245 lines
10 KiB
C++
245 lines
10 KiB
C++
/*
|
|
* Copyright (C) 2018-2022 Intel Corporation
|
|
*
|
|
* SPDX-License-Identifier: MIT
|
|
*
|
|
*/
|
|
|
|
#include "shared/source/aub_mem_dump/definitions/aub_services.h"
|
|
#include "shared/source/gen9/hw_cmds_cfl.h"
|
|
#include "shared/source/helpers/constants.h"
|
|
|
|
#include "aubstream/engine_node.h"
|
|
|
|
namespace NEO {
|
|
|
|
const char *HwMapper<IGFX_COFFEELAKE>::abbreviation = "cfl";
|
|
|
|
const PLATFORM CFL::platform = {
|
|
IGFX_COFFEELAKE,
|
|
PCH_UNKNOWN,
|
|
IGFX_GEN9_CORE,
|
|
IGFX_GEN9_CORE,
|
|
PLATFORM_NONE, // default init
|
|
0, // usDeviceID
|
|
9, // usRevId. 0 sets the stepping to A0
|
|
0, // usDeviceID_PCH
|
|
0, // usRevId_PCH
|
|
GTTYPE_UNDEFINED};
|
|
|
|
const RuntimeCapabilityTable CFL::capabilityTable{
|
|
EngineDirectSubmissionInitVec{
|
|
{aub_stream::ENGINE_RCS, {true, true}}}, // directSubmissionEngines
|
|
{0, 0, 0, 0, false, false, false, false}, // kmdNotifyProperties
|
|
MemoryConstants::max48BitAddress, // gpuAddressSpace
|
|
0, // sharedSystemMemCapabilities
|
|
83.333, // defaultProfilingTimerResolution
|
|
MemoryConstants::pageSize, // requiredPreemptionSurfaceSize
|
|
"core", // platformType
|
|
"", // deviceName
|
|
PreemptionMode::MidThread, // defaultPreemptionMode
|
|
aub_stream::ENGINE_RCS, // defaultEngineType
|
|
0, // maxRenderFrequency
|
|
30, // clVersionSupport
|
|
CmdServicesMemTraceVersion::DeviceValues::Cfl, // aubDeviceId
|
|
0, // extraQuantityThreadsPerEU
|
|
64, // slmSize
|
|
sizeof(CFL::GRF), // grfSize
|
|
36u, // timestampValidBits
|
|
32u, // kernelTimestampValidBits
|
|
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, // sourceLevelDebuggerSupported
|
|
true, // supportsVme
|
|
false, // supportCacheFlushAfterWalker
|
|
true, // supportsImages
|
|
false, // supportsDeviceEnqueue
|
|
true, // supportsPipes
|
|
true, // supportsOcl21Features
|
|
false, // supportsOnDemandPageFaults
|
|
true, // supportsIndependentForwardProgress
|
|
true, // hostPtrTrackingEnabled
|
|
true, // levelZeroSupported
|
|
true, // isIntegratedDevice
|
|
true, // supportsMediaBlock
|
|
false, // p2pAccessSupported
|
|
false, // p2pAtomicAccessSupported
|
|
false, // fusedEuEnabled
|
|
false // l0DebuggerSupported;
|
|
};
|
|
|
|
WorkaroundTable CFL::workaroundTable = {};
|
|
FeatureTable CFL::featureTable = {};
|
|
|
|
void CFL::setupFeatureAndWorkaroundTable(HardwareInfo *hwInfo) {
|
|
FeatureTable *featureTable = &hwInfo->featureTable;
|
|
WorkaroundTable *workaroundTable = &hwInfo->workaroundTable;
|
|
|
|
featureTable->flags.ftrGpGpuMidBatchPreempt = true;
|
|
featureTable->flags.ftrGpGpuThreadGroupLevelPreempt = true;
|
|
featureTable->flags.ftrL3IACoherency = true;
|
|
featureTable->flags.ftrGpGpuMidThreadLevelPreempt = true;
|
|
featureTable->flags.ftrPPGTT = true;
|
|
featureTable->flags.ftrSVM = true;
|
|
featureTable->flags.ftrIA32eGfxPTEs = true;
|
|
featureTable->flags.ftrDisplayYTiling = true;
|
|
featureTable->flags.ftrTranslationTable = true;
|
|
featureTable->flags.ftrUserModeTranslationTable = true;
|
|
featureTable->flags.ftrFbc = true;
|
|
featureTable->flags.ftrTileY = true;
|
|
|
|
workaroundTable->flags.waSendMIFLUSHBeforeVFE = true;
|
|
workaroundTable->flags.waMsaa8xTileYDepthPitchAlignment = true;
|
|
workaroundTable->flags.waLosslessCompressionSurfaceStride = true;
|
|
workaroundTable->flags.waFbcLinearSurfaceStride = true;
|
|
workaroundTable->flags.wa4kAlignUVOffsetNV12LinearSurface = true;
|
|
workaroundTable->flags.waSamplerCacheFlushBetweenRedescribedSurfaceReads = true;
|
|
}
|
|
|
|
void CFL::setupHardwareInfoBase(HardwareInfo *hwInfo, bool setupFeatureTableAndWorkaroundTable) {
|
|
GT_SYSTEM_INFO *gtSysInfo = &hwInfo->gtSystemInfo;
|
|
gtSysInfo->ThreadCount = gtSysInfo->EUCount * CFL::threadsPerEu;
|
|
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 CflHw1x2x6::hwInfo = {
|
|
&CFL::platform,
|
|
&CFL::featureTable,
|
|
&CFL::workaroundTable,
|
|
&CflHw1x2x6::gtSystemInfo,
|
|
CFL::capabilityTable,
|
|
};
|
|
GT_SYSTEM_INFO CflHw1x2x6::gtSystemInfo = {0};
|
|
void CflHw1x2x6::setupHardwareInfo(HardwareInfo *hwInfo, bool setupFeatureTableAndWorkaroundTable) {
|
|
CFL::setupHardwareInfoBase(hwInfo, setupFeatureTableAndWorkaroundTable);
|
|
|
|
GT_SYSTEM_INFO *gtSysInfo = &hwInfo->gtSystemInfo;
|
|
gtSysInfo->SliceCount = 1;
|
|
gtSysInfo->L3CacheSizeInKb = 384;
|
|
gtSysInfo->L3BankCount = 2;
|
|
gtSysInfo->MaxFillRate = 8;
|
|
};
|
|
|
|
const HardwareInfo CflHw1x3x6::hwInfo = {
|
|
&CFL::platform,
|
|
&CFL::featureTable,
|
|
&CFL::workaroundTable,
|
|
&CflHw1x3x6::gtSystemInfo,
|
|
CFL::capabilityTable,
|
|
};
|
|
|
|
GT_SYSTEM_INFO CflHw1x3x6::gtSystemInfo = {0};
|
|
void CflHw1x3x6::setupHardwareInfo(HardwareInfo *hwInfo, bool setupFeatureTableAndWorkaroundTable) {
|
|
CFL::setupHardwareInfoBase(hwInfo, setupFeatureTableAndWorkaroundTable);
|
|
|
|
GT_SYSTEM_INFO *gtSysInfo = &hwInfo->gtSystemInfo;
|
|
gtSysInfo->SliceCount = 1;
|
|
gtSysInfo->L3CacheSizeInKb = 768;
|
|
gtSysInfo->L3BankCount = 4;
|
|
gtSysInfo->MaxFillRate = 8;
|
|
};
|
|
|
|
const HardwareInfo CflHw1x3x8::hwInfo = {
|
|
&CFL::platform,
|
|
&CFL::featureTable,
|
|
&CFL::workaroundTable,
|
|
&CflHw1x3x8::gtSystemInfo,
|
|
CFL::capabilityTable,
|
|
};
|
|
|
|
GT_SYSTEM_INFO CflHw1x3x8::gtSystemInfo = {0};
|
|
void CflHw1x3x8::setupHardwareInfo(HardwareInfo *hwInfo, bool setupFeatureTableAndWorkaroundTable) {
|
|
CFL::setupHardwareInfoBase(hwInfo, setupFeatureTableAndWorkaroundTable);
|
|
|
|
GT_SYSTEM_INFO *gtSysInfo = &hwInfo->gtSystemInfo;
|
|
gtSysInfo->SliceCount = 1;
|
|
gtSysInfo->L3CacheSizeInKb = 768;
|
|
gtSysInfo->L3BankCount = 4;
|
|
gtSysInfo->MaxFillRate = 8;
|
|
};
|
|
|
|
const HardwareInfo CflHw2x3x8::hwInfo = {
|
|
&CFL::platform,
|
|
&CFL::featureTable,
|
|
&CFL::workaroundTable,
|
|
&CflHw2x3x8::gtSystemInfo,
|
|
CFL::capabilityTable,
|
|
};
|
|
GT_SYSTEM_INFO CflHw2x3x8::gtSystemInfo = {0};
|
|
void CflHw2x3x8::setupHardwareInfo(HardwareInfo *hwInfo, bool setupFeatureTableAndWorkaroundTable) {
|
|
CFL::setupHardwareInfoBase(hwInfo, setupFeatureTableAndWorkaroundTable);
|
|
|
|
GT_SYSTEM_INFO *gtSysInfo = &hwInfo->gtSystemInfo;
|
|
gtSysInfo->SliceCount = 2;
|
|
gtSysInfo->L3CacheSizeInKb = 1536;
|
|
gtSysInfo->L3BankCount = 8;
|
|
gtSysInfo->MaxFillRate = 16;
|
|
};
|
|
|
|
const HardwareInfo CflHw3x3x8::hwInfo = {
|
|
&CFL::platform,
|
|
&CFL::featureTable,
|
|
&CFL::workaroundTable,
|
|
&CflHw3x3x8::gtSystemInfo,
|
|
CFL::capabilityTable,
|
|
};
|
|
GT_SYSTEM_INFO CflHw3x3x8::gtSystemInfo = {0};
|
|
void CflHw3x3x8::setupHardwareInfo(HardwareInfo *hwInfo, bool setupFeatureTableAndWorkaroundTable) {
|
|
CFL::setupHardwareInfoBase(hwInfo, setupFeatureTableAndWorkaroundTable);
|
|
|
|
GT_SYSTEM_INFO *gtSysInfo = &hwInfo->gtSystemInfo;
|
|
gtSysInfo->SliceCount = 3;
|
|
gtSysInfo->L3CacheSizeInKb = 2304;
|
|
gtSysInfo->L3BankCount = 12;
|
|
gtSysInfo->MaxFillRate = 24;
|
|
};
|
|
|
|
const HardwareInfo CFL::hwInfo = CflHw1x3x6::hwInfo;
|
|
const uint64_t CFL::defaultHardwareInfoConfig = 0x100030006;
|
|
|
|
void setupCFLHardwareInfoImpl(HardwareInfo *hwInfo, bool setupFeatureTableAndWorkaroundTable, uint64_t hwInfoConfig) {
|
|
if (hwInfoConfig == 0x100030008) {
|
|
CflHw1x3x8::setupHardwareInfo(hwInfo, setupFeatureTableAndWorkaroundTable);
|
|
} else if (hwInfoConfig == 0x200030008) {
|
|
CflHw2x3x8::setupHardwareInfo(hwInfo, setupFeatureTableAndWorkaroundTable);
|
|
} else if (hwInfoConfig == 0x300030008) {
|
|
CflHw3x3x8::setupHardwareInfo(hwInfo, setupFeatureTableAndWorkaroundTable);
|
|
} else if (hwInfoConfig == 0x100020006) {
|
|
CflHw1x2x6::setupHardwareInfo(hwInfo, setupFeatureTableAndWorkaroundTable);
|
|
} else if (hwInfoConfig == 0x100030006) {
|
|
CflHw1x3x6::setupHardwareInfo(hwInfo, setupFeatureTableAndWorkaroundTable);
|
|
} else if (hwInfoConfig == 0x0) {
|
|
// Default config
|
|
CflHw1x3x6::setupHardwareInfo(hwInfo, setupFeatureTableAndWorkaroundTable);
|
|
} else {
|
|
UNRECOVERABLE_IF(true);
|
|
}
|
|
}
|
|
|
|
void (*CFL::setupHardwareInfo)(HardwareInfo *, bool, uint64_t) = setupCFLHardwareInfoImpl;
|
|
} // namespace NEO
|