diff options
| -rw-r--r-- | instantiations.cc | 113 |
1 files changed, 66 insertions, 47 deletions
diff --git a/instantiations.cc b/instantiations.cc index 345add5..a8150a0 100644 --- a/instantiations.cc +++ b/instantiations.cc @@ -78,12 +78,13 @@ namespace vecmathlib { +#include <cassert> #include <cstdlib> using namespace std; // Vector size; this is system-specific and needs to be manually // adapted -#define VECSIZE 4 +const int VECSIZE = 4; using namespace vecmathlib; typedef realvec<double,VECSIZE> doubleV; @@ -100,16 +101,16 @@ typedef realvec<double,VECSIZE> doubleV; // Simple, naive loop adding two arrays extern "C" -void loop_add(double* a_, - double* b_, - double* c_, +void loop_add(double* a, + double* b, + double* c, ptrdiff_t n) { - doubleV* a = (doubleV*)a_; - doubleV* b = (doubleV*)b_; - doubleV* c = (doubleV*)c_; - for (ptrdiff_t i=0; i<n/VECSIZE; ++i) { - a[i] = b[i] + c[i]; + for (ptrdiff_t i=0; i<n; i+=VECSIZE) { + doubleV tmpb = doubleV::loadu(&b[i]); + doubleV tmpc = doubleV::loadu(&c[i]); + doubleV tmpa = tmpb + tmpc; + storeu(tmpa, &a[i]); } } @@ -117,16 +118,16 @@ void loop_add(double* a_, // Declare pointers as restrict extern "C" -void loop_add_restrict(double *restrict a_, - double *restrict b_, - double *restrict c_, +void loop_add_restrict(double *restrict a, + double *restrict b, + double *restrict c, ptrdiff_t n) { - doubleV *restrict a = (doubleV*)a_; - doubleV *restrict b = (doubleV*)b_; - doubleV *restrict c = (doubleV*)c_; - for (ptrdiff_t i=0; i<n/VECSIZE; ++i) { - a[i] = b[i] + c[i]; + for (ptrdiff_t i=0; i<n; i+=VECSIZE) { + doubleV tmpb = doubleV::loadu(&b[i]); + doubleV tmpc = doubleV::loadu(&c[i]); + doubleV tmpa = tmpb + tmpc; + storeu(tmpa, &a[i]); } } @@ -134,16 +135,16 @@ void loop_add_restrict(double *restrict a_, // Declare pointers as restrict and aligned extern "C" -void loop_add_aligned(double *restrict a_, - double *restrict b_, - double *restrict c_, +void loop_add_aligned(double *restrict a, + double *restrict b, + double *restrict c, ptrdiff_t n) { - doubleV *restrict a = (doubleV*)__builtin_assume_aligned(a_, sizeof *a); - doubleV *restrict b = (doubleV*)__builtin_assume_aligned(b_, sizeof *b); - doubleV *restrict c = (doubleV*)__builtin_assume_aligned(c_, sizeof *c); - for (ptrdiff_t i=0; i<n/VECSIZE; ++i) { - a[i] = b[i] + c[i]; + for (ptrdiff_t i=0; i<n; i+=VECSIZE) { + doubleV tmpb = doubleV::loada(&b[i]); + doubleV tmpc = doubleV::loada(&c[i]); + doubleV tmpa = tmpb + tmpc; + storea(tmpa, &a[i]); } } @@ -151,15 +152,15 @@ void loop_add_aligned(double *restrict a_, // Reduction loop extern "C" -double loop_dot_reduce(double *restrict a_, - double *restrict b_, +double loop_dot_reduce(double *restrict a, + double *restrict b, ptrdiff_t n) { - doubleV *restrict a = (doubleV*)__builtin_assume_aligned(a_, sizeof *a); - doubleV *restrict b = (doubleV*)__builtin_assume_aligned(b_, sizeof *b); doubleV sumV = 0.0; - for (ptrdiff_t i=0; i<n/VECSIZE; ++i) { - sumV += a[i] * b[i]; + for (ptrdiff_t i=0; i<n; i+=VECSIZE) { + doubleV tmpa = doubleV::loada(&a[i]); + doubleV tmpb = doubleV::loada(&b[i]); + sumV += tmpa * tmpb; } return sum(sumV); } @@ -168,16 +169,16 @@ double loop_dot_reduce(double *restrict a_, // Loop with a simple if condition (fmax) extern "C" -void loop_if_simple(double *restrict a_, - double *restrict b_, - double *restrict c_, +void loop_if_simple(double *restrict a, + double *restrict b, + double *restrict c, ptrdiff_t n) { - doubleV *restrict a = (doubleV*)__builtin_assume_aligned(a_, sizeof *a); - doubleV *restrict b = (doubleV*)__builtin_assume_aligned(b_, sizeof *b); - doubleV *restrict c = (doubleV*)__builtin_assume_aligned(c_, sizeof *c); - for (ptrdiff_t i=0; i<n/VECSIZE; ++i) { - a[i] = ifthen(b[i] > c[i], b[i], c[i]); + for (ptrdiff_t i=0; i<n; i+=VECSIZE) { + doubleV tmpb = doubleV::loada(&b[i]); + doubleV tmpc = doubleV::loada(&c[i]); + doubleV tmpa = ifthen(tmpb > tmpc, tmpb, tmpc); + storea(tmpa, &a[i]); } } @@ -185,15 +186,33 @@ void loop_if_simple(double *restrict a_, // Loop with a complex if condition (select) extern "C" -void loop_if(double *restrict a_, - double *restrict b_, - double *restrict c_, +void loop_if(double *restrict a, + double *restrict b, + double *restrict c, ptrdiff_t n) { - doubleV *restrict a = (doubleV*)__builtin_assume_aligned(a_, sizeof *a); - doubleV *restrict b = (doubleV*)__builtin_assume_aligned(b_, sizeof *b); - doubleV *restrict c = (doubleV*)__builtin_assume_aligned(c_, sizeof *c); - for (ptrdiff_t i=0; i<n/VECSIZE; ++i) { - a[i] = ifthen(b[i] > doubleV(0.0), b[i] * c[i], doubleV(1.0)); + for (ptrdiff_t i=0; i<n; i+=VECSIZE) { + doubleV tmpb = doubleV::loada(&b[i]); + doubleV tmpc = doubleV::loada(&c[i]); + doubleV tmpa = ifthen(tmpb > doubleV(0.0), tmpb * tmpc, doubleV(1.0)); + storea(tmpa, &a[i]); + } +} + + + +// Skip ghost points +extern "C" +void loop_add_masked(double *restrict a, + double *restrict b, + double *restrict c, + ptrdiff_t n) +{ + for (doubleV::mask_t mask(1, n-1, 0); mask; ++mask) { + ptrdiff_t i = mask.index(); + doubleV tmpb = doubleV::loada(&b[i]); + doubleV tmpc = doubleV::loada(&c[i]); + doubleV tmpa = tmpb + tmpc; + storea(tmpa, &a[i], mask); } } |
