@@ -163,6 +163,7 @@ enum gf_mask {
GF_OMP_FOR_COMBINED = 1 << 4,
GF_OMP_FOR_COMBINED_INTO = 1 << 5,
GF_OMP_FOR_GRID_PHONY = 1 << 6,
+ GF_OMP_FOR_GRID_GROUP_ITER = 1 << 7,
GF_OMP_TARGET_KIND_MASK = (1 << 4) - 1,
GF_OMP_TARGET_KIND_REGION = 0,
GF_OMP_TARGET_KIND_DATA = 1,
@@ -5124,6 +5125,26 @@ gimple_omp_for_set_grid_phony (gomp_for *omp_for, bool value)
omp_for->subcode &= ~GF_OMP_FOR_GRID_PHONY;
}
+/* Return true if iterations of a grid OMP_FOR statement correspond to HSA
+ groups. */
+
+static inline bool
+gimple_omp_for_grid_group_iter (const gomp_for *omp_for)
+{
+ return (gimple_omp_subcode (omp_for) & GF_OMP_FOR_GRID_GROUP_ITER) != 0;
+}
+
+/* Set group_iter flag of OMP_FOR to VALUE. */
+
+static inline void
+gimple_omp_for_set_grid_group_iter (gomp_for *omp_for, bool value)
+{
+ if (value)
+ omp_for->subcode |= GF_OMP_FOR_GRID_GROUP_ITER;
+ else
+ omp_for->subcode &= ~GF_OMP_FOR_GRID_GROUP_ITER;
+}
+
/* Return the clauses associated with OMP_PARALLEL GS. */
static inline tree
@@ -3297,8 +3297,8 @@ check_omp_nesting_restrictions (gimple *stmt, omp_context *ctx)
else if (gimple_code (ctx->stmt) == GIMPLE_OMP_TEAMS)
{
if ((gimple_code (stmt) != GIMPLE_OMP_FOR
- || (gimple_omp_for_kind (stmt)
- != GF_OMP_FOR_KIND_DISTRIBUTE))
+ || ((gimple_omp_for_kind (stmt) != GF_OMP_FOR_KIND_DISTRIBUTE)
+ && (gimple_omp_for_kind (stmt) != GF_OMP_FOR_KIND_GRID_LOOP)))
&& gimple_code (stmt) != GIMPLE_OMP_PARALLEL)
{
error_at (gimple_location (stmt),
@@ -13505,10 +13505,12 @@ expand_omp_target (struct omp_region *region)
}
/* Expand KFOR loop as a GPGPU kernel, i.e. as a body only with iteration
- variable derived from the thread number. */
+ variable derived from the thread number. INTRA_GROUP means this is an
+ expansion of a loop iterating over work-items within a separate iteration
+ over groups. */
static void
-grid_expand_omp_for_loop (struct omp_region *kfor)
+grid_expand_omp_for_loop (struct omp_region *kfor, bool intra_group)
{
gimple_stmt_iterator gsi;
gomp_for *for_stmt = as_a <gomp_for *> (last_stmt (kfor->entry));
@@ -13522,6 +13524,7 @@ grid_expand_omp_for_loop (struct omp_region *kfor)
struct omp_for_data fd;
+ remove_edge (BRANCH_EDGE (kfor->entry));
basic_block body_bb = FALLTHRU_EDGE (kfor->entry)->dest;
gcc_assert (kfor->cont);
@@ -13542,9 +13545,22 @@ grid_expand_omp_for_loop (struct omp_region *kfor)
true, NULL_TREE, true, GSI_SAME_STMT);
step = force_gimple_operand_gsi (&gsi, fold_convert (itype, step),
true, NULL_TREE, true, GSI_SAME_STMT);
- tree threadid = build_call_expr (builtin_decl_explicit
- (BUILT_IN_HSA_GET_WORKITEM_ABSID), 1,
- build_int_cstu (unsigned_type_node, dim));
+ tree threadid;
+ if (gimple_omp_for_grid_group_iter (for_stmt))
+ {
+ gcc_checking_assert (!intra_group);
+ threadid = build_call_expr (builtin_decl_explicit
+ (BUILT_IN_HSA_GET_WORKGROUP_ID), 1,
+ build_int_cstu (unsigned_type_node, dim));
+ }
+ else if (intra_group)
+ threadid = build_call_expr (builtin_decl_explicit
+ (BUILT_IN_HSA_GET_WORKITEM_ID), 1,
+ build_int_cstu (unsigned_type_node, dim));
+ else
+ threadid = build_call_expr (builtin_decl_explicit
+ (BUILT_IN_HSA_GET_WORKITEM_ABSID), 1,
+ build_int_cstu (unsigned_type_node, dim));
threadid = fold_convert (itype, threadid);
threadid = force_gimple_operand_gsi (&gsi, threadid, true, NULL_TREE,
true, GSI_SAME_STMT);
@@ -13573,10 +13589,12 @@ grid_expand_omp_for_loop (struct omp_region *kfor)
&& gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_CONTINUE);
gsi_remove (&gsi, true);
- /* Replace the GIMPLE_OMP_RETURN with a real return. */
+ /* Replace the GIMPLE_OMP_RETURN with a barrier, if necessary. */
gsi = gsi_last_bb (kfor->exit);
gcc_assert (!gsi_end_p (gsi)
&& gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_RETURN);
+ if (intra_group)
+ gsi_insert_before (&gsi, build_omp_barrier (NULL_TREE), GSI_SAME_STMT);
gsi_remove (&gsi, true);
/* Fixup the much simpler CFG. */
@@ -13657,11 +13675,29 @@ grid_expand_target_grid_body (struct omp_region *target)
struct omp_region *kfor = *pp;
gcc_assert (kfor);
- gcc_assert (gimple_omp_for_kind (last_stmt ((kfor)->entry))
- == GF_OMP_FOR_KIND_GRID_LOOP);
+ gomp_for *for_stmt = as_a <gomp_for *> (last_stmt (kfor->entry));
+ gcc_assert (gimple_omp_for_kind (for_stmt) == GF_OMP_FOR_KIND_GRID_LOOP);
*pp = kfor->next;
if (kfor->inner)
- expand_omp (kfor->inner);
+ {
+ if (gimple_omp_for_grid_group_iter (for_stmt))
+ {
+ struct omp_region **next_pp;
+ for (pp = &kfor->inner; *pp; pp = next_pp)
+ {
+ next_pp = &(*pp)->next;
+ if ((*pp)->type != GIMPLE_OMP_FOR)
+ continue;
+ gomp_for *inner = as_a <gomp_for *> (last_stmt ((*pp)->entry));
+ gcc_assert (gimple_omp_for_kind (inner)
+ == GF_OMP_FOR_KIND_GRID_LOOP);
+ grid_expand_omp_for_loop (*pp, true);
+ *pp = (*pp)->next;
+ next_pp = pp;
+ }
+ }
+ expand_omp (kfor->inner);
+ }
if (gpukernel->inner)
expand_omp (gpukernel->inner);
@@ -13690,8 +13726,7 @@ grid_expand_target_grid_body (struct omp_region *target)
struct function *kern_cfun = DECL_STRUCT_FUNCTION (kern_fndecl);
kern_cfun->curr_properties = cfun->curr_properties;
- remove_edge (BRANCH_EDGE (kfor->entry));
- grid_expand_omp_for_loop (kfor);
+ grid_expand_omp_for_loop (kfor, false);
/* Remove the omp for statement */
gimple_stmt_iterator gsi = gsi_last_bb (gpukernel->entry);
@@ -17164,60 +17199,90 @@ lower_omp (gimple_seq *body, omp_context *ctx)
input_location = saved_location;
}
-/* Returen true if STMT is an assignment of a register-type into a local
- VAR_DECL. */
+/* Structure describing the basic properties of the loop we ara analyzing
+ whether it can be gridified and when it is gridified. */
+
+struct grid_prop
+{
+ /* True when we are doing tiling gridification, i.e. when there is a distinct
+ distribute loop over groups and a loop construct over work-items. False
+ when distribute and parallel for loops form a combined construct. */
+ bool tiling;
+ /* Location of the target construct for optimization information
+ messages. */
+ location_t target_loc;
+ /* The collapse clause of the involved loops. Collapse value of all of them
+ must be the same for gridification to take place. */
+ size_t collapse;
+ /* Group sizes, if requested by the user or NULL if not requested. */
+ tree group_sizes[3];
+};
+
+#define GRID_MISSED_MSG_PREFIX "Will not turn target construct into a " \
+ "gridified GPGPU kernel because "
+
+/* Return true if STMT is an assignment of a register-type into a local
+ VAR_DECL. If GRID is non-NULL, the assignment additionally must not be to
+ any of the trees specifying group sizes there. */
static bool
-grid_reg_assignment_to_local_var_p (gimple *stmt)
+grid_safe_assignment_p (gimple *stmt, grid_prop *grid)
{
gassign *assign = dyn_cast <gassign *> (stmt);
if (!assign)
return false;
+ if (gimple_clobber_p (assign))
+ return true;
tree lhs = gimple_assign_lhs (assign);
if (TREE_CODE (lhs) != VAR_DECL
|| !is_gimple_reg_type (TREE_TYPE (lhs))
|| is_global_var (lhs))
return false;
+ if (grid)
+ for (unsigned i = 0; i < grid->collapse; i++)
+ if (lhs == grid->group_sizes[i])
+ return false;
return true;
}
/* Return true if all statements in SEQ are assignments to local register-type
- variables. */
+ variables that do not hold group size information. */
static bool
-grid_seq_only_contains_local_assignments (gimple_seq seq)
+grid_seq_only_contains_local_assignments (gimple_seq seq, grid_prop *grid)
{
if (!seq)
return true;
gimple_stmt_iterator gsi;
for (gsi = gsi_start (seq); !gsi_end_p (gsi); gsi_next (&gsi))
- if (!grid_reg_assignment_to_local_var_p (gsi_stmt (gsi)))
+ if (!grid_safe_assignment_p (gsi_stmt (gsi), grid))
return false;
return true;
}
-/* Scan statements in SEQ and call itself recursively on any bind. If during
- whole search only assignments to register-type local variables and one
- single OMP statement is encountered, return true, otherwise return false.
- RET is where we store any OMP statement encountered. TARGET_LOC and NAME
- are used for dumping a note about a failure. */
+/* Scan statements in SEQ and call itself recursively on any bind. GRID
+ describes hitherto discovered properties of the loop that is evaluated for
+ possible gridification. If during whole search only assignments to
+ register-type local variables (that do not overwrite group size information)
+ and one single OMP statement is encountered, return true, otherwise return
+ false. RET is where we store any OMP statement encountered. */
static bool
-grid_find_single_omp_among_assignments_1 (gimple_seq seq, location_t target_loc,
- const char *name, gimple **ret)
+grid_find_single_omp_among_assignments_1 (gimple_seq seq, grid_prop *grid,
+ const char *name, gimple **ret)
{
gimple_stmt_iterator gsi;
for (gsi = gsi_start (seq); !gsi_end_p (gsi); gsi_next (&gsi))
{
gimple *stmt = gsi_stmt (gsi);
- if (grid_reg_assignment_to_local_var_p (stmt))
+ if (grid_safe_assignment_p (stmt, grid))
continue;
if (gbind *bind = dyn_cast <gbind *> (stmt))
{
if (!grid_find_single_omp_among_assignments_1 (gimple_bind_body (bind),
- target_loc, name, ret))
+ grid, name, ret))
return false;
}
else if (is_gimple_omp (stmt))
@@ -17225,10 +17290,18 @@ grid_find_single_omp_among_assignments_1 (gimple_seq seq, location_t target_loc,
if (*ret)
{
if (dump_enabled_p ())
- dump_printf_loc (MSG_MISSED_OPTIMIZATION, target_loc,
- "Will not turn target construct into a simple "
- "GPGPU kernel because %s construct contains "
- "multiple OpenMP constructs\n", name);
+ {
+ dump_printf_loc (MSG_MISSED_OPTIMIZATION, grid->target_loc,
+ GRID_MISSED_MSG_PREFIX "%s construct "
+ "contains multiple OpenMP constructs\n",
+ name);
+ dump_printf_loc (MSG_NOTE, gimple_location (*ret),
+ "The first OpenMP construct within "
+ "a parallel\n");
+ dump_printf_loc (MSG_NOTE, gimple_location (stmt),
+ "The second OpenMP construct within "
+ "a parallel\n");
+ }
return false;
}
*ret = stmt;
@@ -17236,10 +17309,14 @@ grid_find_single_omp_among_assignments_1 (gimple_seq seq, location_t target_loc,
else
{
if (dump_enabled_p ())
- dump_printf_loc (MSG_MISSED_OPTIMIZATION, target_loc,
- "Will not turn target construct into a simple "
- "GPGPU kernel because %s construct contains "
- "a complex statement\n", name);
+ {
+ dump_printf_loc (MSG_MISSED_OPTIMIZATION, grid->target_loc,
+ GRID_MISSED_MSG_PREFIX "%s construct contains "
+ "a complex statement\n", name);
+ dump_printf_loc (MSG_NOTE, gimple_location (stmt),
+ "This statement cannot be analyzed for "
+ "gridification\n");
+ }
return false;
}
}
@@ -17247,33 +17324,32 @@ grid_find_single_omp_among_assignments_1 (gimple_seq seq, location_t target_loc,
}
/* Scan statements in SEQ and make sure that it and any binds in it contain
- only assignments to local register-type variables and one OMP construct. If
- so, return that construct, otherwise return NULL. If dumping is enabled and
- function fails, use TARGET_LOC and NAME to dump a note with the reason for
- failure. */
+ only assignments to local register-type variables (that do not overwrite
+ group size information) and one OMP construct. If so, return that
+ construct, otherwise return NULL. GRID describes hitherto discovered
+ properties of the loop that is evaluated for possible gridification. If
+ dumping is enabled and function fails, use NAME to dump a note with the
+ reason for failure. */
static gimple *
-grid_find_single_omp_among_assignments (gimple_seq seq, location_t target_loc,
+grid_find_single_omp_among_assignments (gimple_seq seq, grid_prop *grid,
const char *name)
{
if (!seq)
{
if (dump_enabled_p ())
- dump_printf_loc (MSG_MISSED_OPTIMIZATION, target_loc,
- "Will not turn target construct into a simple "
- "GPGPU kernel because %s construct has empty "
- "body\n",
+ dump_printf_loc (MSG_MISSED_OPTIMIZATION, grid->target_loc,
+ GRID_MISSED_MSG_PREFIX "%s construct has empty body\n",
name);
return NULL;
}
gimple *ret = NULL;
- if (grid_find_single_omp_among_assignments_1 (seq, target_loc, name, &ret))
+ if (grid_find_single_omp_among_assignments_1 (seq, grid, name, &ret))
{
if (!ret && dump_enabled_p ())
- dump_printf_loc (MSG_MISSED_OPTIMIZATION, target_loc,
- "Will not turn target construct into a simple "
- "GPGPU kernel because %s construct does not contain"
+ dump_printf_loc (MSG_MISSED_OPTIMIZATION, grid->target_loc,
+ GRID_MISSED_MSG_PREFIX "%s construct does not contain"
"any other OpenMP construct\n", name);
return ret;
}
@@ -17333,157 +17409,128 @@ grid_find_ungridifiable_statement (gimple_stmt_iterator *gsi,
return NULL;
}
-
-/* If TARGET follows a pattern that can be turned into a gridified GPGPU
- kernel, return true, otherwise return false. In the case of success, also
- fill in GROUP_SIZE_P with the requested group size or NULL if there is
- none. */
+/* Examine clauses of omp parallel statement PAR and if any prevents
+ gridification, issue a missed-optimization diagnostics and return false,
+ otherwise return true. GRID describes hitherto discovered properties of the
+ loop that is evaluated for possible gridification. */
static bool
-grid_target_follows_gridifiable_pattern (gomp_target *target, tree *group_size_p)
+grid_parallel_clauses_gridifiable (gomp_parallel *par, location_t tloc)
{
- if (gimple_omp_target_kind (target) != GF_OMP_TARGET_KIND_REGION)
- return false;
-
- location_t tloc = gimple_location (target);
- gimple *stmt
- = grid_find_single_omp_among_assignments (gimple_omp_body (target),
- tloc, "target");
- if (!stmt)
- return false;
- gomp_teams *teams = dyn_cast <gomp_teams *> (stmt);
- tree group_size = NULL;
- if (!teams)
- {
- dump_printf_loc (MSG_MISSED_OPTIMIZATION, tloc,
- "Will not turn target construct into a simple "
- "GPGPU kernel because it does not have a sole teams "
- "construct in it.\n");
- return false;
- }
-
- tree clauses = gimple_omp_teams_clauses (teams);
+ tree clauses = gimple_omp_parallel_clauses (par);
while (clauses)
{
switch (OMP_CLAUSE_CODE (clauses))
{
- case OMP_CLAUSE_NUM_TEAMS:
+ case OMP_CLAUSE_NUM_THREADS:
if (dump_enabled_p ())
- dump_printf_loc (MSG_MISSED_OPTIMIZATION, tloc,
- "Will not turn target construct into a "
- "gridified GPGPU kernel because we cannot "
- "handle num_teams clause of teams "
- "construct\n ");
+ {
+ dump_printf_loc (MSG_MISSED_OPTIMIZATION, tloc,
+ GRID_MISSED_MSG_PREFIX "because there is "
+ "a num_threads clause of the parallel "
+ "construct\n");
+ dump_printf_loc (MSG_NOTE, gimple_location (par),
+ "Parallel construct has a num_threads clause\n");
+ }
return false;
case OMP_CLAUSE_REDUCTION:
if (dump_enabled_p ())
- dump_printf_loc (MSG_MISSED_OPTIMIZATION, tloc,
- "Will not turn target construct into a "
- "gridified GPGPU kernel because a reduction "
- "clause is present\n ");
+ {
+ dump_printf_loc (MSG_MISSED_OPTIMIZATION, tloc,
+ GRID_MISSED_MSG_PREFIX "a reduction clause"
+ "is present\n ");
+ dump_printf_loc (MSG_NOTE, gimple_location (par),
+ "Parallel construct has a reduction clause\n");
+ }
return false;
case OMP_CLAUSE_LASTPRIVATE:
if (dump_enabled_p ())
- dump_printf_loc (MSG_MISSED_OPTIMIZATION, tloc,
- "Will not turn target construct into a "
- "gridified GPGPU kernel because a lastprivate "
- "clause is present\n ");
+ {
+ dump_printf_loc (MSG_MISSED_OPTIMIZATION, tloc,
+ GRID_MISSED_MSG_PREFIX "a lastprivate "
+ "clause is present\n ");
+ dump_printf_loc (MSG_NOTE, gimple_location (par),
+ "Parallel construct has a lastprivate clause\n");
+ }
return false;
- case OMP_CLAUSE_THREAD_LIMIT:
- group_size = OMP_CLAUSE_OPERAND (clauses, 0);
- break;
-
default:
break;
}
clauses = OMP_CLAUSE_CHAIN (clauses);
}
+ return true;
+}
- stmt = grid_find_single_omp_among_assignments (gimple_omp_body (teams), tloc,
- "teams");
- if (!stmt)
- return false;
- gomp_for *dist = dyn_cast <gomp_for *> (stmt);
- if (!dist)
- {
- dump_printf_loc (MSG_MISSED_OPTIMIZATION, tloc,
- "Will not turn target construct into a simple "
- "GPGPU kernel because the teams construct does not have "
- "a sole distribute construct in it.\n");
- return false;
- }
+/* Examine clauses and the body of omp loop statement GFOR and if something
+ prevents gridification, issue a missed-optimization diagnostics and return
+ false, otherwise return true. GRID describes hitherto discovered properties
+ of the loop that is evaluated for possible gridification. */
- gcc_assert (gimple_omp_for_kind (dist) == GF_OMP_FOR_KIND_DISTRIBUTE);
- if (!gimple_omp_for_combined_p (dist))
- {
- if (dump_enabled_p ())
- dump_printf_loc (MSG_MISSED_OPTIMIZATION, tloc,
- "Will not turn target construct into a gridified GPGPU "
- "kernel because we cannot handle a standalone "
- "distribute construct\n ");
- return false;
- }
- if (dist->collapse > 3)
+static bool
+grid_inner_loop_gridifiable_p (gomp_for *gfor, grid_prop *grid)
+{
+ if (!grid_seq_only_contains_local_assignments (gimple_omp_for_pre_body (gfor),
+ grid))
{
if (dump_enabled_p ())
- dump_printf_loc (MSG_MISSED_OPTIMIZATION, tloc,
- "Will not turn target construct into a gridified GPGPU "
- "kernel because the distribute construct contains "
- "collapse clause with parameter greater than 3\n");
- return false;
- }
- struct omp_for_data fd;
- extract_omp_for_data (dist, &fd, NULL);
- if (fd.chunk_size)
- {
- if (group_size && !operand_equal_p (group_size, fd.chunk_size, 0))
{
- if (dump_enabled_p ())
- dump_printf_loc (MSG_MISSED_OPTIMIZATION, tloc,
- "Will not turn target construct into a "
- "gridified GPGPU kernel because the teams "
- "thread limit is different from distribute "
- "schedule chunk\n");
- return false;
+ dump_printf_loc (MSG_MISSED_OPTIMIZATION, grid->target_loc,
+ GRID_MISSED_MSG_PREFIX "the inner loop "
+ "loop bounds computation contains a complex "
+ "statement\n");
+ dump_printf_loc (MSG_NOTE, gimple_location (gfor),
+ "Loop construct cannot be analyzed for "
+ "gridification\n");
}
- group_size = fd.chunk_size;
+ return false;
}
- stmt = grid_find_single_omp_among_assignments (gimple_omp_body (dist), tloc,
- "distribute");
- gomp_parallel *par;
- if (!stmt || !(par = dyn_cast <gomp_parallel *> (stmt)))
- return false;
- clauses = gimple_omp_parallel_clauses (par);
+ tree clauses = gimple_omp_for_clauses (gfor);
while (clauses)
{
switch (OMP_CLAUSE_CODE (clauses))
{
- case OMP_CLAUSE_NUM_THREADS:
- if (dump_enabled_p ())
- dump_printf_loc (MSG_MISSED_OPTIMIZATION, tloc,
- "Will not turn target construct into a gridified"
- "GPGPU kernel because there is a num_threads "
- "clause of the parallel construct\n");
- return false;
+ case OMP_CLAUSE_SCHEDULE:
+ if (OMP_CLAUSE_SCHEDULE_KIND (clauses) != OMP_CLAUSE_SCHEDULE_AUTO)
+ {
+ if (dump_enabled_p ())
+ {
+ dump_printf_loc (MSG_MISSED_OPTIMIZATION, grid->target_loc,
+ GRID_MISSED_MSG_PREFIX "the inner loop "
+ "has a non-automatic schedule clause\n");
+ dump_printf_loc (MSG_NOTE, gimple_location (gfor),
+ "Loop construct has a non automatic "
+ "schedule clause\n");
+ }
+ return false;
+ }
+ break;
case OMP_CLAUSE_REDUCTION:
if (dump_enabled_p ())
- dump_printf_loc (MSG_MISSED_OPTIMIZATION, tloc,
- "Will not turn target construct into a "
- "gridified GPGPU kernel because a reduction "
- "clause is present\n ");
+ {
+ dump_printf_loc (MSG_MISSED_OPTIMIZATION, grid->target_loc,
+ GRID_MISSED_MSG_PREFIX "a reduction "
+ "clause is present\n ");
+ dump_printf_loc (MSG_NOTE, gimple_location (gfor),
+ "Loop construct has a reduction schedule "
+ "clause\n");
+ }
return false;
case OMP_CLAUSE_LASTPRIVATE:
if (dump_enabled_p ())
- dump_printf_loc (MSG_MISSED_OPTIMIZATION, tloc,
- "Will not turn target construct into a "
- "gridified GPGPU kernel because a lastprivate "
- "clause is present\n ");
+ {
+ dump_printf_loc (MSG_MISSED_OPTIMIZATION, grid->target_loc,
+ GRID_MISSED_MSG_PREFIX "a lastprivate "
+ "clause is present\n ");
+ dump_printf_loc (MSG_NOTE, gimple_location (gfor),
+ "Loop construct has a lastprivate schedule "
+ "clause\n");
+ }
return false;
default:
@@ -17491,8 +17538,56 @@ grid_target_follows_gridifiable_pattern (gomp_target *target, tree *group_size_p
}
clauses = OMP_CLAUSE_CHAIN (clauses);
}
+ struct walk_stmt_info wi;
+ memset (&wi, 0, sizeof (wi));
+ if (walk_gimple_seq (gimple_omp_body (gfor),
+ grid_find_ungridifiable_statement,
+ NULL, &wi))
+ {
+ gimple *bad = (gimple *) wi.info;
+ if (dump_enabled_p ())
+ {
+ if (is_gimple_call (bad))
+ dump_printf_loc (MSG_MISSED_OPTIMIZATION, grid->target_loc,
+ GRID_MISSED_MSG_PREFIX "the inner loop contains "
+ "call to a noreturn function\n");
+ else if (gimple_code (bad) == GIMPLE_OMP_FOR)
+ dump_printf_loc (MSG_MISSED_OPTIMIZATION, grid->target_loc,
+ GRID_MISSED_MSG_PREFIX "the inner loop contains "
+ "a simd construct\n");
+ else
+ dump_printf_loc (MSG_MISSED_OPTIMIZATION, grid->target_loc,
+ GRID_MISSED_MSG_PREFIX "the inner loop contains "
+ "statement %s which cannot be transformed\n",
+ gimple_code_name[(int) gimple_code (bad)]);
+ dump_printf_loc (MSG_NOTE, gimple_location (bad),
+ "This statement cannot be analyzed for "
+ "gridification\n");
+ }
+ return false;
+ }
+ return true;
+}
- stmt = grid_find_single_omp_among_assignments (gimple_omp_body (par), tloc,
+/* Given distribute omp construct represented by DIST, which in the original
+ source forms a compound construct with a looping construct, return true if it
+ can be turned into a gridified GPGPU kernel. Otherwise return false. GRID
+ describes hitherto discovered properties of the loop that is evaluated for
+ possible gridification. */
+
+static bool
+grid_dist_follows_simple_pattern (gomp_for *dist, grid_prop *grid)
+{
+ location_t tloc = grid->target_loc;
+ gimple *stmt = grid_find_single_omp_among_assignments (gimple_omp_body (dist),
+ grid, "distribute");
+ gomp_parallel *par;
+ if (!stmt
+ || !(par = dyn_cast <gomp_parallel *> (stmt))
+ || !grid_parallel_clauses_gridifiable (par, tloc))
+ return false;
+
+ stmt = grid_find_single_omp_among_assignments (gimple_omp_body (par), grid,
"parallel");
gomp_for *gfor;
if (!stmt || !(gfor = dyn_cast <gomp_for *> (stmt)))
@@ -17502,101 +17597,441 @@ grid_target_follows_gridifiable_pattern (gomp_target *target, tree *group_size_p
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, tloc,
- "Will not turn target construct into a gridified GPGPU "
- "kernel because the inner loop is not a simple for "
- "loop\n");
+ GRID_MISSED_MSG_PREFIX "the inner loop is not "
+ "a simple for loop\n");
return false;
}
- if (gfor->collapse > 3)
+ gcc_assert (gimple_omp_for_collapse (gfor) == grid->collapse);
+
+ if (!grid_inner_loop_gridifiable_p (gfor, grid))
+ return false;
+
+ return true;
+}
+
+/* Given an omp loop statement GFOR, return true if it can participate in
+ tiling gridification, i.e. in one where the distribute and parallel for
+ loops do not form a compound statement. GRID describes hitherto discovered
+ properties of the loop that is evaluated for possible gridification. */
+
+static bool
+grid_gfor_follows_tiling_pattern (gomp_for *gfor, grid_prop *grid)
+{
+ if (gimple_omp_for_kind (gfor) != GF_OMP_FOR_KIND_FOR)
{
if (dump_enabled_p ())
- dump_printf_loc (MSG_MISSED_OPTIMIZATION, tloc,
- "Will not turn target construct into a gridified GPGPU "
- "kernel because the inner loop contains collapse "
- "clause with parameter greater than 3\n");
+ {
+ dump_printf_loc (MSG_MISSED_OPTIMIZATION, grid->target_loc,
+ GRID_MISSED_MSG_PREFIX "an inner loop is not "
+ "a simple for loop\n");
+ dump_printf_loc (MSG_NOTE, gimple_location (gfor),
+ "This statement is not a simple for loop\n");
+ }
return false;
}
- if (!grid_seq_only_contains_local_assignments (gimple_omp_for_pre_body (gfor)))
+ if (!grid_inner_loop_gridifiable_p (gfor, grid))
+ return false;
+
+ if (gimple_omp_for_collapse (gfor) != grid->collapse)
{
if (dump_enabled_p ())
- dump_printf_loc (MSG_MISSED_OPTIMIZATION, tloc,
- "Will not turn target construct into a gridified GPGPU "
- "kernel because the inner loop pre_body contains"
- "a complex instruction\n");
+ {
+ dump_printf_loc (MSG_MISSED_OPTIMIZATION, grid->target_loc,
+ GRID_MISSED_MSG_PREFIX "an inner loop does not "
+ "have use the same collapse clause\n");
+ dump_printf_loc (MSG_NOTE, gimple_location (gfor),
+ "Loop construct uses a different collapse clause\n");
+ }
return false;
}
- clauses = gimple_omp_for_clauses (gfor);
- while (clauses)
+ struct omp_for_data fd;
+ struct omp_for_data_loop *loops
+ = (struct omp_for_data_loop *)alloca (grid->collapse
+ * sizeof (struct omp_for_data_loop));
+ extract_omp_for_data (gfor, &fd, loops);
+ for (unsigned i = 0; i < grid->collapse; i++)
{
- switch (OMP_CLAUSE_CODE (clauses))
+ tree itype, type = TREE_TYPE (fd.loops[i].v);
+ if (POINTER_TYPE_P (type))
+ itype = signed_type_for (type);
+ else
+ itype = type;
+
+ tree n1 = fold_convert (itype, fd.loops[i].n1);
+ tree n2 = fold_convert (itype, fd.loops[i].n2);
+ tree t = build_int_cst (itype,
+ (fd.loops[i].cond_code == LT_EXPR ? -1 : 1));
+ t = fold_build2 (PLUS_EXPR, itype, fd.loops[i].step, t);
+ t = fold_build2 (PLUS_EXPR, itype, t, n2);
+ t = fold_build2 (MINUS_EXPR, itype, t, n1);
+ if (TYPE_UNSIGNED (itype) && fd.loops[i].cond_code == GT_EXPR)
+ t = fold_build2 (TRUNC_DIV_EXPR, itype,
+ fold_build1 (NEGATE_EXPR, itype, t),
+ fold_build1 (NEGATE_EXPR, itype, fd.loops[i].step));
+ else
+ t = fold_build2 (TRUNC_DIV_EXPR, itype, t, fd.loops[i].step);
+
+ if (!operand_equal_p (grid->group_sizes[i], t, 0))
{
- case OMP_CLAUSE_SCHEDULE:
- if (OMP_CLAUSE_SCHEDULE_KIND (clauses) != OMP_CLAUSE_SCHEDULE_AUTO)
+ if (dump_enabled_p ())
+ {
+ dump_printf_loc (MSG_MISSED_OPTIMIZATION, grid->target_loc,
+ GRID_MISSED_MSG_PREFIX "the distribute and "
+ "an internal loop do not agree on tile size\n");
+ dump_printf_loc (MSG_NOTE, gimple_location (gfor),
+ "Loop construct does not seem to loop over "
+ "a tile size\n");
+ }
+ return false;
+ }
+ }
+ return true;
+}
+
+/* Facing a call to FNDECL in the body of a distribute construct, return true
+ if we can handle it or false if it precludes gridification. */
+
+static bool
+grid_call_permissible_in_distribute_p (tree fndecl)
+{
+ if (DECL_PURE_P (fndecl) || TREE_READONLY (fndecl))
+ return true;
+
+ const char *name = IDENTIFIER_POINTER (DECL_NAME (fndecl));
+ if (strstr (name, "omp_") != name)
+ return false;
+
+ if ((strcmp (name, "omp_get_thread_num") == 0)
+ || (strcmp (name, "omp_get_num_threads") == 0)
+ || (strcmp (name, "omp_get_num_teams") == 0)
+ || (strcmp (name, "omp_get_team_num") == 0)
+ || (strcmp (name, "omp_get_level") == 0)
+ || (strcmp (name, "omp_get_active_level") == 0)
+ || (strcmp (name, "omp_in_parallel") == 0))
+ return true;
+
+ return false;
+}
+
+/* Facing a call satisfying grid_call_permissible_in_distribute_p in the body
+ of a distribute construct that is pointed at by GSI, modify it as necessary
+ for gridification. If the statement itself got removed, return true. */
+
+static bool
+grid_handle_call_in_distribute (gimple_stmt_iterator *gsi)
+{
+ gimple *stmt = gsi_stmt (*gsi);
+ tree fndecl = gimple_call_fndecl (stmt);
+ gcc_checking_assert (stmt);
+ if (DECL_PURE_P (fndecl) || TREE_READONLY (fndecl))
+ return false;
+
+ const char *name = IDENTIFIER_POINTER (DECL_NAME (fndecl));
+ if ((strcmp (name, "omp_get_thread_num") == 0)
+ || (strcmp (name, "omp_get_level") == 0)
+ || (strcmp (name, "omp_get_active_level") == 0)
+ || (strcmp (name, "omp_in_parallel") == 0))
+ {
+ tree lhs = gimple_call_lhs (stmt);
+ if (lhs)
+ {
+ gassign *assign
+ = gimple_build_assign (lhs, build_zero_cst (TREE_TYPE (lhs)));
+ gsi_insert_before (gsi, assign, GSI_SAME_STMT);
+ }
+ gsi_remove (gsi, true);
+ return true;
+ }
+
+ /* The rest of the omp functions can stay as they are, HSA back-end will
+ handle them correctly. */
+ gcc_checking_assert ((strcmp (name, "omp_get_num_threads") == 0)
+ || (strcmp (name, "omp_get_num_teams") == 0)
+ || (strcmp (name, "omp_get_team_num") == 0));
+ return false;
+}
+
+/* Given a sequence of statements within a distribute omp construct or a
+ parallel construct, which in the original source does not form a compound
+ construct with a looping construct, return true if it does not prevent us
+ from turning it into a gridified GPGPU kernel. Otherwise return false. GRID
+ describes hitherto discovered properties of the loop that is evaluated for
+ possible gridification. IN_PARALLEL must be true if seq is within a
+ parallel construct and flase if it is only within a distribute
+ construct. */
+
+static bool
+grid_dist_follows_tiling_pattern (gimple_seq seq, grid_prop *grid,
+ bool in_parallel)
+{
+ gimple_stmt_iterator gsi;
+ for (gsi = gsi_start (seq); !gsi_end_p (gsi); gsi_next (&gsi))
+ {
+ gimple *stmt = gsi_stmt (gsi);
+
+ if (grid_safe_assignment_p (stmt, grid)
+ || gimple_code (stmt) == GIMPLE_GOTO
+ || gimple_code (stmt) == GIMPLE_LABEL
+ || gimple_code (stmt) == GIMPLE_COND)
+ continue;
+ else if (gbind *bind = dyn_cast <gbind *> (stmt))
+ {
+ if (!grid_dist_follows_tiling_pattern (gimple_bind_body (bind),
+ grid, in_parallel))
+ return false;
+ continue;
+ }
+ else if (gtry *try_stmt = dyn_cast <gtry *> (stmt))
+ {
+ if (gimple_try_kind (try_stmt) == GIMPLE_TRY_CATCH)
{
if (dump_enabled_p ())
- dump_printf_loc (MSG_MISSED_OPTIMIZATION, tloc,
- "Will not turn target construct into a "
- "gridified GPGPU kernel because the inner "
- "loop has a non-automatic scheduling clause\n");
+ {
+ dump_printf_loc (MSG_MISSED_OPTIMIZATION, grid->target_loc,
+ GRID_MISSED_MSG_PREFIX "the distribute "
+ "construct contains a try..catch region\n");
+ dump_printf_loc (MSG_NOTE, gimple_location (try_stmt),
+ "This statement cannot be analyzed for "
+ "tiled gridification\n");
+ }
return false;
}
- break;
+ if (!grid_dist_follows_tiling_pattern (gimple_try_eval (try_stmt),
+ grid, in_parallel))
+ return false;
+ if (!grid_dist_follows_tiling_pattern (gimple_try_cleanup (try_stmt),
+ grid, in_parallel))
+ return false;
+ continue;
+ }
+ else if (is_gimple_call (stmt))
+ {
+ tree fndecl = gimple_call_fndecl (stmt);
+ if (fndecl && grid_call_permissible_in_distribute_p (fndecl))
+ continue;
+
+ if (dump_enabled_p ())
+ {
+ dump_printf_loc (MSG_MISSED_OPTIMIZATION, grid->target_loc,
+ GRID_MISSED_MSG_PREFIX "the distribute "
+ "construct contains a call\n");
+ dump_printf_loc (MSG_NOTE, gimple_location (stmt),
+ "This statement cannot be analyzed for "
+ "tiled gridification\n");
+ }
+ return false;
+ }
+ else if (gomp_parallel *par = dyn_cast <gomp_parallel *> (stmt))
+ {
+ if (in_parallel)
+ {
+ if (dump_enabled_p ())
+ {
+ dump_printf_loc (MSG_MISSED_OPTIMIZATION, grid->target_loc,
+ GRID_MISSED_MSG_PREFIX "a parallel "
+ "construct contains another parallel "
+ "construct\n");
+ dump_printf_loc (MSG_NOTE, gimple_location (stmt),
+ "This parallel construct is nested in "
+ "another one\n");
+ }
+ return false;
+ }
+ if (!grid_parallel_clauses_gridifiable (par, grid->target_loc)
+ || !grid_dist_follows_tiling_pattern (gimple_omp_body (par),
+ grid, true))
+ return false;
+ }
+ else if (gomp_for *gfor = dyn_cast <gomp_for *> (stmt))
+ {
+ if (!in_parallel)
+ {
+ if (dump_enabled_p ())
+ {
+ dump_printf_loc (MSG_MISSED_OPTIMIZATION, grid->target_loc,
+ GRID_MISSED_MSG_PREFIX "a loop "
+ "construct is not nested within a parallel "
+ "construct\n");
+ dump_printf_loc (MSG_NOTE, gimple_location (stmt),
+ "This loop construct is not nested in "
+ "a parallel construct\n");
+ }
+ return false;
+ }
+ if (!grid_gfor_follows_tiling_pattern (gfor, grid))
+ return false;
+ }
+ else
+ {
+ if (dump_enabled_p ())
+ {
+ dump_printf_loc (MSG_MISSED_OPTIMIZATION, grid->target_loc,
+ GRID_MISSED_MSG_PREFIX "the distribute "
+ "construct contains a complex statement\n");
+ dump_printf_loc (MSG_NOTE, gimple_location (stmt),
+ "This statement cannot be analyzed for "
+ "tiled gridification\n");
+ }
+ return false;
+ }
+ }
+ return true;
+}
+
+/* If TARGET follows a pattern that can be turned into a gridified GPGPU
+ kernel, return true, otherwise return false. In the case of success, also
+ fill in GROUP_SIZE_P with the requested group size or NULL if there is
+ none. */
+
+static bool
+grid_target_follows_gridifiable_pattern (gomp_target *target, grid_prop *grid)
+{
+ if (gimple_omp_target_kind (target) != GF_OMP_TARGET_KIND_REGION)
+ return false;
+
+ location_t tloc = gimple_location (target);
+ grid->target_loc = tloc;
+ gimple *stmt
+ = grid_find_single_omp_among_assignments (gimple_omp_body (target),
+ grid, "target");
+ if (!stmt)
+ return false;
+ gomp_teams *teams = dyn_cast <gomp_teams *> (stmt);
+ tree group_size = NULL;
+ if (!teams)
+ {
+ dump_printf_loc (MSG_MISSED_OPTIMIZATION, tloc,
+ GRID_MISSED_MSG_PREFIX "it does not have a sole teams "
+ "construct in it.\n");
+ return false;
+ }
+
+ tree clauses = gimple_omp_teams_clauses (teams);
+ while (clauses)
+ {
+ switch (OMP_CLAUSE_CODE (clauses))
+ {
+ case OMP_CLAUSE_NUM_TEAMS:
+ if (dump_enabled_p ())
+ dump_printf_loc (MSG_MISSED_OPTIMIZATION, tloc,
+ GRID_MISSED_MSG_PREFIX "the teams construct "
+ "contains a num_teams clause\n ");
+ return false;
case OMP_CLAUSE_REDUCTION:
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, tloc,
- "Will not turn target construct into a "
- "gridified GPGPU kernel because a reduction "
+ GRID_MISSED_MSG_PREFIX "a reduction "
"clause is present\n ");
return false;
case OMP_CLAUSE_LASTPRIVATE:
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, tloc,
- "Will not turn target construct into a "
- "gridified GPGPU kernel because a lastprivate "
+ GRID_MISSED_MSG_PREFIX "a lastprivate "
"clause is present\n ");
return false;
+ case OMP_CLAUSE_THREAD_LIMIT:
+ if (!integer_zerop (OMP_CLAUSE_OPERAND (clauses, 0)))
+ group_size = OMP_CLAUSE_OPERAND (clauses, 0);
+ break;
+
default:
break;
}
clauses = OMP_CLAUSE_CHAIN (clauses);
}
- struct walk_stmt_info wi;
- memset (&wi, 0, sizeof (wi));
- if (walk_gimple_seq (gimple_omp_body (gfor),
- grid_find_ungridifiable_statement,
- NULL, &wi))
+ stmt = grid_find_single_omp_among_assignments (gimple_omp_body (teams), grid,
+ "teams");
+ if (!stmt)
+ return false;
+ gomp_for *dist = dyn_cast <gomp_for *> (stmt);
+ if (!dist)
+ {
+ dump_printf_loc (MSG_MISSED_OPTIMIZATION, tloc,
+ GRID_MISSED_MSG_PREFIX "the teams construct does not "
+ "have a single distribute construct in it.\n");
+ return false;
+ }
+
+ gcc_assert (gimple_omp_for_kind (dist) == GF_OMP_FOR_KIND_DISTRIBUTE);
+
+ grid->collapse = gimple_omp_for_collapse (dist);
+ if (grid->collapse > 3)
{
- gimple *bad = (gimple *) wi.info;
if (dump_enabled_p ())
+ dump_printf_loc (MSG_MISSED_OPTIMIZATION, tloc,
+ GRID_MISSED_MSG_PREFIX "the distribute construct "
+ "contains collapse clause with parameter greater "
+ "than 3\n");
+ return false;
+ }
+
+ struct omp_for_data fd;
+ struct omp_for_data_loop *dist_loops
+ = (struct omp_for_data_loop *)alloca (grid->collapse
+ * sizeof (struct omp_for_data_loop));
+ extract_omp_for_data (dist, &fd, dist_loops);
+ if (fd.chunk_size)
+ {
+ if (group_size && !operand_equal_p (group_size, fd.chunk_size, 0))
{
- if (is_gimple_call (bad))
- dump_printf_loc (MSG_MISSED_OPTIMIZATION, tloc,
- "Will not turn target construct into a gridified "
- " GPGPU kernel because the inner loop contains "
- "call to a noreturn function\n");
- if (gimple_code (bad) == GIMPLE_OMP_FOR)
- dump_printf_loc (MSG_MISSED_OPTIMIZATION, tloc,
- "Will not turn target construct into a gridified "
- " GPGPU kernel because the inner loop contains "
- "a simd construct\n");
- else
+ if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, tloc,
- "Will not turn target construct into a gridified "
- "GPGPU kernel because the inner loop contains "
- "statement %s which cannot be transformed\n",
- gimple_code_name[(int) gimple_code (bad)]);
+ GRID_MISSED_MSG_PREFIX "the teams "
+ "thread limit is different from distribute "
+ "schedule chunk\n");
+ return false;
}
+ group_size = fd.chunk_size;
+ }
+ if (group_size && grid->collapse > 1)
+ {
+ if (dump_enabled_p ())
+ dump_printf_loc (MSG_MISSED_OPTIMIZATION, tloc,
+ GRID_MISSED_MSG_PREFIX "group size cannot be "
+ "set using thread_limit or schedule clauses "
+ "when also using a collapse clause greater than 1\n");
return false;
}
- *group_size_p = group_size;
- return true;
+ if (gimple_omp_for_combined_p (dist))
+ {
+ grid->tiling = false;
+ grid->group_sizes[0] = group_size;
+ for (unsigned i = 1; i < grid->collapse; i++)
+ grid->group_sizes[i] = NULL;
+ return grid_dist_follows_simple_pattern (dist, grid);
+ }
+ else
+ {
+ grid->tiling = true;
+ if (group_size)
+ {
+ if (dump_enabled_p ())
+ dump_printf_loc (MSG_MISSED_OPTIMIZATION, tloc,
+ GRID_MISSED_MSG_PREFIX "group size cannot be set "
+ "using thread_limit or schedule clauses when "
+ "distribute and loop constructs do not form "
+ "one combined construct\n");
+ return false;
+ }
+ for (unsigned i = 0; i < grid->collapse; i++)
+ {
+ if (fd.loops[i].cond_code == GT_EXPR)
+ grid->group_sizes[i] = fold_build1 (NEGATE_EXPR,
+ TREE_TYPE (fd.loops[i].step),
+ fd.loops[i].step);
+ else
+ grid->group_sizes[i] = fd.loops[i].step;
+ }
+ return grid_dist_follows_tiling_pattern (gimple_omp_body (dist), grid,
+ false);
+ }
}
/* Operand walker, used to remap pre-body declarations according to a hash map
@@ -17623,15 +18058,62 @@ grid_remap_prebody_decls (tree *tp, int *walk_subtrees, void *data)
return NULL_TREE;
}
+/* Identifiers of segments into which a particular variable should be places
+ when gridifying. */
+
+enum grid_var_segment {GRID_SEGMENT_PRIVATE, GRID_SEGMENT_GROUP,
+ GRID_SEGMENT_GLOBAL};
+
+/* Mark VAR so that it is eventually placed into SEGMENT. Place an artificial
+ builtin call into SEQ that will make sure the variable is always considered
+ address taken. */
+
+static void
+grid_mark_variable_segment (tree var, enum grid_var_segment segment)
+{
+ /* Making a non-addressable variables would require that we re-gimplify all
+ their uses. Fortunately, we do not have to do this because if they are
+ not addressable, it means they are not used in atomic or parallel
+ statements and so relaxed GPU consistency rules mean we can just keep them
+ private. */
+ if (!TREE_ADDRESSABLE (var))
+ return;
+
+ switch (segment)
+ {
+ case GRID_SEGMENT_GROUP:
+ DECL_ATTRIBUTES (var) = tree_cons (get_identifier ("hsa_group_segment"),
+ NULL, DECL_ATTRIBUTES (var));
+ break;
+ case GRID_SEGMENT_GLOBAL:
+ DECL_ATTRIBUTES (var) = tree_cons (get_identifier ("hsa_global_segment"),
+ NULL, DECL_ATTRIBUTES (var));
+ break;
+ default:
+ gcc_unreachable ();
+ }
+
+ if (!TREE_STATIC (var))
+ {
+ TREE_STATIC (var) = 1;
+ varpool_node::finalize_decl (var);
+ }
+
+}
+
/* Copy leading register-type assignments to local variables in SRC to just
before DST, Creating temporaries, adjusting mapping of operands in WI and
remapping operands as necessary. Add any new temporaries to TGT_BIND.
- Return the first statement that does not conform to
- grid_reg_assignment_to_local_var_p or NULL. */
+ Return the first statement that does not conform to grid_safe_assignment_p
+ or NULL. If VAR_SEGMENT is not GRID_SEGMENT_PRIVATE, also mark all
+ variables in traversed bind statements so that they are put into the
+ appropriate segment. */
static gimple *
grid_copy_leading_local_assignments (gimple_seq src, gimple_stmt_iterator *dst,
- gbind *tgt_bind, struct walk_stmt_info *wi)
+ gbind *tgt_bind,
+ enum grid_var_segment var_segment,
+ struct walk_stmt_info *wi)
{
hash_map<tree, tree> *declmap = (hash_map<tree, tree> *) wi->info;
gimple_stmt_iterator gsi;
@@ -17641,13 +18123,17 @@ grid_copy_leading_local_assignments (gimple_seq src, gimple_stmt_iterator *dst,
if (gbind *bind = dyn_cast <gbind *> (stmt))
{
gimple *r = grid_copy_leading_local_assignments
- (gimple_bind_body (bind), dst, tgt_bind, wi);
+ (gimple_bind_body (bind), dst, tgt_bind, var_segment, wi);
+
+ if (var_segment != GRID_SEGMENT_PRIVATE)
+ for (tree var = gimple_bind_vars (bind); var; var = DECL_CHAIN (var))
+ grid_mark_variable_segment (var, var_segment);
if (r)
return r;
else
continue;
}
- if (!grid_reg_assignment_to_local_var_p (stmt))
+ if (!grid_safe_assignment_p (stmt, NULL))
return stmt;
tree lhs = gimple_assign_lhs (as_a <gassign *> (stmt));
tree repl = copy_var_decl (lhs, create_tmp_var_name (NULL),
@@ -17663,43 +18149,108 @@ grid_copy_leading_local_assignments (gimple_seq src, gimple_stmt_iterator *dst,
return NULL;
}
+/* Statement walker function marking all parallels as grid_phony and loops as
+ grid ones representing threads of a particular thread group. */
+
+static tree
+grid_mark_tiling_loops (gimple_stmt_iterator *gsi,
+ bool *handled_ops_p,
+ struct walk_stmt_info *)
+{
+ *handled_ops_p = false;
+ if (gomp_for *loop = dyn_cast <gomp_for *> (gsi_stmt (*gsi)))
+ {
+ *handled_ops_p = true;
+ gimple_omp_for_set_kind (loop, GF_OMP_FOR_KIND_GRID_LOOP);
+ }
+ return NULL_TREE;
+}
+
+/* Statement walker function marking all parallels as grid_phony and loops as
+ grid ones representing threads of a particular thread group. */
+
+static tree
+grid_mark_tiling_parallels_and_loops (gimple_stmt_iterator *gsi,
+ bool *handled_ops_p,
+ struct walk_stmt_info *wi)
+{
+ *handled_ops_p = false;
+ wi->removed_stmt = false;
+ gimple *stmt = gsi_stmt (*gsi);
+ if (gbind *bind = dyn_cast <gbind *> (stmt))
+ {
+ for (tree var = gimple_bind_vars (bind); var; var = DECL_CHAIN (var))
+ grid_mark_variable_segment (var, GRID_SEGMENT_GROUP);
+ }
+ else if (gomp_parallel *parallel = dyn_cast <gomp_parallel *> (stmt))
+ {
+ *handled_ops_p = true;
+ gimple_omp_parallel_set_grid_phony (parallel, true);
+ walk_gimple_seq_mod (gimple_omp_body_ptr (parallel),
+ grid_mark_tiling_loops, NULL, wi);
+ }
+ else if (is_a <gcall *> (stmt))
+ wi->removed_stmt = grid_handle_call_in_distribute (gsi);
+ return NULL_TREE;
+}
+
/* Given freshly copied top level kernel SEQ, identify the individual OMP
- components, mark them as part of kernel and return the inner loop, and copy
- assignment leading to them just before DST, remapping them using WI and
- adding new temporaries to TGT_BIND. */
+ components, mark them as part of kernel, copy assignment leading to them
+ just before DST, remapping them using WI and adding new temporaries to
+ TGT_BIND, and and return the loop that will be used for kernel dispatch. */
static gomp_for *
-grid_process_kernel_body_copy (gimple_seq seq, gimple_stmt_iterator *dst,
+grid_process_kernel_body_copy (grid_prop *grid, gimple_seq seq,
+ gimple_stmt_iterator *dst,
gbind *tgt_bind, struct walk_stmt_info *wi)
{
- gimple *stmt = grid_copy_leading_local_assignments (seq, dst, tgt_bind, wi);
+ gimple *stmt = grid_copy_leading_local_assignments (seq, dst, tgt_bind,
+ GRID_SEGMENT_GLOBAL, wi);
gomp_teams *teams = dyn_cast <gomp_teams *> (stmt);
gcc_assert (teams);
gimple_omp_teams_set_grid_phony (teams, true);
stmt = grid_copy_leading_local_assignments (gimple_omp_body (teams), dst,
- tgt_bind, wi);
+ tgt_bind, GRID_SEGMENT_GLOBAL, wi);
gcc_checking_assert (stmt);
gomp_for *dist = dyn_cast <gomp_for *> (stmt);
gcc_assert (dist);
gimple_seq prebody = gimple_omp_for_pre_body (dist);
if (prebody)
- grid_copy_leading_local_assignments (prebody, dst, tgt_bind, wi);
- gimple_omp_for_set_grid_phony (dist, true);
- stmt = grid_copy_leading_local_assignments (gimple_omp_body (dist), dst,
- tgt_bind, wi);
- gcc_checking_assert (stmt);
+ grid_copy_leading_local_assignments (prebody, dst, tgt_bind,
+ GRID_SEGMENT_GROUP, wi);
- gomp_parallel *parallel = as_a <gomp_parallel *> (stmt);
- gimple_omp_parallel_set_grid_phony (parallel, true);
- stmt = grid_copy_leading_local_assignments (gimple_omp_body (parallel), dst,
- tgt_bind, wi);
- gomp_for *inner_loop = as_a <gomp_for *> (stmt);
- gimple_omp_for_set_kind (inner_loop, GF_OMP_FOR_KIND_GRID_LOOP);
- prebody = gimple_omp_for_pre_body (inner_loop);
- if (prebody)
- grid_copy_leading_local_assignments (prebody, dst, tgt_bind, wi);
+ if (grid->tiling)
+ {
+ gimple_omp_for_set_kind (dist, GF_OMP_FOR_KIND_GRID_LOOP);
+ gimple_omp_for_set_grid_group_iter (dist, true);
- return inner_loop;
+ struct walk_stmt_info wi;
+ memset (&wi, 0, sizeof (wi));
+ walk_gimple_seq_mod (gimple_omp_body_ptr (dist),
+ grid_mark_tiling_parallels_and_loops, NULL, &wi);
+ return dist;
+ }
+ else
+ {
+ gimple_omp_for_set_grid_phony (dist, true);
+ stmt = grid_copy_leading_local_assignments (gimple_omp_body (dist), dst,
+ tgt_bind,
+ GRID_SEGMENT_PRIVATE, wi);
+ gcc_checking_assert (stmt);
+ gomp_parallel *parallel = as_a <gomp_parallel *> (stmt);
+ gimple_omp_parallel_set_grid_phony (parallel, true);
+ stmt = grid_copy_leading_local_assignments (gimple_omp_body (parallel),
+ dst, tgt_bind,
+ GRID_SEGMENT_PRIVATE, wi);
+ gomp_for *inner_loop = as_a <gomp_for *> (stmt);
+ gimple_omp_for_set_kind (inner_loop, GF_OMP_FOR_KIND_GRID_LOOP);
+ prebody = gimple_omp_for_pre_body (inner_loop);
+ if (prebody)
+ grid_copy_leading_local_assignments (prebody, dst, tgt_bind,
+ GRID_SEGMENT_PRIVATE, wi);
+
+ return inner_loop;
+ }
}
/* If TARGET points to a GOMP_TARGET which follows a gridifiable pattern,
@@ -17712,8 +18263,10 @@ grid_attempt_target_gridification (gomp_target *target,
gimple_stmt_iterator *gsi,
gbind *tgt_bind)
{
- tree group_size;
- if (!target || !grid_target_follows_gridifiable_pattern (target, &group_size))
+ /* removed group_size */
+ grid_prop grid;
+ memset (&grid, 0, sizeof (grid));
+ if (!target || !grid_target_follows_gridifiable_pattern (target, &grid))
return;
location_t loc = gimple_location (target);
@@ -17732,8 +18285,8 @@ grid_attempt_target_gridification (gomp_target *target,
wi.info = declmap;
/* Copy assignments in between OMP statements before target, mark OMP
- statements within copy appropriatly. */
- gomp_for *inner_loop = grid_process_kernel_body_copy (kernel_seq, gsi,
+ statements within copy appropriately. */
+ gomp_for *inner_loop = grid_process_kernel_body_copy (&grid, kernel_seq, gsi,
tgt_bind, &wi);
gbind *old_bind = as_a <gbind *> (gimple_seq_first (gimple_omp_body (target)));
@@ -17748,10 +18301,10 @@ grid_attempt_target_gridification (gomp_target *target,
(gimple_bind_body_ptr (as_a <gbind *> (gimple_omp_body (target))),
gpukernel);
- walk_tree (&group_size, grid_remap_prebody_decls, &wi, NULL);
+ for (size_t i = 0; i < grid.collapse; i++)
+ walk_tree (&grid.group_sizes[i], grid_remap_prebody_decls, &wi, NULL);
push_gimplify_context ();
- size_t collapse = gimple_omp_for_collapse (inner_loop);
- for (size_t i = 0; i < collapse; i++)
+ for (size_t i = 0; i < grid.collapse; i++)
{
tree itype, type = TREE_TYPE (gimple_omp_for_index (inner_loop, i));
if (POINTER_TYPE_P (type))
@@ -17765,12 +18318,12 @@ grid_attempt_target_gridification (gomp_target *target,
tree n2 = unshare_expr (gimple_omp_for_final (inner_loop, i));
walk_tree (&n2, grid_remap_prebody_decls, &wi, NULL);
adjust_for_condition (loc, &cond_code, &n2);
- tree step;
- step = get_omp_for_step_from_incr (loc,
- gimple_omp_for_incr (inner_loop, i));
- gimple_seq tmpseq = NULL;
n1 = fold_convert (itype, n1);
n2 = fold_convert (itype, n2);
+
+ tree step
+ = get_omp_for_step_from_incr (loc, gimple_omp_for_incr (inner_loop, i));
+
tree t = build_int_cst (itype, (cond_code == LT_EXPR ? -1 : 1));
t = fold_build2 (PLUS_EXPR, itype, step, t);
t = fold_build2 (PLUS_EXPR, itype, t, n2);
@@ -17781,15 +18334,23 @@ grid_attempt_target_gridification (gomp_target *target,
fold_build1 (NEGATE_EXPR, itype, step));
else
t = fold_build2 (TRUNC_DIV_EXPR, itype, t, step);
+ if (grid.tiling)
+ {
+ if (cond_code == GT_EXPR)
+ step = fold_build1 (NEGATE_EXPR, itype, step);
+ t = fold_build2 (MULT_EXPR, itype, t, step);
+ }
+
tree gs = fold_convert (uint32_type_node, t);
+ gimple_seq tmpseq = NULL;
gimplify_expr (&gs, &tmpseq, NULL, is_gimple_val, fb_rvalue);
if (!gimple_seq_empty_p (tmpseq))
gsi_insert_seq_before (gsi, tmpseq, GSI_SAME_STMT);
tree ws;
- if (i == 0 && group_size)
+ if (grid.group_sizes[i])
{
- ws = fold_convert (uint32_type_node, group_size);
+ ws = fold_convert (uint32_type_node, grid.group_sizes[i]);
tmpseq = NULL;
gimplify_expr (&ws, &tmpseq, NULL, is_gimple_val, fb_rvalue);
if (!gimple_seq_empty_p (tmpseq))
new file mode 100644
@@ -0,0 +1,66 @@
+/* { dg-do compile } */
+/* { dg-require-effective-target offload_hsa } */
+/* { dg-options "-fopenmp -fdump-tree-omplower-details" } */
+
+#define BLOCK_SIZE 16
+
+
+void tiled_sgemm_tt(const int M, const int N, const int K, const float alpha, const float*A, const int LDA,
+ const float*B, const int LDB, const float beta, float*C, const int LDC){
+
+#pragma omp target teams map(to:A[M*K],B[K*N]) map(from:C[M*N])
+#pragma omp distribute collapse(2)
+ for (int C_row_start=0 ; C_row_start < M ; C_row_start+=BLOCK_SIZE)
+ for (int C_col_start=0 ; C_col_start < N ; C_col_start+=BLOCK_SIZE)
+ {
+// Each team has a local copy of these mini matrices
+ float As[BLOCK_SIZE][BLOCK_SIZE];
+ float Bs[BLOCK_SIZE][BLOCK_SIZE];
+#pragma omp parallel
+ {
+ int C_row, C_col;
+ float Cval = 0.0;
+
+ for (int kblock = 0; kblock < K ; kblock += BLOCK_SIZE )
+ {
+#pragma omp for collapse(2)
+ for (int row=0 ; row < BLOCK_SIZE ; row++)
+ for (int col=0 ; col < BLOCK_SIZE ; col++)
+ {
+ C_row = C_row_start + row;
+ C_col = C_col_start + col;
+ if ((C_row < M) && (kblock + col < K))
+ As[row][col] = A[(C_row*LDA)+ kblock + col];
+ else
+ As[row][col] = 0;
+ if ((kblock + row < K) && C_col < N)
+ Bs[row][col] = B[((kblock+row)*LDB)+ C_col];
+ else
+ Bs[row][col] = 0;
+ }
+
+#pragma omp for collapse(2)
+ for (int row=0 ; row < BLOCK_SIZE ; row++)
+ for (int col=0 ; col < BLOCK_SIZE ; col++)
+ {
+ for (int e = 0; e < BLOCK_SIZE; ++e)
+ Cval += As[row][e] * Bs[e][col];
+ }
+ } /* End for kblock .. */
+
+
+#pragma omp for collapse(2)
+ for (int row=0 ; row < BLOCK_SIZE ; row++)
+ for (int col=0 ; col < BLOCK_SIZE ; col++)
+ {
+ C_row = C_row_start + row;
+ C_col = C_col_start + col;
+ if ((C_row < M) && (C_col < N))
+ C[(C_row*LDC)+C_col] = alpha*Cval + beta*C[(C_row*LDC)+C_col];
+
+ }
+ } /* end parallel */
+ } /* end target teams distribute */
+}
+
+/* { dg-final { scan-tree-dump "Target construct will be turned into a gridified GPGPU kernel" "omplower" } } */
new file mode 100644
@@ -0,0 +1,68 @@
+/* { dg-do compile } */
+/* { dg-require-effective-target offload_hsa } */
+/* { dg-options "-fopenmp -fdump-tree-omplower-details" } */
+
+#define BLOCK_SIZE 16
+
+void tiled_sgemm_tt(const int M, const int N, const int K, const float alpha, const float*A, const int LDA,
+ const float*B, const int LDB, const float beta, float*C, const int LDC)
+{
+#pragma omp target teams map(to:A[M*K],B[K*N]) map(from:C[M*N])
+#pragma omp distribute collapse(2)
+ for (int C_row_start=0 ; C_row_start < M ; C_row_start+=BLOCK_SIZE)
+ for (int C_col_start=0 ; C_col_start < N ; C_col_start+=BLOCK_SIZE)
+ {
+ float As[BLOCK_SIZE][BLOCK_SIZE];
+ float Bs[BLOCK_SIZE][BLOCK_SIZE];
+ float Cs[BLOCK_SIZE][BLOCK_SIZE];
+ int C_row, C_col;
+
+#pragma omp parallel for collapse(2)
+ for (int row=0 ; row < BLOCK_SIZE ; row++)
+ for (int col=0 ; col < BLOCK_SIZE ; col++)
+ {
+ Cs[row][col] = 0.0;
+ }
+
+
+ for (int kblock = 0; kblock < K ; kblock += BLOCK_SIZE )
+ {
+#pragma omp parallel for collapse(2)
+ for (int row=0 ; row < BLOCK_SIZE ; row++)
+ for (int col=0 ; col < BLOCK_SIZE ; col++)
+ {
+ C_row = C_row_start + row;
+ C_col = C_col_start + col;
+ if ((C_row < M) && (kblock + col < K))
+ As[row][col] = A[(C_row*LDA)+ kblock + col];
+ else
+ As[row][col] = 0;
+ if ((kblock + row < K) && C_col < N)
+ Bs[row][col] = B[((kblock+row)*LDB)+ C_col];
+ else
+ Bs[row][col] = 0;
+ }
+
+#pragma omp parallel for collapse(2)
+ for (int row=0 ; row < BLOCK_SIZE ; row++)
+ for (int col=0 ; col < BLOCK_SIZE ; col++)
+ {
+ for (int e = 0; e < BLOCK_SIZE; ++e)
+ Cs[row][col] += As[row][e] * Bs[e][col];
+ }
+ } /* End for kblock .. */
+
+
+#pragma omp parallel for collapse(2)
+ for (int row=0 ; row < BLOCK_SIZE ; row++)
+ for (int col=0 ; col < BLOCK_SIZE ; col++)
+ {
+ C_row = C_row_start + row;
+ C_col = C_col_start + col;
+ if ((C_row < M) && (C_col < N))
+ C[(C_row*LDC)+C_col] = alpha*Cs[row][col] + beta*C[(C_row*LDC)+C_col];
+ }
+ } /* End distribute */
+}
+
+/* { dg-final { scan-tree-dump "Target construct will be turned into a gridified GPGPU kernel" "omplower" } } */
new file mode 100644
@@ -0,0 +1,212 @@
+/*
+
+ matmul.c : Matrix Multiplication with tiling for openmp4 example
+
+*/
+
+#include <stdlib.h>
+#include <math.h>
+
+#define BLOCK_SIZE 16
+/*
+ #define BLOCK_SIZE 32
+*/
+#define NSECPERSEC 1000000000L
+
+typedef struct {
+ int width;
+ int height;
+ int stride;
+ int hpad;
+ float* elements;
+} Matrix;
+
+/* Correctly extract the number of nanoseconds from the two time structures */
+long int get_nanosecs( struct timespec start_time, struct timespec end_time) {
+ long int nanosecs;
+ if ((end_time.tv_nsec-start_time.tv_nsec)<0) nanosecs =
+ ((((long int) end_time.tv_sec- (long int) start_time.tv_sec )-1)*NSECPERSEC ) +
+ ( NSECPERSEC + (long int) end_time.tv_nsec - (long int) start_time.tv_nsec) ;
+ else nanosecs =
+ (((long int) end_time.tv_sec- (long int) start_time.tv_sec )*NSECPERSEC ) +
+ ( (long int) end_time.tv_nsec - (long int) start_time.tv_nsec );
+ return nanosecs;
+}
+
+void simple_sgemm_tt(const int M,const int N,const int K,const float alpha, const float* A,const int LDA,
+ const float* B,const int LDB, const float beta,float* C, const int LDC) ;
+void simple_sgemm_tn(const int M,const int N,const int K,const float alpha, const float* A,const int LDA,
+ const float* B,const int LDB, const float beta,float* C, const int LDC) ;
+void tiled_sgemm_tt(const int M,const int N,const int K,const float alpha, const float*A, const int LDA,
+ const float* B,const int LDB, const float beta,float* C, const int LDC) ;
+
+int verify(float* v_res, float* v_ref, int len) {
+ int passed = 1;
+ int i;
+ for (i = 0; i < len; ++i) {
+ if (fabs(v_res[i] - v_ref[i]) > 0.001*v_ref[i]) {
+ __builtin_abort ();
+ }
+ }
+ return passed;
+}
+
+
+int main(int argc, char* argv[]){
+
+ Matrix A,B,Bt,C,Cref;
+ int a1,a2,a3,i,j;
+ struct timespec start_time1, end_time1;
+ struct timespec start_time2, end_time2;
+ long int nanosecs,total_ops;
+ float gflopsTiled,gflopsCPU;
+
+ a1 = 35;
+ a2 = 28;
+ a3 = 47;
+
+ A.height = a1;
+ A.width = a2;
+ A.stride = (((A.width-1)/BLOCK_SIZE)+1) * BLOCK_SIZE;
+ A.hpad = (((A.height-1)/BLOCK_SIZE)+1) * BLOCK_SIZE;
+ A.elements = (float*)malloc(A.stride * A.hpad* sizeof(float));
+
+ B.height = a2;
+ B.width = a3;
+ B.stride = (((B.width-1)/BLOCK_SIZE)+1) * BLOCK_SIZE;
+ B.hpad = (((B.height-1)/BLOCK_SIZE)+1) * BLOCK_SIZE;
+ B.elements = (float*)malloc(B.stride * B.hpad * sizeof(float));
+
+ /* Bt is same as B but stored in column-major order */
+ Bt.height = B.height;
+ Bt.width = B.width;
+ Bt.stride = B.stride;
+ Bt.hpad = B.hpad;
+ Bt.elements = (float*)malloc(Bt.stride * Bt.hpad * sizeof(float));
+
+ C.height = a1;
+ C.width = a3;
+ C.stride = (((C.width-1)/BLOCK_SIZE)+1) * BLOCK_SIZE;
+ C.hpad = (((C.height-1)/BLOCK_SIZE)+1) * BLOCK_SIZE;
+ C.elements = (float*)malloc(C.stride * C.hpad * sizeof(float));
+
+ Cref.height = a1;
+ Cref.width = a3;
+ Cref.stride = (((Cref.width-1)/BLOCK_SIZE)+1) * BLOCK_SIZE;
+ Cref.hpad = (((Cref.height-1)/BLOCK_SIZE)+1) * BLOCK_SIZE;
+ Cref.elements = (float*)malloc(Cref.stride * Cref.hpad * sizeof(float));
+
+ for(i = 0; i < A.hpad ; i++)
+ for(j = 0; j < A.stride; j++) {
+ if (( j<A.width ) && (i<A.height)) {
+ A.elements[i*A.stride + j] = (i % 3);
+ } else {
+ A.elements[i*A.stride + j] = 0.0;
+ }
+ }
+
+ /* Initialize B and Bt */
+ for(i = 0; i < B.hpad ; i++)
+ for(j = 0; j < B.stride; j++) {
+ if (( j<B.width ) && (i<B.height)) {
+ B.elements[i*B.stride+j] = (j % 2);
+ Bt.elements[j*Bt.stride+i] = B.elements[i*B.stride+j] ;
+ } else {
+ B.elements[i*B.stride+j] = 0.0;
+ Bt.elements[j*Bt.stride+i] = 0.0;
+ }
+ }
+
+ /* zero C, and Cref */
+ for(i = 0; i < C.hpad; i++)
+ for(j = 0; j < C.stride; j++) {
+ C.elements[i*C.stride+j] = 0.0;
+ Cref.elements[i*Cref.stride+j] = 0.0;
+ }
+
+ simple_sgemm_tt(A.height,B.width,B.height,1.0,A.elements,A.stride,B.elements,B.stride,1.0,Cref.elements,Cref.stride);
+ tiled_sgemm_tt(A.height,B.width,B.height,1.0,A.elements,A.stride,B.elements,B.stride,1.0,C.elements,C.stride);
+
+ verify(C.elements, Cref.elements, C.height * C.stride);
+ return 0;
+}
+
+void simple_sgemm_tt(const int M,const int N,const int K,const float alpha, const float* A,const int LDA,
+const float* B,const int LDB, const float beta,float* C, const int LDC) {
+ /* A,B, and C are in row-major order */
+ int c_row,c_col,inner;
+ float sum;
+ for (c_col = 0 ; c_col<N; c_col++ ) {
+ for (c_row = 0 ; c_row<M; c_row++ ) {
+ sum = 0.0 ;
+ for (inner = 0 ; inner<K; inner++ ) {
+ sum += A[c_row*LDA + inner] * B[inner*LDB + c_col] ;
+ }
+ C[c_row*LDC + c_col] = alpha*sum + beta*C[ c_row*LDC + c_col] ;
+ }
+ }
+}
+
+/***************************
+
+ tiled_sgemm_tt: Tiled matrix multiplication:
+
+***************************/
+
+void tiled_sgemm_tt(const int M, const int N, const int K, const float alpha, const float*A, const int LDA,
+ const float*B, const int LDB, const float beta, float*C, const int LDC){
+
+#pragma omp target teams map(to:A[M*K],B[K*N]) map(from:C[M*N])
+#pragma omp distribute collapse(2)
+ for (int C_row_start=0 ; C_row_start < M ; C_row_start+=BLOCK_SIZE)
+ for (int C_col_start=0 ; C_col_start < N ; C_col_start+=BLOCK_SIZE)
+ {
+// Each team has a local copy of these mini matrices
+ float As[BLOCK_SIZE][BLOCK_SIZE];
+ float Bs[BLOCK_SIZE][BLOCK_SIZE];
+#pragma omp parallel
+ {
+ int C_row, C_col;
+ float Cval = 0.0;
+
+ for (int kblock = 0; kblock < K ; kblock += BLOCK_SIZE )
+ {
+#pragma omp for collapse(2)
+ for (int row=0 ; row < BLOCK_SIZE ; row++)
+ for (int col=0 ; col < BLOCK_SIZE ; col++)
+ {
+ C_row = C_row_start + row;
+ C_col = C_col_start + col;
+ if ((C_row < M) && (kblock + col < K))
+ As[row][col] = A[(C_row*LDA)+ kblock + col];
+ else
+ As[row][col] = 0;
+ if ((kblock + row < K) && C_col < N)
+ Bs[row][col] = B[((kblock+row)*LDB)+ C_col];
+ else
+ Bs[row][col] = 0;
+ }
+
+#pragma omp for collapse(2)
+ for (int row=0 ; row < BLOCK_SIZE ; row++)
+ for (int col=0 ; col < BLOCK_SIZE ; col++)
+ {
+ for (int e = 0; e < BLOCK_SIZE; ++e)
+ Cval += As[row][e] * Bs[e][col];
+ }
+ } /* End for kblock .. */
+
+
+#pragma omp for collapse(2)
+ for (int row=0 ; row < BLOCK_SIZE ; row++)
+ for (int col=0 ; col < BLOCK_SIZE ; col++)
+ {
+ C_row = C_row_start + row;
+ C_col = C_col_start + col;
+ if ((C_row < M) && (C_col < N))
+ C[(C_row*LDC)+C_col] = alpha*Cval + beta*C[(C_row*LDC)+C_col];
+
+ }
+ } /* end parallel */
+ } /* end target teams distribute */
+}
new file mode 100644
@@ -0,0 +1,258 @@
+/*
+
+ matmul.c : Matrix Multiplication with tiling for openmp4 example
+
+*/
+
+#include <stdlib.h>
+#include <math.h>
+
+#define BLOCK_SIZE 16
+/*
+ #define BLOCK_SIZE 32
+*/
+#define NSECPERSEC 1000000000L
+
+typedef struct {
+ int width;
+ int height;
+ int stride;
+ int hpad;
+ float* elements;
+} Matrix;
+
+/* Correctly extract the number of nanoseconds from the two time structures */
+long int get_nanosecs( struct timespec start_time, struct timespec end_time) {
+ long int nanosecs;
+ if ((end_time.tv_nsec-start_time.tv_nsec)<0) nanosecs =
+ ((((long int) end_time.tv_sec- (long int) start_time.tv_sec )-1)*NSECPERSEC ) +
+ ( NSECPERSEC + (long int) end_time.tv_nsec - (long int) start_time.tv_nsec) ;
+ else nanosecs =
+ (((long int) end_time.tv_sec- (long int) start_time.tv_sec )*NSECPERSEC ) +
+ ( (long int) end_time.tv_nsec - (long int) start_time.tv_nsec );
+ return nanosecs;
+}
+
+void simple_sgemm_tt(const int M,const int N,const int K,const float alpha, const float* A,const int LDA,
+ const float* B,const int LDB, const float beta,float* C, const int LDC) ;
+void simple_sgemm_tn(const int M,const int N,const int K,const float alpha, const float* A,const int LDA,
+ const float* B,const int LDB, const float beta,float* C, const int LDC) ;
+void tiled_sgemm_tt(const int M,const int N,const int K,const float alpha, const float*A, const int LDA,
+ const float* B,const int LDB, const float beta,float* C, const int LDC) ;
+
+int verify(float* v_res, float* v_ref, int len) {
+ int passed = 1;
+ int i;
+ for (i = 0; i < len; ++i) {
+ if (fabs(v_res[i] - v_ref[i]) > 0.001*v_ref[i]) {
+ __builtin_abort ();
+ }
+ }
+ return passed;
+}
+
+
+int main(int argc, char* argv[]){
+
+ Matrix A,B,Bt,C,Cref;
+ int a1,a2,a3,i,j;
+ struct timespec start_time1, end_time1;
+ struct timespec start_time2, end_time2;
+ long int nanosecs,total_ops;
+ float gflopsTiled,gflopsCPU;
+
+ a1 = 35;
+ a2 = 28;
+ a3 = 47;
+
+ A.height = a1;
+ A.width = a2;
+ A.stride = (((A.width-1)/BLOCK_SIZE)+1) * BLOCK_SIZE;
+ A.hpad = (((A.height-1)/BLOCK_SIZE)+1) * BLOCK_SIZE;
+ A.elements = (float*)malloc(A.stride * A.hpad* sizeof(float));
+
+ B.height = a2;
+ B.width = a3;
+ B.stride = (((B.width-1)/BLOCK_SIZE)+1) * BLOCK_SIZE;
+ B.hpad = (((B.height-1)/BLOCK_SIZE)+1) * BLOCK_SIZE;
+ B.elements = (float*)malloc(B.stride * B.hpad * sizeof(float));
+
+ /* Bt is same as B but stored in column-major order */
+ Bt.height = B.height;
+ Bt.width = B.width;
+ Bt.stride = B.stride;
+ Bt.hpad = B.hpad;
+ Bt.elements = (float*)malloc(Bt.stride * Bt.hpad * sizeof(float));
+
+ C.height = a1;
+ C.width = a3;
+ C.stride = (((C.width-1)/BLOCK_SIZE)+1) * BLOCK_SIZE;
+ C.hpad = (((C.height-1)/BLOCK_SIZE)+1) * BLOCK_SIZE;
+ C.elements = (float*)malloc(C.stride * C.hpad * sizeof(float));
+
+ Cref.height = a1;
+ Cref.width = a3;
+ Cref.stride = (((Cref.width-1)/BLOCK_SIZE)+1) * BLOCK_SIZE;
+ Cref.hpad = (((Cref.height-1)/BLOCK_SIZE)+1) * BLOCK_SIZE;
+ Cref.elements = (float*)malloc(Cref.stride * Cref.hpad * sizeof(float));
+
+ for(i = 0; i < A.hpad ; i++)
+ for(j = 0; j < A.stride; j++) {
+ if (( j<A.width ) && (i<A.height)) {
+ A.elements[i*A.stride + j] = (i % 3);
+ } else {
+ A.elements[i*A.stride + j] = 0.0;
+ }
+ }
+
+ /* Initialize B and Bt */
+ for(i = 0; i < B.hpad ; i++)
+ for(j = 0; j < B.stride; j++) {
+ if (( j<B.width ) && (i<B.height)) {
+ B.elements[i*B.stride+j] = (j % 2);
+ Bt.elements[j*Bt.stride+i] = B.elements[i*B.stride+j] ;
+ } else {
+ B.elements[i*B.stride+j] = 0.0;
+ Bt.elements[j*Bt.stride+i] = 0.0;
+ }
+ }
+
+ /* zero C, and Cref */
+ for(i = 0; i < C.hpad; i++)
+ for(j = 0; j < C.stride; j++) {
+ C.elements[i*C.stride+j] = 0.0;
+ Cref.elements[i*Cref.stride+j] = 0.0;
+ }
+
+ simple_sgemm_tt(A.height,B.width,B.height,1.0,A.elements,A.stride,B.elements,B.stride,1.0,Cref.elements,Cref.stride);
+ tiled_sgemm_tt(A.height,B.width,B.height,1.0,A.elements,A.stride,B.elements,B.stride,1.0,C.elements,C.stride);
+
+ verify(C.elements, Cref.elements, C.height * C.stride);
+ return 0;
+}
+
+void simple_sgemm_tt(const int M,const int N,const int K,const float alpha, const float* A,const int LDA,
+const float* B,const int LDB, const float beta,float* C, const int LDC) {
+ /* A,B, and C are in row-major order */
+ int c_row,c_col,inner;
+ float sum;
+ for (c_col = 0 ; c_col<N; c_col++ ) {
+ for (c_row = 0 ; c_row<M; c_row++ ) {
+ sum = 0.0 ;
+ for (inner = 0 ; inner<K; inner++ ) {
+ sum += A[c_row*LDA + inner] * B[inner*LDB + c_col] ;
+ }
+ C[c_row*LDC + c_col] = alpha*sum + beta*C[ c_row*LDC + c_col] ;
+ }
+ }
+}
+
+/***************************
+
+ tiled_sgemm_tt: Tiled matrix multiplication:
+
+***************************/
+
+void tiled_sgemm_tt(const int M, const int N, const int K, const float alpha, const float*A, const int LDA,
+ const float*B, const int LDB, const float beta, float*C, const int LDC){
+
+#pragma omp target teams map(to:A[M*K],B[K*N]) map(from:C[M*N])
+#pragma omp distribute collapse(2)
+ for (int C_row_start=0 ; C_row_start < M ; C_row_start+=BLOCK_SIZE) {
+ for (int C_col_start=0 ; C_col_start < N ; C_col_start+=BLOCK_SIZE) {
+
+// We now have M/BLOCK_SIZE * N/BLOCK_SIZE teams = (M*N)/(BLOCK_SIZE*BLOCK_SIZE)
+// The grid global dimensions are M,N,1
+// The grid local dimensions are BLOCK_SIZE,BLOCK_SIZE,1
+
+// -------------------------------------------------------------------
+// The rest of this code forms the HSAIL kernel with the
+// pairs of "paralell for collapse(2)" loops repalced with a barrier.
+// The kernel initializes these values
+// C_row_start = get_group_id(0) * BLOCK_SIZE
+// C_col_start = get_group_id(1) * BLOCK_SIZE
+// row=get_local_id(0)
+// col=get_local_id(1)
+// -------------------------------------------------------------------
+
+// Each team has a local copy of these mini matrices
+ float As[BLOCK_SIZE][BLOCK_SIZE];
+ float Bs[BLOCK_SIZE][BLOCK_SIZE];
+ float Cs[BLOCK_SIZE][BLOCK_SIZE];
+ int C_row, C_col;
+
+ /* Zero Cs for this BLOCK */
+// - - - - - - - - - - - - - - - - - - - -
+// REPLACE NEXT THREE LINES WITH A BARRIER
+#pragma omp parallel for collapse(2)
+ for (int row=0 ; row < BLOCK_SIZE ; row++) {
+ for (int col=0 ; col < BLOCK_SIZE ; col++) {
+// END BARRIER
+// - - - - - - - - - - - - - - - - - - - -
+ Cs[row][col] = 0.0;
+ }
+ }
+
+ // This kblock loop is run on the master thread of each team
+ for (int kblock = 0; kblock < K ; kblock += BLOCK_SIZE ) {
+
+ // Copy global memory values to local memory
+// - - - - - - - - - - - - - - - - - - - -
+// REPLACE NEXT THREE LINES WITH A BARRIER
+#pragma omp parallel for collapse(2)
+ for (int row=0 ; row < BLOCK_SIZE ; row++) {
+ for (int col=0 ; col < BLOCK_SIZE ; col++) {
+// END BARRIER
+// - - - - - - - - - - - - - - - - - - - -
+ C_row = C_row_start + row;
+ C_col = C_col_start + col;
+ if ((C_row < M) && (kblock + col < K))
+ As[row][col] = A[(C_row*LDA)+ kblock + col];
+ else
+ As[row][col] = 0;
+ if ((kblock + row < K) && C_col < N)
+ Bs[row][col] = B[((kblock+row)*LDB)+ C_col];
+ else
+ Bs[row][col] = 0;
+ }
+ }
+
+ // Calculate Cs <- Sum(As X Bs) across all kblocks
+// - - - - - - - - - - - - - - - - - - - -
+// REPLACE NEXT THREE LINES WITH A BARRIER
+#pragma omp parallel for collapse(2)
+ for (int row=0 ; row < BLOCK_SIZE ; row++) {
+ for (int col=0 ; col < BLOCK_SIZE ; col++) {
+// END BARRIER
+// - - - - - - - - - - - - - - - - - - - -
+ for (int e = 0; e < BLOCK_SIZE; ++e)
+ Cs[row][col] += As[row][e] * Bs[e][col];
+ }
+ }
+
+ } /* End for kblock .. */
+
+
+ // Scale Update actual C from Cs
+// - - - - - - - - - - - - - - - - - - - -
+// REPLACE NEXT THREE LINES WITH A BARRIER
+#pragma omp parallel for collapse(2)
+ for (int row=0 ; row < BLOCK_SIZE ; row++) {
+ for (int col=0 ; col < BLOCK_SIZE ; col++) {
+// END BARRIER
+// - - - - - - - - - - - - - - - - - - - -
+ C_row = C_row_start + row;
+ C_col = C_col_start + col;
+ if ((C_row < M) && (C_col < N)) {
+ C[(C_row*LDC)+C_col] = alpha*Cs[row][col] + beta*C[(C_row*LDC)+C_col];
+ }
+ }
+ }
+
+// -------------------------------------------------------------------
+// This is the end of the kernel
+
+ }
+ }
+
+}
Hi, the patch below is the main part of the series that enhances the existing gridification code to pattern-match sequences of constructs in which the distribute, parallel and a loop constructs are not in one simple combined construct but the step in the distribute loop exactly matches the iteration size of (possibly many) inner loop constructs. It also checks other conditions, most notably that an unknown non-const and non-pure function is not called as part of the distribute loop but outside of a normal loop construct. If this pattern is matched, iterations of the distribute loop are converted to HSA groups and iterations of inner loop constructs to individual work-items within those groups. In the code, the inner loops are converted to a barrier. I'll commit this to the hsa branch in a few moments and will re-submit it for trunk at some point in summer. Thanks, Martin 2016-06-09 Martin Jambor <mjambor@suse.cz> gcc/ * gimple.h (enum gf_mask): New element GF_OMP_FOR_GRID_GROUP_ITER. (gimple_omp_for_grid_group_iter): New function. (gimple_omp_for_set_grid_group_iter): Likewise. * omp-low.c (check_omp_nesting_restrictions): Allow kernel loop in place of a distribute one. (grid_expand_omp_for_loop): New parameter specifying whether the loop is an intra-group one. If so or if the loop is over groups, use the respective builtins for expansion. Emit barriers for intra-group ones. Moved a branch removeal here from grid_expand_target_grid_body. (grid_expand_target_grid_body): If the loop iterates over groups, find the intra-group ones, expand them and remove them from the OMP construct tree. (grid_prop): New type. (grid_safe_assignment_p): New parameter grid, do not consider safe assignments to variables holding group sizes. (grid_seq_only_contains_local_assignments): New parameter grid, pass it to grid_safe_assignment_p. (grid_find_single_omp_among_assignments_1): Likewise. Also tkae target location from grid. Emit more missed-optimizations information. (grid_find_single_omp_among_assignments): Likewise. (grid_parallel_clauses_gridifiable): New function. (grid_inner_loop_gridifiable_p): Likewise. (grid_dist_follows_simple_pattern): Likewise. (grid_gfor_follows_tiling_pattern): Likewise. (grid_call_permissible_in_distribute_p): Likewise. (grid_handle_call_in_distribute): Likewise. (grid_dist_follows_tiling_pattern): Likewise. (grid_target_follows_gridifiable_pattern): New parameter grid. If distribute is not in a combined construct, attempt tiled gridification. Parts of simple gridification moved to the new functions. (grid_var_segment): New enum. (grid_mark_variable_segment): New function. (grid_copy_leading_local_assignments): Also call grid_mark_variable_segment if requested by a new parameter. (grid_mark_tiling_loops): New function. (grid_mark_tiling_parallels_and_loops): Likewise. (grid_process_kernel_body_copy): Also handle tiled grids. (grid_attempt_target_gridification): new variable grid to be passed around. Get group sizes from it. gcc/testsuite/ * c-c++-common/gomp/gridify-2.c: New test. * c-c++-common/gomp/gridify-3.c: Likewise. libgomp/ * testsuite/libgomp.hsa.c/tiling-1.c: New test. * testsuite/libgomp.hsa.c/tiling-2.c: Likewise. --- gcc/gimple.h | 21 + gcc/omp-low.c | 1057 ++++++++++++++++++++------- gcc/testsuite/c-c++-common/gomp/gridify-2.c | 66 ++ gcc/testsuite/c-c++-common/gomp/gridify-3.c | 68 ++ libgomp/testsuite/libgomp.hsa.c/tiling-1.c | 212 ++++++ libgomp/testsuite/libgomp.hsa.c/tiling-2.c | 258 +++++++ 6 files changed, 1434 insertions(+), 248 deletions(-) create mode 100644 gcc/testsuite/c-c++-common/gomp/gridify-2.c create mode 100644 gcc/testsuite/c-c++-common/gomp/gridify-3.c create mode 100644 libgomp/testsuite/libgomp.hsa.c/tiling-1.c create mode 100644 libgomp/testsuite/libgomp.hsa.c/tiling-2.c