#include "timemory/ert/configuration.hpp"

Collaboration diagram for tim::ert::configuration< DeviceT, Tp, CounterT >:

Public Types
using	this_type = configuration< DeviceT, Tp, CounterT >

using	ert_data_t = exec_data< CounterT >

using	device_t = DeviceT

using	counter_t = CounterT

using	ert_counter_t = counter< device_t, Tp, counter_t >

using	ert_data_ptr_t = std::shared_ptr< ert_data_t >

using	executor_func_t = std::function< ert_counter_t(ert_data_ptr_t)>

using	get_uint64_t = std::function< uint64_t()>

using	skip_ops_t = std::unordered_set< size_t >

using	get_skip_ops_t = std::function< skip_ops_t()>

Static Public Member Functions
static get_uint64_t &	get_num_threads ()

static get_uint64_t &	get_num_streams ()

static get_uint64_t &	get_grid_size ()

static get_uint64_t &	get_block_size ()

static get_uint64_t &	get_alignment ()

static get_uint64_t &	get_min_working_size ()

static get_uint64_t &	get_max_data_size ()

static get_skip_ops_t &	get_skip_ops ()

template<typename Dev = DeviceT, enable_if_t< std::is_same< Dev, device::cpu >::value, int > = 0>
static void	configure (uint64_t nthreads, uint64_t alignment=sizeof(Tp), uint64_t nstreams=0, uint64_t block_size=0, uint64_t grid_size=0)
	configure the number of threads, number of streams, block size, grid size, and alignment More...

template<typename Dev = DeviceT, enable_if_t< std::is_same< Dev, device::gpu >::value, int > = 0>
static void	configure (uint64_t nthreads, uint64_t alignment=sizeof(Tp), uint64_t nstreams=1, uint64_t block_size=1024, uint64_t grid_size=0)
	configure the number of threads, number of streams, block size, grid size, and alignment More...

static executor_func_t &	get_executor ()

Detailed Description

template<typename DeviceT, typename Tp, typename CounterT>
struct tim::ert::configuration< DeviceT, Tp, CounterT >

Definition at line 62 of file configuration.hpp.

Member Typedef Documentation

◆ counter_t

template<typename DeviceT , typename Tp , typename CounterT >

using tim::ert::configuration< DeviceT, Tp, CounterT >::counter_t = CounterT

Definition at line 67 of file configuration.hpp.

◆ device_t

template<typename DeviceT , typename Tp , typename CounterT >

using tim::ert::configuration< DeviceT, Tp, CounterT >::device_t = DeviceT

Definition at line 66 of file configuration.hpp.

◆ ert_counter_t

template<typename DeviceT , typename Tp , typename CounterT >

using tim::ert::configuration< DeviceT, Tp, CounterT >::ert_counter_t = counter<device_t, Tp, counter_t>

Definition at line 68 of file configuration.hpp.

◆ ert_data_ptr_t

template<typename DeviceT , typename Tp , typename CounterT >

using tim::ert::configuration< DeviceT, Tp, CounterT >::ert_data_ptr_t = std::shared_ptr<ert_data_t>

Definition at line 69 of file configuration.hpp.

◆ ert_data_t

template<typename DeviceT , typename Tp , typename CounterT >

using tim::ert::configuration< DeviceT, Tp, CounterT >::ert_data_t = exec_data<CounterT>

Definition at line 65 of file configuration.hpp.

◆ executor_func_t

template<typename DeviceT , typename Tp , typename CounterT >

using tim::ert::configuration< DeviceT, Tp, CounterT >::executor_func_t = std::function<ert_counter_t(ert_data_ptr_t)>

Definition at line 70 of file configuration.hpp.

◆ get_skip_ops_t

template<typename DeviceT , typename Tp , typename CounterT >

using tim::ert::configuration< DeviceT, Tp, CounterT >::get_skip_ops_t = std::function<skip_ops_t()>

Definition at line 73 of file configuration.hpp.

◆ get_uint64_t

template<typename DeviceT , typename Tp , typename CounterT >

using tim::ert::configuration< DeviceT, Tp, CounterT >::get_uint64_t = std::function<uint64_t()>

Definition at line 71 of file configuration.hpp.

◆ skip_ops_t

template<typename DeviceT , typename Tp , typename CounterT >

using tim::ert::configuration< DeviceT, Tp, CounterT >::skip_ops_t = std::unordered_set<size_t>

Definition at line 72 of file configuration.hpp.

◆ this_type

template<typename DeviceT , typename Tp , typename CounterT >

using tim::ert::configuration< DeviceT, Tp, CounterT >::this_type = configuration<DeviceT, Tp, CounterT>

Definition at line 64 of file configuration.hpp.

Member Function Documentation

◆ configure() [1/2]

template<typename DeviceT , typename Tp , typename CounterT >

template<typename Dev = DeviceT, enable_if_t< std::is_same< Dev, device::cpu >::value, int > = 0>

static void tim::ert::configuration< DeviceT, Tp, CounterT >::configure	(	uint64_t	nthreads,
		uint64_t	alignment = `sizeof(Tp)`,
		uint64_t	nstreams = `0`,
		uint64_t	block_size = `0`,
		uint64_t	grid_size = `0`
	)

inlinestatic

configure the number of threads, number of streams, block size, grid size, and alignment

Definition at line 180 of file configuration.hpp.

    {
        get_num_threads() = [=]() -> uint64_t { return nthreads; };
        get_num_streams() = [=]() -> uint64_t { return nstreams; };
        get_grid_size()   = [=]() -> uint64_t { return grid_size; };
        get_block_size()  = [=]() -> uint64_t { return block_size; };
        get_alignment()   = [=]() -> uint64_t { return alignment; };
    }

◆ configure() [2/2]

template<typename DeviceT , typename Tp , typename CounterT >

template<typename Dev = DeviceT, enable_if_t< std::is_same< Dev, device::gpu >::value, int > = 0>

static void tim::ert::configuration< DeviceT, Tp, CounterT >::configure	(	uint64_t	nthreads,
		uint64_t	alignment = `sizeof(Tp)`,
		uint64_t	nstreams = `1`,
		uint64_t	block_size = `1024`,
		uint64_t	grid_size = `0`
	)

inlinestatic

configure the number of threads, number of streams, block size, grid size, and alignment

Definition at line 196 of file configuration.hpp.

    {
        get_num_threads() = [=]() -> uint64_t { return nthreads; };
        get_num_streams() = [=]() -> uint64_t { return nstreams; };
        get_grid_size()   = [=]() -> uint64_t { return grid_size; };
        get_block_size()  = [=]() -> uint64_t { return block_size; };
        get_alignment()   = [=]() -> uint64_t { return alignment; };
    }

◆ get_alignment()

template<typename DeviceT , typename Tp , typename CounterT >

static get_uint64_t & tim::ert::configuration< DeviceT, Tp, CounterT >::get_alignment ( )

inlinestatic

Definition at line 116 of file configuration.hpp.

    {
        static get_uint64_t _instance = []() {
            return std::max<uint64_t>(settings::ert_alignment(), 8 * sizeof(Tp));
        };
        return _instance;
    }

References tim::ert_alignment.

Referenced by tim::ert::configuration< DeviceT, Tp, CounterT >::configure(), and tim::ert::configuration< DeviceT, Tp, CounterT >::get_executor().

◆ get_block_size()

template<typename DeviceT , typename Tp , typename CounterT >

static get_uint64_t & tim::ert::configuration< DeviceT, Tp, CounterT >::get_block_size ( )

inlinestatic

Definition at line 108 of file configuration.hpp.

    {
        static get_uint64_t _instance = []() { return settings::ert_block_size(); };
        return _instance;
    }

References tim::ert_block_size.

Referenced by tim::ert::configuration< DeviceT, Tp, CounterT >::configure(), and tim::ert::configuration< DeviceT, Tp, CounterT >::get_executor().

◆ get_executor()

template<typename DeviceT , typename Tp , typename CounterT >

static executor_func_t & tim::ert::configuration< DeviceT, Tp, CounterT >::get_executor ( )

inlinestatic

Definition at line 209 of file configuration.hpp.

    {
        static executor_func_t _instance = [](ert_data_ptr_t data) {
            using lli = long long int;
            // configuration sizes
            auto _mws_size   = get_min_working_size()();
            auto _max_size   = get_max_data_size()();
            auto _num_thread = get_num_threads()();
            auto _num_stream = get_num_streams()();
            auto _grid_size  = get_grid_size()();
            auto _block_size = get_block_size()();
            auto _align_size = get_alignment()();
            auto _skip_ops   = get_skip_ops()();
 
            // execution parameters
            exec_params params(_mws_size, _max_size, _num_thread, _num_stream, _grid_size,
                               _block_size);
            // operation _counter instance
            ert_counter_t _counter(params, data, _align_size);
 
            // set bytes per element
            _counter.bytes_per_element = sizeof(Tp);
            // set number of memory accesses per element from two functions
            _counter.memory_accesses_per_element = 2;
 
            for(const auto& itr : _skip_ops)
                _counter.add_skip_ops(itr);
 
            auto dtype = demangle(typeid(Tp).name());
 
            std::string _dev_name{};
            if(std::is_same<DeviceT, device::cpu>::value)
                _dev_name = "[device::cpu]";
            else if(std::is_same<DeviceT, device::gpu>::value)
                _dev_name = "[device::gpu]";
 
            if(settings::debug() || settings::verbose() > 0)
            {
                printf("[ert::executor]%s> "
                       "working-set = %lli, max-size = %lli, num-thread = %lli, "
                       "num-stream = "
                       "%lli, grid-size = %lli, block-size = %lli, align-size = %lli, "
                       "data-type "
                       "= %s\n",
                       _dev_name.c_str(), (lli) _mws_size, (lli) _max_size,
                       (lli) _num_thread, (lli) _num_stream, (lli) _grid_size,
                       (lli) _block_size, (lli) _align_size, dtype.c_str());
            }
 
            return _counter;
        };
        return _instance;
    }

◆ get_grid_size()

template<typename DeviceT , typename Tp , typename CounterT >

static get_uint64_t & tim::ert::configuration< DeviceT, Tp, CounterT >::get_grid_size ( )

inlinestatic

Definition at line 100 of file configuration.hpp.

    {
        static get_uint64_t _instance = []() { return settings::ert_grid_size(); };
        return _instance;
    }

References tim::ert_grid_size.

Referenced by tim::ert::configuration< DeviceT, Tp, CounterT >::configure(), and tim::ert::configuration< DeviceT, Tp, CounterT >::get_executor().

◆ get_max_data_size()

template<typename DeviceT , typename Tp , typename CounterT >

static get_uint64_t & tim::ert::configuration< DeviceT, Tp, CounterT >::get_max_data_size ( )

inlinestatic

Definition at line 140 of file configuration.hpp.

    {
        static get_uint64_t _instance = []() -> uint64_t {
            if(settings::ert_max_data_size() > 0)
                return settings::ert_max_data_size();
            static constexpr bool is_gpu = device::is_gpu<DeviceT>::value;
            if(is_gpu)
            {
                return settings::ert_max_data_size_gpu();
            }
            {
                return 2 * ert::cache_size::get_max();
            }
        };
        return _instance;
    }

References tim::ert_max_data_size, and tim::ert_max_data_size_gpu.

Referenced by tim::ert::configuration< DeviceT, Tp, CounterT >::get_executor().

◆ get_min_working_size()

template<typename DeviceT , typename Tp , typename CounterT >

static get_uint64_t & tim::ert::configuration< DeviceT, Tp, CounterT >::get_min_working_size ( )

inlinestatic

Definition at line 126 of file configuration.hpp.

    {
        static get_uint64_t _instance = []() {
            if(settings::ert_min_working_size() > 0)
                return settings::ert_min_working_size();
            static constexpr bool is_gpu = device::is_gpu<DeviceT>::value;
            return (is_gpu) ? settings::ert_min_working_size_gpu()
                            : settings::ert_min_working_size_cpu();
        };
        return _instance;
    }

References tim::ert_min_working_size, tim::ert_min_working_size_cpu, and tim::ert_min_working_size_gpu.

Referenced by tim::ert::configuration< DeviceT, Tp, CounterT >::get_executor().

◆ get_num_streams()

template<typename DeviceT , typename Tp , typename CounterT >

static get_uint64_t & tim::ert::configuration< DeviceT, Tp, CounterT >::get_num_streams ( )

inlinestatic

Definition at line 92 of file configuration.hpp.

    {
        static get_uint64_t _instance = []() { return settings::ert_num_streams(); };
        return _instance;
    }

References tim::ert_num_streams.

Referenced by tim::ert::configuration< DeviceT, Tp, CounterT >::configure(), and tim::ert::configuration< DeviceT, Tp, CounterT >::get_executor().

◆ get_num_threads()

template<typename DeviceT , typename Tp , typename CounterT >

static get_uint64_t & tim::ert::configuration< DeviceT, Tp, CounterT >::get_num_threads ( )

inlinestatic

Definition at line 77 of file configuration.hpp.

    {
        static get_uint64_t _instance = []() {
            if(settings::ert_num_threads() > 0)
                return settings::ert_num_threads();
            // for checking if gpu
            static constexpr bool is_gpu = device::is_gpu<DeviceT>::value;
            return (is_gpu) ? settings::ert_num_threads_gpu()
                            : settings::ert_num_threads_cpu();
        };
        return _instance;
    }

References tim::ert_num_threads, tim::ert_num_threads_cpu, and tim::ert_num_threads_gpu.

Referenced by tim::ert::configuration< DeviceT, Tp, CounterT >::configure(), and tim::ert::configuration< DeviceT, Tp, CounterT >::get_executor().

◆ get_skip_ops()

template<typename DeviceT , typename Tp , typename CounterT >

static get_skip_ops_t & tim::ert::configuration< DeviceT, Tp, CounterT >::get_skip_ops ( )

inlinestatic

Definition at line 159 of file configuration.hpp.

    {
        static get_skip_ops_t _instance = []() {
            auto       _skipstr    = settings::ert_skip_ops();
            auto       _skipstrvec = delimit(_skipstr, ",; \t");
            skip_ops_t _result;
            for(const auto& itr : _skipstrvec)
            {
                if(itr.find_first_not_of("0123456789") == std::string::npos)
                    _result.insert(atol(itr.c_str()));
            }
            return _result;
        };
        return _instance;
    }

References tim::delimit(), and tim::ert_skip_ops.

Referenced by tim::ert::configuration< DeviceT, Tp, CounterT >::get_executor().

Member Data Documentation

◆ alignment

template<typename DeviceT , typename Tp , typename CounterT >

get_uint64_t tim::ert::configuration< DeviceT, Tp, CounterT >::alignment = this_type::get_alignment()

Definition at line 269 of file configuration.hpp.

Referenced by tim::ert::configuration< DeviceT, Tp, CounterT >::configure().

◆ block_size

template<typename DeviceT , typename Tp , typename CounterT >

get_uint64_t tim::ert::configuration< DeviceT, Tp, CounterT >::block_size = this_type::get_block_size()

Definition at line 271 of file configuration.hpp.

Referenced by tim::ert::configuration< DeviceT, Tp, CounterT >::configure().

◆ executor

template<typename DeviceT , typename Tp , typename CounterT >

executor_func_t tim::ert::configuration< DeviceT, Tp, CounterT >::executor = this_type::get_executor()

Definition at line 272 of file configuration.hpp.

◆ grid_size

template<typename DeviceT , typename Tp , typename CounterT >

get_uint64_t tim::ert::configuration< DeviceT, Tp, CounterT >::grid_size = this_type::get_grid_size()

Definition at line 270 of file configuration.hpp.

Referenced by tim::ert::configuration< DeviceT, Tp, CounterT >::configure().

◆ max_data_size

template<typename DeviceT , typename Tp , typename CounterT >

get_uint64_t tim::ert::configuration< DeviceT, Tp, CounterT >::max_data_size = this_type::get_max_data_size()

Definition at line 268 of file configuration.hpp.

◆ min_working_size

template<typename DeviceT , typename Tp , typename CounterT >

get_uint64_t tim::ert::configuration< DeviceT, Tp, CounterT >::min_working_size = this_type::get_min_working_size()

Definition at line 267 of file configuration.hpp.

◆ num_streams

template<typename DeviceT , typename Tp , typename CounterT >

get_uint64_t tim::ert::configuration< DeviceT, Tp, CounterT >::num_streams = this_type::get_num_streams()

Definition at line 266 of file configuration.hpp.

◆ num_threads

template<typename DeviceT , typename Tp , typename CounterT >

get_uint64_t tim::ert::configuration< DeviceT, Tp, CounterT >::num_threads = this_type::get_num_threads()

Definition at line 265 of file configuration.hpp.

◆ verbose

template<typename DeviceT , typename Tp , typename CounterT >

bool tim::ert::configuration< DeviceT, Tp, CounterT >::verbose = false

Definition at line 264 of file configuration.hpp.

The documentation for this struct was generated from the following file:

timemory/ert/configuration.hpp

Public Attributes
bool	verbose = false

get_uint64_t	num_threads = this_type::get_num_threads()

get_uint64_t	num_streams = this_type::get_num_streams()

get_uint64_t	min_working_size = this_type::get_min_working_size()

get_uint64_t	max_data_size = this_type::get_max_data_size()

get_uint64_t	alignment = this_type::get_alignment()

get_uint64_t	grid_size = this_type::get_grid_size()

get_uint64_t	block_size = this_type::get_block_size()

executor_func_t	executor = this_type::get_executor()

Public Types

Static Public Member Functions

Public Attributes

Detailed Description

Member Typedef Documentation

◆ counter_t

◆ device_t

◆ ert_counter_t

◆ ert_data_ptr_t

◆ ert_data_t

◆ executor_func_t

◆ get_skip_ops_t

◆ get_uint64_t

◆ skip_ops_t

◆ this_type

Member Function Documentation

◆ configure() [1/2]

◆ configure() [2/2]

◆ get_alignment()

◆ get_block_size()

◆ get_executor()

◆ get_grid_size()

◆ get_max_data_size()

◆ get_min_working_size()

◆ get_num_streams()

◆ get_num_threads()

◆ get_skip_ops()

Member Data Documentation

◆ alignment

◆ block_size

◆ executor

◆ grid_size

◆ max_data_size

◆ min_working_size

◆ num_streams

◆ num_threads

◆ verbose