SWDEV-353941 - Fix hipMemset latency issue for hipMallocManaged

- In case of HMM, use blit kernel instead of CPU memcpy for hipMemset

Change-Id: I89bfc96ff01a2375ed8df1b1c6bc05357dea84f7


[ROCm/clr commit: f097cda948]
This commit is contained in:
Rakesh Roy
2022-09-05 23:17:30 +05:30
committed by Rakesh Roy
orang tua 26d51f2c90
melakukan 9cdd929342
2 mengubah file dengan 22 tambahan dan 2 penghapusan
@@ -1872,11 +1872,21 @@ void VirtualGPU::submitFillMemory(amd::FillMemoryCommand& cmd) {
amd::BufferRect rect;
rect.create(static_cast<size_t*>(origin), static_cast<size_t*>(region),
pitch, 0);
bool force_blit = false;
if (amd::IS_HIP) {
constexpr uint32_t kManagedAlloc = (CL_MEM_SVM_FINE_GRAIN_BUFFER | CL_MEM_ALLOC_HOST_PTR);
// In case of HMM, use blit kernel instead of CPU memcpy
if ((cmd.memory().getMemFlags() & kManagedAlloc) == kManagedAlloc) {
force_blit = true;
}
}
for (size_t slice = 0; slice < depth; slice++) {
for (size_t row = 0; row < height; row++) {
const size_t rowOffset = rect.offset(0, row, slice);
if (!fillMemory(cmd.type(), &cmd.memory(), cmd.pattern(), cmd.patternSize(),
amd::Coord3D{rowOffset, 0, 0}, amd::Coord3D{width, 1, 1})) {
amd::Coord3D{rowOffset, 0, 0}, amd::Coord3D{width, 1, 1}, force_blit)) {
cmd.setStatus(CL_INVALID_OPERATION);
}
}
@@ -2296,8 +2296,18 @@ void VirtualGPU::submitFillMemory(amd::FillMemoryCommand& cmd) {
amd::ScopedLock lock(execution());
profilingBegin(cmd);
bool force_blit = false;
if (amd::IS_HIP) {
constexpr uint32_t kManagedAlloc = (CL_MEM_SVM_FINE_GRAIN_BUFFER | CL_MEM_ALLOC_HOST_PTR);
// In case of HMM, use blit kernel instead of CPU memcpy
if ((cmd.memory().getMemFlags() & kManagedAlloc) == kManagedAlloc) {
force_blit = true;
}
}
if (!fillMemory(cmd.type(), &cmd.memory(), cmd.pattern(), cmd.patternSize(),
cmd.surface(), cmd.origin(), cmd.size())) {
cmd.surface(), cmd.origin(), cmd.size(), force_blit)) {
cmd.setStatus(CL_INVALID_OPERATION);
}
profilingEnd(cmd);