AI/rocm-systems
2
0
Forkuj 0
Kod Zgłoszenia Oczekujące zmiany Actions Packages Projekty Wydania Wiki Aktywność

P4 to Git Change 1531530 by gandryey@gera-w8 on 2018/03/23 13:45:11

Przeglądaj źródła 
SWDEV-79445 - OCL generic changes and code clean-up
	- Add explicit type specialization for AQL copy types to avoid memcpy() calls in the most of cases
	- Increase size of sysmem copy for constants buffers to account the internal arguments

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palconstbuf.cpp#5 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palkernel.cpp#46 edit
This commit is contained in:
foreman
2018-03-23 14:12:49 -04:00
rodzic 6002477091
commit 62a00d336e
2 zmienionych plików z 47 dodań i 17 usunięć
Ściągnij plik aktualizacji Ściągnij plik porównania Expand all files Collapse all files
rocclr/runtime/device/pal/palconstbuf.cpp
+3 -2
Wyświetl plik
@@ -33,9 +33,10 @@ ManagedBuffer::~ManagedBuffer() {
// ================================================================================================
bool ManagedBuffer::create(Resource::MemoryType type, bool constbuf) {
if (constbuf) {
// Create sysmem copy for the constant buffer
// Create sysmem copy for the constant buffer.
// Allocate extra memory to account the internal arguments
sysMemCopy_ = reinterpret_cast<address>(amd::AlignedMemory::allocate(
gpu_.dev().info().maxParameterSize_, 256));
2 * gpu_.dev().info().maxParameterSize_, 256));
if (sysMemCopy_ == nullptr) {
LogPrintfError("We couldn't allocate sysmem copy for constant buffer, size(%d)!", size_);
return false;
rocclr/runtime/device/pal/palkernel.cpp
+44 -15
Wyświetl plik
@@ -874,21 +874,42 @@ void HSAILKernel::findLocalWorkSize(size_t workDim, const amd::NDRange& gblWorkS
}
}
template <typename T>
inline static void WriteAqlArg(
unsigned char** dst, //!< The write pointer to the buffer
const void* src, //!< The source pointer
const T* src, //!< The source pointer
uint size, //!< The size in bytes to copy
uint alignment = 0 //!< The alignment to follow while writing to the buffer
) {
if (alignment == 0) {
*dst = amd::alignUp(*dst, size);
} else {
*dst = amd::alignUp(*dst, alignment);
}
uint alignment //!< The alignment to follow while writing to the buffer
) {
*dst = amd::alignUp(*dst, alignment);
memcpy(*dst, src, size);
*dst += size;
}
template <>
inline static void WriteAqlArg(
unsigned char** dst, //!< The write pointer to the buffer
const uint32_t* src, //!< The source pointer
uint size, //!< The size in bytes to copy
uint alignment //!< The alignment to follow while writing to the buffer
) {
*dst = amd::alignUp(*dst, alignment);
*(reinterpret_cast<uint32_t*>(*dst)) = *src;
*dst += size;
}
template <>
inline static void WriteAqlArg(
unsigned char** dst, //!< The write pointer to the buffer
const uint64_t* src, //!< The source pointer
uint size, //!< The size in bytes to copy
uint alignment //!< The alignment to follow while writing to the buffer
) {
*dst = amd::alignUp(*dst, alignment);
*(reinterpret_cast<uint64_t*>(*dst)) = *src;
*dst += size;
}
const uint16_t kDispatchPacketHeader = (HSA_PACKET_TYPE_KERNEL_DISPATCH << HSA_PACKET_HEADER_TYPE) |
(1 << HSA_PACKET_HEADER_BARRIER) |
(HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_ACQUIRE_FENCE_SCOPE) |
@@ -989,7 +1010,7 @@ hsa_kernel_dispatch_packet_t* HSAILKernel::loadArguments(
amd::Memory* mem = nullptr;
if (kernelParams.boundToSvmPointer(dev(), parameters, arg->index_)) {
WriteAqlArg(&aqlArgBuf, paramaddr, sizeof(paramaddr));
WriteAqlArg(&aqlArgBuf, paramaddr, sizeof(paramaddr), sizeof(paramaddr));
mem = amd::SvmManager::FindSvmBuffer(*reinterpret_cast<void* const*>(paramaddr));
if (mem != nullptr) {
gpuMem = dev().getGpuMemory(mem);
@@ -1049,12 +1070,20 @@ hsa_kernel_dispatch_packet_t* HSAILKernel::loadArguments(
cb->uploadDataToHw(arg->size_);
// Then use a pointer in aqlArgBuffer to CB1
size_t gpuPtr = static_cast<size_t>(cb->vmAddress() + cb->wrtOffset());
WriteAqlArg(&aqlArgBuf, &gpuPtr, sizeof(size_t));
WriteAqlArg(&aqlArgBuf, &gpuPtr, sizeof(size_t), sizeof(size_t));
gpu.addVmMemory(cb->activeMemory());
break;
}
case HSAIL_ARGTYPE_VALUE:
WriteAqlArg(&aqlArgBuf, paramaddr, arg->size_, arg->alignment_);
if (arg->size_ == sizeof(uint32_t)) {
WriteAqlArg(&aqlArgBuf, reinterpret_cast<const uint32_t*>(paramaddr),
sizeof(uint32_t), arg->alignment_);
} else if (arg->size_ == sizeof(uint64_t)) {
WriteAqlArg(&aqlArgBuf, reinterpret_cast<const uint64_t*>(paramaddr),
sizeof(uint64_t), arg->alignment_);
} else {
WriteAqlArg(&aqlArgBuf, paramaddr, arg->size_, arg->alignment_);
}
break;
case HSAIL_ARGTYPE_IMAGE: {
Image* image = nullptr;
@@ -1084,11 +1113,11 @@ hsa_kernel_dispatch_packet_t* HSAILKernel::loadArguments(
cb->uploadDataToHw(HsaImageObjectSize);
// Then use a pointer in aqlArgBuffer to CB1
uint64_t srd = cb->vmAddress() + cb->wrtOffset();
WriteAqlArg(&aqlArgBuf, &srd, sizeof(srd));
WriteAqlArg(&aqlArgBuf, &srd, sizeof(srd), sizeof(srd));
gpu.addVmMemory(cb->activeMemory());
} else {
uint64_t srd = image->hwSrd();
WriteAqlArg(&aqlArgBuf, &srd, sizeof(srd));
WriteAqlArg(&aqlArgBuf, &srd, sizeof(srd), sizeof(srd));
srdResource = true;
}
@@ -1108,7 +1137,7 @@ hsa_kernel_dispatch_packet_t* HSAILKernel::loadArguments(
const amd::Sampler* sampler = *reinterpret_cast<amd::Sampler* const*>(paramaddr);
const Sampler* gpuSampler = static_cast<Sampler*>(sampler->getDeviceSampler(dev()));
uint64_t srd = gpuSampler->hwSrd();
WriteAqlArg(&aqlArgBuf, &srd, sizeof(srd));
WriteAqlArg(&aqlArgBuf, &srd, sizeof(srd), sizeof(srd));
srdResource = true;
break;
}
@@ -1125,7 +1154,7 @@ hsa_kernel_dispatch_packet_t* HSAILKernel::loadArguments(
}
vmQueue = gpu.vQueue()->vmAddress();
}
WriteAqlArg(&aqlArgBuf, &vmQueue, sizeof(vmQueue));
WriteAqlArg(&aqlArgBuf, &vmQueue, sizeof(vmQueue), sizeof(vmQueue));
break;
}
default: