intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-12-26 15:03:02 +08:00
Code Issues Packages Projects Releases Wiki Activity

Add new SVM types

Browse Source 
Related-To: NEO-2917

Change-Id: Ica127129799c1e617a326a110348c2f70160b15c
Signed-off-by: Zbigniew Zdanowicz <zbigniew.zdanowicz@intel.com>
This commit is contained in:
Zbigniew Zdanowicz
2019-03-06 16:35:21 +01:00
committed by sys_ocldev
parent 653986aea1
commit 971cbd55f3
34 changed files with 1019 additions and 180 deletions
Download Patch File Download Diff File Expand all files Collapse all files

188
runtime/command_queue/enqueue_svm.h
View File

@@ -6,9 +6,11 @@
*/
#pragma once
#include "runtime/built_ins/built_ins.h"
#include "runtime/command_queue/command_queue_hw.h"
#include "runtime/command_queue/enqueue_common.h"
#include "runtime/event/event.h"
#include "runtime/memory_manager/surface.h"
#include "runtime/memory_manager/svm_memory_manager.h"
#include <new>
@@ -67,25 +69,78 @@ cl_int CommandQueueHw<GfxFamily>::enqueueSVMMap(cl_bool blockingMap,
const cl_event *eventWaitList,
cl_event *event) {
NEO::GraphicsAllocation *svmAllocation = context->getSVMAllocsManager()->getSVMAlloc(svmPtr);
if (svmAllocation == nullptr) {
auto svmData = context->getSVMAllocsManager()->getSVMAlloc(svmPtr);
if (svmData == nullptr) {
return CL_INVALID_VALUE;
}
bool blocking = blockingMap == CL_TRUE;
NullSurface s;
Surface *surfaces[] = {&s};
if (context->isProvidingPerformanceHints()) {
context->providePerformanceHint(CL_CONTEXT_DIAGNOSTICS_LEVEL_GOOD_INTEL, CL_ENQUEUE_SVM_MAP_DOESNT_REQUIRE_COPY_DATA, svmPtr);
if (svmData->gpuAllocation->getAllocationType() == GraphicsAllocation::AllocationType::SVM_ZERO_COPY) {
NullSurface s;
Surface *surfaces[] = {&s};
if (context->isProvidingPerformanceHints()) {
context->providePerformanceHint(CL_CONTEXT_DIAGNOSTICS_LEVEL_GOOD_INTEL, CL_ENQUEUE_SVM_MAP_DOESNT_REQUIRE_COPY_DATA, svmPtr);
}
enqueueHandler<CL_COMMAND_SVM_MAP>(surfaces,
blocking,
MultiDispatchInfo(),
numEventsInWaitList,
eventWaitList,
event);
return CL_SUCCESS;
} else {
auto svmOperation = context->getSVMAllocsManager()->getSvmMapOperation(svmPtr);
if (svmOperation) {
NullSurface s;
Surface *surfaces[] = {&s};
enqueueHandler<CL_COMMAND_SVM_MAP>(surfaces,
blocking,
MultiDispatchInfo(),
numEventsInWaitList,
eventWaitList,
event);
return CL_SUCCESS;
}
MultiDispatchInfo dispatchInfo;
auto &builder = getDevice().getExecutionEnvironment()->getBuiltIns()->getBuiltinDispatchInfoBuilder(EBuiltInOps::CopyBufferToBuffer,
this->getContext(), this->getDevice());
BuiltInOwnershipWrapper builtInLock(builder, this->context);
GeneralSurface dstSurface(svmData->cpuAllocation);
GeneralSurface srcSurface(svmData->gpuAllocation);
Surface *surfaces[] = {&dstSurface, &srcSurface};
void *svmBasePtr = svmData->cpuAllocation->getUnderlyingBuffer();
size_t svmOffset = ptrDiff(svmPtr, svmBasePtr);
BuiltinDispatchInfoBuilder::BuiltinOpParams dc;
dc.dstPtr = reinterpret_cast<void *>(svmData->cpuAllocation->getGpuAddressToPatch());
dc.dstSvmAlloc = svmData->cpuAllocation;
dc.dstOffset = {svmOffset, 0, 0};
dc.srcPtr = reinterpret_cast<void *>(svmData->gpuAllocation->getGpuAddressToPatch());
dc.srcSvmAlloc = svmData->gpuAllocation;
dc.srcOffset = {svmOffset, 0, 0};
dc.size = {size, 0, 0};
builder.buildDispatchInfos(dispatchInfo, dc);
enqueueHandler<CL_COMMAND_READ_BUFFER>(
surfaces,
blocking,
dispatchInfo,
numEventsInWaitList,
eventWaitList,
event);
if (event) {
castToObjectOrAbort<Event>(*event)->setCmdType(CL_COMMAND_SVM_MAP);
}
bool readOnlyMap = SVMAllocsManager::mapFlagIsReadOnly(mapFlags);
context->getSVMAllocsManager()->insertSvmMapOperation(svmPtr, size, svmBasePtr, svmOffset, readOnlyMap);
return CL_SUCCESS;
}
enqueueHandler<CL_COMMAND_SVM_MAP>(surfaces,
blockingMap ? true : false,
MultiDispatchInfo(),
numEventsInWaitList,
eventWaitList,
event);
return CL_SUCCESS;
}
template <typename GfxFamily>
@@ -94,21 +149,82 @@ cl_int CommandQueueHw<GfxFamily>::enqueueSVMUnmap(void *svmPtr,
const cl_event *eventWaitList,
cl_event *event) {
NEO::GraphicsAllocation *svmAllocation = context->getSVMAllocsManager()->getSVMAlloc(svmPtr);
if (svmAllocation == nullptr) {
auto svmData = context->getSVMAllocsManager()->getSVMAlloc(svmPtr);
if (svmData == nullptr) {
return CL_INVALID_VALUE;
}
NullSurface s;
Surface *surfaces[] = {&s};
enqueueHandler<CL_COMMAND_SVM_UNMAP>(surfaces,
false,
MultiDispatchInfo(),
numEventsInWaitList,
eventWaitList,
event);
if (svmData->gpuAllocation->getAllocationType() == GraphicsAllocation::AllocationType::SVM_ZERO_COPY) {
NullSurface s;
Surface *surfaces[] = {&s};
enqueueHandler<CL_COMMAND_SVM_UNMAP>(surfaces,
false,
MultiDispatchInfo(),
numEventsInWaitList,
eventWaitList,
event);
return CL_SUCCESS;
return CL_SUCCESS;
} else {
auto svmOperation = context->getSVMAllocsManager()->getSvmMapOperation(svmPtr);
if (!svmOperation) {
NullSurface s;
Surface *surfaces[] = {&s};
enqueueHandler<CL_COMMAND_SVM_UNMAP>(surfaces,
false,
MultiDispatchInfo(),
numEventsInWaitList,
eventWaitList,
event);
return CL_SUCCESS;
}
if (svmOperation->readOnlyMap) {
NullSurface s;
Surface *surfaces[] = {&s};
enqueueHandler<CL_COMMAND_SVM_UNMAP>(surfaces,
false,
MultiDispatchInfo(),
numEventsInWaitList,
eventWaitList,
event);
context->getSVMAllocsManager()->removeSvmMapOperation(svmPtr);
return CL_SUCCESS;
}
MultiDispatchInfo dispatchInfo;
auto &builder = getDevice().getExecutionEnvironment()->getBuiltIns()->getBuiltinDispatchInfoBuilder(EBuiltInOps::CopyBufferToBuffer,
this->getContext(), this->getDevice());
BuiltInOwnershipWrapper builtInLock(builder, this->context);
GeneralSurface dstSurface(svmData->gpuAllocation);
GeneralSurface srcSurface(svmData->cpuAllocation);
Surface *surfaces[] = {&dstSurface, &srcSurface};
BuiltinDispatchInfoBuilder::BuiltinOpParams dc;
dc.dstPtr = reinterpret_cast<void *>(svmData->gpuAllocation->getGpuAddressToPatch());
dc.dstSvmAlloc = svmData->gpuAllocation;
dc.dstOffset = {svmOperation->offset, 0, 0};
dc.srcPtr = reinterpret_cast<void *>(svmData->cpuAllocation->getGpuAddressToPatch());
dc.srcSvmAlloc = svmData->cpuAllocation;
dc.srcOffset = {svmOperation->offset, 0, 0};
dc.size = {svmOperation->regionSize, 0, 0};
builder.buildDispatchInfos(dispatchInfo, dc);
enqueueHandler<CL_COMMAND_READ_BUFFER>(
surfaces,
false,
dispatchInfo,
numEventsInWaitList,
eventWaitList,
event);
if (event) {
castToObjectOrAbort<Event>(*event)->setCmdType(CL_COMMAND_SVM_UNMAP);
}
context->getSVMAllocsManager()->removeSvmMapOperation(svmPtr);
return CL_SUCCESS;
}
}
template <typename GfxFamily>
@@ -157,9 +273,9 @@ cl_int CommandQueueHw<GfxFamily>::enqueueSVMMemcpy(cl_bool blockingCopy,
const cl_event *eventWaitList,
cl_event *event) {
GraphicsAllocation *pDstSvmAlloc = context->getSVMAllocsManager()->getSVMAlloc(dstPtr);
GraphicsAllocation *pSrcSvmAlloc = context->getSVMAllocsManager()->getSVMAlloc(srcPtr);
if ((pDstSvmAlloc == nullptr) || (pSrcSvmAlloc == nullptr)) {
auto dstSvmData = context->getSVMAllocsManager()->getSVMAlloc(dstPtr);
auto srcSvmData = context->getSVMAllocsManager()->getSVMAlloc(srcPtr);
if ((dstSvmData == nullptr) || (srcSvmData == nullptr)) {
return CL_INVALID_VALUE;
}
@@ -173,14 +289,14 @@ cl_int CommandQueueHw<GfxFamily>::enqueueSVMMemcpy(cl_bool blockingCopy,
BuiltinDispatchInfoBuilder::BuiltinOpParams operationParams;
operationParams.srcPtr = const_cast<void *>(srcPtr);
operationParams.dstPtr = dstPtr;
operationParams.srcSvmAlloc = pSrcSvmAlloc;
operationParams.dstSvmAlloc = pDstSvmAlloc;
operationParams.srcSvmAlloc = srcSvmData->gpuAllocation;
operationParams.dstSvmAlloc = dstSvmData->gpuAllocation;
operationParams.srcOffset = {0, 0, 0};
operationParams.dstOffset = {0, 0, 0};
operationParams.size = {size, 0, 0};
builder.buildDispatchInfos(dispatchInfo, operationParams);
GeneralSurface s1(pSrcSvmAlloc), s2(pDstSvmAlloc);
GeneralSurface s1(srcSvmData->gpuAllocation), s2(dstSvmData->gpuAllocation);
Surface *surfaces[] = {&s1, &s2};
enqueueHandler<CL_COMMAND_SVM_MEMCPY>(
@@ -203,8 +319,8 @@ cl_int CommandQueueHw<GfxFamily>::enqueueSVMMemFill(void *svmPtr,
const cl_event *eventWaitList,
cl_event *event) {
NEO::GraphicsAllocation *pSvmAlloc = context->getSVMAllocsManager()->getSVMAlloc(svmPtr);
if (pSvmAlloc == nullptr) {
auto svmData = context->getSVMAllocsManager()->getSVMAlloc(svmPtr);
if (svmData == nullptr) {
return CL_INVALID_VALUE;
}
@@ -243,12 +359,12 @@ cl_int CommandQueueHw<GfxFamily>::enqueueSVMMemFill(void *svmPtr,
patternAllocation->getUnderlyingBuffer(), patternAllocation, false, false, true);
operationParams.srcMemObj = &patternMemObj;
operationParams.dstPtr = svmPtr;
operationParams.dstSvmAlloc = pSvmAlloc;
operationParams.dstSvmAlloc = svmData->gpuAllocation;
operationParams.dstOffset = {0, 0, 0};
operationParams.size = {size, 0, 0};
builder.buildDispatchInfos(dispatchInfo, operationParams);
GeneralSurface s1(pSvmAlloc);
GeneralSurface s1(svmData->gpuAllocation);
GeneralSurface s2(patternAllocation);
Surface *surfaces[] = {&s1, &s2};