Module: ppl/ppl Branch: floating_point Commit: d8d6e899f5cac4d00d37c1ebfbeeb199cbbbb1bb URL: http://www.cs.unipr.it/git/gitweb.cgi?p=ppl/ppl.git;a=commit;h=d8d6e899f5cac...

Author: Fabio Biselli fabio.biselli@studenti.unipr.it Date: Fri Oct 16 11:53:37 2009 +0200

Added BASE field to floating point struct. Modified absolute/relative error definition using FP_Format::BASE. Modified tests using FP_Format::BASE.

---

src/Float.defs.hh | 9 +++++++++ src/Floating_Point_Expression.defs.hh | 17 +++++++++++------ src/Floating_Point_Expression.templates.hh | 11 +++++++---- tests/Floating_Point_Expression/Makefile.am | 3 +-- .../floatingpointexpr1.cc | 10 +++++----- 5 files changed, 33 insertions(+), 17 deletions(-)

diff --git a/src/Float.defs.hh b/src/Float.defs.hh index f0eb5f8..2030f55 100644 --- a/src/Float.defs.hh +++ b/src/Float.defs.hh @@ -49,6 +49,7 @@ struct float_ieee754_half { static const uint16_t NEG_INF = 0x7c00; static const uint16_t POS_ZERO = 0x0000; static const uint16_t NEG_ZERO = 0x8000; + static const unsigned int BASE = 2; static const unsigned int EXPONENT_BITS = 5; static const unsigned int MANTISSA_BITS = 10; static const int EXPONENT_MAX = (1 << (EXPONENT_BITS - 1)) - 1; @@ -80,6 +81,7 @@ struct float_ieee754_single { static const uint32_t NEG_INF = 0xff800000; static const uint32_t POS_ZERO = 0x00000000; static const uint32_t NEG_ZERO = 0x80000000; + static const unsigned int BASE = 2; static const unsigned int EXPONENT_BITS = 8; static const unsigned int MANTISSA_BITS = 23; static const int EXPONENT_MAX = (1 << (EXPONENT_BITS - 1)) - 1; @@ -124,6 +126,7 @@ struct float_ieee754_double { static const uint32_t LSP_INF = 0; static const uint32_t LSP_ZERO = 0; static const uint32_t LSP_MAX = 0xffffffff; + static const unsigned int BASE = 2; static const unsigned int EXPONENT_BITS = 11; static const unsigned int MANTISSA_BITS = 52; static const int EXPONENT_MAX = (1 << (EXPONENT_BITS - 1)) - 1; @@ -154,6 +157,7 @@ struct float_ibm_single { static const uint32_t NEG_INF = 0xff000000; static const uint32_t POS_ZERO = 0x00000000; static const uint32_t NEG_ZERO = 0x80000000; + static const unsigned int BASE = 16; static const unsigned int EXPONENT_BITS = 7; static const unsigned int MANTISSA_BITS = 24; static const int EXPONENT_BIAS = 64; @@ -177,6 +181,9 @@ struct float_ibm_single { #endif // defined(PPL_DOXYGEN_INCLUDE_IMPLEMENTATION_DETAILS)

struct float_ibm_double { + // FIXME: BASE must be 16, 16^56 should return an overflow + // and absolute/relative error are not sound. + static const unsigned int BASE = 2; static const unsigned int EXPONENT_BITS = 7; static const unsigned int MANTISSA_BITS = 56; static const int EXPONENT_BIAS = 64; @@ -203,6 +210,7 @@ struct float_intel_double_extended { static const uint64_t LSP_ZERO = 0; static const uint64_t LSP_DMAX = 0x7fffffffffffffffULL; static const uint64_t LSP_NMAX = 0xffffffffffffffffULL; + static const unsigned int BASE = 2; static const unsigned int EXPONENT_BITS = 15; static const unsigned int MANTISSA_BITS = 63; static const int EXPONENT_MAX = (1 << (EXPONENT_BITS - 1)) - 1; @@ -241,6 +249,7 @@ struct float_ieee754_quad { static const uint64_t LSP_INF = 0; static const uint64_t LSP_ZERO = 0; static const uint64_t LSP_MAX = 0xffffffffffffffffULL; + static const unsigned int BASE = 2; static const unsigned int EXPONENT_BITS = 15; static const unsigned int MANTISSA_BITS = 112; static const int EXPONENT_MAX = (1 << (EXPONENT_BITS - 1)) - 1; diff --git a/src/Floating_Point_Expression.defs.hh b/src/Floating_Point_Expression.defs.hh index fb94f2d..a66190d 100644 --- a/src/Floating_Point_Expression.defs.hh +++ b/src/Floating_Point_Expression.defs.hh @@ -190,12 +190,17 @@ public:

template <typename FP_Interval_Type, typename FP_Format> typename Floating_Point_Expression<FP_Interval_Type, FP_Format>::boundary_type -Floating_Point_Expression<FP_Interval_Type, FP_Format>::absolute_error = - std::max( - static_cast<typename Floating_Point_Expression<FP_Interval_Type, FP_Format>::boundary_type> - (pow(2, static_cast<typename Floating_Point_Expression<FP_Interval_Type, FP_Format>::boundary_type>(1) - FP_Format::EXPONENT_BIAS - FP_Format::MANTISSA_BITS)) - , - std::numeric_limits<typename Floating_Point_Expression<FP_Interval_Type, FP_Format>::boundary_type>::denorm_min()); +Floating_Point_Expression<FP_Interval_Type, FP_Format>::absolute_error = +std::max( + static_cast<typename Floating_Point_Expression<FP_Interval_Type, FP_Format> + ::boundary_type>(pow(FP_Format::BASE, static_cast<typename + Floating_Point_Expression<FP_Interval_Type, FP_Format> + ::boundary_type>(1) - FP_Format + ::EXPONENT_BIAS - FP_Format + ::MANTISSA_BITS)), + std::numeric_limits<typename + Floating_Point_Expression<FP_Interval_Type, FP_Format> + ::boundary_type>::denorm_min());

} // namespace Parma_Polyhedra_Library

diff --git a/src/Floating_Point_Expression.templates.hh b/src/Floating_Point_Expression.templates.hh index 0343bff..11168e0 100644 --- a/src/Floating_Point_Expression.templates.hh +++ b/src/Floating_Point_Expression.templates.hh @@ -35,9 +35,12 @@ Floating_Point_Expression<FP_Interval_Type, FP_Format> ::relative_error(const FP_Linear_Form& lf, FP_Linear_Form& result) { /* FIXME: here we assume that boundary_type can represent (2)^(-FP_Format::MANTISSA_BITS) precisely. */ - FP_Interval_Type error_propagator(-pow(2, -static_cast<typename Floating_Point_Expression<FP_Interval_Type, FP_Format>::boundary_type>(FP_Format::MANTISSA_BITS))); - // FIXME: this may be incorrect for some policies. - error_propagator.join_assign(FP_Interval_Type(pow(2, -static_cast<typename Floating_Point_Expression<FP_Interval_Type, FP_Format>::boundary_type>(FP_Format::MANTISSA_BITS)))); + FP_Interval_Type error_propagator(-pow(FP_Format::BASE, + -static_cast<typename Floating_Point_Expression<FP_Interval_Type, FP_Format> + ::boundary_type>(FP_Format::MANTISSA_BITS))); + error_propagator.join_assign(FP_Interval_Type(pow(FP_Format::BASE, + -static_cast<typename Floating_Point_Expression<FP_Interval_Type, FP_Format> + ::boundary_type>(FP_Format::MANTISSA_BITS))));

// Handle the inhomogeneous term. const FP_Interval_Type* current_term = &lf.inhomogeneous_term(); @@ -52,7 +55,7 @@ Floating_Point_Expression<FP_Interval_Type, FP_Format> for (dimension_type i = 0; i < dimension; ++i) { current_term = &lf.coefficient(Variable(i)); current_multiplier = FP_Interval_Type(std::max(abs(current_term->lower()), - abs(current_term->upper()))); + abs(current_term->upper()))); current_result_term = FP_Linear_Form(Variable(i)); current_result_term *= current_multiplier; current_result_term *= error_propagator; diff --git a/tests/Floating_Point_Expression/Makefile.am b/tests/Floating_Point_Expression/Makefile.am index 9755e4b..7096075 100644 --- a/tests/Floating_Point_Expression/Makefile.am +++ b/tests/Floating_Point_Expression/Makefile.am @@ -51,9 +51,8 @@ $(top_builddir)/src/libppl.la \ @extra_libraries@

ORIGINAL_TESTS = \ -digitalfilters1 +digitalfilters1

-#digitalfilters1 #bdshape2 \ #bdshape1 \ #floatingpointexpr1 \ diff --git a/tests/Floating_Point_Expression/floatingpointexpr1.cc b/tests/Floating_Point_Expression/floatingpointexpr1.cc index c4322eb..c8facf4 100644 --- a/tests/Floating_Point_Expression/floatingpointexpr1.cc +++ b/tests/Floating_Point_Expression/floatingpointexpr1.cc @@ -101,7 +101,7 @@ test04() { Variable A(0); Variable B(1); FP_Linear_Form known_result = FP_Linear_Form(A); - ANALYZER_FP_FORMAT exp = pow(2, + ANALYZER_FP_FORMAT exp = pow(ANALYZED_FP_FORMAT::BASE, -static_cast<ANALYZER_FP_FORMAT>(ANALYZED_FP_FORMAT::MANTISSA_BITS)); tmp = FP_Interval(1); tmp -= exp; @@ -135,7 +135,7 @@ test05() { Variable A(0); Variable B(1); FP_Linear_Form known_result = FP_Linear_Form(A); - ANALYZER_FP_FORMAT exp = pow(2, + ANALYZER_FP_FORMAT exp = pow(ANALYZED_FP_FORMAT::BASE, -static_cast<ANALYZER_FP_FORMAT>(ANALYZED_FP_FORMAT::MANTISSA_BITS)); tmp = FP_Interval(1); tmp -= exp; @@ -169,7 +169,7 @@ test06() {

tmp = FP_Interval(-FP_Expression::absolute_error); tmp.join_assign(FP_Expression::absolute_error); - ANALYZER_FP_FORMAT exp = pow(2, + ANALYZER_FP_FORMAT exp = pow(ANALYZED_FP_FORMAT::BASE, -static_cast<ANALYZER_FP_FORMAT>((ANALYZED_FP_FORMAT::MANTISSA_BITS-1))); FP_Interval coeff = FP_Interval(2); coeff -= exp; @@ -202,7 +202,7 @@ test07() {

tmp = FP_Interval(-FP_Expression::absolute_error); tmp.join_assign(FP_Expression::absolute_error); - ANALYZER_FP_FORMAT exp = pow(2, + ANALYZER_FP_FORMAT exp = pow(ANALYZED_FP_FORMAT::BASE, -static_cast<ANALYZER_FP_FORMAT>((ANALYZED_FP_FORMAT::MANTISSA_BITS+1))); FP_Interval coeff = FP_Interval(1 / 2.0); coeff -= exp; @@ -234,7 +234,7 @@ test08() {

Variable A(0); FP_Linear_Form known_result = FP_Linear_Form(A); - ANALYZER_FP_FORMAT exp = pow(2, + ANALYZER_FP_FORMAT exp = pow(ANALYZED_FP_FORMAT::BASE, -static_cast<ANALYZER_FP_FORMAT>((ANALYZED_FP_FORMAT::MANTISSA_BITS+1))); tmp = FP_Interval(-1 / 2.0); tmp -= exp;