AI/rocm-systems
2
0 Garpu 0
Kode Masalah Tarik Permintaan Actions Packages Projects Rilis Wiki Kegiatan

SWDEV-234684 - hipmemcpy optimization does not work in tests

Telusuri Sumber 
Change-Id: I899d172c5b2af88c796fe9a36f97d15ac45caf94
This commit is contained in:
Alex Xie committed 2020-05-05 15:58:03 -04:00
1 orang tua 0fbc0a895b
melakukan bfbc8cd09b
2 files changed
+5 -6
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

rocclr/device/rocm/rocblit.cpp
+4 -4
Melihat File
@@ -1658,8 +1658,8 @@ bool KernelBlitManager::readBuffer(device::Memory& srcMemory, void* dstHost,
if ((srcMemory.owner()->getHostMem() == nullptr) && (srcMemory.owner()->getSvmPtr() != nullptr)) {
// CPU read ahead, hence release GPU memory
gpu().releaseGpuMemoryFence();
void* src = srcMemory.owner()->getSvmPtr();
std::memcpy(dstHost, src, size[0]);
char* src = reinterpret_cast<char*>(srcMemory.owner()->getSvmPtr());
std::memcpy(dstHost, src + origin[0], size[0]);
// Set HASPENDINGDISPATCH_ FLAG. That will force L2 invalidation on flush
gpu().hasPendingDispatch();
return true;
@@ -1763,8 +1763,8 @@ bool KernelBlitManager::writeBuffer(const void* srcHost, device::Memory& dstMemo
if ((dstMemory.owner()->getHostMem() == nullptr) && (dstMemory.owner()->getSvmPtr() != nullptr)) {
// CPU read ahead, hence release GPU memory
gpu().releaseGpuMemoryFence();
void* dst = dstMemory.owner()->getSvmPtr();
std::memcpy(dst, srcHost, size[0]);
char* dst = reinterpret_cast<char*>(dstMemory.owner()->getSvmPtr());
std::memcpy(dst + origin[0], srcHost, size[0]);
// Set HASPENDINGDISPATCH_ FLAG. Then releaseGpuMemoryFence() will use barrier to invalidate cache
gpu().hasPendingDispatch();
gpu().releaseGpuMemoryFence();