mirror of
https://git.ffmpeg.org/ffmpeg.git
synced 2026-10-28 10:04:37 +08:00
avfilter/vf_colordetectdsp: rewrite C implementations as macros
In the process, also add the range assumption and cast to the detect_alpha intermediates, which seems to help codegen: checkasm: - CPU: AMD Ryzen 9 9950X3D 16-Core Processor (00B40F40) - Timing source: x86 (rdtsc) - Bench duration: 100000 µs per function (442374016 cycles) - Random seed: 2809724650 Benchmark results: (old) name cycles (vs ref) detect_alpha_8_full_c: 7582.7 detect_alpha_8_full_off_c: 7587.4 detect_alpha_8_limited_c: 13306.7 detect_alpha_16_full_c: 3482.5 detect_alpha_16_full_off_c: 3480.4 detect_alpha_16_limited_c: 8055.1 detect_range_8_c: 1271.4 detect_range_16_c: 1213.0 Benchmark results: (new) name cycles (vs ref) detect_alpha_8_full_c: 4716.6 detect_alpha_8_full_off_c: 4715.4 detect_alpha_8_limited_c: 7943.2 detect_alpha_16_full_c: 2920.2 detect_alpha_16_full_off_c: 2918.9 detect_alpha_16_limited_c: 9952.0 detect_range_8_c: 1253.1 detect_range_16_c: 1194.8 The ~20% slowdown for detect_alpha_16_limited_c is because the previous version of the code short-circuited this condition, which is faster when the compiler is not able to vectorize the loop body and falls back to a pure scalar loop. However, I don't think optimizing around this is necessary as it's a pure compiler optimization question. In particular, it actually _prevents_ better compilers from being able to vectorize this in the future. (As an example, doing the same short-circuiting for 8-bit would be a 2x regression on my end) People running into performance issues should just use the SIMD versions. Sponsored-by: nxtedition AB Signed-off-by: Niklas Haas <git@haasn.dev>
This commit is contained in:
Niklas Haas authored and Niklas Haas committed
1 parent
912208af28
commit
17e6ed9dc0
1 file changed
+85
-133
+85
-133
@@ -54,146 +54,98 @@ void ff_color_detect_dsp_init_aarch64(FFColorDetectDSPContext *dsp, int depth,
|
||||
void ff_color_detect_dsp_init_x86(FFColorDetectDSPContext *dsp, int depth,
|
||||
int offset, enum AVColorRange color_range);
|
||||
|
||||
static inline int ff_detect_range_impl_c(const uint8_t *data, ptrdiff_t stride,
|
||||
ptrdiff_t width, ptrdiff_t height,
|
||||
uint8_t mpeg_min, uint8_t mpeg_max)
|
||||
{
|
||||
while (height--) {
|
||||
uint8_t cond = 0;
|
||||
for (int x = 0; x < width; x++) {
|
||||
const uint8_t val = data[x];
|
||||
cond |= val < mpeg_min || val > mpeg_max;
|
||||
}
|
||||
if (cond)
|
||||
return 1;
|
||||
data += stride;
|
||||
}
|
||||
|
||||
return 0;
|
||||
#define DECL_DETECT_RANGE_IMPL(TYPE, NAME) \
|
||||
static inline int NAME(const uint8_t* data, ptrdiff_t stride, \
|
||||
ptrdiff_t width, ptrdiff_t height, \
|
||||
int mpeg_min, int mpeg_max) \
|
||||
{ \
|
||||
const TYPE min = mpeg_min; \
|
||||
const TYPE max = mpeg_max; \
|
||||
av_assume(min == mpeg_min); \
|
||||
av_assume(max == mpeg_max); \
|
||||
\
|
||||
while (height--) { \
|
||||
const TYPE *row = (const TYPE *) data; \
|
||||
uint8_t cond = 0; \
|
||||
for (int x = 0; x < width; x++) { \
|
||||
const TYPE val = row[x]; \
|
||||
cond |= val < min || val > max; \
|
||||
} \
|
||||
if (cond) \
|
||||
return 1; \
|
||||
data += stride; \
|
||||
} \
|
||||
\
|
||||
return 0; \
|
||||
}
|
||||
|
||||
static inline int ff_detect_range_c(const uint8_t *data, ptrdiff_t stride,
|
||||
ptrdiff_t width, ptrdiff_t height,
|
||||
int mpeg_min, int mpeg_max)
|
||||
{
|
||||
av_assume(mpeg_min >= 0 && mpeg_min <= UINT8_MAX);
|
||||
av_assume(mpeg_max >= 0 && mpeg_max <= UINT8_MAX);
|
||||
return ff_detect_range_impl_c(data, stride, width, height, mpeg_min, mpeg_max);
|
||||
DECL_DETECT_RANGE_IMPL(uint8_t, ff_detect_range_c)
|
||||
DECL_DETECT_RANGE_IMPL(uint16_t, ff_detect_range16_c)
|
||||
|
||||
#define DECL_DETECT_ALPHA_FULL_IMPL(TYPE, INTER, NAME) \
|
||||
static inline int NAME(const uint8_t* color, ptrdiff_t color_stride, \
|
||||
const uint8_t* alpha, ptrdiff_t alpha_stride, \
|
||||
ptrdiff_t width, ptrdiff_t height, int alpha_max, \
|
||||
int mpeg_range, int offset) \
|
||||
{ \
|
||||
const TYPE max = alpha_max; \
|
||||
const INTER off = offset; \
|
||||
av_assume(max == alpha_max); \
|
||||
av_assume(off == offset); \
|
||||
(void) mpeg_range; \
|
||||
\
|
||||
uint8_t transparent = 0; \
|
||||
while (height--) { \
|
||||
const TYPE *col = (const TYPE *) color; \
|
||||
const TYPE *alp = (const TYPE *) alpha; \
|
||||
uint8_t straight = 0; \
|
||||
for (int x = 0; x < width; x++) { \
|
||||
straight |= col[x] > alp[x] + off; \
|
||||
transparent |= alp[x] != max; \
|
||||
} \
|
||||
if (straight) \
|
||||
return FF_ALPHA_STRAIGHT; \
|
||||
color += color_stride; \
|
||||
alpha += alpha_stride; \
|
||||
} \
|
||||
return transparent ? FF_ALPHA_TRANSPARENT : 0; \
|
||||
}
|
||||
|
||||
static inline int ff_detect_range16_impl_c(const uint8_t *data, ptrdiff_t stride,
|
||||
ptrdiff_t width, ptrdiff_t height,
|
||||
uint16_t mpeg_min, uint16_t mpeg_max)
|
||||
{
|
||||
while (height--) {
|
||||
const uint16_t *data16 = (const uint16_t *) data;
|
||||
uint8_t cond = 0;
|
||||
for (int x = 0; x < width; x++) {
|
||||
const uint16_t val = data16[x];
|
||||
cond |= val < mpeg_min || val > mpeg_max;
|
||||
}
|
||||
if (cond)
|
||||
return 1;
|
||||
data += stride;
|
||||
}
|
||||
DECL_DETECT_ALPHA_FULL_IMPL(uint8_t, uint16_t, ff_detect_alpha_full_c)
|
||||
DECL_DETECT_ALPHA_FULL_IMPL(uint16_t, uint32_t, ff_detect_alpha16_full_c)
|
||||
|
||||
return 0;
|
||||
#define DECL_DETECT_ALPHA_LIMITED_IMPL(TYPE, INTER, NAME) \
|
||||
static inline int NAME(const uint8_t* color, ptrdiff_t color_stride, \
|
||||
const uint8_t* alpha, ptrdiff_t alpha_stride, \
|
||||
ptrdiff_t width, ptrdiff_t height, int alpha_max, \
|
||||
int mpeg_range, int offset) \
|
||||
{ \
|
||||
const TYPE max = alpha_max; \
|
||||
const INTER off = offset; \
|
||||
const INTER range = mpeg_range; \
|
||||
av_assume(max == alpha_max); \
|
||||
av_assume(off == offset); \
|
||||
av_assume(range == mpeg_range); \
|
||||
\
|
||||
uint8_t transparent = 0; \
|
||||
while (height--) { \
|
||||
const TYPE *col = (const TYPE *) color; \
|
||||
const TYPE *alp = (const TYPE *) alpha; \
|
||||
uint8_t straight = 0; \
|
||||
for (int x = 0; x < width; x++) { \
|
||||
straight |= (INTER) max * col[x] - off > range * alp[x]; \
|
||||
transparent |= alp[x] != max; \
|
||||
} \
|
||||
if (straight) \
|
||||
return FF_ALPHA_STRAIGHT; \
|
||||
color += color_stride; \
|
||||
alpha += alpha_stride; \
|
||||
} \
|
||||
return transparent ? FF_ALPHA_TRANSPARENT : 0; \
|
||||
}
|
||||
|
||||
static inline int ff_detect_range16_c(const uint8_t *data, ptrdiff_t stride,
|
||||
ptrdiff_t width, ptrdiff_t height,
|
||||
int mpeg_min, int mpeg_max)
|
||||
{
|
||||
av_assume(mpeg_min >= 0 && mpeg_min <= UINT16_MAX);
|
||||
av_assume(mpeg_max >= 0 && mpeg_max <= UINT16_MAX);
|
||||
return ff_detect_range16_impl_c(data, stride, width, height, mpeg_min, mpeg_max);
|
||||
}
|
||||
|
||||
static inline int
|
||||
ff_detect_alpha_full_c(const uint8_t *color, ptrdiff_t color_stride,
|
||||
const uint8_t *alpha, ptrdiff_t alpha_stride,
|
||||
ptrdiff_t width, ptrdiff_t height,
|
||||
int alpha_max, int mpeg_range, int offset)
|
||||
{
|
||||
uint8_t transparent = 0;
|
||||
while (height--) {
|
||||
uint8_t straight = 0;
|
||||
for (int x = 0; x < width; x++) {
|
||||
straight |= color[x] > alpha[x] + offset;
|
||||
transparent |= alpha[x] != alpha_max;
|
||||
}
|
||||
if (straight)
|
||||
return FF_ALPHA_STRAIGHT;
|
||||
color += color_stride;
|
||||
alpha += alpha_stride;
|
||||
}
|
||||
return transparent ? FF_ALPHA_TRANSPARENT : 0;
|
||||
}
|
||||
|
||||
static inline int
|
||||
ff_detect_alpha_limited_c(const uint8_t *color, ptrdiff_t color_stride,
|
||||
const uint8_t *alpha, ptrdiff_t alpha_stride,
|
||||
ptrdiff_t width, ptrdiff_t height,
|
||||
int alpha_max, int mpeg_range, int offset)
|
||||
{
|
||||
uint8_t transparent = 0;
|
||||
while (height--) {
|
||||
uint8_t straight = 0;
|
||||
for (int x = 0; x < width; x++) {
|
||||
straight |= alpha_max * color[x] - offset > mpeg_range * alpha[x];
|
||||
transparent |= alpha[x] != alpha_max;
|
||||
}
|
||||
if (straight)
|
||||
return FF_ALPHA_STRAIGHT;
|
||||
color += color_stride;
|
||||
alpha += alpha_stride;
|
||||
}
|
||||
return transparent ? FF_ALPHA_TRANSPARENT : 0;
|
||||
}
|
||||
|
||||
static inline int
|
||||
ff_detect_alpha16_full_c(const uint8_t *color, ptrdiff_t color_stride,
|
||||
const uint8_t *alpha, ptrdiff_t alpha_stride,
|
||||
ptrdiff_t width, ptrdiff_t height,
|
||||
int alpha_max, int mpeg_range, int offset)
|
||||
{
|
||||
uint8_t transparent = 0;
|
||||
while (height--) {
|
||||
const uint16_t *color16 = (const uint16_t *) color;
|
||||
const uint16_t *alpha16 = (const uint16_t *) alpha;
|
||||
uint8_t straight = 0;
|
||||
for (int x = 0; x < width; x++) {
|
||||
straight |= color16[x] > alpha16[x] + offset;
|
||||
transparent |= alpha16[x] != alpha_max;
|
||||
}
|
||||
if (straight)
|
||||
return FF_ALPHA_STRAIGHT;
|
||||
color += color_stride;
|
||||
alpha += alpha_stride;
|
||||
}
|
||||
return transparent ? FF_ALPHA_TRANSPARENT : 0;
|
||||
}
|
||||
|
||||
static inline int
|
||||
ff_detect_alpha16_limited_c(const uint8_t *color, ptrdiff_t color_stride,
|
||||
const uint8_t *alpha, ptrdiff_t alpha_stride,
|
||||
ptrdiff_t width, ptrdiff_t height,
|
||||
int alpha_max, int mpeg_range, int offset)
|
||||
{
|
||||
uint8_t transparent = 0;
|
||||
while (height--) {
|
||||
const uint16_t *color16 = (const uint16_t *) color;
|
||||
const uint16_t *alpha16 = (const uint16_t *) alpha;
|
||||
for (int x = 0; x < width; x++) {
|
||||
if ((int64_t) alpha_max * color16[x] - offset > (int64_t) mpeg_range * alpha16[x])
|
||||
return FF_ALPHA_STRAIGHT;
|
||||
transparent |= alpha16[x] != alpha_max;
|
||||
}
|
||||
color += color_stride;
|
||||
alpha += alpha_stride;
|
||||
}
|
||||
return transparent ? FF_ALPHA_TRANSPARENT : 0;
|
||||
}
|
||||
DECL_DETECT_ALPHA_LIMITED_IMPL(uint8_t, int32_t, ff_detect_alpha_limited_c)
|
||||
DECL_DETECT_ALPHA_LIMITED_IMPL(uint16_t, int64_t, ff_detect_alpha16_limited_c)
|
||||
|
||||
static av_cold inline void
|
||||
ff_color_detect_dsp_init(FFColorDetectDSPContext *dsp, int depth, int offset,
|
||||
|
||||
Reference in new issue
Block a user