AI/rocm-systems
2
0
Çatalla 0
Kod Konular Değişiklik İstekleri İşlemler Paketler Projeler Sürüm Viki Aktivite

SWDEV-508776 - support VGPRs validation (#274)

Kaynağa Gözat 
Clarify some VGPRs terms description.
Fix some wrong query logics of availableVGPRs_ and
availableRegistersPerCU_ in device info.
Add hipDeviceAttributeMaxAvailableVgprsPerThread
attribute query.
Remove hardcoding of following
 info_.vgprAllocGranularity_
 info_.vgprsPerSimd_

[ROCm/clr commit: 397f303d97]
Bu işleme şunda yer alıyor:
Sang, Tao
2025-07-09 10:46:52 -04:00
işlemeyi yapan: GitHub
ebeveyn 84e9845897
işleme 7fd322af90
7 değiştirilmiş dosya ile 32 ekleme ve 93 silme
Yama Dosyasını İndir Diff Dosyasını İndir Tüm dosyaları genişlet Tüm dosyaları daralt
projects/clr/hipamd/src/hip_device_runtime.cpp
+3
Dosyayı Görüntüle
@@ -448,6 +448,9 @@ hipError_t hipDeviceGetAttribute(int* pi, hipDeviceAttribute_t attr, int device)
case hipDeviceAttributeNumberOfXccs:
*pi = static_cast<int>(g_devices[device]->devices()[0]->info().numberOfXccs_);
break;
case hipDeviceAttributeMaxAvailableVgprsPerThread:
*pi = static_cast<int>(g_devices[device]->devices()[0]->info().availableVGPRs_);
break;
default:
HIP_RETURN(hipErrorInvalidValue);
}
projects/clr/hipamd/src/hip_platform.cpp
+1 -1
Dosyayı Görüntüle
@@ -397,7 +397,7 @@ hipError_t ihipOccupancyMaxActiveBlocksPerMultiprocessor(
}
if (VgprWaves == 0) {
// This should not happen ideally, but in case the usedVGPRs_/availableVGPRs_ values are
// This should not happen ideally, but in case the value is
// incorrect, it can lead to a crash. By returning error, API can exit gracefully.
return hipErrorUnknown;
}
projects/clr/rocclr/device/device.hpp
+7 -7
Dosyayı Görüntüle
@@ -552,10 +552,12 @@ struct Info : public amd::EmbeddedObject {
uint32_t wavefrontWidth_;
//! Available number of SGPRs
uint32_t availableSGPRs_;
//! Available number of VGPRs
uint32_t availableVGPRs_;
//! Available number of registers per CU
uint32_t availableRegistersPerCU_;
uint32_t availableVGPRs_; //!< Number of addressable VGPRs per thread in DWORDs
uint32_t vgprsPerSimd_; //!< Number of VGPRs per SIMD in DWORDs
uint32_t vgprAllocGranularity_; //!< Number of VGPRs allocation granularity per thread in DWORDs
uint32_t availableRegistersPerCU_; //!< Number of VGPRs per CU in DWORDs
//! Number of global memory channels
uint32_t globalMemChannels_;
//! Number of banks in each global memory channel
@@ -650,9 +652,7 @@ struct Info : public amd::EmbeddedObject {
uint32_t driverNodeId_;
//! Number of Physical SGPRs per SIMD
uint32_t sgprsPerSimd_;
//! Number of VGPRs per SIMD
uint32_t vgprsPerSimd_;
uint32_t vgprAllocGranularity_;
uint32_t numSDMAengines_; //!< Number of available SDMA engines
uint32_t luidLowPart_; //!< Luid low 4 bytes, available in Windows only
projects/clr/rocclr/device/devkernel.cpp
+1 -1
Dosyayı Görüntüle
@@ -619,7 +619,7 @@ Kernel::Kernel(const amd::Device& dev, const std::string& name, const Program& p
workGroupInfo_.usedGPRs_ = 0;
workGroupInfo_.availableSGPRs_ = 0;
workGroupInfo_.usedSGPRs_ = 0;
workGroupInfo_.availableVGPRs_ = 0;
workGroupInfo_.availableVGPRs_ = dev.info().availableVGPRs_;
workGroupInfo_.usedVGPRs_ = 0;
workGroupInfo_.availableLDSSize_ = 0;
workGroupInfo_.usedLDSSize_ = 0;
projects/clr/rocclr/device/devkernel.hpp
+2 -2
Dosyayı Görüntüle
@@ -218,8 +218,8 @@ class Kernel : public amd::HeapObject {
size_t usedGPRs_; //!< GPRs used by the program
size_t availableSGPRs_; //!< SGPRs available to the program
size_t usedSGPRs_; //!< SGPRs used by the program
size_t availableVGPRs_; //!< VGPRs available to the program
size_t usedVGPRs_; //!< VGPRs used by the program
size_t availableVGPRs_; //!< VGPRs addressable to the program per thread in DWORDs
size_t usedVGPRs_; //!< VGPRs used by the program per thread in DWORDs
size_t availableLDSSize_; //!< available LDS size
size_t usedLDSSize_; //!< used LDS size
size_t availableStackSize_; //!< available stack size
projects/clr/rocclr/device/pal/paldevice.cpp
+2 -1
Dosyayı Görüntüle
@@ -678,8 +678,9 @@ void NullDevice::fillDeviceInfo(const Pal::DeviceProperties& palProp,
static_cast<size_t>(palProp.gpuMemoryProperties.virtualMemAllocGranularity);
info_.vgprAllocGranularity_ = palProp.gfxipProperties.shaderCore.vgprAllocGranularity;
info_.vgprsPerSimd_ = palProp.gfxipProperties.shaderCore.vgprsPerSimd;
info_.availableVGPRs_ = palProp.gfxipProperties.shaderCore.numAvailableVgprs;
info_.sgprsPerSimd_ = palProp.gfxipProperties.shaderCore.sgprsPerSimd;
info_.availableRegistersPerCU_ = info_.vgprsPerSimd_ * info_.simdPerCU_ * 32;
info_.availableRegistersPerCU_ = info_.vgprsPerSimd_ * info_.simdPerCU_ * info_.wavefrontWidth_;
#if IS_WINDOWS
info_.luidLowPart_ = palProp.osProperties.luidLowPart;
info_.luidHighPart_ = palProp.osProperties.luidHighPart;
projects/clr/rocclr/device/rocm/rocdevice.cpp
+16 -81
Dosyayı Görüntüle
@@ -1634,12 +1634,21 @@ bool Device::populateOCLDeviceConstants() {
// Get Values from from Comgr
amd_comgr_metadata_node_t isaMeta;
if (getIsaMeta(std::move(isa().isaName()), isaMeta)) {
std::string vgprValue;
info_.availableVGPRs_ = (getValueFromIsaMeta(isaMeta, "AddressableNumVGPRs", vgprValue))
? (atoi(vgprValue.c_str()) * info_.simdPerCU_)
: 0;
std::string addressableNumVGPRs, totalNumVGPRs, vGPRAllocGranule;
info_.availableVGPRs_ = getValueFromIsaMeta(isaMeta, "AddressableNumVGPRs",
addressableNumVGPRs) ? atoi(addressableNumVGPRs.c_str()) : 0;
info_.vgprsPerSimd_ = getValueFromIsaMeta(isaMeta, "TotalNumVGPRs",
totalNumVGPRs) ? atoi(totalNumVGPRs.c_str()) : 0;
info_.vgprAllocGranularity_ = getValueFromIsaMeta(isaMeta, "VGPRAllocGranule",
vGPRAllocGranule) ? atoi(vGPRAllocGranule.c_str()) : 0;
info_.availableRegistersPerCU_ = info_.availableVGPRs_ * 64; // 64 registers per VGPR
info_.availableRegistersPerCU_ = info_.vgprsPerSimd_ * info_.simdPerCU_ *
info_.wavefrontWidth_;
ClPrint(amd::LOG_INFO, amd::LOG_INIT,
"addressableNumVGPRs=%u, totalNumVGPRs=%u, vGPRAllocGranule=%u," \
" availableRegistersPerCU_=%u",
info_.availableVGPRs_, info_.vgprsPerSimd_, info_.vgprAllocGranularity_,
info_.availableRegistersPerCU_);
std::string sgprValue;
info_.availableSGPRs_ = (getValueFromIsaMeta(isaMeta, "AddressableNumSGPRs", sgprValue))
@@ -1648,6 +1657,8 @@ bool Device::populateOCLDeviceConstants() {
if (!releaseIsaMeta(isaMeta)) {
LogInfo("Can not release the isa meta node");
}
} else {
ClPrint(amd::LOG_ERROR, amd::LOG_INIT, "getIsaMeta(%s) failed!", isa().isaName().c_str());
}
// Generic support for HMM interfaces
@@ -1694,82 +1705,6 @@ bool Device::populateOCLDeviceConstants() {
}
HIP_MEM_POOL_USE_VM &= info_.virtualMemoryManagement_;
switch (isa().versionMajor()) {
case (12):
if (isa().versionMinor() == 0) {
switch (isa().versionStepping()) {
case (0):
case (1):
default:
info_.vgprAllocGranularity_ = 24;
info_.vgprsPerSimd_ = 1536;
break;
}
}
break;
case (11):
if (isa().versionMinor() == 0) {
switch (isa().versionStepping()) {
case (0):
case (1):
case (5):
info_.vgprAllocGranularity_ = 24;
info_.vgprsPerSimd_ = 1536;
break;
case (2):
case (3):
default:
info_.vgprAllocGranularity_ = 16;
info_.vgprsPerSimd_ = 1024;
break;
}
} else if (isa().versionMinor() == 5) {
switch (isa().versionStepping()) {
case (1):
info_.vgprAllocGranularity_ = 24;
info_.vgprsPerSimd_ = 1536;
break;
default:
info_.vgprAllocGranularity_ = 16;
info_.vgprsPerSimd_ = 1024;
break;
}
} else {
info_.vgprAllocGranularity_ = 16;
info_.vgprsPerSimd_ = 1024;
}
break;
case (10):
switch (isa().versionMinor()) {
case (0):
case (1):
info_.vgprAllocGranularity_ = 8;
info_.vgprsPerSimd_ = 1024;
break;
case (3):
default:
info_.vgprAllocGranularity_ = 16;
info_.vgprsPerSimd_ = 1024;
break;
}
break;
case (9):
if ((isa().versionMinor() == 0 && isa().versionStepping() == 10) ||
isa().versionMinor() == 4 || isa().versionMinor() == 5) {
info_.vgprAllocGranularity_ = 8;
info_.vgprsPerSimd_ = 512;
} else {
info_.vgprAllocGranularity_ = 4;
info_.vgprsPerSimd_ = 256;
}
break;
default:
// For gfx<=8
info_.vgprAllocGranularity_ = 4;
info_.vgprsPerSimd_ = 256;
break;
}
if (isa().versionMajor() < 8) {
info_.sgprsPerSimd_ = 512;
} else if (isa().versionMajor() < 10) {