Skip to content

Gemm #50

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
imre-palik wants to merge 14 commits into
base: feature/ublas00004_simd_gemm
Choose a base branch
from
Open

Gemm #50

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
14 commits
Select commit Hold shift + click to select a range
5a8731e
Fix bug in weak equality check
ShangtongZhang Mar 13, 2016
b40a65f
Merge pull request #44 from ShangtongZhang/dev
nasosi Mar 16, 2016
986547a
Add norm_2_square
rdwampler Mar 28, 2016
e88fc29
Add documentation for norm_2_square
rdwampler Mar 28, 2016
08d2084
Merge pull request #47 from rdwampler/develop
nasosi Mar 31, 2016
ad76936
Fix compilation error when BOOST_UBLAS_SCALED_NORM is defined
rdwampler Apr 7, 2016
9d30e38
Merge pull request #49 from rdwampler/scaled-norm
nasosi Apr 11, 2016
5dc64f9
ublas: improved dense matrix multiplication performance
imre-palik Feb 11, 2016
b20cfb0
boost::ublas: gcc support for optimal matrix multiplication
imre-palik Feb 27, 2016
cff83d1
boost::ublas increasing the range of BLAS level 3 benchmarks
imre-palik Feb 28, 2016
93545c0
ublas::gemm adding support for clang
imre-palik Mar 6, 2016
8ceb6c0
ublas::gemm making possible to disable gemm as the default dense matr…
imre-palik Mar 13, 2016
752bc93
ublas:gemm kernels for complex multiplication
imre-palik Mar 19, 2016
4f4d251
ublas::gemm support for icc
imre-palik Apr 3, 2016
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion benchmarks/bench1/Jamfile.v2
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7,4 +7,4 @@

exe bench1
: bench1.cpp bench11.cpp bench12.cpp bench13.cpp
;
: : <optimization>speed ;
14 changes: 10 additions & 4 deletions benchmarks/bench1/bench1.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -76,22 +76,28 @@ void do_bench (std::string type_string, int scale)
header (type_string + ", 3");
bench_1<scalar, 3> () (1000000 * scale);
bench_2<scalar, 3> () (300000 * scale);
bench_3<scalar, 3> () (100000 * scale);
bench_3<scalar, 3> () (3000000 * scale);

header (type_string + ", 10");
bench_1<scalar, 10> () (300000 * scale);
bench_2<scalar, 10> () (30000 * scale);
bench_3<scalar, 10> () (3000 * scale);
bench_3<scalar, 10> () (100000 * scale);

header (type_string + ", 30");
bench_1<scalar, 30> () (100000 * scale);
bench_2<scalar, 30> () (3000 * scale);
bench_3<scalar, 30> () (100 * scale);
bench_3<scalar, 30> () (30000 * scale);

header (type_string + ", 100");
bench_1<scalar, 100> () (30000 * scale);
bench_2<scalar, 100> () (300 * scale);
bench_3<scalar, 100> () (3 * scale);
bench_3<scalar, 100> () (1000 * scale);

header (type_string + ", 300");
bench_3<scalar, 300> () (30 * scale);

header (type_string + ", 1000");
bench_3<scalar, 1000> () (1 * scale);
}

int main (int argc, char *argv []) {
Expand Down
8 changes: 8 additions & 0 deletions benchmarks/bench1/bench13.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -166,13 +166,17 @@ template struct bench_3<float, 3>;
template struct bench_3<float, 10>;
template struct bench_3<float, 30>;
template struct bench_3<float, 100>;
template struct bench_3<float, 300>;
template struct bench_3<float, 1000>;
#endif

#ifdef USE_DOUBLE
template struct bench_3<double, 3>;
template struct bench_3<double, 10>;
template struct bench_3<double, 30>;
template struct bench_3<double, 100>;
template struct bench_3<double, 300>;
template struct bench_3<double, 1000>;
#endif

#ifdef USE_STD_COMPLEX
Expand All @@ -181,12 +185,16 @@ template struct bench_3<std::complex<float>, 3>;
template struct bench_3<std::complex<float>, 10>;
template struct bench_3<std::complex<float>, 30>;
template struct bench_3<std::complex<float>, 100>;
template struct bench_3<std::complex<float>, 300>;
template struct bench_3<std::complex<float>, 1000>;
#endif

#ifdef USE_DOUBLE
template struct bench_3<std::complex<double>, 3>;
template struct bench_3<std::complex<double>, 10>;
template struct bench_3<std::complex<double>, 30>;
template struct bench_3<std::complex<double>, 100>;
template struct bench_3<std::complex<double>, 300>;
template struct bench_3<std::complex<double>, 1000>;
#endif
#endif
3 changes: 2 additions & 1 deletion benchmarks/bench3/Jamfile.v2
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7,4 +7,5 @@

exe bench3
: bench3.cpp bench31.cpp bench32.cpp bench33.cpp
;
: : <optimization>speed ;

14 changes: 10 additions & 4 deletions benchmarks/bench3/bench3.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -76,22 +76,28 @@ void do_bench (std::string type_string, int scale)
header (type_string + ", 3");
bench_1<scalar, 3> () (1000000 * scale);
bench_2<scalar, 3> () (300000 * scale);
bench_3<scalar, 3> () (100000 * scale);
bench_3<scalar, 3> () (3000000 * scale);

header (type_string + ", 10");
bench_1<scalar, 10> () (300000 * scale);
bench_2<scalar, 10> () (30000 * scale);
bench_3<scalar, 10> () (3000 * scale);
bench_3<scalar, 10> () (100000 * scale);

header (type_string + ", 30");
bench_1<scalar, 30> () (100000 * scale);
bench_2<scalar, 30> () (3000 * scale);
bench_3<scalar, 30> () (100 * scale);
bench_3<scalar, 30> () (30000 * scale);

header (type_string + ", 100");
bench_1<scalar, 100> () (30000 * scale);
bench_2<scalar, 100> () (300 * scale);
bench_3<scalar, 100> () (3 * scale);
bench_3<scalar, 100> () (1000 * scale);

header (type_string + ", 300");
bench_3<scalar, 300> () (30 * scale);

header (type_string + ", 1000");
bench_3<scalar, 1000> () (1 * scale);
}

int main (int argc, char *argv []) {
Expand Down
8 changes: 8 additions & 0 deletions benchmarks/bench3/bench33.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -172,13 +172,17 @@ template struct bench_3<float, 3>;
template struct bench_3<float, 10>;
template struct bench_3<float, 30>;
template struct bench_3<float, 100>;
template struct bench_3<float, 300>;
template struct bench_3<float, 1000>;
#endif

#ifdef USE_DOUBLE
template struct bench_3<double, 3>;
template struct bench_3<double, 10>;
template struct bench_3<double, 30>;
template struct bench_3<double, 100>;
template struct bench_3<double, 300>;
template struct bench_3<double, 1000>;
#endif

#ifdef USE_STD_COMPLEX
Expand All @@ -187,12 +191,16 @@ template struct bench_3<std::complex<float>, 3>;
template struct bench_3<std::complex<float>, 10>;
template struct bench_3<std::complex<float>, 30>;
template struct bench_3<std::complex<float>, 100>;
template struct bench_3<std::complex<float>, 300>;
template struct bench_3<std::complex<float>, 1000>;
#endif

#ifdef USE_DOUBLE
template struct bench_3<std::complex<double>, 3>;
template struct bench_3<std::complex<double>, 10>;
template struct bench_3<std::complex<double>, 30>;
template struct bench_3<std::complex<double>, 100>;
template struct bench_3<std::complex<double>, 300>;
template struct bench_3<std::complex<double>, 1000>;
#endif
#endif
3 changes: 2 additions & 1 deletion doc/operations_overview.html
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -87,7 +87,8 @@ <h3>norms</h3>

<pre><code>
t = norm_inf(v); i = index_norm_inf(v);
t = norm_1(v); t = norm_2(v);
t = norm_1(v); t = norm_2(v);
t = norm_2_square(v);
t = norm_inf(A); i = index_norm_inf(A);
t = norm_1(A); t = norm_frobenius(A);
</code></pre>
Expand Down
5 changes: 5 additions & 0 deletions doc/vector_expression.html
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -847,6 +847,11 @@ <h4>Prototypes</h4>
typename vector_scalar_unary_traits&lt;E, vector_norm_2&lt;typename E::value_type&gt; &gt;::result_type
norm_2 (const vector_expression&lt;E&gt; &amp;e);

// norm_2_square v = sum (v [i] * v [i])
template&lt;class E&gt;
typename vector_scalar_unary_traits&lt;E, vector_norm_2_square&lt;typename E::value_type&gt; &gt;::result_type
norm_2_square (const vector_expression&lt;E&gt; &amp;e);

// norm_inf v = max (abs (v [i]))
template&lt;class E&gt;
typename vector_scalar_unary_traits&lt;E, vector_norm_inf&lt;typename E::value_type&gt; &gt;::result_type
Expand Down
121 changes: 121 additions & 0 deletions include/boost/numeric/ublas/detail/block_sizes.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,121 @@
//
// Copyright (c) 2016
// Michael Lehn, Imre Palik
//
// 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)

#ifndef _BOOST_UBLAS_BLOCK_SIZES_
#define _BOOST_UBLAS_BLOCK_SIZES_

#include <boost/numeric/ublas/detail/vector.hpp>

namespace boost { namespace numeric { namespace ublas { namespace detail {

template <typename T>
struct prod_block_size {
static const unsigned vector_length = _BOOST_UBLAS_VECTOR_SIZE/sizeof(T); // Number of elements in a vector register
static const unsigned mc = 256;
static const unsigned kc = 512; // stripe length
static const unsigned nc = (4096/(3 * vector_length)) * (3 * vector_length);
static const unsigned mr = 4; // stripe width for lhs
static const unsigned nr = 3 * vector_length; // stripe width for rhs
static const unsigned align = 64; // align temporary arrays to this boundary
static const unsigned limit = 14; // Use gemm from this size
BOOST_STATIC_ASSERT_MSG(mc>0 && kc>0 && nc>0 && mr>0 && nr>0, "Invalid block size.");
BOOST_STATIC_ASSERT_MSG(mc % mr == 0, "MC must be a multiple of MR.");
BOOST_STATIC_ASSERT_MSG(nc % nr == 0, "NC must be a multiple of NR.");
};

template <>
struct prod_block_size<float> {
static const unsigned mc = 256;
static const unsigned kc = 512; // stripe length
static const unsigned nc = 4096;
static const unsigned mr = 4; // stripe width for lhs
static const unsigned nr = 16; // stripe width for rhs
static const unsigned align = 64; // align temporary arrays to this boundary
static const unsigned limit = 14; // Use gemm from this size
static const unsigned vector_length = _BOOST_UBLAS_VECTOR_SIZE/sizeof(float); // Number of elements in a vector register
BOOST_STATIC_ASSERT_MSG(mc>0 && kc>0 && nc>0 && mr>0 && nr>0, "Invalid block size.");
BOOST_STATIC_ASSERT_MSG(mc % mr == 0, "MC must be a multiple of MR.");
BOOST_STATIC_ASSERT_MSG(nc % nr == 0, "NC must be a multiple of NR.");
};

template <>
struct prod_block_size<long double> {
static const unsigned mc = 256;
static const unsigned kc = 512; // stripe length
static const unsigned nc = 4096;
static const unsigned mr = 1; // stripe width for lhs
static const unsigned nr = 4; // stripe width for rhs
static const unsigned align = 64; // align temporary arrays to this boundary
static const unsigned limit = 42; // Use gemm from this size
static const unsigned vector_length = 1; // Number of elements in a vector register
BOOST_STATIC_ASSERT_MSG(mc>0 && kc>0 && nc>0 && mr>0 && nr>0, "Invalid block size.");
BOOST_STATIC_ASSERT_MSG(mc % mr == 0, "MC must be a multiple of MR.");
BOOST_STATIC_ASSERT_MSG(nc % nr == 0, "NC must be a multiple of NR.");
};

template <typename T>
struct prod_block_size<std::complex<T> > {
static const unsigned vector_length = _BOOST_UBLAS_VECTOR_SIZE/sizeof(T); // Number of elements in a vector register
static const unsigned mc = 255;
static const unsigned kc = 512; // stripe length
static const unsigned nc = 4096;
static const unsigned mr = 3; // stripe width for lhs
static const unsigned nr = vector_length; // stripe width for rhs
static const unsigned align = 64; // align temporary arrays to this boundary
static const unsigned limit = 23; // Use gemm from this size
BOOST_STATIC_ASSERT_MSG(mc>0 && kc>0 && nc>0 && mr>0 && nr>0, "Invalid block size.");
BOOST_STATIC_ASSERT_MSG(mc % mr == 0, "MC must be a multiple of MR.");
BOOST_STATIC_ASSERT_MSG(nc % nr == 0, "NC must be a multiple of NR.");
};

template <>
struct prod_block_size<std::complex<long double> > {
static const unsigned mc = 256;
static const unsigned kc = 512; // stripe length
static const unsigned nc = 4096;
static const unsigned mr = 1; // stripe width for lhs
static const unsigned nr = 1; // stripe width for rhs
static const unsigned align = 64; // align temporary arrays to this boundary
static const unsigned limit = 68; // Use gemm from this size
static const unsigned vector_length = 1; // Number of elements in a vector register
BOOST_STATIC_ASSERT_MSG(mc>0 && kc>0 && nc>0 && mr>0 && nr>0, "Invalid block size.");
BOOST_STATIC_ASSERT_MSG(mc % mr == 0, "MC must be a multiple of MR.");
BOOST_STATIC_ASSERT_MSG(nc % nr == 0, "NC must be a multiple of NR.");
};

template<typename T>
struct is_blocksize {
struct fallback { static const int nr = 0; };
struct derived : T, fallback {};
template<int C1>
struct nonr {
static const bool value = false;
typedef false_type type;
};

template<typename C> static char (&f(typename nonr<C::nr>::type*))[1];
template<typename C> static char (&f(...))[2];

static bool const value = sizeof(f<derived>(0)) == 2;
};

template<typename T>
struct check_blocksize {
BOOST_STATIC_ASSERT_MSG(T::mc>0 && T::kc>0 && T::nc>0 && T::mr>0 && T::nr>0,
"Invalid block size.");
BOOST_STATIC_ASSERT_MSG(T::mc % T::mr == 0,
"MC must be a multiple of MR.");
BOOST_STATIC_ASSERT_MSG(T::nc % T::nr == 0,
"NC must be a multiple of NR.");
BOOST_STATIC_ASSERT_MSG(T::vector_length <= 1 || T::nr % T::vector_length == 0,
"NR must be a multiple of vector size");
BOOST_STATIC_ASSERT_MSG(T::limit >= 2,
"Minimum matrix size for gemm is 2*2");
};
}}}}
#endif
Loading