void add_iteration_lists(Graph g)
{
Node node,child,parent,child1,child2;
int i,j,k,l;
MARKED_1(g -> source) = TRUE;
for(node = g -> source->next; node != NULL; node = node -> next) {
for(i = 0; i < ListSize(node -> predecessors); i++) {
parent = (Node) ListIndex(node -> predecessors,i);
if (MARKED_1(parent) == FALSE) {
for(j=0; j < ListSize(parent -> successors); j++) {
child = (Node) ListIndex(parent->successors,j);
if ((strcmp(child->name,node->name) != 0) && (ORDER(child) > ORDER(parent)))
{
ListPut(ITER_END_NODES(node),child);
}
}
}
}
MARKED_1(node) = TRUE;
for(k = 0; k < ListSize(node -> successors); k++) {
child1 = (Node) ListIndex(node -> successors,k);
if (MARKED_1(child1) == TRUE) {
for(l = 0; l < ListSize(node -> successors); l++) {
child2 = (Node) ListIndex(node -> successors,l);
if (MARKED_1(child2) == FALSE) {
ListPut(ITER_START_NODES(child2),child1);
}
}
}
}
}
}
void handle_selection(Node n)
{
int i,j;
Node parent;
Node child;
if ((n -> predecessors == NULL) || (MARKED_3(n) == TRUE))
return;
MARKED_3(n) = TRUE;
for(i = 0; i < ListSize(n -> predecessors); i++) {
parent = (Node) ListIndex(n -> predecessors,i);
if ((parent -> type) == SELECTION) {
for(j=0; j < ListSize(parent -> successors); j++) {
child = (Node) ListIndex(parent -> successors,j);
if (strcmp((child->name),n->name) != 0) {
mark_successors(child,ACT_NONE);
}
}
}
if (ORDER(n) > ORDER(parent))
handle_selection(parent);
}
return;
}
void
check ()
{
mpfi_t interval;
long a, b, c;
int cmp;
int i;
i = 1;
a = LONG_MAX;
while (a >>= 1) i++;
mpfi_init2 (interval, i);
for (i = 0; i <= 1000; ++i) {
/* random numbers a < b < c */
a = random_si ();
b = random_si ();
c = random_si ();
ORDER (a, b);
ORDER (a, c);
ORDER (b, c);
mpfi_interv_si (interval, b, c);
cmp = mpfi_cmp_si (interval, a);
if (cmp < 0 || (cmp == 0 && a != b)) {
print_error (a, interval);
}
mpfr_set_nan (&(interval->right));
if (mpfi_cmp_si (interval, a) != 1) {
print_error (a, interval);
}
mpfi_interv_si (interval, a, c);
if (mpfi_cmp_si (interval, b) != 0) {
print_error (b, interval);
}
mpfr_set_nan (&(interval->right));
if (mpfi_cmp_si (interval, b) != 1) {
print_error (b, interval);
}
mpfi_interv_si (interval, a, b);
cmp = mpfi_cmp_si (interval, c);
if (cmp > 0 || (cmp == 0 && c != b)) {
print_error (c, interval);
}
mpfr_set_nan (&(interval->right));
if (mpfi_cmp_si (interval, c) != 1) {
print_error (c, interval);
}
}
mpfi_clear (interval);
}
/* update ringlist-tree */
void update_tree(int i, int j) {
baum *rli, *rlj, *tempb;
if ( abs(i) < GSV.len) { /* >> single basepair move */
if ((i > 0) && (j > 0)) { /* insert */
rli = &rl[i-1];
rlj = &rl[j-1];
close_bp_en(rli, rlj);
}
else if ((i < 0)&&(j < 0)) { /* delete */
i = -i;
rli = &rl[i-1];
open_bp_en(rli);
}
else { /* shift */
if (i > 0) { /* i remains the same, j shifts */
j=-j;
rli=&rl[i-1];
rlj=&rl[j-1];
open_bp_en(rli);
ORDER(rli, rlj);
close_bp_en(rli, rlj);
}
else { /* j remains the same, i shifts */
baum *old_rli;
i = -i;
rli = &rl[i-1];
rlj = &rl[j-1];
old_rli = rlj->up;
open_bp_en(old_rli);
ORDER(rli, rlj);
close_bp_en(rli, rlj);
}
}
} /* << single basepair move */
else { /* >> double basepair move */
if ((i > 0) && (j > 0)) { /* insert */
rli = &rl[i-GSV.len-2];
rlj = &rl[j-GSV.len-2];
close_bp_en(rli->next, rlj->prev);
close_bp_en(rli, rlj);
}
else if ((i < 0)&&(j < 0)) { /* delete */
i = -i;
rli = &rl[i-GSV.len-2];
open_bp_en(rli);
open_bp_en(rli->next);
}
} /* << double basepair move */
}
int PARMCI_AccV( int op, /* oeration code */
void *scale, /*scaling factor for accumulate */
armci_giov_t darr[], /* descriptor array */
int len, /* length of descriptor array */
int proc /* remote process(or) ID */
)
{
int rc=0, i,direct=0;
if(len<1) return FAIL;
for(i=0;i<len;i++){
if(darr[i].src_ptr_array==NULL ||darr[i].dst_ptr_array==NULL)return FAIL2;
if(darr[i].bytes<1)return FAIL3;
if(darr[i].ptr_array_len <1) return FAIL4;
}
if(proc<0 || proc >= armci_nproc)return FAIL5;
ORDER(op,proc); /* ensure ordering */
direct=SAMECLUSNODE(proc);
# if defined(ACC_COPY) && !defined(ACC_SMP)
if(armci_me != proc) direct=0;
# error "grrr"
# endif
if(direct) {
rc = armci_acc_vector( op, scale, darr, len, proc);
} else {
DO_FENCE(proc,SERVER_PUT);
rc = armci_pack_vector(op, scale, darr, len, proc,NULL);
}
if(rc) return FAIL6;
else return 0;
}
int PARMCI_GetV( armci_giov_t darr[], /* descriptor array */
int len, /* length of descriptor array */
int proc /* remote process(or) ID */
)
{
int rc=0, i,direct=1;
if(len<1) return FAIL;
for(i=0;i<len;i++){
if(darr[i].src_ptr_array==NULL ||darr[i].dst_ptr_array==NULL)return FAIL2;
if(darr[i].bytes<1)return FAIL3;
if(darr[i].ptr_array_len <1) return FAIL4;
}
if(proc<0 || proc >= armci_nproc)return FAIL5;
ORDER(GET,proc); /* ensure ordering */
#ifndef QUADRICS
direct=SAMECLUSNODE(proc);
#endif
if(direct){
if(!SAMECLUSNODE(proc))DO_FENCE(proc,DIRECT_GET);
rc = armci_copy_vector(GET, darr, len, proc);
}
else{
DO_FENCE(proc,SERVER_GET);
rc = armci_pack_vector(GET, NULL, darr, len, proc,NULL);
}
if(rc) return FAIL6;
else return 0;
}
//===========================================================================
// PIT_CheckLine
// Adjusts tmfloorz and tmceilingz as lines are contacted.
//===========================================================================
static boolean PIT_CheckLine(line_t *ld, void *parm)
{
checkpos_data_t *tm = parm;
fixed_t bbox[4];
// Setup the bounding box for the line.
ORDER(ld->v1->x, ld->v2->x, bbox[BOXLEFT], bbox[BOXRIGHT]);
ORDER(ld->v1->y, ld->v2->y, bbox[BOXBOTTOM], bbox[BOXTOP]);
if(tm->box[BOXRIGHT] <= bbox[BOXLEFT] || tm->box[BOXLEFT] >= bbox[BOXRIGHT]
|| tm->box[BOXTOP] <= bbox[BOXBOTTOM] ||
tm->box[BOXBOTTOM] >= bbox[BOXTOP])
return true;
if(P_BoxOnLineSide(tm->box, ld) != -1)
return true;
// A line has been hit.
tm->thing->wallhit = true;
if(!ld->backsector)
return false; // One sided line, can't go through.
if(!(tm->thing->ddflags & DDMF_MISSILE))
{
if(ld->flags & ML_BLOCKING)
return false; // explicitly blocking everything
}
// set openrange, opentop, openbottom.
P_LineOpening(ld);
// adjust floor / ceiling heights.
if(opentop < tm->ceilingz)
tm->ceilingz = opentop;
if(openbottom > tm->floorz)
tm->floorz = openbottom;
if(lowfloor < tm->dropoffz)
tm->dropoffz = lowfloor;
tm->thing->wallhit = false;
return true;
}
bool OSOrderedSet::setObject(const OSMetaClassBase *anObject )
{
unsigned int i;
// queue it behind those with same priority
for( i = 0;
(i < count) && (ORDER(array[i].obj, anObject) >= 0);
i++ ) {}
return( setObject(i, anObject));
}
// draw a triangle
void ArtDisplayDevice::triangle(const float *a, const float *b, const float *c, const float *n1,
const float *n2, const float *n3) {
float vec1[3], vec2[3], vec3[3];
float nor1[3], nor2[3], nor3[3];
// transform the world coordinates
(transMat.top()).multpoint3d(a, vec1);
(transMat.top()).multpoint3d(b, vec2);
(transMat.top()).multpoint3d(c, vec3);
// and the normals
(transMat.top()).multnorm3d(n1, nor1);
(transMat.top()).multnorm3d(n2, nor2);
(transMat.top()).multnorm3d(n3, nor3);
// draw the triangle
fprintf(outfile, "polygon {\n");
fprintf(outfile, "colour %f,%f,%f\n",
matData[colorIndex][0],
matData[colorIndex][1],
matData[colorIndex][2]);
fprintf(outfile, "vertex (%f,%f,%f),(%f,%f,%f)\n",
ORDER(vec1[0], vec1[1], vec1[2]), // point one
ORDER(nor1[0], nor1[1], nor1[2]));
fprintf(outfile, "vertex (%f,%f,%f),(%f,%f,%f)\n",
ORDER(vec2[0], vec2[1], vec2[2]), // point two
ORDER(nor2[0], nor2[1], nor2[2]));
fprintf(outfile, "vertex (%f,%f,%f),(%f,%f,%f)\n",
ORDER(vec3[0], vec3[1], vec3[2]), // point three
ORDER(nor3[0], nor3[1], nor3[2]));
fprintf(outfile, "}\n");
}
// draw a cone
void ArtDisplayDevice::cone(float *a, float *b, float r) {
float vec1[3], vec2[3];
// transform the world coordinates
(transMat.top()).multpoint3d(a, vec1);
(transMat.top()).multpoint3d(b, vec2);
fprintf(outfile, "cone {\n");
fprintf(outfile, "colour %f,%f,%f\n",
matData[colorIndex][0],
matData[colorIndex][1],
matData[colorIndex][2]);
// second point
fprintf(outfile, "vertex(%f,%f,%f)\n",
ORDER(vec2[0], vec2[1], vec2[2]));
// first point
fprintf(outfile, "center(%f,%f,%f)\n",
ORDER(vec1[0], vec1[1], vec1[2]));
// radius
fprintf(outfile, "radius %f\n}\n", scale_radius(r));
}
// draw a point
void ArtDisplayDevice::point(float * spdata) {
float vec[3];
// transform the world coordinates
(transMat.top()).multpoint3d(spdata, vec);
// draw the sphere
fprintf(outfile, "sphere {\ncolour %f,%f,%f\n",
matData[colorIndex][0],
matData[colorIndex][1],
matData[colorIndex][2]);
fprintf(outfile, "radius %f\n", float(lineWidth) * DEFAULT_RADIUS);
fprintf(outfile, "center (%f,%f,%f)\n}\n", ORDER(vec[0], vec[1], vec[2]));
}
// draw a sphere
void ArtDisplayDevice::sphere(float * spdata) {
float vec[3];
float radius;
// transform the world coordinates
(transMat.top()).multpoint3d(spdata, vec);
radius = scale_radius(spdata[3]);
// draw the sphere
fprintf(outfile, "sphere {\ncolour %f,%f,%f\n",
matData[colorIndex][0],
matData[colorIndex][1],
matData[colorIndex][2]);
fprintf(outfile, "radius %f\n", radius);
fprintf(outfile, "center (%f,%f,%f)\n}\n", ORDER(vec[0], vec[1], vec[2]));
}
// Read a value; use only after gpio_adc_sample() returns zero
uint16_t
gpio_adc_read(struct gpio_adc g)
{
adc_status.chan |= ADC_DONE;
// Perform median filter on 5 read samples
uint16_t *p = adc_status.samples;
uint32_t v0 = p[0], v4 = p[1], v1 = p[2], v3 = p[3], v2 = p[4];
ORDER(v0, v4);
ORDER(v1, v3);
ORDER(v0, v1);
ORDER(v3, v4);
ORDER(v1, v3);
ORDER(v1, v2);
ORDER(v2, v3);
return v2;
}
Node make_node(char *name, vm_act_state state, int type,int order)
{
Node n = (Node) malloc (sizeof(struct node));
n -> data = (void *) malloc (sizeof (struct data));
n->name = name;
STATE(n) = state;
n->type = type;
n->script = "script";
n -> predecessors = NULL;
n -> successors = NULL;
ITER_START(n) = 0;
ITER_END(n) = 0;
MARKED_0(n) = 0;
MARKED_1(n) = 0;
MARKED_2(n) = 0;
MARKED_3(n) = 0;
MARKED_4(n) = 0;
ITER_START_NODES(n) = ListCreate();
ITER_END_NODES(n) = ListCreate();
SUPER_NODES(n) = ListCreate();
ORDER(n) = order;
return n;
}
void initialize_graph(Graph g, int pid)
{
Node n;
int i = 0;
for(n = g -> source; n != NULL; n = n -> next) {
n -> data = (void *) malloc (sizeof (struct data));
sanitize_node(n);
PID(n) = pid;
STATE(n) = ACT_NONE;
ORDER(n) = i;
i++;
ITER_START(n) = FALSE;
ITER_END(n) = FALSE;
ITER_START_NODES(n) = ListCreate();
ITER_END_NODES(n) = ListCreate();
SUPER_NODES(n) = ListCreate();
}
add_super_node_lists(g); /* add the node lists */
mark_for_iteration(g); /* mark beginning and end of iterations */
add_iteration_lists(g); /* add the iteration lists */
sanitize(g); /* sanitize markers */
mark_successors(g->source->next,ACT_READY);
}
// draw a line (cylinder) from a to b
void ArtDisplayDevice::line(float *a, float *b) {
int i, j, test;
float dirvec[3], unitdirvec[3];
float from[3], to[3], tmp1[3], tmp2[3];
float len;
if(lineStyle == ::SOLIDLINE ) {
// transform the world coordinates
(transMat.top()).multpoint3d(a, from);
(transMat.top()).multpoint3d(b, to);
// draw the cylinder
fprintf(outfile, "cylinder {\n");
fprintf(outfile, "colour %f,%f,%f\n",
matData[colorIndex][0],
matData[colorIndex][1],
matData[colorIndex][2]);
fprintf(outfile, "center(%f,%f,%f)\n",
ORDER(from[0], from[1], from[2])); // first point
fprintf(outfile, "center(%f,%f,%f)\n",
ORDER(to[0], to[1], to[2])); // second point
fprintf(outfile, "radius %f\n}\n",
float(lineWidth)*DEFAULT_RADIUS); // radius
} else if (lineStyle == ::DASHEDLINE ) {
// transform the world coordinates
(transMat.top()).multpoint3d(a, tmp1);
(transMat.top()).multpoint3d(b, tmp2);
// how to create a dashed line
for(i=0;i<3;i++) {
dirvec[i] = tmp2[i] - tmp1[i]; // vector from a to b
}
len = sqrtf( dirvec[0]*dirvec[0] + dirvec[1]*dirvec[1] + dirvec[2]*dirvec[2] );
for(i=0;i<3;i++) {
unitdirvec[i] = dirvec[i] / sqrtf(len); // unit vector pointing from a to b
}
test = 1;
i = 0;
while( test == 1 ) {
for(j=0;j<3;j++) {
from[j] = tmp1[j] + (2*i)*DASH_LENGTH*unitdirvec[j];
to[j] = tmp1[j] + (2*i + 1)*DASH_LENGTH*unitdirvec[j];
}
if( fabsf(tmp1[0] - to[0]) >= fabsf(dirvec[0]) ) {
for(j=0;j<3;j++) {
to[j] = tmp2[j];
}
test = 0;
}
// draw the cylinder
fprintf(outfile, "cylinder {\n");
fprintf(outfile, "colour %f,%f,%f\n",
matData[colorIndex][0],
matData[colorIndex][1],
matData[colorIndex][2]);
// first point
fprintf(outfile, "center(%f,%f,%f)\n",
ORDER(from[0], from[1], from[2]));
// second point
fprintf(outfile, "center(%f,%f,%f)\n",
ORDER(to[0], to[1], to[2]));
// radius
fprintf(outfile, "radius %f\n}\n",
float(lineWidth)*DEFAULT_RADIUS);
i++;
}
} else {
msgErr << "ArtDisplayDevice: Unknown line style " << lineStyle << sendmsg;
}
}
void ap_physical_gradients(double* restrict C, double* restrict B,
double* restrict ref_dx, double* restrict ref_dy,
LAPACKINDEX num_points,
double* restrict dx, double* restrict dy)
{
double dx_unmapped[21*num_points];
double dy_unmapped[21*num_points];
int i;
/* stuff for DGEMM */
LAPACKINDEX i_twentyone = 21;
/* Calculate the physical-to-reference mapping. */
const double B_det_inv = 1/(B[ORDER(0, 0, 2, 2)]*B[ORDER(1, 1, 2, 2)] -
B[ORDER(0, 1, 2, 2)]*B[ORDER(1, 0, 2, 2)]);
const double B_inv00 = B_det_inv*B[ORDER(1, 1, 2, 2)];
const double B_inv01 = -B_det_inv*B[ORDER(0, 1, 2, 2)];
const double B_inv10 = -B_det_inv*B[ORDER(1, 0, 2, 2)];
const double B_inv11 = B_det_inv*B[ORDER(0, 0, 2, 2)];
/*
* Perform the transformation using B inverse. This is equivalent to
* putting the reference values in long columns side-by-side and
* multiplying by B_inv.
*/
for (i = 0; i < 21*num_points; i++) {
dx_unmapped[i] = B_inv00*ref_dx[i] + B_inv10*ref_dy[i];
dy_unmapped[i] = B_inv01*ref_dx[i] + B_inv11*ref_dy[i];
}
SInt32 OSOrderedSet::orderObject( const OSMetaClassBase * anObject )
{
return( ORDER( anObject, 0 ));
}
struct fpn *
fpu_add(struct fpemu *fe)
{
struct fpn *x = &fe->fe_f1, *y = &fe->fe_f2, *r;
u_int r0, r1, r2, r3;
int rd;
/*
* Put the `heavier' operand on the right (see fpu_emu.h).
* Then we will have one of the following cases, taken in the
* following order:
*
* - y = NaN. Implied: if only one is a signalling NaN, y is.
* The result is y.
* - y = Inf. Implied: x != NaN (is 0, number, or Inf: the NaN
* case was taken care of earlier).
* If x = -y, the result is NaN. Otherwise the result
* is y (an Inf of whichever sign).
* - y is 0. Implied: x = 0.
* If x and y differ in sign (one positive, one negative),
* the result is +0 except when rounding to -Inf. If same:
* +0 + +0 = +0; -0 + -0 = -0.
* - x is 0. Implied: y != 0.
* Result is y.
* - other. Implied: both x and y are numbers.
* Do addition a la Hennessey & Patterson.
*/
DPRINTF(FPE_REG, ("fpu_add:\n"));
DUMPFPN(FPE_REG, x);
DUMPFPN(FPE_REG, y);
DPRINTF(FPE_REG, ("=>\n"));
ORDER(x, y);
if (ISNAN(y)) {
fe->fe_cx |= FPSCR_VXSNAN;
DUMPFPN(FPE_REG, y);
return (y);
}
if (ISINF(y)) {
if (ISINF(x) && x->fp_sign != y->fp_sign) {
fe->fe_cx |= FPSCR_VXISI;
return (fpu_newnan(fe));
}
DUMPFPN(FPE_REG, y);
return (y);
}
rd = ((fe->fe_fpscr) & FPSCR_RN);
if (ISZERO(y)) {
if (rd != FSR_RD_RM) /* only -0 + -0 gives -0 */
y->fp_sign &= x->fp_sign;
else /* any -0 operand gives -0 */
y->fp_sign |= x->fp_sign;
DUMPFPN(FPE_REG, y);
return (y);
}
if (ISZERO(x)) {
DUMPFPN(FPE_REG, y);
return (y);
}
/*
* We really have two numbers to add, although their signs may
* differ. Make the exponents match, by shifting the smaller
* number right (e.g., 1.011 => 0.1011) and increasing its
* exponent (2^3 => 2^4). Note that we do not alter the exponents
* of x and y here.
*/
r = &fe->fe_f3;
r->fp_class = FPC_NUM;
if (x->fp_exp == y->fp_exp) {
r->fp_exp = x->fp_exp;
r->fp_sticky = 0;
} else {
if (x->fp_exp < y->fp_exp) {
/*
* Try to avoid subtract case iii (see below).
* This also guarantees that x->fp_sticky = 0.
*/
SWAP(x, y);
}
/* now x->fp_exp > y->fp_exp */
r->fp_exp = x->fp_exp;
r->fp_sticky = fpu_shr(y, x->fp_exp - y->fp_exp);
}
r->fp_sign = x->fp_sign;
if (x->fp_sign == y->fp_sign) {
FPU_DECL_CARRY
/*
* The signs match, so we simply add the numbers. The result
* may be `supernormal' (as big as 1.111...1 + 1.111...1, or
* 11.111...0). If so, a single bit shift-right will fix it
* (but remember to adjust the exponent).
*/
/* r->fp_mant = x->fp_mant + y->fp_mant */
FPU_ADDS(r->fp_mant[3], x->fp_mant[3], y->fp_mant[3]);
FPU_ADDCS(r->fp_mant[2], x->fp_mant[2], y->fp_mant[2]);
FPU_ADDCS(r->fp_mant[1], x->fp_mant[1], y->fp_mant[1]);
FPU_ADDC(r0, x->fp_mant[0], y->fp_mant[0]);
if ((r->fp_mant[0] = r0) >= FP_2) {
(void) fpu_shr(r, 1);
r->fp_exp++;
}
} else {
FPU_DECL_CARRY
/*
* The signs differ, so things are rather more difficult.
* H&P would have us negate the negative operand and add;
* this is the same as subtracting the negative operand.
* This is quite a headache. Instead, we will subtract
* y from x, regardless of whether y itself is the negative
* operand. When this is done one of three conditions will
* hold, depending on the magnitudes of x and y:
* case i) |x| > |y|. The result is just x - y,
* with x's sign, but it may need to be normalized.
* case ii) |x| = |y|. The result is 0 (maybe -0)
* so must be fixed up.
* case iii) |x| < |y|. We goofed; the result should
* be (y - x), with the same sign as y.
* We could compare |x| and |y| here and avoid case iii,
* but that would take just as much work as the subtract.
* We can tell case iii has occurred by an overflow.
*
* N.B.: since x->fp_exp >= y->fp_exp, x->fp_sticky = 0.
*/
/* r->fp_mant = x->fp_mant - y->fp_mant */
FPU_SET_CARRY(y->fp_sticky);
FPU_SUBCS(r3, x->fp_mant[3], y->fp_mant[3]);
FPU_SUBCS(r2, x->fp_mant[2], y->fp_mant[2]);
FPU_SUBCS(r1, x->fp_mant[1], y->fp_mant[1]);
FPU_SUBC(r0, x->fp_mant[0], y->fp_mant[0]);
if (r0 < FP_2) {
/* cases i and ii */
if ((r0 | r1 | r2 | r3) == 0) {
/* case ii */
r->fp_class = FPC_ZERO;
r->fp_sign = rd == FSR_RD_RM;
return (r);
}
} else {
/*
* Oops, case iii. This can only occur when the
* exponents were equal, in which case neither
* x nor y have sticky bits set. Flip the sign
* (to y's sign) and negate the result to get y - x.
*/
#ifdef DIAGNOSTIC
if (x->fp_exp != y->fp_exp || r->fp_sticky)
panic("fpu_add");
#endif
r->fp_sign = y->fp_sign;
FPU_SUBS(r3, 0, r3);
FPU_SUBCS(r2, 0, r2);
FPU_SUBCS(r1, 0, r1);
FPU_SUBC(r0, 0, r0);
}
r->fp_mant[3] = r3;
r->fp_mant[2] = r2;
r->fp_mant[1] = r1;
r->fp_mant[0] = r0;
if (r0 < FP_1)
fpu_norm(r);
}
DUMPFPN(FPE_REG, r);
return (r);
}
void ap_matrix_betaplane(double* restrict C, double* restrict B,
double* restrict ref_values,
double* restrict ref_dx, double* restrict ref_dy,
double* restrict weights,
LAPACKINDEX num_points, double* restrict betaplane)
{
int i;
double values[21*num_points];
double dx[21*num_points];
double dy[21*num_points];
double weights_scaled[num_points];
/* stuff for DGEMM. */
LAPACKINDEX i_twentyone = 21;
const double jacobian = fabs(B[ORDER(0, 0, 2, 2)]*B[ORDER(1, 1, 2, 2)] -
B[ORDER(0, 1, 2, 2)]*B[ORDER(1, 0, 2, 2)]);
ap_physical_gradients(C, B, ref_dx, ref_dy, num_points, dx, dy);
ap_physical_values(C, ref_values, num_points, values);
/* scale the weights by the jacobian. */
for (i = 0; i < num_points; i++) {
weights_scaled[i] = weights[i]*jacobian;
}
/*
* scale the function values by the weights and determinant. Then
* perform matrix multiplication.
*/
ap_diagonal_multiply(21, num_points, values, weights_scaled);
struct fpn *
fpu_div(struct fpemu *fe)
{
struct fpn *x = &fe->fe_f1, *y = &fe->fe_f2;
u_int q, bit;
u_int r0, r1, r2, r3, d0, d1, d2, d3, y0, y1, y2, y3;
FPU_DECL_CARRY
/*
* Since divide is not commutative, we cannot just use ORDER.
* Check either operand for NaN first; if there is at least one,
* order the signalling one (if only one) onto the right, then
* return it. Otherwise we have the following cases:
*
* Inf / Inf = NaN, plus NV exception
* Inf / num = Inf [i.e., return x]
* Inf / 0 = Inf [i.e., return x]
* 0 / Inf = 0 [i.e., return x]
* 0 / num = 0 [i.e., return x]
* 0 / 0 = NaN, plus NV exception
* num / Inf = 0
* num / num = num (do the divide)
* num / 0 = Inf, plus DZ exception
*/
DPRINTF(FPE_REG, ("fpu_div:\n"));
DUMPFPN(FPE_REG, x);
DUMPFPN(FPE_REG, y);
DPRINTF(FPE_REG, ("=>\n"));
if (ISNAN(x) || ISNAN(y)) {
ORDER(x, y);
fe->fe_cx |= FPSCR_VXSNAN;
DUMPFPN(FPE_REG, y);
return (y);
}
/*
* Need to split the following out cause they generate different
* exceptions.
*/
if (ISINF(x)) {
if (x->fp_class == y->fp_class) {
fe->fe_cx |= FPSCR_VXIDI;
return (fpu_newnan(fe));
}
DUMPFPN(FPE_REG, x);
return (x);
}
if (ISZERO(x)) {
fe->fe_cx |= FPSCR_ZX;
if (x->fp_class == y->fp_class) {
fe->fe_cx |= FPSCR_VXZDZ;
return (fpu_newnan(fe));
}
DUMPFPN(FPE_REG, x);
return (x);
}
/* all results at this point use XOR of operand signs */
x->fp_sign ^= y->fp_sign;
if (ISINF(y)) {
x->fp_class = FPC_ZERO;
DUMPFPN(FPE_REG, x);
return (x);
}
if (ISZERO(y)) {
fe->fe_cx = FPSCR_ZX;
x->fp_class = FPC_INF;
DUMPFPN(FPE_REG, x);
return (x);
}
/*
* Macros for the divide. See comments at top for algorithm.
* Note that we expand R, D, and Y here.
*/
#define SUBTRACT /* D = R - Y */ \
FPU_SUBS(d3, r3, y3); FPU_SUBCS(d2, r2, y2); \
FPU_SUBCS(d1, r1, y1); FPU_SUBC(d0, r0, y0)
#define NONNEGATIVE /* D >= 0 */ \
((int)d0 >= 0)
#ifdef FPU_SHL1_BY_ADD
#define SHL1 /* R <<= 1 */ \
FPU_ADDS(r3, r3, r3); FPU_ADDCS(r2, r2, r2); \
FPU_ADDCS(r1, r1, r1); FPU_ADDC(r0, r0, r0)
#else
#define SHL1 \
r0 = (r0 << 1) | (r1 >> 31), r1 = (r1 << 1) | (r2 >> 31), \
r2 = (r2 << 1) | (r3 >> 31), r3 <<= 1
#endif
#define LOOP /* do ... while (bit >>= 1) */ \
do { \
SHL1; \
SUBTRACT; \
if (NONNEGATIVE) { \
q |= bit; \
r0 = d0, r1 = d1, r2 = d2, r3 = d3; \
} \
} while ((bit >>= 1) != 0)
#define WORD(r, i) /* calculate r->fp_mant[i] */ \
q = 0; \
bit = 1 << 31; \
LOOP; \
(x)->fp_mant[i] = q
/* Setup. Note that we put our result in x. */
r0 = x->fp_mant[0];
r1 = x->fp_mant[1];
r2 = x->fp_mant[2];
r3 = x->fp_mant[3];
y0 = y->fp_mant[0];
y1 = y->fp_mant[1];
y2 = y->fp_mant[2];
y3 = y->fp_mant[3];
bit = FP_1;
SUBTRACT;
if (NONNEGATIVE) {
x->fp_exp -= y->fp_exp;
r0 = d0, r1 = d1, r2 = d2, r3 = d3;
q = bit;
bit >>= 1;
} else {
/*
* The multiplication algorithm for normal numbers is as follows:
*
* The fraction of the product is built in the usual stepwise fashion.
* Each step consists of shifting the accumulator right one bit
* (maintaining any guard bits) and, if the next bit in y is set,
* adding the multiplicand (x) to the accumulator. Then, in any case,
* we advance one bit leftward in y. Algorithmically:
*
* A = 0;
* for (bit = 0; bit < FP_NMANT; bit++) {
* sticky |= A & 1, A >>= 1;
* if (Y & (1 << bit))
* A += X;
* }
*
* (X and Y here represent the mantissas of x and y respectively.)
* The resultant accumulator (A) is the product's mantissa. It may
* be as large as 11.11111... in binary and hence may need to be
* shifted right, but at most one bit.
*
* Since we do not have efficient multiword arithmetic, we code the
* accumulator as four separate words, just like any other mantissa.
* We use local variables in the hope that this is faster than memory.
* We keep x->fp_mant in locals for the same reason.
*
* In the algorithm above, the bits in y are inspected one at a time.
* We will pick them up 32 at a time and then deal with those 32, one
* at a time. Note, however, that we know several things about y:
*
* - the guard and round bits at the bottom are sure to be zero;
*
* - often many low bits are zero (y is often from a single or double
* precision source);
*
* - bit FP_NMANT-1 is set, and FP_1*2 fits in a word.
*
* We can also test for 32-zero-bits swiftly. In this case, the center
* part of the loop---setting sticky, shifting A, and not adding---will
* run 32 times without adding X to A. We can do a 32-bit shift faster
* by simply moving words. Since zeros are common, we optimize this case.
* Furthermore, since A is initially zero, we can omit the shift as well
* until we reach a nonzero word.
*/
struct fpn *
fpu_mul(struct fpemu *fe)
{
struct fpn *x = &fe->fe_f1, *y = &fe->fe_f2;
u_int a3, a2, a1, a0, x3, x2, x1, x0, bit, m;
int sticky;
FPU_DECL_CARRY;
/*
* Put the `heavier' operand on the right (see fpu_emu.h).
* Then we will have one of the following cases, taken in the
* following order:
*
* - y = NaN. Implied: if only one is a signalling NaN, y is.
* The result is y.
* - y = Inf. Implied: x != NaN (is 0, number, or Inf: the NaN
* case was taken care of earlier).
* If x = 0, the result is NaN. Otherwise the result
* is y, with its sign reversed if x is negative.
* - x = 0. Implied: y is 0 or number.
* The result is 0 (with XORed sign as usual).
* - other. Implied: both x and y are numbers.
* The result is x * y (XOR sign, multiply bits, add exponents).
*/
DPRINTF(FPE_REG, ("fpu_mul:\n"));
DUMPFPN(FPE_REG, x);
DUMPFPN(FPE_REG, y);
DPRINTF(FPE_REG, ("=>\n"));
ORDER(x, y);
if (ISNAN(y)) {
y->fp_sign ^= x->fp_sign;
fe->fe_cx |= FPSCR_VXSNAN;
DUMPFPN(FPE_REG, y);
return (y);
}
if (ISINF(y)) {
if (ISZERO(x)) {
fe->fe_cx |= FPSCR_VXIMZ;
return (fpu_newnan(fe));
}
y->fp_sign ^= x->fp_sign;
DUMPFPN(FPE_REG, y);
return (y);
}
if (ISZERO(x)) {
x->fp_sign ^= y->fp_sign;
DUMPFPN(FPE_REG, x);
return (x);
}
/*
* Setup. In the code below, the mask `m' will hold the current
* mantissa byte from y. The variable `bit' denotes the bit
* within m. We also define some macros to deal with everything.
*/
x3 = x->fp_mant[3];
x2 = x->fp_mant[2];
x1 = x->fp_mant[1];
x0 = x->fp_mant[0];
sticky = a3 = a2 = a1 = a0 = 0;
#define ADD /* A += X */ \
FPU_ADDS(a3, a3, x3); \
FPU_ADDCS(a2, a2, x2); \
FPU_ADDCS(a1, a1, x1); \
FPU_ADDC(a0, a0, x0)
#define SHR1 /* A >>= 1, with sticky */ \
sticky |= a3 & 1, a3 = (a3 >> 1) | (a2 << 31), \
a2 = (a2 >> 1) | (a1 << 31), a1 = (a1 >> 1) | (a0 << 31), a0 >>= 1
#define SHR32 /* A >>= 32, with sticky */ \
sticky |= a3, a3 = a2, a2 = a1, a1 = a0, a0 = 0
#define STEP /* each 1-bit step of the multiplication */ \
SHR1; if (bit & m) { ADD; }; bit <<= 1
/*
* We are ready to begin. The multiply loop runs once for each
* of the four 32-bit words. Some words, however, are special.
* As noted above, the low order bits of Y are often zero. Even
* if not, the first loop can certainly skip the guard bits.
* The last word of y has its highest 1-bit in position FP_NMANT-1,
* so we stop the loop when we move past that bit.
*/
if ((m = y->fp_mant[3]) == 0) {
/* SHR32; */ /* unneeded since A==0 */
} else {
bit = 1 << FP_NG;
do {
STEP;
} while (bit != 0);
}
if ((m = y->fp_mant[2]) == 0) {
SHR32;
} else {
bit = 1;
do {
STEP;
} while (bit != 0);
}
if ((m = y->fp_mant[1]) == 0) {
SHR32;
} else {
bit = 1;
do {
STEP;
} while (bit != 0);
}
m = y->fp_mant[0]; /* definitely != 0 */
bit = 1;
do {
STEP;
} while (bit <= m);
/*
* Done with mantissa calculation. Get exponent and handle
* 11.111...1 case, then put result in place. We reuse x since
* it already has the right class (FP_NUM).
*/
m = x->fp_exp + y->fp_exp;
if (a0 >= FP_2) {
SHR1;
m++;
}
x->fp_sign ^= y->fp_sign;
x->fp_exp = m;
x->fp_sticky = sticky;
x->fp_mant[3] = a3;
x->fp_mant[2] = a2;
x->fp_mant[1] = a1;
x->fp_mant[0] = a0;
DUMPFPN(FPE_REG, x);
return (x);
}
