||
|
||||
return (error_count[0] == 0 && error_count[1] == 0 && error_count[2] == 0)
|
||||
? EXIT_SUCCESS
|
||||
: EXIT_FAILURE;
|
||||
return (error_count[0] == 0 && error_count[1] == 0 && error_count[2] == 0) ? EXIT_SUCCESS : EXIT_FAILURE;
|
||||
}
|
||||
|
||||
@@ -38,43 +38,44 @@
|
||||
// If ANY one of the threads (within the warp) of the predicated condition
|
||||
// returns a non-zero value, then all threads within this warp will return a
|
||||
// non-zero value
|
||||
__global__ void VoteAnyKernel1(unsigned int *input, unsigned int *result,
|
||||
int size) {
|
||||
int tx = threadIdx.x;
|
||||
__global__ void VoteAnyKernel1(unsigned int *input, unsigned int *result, int size)
|
||||
{
|
||||
int tx = threadIdx.x;
|
||||
|
||||
int mask = 0xffffffff;
|
||||
result[tx] = __any_sync(mask, input[tx]);
|
||||
int mask = 0xffffffff;
|
||||
result[tx] = __any_sync(mask, input[tx]);
|
||||
}
|
||||
|
||||
// Kernel #2 tests the across-the-warp vote(all) intrinsic.
|
||||
// If ALL of the threads (within the warp) of the predicated condition returns
|
||||
// a non-zero value, then all threads within this warp will return a non-zero
|
||||
// value
|
||||
__global__ void VoteAllKernel2(unsigned int *input, unsigned int *result,
|
||||
int size) {
|
||||
int tx = threadIdx.x;
|
||||
__global__ void VoteAllKernel2(unsigned int *input, unsigned int *result, int size)
|
||||
{
|
||||
int tx = threadIdx.x;
|
||||
|
||||
int mask = 0xffffffff;
|
||||
result[tx] = __all_sync(mask, input[tx]);
|
||||
int mask = 0xffffffff;
|
||||
result[tx] = __all_sync(mask, input[tx]);
|
||||
}
|
||||
|
||||
// Kernel #3 is a directed test for the across-the-warp vote(all) intrinsic.
|
||||
// This kernel will test for conditions across warps, and within half warps
|
||||
__global__ void VoteAnyKernel3(bool *info, int warp_size) {
|
||||
int tx = threadIdx.x;
|
||||
unsigned int mask = 0xffffffff;
|
||||
bool *offs = info + (tx * 3);
|
||||
__global__ void VoteAnyKernel3(bool *info, int warp_size)
|
||||
{
|
||||
int tx = threadIdx.x;
|
||||
unsigned int mask = 0xffffffff;
|
||||
bool *offs = info + (tx * 3);
|
||||
|
||||
// The following should hold true for the second and third warp
|
||||
*offs = __any_sync(mask, (tx >= (warp_size * 3) / 2));
|
||||
// The following should hold true for the "upper half" of the second warp,
|
||||
// and all of the third warp
|
||||
*(offs + 1) = (tx >= (warp_size * 3) / 2 ? true : false);
|
||||
// The following should hold true for the second and third warp
|
||||
*offs = __any_sync(mask, (tx >= (warp_size * 3) / 2));
|
||||
// The following should hold true for the "upper half" of the second warp,
|
||||
// and all of the third warp
|
||||
*(offs + 1) = (tx >= (warp_size * 3) / 2 ? true : false);
|
||||
|
||||
// The following should hold true for the third warp only
|
||||
if (__all_sync(mask, (tx >= (warp_size * 3) / 2))) {
|
||||
*(offs + 2) = true;
|
||||
}
|
||||
// The following should hold true for the third warp only
|
||||
if (__all_sync(mask, (tx >= (warp_size * 3) / 2))) {
|
||||
*(offs + 2) = true;
|
||||
}
|
||||
}
|
||||
|
||||
#endif
|
||||
|
||||
Reference in New Issue
Block a user