diff options
Diffstat (limited to 'chill/src/loop_tile.cc')
| -rw-r--r-- | chill/src/loop_tile.cc | 630 |
1 files changed, 630 insertions, 0 deletions
diff --git a/chill/src/loop_tile.cc b/chill/src/loop_tile.cc new file mode 100644 index 0000000..ad1d3b7 --- /dev/null +++ b/chill/src/loop_tile.cc @@ -0,0 +1,630 @@ +/* + * loop_tile.cc + * + * Created on: Nov 12, 2012 + * Author: anand + */ + +#include <codegen.h> +#include "loop.hh" +#include "omegatools.hh" +#include "ir_code.hh" +#include "chill_error.hh" + +using namespace omega; + + + + +void Loop::tile(int stmt_num, int level, int tile_size, int outer_level, + TilingMethodType method, int alignment_offset, int alignment_multiple) { + // check for sanity of parameters + if (tile_size < 0) + throw std::invalid_argument("invalid tile size"); + if (alignment_multiple < 1 || alignment_offset < 0) + throw std::invalid_argument("invalid alignment for tile"); + if (stmt_num < 0 || stmt_num >= stmt.size()) + throw std::invalid_argument("invalid statement " + to_string(stmt_num)); + if (level <= 0) + throw std::invalid_argument("invalid loop level " + to_string(level)); + if (level > stmt[stmt_num].loop_level.size()) + throw std::invalid_argument( + "there is no loop level " + to_string(level) + " for statement " + + to_string(stmt_num)); + if (outer_level <= 0 || outer_level > level) + throw std::invalid_argument( + "invalid tile controlling loop level " + + to_string(outer_level)); + + // invalidate saved codegen computation + delete last_compute_cgr_; + last_compute_cgr_ = NULL; + delete last_compute_cg_; + last_compute_cg_ = NULL; + + int dim = 2 * level - 1; + int outer_dim = 2 * outer_level - 1; + std::vector<int> lex = getLexicalOrder(stmt_num); + std::set<int> same_tiled_loop = getStatements(lex, dim - 1); + std::set<int> same_tile_controlling_loop = getStatements(lex, + outer_dim - 1); + + for (std::set<int>::iterator i = same_tiled_loop.begin(); + i != same_tiled_loop.end(); i++) { + for (DependenceGraph::EdgeList::iterator j = + dep.vertex[*i].second.begin(); j != dep.vertex[*i].second.end(); + j++) { + if (same_tiled_loop.find(j->first) != same_tiled_loop.end()) + for (int k = 0; k < j->second.size(); k++) { + DependenceVector dv = j->second[k]; + int dim2 = level - 1; + if ((dv.type != DEP_CONTROL) && (dv.type != DEP_UNKNOWN)) { + while (stmt[*i].loop_level[dim2].type == LoopLevelTile) { + dim2 = stmt[*i].loop_level[dim2].payload - 1; + } + dim2 = stmt[*i].loop_level[dim2].payload; + + if (dv.hasNegative(dim2) && (!dv.quasi)) { + for (int l = outer_level; l < level; l++) + if (stmt[*i].loop_level[l - 1].type + != LoopLevelTile) { + if (dv.isCarried( + stmt[*i].loop_level[l - 1].payload) + && dv.hasPositive( + stmt[*i].loop_level[l - 1].payload)) + throw loop_error( + "loop error: Tiling is illegal, dependence violation!"); + } else { + + int dim3 = l - 1; + while (stmt[*i].loop_level[l - 1].type + != LoopLevelTile) { + dim3 = + stmt[*i].loop_level[l - 1].payload + - 1; + + } + + dim3 = stmt[*i].loop_level[l - 1].payload; + if (dim3 < level - 1) + if (dv.isCarried(dim3) + && dv.hasPositive(dim3)) + throw loop_error( + "loop error: Tiling is illegal, dependence violation!"); + } + } + } + } + } + } + // special case for no tiling + if (tile_size == 0) { + for (std::set<int>::iterator i = same_tile_controlling_loop.begin(); + i != same_tile_controlling_loop.end(); i++) { + Relation r(stmt[*i].xform.n_out(), stmt[*i].xform.n_out() + 2); + F_And *f_root = r.add_and(); + for (int j = 1; j <= 2 * outer_level - 1; j++) { + EQ_Handle h = f_root->add_EQ(); + h.update_coef(r.input_var(j), 1); + h.update_coef(r.output_var(j), -1); + } + EQ_Handle h1 = f_root->add_EQ(); + h1.update_coef(r.output_var(2 * outer_level), 1); + EQ_Handle h2 = f_root->add_EQ(); + h2.update_coef(r.output_var(2 * outer_level + 1), 1); + for (int j = 2 * outer_level; j <= stmt[*i].xform.n_out(); j++) { + EQ_Handle h = f_root->add_EQ(); + h.update_coef(r.input_var(j), 1); + h.update_coef(r.output_var(j + 2), -1); + } + + stmt[*i].xform = Composition(copy(r), stmt[*i].xform); + } + } + // normal tiling + else { + std::set<int> private_stmt; + for (std::set<int>::iterator i = same_tile_controlling_loop.begin(); + i != same_tile_controlling_loop.end(); i++) { +// if (same_tiled_loop.find(*i) == same_tiled_loop.end() && !is_single_iteration(getNewIS(*i), dim)) +// same_tiled_loop.insert(*i); + + // should test dim's value directly but it is ok for now +// if (same_tiled_loop.find(*i) == same_tiled_loop.end() && get_const(stmt[*i].xform, dim+1, Output_Var) == posInfinity) + if (same_tiled_loop.find(*i) == same_tiled_loop.end() + && overflow.find(*i) != overflow.end()) + private_stmt.insert(*i); + } + + // extract the union of the iteration space to be considered + Relation hull; + /*{ + Tuple < Relation > r_list; + Tuple<int> r_mask; + + for (std::set<int>::iterator i = same_tile_controlling_loop.begin(); + i != same_tile_controlling_loop.end(); i++) + if (private_stmt.find(*i) == private_stmt.end()) { + Relation r = project_onto_levels(getNewIS(*i), dim + 1, + true); + for (int j = outer_dim; j < dim; j++) + r = Project(r, j + 1, Set_Var); + for (int j = 0; j < outer_dim; j += 2) + r = Project(r, j + 1, Set_Var); + r_list.append(r); + r_mask.append(1); + } + + hull = Hull(r_list, r_mask, 1, true); + }*/ + + { + std::vector<Relation> r_list; + + for (std::set<int>::iterator i = same_tile_controlling_loop.begin(); + i != same_tile_controlling_loop.end(); i++) + if (private_stmt.find(*i) == private_stmt.end()) { + Relation r = getNewIS(*i); + for (int j = dim + 2; j <= r.n_set(); j++) + r = Project(r, r.set_var(j)); + for (int j = outer_dim; j < dim; j++) + r = Project(r, j + 1, Set_Var); + for (int j = 0; j < outer_dim; j += 2) + r = Project(r, j + 1, Set_Var); + r.simplify(2, 4); + r_list.push_back(r); + } + + hull = SimpleHull(r_list); + // hull = Hull(r_list, std::vector<bool>(r_list.size(), true), 1, true); + } + + // extract the bound of the dimension to be tiled + Relation bound = get_loop_bound(hull, dim); + if (!bound.has_single_conjunct()) { + // further simplify the bound + hull = Approximate(hull); + bound = get_loop_bound(hull, dim); + + int i = outer_dim - 2; + while (!bound.has_single_conjunct() && i >= 0) { + hull = Project(hull, i + 1, Set_Var); + bound = get_loop_bound(hull, dim); + i -= 2; + } + + if (!bound.has_single_conjunct()) + throw loop_error("cannot handle tile bounds"); + } + + // separate lower and upper bounds + std::vector<GEQ_Handle> lb_list, ub_list; + { + Conjunct *c = bound.query_DNF()->single_conjunct(); + for (GEQ_Iterator gi(c->GEQs()); gi; gi++) { + int coef = (*gi).get_coef(bound.set_var(dim + 1)); + if (coef < 0) + ub_list.push_back(*gi); + else if (coef > 0) + lb_list.push_back(*gi); + } + } + if (lb_list.size() == 0) + throw loop_error( + "unable to calculate tile controlling loop lower bound"); + if (ub_list.size() == 0) + throw loop_error( + "unable to calculate tile controlling loop upper bound"); + + // find the simplest lower bound for StridedTile or simplest iteration count for CountedTile + int simplest_lb = 0, simplest_ub = 0; + if (method == StridedTile) { + int best_cost = INT_MAX; + for (int i = 0; i < lb_list.size(); i++) { + int cost = 0; + for (Constr_Vars_Iter ci(lb_list[i]); ci; ci++) { + switch ((*ci).var->kind()) { + case Input_Var: { + cost += 5; + break; + } + case Global_Var: { + cost += 2; + break; + } + default: + cost += 15; + break; + } + } + + if (cost < best_cost) { + best_cost = cost; + simplest_lb = i; + } + } + } else if (method == CountedTile) { + std::map<Variable_ID, coef_t> s1, s2, s3; + int best_cost = INT_MAX; + for (int i = 0; i < lb_list.size(); i++) + for (int j = 0; j < ub_list.size(); j++) { + int cost = 0; + + for (Constr_Vars_Iter ci(lb_list[i]); ci; ci++) { + switch ((*ci).var->kind()) { + case Input_Var: { + s1[(*ci).var] += (*ci).coef; + break; + } + case Global_Var: { + s2[(*ci).var] += (*ci).coef; + break; + } + case Exists_Var: + case Wildcard_Var: { + s3[(*ci).var] += (*ci).coef; + break; + } + default: + cost = INT_MAX - 2; + break; + } + } + + for (Constr_Vars_Iter ci(ub_list[j]); ci; ci++) { + switch ((*ci).var->kind()) { + case Input_Var: { + s1[(*ci).var] += (*ci).coef; + break; + } + case Global_Var: { + s2[(*ci).var] += (*ci).coef; + break; + } + case Exists_Var: + case Wildcard_Var: { + s3[(*ci).var] += (*ci).coef; + break; + } + default: + if (cost == INT_MAX - 2) + cost = INT_MAX - 1; + else + cost = INT_MAX - 3; + break; + } + } + + if (cost == 0) { + for (std::map<Variable_ID, coef_t>::iterator k = + s1.begin(); k != s1.end(); k++) + if ((*k).second != 0) + cost += 5; + for (std::map<Variable_ID, coef_t>::iterator k = + s2.begin(); k != s2.end(); k++) + if ((*k).second != 0) + cost += 2; + for (std::map<Variable_ID, coef_t>::iterator k = + s3.begin(); k != s3.end(); k++) + if ((*k).second != 0) + cost += 15; + } + + if (cost < best_cost) { + best_cost = cost; + simplest_lb = i; + simplest_ub = j; + } + } + } + + // prepare the new transformation relations + for (std::set<int>::iterator i = same_tile_controlling_loop.begin(); + i != same_tile_controlling_loop.end(); i++) { + Relation r(stmt[*i].xform.n_out(), stmt[*i].xform.n_out() + 2); + F_And *f_root = r.add_and(); + for (int j = 0; j < outer_dim - 1; j++) { + EQ_Handle h = f_root->add_EQ(); + h.update_coef(r.output_var(j + 1), 1); + h.update_coef(r.input_var(j + 1), -1); + } + + for (int j = outer_dim - 1; j < stmt[*i].xform.n_out(); j++) { + EQ_Handle h = f_root->add_EQ(); + h.update_coef(r.output_var(j + 3), 1); + h.update_coef(r.input_var(j + 1), -1); + } + + EQ_Handle h = f_root->add_EQ(); + h.update_coef(r.output_var(outer_dim), 1); + h.update_const(-lex[outer_dim - 1]); + + stmt[*i].xform = Composition(r, stmt[*i].xform); + } + + // add tiling constraints. + for (std::set<int>::iterator i = same_tile_controlling_loop.begin(); + i != same_tile_controlling_loop.end(); i++) { + F_And *f_super_root = stmt[*i].xform.and_with_and(); + F_Exists *f_exists = f_super_root->add_exists(); + F_And *f_root = f_exists->add_and(); + + // create a lower bound variable for easy formula creation later + Variable_ID aligned_lb; + { + Variable_ID lb = f_exists->declare(); + coef_t coef = lb_list[simplest_lb].get_coef( + bound.set_var(dim + 1)); + if (coef == 1) { // e.g. if i >= m+5, then LB = m+5 + EQ_Handle h = f_root->add_EQ(); + h.update_coef(lb, 1); + for (Constr_Vars_Iter ci(lb_list[simplest_lb]); ci; ci++) { + switch ((*ci).var->kind()) { + case Input_Var: { + int pos = (*ci).var->get_position(); + if (pos != dim + 1) + h.update_coef(stmt[*i].xform.output_var(pos), + (*ci).coef); + break; + } + case Global_Var: { + Global_Var_ID g = (*ci).var->get_global_var(); + Variable_ID v; + if (g->arity() == 0) + v = stmt[*i].xform.get_local(g); + else + v = stmt[*i].xform.get_local(g, + (*ci).var->function_of()); + h.update_coef(v, (*ci).coef); + break; + } + default: + throw loop_error("cannot handle tile bounds"); + } + } + h.update_const(lb_list[simplest_lb].get_const()); + } else { // e.g. if 2i >= m+5, then m+5 <= 2*LB < m+5+2 + GEQ_Handle h1 = f_root->add_GEQ(); + GEQ_Handle h2 = f_root->add_GEQ(); + for (Constr_Vars_Iter ci(lb_list[simplest_lb]); ci; ci++) { + switch ((*ci).var->kind()) { + case Input_Var: { + int pos = (*ci).var->get_position(); + if (pos == dim + 1) { + h1.update_coef(lb, (*ci).coef); + h2.update_coef(lb, -(*ci).coef); + } else { + h1.update_coef(stmt[*i].xform.output_var(pos), + (*ci).coef); + h2.update_coef(stmt[*i].xform.output_var(pos), + -(*ci).coef); + } + break; + } + case Global_Var: { + Global_Var_ID g = (*ci).var->get_global_var(); + Variable_ID v; + if (g->arity() == 0) + v = stmt[*i].xform.get_local(g); + else + v = stmt[*i].xform.get_local(g, + (*ci).var->function_of()); + h1.update_coef(v, (*ci).coef); + h2.update_coef(v, -(*ci).coef); + break; + } + default: + throw loop_error("cannot handle tile bounds"); + } + } + h1.update_const(lb_list[simplest_lb].get_const()); + h2.update_const(-lb_list[simplest_lb].get_const()); + h2.update_const(coef - 1); + } + + Variable_ID offset_lb; + if (alignment_offset == 0) + offset_lb = lb; + else { + EQ_Handle h = f_root->add_EQ(); + offset_lb = f_exists->declare(); + h.update_coef(offset_lb, 1); + h.update_coef(lb, -1); + h.update_const(alignment_offset); + } + + if (alignment_multiple == 1) { // trivial + aligned_lb = offset_lb; + } else { // e.g. to align at 4, aligned_lb = 4*alpha && LB-4 < 4*alpha <= LB + aligned_lb = f_exists->declare(); + Variable_ID e = f_exists->declare(); + + EQ_Handle h = f_root->add_EQ(); + h.update_coef(aligned_lb, 1); + h.update_coef(e, -alignment_multiple); + + GEQ_Handle h1 = f_root->add_GEQ(); + GEQ_Handle h2 = f_root->add_GEQ(); + h1.update_coef(e, alignment_multiple); + h2.update_coef(e, -alignment_multiple); + h1.update_coef(offset_lb, -1); + h2.update_coef(offset_lb, 1); + h1.update_const(alignment_multiple - 1); + } + } + + // create an upper bound variable for easy formula creation later + Variable_ID ub = f_exists->declare(); + { + coef_t coef = -ub_list[simplest_ub].get_coef( + bound.set_var(dim + 1)); + if (coef == 1) { // e.g. if i <= m+5, then UB = m+5 + EQ_Handle h = f_root->add_EQ(); + h.update_coef(ub, -1); + for (Constr_Vars_Iter ci(ub_list[simplest_ub]); ci; ci++) { + switch ((*ci).var->kind()) { + case Input_Var: { + int pos = (*ci).var->get_position(); + if (pos != dim + 1) + h.update_coef(stmt[*i].xform.output_var(pos), + (*ci).coef); + break; + } + case Global_Var: { + Global_Var_ID g = (*ci).var->get_global_var(); + Variable_ID v; + if (g->arity() == 0) + v = stmt[*i].xform.get_local(g); + else + v = stmt[*i].xform.get_local(g, + (*ci).var->function_of()); + h.update_coef(v, (*ci).coef); + break; + } + default: + throw loop_error("cannot handle tile bounds"); + } + } + h.update_const(ub_list[simplest_ub].get_const()); + } else { // e.g. if 2i <= m+5, then m+5-2 < 2*UB <= m+5 + GEQ_Handle h1 = f_root->add_GEQ(); + GEQ_Handle h2 = f_root->add_GEQ(); + for (Constr_Vars_Iter ci(ub_list[simplest_ub]); ci; ci++) { + switch ((*ci).var->kind()) { + case Input_Var: { + int pos = (*ci).var->get_position(); + if (pos == dim + 1) { + h1.update_coef(ub, -(*ci).coef); + h2.update_coef(ub, (*ci).coef); + } else { + h1.update_coef(stmt[*i].xform.output_var(pos), + -(*ci).coef); + h2.update_coef(stmt[*i].xform.output_var(pos), + (*ci).coef); + } + break; + } + case Global_Var: { + Global_Var_ID g = (*ci).var->get_global_var(); + Variable_ID v; + if (g->arity() == 0) + v = stmt[*i].xform.get_local(g); + else + v = stmt[*i].xform.get_local(g, + (*ci).var->function_of()); + h1.update_coef(v, -(*ci).coef); + h2.update_coef(v, (*ci).coef); + break; + } + default: + throw loop_error("cannot handle tile bounds"); + } + } + h1.update_const(-ub_list[simplest_ub].get_const()); + h2.update_const(ub_list[simplest_ub].get_const()); + h1.update_const(coef - 1); + } + } + + // insert tile controlling loop constraints + if (method == StridedTile) { // e.g. ii = LB + 32 * alpha && alpha >= 0 + Variable_ID e = f_exists->declare(); + GEQ_Handle h1 = f_root->add_GEQ(); + h1.update_coef(e, 1); + + EQ_Handle h2 = f_root->add_EQ(); + h2.update_coef(stmt[*i].xform.output_var(outer_dim + 1), 1); + h2.update_coef(e, -tile_size); + h2.update_coef(aligned_lb, -1); + } else if (method == CountedTile) { // e.g. 0 <= ii < ceiling((UB-LB+1)/32) + GEQ_Handle h1 = f_root->add_GEQ(); + h1.update_coef(stmt[*i].xform.output_var(outer_dim + 1), 1); + + GEQ_Handle h2 = f_root->add_GEQ(); + h2.update_coef(stmt[*i].xform.output_var(outer_dim + 1), + -tile_size); + h2.update_coef(aligned_lb, -1); + h2.update_coef(ub, 1); + } + + // special care for private statements like overflow assignment + if (private_stmt.find(*i) != private_stmt.end()) { // e.g. ii <= UB + GEQ_Handle h = f_root->add_GEQ(); + h.update_coef(stmt[*i].xform.output_var(outer_dim + 1), -1); + h.update_coef(ub, 1); + } + // if (private_stmt.find(*i) != private_stmt.end()) { + // if (stmt[*i].xform.n_out() > dim+3) { // e.g. ii <= UB && i = ii + // GEQ_Handle h = f_root->add_GEQ(); + // h.update_coef(stmt[*i].xform.output_var(outer_dim+1), -1); + // h.update_coef(ub, 1); + + // stmt[*i].xform = Project(stmt[*i].xform, dim+3, Output_Var); + // f_root = stmt[*i].xform.and_with_and(); + // EQ_Handle h1 = f_root->add_EQ(); + // h1.update_coef(stmt[*i].xform.output_var(dim+3), 1); + // h1.update_coef(stmt[*i].xform.output_var(outer_dim+1), -1); + // } + // else if (method == StridedTile) { // e.g. ii <= UB since i does not exist + // GEQ_Handle h = f_root->add_GEQ(); + // h.update_coef(stmt[*i].xform.output_var(outer_dim+1), -1); + // h.update_coef(ub, 1); + // } + // } + + // restrict original loop index inside the tile + else { + if (method == StridedTile) { // e.g. ii <= i < ii + tile_size + GEQ_Handle h1 = f_root->add_GEQ(); + h1.update_coef(stmt[*i].xform.output_var(dim + 3), 1); + h1.update_coef(stmt[*i].xform.output_var(outer_dim + 1), + -1); + + GEQ_Handle h2 = f_root->add_GEQ(); + h2.update_coef(stmt[*i].xform.output_var(dim + 3), -1); + h2.update_coef(stmt[*i].xform.output_var(outer_dim + 1), 1); + h2.update_const(tile_size - 1); + } else if (method == CountedTile) { // e.g. LB+32*ii <= i < LB+32*ii+tile_size + GEQ_Handle h1 = f_root->add_GEQ(); + h1.update_coef(stmt[*i].xform.output_var(outer_dim + 1), + -tile_size); + h1.update_coef(stmt[*i].xform.output_var(dim + 3), 1); + h1.update_coef(aligned_lb, -1); + + GEQ_Handle h2 = f_root->add_GEQ(); + h2.update_coef(stmt[*i].xform.output_var(outer_dim + 1), + tile_size); + h2.update_coef(stmt[*i].xform.output_var(dim + 3), -1); + h2.update_const(tile_size - 1); + h2.update_coef(aligned_lb, 1); + } + } + } + } + + // update loop level information + for (std::set<int>::iterator i = same_tile_controlling_loop.begin(); + i != same_tile_controlling_loop.end(); i++) { + for (int j = 1; j <= stmt[*i].loop_level.size(); j++) + switch (stmt[*i].loop_level[j - 1].type) { + case LoopLevelOriginal: + break; + case LoopLevelTile: + if (stmt[*i].loop_level[j - 1].payload >= outer_level) + stmt[*i].loop_level[j - 1].payload++; + break; + default: + throw loop_error( + "unknown loop level type for statement " + + to_string(*i)); + } + + LoopLevel ll; + ll.type = LoopLevelTile; + ll.payload = level + 1; + ll.parallel_level = 0; + stmt[*i].loop_level.insert( + stmt[*i].loop_level.begin() + (outer_level - 1), ll); + } +} + |