[
Date Prev][
Date Next][
Thread Prev][
Thread Next][
Date Index][
Thread Index]
[vsipl++] [patch] matrix transpose products using CML
- To: VSIPL++ Developers List <vsipl++@xxxxxxxxxxxxxxxx>
- Subject: [vsipl++] [patch] matrix transpose products using CML
- From: Don McCoy <don@xxxxxxxxxxxxxxxx>
- Date: Tue, 20 May 2008 11:55:14 -0600
As attached. Note the expanded test coverage using dimension
orderings. This is due to the fact that for row-major matrices 'r' and 'a'
r = prodt(a, b) // if 'b' is row-major
is equivalent to
r = prod(a, b) // if 'b' is col-major
Ok to commit?
--
Don McCoy
don (at) CodeSourcery
(888) 776-0262 / (650) 331-3385, x712
2008-05-20 Don McCoy <don@xxxxxxxxxxxxxxxx>
* src/vsip/core/matvec_prod.hpp: Fixed bug in size passed for transpose
product case.
* src/vsip/opt/cbe/ppu/alf.hpp: Removed conditionals around ALF
initialization -- now uses cml_init/fini exclusively.
* src/vsip/opt/cbe/cml/matvec.hpp: Modified evaluator to handle
transpose cases.
* src/vsip/opt/cbe/cml/prod.hpp: Added bindings for transpose cases.
* tests/matvec-prod.cpp: Modified transpose product tests to run
them with different orderings, as is done with normal products.
Only added orderings relevant to CML so as not to expand the
test coverage unnecessarily. Also fixed the test to make it call
prodt() directly instead of prod(a, trans(b)).
Index: src/vsip/core/matvec_prod.hpp
===================================================================
--- src/vsip/core/matvec_prod.hpp (revision 208077)
+++ src/vsip/core/matvec_prod.hpp (working copy)
@@ -446,7 +446,7 @@
{
typedef typename Promotion<T0, T1>::type return_type;
- Matrix<return_type> r(m0.size(0), m1.size(1));
+ Matrix<return_type> r(m0.size(0), m1.size(0));
impl::generic_prod(m0, trans(m1), r);
Index: src/vsip/opt/cbe/ppu/alf.hpp
===================================================================
--- src/vsip/opt/cbe/ppu/alf.hpp (revision 208077)
+++ src/vsip/opt/cbe/ppu/alf.hpp (working copy)
@@ -19,10 +19,7 @@
# error "vsip/opt files cannot be used as part of the reference impl."
#endif
-#ifdef VSIP_IMPL_HAVE_CML
#include <cml/ppu/cml.h>
-#endif
-
#include <vsip/core/metaprogramming.hpp>
#include <vsip/support.hpp>
#include "alf.h"
@@ -149,7 +146,6 @@
ALF(unsigned int num_accelerators)
: num_accelerators_(0)
{
-#ifdef VSIP_IMPL_HAVE_CML
int argc = 3;
char* argv[3];
char number[256];
@@ -160,20 +156,10 @@
cml_init_argv(&argc, argv);
alf_ = cml_impl_alf_handle();
num_accelerators_ = cml_impl_alf_num_spes();
-#else
- int status = alf_init(0, &alf_);
- assert(status >= 0);
-
- set_num_accelerators(num_accelerators);
-#endif
}
~ALF()
{
-#ifdef VSIP_IMPL_HAVE_CML
cml_fini();
-#else
- alf_exit(&alf_, ALF_EXIT_POLICY_WAIT, -1);
-#endif
}
void set_num_accelerators(unsigned int n)
{
Index: src/vsip/opt/cbe/cml/matvec.hpp
===================================================================
--- src/vsip/opt/cbe/cml/matvec.hpp (revision 208077)
+++ src/vsip/opt/cbe/cml/matvec.hpp (working copy)
@@ -61,10 +61,9 @@
// check that the complex layouts are equal
Is_split_block<Block0>::value == Is_split_block<Block1>::value &&
Is_split_block<Block0>::value == Is_split_block<Block2>::value &&
- // check that the layout is row-major
+ // check that the layout is row-major for the first input and the output
Type_equal<order0_type, row2_type>::value &&
Type_equal<order1_type, row2_type>::value &&
- Type_equal<order2_type, row2_type>::value &&
// check that direct access is supported
Ext_data_cost<Block0>::value == 0 &&
Ext_data_cost<Block1>::value == 0 &&
@@ -76,28 +75,49 @@
Ext_data<Block1> ext_a(const_cast<Block1&>(a));
Ext_data<Block2> ext_b(const_cast<Block2&>(b));
+ // For 'b', the dimension with the smallest stride must be one,
+ // which depends on whether it is row- or column-major.
+ bool is_b_row = Type_equal<order2_type, row2_type>::value;
+ stride_type b_stride = is_b_row ? ext_b.stride(1) : ext_b.stride(0);
+
return
// ensure the data is unit-stide
( ext_r.stride(1) == 1 &&
ext_a.stride(1) == 1 &&
- ext_b.stride(1) == 1 );
+ b_stride == 1 );
}
static void exec(Block0& r, Block1 const& a, Block2 const& b)
{
- typedef typename Block_layout<Block0>::complex_type complex_type;
-
Ext_data<Block0> ext_r(const_cast<Block0&>(r));
Ext_data<Block1> ext_a(const_cast<Block1&>(a));
Ext_data<Block2> ext_b(const_cast<Block2&>(b));
- cml::mprod(
- ext_a.data(), ext_a.stride(0),
- ext_b.data(), ext_b.stride(0),
- ext_r.data(), ext_r.stride(0),
- a.size(2, 0), // M
- a.size(2, 1), // N
- b.size(2, 1) ); // P
+ // Either row- or column-major layouts are supported for
+ // the second input by mapping them to the normal product
+ // or transpose product respectively.
+ if (Type_equal<order2_type, row2_type>::value)
+ {
+ cml::mprod(
+ ext_a.data(), ext_a.stride(0),
+ ext_b.data(), ext_b.stride(0),
+ ext_r.data(), ext_r.stride(0),
+ a.size(2, 0), // M
+ a.size(2, 1), // N
+ b.size(2, 1) ); // P
+ }
+ else if (Type_equal<order2_type, col2_type>::value)
+ {
+ cml::mprodt(
+ ext_a.data(), ext_a.stride(0),
+ ext_b.data(), ext_b.stride(1),
+ ext_r.data(), ext_r.stride(0),
+ a.size(2, 0), // M
+ a.size(2, 1), // N
+ b.size(2, 1) ); // P
+ }
+ else
+ assert(0);
}
};
Index: src/vsip/opt/cbe/cml/prod.hpp
===================================================================
--- src/vsip/opt/cbe/cml/prod.hpp (revision 208077)
+++ src/vsip/opt/cbe/cml/prod.hpp (working copy)
@@ -46,55 +46,58 @@
// This macro supports scalar and interleaved complex types
-#define VSIP_IMPL_CML_MPROD(T, FCN) \
-inline void \
-mprod( \
- T *a, int lda, \
- T *b, int ldb, \
- T *z, int ldz, \
- int m, int n, int p) \
-{ \
- typedef Scalar_of<T>::type CML_T; \
- FCN( \
- reinterpret_cast<CML_T*>(a), \
- static_cast<ptrdiff_t>(lda), \
- reinterpret_cast<CML_T*>(b), \
- static_cast<ptrdiff_t>(ldb), \
- reinterpret_cast<CML_T*>(z), \
- static_cast<ptrdiff_t>(ldz), \
- static_cast<size_t>(m), \
- static_cast<size_t>(n), \
- static_cast<size_t>(p) ); \
-}
+#define VSIP_IMPL_CML_MPROD(T, FCN, CML_FCN) \
+ inline void \
+ FCN( \
+ T *a, int lda, \
+ T *b, int ldb, \
+ T *z, int ldz, \
+ int m, int n, int p) \
+ { \
+ typedef Scalar_of<T>::type CML_T; \
+ CML_FCN( \
+ reinterpret_cast<CML_T*>(a), \
+ static_cast<ptrdiff_t>(lda), \
+ reinterpret_cast<CML_T*>(b), \
+ static_cast<ptrdiff_t>(ldb), \
+ reinterpret_cast<CML_T*>(z), \
+ static_cast<ptrdiff_t>(ldz), \
+ static_cast<size_t>(m), \
+ static_cast<size_t>(n), \
+ static_cast<size_t>(p) ); \
+ }
-VSIP_IMPL_CML_MPROD(float, cml_mprod1_f)
-VSIP_IMPL_CML_MPROD(std::complex<float>, cml_cmprod1_f)
+VSIP_IMPL_CML_MPROD(float, mprod, cml_mprod1_f)
+VSIP_IMPL_CML_MPROD(float, mprodt, cml_mprodt1_f)
+VSIP_IMPL_CML_MPROD(std::complex<float>, mprod, cml_cmprod1_f)
+VSIP_IMPL_CML_MPROD(std::complex<float>, mprodt, cml_cmprodt1_f)
#undef VSIP_IMPL_CML_MPROD
// This version is for split complex only.
-#define VSIP_IMPL_CML_ZMPROD(T, FCN) \
-inline void \
-mprod( \
- std::pair<T*, T*> a, int lda, \
- std::pair<T*, T*> b, int ldb, \
- std::pair<T*, T*> z, int ldz, \
- int m, int n, int p) \
-{ \
- FCN( \
- a.first, a.second, \
- static_cast<ptrdiff_t>(lda), \
- b.first, b.second, \
- static_cast<ptrdiff_t>(ldb), \
- z.first, z.second, \
- static_cast<ptrdiff_t>(ldz), \
- static_cast<size_t>(m), \
- static_cast<size_t>(n), \
- static_cast<size_t>(p) ); \
-}
+#define VSIP_IMPL_CML_ZMPROD(T, FCN, CML_FCN) \
+ inline void \
+ FCN( \
+ std::pair<T*, T*> a, int lda, \
+ std::pair<T*, T*> b, int ldb, \
+ std::pair<T*, T*> z, int ldz, \
+ int m, int n, int p) \
+ { \
+ CML_FCN( \
+ a.first, a.second, \
+ static_cast<ptrdiff_t>(lda), \
+ b.first, b.second, \
+ static_cast<ptrdiff_t>(ldb), \
+ z.first, z.second, \
+ static_cast<ptrdiff_t>(ldz), \
+ static_cast<size_t>(m), \
+ static_cast<size_t>(n), \
+ static_cast<size_t>(p) ); \
+ }
-VSIP_IMPL_CML_ZMPROD(float, cml_zmprod1_f)
+VSIP_IMPL_CML_ZMPROD(float, mprod, cml_zmprod1_f)
+VSIP_IMPL_CML_ZMPROD(float, mprodt, cml_zmprodt1_f)
#undef VSIP_IMPL_CML_ZMPROD
Index: tests/matvec-prod.cpp
===================================================================
--- tests/matvec-prod.cpp (revision 208077)
+++ tests/matvec-prod.cpp (working copy)
@@ -557,24 +557,31 @@
template <typename T0,
- typename T1>
+ typename T1,
+ typename OrderR,
+ typename Order0,
+ typename Order1>
void
test_prodt_rand(length_type m, length_type n, length_type k)
{
typedef typename Promotion<T0, T1>::type return_type;
typedef typename vsip::impl::Scalar_of<return_type>::type scalar_type;
- Matrix<T0> a(m, n);
- Matrix<T1> b(k, n);
- Matrix<return_type> res1(m, k);
- Matrix<return_type> chk(m, k);
+ typedef Dense<2, T0, Order0> block0_type;
+ typedef Dense<2, T1, Order1> block1_type;
+ typedef Dense<2, return_type, OrderR> blockR_type;
+
+ Matrix<T0, block0_type> a(m, n);
+ Matrix<T1, block1_type> b(k, n);
+ Matrix<return_type, blockR_type> res1(m, k);
+ Matrix<return_type, blockR_type> chk(m, k);
Matrix<scalar_type> gauge(m, k);
randm(a);
randm(b);
// Test matrix-matrix prod for transpose
- res1 = prod(a, trans(b));
+ res1 = prodt(a, b);
chk = ref::prod(a, trans(b));
gauge = ref::prod(mag(a), mag(trans(b)));
@@ -624,8 +631,33 @@
template <typename T0,
+ typename T1,
+ typename OrderR,
+ typename Order0,
+ typename Order1>
+void
+prodt_types_with_order()
+{
+ test_prodt_rand<T0, T1, OrderR, Order0, Order1>(5, 5, 5);
+ test_prodt_rand<T0, T1, OrderR, Order0, Order1>(5, 7, 9);
+ test_prodt_rand<T0, T1, OrderR, Order0, Order1>(9, 5, 7);
+ test_prodt_rand<T0, T1, OrderR, Order0, Order1>(9, 7, 5);
+}
+
+
+template <typename T0,
typename T1>
void
+prodt_cases_with_order()
+{
+ prodt_types_with_order<T0, T1, row2_type, row2_type, row2_type>();
+ prodt_types_with_order<T0, T1, row2_type, row2_type, col2_type>();
+}
+
+
+template <typename T0,
+ typename T1>
+void
prod_cases()
{
test_prod_mv<T0, T1>(5, 7);
@@ -633,12 +665,9 @@
test_prod3_rand<T0, T1>();
test_prod4_rand<T0, T1>();
- test_prodt_rand<T0, T1>(5, 5, 5);
- test_prodt_rand<T0, T1>(5, 7, 9);
- test_prodt_rand<T0, T1>(9, 5, 7);
- test_prodt_rand<T0, T1>(9, 7, 5);
prod_cases_with_order<T0, T1>();
+ prodt_cases_with_order<T0, T1>();
}
@@ -669,7 +698,6 @@
-
/***********************************************************************
Main
***********************************************************************/
