yadif: restore speed of the C filtering code

Always use the special filter for the first and last 3 columns (only).

Changes made in 64ed397 slowed the filter to just under 3/4 of what it
was.  This commit restores the speed while maintaining identical output.

For reference, on my Athlon64:
1733222 decicycles in old
2358563 decicycles in new
1727558 decicycles in this

Signed-off-by: Anton Khirnov <anton@khirnov.net>
This commit is contained in:
James Darnley 2013-03-10 15:08:50 +01:00 committed by Anton Khirnov
parent 252c0bfdc0
commit b0ef0ae776
3 changed files with 37 additions and 49 deletions

View File

@ -33,14 +33,17 @@
#include <assert.h> #include <assert.h>
#define CHECK(j)\ #define CHECK(j)\
{ int score = FFABS(cur[mrefs + off_left + (j)] - cur[prefs + off_left - (j)])\ { int score = FFABS(cur[mrefs - 1 + (j)] - cur[prefs - 1 - (j)])\
+ FFABS(cur[mrefs +(j)] - cur[prefs -(j)])\ + FFABS(cur[mrefs +(j)] - cur[prefs -(j)])\
+ FFABS(cur[mrefs + off_right + (j)] - cur[prefs + off_right - (j)]);\ + FFABS(cur[mrefs + 1 + (j)] - cur[prefs + 1 - (j)]);\
if (score < spatial_score) {\ if (score < spatial_score) {\
spatial_score= score;\ spatial_score= score;\
spatial_pred= (cur[mrefs +(j)] + cur[prefs -(j)])>>1;\ spatial_pred= (cur[mrefs +(j)] + cur[prefs -(j)])>>1;\
#define FILTER(start, end) \ /* The is_not_edge argument here controls when the code will enter a branch
* which reads up to and including x-3 and x+3. */
#define FILTER(start, end, is_not_edge) \
for (x = start; x < end; x++) { \ for (x = start; x < end; x++) { \
int c = cur[mrefs]; \ int c = cur[mrefs]; \
int d = (prev2[0] + next2[0])>>1; \ int d = (prev2[0] + next2[0])>>1; \
@ -50,12 +53,10 @@
int temporal_diff2 =(FFABS(next[mrefs] - c) + FFABS(next[prefs] - e) )>>1; \ int temporal_diff2 =(FFABS(next[mrefs] - c) + FFABS(next[prefs] - e) )>>1; \
int diff = FFMAX3(temporal_diff0 >> 1, temporal_diff1, temporal_diff2); \ int diff = FFMAX3(temporal_diff0 >> 1, temporal_diff1, temporal_diff2); \
int spatial_pred = (c+e) >> 1; \ int spatial_pred = (c+e) >> 1; \
int off_right = (x < w - 1) ? 1 : -1;\
int off_left = x ? -1 : 1;\
int spatial_score = FFABS(cur[mrefs + off_left] - cur[prefs + off_left]) + FFABS(c-e) \
+ FFABS(cur[mrefs + off_right] - cur[prefs + off_right]) - 1; \
\ \
if (x > 2 && x < w - 3) {\ if (is_not_edge) {\
int spatial_score = FFABS(cur[mrefs - 1] - cur[prefs - 1]) + FFABS(c-e) \
+ FFABS(cur[mrefs + 1] - cur[prefs + 1]) - 1; \
CHECK(-1) CHECK(-2) }} }} \ CHECK(-1) CHECK(-2) }} }} \
CHECK( 1) CHECK( 2) }} }} \ CHECK( 1) CHECK( 2) }} }} \
}\ }\
@ -96,12 +97,15 @@ static void filter_line_c(void *dst1,
uint8_t *prev2 = parity ? prev : cur ; uint8_t *prev2 = parity ? prev : cur ;
uint8_t *next2 = parity ? cur : next; uint8_t *next2 = parity ? cur : next;
FILTER(0, w) /* The function is called with the pointers already pointing to data[3] and
* with 6 subtracted from the width. This allows the FILTER macro to be
* called so that it processes all the pixels normally. A constant value of
* true for is_not_edge lets the compiler ignore the if statement. */
FILTER(0, w, 1)
} }
static void filter_edges(void *dst1, void *prev1, void *cur1, void *next1, static void filter_edges(void *dst1, void *prev1, void *cur1, void *next1,
int w, int prefs, int mrefs, int parity, int mode, int w, int prefs, int mrefs, int parity, int mode)
int l_edge)
{ {
uint8_t *dst = dst1; uint8_t *dst = dst1;
uint8_t *prev = prev1; uint8_t *prev = prev1;
@ -111,7 +115,9 @@ static void filter_edges(void *dst1, void *prev1, void *cur1, void *next1,
uint8_t *prev2 = parity ? prev : cur ; uint8_t *prev2 = parity ? prev : cur ;
uint8_t *next2 = parity ? cur : next; uint8_t *next2 = parity ? cur : next;
FILTER(0, l_edge) /* Only edge pixels need to be processed here. A constant value of false
* for is_not_edge should let the compiler ignore the whole branch. */
FILTER(0, 3, 0)
dst = (uint8_t*)dst1 + w - 3; dst = (uint8_t*)dst1 + w - 3;
prev = (uint8_t*)prev1 + w - 3; prev = (uint8_t*)prev1 + w - 3;
@ -120,7 +126,7 @@ static void filter_edges(void *dst1, void *prev1, void *cur1, void *next1,
prev2 = (uint8_t*)(parity ? prev : cur); prev2 = (uint8_t*)(parity ? prev : cur);
next2 = (uint8_t*)(parity ? cur : next); next2 = (uint8_t*)(parity ? cur : next);
FILTER(w - 3, w) FILTER(w - 3, w, 0)
} }
@ -139,12 +145,11 @@ static void filter_line_c_16bit(void *dst1,
mrefs /= 2; mrefs /= 2;
prefs /= 2; prefs /= 2;
FILTER(0, w) FILTER(0, w, 1)
} }
static void filter_edges_16bit(void *dst1, void *prev1, void *cur1, void *next1, static void filter_edges_16bit(void *dst1, void *prev1, void *cur1, void *next1,
int w, int prefs, int mrefs, int parity, int mode, int w, int prefs, int mrefs, int parity, int mode)
int l_edge)
{ {
uint16_t *dst = dst1; uint16_t *dst = dst1;
uint16_t *prev = prev1; uint16_t *prev = prev1;
@ -154,7 +159,7 @@ static void filter_edges_16bit(void *dst1, void *prev1, void *cur1, void *next1,
uint16_t *prev2 = parity ? prev : cur ; uint16_t *prev2 = parity ? prev : cur ;
uint16_t *next2 = parity ? cur : next; uint16_t *next2 = parity ? cur : next;
FILTER(0, l_edge) FILTER(0, 3, 0)
dst = (uint16_t*)dst1 + w - 3; dst = (uint16_t*)dst1 + w - 3;
prev = (uint16_t*)prev1 + w - 3; prev = (uint16_t*)prev1 + w - 3;
@ -163,7 +168,7 @@ static void filter_edges_16bit(void *dst1, void *prev1, void *cur1, void *next1,
prev2 = (uint16_t*)(parity ? prev : cur); prev2 = (uint16_t*)(parity ? prev : cur);
next2 = (uint16_t*)(parity ? cur : next); next2 = (uint16_t*)(parity ? cur : next);
FILTER(w - 3, w) FILTER(w - 3, w, 0)
} }
static void filter(AVFilterContext *ctx, AVFrame *dstpic, static void filter(AVFilterContext *ctx, AVFrame *dstpic,
@ -177,7 +182,7 @@ static void filter(AVFilterContext *ctx, AVFrame *dstpic,
int h = dstpic->height; int h = dstpic->height;
int refs = yadif->cur->linesize[i]; int refs = yadif->cur->linesize[i];
int df = (yadif->csp->comp[i].depth_minus1 + 8) / 8; int df = (yadif->csp->comp[i].depth_minus1 + 8) / 8;
int l_edge, l_edge_pix; int pix_3 = 3 * df;
if (i == 1 || i == 2) { if (i == 1 || i == 2) {
/* Why is this not part of the per-plane description thing? */ /* Why is this not part of the per-plane description thing? */
@ -188,8 +193,6 @@ static void filter(AVFilterContext *ctx, AVFrame *dstpic,
/* filtering reads 3 pixels to the left/right; to avoid invalid reads, /* filtering reads 3 pixels to the left/right; to avoid invalid reads,
* we need to call the c variant which avoids this for border pixels * we need to call the c variant which avoids this for border pixels
*/ */
l_edge = yadif->req_align;
l_edge_pix = l_edge / df;
for (y = 0; y < h; y++) { for (y = 0; y < h; y++) {
if ((y ^ parity) & 1) { if ((y ^ parity) & 1) {
@ -198,22 +201,15 @@ static void filter(AVFilterContext *ctx, AVFrame *dstpic,
uint8_t *next = &yadif->next->data[i][y * refs]; uint8_t *next = &yadif->next->data[i][y * refs];
uint8_t *dst = &dstpic->data[i][y * dstpic->linesize[i]]; uint8_t *dst = &dstpic->data[i][y * dstpic->linesize[i]];
int mode = y == 1 || y + 2 == h ? 2 : yadif->mode; int mode = y == 1 || y + 2 == h ? 2 : yadif->mode;
if (yadif->req_align) { yadif->filter_line(dst + pix_3, prev + pix_3, cur + pix_3,
yadif->filter_line(dst + l_edge, prev + l_edge, cur + l_edge, next + pix_3, w - 6,
next + l_edge, w - l_edge_pix - 3, y + 1 < h ? refs : -refs,
y + 1 < h ? refs : -refs, y ? -refs : refs,
y ? -refs : refs, parity ^ tff, mode);
parity ^ tff, mode); yadif->filter_edges(dst, prev, cur, next, w,
yadif->filter_edges(dst, prev, cur, next, w, y + 1 < h ? refs : -refs,
y + 1 < h ? refs : -refs, y ? -refs : refs,
y ? -refs : refs, parity ^ tff, mode);
parity ^ tff, mode, l_edge_pix);
} else {
yadif->filter_line(dst, prev, cur, next + l_edge, w,
y + 1 < h ? refs : -refs,
y ? -refs : refs,
parity ^ tff, mode);
}
} else { } else {
memcpy(&dstpic->data[i][y * dstpic->linesize[i]], memcpy(&dstpic->data[i][y * dstpic->linesize[i]],
&yadif->cur->data[i][y * refs], w * df); &yadif->cur->data[i][y * refs], w * df);

View File

@ -42,18 +42,12 @@ av_cold void ff_yadif_init_x86(YADIFContext *yadif)
int cpu_flags = av_get_cpu_flags(); int cpu_flags = av_get_cpu_flags();
#if ARCH_X86_32 #if ARCH_X86_32
if (EXTERNAL_MMXEXT(cpu_flags)) { if (EXTERNAL_MMXEXT(cpu_flags))
yadif->filter_line = ff_yadif_filter_line_mmxext; yadif->filter_line = ff_yadif_filter_line_mmxext;
yadif->req_align = 8;
}
#endif /* ARCH_X86_32 */ #endif /* ARCH_X86_32 */
if (EXTERNAL_SSE2(cpu_flags)) { if (EXTERNAL_SSE2(cpu_flags))
yadif->filter_line = ff_yadif_filter_line_sse2; yadif->filter_line = ff_yadif_filter_line_sse2;
yadif->req_align = 16; if (EXTERNAL_SSSE3(cpu_flags))
}
if (EXTERNAL_SSSE3(cpu_flags)) {
yadif->filter_line = ff_yadif_filter_line_ssse3; yadif->filter_line = ff_yadif_filter_line_ssse3;
yadif->req_align = 16;
}
#endif /* HAVE_YASM */ #endif /* HAVE_YASM */
} }

View File

@ -55,13 +55,11 @@ typedef struct YADIFContext {
/** /**
* Required alignment for filter_line * Required alignment for filter_line
*/ */
int req_align;
void (*filter_line)(void *dst, void (*filter_line)(void *dst,
void *prev, void *cur, void *next, void *prev, void *cur, void *next,
int w, int prefs, int mrefs, int parity, int mode); int w, int prefs, int mrefs, int parity, int mode);
void (*filter_edges)(void *dst, void *prev, void *cur, void *next, void (*filter_edges)(void *dst, void *prev, void *cur, void *next,
int w, int prefs, int mrefs, int parity, int mode, int w, int prefs, int mrefs, int parity, int mode);
int l_edge);
const AVPixFmtDescriptor *csp; const AVPixFmtDescriptor *csp;
int eof; int eof;