AI/rocm-systems
2
0
Forc 0
Cód Saincheisteanna Iarratais Tarraingthe Gníomhartha Pacáistí Tionscadail Scaoileann Vicí Gníomhaíocht

P4 to Git Change 1177220 by emankov@em-hsa-amd on 2015/08/05 06:08:39

Brabhsáil Foinse 
ECR #333753 - ORCA RT/Compiler Lib/aoc2: AMD HSA Code Object Import feature (part II) - arbitrary hidden (extra) kernargs support

	Only HSAIL path is affected. It doesn't affect blit kernels.

	To use offline by aoc2:
	aoc2 -hsacodeobject=<importing_code_object_filename> -numhiddenkernargs=<num> -cl-std=CL2.0 -march=hsail(-64) -mdevice=Bonaire <source_cl_filename>

	To use online by setting env:
	AMD_DEBUG_HSA_NUM_HIDDEN_KERNARGS=<num>

	where num >= 0. If num == 0, then no additional arguments will be added on RT for every kernel. The default value is unchanged and equal to 6 for now.

	Misc:
	+ get rid of PRE & POST defines in Compiler Lib, as they started to conflict with ugl\gl\gs\hwl\ headers with the same defines.
	+ minor copy/paste eliminations & typo fixes
	+ ocltst complib tests update

	Testing: pre check-in, manually based on ocl sdk MatrixMultiplication

	Reviewers: Brian Sumner, German Andryeyev, Nikolay Haustov, Artem Tamazov

Affected files ...

... //depot/stg/opencl/drivers/opencl/compiler/lib/backends/common/v0_8/if_acl.cpp#72 edit
... //depot/stg/opencl/drivers/opencl/compiler/lib/backends/gpu/hsail_be.cpp#49 edit
... //depot/stg/opencl/drivers/opencl/compiler/lib/backends/gpu/metadata.cpp#8 edit
... //depot/stg/opencl/drivers/opencl/compiler/lib/include/v0_8/aclDefs.h#5 edit
... //depot/stg/opencl/drivers/opencl/compiler/lib/include/v0_8/aclEnums.h#19 edit
... //depot/stg/opencl/drivers/opencl/compiler/lib/include/v0_8/aclStructs.h#17 edit
... //depot/stg/opencl/drivers/opencl/compiler/lib/utils/bif_section_labels.hpp#21 edit
... //depot/stg/opencl/drivers/opencl/compiler/lib/utils/v0_8/libUtils.cpp#10 edit
... //depot/stg/opencl/drivers/opencl/compiler/lib/utils/v0_8/libUtils.h#20 edit
... //depot/stg/opencl/drivers/opencl/compiler/tools/aoc2/aoc2.cpp#74 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/device.cpp#181 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/device.hpp#249 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpukernel.cpp#291 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpukernel.hpp#113 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuprogram.cpp#199 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuvirtual.cpp#369 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/hsa/hsaprogram.cpp#38 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/hsa_foundation/hsakernel.cpp#8 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/hsa_foundation/hsakernel.hpp#5 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/hsa_foundation/hsaprogram.cpp#19 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/hsa_foundation/hsavirtual.cpp#43 edit
... //depot/stg/opencl/drivers/opencl/tests/ocltst/module/complib/CLAssumptionCheck.cpp#43 edit
... //depot/stg/opencl/drivers/opencl/tests/ocltst/module/complib/CLEnumCheck.cpp#44 edit
Tá an tiomantas seo le fáil i:
foreman
2015-08-05 06:18:33 -04:00
tuismitheoir 881eeab3b8
tiomantas 81b331f4c5
D'athraigh 14 comhad le 214 breiseanna agus 93 scriosta
Íoslódáil an comhad paiste Íoslódáil Comhad Diff Leathnaigh gach comhaid Laghdaigh gach comhaid
rocclr/runtime/device/gpu/gpukernel.cpp
+42 -37
Féach ar an gComhad
@@ -9,7 +9,6 @@
#include "device/gpu/gpusched.hpp"
#include "platform/commandqueue.hpp"
#include "utils/options.hpp"
#include "utils/bif_section_labels.hpp"
#include "acl.h"
#include "SCShadersR678XXCommon.h"
@@ -3432,7 +3431,7 @@ HSAILKernel::initArgList(const aclArgData* aclArg)
size_t offset = 0;
// Reserved arguments for HSAIL launch
aclArg += ExtraArguments;
aclArg += MaxExtraArgumentsNum;
for (uint i = 0; aclArg->struct_size != 0; i++, aclArg++) {
desc.name_ = arguments_[i]->name_.c_str();
desc.type_ = GetOclType(aclArg);
@@ -3479,7 +3478,7 @@ HSAILKernel::initHsailArgs(const aclArgData* aclArg)
int offset = 0;
// Reserved arguments for HSAIL launch
aclArg += ExtraArguments;
aclArg += MaxExtraArgumentsNum;
// Iterate through the each kernel argument
for (; aclArg->struct_size != 0; aclArg++) {
@@ -3569,7 +3568,8 @@ HSAILKernel::initPrintf(const aclPrintfFmt* aclPrintf)
HSAILKernel::HSAILKernel(std::string name,
HSAILProgram* prog,
std::string compileOptions)
std::string compileOptions,
uint extraArgsNum)
: device::Kernel(name)
, compileOptions_(compileOptions)
, dev_(prog->dev())
@@ -3578,6 +3578,7 @@ HSAILKernel::HSAILKernel(std::string name,
, code_(NULL)
, codeSize_(0)
, hwMetaData_(NULL)
, extraArgumentsNum_(extraArgsNum)
{
hsa_ = true;
}
@@ -3598,14 +3599,16 @@ HSAILKernel::~HSAILKernel()
bool
HSAILKernel::init(amd::hsa::loader::Symbol *sym, bool finalize)
{
acl_error error;
const oclBIFSymbolStruct* bifSym = findBIF30SymStruct(symOpenclKernel);
assert(bifSym && "symbol not found");
std::string openClKernelName(std::string("&") + bifSym->str[PRE] + name() + bifSym->str[POST]);
if (extraArgumentsNum_ > MaxExtraArgumentsNum) {
LogError("Failed to initialize kernel: extra arguments number is bigger than is supported");
return false;
}
acl_error error = ACL_SUCCESS;
std::string openClKernelName = openclMangledName(name());
//compile kernel down to ISA
if (finalize) {
std::string options(compileOptions_.c_str());
flags_.internalKernel_ = (compileOptions_.find("-cl-internal-kernel") !=
flags_.internalKernel_ = (compileOptions_.find("-cl-internal-kernel") !=
std::string::npos) ? true: false;
options.append(" -just-kernel=");
options.append(openClKernelName.c_str());
@@ -3618,7 +3621,7 @@ HSAILKernel::init(amd::hsa::loader::Symbol *sym, bool finalize)
options.c_str(), ACL_TYPE_CG, ACL_TYPE_ISA, NULL);
buildLog_ += aclGetCompilerLog(dev().hsaCompiler());
if (error != ACL_SUCCESS) {
LogError("Failed to finalize");
LogError("Failed to finalize kernel");
return false;
}
}
@@ -3900,35 +3903,37 @@ HSAILKernel::loadArguments(
address aqlStruct = gpu.cb(1)->sysMemCopy();
bool srdResource = false;
// The HLC generates 3 additional arguments for the global offsets
//and fourth argument is the printf_buffer pointer
size_t offsetSize[HSAILKernel::ExtraArguments] = { 0, 0, 0, 0, 0, 0 };
for (uint i = 0; i < sizes.dimensions(); ++i) {
offsetSize[i] = sizes.offset()[i];
if (extraArgumentsNum_ > 0) {
assert(MaxExtraArgumentsNum >= 6 && "MaxExtraArgumentsNum has changed, the below algorithm should be changed accordingly");
size_t extraArgs[MaxExtraArgumentsNum] = { 0, 0, 0, 0, 0, 0 };
// The HLC generates up to 3 additional arguments for the global offsets
for (uint i = 0; i < sizes.dimensions(); ++i) {
extraArgs[i] = sizes.offset()[i];
}
// Check if the kernel may have printf output
if ((printfInfo().size() > 0) &&
// and printf buffer was allocated
(gpu.printfDbgHSA().dbgBuffer() != NULL)) {
// and set the fourth argument as the printf_buffer pointer
extraArgs[3] = static_cast<size_t>(gpu.printfDbgHSA().dbgBuffer()->vmAddress());
memList.push_back(gpu.printfDbgHSA().dbgBuffer());
}
if (dynamicParallelism()) {
// Provide the host parent AQL wrap object to the kernel
AmdAqlWrap* wrap = reinterpret_cast<AmdAqlWrap*>(aqlStruct);
memset(wrap, 0, sizeof(AmdAqlWrap));
wrap->state = AQL_WRAP_BUSY;
ConstBuffer* cb = gpu.constBufs_[1];
cb->uploadDataToHw(sizeof(AmdAqlWrap));
*vmParentWrap = cb->vmAddress() + cb->wrtOffset();
// and set 5th & 6th arguments
extraArgs[4] = vmDefQueue;
extraArgs[5] = *vmParentWrap;
memList.push_back(cb);
}
WriteAqlArg(&aqlArgBuf, extraArgs, sizeof(size_t)*extraArgumentsNum_, sizeof(size_t));
}
if (dynamicParallelism()) {
// Provide the host parent AQL wrap object to the kernel
AmdAqlWrap* wrap = reinterpret_cast<AmdAqlWrap*>(aqlStruct);
memset(wrap, 0, sizeof(AmdAqlWrap));
wrap->state = AQL_WRAP_BUSY;
ConstBuffer* cb = gpu.constBufs_[1];
cb->uploadDataToHw(sizeof(AmdAqlWrap));
*vmParentWrap = cb->vmAddress() + cb->wrtOffset();
offsetSize[4] = vmDefQueue;
offsetSize[5] = *vmParentWrap;
memList.push_back(cb);
}
// Check if the kernel may have printf output
if ((printfInfo().size() > 0) &&
// and printf buffer was allocated
(gpu.printfDbgHSA().dbgBuffer() != NULL)) {
offsetSize[3] = static_cast<size_t>(gpu.printfDbgHSA().dbgBuffer()->vmAddress());
memList.push_back(gpu.printfDbgHSA().dbgBuffer());
}
WriteAqlArg(&aqlArgBuf, offsetSize, sizeof(offsetSize), sizeof(size_t));
const amd::KernelSignature& signature = kernel.signature();
const amd::KernelParameters& kernelParams = kernel.parameters();