[Git][freetype/freetype][master] [truetype] Reduce heap allocation of `deltaSet` variation data.

[Werner Lemberg (@wl)]

Werner Lemberg pushed to branch master at FreeType / FreeType

Commits:

• 115e9275
  by Seigo Nonaka at 2023-05-19T06:50:28+02:00

  [truetype] Reduce heap allocation of `deltaSet` variation data.
  
  `deltaSet` is an array of packed integers that can be 32 bits, 16 bits, or
  8 bits.  Before this change, these values were unpacked to 32-bit integers.
  However, this can cause big heap allocations, e.g., around 500 KByte for
  'NotoSansCJK'.  To reduce this amount, store the packed integers and unpack
  them just before passing to the calculation.  At calculation time, due to
  the variable length of region indices, temporary heap allocations are
  necessary.  This heap allocation is not negligible and visible in `ftbench`
  results.  So, use stack-allocated arrays for short array calculations.
  
  Fixes #1230.
  
  * include/freetype/internal/ftmmtypes.h (GX_ItemVarDataRec): New fields
  `wordDeltaCount` and `longWords`.
  
  * src/truetype/ttgxvar.c (tt_var_load_item_variation_store): Load packed
  data.
  (tt_var_get_item_delta): Unpack data before applying.
  

2 changed files:

• include/freetype/internal/ftmmtypes.h
• src/truetype/ttgxvar.c

Changes:

include/freetype/internal/ftmmtypes.h

---

...	...	@@ -28,13 +28,19 @@ FT_BEGIN_HEADER
28	28
29	29	typedef struct GX_ItemVarDataRec_
30	30	{
31		- FT_UInt itemCount; /* number of delta sets per item */
32		- FT_UInt regionIdxCount; /* number of region indices */
33		- FT_UInt* regionIndices; /* array of `regionCount' indices; */
34		- /* these index `varRegionList' */
35		- FT_ItemVarDelta* deltaSet; /* array of `itemCount' deltas */
36		- /* use `innerIndex' for this array */
37		-
	31	+ FT_UInt itemCount; /* Number of delta sets per item. */
	32	+ FT_UInt regionIdxCount; /* Number of region indices. */
	33	+ FT_UInt* regionIndices; /* Array of `regionCount` indices; */
	34	+ /* these index `varRegionList`. */
	35	+ FT_Byte* deltaSet; /* Array of `itemCount` deltas; */
	36	+ /* use `innerIndex` for this array. */
	37	+ FT_UShort wordDeltaCount; /* Number of the first 32-bit ints */
	38	+ /* or 16-bit ints of `deltaSet` */
	39	+ /* depending on `longWords`. */
	40	+ FT_Bool longWords; /* If true, `deltaSet` is a 32-bit */
	41	+ /* array followed by a 16-bit */
	42	+ /* array, otherwise a 16-bit array */
	43	+ /* followed by an 8-bit array. */
38	44	} GX_ItemVarDataRec, *GX_ItemVarData;
39	45
40	46

src/truetype/ttgxvar.c

---

...	...	@@ -509,7 +509,7 @@
509	509	FT_UShort axis_count;
510	510	FT_UInt region_count;
511	511
512		- FT_UInt i, j, k;
	512	+ FT_UInt i, j;
513	513	FT_Bool long_words;
514	514
515	515	GX_Blend blend = ttface->blend;
...	...	@@ -624,6 +624,7 @@
624	624	FT_UInt item_count;
625	625	FT_UInt word_delta_count;
626	626	FT_UInt region_idx_count;
	627	+ FT_UInt per_region_size;
627	628
628	629
629	630	if ( FT_STREAM_SEEK( offset + dataOffsetArray[i] ) )
...	...	@@ -660,6 +661,8 @@
660	661	if ( FT_NEW_ARRAY( varData->regionIndices, region_idx_count ) )
661	662	goto Exit;
662	663	varData->regionIdxCount = region_idx_count;
	664	+ varData->wordDeltaCount = word_delta_count;
	665	+ varData->longWords = long_words;
663	666
664	667	for ( j = 0; j < varData->regionIdxCount; j++ )
665	668	{
...	...	@@ -675,37 +678,22 @@
675	678	}
676	679	}
677	680
678		- /* Parse delta set. */
679		- /* */
680		- /* On input, deltas are (word_delta_count + region_idx_count) bytes */
681		- /* each if `long_words` isn't set, and twice as much otherwise. */
682		- /* */
683		- /* On output, deltas are expanded to `region_idx_count` shorts each. */
684		- if ( FT_NEW_ARRAY( varData->deltaSet, item_count * region_idx_count ) )
685		- goto Exit;
686		- varData->itemCount = item_count;
	681	+ per_region_size = word_delta_count + region_idx_count;
	682	+ if ( long_words )
	683	+ per_region_size *= 2;
687	684
688		- for ( j = 0; j < item_count * region_idx_count; )
	685	+ if ( FT_NEW_ARRAY( varData->deltaSet, per_region_size * item_count ) )
	686	+ goto Exit;
	687	+ if ( FT_Stream_Read( stream,
	688	+ varData->deltaSet,
	689	+ per_region_size * item_count ) )
689	690	{
690		- if ( long_words )
691		- {
692		- for ( k = 0; k < word_delta_count; k++, j++ )
693		- if ( FT_READ_LONG( varData->deltaSet[j] ) )
694		- goto Exit;
695		- for ( ; k < region_idx_count; k++, j++ )
696		- if ( FT_READ_SHORT( varData->deltaSet[j] ) )
697		- goto Exit;
698		- }
699		- else
700		- {
701		- for ( k = 0; k < word_delta_count; k++, j++ )
702		- if ( FT_READ_SHORT( varData->deltaSet[j] ) )
703		- goto Exit;
704		- for ( ; k < region_idx_count; k++, j++ )
705		- if ( FT_READ_CHAR( varData->deltaSet[j] ) )
706		- goto Exit;
707		- }
	691	+ FT_TRACE2(( "deltaSet read failed." ));
	692	+ error = FT_THROW( Invalid_Table );
	693	+ goto Exit;
708	694	}
	695	+
	696	+ varData->itemCount = item_count;
709	697	}
710	698
711	699	Exit:
...	...	@@ -1005,11 +993,16 @@
1005	993	FT_Error error = FT_Err_Ok;
1006	994
1007	995	GX_ItemVarData varData;
1008		- FT_ItemVarDelta* deltaSet;
	996	+ FT_ItemVarDelta* deltaSet = NULL;
	997	+ FT_ItemVarDelta deltaSetStack[16];
	998	+
	999	+ FT_Fixed* scalars = NULL;
	1000	+ FT_Fixed scalarsStack[16];
1009	1001
1010	1002	FT_UInt master, j;
1011		- FT_Fixed* scalars = NULL;
1012		- FT_ItemVarDelta returnValue;
	1003	+ FT_ItemVarDelta returnValue = 0;
	1004	+ FT_UInt per_region_size;
	1005	+ FT_Byte* bytes;
1013	1006
1014	1007
1015	1008	if ( !ttface->blend || !ttface->blend->normalizedcoords )
...	...	@@ -1026,15 +1019,48 @@
1026	1019	if ( outerIndex >= itemStore->dataCount )
1027	1020	return 0; /* Out of range. */
1028	1021
1029		- varData = &itemStore->varData[outerIndex];
1030		- deltaSet = FT_OFFSET( varData->deltaSet,
1031		- varData->regionIdxCount * innerIndex );
	1022	+ varData = &itemStore->varData[outerIndex];
1032	1023
1033	1024	if ( innerIndex >= varData->itemCount )
1034	1025	return 0; /* Out of range. */
1035	1026
1036		- if ( FT_QNEW_ARRAY( scalars, varData->regionIdxCount ) )
1037		- return 0;
	1027	+ if ( varData->regionIdxCount < 16 )
	1028	+ {
	1029	+ deltaSet = deltaSetStack;
	1030	+ scalars = scalarsStack;
	1031	+ }
	1032	+ else
	1033	+ {
	1034	+ if ( FT_QNEW_ARRAY( deltaSet, varData->regionIdxCount ) )
	1035	+ goto Exit;
	1036	+ if ( FT_QNEW_ARRAY( scalars, varData->regionIdxCount ) )
	1037	+ goto Exit;
	1038	+ }
	1039	+
	1040	+ /* Parse delta set. */
	1041	+ /* */
	1042	+ /* Deltas are (word_delta_count + region_idx_count) bytes each */
	1043	+ /* if `longWords` isn't set, and twice as much otherwise. */
	1044	+ per_region_size = varData->wordDeltaCount + varData->regionIdxCount;
	1045	+ if ( varData->longWords )
	1046	+ per_region_size *= 2;
	1047	+
	1048	+ bytes = varData->deltaSet + per_region_size * innerIndex;
	1049	+
	1050	+ if ( varData->longWords )
	1051	+ {
	1052	+ for ( master = 0; master < varData->wordDeltaCount; master++ )
	1053	+ deltaSet[master] = FT_NEXT_LONG( bytes );
	1054	+ for ( ; master < varData->regionIdxCount; master++ )
	1055	+ deltaSet[master] = FT_NEXT_SHORT( bytes );
	1056	+ }
	1057	+ else
	1058	+ {
	1059	+ for ( master = 0; master < varData->wordDeltaCount; master++ )
	1060	+ deltaSet[master] = FT_NEXT_SHORT( bytes );
	1061	+ for ( ; master < varData->regionIdxCount; master++ )
	1062	+ deltaSet[master] = FT_NEXT_CHAR( bytes );
	1063	+ }
1038	1064
1039	1065	/* outer loop steps through master designs to be blended */
1040	1066	for ( master = 0; master < varData->regionIdxCount; master++ )
...	...	@@ -1109,7 +1135,11 @@
1109	1135	*/
1110	1136	returnValue = FT_MulAddFix( scalars, deltaSet, varData->regionIdxCount );
1111	1137
1112		- FT_FREE( scalars );
	1138	+ Exit:
	1139	+ if ( scalars != scalarsStack )
	1140	+ FT_FREE( scalars );
	1141	+ if ( deltaSet != deltaSetStack )
	1142	+ FT_FREE( deltaSet );
1113	1143
1114	1144	return returnValue;
1115	1145	}

—
View it on GitLab.
You're receiving this email because of your account on gitlab.freedesktop.org. Manage all notifications · Help