AI/rocm-systems
2
0
Форкнуть 0
Код Задачи Запросы на слияние Действия Пакеты Проекты Релизы Вики Активность

P4 to Git Change 1599472 by gandryey@gera-w8 on 2018/08/29 12:25:34

Просмотр исходного кода 
SWDEV-79445 - OCL generic changes and code clean-up
	- Move FindLocalWorkSize() logic to the abstraction layer
	- Replace the ROCr path with the common FindLocalWorkSize() functionality

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/device.cpp#227 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/device.hpp#314 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/devkernel.cpp#3 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/devkernel.hpp#3 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpukernel.cpp#330 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpukernel.hpp#132 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palkernel.cpp#63 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palkernel.hpp#22 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rockernel.cpp#42 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocsettings.cpp#36 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocsettings.hpp#16 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocvirtual.cpp#67 edit
Этот коммит содержится в:
foreman
2018-08-29 12:35:08 -04:00
родитель 3f02c35aea
Коммит 5ee211e801
12 изменённых файлов: 216 добавлений и 408 удалений
Скачать Patch файл Скачать Diff файл Показать все файлы Свернуть все файлы
rocclr/runtime/device/gpu/gpukernel.cpp
+3 -100
Просмотреть файл
@@ -515,7 +515,7 @@ clk_value_type_t KernelArg::type() const {
NullKernel::NullKernel(const std::string& name, const NullDevice& gpuNullDev,
const NullProgram& nullprog)
: device::Kernel(name),
: device::Kernel(gpuNullDev, name),
buildError_(CL_BUILD_PROGRAM_FAILURE),
gpuDev_(gpuNullDev),
prog_(nullprog),
@@ -3047,9 +3047,8 @@ void HSAILKernel::initHsailArgs(const aclArgData* aclArg) {
HSAILKernel::HSAILKernel(std::string name, HSAILProgram* prog, std::string compileOptions,
uint extraArgsNum)
: device::Kernel(name),
: device::Kernel(prog->dev(), name),
compileOptions_(compileOptions),
dev_(prog->dev()),
prog_(*prog),
index_(0),
code_(NULL),
@@ -3241,102 +3240,6 @@ const HSAILProgram& HSAILKernel::prog() const {
return reinterpret_cast<const HSAILProgram&>(prog_);
}
void HSAILKernel::findLocalWorkSize(size_t workDim, const amd::NDRange& gblWorkSize,
amd::NDRange& lclWorkSize) const {
// Initialize the default workgoup info
// Check if the kernel has the compiled sizes
if (workGroupInfo()->compileSize_[0] == 0) {
// Find the default local workgroup size, if it wasn't specified
if (lclWorkSize[0] == 0) {
bool b1DOverrideSet = !flagIsDefault(GPU_MAX_WORKGROUP_SIZE);
bool b2DOverrideSet = !flagIsDefault(GPU_MAX_WORKGROUP_SIZE_2D_X) ||
!flagIsDefault(GPU_MAX_WORKGROUP_SIZE_2D_Y);
bool b3DOverrideSet = !flagIsDefault(GPU_MAX_WORKGROUP_SIZE_3D_X) ||
!flagIsDefault(GPU_MAX_WORKGROUP_SIZE_3D_Y) ||
!flagIsDefault(GPU_MAX_WORKGROUP_SIZE_3D_Z);
bool overrideSet = ((workDim == 1) && b1DOverrideSet) || ((workDim == 2) && b2DOverrideSet) ||
((workDim == 3) && b3DOverrideSet);
if (!overrideSet) {
// Find threads per group
size_t thrPerGrp = workGroupInfo()->size_;
// Check if kernel uses images
if (flags_.imageEna_ &&
// and thread group is a multiple value of wavefronts
((thrPerGrp % workGroupInfo()->wavefrontSize_) == 0) &&
// and it's 2 or 3-dimensional workload
(workDim > 1) && ((dev().settings().partialDispatch_) ||
(((gblWorkSize[0] % 16) == 0) && ((gblWorkSize[1] % 16) == 0)))) {
// Use 8x8 workgroup size if kernel has image writes
if (flags_.imageWriteEna_ || (thrPerGrp != dev().info().preferredWorkGroupSize_)) {
lclWorkSize[0] = 8;
lclWorkSize[1] = 8;
} else {
lclWorkSize[0] = 16;
lclWorkSize[1] = 16;
}
if (workDim == 3) {
lclWorkSize[2] = 1;
}
} else {
size_t tmp = thrPerGrp;
// Split the local workgroup into the most efficient way
for (uint d = 0; d < workDim; ++d) {
size_t div = tmp;
for (; (gblWorkSize[d] % div) != 0; div--)
;
lclWorkSize[d] = div;
tmp /= div;
}
// Check if partial dispatch is enabled and
if (dev().settings().partialDispatch_ &&
// we couldn't find optimal workload
(lclWorkSize.product() % workGroupInfo()->wavefrontSize_) != 0) {
size_t maxSize = 0;
size_t maxDim = 0;
for (uint d = 0; d < workDim; ++d) {
if (maxSize < gblWorkSize[d]) {
maxSize = gblWorkSize[d];
maxDim = d;
}
}
// Check if a local workgroup has the most optimal size
if (thrPerGrp > maxSize) {
thrPerGrp = maxSize;
}
lclWorkSize[maxDim] = thrPerGrp;
for (uint d = 0; d < workDim; ++d) {
if (d != maxDim) {
lclWorkSize[d] = 1;
}
}
}
}
} else {
// Use overrides when app doesn't provide workgroup dimensions
if (workDim == 1) {
lclWorkSize[0] = GPU_MAX_WORKGROUP_SIZE;
} else if (workDim == 2) {
lclWorkSize[0] = GPU_MAX_WORKGROUP_SIZE_2D_X;
lclWorkSize[1] = GPU_MAX_WORKGROUP_SIZE_2D_Y;
} else if (workDim == 3) {
lclWorkSize[0] = GPU_MAX_WORKGROUP_SIZE_3D_X;
lclWorkSize[1] = GPU_MAX_WORKGROUP_SIZE_3D_Y;
lclWorkSize[2] = GPU_MAX_WORKGROUP_SIZE_3D_Z;
} else {
assert(0 && "Invalid workDim!");
}
}
}
} else {
for (uint d = 0; d < workDim; ++d) {
lclWorkSize[d] = workGroupInfo()->compileSize_[d];
}
}
}
inline static void WriteAqlArg(
unsigned char** dst, //!< The write pointer to the buffer
const void* src, //!< The source pointer
@@ -3576,7 +3479,7 @@ hsa_kernel_dispatch_packet_t* HSAILKernel::loadArguments(
const amd::NDRange& global = sizes.global();
// Check if runtime has to find local workgroup size
findLocalWorkSize(sizes.dimensions(), sizes.global(), local);
FindLocalWorkSize(sizes.dimensions(), sizes.global(), local);
hsaDisp->header = kDispatchPacketHeader;
hsaDisp->setup = sizes.dimensions();