compute-runtime/runtime/gen9/hw_info_cfl.inl

287 lines
12 KiB
Plaintext
Raw Normal View History

/*
* Copyright (C) 2018-2020 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
*/
#include "core/gen9/hw_cmds.h"
#include "core/memory_manager/memory_constants.h"
#include "runtime/aub_mem_dump/aub_services.h"
#include "engine_node.h"
namespace NEO {
const char *HwMapper<IGFX_COFFEELAKE>::abbreviation = "cfl";
bool isSimulationCFL(unsigned short deviceId) {
return false;
};
const PLATFORM CFL::platform = {
IGFX_COFFEELAKE,
PCH_UNKNOWN,
IGFX_GEN9_CORE,
IGFX_GEN9_CORE,
PLATFORM_NONE, // default init
0, // usDeviceID
0, // usRevId. 0 sets the stepping to A0
0, // usDeviceID_PCH
0, // usRevId_PCH
GTTYPE_UNDEFINED};
const RuntimeCapabilityTable CFL::capabilityTable{
{0, 0, 0, false, false, false}, // kmdNotifyProperties
MemoryConstants::max48BitAddress, // gpuAddressSpace
83.333, // defaultProfilingTimerResolution
MemoryConstants::pageSize, // requiredPreemptionSurfaceSize
&isSimulationCFL, // isSimulation
PreemptionMode::MidThread, // defaultPreemptionMode
aub_stream::ENGINE_RCS, // defaultEngineType
0, // maxRenderFrequency
21, // clVersionSupport
CmdServicesMemTraceVersion::DeviceValues::Cfl, // aubDeviceId
0, // extraQuantityThreadsPerEU
64, // slmSize
sizeof(CFL::GRF), // grfSize
false, // blitterOperationsSupported
true, // ftrSupportsInteger64BitAtomics
true, // ftrSupportsFP64
true, // ftrSupports64BitMath
true, // ftrSvm
true, // ftrSupportsCoherency
true, // ftrSupportsVmeAvcTextureSampler
false, // ftrSupportsVmeAvcPreemption
false, // ftrRenderCompressedBuffers
false, // ftrRenderCompressedImages
true, // ftr64KBpages
true, // instrumentationEnabled
true, // forceStatelessCompilationFor32Bit
"core", // platformType
true, // sourceLevelDebuggerSupported
true, // supportsVme
false, // supportCacheFlushAfterWalker
true, // supportsImages
true, // supportsDeviceEnqueue
true // hostPtrTrackingEnabled
};
WorkaroundTable CFL::workaroundTable = {};
FeatureTable CFL::featureTable = {};
void CFL::setupFeatureAndWorkaroundTable(HardwareInfo *hwInfo) {
FeatureTable *featureTable = &hwInfo->featureTable;
WorkaroundTable *workaroundTable = &hwInfo->workaroundTable;
featureTable->ftrGpGpuMidBatchPreempt = true;
featureTable->ftrGpGpuThreadGroupLevelPreempt = true;
featureTable->ftrL3IACoherency = true;
featureTable->ftrVEBOX = true;
featureTable->ftrGpGpuMidThreadLevelPreempt = false;
featureTable->ftr3dMidBatchPreempt = true;
featureTable->ftr3dObjectLevelPreempt = true;
featureTable->ftrPerCtxtPreemptionGranularityControl = true;
featureTable->ftrPPGTT = true;
featureTable->ftrSVM = true;
featureTable->ftrIA32eGfxPTEs = true;
featureTable->ftrDisplayYTiling = true;
featureTable->ftrTranslationTable = true;
featureTable->ftrUserModeTranslationTable = true;
featureTable->ftrEnableGuC = true;
featureTable->ftrFbc = true;
featureTable->ftrFbc2AddressTranslation = true;
featureTable->ftrFbcBlitterTracking = true;
featureTable->ftrFbcCpuTracking = true;
featureTable->ftrTileY = true;
workaroundTable->waEnablePreemptionGranularityControlByUMD = true;
workaroundTable->waSendMIFLUSHBeforeVFE = true;
workaroundTable->waReportPerfCountUseGlobalContextID = true;
workaroundTable->waMsaa8xTileYDepthPitchAlignment = true;
workaroundTable->waLosslessCompressionSurfaceStride = true;
workaroundTable->waFbcLinearSurfaceStride = true;
workaroundTable->wa4kAlignUVOffsetNV12LinearSurface = true;
workaroundTable->waSamplerCacheFlushBetweenRedescribedSurfaceReads = true;
}
const HardwareInfo CFL_1x2x6::hwInfo = {
&CFL::platform,
&CFL::featureTable,
&CFL::workaroundTable,
&CFL_1x2x6::gtSystemInfo,
CFL::capabilityTable,
};
GT_SYSTEM_INFO CFL_1x2x6::gtSystemInfo = {0};
void CFL_1x2x6::setupHardwareInfo(HardwareInfo *hwInfo, bool setupFeatureTableAndWorkaroundTable) {
GT_SYSTEM_INFO *gtSysInfo = &hwInfo->gtSystemInfo;
gtSysInfo->ThreadCount = gtSysInfo->EUCount * CFL::threadsPerEu;
gtSysInfo->SliceCount = 1;
gtSysInfo->L3CacheSizeInKb = 384;
gtSysInfo->L3BankCount = 2;
gtSysInfo->MaxFillRate = 8;
gtSysInfo->TotalVsThreads = 336;
gtSysInfo->TotalHsThreads = 336;
gtSysInfo->TotalDsThreads = 336;
gtSysInfo->TotalGsThreads = 336;
gtSysInfo->TotalPsThreadsWindowerRange = 64;
gtSysInfo->CsrSizeInMb = 8;
gtSysInfo->MaxEuPerSubSlice = CFL::maxEuPerSubslice;
gtSysInfo->MaxSlicesSupported = CFL::maxSlicesSupported;
gtSysInfo->MaxSubSlicesSupported = CFL::maxSubslicesSupported;
gtSysInfo->IsL3HashModeEnabled = false;
gtSysInfo->IsDynamicallyPopulated = false;
if (setupFeatureTableAndWorkaroundTable) {
setupFeatureAndWorkaroundTable(hwInfo);
}
};
const HardwareInfo CFL_1x3x6::hwInfo = {
&CFL::platform,
&CFL::featureTable,
&CFL::workaroundTable,
&CFL_1x3x6::gtSystemInfo,
CFL::capabilityTable,
};
GT_SYSTEM_INFO CFL_1x3x6::gtSystemInfo = {0};
void CFL_1x3x6::setupHardwareInfo(HardwareInfo *hwInfo, bool setupFeatureTableAndWorkaroundTable) {
GT_SYSTEM_INFO *gtSysInfo = &hwInfo->gtSystemInfo;
gtSysInfo->ThreadCount = gtSysInfo->EUCount * CFL::threadsPerEu;
gtSysInfo->SliceCount = 1;
gtSysInfo->L3CacheSizeInKb = 768;
gtSysInfo->L3BankCount = 4;
gtSysInfo->MaxFillRate = 8;
gtSysInfo->TotalVsThreads = 336;
gtSysInfo->TotalHsThreads = 336;
gtSysInfo->TotalDsThreads = 336;
gtSysInfo->TotalGsThreads = 336;
gtSysInfo->TotalPsThreadsWindowerRange = 64;
gtSysInfo->CsrSizeInMb = 8;
gtSysInfo->MaxEuPerSubSlice = CFL::maxEuPerSubslice;
gtSysInfo->MaxSlicesSupported = CFL::maxSlicesSupported;
gtSysInfo->MaxSubSlicesSupported = CFL::maxSubslicesSupported;
gtSysInfo->IsL3HashModeEnabled = false;
gtSysInfo->IsDynamicallyPopulated = false;
if (setupFeatureTableAndWorkaroundTable) {
setupFeatureAndWorkaroundTable(hwInfo);
}
};
const HardwareInfo CFL_1x3x8::hwInfo = {
&CFL::platform,
&CFL::featureTable,
&CFL::workaroundTable,
&CFL_1x3x8::gtSystemInfo,
CFL::capabilityTable,
};
GT_SYSTEM_INFO CFL_1x3x8::gtSystemInfo = {0};
void CFL_1x3x8::setupHardwareInfo(HardwareInfo *hwInfo, bool setupFeatureTableAndWorkaroundTable) {
GT_SYSTEM_INFO *gtSysInfo = &hwInfo->gtSystemInfo;
gtSysInfo->ThreadCount = gtSysInfo->EUCount * CFL::threadsPerEu;
gtSysInfo->SliceCount = 1;
gtSysInfo->L3CacheSizeInKb = 768;
gtSysInfo->L3BankCount = 4;
gtSysInfo->MaxFillRate = 8;
gtSysInfo->TotalVsThreads = 336;
gtSysInfo->TotalHsThreads = 336;
gtSysInfo->TotalDsThreads = 336;
gtSysInfo->TotalGsThreads = 336;
gtSysInfo->TotalPsThreadsWindowerRange = 64;
gtSysInfo->CsrSizeInMb = 8;
gtSysInfo->MaxEuPerSubSlice = CFL::maxEuPerSubslice;
gtSysInfo->MaxSlicesSupported = CFL::maxSlicesSupported;
gtSysInfo->MaxSubSlicesSupported = CFL::maxSubslicesSupported;
gtSysInfo->IsL3HashModeEnabled = false;
gtSysInfo->IsDynamicallyPopulated = false;
if (setupFeatureTableAndWorkaroundTable) {
setupFeatureAndWorkaroundTable(hwInfo);
}
};
const HardwareInfo CFL_2x3x8::hwInfo = {
&CFL::platform,
&CFL::featureTable,
&CFL::workaroundTable,
&CFL_2x3x8::gtSystemInfo,
CFL::capabilityTable,
};
GT_SYSTEM_INFO CFL_2x3x8::gtSystemInfo = {0};
void CFL_2x3x8::setupHardwareInfo(HardwareInfo *hwInfo, bool setupFeatureTableAndWorkaroundTable) {
GT_SYSTEM_INFO *gtSysInfo = &hwInfo->gtSystemInfo;
gtSysInfo->ThreadCount = gtSysInfo->EUCount * CFL::threadsPerEu;
gtSysInfo->SliceCount = 2;
gtSysInfo->L3CacheSizeInKb = 1536;
gtSysInfo->L3BankCount = 8;
gtSysInfo->MaxFillRate = 16;
gtSysInfo->TotalVsThreads = 336;
gtSysInfo->TotalHsThreads = 336;
gtSysInfo->TotalDsThreads = 336;
gtSysInfo->TotalGsThreads = 336;
gtSysInfo->TotalPsThreadsWindowerRange = 64;
gtSysInfo->CsrSizeInMb = 8;
gtSysInfo->MaxEuPerSubSlice = CFL::maxEuPerSubslice;
gtSysInfo->MaxSlicesSupported = CFL::maxSlicesSupported;
gtSysInfo->MaxSubSlicesSupported = CFL::maxSubslicesSupported;
gtSysInfo->IsL3HashModeEnabled = false;
gtSysInfo->IsDynamicallyPopulated = false;
if (setupFeatureTableAndWorkaroundTable) {
setupFeatureAndWorkaroundTable(hwInfo);
}
};
const HardwareInfo CFL_3x3x8::hwInfo = {
&CFL::platform,
&CFL::featureTable,
&CFL::workaroundTable,
&CFL_3x3x8::gtSystemInfo,
CFL::capabilityTable,
};
GT_SYSTEM_INFO CFL_3x3x8::gtSystemInfo = {0};
void CFL_3x3x8::setupHardwareInfo(HardwareInfo *hwInfo, bool setupFeatureTableAndWorkaroundTable) {
GT_SYSTEM_INFO *gtSysInfo = &hwInfo->gtSystemInfo;
gtSysInfo->ThreadCount = gtSysInfo->EUCount * CFL::threadsPerEu;
gtSysInfo->SliceCount = 3;
gtSysInfo->L3CacheSizeInKb = 2304;
gtSysInfo->L3BankCount = 12;
gtSysInfo->MaxFillRate = 24;
gtSysInfo->TotalVsThreads = 336;
gtSysInfo->TotalHsThreads = 336;
gtSysInfo->TotalDsThreads = 336;
gtSysInfo->TotalGsThreads = 336;
gtSysInfo->TotalPsThreadsWindowerRange = 64;
gtSysInfo->CsrSizeInMb = 8;
gtSysInfo->MaxEuPerSubSlice = CFL::maxEuPerSubslice;
gtSysInfo->MaxSlicesSupported = CFL::maxSlicesSupported;
gtSysInfo->MaxSubSlicesSupported = CFL::maxSubslicesSupported;
gtSysInfo->IsL3HashModeEnabled = false;
gtSysInfo->IsDynamicallyPopulated = false;
if (setupFeatureTableAndWorkaroundTable) {
setupFeatureAndWorkaroundTable(hwInfo);
}
};
const HardwareInfo CFL::hwInfo = CFL_1x3x6::hwInfo;
const uint64_t CFL::defaultHardwareInfoConfig = 0x100030006;
void setupCFLHardwareInfoImpl(HardwareInfo *hwInfo, bool setupFeatureTableAndWorkaroundTable, uint64_t hwInfoConfig) {
if (hwInfoConfig == 0x100030008) {
CFL_1x3x8::setupHardwareInfo(hwInfo, setupFeatureTableAndWorkaroundTable);
} else if (hwInfoConfig == 0x200030008) {
CFL_2x3x8::setupHardwareInfo(hwInfo, setupFeatureTableAndWorkaroundTable);
} else if (hwInfoConfig == 0x300030008) {
CFL_3x3x8::setupHardwareInfo(hwInfo, setupFeatureTableAndWorkaroundTable);
} else if (hwInfoConfig == 0x100020006) {
CFL_1x2x6::setupHardwareInfo(hwInfo, setupFeatureTableAndWorkaroundTable);
} else if (hwInfoConfig == 0x100030006) {
CFL_1x3x6::setupHardwareInfo(hwInfo, setupFeatureTableAndWorkaroundTable);
} else if (hwInfoConfig == 0x0) {
// Default config
CFL_1x3x6::setupHardwareInfo(hwInfo, setupFeatureTableAndWorkaroundTable);
} else {
UNRECOVERABLE_IF(true);
}
}
void (*CFL::setupHardwareInfo)(HardwareInfo *, bool, uint64_t) = setupCFLHardwareInfoImpl;
} // namespace NEO