Skip to content

GitLab

Commit ab215a89 authored by Andreas Marek's avatar Andreas Marek
Browse files

Fix some errors in real Neon

parent 4d0b0ab1
Hide whitespace changes
Inline Side-by-side
configure.ac
View file @ ab215a89
......@@ -227,9 +227,9 @@ fi
dnl check which MPI binray invokes a MPI job
if test x"$with_mpi" = x"yes"; then
AC_CHECK_PROGS([MPI_BINARY], [mpiexec.hydra mpiexec mpirun poe runjob srun], [no])
AC_CHECK_PROGS([MPI_BINARY], [mpiexec.hydra mpiexec mpirun poe runjob srun aprun], [no])
if test x"$MPI_BINARY" = x"no"; then
AC_MSG_ERROR([Could not find either of the MPI binaries: mpiexec.hydra, mpiexec, mpirun, poe, runjob, srun])
AC_MSG_ERROR([Could not find either of the MPI binaries: mpiexec.hydra, mpiexec, mpirun, poe, runjob, srun, aprun])
fi
fi
......@@ -929,7 +929,7 @@ if test x"${need_neon_arch64}" = x"yes"; then
#include <arm_neon.h>
int main(int argc, char **argv) {
__Float64x2_t x1, x2, x3, x4;
x4 = vfmaq_64(x1, x2, x3);
x4 = vfmaq_f64(x1, x2, x3);
return 0;
}
])],
......
src/elpa2/compute_hh_trafo.F90
View file @ ab215a89
......@@ -2192,7 +2192,7 @@
#else
call quad_hh_trafo_&
&MATH_DATATYPE&
&neon_arch64_4hv_&
&_neon_arch64_4hv_&
&PRECISION&
& (c_loc(a(1,jj+off+a_off-3,istripe)), w, &
nbw, nl, stripe_width, nbw)
......
src/elpa2/kernels/real_128bit_256bit_512bit_BLOCK_template.c
View file @ ab215a89
......@@ -381,7 +381,7 @@
#endif
 
 
#if VEC_SET == 128 || VEC_SET == 1281 || VEC_SET == NEON_ARCH64_128
#if VEC_SET == 128 || VEC_SET == 1281 || VEC_SET == 1282 || VEC_SET == NEON_ARCH64_128
//Forward declaration
#ifdef DOUBLE_PRECISION_REAL
#undef ROW_LENGTH
......@@ -391,7 +391,7 @@
#undef ROW_LENGTH
#define ROW_LENGTH 4
#endif
#endif /* VEC_SET == 128 || VEC_SET == 1281 || VEC_SET == NEON_ARCH64_128 */
#endif /* VEC_SET == 128 || VEC_SET == 1281 || VEC_SET == 1282 || VEC_SET == NEON_ARCH64_128 */
 
#if VEC_SET == 256
#ifdef DOUBLE_PRECISION_REAL
......@@ -425,7 +425,7 @@ __forceinline void CONCAT_8ARGS(hh_trafo_kernel_,ROW_LENGTH,_,SIMD_SET,_,BLOCK,h
DATA_TYPE_PTR scalarprods);
#endif
 
#if VEC_SET == 128 || VEC_SET == 1281 || VEC_SET == NEON_ARCH64_128
#if VEC_SET == 128 || VEC_SET == 1281 || VEC_SET == 1282 || VEC_SET == NEON_ARCH64_128
#ifdef DOUBLE_PRECISION_REAL
#undef ROW_LENGTH
#define ROW_LENGTH 4
......@@ -434,7 +434,7 @@ __forceinline void CONCAT_8ARGS(hh_trafo_kernel_,ROW_LENGTH,_,SIMD_SET,_,BLOCK,h
#undef ROW_LENGTH
#define ROW_LENGTH 8
#endif
#endif /* VEC_SET == 128 || VEC_SET == 1281 || VEC_SET == NEON_ARCH64_128 */
#endif /* VEC_SET == 128 || VEC_SET == 1281 || VEC_SET == 1282 || VEC_SET == NEON_ARCH64_128 */
 
#if VEC_SET == 256
#ifdef DOUBLE_PRECISION_REAL
......@@ -468,7 +468,7 @@ __forceinline void CONCAT_8ARGS(hh_trafo_kernel_,ROW_LENGTH,_,SIMD_SET,_,BLOCK,h
DATA_TYPE_PTR scalarprods);
#endif
 
#if VEC_SET == 128 || VEC_SET == 1281 || VEC_SET == NEON_ARCH64_128
#if VEC_SET == 128 || VEC_SET == 1281 || VEC_SET == 1282 || VEC_SET == NEON_ARCH64_128
#ifdef DOUBLE_PRECISION_REAL
#undef ROW_LENGTH
#define ROW_LENGTH 6
......@@ -477,7 +477,7 @@ __forceinline void CONCAT_8ARGS(hh_trafo_kernel_,ROW_LENGTH,_,SIMD_SET,_,BLOCK,h
#undef ROW_LENGTH
#define ROW_LENGTH 12
#endif
#endif /* VEC_SET == 128 || VEC_SET == 1281 || VEC_SET == NEON_ARCH64_128 */
#endif /* VEC_SET == 128 || VEC_SET == 1281 || VEC_SET == 1282 || VEC_SET == NEON_ARCH64_128 */
 
#if VEC_SET == 256
#ifdef DOUBLE_PRECISION_REAL
......@@ -512,7 +512,7 @@ __forceinline void CONCAT_8ARGS(hh_trafo_kernel_,ROW_LENGTH,_,SIMD_SET,_,BLOCK,h
DATA_TYPE_PTR scalarprods);
#endif
 
#if VEC_SET == 128 || VEC_SET == 1281 || VEC_SET == NEON_ARCH64_128
#if VEC_SET == 128 || VEC_SET == 1281 || VEC_SET == 1282 || VEC_SET == NEON_ARCH64_128
#ifdef DOUBLE_PRECISION_REAL
#undef ROW_LENGTH
#define ROW_LENGTH 8
......@@ -521,7 +521,7 @@ __forceinline void CONCAT_8ARGS(hh_trafo_kernel_,ROW_LENGTH,_,SIMD_SET,_,BLOCK,h
#undef ROW_LENGTH
#define ROW_LENGTH 16
#endif
#endif /* VEC_SET == 128 || VEC_SET == 1281 || VEC_SET == NEON_ARCH64_128 */
#endif /* VEC_SET == 128 || VEC_SET == 1281 || VEC_SET == 1282 || VEC_SET == NEON_ARCH64_128 */
 
#if VEC_SET == 256
#ifdef DOUBLE_PRECISION_REAL
......@@ -556,7 +556,7 @@ __forceinline void CONCAT_8ARGS(hh_trafo_kernel_,ROW_LENGTH,_,SIMD_SET,_,BLOCK,h
DATA_TYPE_PTR scalarprods);
#endif
 
#if VEC_SET == 128 || VEC_SET == 1281 || VEC_SET == NEON_ARCH64_128
#if VEC_SET == 128 || VEC_SET == 1281 || VEC_SET == 1282 || VEC_SET == NEON_ARCH64_128
#ifdef DOUBLE_PRECISION_REAL
#undef ROW_LENGTH
#define ROW_LENGTH 10
......@@ -565,7 +565,7 @@ __forceinline void CONCAT_8ARGS(hh_trafo_kernel_,ROW_LENGTH,_,SIMD_SET,_,BLOCK,h
#undef ROW_LENGTH
#define ROW_LENGTH 20
#endif
#endif /* VEC_SET == 128 || VEC_SET == 1281 || VEC_SET == NEON_ARCH64_128 */
#endif /* VEC_SET == 128 || VEC_SET == 1281 || VEC_SET == 1282 || VEC_SET == NEON_ARCH64_128 */
 
#if VEC_SET == 256
#ifdef DOUBLE_PRECISION_REAL
......@@ -600,7 +600,7 @@ __forceinline void CONCAT_8ARGS(hh_trafo_kernel_,ROW_LENGTH,_,SIMD_SET,_,BLOCK,h
DATA_TYPE_PTR scalarprods);
#endif
 
#if VEC_SET == 128 || VEC_SET == 1281 || VEC_SET == NEON_ARCH64_128
#if VEC_SET == 128 || VEC_SET == 1281 || VEC_SET == 1282 || VEC_SET == NEON_ARCH64_128
#ifdef DOUBLE_PRECISION_REAL
#undef ROW_LENGTH
#define ROW_LENGTH 12
......@@ -609,7 +609,7 @@ __forceinline void CONCAT_8ARGS(hh_trafo_kernel_,ROW_LENGTH,_,SIMD_SET,_,BLOCK,h
#undef ROW_LENGTH
#define ROW_LENGTH 24
#endif
#endif /* VEC_SET == 128 || VEC_SET == 1281 || VEC_SET == NEON_ARCH64_128 */
#endif /* VEC_SET == 128 || VEC_SET == 1281 || VEC_SET == 1282 || VEC_SET == NEON_ARCH64_128 */
 
#if VEC_SET == 256
#ifdef DOUBLE_PRECISION_REAL
......@@ -982,7 +982,7 @@ void CONCAT_7ARGS(PREFIX,_hh_trafo_real_,SIMD_SET,_,BLOCK,hv_,WORD_LENGTH) (DATA
/*
!f>#ifdef HAVE_NEON_ARCH64_SSE
!f> interface
!f> subroutine hexa_hh_trafo_real_NEON64_6hv_double(q, hh, pnb, pnq, pldq, pldh) &
!f> subroutine hexa_hh_trafo_real_NEON_ARCH64_6hv_double(q, hh, pnb, pnq, pldq, pldh) &
!f> bind(C, name="hexa_hh_trafo_real_NEON_ARCH64_6hv_double")
!f> use, intrinsic :: iso_c_binding
!f> integer(kind=c_int) :: pnb, pnq, pldq, pldh
......@@ -1744,14 +1744,14 @@ void CONCAT_7ARGS(PREFIX,_hh_trafo_real_,SIMD_SET,_,BLOCK,hv_,WORD_LENGTH) (DATA
}
 
#undef ROW_LENGTH
#if VEC_SET == 128 || VEC_SET == 1281 || VEC_SET == NEON_ARCH64_128
#if VEC_SET == 128 || VEC_SET == 1281 || VEC_SET == 1282 || VEC_SET == NEON_ARCH64_128
#ifdef DOUBLE_PRECISION_REAL
#define ROW_LENGTH 12
#endif
#ifdef SINGLE_PRECISION_REAL
#define ROW_LENGTH 24
#endif
#endif /* VEC_SET == 128 || VEC_SET == 1281 || VEC_SET == NEON_ARCH64_128 */
#endif /* VEC_SET == 128 || VEC_SET == 1281 || VEC_SET == 1282 || VEC_SET == NEON_ARCH64_128 */
 
#if VEC_SET == 256
#ifdef DOUBLE_PRECISION_REAL
......@@ -3312,7 +3312,7 @@ __forceinline void CONCAT_8ARGS(hh_trafo_kernel_,ROW_LENGTH,_,SIMD_SET,_,BLOCK,h
#endif
 
#if VEC_SET == 1281 || VEC_SET == NEON_ARCH64_128
h1 = SIMD_NEG(tau1);
h1 = _SIMD_NEG(tau1);
#endif
 
#if VEC_SET == 512
......@@ -5040,14 +5040,14 @@ __forceinline void CONCAT_8ARGS(hh_trafo_kernel_,ROW_LENGTH,_,SIMD_SET,_,BLOCK,h
}
 
#undef ROW_LENGTH
#if VEC_SET == 128 || VEC_SET == 1281
#if VEC_SET == 128 || VEC_SET == 1281 || VEC_SET == 1282 || VEC_SET == NEON_ARCH64_128
#ifdef DOUBLE_PRECISION_REAL
#define ROW_LENGTH 10
#endif
#ifdef SINGLE_PRECISION_REAL
#define ROW_LENGTH 20
#endif
#endif /* VEC_SET == 128 || VEC_SET == 1281 */
#endif /* VEC_SET == 128 || VEC_SET == 1281 || VEC_SET == 1282 || VEC_SET == NEON_ARCH64_128 */
 
#if VEC_SET == 256
#ifdef DOUBLE_PRECISION_REAL
......@@ -8057,14 +8057,14 @@ __forceinline void CONCAT_8ARGS(hh_trafo_kernel_,ROW_LENGTH,_,SIMD_SET,_,BLOCK,h
 
 
#undef ROW_LENGTH
#if VEC_SET == 128 || VEC_SET == 1281
#if VEC_SET == 128 || VEC_SET == 1281 || VEC_SET == 1282 || VEC_SET == NEON_ARCH64_128
#ifdef DOUBLE_PRECISION_REAL
#define ROW_LENGTH 8
#endif
#ifdef SINGLE_PRECISION_REAL
#define ROW_LENGTH 16
#endif
#endif /* VEC_SET == 128 || VEC_SET == 1281 */
#endif /* VEC_SET == 128 || VEC_SET == 1281 || VEC_SET == 1282 || VEC_SET == NEON_ARCH64_128 */
 
#if VEC_SET == 256
#ifdef DOUBLE_PRECISION_REAL
......@@ -10781,14 +10781,14 @@ __forceinline void CONCAT_8ARGS(hh_trafo_kernel_,ROW_LENGTH,_,SIMD_SET,_,BLOCK,h
}
 
#undef ROW_LENGTH
#if VEC_SET == 128 || VEC_SET == 1281
#if VEC_SET == 128 || VEC_SET == 1281 || VEC_SET == 1282 || VEC_SET == NEON_ARCH64_128
#ifdef DOUBLE_PRECISION_REAL
#define ROW_LENGTH 6
#endif
#ifdef SINGLE_PRECISION_REAL
#define ROW_LENGTH 12
#endif
#endif /* VEC_SET == 128 || VEC_SET == 1281 */
#endif /* VEC_SET == 128 || VEC_SET == 1281 || VEC_SET == 1282 || VEC_SET == NEON_ARCH64_128 */
 
#if VEC_SET == 256
#ifdef DOUBLE_PRECISION_REAL
......@@ -11956,7 +11956,7 @@ __forceinline void CONCAT_8ARGS(hh_trafo_kernel_,ROW_LENGTH,_,SIMD_SET,_,BLOCK,h
#endif
 
#if VEC_SET == 1281 || VEC_SET == NEON_ARCH64_128
h1 = _(tau1);
h1 = _SIMD_NEG(tau1);
#endif
 
#if VEC_SET == 512
......@@ -13238,7 +13238,7 @@ __forceinline void CONCAT_8ARGS(hh_trafo_kernel_,ROW_LENGTH,_,SIMD_SET,_,BLOCK,h
 
 
#undef ROW_LENGTH
#if VEC_SET == 128 || VEC_SET == 1281
#if VEC_SET == 128 || VEC_SET == 1281 || VEC_SET == 1282 || VEC_SET == NEON_ARCH64_128
#ifdef DOUBLE_PRECISION_REAL
#undef ROW_LENGTH
#define ROW_LENGTH 4
......@@ -13247,7 +13247,7 @@ __forceinline void CONCAT_8ARGS(hh_trafo_kernel_,ROW_LENGTH,_,SIMD_SET,_,BLOCK,h
#undef ROW_LENGTH
#define ROW_LENGTH 8
#endif
#endif /* VEC_SET == 128 || VEC_SET == 1281 */
#endif /* VEC_SET == 128 || VEC_SET == 1281 || VEC_SET == 1282 || VEC_SET == NEON_ARCH64_128 */
 
#if VEC_SET == 256
#ifdef DOUBLE_PRECISION_REAL
......@@ -15409,7 +15409,7 @@ __forceinline void CONCAT_8ARGS(hh_trafo_kernel_,ROW_LENGTH,_,SIMD_SET,_,BLOCK,h
}
 
#undef ROW_LENGTH
#if VEC_SET == 128 || VEC_SET == 1281
#if VEC_SET == 128 || VEC_SET == 1281 || VEC_SET == 1282 || VEC_SET == NEON_ARCH64_128
#ifdef DOUBLE_PRECISION_REAL
#undef ROW_LENGTH
#define ROW_LENGTH 2
......@@ -15418,7 +15418,7 @@ __forceinline void CONCAT_8ARGS(hh_trafo_kernel_,ROW_LENGTH,_,SIMD_SET,_,BLOCK,h
#undef ROW_LENGTH
#define ROW_LENGTH 4
#endif
#endif /* VEC_SET == 128 || VEC_SET == 1281 */
#endif /* VEC_SET == 128 || VEC_SET == 1281 || VEC_SET == 1282 || VEC_SET == NEON_ARCH64_128 */
 
#if VEC_SET == 256
#ifdef DOUBLE_PRECISION_REAL
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment