generic-library/webp
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

added C-level optimization for DecodeAlphaData function

Browse Source 
Copies with short distances of 1,2 and 4 are specialized.

up to 10-14% faster alpha decoding.

Change-Id: I9708e98193910bfaf8ef43091f3fdea73b63896d
This commit is contained in:
Djordje Pesut 2014-08-28 16:53:33 +02:00 committed by skal
parent 187d379db6
commit a59562283f
1 changed files with 70 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

71
src/dec/vp8l.c
View File

@ -745,6 +745,39 @@ static void ExtractPalettedAlphaRows(VP8LDecoder* const dec, int row) {
dec->last_row_ = dec->last_out_row_ = row;
}
// cyclic rotation of pattern word
#if defined(WORDS_BIGENDIAN)
#define ROTATE8b(V) do { \
(V) = (((V) & 0xff000000u) >> 24) | ((V) << 8); \
} while (0)
#else
#define ROTATE8b(V) do { \
(V) = (((V) & 0xffu) << 24) | ((V) >> 8); \
} while (0)
#endif
// copy 1, 2 or 4-bytes pattern
#define COPY_SMALL_PATTERN() do { \
int ilength = length; \
uint32_t* pdata; \
int j = 0; \
while ((uintptr_t)pdata1 & 3) { \
*pdata1++ = pdata2[j]; \
ROTATE8b(temp1); \
++j; \
} \
ilength -= j; \
pdata = (uint32_t*)pdata1; \
for (i = 0; i < (ilength >> 2); ++i) { \
pdata[i] = temp1; \
} \
pdata1 = (uint8_t*)pdata; \
pdata2 += j; \
for (i <<= 2; i < ilength; ++i) { \
pdata1[i] = pdata2[i]; \
} \
} while (0)
static int DecodeAlphaData(VP8LDecoder* const dec, uint8_t* const data,
int width, int height, int last_row) {
int ok = 1;
@ -791,8 +824,41 @@ static int DecodeAlphaData(VP8LDecoder* const dec, uint8_t* const data,
dist_code = GetCopyDistance(dist_symbol, br);
dist = PlaneCodeToDistance(width, dist_code);
if (pos >= dist && end - pos >= length) {
uint8_t* pdata1 = data + pos;
const uint8_t* pdata2 = pdata1 - dist;
int i;
for (i = 0; i < length; ++i) data[pos + i] = data[pos + i - dist];
if (length >= 8) {
uint32_t temp1;
switch (dist) {
case 1:
temp1 = pdata1[-1];
#if defined(__arm__) || defined(_M_ARM) // arm doesn't like multiply that much
temp1 |= temp1 << 8;
temp1 |= temp1 << 16;
#else
temp1 = 0x01010101u * temp1;
#endif
break;
case 2:
temp1 = ((uint16_t*)pdata1)[-1];
#if defined(__arm__) || defined(_M_ARM)
temp1 |= temp1 << 16;
#else
temp1 = 0x00010001u * temp1;
#endif
break;
case 4:
temp1 = ((uint32_t*)pdata1)[-1];
break;
default:
goto Copy;
break;
}
COPY_SMALL_PATTERN();
} else {
Copy:
for (i = 0; i < length; ++i) pdata1[i] = pdata2[i];
}
} else {
ok = 0;
goto End;
@ -831,6 +897,9 @@ static int DecodeAlphaData(VP8LDecoder* const dec, uint8_t* const data,
return ok;
}
#undef COPY_PATTERN
#undef ROTATE8b
static int DecodeImageData(VP8LDecoder* const dec, uint32_t* const data,
int width, int height, int last_row,
ProcessRowsFunc process_func) {