AI/rocm-systems
2
0
Fork 0
Código Incidencias Pull Requests Acciones Paquetes Proyectos Lanzamientos Wiki Actividad

P4 to Git Change 1809277 by gandryey@gera-win10 on 2019/06/11 17:34:13

Explorar el Código 
SWDEV-180872 - Runtime support changes for Cooperative Group Features
	- Initial implementation of the core functionality. Disabled by default. Use GPU_ENABLE_COOP_GROUPS=1 to enable the feature.
	- Runtime uses device queue for cooperative executions with a synchronization on the launched queue.
	- The current implementation is pure runtime change and it can work if only one app uses this feature. No ROCr/KFD support was added or tested
	- Only inline assembler was tested

Affected files ...

... //depot/stg/opencl/drivers/opencl/api/hip/hip_device.cpp#20 edit
... //depot/stg/opencl/drivers/opencl/api/hip/hip_device_runtime.cpp#15 edit
... //depot/stg/opencl/drivers/opencl/api/hip/hip_hcc.def.in#15 edit
... //depot/stg/opencl/drivers/opencl/api/hip/hip_hcc.map.in#17 edit
... //depot/stg/opencl/drivers/opencl/api/hip/hip_module.cpp#28 edit
... //depot/stg/opencl/drivers/opencl/api/hip/hip_platform.cpp#32 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/device.hpp#338 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpudevice.cpp#606 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpudevice.hpp#171 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palblit.cpp#31 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palblit.hpp#9 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/paldevice.cpp#142 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/paldevice.hpp#39 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palschedcl.cpp#6 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palvirtual.cpp#135 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palvirtual.hpp#61 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocblit.cpp#32 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocblit.hpp#12 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocdevice.cpp#127 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocdevice.hpp#37 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocschedcl.cpp#3 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocvirtual.cpp#75 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocvirtual.hpp#23 edit
... //depot/stg/opencl/drivers/opencl/runtime/platform/command.cpp#94 edit
... //depot/stg/opencl/drivers/opencl/runtime/platform/command.hpp#92 edit
... //depot/stg/opencl/drivers/opencl/runtime/utils/flags.hpp#311 edit
Este commit está contenido en:
foreman
2019-06-12 10:00:38 -04:00
padre 21b1c1a183
commit 17943639e4
Se han modificado 6 ficheros con 155 adiciones y 9 borrados
Descargar archivo de parche Descargar archivo de diferencias Expandir todos los archivos Contraer todos los archivos
api/hip/hip_platform.cpp
+77
Ver fichero
@@ -448,6 +448,69 @@ hipError_t ihipCreateGlobalVarObj(const char* name, hipModule_t hmod, amd::Memor
}
namespace hip_impl {
hipError_t ihipOccupancyMaxActiveBlocksPerMultiprocessor(int* numBlocks,
const void* f,
int blockSize,
size_t dynamicSMemSize)
{
HIP_INIT_API(f, blockSize, dynamicSMemSize);
int deviceId = ihipGetDevice();
hipFunction_t func = PlatformState::instance().getFunc(f, deviceId);
if (func == nullptr) {
HIP_RETURN(hipErrorUnknown);
}
hip::Function* function = hip::Function::asFunction(func);
amd::Kernel* kernel = function->function_;
if (!kernel) {
HIP_RETURN(hipErrorOutOfMemory);
}
if (blockSize == 0) {
HIP_RETURN(hipErrorInvalidValue);
}
amd::Device* device = hip::getCurrentContext()->devices()[0];
const device::Kernel::WorkGroupInfo* wrkGrpInfo = kernel->getDeviceKernel(*device)->workGroupInfo();
// Find threads accupancy per CU => simd_per_cu * GPR usage
constexpr size_t MaxWavesPerSimd = 8; // Limited by SPI 32 per CU, hence 8 per SIMD
size_t alu_accupancy = device->info().simdPerCU_ *
std::min(MaxWavesPerSimd, (wrkGrpInfo->availableVGPRs_ / amd::alignUp(wrkGrpInfo->usedVGPRs_, 4)));
alu_accupancy *= wrkGrpInfo->wavefrontSize_;
// Calculate blocks occupancy per CU
*numBlocks = alu_accupancy / amd::alignUp(blockSize, wrkGrpInfo->wavefrontSize_);
size_t total_used_lds = wrkGrpInfo->usedLDSSize_ + dynamicSMemSize;
if (total_used_lds != 0) {
// Calculate LDS occupacy per CU. lds_per_cu / (static_lsd + dynamic_lds)
int lds_occupancy = static_cast<int>(device->info().localMemSize_ / total_used_lds);
*numBlocks = std::min(*numBlocks, lds_occupancy);
}
HIP_RETURN(hipSuccess);
}
}
hipError_t hipOccupancyMaxActiveBlocksPerMultiprocessor(int* numBlocks,
const void* f,
int blockSize,
size_t dynamicSMemSize)
{
HIP_RETURN(hip_impl::ihipOccupancyMaxActiveBlocksPerMultiprocessor(numBlocks, f, blockSize, dynamicSMemSize));
}
hipError_t hipOccupancyMaxActiveBlocksPerMultiprocessorWithFlags(int* numBlocks,
const void* f,
int blockSize,
size_t dynamicSMemSize,
unsigned int flags)
{
HIP_RETURN(hip_impl::ihipOccupancyMaxActiveBlocksPerMultiprocessor(numBlocks, f, blockSize, dynamicSMemSize));
}
#if defined(ATI_OS_LINUX)
namespace hip_impl {
@@ -668,6 +731,20 @@ void hipLaunchKernelGGLImpl(
sharedMemBytes, stream, nullptr, kernarg);
}
void hipLaunchCooperativeKernelGGLImpl(
uintptr_t function_address,
const dim3& numBlocks,
const dim3& dimBlocks,
uint32_t sharedMemBytes,
hipStream_t stream,
void** kernarg)
{
HIP_INIT();
hipLaunchCooperativeKernel(reinterpret_cast<void*>(function_address),
numBlocks, dimBlocks, kernarg, sharedMemBytes, stream);
}
}
// conversion routines between float and half precision