P4 to Git Change 1458879 by gandryey@gera-w8 on 2017/09/14 11:41:44
SWDEV-129129 - [[CQE OCL][Vega vs Fiji] Upto 12% Performance drop observed on VEGA10 compared to FIJI while running BlackMagic Davinci Resolve
More benchmark tuning:
- Keep system memory locked in the resource cache. That removes huge amount of lock/unlock calls to OS due to the resource creation and destruciton
- Reduce the command buffer size to 256 commands and incrrease the amount of CBs to 16
- Increase the amount of resident resources to 2048
Affected files ...
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpudevice.cpp#574 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palmemory.cpp#17 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palresource.cpp#37 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palvirtual.cpp#58 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palvirtual.hpp#31 edit
[ROCm/clr commit: 4066449a8b]
This commit is contained in:
@@ -1337,6 +1337,12 @@ gpu::Memory* Device::createBuffer(amd::Memory& owner, bool directAccess) const {
|
||||
(owner.forceSysMemAlloc() || (owner.getMemFlags() & CL_MEM_SVM_FINE_GRAIN_BUFFER))
|
||||
? Resource::Remote
|
||||
: Resource::Local;
|
||||
|
||||
// Check if runtime can force a tiny buffer into USWC memory
|
||||
if ((size <= (GPU_MAX_REMOTE_MEM_SIZE * Ki)) && (type == Resource::Local) &&
|
||||
(owner.getMemFlags() & CL_MEM_READ_ONLY)) {
|
||||
type = Resource::RemoteUSWC;
|
||||
}
|
||||
|
||||
if (owner.getMemFlags() & CL_MEM_BUS_ADDRESSABLE_AMD) {
|
||||
type = Resource::BusAddressable;
|
||||
|
||||
Reference in New Issue
Block a user