ファイル
rocm-systems/projects/clr/rocclr/platform/commandqueue.cpp
T
SaleelK 6b28faa532 clr: Implement per-stream SDMA engine affinity for improved copy performance (#2480)
Problem:
The existing SDMA engine selection logic had several issues:
1. Same VirtualGPU/stream could use different SDMA engines for consecutive
   async copies since copy_engine_status may report engines as busy
2. Busy and Preferred engine check for every copy
3. No global tracking of which VirtualGPU uses which engine, leading to
   suboptimal resource allocation

Solution:
Implemented a global SDMA engine allocator with per-stream affinity:

- Added Device::SdmaEngineAllocator to manage VirtualGPU → engine assignments
  * Maintains global map of active assignments
  * Enforces exclusivity: different streams use different engines (except
    inter-GPU copies where preferred engines are prioritized for optimal
    hardware paths like XGMI links)
  * Thread-safe allocation/release with Monitor lock

- Modified VirtualGPU to cache assigned engine locally (assigned_sdma_engine_)
  for fast lookup without map access on hot path

- Refactored rocrCopyBuffer() to:
  1. Check local cached engine first → use if assigned
  2. Call AllocateSdmaEngine() if not assigned → cache result

- Moved HSA API queries (memory_copy_engine_status, memory_get_preferred_copy_engine)
  into AllocateEngine() for cleaner separation of concerns

- Engine release on HostQueue::finish() instead of only VirtualGPU destruction
  * Improves engine utilization by releasing earlier
  * Added virtual ReleaseSdmaEngines() method to device::VirtualDevice

- Added future path for simple round-robin allocation (kUseSimpleRR) for
  next-gen GPUs with uniform SDMA bandwidth (disabled by default)

Cleanup:
- Removed selectSdmaEngine() helper (logic moved to allocator)
- Removed getSdmaRWMasks() (allocator accesses maxSdmaReadMask_/WriteMask_ directly)
- Removed unused sdmaEngineReadMask_/WriteMask_ member variables from DmaBlitManager

Benefits:
- Ensures consistent per-stream SDMA engine usage
- Prevents cross-stream contention and engine thrashing
- Prioritizes hardware-optimal paths for inter-GPU transfers
- Better resource utilization through earlier release
- Cleaner, more maintainable code structure
2026-01-07 19:37:45 -08:00

398 行
13 KiB
C++

/* Copyright (c) 2012 - 2021 Advanced Micro Devices, Inc.
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in
all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
THE SOFTWARE. */
#include "commandqueue.hpp"
#include "thread/monitor.hpp"
#include "device/device.hpp"
#include "platform/context.hpp"
#include "utils/flags.hpp"
/*!
* \file commandQueue.cpp
* \brief Definitions for HostQueue object.
*
* \author Laurent Morichetti
* \date October 2008
*/
namespace amd {
HostQueue::HostQueue(Context& context, Device& device, cl_command_queue_properties props,
uint queueRTCUs, Priority priority, const std::vector<uint32_t>& cuMask)
: CommandQueue(context, device, props, device.info().queueProperties_, queueRTCUs, priority,
cuMask),
lastEnqueueCommand_(nullptr),
head_(nullptr),
tail_(nullptr),
isActive_(false),
sync_policy_(amd::SyncPolicy::Auto) {
if (GPU_FORCE_QUEUE_PROFILING) {
properties().set(CL_QUEUE_PROFILING_ENABLE);
}
if (AMD_DIRECT_DISPATCH) {
// Initialize the queue
thread_.Init(this);
} else {
if (thread_.state() >= Thread::INITIALIZED) {
ScopedLock sl(queueLock_);
thread_.start(this);
// wait for HostQueue::loop() to update acceptingCommands_ as true
while (!thread_.acceptingCommands_) {
queueLock_.wait();
}
}
}
}
bool HostQueue::terminate() {
// In case of force destroy skip checking on the last command
if (AMD_DIRECT_DISPATCH) {
if (!forceDestroy_ && vdev() != nullptr) {
// If the queue still has the last command, then wait and release it
// We must be in protected way to get last command when calling
// awaitCompletion() where lastCommand will be released and possibly
// destroyed.
Command* lastCommand = getLastQueuedCommand(true);
if (lastCommand != nullptr) {
// Check if CPU batch wasn't flushed for completion with the last command
if (GetSubmissionBatchSize() != 0) {
auto command = new Marker(*this, false);
if (command != nullptr) {
ClPrint(LOG_DETAIL_DEBUG, LOG_CMD, "Marker queued to ensure finish");
command->enqueue();
lastCommand->release();
lastCommand = command;
}
}
if (device_.gpu_error_ == CL_SUCCESS) {
lastCommand->awaitCompletion();
}
// Note that if lastCommand isn't a marker, it may not be lastEnqueueCommand_ now
// after lastCommand->awaitCompletion() is called.
if (lastEnqueueCommand_ != nullptr) {
lastEnqueueCommand_->release(); // lastEnqueueCommand_ should be a marker
lastEnqueueCommand_ = nullptr;
}
lastCommand->release();
}
}
thread_.Release();
thread_.acceptingCommands_ = false;
} else {
if (Os::isThreadAlive(thread_)) {
Command* marker = nullptr;
// Send a finish if the queue is still accepting commands.
if (lastEnqueueCommand_ != nullptr || !amd::IS_HIP) {
ScopedLock sl(queueLock_);
if (thread_.acceptingCommands_) {
marker = new Marker(*this, false);
if (marker != nullptr) {
append(*marker);
queueLock_.notify();
}
}
}
if (marker != nullptr) {
if (marker->notifyCmdQueue()) {
while (marker->status() > CL_COMPLETE && Os::isThreadAlive(thread_)) {
amd::Os::yield();
}
}
marker->release();
}
// Wake-up the command loop, so it can exit
{
ScopedLock sl(queueLock_);
thread_.acceptingCommands_ = false;
queueLock_.notify();
}
// FIXME_lmoriche: fix termination handshake
while (thread_.state() < Thread::FINISHED && Os::isThreadAlive(thread_)) {
Os::yield();
}
}
}
if (Agent::shouldPostCommandQueueEvents()) {
Agent::postCommandQueueFree(as_cl(this->asCommandQueue()));
}
device_.removeFromActiveQueues(this);
return true;
}
void HostQueue::finishCommand(Command* command) {
if (command == nullptr) {
command = getLastQueuedCommand(true);
if (command != nullptr) {
ClPrint(LOG_DETAIL_DEBUG, LOG_CMD, "No command, awaiting complete status on host");
command->awaitCompletion();
command->release();
}
return;
}
// Check hardware event status for the specific command
static constexpr bool kWaitCompletion = true;
if (!device().IsHwEventReady(command->event(), kWaitCompletion)) {
ClPrint(LOG_DETAIL_DEBUG, LOG_CMD, "No HW event, awaiting complete status on host");
command->awaitCompletion();
}
}
void HostQueue::finish(bool cpu_wait) {
Command* command = nullptr;
size_t minBatchSize = 0;
if (IS_HIP) {
minBatchSize = DEBUG_CLR_BATCH_CPU_SYNC_SIZE;
command = getLastQueuedCommand(true);
if (command == nullptr) {
ClPrint(LOG_DEBUG, LOG_CMD, "No command awaiting completion on host");
return;
}
// Force blocking wait if requested. That allows to avoid a build up of unreleased CPU commands
if ((DEBUG_HIP_BLOCK_SYNC > 0) &&
(vdev()->QueuedAsyncHandlers().load() > DEBUG_HIP_BLOCK_SYNC)) {
cpu_wait = true;
}
} else {
// Force CPU wait for OpenCL, since the tests may check OCL command status after finish
cpu_wait = true;
}
size_t batchSize = GetSubmissionBatchSize();
ClPrint(LOG_DETAIL_DEBUG, LOG_CMD,
"finish() called with batch size: %zu, cpu_wait: %d, "
"fence dirty: %d",
batchSize, cpu_wait, vdev()->isFenceDirty());
// Force marker if the batch wasn't sent for CPU update or fence is dirty
if (nullptr == command || (batchSize != 0)|| vdev()->isFenceDirty()) {
if (nullptr != command) {
command->release();
}
const Command::EventWaitList nullWaitList = {};
// Send a finish to make sure we finished all commands
command = new Marker(*this, false, nullWaitList, nullptr, batchSize < minBatchSize);
if (command == NULL) {
return;
}
command->enqueue();
}
// Check HW status of the ROCclr event. Note: not all ROCclr modes support HW status
static constexpr bool kWaitCompletion = true;
if (cpu_wait || !device().IsHwEventReady(command->event(), kWaitCompletion, GetSyncPolicy())) {
ClPrint(LOG_DETAIL_DEBUG, LOG_CMD,
"No HW event or batch size is less than %zu, "
"await command completion",
minBatchSize);
command->awaitCompletion();
}
if (IS_HIP) {
ScopedLock sl(vdev()->execution());
ScopedLock l(lastCmdLock_);
// Runtime can clear the last command only if no other submissions occured
// during finish()
if (command == lastEnqueueCommand_) {
device_.removeFromActiveQueues(this);
// Under Windows runtime can't destroy objects in the callback thread.
// Also runtime should force interrupt before any destroy. Hence, if it was just gpu wait,
// then keep the lastEnqueueCommand_ for the interrupt handling.
if (IS_LINUX || cpu_wait || GPU_ENABLE_PAL != 0) {
lastEnqueueCommand_->release();
lastEnqueueCommand_ = nullptr;
}
}
}
// Release SDMA engine assignments
vdev()->ReleaseSdmaEngines();
// Release all HW queues, which are idle or nearly idle
vdev()->ReleaseAllHwQueues();
command->release();
ClPrint(LOG_DEBUG, LOG_CMD, "All commands finished for host queue : %p", this);
}
void HostQueue::loop(device::VirtualDevice* virtualDevice) {
// Notify the caller that the queue is ready to accept commands.
{
ScopedLock sl(queueLock_);
// Notify HostQueue() that acceptingCommands_ is updated to true
thread_.acceptingCommands_ = true;
queueLock_.notify();
}
// Create a command batch with all the commands present in the queue.
Command* head = NULL;
Command* tail = NULL;
while (true) {
// Get one command from the queue
Command* command = queue_.dequeue();
if (command == NULL) {
ScopedLock sl(queueLock_);
while ((command = queue_.dequeue()) == NULL) {
if (!thread_.acceptingCommands_) {
return;
}
queueLock_.wait();
}
}
command->retain();
// Process the command's event wait list.
const Command::EventWaitList& events = command->eventWaitList();
bool dependencyFailed = false;
ClPrint(LOG_DETAIL_DEBUG, LOG_CMD, "Command (%s) processing: %p ,events.size(): %d",
amd::activity_prof::getOclCommandKindString(command->type()), command, events.size());
for (const auto& it : events) {
// Only wait if the command is enqueued into another queue.
if (it->command().queue() != this) {
// Runtime has to flush the current batch only if the dependent wait is blocking
if (it->command().status() != CL_COMPLETE) {
ClPrint(LOG_DETAIL_DEBUG, LOG_CMD, "Command (%s) %p awaiting event: %p",
amd::activity_prof::getOclCommandKindString(command->type()), command, it);
virtualDevice->flush(head, true);
tail = head = NULL;
dependencyFailed |= !it->awaitCompletion();
}
}
}
// Insert the command to the linked list.
if (NULL == head) { // if the list is empty
head = tail = command;
} else {
tail->setNext(command);
tail = command;
}
if (dependencyFailed) {
command->setStatus(CL_EXEC_STATUS_ERROR_FOR_EVENTS_IN_WAIT_LIST);
continue;
}
ClPrint(LOG_DETAIL_DEBUG, LOG_CMD, "Command (%s) submitted: %p",
amd::activity_prof::getOclCommandKindString(command->type()), command);
command->setStatus(CL_SUBMITTED);
// Submit to the device queue.
command->submit(*virtualDevice);
// if this is a user invisible marker with a waiting event, then flush
if (0 == command->type()) {
virtualDevice->flush(head);
tail = head = NULL;
}
} // while (true) {
}
void HostQueue::append(Command& command) {
// We retain the command here. It will be released when its status
// changes to CL_COMPLETE
if ((command.getWaitBits() & 0x1) != 0) {
finish();
}
command.retain();
command.setStatus(CL_QUEUED);
queue_.enqueue(&command);
if (!IS_HIP) {
return;
}
// Set last submitted command
Command* prevLastEnqueueCommand = nullptr;
// Attach only real commands and skip internal notifications for CPU queue
if (command.waitingEvent() == nullptr) {
command.retain();
// lastCmdLock_ ensures that lastEnqueueCommand() can retain the command before it is swapped
// out. We want to keep this critical section as short as possible, so the command should be
// released outside this section.
ScopedLock l(lastCmdLock_);
prevLastEnqueueCommand = lastEnqueueCommand_;
lastEnqueueCommand_ = &command;
}
if (prevLastEnqueueCommand != nullptr) {
prevLastEnqueueCommand->release();
} else {
// The queue becomes active. Add it to the set of activeQueues.
device_.addToActiveQueues(this);
}
}
bool HostQueue::isEmpty() {
// Get a snapshot of queue size
return queue_.empty();
}
Command* HostQueue::getLastQueuedCommand(bool retain) {
if (AMD_DIRECT_DISPATCH) {
// The batch update must be lock protected to avoid a race condition
// when multiple threads submit/flush/update the batch at the same time
ScopedLock sl(vdev()->execution());
// Since the lastCmdLock_ is acquired, it is safe to read and retain the lastEnqueueCommand.
// It is guaranteed that the pointer will not change.
if (retain && lastEnqueueCommand_ != nullptr) {
lastEnqueueCommand_->retain();
}
return lastEnqueueCommand_;
} else {
// Get last submitted command
ScopedLock l(lastCmdLock_);
// Since the lastCmdLock_ is acquired, it is safe to read and retain the lastEnqueueCommand.
// It is guaranteed that the pointer will not change.
if (retain && lastEnqueueCommand_ != nullptr) {
lastEnqueueCommand_->retain();
}
return lastEnqueueCommand_;
}
}
DeviceQueue::~DeviceQueue() {
delete virtualDevice_;
ScopedLock lock(context().lock());
context().removeDeviceQueue(device(), this);
}
bool DeviceQueue::create() {
const bool defaultDeviceQueue = properties().test(CL_QUEUE_ON_DEVICE_DEFAULT);
bool result = false;
virtualDevice_ = device().createVirtualDevice(this);
if (virtualDevice_ != NULL) {
result = true;
context().addDeviceQueue(device(), this, defaultDeviceQueue);
}
return result;
}
} // namespace amd