/* decrease indentation and close for loop */
void
mknb_end_loop(void)
{
mknb_indent_level--;
if(mknb_fortran)
mknb_code("end do");
else
mknb_code("}");
mknb_code("");
}
/* integer constant */
void
mknb_declare_const_int(char *name, int value)
{
char type_name[255];
sprintf(type_name, "%-13s", mknb_fortran ? "integer*4" : "const int");
if(mknb_fortran) {
mknb_code("%s %s", type_name ,name);
mknb_code(" parameter (%s = %d)",name,value);
} else
mknb_code("%s %s = %d;", type_name ,name, value);
}
void
mknb_do_else()
{
mknb_indent_level--;
if(mknb_fortran)
mknb_code("else");
else
{
mknb_code("}");
mknb_code("else");
mknb_code("{");
}
mknb_indent_level++;
}
void
mknb_start_if(char *cond)
{
mknb_code("");
if(mknb_fortran)
mknb_code("if (%s) then",cond);
else
{
mknb_code("if(%s)", cond);
mknb_code("{");
}
mknb_indent_level++;
}
/* Declare a constant fp variable */
void
mknb_declare_const_real(char *name, double value)
{
char type_name[255];
if(mknb_fortran) {
sprintf(type_name, "%-13s", mknb_double ? "real*8" : "real*4");
mknb_code("%s %s",type_name,name);
mknb_code(" parameter (%s = %f)",name,value);
} else {
sprintf(type_name, "%-13s",
mknb_double ? "const double" : "const float");
mknb_code("%s %s = %.16f;",type_name,name,value);
}
}
/* Start a for loop and increase indentation.*/
void
mknb_start_loop(char *lvar,char *from,char *to)
{
mknb_code("");
if(mknb_fortran)
{
mknb_code("do %s=%s,%s",lvar,from,to);
}
else
{
mknb_code("for(%s=%s; (%s<%s); %s++)", lvar,from,lvar,to,lvar);
mknb_code("{");
}
mknb_indent_level++;
}
/* Arbitrary declaration */
void
mknb_declare_other(char *type_name,char *name)
{
char tmp[255];
sprintf(tmp,"%-13s", type_name);
mknb_code("%s %s%s", tmp, name, mknb_fortran ? "" : ";");
}
/* 4-byte Integer (same size as single precision fp) */
void
mknb_declare_int4(char *name)
{
char type_name[255];
sprintf(type_name, "%-13s", mknb_fortran ? "integer*4" : "int");
mknb_code("%s %s%s", type_name ,name, mknb_fortran ? "" : ";");
}
/* Declare a single-precision floating point var. */
void
mknb_declare_real4(char *name)
{
char type_name[255];
sprintf(type_name, "%-13s", mknb_fortran ? "real*4" : "float");
mknb_code("%s %s%s", type_name ,name, mknb_fortran ? "" : ";");
}
/* Define a new floating-point variable */
void
mknb_declare_real(char *name)
{
char type_name[255];
if(mknb_fortran)
sprintf(type_name, "%-13s", mknb_double ? "real*8" : "real*4");
else
sprintf(type_name, "%-13s", mknb_double ? "double" : "float");
mknb_code("%s %s%s",type_name,name, mknb_fortran ? "" : ";");
}
void
mknb_outerloop(void) {
int i,nflops = 0;
char tmp[255];
int indent;
if(mknb_options.threads) {
mknb_comment("Loop over thread workunits");
if(mknb_fortran) {
char space[25];
indent = MKNB_FORTRAN_INDENT_STEP*mknb_indent_level;
for(i=0 ; i<indent ; i++)
space[i]=' ';
space[i]=0;
fprintf(mknb_output,
" 10 %scall f77kernelsync(mtx,count,nri,nthreads,nn0,nn1)\n",space);
/* since f77 use call-by-reference we can send the pointer of the
* count variable and the mutex to c without fortran knowing about it!
*/
mknb_indent_level++;
mknb_code("if(nn1.gt.nri) nn1=nri");
mknb_comment("Start outer loop over neighborlists");
mknb_start_loop("n", "nn0+1", "nn1");
} else {
/* C */
mknb_code("");
mknb_code("do");
mknb_code("{");
mknb_indent_level++;
mknb_code("gmx_thread_mutex_lock((gmx_thread_mutex_t *)mtx);");
mknb_assign("nn0","*count");
mknb_comment("Take successively smaller chunks (at least 10 lists)");
mknb_assign("nn1","nn0+(nri-nn0)/(2*nthreads)+10");
/* take sucessively smaller chunks */
mknb_assign("*count","nn1");
mknb_code("gmx_thread_mutex_unlock((gmx_thread_mutex_t *)mtx);");
mknb_code("if(nn1>nri) nn1=nri;");
mknb_comment("Start outer loop over neighborlists");
mknb_start_loop("n", "nn0", "nn1");
}
} else {
mknb_comment("Start outer loop over neighborlists");
mknb_start_loop("n", (mknb_fortran) ? "1" : "0", "nri");
}
/* load shift index and then shift vector for this list */
nflops += mknb_load_shift_vector();
mknb_comment("Load limits for loop over neighbors");
mknb_assign("nj0", "%s%s", mknb_array("jindex","n"),
(mknb_fortran) ? "+1" : "");
mknb_assign("nj1", mknb_array("jindex","n+1"));
mknb_comment("Get outer coordinate index");
mknb_assign("ii", "%s%s", mknb_array("iinr","n"),
(mknb_fortran) ? "+1" : "");
mknb_assign("ii3", "3*ii%s", (mknb_fortran) ? "-2" : "");
nflops += mknb_load_outer_coordinates();
nflops += mknb_load_outer_parameters();
nflops += mknb_zero_outer_potential();
nflops += mknb_zero_outer_forces();
/* do the inner loop ( separate nflops, so no return value) */
mknb_innerloop();
nflops += mknb_update_outer_forces();
nflops += mknb_update_outer_potential();
/* Update innerloop counter (for accounting) if we did threads */
mknb_comment("Increment number of inner iterations");
if(mknb_options.threads)
mknb_assign("ninner","ninner + nj1 - nj0");
/* close outer loop */
sprintf(tmp,"Outer loop uses %d flops/iteration",nflops);
mknb_comment(tmp);
mknb_end_loop();
if(mknb_options.threads) {
mknb_comment("Increment number of outer iterations");
mknb_assign("nouter","nouter + nn1 - nn0");
mknb_indent_level--;
if(mknb_fortran)
mknb_code("if(nn1.lt.nri) goto 10");
else
{
mknb_code("}");
mknb_code("while (nn1<nri);");
mknb_code("");
}
}
}
