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

P4 to Git Change 1991259 by gandryey@gera-win10 on 2019/08/29 18:45:50

Browse Source 
SWDEV-200614 - [Schneider] Crash in Agisoft when run in mGPU environment
	-  Add a workaround for memory pinning path. It will perform 2-step copy to make sure memory pinning doesn't occur on the first unaligned page, because in Windows memory manager can have CPU access to the allocation header in another thread and a race condition is possible
	- change some default setting for staging and pinned paths, because PCIE gen3 performance.

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palsettings.cpp#96 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palvirtual.cpp#150 edit
... //depot/stg/opencl/drivers/opencl/runtime/utils/flags.hpp#317 edit


[ROCm/clr commit: daa0e96062]
This commit is contained in:
foreman
2019-08-29 18:58:53 -04:00
parent 6b9ede7d26
commit f9686edf91
3 changed files with 54 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/clr/rocclr/runtime/device/pal/palvirtual.cpp
+44 -3
View File
@@ -1145,8 +1145,28 @@ void VirtualGPU::submitReadMemory(amd::ReadMemoryCommand& vcmd) {
result = blitMgr().copyBuffer(*memory, *hostMemory, origin, dstOrigin, size,
vcmd.isEntireMemory());
} else {
result =
blitMgr().readBuffer(*memory, vcmd.destination(), origin, size, vcmd.isEntireMemory());
// The logic below will perform 2 step copy to make sure memory pinning doesn't
// occur on the first unaligned page, because in Windows memory manager can
// have CPU access to the allocation header in another thread
// and a race condition is possible.
char* tmpHost = amd::alignUp(
reinterpret_cast<char*>(vcmd.destination()), PinnedMemoryAlignment);
// Find the partial size for unaligned copy
size_t partial = tmpHost - reinterpret_cast<char*>(vcmd.destination());
result = true;
// Check if it's staging copy, then ignore unaligned address
if (size[0] <= dev().settings().pinnedMinXferSize_) {
partial = size[0];
}
// Make first step transfer
if (partial > 0) {
result = blitMgr().readBuffer(*memory, vcmd.destination(), origin, partial);
}
// Second step transfer if something left to copy
if (partial < size[0]) {
result &= blitMgr().readBuffer(*memory, tmpHost, origin[0] + partial, size[0] - partial);
}
}
if (nullptr != bufferFromImage) {
bufferFromImage->release();
@@ -1258,7 +1278,28 @@ void VirtualGPU::submitWriteMemory(amd::WriteMemoryCommand& vcmd) {
result = blitMgr().copyBuffer(*hostMemory, *memory, srcOrigin, origin, size,
vcmd.isEntireMemory());
} else {
result = blitMgr().writeBuffer(vcmd.source(), *memory, origin, size, vcmd.isEntireMemory());
// The logic below will perform 2 step copy to make sure memory pinning doesn't
// occur on the first unaligned page, because in Windows memory manager can
// have CPU access to the allocation header in another thread
// and a race condition is possible.
const char* tmpHost =
amd::alignUp(reinterpret_cast<const char*>(vcmd.source()), PinnedMemoryAlignment);
// Find the partial size for unaligned copy
size_t partial = tmpHost - reinterpret_cast<const char*>(vcmd.source());
result = true;
// Check if it's staging copy, then ignore unaligned address
if (size[0] <= dev().settings().pinnedMinXferSize_) {
partial = size[0];
}
// Make first step transfer
if (partial > 0) {
result = blitMgr().writeBuffer(vcmd.source(), *memory, origin, partial);
}
// Second step transfer if something left to copy
if (partial < size[0]) {
result &= blitMgr().writeBuffer(tmpHost, *memory, origin[0] + partial, size[0] - partial);
}
}
if (nullptr != bufferFromImage) {
bufferFromImage->release();