intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-12-24 21:18:24 +08:00
Code Issues Packages Projects Releases Wiki Activity

Program GPU atomics on stateless kernels for L0

Browse Source 
Signed-off-by: Jaime Arteaga <jaime.a.arteaga.molina@intel.com>
This commit is contained in:
Jaime Arteaga
2021-06-15 02:02:59 +00:00
committed by Compute-Runtime-Automation
parent c9c752526e
commit a481c28e55
4 changed files with 45 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
shared/source/command_container/command_encoder.h
View File

@@ -41,6 +41,7 @@ struct EncodeDispatchKernel {
Device *device,
PreemptionMode preemptionMode,
bool &requiresUncachedMocs,
bool useGlobalAtomics,
uint32_t &partitionCount,
bool isInternal);
@@ -212,7 +213,7 @@ template <typename GfxFamily>
struct EncodeStateBaseAddress {
using STATE_BASE_ADDRESS = typename GfxFamily::STATE_BASE_ADDRESS;
static void encode(CommandContainer &container, STATE_BASE_ADDRESS &sbaCmd);
static void encode(CommandContainer &container, STATE_BASE_ADDRESS &sbaCmd, uint32_t statelessMocsIndex);
static void encode(CommandContainer &container, STATE_BASE_ADDRESS &sbaCmd, uint32_t statelessMocsIndex, bool useGlobalAtomics);
};
template <typename GfxFamily>

12
shared/source/command_container/command_encoder_bdw_plus.inl
View File

@@ -25,8 +25,8 @@ namespace NEO {
template <typename Family>
void EncodeDispatchKernel<Family>::encode(CommandContainer &container,
const void *pThreadGroupDimensions, bool isIndirect, bool isPredicate, DispatchKernelEncoderI *dispatchInterface,
uint64_t eventAddress, bool isTimestampEvent, bool L3FlushEnable, Device *device, PreemptionMode preemptionMode, bool &requiresUncachedMocs,
uint32_t &partitionCount, bool isInternal) {
uint64_t eventAddress, bool isTimestampEvent, bool L3FlushEnable, Device *device, PreemptionMode preemptionMode,
bool &requiresUncachedMocs, bool useGlobalAtomics, uint32_t &partitionCount, bool isInternal) {
using MEDIA_STATE_FLUSH = typename Family::MEDIA_STATE_FLUSH;
using MEDIA_INTERFACE_DESCRIPTOR_LOAD = typename Family::MEDIA_INTERFACE_DESCRIPTOR_LOAD;
@@ -168,7 +168,7 @@ void EncodeDispatchKernel<Family>::encode(CommandContainer &container,
auto gmmHelper = container.getDevice()->getGmmHelper();
uint32_t statelessMocsIndex =
requiresUncachedMocs ? (gmmHelper->getMOCS(GMM_RESOURCE_USAGE_OCL_BUFFER_CACHELINE_MISALIGNED) >> 1) : (gmmHelper->getMOCS(GMM_RESOURCE_USAGE_OCL_BUFFER) >> 1);
EncodeStateBaseAddress<Family>::encode(container, sba, statelessMocsIndex);
EncodeStateBaseAddress<Family>::encode(container, sba, statelessMocsIndex, false);
container.setDirtyStateForAllHeaps(false);
requiresUncachedMocs = false;
}
@@ -351,11 +351,11 @@ template <typename Family>
void EncodeStateBaseAddress<Family>::encode(CommandContainer &container, STATE_BASE_ADDRESS &sbaCmd) {
auto gmmHelper = container.getDevice()->getRootDeviceEnvironment().getGmmHelper();
uint32_t statelessMocsIndex = (gmmHelper->getMOCS(GMM_RESOURCE_USAGE_OCL_BUFFER) >> 1);
EncodeStateBaseAddress<Family>::encode(container, sbaCmd, statelessMocsIndex);
EncodeStateBaseAddress<Family>::encode(container, sbaCmd, statelessMocsIndex, false);
}
template <typename Family>
void EncodeStateBaseAddress<Family>::encode(CommandContainer &container, STATE_BASE_ADDRESS &sbaCmd, uint32_t statelessMocsIndex) {
void EncodeStateBaseAddress<Family>::encode(CommandContainer &container, STATE_BASE_ADDRESS &sbaCmd, uint32_t statelessMocsIndex, bool useGlobalAtomics) {
EncodeWA<Family>::encodeAdditionalPipelineSelect(*container.getDevice(), *container.getCommandStream(), true);
auto gmmHelper = container.getDevice()->getGmmHelper();
@@ -376,7 +376,7 @@ void EncodeStateBaseAddress<Family>::encode(CommandContainer &container, STATE_B
gmmHelper,
false,
MemoryCompressionState::NotApplicable,
false,
useGlobalAtomics,
1u);
auto pCmd = reinterpret_cast<STATE_BASE_ADDRESS *>(container.getCommandStream()->getSpace(sizeof(STATE_BASE_ADDRESS)));