diff --git a/client_example/14_instance_id/CMakeLists.txt b/client_example/14_instance_id/CMakeLists.txt new file mode 100644 index 0000000000..87b2a9a0cb --- /dev/null +++ b/client_example/14_instance_id/CMakeLists.txt @@ -0,0 +1,2 @@ +add_executable(client_batchnorm_fwd_instance_id batchnorm_fwd_instance_id.cpp) +target_link_libraries(client_batchnorm_fwd_instance_id PRIVATE composable_kernel::device_operations) diff --git a/client_example/14_instance_id/batchnorm_fwd_instance_id.cpp b/client_example/14_instance_id/batchnorm_fwd_instance_id.cpp new file mode 100644 index 0000000000..9cfeee1cfe --- /dev/null +++ b/client_example/14_instance_id/batchnorm_fwd_instance_id.cpp @@ -0,0 +1,206 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include +#include +#include +#include +#include + +#include "ck/ck.hpp" +#include "ck/tensor_operation/gpu/device/device_reduce.hpp" +#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" + +#include "ck/library/tensor_operation_instance/gpu/batchnorm_forward.hpp" + +using XDataType = float; +using YDataType = float; +using AccDataType = float; +using ScaleDataType = AccDataType; +using BiasDataType = AccDataType; +using MeanVarDataType = AccDataType; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; + +constexpr int Rank = 4; +constexpr int NumBatchNormReduceDim = 3; + +const double epsilon = std::numeric_limits::epsilon(); +const double averageFactor = 0.1; + +struct SimpleDeviceMem +{ + SimpleDeviceMem() = delete; + + SimpleDeviceMem(std::size_t mem_size) : p_mem_{} + { + (void)hipMalloc(static_cast(&p_mem_), mem_size); + } + + void* GetDeviceBuffer() { return p_mem_; } + + ~SimpleDeviceMem() { (void)hipFree(p_mem_); } + + void* p_mem_; +}; + +// In the actual application, the instance index and name are usually from the perf db +static int instance_index = -1; +static std::string instance_name; + +int main(int argc, char* argv[]) +{ + std::array xyLengths{16, 8, 128, 256}; + std::array xyStrides{8 * 128 * 256, 128 * 256, 256, 1}; + std::array scaleBiasMeanVarLengths{256}; + std::array scaleBiasMeanVarStrides{1}; + std::array reduceDims{0, 1, 2}; + + ck::index_t numXYElement = + std::accumulate(xyLengths.begin(), xyLengths.end(), 1, std::multiplies()); + + ck::index_t numScaleBiasMeanVarElement = std::accumulate(scaleBiasMeanVarLengths.begin(), + scaleBiasMeanVarLengths.end(), + 1, + std::multiplies()); + + SimpleDeviceMem x(sizeof(XDataType) * numXYElement); + SimpleDeviceMem y(sizeof(YDataType) * numXYElement); + SimpleDeviceMem scale(sizeof(ScaleDataType) * numScaleBiasMeanVarElement); + SimpleDeviceMem bias(sizeof(BiasDataType) * numScaleBiasMeanVarElement); + SimpleDeviceMem mean(sizeof(MeanVarDataType) * numScaleBiasMeanVarElement); + SimpleDeviceMem invVariance(sizeof(MeanVarDataType) * numScaleBiasMeanVarElement); + + using DeviceOp = ck::tensor_operation::device::DeviceBatchNormFwd; + + const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< + DeviceOp>::GetInstances(); + + std::cout << "found " << op_ptrs.size() << " instances" << std::endl; + + bool found = false; + int best_op_index = -1; + float best_ave_time = std::numeric_limits::max(); + + // profile device operation instances and save the best performant instance index and instance + // name + std::cout << "Run all instances and do timing" << std::endl; + + for(int i = 0; i < op_ptrs.size(); ++i) + { + auto& op_ptr = op_ptrs[i]; + + auto argument_ptr = op_ptr->MakeArgumentPointer(xyLengths, + xyStrides, + xyStrides, + reduceDims, + scaleBiasMeanVarLengths, + scaleBiasMeanVarStrides, + scaleBiasMeanVarStrides, + scaleBiasMeanVarStrides, + x.GetDeviceBuffer(), + scale.GetDeviceBuffer(), + bias.GetDeviceBuffer(), + epsilon, + PassThrough{}, + y.GetDeviceBuffer(), + mean.GetDeviceBuffer(), + invVariance.GetDeviceBuffer(), + averageFactor, + nullptr, + nullptr); + + auto invoker_ptr = op_ptr->MakeInvokerPointer(); + + if(op_ptr->IsSupportedArgument(argument_ptr.get())) + { + size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get()); + + SimpleDeviceMem workspace(workspace_sz); + + op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer()); + + float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true}); + + if(ave_time < best_ave_time) + { + found = true; + best_op_index = i; + best_ave_time = ave_time; + } + } + } + + if(found) + { + instance_index = best_op_index; + instance_name = op_ptrs[instance_index]->GetTypeIdHashCode(); + }; + + // simulate the execution of the operation when the instance index and name are available + const auto op_ptrs_2 = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< + DeviceOp>::GetInstances(); + + if(instance_index >= 0 && instance_index < op_ptrs_2.size()) + { + auto& op_ptr = op_ptrs_2[instance_index]; + + if(op_ptr->GetTypeIdHashCode() == instance_name) + { + + auto argument_ptr = op_ptr->MakeArgumentPointer(xyLengths, + xyStrides, + xyStrides, + reduceDims, + scaleBiasMeanVarLengths, + scaleBiasMeanVarStrides, + scaleBiasMeanVarStrides, + scaleBiasMeanVarStrides, + x.GetDeviceBuffer(), + scale.GetDeviceBuffer(), + bias.GetDeviceBuffer(), + epsilon, + PassThrough{}, + y.GetDeviceBuffer(), + mean.GetDeviceBuffer(), + invVariance.GetDeviceBuffer(), + averageFactor, + nullptr, + nullptr); + + auto invoker_ptr = op_ptr->MakeInvokerPointer(); + + if(op_ptr->IsSupportedArgument(argument_ptr.get())) + { + size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get()); + + SimpleDeviceMem workspace(workspace_sz); + + op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer()); + + float exec_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true}); + + size_t num_bytes = numXYElement * (sizeof(XDataType) + sizeof(YDataType)) + + numScaleBiasMeanVarElement * + (sizeof(ScaleDataType) + sizeof(BiasDataType) + + sizeof(MeanVarDataType) + sizeof(MeanVarDataType)); + + float gb_per_sec = num_bytes / 1.E6 / exec_time; + + std::cout << "Kernel execution time: " << std::setw(10) << exec_time + << " ms, effective data transfer bandwidth: " << gb_per_sec << " GB/s" + << std::endl; + } + }; + } + + return 0; +} diff --git a/include/ck/tensor_operation/gpu/device/device_base.hpp b/include/ck/tensor_operation/gpu/device/device_base.hpp index 65906bd03c..8137098eac 100644 --- a/include/ck/tensor_operation/gpu/device/device_base.hpp +++ b/include/ck/tensor_operation/gpu/device/device_base.hpp @@ -5,6 +5,7 @@ #include #include +#include #include "ck/stream_config.hpp" @@ -46,6 +47,17 @@ struct BaseOperator virtual bool IsSupportedArgument(const BaseArgument*) { return false; } virtual std::string GetTypeString() const { return ""; } + virtual std::string GetTypeIdName() const { return typeid(*this).name(); } + + virtual std::string GetTypeIdHashCode() const + { + std::ostringstream oss; + + oss << std::hex << typeid(*this).hash_code(); + + return oss.str(); + }; + virtual size_t GetWorkSpaceSize(const BaseArgument*) const { return 0; } virtual void SetWorkSpacePointer(BaseArgument* p_arg, void* p_workspace) const