intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-12-27 07:44:16 +08:00
Code Issues Packages Projects Releases Wiki Activity

performance: use BCS for transfers if CCS is busy

Browse Source 
Related-To: NEO-11501

Also, if device is iGPU, don't use staging buffers
in that case.

Signed-off-by: Szymon Morek <szymon.morek@intel.com>
This commit is contained in:
Szymon Morek
2024-07-19 10:25:15 +00:00
committed by Compute-Runtime-Automation
parent 17380dcbf3
commit 39ec7facee
18 changed files with 211 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files

23
opencl/source/command_queue/command_queue.cpp
View File

@@ -302,8 +302,12 @@ CommandStreamReceiver &CommandQueue::selectCsrForBuiltinOperation(const CsrSelec
case TransferDirection::hostToHost:
case TransferDirection::hostToLocal:
case TransferDirection::localToHost: {
preferBcs = true;
auto isWriteToImageFromBuffer = args.dstResource.image && args.dstResource.image->isImageFromBuffer();
auto &productHelper = device->getProductHelper();
preferBcs = device->getRootDeviceEnvironment().isWddmOnLinux() || productHelper.blitEnqueuePreferred(isWriteToImageFromBuffer);
if (debugManager.flags.EnableBlitterForEnqueueOperations.get() == 1) {
preferBcs = true;
}
auto preferredBCSType = true;
if (debugManager.flags.AssignBCSAtEnqueue.get() != -1) {
@@ -322,6 +326,11 @@ CommandStreamReceiver &CommandQueue::selectCsrForBuiltinOperation(const CsrSelec
device->getSelectorCopyEngine(), false);
}
}
if (!preferBcs && isOOQEnabled() && getGpgpuCommandStreamReceiver().isBusy()) {
// If CCS is preferred but it's OOQ and compute engine is busy, select BCS instead
preferBcs = true;
}
break;
}
default:
@@ -1093,8 +1102,7 @@ bool CommandQueue::queueDependenciesClearRequired() const {
}
bool CommandQueue::blitEnqueueAllowed(const CsrSelectionArgs &args) const {
auto isWriteToImageFromBuffer = args.dstResource.image && args.dstResource.image->isImageFromBuffer();
bool blitEnqueueAllowed = ((device->getRootDeviceEnvironment().isWddmOnLinux() || device->getRootDeviceEnvironment().getProductHelper().blitEnqueueAllowed(isWriteToImageFromBuffer)) && getGpgpuCommandStreamReceiver().peekTimestampPacketWriteEnabled()) || this->isCopyOnly;
bool blitEnqueueAllowed = getGpgpuCommandStreamReceiver().peekTimestampPacketWriteEnabled() || this->isCopyOnly;
if (debugManager.flags.EnableBlitterForEnqueueOperations.get() != -1) {
blitEnqueueAllowed = debugManager.flags.EnableBlitterForEnqueueOperations.get();
}
@@ -1611,6 +1619,13 @@ bool CommandQueue::isValidForStagingBufferCopy(Device &device, void *dstPtr, con
// Direct transfer from mapped allocation is faster than staging buffer
return false;
}
auto rootDeviceIndex = device.getRootDeviceIndex();
auto isLocalMem = device.getMemoryManager()->isLocalMemorySupported(rootDeviceIndex);
if (isOOQEnabled() && getGpgpuCommandStreamReceiver().isBusy() && !isLocalMem) {
// It's not beneficial to make copy through staging buffers if it's OOQ,
// compute engine is busy and device is iGPU.
return false;
}
CsrSelectionArgs csrSelectionArgs{CL_COMMAND_SVM_MEMCPY, nullptr};
csrSelectionArgs.direction = TransferDirection::hostToLocal;
auto csr = &selectCsrForBuiltinOperation(csrSelectionArgs);