AI/rocm-systems
2
0
Fork 0
Κώδικας Ζητήματα Pull Requests Δράσεις Πακέτα Έργα Κυκλοφορίες Wiki Δραστηριότητα

P4 to Git Change 1101352 by gandryey@gera-dev-w7 on 2014/11/28 18:03:18

Περιήγηση Στο Πηγαίο 
ECR #304775 - Make optimization for read map of USWC memory
	- If runtime detects USWC map with read operation, then it will switch to indirect map. This should improve map-read  performance on APU(s)  when USWC memory is used instead of frame buffer

Affected files ...

... //depot/stg/opencl/drivers/opencl/api/opencl/amdocl/cl_memobj.cpp#72 edit
... //depot/stg/opencl/drivers/opencl/api/opencl/amdocl/cl_svm.cpp#8 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/cpu/cpudevice.cpp#269 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/cpu/cpudevice.hpp#89 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/device.cpp#172 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/device.hpp#234 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpudevice.cpp#486 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpudevice.hpp#134 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpumemory.cpp#112 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpumemory.hpp#43 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuvirtual.cpp#340 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/hsa/hsadevice.cpp#88 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/hsa/hsadevice.hpp#45 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/hsa/hsamemory.cpp#42 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/hsa/hsamemory.hpp#27 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/hsa/hsavirtual.cpp#98 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/hsa_foundation/hsadevice.cpp#21 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/hsa_foundation/hsadevice.hpp#7 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/hsa_foundation/hsamemory.cpp#6 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/hsa_foundation/hsamemory.hpp#5 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/hsa_foundation/hsavirtual.cpp#26 edit


[ROCm/clr commit: 6d464be252]
Αυτή η υποβολή περιλαμβάνεται σε:
foreman
2014-11-28 18:11:36 -05:00
γονέας de5e6ef163
υποβολή f43897e51e
14 αρχεία άλλαξαν με 75 προσθήκες και 65 διαγραφές
Λήψη Αρχείου Patch Λήψη Αρχείου Diff Ανάπτυξη όλων των αρχείων Σύμπτυξη όλων των αρχείων
projects/clr/rocclr/runtime/device/gpu/gpudevice.cpp
+2 -1
Προβολή Αρχείου
@@ -2098,6 +2098,7 @@ Device::allocMapTarget(
amd::Memory& mem,
const amd::Coord3D& origin,
const amd::Coord3D& region,
uint mapFlags,
size_t* rowPitch,
size_t* slicePitch)
{
@@ -2109,7 +2110,7 @@ Device::allocMapTarget(
}
// Pass request over to memory
return memory->allocMapTarget(origin, region, rowPitch, slicePitch);
return memory->allocMapTarget(origin, region, mapFlags, rowPitch, slicePitch);
}
bool
projects/clr/rocclr/runtime/device/gpu/gpudevice.hpp
+2
Προβολή Αρχείου
@@ -111,6 +111,7 @@ public:
amd::Memory& mem, //!< Abstraction layer memory object
const amd::Coord3D& origin, //!< The map location in memory
const amd::Coord3D& region, //!< The map region in memory
uint mapFlags, //!< Map flags
size_t* rowPitch = NULL, //!< Row pitch for the mapped memory
size_t* slicePitch = NULL //!< Slice for the mapped memory
) { return NULL; }
@@ -437,6 +438,7 @@ public:
amd::Memory& mem, //!< Abstraction layer memory object
const amd::Coord3D& origin, //!< The map location in memory
const amd::Coord3D& region, //!< The map region in memory
uint mapFlags, //!< Map flags
size_t* rowPitch = NULL, //!< Row pitch for the mapped memory
size_t* slicePitch = NULL //!< Slice for the mapped memory
);
projects/clr/rocclr/runtime/device/gpu/gpumemory.cpp
+6 -2
Προβολή Αρχείου
@@ -906,6 +906,7 @@ void*
Memory::allocMapTarget(
const amd::Coord3D& origin,
const amd::Coord3D& region,
uint mapFlags,
size_t* rowPitch,
size_t* slicePitch)
{
@@ -921,7 +922,8 @@ Memory::allocMapTarget(
incIndMapCount();
// If host memory exists, use it
if (owner()->getHostMem() != NULL) {
if ((owner()->getHostMem() != NULL) &&
(isCacheable() || !isHostMemDirectAccess() || !(mapFlags & CL_MAP_READ))) {
mapAddress = reinterpret_cast<address>(owner()->getHostMem());
}
// If resource is a persistent allocation, we can use it directly
@@ -1226,6 +1228,7 @@ void*
Image::allocMapTarget(
const amd::Coord3D& origin,
const amd::Coord3D& region,
uint mapFlags,
size_t* rowPitch,
size_t* slicePitch)
{
@@ -1245,7 +1248,8 @@ Image::allocMapTarget(
incIndMapCount();
// If host memory exists, use it
if (owner()->getHostMem() != NULL) {
if ((owner()->getHostMem() != NULL) &&
(isCacheable() || !isHostMemDirectAccess() || !(mapFlags & CL_MAP_READ))) {
useRemoteResource = false;
mapAddress = reinterpret_cast<address>(owner()->getHostMem());
amd::Image* amdImage = owner()->asImage();
projects/clr/rocclr/runtime/device/gpu/gpumemory.hpp
+2
Προβολή Αρχείου
@@ -121,6 +121,7 @@ public:
virtual void* allocMapTarget(
const amd::Coord3D& origin, //!< The map location in memory
const amd::Coord3D& region, //!< The map region in memory
uint mapFlags, //!< Map flags
size_t* rowPitch = NULL, //!< Row pitch for the mapped memory
size_t* slicePitch = NULL //!< Slice for the mapped memory
);
@@ -288,6 +289,7 @@ public:
virtual void* allocMapTarget(
const amd::Coord3D& origin, //!< The map location in memory
const amd::Coord3D& region, //!< The map region in memory
uint mapFlags, //!< Map flags
size_t* rowPitch = NULL, //!< Row pitch for the mapped memory
size_t* slicePitch = NULL //!< Slice for the mapped memory
);
projects/clr/rocclr/runtime/device/gpu/gpuvirtual.cpp
+28 -47
Προβολή Αρχείου
@@ -1085,14 +1085,15 @@ VirtualGPU::submitMapMemory(amd::MapMemoryCommand& vcmd)
gpu::Memory* memory = dev().getGpuMemory(&vcmd.memory());
// Save write map info for unmap copy
if (vcmd.mapFlags() & (CL_MAP_WRITE | CL_MAP_WRITE_INVALIDATE_REGION)) {
memory->saveWriteMapInfo(vcmd.origin(),
vcmd.size(), vcmd.isEntireMemory());
}
// Save map info for unmap operation
memory->saveMapInfo(vcmd.origin(), vcmd.size(),
vcmd.mapFlags(), vcmd.isEntireMemory());
// If we have host memory, use it
if (memory->owner()->getHostMem() != NULL) {
if ((memory->owner()->getHostMem() != NULL) &&
(memory->isCacheable() ||
!memory->isHostMemDirectAccess() ||
!(vcmd.mapFlags() & CL_MAP_READ))) {
if (!memory->isHostMemDirectAccess()) {
// Make sure GPU finished operation before
// synchronization with the backing store
@@ -1177,7 +1178,10 @@ VirtualGPU::submitUnmapMemory(amd::UnmapMemoryCommand& vcmd)
amd::Memory* owner = memory->owner();
// We used host memory
if (owner->getHostMem() != NULL) {
if ((owner->getHostMem() != NULL) &&
(memory->isCacheable() ||
!memory->isHostMemDirectAccess() ||
!memory->isUnmapRead())) {
if (memory->isUnmapWrite()) {
// Target is the backing store, so sync
owner->signalWrite(NULL);
@@ -1254,8 +1258,8 @@ VirtualGPU::submitUnmapMemory(amd::UnmapMemoryCommand& vcmd)
vcmd.setStatus(CL_INVALID_VALUE);
}
// Clear read only flag
memory->clearUnmapWrite();
// Clear unmap flags
memory->clearUnmapFlags();
profilingEnd(vcmd);
}
@@ -1357,31 +1361,20 @@ VirtualGPU::submitSvmMapMemory(amd::SvmMapMemoryCommand& vcmd)
profilingBegin(vcmd, true);
//check if the ptr is in the svm space
amd::Memory* svmMem = vcmd.getSvmMem();
if (NULL == svmMem) {
LogWarning("wrong svm address ");
vcmd.setStatus(CL_INVALID_VALUE);
return;
}
// Make sure we have memory for the command execution
gpu::Memory* memory = dev().getGpuMemory(svmMem);
gpu::Memory* memory = dev().getGpuMemory(vcmd.getSvmMem());
if (vcmd.mapFlags() & (CL_MAP_WRITE | CL_MAP_WRITE_INVALIDATE_REGION)) {
memory->saveWriteMapInfo(vcmd.origin(), vcmd.size(), vcmd.isEntireMemory());
}
memory->saveMapInfo(vcmd.origin(), vcmd.size(),
vcmd.mapFlags(), vcmd.isEntireMemory());
if (memory->mapMemory() != NULL) {
if (vcmd.mapFlags() & (CL_MAP_READ | CL_MAP_WRITE)) {
amd::Coord3D dstOrigin(0, 0, 0);
if (memory->cal()->buffer_) {
if (!blitMgr().copyBuffer(*memory,
*memory->mapMemory(), vcmd.origin(), dstOrigin,
vcmd.size(), vcmd.isEntireMemory())) {
LogError("submitSVMMapMemory() - copy failed");
vcmd.setStatus(CL_MAP_FAILURE);
}
assert(memory->cal()->buffer_ && "SVM memory can't be an image");
if (!blitMgr().copyBuffer(*memory, *memory->mapMemory(),
vcmd.origin(), dstOrigin, vcmd.size(), vcmd.isEntireMemory())) {
LogError("submitSVMMapMemory() - copy failed");
vcmd.setStatus(CL_MAP_FAILURE);
}
}
}
@@ -1399,30 +1392,18 @@ VirtualGPU::submitSvmUnmapMemory(amd::SvmUnmapMemoryCommand& vcmd)
amd::ScopedLock lock(execution());
profilingBegin(vcmd, true);
amd::Memory* svmMem = vcmd.getSvmMem();
if (NULL == svmMem) {
LogWarning("wrong svm address ");
vcmd.setStatus(CL_INVALID_VALUE);
return;
}
gpu::Memory* memory = dev().getGpuMemory(svmMem);
gpu::Memory* memory = dev().getGpuMemory(vcmd.getSvmMem());
if (memory->mapMemory() != NULL) {
if (memory->isUnmapWrite()) {
amd::Coord3D srcOrigin(0, 0, 0);
// Target is a remote resource, so copy
assert(memory->mapMemory() != NULL);
if (memory->cal()->buffer_) {
if (!blitMgr().copyBuffer(
*memory->mapMemory(), *memory,
srcOrigin,
memory->writeMapInfo()->origin_,
memory->writeMapInfo()->region_,
memory->writeMapInfo()->entire_)) {
LogError("submitUnmapMemory() - copy failed");
vcmd.setStatus(CL_OUT_OF_RESOURCES);
}
assert(memory->cal()->buffer_ && "SVM memory can't be an image");
if (!blitMgr().copyBuffer(*memory->mapMemory(), *memory, srcOrigin,
memory->writeMapInfo()->origin_, memory->writeMapInfo()->region_,
memory->writeMapInfo()->entire_)) {
LogError("submitSvmUnmapMemory() - copy failed");
vcmd.setStatus(CL_OUT_OF_RESOURCES);
}
}
}