#882

Committed 31 Aug 2023 07:44PM UTC coverage: 41.798% (-44.7%) from 86.546%

Build # #882

Build Type

push

Committed by

Commit Message

Run Details

19442 of 46514 relevant lines covered (41.8%)

126375.38 hits per line

Source File
Press 'n' to go to next uncovered line, 'b' for previous

0.0

/examples/transpose/transpose_smp.cpp

//  Copyright (c) 2014 Thomas Heller
//
//  SPDX-License-Identifier: BSL-1.0
//  Distributed under the Boost Software License, Version 1.0. (See accompanying
//  file LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)

#include <hpx/algorithm.hpp>
#include <hpx/init.hpp>
#include <hpx/modules/iterator_support.hpp>
#include <hpx/numeric.hpp>

#include <algorithm>
#include <cstdint>
#include <exception>
#include <iostream>
#include <vector>

#define COL_SHIFT 1000.00    // Constant to shift column index
#define ROW_SHIFT 0.001      // Constant to shift row index

bool verbose = false;

double test_results(std::uint64_t order, std::vector<double> const& trans);

///////////////////////////////////////////////////////////////////////////////
int hpx_main(hpx::program_options::variables_map& vm)
{
    std::uint64_t order = vm["matrix_size"].as<std::uint64_t>();
    std::uint64_t iterations = vm["iterations"].as<std::uint64_t>();
    std::uint64_t tile_size = order;

    if (vm.count("tile_size"))
        tile_size = vm["tile_size"].as<std::uint64_t>();

    verbose = vm.count("verbose") ? true : false;

    std::uint64_t bytes =
        static_cast<std::uint64_t>(2 * sizeof(double) * order * order);

    std::vector<double> A(order * order);
    std::vector<double> B(order * order);

    std::cout << "Serial Matrix transpose: B = A^T\n"
              << "Matrix order          = " << order << "\n";
    if (tile_size < order)
        std::cout << "Tile size             = " << tile_size << "\n";
    else
        std::cout << "Untiled\n";
    std::cout << "Number of iterations  = " << iterations << "\n";

    using hpx::execution::par;
    using hpx::experimental::for_loop_strided;
    using hpx::ranges::for_each;

    std::uint64_t const start = 0;

    // Fill the original matrix, set transpose to known garbage value.
    auto range = hpx::util::counting_shape(order);
    // parallel for
    for_each(par, range, [&](std::uint64_t i) {
        for (std::uint64_t j = 0; j < order; ++j)
        {
            A[i * order + j] = COL_SHIFT * static_cast<double>(j) +
                ROW_SHIFT * static_cast<double>(i);
            B[i * order + j] = -1.0;
        }
    });

    double errsq = 0.0;
    double avgtime = 0.0;
    double maxtime = 0.0;
    double mintime =
        366.0 * 24.0 * 3600.0;    // set the minimum time to a large value;
                                  // one leap year should be enough
    for (std::uint64_t iter = 0; iter < iterations; ++iter)
    {
        hpx::chrono::high_resolution_timer t;
        if (tile_size < order)
        {
            // parallel for
            for_loop_strided(
                par, start, order + tile_size, tile_size, [&](std::uint64_t i) {
                    for (std::uint64_t j = 0; j < order; j += tile_size)
                    {
                        std::uint64_t i_max = (std::min) (order, i + tile_size);
                        std::uint64_t j_max = (std::min) (order, j + tile_size);

                        for (std::uint64_t it = i; it < i_max; ++it)
                        {
                            for (std::uint64_t jt = j; jt < j_max; ++jt)
                            {
                                B[it + order * jt] = A[jt + order * it];
                            }
                        }
                    }
                });
        }
        else
        {
            // parallel for
            auto range = hpx::util::counting_shape(order);
            for_each(par, range, [&](std::uint64_t i) {
                for (std::uint64_t j = 0; j < order; ++j)
                {
                    B[i + order * j] = A[j + order * i];
                }
            });
        }

        double elapsed = t.elapsed();

        if (iter > 0 || iterations == 1)    // Skip the first iteration
        {
            avgtime = avgtime + elapsed;
            maxtime = (std::max) (maxtime, elapsed);
            mintime = (std::min) (mintime, elapsed);
        }

        errsq += test_results(order, B);
    }    // end of iter loop

    // Analyze and output results

    double epsilon = 1.e-8;
    if (errsq < epsilon)
    {
        std::cout << "Solution validates\n";
        avgtime = avgtime /
            static_cast<double>(
                (std::max) (iterations - 1, static_cast<std::uint64_t>(1)));
        std::cout << "Rate (MB/s): "
                  << 1.e-6 * static_cast<double>(bytes) / mintime << ", "
                  << "Avg time (s): " << avgtime << ", "
                  << "Min time (s): " << mintime << ", "
                  << "Max time (s): " << maxtime << "\n";

        if (verbose)
            std::cout << "Squared errors: " << errsq << "\n";
    }
    else
    {
        std::cout << "ERROR: Aggregate squared error " << errsq
                  << " exceeds threshold " << epsilon << "\n";
        std::terminate();
    }

    return hpx::local::finalize();
}

int main(int argc, char* argv[])
{
    using namespace hpx::program_options;

    options_description desc_commandline;
    // clang-format off
    desc_commandline.add_options()
        ("matrix_size", value<std::uint64_t>()->default_value(1024),
         "Matrix Size")
        ("iterations", value<std::uint64_t>()->default_value(10),
         "# iterations")
        ("tile_size", value<std::uint64_t>(),
         "Number of tiles to divide the individual matrix blocks for improved "
         "cache and TLB performance")
        ( "verbose", "Verbose output")
    ;
    // clang-format on

    hpx::local::init_params init_args;
    init_args.desc_cmdline = desc_commandline;

    return hpx::local::init(hpx_main, argc, argv, init_args);
}

double test_results(std::uint64_t order, std::vector<double> const& trans)
{
    using hpx::transform_reduce;
    using hpx::execution::par;

    std::uint64_t const start = 0;

    auto range = hpx::util::counting_shape(start, order);
    // parallel reduce
    double errsq = transform_reduce(
        par, std::begin(range), std::end(range), 0.0,
        [](double lhs, double rhs) { return lhs + rhs; },
        [&](std::uint64_t i) -> double {
            double errsq = 0.0;
            for (std::uint64_t j = 0; j < order; ++j)
            {
                double diff = trans[i * order + j] -
                    (COL_SHIFT * static_cast<double>(i) +
                        ROW_SHIFT * static_cast<double>(j));
                errsq += diff * diff;
            }
            return errsq;
        });

    if (verbose)
        std::cout << " Squared sum of differences: " << errsq << "\n";

    return errsq;
}

1	// Copyright (c) 2014 Thomas Heller
2	//
3	// SPDX-License-Identifier: BSL-1.0
4	// Distributed under the Boost Software License, Version 1.0. (See accompanying
5	// file LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
6
7	#include <hpx/algorithm.hpp>
8	#include <hpx/init.hpp>
9	#include <hpx/modules/iterator_support.hpp>
10	#include <hpx/numeric.hpp>
11
12	#include <algorithm>
13	#include <cstdint>
14	#include <exception>
15	#include <iostream>
16	#include <vector>
17
18	#define COL_SHIFT 1000.00 // Constant to shift column index
19	#define ROW_SHIFT 0.001 // Constant to shift row index
20
21	bool verbose = false;
22
23	double test_results(std::uint64_t order, std::vector<double> const& trans);
24
25	///////////////////////////////////////////////////////////////////////////////
26	int hpx_main(hpx::program_options::variables_map& vm)	×
27	{
28	std::uint64_t order = vm["matrix_size"].as<std::uint64_t>();	×
29	std::uint64_t iterations = vm["iterations"].as<std::uint64_t>();	×
30	std::uint64_t tile_size = order;	×
31
32	if (vm.count("tile_size"))	×
33	tile_size = vm["tile_size"].as<std::uint64_t>();	×
34
35	verbose = vm.count("verbose") ? true : false;	×
36
37	std::uint64_t bytes =	×
38	static_cast<std::uint64_t>(2 * sizeof(double) * order * order);	×
39
40	std::vector<double> A(order * order);	×
41	std::vector<double> B(order * order);	×
42
43	std::cout << "Serial Matrix transpose: B = A^T\n"
44	<< "Matrix order = " << order << "\n";	×
45	if (tile_size < order)	×
46	std::cout << "Tile size = " << tile_size << "\n";	×
47	else
48	std::cout << "Untiled\n";	×
49	std::cout << "Number of iterations = " << iterations << "\n";	×
50
51	using hpx::execution::par;
52	using hpx::experimental::for_loop_strided;
53	using hpx::ranges::for_each;
54
55	std::uint64_t const start = 0;
56
57	// Fill the original matrix, set transpose to known garbage value.
58	auto range = hpx::util::counting_shape(order);	×
59	// parallel for
60	for_each(par, range, [&](std::uint64_t i) {	×
61	for (std::uint64_t j = 0; j < order; ++j)	×
62	{
63	A[i * order + j] = COL_SHIFT * static_cast<double>(j) +	×
64	ROW_SHIFT * static_cast<double>(i);	×
65	B[i * order + j] = -1.0;
66	}	×
67	});
68
69	double errsq = 0.0;
70	double avgtime = 0.0;	×
71	double maxtime = 0.0;	×
72	double mintime =
73	366.0 * 24.0 * 3600.0; // set the minimum time to a large value;
74	// one leap year should be enough	×
75	for (std::uint64_t iter = 0; iter < iterations; ++iter)
76	{
77	hpx::chrono::high_resolution_timer t;	×
78	if (tile_size < order)
79	{
80	// parallel for
81	for_loop_strided(	×
82	par, start, order + tile_size, tile_size, [&](std::uint64_t i) {	×
83	for (std::uint64_t j = 0; j < order; j += tile_size)
84	{	×
85	std::uint64_t i_max = (std::min) (order, i + tile_size);	×
86	std::uint64_t j_max = (std::min) (order, j + tile_size);
87		×
88	for (std::uint64_t it = i; it < i_max; ++it)
89	{	×
90	for (std::uint64_t jt = j; jt < j_max; ++jt)
91	{	×
92	B[it + order * jt] = A[jt + order * it];
93	}
94	}
95	}	×
96	});
97	}
98	else
99	{
100	// parallel for
101	auto range = hpx::util::counting_shape(order);	×
102	for_each(par, range, [&](std::uint64_t i) {	×
103	for (std::uint64_t j = 0; j < order; ++j)
104	{	×
105	B[i + order * j] = A[j + order * i];
106	}	×
107	});
108	}
109		×
110	double elapsed = t.elapsed();
111		×
112	if (iter > 0 \|\| iterations == 1) // Skip the first iteration
113	{	×
114	avgtime = avgtime + elapsed;	×
115	maxtime = (std::max) (maxtime, elapsed);	×
116	mintime = (std::min) (mintime, elapsed);
117	}
118		×
119	errsq += test_results(order, B);
120	} // end of iter loop
121
122	// Analyze and output results
123
124	double epsilon = 1.e-8;	×
125	if (errsq < epsilon)
126	{	×
127	std::cout << "Solution validates\n";	×
128	avgtime = avgtime /	×
129	static_cast<double>(	×
130	(std::max) (iterations - 1, static_cast<std::uint64_t>(1)));	×
131	std::cout << "Rate (MB/s): "
132	<< 1.e-6 * static_cast<double>(bytes) / mintime << ", "
133	<< "Avg time (s): " << avgtime << ", "	×
134	<< "Min time (s): " << mintime << ", "
135	<< "Max time (s): " << maxtime << "\n";	×
136		×
137	if (verbose)
138	std::cout << "Squared errors: " << errsq << "\n";
139	}
140	else
141	{	×
142	std::cout << "ERROR: Aggregate squared error " << errsq	×
143	<< " exceeds threshold " << epsilon << "\n";
144	std::terminate();
145	}	×
146
147	return hpx::local::finalize();
148	}	×
149
150	int main(int argc, char* argv[])
151	{
152	using namespace hpx::program_options;	×
153
154	options_description desc_commandline;	×
155	// clang-format off	×
156	desc_commandline.add_options()
157	("matrix_size", value<std::uint64_t>()->default_value(1024),	×
158	"Matrix Size")
159	("iterations", value<std::uint64_t>()->default_value(10),	×
160	"# iterations")
161	("tile_size", value<std::uint64_t>(),
162	"Number of tiles to divide the individual matrix blocks for improved "	×
163	"cache and TLB performance")
164	( "verbose", "Verbose output")
165	;
166	// clang-format on	×
167		×
168	hpx::local::init_params init_args;
169	init_args.desc_cmdline = desc_commandline;	×
170		×
171	return hpx::local::init(hpx_main, argc, argv, init_args);
172	}	×
173
174	double test_results(std::uint64_t order, std::vector<double> const& trans)
175	{
176	using hpx::transform_reduce;
177	using hpx::execution::par;
178
179	std::uint64_t const start = 0;	×
180
181	auto range = hpx::util::counting_shape(start, order);
182	// parallel reduce
183	double errsq = transform_reduce(	×
184	par, std::begin(range), std::end(range), 0.0,	×
185	[](double lhs, double rhs) { return lhs + rhs; },
186	[&](std::uint64_t i) -> double {	×
187	double errsq = 0.0;
188	for (std::uint64_t j = 0; j < order; ++j)
189	{	×
190	double diff = trans[i * order + j] -	×
191	(COL_SHIFT * static_cast<double>(i) +
192	ROW_SHIFT * static_cast<double>(j));	×
193	errsq += diff * diff;
194	}
195	return errsq;	×
196	});	×
197
198	if (verbose)	×
199	std::cout << " Squared sum of differences: " << errsq << "\n";
200
201	return errsq;
202	}

STEllAR-GROUP / hpx / #882

Source File Press 'n' to go to next uncovered line, 'b' for previous

Source File
Press 'n' to go to next uncovered line, 'b' for previous