[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
[freetype2] master 115e92754: [truetype] Reduce heap allocation of `delt
From: |
Werner Lemberg |
Subject: |
[freetype2] master 115e92754: [truetype] Reduce heap allocation of `deltaSet` variation data. |
Date: |
Fri, 19 May 2023 00:52:44 -0400 (EDT) |
branch: master
commit 115e927540dba128980dd734dadeb06aa7b0f4d8
Author: Seigo Nonaka <nona@google.com>
Commit: Werner Lemberg <wl@gnu.org>
[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.
---
include/freetype/internal/ftmmtypes.h | 20 ++++---
src/truetype/ttgxvar.c | 106 ++++++++++++++++++++++------------
2 files changed, 81 insertions(+), 45 deletions(-)
diff --git a/include/freetype/internal/ftmmtypes.h
b/include/freetype/internal/ftmmtypes.h
index b7c66c35d..c4b21d614 100644
--- a/include/freetype/internal/ftmmtypes.h
+++ b/include/freetype/internal/ftmmtypes.h
@@ -28,13 +28,19 @@ FT_BEGIN_HEADER
typedef struct GX_ItemVarDataRec_
{
- FT_UInt itemCount; /* number of delta sets per item */
- FT_UInt regionIdxCount; /* number of region indices */
- FT_UInt* regionIndices; /* array of `regionCount' indices; */
- /* these index `varRegionList' */
- FT_ItemVarDelta* deltaSet; /* array of `itemCount' deltas */
- /* use `innerIndex' for this array */
-
+ FT_UInt itemCount; /* Number of delta sets per item. */
+ FT_UInt regionIdxCount; /* Number of region indices. */
+ FT_UInt* regionIndices; /* Array of `regionCount` indices; */
+ /* these index `varRegionList`. */
+ FT_Byte* deltaSet; /* Array of `itemCount` deltas; */
+ /* use `innerIndex` for this array. */
+ FT_UShort wordDeltaCount; /* Number of the first 32-bit ints */
+ /* or 16-bit ints of `deltaSet` */
+ /* depending on `longWords`. */
+ FT_Bool longWords; /* If true, `deltaSet` is a 32-bit */
+ /* array followed by a 16-bit */
+ /* array, otherwise a 16-bit array */
+ /* followed by an 8-bit array. */
} GX_ItemVarDataRec, *GX_ItemVarData;
diff --git a/src/truetype/ttgxvar.c b/src/truetype/ttgxvar.c
index 04caff7b0..8c713f1b6 100644
--- a/src/truetype/ttgxvar.c
+++ b/src/truetype/ttgxvar.c
@@ -509,7 +509,7 @@
FT_UShort axis_count;
FT_UInt region_count;
- FT_UInt i, j, k;
+ FT_UInt i, j;
FT_Bool long_words;
GX_Blend blend = ttface->blend;
@@ -624,6 +624,7 @@
FT_UInt item_count;
FT_UInt word_delta_count;
FT_UInt region_idx_count;
+ FT_UInt per_region_size;
if ( FT_STREAM_SEEK( offset + dataOffsetArray[i] ) )
@@ -660,6 +661,8 @@
if ( FT_NEW_ARRAY( varData->regionIndices, region_idx_count ) )
goto Exit;
varData->regionIdxCount = region_idx_count;
+ varData->wordDeltaCount = word_delta_count;
+ varData->longWords = long_words;
for ( j = 0; j < varData->regionIdxCount; j++ )
{
@@ -675,37 +678,22 @@
}
}
- /* Parse delta set. */
- /* */
- /* On input, deltas are (word_delta_count + region_idx_count) bytes */
- /* each if `long_words` isn't set, and twice as much otherwise. */
- /* */
- /* On output, deltas are expanded to `region_idx_count` shorts each. */
- if ( FT_NEW_ARRAY( varData->deltaSet, item_count * region_idx_count ) )
- goto Exit;
- varData->itemCount = item_count;
+ per_region_size = word_delta_count + region_idx_count;
+ if ( long_words )
+ per_region_size *= 2;
- for ( j = 0; j < item_count * region_idx_count; )
+ if ( FT_NEW_ARRAY( varData->deltaSet, per_region_size * item_count ) )
+ goto Exit;
+ if ( FT_Stream_Read( stream,
+ varData->deltaSet,
+ per_region_size * item_count ) )
{
- if ( long_words )
- {
- for ( k = 0; k < word_delta_count; k++, j++ )
- if ( FT_READ_LONG( varData->deltaSet[j] ) )
- goto Exit;
- for ( ; k < region_idx_count; k++, j++ )
- if ( FT_READ_SHORT( varData->deltaSet[j] ) )
- goto Exit;
- }
- else
- {
- for ( k = 0; k < word_delta_count; k++, j++ )
- if ( FT_READ_SHORT( varData->deltaSet[j] ) )
- goto Exit;
- for ( ; k < region_idx_count; k++, j++ )
- if ( FT_READ_CHAR( varData->deltaSet[j] ) )
- goto Exit;
- }
+ FT_TRACE2(( "deltaSet read failed." ));
+ error = FT_THROW( Invalid_Table );
+ goto Exit;
}
+
+ varData->itemCount = item_count;
}
Exit:
@@ -1005,11 +993,16 @@
FT_Error error = FT_Err_Ok;
GX_ItemVarData varData;
- FT_ItemVarDelta* deltaSet;
+ FT_ItemVarDelta* deltaSet = NULL;
+ FT_ItemVarDelta deltaSetStack[16];
+
+ FT_Fixed* scalars = NULL;
+ FT_Fixed scalarsStack[16];
FT_UInt master, j;
- FT_Fixed* scalars = NULL;
- FT_ItemVarDelta returnValue;
+ FT_ItemVarDelta returnValue = 0;
+ FT_UInt per_region_size;
+ FT_Byte* bytes;
if ( !ttface->blend || !ttface->blend->normalizedcoords )
@@ -1026,15 +1019,48 @@
if ( outerIndex >= itemStore->dataCount )
return 0; /* Out of range. */
- varData = &itemStore->varData[outerIndex];
- deltaSet = FT_OFFSET( varData->deltaSet,
- varData->regionIdxCount * innerIndex );
+ varData = &itemStore->varData[outerIndex];
if ( innerIndex >= varData->itemCount )
return 0; /* Out of range. */
- if ( FT_QNEW_ARRAY( scalars, varData->regionIdxCount ) )
- return 0;
+ if ( varData->regionIdxCount < 16 )
+ {
+ deltaSet = deltaSetStack;
+ scalars = scalarsStack;
+ }
+ else
+ {
+ if ( FT_QNEW_ARRAY( deltaSet, varData->regionIdxCount ) )
+ goto Exit;
+ if ( FT_QNEW_ARRAY( scalars, varData->regionIdxCount ) )
+ goto Exit;
+ }
+
+ /* Parse delta set. */
+ /* */
+ /* Deltas are (word_delta_count + region_idx_count) bytes each */
+ /* if `longWords` isn't set, and twice as much otherwise. */
+ per_region_size = varData->wordDeltaCount + varData->regionIdxCount;
+ if ( varData->longWords )
+ per_region_size *= 2;
+
+ bytes = varData->deltaSet + per_region_size * innerIndex;
+
+ if ( varData->longWords )
+ {
+ for ( master = 0; master < varData->wordDeltaCount; master++ )
+ deltaSet[master] = FT_NEXT_LONG( bytes );
+ for ( ; master < varData->regionIdxCount; master++ )
+ deltaSet[master] = FT_NEXT_SHORT( bytes );
+ }
+ else
+ {
+ for ( master = 0; master < varData->wordDeltaCount; master++ )
+ deltaSet[master] = FT_NEXT_SHORT( bytes );
+ for ( ; master < varData->regionIdxCount; master++ )
+ deltaSet[master] = FT_NEXT_CHAR( bytes );
+ }
/* outer loop steps through master designs to be blended */
for ( master = 0; master < varData->regionIdxCount; master++ )
@@ -1109,7 +1135,11 @@
*/
returnValue = FT_MulAddFix( scalars, deltaSet, varData->regionIdxCount );
- FT_FREE( scalars );
+ Exit:
+ if ( scalars != scalarsStack )
+ FT_FREE( scalars );
+ if ( deltaSet != deltaSetStack )
+ FT_FREE( deltaSet );
return returnValue;
}
[Prev in Thread] |
Current Thread |
[Next in Thread] |
- [freetype2] master 115e92754: [truetype] Reduce heap allocation of `deltaSet` variation data.,
Werner Lemberg <=