op_plan *op_plan_core(char const *name, op_set set, int part_size, int nargs,
op_arg *args, int ninds, int *inds, int staging) {
// set exec length
int exec_length = set->size;
for (int i = 0; i < nargs; i++) {
if (args[i].opt && args[i].idx != -1 && args[i].acc != OP_READ) {
exec_length += set->exec_size;
break;
}
}
/* first look for an existing execution plan */
int ip = 0, match = 0;
while (match == 0 && ip < OP_plan_index) {
if ((strcmp(name, OP_plans[ip].name) == 0) && (set == OP_plans[ip].set) &&
(nargs == OP_plans[ip].nargs) && (ninds == OP_plans[ip].ninds) &&
(part_size == OP_plans[ip].part_size)) {
match = 1;
for (int m = 0; m < nargs; m++) {
if (args[m].dat != NULL && OP_plans[ip].dats[m] != NULL)
match = match && (args[m].dat->size == OP_plans[ip].dats[m]->size) &&
(args[m].dat->dim == OP_plans[ip].dats[m]->dim) &&
(args[m].map == OP_plans[ip].maps[m]) &&
(args[m].idx == OP_plans[ip].idxs[m]) &&
(args[m].acc == OP_plans[ip].accs[m]);
else
match = match && (args[m].dat == OP_plans[ip].dats[m]) &&
(args[m].map == OP_plans[ip].maps[m]) &&
(args[m].idx == OP_plans[ip].idxs[m]) &&
(args[m].acc == OP_plans[ip].accs[m]);
}
}
ip++;
}
if (match) {
ip--;
if (OP_diags > 3)
printf(" old execution plan #%d\n", ip);
OP_plans[ip].count++;
return &(OP_plans[ip]);
} else {
if (OP_diags > 1)
printf(" new execution plan #%d for kernel %s\n", ip, name);
}
double wall_t1, wall_t2, cpu_t1, cpu_t2;
op_timers_core(&cpu_t1, &wall_t1);
/* work out worst case shared memory requirement per element */
int halo_exchange = 0;
for (int i = 0; i < nargs; i++) {
if (args[i].opt && args[i].idx != -1 && args[i].acc != OP_WRITE &&
args[i].acc != OP_INC) {
halo_exchange = 1;
break;
}
}
int maxbytes = 0;
for (int m = 0; m < nargs; m++) {
if (args[m].opt && inds[m] >= 0) {
if ((staging == OP_STAGE_INC && args[m].acc == OP_INC) ||
(staging == OP_STAGE_ALL || staging == OP_STAGE_PERMUTE))
maxbytes += args[m].dat->size;
}
}
/* set blocksize and number of blocks; adaptive size based on 48kB of shared
* memory */
int bsize = part_size; // blocksize
if (bsize == 0 && maxbytes > 0)
bsize = MAX((24 * 1024 / (64 * maxbytes)) * 64,
256); // 48kB exactly is too much, make it 24
else if (bsize == 0 && maxbytes == 0)
bsize = 256;
// If we do 1 level of coloring, do it in one go
if (staging == OP_COLOR2)
bsize = exec_length;
int nblocks = 0;
int indirect_reduce = 0;
for (int m = 0; m < nargs; m++) {
indirect_reduce |=
(args[m].acc != OP_READ && args[m].argtype == OP_ARG_GBL);
}
indirect_reduce &= (ninds > 0);
/* Work out indirection arrays for OP_INCs */
int ninds_staged = 0; // number of distinct (unique dat) indirect incs
int *inds_staged = (int *)op_malloc(nargs * sizeof(int));
int *inds_to_inds_staged = (int *)op_malloc(ninds * sizeof(int));
for (int i = 0; i < nargs; i++)
inds_staged[i] = -1;
for (int i = 0; i < ninds; i++)
inds_to_inds_staged[i] = -1;
for (int i = 0; i < nargs; i++) {
if (inds[i] >= 0 &&
((staging == OP_STAGE_INC && args[i].acc == OP_INC) ||
(staging == OP_STAGE_ALL || staging == OP_STAGE_PERMUTE))) {
if (inds_to_inds_staged[inds[i]] == -1) {
inds_to_inds_staged[inds[i]] = ninds_staged;
inds_staged[i] = ninds_staged;
ninds_staged++;
} else {
inds_staged[i] = inds_to_inds_staged[inds[i]];
}
}
}
int *invinds_staged = (int *)op_malloc(ninds_staged * sizeof(int));
for (int i = 0; i < ninds_staged; i++)
invinds_staged[i] = -1;
for (int i = 0; i < nargs; i++)
if (inds[i] >= 0 &&
((staging == OP_STAGE_INC && args[i].acc == OP_INC) ||
(staging == OP_STAGE_ALL || staging == OP_STAGE_PERMUTE)) &&
invinds_staged[inds_staged[i]] == -1)
invinds_staged[inds_staged[i]] = i;
int prev_offset = 0;
int next_offset = 0;
while (next_offset < exec_length) {
prev_offset = next_offset;
if (prev_offset + bsize >= set->core_size && prev_offset < set->core_size) {
next_offset = set->core_size;
} else if (prev_offset + bsize >= set->size && prev_offset < set->size &&
indirect_reduce) {
next_offset = set->size;
} else if (prev_offset + bsize >= exec_length &&
prev_offset < exec_length) {
next_offset = exec_length;
} else {
next_offset = prev_offset + bsize;
}
nblocks++;
}
// If we do 1 level of coloring, we have a single "block"
if (staging == OP_COLOR2) {
nblocks = 1;
prev_offset = 0;
next_offset = exec_length;
};
/* enlarge OP_plans array if needed */
if (ip == OP_plan_max) {
// printf("allocating more memory for OP_plans %d\n", OP_plan_max);
OP_plan_max += 10;
OP_plans = (op_plan *)op_realloc(OP_plans, OP_plan_max * sizeof(op_plan));
if (OP_plans == NULL) {
printf(" op_plan error -- error reallocating memory for OP_plans\n");
exit(-1);
}
}
/* allocate memory for new execution plan and store input arguments */
OP_plans[ip].dats = (op_dat *)op_malloc(nargs * sizeof(op_dat));
OP_plans[ip].idxs = (int *)op_malloc(nargs * sizeof(int));
OP_plans[ip].optflags = (int *)op_malloc(nargs * sizeof(int));
OP_plans[ip].maps = (op_map *)op_malloc(nargs * sizeof(op_map));
OP_plans[ip].accs = (op_access *)op_malloc(nargs * sizeof(op_access));
OP_plans[ip].inds_staged =
(op_access *)op_malloc(ninds_staged * sizeof(op_access));
OP_plans[ip].nthrcol = (int *)op_malloc(nblocks * sizeof(int));
OP_plans[ip].thrcol = (int *)op_malloc(exec_length * sizeof(int));
OP_plans[ip].col_reord = (int *)op_malloc((exec_length + 16) * sizeof(int));
OP_plans[ip].col_offsets = NULL;
OP_plans[ip].offset = (int *)op_malloc(nblocks * sizeof(int));
OP_plans[ip].ind_maps = (int **)op_malloc(ninds_staged * sizeof(int *));
OP_plans[ip].ind_offs =
(int *)op_malloc(nblocks * ninds_staged * sizeof(int));
OP_plans[ip].ind_sizes =
(int *)op_malloc(nblocks * ninds_staged * sizeof(int));
OP_plans[ip].nindirect = (int *)op_calloc(ninds, sizeof(int));
OP_plans[ip].loc_maps = (short **)op_malloc(nargs * sizeof(short *));
OP_plans[ip].nelems = (int *)op_malloc(nblocks * sizeof(int));
OP_plans[ip].ncolblk =
(int *)op_calloc(exec_length, sizeof(int)); /* max possibly needed */
OP_plans[ip].blkmap = (int *)op_calloc(nblocks, sizeof(int));
int *offsets = (int *)op_malloc((ninds_staged + 1) * sizeof(int));
offsets[0] = 0;
for (int m = 0; m < ninds_staged; m++) {
int count = 0;
for (int m2 = 0; m2 < nargs; m2++)
if (inds_staged[m2] == m)
count++;
offsets[m + 1] = offsets[m] + count;
}
OP_plans[ip].ind_map =
(int *)op_malloc(offsets[ninds_staged] * exec_length * sizeof(int));
for (int m = 0; m < ninds_staged; m++) {
OP_plans[ip].ind_maps[m] = &OP_plans[ip].ind_map[exec_length * offsets[m]];
}
free(offsets);
int counter = 0;
for (int m = 0; m < nargs; m++) {
if (inds_staged[m] >= 0)
counter++;
else
OP_plans[ip].loc_maps[m] = NULL;
OP_plans[ip].dats[m] = args[m].dat;
OP_plans[ip].idxs[m] = args[m].idx;
OP_plans[ip].optflags[m] = args[m].opt;
OP_plans[ip].maps[m] = args[m].map;
OP_plans[ip].accs[m] = args[m].acc;
}
OP_plans[ip].loc_map =
(short *)op_malloc(counter * exec_length * sizeof(short));
counter = 0;
for (int m = 0; m < nargs; m++) {
if (inds_staged[m] >= 0) {
OP_plans[ip].loc_maps[m] = &OP_plans[ip].loc_map[exec_length * (counter)];
counter++;
}
}
OP_plans[ip].name = name;
OP_plans[ip].set = set;
OP_plans[ip].nargs = nargs;
OP_plans[ip].ninds = ninds;
OP_plans[ip].ninds_staged = ninds_staged;
OP_plans[ip].part_size = part_size;
OP_plans[ip].nblocks = nblocks;
OP_plans[ip].ncolors_core = 0;
OP_plans[ip].ncolors_owned = 0;
OP_plans[ip].count = 1;
OP_plans[ip].inds_staged = inds_staged;
OP_plan_index++;
/* define aliases */
op_dat *dats = OP_plans[ip].dats;
int *idxs = OP_plans[ip].idxs;
op_map *maps = OP_plans[ip].maps;
op_access *accs = OP_plans[ip].accs;
int *offset = OP_plans[ip].offset;
int *nelems = OP_plans[ip].nelems;
int **ind_maps = OP_plans[ip].ind_maps;
int *ind_offs = OP_plans[ip].ind_offs;
int *ind_sizes = OP_plans[ip].ind_sizes;
int *nindirect = OP_plans[ip].nindirect;
/* allocate working arrays */
uint **work;
work = (uint **)op_malloc(ninds * sizeof(uint *));
for (int m = 0; m < ninds; m++) {
int m2 = 0;
while (inds[m2] != m)
m2++;
if (args[m2].opt == 0) {
work[m] = NULL;
continue;
}
int to_size = (maps[m2]->to)->exec_size + (maps[m2]->to)->nonexec_size +
(maps[m2]->to)->size;
work[m] = (uint *)op_malloc(to_size * sizeof(uint));
}
int *work2;
work2 =
(int *)op_malloc(nargs * bsize * sizeof(int)); /* max possibly needed */
/* process set one block at a time */
float total_colors = 0;
prev_offset = 0;
next_offset = 0;
for (int b = 0; b < nblocks; b++) {
prev_offset = next_offset;
if (prev_offset + bsize >= set->core_size && prev_offset < set->core_size) {
next_offset = set->core_size;
} else if (prev_offset + bsize >= set->size && prev_offset < set->size &&
indirect_reduce) {
next_offset = set->size;
} else if (prev_offset + bsize >= exec_length &&
prev_offset < exec_length) {
next_offset = exec_length;
} else {
next_offset = prev_offset + bsize;
}
if (staging == OP_COLOR2) {
prev_offset = 0;
next_offset = exec_length;
};
int bs = next_offset - prev_offset;
offset[b] = prev_offset; /* offset for block */
nelems[b] = bs; /* size of block */
/* loop over indirection sets */
for (int m = 0; m < ninds; m++) {
int m2 = 0;
while (inds[m2] != m)
m2++;
int m3 = inds_staged[m2];
if (m3 < 0)
continue;
if (args[m2].opt == 0) {
if (b == 0) {
ind_offs[m3 + b * ninds_staged] = 0;
ind_sizes[m3 + b * ninds_staged] = 0;
} else {
ind_offs[m3 + b * ninds_staged] =
ind_offs[m3 + (b - 1) * ninds_staged];
ind_sizes[m3 + b * ninds_staged] = 0;
}
continue;
}
/* build the list of elements indirectly referenced in this block */
int ne = 0; /* number of elements */
for (int m2 = 0; m2 < nargs; m2++) {
if (inds[m2] == m) {
for (int e = prev_offset; e < next_offset; e++)
work2[ne++] = maps[m2]->map[idxs[m2] + e * maps[m2]->dim];
}
}
/* sort them, then eliminate duplicates */
qsort(work2, ne, sizeof(int), comp);
int nde = 0;
int p = 0;
while (p < ne) {
work2[nde] = work2[p];
while (p < ne && work2[p] == work2[nde])
p++;
nde++;
}
ne = nde; /* number of distinct elements */
/*
if (OP_diags > 5) { printf(" indirection set %d: ",m); for (int e=0;
e<ne; e++) printf("
%d",work2[e]); printf(" \n"); } */
/* store mapping and renumbered mappings in execution plan */
for (int e = 0; e < ne; e++) {
ind_maps[m3][nindirect[m]++] = work2[e];
work[m][work2[e]] = e; // inverse mapping
}
for (int m2 = 0; m2 < nargs; m2++) {
if (inds[m2] == m) {
for (int e = prev_offset; e < next_offset; e++)
OP_plans[ip].loc_maps[m2][e] =
(short)(work[m][maps[m2]->map[idxs[m2] + e * maps[m2]->dim]]);
}
}
if (b == 0) {
ind_offs[m3 + b * ninds_staged] = 0;
ind_sizes[m3 + b * ninds_staged] = nindirect[m];
} else {
ind_offs[m3 + b * ninds_staged] =
ind_offs[m3 + (b - 1) * ninds_staged] +
ind_sizes[m3 + (b - 1) * ninds_staged];
ind_sizes[m3 + b * ninds_staged] =
nindirect[m] - ind_offs[m3 + b * ninds_staged];
}
}
/* now colour main set elements */
for (int e = prev_offset; e < next_offset; e++)
OP_plans[ip].thrcol[e] = -1;
int repeat = 1;
int ncolor = 0;
int ncolors = 0;
while (repeat) {
repeat = 0;
for (int m = 0; m < nargs; m++) {
if (inds[m] >= 0 && args[m].opt)
for (int e = prev_offset; e < next_offset; e++)
work[inds[m]][maps[m]->map[idxs[m] + e * maps[m]->dim]] =
0; /* zero out color array */
}
for (int e = prev_offset; e < next_offset; e++) {
if (OP_plans[ip].thrcol[e] == -1) {
int mask = 0;
if (staging == OP_COLOR2 && halo_exchange && e >= set->core_size &&
ncolor == 0)
mask = 1;
for (int m = 0; m < nargs; m++)
if (inds[m] >= 0 && (accs[m] == OP_INC || accs[m] == OP_RW) &&
args[m].opt)
mask |=
work[inds[m]]
[maps[m]->map[idxs[m] +
e * maps[m]->dim]]; /* set bits of mask */
int color = ffs(~mask) - 1; /* find first bit not set */
if (color == -1) { /* run out of colors on this pass */
repeat = 1;
} else {
OP_plans[ip].thrcol[e] = ncolor + color;
mask = 1 << color;
ncolors = MAX(ncolors, ncolor + color + 1);
for (int m = 0; m < nargs; m++)
if (inds[m] >= 0 && (accs[m] == OP_INC || accs[m] == OP_RW) &&
args[m].opt)
work[inds[m]][maps[m]->map[idxs[m] + e * maps[m]->dim]] |=
mask; /* set color bit */
}
}
}
ncolor += 32; /* increment base level */
}
OP_plans[ip].nthrcol[b] =
ncolors; /* number of thread colors in this block */
total_colors += ncolors;
// if(ncolors>1) printf(" number of colors in this block = %d \n",ncolors);
}
/* create element permutation by color */
if (staging == OP_STAGE_PERMUTE || staging == OP_COLOR2) {
int size_of_col_offsets = 0;
for (int b = 0; b < nblocks; b++) {
size_of_col_offsets += OP_plans[ip].nthrcol[b] + 1;
}
// allocate
OP_plans[ip].col_offsets = (int **)op_malloc(nblocks * sizeof(int *));
int *col_offsets = (int *)op_malloc(size_of_col_offsets * sizeof(int *));
size_of_col_offsets = 0;
op_keyvalue *kv = (op_keyvalue *)op_malloc(bsize * sizeof(op_keyvalue));
for (int b = 0; b < nblocks; b++) {
int ncolor = OP_plans[ip].nthrcol[b];
for (int e = 0; e < nelems[b]; e++) {
kv[e].key = OP_plans[ip].thrcol[offset[b] + e];
kv[e].value = e;
}
qsort(kv, nelems[b], sizeof(op_keyvalue), comp2);
OP_plans[ip].col_offsets[b] = col_offsets + size_of_col_offsets;
OP_plans[ip].col_offsets[b][0] = 0;
size_of_col_offsets += (ncolor + 1);
// Set up permutation and pointers to beginning of each color
ncolor = 0;
for (int e = 0; e < nelems[b]; e++) {
OP_plans[ip].thrcol[offset[b] + e] = kv[e].key;
OP_plans[ip].col_reord[offset[b] + e] = kv[e].value;
if (e > 0)
if (kv[e].key > kv[e - 1].key) {
ncolor++;
OP_plans[ip].col_offsets[b][ncolor] = e;
}
}
OP_plans[ip].col_offsets[b][ncolor + 1] = nelems[b];
}
for (int i = exec_length; i < exec_length + 16; i++)
OP_plans[ip].col_reord[i] = 0;
}
/* color the blocks, after initialising colors to 0 */
int *blk_col;
blk_col = (int *)op_malloc(nblocks * sizeof(int));
for (int b = 0; b < nblocks; b++)
blk_col[b] = -1;
int repeat = 1;
int ncolor = 0;
int ncolors = 0;
while (repeat) {
repeat = 0;
for (int m = 0; m < nargs; m++) {
if (inds[m] >= 0 && args[m].opt) {
int to_size = (maps[m]->to)->exec_size + (maps[m]->to)->nonexec_size +
(maps[m]->to)->size;
for (int e = 0; e < to_size; e++)
work[inds[m]][e] = 0; // zero out color arrays
}
}
prev_offset = 0;
next_offset = 0;
for (int b = 0; b < nblocks; b++) {
prev_offset = next_offset;
if (prev_offset + bsize >= set->core_size &&
prev_offset < set->core_size) {
next_offset = set->core_size;
} else if (prev_offset + bsize >= set->size && prev_offset < set->size &&
indirect_reduce) {
next_offset = set->size;
} else if (prev_offset + bsize >= exec_length &&
prev_offset < exec_length) {
next_offset = exec_length;
} else {
next_offset = prev_offset + bsize;
}
if (blk_col[b] == -1) { // color not yet assigned to block
uint mask = 0;
if (next_offset > set->core_size) { // should not use block colors from
// the core set when doing the
// non_core ones
if (prev_offset <= set->core_size)
OP_plans[ip].ncolors_core = ncolors;
for (int shifter = 0; shifter < OP_plans[ip].ncolors_core; shifter++)
mask |= 1 << shifter;
if (prev_offset == set->size && indirect_reduce)
OP_plans[ip].ncolors_owned = ncolors;
for (int shifter = OP_plans[ip].ncolors_core;
indirect_reduce && shifter < OP_plans[ip].ncolors_owned;
shifter++)
mask |= 1 << shifter;
}
for (int m = 0; m < nargs; m++) {
if (inds[m] >= 0 && (accs[m] == OP_INC || accs[m] == OP_RW) &&
args[m].opt)
for (int e = prev_offset; e < next_offset; e++)
mask |= work[inds[m]]
[maps[m]->map[idxs[m] + e * maps[m]->dim]]; // set
// bits of
// mask
}
int color = ffs(~mask) - 1; // find first bit not set
if (color == -1) { // run out of colors on this pass
repeat = 1;
} else {
blk_col[b] = ncolor + color;
mask = 1 << color;
ncolors = MAX(ncolors, ncolor + color + 1);
for (int m = 0; m < nargs; m++) {
if (inds[m] >= 0 && (accs[m] == OP_INC || accs[m] == OP_RW) &&
args[m].opt)
for (int e = prev_offset; e < next_offset; e++)
work[inds[m]][maps[m]->map[idxs[m] + e * maps[m]->dim]] |= mask;
}
}
}
}
ncolor += 32; // increment base level
}
/* store block mapping and number of blocks per color */
if (indirect_reduce && OP_plans[ip].ncolors_owned == 0)
OP_plans[ip].ncolors_owned =
ncolors; // no MPI, so get the reduction arrays after everyting is done
OP_plans[ip].ncolors = ncolors;
if (staging == OP_COLOR2)
OP_plans[ip].ncolors = OP_plans[ip].nthrcol[0];
/*for(int col = 0; col = OP_plans[ip].ncolors;col++) //should initialize to
zero because op_calloc returns garbage!!
{
OP_plans[ip].ncolblk[col] = 0;
}*/
for (int b = 0; b < nblocks; b++)
OP_plans[ip].ncolblk[blk_col[b]]++; // number of blocks of each color
for (int c = 1; c < ncolors; c++)
OP_plans[ip].ncolblk[c] += OP_plans[ip].ncolblk[c - 1]; // cumsum
for (int c = 0; c < ncolors; c++)
work2[c] = 0;
for (int b = 0; b < nblocks; b++) {
int c = blk_col[b];
int b2 = work2[c]; // number of preceding blocks of this color
if (c > 0)
b2 += OP_plans[ip].ncolblk[c - 1]; // plus previous colors
OP_plans[ip].blkmap[b2] = b;
work2[c]++; // increment counter
}
for (int c = ncolors - 1; c > 0; c--)
OP_plans[ip].ncolblk[c] -= OP_plans[ip].ncolblk[c - 1]; // undo cumsum
/* reorder blocks by color? */
/* work out shared memory requirements */
OP_plans[ip].nsharedCol = (int *)op_malloc(ncolors * sizeof(int));
float total_shared = 0;
for (int col = 0; col < ncolors; col++) {
OP_plans[ip].nsharedCol[col] = 0;
for (int b = 0; b < nblocks; b++) {
if (blk_col[b] == col) {
int nbytes = 0;
for (int m = 0; m < ninds_staged; m++) {
int m2 = 0;
while (inds_staged[m2] != m)
m2++;
if (args[m2].opt == 0)
continue;
nbytes +=
ROUND_UP_64(ind_sizes[m + b * ninds_staged] * dats[m2]->size);
}
OP_plans[ip].nsharedCol[col] =
MAX(OP_plans[ip].nsharedCol[col], nbytes);
total_shared += nbytes;
}
}
}
OP_plans[ip].nshared = 0;
total_shared = 0;
for (int b = 0; b < nblocks; b++) {
int nbytes = 0;
for (int m = 0; m < ninds_staged; m++) {
int m2 = 0;
while (inds_staged[m2] != m)
m2++;
if (args[m2].opt == 0)
continue;
nbytes += ROUND_UP_64(ind_sizes[m + b * ninds_staged] * dats[m2]->size);
}
OP_plans[ip].nshared = MAX(OP_plans[ip].nshared, nbytes);
total_shared += nbytes;
}
/* work out total bandwidth requirements */
OP_plans[ip].transfer = 0;
OP_plans[ip].transfer2 = 0;
float transfer3 = 0;
if (staging != OP_COLOR2 && staging != OP_STAGE_INC) {
for (int b = 0; b < nblocks; b++) {
for (int m = 0; m < nargs; m++) // for each argument
{
if (args[m].opt) {
if (inds[m] < 0) // if it is directly addressed
{
float fac = 2.0f;
if (accs[m] == OP_READ ||
accs[m] == OP_WRITE) // if you only read or write it
fac = 1.0f;
if (dats[m] != NULL) {
OP_plans[ip].transfer +=
fac * nelems[b] * dats[m]->size; // cost of reading it all
OP_plans[ip].transfer2 += fac * nelems[b] * dats[m]->size;
transfer3 += fac * nelems[b] * dats[m]->size;
}
} else // if it is indirectly addressed: cost of reading the pointer
// to it
{
OP_plans[ip].transfer += nelems[b] * sizeof(short);
OP_plans[ip].transfer2 += nelems[b] * sizeof(short);
transfer3 += nelems[b] * sizeof(short);
}
}
}
for (int m = 0; m < ninds; m++) // for each indirect mapping
{
int m2 = 0;
while (inds[m2] != m) // find the first argument that uses this mapping
m2++;
if (args[m2].opt == 0)
continue;
float fac = 2.0f;
if (accs[m2] == OP_READ || accs[m2] == OP_WRITE) // only read it
fac = 1.0f;
if (staging == OP_STAGE_INC && accs[m2] != OP_INC) {
OP_plans[ip].transfer += 1;
OP_plans[ip].transfer2 += 1;
continue;
}
OP_plans[ip].transfer +=
fac * ind_sizes[m + b * ninds] *
dats[m2]->size; // simply read all data one by one
/* work out how many cache lines are used by indirect addressing */
int i_map, l_new, l_old;
int e0 = ind_offs[m + b * ninds]; // where it starts
int e1 = e0 + ind_sizes[m + b * ninds]; // where it ends
l_old = -1;
for (int e = e0; e < e1;
e++) // iterate through every indirectly accessed data element
{
i_map = ind_maps[m][e]; // the pointer to the data element
l_new = (i_map * dats[m2]->size) /
OP_cache_line_size; // which cache line it is on (full size,
// dim*sizeof(type))
if (l_new > l_old) // if it is on a further cache line (that is not
// yet loaded, - i_map is ordered)
OP_plans[ip].transfer2 +=
fac * OP_cache_line_size; // load the cache line
l_old = l_new;
l_new = ((i_map + 1) * dats[m2]->size - 1) /
OP_cache_line_size; // the last byte of the data
OP_plans[ip].transfer2 += fac * (l_new - l_old) *
OP_cache_line_size; // again, if not loaded,
// load it (can be
// multiple cache lines)
l_old = l_new;
}
l_old = -1;
for (int e = e0; e < e1; e++) {
i_map = ind_maps[m][e]; // pointer to the data element
l_new = (i_map * dats[m2]->size) /
(dats[m2]->dim * OP_cache_line_size); // which cache line the
// first dimension of
// the data is on
if (l_new > l_old)
transfer3 +=
fac * dats[m2]->dim *
OP_cache_line_size; // if not loaded yet, load all cache lines
l_old = l_new;
l_new =
((i_map + 1) * dats[m2]->size - 1) /
(dats[m2]->dim * OP_cache_line_size); // primitve type's last byte
transfer3 += fac * (l_new - l_old) * dats[m2]->dim *
OP_cache_line_size; // load it
l_old = l_new;
}
/* also include mappings to load/store data */
fac = 1.0f;
if (accs[m2] == OP_RW)
fac = 2.0f;
OP_plans[ip].transfer += fac * ind_sizes[m + b * ninds] * sizeof(int);
OP_plans[ip].transfer2 += fac * ind_sizes[m + b * ninds] * sizeof(int);
transfer3 += fac * ind_sizes[m + b * ninds] * sizeof(int);
}
}
}
/* print out useful information */
if (OP_diags > 1) {
printf(" number of blocks = %d \n", nblocks);
printf(" number of block colors = %d \n", OP_plans[ip].ncolors);
printf(" maximum block size = %d \n", bsize);
printf(" average thread colors = %.2f \n", total_colors / nblocks);
printf(" shared memory required = ");
for (int i = 0; i < ncolors - 1; i++)
printf(" %.2f KB,", OP_plans[ip].nsharedCol[i] / 1024.0f);
printf(" %.2f KB\n", OP_plans[ip].nsharedCol[ncolors - 1] / 1024.0f);
printf(" average data reuse = %.2f \n",
maxbytes * (exec_length / total_shared));
printf(" data transfer (used) = %.2f MB \n",
OP_plans[ip].transfer / (1024.0f * 1024.0f));
printf(" data transfer (total) = %.2f MB \n",
OP_plans[ip].transfer2 / (1024.0f * 1024.0f));
printf(" SoA/AoS transfer ratio = %.2f \n\n",
transfer3 / OP_plans[ip].transfer2);
}
/* validate plan info */
op_plan_check(OP_plans[ip], ninds_staged, inds_staged);
/* free work arrays */
for (int m = 0; m < ninds; m++)
free(work[m]);
free(work);
free(work2);
free(blk_col);
free(inds_to_inds_staged);
free(invinds_staged);
op_timers_core(&cpu_t2, &wall_t2);
for (int i = 0; i < OP_kern_max; i++) {
if (strcmp(name, OP_kernels[i].name) == 0) {
OP_kernels[i].plan_time += wall_t2 - wall_t1;
break;
}
}
/* return pointer to plan */
OP_plan_time += wall_t2 - wall_t1;
return &(OP_plans[ip]);
}
op_plan * op_plan_core(char const *name, op_set set, int part_size,
int nargs, op_arg *args, int ninds, int *inds){
// first look for an existing execution plan
int ip=0, match=0;
while (match==0 && ip<OP_plan_index) {
if ( (strcmp(name, OP_plans[ip].name) == 0)
&& (set == OP_plans[ip].set )
&& (nargs == OP_plans[ip].nargs )
&& (ninds == OP_plans[ip].ninds )
&& (part_size == OP_plans[ip].part_size ) ) {
match = 1;
for (int m=0; m<nargs; m++) {
match = match && (args[m].dat == OP_plans[ip].dats[m])
&& (args[m].map == OP_plans[ip].maps[m])
&& (args[m].idx == OP_plans[ip].idxs[m])
&& (args[m].acc == OP_plans[ip].accs[m]);
}
}
ip++;
}
if (match) {
ip--;
if (OP_diags > 3) printf(" old execution plan #%d\n",ip);
OP_plans[ip].count++;
return &(OP_plans[ip]);
}
else {
if (OP_diags > 1) printf(" new execution plan #%d for kernel %s\n",ip,name);
}
// work out worst case shared memory requirement per element
int maxbytes = 0;
for (int m=0; m<nargs; m++) {
if (inds[m]>=0) maxbytes += args[m].dat->size;
}
// set blocksize and number of blocks;
// adaptive size based on 48kB of shared memory
int bsize = part_size; // blocksize
if (bsize==0) bsize = (48*1024/(64*maxbytes))*64;
int nblocks = (set->size-1)/bsize + 1;
// enlarge OP_plans array if needed
if (ip==OP_plan_max) {
// printf("allocating more memory for OP_plans \n");
OP_plan_max += 10;
OP_plans = (op_plan *) realloc(OP_plans,OP_plan_max*sizeof(op_plan));
if (OP_plans==NULL) {
printf(" op_plan error -- error reallocating memory for OP_plans\n");
exit(-1);
}
}
// allocate memory for new execution plan and store input arguments
OP_plans[ip].dats = (op_dat *)malloc(nargs*sizeof(op_dat));
OP_plans[ip].idxs = (int *)malloc(nargs*sizeof(int));
OP_plans[ip].maps = (op_map *)malloc(nargs*sizeof(op_map));
OP_plans[ip].accs = (op_access *)malloc(nargs*sizeof(op_access));
OP_plans[ip].nthrcol = (int *)malloc(nblocks*sizeof(int));
OP_plans[ip].thrcol = (int *)malloc(set->size*sizeof(int));
OP_plans[ip].offset = (int *)malloc(nblocks*sizeof(int));
OP_plans[ip].ind_maps = (int **)malloc(ninds*sizeof(int *));
OP_plans[ip].ind_offs = (int *)malloc(nblocks*ninds*sizeof(int));
OP_plans[ip].ind_sizes = (int *)malloc(nblocks*ninds*sizeof(int));
OP_plans[ip].nindirect = (int *)calloc(ninds,sizeof(int));
OP_plans[ip].loc_maps = (short **)malloc(nargs*sizeof(short *));
OP_plans[ip].nelems = (int *)malloc(nblocks*sizeof(int));
OP_plans[ip].ncolblk = (int *)calloc(set->size,sizeof(int)); // max possibly needed
OP_plans[ip].blkmap = (int *)calloc(nblocks,sizeof(int));
for (int m=0; m<ninds; m++) {
int count = 0;
for (int m2=0; m2<nargs; m2++)
if (inds[m2]==m) count++;
OP_plans[ip].ind_maps[m] = (int *)malloc(count*set->size*sizeof(int));
}
for (int m=0; m<nargs; m++) {
if (inds[m]>=0)
OP_plans[ip].loc_maps[m] = (short *)malloc(set->size*sizeof(short));
else
OP_plans[ip].loc_maps[m] = NULL;
OP_plans[ip].dats[m] = args[m].dat;
OP_plans[ip].idxs[m] = args[m].idx;
OP_plans[ip].maps[m] = args[m].map;
OP_plans[ip].accs[m] = args[m].acc;
}
OP_plans[ip].name = name;
OP_plans[ip].set = set;
OP_plans[ip].nargs = nargs;
OP_plans[ip].ninds = ninds;
OP_plans[ip].part_size = part_size;
OP_plans[ip].nblocks = nblocks;
OP_plans[ip].count = 1;
OP_plan_index++;
// define aliases
op_dat *dats = OP_plans[ip].dats;
int *idxs = OP_plans[ip].idxs;
op_map *maps = OP_plans[ip].maps;
op_access *accs = OP_plans[ip].accs;
int *offset = OP_plans[ip].offset;
int *nelems = OP_plans[ip].nelems;
int **ind_maps = OP_plans[ip].ind_maps;
int *ind_offs = OP_plans[ip].ind_offs;
int *ind_sizes = OP_plans[ip].ind_sizes;
int *nindirect = OP_plans[ip].nindirect;
// allocate working arrays
uint **work;
work = (uint **)malloc(ninds*sizeof(uint *));
for (int m=0; m<ninds; m++) {
int m2 = 0;
while(inds[m2]!=m) m2++;
work[m] = (uint *)malloc((maps[m2]->to)->size*sizeof(uint));
}
int *work2;
work2 = (int *)malloc(nargs*bsize*sizeof(int)); // max possibly needed
// process set one block at a time
float total_colors = 0;
for (int b=0; b<nblocks; b++) {
int bs = MIN(bsize, set->size - b*bsize);
offset[b] = b*bsize; // offset for block
nelems[b] = bs; // size of block
// loop over indirection sets
for (int m=0; m<ninds; m++) {
// build the list of elements indirectly referenced in this block
int ne = 0; // number of elements
for (int m2=0; m2<nargs; m2++) {
if (inds[m2]==m) {
for (int e=b*bsize; e<b*bsize+bs; e++)
work2[ne++] = maps[m2]->map[idxs[m2]+e*maps[m2]->dim];
}
}
// sort them, then eliminate duplicates
qsort(work2,ne,sizeof(int),comp);
int e = 0;
int p = 0;
while (p<ne) {
work2[e] = work2[p];
while (p<ne && work2[p]==work2[e]) p++;
e++;
}
ne = e; // number of distinct elements
/*
if (OP_diags > 5) {
printf(" indirection set %d: ",m);
for (int e=0; e<ne; e++) printf(" %d",work2[e]);
printf(" \n");
}
*/
// store mapping and renumbered mappings in execution plan
for (int e=0; e<ne; e++) {
ind_maps[m][nindirect[m]++] = work2[e];
work[m][work2[e]] = e; // inverse mapping
}
for (int m2=0; m2<nargs; m2++) {
if (inds[m2]==m) {
for (int e=b*bsize; e<b*bsize+bs; e++)
OP_plans[ip].loc_maps[m2][e] =
work[m][maps[m2]->map[idxs[m2]+e*maps[m2]->dim]];
}
}
if (b==0) {
ind_offs[m+b*ninds] = 0;
ind_sizes[m+b*ninds] = nindirect[m];
}
else {
ind_offs[m+b*ninds] = ind_offs[m+(b-1)*ninds]
+ ind_sizes[m+(b-1)*ninds];
ind_sizes[m+b*ninds] = nindirect[m]
- ind_offs[m+b*ninds];
}
}
// now colour main set elements
for (int e=b*bsize; e<b*bsize+bs; e++) OP_plans[ip].thrcol[e]=-1;
int repeat = 1;
int ncolor = 0;
int ncolors = 0;
while (repeat) {
repeat = 0;
for (int m=0; m<nargs; m++) {
if (inds[m]>=0)
for (int e=b*bsize; e<b*bsize+bs; e++)
work[inds[m]][maps[m]->map[idxs[m]+e*maps[m]->dim]] = 0; // zero out color array
}
for (int e=b*bsize; e<b*bsize+bs; e++) {
if (OP_plans[ip].thrcol[e] == -1) {
int mask = 0;
for (int m=0; m<nargs; m++)
if (inds[m]>=0 && accs[m]==OP_INC)
mask |= work[inds[m]][maps[m]->map[idxs[m]+e*maps[m]->dim]]; // set bits of mask
int color = ffs(~mask) - 1; // find first bit not set
if (color==-1) { // run out of colors on this pass
repeat = 1;
}
else {
OP_plans[ip].thrcol[e] = ncolor+color;
mask = 1 << color;
ncolors = MAX(ncolors, ncolor+color+1);
for (int m=0; m<nargs; m++)
if (inds[m]>=0 && accs[m]==OP_INC)
work[inds[m]][maps[m]->map[idxs[m]+e*maps[m]->dim]] |= mask; // set color bit
}
}
}
ncolor += 32; // increment base level
}
OP_plans[ip].nthrcol[b] = ncolors; // number of thread colors in this block
total_colors += ncolors;
// if(ncolors>1) printf(" number of colors in this block = %d \n",ncolors);
// reorder elements by color?
}
// colour the blocks, after initialising colors to 0
int *blk_col;
blk_col = (int *) malloc(nblocks*sizeof(int));
for (int b=0; b<nblocks; b++) blk_col[b] = -1;
int repeat = 1;
int ncolor = 0;
int ncolors = 0;
while (repeat) {
repeat = 0;
for (int m=0; m<nargs; m++) {
if (inds[m]>=0)
for (int e=0; e<(maps[m]->to)->size; e++)
work[inds[m]][e] = 0; // zero out color arrays
}
for (int b=0; b<nblocks; b++) {
if (blk_col[b] == -1) { // color not yet assigned to block
int bs = MIN(bsize, set->size - b*bsize);
uint mask = 0;
for (int m=0; m<nargs; m++) {
if (inds[m]>=0 && accs[m]==OP_INC)
for (int e=b*bsize; e<b*bsize+bs; e++)
mask |= work[inds[m]][maps[m]->map[idxs[m]+e*maps[m]->dim]]; // set bits of mask
}
int color = ffs(~mask) - 1; // find first bit not set
if (color==-1) { // run out of colors on this pass
repeat = 1;
}
else {
blk_col[b] = ncolor + color;
mask = 1 << color;
ncolors = MAX(ncolors, ncolor+color+1);
for (int m=0; m<nargs; m++) {
if (inds[m]>=0 && accs[m]==OP_INC)
for (int e=b*bsize; e<b*bsize+bs; e++)
work[inds[m]][maps[m]->map[idxs[m]+e*maps[m]->dim]] |= mask;
}
}
}
}
ncolor += 32; // increment base level
}
// store block mapping and number of blocks per color
OP_plans[ip].ncolors = ncolors;
for (int b=0; b<nblocks; b++)
OP_plans[ip].ncolblk[blk_col[b]]++; // number of blocks of each color
for (int c=1; c<ncolors; c++)
OP_plans[ip].ncolblk[c] += OP_plans[ip].ncolblk[c-1]; // cumsum
for (int c=0; c<ncolors; c++) work2[c]=0;
for (int b=0; b<nblocks; b++) {
int c = blk_col[b];
int b2 = work2[c]; // number of preceding blocks of this color
if (c>0) b2 += OP_plans[ip].ncolblk[c-1]; // plus previous colors
OP_plans[ip].blkmap[b2] = b;
work2[c]++; // increment counter
}
for (int c=ncolors-1; c>0; c--)
OP_plans[ip].ncolblk[c] -= OP_plans[ip].ncolblk[c-1]; // undo cumsum
// reorder blocks by color?
// work out shared memory requirements
OP_plans[ip].nshared = 0;
float total_shared = 0;
for (int b=0; b<nblocks; b++) {
int nbytes = 0;
for (int m=0; m<ninds; m++) {
int m2 = 0;
while(inds[m2]!=m) m2++;
nbytes += ROUND_UP(ind_sizes[m+b*ninds]*dats[m2]->size);
}
OP_plans[ip].nshared = MAX(OP_plans[ip].nshared,nbytes);
total_shared += nbytes;
}
// work out total bandwidth requirements
OP_plans[ip].transfer = 0;
OP_plans[ip].transfer2 = 0;
float transfer3 = 0;
for (int b=0; b<nblocks; b++) {
for (int m=0; m<nargs; m++) {
if (inds[m]<0) {
float fac = 2.0f;
if (accs[m]==OP_READ) fac = 1.0f;
if (dats[m]!=NULL) {
OP_plans[ip].transfer += fac*nelems[b]*dats[m]->size;
OP_plans[ip].transfer2 += fac*nelems[b]*dats[m]->size;
transfer3 += fac*nelems[b]*dats[m]->size;
}
}
else {
OP_plans[ip].transfer += nelems[b]*sizeof(short);
OP_plans[ip].transfer2 += nelems[b]*sizeof(short);
transfer3 += nelems[b]*sizeof(short);
}
}
for (int m=0; m<ninds; m++) {
int m2 = 0;
while(inds[m2]!=m) m2++;
float fac = 2.0f;
if (accs[m2]==OP_READ) fac = 1.0f;
OP_plans[ip].transfer +=
fac*ind_sizes[m+b*ninds]*dats[m2]->size;
// work out how many cache lines are used by indirect addressing
int i_map, l_new, l_old;
int e0 = ind_offs[m+b*ninds];
int e1 = e0 + ind_sizes[m+b*ninds];
l_old = -1;
for (int e=e0; e<e1; e++) {
i_map = ind_maps[m][e];
l_new = (i_map*dats[m2]->size)/OP_cache_line_size;
if (l_new>l_old) OP_plans[ip].transfer2 += fac*OP_cache_line_size;
l_old = l_new;
l_new = ((i_map+1)*dats[m2]->size-1)/OP_cache_line_size;
OP_plans[ip].transfer2 += fac*(l_new-l_old)*OP_cache_line_size;
l_old = l_new;
}
l_old = -1;
for (int e=e0; e<e1; e++) {
i_map = ind_maps[m][e];
l_new = (i_map*dats[m2]->size)/(dats[m2]->dim*OP_cache_line_size);
if (l_new>l_old) transfer3 += fac*dats[m2]->dim*OP_cache_line_size;
l_old = l_new;
l_new = ((i_map+1)*dats[m2]->size-1)/(dats[m2]->dim*OP_cache_line_size);
transfer3 += fac*(l_new-l_old)*dats[m2]->dim*OP_cache_line_size;
l_old = l_new;
}
// also include mappings to load/store data
fac = 1.0f;
if (accs[m2]==OP_RW) fac = 2.0f;
OP_plans[ip].transfer += fac*ind_sizes[m+b*ninds]*sizeof(int);
OP_plans[ip].transfer2 += fac*ind_sizes[m+b*ninds]*sizeof(int);
transfer3 += fac*ind_sizes[m+b*ninds]*sizeof(int);
}
}
// print out useful information
if (OP_diags>1) {
printf(" number of blocks = %d \n",nblocks);
printf(" number of block colors = %d \n",OP_plans[ip].ncolors);
printf(" maximum block size = %d \n",bsize);
printf(" average thread colors = %.2f \n",total_colors/nblocks);
printf(" shared memory required = %.2f KB \n",OP_plans[ip].nshared/1024.0f);
printf(" average data reuse = %.2f \n",maxbytes*(set->size/total_shared));
printf(" data transfer (used) = %.2f MB \n",
OP_plans[ip].transfer/(1024.0f*1024.0f));
printf(" data transfer (total) = %.2f MB \n",
OP_plans[ip].transfer2/(1024.0f*1024.0f));
printf(" SoA/AoS transfer ratio = %.2f \n\n",
transfer3 / OP_plans[ip].transfer2);
}
// validate plan info
op_plan_check(OP_plans[ip],ninds,inds);
// free work arrays
for (int m=0; m<ninds; m++) free(work[m]);
free(work);
free(work2);
free(blk_col);
// return pointer to plan
return &(OP_plans[ip]);
}
op_plan *op_plan_core(char const *name, op_set set, int set_offset, int part_size,
int nargs, op_arg *args, int ninds, int *inds )
{
//set exec length
int exec_length = set->size;
for(int i = 0; i< nargs; i++)
{
if(args[i].idx != -1 && args[i].acc != OP_READ )
{
exec_length += set->exec_size;
break;
}
}
/* first look for an existing execution plan */
int ip = 0, match = 0;
while ( match == 0 && ip < OP_plan_index )
{
if ( ( strcmp ( name, OP_plans[ip].name ) == 0 )
&& ( set == OP_plans[ip].set )
&& ( set_offset == OP_plans[ip].set_offset )
&& ( nargs == OP_plans[ip].nargs )
&& ( ninds == OP_plans[ip].ninds )
&& ( part_size == OP_plans[ip].part_size ) )
{
match = op_plan_args_match ( nargs, args, OP_plans[ip] );
}
ip++;
}
if ( match )
{
ip--;
if ( OP_diags > 3 )
printf ( " old execution plan #%d\n", ip );
OP_plans[ip].count++;
return &( OP_plans[ip] );
}
else
{
if ( OP_diags > 1 )
printf ( " new execution plan #%d for kernel %s\n", ip, name );
}
/* work out worst case shared memory requirement per element */
int maxbytes = 0;
for ( int m = 0; m < nargs; m++ )
{
if ( inds[m] >= 0 )
maxbytes += args[m].dat->size;
}
/* set blocksize and number of blocks; adaptive size based on 48kB of shared memory */
int bsize = part_size; // blocksize
if ( bsize == 0 )
bsize = ( 48 * 1024 / ( 64 * maxbytes ) ) * 64;
int nblocks = ( exec_length - 1 ) / bsize + 1;
/* enlarge OP_plans array if needed */
if ( ip == OP_plan_max )
{
//printf("allocating more memory for OP_plans %d\n", OP_plan_max);
OP_plan_max += 10;
OP_plans = ( op_plan * ) realloc ( OP_plans, OP_plan_max * sizeof ( op_plan ) );
if ( OP_plans == NULL )
{
printf ( " op_plan error -- error reallocating memory for OP_plans\n" );
exit ( -1 );
}
}
/* allocate memory for new execution plan and store input arguments */
OP_plans[ip].dats = ( op_dat * ) malloc ( nargs * sizeof ( op_dat ) );
OP_plans[ip].mats = ( op_mat * ) malloc ( nargs * sizeof ( op_mat ) );
OP_plans[ip].idxs = ( int * ) malloc ( nargs * sizeof ( int ) );
OP_plans[ip].idxs2 = ( int * ) malloc ( nargs * sizeof ( int ) );
OP_plans[ip].maps = ( op_map * ) malloc ( nargs * sizeof ( op_map ) );
OP_plans[ip].maps2 = ( op_map * ) malloc ( nargs * sizeof ( op_map ) );
OP_plans[ip].accs = ( op_access * ) malloc ( nargs * sizeof ( op_access ) );
OP_plans[ip].nthrcol = ( int * ) malloc ( nblocks * sizeof ( int ) );
OP_plans[ip].thrcol = ( int * ) malloc ( exec_length * sizeof ( int ) );
OP_plans[ip].offset = ( int * ) malloc ( nblocks * sizeof ( int ) );
OP_plans[ip].ind_maps = ( int ** ) malloc ( ninds * sizeof ( int * ) );
OP_plans[ip].ind_offs = ( int * ) malloc ( nblocks * ninds * sizeof ( int ) );
OP_plans[ip].ind_sizes = ( int * ) malloc ( nblocks * ninds * sizeof ( int ) );
OP_plans[ip].nindirect = ( int * ) calloc ( ninds, sizeof ( int ) );
OP_plans[ip].loc_maps = ( short ** ) malloc ( nargs * sizeof ( short * ) );
OP_plans[ip].nelems = ( int * ) malloc ( nblocks * sizeof ( int ) );
OP_plans[ip].ncolblk = ( int * ) calloc ( exec_length, sizeof ( int ) ); /* max possibly needed */
OP_plans[ip].blkmap = ( int * ) calloc ( nblocks, sizeof ( int ) );
for ( int m = 0; m < ninds; m++ )
{
int count = 0;
for ( int m2 = 0; m2 < nargs; m2++ )
if ( inds[m2] == m )
count++;
OP_plans[ip].ind_maps[m] = ( int * ) malloc ( count * exec_length * sizeof ( int ) );
}
for ( int m = 0; m < nargs; m++ )
{
if ( inds[m] >= 0 )
OP_plans[ip].loc_maps[m] = ( short * ) malloc ( exec_length * sizeof ( short ) );
else
OP_plans[ip].loc_maps[m] = NULL;
OP_plans[ip].dats[m] = args[m].dat;
OP_plans[ip].idxs[m] = args[m].idx;
OP_plans[ip].maps[m] = args[m].map;
OP_plans[ip].accs[m] = args[m].acc;
}
OP_plans[ip].name = name;
OP_plans[ip].set = set;
OP_plans[ip].set_offset = set_offset;
OP_plans[ip].nargs = nargs;
OP_plans[ip].ninds = ninds;
OP_plans[ip].part_size = part_size;
OP_plans[ip].nblocks = nblocks;
OP_plans[ip].count = 1;
OP_plan_index++;
/* define aliases */
op_dat * dats = OP_plans[ip].dats;
int * idxs = OP_plans[ip].idxs;
op_map * maps = OP_plans[ip].maps;
op_access * accs = OP_plans[ip].accs;
int * offset = OP_plans[ip].offset;
int * nelems = OP_plans[ip].nelems;
int ** ind_maps = OP_plans[ip].ind_maps;
int * ind_offs = OP_plans[ip].ind_offs;
int * ind_sizes = OP_plans[ip].ind_sizes;
int * nindirect = OP_plans[ip].nindirect;
/* allocate working arrays */
//printf("ninds = %d\n",ninds);
uint **work;
work = (uint **)malloc(ninds * sizeof(uint *));
for ( int m = 0; m < ninds; m++ )
{
int m2 = 0;
while ( inds[m2] != m )
m2++;
int to_size = (maps[m2]->to)->exec_size + (maps[m2]->to)->nonexec_size + (maps[m2]->to)->size;
work[m] = ( uint * )malloc( to_size * sizeof (uint));
}
int *work2;
work2 = ( int * ) malloc ( nargs * bsize * sizeof ( int ) ); /* max possibly needed */
/* process set one block at a time */
float total_colors = 0;
for ( int b = 0; b < nblocks; b++ )
{
int bs = MIN ( bsize, exec_length - b * bsize );/****/
offset[b] = b * bsize; /* offset for block */
nelems[b] = bs; /* size of block */
/* loop over indirection sets */
for ( int m = 0; m < ninds; m++ )
{
/* build the list of elements indirectly referenced in this block */
int ne = 0;/* number of elements */
for ( int m2 = 0; m2 < nargs; m2++ )
{
if ( inds[m2] == m )
{
for ( int e = set_offset + b * bsize; e < set_offset + b * bsize + bs; e++ )
work2[ne++] = maps[m2]->map[idxs[m2] + e * maps[m2]->dim];
}
}
/* sort them, then eliminate duplicates */
qsort(work2, ne, sizeof(int), comp);
int nde = 0;
int p = 0;
while ( p < ne )
{
work2[nde] = work2[p];
while ( p < ne && work2[p] == work2[nde] )
p++;
nde++;
}
ne = nde; /* number of distinct elements */
/*
if (OP_diags > 5) { printf(" indirection set %d: ",m); for (int e=0; e<ne; e++) printf("
%d",work2[e]); printf(" \n"); } */
/* store mapping and renumbered mappings in execution plan */
for ( int e = 0; e < ne; e++ )
{
ind_maps[m][nindirect[m]++] = work2[e];
work[m][work2[e]] = e; // inverse mapping
}
for ( int m2 = 0; m2 < nargs; m2++ )
{
if ( inds[m2] == m )
{
for ( int e = b * bsize; e < b * bsize + bs; e++ )
OP_plans[ip].loc_maps[m2][e] = (short)(work[m][maps[m2]->map[idxs[m2] + (set_offset + e) * maps[m2]->dim]]);
}
}
if ( b == 0 )
{
ind_offs[m + b * ninds] = 0;
ind_sizes[m + b * ninds] = nindirect[m];
}
else
{
ind_offs[m + b * ninds] = ind_offs[m + ( b - 1 ) * ninds]
+ ind_sizes[m + ( b - 1 ) * ninds];
ind_sizes[m + b * ninds] = nindirect[m] - ind_offs[m + b * ninds];
}
}
/* now colour main set elements */
for ( int e = b * bsize; e < b * bsize + bs; e++ )
OP_plans[ip].thrcol[e] = -1;
int repeat = 1;
int ncolor = 0;
int ncolors = 0;
while ( repeat )
{
repeat = 0;
for ( int m = 0; m < nargs; m++ )
{
if ( inds[m] >= 0 )
for ( int e = b * bsize; e < b * bsize + bs; e++ )
work[inds[m]][maps[m]->map[idxs[m] + e * maps[m]->dim]] = 0; /* zero out color array */
//work[inds[m]][maps[m]->map[idxs[m] + (set_offset+e) * maps[m]->dim]] = 0; /* zero out color array */
}
for ( int e = b * bsize; e < b * bsize + bs; e++ )
{
if ( OP_plans[ip].thrcol[e] == -1 )
{
int mask = 0;
for ( int m = 0; m < nargs; m++ )
if ( inds[m] >= 0 && accs[m] == OP_INC )
mask |= work[inds[m]][maps[m]->map[idxs[m] + (set_offset+e) * maps[m]->dim]]; /* set bits of mask
*/
int color = ffs ( ~mask ) - 1; /* find first bit not set */
if ( color == -1 )
{ /* run out of colors on this pass */
repeat = 1;
}
else
{
OP_plans[ip].thrcol[e] = ncolor + color;
mask = 1 << color;
ncolors = MAX ( ncolors, ncolor + color + 1 );
for ( int m = 0; m < nargs; m++ )
if ( inds[m] >= 0 && accs[m] == OP_INC )
work[inds[m]][maps[m]->map[idxs[m] + (set_offset+e) * maps[m]->dim]] |= mask; /* set color bit */
}
}
}
ncolor += 32; /* increment base level */
}
OP_plans[ip].nthrcol[b] = ncolors; /* number of thread colors in this block */
total_colors += ncolors;
//if(ncolors>1) printf(" number of colors in this block = %d \n",ncolors);
/* reorder elements by color? */
}
/* colour the blocks, after initialising colors to 0 */
int * blk_col;
blk_col = ( int * ) malloc ( nblocks * sizeof ( int ) );
for ( int b = 0; b < nblocks; b++ )
blk_col[b] = -1;
int repeat = 1;
int ncolor = 0;
int ncolors = 0;
while ( repeat )
{
repeat = 0;
for ( int m = 0; m < nargs; m++ )
{
if ( inds[m] >= 0 )
{
int to_size = (maps[m]->to)->exec_size + (maps[m]->to)->nonexec_size + (maps[m]->to)->size;
for ( int e = 0; e < to_size; e++ )
work[inds[m]][e] = 0; // zero out color arrays
}
}
for ( int b = 0; b < nblocks; b++ )
{
if ( blk_col[b] == -1 )
{ // color not yet assigned to block
int bs = MIN( bsize, exec_length - b * bsize );
uint mask = 0;
for ( int m = 0; m < nargs; m++ )
{
if ( inds[m] >= 0 && accs[m] == OP_INC )
for ( int e = b * bsize; e < b * bsize + bs; e++ )
mask |= work[inds[m]][maps[m]->map[idxs[m] +
(set_offset + e) * maps[m]->dim]]; // set bits of mask
}
int color = ffs( ~mask ) - 1; // find first bit not set
if ( color == -1 )
{ //run out of colors on this pass
repeat = 1;
}
else
{
blk_col[b] = ncolor + color;
mask = 1 << color;
ncolors = MAX( ncolors, ncolor + color + 1 );
for ( int m = 0; m < nargs; m++ )
{
if ( inds[m] >= 0 && accs[m] == OP_INC )
for ( int e = b * bsize; e < b * bsize + bs; e++ )
work[inds[m]][maps[m]->map[idxs[m] +
(set_offset+e) * maps[m]->dim]] |= mask;
}
}
}
}
ncolor += 32; // increment base level
}
/* store block mapping and number of blocks per color */
OP_plans[ip].ncolors = ncolors;
/*for(int col = 0; col = OP_plans[ip].ncolors;col++) //should initialize to zero because calloc returns garbage!!
{
OP_plans[ip].ncolblk[col] = 0;
}*/
for ( int b = 0; b < nblocks; b++ )
OP_plans[ip].ncolblk[blk_col[b]]++; // number of blocks of each color
for ( int c = 1; c < ncolors; c++ )
OP_plans[ip].ncolblk[c] += OP_plans[ip].ncolblk[c - 1]; // cumsum
for ( int c = 0; c < ncolors; c++ )
work2[c] = 0;
for ( int b = 0; b < nblocks; b++ )
{
int c = blk_col[b];
int b2 = work2[c]; // number of preceding blocks of this color
if ( c > 0 )
b2 += OP_plans[ip].ncolblk[c - 1]; // plus previous colors
OP_plans[ip].blkmap[b2] = b;
work2[c]++; //increment counter
}
for ( int c = ncolors - 1; c > 0; c-- )
OP_plans[ip].ncolblk[c] -= OP_plans[ip].ncolblk[c - 1]; // undo cumsum
/* reorder blocks by color? */
/* work out shared memory requirements */
OP_plans[ip].nshared = 0;
float total_shared = 0;
for ( int b = 0; b < nblocks; b++ )
{
int nbytes = 0;
for ( int m = 0; m < ninds; m++ )
{
int m2 = 0;
while ( inds[m2] != m )
m2++;
nbytes += ROUND_UP ( ind_sizes[m + b * ninds] * dats[m2]->size );
}
OP_plans[ip].nshared = MAX ( OP_plans[ip].nshared, nbytes );
total_shared += nbytes;
}
/* work out total bandwidth requirements */
OP_plans[ip].transfer = 0;
OP_plans[ip].transfer2 = 0;
float transfer3 = 0;
for ( int b = 0; b < nblocks; b++ )
{
for ( int m = 0; m < nargs; m++ ) //for each argument
{
if ( inds[m] < 0 ) //if it is directly addressed
{
float fac = 2.0f;
if ( accs[m] == OP_READ ) //if you only read it - only write???
fac = 1.0f;
if ( dats[m] != NULL )
{
OP_plans[ip].transfer += fac * nelems[b] * dats[m]->size; //cost of reading it all
OP_plans[ip].transfer2 += fac * nelems[b] * dats[m]->size;
transfer3 += fac * nelems[b] * dats[m]->size;
}
}
else //if it is indirectly addressed: cost of reading the pointer to it
{
OP_plans[ip].transfer += nelems[b] * sizeof ( short );
OP_plans[ip].transfer2 += nelems[b] * sizeof ( short );
transfer3 += nelems[b] * sizeof ( short );
}
}
for ( int m = 0; m < ninds; m++ ) //for each indirect mapping
{
int m2 = 0;
while ( inds[m2] != m ) //find the first argument that uses this mapping
m2++;
float fac = 2.0f;
if ( accs[m2] == OP_READ ) //only read it (write??)
fac = 1.0f;
OP_plans[ip].transfer += fac * ind_sizes[m + b * ninds] * dats[m2]->size; //simply read all data one by one
/* work out how many cache lines are used by indirect addressing */
int i_map, l_new, l_old;
int e0 = ind_offs[m + b * ninds]; //where it starts
int e1 = e0 + ind_sizes[m + b * ninds]; //where it ends
l_old = -1;
for ( int e = e0; e < e1; e++ ) //iterate through every indirectly accessed data element
{
i_map = ind_maps[m][e]; //the pointer to the data element
l_new = ( i_map * dats[m2]->size ) / OP_cache_line_size; //which cache line it is on (full size, dim*sizeof(type))
if ( l_new > l_old ) //if it is on a further cache line (that is not yet loaded, - i_map is ordered)
OP_plans[ip].transfer2 += fac * OP_cache_line_size; //load the cache line
l_old = l_new;
l_new = ( ( i_map + 1 ) * dats[m2]->size - 1 ) / OP_cache_line_size; //the last byte of the data
OP_plans[ip].transfer2 += fac * ( l_new - l_old ) * OP_cache_line_size; //again, if not loaded, load it (can be multiple cache lines)
l_old = l_new;
}
l_old = -1;
for ( int e = e0; e < e1; e++ )
{
i_map = ind_maps[m][e]; //pointer to the data element
l_new = ( i_map * dats[m2]->size ) / ( dats[m2]->dim * OP_cache_line_size ); //which cache line the first dimension of the data is on
if ( l_new > l_old )
transfer3 += fac * dats[m2]->dim * OP_cache_line_size; //if not loaded yet, load all cache lines
l_old = l_new;
l_new = ( ( i_map + 1 ) * dats[m2]->size - 1 ) / ( dats[m2]->dim * OP_cache_line_size ); //primitve type's last byte
transfer3 += fac * ( l_new - l_old ) * dats[m2]->dim * OP_cache_line_size; //load it
l_old = l_new;
}
/* also include mappings to load/store data */
fac = 1.0f;
if ( accs[m2] == OP_RW )
fac = 2.0f;
OP_plans[ip].transfer += fac * ind_sizes[m + b * ninds] * sizeof ( int );
OP_plans[ip].transfer2 += fac * ind_sizes[m + b * ninds] * sizeof ( int );
transfer3 += fac * ind_sizes[m + b * ninds] * sizeof ( int );
}
}
/* print out useful information */
if ( OP_diags > 1 )
{
printf( " number of blocks = %d \n", nblocks );
printf( " number of block colors = %d \n", OP_plans[ip].ncolors );
printf( " maximum block size = %d \n", bsize );
printf( " average thread colors = %.2f \n", total_colors / nblocks );
printf( " shared memory required = %.2f KB \n", OP_plans[ip].nshared / 1024.0f );
printf( " average data reuse = %.2f \n", maxbytes * ( exec_length / total_shared ) );
printf( " data transfer (used) = %.2f MB \n",
OP_plans[ip].transfer / ( 1024.0f * 1024.0f ) );
printf( " data transfer (total) = %.2f MB \n",
OP_plans[ip].transfer2 / ( 1024.0f * 1024.0f ) );
printf( " SoA/AoS transfer ratio = %.2f \n\n", transfer3 / OP_plans[ip].transfer2 );
}
/* validate plan info */
op_plan_check ( OP_plans[ip], ninds, inds );
/* free work arrays */
for ( int m = 0; m < ninds; m++ )
free ( work[m] );
free ( work );
free ( work2 );
free ( blk_col );
/* return pointer to plan */
return &( OP_plans[ip] );
}
