diff mbox

[GRAPHITE] Ping: Loop unroll and jam optimization

Message ID 1946236345.2179089.1416049070123.JavaMail.zimbra@inria.fr
State New
Headers show

Commit Message

Mircea Namolaru Nov. 15, 2014, 10:57 a.m. UTC 
The close of stage 1 is getting close (very close). Even there is not so much new code (basically
the new code computes the separation class option for AST build), I am not sure that the patch 
qualify for stage 2.

There is very nice code generated by unroll-and-jam (stride mining) for small kernels both for constant 
or non-constant bound loops, and is an argument for the new isl based code generator. Otherwise I'm afraid 
that the code generated looks very similar with the cloog generated one, an inner loop
with bounds of min/max that GCC doesn't further optimize, preventing perceived advantages of 
strip mining (register reuse and scalar reduction, instruction scheduling etc).

ok for trunk ?

Thanks, Mircea
2014-11-12  Mircea Namolaru  <mircea.namolaru@inria.fr>

	* common.opt (flag_loop_unroll_and_jam): New flag for unroll and jam.
	* params.def (PARAM_LOOP_UNROLL_JAM_SIZE)
	(PARAM_LOOP_UNROLL_JAM_DEPTH): Parameters for unroll and jam flag.
	* graphite-poly.h (struct poly_bb:map_sepclass): New field
	* graphite-poly.c (new_poly_bb): New field intialization.
	* graphite-isl-ast-to-gimple.c (generate_luj_sepclass_opt,
	generate_luj_sepclass,generate_luj_options):
        New functions to set AST options for unroll and jam.
        * graphite-isl-ast-to-gimple.c (set_options,scop_to_isl_ast):
        Added support for unroll and jam options.
	* graphite-optimize-isl.c (getPrevectorMap_full): New function for
        sepating class map.
	* graphite-optimize-isl.c (optimize_isl,apply_schedule_map_to_scop,
	getScheduleMap,getScheduleForBand,getScheduleForBandList):
	Added support for the separating class map.
        * graphite.c: Support for unroll and jam flag.
	* graphite-poly.c: Likewise.
	* toplev.c: Likewise.

Comments

Richard Biener Nov. 17, 2014, 10:02 a.m. UTC | #1 
On Sat, Nov 15, 2014 at 11:57 AM, Mircea Namolaru
<mircea.namolaru@inria.fr> wrote:
> The close of stage 1 is getting close (very close). Even there is not so much new code (basically
> the new code computes the separation class option for AST build), I am not sure that the patch
> qualify for stage 2.
>
> There is very nice code generated by unroll-and-jam (stride mining) for small kernels both for constant
> or non-constant bound loops, and is an argument for the new isl based code generator. Otherwise I'm afraid
> that the code generated looks very similar with the cloog generated one, an inner loop
> with bounds of min/max that GCC doesn't further optimize, preventing perceived advantages of
> strip mining (register reuse and scalar reduction, instruction scheduling etc).
>
> ok for trunk ?

New optimization flags and new params need documentation in
gcc/doc/invoke.texi.

The description of the --params suggest they provide fixed values - is
there no way to autodetect sensible values with a cost-model?  I
hardly doubt that you can find two fixed values that apply for a whole
program...

Richard.

> Thanks, Mircea
>
Mircea Namolaru Nov. 17, 2014, 11:04 p.m. UTC | #2 
> New optimization flags and new params need documentation in
> gcc/doc/invoke.texi.
> 

Thanks. Added description in invoke.texi. The patch is in trunk.

> The description of the --params suggest they provide fixed values - is
> there no way to autodetect sensible values with a cost-model?  I
> hardly doubt that you can find two fixed values that apply for a whole
> program...

There are a lot of models/heuristics, but in the general case tile sizes remain
kind of an open problem. For the time being, this option and its parameters are intended 
mostly to compiler developers interested to bring the loop to a specific form enabling
further optimizations.

Mircea.
diff mbox

Patch

Index: gcc/graphite-poly.h
===================================================================
--- gcc/graphite-poly.h	(revision 217013)
+++ gcc/graphite-poly.h	(working copy)
@@ -349,6 +349,9 @@ 
   poly_scattering_p _saved;
   isl_map *saved;
 
+  /* For tiling, the map for computing the separating class.  */
+  isl_map *map_sepclass;
+
   /* True when this PBB contains only a reduction statement.  */
   bool is_reduction;
 };
Index: gcc/graphite.c
===================================================================
--- gcc/graphite.c	(revision 217013)
+++ gcc/graphite.c	(working copy)
@@ -383,7 +383,8 @@ 
       || flag_loop_strip_mine
       || flag_graphite_identity
       || flag_loop_parallelize_all
-      || flag_loop_optimize_isl)
+      || flag_loop_optimize_isl
+      || flag_loop_unroll_jam)
     flag_graphite = 1;
 
   return flag_graphite != 0;
Index: gcc/common.opt
===================================================================
--- gcc/common.opt	(revision 217013)
+++ gcc/common.opt	(working copy)
@@ -1328,6 +1328,10 @@ 
 Common Report Var(flag_loop_block) Optimization
 Enable Loop Blocking transformation
 
+floop-unroll-and-jam
+Common Report Var(flag_loop_unroll_jam) Optimization
+Enable Loop Unroll Jam transformation
+ 
 fgnu-tm
 Common Report Var(flag_tm)
 Enable support for GNU transactional memory
Index: gcc/graphite-optimize-isl.c
===================================================================
--- gcc/graphite-optimize-isl.c	(revision 217013)
+++ gcc/graphite-optimize-isl.c	(working copy)
@@ -186,7 +186,7 @@ 
   PartialSchedule = isl_band_get_partial_schedule (Band);
   *Dimensions = isl_band_n_member (Band);
 
-  if (DisableTiling)
+  if (DisableTiling || flag_loop_unroll_jam)
     return PartialSchedule;
 
   /* It does not make any sense to tile a band with just one dimension.  */
@@ -241,7 +241,9 @@ 
    constant number of iterations, if the number of loop iterations at
    DimToVectorize can be devided by VectorWidth. The default VectorWidth is
    currently constant and not yet target specific. This function does not reason
-   about parallelism.  */
+   about parallelism.
+
+  */
 static isl_map *
 getPrevectorMap (isl_ctx *ctx, int DimToVectorize,
 		 int ScheduleDimensions,
@@ -305,8 +307,98 @@ 
   isl_constraint_set_constant_si (c, VectorWidth - 1);
   TilingMap = isl_map_add_constraint (TilingMap, c);
 
-  isl_map_dump (TilingMap);
+  return TilingMap;
+}
 
+/* Compute an auxiliary map to getPrevectorMap, for computing the separating 
+   class defined by full tiles.  Used in graphite_isl_ast_to_gimple.c to set the 
+   corresponding option for AST build.
+
+   The map (for VectorWidth=4):
+
+   [i,j] -> [it,j,ip] : it % 4 = 0 and it <= ip <= it + 3 and it + 3 = i and
+                        ip >= 0
+
+   The image of this map is the separation class. The range of this map includes
+   all the i that are multiple of 4 in the domain beside the greater one. 
+    
+ */ 
+static isl_map *
+getPrevectorMap_full (isl_ctx *ctx, int DimToVectorize,
+		 int ScheduleDimensions,
+		 int VectorWidth)
+{
+  isl_space *Space;
+  isl_local_space *LocalSpace, *LocalSpaceRange;
+  isl_set *Modulo;
+  isl_map *TilingMap;
+  isl_constraint *c;
+  isl_aff *Aff;
+  int PointDimension; /* ip */
+  int TileDimension;  /* it */
+  isl_val *VectorWidthMP;
+  int i;
+
+  /* assert (0 <= DimToVectorize && DimToVectorize < ScheduleDimensions);*/
+
+  Space = isl_space_alloc (ctx, 0, ScheduleDimensions, ScheduleDimensions + 1);
+  TilingMap = isl_map_universe (isl_space_copy (Space));
+  LocalSpace = isl_local_space_from_space (Space);
+  PointDimension = ScheduleDimensions;
+  TileDimension = DimToVectorize;
+
+  /* Create an identity map for everything except DimToVectorize and the 
+     point loop. */
+  for (i = 0; i < ScheduleDimensions; i++)
+    {
+      if (i == DimToVectorize)
+        continue;
+
+      c = isl_equality_alloc (isl_local_space_copy (LocalSpace));
+
+      isl_constraint_set_coefficient_si (c, isl_dim_in, i, -1);
+      isl_constraint_set_coefficient_si (c, isl_dim_out, i, 1);
+
+      TilingMap = isl_map_add_constraint (TilingMap, c);
+    }
+
+  /* it % 'VectorWidth' = 0  */
+  LocalSpaceRange = isl_local_space_range (isl_local_space_copy (LocalSpace));
+  Aff = isl_aff_zero_on_domain (LocalSpaceRange);
+  Aff = isl_aff_set_constant_si (Aff, VectorWidth);
+  Aff = isl_aff_set_coefficient_si (Aff, isl_dim_in, TileDimension, 1);
+
+  VectorWidthMP = isl_val_int_from_si (ctx, VectorWidth);
+  Aff = isl_aff_mod_val (Aff, VectorWidthMP);
+  Modulo = isl_pw_aff_zero_set (isl_pw_aff_from_aff (Aff));
+  TilingMap = isl_map_intersect_range (TilingMap, Modulo);
+
+  /* it + ('VectorWidth' - 1) = i0  */
+  c = isl_equality_alloc (isl_local_space_copy(LocalSpace));
+  isl_constraint_set_coefficient_si (c, isl_dim_out, TileDimension,-1);
+  isl_constraint_set_coefficient_si (c, isl_dim_in, TileDimension, 1);
+  isl_constraint_set_constant_si (c, -VectorWidth + 1);
+  TilingMap = isl_map_add_constraint (TilingMap, c);
+
+  /* ip >= 0 */
+  c = isl_inequality_alloc (isl_local_space_copy (LocalSpace));
+  isl_constraint_set_coefficient_si (c, isl_dim_out, PointDimension, 1);
+  isl_constraint_set_constant_si (c, 0);
+  TilingMap = isl_map_add_constraint (TilingMap, c);
+
+  /* it <= ip */
+  c = isl_inequality_alloc (isl_local_space_copy (LocalSpace));
+  isl_constraint_set_coefficient_si (c, isl_dim_out, TileDimension, -1);
+  isl_constraint_set_coefficient_si (c, isl_dim_out, PointDimension, 1);
+  TilingMap = isl_map_add_constraint (TilingMap, c);
+
+  /* ip <= it + ('VectorWidth' - 1) */
+  c = isl_inequality_alloc (LocalSpace);
+  isl_constraint_set_coefficient_si (c, isl_dim_out, TileDimension, 1);
+  isl_constraint_set_coefficient_si (c, isl_dim_out, PointDimension, -1);
+  isl_constraint_set_constant_si (c, VectorWidth - 1);
+  TilingMap = isl_map_add_constraint (TilingMap, c);
+
   return TilingMap;
 }
 
@@ -316,9 +408,11 @@ 
     
    We walk recursively the forest of bands to combine the schedules of the
    individual bands to the overall schedule. In case tiling is requested,
-   the individual bands are tiled.  */
+   the individual bands are tiled.
+   For unroll and jam the map the schedule for full tiles of the unrolled
+   dimnesion is computed.  */
 static isl_union_map *
-getScheduleForBandList (isl_band_list *BandList)
+getScheduleForBandList (isl_band_list *BandList, isl_union_map **map_sepcl)
 {
   int NumBands, i;
   isl_union_map *Schedule;
@@ -335,55 +429,87 @@ 
       int ScheduleDimensions;
       isl_space *Space;
 
+      isl_union_map *PartialSchedule_f;
+
       Band = isl_band_list_get_band (BandList, i);
       PartialSchedule = getScheduleForBand (Band, &ScheduleDimensions);
       Space = isl_union_map_get_space (PartialSchedule);
 
+      PartialSchedule_f = NULL;
+
       if (isl_band_has_children (Band))
 	{
 	  isl_band_list *Children;
 	  isl_union_map *SuffixSchedule;
 
 	  Children = isl_band_get_children (Band);
-	  SuffixSchedule = getScheduleForBandList (Children);
+	  SuffixSchedule = getScheduleForBandList (Children, map_sepcl);
 	  PartialSchedule = isl_union_map_flat_range_product (PartialSchedule,
 							      SuffixSchedule);
 	  isl_band_list_free (Children);
 	}
-      else if (EnablePollyVector)
+      else if (EnablePollyVector || flag_loop_unroll_jam)
 	{
+	  int i;
+	  int depth;
+ 
+ 	  depth = PARAM_VALUE (PARAM_LOOP_UNROLL_JAM_DEPTH);
+  
 	  for (i = ScheduleDimensions - 1 ;  i >= 0 ; i--)
 	    {
+	      if (flag_loop_unroll_jam && (i != (ScheduleDimensions - depth)))
+		continue;
+
 	      if (isl_band_member_is_zero_distance (Band, i))
 		{
 		  isl_map *TileMap;
 		  isl_union_map *TileUMap;
+		  int stride;
 
-		  TileMap = getPrevectorMap (ctx, i, ScheduleDimensions, 4);
+                  stride = PARAM_VALUE (PARAM_LOOP_UNROLL_JAM_SIZE);    
+
+		  TileMap = getPrevectorMap_full (ctx, i, ScheduleDimensions, 
+						  stride); 
+ 		  TileUMap = isl_union_map_from_map (TileMap);
+		  TileUMap = isl_union_map_align_params
+		    (TileUMap, isl_space_copy (Space));
+		  PartialSchedule_f = isl_union_map_apply_range
+		    (isl_union_map_copy (PartialSchedule), TileUMap);
+
+		  TileMap = getPrevectorMap (ctx, i, ScheduleDimensions, stride);
 		  TileUMap = isl_union_map_from_map (TileMap);
 		  TileUMap = isl_union_map_align_params
 		    (TileUMap, isl_space_copy (Space));
 		  PartialSchedule = isl_union_map_apply_range
 		    (PartialSchedule, TileUMap);
 		  break;
-		}
+		}	
 	    }
 	}
-
       Schedule = isl_union_map_union (Schedule, PartialSchedule);
 
       isl_band_free (Band);
       isl_space_free (Space);
+
+      if (!flag_loop_unroll_jam)
+	continue;
+
+      if (PartialSchedule_f)
+	*map_sepcl = isl_union_map_union (*map_sepcl, 
+					  PartialSchedule_f);
+      else
+        *map_sepcl = isl_union_map_union (*map_sepcl, 
+         				  isl_union_map_copy (PartialSchedule));
     }
 
   return Schedule;
 }
 
 static isl_union_map *
-getScheduleMap (isl_schedule *Schedule)
+getScheduleMap (isl_schedule *Schedule, isl_union_map **map_sepcl)
 {
   isl_band_list *BandList = isl_schedule_get_band_forest (Schedule);
-  isl_union_map *ScheduleMap = getScheduleForBandList (BandList);
+  isl_union_map *ScheduleMap = getScheduleForBandList (BandList, map_sepcl);
   isl_band_list_free (BandList);
   return ScheduleMap;
 }
@@ -398,7 +524,7 @@ 
 }
 
 static void
-apply_schedule_map_to_scop (scop_p scop, isl_union_map *schedule_map)
+apply_schedule_map_to_scop (scop_p scop, isl_union_map *schedule_map, bool sepcl)
 {
   int i;
   poly_bb_p pbb;
@@ -413,8 +539,15 @@ 
 	(isl_union_map_copy (schedule_map),
 	 isl_union_set_from_set (domain));
       isl_union_map_foreach_map (stmtBand, getSingleMap, &stmtSchedule);
-      isl_map_free (pbb->transformed);
-      pbb->transformed = stmtSchedule;
+
+      if (!sepcl)
+	{
+	  isl_map_free (pbb->transformed);
+	  pbb->transformed = stmtSchedule;
+	}
+      else
+	  pbb->map_sepclass = stmtSchedule;
+
       isl_union_map_free (stmtBand);
     }
 }
@@ -429,6 +562,7 @@ 
   isl_union_set *domain;
   isl_union_map *validity, *proximity, *dependences;
   isl_union_map *schedule_map;
+  isl_union_map *schedule_map_f;
 
   domain = scop_get_domains (scop);
   dependences = scop_get_dependences (scop);
@@ -450,9 +584,13 @@ 
   if (!schedule)
     return false;
 
-  schedule_map = getScheduleMap (schedule);
+  schedule_map_f = isl_union_map_empty (isl_space_params_alloc (scop->ctx, 0));
+  schedule_map = getScheduleMap (schedule, &schedule_map_f);
 
-  apply_schedule_map_to_scop (scop, schedule_map);
+  apply_schedule_map_to_scop (scop, schedule_map, false);
+  if (!isl_union_map_is_empty (schedule_map_f))
+    apply_schedule_map_to_scop (scop, schedule_map_f, true);
+  isl_union_map_free (schedule_map_f);
 
   isl_schedule_free (schedule);
   isl_union_map_free (schedule_map);
Index: gcc/graphite-isl-ast-to-gimple.c
===================================================================
--- gcc/graphite-isl-ast-to-gimple.c	(revision 217013)
+++ gcc/graphite-isl-ast-to-gimple.c	(working copy)
@@ -831,6 +831,92 @@ 
   return schedule;
 }
 
+/* Set the separation_class option for unroll and jam. */
+
+static __isl_give isl_union_map *
+generate_luj_sepclass_opt (scop_p scop, __isl_take isl_union_set *domain, 
+			int dim, int cl)
+{
+  isl_map  *map;
+  isl_space *space, *space_sep;
+  isl_ctx *ctx;
+  isl_union_map *mapu;
+  int nsched = get_max_schedule_dimensions (scop);
+ 
+  ctx = scop->ctx;
+  space_sep = isl_space_alloc (ctx, 0, 1, 1);
+  space_sep = isl_space_wrap (space_sep);
+  space_sep = isl_space_set_tuple_name (space_sep, isl_dim_set,
+				        "separation_class");
+  space = isl_set_get_space (scop->context);
+  space_sep = isl_space_align_params (space_sep, isl_space_copy(space));
+  space = isl_space_map_from_domain_and_range (space, space_sep);
+  space = isl_space_add_dims (space,isl_dim_in, nsched);
+  map = isl_map_universe (space);
+  isl_map_fix_si (map,isl_dim_out,0,dim);
+  isl_map_fix_si (map,isl_dim_out,1,cl);
+
+  mapu = isl_union_map_intersect_domain (isl_union_map_from_map (map), 
+					 domain);
+  return (mapu);
+}
+
+/* Compute the separation class for loop unroll and jam.  */
+
+static __isl_give isl_union_set *
+generate_luj_sepclass (scop_p scop)
+{
+  int i;
+  poly_bb_p pbb;
+  isl_union_set *domain_isl;
+
+  domain_isl = isl_union_set_empty (isl_set_get_space (scop->context));
+
+  FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
+    {
+      isl_set *bb_domain;
+      isl_set *bb_domain_s;
+
+      if (pbb->map_sepclass == NULL)
+	continue;
+
+      if (isl_set_is_empty (pbb->domain))
+	continue;
+
+      bb_domain = isl_set_copy (pbb->domain);
+      bb_domain_s = isl_set_apply (bb_domain, pbb->map_sepclass);
+      pbb->map_sepclass = NULL;
+
+      domain_isl =
+	isl_union_set_union (domain_isl, isl_union_set_from_set (bb_domain_s));
+    }
+
+  return domain_isl;
+}
+
+/* Set the AST built options for loop unroll and jam. */
+ 
+static __isl_give isl_union_map *
+generate_luj_options (scop_p scop)
+{
+  isl_union_set *domain_isl;
+  isl_union_map *options_isl_ss;
+  isl_union_map *options_isl =
+    isl_union_map_empty (isl_set_get_space (scop->context));
+  int dim = get_max_schedule_dimensions (scop) - 1;
+  int dim1 = dim - PARAM_VALUE (PARAM_LOOP_UNROLL_JAM_DEPTH);
+
+  if (!flag_loop_unroll_jam)
+    return options_isl;
+
+  domain_isl = generate_luj_sepclass (scop);
+
+  options_isl_ss = generate_luj_sepclass_opt (scop, domain_isl, dim1, 0);
+  options_isl = isl_union_map_union (options_isl, options_isl_ss);
+
+  return options_isl;
+}
+
 /* Generates a schedule, which specifies an order used to
    visit elements in a domain.  */
 
@@ -879,11 +965,13 @@ 
 }
 
 /* Set the separate option for all dimensions.
-   This helps to reduce control overhead.  */
+   This helps to reduce control overhead.
+   Set the options for unroll and jam.  */
 
 static __isl_give isl_ast_build *
 set_options (__isl_take isl_ast_build *control,
-	     __isl_keep isl_union_map *schedule)
+	     __isl_keep isl_union_map *schedule,
+	     __isl_take isl_union_map *opt_luj)
 {
   isl_ctx *ctx = isl_union_map_get_ctx (schedule);
   isl_space *range_space = isl_space_set_alloc (ctx, 0, 1);
@@ -894,6 +982,9 @@ 
   isl_union_set *domain = isl_union_map_range (isl_union_map_copy (schedule));
   domain = isl_union_set_universe (domain);
   isl_union_map *options = isl_union_map_from_domain_and_range (domain, range);
+
+  options = isl_union_map_union (options, opt_luj);
+
   return isl_ast_build_set_options (control, options);
 }
 
@@ -907,9 +998,14 @@ 
   isl_options_set_ast_build_atomic_upper_bound (scop->ctx, true);
 
   add_parameters_to_ivs_params (scop, ip);
+
+  isl_union_map *options_luj = generate_luj_options (scop);
+
   isl_union_map *schedule_isl = generate_isl_schedule (scop);
   isl_ast_build *context_isl = generate_isl_context (scop);
-  context_isl = set_options (context_isl, schedule_isl);
+
+  context_isl = set_options (context_isl, schedule_isl, options_luj);
+
   isl_union_map *dependences = NULL;
   if (flag_loop_parallelize_all)
   {
Index: gcc/toplev.c
===================================================================
--- gcc/toplev.c	(revision 217013)
+++ gcc/toplev.c	(working copy)
@@ -1302,11 +1302,12 @@ 
       || flag_loop_block
       || flag_loop_interchange
       || flag_loop_strip_mine
-      || flag_loop_parallelize_all)
+      || flag_loop_parallelize_all
+      || flag_loop_unroll_jam)
     sorry ("Graphite loop optimizations cannot be used (ISL is not available)" 
 	   "(-fgraphite, -fgraphite-identity, -floop-block, "
 	   "-floop-interchange, -floop-strip-mine, -floop-parallelize-all, "
-	   "and -ftree-loop-linear)");
+	   "-floop-unroll-and-jam, and -ftree-loop-linear)");
 #endif
 
   /* One region RA really helps to decrease the code size.  */
Index: gcc/params.def
===================================================================
--- gcc/params.def	(revision 217013)
+++ gcc/params.def	(working copy)
@@ -847,6 +847,21 @@ 
 	  "size of tiles for loop blocking",
 	  51, 0, 0)
 
+/* Size of unrolling factor for unroll-and-jam.  */
+ 
+DEFPARAM (PARAM_LOOP_UNROLL_JAM_SIZE,
+ 	  "loop-unroll-jam-size",
+ 	  "size of unrolling factor for unroll-and-jam",
+ 	  4, 0, 0)
+
+/* Size of the band formed by the strip mined dimension and the most inner one for unroll-and-jam.  */
+ 
+DEFPARAM (PARAM_LOOP_UNROLL_JAM_DEPTH,
+ 	  "loop-unroll-jam-depth",
+ 	  "depth of unrolled loop for unroll-and-jam",
+ 	  2, 0, 0)
+
+
 /* Maximal number of parameters that we allow in a SCoP.  */
 
 DEFPARAM (PARAM_GRAPHITE_MAX_NB_SCOP_PARAMS,
Index: gcc/graphite-poly.c
===================================================================
--- gcc/graphite-poly.c	(revision 217013)
+++ gcc/graphite-poly.c	(working copy)
@@ -276,7 +276,7 @@ 
 
   /* This pass needs to be run at the final stage, as it does not
      update the lst.  */
-  if (flag_loop_optimize_isl)
+  if (flag_loop_optimize_isl || flag_loop_unroll_jam)
     transform_done |= optimize_isl (scop);
 
   return transform_done;
@@ -327,6 +327,7 @@ 
   pbb->schedule = NULL;
   pbb->transformed = NULL;
   pbb->saved = NULL;
+  pbb->map_sepclass = NULL;
   PBB_SCOP (pbb) = scop;
   pbb_set_black_box (pbb, black_box);
   PBB_TRANSFORMED (pbb) = NULL;