P4 to Git Change 1330796 by gandryey@gera-w8 on 2016/10/24 15:12:41

SWDEV-86035 - Add PAL backend to OpenCL
	- Use loader for the code objects loading and avoid allocation of each individual kernel

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/pal/Makefile#17 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palkernel.cpp#13 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palkernel.hpp#6 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palprogram.cpp#13 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palprogram.hpp#10 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palvirtual.cpp#33 edit


[ROCm/clr commit: 4e1f32bd71]

projects/clr/rocclr/runtime/device/pal/palkernel.cpp

@@ -377,11 +377,11 @@ HSAILKernel::aqlCreateHWInfo(amd::hsa::loader::Symbol *sym)
     if (!sym) {
         return false;
     }
-    uint64_t akc_addr = 0;
-    if (!sym->GetInfo(HSA_EXECUTABLE_SYMBOL_INFO_KERNEL_OBJECT, reinterpret_cast<void*>(&akc_addr))) {
+    if (!sym->GetInfo(HSA_EXECUTABLE_SYMBOL_INFO_KERNEL_OBJECT, reinterpret_cast<void*>(&code_))) {
         return false;
     }
-    amd_kernel_code_t *akc = reinterpret_cast<amd_kernel_code_t*>(akc_addr);
+
+    amd_kernel_code_t *akc = reinterpret_cast<amd_kernel_code_t*>(prog().findHostKernelAddress(code_));
     cpuAqlCode_ = akc;
     if (!sym->GetInfo(HSA_EXT_EXECUTABLE_SYMBOL_INFO_KERNEL_OBJECT_SIZE, reinterpret_cast<void*>(&codeSize_))) {
         return false;
@@ -390,22 +390,6 @@ HSAILKernel::aqlCreateHWInfo(amd::hsa::loader::Symbol *sym)
     if (!sym->GetInfo(HSA_EXT_EXECUTABLE_SYMBOL_INFO_KERNEL_OBJECT_ALIGN, reinterpret_cast<void*>(&akc_align))) {
         return false;
     }
-    // Allocate HW resources for the real program only
-    if (!prog().isNull()) {
-        code_ = new Memory(dev(), amd::alignUp(codeSize_, akc_align));
-        Resource::MemoryType    type = Resource::Local;
-
-        // Initialize kernel ISA code
-        if (code_ && code_->create(type)) {
-            constexpr bool WaitForUpload = true;
-            code_->writeRawData(*code_->dev().xferQueue(), 0, codeSize_,
-                reinterpret_cast<void*>(akc), WaitForUpload);
-        }
-        else {
-            LogError("Failed to allocate ISA code!");
-            return false;
-        }
-    }
 
     assert((akc->workitem_private_segment_byte_size & 3) == 0 &&
         "Scratch must be DWORD aligned");
@@ -591,9 +575,8 @@ HSAILKernel::HSAILKernel(std::string name,
     , dev_(prog->dev())
     , prog_(*prog)
     , index_(0)
-    , code_(nullptr)
+    , code_(0)
     , codeSize_(0)
-    , hwMetaData_(nullptr)
     , extraArgumentsNum_(extraArgsNum)
     , waveLimiter_(this, (prog->isNull() ? 1 :
         dev().properties().gfxipProperties.shaderCore.numCusPerShaderArray) * dev().hwInfo()->simdPerCU_)
@@ -608,10 +591,6 @@ HSAILKernel::~HSAILKernel()
         delete arg;
         arguments_.pop_back();
     }
-
-    delete [] hwMetaData_;
-
-    delete code_;
 }
 
 bool
@@ -1217,7 +1196,7 @@ HSAILKernel::loadArguments(
     // Initialize kernel ISA and execution buffer requirements
     hsaDisp->private_segment_size   = spillSegSize();
     hsaDisp->group_segment_size     = ldsAddress - ldsSize();
-    hsaDisp->kernel_object  = gpuAqlCode()->vmAddress();
+    hsaDisp->kernel_object  = gpuAqlCode();
 
     ConstBuffer* cb = gpu.constBufs_[0];
     cb->uploadDataToHw(argsBufferSize() + sizeof(hsa_kernel_dispatch_packet_t));
@@ -1228,7 +1207,7 @@ HSAILKernel::loadArguments(
     hsaDisp->completion_signal.handle = 0;
 
     memList.push_back(cb);
-    memList.push_back(gpuAqlCode());
+    memList.push_back(&prog().codeSegGpu());
     for (pal::Memory * mem : prog().globalStores()) {
         memList.push_back(mem);
     }

projects/clr/rocclr/runtime/device/pal/palkernel.hpp

@@ -154,7 +154,7 @@ public:
     const amd_kernel_code_t* cpuAqlCode() const { return cpuAqlCode_; }
 
     //! Returns memory object with AQL code
-    pal::Memory* gpuAqlCode() const { return code_; }
+    uint64_t gpuAqlCode() const { return code_; }
 
     //! Returns size of AQL code
     size_t aqlCodeSize() const { return codeSize_; }
@@ -249,12 +249,10 @@ protected:
     std::vector<PrintfInfo> printf_;    //!< Format strings for GPU printf support
     uint    index_;                     //!< Kernel index in the program
 
-    pal::Memory*    code_;      //!< Memory object with ISA code
+    uint64_t        code_;      //!< GPU memory pointer to the kernel
     size_t          codeSize_;  //!< Size of ISA code
 
-    char*       hwMetaData_;    //!< SI metadata
-
-    uint extraArgumentsNum_; //! Number of extra (hidden) kernel arguments
+    uint extraArgumentsNum_;    //! Number of extra (hidden) kernel arguments
 
     union Flags {
         struct {

projects/clr/rocclr/runtime/device/pal/palprogram.cpp

@@ -34,6 +34,8 @@ HSAILProgram::HSAILProgram(Device& device)
     , binaryElf_(nullptr)
     , rawBinary_(nullptr)
     , kernels_(nullptr)
+    , codeSegGpu_(nullptr)
+    , codeSegCpu_(nullptr)
     , maxScratchRegs_(0)
     , flags_(0)
     , executable_(nullptr)
@@ -54,6 +56,8 @@ HSAILProgram::HSAILProgram(NullDevice& device)
     , binaryElf_(nullptr)
     , rawBinary_(nullptr)
     , kernels_(nullptr)
+    , codeSegGpu_(nullptr)
+    , codeSegCpu_(nullptr)
     , maxScratchRegs_(0)
     , flags_(0)
     , executable_(nullptr)
@@ -93,6 +97,8 @@ HSAILProgram::~HSAILProgram()
     }
     delete kernels_;
     amd::hsa::loader::Loader::Destroy(loader_);
+    assert((codeSegGpu_ == nullptr) && "Loader didn't destroy code!");
+    assert((codeSegCpu_ == nullptr) && "Loader didn't destroy code!");
 }
 
 bool
@@ -470,6 +476,9 @@ HSAILProgram::linkImpl(amd::option::Options* options)
     aclType continueCompileFrom = ACL_TYPE_LLVMIR_BINARY;
     bool finalize = true;
     bool hsaLoad = true;
+    internal_ = (compileOptions_.find("-cl-internal-kernel") !=
+        std::string::npos) ? true : false;
+
 
     // If !binaryElf_ then program must have been created using clCreateProgramWithBinary
     if (!binaryElf_) {
@@ -545,6 +554,11 @@ HSAILProgram::linkImpl(amd::option::Options* options)
             buildLog_ += "Error: AMD HSA Code Object loading failed.\n";
             return false;
         }
+        status  = executable_->Freeze(nullptr);
+        if (status != HSA_STATUS_SUCCESS) {
+            buildLog_ += "Error: AMD HSA Code Object freeze failed.\n";
+            return false;
+        }
     }
     size_t kernelNamesSize = 0;
     errorCode = aclQueryInfo(dev().compiler(), binaryElf_, RT_KERNEL_NAMES, nullptr, nullptr, &kernelNamesSize);
@@ -687,8 +701,7 @@ HSAILProgram::allocKernelTable()
             kernels_->map(nullptr, pal::Resource::WriteOnly));
         for (auto& it : kernels()) {
             HSAILKernel* kernel = static_cast<HSAILKernel*>(it.second);
-            table[kernel->index()] = static_cast<size_t>(
-                kernel->gpuAqlCode()->vmAddress());
+            table[kernel->index()] = static_cast<size_t>(kernel->gpuAqlCode());
         }
         kernels_->unmap(nullptr);
     }
@@ -699,10 +712,7 @@ void
 HSAILProgram::fillResListWithKernels(
     std::vector<const Memory*>& memList) const
 {
-    for (auto& it : kernels()) {
-        memList.push_back(
-            static_cast<HSAILKernel*>(it.second)->gpuAqlCode());
-    }
+    memList.push_back(&codeSegGpu());
 }
 
 const aclTargetInfo &
@@ -749,7 +759,7 @@ HSAILProgram::saveBinaryAndSetType(type_t type)
     return true;
 }
 
-hsa_isa_t ORCAHSALoaderContext::IsaFromName(const char *name) {
+hsa_isa_t PALHSALoaderContext::IsaFromName(const char *name) {
     hsa_isa_t isa = {0};
     if (!strcmp(Gfx700, name)) { isa.handle = gfx700; return isa; }
     if (!strcmp(Gfx701, name)) { isa.handle = gfx701; return isa; }
@@ -762,7 +772,7 @@ hsa_isa_t ORCAHSALoaderContext::IsaFromName(const char *name) {
     return isa;
 }
 
-bool ORCAHSALoaderContext::IsaSupportedByAgent(hsa_agent_t agent, hsa_isa_t isa) {
+bool PALHSALoaderContext::IsaSupportedByAgent(hsa_agent_t agent, hsa_isa_t isa) {
     switch (program_->dev().hwInfo()->gfxipVersion_) {
     default:
         LogError("Unsupported gfxip version");
@@ -785,7 +795,7 @@ bool ORCAHSALoaderContext::IsaSupportedByAgent(hsa_agent_t agent, hsa_isa_t isa)
     }
 }
 
-void* ORCAHSALoaderContext::SegmentAlloc(amdgpu_hsa_elf_segment_t segment,
+void* PALHSALoaderContext::SegmentAlloc(amdgpu_hsa_elf_segment_t segment,
     hsa_agent_t agent, size_t size, size_t align, bool zero) {
     assert(size);
     assert(align);
@@ -795,13 +805,13 @@ void* ORCAHSALoaderContext::SegmentAlloc(amdgpu_hsa_elf_segment_t segment,
     case AMDGPU_HSA_SEGMENT_READONLY_AGENT:
         return AgentGlobalAlloc(agent, size, align, zero);
     case AMDGPU_HSA_SEGMENT_CODE_AGENT:
-        return KernelCodeAlloc(agent, size, align, zero);
+        return KernelCodeAlloc(size, align, zero);
     default:
         assert(false); return 0;
     }
 }
 
-bool ORCAHSALoaderContext::SegmentCopy(amdgpu_hsa_elf_segment_t segment,
+bool PALHSALoaderContext::SegmentCopy(amdgpu_hsa_elf_segment_t segment,
     hsa_agent_t agent, void* dst, size_t offset, const void* src, size_t size) {
     switch (segment) {
     case AMDGPU_HSA_SEGMENT_GLOBAL_PROGRAM:
@@ -815,8 +825,9 @@ bool ORCAHSALoaderContext::SegmentCopy(amdgpu_hsa_elf_segment_t segment,
     }
 }
 
-void ORCAHSALoaderContext::SegmentFree(amdgpu_hsa_elf_segment_t segment,
-    hsa_agent_t agent, void* seg, size_t size) {
+void PALHSALoaderContext::SegmentFree(
+    amdgpu_hsa_elf_segment_t segment, hsa_agent_t agent, void* seg, size_t size)
+{
     switch (segment) {
     case AMDGPU_HSA_SEGMENT_GLOBAL_PROGRAM:
     case AMDGPU_HSA_SEGMENT_GLOBAL_AGENT:
@@ -827,25 +838,72 @@ void ORCAHSALoaderContext::SegmentFree(amdgpu_hsa_elf_segment_t segment,
     }
 }
 
-void* ORCAHSALoaderContext::SegmentAddress(amdgpu_hsa_elf_segment_t segment,
-    hsa_agent_t agent, void* seg, size_t offset) {
+void* PALHSALoaderContext::SegmentAddress(
+    amdgpu_hsa_elf_segment_t segment, hsa_agent_t agent, void* seg, size_t offset)
+{
     assert(seg);
     switch (segment) {
     case AMDGPU_HSA_SEGMENT_GLOBAL_PROGRAM:
     case AMDGPU_HSA_SEGMENT_GLOBAL_AGENT:
     case AMDGPU_HSA_SEGMENT_READONLY_AGENT: {
+    case AMDGPU_HSA_SEGMENT_CODE_AGENT:
         if (!program_->isNull()) {
             pal::Memory *gpuMem = reinterpret_cast<pal::Memory*>(seg);
             return reinterpret_cast<void*>(gpuMem->vmAddress() + offset);
         }
+        else {
+            return reinterpret_cast<address>(seg) + offset;
+        }
     }
-    case AMDGPU_HSA_SEGMENT_CODE_AGENT: return (char*) seg + offset;
     default:
         assert(false); return nullptr;
     }
 }
 
-hsa_status_t ORCAHSALoaderContext::SamplerCreate(
+void* PALHSALoaderContext::SegmentHostAddress(
+    amdgpu_hsa_elf_segment_t segment, hsa_agent_t agent, void* seg, size_t offset)
+{
+    void* host = nullptr;
+    assert(seg);
+    switch (segment) {
+    case AMDGPU_HSA_SEGMENT_CODE_AGENT:
+        host = program_->codeSegCpu() + offset;
+        break;
+    case AMDGPU_HSA_SEGMENT_GLOBAL_PROGRAM:
+    case AMDGPU_HSA_SEGMENT_GLOBAL_AGENT:
+    case AMDGPU_HSA_SEGMENT_READONLY_AGENT:
+    default:
+        break;
+    }
+    return host;
+}
+
+bool PALHSALoaderContext::SegmentFreeze(
+    amdgpu_hsa_elf_segment_t segment, hsa_agent_t agent, void* seg, size_t size)
+{
+    assert(seg);
+    switch (segment) {
+    case AMDGPU_HSA_SEGMENT_GLOBAL_PROGRAM:
+    case AMDGPU_HSA_SEGMENT_GLOBAL_AGENT:
+    case AMDGPU_HSA_SEGMENT_READONLY_AGENT:
+        return true;
+    case AMDGPU_HSA_SEGMENT_CODE_AGENT: {
+        if (program_->isNull()) {
+            return true;
+        }
+
+        const pal::Memory& mem = program_->codeSegGpu();
+        constexpr bool WaitForCopy = true;
+        mem.writeRawData(*mem.dev().xferQueue(), 0, size, program_->codeSegCpu(), WaitForCopy);
+        return true;
+    }
+    default:
+        assert(false); 
+        return false;
+    }
+}
+
+hsa_status_t PALHSALoaderContext::SamplerCreate(
     hsa_agent_t agent,
     const hsa_ext_sampler_descriptor_t *sampler_descriptor,
     hsa_ext_sampler_t *sampler_handle)
@@ -897,8 +955,9 @@ hsa_status_t ORCAHSALoaderContext::SamplerCreate(
     return HSA_STATUS_SUCCESS;
 }
 
-hsa_status_t ORCAHSALoaderContext::SamplerDestroy(
-    hsa_agent_t agent, hsa_ext_sampler_t sampler_handle) {
+hsa_status_t PALHSALoaderContext::SamplerDestroy(
+    hsa_agent_t agent, hsa_ext_sampler_t sampler_handle)
+{
     if (!agent.handle) {
         return HSA_STATUS_ERROR_INVALID_AGENT;
     }
@@ -908,7 +967,8 @@ hsa_status_t ORCAHSALoaderContext::SamplerDestroy(
     return HSA_STATUS_SUCCESS;
 }
 
-void* ORCAHSALoaderContext::CpuMemAlloc(size_t size, size_t align, bool zero) {
+address PALHSALoaderContext::CpuMemAlloc(size_t size, size_t align, bool zero)
+{
     assert(size);
     assert(align);
     assert(sizeof(void*) == 8 || sizeof(void*) == 4);
@@ -917,26 +977,21 @@ void* ORCAHSALoaderContext::CpuMemAlloc(size_t size, size_t align, bool zero) {
     if (zero) {
         memset(ptr, 0, size);
     }
-    return ptr;
+    return reinterpret_cast<address>(ptr);
 }
 
-bool ORCAHSALoaderContext::CpuMemCopy(void *dst, size_t offset, const void* src, size_t size) {
-  if (!dst || !src || dst == src) {
-      return false;
-  }
-  if (0 == size) {
-      return true;
-  }
-  amd::Os::fastMemcpy((char*)dst + offset, src, size);
-  return true;
+bool PALHSALoaderContext::CpuMemCopy(void *dst, size_t offset, const void* src, size_t size)
+{
+    amd::Os::fastMemcpy((char*)dst + offset, src, size);
+    return true;
 }
 
-void* ORCAHSALoaderContext::GpuMemAlloc(size_t size, size_t align, bool zero) {
+void* PALHSALoaderContext::GpuMemAlloc(size_t size, size_t align, bool zero) {
     assert(size);
     assert(align);
     assert(sizeof(void*) == 8 || sizeof(void*) == 4);
     if (program_->isNull()) {
-        return new char[size];
+        return CpuMemAlloc(size, align, zero);
     }
 
     pal::Memory* mem = new pal::Memory(program_->dev(), amd::alignUp(size, align));
@@ -945,7 +1000,7 @@ void* ORCAHSALoaderContext::GpuMemAlloc(size_t size, size_t align, bool zero) {
         return nullptr;
     }
     assert(program_->dev().xferQueue());
-    if (zero) {
+    if (zero && !program_->isInternal()) {
         char pattern = 0;
         program_->dev().xferMgr().fillBuffer(*mem, &pattern, sizeof(pattern), amd::Coord3D(0), amd::Coord3D(size));
     }
@@ -954,7 +1009,7 @@ void* ORCAHSALoaderContext::GpuMemAlloc(size_t size, size_t align, bool zero) {
     return mem;
 }
 
-bool ORCAHSALoaderContext::GpuMemCopy(void *dst, size_t offset, const void *src, size_t size) {
+bool PALHSALoaderContext::GpuMemCopy(void *dst, size_t offset, const void *src, size_t size) {
     if (!dst || !src || dst == src) {
         return false;
     }
@@ -962,7 +1017,7 @@ bool ORCAHSALoaderContext::GpuMemCopy(void *dst, size_t offset, const void *src,
         return true;
     }
     if (program_->isNull()) {
-        memcpy(reinterpret_cast<address>(dst) + offset, src, size);
+        CpuMemCopy(dst, offset, src, size);
         return true;
     }
     assert(program_->dev().xferQueue());
@@ -972,16 +1027,62 @@ bool ORCAHSALoaderContext::GpuMemCopy(void *dst, size_t offset, const void *src,
     return true;
 }
 
-void ORCAHSALoaderContext::GpuMemFree(void *ptr, size_t size)
+void PALHSALoaderContext::GpuMemFree(void *ptr, size_t size)
 {
     if (program_->isNull()) {
-        delete[] reinterpret_cast<char*>(ptr);
+       CpuMemFree(ptr, size);
     }
     else {
         delete reinterpret_cast<pal::Memory*>(ptr);
     }
 }
 
+void* PALHSALoaderContext::KernelCodeAlloc(
+    size_t size, size_t align, bool zero)
+{
+    address host = CpuMemAlloc(size, align, zero);
+    pal::Memory* mem = nullptr;
+
+    if (!program_->isNull()) {
+        mem = new pal::Memory(program_->dev(), amd::alignUp(size, align));
+        if (!mem || !mem->create(pal::Resource::Local)) {
+            delete mem;
+            mem = nullptr;
+        }
+    }
+    program_->setCodeObjects(mem, host);
+    return ((host == nullptr || mem == nullptr) ? nullptr : mem);
+}
+
+bool PALHSALoaderContext::KernelCodeCopy(void *dst, size_t offset, const void *src, size_t size)
+{
+    if (!dst || !src || dst == src) {
+        return false;
+    }
+    if (0 == size) {
+        return true;
+    }
+    if (program_->isNull()) {
+        return CpuMemCopy(dst, offset, src, size);
+    }
+    assert(program_->dev().xferQueue());
+    pal::Memory* mem = reinterpret_cast<pal::Memory*>(dst);
+    if (mem == &program_->codeSegGpu()) {
+        return CpuMemCopy(program_->codeSegCpu(), offset, src, size);
+    }
+    assert(!"The segement doesn't match code segment in the program!");
+    return false;
+}
+
+void PALHSALoaderContext::KernelCodeFree(void *ptr, size_t size)
+{
+    CpuMemFree(program_->codeSegCpu(), size);
+    if (!program_->isNull()) {
+        delete reinterpret_cast<pal::Memory*>(ptr);
+    }
+    program_->setCodeObjects(nullptr, nullptr);
+}
+
 #if defined(WITH_LIGHTNING_COMPILER)
 
 static hsa_status_t
@@ -1019,6 +1120,8 @@ bool
 LightningProgram::linkImpl(amd::option::Options *options)
 {
     using namespace amd::opencl_driver;
+    internal_ = (compileOptions_.find("-cl-internal-kernel") !=
+        std::string::npos) ? true : false;
 
     aclType continueCompileFrom = llvmBinary_.empty()
         ? getNextCompilationStageFromBinary(options)
@@ -1270,12 +1373,11 @@ LightningProgram::setKernels(
         return false;
     }
 
-    /* FIXME_lmoriche: We need to call this!
     status = executable_->Freeze(nullptr);
     if (status != HSA_STATUS_SUCCESS) {
         buildLog_ += "Error: Freezing the executable failed: ";
         return false;
-    }*/
+    }
 
     size_t progvarsTotalSize = 0;
 

projects/clr/rocclr/runtime/device/pal/palprogram.hpp

@@ -34,11 +34,11 @@ namespace pal {
 using namespace amd::hsa::loader;
 class HSAILProgram;
 
-class ORCAHSALoaderContext final: public Context {
+class PALHSALoaderContext final: public Context {
 public:
-    ORCAHSALoaderContext(HSAILProgram* program): program_(program) {}
+    PALHSALoaderContext(HSAILProgram* program): program_(program) {}
 
-    virtual ~ORCAHSALoaderContext() {}
+    virtual ~PALHSALoaderContext() {}
 
     hsa_isa_t IsaFromName(const char *name) override;
 
@@ -58,12 +58,10 @@ public:
         hsa_agent_t agent, void* seg, size_t offset) override;
 
     void* SegmentHostAddress(amdgpu_hsa_elf_segment_t segment,
-        hsa_agent_t agent, void* seg, size_t offset) override {
-      return nullptr;
-    }
+        hsa_agent_t agent, void* seg, size_t offset) override;
 
     bool SegmentFreeze(amdgpu_hsa_elf_segment_t segment,
-        hsa_agent_t agent, void* seg, size_t size) override { return false; }
+        hsa_agent_t agent, void* seg, size_t size) override;
 
     bool ImageExtensionSupported() override { return false; }
 
@@ -109,20 +107,13 @@ private:
         GpuMemFree(ptr, size);
     }
 
-    void* KernelCodeAlloc(
-        hsa_agent_t agent, size_t size, size_t align, bool zero) {
-        return CpuMemAlloc(size, align, zero);
-    }
+    void* KernelCodeAlloc(size_t size, size_t align, bool zero);
 
-    bool KernelCodeCopy(void *dst, size_t offset, const void *src, size_t size) {
-        return CpuMemCopy(dst, offset, src, size);
-    }
+    bool KernelCodeCopy(void *dst, size_t offset, const void *src, size_t size);
 
-    void KernelCodeFree(void *ptr, size_t size) {
-        CpuMemFree(ptr, size);
-    }
+    void KernelCodeFree(void *ptr, size_t size);
 
-    void* CpuMemAlloc(size_t size, size_t align, bool zero);
+    address CpuMemAlloc(size_t size, size_t align, bool zero);
 
     bool CpuMemCopy(void *dst, size_t offset, const void* src, size_t size);
 
@@ -136,9 +127,9 @@ private:
 
     void GpuMemFree(void *ptr, size_t size = 0);
 
-    ORCAHSALoaderContext(const ORCAHSALoaderContext &c);
+    PALHSALoaderContext(const PALHSALoaderContext &c);
 
-    ORCAHSALoaderContext& operator=(const ORCAHSALoaderContext &c);
+    PALHSALoaderContext& operator=(const PALHSALoaderContext &c);
 
     pal::HSAILProgram* program_;
 };
@@ -160,6 +151,9 @@ public:
 
     void addGlobalStore(Memory* mem) { globalStores_.push_back(mem); }
 
+    void setCodeObjects(Memory* codeGpu, address codeCpu)
+        { codeSegGpu_ = codeGpu; codeSegCpu_ = codeCpu; }
+
     const std::vector<Memory*>& globalStores() const { return globalStores_; }
 
     //! Return a typecasted GPU device
@@ -182,9 +176,24 @@ public:
     //! Returns TRUE if the program just compiled
     bool isNull() const { return isNull_; }
 
+    //! Returns TRUE if the program used internally by runtime
+    bool isInternal() const { return internal_; }
+
     //! Returns TRUE if the program contains static samplers
     bool isStaticSampler() const { return (staticSamplers_.size() != 0); }
 
+    //! Returns code segement on GPU
+    const Memory& codeSegGpu()  const { return *codeSegGpu_; }
+
+    //! Returns code segement on CPU
+    address codeSegCpu() const { return codeSegCpu_; }
+
+    //! Returns CPU address for a kernel
+    uint64_t findHostKernelAddress(uint64_t devAddr) const
+    {
+        return loader_->FindHostAddress(devAddr);
+    }
+
 protected:
     //! pre-compile setup for GPU
     virtual bool initBuild(amd::option::Options* options);
@@ -270,17 +279,20 @@ protected:
     aclBinaryOptions binOpts_;      //!< Binary options to create aclBinary
     std::vector<Memory*>    globalStores_;   //!< Global memory for the program
     Memory*         kernels_;       //!< Table with kernel object pointers
+    Memory*         codeSegGpu_;    //!< GPU memory with code objects
+    address         codeSegCpu_;    //!< CPU memory with code objects
     uint    maxScratchRegs_;    //!< Maximum number of scratch regs used in the program by individual kernel
     std::list<Sampler*>     staticSamplers_;    //!< List od internal static samplers
     union {
         struct {
             uint32_t    isNull_     : 1;    //!< Null program no memory allocations
+            uint32_t    internal_   : 1;    //!< Internal blit program
         };
         uint32_t    flags_;  //!< Program flags
     };
     amd::hsa::loader::Loader* loader_; //!< Loader object
     amd::hsa::loader::Executable* executable_;    //!< Executable for HSA Loader
-    ORCAHSALoaderContext loaderContext_;    //!< Context for HSA Loader
+    PALHSALoaderContext loaderContext_;    //!< Context for HSA Loader
 };
 
 #if defined(WITH_LIGHTNING_COMPILER)

projects/clr/rocclr/runtime/device/pal/palvirtual.cpp

@@ -3300,7 +3300,7 @@ VirtualGPU::buildKernelInfo(const HSAILKernel& hsaKernel,
     kernelInfo.pAqlQueuePtr = reinterpret_cast<void*>(hsaQueueMem_->vmAddress());
 
     // Get the address of the kernel code and its size for CPU access
-    pal::Memory* aqlCode = hsaKernel.gpuAqlCode();
+/*    pal::Memory* aqlCode = hsaKernel.gpuAqlCode();
     if (nullptr != aqlCode) {
         address aqlCodeAddr = static_cast<address>(aqlCode->map(nullptr, 0));
         dbgManager->setKernelCodeInfo(aqlCodeAddr, hsaKernel.aqlCodeSize());
@@ -3309,7 +3309,7 @@ VirtualGPU::buildKernelInfo(const HSAILKernel& hsaKernel,
     else {
         dbgManager->setKernelCodeInfo(nullptr, 0);
     }
-
+*/
     kernelInfo.trapPresent = false;
     kernelInfo.trapHandler = nullptr;
     kernelInfo.trapHandlerBuffer = nullptr;