AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

SWDEV-423317 - Enable GPU wait for hip sync calls

Browse Source 
hipStreamSynchronize and hipDeviceSynchronize won't longer wait
for CPU commands in DD mode

Change-Id: I079c8bbfc34ddc6d3e2d74c92a34665877e512a5


[ROCm/clr commit: fbea58ba11]
This commit is contained in:
German Andryeyev
2023-09-21 15:47:56 -04:00
parent f62807d411
commit 2d492a201b
5 changed files with 13 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/clr/hipamd/src/hip_device_runtime.cpp
+2 -1
View File
@@ -520,7 +520,8 @@ hipError_t hipDeviceSetSharedMemConfig ( hipSharedMemConfig config ) {
hipError_t hipDeviceSynchronize() { hipError_t hipDeviceSynchronize() {
HIP_INIT_API(hipDeviceSynchronize); HIP_INIT_API(hipDeviceSynchronize);
hip::Stream::SyncAllStreams(hip::getCurrentDevice()->deviceId()); constexpr bool kDontWaitForCpu = false;
hip::Stream::SyncAllStreams(hip::getCurrentDevice()->deviceId(), kDontWaitForCpu);
HIP_RETURN(hipSuccess); HIP_RETURN(hipSuccess);
} }
projects/clr/hipamd/src/hip_internal.hpp
+1 -1
View File
@@ -293,7 +293,7 @@ namespace hip {
const std::vector<uint32_t> GetCUMask() const { return cuMask_; } const std::vector<uint32_t> GetCUMask() const { return cuMask_; }
/// Sync all streams /// Sync all streams
static void SyncAllStreams(int deviceId); static void SyncAllStreams(int deviceId, bool cpu_wait = true);
/// Check whether any blocking stream running /// Check whether any blocking stream running
static bool StreamCaptureBlocking(); static bool StreamCaptureBlocking();
projects/clr/hipamd/src/hip_stream.cpp
+4 -3
View File
@@ -122,7 +122,7 @@ int Stream::DeviceId(const hipStream_t hStream) {
} }
// ================================================================================================ // ================================================================================================
void Stream::SyncAllStreams(int deviceId) { void Stream::SyncAllStreams(int deviceId, bool cpu_wait) {
// Make a local copy to avoid stalls for GPU finish with multiple threads // Make a local copy to avoid stalls for GPU finish with multiple threads
std::vector<hip::Stream*> streams; std::vector<hip::Stream*> streams;
streams.reserve(streamSet.size()); streams.reserve(streamSet.size());
@@ -136,7 +136,7 @@ void Stream::SyncAllStreams(int deviceId) {
} }
} }
for (auto it : streams) { for (auto it : streams) {
it->finish(); it->finish(cpu_wait);
it->release(); it->release();
} }
} }
@@ -442,8 +442,9 @@ hipError_t hipStreamSynchronize_common(hipStream_t stream) {
} }
} }
bool wait = (stream == nullptr) ? true : false; bool wait = (stream == nullptr) ? true : false;
constexpr bool kDontWaitForCpu = false;
// Wait for the current host queue // Wait for the current host queue
hip::getStream(stream, wait)->finish(); hip::getStream(stream, wait)->finish(kDontWaitForCpu);
return hipSuccess; return hipSuccess;
} }
projects/clr/rocclr/platform/commandqueue.cpp
+4 -3
View File
@@ -113,7 +113,7 @@ bool HostQueue::terminate() {
return true; return true;
} }
void HostQueue::finish() { void HostQueue::finish(bool cpu_wait) {
Command* command = nullptr; Command* command = nullptr;
if (IS_HIP) { if (IS_HIP) {
command = getLastQueuedCommand(true); command = getLastQueuedCommand(true);
@@ -121,7 +121,8 @@ void HostQueue::finish() {
return; return;
} }
} }
if (nullptr == command || vdev()->isHandlerPending() || vdev()->isFenceDirty()) { if (nullptr == command || command->type() != CL_COMMAND_MARKER ||
vdev()->isHandlerPending() || vdev()->isFenceDirty()) {
if (nullptr != command) { if (nullptr != command) {
command->release(); command->release();
} }
@@ -135,7 +136,7 @@ void HostQueue::finish() {
} }
// Check HW status of the ROCcrl event. Note: not all ROCclr modes support HW status // Check HW status of the ROCcrl event. Note: not all ROCclr modes support HW status
static constexpr bool kWaitCompletion = true; static constexpr bool kWaitCompletion = true;
if (!device().IsHwEventReady(command->event(), kWaitCompletion)) { if (cpu_wait || !device().IsHwEventReady(command->event(), kWaitCompletion)) {
ClPrint(LOG_DEBUG, LOG_CMD, "HW Event not ready, awaiting completion instead"); ClPrint(LOG_DEBUG, LOG_CMD, "HW Event not ready, awaiting completion instead");
command->awaitCompletion(); command->awaitCompletion();
} }
projects/clr/rocclr/platform/commandqueue.hpp
+1 -1
View File
@@ -231,7 +231,7 @@ class HostQueue : public CommandQueue {
} }
//! Finish all queued commands //! Finish all queued commands
void finish(); void finish(bool cpu_wait = false);
//! Check if hostQueue empty snapshot //! Check if hostQueue empty snapshot
bool isEmpty(); bool isEmpty();