Sync v13.5-public from internal repo

This commit is contained in:
cuda-samples sync
2026-11-04 20:44:00 +00:00
parent 5443602d89
commit a9315c55fb
53 changed files with 1750 additions and 183 deletions

View File

@@ -89,8 +89,10 @@ template <> __device__ double getSofteningSquared<double>() { return soften
template <typename T> struct DeviceData
{
T *dPos[2]; // mapped host pointers
T *dReadScratch;
T *dVel;
cudaEvent_t event;
cudaStream_t stream;
unsigned int offset;
unsigned int numBodies;
};
@@ -210,14 +212,22 @@ void integrateNbodySystem(DeviceData<T> *deviceData,
bool bUsePBO)
{
if (bUsePBO) {
if (numDevices > 1) {
checkCudaErrors(cudaSetDevice(0));
}
checkCudaErrors(cudaGraphicsResourceSetMapFlags(pgres[currentRead], cudaGraphicsMapFlagsReadOnly));
checkCudaErrors(cudaGraphicsResourceSetMapFlags(pgres[1 - currentRead], cudaGraphicsMapFlagsWriteDiscard));
checkCudaErrors(cudaGraphicsMapResources(2, pgres, 0));
checkCudaErrors(cudaGraphicsMapResources(2, pgres, deviceData[0].stream));
size_t bytes;
checkCudaErrors(cudaGraphicsResourceGetMappedPointer(
(void **)&(deviceData[0].dPos[currentRead]), &bytes, pgres[currentRead]));
checkCudaErrors(cudaGraphicsResourceGetMappedPointer(
(void **)&(deviceData[0].dPos[1 - currentRead]), &bytes, pgres[1 - currentRead]));
// Peer copies and non-owner writes use deviceData[0].dPos directly.
// Those pointers are only valid after this map; other devices' streams
// do not wait for it.
checkCudaErrors(cudaStreamSynchronize(deviceData[0].stream));
}
for (unsigned int dev = 0; dev != numDevices; dev++) {
@@ -228,21 +238,33 @@ void integrateNbodySystem(DeviceData<T> *deviceData,
int numBlocks = (deviceData[dev].numBodies + blockSize - 1) / blockSize;
int numTiles = (numBodies + blockSize - 1) / blockSize;
int sharedMemSize = blockSize * 4 * sizeof(T); // 4 floats for pos
// GPU0 owns both ping-pong position buffers and velocity. Always pass
// this map's GPU0 pointers into the kernel so a non-owner never writes
// through a pre-map alias. Local scratch is the only per-device buffer.
T *readPos = deviceData[0].dPos[currentRead];
integrateBodies<T>
<<<numBlocks, blockSize, sharedMemSize>>>((typename vec4<T>::Type *)deviceData[dev].dPos[1 - currentRead],
(typename vec4<T>::Type *)deviceData[dev].dPos[currentRead],
(typename vec4<T>::Type *)deviceData[dev].dVel,
deviceData[dev].offset,
deviceData[dev].numBodies,
deltaTime,
damping,
numTiles);
if (deviceData[dev].dReadScratch != nullptr) {
checkCudaErrors(cudaMemcpyPeerAsync(deviceData[dev].dReadScratch,
dev,
deviceData[0].dPos[currentRead],
0,
numBodies * 4 * sizeof(T),
deviceData[dev].stream));
readPos = deviceData[dev].dReadScratch;
}
integrateBodies<T><<<numBlocks, blockSize, sharedMemSize, deviceData[dev].stream>>>(
(typename vec4<T>::Type *)deviceData[0].dPos[1 - currentRead],
(typename vec4<T>::Type *)readPos,
(typename vec4<T>::Type *)deviceData[0].dVel,
deviceData[dev].offset,
deviceData[dev].numBodies,
deltaTime,
damping,
numTiles);
if (numDevices > 1) {
checkCudaErrors(cudaEventRecord(deviceData[dev].event));
// MJH: Hack on older driver versions to force kernel launches to flush!
cudaStreamQuery(0);
checkCudaErrors(cudaEventRecord(deviceData[dev].event, deviceData[dev].stream));
}
// check if kernel invocation generated an error
@@ -250,13 +272,23 @@ void integrateNbodySystem(DeviceData<T> *deviceData,
}
if (numDevices > 1) {
// Single-GPU stays on the legacy default stream (see _initialize())
// and keeps queuing kernels without a host wait here. Multi-GPU
// kernels run on per-device non-blocking streams; wait so the next
// step's peer copy and any PBO unmap see completed writes.
for (unsigned int dev = 0; dev < numDevices; dev++) {
checkCudaErrors(cudaEventSynchronize(deviceData[dev].event));
}
}
if (bUsePBO) {
checkCudaErrors(cudaGraphicsUnmapResources(2, pgres, 0));
if (numDevices > 1) {
checkCudaErrors(cudaSetDevice(0));
}
checkCudaErrors(cudaGraphicsUnmapResources(2, pgres, deviceData[0].stream));
// GL (windowed display) may use the PBO as soon as update() returns.
checkCudaErrors(cudaStreamSynchronize(deviceData[0].stream));
}
}