Rename SSE assembly kernels

src/compute_hh_trafo.X90

@@ -488,12 +488,14 @@
                  &MATH_DATATYPE&
 		 &_&
 		 &PRECISION&
+		 &_sse_assembly&
 		 & (c_loc(a(1,j+off+a_off-1,istripe,my_thread)), w, nbw, nl, stripe_width, nbw)
 #else
                  call double_hh_trafo_&
 		 &MATH_DATATYPE&
 		 &_&
 		 &PRECISION&
+		 &_sse_assembly&
 		 & (c_loc(a(1,j+off+a_off-1,istripe)), w, nbw, nl, stripe_width, nbw)
 #endif
                enddo
@@ -517,12 +519,14 @@
 	      &MATH_DATATYPE&
 	      &_&
 	      &PRECISION&
+	      &_sse_assembly&
 	      & (a(1,j+off+a_off,istripe,my_thread), bcast_buffer(1,j+off),nbw,nl,stripe_width)
 #else
               call single_hh_trafo_&
 	      &MATH_DATATYPE&
 	      &_&
 	      &PRECISION&
+	      &_sse_assembly&
 	      & (a(1,j+off+a_off,istripe), bcast_buffer(1,j+off),nbw,nl,stripe_width)
 #endif
             enddo

src/elpa2_kernels/elpa2_kernels_asm_x86_64_double_precision.s

@@ -54,8 +54,8 @@
 # distributed along with the original code in the file "COPYING".
 #
 # --------------------------------------------------------------------------------------------------
-        .globl double_hh_trafo_real_double
-        .globl single_hh_trafo_complex_double
+        .globl double_hh_trafo_real_double_sse_assembly
+        .globl single_hh_trafo_complex_double_sse_assembly
         .text
 
 #-------------------------------------------------------------------------------
@@ -367,7 +367,7 @@
 #-------------------------------------------------------------------------------
 # FORTRAN Interface:
 #
-# subroutine double_hh_trafo_real_double(q, hh, nb, nq, ldq, ldh)
+# subroutine double_hh_trafo_real_double_sse_assembly(q, hh, nb, nq, ldq, ldh)
 #
 #   integer, intent(in) :: nb, nq, ldq, ldh
 #   real*8, intent(inout) :: q(ldq,*)
@@ -385,7 +385,8 @@
 
 #!f>#ifdef WITH_REAL_SSE_ASSEMBLY_KERNEL
 #!f>  interface
-#!f>    subroutine double_hh_trafo_real_double(q, hh, nb, nq, ldq, ldh) bind(C,name="double_hh_trafo_real_double")
+#!f>    subroutine double_hh_trafo_real_double_sse_assembly(q, hh, nb, nq, ldq, ldh) &
+#!f>      bind(C,name="double_hh_trafo_real_double_sse_assembly")
 #!f>      use, intrinsic :: iso_c_binding
 #!f>      integer(kind=c_int)  :: nb, nq, ldq, ldh
 #!f>      type(c_ptr), value   :: q
@@ -394,7 +395,7 @@
 #!f>  end interface
 #!f>#endif
         .align    16,0x90
-double_hh_trafo_real_double:
+double_hh_trafo_real_double_sse_assembly:
 
         # Get integer parameters into corresponding registers
 
@@ -695,7 +696,7 @@ return1:
 #-------------------------------------------------------------------------------
 # FORTRAN Interface:
 #
-# subroutine single_hh_trafo_complex_double(q, hh, nb, nq, ldq)
+# subroutine single_hh_trafo_complex_double_sse_assembly(q, hh, nb, nq, ldq)
 #
 #   integer, intent(in) :: nb, nq, ldq
 #   complex*16, intent(inout) :: q(ldq,*)
@@ -711,7 +712,8 @@ return1:
 #-------------------------------------------------------------------------------
 #!f>#ifdef WITH_COMPLEX_SSE_ASSEMBLY_KERNEL
 #!f>  interface
-#!f>    subroutine single_hh_trafo_complex_double(q, hh, nb, nq, ldq) bind(C,name="single_hh_trafo_complex_double")
+#!f>    subroutine single_hh_trafo_complex_double_sse_assembly(q, hh, nb, nq, ldq) &
+#!f>      bind(C,name="single_hh_trafo_complex_double_sse_assembly")
 #!f>      use, intrinsic :: iso_c_binding
 #!f>      integer(kind=c_int)     :: nb, nq, ldq
 #!f>      complex(kind=c_double_complex)  :: q(*)
@@ -720,7 +722,7 @@ return1:
 #!f>  end interface
 #!f>#endif
         .align    16,0x90
-single_hh_trafo_complex_double:
+single_hh_trafo_complex_double_sse_assembly:
 
         # Get integer parameters into corresponding registers
 

src/elpa2_kernels/elpa2_kernels_asm_x86_64_single_precision.s

@@ -41,8 +41,8 @@
 #
 # Author: Andreas Marek, MPCDF
 
-        .globl double_hh_trafo_real_single
-        .globl single_hh_trafo_complex_single
+        .globl double_hh_trafo_real_single_sse_assembly
+        .globl single_hh_trafo_complex_single_sse_assembly
 
 	.text
 #-------------------------------------------------------------------------------
@@ -376,7 +376,7 @@
 #-------------------------------------------------------------------------------
 # FORTRAN Interface:
 #
-# subroutine double_hh_trafo_real_single(q, hh, nb, nq, ldq, ldh)
+# subroutine double_hh_trafo_real_single_sse_assembly(q, hh, nb, nq, ldq, ldh)
 #
 #   integer, intent(in) :: nb, nq, ldq, ldh
 #   real*8, intent(inout) :: q(ldq,*)
@@ -394,7 +394,8 @@
 #!f>#ifdef WITH_REAL_SSE_ASSEMBLY_KERNEL
 #!f>#ifdef WANT_SINGLE_PRECISION_REAL
 #!f>  interface
-#!f>    subroutine double_hh_trafo_real_single(q, hh, nb, nq, ldq, ldh) bind(C,name="double_hh_trafo_real_single")
+#!f>    subroutine double_hh_trafo_real_single_sse_assembly(q, hh, nb, nq, ldq, ldh) &
+#!f>      bind(C,name="double_hh_trafo_real_single_sse_assembly")
 #!f>      use, intrinsic :: iso_c_binding
 #!f>      integer(kind=c_int) :: nb, nq, ldq, ldh
 #!f>      type(c_ptr), value  :: q
@@ -404,7 +405,7 @@
 #!f>#endif
 #!f>#endif
         .align    16,0x90
-double_hh_trafo_real_single:
+double_hh_trafo_real_single_sse_assembly:
 
         # Get integer parameters into corresponding registers
 
@@ -714,7 +715,7 @@ return1:
 #-------------------------------------------------------------------------------
 # FORTRAN Interface:
 #
-# subroutine single_hh_trafo_complex_single(q, hh, nb, nq, ldq)
+# subroutine single_hh_trafo_complex_single_sse_assembly(q, hh, nb, nq, ldq)
 #
 #   integer, intent(in) :: nb, nq, ldq
 #   complex(kind=c_float_complex), intent(inout) :: q(ldq,*)
@@ -731,7 +732,8 @@ return1:
 #!f>#ifdef WITH_COMPLEX_SSE_ASSEMBLY_KERNEL
 #!f>#ifdef WANT_SINGLE_PRECISION_COMPLEX
 #!f>  interface
-#!f>    subroutine single_hh_trafo_complex_single(q, hh, nb, nq, ldq) bind(C,name="single_hh_trafo_complex_single")
+#!f>    subroutine single_hh_trafo_complex_single_sse_assembly(q, hh, nb, nq, ldq) &
+#!f>      bind(C,name="single_hh_trafo_complex_single_sse_assembly")
 #!f>      use, intrinsic :: iso_c_binding
 #!f>      integer(kind=c_int)   :: nb, nq, ldq
 #!f>      complex(kind=c_float_complex) :: q(*)
@@ -742,7 +744,7 @@ return1:
 #!f>#endif
 
         .align    16,0x90
-single_hh_trafo_complex_single:
+single_hh_trafo_complex_single_sse_assembly:
 
         # Get integer parameters into corresponding registers