Re: [hsa 7/12] Disabling the vectorizer for GPU kernels/functions

[Martin Jambor <mjambor at suse dot cz>]

On Fri, Nov 06, 2015 at 09:38:21AM +0100, Richard Biener wrote:
> On Thu, 5 Nov 2015, Martin Jambor wrote:
> 
> > Hi,
> > 
> > in the previous email I wrote we need to "change behavior" of a few
> > optimization passes.  One was the flattening of GPU functions and the
> > other two are in the patch below.  It all comes to that, at the
> > moment, we need to switch off the vectorizer (only for the GPU
> > functions, of course).
> > 
> > We are actually quite close to being able to handle gimple vector
> > input in HSA back-end but not all the way yet, and before allowing the
> > vectorizer again, we will have to make sure it never produces vectors
> > bigger than 128bits (in GPU functions).
> 
> Hmm.  I'd rather have this modify
> DECL_FUNCTION_SPECIFIC_OPTIMIZATION of the hsa function to get this
> effect.  I think I mentioned this to the OACC guys as well for a
> similar needs of them.

I see, that is a good idea.  I have reverted changes to
tree-ssa-loop.c and tree-vectorizer.c and on top of that committed the
following patch to the branch which makes modifications to HSA fndecls
at a more convenient spot and disables vectorization in the following
way:

  tree gdecl = gpu->decl;
  tree fn_opts = DECL_FUNCTION_SPECIFIC_OPTIMIZATION (gdecl);
  if (fn_opts == NULL_TREE)
    fn_opts = optimization_default_node;
  fn_opts = copy_node (fn_opts);
  TREE_OPTIMIZATION (fn_opts)->x_flag_tree_loop_vectorize = false;
  TREE_OPTIMIZATION (fn_opts)->x_flag_tree_slp_vectorize = false;
  DECL_FUNCTION_SPECIFIC_OPTIMIZATION (gdecl) = fn_opts;

I hope that is what you meant.  I have also verified that it works.

Thanks,

Martin


2015-11-10  Martin Jambor  <mjambor@suse.cz>

	* hsa.h (hsa_summary_t): Add a comment to method link_functions.
	(hsa_summary_t::link_functions): Moved...
	* hsa.c (hsa_summary_t::link_functions): ...here.  Added common fndecl
	modifications.
	Include stringpool.h.
	* ipa-hsa.c (process_hsa_functions): Do not add flatten attribute
	here.  Fixed comments.

diff --git a/gcc/hsa.c b/gcc/hsa.c
index ab05a1d..e63be95 100644
--- a/gcc/hsa.c
+++ b/gcc/hsa.c
@@ -34,6 +34,7 @@ along with GCC; see the file COPYING3.  If not see
 #include "alloc-pool.h"
 #include "cgraph.h"
 #include "print-tree.h"
+#include "stringpool.h"
 #include "symbol-summary.h"
 #include "hsa.h"
 
@@ -693,6 +694,40 @@ hsa_get_declaration_name (tree decl)
   return NULL;
 }
 
+/* Couple GPU and HOST as gpu-specific and host-specific implementation of the
+   same function.  KIND determines whether GPU is a host-invokable kernel or
+   gpu-callable function.  */
+
+inline void
+hsa_summary_t::link_functions (cgraph_node *gpu, cgraph_node *host,
+			       hsa_function_kind kind)
+{
+  hsa_function_summary *gpu_summary = get (gpu);
+  hsa_function_summary *host_summary = get (host);
+
+  gpu_summary->m_kind = kind;
+  host_summary->m_kind = kind;
+
+  gpu_summary->m_gpu_implementation_p = true;
+  host_summary->m_gpu_implementation_p = false;
+
+  gpu_summary->m_binded_function = host;
+  host_summary->m_binded_function = gpu;
+
+  tree gdecl = gpu->decl;
+  DECL_ATTRIBUTES (gdecl)
+    = tree_cons (get_identifier ("flatten"), NULL_TREE,
+		 DECL_ATTRIBUTES (gdecl));
+
+  tree fn_opts = DECL_FUNCTION_SPECIFIC_OPTIMIZATION (gdecl);
+  if (fn_opts == NULL_TREE)
+    fn_opts = optimization_default_node;
+  fn_opts = copy_node (fn_opts);
+  TREE_OPTIMIZATION (fn_opts)->x_flag_tree_loop_vectorize = false;
+  TREE_OPTIMIZATION (fn_opts)->x_flag_tree_slp_vectorize = false;
+  DECL_FUNCTION_SPECIFIC_OPTIMIZATION (gdecl) = fn_opts;
+}
+
 /* Add a HOST function to HSA summaries.  */
 
 void
diff --git a/gcc/hsa.h b/gcc/hsa.h
index 025de67..b6855ea 100644
--- a/gcc/hsa.h
+++ b/gcc/hsa.h
@@ -1161,27 +1161,14 @@ public:
   hsa_summary_t (symbol_table *table):
     function_summary<hsa_function_summary *> (table) { }
 
+  /* Couple GPU and HOST as gpu-specific and host-specific implementation of
+     the same function.  KIND determines whether GPU is a host-invokable kernel
+     or gpu-callable function.  */
+
   void link_functions (cgraph_node *gpu, cgraph_node *host,
 		       hsa_function_kind kind);
 };
 
-inline void
-hsa_summary_t::link_functions (cgraph_node *gpu, cgraph_node *host,
-			       hsa_function_kind kind)
-{
-  hsa_function_summary *gpu_summary = get (gpu);
-  hsa_function_summary *host_summary = get (host);
-
-  gpu_summary->m_kind = kind;
-  host_summary->m_kind = kind;
-
-  gpu_summary->m_gpu_implementation_p = true;
-  host_summary->m_gpu_implementation_p = false;
-
-  gpu_summary->m_binded_function = host;
-  host_summary->m_binded_function = gpu;
-}
-
 /* in hsa.c */
 extern struct hsa_function_representation *hsa_cfun;
 extern hash_map <tree, vec <const char *> *> *hsa_decl_kernel_dependencies;
diff --git a/gcc/ipa-hsa.c b/gcc/ipa-hsa.c
index b4cb58e..d77fa6b 100644
--- a/gcc/ipa-hsa.c
+++ b/gcc/ipa-hsa.c
@@ -90,16 +90,12 @@ process_hsa_functions (void)
 	  cgraph_node *clone = node->create_virtual_clone
 	    (vec <cgraph_edge *> (), NULL, NULL, "hsa");
 	  TREE_PUBLIC (clone->decl) = TREE_PUBLIC (node->decl);
-	  if (s->m_kind == HSA_KERNEL)
-	    DECL_ATTRIBUTES (clone->decl)
-	      = tree_cons (get_identifier ("flatten"), NULL_TREE,
-			   DECL_ATTRIBUTES (clone->decl));
 
 	  clone->force_output = true;
 	  hsa_summaries->link_functions (clone, node, s->m_kind);
 
 	  if (dump_file)
-	    fprintf (dump_file, "HSA creates a new clone: %s, type: %s\n",
+	    fprintf (dump_file, "Created a new HSA clone: %s, type: %s\n",
 		     clone->name (),
 		     s->m_kind == HSA_KERNEL ? "kernel" : "function");
 	}
@@ -116,7 +112,7 @@ process_hsa_functions (void)
 	  hsa_summaries->link_functions (clone, node, HSA_FUNCTION);
 
 	  if (dump_file)
-	    fprintf (dump_file, "HSA creates a new function clone: %s\n",
+	    fprintf (dump_file, "Created a new HSA function clone: %s\n",
 		     clone->name ());
 	}
     }