intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-12-22 10:17:01 +08:00
Code Issues Packages Projects Releases Wiki Activity

Add dedicated map allocation

Browse Source 
Related-To: NEO-2917

Change-Id: Ieeca40f5faf29433a5c464d2c3ca3b8910695a9b
Signed-off-by: Zbigniew Zdanowicz <zbigniew.zdanowicz@intel.com>
This commit is contained in:
Zbigniew Zdanowicz
2019-04-08 14:49:35 +02:00
committed by sys_ocldev
parent 91a64c8518
commit e201725dd5
45 changed files with 287 additions and 77 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
runtime/command_queue/command_queue.cpp
View File

@@ -324,7 +324,7 @@ cl_int CommandQueue::enqueueWriteMemObjForUnmap(MemObj *memObj, void *mappedPtr,
if (!unmapInfo.readOnly) {
if (memObj->peekClMemObjType() == CL_MEM_OBJECT_BUFFER) {
auto buffer = castToObject<Buffer>(memObj);
retVal = enqueueWriteBuffer(buffer, CL_TRUE, unmapInfo.offset[0], unmapInfo.size[0], mappedPtr,
retVal = enqueueWriteBuffer(buffer, CL_TRUE, unmapInfo.offset[0], unmapInfo.size[0], mappedPtr, memObj->getMapAllocation(),
eventsRequest.numEventsInWaitList, eventsRequest.eventWaitList, eventsRequest.outEvent);
} else {
auto image = castToObjectOrAbort<Image>(memObj);
@@ -360,8 +360,12 @@ cl_int CommandQueue::enqueueWriteMemObjForUnmap(MemObj *memObj, void *mappedPtr,
}
void *CommandQueue::enqueueReadMemObjForMap(TransferProperties &transferProperties, EventsRequest &eventsRequest, cl_int &errcodeRet) {
void *returnPtr = ptrOffset(transferProperties.memObj->getBasePtrForMap(),
transferProperties.memObj->calculateOffsetForMapping(transferProperties.offset) + transferProperties.mipPtrOffset);
void *basePtr = transferProperties.memObj->getBasePtrForMap();
size_t mapPtrOffset = transferProperties.memObj->calculateOffsetForMapping(transferProperties.offset) + transferProperties.mipPtrOffset;
if (transferProperties.memObj->peekClMemObjType() == CL_MEM_OBJECT_BUFFER) {
mapPtrOffset += transferProperties.memObj->getOffset();
}
void *returnPtr = ptrOffset(basePtr, mapPtrOffset);
if (!transferProperties.memObj->addMappedPtr(returnPtr, transferProperties.memObj->calculateMappedPtrLength(transferProperties.size),
transferProperties.mapFlags, transferProperties.size, transferProperties.offset, transferProperties.mipLevel)) {
@@ -371,8 +375,9 @@ void *CommandQueue::enqueueReadMemObjForMap(TransferProperties &transferProperti
if (transferProperties.memObj->peekClMemObjType() == CL_MEM_OBJECT_BUFFER) {
auto buffer = castToObject<Buffer>(transferProperties.memObj);
errcodeRet = enqueueReadBuffer(buffer, transferProperties.blocking, transferProperties.offset[0], transferProperties.size[0], returnPtr,
eventsRequest.numEventsInWaitList, eventsRequest.eventWaitList, eventsRequest.outEvent);
errcodeRet = enqueueReadBuffer(buffer, transferProperties.blocking, transferProperties.offset[0], transferProperties.size[0],
returnPtr, transferProperties.memObj->getMapAllocation(), eventsRequest.numEventsInWaitList,
eventsRequest.eventWaitList, eventsRequest.outEvent);
} else {
auto image = castToObjectOrAbort<Image>(transferProperties.memObj);
size_t readOrigin[4] = {transferProperties.offset[0], transferProperties.offset[1], transferProperties.offset[2], 0};

2
runtime/command_queue/command_queue.h
View File

@@ -207,6 +207,7 @@ class CommandQueue : public BaseObject<_cl_command_queue> {
virtual cl_int enqueueReadBuffer(Buffer *buffer, cl_bool blockingRead,
size_t offset, size_t size, void *ptr,
GraphicsAllocation *mapAllocation,
cl_uint numEventsInWaitList,
const cl_event *eventWaitList,
cl_event *event) {
@@ -224,6 +225,7 @@ class CommandQueue : public BaseObject<_cl_command_queue> {
virtual cl_int enqueueWriteBuffer(Buffer *buffer, cl_bool blockingWrite,
size_t offset, size_t cb, const void *ptr,
GraphicsAllocation *mapAllocation,
cl_uint numEventsInWaitList,
const cl_event *eventWaitList,
cl_event *event) {

2
runtime/command_queue/command_queue_hw.h
View File

@@ -197,6 +197,7 @@ class CommandQueueHw : public CommandQueue {
size_t offset,
size_t size,
void *ptr,
GraphicsAllocation *mapAllocation,
cl_uint numEventsInWaitList,
const cl_event *eventWaitList,
cl_event *event) override;
@@ -231,6 +232,7 @@ class CommandQueueHw : public CommandQueue {
size_t offset,
size_t cb,
const void *ptr,
GraphicsAllocation *mapAllocation,
cl_uint numEventsInWaitList,
const cl_event *eventWaitList,
cl_event *event) override;

25
runtime/command_queue/enqueue_read_buffer.h
View File

@@ -28,6 +28,7 @@ cl_int CommandQueueHw<GfxFamily>::enqueueReadBuffer(
size_t offset,
size_t size,
void *ptr,
GraphicsAllocation *mapAllocation,
cl_uint numEventsInWaitList,
const cl_event *eventWaitList,
cl_event *event) {
@@ -84,21 +85,29 @@ cl_int CommandQueueHw<GfxFamily>::enqueueReadBuffer(
auto &builder = getDevice().getExecutionEnvironment()->getBuiltIns()->getBuiltinDispatchInfoBuilder(EBuiltInOps::CopyBufferToBuffer,
this->getContext(), this->getDevice());
BuiltInOwnershipWrapper builtInLock(builder, this->context);
void *dstPtr = ptr;
MemObjSurface bufferSurf(buffer);
HostPtrSurface hostPtrSurf(dstPtr, size);
Surface *surfaces[] = {&bufferSurf, &hostPtrSurf};
GeneralSurface mapSurface;
Surface *surfaces[] = {&bufferSurf, nullptr};
if (size != 0) {
bool status = getCommandStreamReceiver().createAllocationForHostSurface(hostPtrSurf, true);
if (!status) {
return CL_OUT_OF_RESOURCES;
if (mapAllocation) {
surfaces[1] = &mapSurface;
mapSurface.setGraphicsAllocation(mapAllocation);
//get offset between base cpu ptr of map allocation and dst ptr
size_t dstOffset = ptrDiff(dstPtr, mapAllocation->getUnderlyingBuffer());
dstPtr = reinterpret_cast<void *>(mapAllocation->getGpuAddress() + dstOffset);
} else {
surfaces[1] = &hostPtrSurf;
if (size != 0) {
bool status = getCommandStreamReceiver().createAllocationForHostSurface(hostPtrSurf, true);
if (!status) {
return CL_OUT_OF_RESOURCES;
}
dstPtr = reinterpret_cast<void *>(hostPtrSurf.getAllocation()->getGpuAddress());
}
dstPtr = reinterpret_cast<void *>(hostPtrSurf.getAllocation()->getGpuAddress());
}
void *alignedDstPtr = alignDown(dstPtr, 4);
size_t dstPtrOffset = ptrDiff(dstPtr, alignedDstPtr);

24
runtime/command_queue/enqueue_write_buffer.h
View File

@@ -27,6 +27,7 @@ cl_int CommandQueueHw<GfxFamily>::enqueueWriteBuffer(
size_t offset,
size_t size,
const void *ptr,
GraphicsAllocation *mapAllocation,
cl_uint numEventsInWaitList,
const cl_event *eventWaitList,
cl_event *event) {
@@ -89,16 +90,25 @@ cl_int CommandQueueHw<GfxFamily>::enqueueWriteBuffer(
HostPtrSurface hostPtrSurf(srcPtr, size, true);
MemObjSurface bufferSurf(buffer);
Surface *surfaces[] = {&bufferSurf, &hostPtrSurf};
GeneralSurface mapSurface;
Surface *surfaces[] = {&bufferSurf, nullptr};
if (size != 0) {
bool status = getCommandStreamReceiver().createAllocationForHostSurface(hostPtrSurf, false);
if (!status) {
return CL_OUT_OF_RESOURCES;
if (mapAllocation) {
surfaces[1] = &mapSurface;
mapSurface.setGraphicsAllocation(mapAllocation);
//get offset between base cpu ptr of map allocation and dst ptr
size_t srcOffset = ptrDiff(srcPtr, mapAllocation->getUnderlyingBuffer());
srcPtr = reinterpret_cast<void *>(mapAllocation->getGpuAddress() + srcOffset);
} else {
surfaces[1] = &hostPtrSurf;
if (size != 0) {
bool status = getCommandStreamReceiver().createAllocationForHostSurface(hostPtrSurf, false);
if (!status) {
return CL_OUT_OF_RESOURCES;
}
srcPtr = reinterpret_cast<void *>(hostPtrSurf.getAllocation()->getGpuAddress());
}
srcPtr = reinterpret_cast<void *>(hostPtrSurf.getAllocation()->getGpuAddress());
}
void *alignedSrcPtr = alignDown(srcPtr, 4);
size_t srcPtrOffset = ptrDiff(srcPtr, alignedSrcPtr);