From 5787f039930c0deb0697d5cf99515810eba58c4a Mon Sep 17 00:00:00 2001 From: foreman Date: Tue, 2 Oct 2018 18:50:36 -0400 Subject: [PATCH] P4 to Git Change 1613773 by gandryey@gera-w8 on 2018/10/02 18:37:59 SWDEV-79445 - OCL generic changes and code clean-up Program compilation clean-up: - Remove !defined(WITH_LIGHTNING_COMPILER), but appprofiler still requires more changes. Affected files ... ... //depot/stg/opencl/drivers/opencl/runtime/device/device.cpp#232 edit ... //depot/stg/opencl/drivers/opencl/runtime/device/devkernel.cpp#8 edit ... //depot/stg/opencl/drivers/opencl/runtime/device/devkernel.hpp#5 edit ... //depot/stg/opencl/drivers/opencl/runtime/device/devprogram.cpp#10 edit ... //depot/stg/opencl/drivers/opencl/runtime/device/pal/paldevice.cpp#112 edit ... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palkernel.cpp#68 edit ... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palprogram.cpp#81 edit ... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palsettings.cpp#59 edit ... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocdevice.cpp#101 edit ... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocsettings.cpp#38 edit [ROCm/clr commit: 8ee6142811a516d28e74280180c21c2a7b6de376] --- projects/clr/rocclr/runtime/device/device.cpp | 11 ++- .../clr/rocclr/runtime/device/devkernel.cpp | 22 +++--- .../clr/rocclr/runtime/device/devkernel.hpp | 2 +- .../clr/rocclr/runtime/device/devprogram.cpp | 70 ++++++++++++------- .../rocclr/runtime/device/pal/paldevice.cpp | 54 +++++++------- .../rocclr/runtime/device/pal/palkernel.cpp | 4 +- .../rocclr/runtime/device/pal/palprogram.cpp | 18 ++--- .../rocclr/runtime/device/pal/palsettings.cpp | 20 +++--- .../rocclr/runtime/device/rocm/rocdevice.cpp | 23 +++--- .../runtime/device/rocm/rocsettings.cpp | 17 +++-- 10 files changed, 125 insertions(+), 116 deletions(-) diff --git a/projects/clr/rocclr/runtime/device/device.cpp b/projects/clr/rocclr/runtime/device/device.cpp index e19f4f3a6b..fadd8fe272 100644 --- a/projects/clr/rocclr/runtime/device/device.cpp +++ b/projects/clr/rocclr/runtime/device/device.cpp @@ -119,13 +119,10 @@ bool Device::BlitProgram::create(amd::Device* device, const char* extraKernels, } // Build all kernels - std::string opt = - "-cl-internal-kernel " -#if !defined(WITH_LIGHTNING_COMPILER) - "-Wf,--force_disable_spir -fno-lib-no-inline " - "-fno-sc-keep-calls " -#endif // !defined(WITH_LIGHTNING_COMPILER) - ; + std::string opt = "-cl-internal-kernel "; + if (!device->settings().useLightning_) { + opt += "-Wf,--force_disable_spir -fno-lib-no-inline -fno-sc-keep-calls "; + } if (extraOptions != nullptr) { opt += extraOptions; diff --git a/projects/clr/rocclr/runtime/device/devkernel.cpp b/projects/clr/rocclr/runtime/device/devkernel.cpp index 33438dfc8f..8c0aa82d7f 100644 --- a/projects/clr/rocclr/runtime/device/devkernel.cpp +++ b/projects/clr/rocclr/runtime/device/devkernel.cpp @@ -280,7 +280,7 @@ static inline uint32_t GetOclArgumentTypeOCL(const KernelArgMD& lcArg, bool* isH } #endif // ================================================================================================ -#if defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#if defined(WITH_COMPILER_LIB) static inline uint32_t GetOclArgumentTypeOCL(const aclArgData* argInfo, bool* isHidden) { if (argInfo->argStr[0] == '_' && argInfo->argStr[1] == '.') { *isHidden = true; @@ -414,7 +414,7 @@ static inline clk_value_type_t GetOclTypeOCL(const KernelArgMD& lcArg, size_t si } #endif // ================================================================================================ -#if defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#if defined(WITH_COMPILER_LIB) static inline clk_value_type_t GetOclTypeOCL(const aclArgData* argInfo, size_t size = 0) { uint sizeType; uint numElements; @@ -497,7 +497,7 @@ static inline size_t GetArgAlignmentOCL(const KernelArgMD& lcArg) { return lcArg #endif // ================================================================================================ -#if defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#if defined(WITH_COMPILER_LIB) static inline size_t GetArgAlignmentOCL(const aclArgData* argInfo) { switch (argInfo->type) { case ARG_TYPE_POINTER: @@ -551,7 +551,7 @@ static inline size_t GetArgPointeeAlignmentOCL(const KernelArgMD& lcArg) { #endif // ================================================================================================ -#if defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#if defined(WITH_COMPILER_LIB) static inline size_t GetArgPointeeAlignmentOCL(const aclArgData* argInfo) { if (argInfo->type == ARG_TYPE_POINTER) { return argInfo->arg.pointer.align; @@ -578,7 +578,7 @@ static inline bool GetReadOnlyOCL(const KernelArgMD& lcArg) { #endif // ================================================================================================ -#if defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#if defined(WITH_COMPILER_LIB) static inline bool GetReadOnlyOCL(const aclArgData* argInfo) { if (argInfo->type == ARG_TYPE_POINTER) { return (argInfo->arg.pointer.type == ACCESS_TYPE_RO) ? true : false; @@ -596,7 +596,7 @@ static inline int GetArgSizeOCL(const KernelArgMD& lcArg) { return lcArg.mSize; #endif // ================================================================================================ -#if defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#if defined(WITH_COMPILER_LIB) inline static int GetArgSizeOCL(const aclArgData* argInfo) { switch (argInfo->type) { case ARG_TYPE_POINTER: @@ -659,7 +659,7 @@ static inline cl_kernel_arg_address_qualifier GetOclAddrQualOCL(const KernelArgM #endif // ================================================================================================ -#if defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#if defined(WITH_COMPILER_LIB) static inline cl_kernel_arg_address_qualifier GetOclAddrQualOCL(const aclArgData* argInfo) { if (argInfo->type == ARG_TYPE_POINTER) { switch (argInfo->arg.pointer.memory) { @@ -708,7 +708,7 @@ static inline cl_kernel_arg_access_qualifier GetOclAccessQualOCL(const KernelArg #endif // ================================================================================================ -#if defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#if defined(WITH_COMPILER_LIB) static inline cl_kernel_arg_access_qualifier GetOclAccessQualOCL(const aclArgData* argInfo) { if (argInfo->type == ARG_TYPE_IMAGE) { switch (argInfo->arg.image.type) { @@ -749,7 +749,7 @@ static inline cl_kernel_arg_type_qualifier GetOclTypeQualOCL(const KernelArgMD& #endif // ================================================================================================ -#if defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#if defined(WITH_COMPILER_LIB) static inline cl_kernel_arg_type_qualifier GetOclTypeQualOCL(const aclArgData* argInfo) { cl_kernel_arg_type_qualifier rv = CL_KERNEL_ARG_TYPE_NONE; if (argInfo->type == ARG_TYPE_POINTER) { @@ -858,7 +858,7 @@ void Kernel::InitParameters(const KernelMD& kernelMD, uint32_t argBufferSize) { #endif // ================================================================================================ -#if defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#if defined(WITH_COMPILER_LIB) void Kernel::InitParameters(const aclArgData* aclArg, uint32_t argBufferSize) { // Iterate through the arguments and insert into parameterList device::Kernel::parameters_t params; @@ -1032,7 +1032,7 @@ void Kernel::InitPrintf(const std::vector& printfInfoStrings) { #endif // defined(WITH_LIGHTNING_COMPILER) // ================================================================================================ -#if defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#if defined(WITH_COMPILER_LIB) void Kernel::InitPrintf(const aclPrintfFmt* aclPrintf) { PrintfInfo info; uint index = 0; diff --git a/projects/clr/rocclr/runtime/device/devkernel.hpp b/projects/clr/rocclr/runtime/device/devkernel.hpp index 59f7733fa9..d3ec08dcae 100644 --- a/projects/clr/rocclr/runtime/device/devkernel.hpp +++ b/projects/clr/rocclr/runtime/device/devkernel.hpp @@ -224,7 +224,7 @@ class Kernel : public amd::HeapObject { //! Initializes HSAIL Printf metadata and info for LC void InitPrintf(const std::vector& printfInfoStrings); #endif -#if defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#if defined(WITH_COMPILER_LIB) void InitParameters( const aclArgData* aclArg, //!< List of ACL arguments uint32_t argBufferSize diff --git a/projects/clr/rocclr/runtime/device/devprogram.cpp b/projects/clr/rocclr/runtime/device/devprogram.cpp index e531993025..88d0abf88d 100644 --- a/projects/clr/rocclr/runtime/device/devprogram.cpp +++ b/projects/clr/rocclr/runtime/device/devprogram.cpp @@ -16,7 +16,7 @@ #include "libraries.amdgcn.inc" #include "opencl1.2-c.amdgcn.inc" #include "opencl2.0-c.amdgcn.inc" -#endif // !defined(WITH_LIGHTNING_COMPILER) +#endif // defined(WITH_LIGHTNING_COMPILER) #include #include @@ -387,7 +387,7 @@ static void logFunction(const char* msg, size_t size) { bool Program::compileImplHSAIL(const std::string& sourceCode, const std::vector& headers, const char** headerIncludeNames, amd::option::Options* options) { -#if defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#if defined(WITH_COMPILER_LIB) acl_error errorCode; aclTargetInfo target; @@ -479,7 +479,7 @@ bool Program::compileImplHSAIL(const std::string& sourceCode, // Save the binary in the interface class saveBinaryAndSetType(TYPE_COMPILED); -#endif // defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#endif // defined(WITH_COMPILER_LIB) return true; } @@ -592,7 +592,7 @@ bool Program::linkImplLC(const std::vector& inputPrograms, // ================================================================================================ bool Program::linkImplHSAIL(const std::vector& inputPrograms, amd::option::Options* options, bool createLibrary) { -#if defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#if defined(WITH_COMPILER_LIB) acl_error errorCode; // For each program we need to extract the LLVMIR and create @@ -678,7 +678,7 @@ bool Program::linkImplHSAIL(const std::vector& inputPrograms, return linkImpl(options); #else return false; -#endif // defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#endif // defined(WITH_COMPILER_LIB) } // ================================================================================================ @@ -931,7 +931,7 @@ bool Program::linkImplLC(amd::option::Options* options) { // ================================================================================================ bool Program::linkImplHSAIL(amd::option::Options* options) { -#if defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#if defined(WITH_COMPILER_LIB) acl_error errorCode; bool finalize = true; internal_ = (compileOptions_.find("-cl-internal-kernel") != std::string::npos) ? true : false; @@ -1020,7 +1020,7 @@ bool Program::linkImplHSAIL(amd::option::Options* options) { return true; #else return false; -#endif // defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#endif // defined(WITH_COMPILER_LIB) } // ================================================================================================ @@ -1420,14 +1420,14 @@ std::string Program::ProcessOptions(amd::option::Options* options) { optionsStr.append(opts.str()); } -#if !defined(WITH_LIGHTNING_COMPILER) - if (!device().settings().singleFpDenorm_) { - optionsStr.append(" -cl-denorms-are-zero"); - } + if (!device().settings().useLightning_) { + if (!device().settings().singleFpDenorm_) { + optionsStr.append(" -cl-denorms-are-zero"); + } - // Check if the host is 64 bit or 32 bit - LP64_ONLY(optionsStr.append(" -m64")); -#endif // !defined(WITH_LIGHTNING_COMPILER) + // Check if the host is 64 bit or 32 bit + LP64_ONLY(optionsStr.append(" -m64")); + } // Tokenize the extensions string into a vector of strings std::istringstream istrstr(device().info().extensions_); @@ -1518,6 +1518,15 @@ bool Program::getCompileOptionsAtLinking(const std::vector& inputProgr return true; } +// ================================================================================================ +bool isSPIRVMagicL(const void* Image, size_t Length) { + const unsigned SPRVMagicNumber = 0x07230203; + if (Image == nullptr || Length < sizeof(unsigned)) + return false; + auto Magic = static_cast(Image); + return *Magic == SPRVMagicNumber; +} + // ================================================================================================ bool Program::initClBinary(const char* binaryIn, size_t size) { if (!initClBinary()) { @@ -1533,10 +1542,18 @@ bool Program::initClBinary(const char* binaryIn, size_t size) { // unencrypted int encryptCode = 0; char* decryptedBin = nullptr; + bool isSPIRV = false; + bool isBc = false; -#if !defined(WITH_LIGHTNING_COMPILER) - bool isSPIRV = isSPIRVMagic(binaryIn, size); - if (isSPIRV || isBcMagic(binaryIn)) { +#if defined(WITH_COMPILER_LIB) + if (!device().settings().useLightning_) { + isSPIRV = isSPIRVMagicL(binaryIn, size); + isBc = isBcMagic(binaryIn); + } +#endif // defined(WITH_COMPILER_LIB) + + if (isSPIRV || isBc) { +#if defined(WITH_COMPILER_LIB) acl_error err = ACL_SUCCESS; aclBinaryOptions binOpts = {0}; binOpts.struct_size = sizeof(binOpts); @@ -1580,9 +1597,8 @@ bool Program::initClBinary(const char* binaryIn, size_t size) { aclBinaryFini(aclbin_v30); aclBinaryFini(aclbin_v21); } - } else -#endif // !defined(WITH_LIGHTNING_COMPILER) - { +#endif // defined(WITH_COMPILER_LIB) + } else { size_t decryptedSize; if (!clBinary()->decryptElf(binaryIn, size, &decryptedBin, &decryptedSize, &encryptCode)) { return false; @@ -1705,9 +1721,9 @@ aclType Program::getCompilationStagesFromBinary(std::vector& completeSt default: break; } -#endif // !defined(WITH_LIGHTNING_COMPILER) +#endif // defined(WITH_LIGHTNING_COMPILER) } else { -#if defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#if defined(WITH_COMPILER_LIB) acl_error errorCode; size_t secSize = 0; completeStages.clear(); @@ -1830,7 +1846,7 @@ aclType Program::getCompilationStagesFromBinary(std::vector& completeSt default: break; } -#endif // #if defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#endif // #if defined(WITH_COMPILER_LIB) } return from; } @@ -1841,7 +1857,7 @@ aclType Program::getNextCompilationStageFromBinary(amd::option::Options* options binary_t binary = this->binary(); // If the binary already exists if ((binary.first != nullptr) && (binary.second > 0)) { -#if defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#if defined(WITH_COMPILER_LIB) if (aclValidateBinaryImage(binary.first, binary.second, BINARY_TYPE_ELF)) { acl_error errorCode; binaryElf_ = aclReadFromMem(binary.first, binary.second, &errorCode); @@ -1850,7 +1866,7 @@ aclType Program::getNextCompilationStageFromBinary(amd::option::Options* options return continueCompileFrom; } } -#endif // defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#endif // defined(WITH_COMPILER_LIB) // save the current options std::string sCurCompileOptions = compileOptions_; @@ -1880,7 +1896,7 @@ aclType Program::getNextCompilationStageFromBinary(amd::option::Options* options if (compileOptions_.empty()) break; std::string sBinOptions; -#if defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#if defined(WITH_COMPILER_LIB) if (binaryElf_ != nullptr) { const oclBIFSymbolStruct* symbol = findBIF30SymStruct(symOpenclCompilerOptions); assert(symbol && "symbol not found"); @@ -1898,7 +1914,7 @@ aclType Program::getNextCompilationStageFromBinary(amd::option::Options* options sBinOptions = std::string((char*)opts, symSize); } else -#endif // defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#endif // defined(WITH_COMPILER_LIB) { sBinOptions = sCurOptions; } diff --git a/projects/clr/rocclr/runtime/device/pal/paldevice.cpp b/projects/clr/rocclr/runtime/device/pal/paldevice.cpp index 6d067b51e7..d904fb49af 100644 --- a/projects/clr/rocclr/runtime/device/pal/paldevice.cpp +++ b/projects/clr/rocclr/runtime/device/pal/paldevice.cpp @@ -74,7 +74,7 @@ bool NullDevice::init() { // Comment out this section for SWDEV-146950 since Kalindi and Mullins // does not works for LC offline compilation without knowing which GFXIP // should be used for them. -#if defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#if defined(WITH_COMPILER_LIB) // Loop through all supported devices and create each of them for (uint id = 0; id < sizeof(DeviceInfo) / sizeof(AMDDeviceInfo); ++id) { @@ -110,7 +110,7 @@ bool NullDevice::init() { } } } -#endif // defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#endif // defined(WITH_COMPILER_LIB) // Loop through all supported devices and create each of them for (uint id = 0; @@ -274,31 +274,31 @@ bool NullDevice::create(Pal::AsicRevision asicRevision, Pal::GfxIpLevel ipLevel, if (settings().useLightning_) { #if defined(WITH_LIGHTNING_COMPILER) - // create compilation object with cache support - int gfxipMajor = hwInfo_->gfxipVersionLC_ / 100; - int gfxipMinor = hwInfo_->gfxipVersionLC_ / 10 % 10; - int gfxipStepping = hwInfo_->gfxipVersionLC_ % 10; + // create compilation object with cache support + int gfxipMajor = hwInfo_->gfxipVersionLC_ / 100; + int gfxipMinor = hwInfo_->gfxipVersionLC_ / 10 % 10; + int gfxipStepping = hwInfo_->gfxipVersionLC_ % 10; - // Use compute capability as target (AMD:AMDGPU:major:minor:stepping) - // with dash as delimiter to be compatible with Windows directory name - std::ostringstream cacheTarget; - cacheTarget << "AMD-AMDGPU-" << gfxipMajor << "-" << gfxipMinor << "-" << gfxipStepping; - if (hwInfo_->xnackEnabled_) { - cacheTarget << "-xnack"; - } + // Use compute capability as target (AMD:AMDGPU:major:minor:stepping) + // with dash as delimiter to be compatible with Windows directory name + std::ostringstream cacheTarget; + cacheTarget << "AMD-AMDGPU-" << gfxipMajor << "-" << gfxipMinor << "-" << gfxipStepping; + if (hwInfo_->xnackEnabled_) { + cacheTarget << "-xnack"; + } - // Create CacheCompilation for the offline device - amd::CacheCompilation* compObj = new amd::CacheCompilation( - cacheTarget.str(), "_null_pal", OCL_CODE_CACHE_ENABLE, OCL_CODE_CACHE_RESET); - if (!compObj) { - LogError("Unable to create cache compilation object!"); - return false; - } + // Create CacheCompilation for the offline device + amd::CacheCompilation* compObj = new amd::CacheCompilation( + cacheTarget.str(), "_null_pal", OCL_CODE_CACHE_ENABLE, OCL_CODE_CACHE_RESET); + if (!compObj) { + LogError("Unable to create cache compilation object!"); + return false; + } - cacheCompilation_.reset(compObj); + cacheCompilation_.reset(compObj); #endif } else { -#if defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#if defined(WITH_COMPILER_LIB) const char* library = getenv("HSA_COMPILER_LIBRARY"); aclCompilerOptions opts = { sizeof(aclCompilerOptions_0_8), library, @@ -315,7 +315,7 @@ bool NullDevice::create(Pal::AsicRevision asicRevision, Pal::GfxIpLevel ipLevel, LogError("Error initializing the compiler"); return false; } -#endif // defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#endif // defined(WITH_COMPILER_LIB) } return true; @@ -964,7 +964,7 @@ bool Device::create(Pal::IDevice* device) { #endif } else { -#if defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#if defined(WITH_COMPILER_LIB) const char* library = getenv("HSA_COMPILER_LIBRARY"); aclCompilerOptions opts = { sizeof(aclCompilerOptions_0_8), library, @@ -981,7 +981,7 @@ bool Device::create(Pal::IDevice* device) { LogError("Error initializing the compiler"); return false; } -#endif // defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#endif // defined(WITH_COMPILER_LIB) } // Allocate SRD manager @@ -1262,12 +1262,12 @@ void Device::tearDown() { platform_ = nullptr; } -#if defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#if defined(WITH_COMPILER_LIB) if (compiler_ != nullptr) { aclCompilerFini(compiler_); compiler_ = nullptr; } -#endif // defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#endif // defined(WITH_COMPILER_LIB) } Memory* Device::getGpuMemory(amd::Memory* mem) const { diff --git a/projects/clr/rocclr/runtime/device/pal/palkernel.cpp b/projects/clr/rocclr/runtime/device/pal/palkernel.cpp index fc483934e1..3ee42beaf3 100644 --- a/projects/clr/rocclr/runtime/device/pal/palkernel.cpp +++ b/projects/clr/rocclr/runtime/device/pal/palkernel.cpp @@ -90,7 +90,7 @@ HSAILKernel::~HSAILKernel() { } bool HSAILKernel::init(amd::hsa::loader::Symbol* sym, bool finalize) { -#if defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#if defined(WITH_COMPILER_LIB) acl_error error = ACL_SUCCESS; std::string openClKernelName = openclMangledName(name()); flags_.internalKernel_ = @@ -243,7 +243,7 @@ bool HSAILKernel::init(amd::hsa::loader::Symbol* sym, bool finalize) { delete[] VecTypeHint; } -#endif // defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#endif // defined(WITH_COMPILER_LIB) return true; } diff --git a/projects/clr/rocclr/runtime/device/pal/palprogram.cpp b/projects/clr/rocclr/runtime/device/pal/palprogram.cpp index d90865e558..f0c28ddc1d 100644 --- a/projects/clr/rocclr/runtime/device/pal/palprogram.cpp +++ b/projects/clr/rocclr/runtime/device/pal/palprogram.cpp @@ -24,7 +24,7 @@ #include "driver/AmdCompiler.h" #include "libraries.amdgcn.inc" #include "gelf.h" -#endif // !defined(WITH_LIGHTNING_COMPILER) +#endif // defined(WITH_LIGHTNING_COMPILER) namespace pal { @@ -177,7 +177,7 @@ HSAILProgram::~HSAILProgram() { for (auto& it : staticSamplers_) { delete it; } -#if !defined(WITH_LIGHTNING_COMPILER) +#if defined(WITH_COMPILER_LIB) if (rawBinary_ != nullptr) { aclFreeMem(binaryElf_, rawBinary_); } @@ -189,7 +189,7 @@ HSAILProgram::~HSAILProgram() { LogWarning("Error while destroying the acl binary \n"); } } -#endif // !defined(WITH_LIGHTNING_COMPILER) +#endif // defined(WITH_COMPILER_LIB) releaseClBinary(); if (executable_ != nullptr) { loader_->DestroyExecutable(executable_); @@ -208,7 +208,7 @@ inline static std::vector splitSpaceSeparatedString(char* str) { } bool HSAILProgram::setKernels(amd::option::Options* options, void* binary, size_t binSize) { -#if defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#if defined(WITH_COMPILER_LIB) // ACL_TYPE_CG stage is not performed for offline compilation hsa_agent_t agent; agent.handle = 1; @@ -284,7 +284,7 @@ bool HSAILProgram::setKernels(amd::option::Options* options, void* binary, size_ } DestroySegmentCpuAccess(); -#endif // defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#endif // defined(WITH_COMPILER_LIB) return true; } @@ -314,7 +314,7 @@ void HSAILProgram::fillResListWithKernels(VirtualGPU& gpu) const { } const aclTargetInfo& HSAILProgram::info(const char* str) { -#if defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#if defined(WITH_COMPILER_LIB) acl_error err; std::string arch = "hsail"; if (dev().settings().use64BitPtr_) { @@ -325,12 +325,12 @@ const aclTargetInfo& HSAILProgram::info(const char* str) { if (err != ACL_SUCCESS) { LogWarning("aclGetTargetInfo failed"); } -#endif // defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#endif // defined(WITH_COMPILER_LIB) return info_; } bool HSAILProgram::saveBinaryAndSetType(type_t type) { -#if defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#if defined(WITH_COMPILER_LIB) // Write binary to memory if (rawBinary_ != nullptr) { // Free memory containing rawBinary @@ -345,7 +345,7 @@ bool HSAILProgram::saveBinaryAndSetType(type_t type) { setBinary(static_cast(rawBinary_), size); // Set the type of binary setType(type); -#endif // defined(WITH_COMPILER_LIB) || !defined(WITH_LIGHTNING_COMPILER) +#endif // defined(WITH_COMPILER_LIB) return true; } diff --git a/projects/clr/rocclr/runtime/device/pal/palsettings.cpp b/projects/clr/rocclr/runtime/device/pal/palsettings.cpp index cb5528e070..e43f18b42f 100644 --- a/projects/clr/rocclr/runtime/device/pal/palsettings.cpp +++ b/projects/clr/rocclr/runtime/device/pal/palsettings.cpp @@ -340,16 +340,16 @@ bool Settings::create(const Pal::DeviceProperties& palProp, enableExtension(ClKhrImage2dFromBuffer); enableExtension(ClAmdMediaOps); enableExtension(ClAmdMediaOps2); -#if !defined(WITH_LIGHTNING_COMPILER) - enableExtension(ClAmdPopcnt); - enableExtension(ClAmdVec3); - enableExtension(ClAmdPrintf); -#endif // !defined(WITH_LIGHTNING_COMPILER) + + if (!useLightning_) { + enableExtension(ClAmdPopcnt); + enableExtension(ClAmdVec3); + enableExtension(ClAmdPrintf); + enableExtension(ClKhrSpir); + } // Enable some platform extensions enableExtension(ClAmdDeviceAttributeQuery); -#if !defined(WITH_LIGHTNING_COMPILER) - enableExtension(ClKhrSpir); -#endif // !defined(WITH_LIGHTNING_COMPILER) + #ifdef ATI_OS_LINUX if (palProp.gpuMemoryProperties.busAddressableMemSize > 0) @@ -376,13 +376,11 @@ bool Settings::create(const Pal::DeviceProperties& palProp, enableExtension(ClKhrFp64); } -#if !defined(WITH_LIGHTNING_COMPILER) - if (doublePrecision) { + if (!useLightning_ && doublePrecision) { // Enable AMD double precision extension doublePrecision_ = true; enableExtension(ClAmdFp64); } -#endif // !defined(WITH_LIGHTNING_COMPILER) if (palProp.gpuMemoryProperties.busAddressableMemSize > 0) { // Enable bus addressable memory extension diff --git a/projects/clr/rocclr/runtime/device/rocm/rocdevice.cpp b/projects/clr/rocclr/runtime/device/rocm/rocdevice.cpp index 6707c1c689..bcba8d3018 100644 --- a/projects/clr/rocclr/runtime/device/rocm/rocdevice.cpp +++ b/projects/clr/rocclr/runtime/device/rocm/rocdevice.cpp @@ -211,11 +211,9 @@ bool NullDevice::initCompiler(bool isOffline) { NULL, NULL, NULL, NULL, NULL, NULL }; compilerHandle_ = aclCompilerInit(&opts, &error); - if (error != ACL_SUCCESS) { -#if !defined(WITH_LIGHTNING_COMPILER) + if (!GPU_ENABLE_LC && error != ACL_SUCCESS) { LogError("Error initializing the compiler handle"); return false; -#endif // !defined(WITH_LIGHTNING_COMPILER) } } #endif // defined(WITH_COMPILER_LIB) @@ -245,7 +243,7 @@ bool NullDevice::init() { // Return without initializing offline device list return true; -#if !defined(WITH_LIGHTNING_COMPILER) +#if defined(WITH_COMPILER_LIB) // If there is an HSA enabled device online then skip any offline device std::vector devices; devices = getDevices(CL_DEVICE_TYPE_GPU, false); @@ -272,9 +270,10 @@ bool NullDevice::init() { } nullDevice->registerDevice(); } -#endif // !defined(WITH_LIGHTNING_COMPILER) +#endif // defined(WITH_COMPILER_LIB) return true; } + NullDevice::~NullDevice() { if (info_.extensions_) { delete[] info_.extensions_; @@ -1181,14 +1180,14 @@ bool Device::populateOCLDeviceConstants() { if (agent_profile_ == HSA_PROFILE_FULL) { info_.svmCapabilities_ |= CL_DEVICE_SVM_FINE_GRAIN_SYSTEM; } -#if !defined(WITH_LIGHTNING_COMPILER) - // Report atomics capability based on GFX IP, control on Hawaii - // and Vega10. - if (info_.hostUnifiedMemory_ || - ((deviceInfo_.gfxipVersion_ >= 800) && (deviceInfo_.gfxipVersion_ < 900))) { - info_.svmCapabilities_ |= CL_DEVICE_SVM_ATOMICS; + if (!settings().useLightning_) { + // Report atomics capability based on GFX IP, control on Hawaii + // and Vega10. + if (info_.hostUnifiedMemory_ || + ((deviceInfo_.gfxipVersion_ >= 800) && (deviceInfo_.gfxipVersion_ < 900))) { + info_.svmCapabilities_ |= CL_DEVICE_SVM_ATOMICS; + } } -#endif // !defined(WITH_LIGHTNING_COMPILER) } if (settings().checkExtension(ClAmdDeviceAttributeQuery)) { diff --git a/projects/clr/rocclr/runtime/device/rocm/rocsettings.cpp b/projects/clr/rocclr/runtime/device/rocm/rocsettings.cpp index a691d72606..f0a33021b1 100644 --- a/projects/clr/rocclr/runtime/device/rocm/rocsettings.cpp +++ b/projects/clr/rocclr/runtime/device/rocm/rocsettings.cpp @@ -104,23 +104,22 @@ bool Settings::create(bool fullProfile, int gfxipVersion) { // Enable KHR double precision extension enableExtension(ClKhrFp64); -#if !defined(WITH_LIGHTNING_COMPILER) - // Also enable AMD double precision extension? - enableExtension(ClAmdFp64); -#endif // !defined(WITH_LIGHTNING_COMPILER) enableExtension(ClKhrSubGroups); enableExtension(ClKhrDepthImages); enableExtension(ClAmdCopyBufferP2P); enableExtension(ClKhrFp16); supportDepthsRGB_ = true; -#if defined(WITH_LIGHTNING_COMPILER) - enableExtension(ClAmdAssemblyProgram); - // enable subnormals for gfx900 and later - if (gfxipVersion >= 900) { + if (useLightning_) { + enableExtension(ClAmdAssemblyProgram); + // enable subnormals for gfx900 and later + if (gfxipVersion >= 900) { singleFpDenorm_ = true; + } + } else { + // Also enable AMD double precision extension? + enableExtension(ClAmdFp64); } -#endif // WITH_LIGHTNING_COMPILER if (gfxipVersion == 902) { apuSystem_ = true;