intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2026-01-05 09:09:04 +08:00
Code Issues Packages Projects Releases Wiki Activity

Change DG2 l1 cache policy to WB

Browse Source 
With compiler LSC WAs this gives better performance.

If debugger is active, policy will not be changed ie.
will be WBP.

Related-To: NEO-7003

Signed-off-by: Dominik Dabek <dominik.dabek@intel.com>
This commit is contained in:
Dominik Dabek
2022-08-23 11:48:18 +00:00
committed by Compute-Runtime-Automation
parent 2d949a9f3d
commit a820e73dd7
48 changed files with 447 additions and 128 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
shared/source/command_container/command_encoder.h
View File

@@ -283,7 +283,7 @@ struct EncodeSurfaceState {
static void encodeBuffer(EncodeSurfaceStateArgs &args);
static void encodeExtraBufferParams(EncodeSurfaceStateArgs &args);
static void encodeImplicitScalingParams(const EncodeSurfaceStateArgs &args);
static void encodeExtraCacheSettings(R_SURFACE_STATE *surfaceState, const HardwareInfo &hwInfo);
static void encodeExtraCacheSettings(R_SURFACE_STATE *surfaceState, const EncodeSurfaceStateArgs &args);
static void appendBufferSurfaceState(EncodeSurfaceStateArgs &args);
static constexpr uintptr_t getSurfaceBaseAddressAlignmentMask() {

2
shared/source/command_container/command_encoder.inl
View File

@@ -467,7 +467,7 @@ size_t EncodeSurfaceState<Family>::pushBindingTableAndSurfaceStates(IndirectHeap
}
template <typename Family>
inline void EncodeSurfaceState<Family>::encodeExtraCacheSettings(R_SURFACE_STATE *surfaceState, const HardwareInfo &hwInfo) {}
inline void EncodeSurfaceState<Family>::encodeExtraCacheSettings(R_SURFACE_STATE *surfaceState, const EncodeSurfaceStateArgs &args) {}
template <typename Family>
void EncodeSurfaceState<Family>::setImageAuxParamsForCCS(R_SURFACE_STATE *surfaceState, Gmm *gmm) {

6
shared/source/command_container/command_encoder_bdw_and_later.inl
View File

@@ -371,6 +371,7 @@ void EncodeStateBaseAddress<Family>::encode(EncodeStateBaseAddressArgs<Family> &
auto dsh = args.container->isHeapDirty(HeapType::DYNAMIC_STATE) ? args.container->getIndirectHeap(HeapType::DYNAMIC_STATE) : nullptr;
auto ioh = args.container->isHeapDirty(HeapType::INDIRECT_OBJECT) ? args.container->getIndirectHeap(HeapType::INDIRECT_OBJECT) : nullptr;
auto ssh = args.container->isHeapDirty(HeapType::SURFACE_STATE) ? args.container->getIndirectHeap(HeapType::SURFACE_STATE) : nullptr;
auto isDebuggerActive = device.isDebuggerActive() || device.getDebugger() != nullptr;
StateBaseAddressHelperArgs<Family> stateBaseAddressHelperArgs = {
0, // generalStateBase
@@ -391,7 +392,8 @@ void EncodeStateBaseAddress<Family>::encode(EncodeStateBaseAddressArgs<Family> &
false, // isMultiOsContextCapable
args.useGlobalAtomics, // useGlobalAtomics
false, // areMultipleSubDevicesInContext
false // overrideSurfaceStateBaseAddress
false, // overrideSurfaceStateBaseAddress
isDebuggerActive // isDebuggerActive
};
StateBaseAddressHelper<Family>::programStateBaseAddress(stateBaseAddressHelperArgs);
@@ -457,7 +459,7 @@ inline void EncodeWA<GfxFamily>::adjustCompressionFormatForPlanarImage(uint32_t
template <typename GfxFamily>
inline void EncodeSurfaceState<GfxFamily>::encodeExtraBufferParams(EncodeSurfaceStateArgs &args) {
auto surfaceState = reinterpret_cast<R_SURFACE_STATE *>(args.outMemory);
encodeExtraCacheSettings(surfaceState, *args.gmmHelper->getHardwareInfo());
encodeExtraCacheSettings(surfaceState, args);
}
template <typename GfxFamily>

10
shared/source/command_container/command_encoder_xe_hpg_core_and_later.inl
View File

@@ -13,16 +13,16 @@
namespace NEO {
template <>
void EncodeSurfaceState<Family>::encodeExtraCacheSettings(R_SURFACE_STATE *surfaceState, const HardwareInfo &hwInfo) {
void EncodeSurfaceState<Family>::encodeExtraCacheSettings(R_SURFACE_STATE *surfaceState, const EncodeSurfaceStateArgs &args) {
using L1_CACHE_POLICY = typename R_SURFACE_STATE::L1_CACHE_POLICY;
const auto &hwInfo = *args.gmmHelper->getHardwareInfo();
auto hwInfoConfig = HwInfoConfig::get(hwInfo.platform.eProductFamily);
auto cachePolicy = static_cast<L1_CACHE_POLICY>(hwInfoConfig->getL1CachePolicy());
surfaceState->setL1CachePolicyL1CacheControl(cachePolicy);
auto cachePolicy = static_cast<L1_CACHE_POLICY>(hwInfoConfig->getL1CachePolicy(args.isDebuggerActive));
if (DebugManager.flags.OverrideL1CacheControlInSurfaceState.get() != -1 &&
DebugManager.flags.ForceAllResourcesUncached.get() == false) {
surfaceState->setL1CachePolicyL1CacheControl(static_cast<L1_CACHE_POLICY>(DebugManager.flags.OverrideL1CacheControlInSurfaceState.get()));
cachePolicy = static_cast<L1_CACHE_POLICY>(DebugManager.flags.OverrideL1CacheControlInSurfaceState.get());
}
surfaceState->setL1CachePolicyL1CacheControl(cachePolicy);
}
template <typename GfxFamily>

11
shared/source/command_container/command_encoder_xehp_and_later.inl
View File

@@ -473,11 +473,13 @@ void EncodeStateBaseAddress<Family>::setSbaAddressesForDebugger(NEO::Debugger::S
template <typename Family>
void EncodeStateBaseAddress<Family>::encode(EncodeStateBaseAddressArgs<Family> &args) {
auto gmmHelper = args.container->getDevice()->getRootDeviceEnvironment().getGmmHelper();
auto &device = *args.container->getDevice();
auto gmmHelper = device.getRootDeviceEnvironment().getGmmHelper();
auto dsh = args.container->isHeapDirty(HeapType::DYNAMIC_STATE) ? args.container->getIndirectHeap(HeapType::DYNAMIC_STATE) : nullptr;
auto ioh = args.container->isHeapDirty(HeapType::INDIRECT_OBJECT) ? args.container->getIndirectHeap(HeapType::INDIRECT_OBJECT) : nullptr;
auto ssh = args.container->isHeapDirty(HeapType::SURFACE_STATE) ? args.container->getIndirectHeap(HeapType::SURFACE_STATE) : nullptr;
auto isDebuggerActive = device.isDebuggerActive() || device.getDebugger() != nullptr;
StateBaseAddressHelperArgs<Family> stateBaseAddressHelperArgs = {
0, // generalStateBase
@@ -498,7 +500,8 @@ void EncodeStateBaseAddress<Family>::encode(EncodeStateBaseAddressArgs<Family> &
args.multiOsContextCapable, // isMultiOsContextCapable
args.useGlobalAtomics, // useGlobalAtomics
false, // areMultipleSubDevicesInContext
false // overrideSurfaceStateBaseAddress
false, // overrideSurfaceStateBaseAddress
isDebuggerActive // isDebuggerActive
};
StateBaseAddressHelper<Family>::programStateBaseAddress(stateBaseAddressHelperArgs);
@@ -506,7 +509,7 @@ void EncodeStateBaseAddress<Family>::encode(EncodeStateBaseAddressArgs<Family> &
auto cmdSpace = StateBaseAddressHelper<Family>::getSpaceForSbaCmd(*args.container->getCommandStream());
*cmdSpace = args.sbaCmd;
auto &hwInfo = args.container->getDevice()->getHardwareInfo();
auto &hwInfo = device.getHardwareInfo();
auto &hwInfoConfig = *HwInfoConfig::get(hwInfo.platform.eProductFamily);
if (hwInfoConfig.isAdditionalStateBaseAddressWARequired(hwInfo)) {
cmdSpace = StateBaseAddressHelper<Family>::getSpaceForSbaCmd(*args.container->getCommandStream());
@@ -630,7 +633,7 @@ void EncodeSurfaceState<Family>::encodeExtraBufferParams(EncodeSurfaceStateArgs
surfaceState->setMemoryObjectControlState(args.gmmHelper->getMOCS(GMM_RESOURCE_USAGE_OCL_BUFFER_CONST));
}
encodeExtraCacheSettings(surfaceState, *args.gmmHelper->getHardwareInfo());
encodeExtraCacheSettings(surfaceState, args);
encodeImplicitScalingParams(args);

1
shared/source/command_container/definitions/encode_surface_state_args_base.h
View File

@@ -31,6 +31,7 @@ struct EncodeSurfaceStateArgsBase {
bool useGlobalAtomics = false;
bool areMultipleSubDevicesInContext = false;
bool implicitScaling = false;
bool isDebuggerActive = false;
protected:
EncodeSurfaceStateArgsBase() = default;

3
shared/source/command_stream/command_stream_receiver_hw_base.inl
View File

@@ -428,7 +428,8 @@ CompletionStamp CommandStreamReceiverHw<GfxFamily>::flushTask(
isMultiOsContextCapable(), // isMultiOsContextCapable
dispatchFlags.useGlobalAtomics, // useGlobalAtomics
dispatchFlags.areMultipleSubDevicesInContext, // areMultipleSubDevicesInContext
false // overrideSurfaceStateBaseAddress
false, // overrideSurfaceStateBaseAddress
debuggingEnabled || device.isDebuggerActive() // isDebuggerActive
};
StateBaseAddressHelper<GfxFamily>::programStateBaseAddress(args);

6
shared/source/helpers/cache_policy.h
View File

@@ -21,13 +21,13 @@ bool isL3Capable(const GraphicsAllocation &graphicsAllocation);
template <PRODUCT_FAMILY gfxProduct>
struct L1CachePolicyHelper {
static const char *getCachingPolicyOptions();
static const char *getCachingPolicyOptions(bool isDebuggerActive);
static uint32_t getDefaultL1CachePolicy();
static uint32_t getDefaultL1CachePolicy(bool isDebuggerActive);
static uint32_t getUncachedL1CachePolicy();
static uint32_t getL1CachePolicy();
static uint32_t getL1CachePolicy(bool isDebuggerActive);
};
} // namespace NEO

4
shared/source/helpers/cache_policy_base.inl
View File

@@ -12,14 +12,14 @@
namespace NEO {
template <PRODUCT_FAMILY gfxProduct>
uint32_t L1CachePolicyHelper<gfxProduct>::getL1CachePolicy() {
uint32_t L1CachePolicyHelper<gfxProduct>::getL1CachePolicy(bool isDebuggerActive) {
if (DebugManager.flags.ForceAllResourcesUncached.get()) {
return L1CachePolicyHelper<gfxProduct>::getUncachedL1CachePolicy();
}
if (DebugManager.flags.OverrideL1CachePolicyInSurfaceStateAndStateless.get() != -1) {
return DebugManager.flags.OverrideL1CachePolicyInSurfaceStateAndStateless.get();
}
return L1CachePolicyHelper<gfxProduct>::getDefaultL1CachePolicy();
return L1CachePolicyHelper<gfxProduct>::getDefaultL1CachePolicy(isDebuggerActive);
}
} // namespace NEO

4
shared/source/helpers/cache_policy_bdw_and_later.inl
View File

@@ -11,12 +11,12 @@
namespace NEO {
template <PRODUCT_FAMILY gfxProduct>
const char *L1CachePolicyHelper<gfxProduct>::getCachingPolicyOptions() {
const char *L1CachePolicyHelper<gfxProduct>::getCachingPolicyOptions(bool isDebuggerActive) {
return nullptr;
}
template <PRODUCT_FAMILY gfxProduct>
uint32_t L1CachePolicyHelper<gfxProduct>::getDefaultL1CachePolicy() {
uint32_t L1CachePolicyHelper<gfxProduct>::getDefaultL1CachePolicy(bool isDebuggerActive) {
return 0u;
}

6
shared/source/helpers/cache_policy_dg2_and_later.inl
View File

@@ -12,14 +12,14 @@
namespace NEO {
template <PRODUCT_FAMILY gfxProduct>
const char *L1CachePolicyHelper<gfxProduct>::getCachingPolicyOptions() {
const char *L1CachePolicyHelper<gfxProduct>::getCachingPolicyOptions(bool isDebuggerActive) {
using GfxFamily = typename HwMapper<gfxProduct>::GfxFamily;
static constexpr const char *writeBackCachingPolicy = "-cl-store-cache-default=7 -cl-load-cache-default=4";
static constexpr const char *writeByPassCachingPolicy = "-cl-store-cache-default=2 -cl-load-cache-default=4";
static constexpr const char *uncachedCachingPolicy = "-cl-store-cache-default=1 -cl-load-cache-default=1";
switch (L1CachePolicyHelper<gfxProduct>::getL1CachePolicy()) {
switch (L1CachePolicyHelper<gfxProduct>::getL1CachePolicy(isDebuggerActive)) {
case GfxFamily::STATE_BASE_ADDRESS::L1_CACHE_POLICY_WBP:
return writeByPassCachingPolicy;
case GfxFamily::STATE_BASE_ADDRESS::L1_CACHE_POLICY_WB:
@@ -32,7 +32,7 @@ const char *L1CachePolicyHelper<gfxProduct>::getCachingPolicyOptions() {
}
template <PRODUCT_FAMILY gfxProduct>
uint32_t L1CachePolicyHelper<gfxProduct>::getDefaultL1CachePolicy() {
uint32_t L1CachePolicyHelper<gfxProduct>::getDefaultL1CachePolicy(bool isDebuggerActive) {
using GfxFamily = typename HwMapper<gfxProduct>::GfxFamily;
return GfxFamily::STATE_BASE_ADDRESS::L1_CACHE_POLICY_WBP;
}

4
shared/source/helpers/compiler_hw_info_config.h
View File

@@ -30,7 +30,7 @@ class CompilerHwInfoConfig {
virtual bool isForceToStatelessRequired() const = 0;
virtual void adjustHwInfoForIgc(HardwareInfo &hwInfo) const = 0;
virtual void setProductConfigForHwInfo(HardwareInfo &hwInfo, AheadOfTimeConfig config) const = 0;
virtual const char *getCachingPolicyOptions() const = 0;
virtual const char *getCachingPolicyOptions(bool isDebuggerActive) const = 0;
};
template <PRODUCT_FAMILY gfxProduct>
@@ -47,7 +47,7 @@ class CompilerHwInfoConfigHw : public CompilerHwInfoConfig {
bool isForceToStatelessRequired() const override;
void adjustHwInfoForIgc(HardwareInfo &hwInfo) const override;
void setProductConfigForHwInfo(HardwareInfo &hwInfo, AheadOfTimeConfig config) const override;
const char *getCachingPolicyOptions() const override;
const char *getCachingPolicyOptions(bool isDebuggerActive) const override;
protected:
CompilerHwInfoConfigHw() = default;

4
shared/source/helpers/compiler_hw_info_config_base.inl
View File

@@ -26,8 +26,8 @@ void CompilerHwInfoConfigHw<gfxProduct>::adjustHwInfoForIgc(HardwareInfo &hwInfo
}
template <PRODUCT_FAMILY gfxProduct>
const char *CompilerHwInfoConfigHw<gfxProduct>::getCachingPolicyOptions() const {
return L1CachePolicyHelper<gfxProduct>::getCachingPolicyOptions();
const char *CompilerHwInfoConfigHw<gfxProduct>::getCachingPolicyOptions(bool isDebuggerActive) const {
return L1CachePolicyHelper<gfxProduct>::getCachingPolicyOptions(isDebuggerActive);
};
} // namespace NEO

3
shared/source/helpers/state_base_address.h
View File

@@ -45,6 +45,7 @@ struct StateBaseAddressHelperArgs {
bool useGlobalAtomics = false;
bool areMultipleSubDevicesInContext = false;
bool overrideSurfaceStateBaseAddress = false;
bool isDebuggerActive = false;
};
template <typename GfxFamily>
@@ -60,7 +61,7 @@ struct StateBaseAddressHelper {
static void appendStateBaseAddressParameters(StateBaseAddressHelperArgs<GfxFamily> &args,
bool overrideBindlessSurfaceStateBase);
static void appendExtraCacheSettings(STATE_BASE_ADDRESS *stateBaseAddress, const HardwareInfo *hwInfo);
static void appendExtraCacheSettings(StateBaseAddressHelperArgs<GfxFamily> &args);
static void programBindingTableBaseAddress(LinearStream &commandStream, const IndirectHeap &ssh, GmmHelper *gmmHelper);

2
shared/source/helpers/state_base_address_bdw_and_later.inl
View File

@@ -29,6 +29,6 @@ void StateBaseAddressHelper<GfxFamily>::appendIohParameters(StateBaseAddressHelp
}
template <typename GfxFamily>
void StateBaseAddressHelper<GfxFamily>::appendExtraCacheSettings(STATE_BASE_ADDRESS *stateBaseAddress, const HardwareInfo *hwInfo) {}
void StateBaseAddressHelper<GfxFamily>::appendExtraCacheSettings(StateBaseAddressHelperArgs<GfxFamily> &args) {}
} // namespace NEO

10
shared/source/helpers/state_base_address_xe_hpg_core_and_later.inl
View File

@@ -6,13 +6,13 @@
*/
template <typename GfxFamily>
void StateBaseAddressHelper<GfxFamily>::appendExtraCacheSettings(STATE_BASE_ADDRESS *stateBaseAddress, const HardwareInfo *hwInfo) {
auto hwInfoConfig = HwInfoConfig::get(hwInfo->platform.eProductFamily);
auto cachePolicy = hwInfoConfig->getL1CachePolicy();
stateBaseAddress->setL1CachePolicyL1CacheControl(static_cast<typename STATE_BASE_ADDRESS::L1_CACHE_POLICY>(cachePolicy));
void StateBaseAddressHelper<GfxFamily>::appendExtraCacheSettings(StateBaseAddressHelperArgs<GfxFamily> &args) {
auto hwInfoConfig = HwInfoConfig::get(args.gmmHelper->getHardwareInfo()->platform.eProductFamily);
auto cachePolicy = hwInfoConfig->getL1CachePolicy(args.isDebuggerActive);
args.stateBaseAddressCmd->setL1CachePolicyL1CacheControl(static_cast<typename STATE_BASE_ADDRESS::L1_CACHE_POLICY>(cachePolicy));
if (DebugManager.flags.ForceStatelessL1CachingPolicy.get() != -1 &&
DebugManager.flags.ForceAllResourcesUncached.get() == false) {
stateBaseAddress->setL1CachePolicyL1CacheControl(static_cast<typename STATE_BASE_ADDRESS::L1_CACHE_POLICY>(DebugManager.flags.ForceStatelessL1CachingPolicy.get()));
args.stateBaseAddressCmd->setL1CachePolicyL1CacheControl(static_cast<typename STATE_BASE_ADDRESS::L1_CACHE_POLICY>(DebugManager.flags.ForceStatelessL1CachingPolicy.get()));
}
}

2
shared/source/helpers/state_base_address_xehp_and_later.inl
View File

@@ -78,7 +78,7 @@ void StateBaseAddressHelper<GfxFamily>::appendStateBaseAddressParameters(
args.stateBaseAddressCmd->setStatelessDataPortAccessMemoryObjectControlState(args.gmmHelper->getMOCS(GMM_RESOURCE_USAGE_OCL_BUFFER_CONST));
}
appendExtraCacheSettings(args.stateBaseAddressCmd, args.gmmHelper->getHardwareInfo());
appendExtraCacheSettings(args);
}
template <typename GfxFamily>

4
shared/source/os_interface/hw_info_config.h
View File

@@ -132,7 +132,7 @@ class HwInfoConfig {
virtual bool isCpuCopyNecessary(const void *ptr, MemoryManager *memoryManager) const = 0;
virtual bool isAdjustWalkOrderAvailable(const HardwareInfo &hwInfo) const = 0;
virtual bool isAssignEngineRoundRobinSupported() const = 0;
virtual uint32_t getL1CachePolicy() const = 0;
virtual uint32_t getL1CachePolicy(bool isDebuggerActive) const = 0;
virtual bool isEvictionIfNecessaryFlagSupported() const = 0;
virtual void adjustNumberOfCcs(HardwareInfo &hwInfo) const = 0;
virtual bool isPrefetcherDisablingInDirectSubmissionRequired() const = 0;
@@ -258,7 +258,7 @@ class HwInfoConfigHw : public HwInfoConfig {
bool isCpuCopyNecessary(const void *ptr, MemoryManager *memoryManager) const override;
bool isAdjustWalkOrderAvailable(const HardwareInfo &hwInfo) const override;
bool isAssignEngineRoundRobinSupported() const override;
uint32_t getL1CachePolicy() const override;
uint32_t getL1CachePolicy(bool isDebuggerActive) const override;
bool isEvictionIfNecessaryFlagSupported() const override;
void adjustNumberOfCcs(HardwareInfo &hwInfo) const override;
bool isPrefetcherDisablingInDirectSubmissionRequired() const override;

4
shared/source/os_interface/hw_info_config.inl
View File

@@ -482,8 +482,8 @@ template <PRODUCT_FAMILY gfxProduct>
bool HwInfoConfigHw<gfxProduct>::isAdjustWalkOrderAvailable(const HardwareInfo &hwInfo) const { return false; }
template <PRODUCT_FAMILY gfxProduct>
uint32_t HwInfoConfigHw<gfxProduct>::getL1CachePolicy() const {
return L1CachePolicyHelper<gfxProduct>::getL1CachePolicy();
uint32_t HwInfoConfigHw<gfxProduct>::getL1CachePolicy(bool isDebuggerActive) const {
return L1CachePolicyHelper<gfxProduct>::getL1CachePolicy(isDebuggerActive);
}
template <PRODUCT_FAMILY gfxProduct>

2
shared/source/xe_hp_core/command_encoder_xe_hp_core.cpp
View File

@@ -26,7 +26,7 @@ void EncodeDispatchKernel<Family>::adjustTimestampPacket(WALKER_TYPE &walkerCmd,
}
template <>
inline void EncodeSurfaceState<Family>::encodeExtraCacheSettings(R_SURFACE_STATE *surfaceState, const HardwareInfo &hwInfo) {
inline void EncodeSurfaceState<Family>::encodeExtraCacheSettings(R_SURFACE_STATE *surfaceState, const EncodeSurfaceStateArgs &args) {
}
template <>

2
shared/source/xe_hp_core/state_base_address_xe_hp_core.cpp
View File

@@ -11,7 +11,7 @@
namespace NEO {
template <>
void StateBaseAddressHelper<XeHpFamily>::appendExtraCacheSettings(STATE_BASE_ADDRESS *stateBaseAddress, const HardwareInfo *hwInfo) {
void StateBaseAddressHelper<XeHpFamily>::appendExtraCacheSettings(StateBaseAddressHelperArgs<XeHpFamily> &args) {
}
template struct StateBaseAddressHelper<XeHpFamily>;

7
shared/source/xe_hpg_core/enable_xe_hpg_core.cpp
View File

@@ -18,9 +18,12 @@ namespace NEO {
#ifdef SUPPORT_DG2
template <>
uint32_t L1CachePolicyHelper<IGFX_DG2>::getDefaultL1CachePolicy() {
uint32_t L1CachePolicyHelper<IGFX_DG2>::getDefaultL1CachePolicy(bool isDebuggerActive) {
using GfxFamily = HwMapper<IGFX_DG2>::GfxFamily;
return GfxFamily::STATE_BASE_ADDRESS::L1_CACHE_POLICY_WBP;
if (isDebuggerActive) {
return GfxFamily::STATE_BASE_ADDRESS::L1_CACHE_POLICY_WBP;
}
return GfxFamily::STATE_BASE_ADDRESS::L1_CACHE_POLICY_WB;
}
template struct L1CachePolicyHelper<IGFX_DG2>;