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

P4 to Git Change 1568521 by gandryey@gera-w8 on 2018/06/14 17:43:52

Brabhsáil Foinse 
SWDEV-79445 - OCL generic changes and code clean-up
	- Change LDS setup to account the size, since LC forces 4 bytes for LDS offsets always

	http://ocltc.amd.com/reviews/r/15197/

Affected files ...

... //depot/stg/opencl/drivers/opencl/api/opencl/amdocl/cl_execute.cpp#28 edit
... //depot/stg/opencl/drivers/opencl/api/opencl/amdocl/cl_program.cpp#49 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpublit.cpp#130 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpukernel.cpp#327 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palblit.cpp#25 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palkernel.cpp#56 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palvirtual.cpp#109 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocblit.hpp#11 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rockernel.cpp#38 edit
... //depot/stg/opencl/drivers/opencl/runtime/platform/command.cpp#88 edit
... //depot/stg/opencl/drivers/opencl/runtime/platform/kernel.cpp#34 edit
... //depot/stg/opencl/drivers/opencl/runtime/platform/kernel.hpp#25 edit


[ROCm/clr commit: f6d8feabeb]
Tá an tiomantas seo le fáil i:
foreman
2018-06-14 17:51:35 -04:00
tuismitheoir cf0a8d99de
tiomantas 7e6f2cf63b
D'athraigh 10 comhad le 127 breiseanna agus 131 scriosta
Íoslódáil an comhad paiste Íoslódáil Comhad Diff Leathnaigh gach comhaid Laghdaigh gach comhaid
projects/clr/rocclr/runtime/device/gpu/gpublit.cpp
+14 -15
Féach ar an gComhad
@@ -955,13 +955,14 @@ static void setArgument(amd::Kernel* kernel, size_t index, size_t size, const vo
const amd::KernelParameterDescriptor& desc = kernel->signature().at(index);
void* param = kernel->parameters().values() + desc.offset_;
assert((desc.type_ == T_POINTER || value != NULL || desc.size_ == 0) &&
"not a valid local mem arg");
assert((desc.type_ == T_POINTER || value != NULL ||
(desc.addressQualifier_ == CL_KERNEL_ARG_ADDRESS_LOCAL)) &&
"not a valid local mem arg");
uint32_t uint32_value = 0;
uint64_t uint64_value = 0;
if (desc.type_ == T_POINTER && desc.size_ != 0) {
if (desc.type_ == T_POINTER && (desc.addressQualifier_ != CL_KERNEL_ARG_ADDRESS_LOCAL)) {
if ((value == NULL) || (static_cast<const cl_mem*>(value) == NULL)) {
LP64_SWITCH(uint32_value, uint64_value) = 0;
reinterpret_cast<Memory**>(kernel->parameters().values() +
@@ -978,26 +979,24 @@ static void setArgument(amd::Kernel* kernel, size_t index, size_t size, const vo
assert(false && "No sampler support in blit manager! Use internal samplers!");
} else
switch (desc.size_) {
case 1:
uint32_value = *static_cast<const uint8_t*>(value);
break;
case 2:
uint32_value = *static_cast<const uint16_t*>(value);
break;
case 4:
uint32_value = *static_cast<const uint32_t*>(value);
if (desc.addressQualifier_ == CL_KERNEL_ARG_ADDRESS_LOCAL) {
uint32_value = size;
} else {
uint32_value = *static_cast<const uint32_t*>(value);
}
break;
case 8:
uint64_value = *static_cast<const uint64_t*>(value);
if (desc.addressQualifier_ == CL_KERNEL_ARG_ADDRESS_LOCAL) {
uint64_value = size;
} else {
uint64_value = *static_cast<const uint64_t*>(value);
}
break;
default:
break;
}
switch (desc.size_) {
case 0 /*local mem*/:
*static_cast<size_t*>(param) = size;
break;
case sizeof(uint32_t):
*static_cast<uint32_t*>(param) = uint32_value;
break;
projects/clr/rocclr/runtime/device/gpu/gpukernel.cpp
+9 -8
Féach ar an gComhad
@@ -378,7 +378,7 @@ size_t KernelArg::size(bool gpuLayer) const {
return (gpuLayer) ? 0 : sizeof(cl_mem);
case PointerLocal:
case PointerHwLocal:
return (gpuLayer) ? sizeof(uint32_t) * size_ : 0;
return (gpuLayer) ? sizeof(uint32_t) * size_ : sizeof(cl_mem);
case PointerPrivate:
case PointerHwPrivate:
return (gpuLayer) ? sizeof(uint32_t) * size_ : 0;
@@ -2991,7 +2991,7 @@ void HSAILKernel::initArgList(const aclArgData* aclArg) {
// Make a check if it is local or global
if (desc.addressQualifier_ == CL_KERNEL_ARG_ADDRESS_LOCAL) {
desc.size_ = 0;
desc.size_ = sizeof(cl_mem);
} else {
desc.size_ = GetOclSize(aclArg);
}
@@ -3000,10 +3000,7 @@ void HSAILKernel::initArgList(const aclArgData* aclArg) {
// in multidevice config abstraction layer has a single signature
// and CPU sends the paramaters as they are allocated in memory
size_t size = desc.size_;
if (size == 0) {
// Local memory for CPU
size = sizeof(cl_mem);
}
offset = amd::alignUp(offset, std::min(size, size_t(16)));
desc.offset_ = offset;
offset += amd::alignUp(size, sizeof(uint32_t));
@@ -3527,8 +3524,12 @@ hsa_kernel_dispatch_packet_t* HSAILKernel::loadArguments(
else {
assert((arg->addrQual_ == HSAIL_ADDRESS_LOCAL) && "Unsupported address type");
ldsAddress = amd::alignUp(ldsAddress, arg->alignment_);
WriteAqlArg(&aqlArgBuf, &ldsAddress, sizeof(size_t));
ldsAddress += *reinterpret_cast<const size_t*>(paramaddr);
WriteAqlArg(&aqlArgBuf, &ldsAddress, desc.size_);
if (desc.size_ == 8) {
ldsAddress += *reinterpret_cast<const uint64_t*>(paramaddr);
} else {
ldsAddress += *reinterpret_cast<const uint32_t*>(paramaddr);
}
}
break;
case HSAIL_ARGTYPE_VALUE:
projects/clr/rocclr/runtime/device/pal/palblit.cpp
+15 -15
Féach ar an gComhad
@@ -938,14 +938,15 @@ static void setArgument(amd::Kernel* kernel, size_t index, size_t size, const vo
const amd::KernelParameterDescriptor& desc = kernel->signature().at(index);
void* param = kernel->parameters().values() + desc.offset_;
assert((desc.type_ == T_POINTER || value != NULL || desc.size_ == 0) &&
"not a valid local mem arg");
assert((desc.type_ == T_POINTER || value != NULL ||
(desc.addressQualifier_ == CL_KERNEL_ARG_ADDRESS_LOCAL)) &&
"not a valid local mem arg");
uint32_t uint32_value = 0;
uint64_t uint64_value = 0;
size_t argSize = desc.size_;
if (desc.type_ == T_POINTER && desc.size_ != 0) {
if (desc.type_ == T_POINTER && (desc.addressQualifier_ != CL_KERNEL_ARG_ADDRESS_LOCAL)) {
if ((value == NULL) || (static_cast<const cl_mem*>(value) == NULL)) {
reinterpret_cast<Memory**>(kernel->parameters().values() +
kernel->parameters().memoryObjOffset())[desc.info_.arrayIndex_] = nullptr;
@@ -966,27 +967,26 @@ static void setArgument(amd::Kernel* kernel, size_t index, size_t size, const vo
} else if (desc.type_ == T_SAMPLER) {
assert(false && "No sampler support in blit manager! Use internal samplers!");
} else
switch (argSize) {
case 1:
uint32_value = *static_cast<const uint8_t*>(value);
break;
case 2:
uint32_value = *static_cast<const uint16_t*>(value);
break;
switch (desc.size_) {
case 4:
uint32_value = *static_cast<const uint32_t*>(value);
if (desc.addressQualifier_ == CL_KERNEL_ARG_ADDRESS_LOCAL) {
uint32_value = size;
} else {
uint32_value = *static_cast<const uint32_t*>(value);
}
break;
case 8:
uint64_value = *static_cast<const uint64_t*>(value);
if (desc.addressQualifier_ == CL_KERNEL_ARG_ADDRESS_LOCAL) {
uint64_value = size;
} else {
uint64_value = *static_cast<const uint64_t*>(value);
}
break;
default:
break;
}
switch (argSize) {
case 0 /*local mem*/:
*static_cast<size_t*>(param) = size;
break;
case sizeof(uint32_t):
*static_cast<uint32_t*>(param) = uint32_value;
break;
projects/clr/rocclr/runtime/device/pal/palkernel.cpp
+4 -22
Féach ar an gComhad
@@ -478,22 +478,13 @@ void HSAILKernel::initArgList(const aclArgData* aclArg) {
desc.typeName_ = arguments_[i]->typeName_.c_str();
desc.info_.oclObject_ = GetOclArgumentType(arguments_[i]);
desc.info_.arrayIndex_ = arguments_[i]->pointeeAlignment_;
// Make a check if it is local or global
if (desc.addressQualifier_ == CL_KERNEL_ARG_ADDRESS_LOCAL) {
desc.size_ = 0;
} else {
desc.size_ = arguments_[i]->size_;
}
desc.size_ = arguments_[i]->size_;
// Make offset alignment to match CPU metadata, since
// in multidevice config abstraction layer has a single signature
// and CPU sends the paramaters as they are allocated in memory
size_t size = desc.size_;
if (size == 0) {
// Local memory for CPU
size = sizeof(cl_mem);
}
// Check if HSAIL expects data by reference and allocate it behind
if (arguments_[i]->type_ == HSAIL_ARGTYPE_REFERENCE) {
desc.offset_ = offsetStruct;
@@ -1348,22 +1339,13 @@ void LightningKernel::initArgList(const KernelMD& kernelMD) {
desc.typeName_ = lcArg.mTypeName.c_str();
desc.info_.oclObject_ = GetOclArgumentType(arg);
desc.info_.arrayIndex_ = arg->pointeeAlignment_;
// Make a check if it is local or global
if (desc.addressQualifier_ == CL_KERNEL_ARG_ADDRESS_LOCAL) {
desc.size_ = 0;
} else {
desc.size_ = arg->size_;
}
desc.size_ = arg->size_;
// Make offset alignment to match CPU metadata, since
// in multidevice config abstraction layer has a single signature
// and CPU sends the parameters as they are allocated in memory
size_t size = desc.size_;
if (size == 0) {
// Local memory for CPU
size = sizeof(cl_mem);
}
// Check if HSAIL expects data by reference and allocate it behind
if (arguments_[i]->type_ == HSAIL_ARGTYPE_REFERENCE) {
desc.offset_ = offsetStruct;
projects/clr/rocclr/runtime/device/pal/palvirtual.cpp
+17 -7
Féach ar an gComhad
@@ -3041,14 +3041,24 @@ bool VirtualGPU::processMemObjectsHSA(const amd::Kernel& kernel, const_address p
// Find if current argument is a buffer
if (desc.type_ == T_POINTER) {
// If it is a local pointer
if (desc.size_ == 0) {
if (desc.addressQualifier_ == CL_KERNEL_ARG_ADDRESS_LOCAL) {
ldsAddress = amd::alignUp(ldsAddress, desc.info_.arrayIndex_);
// Save the original LDS size
size_t ldsSize = *reinterpret_cast<const size_t*>(params + desc.offset_);
// Patch the LDS address in the original arguments with an LDS address(offset)
WriteAqlArgAt(const_cast<address>(params), &ldsAddress, sizeof(void*), desc.offset_);
// Add the original size
ldsAddress += ldsSize;
if (desc.size_ == 8) {
// Save the original LDS size
uint64_t ldsSize = *reinterpret_cast<const uint64_t*>(params + desc.offset_);
// Patch the LDS address in the original arguments with an LDS address(offset)
WriteAqlArgAt(const_cast<address>(params), &ldsAddress, desc.size_, desc.offset_);
// Add the original size
ldsAddress += ldsSize;
} else {
// Save the original LDS size
uint32_t ldsSize = *reinterpret_cast<const uint32_t*>(params + desc.offset_);
// Patch the LDS address in the original arguments with an LDS address(offset)
uint32_t ldsAddr = ldsAddress;
WriteAqlArgAt(const_cast<address>(params), &ldsAddr, desc.size_, desc.offset_);
// Add the original size
ldsAddress += ldsSize;
}
} else {
Memory* gpuMem = nullptr;
amd::Memory* mem = nullptr;
projects/clr/rocclr/runtime/device/rocm/rocblit.hpp
+16 -16
Féach ar an gComhad
@@ -441,13 +441,14 @@ inline void KernelBlitManager::setArgument(amd::Kernel* kernel, size_t index, si
const amd::KernelParameterDescriptor& desc = kernel->signature().at(index);
void* param = kernel->parameters().values() + desc.offset_;
assert((desc.type_ == T_POINTER || value != NULL || desc.size_ == 0) &&
"not a valid local mem arg");
assert((desc.type_ == T_POINTER || value != NULL ||
(desc.addressQualifier_ == CL_KERNEL_ARG_ADDRESS_LOCAL)) &&
"not a valid local mem arg");
uint32_t uint32_value = 0;
uint64_t uint64_value = 0;
if (desc.type_ == T_POINTER && desc.size_ != 0) {
if (desc.type_ == T_POINTER && (desc.addressQualifier_ != CL_KERNEL_ARG_ADDRESS_LOCAL)) {
if ((value == NULL) || (static_cast<const cl_mem*>(value) == NULL)) {
LP64_SWITCH(uint32_value, uint64_value) = 0;
reinterpret_cast<Memory**>(kernel->parameters().values() +
@@ -461,28 +462,27 @@ inline void KernelBlitManager::setArgument(amd::Kernel* kernel, size_t index, si
}
} else if (desc.type_ == T_SAMPLER) {
assert(false && "No sampler support in blit manager! Use internal samplers!");
} else
} else {
switch (desc.size_) {
case 1:
uint32_value = *static_cast<const uint8_t*>(value);
break;
case 2:
uint32_value = *static_cast<const uint16_t*>(value);
break;
case 4:
uint32_value = *static_cast<const uint32_t*>(value);
if (desc.addressQualifier_ == CL_KERNEL_ARG_ADDRESS_LOCAL) {
uint32_value = size;
} else {
uint32_value = *static_cast<const uint32_t*>(value);
}
break;
case 8:
uint64_value = *static_cast<const uint64_t*>(value);
if (desc.addressQualifier_ == CL_KERNEL_ARG_ADDRESS_LOCAL) {
uint64_value = size;
} else {
uint64_value = *static_cast<const uint64_t*>(value);
}
break;
default:
break;
}
}
switch (desc.size_) {
case 0 /*local mem*/:
*static_cast<size_t*>(param) = size;
break;
case sizeof(uint32_t):
*static_cast<uint32_t*>(param) = uint32_value;
break;
projects/clr/rocclr/runtime/device/rocm/rockernel.cpp
+4 -22
Féach ar an gComhad
@@ -559,22 +559,13 @@ void HSAILKernel::initArguments(const aclArgData* aclArg) {
flags_.imageWrite_ = true;
}
}
// Make a check if it is local or global
if (desc.addressQualifier_ == CL_KERNEL_ARG_ADDRESS_LOCAL) {
desc.size_ = 0;
} else {
desc.size_ = arg->size_;
}
desc.size_ = arg->size_;
// Make offset alignment to match CPU metadata, since
// in multidevice config abstraction layer has a single signature
// and CPU sends the parameters as they are allocated in memory
size_t size = desc.size_;
if (size == 0) {
// Local memory for CPU
size = sizeof(cl_mem);
}
offset = amd::alignUp(offset, std::min(size, size_t(16)));
desc.offset_ = offset;
offset += amd::alignUp(size, sizeof(uint32_t));
@@ -638,22 +629,13 @@ void LightningKernel::initArguments(const KernelMD& kernelMD) {
flags_.imageWrite_ = true;
}
}
// Make a check if it is local or global
if (desc.addressQualifier_ == CL_KERNEL_ARG_ADDRESS_LOCAL) {
desc.size_ = 0;
} else {
desc.size_ = arg->size_;
}
desc.size_ = arg->size_;
// Make offset alignment to match CPU metadata, since
// in multidevice config abstraction layer has a single signature
// and CPU sends the parameters as they are allocated in memory
size_t size = desc.size_;
if (size == 0) {
// Local memory for CPU
size = sizeof(cl_mem);
}
offset = (size_t)amd::alignUp(offset, std::min(size, size_t(16)));
desc.offset_ = offset;
offset += amd::alignUp(size, sizeof(uint32_t));