1 files changed, 110 insertions, 0 deletions
diff --git a/src/vectors-8-VSX.h b/src/vectors-8-VSX.h
new file mode 100644
index 0000000..9d7c17c
--- /dev/null
+++ b/src/vectors-8-VSX.h
@@ -0,0 +1,110 @@
+// Vectorise using IBM's Altivec VSX (Power)
+
+// Use the type vector double directly, without introducing a wrapper class
+// Use macros instead of inline functions
+
+
+
+#include <altivec.h>
+
+
+
+// Vector type corresponding to CCTK_REAL
+#define CCTK_REAL8_VEC vector double
+
+// Number of vector elements in a CCTK_REAL_VEC
+#define CCTK_REAL8_VEC_SIZE 2
+
+
+
+// Create vectors, extract vector elements
+
+#define vec8_set1(a)  (vec_splats(a))
+#define vec8_set(a,b)                           \
+({                                              \
+  CCTK_REAL8_VEC x;                             \
+  x[0]=(a);                                     \
+  x[1]=(b);                                     \
+  x;                                            \
+})
+
+#define vec8_elt0(x) ((x)[0])
+#define vec8_elt1(x) ((x)[1])
+#define vec8_elt(x,d) ((x)[d])
+
+
+
+// Load and store vectors
+
+// Load a vector from memory (aligned and unaligned); this loads from
+// a reference to a scalar
+#define vec8_load(p)  (*(CCTK_REAL8_VEC const*)&(p))
+#define vec8_loadu(p) (*(CCTK_REAL8_VEC const*)&(p))
+
+// Load a vector from memory that may or may not be aligned, as
+// decided by the offset and the vector size
+#define vec8_loadu_maybe(off,p)             (vec8_loadu(p))
+#define vec8_loadu_maybe3(off1,off2,off3,p) (vec8_loadu(p))
+
+// Store a vector to memory (aligned and non-temporal); this stores to
+// a reference to a scalar
+#define vec8_store(p,x)     (*(CCTK_REAL8_VEC*)&(p)=(x))
+#define vec8_storeu(p,x)    (*(CCTK_REAL8_VEC*)&(p)=(x))
+#if 1
+#  define vec8_store_nta(p,x) (*(CCTK_REAL8_VEC*)&(p)=(x))
+#else
+// stvxl instruction doesn't exist for double precision
+#  define vec8_store_nta(p,x) (vec_stl(x,0,(CCTK_REAL8_VEC*)&(p)))
+#endif
+
+// Store a lower or higher partial vector (aligned and non-temporal);
+// the non-temporal hint is probably ignored
+#define vec8_store_nta_partial_lo(p,x,n) ((&(p))[0]=(x)[0])
+#define vec8_store_nta_partial_hi(p,x,n) ((&(p))[1]=(x)[1])
+
+
+
+// Functions and operators
+
+// Operators
+#define k8pos(x) (+(x))
+#define k8neg(x) (-(x))
+
+#define k8add(x,y) ((x)+(y))
+#define k8sub(x,y) ((x)-(y))
+#define k8mul(x,y) ((x)*(y))
+#define k8div(x,y) ((x)/(y))
+
+// Fused multiply-add, defined as [+-]x*y[+-]z
+#define k8madd(x,y,z)  (vec_madd(x,y,z))
+#define k8msub(x,y,z)  (vec_msub(x,y,z))
+#define k8nmadd(x,y,z) (vec_nmadd(x,y,z))
+#define k8nmsub(x,y,z) (vec_nmsub(x,y,z))
+
+// Cheap functions
+#define k8fabs(x)   (vec_abs(x))
+#define k8fmax(x,y) (vec_max(x,y))
+#define k8fmin(x,y) (vec_min(x,y))
+#define k8fnabs(x)  (vec_nabs(x))
+
+#define k8exp(x)                                        \
+({                                                      \
+  CCTK_REAL8_VEC const xexp=(x);                        \
+  vec8_set(exp(vec8_elt0(xexp)), exp(vec8_elt1(xexp))); \
+})
+#define k8log(x)                                        \
+({                                                      \
+  CCTK_REAL8_VEC const xlog=(x);                        \
+  vec8_set(log(vec8_elt0(xlog)), log(vec8_elt1(xlog))); \
+})
+#define k8pow(x,a)                                                      \
+({                                                                      \
+  CCTK_REAL8_VEC const xpow=(x);                                        \
+  CCTK_REAL8 const apow=(a);                                            \
+  vec8_set(pow(vec8_elt0(xpow),apow), pow(vec8_elt1(xpow),apow));       \
+})
+#define k8sqrt(x)                                               \
+({                                                              \
+  CCTK_REAL8_VEC const xsqrt=(x);                               \
+  vec8_set(sqrt(vec8_elt0(xsqrt)), sqrt(vec8_elt1(xsqrt)));     \
+})