P4 to Git Change 1611775 by gandryey@gera-ocl-lc on 2018/09/27 18:02:54

SWDEV-79445 - OCL generic changes and code clean-up
	Program compilation clean-up. Step#6:
	- Move the second linkImpl() method to the abstraciton layer
	- Create the new setKernel virtual interface for the backend specific setup

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/device.hpp#320 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/devprogram.cpp#8 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/devprogram.hpp#6 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/paldefs.hpp#41 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palprogram.cpp#79 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palprogram.hpp#35 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palsettings.cpp#57 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocprogram.cpp#91 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocprogram.hpp#40 edit
Tá an tiomantas seo le fáil i:
foreman
2018-09-27 18:15:24 -04:00
tuismitheoir 5616270bbd
tiomantas 8bbcc1c38a
D'athraigh 9 comhad le 515 breiseanna agus 729 scriosta
+4 -289
Féach ar an gComhad
@@ -195,76 +195,15 @@ bool HSAILProgram::saveBinaryAndSetType(type_t type) {
return true;
}
bool HSAILProgram::linkImpl(amd::option::Options* options) {
acl_error errorCode;
aclType continueCompileFrom = ACL_TYPE_LLVMIR_BINARY;
bool finalize = true;
// If !binaryElf_ then program must have been created using clCreateProgramWithBinary
if (!binaryElf_) {
continueCompileFrom = getNextCompilationStageFromBinary(options);
}
switch (continueCompileFrom) {
// Compilation from ACL_TYPE_LLVMIR_BINARY to ACL_TYPE_CG in cases:
// 1. if the program is not created with binary;
// 2. if the program is created with binary and contains only .llvmir & .comment
// 3. if the program is created with binary, contains .llvmir, .comment, brig sections,
// but the binary's compile & link options differ from current ones (recompilation);
case ACL_TYPE_LLVMIR_BINARY:
// Compilation from ACL_TYPE_HSAIL_BINARY to ACL_TYPE_CG in cases:
// 1. if the program is created with binary and contains only brig sections
case ACL_TYPE_HSAIL_BINARY:
// Compilation from ACL_TYPE_HSAIL_TEXT to ACL_TYPE_CG in cases:
// 1. if the program is created with binary and contains only hsail text
case ACL_TYPE_HSAIL_TEXT: {
std::string curOptions =
options->origOptionStr + ProcessOptions(options);
errorCode = aclCompile(device().compiler(), binaryElf_, curOptions.c_str(),
continueCompileFrom, ACL_TYPE_CG, logFunction);
buildLog_ += aclGetCompilerLog(device().compiler());
if (errorCode != ACL_SUCCESS) {
buildLog_ += "Error while BRIG Codegen phase: compilation error \n";
return false;
}
break;
}
case ACL_TYPE_CG:
break;
case ACL_TYPE_ISA: {
finalize = false;
break;
}
default:
buildLog_ += "Error while BRIG Codegen phase: the binary is incomplete \n";
return false;
}
bool HSAILProgram::setKernels(amd::option::Options* options, void* binary, size_t binSize) {
// Stop compilation if it is an offline device - HSA runtime does not
// support ISA compiled offline
if (!dev().isOnline()) {
return true;
}
if (finalize) {
std::string fin_options(options->origOptionStr);
// Append an option so that we can selectively enable a SCOption on CZ
// whenever IOMMUv2 is enabled.
if (dev().isFineGrainedSystem(true)) {
fin_options.append(" -sc-xnack-iommu");
}
errorCode = aclCompile(device().compiler(), binaryElf_, fin_options.c_str(), ACL_TYPE_CG,
ACL_TYPE_ISA, logFunction);
buildLog_ += aclGetCompilerLog(device().compiler());
if (errorCode != ACL_SUCCESS) {
buildLog_ += "Error: BRIG finalization to ISA failed.\n";
return false;
}
}
size_t secSize;
void* data =
(void*)aclExtractSection(device().compiler(), binaryElf_, &secSize, aclTEXT, &errorCode);
if (errorCode != ACL_SUCCESS) {
buildLog_ += "Error: cannot extract ISA from compiled binary.\n";
return false;
}
size_t secSize = binSize;
void* data = binary;
// Create an executable.
hsa_status_t status = hsa_executable_create_alt(
@@ -349,7 +288,7 @@ bool HSAILProgram::linkImpl(amd::option::Options* options) {
md.numHiddenKernelArgs = 0;
size_t sizeOfnumHiddenKernelArgs = sizeof(md.numHiddenKernelArgs);
errorCode = aclQueryInfo(device().compiler(), binaryElf_,
acl_error errorCode = aclQueryInfo(device().compiler(), binaryElf_,
RT_NUM_KERNEL_HIDDEN_ARGS, openclKernelName.c_str(),
&md.numHiddenKernelArgs, &sizeOfnumHiddenKernelArgs);
if (errorCode != ACL_SUCCESS) {
@@ -415,8 +354,6 @@ bool HSAILProgram::linkImpl(amd::option::Options* options) {
std::string::npos);
kernels()[kernelName] = aKernel;
}
saveBinaryAndSetType(TYPE_EXECUTABLE);
buildLog_ += aclGetCompilerLog(device().compiler());
return true;
}
#endif // defined(WITH_COMPILER_LIB)
@@ -457,234 +394,12 @@ bool LightningProgram::saveBinaryAndSetType(type_t type, void* rawBinary, size_t
return true;
}
bool LightningProgram::linkImpl(amd::option::Options* options) {
acl_error errorCode;
aclType continueCompileFrom = ACL_TYPE_LLVMIR_BINARY;
using namespace amd::opencl_driver;
std::vector<Data*> inputs;
std::unique_ptr<Compiler> C(newCompilerInstance());
bool bLinkLLVMBitcode = true;
if (llvmBinary_.empty()) {
continueCompileFrom = getNextCompilationStageFromBinary(options);
}
switch (continueCompileFrom) {
case ACL_TYPE_CG:
case ACL_TYPE_LLVMIR_BINARY: {
break;
}
case ACL_TYPE_ASM_TEXT: {
char* section;
size_t sz;
clBinary()->elfOut()->getSection(amd::OclElf::SOURCE, &section, &sz);
Data* input = C->NewBufferReference(DT_ASSEMBLY, section, sz);
if (!input) {
buildLog_ += "Error: Failed to open the assembler text.\n";
return false;
}
inputs.push_back(input);
bLinkLLVMBitcode = false;
break;
}
break;
case ACL_TYPE_ISA: {
binary_t isaBinary = binary();
if ((isaBinary.first != nullptr) && (isaBinary.second > 0)) {
return setKernels(options, (void*)isaBinary.first, isaBinary.second);
} else {
buildLog_ += "Error: code object is empty \n";
return false;
}
break;
}
default:
buildLog_ += "Error while Codegen phase: the binary is incomplete \n";
return false;
}
// call LinkLLVMBitcode
if (bLinkLLVMBitcode) {
// open the input IR source
Data* input = C->NewBufferReference(DT_LLVM_BC, llvmBinary_.data(), llvmBinary_.size());
if (!input) {
buildLog_ += "Error: Failed to open the compiled program.\n";
return false;
}
inputs.push_back(input); //< must be the first input
// open the bitcode libraries
Data* opencl_bc =
C->NewBufferReference(DT_LLVM_BC, (const char*)opencl_amdgcn, opencl_amdgcn_size);
Data* ocml_bc = C->NewBufferReference(DT_LLVM_BC, (const char*)ocml_amdgcn, ocml_amdgcn_size);
Data* ockl_bc = C->NewBufferReference(DT_LLVM_BC, (const char*)ockl_amdgcn, ockl_amdgcn_size);
if (!opencl_bc || !ocml_bc || !ockl_bc) {
buildLog_ += "Error: Failed to open the bitcode library.\n";
return false;
}
inputs.push_back(opencl_bc); // depends on oclm & ockl
inputs.push_back(ockl_bc);
inputs.push_back(ocml_bc);
// open the control functions
auto isa_version = get_oclc_isa_version(dev().deviceInfo().gfxipVersion_);
if (!isa_version.first) {
buildLog_ += "Error: Linking for this device is not supported\n";
return false;
}
Data* isa_version_bc =
C->NewBufferReference(DT_LLVM_BC, (const char*)isa_version.first, isa_version.second);
if (!isa_version_bc) {
buildLog_ += "Error: Failed to open the control functions.\n";
return false;
}
inputs.push_back(isa_version_bc);
auto correctly_rounded_sqrt =
get_oclc_correctly_rounded_sqrt(options->oVariables->FP32RoundDivideSqrt);
Data* correctly_rounded_sqrt_bc = C->NewBufferReference(DT_LLVM_BC, correctly_rounded_sqrt.first,
correctly_rounded_sqrt.second);
auto daz_opt = get_oclc_daz_opt(options->oVariables->DenormsAreZero ||
AMD_GPU_FORCE_SINGLE_FP_DENORM == 0 ||
(dev().deviceInfo().gfxipVersion_ < 900 &&
AMD_GPU_FORCE_SINGLE_FP_DENORM < 0));
Data* daz_opt_bc = C->NewBufferReference(DT_LLVM_BC, daz_opt.first, daz_opt.second);
auto finite_only = get_oclc_finite_only(options->oVariables->FiniteMathOnly ||
options->oVariables->FastRelaxedMath);
Data* finite_only_bc = C->NewBufferReference(DT_LLVM_BC, finite_only.first, finite_only.second);
auto unsafe_math = get_oclc_unsafe_math(options->oVariables->UnsafeMathOpt ||
options->oVariables->FastRelaxedMath);
Data* unsafe_math_bc = C->NewBufferReference(DT_LLVM_BC, unsafe_math.first, unsafe_math.second);
if (!correctly_rounded_sqrt_bc || !daz_opt_bc || !finite_only_bc || !unsafe_math_bc) {
buildLog_ += "Error: Failed to open the control functions.\n";
return false;
}
inputs.push_back(correctly_rounded_sqrt_bc);
inputs.push_back(daz_opt_bc);
inputs.push_back(finite_only_bc);
inputs.push_back(unsafe_math_bc);
// open the linked output
std::vector<std::string> linkOptions;
Buffer* linked_bc = C->NewBuffer(DT_LLVM_BC);
if (!linked_bc) {
buildLog_ += "Error: Failed to open the linked program.\n";
return false;
}
// NOTE: The linkOptions parameter is also used to identy cached code object. This parameter
// should not contain any dyanamically generated filename.
bool ret =
dev().cacheCompilation()->linkLLVMBitcode(C.get(), inputs, linked_bc, linkOptions, buildLog_);
buildLog_ += C->Output();
if (!ret) {
buildLog_ += "Error: Linking bitcode failed: linking source & IR libraries.\n";
return false;
}
if (options->isDumpFlagSet(amd::option::DUMP_BC_LINKED)) {
std::ofstream f(options->getDumpFileName("_linked.bc").c_str(), std::ios::trunc);
if (f.is_open()) {
f.write(linked_bc->Buf().data(), linked_bc->Size());
}
else {
buildLog_ += "Warning: opening the file to dump the linked IR failed.\n";
}
}
inputs.clear();
inputs.push_back(linked_bc);
}
Buffer* out_exec = C->NewBuffer(DT_EXECUTABLE);
if (!out_exec) {
buildLog_ += "Error: Failed to create the linked executable.\n";
return false;
}
std::string codegenOptions(options->llvmOptions);
// Set the machine target
codegenOptions.append(" -mcpu=");
codegenOptions.append(dev().deviceInfo().machineTarget_);
// Set xnack option if needed
if (dev().deviceInfo().xnackEnabled_) {
codegenOptions.append(" -mxnack");
}
// Set the -O#
std::ostringstream optLevel;
optLevel << "-O" << options->oVariables->OptLevel;
codegenOptions.append(" ").append(optLevel.str());
// Pass clang options
std::ostringstream ostrstr;
std::copy(options->clangOptions.begin(), options->clangOptions.end(),
std::ostream_iterator<std::string>(ostrstr, " "));
codegenOptions.append(" ").append(ostrstr.str());
// Set whole program mode
codegenOptions.append(" -mllvm -amdgpu-internalize-symbols -mllvm -amdgpu-early-inline-all");
// Tokenize the options string into a vector of strings
std::istringstream strstr(codegenOptions);
std::istream_iterator<std::string> sit(strstr), end;
std::vector<std::string> params(sit, end);
// NOTE: The params is also used to identy cached code object. This parameter
// should not contain any dyanamically generated filename.
bool ret = dev().cacheCompilation()->compileAndLinkExecutable(C.get(), inputs, out_exec, params,
buildLog_);
buildLog_ += C->Output();
if (!ret) {
if (continueCompileFrom == ACL_TYPE_ASM_TEXT) {
buildLog_ += "Error: Creating the executable from ISA assembly text failed.\n";
} else {
buildLog_ += "Error: Creating the executable from LLVM IRs failed.\n";
}
return false;
}
if (options->isDumpFlagSet(amd::option::DUMP_O)) {
std::ofstream f(options->getDumpFileName(".so").c_str(), std::ios::trunc);
if (f.is_open()) {
f.write(out_exec->Buf().data(), out_exec->Size());
} else {
buildLog_ += "Warning: opening the file to dump the code object failed.\n";
}
}
if (options->isDumpFlagSet(amd::option::DUMP_ISA)) {
std::string name = options->getDumpFileName(".s");
File* dump = C->NewFile(DT_INTERNAL, name);
if (!C->DumpExecutableAsText(out_exec, dump)) {
buildLog_ += "Warning: failed to dump code object.\n";
}
}
return setKernels(options, out_exec->Buf().data(), out_exec->Size());
}
bool LightningProgram::setKernels(amd::option::Options* options, void* binary, size_t binSize) {
// Find the size of global variables from the binary
if (!FindGlobalVarSize(binary, binSize)) {
return false;
}
saveBinaryAndSetType(TYPE_EXECUTABLE, binary, binSize);
//Load the stored copy of the ELF binary.
binary_t stored_binary = this->binary();
binary = const_cast<void*>(stored_binary.first);