* [gomp] openacc reduction cleanup
@ 2015-06-30 18:07 Cesar Philippidis
2015-07-01 13:45 ` Cesar Philippidis
0 siblings, 1 reply; 2+ messages in thread
From: Cesar Philippidis @ 2015-06-30 18:07 UTC (permalink / raw)
To: gcc-patches
This patch removes a lot of the crufty code which was necessary for an
arbitrary sized vector reductions. The new plan going forward it to fix
vector_length to a size such that vector loops don't require any
synchronization after the loop have terminated. In the case of nvptx
targets, vector_length = warp_sz, which is currently 32 threads. I'll
follow up this patch with another one which adds support for tree
reductions in vector loops in a later date.
This patch has been applied to gomp-4_0-branch.
Cesar
^ permalink raw reply [flat|nested] 2+ messages in thread
* Re: [gomp] openacc reduction cleanup
2015-06-30 18:07 [gomp] openacc reduction cleanup Cesar Philippidis
@ 2015-07-01 13:45 ` Cesar Philippidis
0 siblings, 0 replies; 2+ messages in thread
From: Cesar Philippidis @ 2015-07-01 13:45 UTC (permalink / raw)
To: gcc-patches
[-- Attachment #1: Type: text/plain, Size: 174 bytes --]
On 06/30/2015 11:01 AM, Cesar Philippidis wrote:
> This patch has been applied to gomp-4_0-branch.
It was brought to my attention that I forgot to attach the patch.
Cesar
[-- Attachment #2: reduction-cleanup.diff --]
[-- Type: text/x-patch, Size: 34409 bytes --]
2015-06-30 Cesar Philippidis <cesar@codesourcery.com>
gcc/
* omp-low.c (oacc_get_reduction_array_id): Remove.
(oacc_parallel_max_reduction_array_size): Remove.
(install_array_var_ganglocal): Remove.
(oacc_outermost_parallel_kernels_context): Remove.
(oacc_inside_routine): Remove.
(is_oacc_multithreaded): Remove.
(oacc_needs_global_memory): Remove.
(oacc_max_threads): Remove.
(scan_sharing_clauses): Don't allocate an array for openacc reductions.
(oacc_lower_reduction_var_helper): Remove.
(lower_reduction_clauses): Handle all OpenACC reductions using atomics.
(oacc_gimple_assign): Remove.
(oacc_initialize_reduction_data): Remove.
(oacc_serial_reduction): Remove.
(oacc_finalize_reduction_data): Remove.
(oacc_process_reduction_data_helper): Remove.
(oacc_process_reduction_data): Remove.
(lower_omp_target): Don't call oacc_process_reduction_data anymore.
diff --git a/gcc/omp-low.c b/gcc/omp-low.c
index 24fac7c..11ac909 100644
--- a/gcc/omp-low.c
+++ b/gcc/omp-low.c
@@ -305,18 +305,6 @@ static basic_block oacc_broadcast (basic_block, basic_block,
*handled_ops_p = false; \
break;
-/* Helper function to get the name of the array containing the partial
- reductions for OpenACC reductions. */
-static const char *
-oacc_get_reduction_array_id (tree node)
-{
- const char *id = IDENTIFIER_POINTER (DECL_NAME (node));
- int len = strlen ("OACC") + strlen (id);
- char *temp_name = XALLOCAVEC (char, len + 1);
- snprintf (temp_name, len + 1, "OACC%s", id);
- return IDENTIFIER_POINTER (get_identifier (temp_name));
-}
-
static bool
is_oacc_parallel (omp_context *ctx)
{
@@ -326,68 +314,6 @@ is_oacc_parallel (omp_context *ctx)
== GF_OMP_TARGET_KIND_OACC_PARALLEL));
}
-/* Determine the maximum number of threads available to the current omp
- context for a parallel OpenACC reduction. */
-
-static tree
-oacc_parallel_max_reduction_array_size (omp_context *ctx, tree clauses,
- gimple_seq *seqp)
-{
- tree nthreads;
- bool gangs, workers, vectors;
- omp_context *oc;
-
- gangs = find_omp_clause (clauses, OMP_CLAUSE_GANG) != NULL_TREE;
- workers = find_omp_clause (clauses, OMP_CLAUSE_WORKER) != NULL_TREE;
- vectors = find_omp_clause (clauses, OMP_CLAUSE_VECTOR) != NULL_TREE;
-
- /* num_gangs, num_workers and vector_length is set in innermost parallel
- gimple stmt. Find that stmt and extract those values. */
-
- for (oc = ctx; !is_oacc_parallel (oc); oc = oc->outer)
- ;
-
- clauses = gimple_omp_target_clauses (oc->stmt);
-
- /* FIXME: this assignment may be useless. */
- nthreads = create_tmp_var (sizetype, NULL);
- gimplify_assign (nthreads, build_int_cst (sizetype, 1), seqp);
-
- if (vectors)
- {
- tree v = find_omp_clause (clauses, OMP_CLAUSE_VECTOR_LENGTH);
- if (v)
- {
- //tree t = fold_convert (sizetype, OMP_CLAUSE_VECTOR_LENGTH_EXPR (v));
- gimplify_assign (nthreads, build_int_cst (sizetype, 1), seqp);
- }
- }
-
- if (workers)
- {
- tree w = find_omp_clause (clauses, OMP_CLAUSE_NUM_WORKERS);
- if (w)
- {
- tree t = fold_convert (sizetype, OMP_CLAUSE_NUM_WORKERS_EXPR (w));
- gimple stmt = gimple_build_assign (nthreads, MULT_EXPR, nthreads, t);
- gimple_seq_add_stmt (seqp, stmt);
- }
- }
-
- if (gangs)
- {
- tree g = find_omp_clause (clauses, OMP_CLAUSE_NUM_GANGS);
- if (g)
- {
- tree t = fold_convert (sizetype, OMP_CLAUSE_NUM_GANGS_EXPR (g));
- gimple stmt = gimple_build_assign(nthreads, MULT_EXPR, nthreads, t);
- gimple_seq_add_stmt (seqp, stmt);
- }
- }
-
- return nthreads;
-}
-
/* Holds offload tables with decls. */
vec<tree, va_gc> *offload_funcs, *offload_vars;
@@ -1583,14 +1509,6 @@ install_var_ganglocal (tree decl, omp_context *ctx)
return new_var;
}
-static tree
-install_array_var_ganglocal (tree decl, tree type, tree size, omp_context *ctx)
-{
- tree ptr = alloc_var_ganglocal (decl, type, ctx, size);
- insert_decl_map (&ctx->cb, decl, ptr);
- return ptr;
-}
-
/* Debugging dumps for parallel regions. */
void dump_omp_region (FILE *, struct omp_region *, int);
void debug_omp_region (struct omp_region *);
@@ -1870,140 +1788,6 @@ fixup_child_record_type (omp_context *ctx)
= build_qualified_type (build_reference_type (type), TYPE_QUAL_RESTRICT);
}
-static omp_context *
-oacc_outermost_parallel_kernels_context (omp_context *ctx)
-{
- omp_context *octx = ctx;
-
- while (octx)
- {
- if (gimple_code (octx->stmt) == GIMPLE_OMP_TARGET)
- {
- int kind = gimple_omp_target_kind (octx->stmt);
- if (kind == GF_OMP_TARGET_KIND_OACC_PARALLEL
- || kind == GF_OMP_TARGET_KIND_OACC_KERNELS)
- return octx;
- }
-
- octx = octx->outer;
- }
-
- return ctx;
-}
-
-/* OpenACC routines contain acc loops without any kernels or parallel
- regions. If the omp_context isn't a descendant of a kernels or
- parallel region, then it must be inside a routine. */
-
-static bool
-oacc_inside_routine (omp_context *ctx)
-{
- omp_context *octx = ctx;
-
- while (octx)
- {
- if (gimple_code (octx->stmt) == GIMPLE_OMP_TARGET)
- {
- int kind = gimple_omp_target_kind (octx->stmt);
- if (kind == GF_OMP_TARGET_KIND_OACC_PARALLEL
- || kind == GF_OMP_TARGET_KIND_OACC_KERNELS)
- return false;
- }
-
- octx = octx->outer;
- }
-
- return true;
-}
-
-/* Returns true if an oacc context is to be executed in parallel. OpenACC
- kernels get translated into OpenACC parallel regions, so don't consider
- them as candidates for multithreaded processing. */
-
-static bool
-is_oacc_multithreaded (tree clauses, omp_context *ctx)
-{
- omp_context *octx = oacc_outermost_parallel_kernels_context (ctx);
-
- /* This is true if thies stmt is inside a routine or a kernels region. */
- if (!is_oacc_parallel (octx))
- return false;
-
- if (find_omp_clause (clauses, OMP_CLAUSE_GANG))
- return true;
-
- if (find_omp_clause (clauses, OMP_CLAUSE_WORKER))
- return true;
-
- if (find_omp_clause (clauses, OMP_CLAUSE_VECTOR))
- return true;
-
- return false;
-}
-
-/* Helper function to determine if a set of clauses uses global memory.
- All reductions involving gangs or are in a parallel construction are
- need atomic. */
-
-static bool
-oacc_needs_global_memory (tree clauses, omp_context *ctx)
-{
- if (is_oacc_parallel (ctx))
- return true;
-
- if (oacc_inside_routine (ctx))
- return true;
-
- return find_omp_clause (clauses, OMP_CLAUSE_GANG) != NULL_TREE;
-}
-
-/* Determine the maximum number of threads used by an openacc loop.
- TODO: Add support for kernels regions. */
-
-static tree
-oacc_max_threads (omp_context *ctx)
-{
- tree nthreads, vector_length, num_gangs, num_workers, clauses, oclauses, c;
- omp_context *octx;
-
- octx = oacc_outermost_parallel_kernels_context (ctx);
- clauses = gimple_omp_for_clauses (ctx->stmt);
- oclauses = gimple_omp_target_clauses (octx->stmt);
-
- num_gangs = build_int_cst (size_type_node, 1);
- num_workers = build_int_cst (size_type_node, 1);
- vector_length = build_int_cst (size_type_node, 1);
- nthreads = build_int_cst (size_type_node, 1);
-
- /* Find num_gangs, num_workers and vector_length. FIXME: Vector reduction
- also need to scan outer loops for worker threads so that memeory is
- allocated for all of the vector lanes in each worker thread. */
- for (c = oclauses; c ; c = OMP_CLAUSE_CHAIN (c))
- {
- tree size = OMP_CLAUSE_OPERAND (c, 0);
-
- if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_NUM_GANGS)
- num_gangs = fold_build1 (NOP_EXPR, size_type_node, size);
- else if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_NUM_WORKERS)
- num_workers = fold_build1 (NOP_EXPR, size_type_node, size);
- else if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_VECTOR_LENGTH)
- vector_length = fold_build1 (NOP_EXPR, size_type_node, size);
- }
-
- /* Calculate nthreads. */
- for (c = clauses; c ; c = OMP_CLAUSE_CHAIN (c))
- {
- if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_GANG)
- nthreads = fold_build2 (MULT_EXPR, sizetype, nthreads, num_gangs);
- else if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_WORKER)
- nthreads = fold_build2 (MULT_EXPR, sizetype, nthreads, num_workers);
- else if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_VECTOR)
- nthreads = fold_build2 (MULT_EXPR, sizetype, nthreads, vector_length);
- }
-
- return nthreads;
-}
-
/* Instantiate decls as necessary in CTX to satisfy the data sharing
specified by CLAUSES. */
@@ -2100,9 +1884,6 @@ scan_sharing_clauses (tree clauses, omp_context *ctx)
case OMP_CLAUSE_LINEAR:
decl = OMP_CLAUSE_DECL (c);
- if (is_gimple_omp_oacc (ctx->stmt))
- scan_array_reductions = true;
-
do_private:
if (is_variable_sized (decl))
{
@@ -2427,43 +2208,7 @@ scan_sharing_clauses (tree clauses, omp_context *ctx)
if (scan_array_reductions)
{
for (c = clauses; c; c = OMP_CLAUSE_CHAIN (c))
- if (is_gimple_omp_oacc (ctx->stmt)
- && OMP_CLAUSE_CODE (c) == OMP_CLAUSE_REDUCTION
- && is_oacc_multithreaded (clauses, ctx)
- && !oacc_needs_global_memory (clauses, ctx))
- {
- /* Create a decl for the reduction array. */
- tree var = OMP_CLAUSE_DECL (c);
- tree type = get_base_type (var);
- tree array;
-
- array = create_tmp_var (type,
- oacc_get_reduction_array_id (var));
-
- omp_context *octx = ctx;
- tree size = fold_build2 (MULT_EXPR, sizetype, TYPE_SIZE_UNIT (type),
- oacc_max_threads (ctx));
-
- /* This array must be inserted into the field map, which is only
- present in the outermost omp context. */
- octx = oacc_outermost_parallel_kernels_context (ctx);
-
- if (octx != NULL && octx->field_map != NULL)
- {
- array = install_array_var_ganglocal (array, type, size, octx);
-
- /* Insert it into the current context. */
- splay_tree_insert (ctx->reduction_map, (splay_tree_key)
- oacc_get_reduction_array_id (var),
- (splay_tree_value) array);
- splay_tree_insert (ctx->reduction_map,
- (splay_tree_key) array,
- (splay_tree_value) array);
- }
- else
- sorry ("Reductions are not supported inside routines yet.");
- }
- else if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_REDUCTION
+ if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_REDUCTION
&& OMP_CLAUSE_REDUCTION_PLACEHOLDER (c))
{
scan_omp (&OMP_CLAUSE_REDUCTION_GIMPLE_INIT (c), ctx);
@@ -4911,58 +4656,6 @@ lower_lastprivate_clauses (tree clauses, tree predicate, gimple_seq *stmt_list,
gimple_seq_add_stmt (stmt_list, gimple_build_label (label));
}
-static void
-oacc_lower_reduction_var_helper (gimple_seq *stmt_seqp, omp_context *ctx,
- tree tid, tree var, tree new_var)
-{
- /* The atomic operation at the end of the reduction creates
- unnecessary write contention on accelerators. To work around
- this, create an array to store the partial reductions. Later, in
- lower_omp_for (for openacc), the values of array will be
- combined. */
-
- tree t = NULL_TREE, array, x;
- tree type = get_base_type (var);
- gimple stmt;
-
- /* Now insert the partial reductions into the array. */
-
- /* Find the reduction array. */
-
- tree ptype = build_pointer_type (type);
-
- t = lookup_oacc_reduction (oacc_get_reduction_array_id (var), ctx);
- t = build_receiver_ref (t, false, ctx->outer);
-
- array = create_tmp_var (ptype);
- gimplify_assign (array, t, stmt_seqp);
-
- tree ptr = create_tmp_var (TREE_TYPE (array));
-
- /* Find the reduction array. */
-
- /* testing a unary conversion. */
- tree offset = create_tmp_var (sizetype);
- gimplify_assign (offset, TYPE_SIZE_UNIT (type),
- stmt_seqp);
- t = create_tmp_var (sizetype);
- gimplify_assign (t, unshare_expr (fold_build1 (NOP_EXPR, sizetype, tid)),
- stmt_seqp);
- stmt = gimple_build_assign (offset, MULT_EXPR, offset, t);
- gimple_seq_add_stmt (stmt_seqp, stmt);
-
- /* Offset expression. Does the POINTER_PLUS_EXPR take care
- of adding sizeof(var) to the array? */
- ptr = create_tmp_var (ptype);
- stmt = gimple_build_assign (unshare_expr (ptr), POINTER_PLUS_EXPR, array,
- offset);
- gimple_seq_add_stmt (stmt_seqp, stmt);
-
- /* Move the local sum to gfc$sum[i]. */
- x = unshare_expr (build_simple_mem_ref (ptr));
- stmt = gimplify_assign (x, new_var, stmt_seqp);
-}
-
/* Returns true if this reduction operation is compatible with atomics. */
static bool
@@ -5057,13 +4750,6 @@ expand_oacc_get_thread_num (gimple_seq *seq, int gwv_bits)
return res;
}
-
-static void
-oacc_finalize_reduction_data (tree clauses, tree nthreads, tree tid,
- gimple_seq *stmt_seqp, omp_context *ctx,
- bool is_accelerator);
-
-
/* Generate code to implement the REDUCTION clauses. OpenACC reductions
are usually executed in parallel, but they fallback to sequential code for
known single-threaded regions. */
@@ -5073,7 +4759,7 @@ lower_reduction_clauses (tree clauses, gimple_seq *stmt_seqp, omp_context *ctx)
{
gimple_seq sub_seq = NULL;
gimple stmt;
- tree x, c, tid = NULL_TREE;
+ tree x, c;
int count = 0;
/* SIMD reductions are handled in lower_rec_input_clauses. */
@@ -5098,16 +4784,6 @@ lower_reduction_clauses (tree clauses, gimple_seq *stmt_seqp, omp_context *ctx)
if (count == 0)
return;
- /* Initialize thread info for OpenACC. */
- if (is_gimple_omp_oacc (ctx->stmt)
- && !(is_oacc_multithreaded (clauses, ctx)
- && oacc_needs_global_memory (clauses, ctx))
- && !is_oacc_parallel (ctx))
- {
- /* Get the current thread id. */
- tid = expand_oacc_get_thread_num (stmt_seqp, ctx->gwv_this);
- }
-
for (c = clauses; c ; c = OMP_CLAUSE_CHAIN (c))
{
tree var, ref, new_var;
@@ -5135,17 +4811,12 @@ lower_reduction_clauses (tree clauses, gimple_seq *stmt_seqp, omp_context *ctx)
if (count == 1 && is_atomic_compatible_reduction (var, ctx))
{
- if (!is_gimple_omp_oacc (ctx->stmt)
- || (is_gimple_omp_oacc (ctx->stmt)
- && (!is_oacc_multithreaded (clauses, ctx)
- || oacc_needs_global_memory (clauses, ctx))))
- {
tree addr = build_fold_addr_expr_loc (clause_loc, ref);
addr = save_expr (addr);
if (is_gimple_omp_oacc (ctx->stmt)
- && !oacc_needs_global_memory (clauses, ctx))
+ && (ctx->gwv_this == 0))
{
/* This reduction is done sequentially in OpenACC by a single
thread. There is no need to use atomics. */
@@ -5163,78 +4834,6 @@ lower_reduction_clauses (tree clauses, gimple_seq *stmt_seqp, omp_context *ctx)
}
return;
- }
- else
- {
- /* The atomic add at the end of the sum creates unnecessary
- write contention on accelerators. To work around this,
- create an array to store the partial reductions. Later, in
- lower_omp_for (for openacc), the values of array will be
- combined. */
-
- tree t, array, nthreads, ptr;
- tree type = get_base_type (var);
- omp_context *octx;
- gimple stmt;
-
- /* Find the total number of threads. A reduction clause
- only appears inside a loop construction or a combined
- parallel and loop construct. */
- nthreads = expand_oacc_get_num_threads (stmt_seqp, ctx->gwv_this);
-
- /* Now insert the partial reductions into the array. */
-
- /* Find the reduction array. */
-
- tree ptype = build_pointer_type (type);
-
- t = lookup_oacc_reduction (oacc_get_reduction_array_id (var), ctx);
-
- /* Find the outermost ctx where the array variable has been
- declared. */
-
- octx = oacc_outermost_parallel_kernels_context (ctx);
-
- if (oacc_needs_global_memory (clauses, ctx))
- {
- t = build_receiver_ref (t, false, octx);
- array = create_tmp_var (ptype, NULL);
- gimplify_assign (array, t, stmt_seqp);
- }
- else
- array = t;
-
- /* Find the reduction array. */
-
- /* testing a unary conversion. */
- tree offset = create_tmp_var (sizetype, NULL);
- gimplify_assign (offset, TYPE_SIZE_UNIT (type),
- stmt_seqp);
- t = create_tmp_var (sizetype, NULL);
- gimplify_assign (t, unshare_expr (fold_build1 (NOP_EXPR, sizetype,
- tid)),
- stmt_seqp);
- stmt = gimple_build_assign (offset, MULT_EXPR, offset, t);
- gimple_seq_add_stmt (stmt_seqp, stmt);
-
- /* Offset expression. Does the POINTER_PLUS_EXPR take care
- of adding sizeof(var) to the array? */
- ptr = create_tmp_var (ptype, NULL);
- stmt = gimple_build_assign (unshare_expr(ptr), POINTER_PLUS_EXPR,
- array, offset);
- gimple_seq_add_stmt (stmt_seqp, stmt);
-
- /* Move the local sum to OACCsum[i]. */
- x = unshare_expr (build_simple_mem_ref (ptr));
- stmt = gimplify_assign (x, new_var, stmt_seqp);
-
- /* Synchronize the threads and finish up the reduction. */
-
- oacc_finalize_reduction_data (clauses, nthreads, tid, stmt_seqp, ctx,
- true);
-
- return;
- }
}
else if (OMP_CLAUSE_REDUCTION_PLACEHOLDER (c))
{
@@ -5253,14 +4852,7 @@ lower_reduction_clauses (tree clauses, gimple_seq *stmt_seqp, omp_context *ctx)
}
else
{
- if (is_gimple_omp_oacc (ctx->stmt)
- && is_oacc_multithreaded (clauses, ctx)
- && !oacc_needs_global_memory (clauses, ctx))
- {
- oacc_lower_reduction_var_helper (stmt_seqp, ctx, tid, var,
- new_var);
- }
- else if (is_oacc_parallel (ctx) && is_reference (var))
+ if (is_oacc_parallel (ctx) && is_reference (var))
{
tree t1, t2;
tree type = TREE_TYPE (new_var);
@@ -5285,11 +4877,6 @@ lower_reduction_clauses (tree clauses, gimple_seq *stmt_seqp, omp_context *ctx)
}
}
- if (is_gimple_omp_oacc (ctx->stmt)
- && is_oacc_multithreaded (clauses, ctx)
- && !oacc_needs_global_memory (clauses, ctx))
- return;
-
stmt = gimple_build_call (builtin_decl_explicit (BUILT_IN_GOMP_ATOMIC_START),
0);
gimple_seq_add_stmt (stmt_seqp, stmt);
@@ -11337,471 +10924,6 @@ make_pass_expand_omp_ssa (gcc::context *ctxt)
return new pass_expand_omp_ssa (ctxt);
}
\f
-/* Routines to lower OMP directives into OMP-GIMPLE. */
-
-/* Helper function to preform, potentially COMPLEX_TYPE, operation and
- convert it to gimple. */
-static void
-oacc_gimple_assign (tree dest, tree_code op, tree src, gimple_seq *seq)
-{
- gimple stmt;
-
- if (TREE_CODE (TREE_TYPE (dest)) != COMPLEX_TYPE)
- {
- stmt = gimple_build_assign (dest, op, dest, src);
- gimple_seq_add_stmt (seq, stmt);
- return;
- }
-
- tree t = create_tmp_var (TREE_TYPE (TREE_TYPE (dest)));
- tree rdest = fold_build1 (REALPART_EXPR, TREE_TYPE (TREE_TYPE (dest)), dest);
- gimplify_assign (t, rdest, seq);
- rdest = t;
-
- t = create_tmp_var (TREE_TYPE (TREE_TYPE (dest)));
- tree idest = fold_build1 (IMAGPART_EXPR, TREE_TYPE (TREE_TYPE (dest)), dest);
- gimplify_assign (t, idest, seq);
- idest = t;
-
- t = create_tmp_var (TREE_TYPE (TREE_TYPE (src)));
- tree rsrc = fold_build1 (REALPART_EXPR, TREE_TYPE (TREE_TYPE (src)), src);
- gimplify_assign (t, rsrc, seq);
- rsrc = t;
-
- t = create_tmp_var (TREE_TYPE (TREE_TYPE (src)));
- tree isrc = fold_build1 (IMAGPART_EXPR, TREE_TYPE (TREE_TYPE (src)), src);
- gimplify_assign (t, isrc, seq);
- isrc = t;
-
- tree r = create_tmp_var (TREE_TYPE (TREE_TYPE (dest)));
- tree i = create_tmp_var (TREE_TYPE (TREE_TYPE (dest)));
- tree result;
-
- if (op == PLUS_EXPR)
- {
- stmt = gimple_build_assign (r, op, rdest, rsrc);
- gimple_seq_add_stmt (seq, stmt);
-
- stmt = gimple_build_assign (i, op, idest, isrc);
- gimple_seq_add_stmt (seq, stmt);
- }
- else if (op == MULT_EXPR)
- {
- /* Let x = a + ib = dest, y = c + id = src.
- x * y = (ac - bd) + i(ad + bc) */
- tree ac = create_tmp_var (TREE_TYPE (TREE_TYPE (dest)));
- tree bd = create_tmp_var (TREE_TYPE (TREE_TYPE (dest)));
- tree ad = create_tmp_var (TREE_TYPE (TREE_TYPE (dest)));
- tree bc = create_tmp_var (TREE_TYPE (TREE_TYPE (dest)));
-
- stmt = gimple_build_assign (ac, MULT_EXPR, rdest, rsrc);
- gimple_seq_add_stmt (seq, stmt);
-
- stmt = gimple_build_assign (bd, MULT_EXPR, idest, isrc);
- gimple_seq_add_stmt (seq, stmt);
-
- stmt = gimple_build_assign (r, MINUS_EXPR, ac, bd);
- gimple_seq_add_stmt (seq, stmt);
-
- stmt = gimple_build_assign (ad, MULT_EXPR, rdest, isrc);
- gimple_seq_add_stmt (seq, stmt);
-
- stmt = gimple_build_assign (bd, MULT_EXPR, idest, rsrc);
- gimple_seq_add_stmt (seq, stmt);
-
- stmt = gimple_build_assign (i, PLUS_EXPR, ad, bc);
- gimple_seq_add_stmt (seq, stmt);
- }
- else
- gcc_unreachable ();
-
- result = build2 (COMPLEX_EXPR, TREE_TYPE (dest), r, i);
- gimplify_assign (dest, result, seq);
-}
-
-/* Helper function to initialize local data for the reduction arrays.
- The reduction arrays need to be placed inside the calling function
- for accelerators, or else the host won't be able to preform the final
- reduction. */
-
-static void
-oacc_initialize_reduction_data (tree clauses, tree nthreads,
- gimple_seq *stmt_seqp, omp_context *ctx)
-{
- tree c, t, oc;
- gimple stmt;
- omp_context *octx;
-
- /* Find the innermost PARALLEL openmp context. This is needed to insert
- new data mappings to the containing parallel/kernels region. */
- octx = oacc_outermost_parallel_kernels_context (ctx);
-
- /* Extract the clauses. */
- oc = gimple_omp_target_clauses (octx->stmt);
-
- /* Find the last parallel/kernels region clause. */
- for (; oc && OMP_CLAUSE_CHAIN (oc); oc = OMP_CLAUSE_CHAIN (oc))
- ;
-
- /* Allocate arrays for each reduction variable. */
- for (c = clauses; c; c = OMP_CLAUSE_CHAIN (c))
- {
- if (OMP_CLAUSE_CODE (c) != OMP_CLAUSE_REDUCTION)
- continue;
-
- if (!oacc_needs_global_memory (clauses, ctx))
- continue;
-
- tree var = OMP_CLAUSE_DECL (c);
- tree type = get_base_type (var);
- tree array = lookup_oacc_reduction (oacc_get_reduction_array_id (var),
- ctx);
- tree size, call;
-
- /* Calculate size of the reduction array. */
- t = create_tmp_var (TREE_TYPE (nthreads));
- stmt = gimple_build_assign (t, MULT_EXPR, nthreads,
- fold_convert (TREE_TYPE (nthreads),
- TYPE_SIZE_UNIT (type)));
- gimple_seq_add_stmt (stmt_seqp, stmt);
-
- size = create_tmp_var (sizetype);
- gimplify_assign (size, fold_build1 (NOP_EXPR, sizetype, t), stmt_seqp);
-
- /* Now allocate memory for it. FIXME: Allocating memory for the
- reduction array may be unnecessary once the final reduction is able
- to be preformed on the accelerator. Instead of allocating memory on
- the host side, it could just be allocated on the accelerator. */
- call = unshare_expr (builtin_decl_explicit (BUILT_IN_ALLOCA));
- stmt = gimple_build_call (call, 1, size);
- gimple_call_set_lhs (stmt, array);
- gimple_seq_add_stmt (stmt_seqp, stmt);
-
- /* Map this array into the accelerator. */
-
- /* Add the reduction array to the list of clauses. */
- /* FIXME: Currently, these variables must be placed in the outer
- most clause so that copy-out works. */
- tree x = array;
- t = build_omp_clause (gimple_location (ctx->stmt), OMP_CLAUSE_MAP);
- OMP_CLAUSE_SET_MAP_KIND (t, GOMP_MAP_FORCE_FROM);
- OMP_CLAUSE_DECL (t) = x;
- OMP_CLAUSE_CHAIN (t) = NULL;
- if (oc)
- OMP_CLAUSE_CHAIN (oc) = t;
- else
- gimple_omp_target_set_clauses (as_a <gomp_target *> (octx->stmt), t);
- OMP_CLAUSE_SIZE (t) = size;
- oc = t;
- }
-}
-
-/* Serial reduction helper. */
-
-static void
-oacc_serial_reduction (tree clauses, tree nthreads, tree tid,
- gimple_seq *stmt_seqp, omp_context *ctx,
- hash_map<tree, tree> &reduction_vars,
- bool is_accelerator)
-{
- tree c, x, var, orig_var, array, loop_header, loop_body, loop_exit, type;
- tree next, reduction_exit, t1, t2;
- gimple stmt;
-
- if (is_accelerator)
- {
- /* Synchronize all of the threads. */
- tree call = builtin_decl_explicit (BUILT_IN_SYNC_SYNCHRONIZE);
- stmt = gimple_build_call (call, 0);
- gimple_seq_add_stmt (stmt_seqp, stmt);
-
- /* Jump to the exit label if tid != 0. Ensure that TID considers
- the global TID (gang, worker and vector) not just the local
- loop-specific TID. */
-
- next = create_artificial_label (UNKNOWN_LOCATION);
- reduction_exit = create_artificial_label (UNKNOWN_LOCATION);
-
- t1 = create_tmp_var (sizetype, NULL);
- t2 = create_tmp_var (sizetype, NULL);
- gimplify_assign (t1, fold_build1 (NOP_EXPR, sizetype, tid),
- stmt_seqp);
- gimplify_assign (t2, fold_build1 (NOP_EXPR, sizetype,
- integer_zero_node),
- stmt_seqp);
- stmt = gimple_build_cond (NE_EXPR, t1, t2, reduction_exit, next);
- gimple_seq_add_stmt (stmt_seqp, stmt);
- gimple_seq_add_stmt (stmt_seqp, gimple_build_label (next));
- }
-
- /* Create for loop.
-
- let var = the original reduction variable
- let array = reduction variable array
-
- for (i = 0; i < nthreads; i++)
- var op= array[i]
- */
-
- loop_header = create_artificial_label (UNKNOWN_LOCATION);
- loop_body = create_artificial_label (UNKNOWN_LOCATION);
- loop_exit = create_artificial_label (UNKNOWN_LOCATION);
-
- /* Create and initialize an index variable. */
- tree ix = create_tmp_var (sizetype);
- gimplify_assign (ix, fold_build1 (NOP_EXPR, sizetype, integer_zero_node),
- stmt_seqp);
-
- /* Insert the loop header label here. */
- gimple_seq_add_stmt (stmt_seqp, gimple_build_label (loop_header));
-
- /* Exit loop if ix >= nthreads. */
- x = create_tmp_var (sizetype);
- gimplify_assign (x, fold_build1 (NOP_EXPR, sizetype, nthreads), stmt_seqp);
- stmt = gimple_build_cond (GE_EXPR, ix, x, loop_exit, loop_body);
- gimple_seq_add_stmt (stmt_seqp, stmt);
-
- /* Insert the loop body label here. */
- gimple_seq_add_stmt (stmt_seqp, gimple_build_label (loop_body));
-
- /* Collapse each reduction array, one element at a time. */
- for (c = clauses; c; c = OMP_CLAUSE_CHAIN (c))
- {
- if (OMP_CLAUSE_CODE (c) != OMP_CLAUSE_REDUCTION)
- continue;
-
- tree_code reduction_code = OMP_CLAUSE_REDUCTION_CODE (c);
-
- /* reduction(-:var) sums up the partial results, so it acts
- identically to reduction(+:var). */
- if (reduction_code == MINUS_EXPR)
- reduction_code = PLUS_EXPR;
-
- /* Set up reduction variable var. */
- orig_var = OMP_CLAUSE_DECL (c);
- type = get_base_type (orig_var);
-
- if (is_accelerator)
- orig_var = lookup_decl_in_outer_ctx (orig_var, ctx);
-
- /* Lookup the local-private reduction variable. */
- var = *reduction_vars.get (orig_var);
-
- array = lookup_oacc_reduction (oacc_get_reduction_array_id
- (OMP_CLAUSE_DECL (c)), ctx);
-
- /* Calculate the array offset. */
- tree offset = create_tmp_var (sizetype);
- gimplify_assign (offset, TYPE_SIZE_UNIT (type), stmt_seqp);
- stmt = gimple_build_assign (offset, MULT_EXPR, offset, ix);
- gimple_seq_add_stmt (stmt_seqp, stmt);
-
- tree ptr = create_tmp_var (TREE_TYPE (array));
- stmt = gimple_build_assign (ptr, POINTER_PLUS_EXPR, array, offset);
- gimple_seq_add_stmt (stmt_seqp, stmt);
-
- /* Extract array[ix] into mem. */
- tree mem = create_tmp_var (type);
- gimplify_assign (mem, build_simple_mem_ref (ptr), stmt_seqp);
-
- /* Find the original reduction variable. */
- if (is_reference (var))
- var = build_simple_mem_ref (var);
-
- tree t = create_tmp_var (type);
-
- x = lang_hooks.decls.omp_clause_assign_op (c, t, var);
- gimplify_and_add (unshare_expr(x), stmt_seqp);
-
- /* var = var op mem */
- switch (OMP_CLAUSE_REDUCTION_CODE (c))
- {
- case TRUTH_ANDIF_EXPR:
- case TRUTH_ORIF_EXPR:
- t = fold_build2 (OMP_CLAUSE_REDUCTION_CODE (c), integer_type_node,
- t, mem);
- gimplify_and_add (t, stmt_seqp);
- break;
- default:
- /* The lhs isn't a gimple_reg when var is COMPLEX_TYPE. */
- oacc_gimple_assign (t, OMP_CLAUSE_REDUCTION_CODE (c), mem,
- stmt_seqp);
- }
-
- t = fold_build1 (NOP_EXPR, TREE_TYPE (var), t);
- x = lang_hooks.decls.omp_clause_assign_op (c, var, t);
- gimplify_and_add (unshare_expr(x), stmt_seqp);
- }
-
- /* Increment the induction variable. */
- tree one = fold_build1 (NOP_EXPR, sizetype, integer_one_node);
- stmt = gimple_build_assign (ix, PLUS_EXPR, ix, one);
- gimple_seq_add_stmt (stmt_seqp, stmt);
-
- /* Go back to the top of the loop. */
- gimple_seq_add_stmt (stmt_seqp, gimple_build_goto (loop_header));
-
- /* Place the loop exit label here. */
- gimple_seq_add_stmt (stmt_seqp, gimple_build_label (loop_exit));
-
- /* Finally, update the original reduction variables. */
- for (c = clauses; c; c = OMP_CLAUSE_CHAIN (c))
- {
- if (OMP_CLAUSE_CODE (c) != OMP_CLAUSE_REDUCTION)
- continue;
-
- /* Set up reduction variable var. */
- orig_var = OMP_CLAUSE_DECL (c);
- if (is_accelerator)
- orig_var = lookup_decl_in_outer_ctx (orig_var, ctx);
- type = get_base_type (orig_var);
-
- var = *reduction_vars.get (orig_var);
- x = lang_hooks.decls.omp_clause_assign_op (c, orig_var, var);
- gimplify_and_add (unshare_expr (x), stmt_seqp);
- }
-
- if (is_accelerator)
- gimple_seq_add_stmt (stmt_seqp, gimple_build_label (reduction_exit));
-}
-
-/* Helper function to finalize local data for the reduction arrays. The
- reduction array needs to be reduced to the original reduction variable.
-
- FIXME: This function assumes that there are vector_length threads in
- total. Also, it assumes that there are at least vector_length iterations
- in the for loop.
-
- FIXME: This function should also take reduction variables which are also
- private.
-*/
-
-static void
-oacc_finalize_reduction_data (tree clauses, tree nthreads, tree tid,
- gimple_seq *stmt_seqp, omp_context *ctx,
- bool is_accelerator)
-{
- gcc_assert (is_gimple_omp_oacc (ctx->stmt));
- tree c, x, var, orig_var, type;
-
- /* Create a private copy of each original reduction variable to preform
- the intermediate reduction result. The original reduction variable
- cannot be used because it may need to be updated at the end of a loop
- or parallel region depending on whether a gang clause is present. */
-
- /* Create a hash_map for the gang-local temporary reduction variables. */
- hash_map<tree, tree> reduction_vars;
-
- for (c = clauses; c; c = OMP_CLAUSE_CHAIN (c))
- {
- if (OMP_CLAUSE_CODE (c) != OMP_CLAUSE_REDUCTION)
- continue;
-
- /* Set up reduction variable var. */
- orig_var = OMP_CLAUSE_DECL (c);
-
- if (is_accelerator)
- orig_var = lookup_decl_in_outer_ctx (orig_var, ctx);
-
- type = get_base_type (orig_var);
- var = create_tmp_var (type, NULL);
- x = lang_hooks.decls.omp_clause_assign_op (c, var, orig_var);
- gimplify_and_add (unshare_expr (x), stmt_seqp);
- reduction_vars.put (orig_var, var);
- }
-
- oacc_serial_reduction (clauses, nthreads, tid, stmt_seqp, ctx,
- reduction_vars, is_accelerator);
-}
-
-static void
-oacc_process_reduction_data_helper (gimple_seq stmt, gimple_seq *in_stmt_seqp,
- gimple_seq *out_stmt_seqp /* TODO */ ATTRIBUTE_UNUSED, tree clauses,
- omp_context *ctx)
-{
- switch (gimple_code (stmt))
- {
- case GIMPLE_OMP_FOR:
- clauses = gimple_omp_for_clauses (stmt);
- break;
- case GIMPLE_OMP_TARGET:
- clauses = gimple_omp_target_clauses (stmt);
- break;
- default:
- clauses = NULL;
- }
-
- /* Allocate an array of num_gangs threads for the parallel reduction, and
- finalize the result when the parallel region exits. */
- if (is_oacc_parallel (ctx)
- && find_omp_clause (clauses, OMP_CLAUSE_REDUCTION))
- {
- tree nthreads;
- nthreads = oacc_parallel_max_reduction_array_size (ctx, clauses,
- in_stmt_seqp);
-
- if (!oacc_needs_global_memory (clauses, ctx))
- oacc_initialize_reduction_data (clauses, nthreads, in_stmt_seqp,
- ctx);
- }
-}
-
-/* Scan through all of the gimple stmts searching for an OMP_FOR_EXPR, and
- scan that for reductions. Allocate memory for the arrays of partial
- reductions and process them as necessary. All memory involving gangs
- is allocated via a memory map, otherwise gang-local shared memory is
- used. */
-
-static void
-oacc_process_reduction_data (gimple_seq *body, gimple_seq *in_stmt_seqp,
- gimple_seq *out_stmt_seqp, omp_context *ctx)
-{
- gimple_stmt_iterator gsi;
- gimple_seq inner = NULL;
- gimple stmt;
-
- gsi = gsi_start (*body);
- while (!gsi_end_p (gsi))
- {
- stmt = gsi_stmt (gsi);
- /* A collapse clause may have inserted a new bind block. */
- if (gbind *bind_stmt = dyn_cast <gbind *> (stmt))
- {
- /* Recursively scan for nested loops. */
- inner = gimple_bind_body (bind_stmt);
- oacc_process_reduction_data (&inner, in_stmt_seqp, out_stmt_seqp,
- ctx);
- }
- else if (gimple_code (stmt) == GIMPLE_OMP_FOR)
- {
- tree clauses, c;
- bool reduction_found = false;
-
- stmt = gsi_stmt (gsi);
- clauses = gimple_omp_for_clauses (stmt);
-
- /* Search for a reduction clause. */
- for (c = clauses; c; c = OMP_CLAUSE_CHAIN (c))
- if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_REDUCTION)
- {
- reduction_found = true;
- break;
- }
-
- if (reduction_found && is_oacc_multithreaded (clauses, ctx)
- && !oacc_needs_global_memory (clauses, ctx))
- oacc_process_reduction_data_helper (stmt, in_stmt_seqp,
- out_stmt_seqp, clauses,
- ctx);
-
- /* Now recursively scan for nested loops. */
- oacc_process_reduction_data (gimple_omp_body_ptr (stmt),
- in_stmt_seqp, out_stmt_seqp, ctx);
- }
- gsi_next (&gsi);
- }
-}
-
/* Routines to lower OpenMP directives into OMP-GIMPLE. */
/* If ctx is a worksharing context inside of a cancellable parallel
@@ -13228,7 +12350,6 @@ lower_omp_target (gimple_stmt_iterator *gsi_p, omp_context *ctx)
if (has_reduction)
{
lower_rec_input_clauses (clauses, &irlist, &orlist, ctx, NULL);
- oacc_process_reduction_data (&tgt_body, &irlist, &orlist, ctx);
lower_reduction_clauses (clauses, &orlist, ctx);
}
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2015-07-01 13:45 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2015-06-30 18:07 [gomp] openacc reduction cleanup Cesar Philippidis
2015-07-01 13:45 ` Cesar Philippidis
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).