* [gomp4] OpenACC loop expand reorg
@ 2015-10-12 13:11 Nathan Sidwell
0 siblings, 0 replies; only message in thread
From: Nathan Sidwell @ 2015-10-12 13:11 UTC (permalink / raw)
To: GCC Patches; +Cc: Jakub Jelinek, Tom de Vries
[-- Attachment #1: Type: text/plain, Size: 1613 bytes --]
I've committed this to gomp4 branch. It reworks the loop expansion code for
OpenACC loops in the following ways
1) Removes OpenACC handling from expand_omp_for_static_{,no}chunk}. These are
thus now OpenMP only. (Jakub, that should reduce conflicts between the two
implemenations)
2) Implements expand_oacc_for, for openacc loops. Loops are expanded using a
new internal fn 'IFN_GOACC_LOOP', which abstracts the chunk size, step size,
initial iteration per-thread iteration value, and per-thread bound.
3) IFN_GOACC_LOOP is lowered in the oacc_device_lower pass. There are 4
variants of the call, and an initial INTEGER_CST arg is used to distinguish use.
(This seemed better than having 4 new internal fns, and matches the IFN_UNIQUE
mechanism).
4) The GOACC_LOOP lowering deals with chunking, and choses whether a compute
axis should assign adjacent iterations to adjacent compute elements (striding),
or assign a compute element to a contiguous span of iterations (contiguous).
You want to stride at the vector level, but be continuous at outer levels, to
maximize cache friendliness (in general). Chunking of size 1 is the same as
striding, and we reduce the former to the latter.
right now, GOACC_LOOP expansion is not device specific. The expansion done
earlier passes in the necessary information. I'm working on a patch to separate
that now, which will complete the transition. With that
(a) kernels will be easier to mark and optimize
(b) the implementation will be device_type friendly, as device-specific choices
will all have been moved to the target compiler.
nathan
[-- Attachment #2: gomp4-loop-reorg.patch --]
[-- Type: text/x-patch, Size: 32356 bytes --]
2015-10-12 Nathan Sidwell <nathan@codesourcery.com>
* omp-low.c (expand_omp_for_static_nochunk): Remove OpenACC
pieces.
(expand_omp_for_static_chunk): Likewise,
(struct oacc_collapse): New.
(expand_oacc_collapse_init, expand_oacc_collapse_vars): New.
(expand_oacc_for): New.
(expand_omp_for): Call expand_oacc_for for OpenACC loops.
(oacc_xform_loop): New.
(execute_oacc_device_lower): Call it.
* internal-fn.def (GOACC_LOOP): New internal fn.
(IFN_GOACC_LOOP_CHUNKS, IFN_GOACC_LOOP_STEP,
IFN_GOACC_LOOP_OFFSET, IFN_GOACC_LOOP_BOUND): New.
* internal-fn.c (expand_GOACC_LOOP): New.
Index: gcc/omp-low.c
===================================================================
--- gcc/omp-low.c (revision 228696)
+++ gcc/omp-low.c (working copy)
@@ -7002,11 +7002,6 @@ expand_omp_for_generic (struct omp_regio
V += STEP;
if (V cond e) goto L1;
L2:
-
- For OpenACC the above is wrapped in an OACC_FORK/OACC_JOIN pair.
- Currently we wrap the whole sequence, but it'd be better to place the
- markers just inside the outer conditional, so they can be entirely
- eliminated if the loop is unreachable.
*/
static void
@@ -7025,9 +7020,8 @@ expand_omp_for_static_nochunk (struct om
tree *counts = NULL;
tree n1, n2, step;
- gcc_checking_assert ((gimple_omp_for_kind (fd->for_stmt)
- != GF_OMP_FOR_KIND_OACC_LOOP)
- || !inner_stmt);
+ gcc_checking_assert (gimple_omp_for_kind (fd->for_stmt)
+ != GF_OMP_FOR_KIND_OACC_LOOP);
itype = type = TREE_TYPE (fd->loop.v);
if (POINTER_TYPE_P (type))
@@ -7126,14 +7120,6 @@ expand_omp_for_static_nochunk (struct om
threadid = builtin_decl_explicit (BUILT_IN_OMP_GET_TEAM_NUM);
threadid = build_call_expr (threadid, 0);
break;
- case GF_OMP_FOR_KIND_OACC_LOOP:
- {
- gimple_seq seq = NULL;
- nthreads = expand_oacc_get_num_threads (&seq, region->gwv_this);
- threadid = expand_oacc_get_thread_num (&seq, region->gwv_this);
- gsi_insert_seq_before (&gsi, seq, GSI_SAME_STMT);
- }
- break;
default:
gcc_unreachable ();
}
@@ -7312,8 +7298,7 @@ expand_omp_for_static_nochunk (struct om
/* Replace the GIMPLE_OMP_RETURN with a barrier, or nothing. */
gsi = gsi_last_bb (exit_bb);
- if (gimple_omp_for_kind (fd->for_stmt) != GF_OMP_FOR_KIND_OACC_LOOP
- && !gimple_omp_return_nowait_p (gsi_stmt (gsi)))
+ if (!gimple_omp_return_nowait_p (gsi_stmt (gsi)))
{
t = gimple_omp_return_lhs (gsi_stmt (gsi));
gsi_insert_after (&gsi, build_omp_barrier (t), GSI_SAME_STMT);
@@ -7437,11 +7422,6 @@ find_phi_with_arg_on_edge (tree arg, edg
trip += 1;
goto L0;
L4:
-
- For OpenACC the above is wrapped in an OACC_FORK/OACC_JOIN pair.
- Currently we wrap the whole sequence, but it'd be better to place the
- markers just inside the outer conditional, so they can be entirely
- eliminated if the loop is unreachable.
*/
static void
@@ -7459,9 +7439,8 @@ expand_omp_for_static_chunk (struct omp_
tree *counts = NULL;
tree n1, n2, step;
- gcc_checking_assert ((gimple_omp_for_kind (fd->for_stmt)
- != GF_OMP_FOR_KIND_OACC_LOOP)
- || !inner_stmt);
+ gcc_checking_assert (gimple_omp_for_kind (fd->for_stmt)
+ != GF_OMP_FOR_KIND_OACC_LOOP);
itype = type = TREE_TYPE (fd->loop.v);
if (POINTER_TYPE_P (type))
@@ -7565,14 +7544,6 @@ expand_omp_for_static_chunk (struct omp_
threadid = builtin_decl_explicit (BUILT_IN_OMP_GET_TEAM_NUM);
threadid = build_call_expr (threadid, 0);
break;
- case GF_OMP_FOR_KIND_OACC_LOOP:
- {
- gimple_seq seq = NULL;
- nthreads = expand_oacc_get_num_threads (&seq, region->gwv_this);
- threadid = expand_oacc_get_thread_num (&seq, region->gwv_this);
- gsi_insert_seq_before (&gsi, seq, GSI_SAME_STMT);
- }
- break;
default:
gcc_unreachable ();
}
@@ -7774,8 +7745,7 @@ expand_omp_for_static_chunk (struct omp_
/* Replace the GIMPLE_OMP_RETURN with a barrier, or nothing. */
gsi = gsi_last_bb (exit_bb);
- if (!gimple_omp_return_nowait_p (gsi_stmt (gsi))
- && gimple_omp_for_kind (fd->for_stmt) != GF_OMP_FOR_KIND_OACC_LOOP)
+ if (!gimple_omp_return_nowait_p (gsi_stmt (gsi)))
{
t = gimple_omp_return_lhs (gsi_stmt (gsi));
gsi_insert_after (&gsi, build_omp_barrier (t), GSI_SAME_STMT);
@@ -8429,6 +8399,555 @@ expand_omp_simd (struct omp_region *regi
cfun->has_simduid_loops = true;
}
+/* Information about members of an OpenACC collapsed loop nest. */
+
+struct oacc_collapse
+{
+ tree base; /* Base value. */
+ tree iters; /* Number of steps. */
+ tree step; /* step size. */
+};
+
+/* Helper for expand_oacc_for. Determine collapsed loop information.
+ Fill in COUNTS array. Emit any initialization code before GSI.
+ Return the calculated outer loop bound of BOUND_TYPE. */
+
+static tree
+expand_oacc_collapse_init (const struct omp_for_data *fd,
+ gimple_stmt_iterator *gsi,
+ oacc_collapse *counts, tree bound_type)
+{
+ tree total = build_int_cst (bound_type, 1);
+ int ix;
+
+ gcc_assert (integer_onep (fd->loop.step));
+ gcc_assert (integer_zerop (fd->loop.n1));
+
+ for (ix = 0; ix != fd->collapse; ix++)
+ {
+ const omp_for_data_loop *loop = &fd->loops[ix];
+
+ tree iter_type = TREE_TYPE (loop->v);
+ tree diff_type = iter_type;
+ tree plus_type = iter_type;
+
+ gcc_assert (loop->cond_code == fd->loop.cond_code);
+
+ if (POINTER_TYPE_P (iter_type))
+ plus_type = sizetype;
+ if (POINTER_TYPE_P (diff_type) || TYPE_UNSIGNED (diff_type))
+ diff_type = signed_type_for (diff_type);
+
+ tree b = loop->n1;
+ tree e = loop->n2;
+ tree s = loop->step;
+ bool up = loop->cond_code == LT_EXPR;
+ tree dir = build_int_cst (diff_type, up ? +1 : -1);
+ bool negating;
+ tree expr;
+
+ b = force_gimple_operand_gsi (gsi, b, true, NULL_TREE,
+ true, GSI_SAME_STMT);
+ e = force_gimple_operand_gsi (gsi, e, true, NULL_TREE,
+ true, GSI_SAME_STMT);
+
+ /* Convert the step, avoiding possible unsigned->signed overflow. */
+ negating = !up && TYPE_UNSIGNED (TREE_TYPE (s));
+ if (negating)
+ s = fold_build1 (NEGATE_EXPR, TREE_TYPE (s), s);
+ s = fold_convert (diff_type, s);
+ if (negating)
+ s = fold_build1 (NEGATE_EXPR, diff_type, s);
+ s = force_gimple_operand_gsi (gsi, s, true, NULL_TREE,
+ true, GSI_SAME_STMT);
+
+ /* Determine the range, avoiding possible unsigned->signed overflow. */
+ negating = !up && TYPE_UNSIGNED (iter_type);
+ expr = fold_build2 (MINUS_EXPR, plus_type,
+ fold_convert (plus_type, negating ? b : e),
+ fold_convert (plus_type, negating ? e : b));
+ expr = fold_convert (diff_type, expr);
+ if (negating)
+ expr = fold_build1 (NEGATE_EXPR, diff_type, expr);
+ tree range = force_gimple_operand_gsi
+ (gsi, expr, true, NULL_TREE, true, GSI_SAME_STMT);
+
+ /* Determine number of iterations. */
+ expr = fold_build2 (MINUS_EXPR, diff_type, range, dir);
+ expr = fold_build2 (PLUS_EXPR, diff_type, expr, s);
+ expr = fold_build2 (TRUNC_DIV_EXPR, diff_type, expr, s);
+
+ tree iters = force_gimple_operand_gsi (gsi, expr, true, NULL_TREE,
+ true, GSI_SAME_STMT);
+
+ counts[ix].base = b;
+ counts[ix].iters = iters;
+ counts[ix].step = s;
+
+ total = fold_build2 (MULT_EXPR, bound_type, total,
+ fold_convert (bound_type, iters));
+ }
+
+ return total;
+}
+
+/* Emit initializers for collapsed loop members. IVAR is the outer
+ loop iteration variable, from which collapsed loop iteration values
+ are calculated. COUNTS array has been initialized by
+ expand_oacc_collapse_inits. */
+
+static void
+expand_oacc_collapse_vars (const struct omp_for_data *fd,
+ gimple_stmt_iterator *gsi,
+ const oacc_collapse *counts, tree ivar)
+{
+ tree ivar_type = TREE_TYPE (ivar);
+
+ /* The most rapidly changing iteration variable is the innermost
+ one. */
+ for (int ix = fd->collapse; ix--;)
+ {
+ const omp_for_data_loop *loop = &fd->loops[ix];
+ const oacc_collapse *collapse = &counts[ix];
+ tree iter_type = TREE_TYPE (loop->v);
+ tree diff_type = TREE_TYPE (collapse->step);
+ tree plus_type = iter_type;
+ enum tree_code plus_code = PLUS_EXPR;
+ tree expr;
+
+ if (POINTER_TYPE_P (iter_type))
+ {
+ plus_code = POINTER_PLUS_EXPR;
+ plus_type = sizetype;
+ }
+
+ expr = build2 (TRUNC_MOD_EXPR, ivar_type, ivar,
+ fold_convert (ivar_type, collapse->iters));
+ expr = build2 (MULT_EXPR, diff_type, fold_convert (diff_type, expr),
+ collapse->step);
+ expr = build2 (plus_code, iter_type, collapse->base,
+ fold_convert (plus_type, expr));
+ expr = force_gimple_operand_gsi (gsi, expr, false, NULL_TREE,
+ true, GSI_SAME_STMT);
+ gassign *ass = gimple_build_assign (loop->v, expr);
+ gsi_insert_before (gsi, ass, GSI_SAME_STMT);
+
+ if (ix)
+ {
+ expr = build2 (TRUNC_DIV_EXPR, ivar_type, ivar,
+ fold_convert (ivar_type, collapse->iters));
+ ivar = force_gimple_operand_gsi (gsi, expr, true, NULL_TREE,
+ true, GSI_SAME_STMT);
+ }
+ }
+}
+
+/* A subroutine of expand_omp_for. Generate code for an OpenACC
+ partitioned loop. The lowering here is abstracted, in that the
+ loop parameters are passed through internal functions, which are
+ further lowered by oacc_device_lower, once we get to the target
+ compiler. The loop is of the form:
+
+ for (V = B; V LTGT E; V += S) {BODY}
+
+ where LTGT is < or >. We may have a specified chunking size, CHUNKING
+ (constant 0 for no chunking) and we will have a GWV partitioning
+ mask, specifying dimensions over which the loop is to be
+ partitioned (see note below). We generate code that looks like:
+
+ <entry_bb> [incoming FALL->body, BRANCH->exit]
+ typedef signedintify (typeof (V)) T; // underlying signed integral type
+ T range = E - B;
+ T chunk_no = 0;
+ T DIR = LTGT == '<' ? +1 : -1;
+ T chunk_max = GOACC_LOOP_CHUNK (dir, range, S, CHUNK_SIZE, GWV);
+ T step = GOACC_LOOP_STEP (dir, range, S, CHUNK_SIZE, GWV);
+
+ <head_bb> [created by splitting end of entry_bb]
+ T offset = GOACC_LOOP_OFFSET (dir, range, S, CHUNK_SIZE, GWV, chunk_no);
+ T bound = GOACC_LOOP_BOUND (dir, range, S, CHUNK_SIZE, GWV, offset);
+ if (!(offset LTGT bound)) goto bottom_bb;
+
+ <body_bb> [incoming]
+ V = B + offset;
+ {BODY}
+
+ <cont_bb> [incoming, may == body_bb FALL->exit_bb, BRANCH->body_bb]
+ offset += step;
+ if (offset LTGT bound) goto body_bb; [*]
+
+ <bottom_bb> [created by splitting start of exit_bb] insert BRANCH->head_bb
+ chunk_no++;
+ if (chunk < chunk_max) goto head_bb;
+
+ <exit_bb> [incoming]
+ V = B + ((range -/+ 1) / S +/- 1) * S [*]
+
+ [*] Needed if V live at end of loop
+
+ Note: CHUNKING & GWV mask are specified explicitly here. This is a
+ transition, and will be specified by a more general mechanism shortly.
+ */
+
+static void
+expand_oacc_for (struct omp_region *region, struct omp_for_data *fd)
+{
+ tree v = fd->loop.v;
+ enum tree_code cond_code = fd->loop.cond_code;
+ enum tree_code plus_code = PLUS_EXPR;
+
+ tree chunk_size = fd->chunk_size;
+ tree gwv = build_int_cst (integer_type_node, region->gwv_this);
+ tree iter_type = TREE_TYPE (v);
+ tree diff_type = iter_type;
+ tree plus_type = iter_type;
+ struct oacc_collapse *counts = NULL;
+
+ gcc_checking_assert (gimple_omp_for_kind (fd->for_stmt)
+ == GF_OMP_FOR_KIND_OACC_LOOP);
+ gcc_assert (!gimple_omp_for_combined_into_p (fd->for_stmt));
+ gcc_assert (cond_code == LT_EXPR || cond_code == GT_EXPR);
+
+ if (POINTER_TYPE_P (iter_type))
+ {
+ plus_code = POINTER_PLUS_EXPR;
+ plus_type = sizetype;
+ }
+ if (POINTER_TYPE_P (diff_type) || TYPE_UNSIGNED (diff_type))
+ diff_type = signed_type_for (diff_type);
+
+ basic_block entry_bb = region->entry; /* BB ending in OMP_FOR */
+ basic_block exit_bb = region->exit; /* BB ending in OMP_RETURN */
+ basic_block cont_bb = region->cont; /* BB ending in OMP_CONTINUE */
+ basic_block bottom_bb = NULL;
+
+ /* entry_bb has two sucessors; the branch edge is to the exit
+ block, fallthrough edge to body. */
+ gcc_assert (EDGE_COUNT (entry_bb->succs) == 2
+ && BRANCH_EDGE (entry_bb)->dest == exit_bb);
+
+ /* If cont_bb non-NULL, it has 2 successors. The branch successor is
+ body_bb, or to a block whose only successor is the body_bb. Its
+ fallthrough successor is the final block (same as the branch
+ successor of the entry_bb). */
+ if (cont_bb)
+ {
+ basic_block body_bb = FALLTHRU_EDGE (entry_bb)->dest;
+ basic_block bed = BRANCH_EDGE (cont_bb)->dest;
+
+ gcc_assert (FALLTHRU_EDGE (cont_bb)->dest == exit_bb);
+ gcc_assert (bed == body_bb || single_succ_edge (bed)->dest == body_bb);
+ }
+ else
+ gcc_assert (!gimple_in_ssa_p (cfun));
+
+ /* The exit block only has entry_bb and cont_bb as predecessors. */
+ gcc_assert (EDGE_COUNT (exit_bb->preds) == 1 + (cont_bb != NULL));
+
+ tree chunk_no;
+ tree chunk_max = NULL_TREE;
+ tree bound, offset;
+ tree step = create_tmp_var (diff_type, ".step");
+ bool up = cond_code == LT_EXPR;
+ tree dir = build_int_cst (diff_type, up ? +1 : -1);
+ bool chunking = chunk_size != NULL_TREE;
+ bool negating;
+
+ /* SSA instances. */
+ tree offset_incr = NULL_TREE;
+ tree offset_init = NULL_TREE;
+
+ gimple_stmt_iterator gsi;
+ gassign *ass;
+ gcall *call;
+ gimple *stmt;
+ tree expr;
+ location_t loc;
+ edge split, be, fte;
+
+ /* Split the end of entry_bb to create head_bb. */
+ split = split_block (entry_bb, last_stmt (entry_bb));
+ basic_block head_bb = split->dest;
+ entry_bb = split->src;
+
+ /* Chunk setup goes at end of entry_bb, replacing the omp_for. */
+ gsi = gsi_last_bb (entry_bb);
+ gomp_for *for_stmt = as_a <gomp_for *> (gsi_stmt (gsi));
+ loc = gimple_location (for_stmt);
+
+ if (gimple_in_ssa_p (cfun))
+ {
+ offset_init = gimple_omp_for_index (for_stmt, 0);
+ gcc_assert (integer_zerop (fd->loop.n1));
+ }
+
+ if (fd->collapse > 1)
+ {
+ counts = XALLOCAVEC (struct oacc_collapse, fd->collapse);
+ tree total = expand_oacc_collapse_init (fd, &gsi, counts,
+ TREE_TYPE (fd->loop.n2));
+
+ if (SSA_VAR_P (fd->loop.n2))
+ {
+ total = force_gimple_operand_gsi (&gsi, total, false, NULL_TREE,
+ true, GSI_SAME_STMT);
+ ass = gimple_build_assign (fd->loop.n2, total);
+ gsi_insert_before (&gsi, ass, GSI_SAME_STMT);
+ }
+
+ }
+
+ tree b = fd->loop.n1;
+ tree e = fd->loop.n2;
+ tree s = fd->loop.step;
+
+ b = force_gimple_operand_gsi (&gsi, b, true, NULL_TREE, true, GSI_SAME_STMT);
+ e = force_gimple_operand_gsi (&gsi, e, true, NULL_TREE, true, GSI_SAME_STMT);
+
+ /* Convert the step, avoiding possible unsigned->signed overflow. */
+ negating = !up && TYPE_UNSIGNED (TREE_TYPE (s));
+ if (negating)
+ s = fold_build1 (NEGATE_EXPR, TREE_TYPE (s), s);
+ s = fold_convert (diff_type, s);
+ if (negating)
+ s = fold_build1 (NEGATE_EXPR, diff_type, s);
+ s = force_gimple_operand_gsi (&gsi, s, true, NULL_TREE, true, GSI_SAME_STMT);
+
+ if (!chunking)
+ chunk_size = integer_zero_node;
+ expr = fold_convert (diff_type, chunk_size);
+ chunk_size = force_gimple_operand_gsi (&gsi, expr, true,
+ NULL_TREE, true, GSI_SAME_STMT);
+
+ /* Determine the range, avoiding possible unsigned->signed overflow. */
+ negating = !up && TYPE_UNSIGNED (iter_type);
+ expr = fold_build2 (MINUS_EXPR, plus_type,
+ fold_convert (plus_type, negating ? b : e),
+ fold_convert (plus_type, negating ? e : b));
+ expr = fold_convert (diff_type, expr);
+ if (negating)
+ expr = fold_build1 (NEGATE_EXPR, diff_type, expr);
+ tree range = force_gimple_operand_gsi (&gsi, expr, true,
+ NULL_TREE, true, GSI_SAME_STMT);
+
+ chunk_no = build_int_cst (diff_type, 0);
+ if (chunking)
+ {
+ gcc_assert (!gimple_in_ssa_p (cfun));
+
+ expr = chunk_no;
+ chunk_max = create_tmp_var (diff_type, ".chunk_max");
+ chunk_no = create_tmp_var (diff_type, ".chunk_no");
+
+ ass = gimple_build_assign (chunk_no, expr);
+ gsi_insert_before (&gsi, ass, GSI_SAME_STMT);
+
+ expr = fold_convert (diff_type, chunk_size);
+ chunk_size = force_gimple_operand_gsi (&gsi, expr, true,
+ NULL_TREE, true, GSI_SAME_STMT);
+
+ call = gimple_build_call_internal (IFN_GOACC_LOOP, 6,
+ build_int_cst (integer_type_node,
+ IFN_GOACC_LOOP_CHUNKS),
+ dir, range, s, chunk_size, gwv);
+ gimple_call_set_lhs (call, chunk_max);
+ gimple_set_location (call, loc);
+ gsi_insert_before (&gsi, call, GSI_SAME_STMT);
+ }
+ else
+ chunk_size = chunk_no;
+
+ call = gimple_build_call_internal (IFN_GOACC_LOOP, 6,
+ build_int_cst (integer_type_node,
+ IFN_GOACC_LOOP_STEP),
+ dir, range, s, chunk_size, gwv);
+ gimple_call_set_lhs (call, step);
+ gimple_set_location (call, loc);
+ gsi_insert_before (&gsi, call, GSI_SAME_STMT);
+
+ /* Remove the GIMPLE_OMP_FOR. */
+ gsi_remove (&gsi, true);
+
+ /* Fixup edges from head_bb */
+ be = BRANCH_EDGE (head_bb);
+ fte = FALLTHRU_EDGE (head_bb);
+ be->flags |= EDGE_FALSE_VALUE;
+ fte->flags ^= EDGE_FALLTHRU | EDGE_TRUE_VALUE;
+
+ basic_block body_bb = fte->dest;
+
+ if (gimple_in_ssa_p (cfun))
+ {
+ gsi = gsi_last_bb (cont_bb);
+ gomp_continue *cont_stmt = as_a <gomp_continue *> (gsi_stmt (gsi));
+
+ offset = gimple_omp_continue_control_use (cont_stmt);
+ offset_incr = gimple_omp_continue_control_def (cont_stmt);
+ }
+ else
+ {
+ offset = create_tmp_var (diff_type, ".offset");
+ offset_init = offset_incr = offset;
+ }
+ bound = create_tmp_var (TREE_TYPE (offset), ".bound");
+
+ /* Loop offset & bound go into head_bb. */
+ gsi = gsi_start_bb (head_bb);
+
+ call = gimple_build_call_internal (IFN_GOACC_LOOP, 7,
+ build_int_cst (integer_type_node,
+ IFN_GOACC_LOOP_OFFSET),
+ dir, range, s,
+ chunk_size, gwv, chunk_no);
+ gimple_call_set_lhs (call, offset_init);
+ gimple_set_location (call, loc);
+ gsi_insert_after (&gsi, call, GSI_CONTINUE_LINKING);
+
+ call = gimple_build_call_internal (IFN_GOACC_LOOP, 7,
+ build_int_cst (integer_type_node,
+ IFN_GOACC_LOOP_BOUND),
+ dir, range, s,
+ chunk_size, gwv, offset_init);
+ gimple_call_set_lhs (call, bound);
+ gimple_set_location (call, loc);
+ gsi_insert_after (&gsi, call, GSI_CONTINUE_LINKING);
+
+ expr = build2 (cond_code, boolean_type_node, offset_init, bound);
+ gsi_insert_after (&gsi, gimple_build_cond_empty (expr),
+ GSI_CONTINUE_LINKING);
+
+ /* V assignment goes into body_bb. */
+ if (!gimple_in_ssa_p (cfun))
+ {
+ gsi = gsi_start_bb (body_bb);
+
+ expr = build2 (plus_code, iter_type, b,
+ fold_convert (plus_type, offset));
+ expr = force_gimple_operand_gsi (&gsi, expr, false, NULL_TREE,
+ true, GSI_SAME_STMT);
+ ass = gimple_build_assign (v, expr);
+ gsi_insert_before (&gsi, ass, GSI_SAME_STMT);
+ if (fd->collapse > 1)
+ expand_oacc_collapse_vars (fd, &gsi, counts, v);
+ }
+
+ /* Loop increment goes into cont_bb. If this is not a loop, we
+ will have spawned threads as if it was, and each one will
+ execute one iteration. The specification is not explicit about
+ whether such constructs are ill-formed or not, and they can
+ occur, especially when noreturn routines are involved. */
+ if (cont_bb)
+ {
+ gsi = gsi_last_bb (cont_bb);
+ gomp_continue *cont_stmt = as_a <gomp_continue *> (gsi_stmt (gsi));
+ loc = gimple_location (cont_stmt);
+
+ /* Increment offset. */
+ if (gimple_in_ssa_p (cfun))
+ expr= build2 (plus_code, iter_type, offset,
+ fold_convert (plus_type, step));
+ else
+ expr = build2 (PLUS_EXPR, diff_type, offset, step);
+ expr = force_gimple_operand_gsi (&gsi, expr, false, NULL_TREE,
+ true, GSI_SAME_STMT);
+ ass = gimple_build_assign (offset_incr, expr);
+ gsi_insert_before (&gsi, ass, GSI_SAME_STMT);
+ expr = build2 (cond_code, boolean_type_node, offset_incr, bound);
+ gsi_insert_before (&gsi, gimple_build_cond_empty (expr), GSI_SAME_STMT);
+
+ /* Remove the GIMPLE_OMP_CONTINUE. */
+ gsi_remove (&gsi, true);
+
+ /* Fixup edges from cont_bb */
+ be = BRANCH_EDGE (cont_bb);
+ fte = FALLTHRU_EDGE (cont_bb);
+ be->flags |= EDGE_TRUE_VALUE;
+ fte->flags ^= EDGE_FALLTHRU | EDGE_FALSE_VALUE;
+
+ if (chunking)
+ {
+ /* Split the beginning of exit_bb to make bottom_bb. We
+ need to insert a nop at the start, because splitting is
+ after a stmt, not before. */
+ gsi = gsi_start_bb (exit_bb);
+ stmt = gimple_build_nop ();
+ gsi_insert_before (&gsi, stmt, GSI_SAME_STMT);
+ split = split_block (exit_bb, stmt);
+ bottom_bb = split->src;
+ exit_bb = split->dest;
+ gsi = gsi_last_bb (bottom_bb);
+
+ /* Chunk increment and test goes into bottom_bb. */
+ expr = build2 (PLUS_EXPR, diff_type, chunk_no,
+ build_int_cst (diff_type, 1));
+ ass = gimple_build_assign (chunk_no, expr);
+ gsi_insert_after (&gsi, ass, GSI_CONTINUE_LINKING);
+
+ /* Chunk test at end of bottom_bb. */
+ expr = build2 (LT_EXPR, boolean_type_node, chunk_no, chunk_max);
+ gsi_insert_after (&gsi, gimple_build_cond_empty (expr),
+ GSI_CONTINUE_LINKING);
+
+ /* Fixup edges from bottom_bb. */
+ split->flags ^= EDGE_FALLTHRU | EDGE_FALSE_VALUE;
+ make_edge (bottom_bb, head_bb, EDGE_TRUE_VALUE);
+ }
+ }
+
+ gsi = gsi_last_bb (exit_bb);
+ gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_RETURN);
+ loc = gimple_location (gsi_stmt (gsi));
+
+ if (!gimple_in_ssa_p (cfun))
+ {
+ /* Insert the final value of V, in case it is live. This is the
+ value for the only thread that survives past the join. */
+ expr = fold_build2 (MINUS_EXPR, diff_type, range, dir);
+ expr = fold_build2 (PLUS_EXPR, diff_type, expr, s);
+ expr = fold_build2 (TRUNC_DIV_EXPR, diff_type, expr, s);
+ expr = fold_build2 (MULT_EXPR, diff_type, expr, s);
+ expr = build2 (plus_code, iter_type, b, fold_convert (plus_type, expr));
+ expr = force_gimple_operand_gsi (&gsi, expr, false, NULL_TREE,
+ true, GSI_SAME_STMT);
+ ass = gimple_build_assign (v, expr);
+ gsi_insert_before (&gsi, ass, GSI_SAME_STMT);
+ }
+
+ /* Remove the OMP_RETURN. */
+ gsi_remove (&gsi, true);
+
+ if (cont_bb)
+ {
+ /* We now have one or two nested loops. Update the loop
+ structures. */
+ struct loop *parent = entry_bb->loop_father;
+ struct loop *body = body_bb->loop_father;
+
+ if (chunking)
+ {
+ struct loop *chunk_loop = alloc_loop ();
+ chunk_loop->header = head_bb;
+ chunk_loop->latch = bottom_bb;
+ add_loop (chunk_loop, parent);
+ parent = chunk_loop;
+ }
+ else if (parent != body)
+ {
+ gcc_assert (body->header == body_bb);
+ gcc_assert (body->latch == cont_bb
+ || single_pred (body->latch) == cont_bb);
+ parent = NULL;
+ }
+
+ if (parent)
+ {
+ struct loop *body_loop = alloc_loop ();
+ body_loop->header = body_bb;
+ body_loop->latch = cont_bb;
+ add_loop (body_loop, parent);
+ }
+ }
+}
/* Expand the OMP loop defined by REGION. */
@@ -8476,6 +8995,11 @@ expand_omp_for (struct omp_region *regio
expand_omp_simd (region, &fd);
else if (gimple_omp_for_kind (fd.for_stmt) == GF_OMP_FOR_KIND_CILKFOR)
expand_cilk_for (region, &fd);
+ else if (gimple_omp_for_kind (fd.for_stmt) == GF_OMP_FOR_KIND_OACC_LOOP)
+ {
+ gcc_assert (!inner_stmt);
+ expand_oacc_for (region, &fd);
+ }
else if (fd.sched_kind == OMP_CLAUSE_SCHEDULE_STATIC
&& !fd.have_ordered)
{
@@ -14731,6 +15255,183 @@ make_pass_late_lower_omp (gcc::context *
return new pass_late_lower_omp (ctxt);
}
+/* Transform IFN_GOACC_LOOP calls to actual code. See
+ expand_oacc_for for where these are generated. At the vector
+ level, we stride loops, such that each member of a warp will
+ operate on adjacent iterations. At the worker and gang level,
+ each gang/warp executes a set of contiguous iterations. Chunking
+ can override this such that each iteration engine executes a
+ contiguous chunk, and then moves on to stride to the next chunk.
+
+ TODO: As with expand_oacc_for, the presence of GWV and CHUNK_SIZE
+ parameters here is an intermediate step. */
+
+static void
+oacc_xform_loop (gcall *call)
+{
+ gimple_stmt_iterator gsi = gsi_for_stmt (call);
+ unsigned code = (unsigned)TREE_INT_CST_LOW (gimple_call_arg (call, 0));
+ tree dir = gimple_call_arg (call, 1);
+ tree range = gimple_call_arg (call, 2);
+ tree step = gimple_call_arg (call, 3);
+ tree chunk_size = gimple_call_arg (call, 4);
+ unsigned mask = (unsigned)TREE_INT_CST_LOW (gimple_call_arg (call, 5));
+ tree lhs = gimple_call_lhs (call);
+ tree type = TREE_TYPE (lhs);
+ tree diff_type = TREE_TYPE (range);
+ tree r = NULL_TREE;
+ gimple_seq seq = NULL;
+ bool chunking, striding;
+ unsigned outer_mask = mask & (~mask + 1); // Outermost partitioning
+ unsigned inner_mask = mask & ~outer_mask; // Inner partitioning (if any)
+
+ if (integer_zerop (chunk_size))
+ {
+ /* If we're at the gang or worker level, we want each to execute
+ a contiguous run of iterations. Otherwise we want each
+ element to stride. */
+ striding = !(outer_mask & (GOMP_DIM_MASK (GOMP_DIM_WORKER)
+ | GOMP_DIM_MASK (GOMP_DIM_GANG)));
+ chunking = false;
+ }
+ else
+ {
+ /* Chunk of size 1 is striding. */
+ striding = integer_onep (chunk_size);
+ chunking = !striding;
+ }
+
+ push_gimplify_context (true);
+
+ switch (code)
+ {
+ default: gcc_unreachable ();
+
+ case IFN_GOACC_LOOP_CHUNKS:
+ if (!chunking)
+ r = build_int_cst (type, 1);
+ else
+ {
+ /* chunk_max
+ = (range - dir) / (chunks * step * num_threads) + dir */
+ tree per = expand_oacc_get_num_threads (&seq, mask);
+ per = fold_convert (type, per);
+ per = fold_build2 (MULT_EXPR, type, per, chunk_size);
+ per = fold_build2 (MULT_EXPR, type, per, step);
+ r = build2 (MINUS_EXPR, type, range, dir);
+ r = build2 (PLUS_EXPR, type, r, per);
+ r = build2 (TRUNC_DIV_EXPR, type, r, per);
+ }
+ break;
+
+ case IFN_GOACC_LOOP_STEP:
+ {
+ /* If striding, step by the entire compute volume, otherwise
+ step by the inner volume. */
+ unsigned volume = striding ? mask : inner_mask;
+
+ r = expand_oacc_get_num_threads (&seq, volume);
+ r = build2 (MULT_EXPR, type, fold_convert (type, r), step);
+ }
+ break;
+
+ case IFN_GOACC_LOOP_OFFSET:
+ if (striding)
+ {
+ r = expand_oacc_get_thread_num (&seq, mask);
+ r = fold_convert (diff_type, r);
+ }
+ else
+ {
+ tree span;
+ tree inner_size = expand_oacc_get_num_threads (&seq, inner_mask);
+ tree outer_size = expand_oacc_get_num_threads (&seq, outer_mask);
+ tree volume = fold_build2 (MULT_EXPR, TREE_TYPE (inner_size),
+ inner_size, outer_size);
+
+ if (chunking)
+ {
+ span = inner_size;
+ span = fold_convert (type, span);
+ span = fold_build2 (MULT_EXPR, diff_type, span, chunk_size);
+ }
+ else
+ {
+ tree per = fold_convert (diff_type, volume);
+ per = fold_build2 (MULT_EXPR, diff_type, per, step);
+
+ span = build2 (MINUS_EXPR, diff_type, range, dir);
+ span = build2 (PLUS_EXPR, diff_type, span, per);
+ span = build2 (TRUNC_DIV_EXPR, diff_type, span, per);
+ span = build2 (MULT_EXPR, diff_type, span, inner_size);
+ }
+
+ r = expand_oacc_get_thread_num (&seq, outer_mask);
+ r = fold_convert (diff_type, r);
+ r = build2 (MULT_EXPR, diff_type, r, span);
+
+ tree inner = expand_oacc_get_thread_num (&seq, inner_mask);
+ inner = fold_convert (diff_type, inner);
+ r = fold_build2 (PLUS_EXPR, diff_type, r, inner);
+
+ if (chunking)
+ {
+ tree chunk = gimple_call_arg (call, 6);
+ tree per = fold_convert (diff_type, volume);
+ per = fold_build2 (MULT_EXPR, diff_type, per, chunk_size);
+ per = build2 (MULT_EXPR, diff_type, per, chunk);
+
+ r = build2 (PLUS_EXPR, diff_type, r, per);
+ }
+ }
+ r = fold_build2 (MULT_EXPR, diff_type, r, step);
+ if (type != diff_type)
+ r = fold_convert (type, r);
+ break;
+
+ case IFN_GOACC_LOOP_BOUND:
+ if (striding)
+ r = range;
+ else
+ {
+ tree offset = gimple_call_arg (call, 6);
+ tree span;
+
+ if (chunking)
+ {
+ span = expand_oacc_get_num_threads (&seq, inner_mask);
+ span = fold_convert (diff_type, span);
+ span = fold_build2 (MULT_EXPR, diff_type, span, chunk_size);
+ }
+ else
+ {
+ tree per = expand_oacc_get_num_threads (&seq, mask);
+ per = fold_convert (diff_type, per);
+ per = build2 (MULT_EXPR, diff_type, per, step);
+ span = build2 (MINUS_EXPR, diff_type, range, dir);
+ span = build2 (PLUS_EXPR, diff_type, span, per);
+ span = build2 (TRUNC_DIV_EXPR, diff_type, span, per);
+ }
+
+ r = fold_build2 (MULT_EXPR, diff_type, span, step);
+
+ r = build2 (PLUS_EXPR, diff_type, r,
+ fold_convert (diff_type, offset));
+ r = build2 (integer_onep (dir) ? MIN_EXPR : MAX_EXPR,
+ diff_type, r, range);
+ }
+ if (diff_type != type)
+ r = fold_convert (type, r);
+ break;
+ }
+
+ gimplify_assign (lhs, r, &seq);
+
+ pop_gimplify_context (NULL);
+
+ gsi_replace_with_seq (&gsi, seq, true);
+}
+
/* Transform oacc_dim_size and oacc_dim_pos internal function calls to
constants, where possible. */
@@ -14874,6 +15575,11 @@ execute_oacc_device_lower ()
rescan = 1;
break;
+ case IFN_GOACC_LOOP:
+ oacc_xform_loop (call);
+ rescan = 1;
+ break;
+
case IFN_GOACC_REDUCTION_SETUP:
case IFN_GOACC_REDUCTION_INIT:
case IFN_GOACC_REDUCTION_FINI:
Index: gcc/internal-fn.c
===================================================================
--- gcc/internal-fn.c (revision 228696)
+++ gcc/internal-fn.c (working copy)
@@ -2052,6 +2052,13 @@ expand_GOACC_REDUCTION_TEARDOWN (gcall *
gcc_unreachable ();
}
+/* This is expanded by oacc_device_lower pass. */
+
+static void
+expand_GOACC_LOOP (gcall *stmt ATTRIBUTE_UNUSED)
+{
+ gcc_unreachable ();
+}
/* Routines to expand each internal function, indexed by function number.
Each routine has the prototype:
Index: gcc/internal-fn.def
===================================================================
--- gcc/internal-fn.def (revision 228696)
+++ gcc/internal-fn.def (working copy)
@@ -106,3 +106,28 @@ DEF_INTERNAL_FN (GOACC_REDUCTION_INIT, E
DEF_INTERNAL_FN (GOACC_REDUCTION_FINI, ECF_NOTHROW, NULL)
DEF_INTERNAL_FN (GOACC_REDUCTION_TEARDOWN, ECF_NOTHROW, NULL)
+/* OpenACC looping abstraction. Allows the precise stepping of
+ the compute geometry over the loop iterations to be deferred until
+ it is known which compiler is generating the code. The action is
+ encoded in a constant first argument.
+
+ CHUNK_MAX = LOOP (CODE_CHUNKS, DIR, RANGE, STEP, CHUNK_SIZE, MASK)
+ STEP = LOOP (CODE_STEP, DIR, RANGE, STEP, CHUNK_SIZE, MASK)
+ OFFSET = LOOP (CODE_OFFSET, DIR, RANGE, STEP, CHUNK_SIZE, MASK, CHUNK_NO)
+ BOUND = LOOP (CODE_BOUND, DIR, RANGE, STEP, CHUNK_SIZE, MASK, OFFSET)
+
+ DIR - +1 for up loop, -1 for down loop
+ RANGE - Range of loop (END - BASE)
+ STEP - iteration step size
+ CHUNKING - size of chunking, (constant zero for no chunking)
+ CHUNK_NO - chunk number
+ MASK - partitioning mask.
+
+ TODO: The partitioning mask and chunk size are a transition stage,
+ they will be removed once the required infrastructure is in place. */
+
+DEF_INTERNAL_FN (GOACC_LOOP, ECF_PURE | ECF_NOTHROW, NULL)
+#define IFN_GOACC_LOOP_CHUNKS 0 /* Number of chunks. */
+#define IFN_GOACC_LOOP_STEP 1 /* Size of each thread's step. */
+#define IFN_GOACC_LOOP_OFFSET 2 /* Initial iteration value. */
+#define IFN_GOACC_LOOP_BOUND 3 /* Limit of iteration value. */
^ permalink raw reply [flat|nested] only message in thread
only message in thread, other threads:[~2015-10-12 13:11 UTC | newest]
Thread overview: (only message) (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2015-10-12 13:11 [gomp4] OpenACC loop expand reorg Nathan Sidwell
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).