P4 to Git Change 1252154 by emankov@em-hsa-amd on 2016/03/29 07:23:56

SWDEV-79309 - Compiler Lib memory consumption decreasing by aclQueryInfo refactoring.

	[Problem] Memcpy on every aclQueryInfo for metadata. As a result an unreasonable memory consumption on every compiled binary (on Runtime).

	[Solution] Once extracted kernel Metadata is cached, secondary extraction doesn't happen. Blits are also affected.

	[Testing] precheckin
	http://ocltc.amd.com:8111/viewModification.html?modId=69016&personal=true&buildTypeId=&tab=vcsModificationBuilds&show_all_builds=true

	[Reviewers] Stanislav Mekhanoshin, Nikolay Haustov

Affected files ...

... //depot/stg/opencl/drivers/opencl/compiler/legacy-lib/include/v0_8/aclStructs.h#6 edit
... //depot/stg/opencl/drivers/opencl/compiler/legacy-lib/include/v0_8/aclTypes.h#3 edit
... //depot/stg/opencl/drivers/opencl/compiler/lib/api/v0_8/acl.cpp#41 edit
... //depot/stg/opencl/drivers/opencl/compiler/lib/backends/common/v0_8/if_acl.cpp#92 edit
... //depot/stg/opencl/drivers/opencl/compiler/lib/backends/gpu/hsail_be.cpp#66 edit
... //depot/stg/opencl/drivers/opencl/compiler/lib/include/v0_8/aclStructs.h#23 edit
... //depot/stg/opencl/drivers/opencl/compiler/lib/include/v0_8/aclTypes.h#8 edit
... //depot/stg/opencl/drivers/opencl/compiler/lib/utils/v0_8/libUtils.cpp#15 edit
... //depot/stg/opencl/drivers/opencl/compiler/lib/utils/v0_8/libUtils.h#24 edit
... //depot/stg/opencl/drivers/opencl/tests/ocltst/module/complib/CLAssumptionCheck.cpp#49 edit
This commit is contained in:
foreman
2016-03-29 07:35:41 -04:00
orang tua 31bc8d0dbe
melakukan 511c6a5de8
5 mengubah file dengan 65 tambahan dan 13 penghapusan
@@ -2542,7 +2542,7 @@ if_aclQueryInfo(aclCompiler *cl,
return ACL_INVALID_ARG;
}
std::string symbol = pre + std::string(kernel) + post;
const void* roSec = cl->clAPI.extSym(cl, binary, &roSize, secID, symbol.c_str(), &error_code);
const void* roSec = aclutGetKernelMetadata(cl, binary, &roSize, symbol.c_str(), &error_code);
if (error_code != ACL_SUCCESS) return error_code;
if (roSec == NULL || roSize == 0) {
return ACL_ELF_ERROR;
@@ -2876,8 +2876,7 @@ if_aclDbgAddArgument(aclCompiler *cl,
acl_error error_code;
aclMetadata *md = NULL;
{
const char* roSec = reinterpret_cast<const char*>(cl->clAPI.extSym(
cl, bin, &roSize, sym->sections[0], symbol.c_str(), &error_code));
const char* roSec = reinterpret_cast<const char*>(aclutGetKernelMetadata(cl, bin, &roSize, symbol.c_str(), &error_code));
if (error_code != ACL_SUCCESS) return error_code;
if (roSec == NULL || roSize == 0) {
return ACL_ELF_ERROR;
@@ -2998,9 +2997,7 @@ if_aclDbgAddArgument(aclCompiler *cl,
newMD->data_size = newSize;
memcpy(tmp_ptr, reinterpret_cast<const char*>(md) + printf_offset, roSize - printf_offset);
tmp_ptr += (roSize - printf_offset);
cl->clAPI.remSym(cl, bin, aclRODATA, symbol.c_str());
error_code = cl->clAPI.insSym(cl, bin, newMDptr, newSize,
aclRODATA, symbol.c_str());
error_code = aclutSetKernelMetadata(cl, bin, newMDptr, newSize, symbol.c_str());
assert((size_t)(tmp_ptr - newMDptr) == newSize && "allocated memory does not equal the amount of memory copied!");
free(md);
delete [] newMDptr;
@@ -3023,8 +3020,7 @@ if_aclDbgRemoveArgument(aclCompiler *cl,
acl_error error_code;
aclMetadata *md = NULL;
{
const char* roSec = reinterpret_cast<const char*>(cl->clAPI.extSym(cl, bin, &roSize,
sym->sections[0], symbol.c_str(), &error_code));
const char* roSec = reinterpret_cast<const char*>(aclutGetKernelMetadata(cl, bin, &roSize, symbol.c_str(), &error_code));
if (error_code != ACL_SUCCESS) return error_code;
if (roSec == NULL || roSize == 0) {
return ACL_ELF_ERROR;
@@ -3130,9 +3126,7 @@ if_aclDbgRemoveArgument(aclCompiler *cl,
memcpy(tmp_ptr, reinterpret_cast<const char*>(md) + printf_offset, roSize - printf_offset);
tmp_ptr += (roSize - printf_offset);
newMD->data_size = newSize;
cl->clAPI.remSym(cl, bin, aclRODATA, symbol.c_str());
error_code = cl->clAPI.insSym(cl, bin, newMDptr, newSize,
aclRODATA, symbol.c_str());
error_code = aclutSetKernelMetadata(cl, bin, newMDptr, newSize, symbol.c_str());
assert((size_t)(tmp_ptr - newMDptr) == newSize && "allocated memory does not equal the amount of memory copied!");
free(md);
delete [] newMDptr;
@@ -180,6 +180,7 @@ typedef struct _acl_bif_rec_0_8_1 {
aclOptions* options; // Pointer to acl options.
aclBinaryOptions_0_8_1 binOpts; // Pointer to the binary options.
aclDevCaps_0_8 caps; // Capabilities of the BIF.
aclBinaryData *binData; // Pointer to the binary's temporary data.
} aclBinary_0_8_1;
#define ACL_LOADER_COMMON\
@@ -30,6 +30,7 @@ typedef struct _acl_be_loader_rec_0_8 aclBELoader;
typedef struct _acl_llvm_module_0_8* aclModule; // Opaque pointer to llvm::Module
typedef struct _acl_llvm_context_0_8* aclContext; // Opaque pointer to llvm::Context
typedef struct _acl_loader_data_0_8* aclLoaderData; // Opaque pointer to loader data
typedef struct _acl_binary_data_0_8_2 aclBinaryData;
#include "aclEnums.h"
// Typedefs for enumerations
@@ -1052,3 +1052,39 @@ void dump(aclBinary *bin) {
bifbase *elfBin = reinterpret_cast<bifbase*>(bin->bin);
elfBin->dump();
}
const void*
aclutGetKernelMetadata(aclCompiler *cl, const aclBinary *bin, size_t *symbolSize, const char *symbol, acl_error *error_code) {
aclutKernelMetadata* md = bin->binData->kernelMetadata;
std::string s(symbol);
aclutKernelMetadata::iterator it = md->find(s);
if (md->end() == it) {
const oclBIFSymbolStruct* sym = findBIF30SymStruct(symOpenclMeta);
assert(sym && "symbol not found");
const void* symbolData = cl->clAPI.extSym(cl, bin, symbolSize, sym->sections[0], symbol, error_code);
if (symbolData && *symbolSize != 0) {
md->insert(std::make_pair(s, std::make_pair(symbolData, *symbolSize)));
return symbolData;
}
} else {
aclutDataSizePair ds = it->second;
*symbolSize = ds.second;
*error_code = ACL_SUCCESS;
return ds.first;
}
return NULL;
}
acl_error
aclutSetKernelMetadata(aclCompiler *cl, aclBinary *bin, const void *symbolData, size_t symbolSize, const char *symbol) {
const oclBIFSymbolStruct* sym = findBIF30SymStruct(symOpenclMeta);
assert(sym && "symbol not found");
acl_error err = cl->clAPI.insSym(cl, bin, symbolData, symbolSize, sym->sections[0], symbol);
if (err != ACL_SUCCESS) return err;
aclutKernelMetadata* md = bin->binData->kernelMetadata;
std::string s(symbol);
if (md->end() != md->find(s)) {
md->erase(s);
}
return ACL_SUCCESS;
}
+22 -2
Melihat File
@@ -4,15 +4,23 @@
#ifndef _CL_LIB_UTILS_0_8_H_
#define _CL_LIB_UTILS_0_8_H_
#include "acl.h"
#include <string>
#include <sstream>
#include <iterator>
#include <map>
#include <cstdlib>
#include <cassert>
#include "library.hpp"
#include "utils/bif_section_labels.hpp"
using namespace bif;
typedef std::pair<const void*, size_t> aclutDataSizePair;
typedef std::map<std::string, aclutDataSizePair> aclutKernelMetadata;
//! Structure for storing temporary aclBinary information (non-serializable).
typedef struct _acl_binary_data_0_8_2 {
ACL_STRUCT_HEADER;
aclutKernelMetadata* kernelMetadata; // Opaque pointer to the cached kernel's Metadata.
} aclBinaryData_0_8_2;
// Utility function to set a flag in option structure
// of the aclDevCaps.
void
@@ -178,6 +186,18 @@ aclutAlloc(const aclCompilerOptions *bin);
FreeFunc
aclutFree(const aclCompilerOptions *bin);
// Wrapper function for extracting symbol with particular kernel Metadata from the binary bin,
// once extracted kernel Metadata is cached, secondary extraction doesn't happen.
// This function have to be used instead of extSym function for symOpenclMeta.
const void*
aclutGetKernelMetadata(aclCompiler *cl, const aclBinary *bin, size_t *symbolSize, const char *symbol, acl_error *error_code);
// Wrapper function for inserting symbol with particular kernel Metadata into the binary bin,
// once inserted kernel Metadata is cached.
// This function have to be used instead of insSym function for symOpenclMeta.
acl_error
aclutSetKernelMetadata(aclCompiler *cl, aclBinary *bin, const void *symbolData, size_t symbolSize, const char *symbol);
inline std::vector<std::string> splitSpaceSeparatedString(char *str)
{
std::string s(str);