max/ffmpeg
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

swscale/x86/yuv2yuvX: Port ff_yuv2yuvX_mmxext to SSE2

Browse Source 
The mmx function performs two registers in parallel;
given the larger register size of SSE2, the same amount
of data can be processed in one register with some speedups.
(Given that this function is used for tail-processing,
not processing more data is important.)

Switching to SSE2 also fixes a bug introduced in
554c2bc708: Since said
commit, only half the dither values were used. This
seems not to matter in practice, as the functions here
use dither only in the following form:
((filtersize-1)*8+dither)>>4. The dither values used
here come from ff_dither_8x8_128 which has the property
that ff_dither_8x8_128[i][j] and ff_dither_8x8_128[i][j+4]
always lead to the same result in the above formula.

Old benchmarks:
yuv2yuvX_8_2_0_512_approximate_c:                     2309.9 ( 1.00x)
yuv2yuvX_8_2_0_512_approximate_mmxext:                 250.2 ( 9.23x)
yuv2yuvX_8_2_0_512_approximate_sse3:                    98.8 (23.39x)
yuv2yuvX_8_2_0_512_approximate_avx2:                    52.9 (43.63x)
yuv2yuvX_8_2_16_512_approximate_c:                    2263.0 ( 1.00x)
yuv2yuvX_8_2_16_512_approximate_mmxext:                245.3 ( 9.22x)
yuv2yuvX_8_2_16_512_approximate_sse3:                  114.3 (19.80x)
yuv2yuvX_8_2_16_512_approximate_avx2:                   85.6 (26.45x)
yuv2yuvX_8_2_32_512_approximate_c:                    2155.8 ( 1.00x)
yuv2yuvX_8_2_32_512_approximate_mmxext:                235.6 ( 9.15x)
yuv2yuvX_8_2_32_512_approximate_sse3:                   93.6 (23.04x)
yuv2yuvX_8_2_32_512_approximate_avx2:                   78.1 (27.60x)
yuv2yuvX_8_2_48_512_approximate_c:                    2084.8 ( 1.00x)
yuv2yuvX_8_2_48_512_approximate_mmxext:                230.2 ( 9.05x)
yuv2yuvX_8_2_48_512_approximate_sse3:                  105.0 (19.85x)
yuv2yuvX_8_2_48_512_approximate_avx2:                   71.9 (29.00x)
yuv2yuvX_8_4_0_512_approximate_c:                     3496.3 ( 1.00x)
yuv2yuvX_8_4_0_512_approximate_mmxext:                 455.0 ( 7.68x)
yuv2yuvX_8_4_0_512_approximate_sse3:                   157.5 (22.20x)
yuv2yuvX_8_4_0_512_approximate_avx2:                    88.4 (39.53x)
yuv2yuvX_8_4_16_512_approximate_c:                    3380.9 ( 1.00x)
yuv2yuvX_8_4_16_512_approximate_mmxext:                440.0 ( 7.68x)
yuv2yuvX_8_4_16_512_approximate_sse3:                  175.0 (19.32x)
yuv2yuvX_8_4_16_512_approximate_avx2:                  134.1 (25.22x)
yuv2yuvX_8_4_32_512_approximate_c:                    3277.6 ( 1.00x)
yuv2yuvX_8_4_32_512_approximate_mmxext:                427.2 ( 7.67x)
yuv2yuvX_8_4_32_512_approximate_sse3:                  149.7 (21.89x)
yuv2yuvX_8_4_32_512_approximate_avx2:                  115.5 (28.37x)
yuv2yuvX_8_4_48_512_approximate_c:                    3167.8 ( 1.00x)
yuv2yuvX_8_4_48_512_approximate_mmxext:                414.9 ( 7.63x)
yuv2yuvX_8_4_48_512_approximate_sse3:                  164.1 (19.31x)
yuv2yuvX_8_4_48_512_approximate_avx2:                  101.2 (31.30x)
yuv2yuvX_8_8_0_512_approximate_c:                     5987.5 ( 1.00x)
yuv2yuvX_8_8_0_512_approximate_mmxext:                 854.1 ( 7.01x)
yuv2yuvX_8_8_0_512_approximate_sse3:                   294.6 (20.32x)
yuv2yuvX_8_8_0_512_approximate_avx2:                   144.1 (41.56x)
yuv2yuvX_8_8_16_512_approximate_c:                    5848.9 ( 1.00x)
yuv2yuvX_8_8_16_512_approximate_mmxext:                834.4 ( 7.01x)
yuv2yuvX_8_8_16_512_approximate_sse3:                  312.1 (18.74x)
yuv2yuvX_8_8_16_512_approximate_avx2:                  214.9 (27.22x)
yuv2yuvX_8_8_32_512_approximate_c:                    5610.1 ( 1.00x)
yuv2yuvX_8_8_32_512_approximate_mmxext:                811.6 ( 6.91x)
yuv2yuvX_8_8_32_512_approximate_sse3:                  277.5 (20.21x)
yuv2yuvX_8_8_32_512_approximate_avx2:                  189.8 (29.55x)
yuv2yuvX_8_8_48_512_approximate_c:                    5415.8 ( 1.00x)
yuv2yuvX_8_8_48_512_approximate_mmxext:                782.3 ( 6.92x)
yuv2yuvX_8_8_48_512_approximate_sse3:                  289.4 (18.72x)
yuv2yuvX_8_8_48_512_approximate_avx2:                  165.3 (32.76x)
yuv2yuvX_8_16_0_512_approximate_c:                   11100.7 ( 1.00x)
yuv2yuvX_8_16_0_512_approximate_mmxext:               1682.1 ( 6.60x)
yuv2yuvX_8_16_0_512_approximate_sse3:                  558.8 (19.86x)
yuv2yuvX_8_16_0_512_approximate_avx2:                  280.1 (39.63x)
yuv2yuvX_8_16_16_512_approximate_c:                  10772.1 ( 1.00x)
yuv2yuvX_8_16_16_512_approximate_mmxext:              1611.0 ( 6.69x)
yuv2yuvX_8_16_16_512_approximate_sse3:                 578.1 (18.63x)
yuv2yuvX_8_16_16_512_approximate_avx2:                 418.8 (25.72x)
yuv2yuvX_8_16_32_512_approximate_c:                  10381.5 ( 1.00x)
yuv2yuvX_8_16_32_512_approximate_mmxext:              1560.4 ( 6.65x)
yuv2yuvX_8_16_32_512_approximate_sse3:                 525.8 (19.74x)
yuv2yuvX_8_16_32_512_approximate_avx2:                 370.7 (28.01x)
yuv2yuvX_8_16_48_512_approximate_c:                  10046.1 ( 1.00x)
yuv2yuvX_8_16_48_512_approximate_mmxext:              1512.4 ( 6.64x)
yuv2yuvX_8_16_48_512_approximate_sse3:                 546.0 (18.40x)
yuv2yuvX_8_16_48_512_approximate_avx2:                 315.0 (31.89x)

New benchmarks:
yuv2yuvX_8_2_0_512_approximate_c:                     2302.5 ( 1.00x)
yuv2yuvX_8_2_0_512_approximate_sse2:                   184.4 (12.49x)
yuv2yuvX_8_2_0_512_approximate_sse3:                   100.1 (23.01x)
yuv2yuvX_8_2_0_512_approximate_avx2:                    54.9 (41.98x)
yuv2yuvX_8_2_16_512_approximate_c:                    2224.6 ( 1.00x)
yuv2yuvX_8_2_16_512_approximate_sse2:                  180.0 (12.36x)
yuv2yuvX_8_2_16_512_approximate_sse3:                  109.5 (20.31x)
yuv2yuvX_8_2_16_512_approximate_avx2:                   81.3 (27.35x)
yuv2yuvX_8_2_32_512_approximate_c:                    2165.3 ( 1.00x)
yuv2yuvX_8_2_32_512_approximate_sse2:                  176.6 (12.26x)
yuv2yuvX_8_2_32_512_approximate_sse3:                   93.7 (23.11x)
yuv2yuvX_8_2_32_512_approximate_avx2:                   73.1 (29.61x)
yuv2yuvX_8_2_48_512_approximate_c:                    2088.0 ( 1.00x)
yuv2yuvX_8_2_48_512_approximate_sse2:                  170.7 (12.23x)
yuv2yuvX_8_2_48_512_approximate_sse3:                  103.4 (20.20x)
yuv2yuvX_8_2_48_512_approximate_avx2:                   69.4 (30.10x)
yuv2yuvX_8_4_0_512_approximate_c:                     3496.8 ( 1.00x)
yuv2yuvX_8_4_0_512_approximate_sse2:                   320.3 (10.92x)
yuv2yuvX_8_4_0_512_approximate_sse3:                   158.8 (22.02x)
yuv2yuvX_8_4_0_512_approximate_avx2:                    86.4 (40.49x)
yuv2yuvX_8_4_16_512_approximate_c:                    3443.5 ( 1.00x)
yuv2yuvX_8_4_16_512_approximate_sse2:                  325.3 (10.59x)
yuv2yuvX_8_4_16_512_approximate_sse3:                  171.9 (20.03x)
yuv2yuvX_8_4_16_512_approximate_avx2:                  123.6 (27.85x)
yuv2yuvX_8_4_32_512_approximate_c:                    3272.2 ( 1.00x)
yuv2yuvX_8_4_32_512_approximate_sse2:                  302.7 (10.81x)
yuv2yuvX_8_4_32_512_approximate_sse3:                  148.9 (21.98x)
yuv2yuvX_8_4_32_512_approximate_avx2:                  110.6 (29.58x)
yuv2yuvX_8_4_48_512_approximate_c:                    3166.3 ( 1.00x)
yuv2yuvX_8_4_48_512_approximate_sse2:                  291.0 (10.88x)
yuv2yuvX_8_4_48_512_approximate_sse3:                  162.9 (19.44x)
yuv2yuvX_8_4_48_512_approximate_avx2:                  102.3 (30.95x)
yuv2yuvX_8_8_0_512_approximate_c:                     5967.6 ( 1.00x)
yuv2yuvX_8_8_0_512_approximate_sse2:                   691.2 ( 8.63x)
yuv2yuvX_8_8_0_512_approximate_sse3:                   294.2 (20.28x)
yuv2yuvX_8_8_0_512_approximate_avx2:                   154.9 (38.52x)
yuv2yuvX_8_8_16_512_approximate_c:                    5780.2 ( 1.00x)
yuv2yuvX_8_8_16_512_approximate_sse2:                  606.2 ( 9.53x)
yuv2yuvX_8_8_16_512_approximate_sse3:                  309.3 (18.69x)
yuv2yuvX_8_8_16_512_approximate_avx2:                  208.7 (27.69x)
yuv2yuvX_8_8_32_512_approximate_c:                    5604.3 ( 1.00x)
yuv2yuvX_8_8_32_512_approximate_sse2:                  592.3 ( 9.46x)
yuv2yuvX_8_8_32_512_approximate_sse3:                  281.1 (19.94x)
yuv2yuvX_8_8_32_512_approximate_avx2:                  185.4 (30.23x)
yuv2yuvX_8_8_48_512_approximate_c:                    5413.7 ( 1.00x)
yuv2yuvX_8_8_48_512_approximate_sse2:                  570.4 ( 9.49x)
yuv2yuvX_8_8_48_512_approximate_sse3:                  294.9 (18.36x)
yuv2yuvX_8_8_48_512_approximate_avx2:                  166.5 (32.51x)
yuv2yuvX_8_16_0_512_approximate_c:                   11099.4 ( 1.00x)
yuv2yuvX_8_16_0_512_approximate_sse2:                 1213.6 ( 9.15x)
yuv2yuvX_8_16_0_512_approximate_sse3:                  563.0 (19.72x)
yuv2yuvX_8_16_0_512_approximate_avx2:                  294.8 (37.65x)
yuv2yuvX_8_16_16_512_approximate_c:                  10718.1 ( 1.00x)
yuv2yuvX_8_16_16_512_approximate_sse2:                1121.2 ( 9.56x)
yuv2yuvX_8_16_16_512_approximate_sse3:                 563.7 (19.01x)
yuv2yuvX_8_16_16_512_approximate_avx2:                 389.5 (27.51x)
yuv2yuvX_8_16_32_512_approximate_c:                  10373.3 ( 1.00x)
yuv2yuvX_8_16_32_512_approximate_sse2:                1096.2 ( 9.46x)
yuv2yuvX_8_16_32_512_approximate_sse3:                 526.7 (19.70x)
yuv2yuvX_8_16_32_512_approximate_avx2:                 354.7 (29.24x)
yuv2yuvX_8_16_48_512_approximate_c:                  10066.9 ( 1.00x)
yuv2yuvX_8_16_48_512_approximate_sse2:                1055.8 ( 9.53x)
yuv2yuvX_8_16_48_512_approximate_sse3:                 527.9 (19.07x)
yuv2yuvX_8_16_48_512_approximate_avx2:                 313.7 (32.09x)

Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
This commit is contained in:
Andreas Rheinhardt
2026-04-22 14:50:22 +02:00
parent 62285be009
commit f5ed254528
3 changed files with 53 additions and 48 deletions
Download Patch File Download Diff File Expand all files Collapse all files
libswscale/x86/swscale.c
+25 -26
View File
@@ -174,19 +174,6 @@ void ff_updateMMXDitherTables(SwsInternal *c, int dstY)
}
}
#define YUV2YUVX_FUNC_MMX(opt, step) \
void ff_yuv2yuvX_ ##opt(const int16_t *filter, int filterSize, int srcOffset, \
uint8_t *dest, int dstW, \
const uint8_t *dither, int offset); \
static void yuv2yuvX_ ##opt(const int16_t *filter, int filterSize, \
const int16_t **src, uint8_t *dest, int dstW, \
const uint8_t *dither, int offset) \
{ \
if(dstW > 0) \
ff_yuv2yuvX_ ##opt(filter, filterSize - 1, 0, dest - offset, dstW + offset, dither, offset); \
return; \
}
#define YUV2YUVX_FUNC(opt, step) \
void ff_yuv2yuvX_ ##opt(const int16_t *filter, int filterSize, int srcOffset, \
uint8_t *dest, int dstW, \
@@ -198,26 +185,36 @@ static void yuv2yuvX_ ##opt(const int16_t *filter, int filterSize, \
int remainder = (dstW % step); \
int pixelsProcessed = dstW - remainder; \
if(((uintptr_t)dest) & 15){ \
yuv2yuvX_mmxext(filter, filterSize, src, dest, dstW, dither, offset); \
yuv2yuvX_sse2(filter, filterSize, src, dest, dstW, dither, offset); \
return; \
} \
if(pixelsProcessed > 0) \
ff_yuv2yuvX_ ##opt(filter, filterSize - 1, 0, dest - offset, pixelsProcessed + offset, dither, offset); \
if(remainder > 0){ \
ff_yuv2yuvX_mmxext(filter, filterSize - 1, pixelsProcessed, dest - offset, pixelsProcessed + remainder + offset, dither, offset); \
ff_yuv2yuvX_sse2(filter, filterSize - 1, pixelsProcessed, dest - offset, pixelsProcessed + remainder + offset, dither, offset); \
} \
return; \
}
#if HAVE_MMXEXT_EXTERNAL
YUV2YUVX_FUNC_MMX(mmxext, 16)
#endif
#if HAVE_SSE2_EXTERNAL
void ff_yuv2yuvX_sse2(const int16_t *filter, int filterSize, int srcOffset,
uint8_t *dest, int dstW,
const uint8_t *dither, int offset);
static void yuv2yuvX_sse2(const int16_t *filter, int filterSize,
const int16_t **src, uint8_t *dest, int dstW,
const uint8_t *dither, int offset)
{
if (dstW > 0)
ff_yuv2yuvX_sse2(filter, filterSize - 1, 0, dest - offset, dstW + offset, dither, offset);
return;
}
#if HAVE_SSE3_EXTERNAL
YUV2YUVX_FUNC(sse3, 32)
#endif
#if HAVE_AVX2_EXTERNAL
YUV2YUVX_FUNC(avx2, 64)
#endif
#endif
#define SCALE_FUNC(filter_n, from_bpc, to_bpc, opt) \
void ff_hscale ## from_bpc ## to ## to_bpc ## _ ## filter_n ## _ ## opt( \
@@ -513,18 +510,20 @@ av_cold void ff_sws_init_swscale_x86(SwsInternal *c)
}
#endif
} else {
#if HAVE_MMXEXT_EXTERNAL
c->use_mmx_vfilter = 1;
c->yuv2planeX = yuv2yuvX_mmxext;
#if HAVE_SSE2_EXTERNAL
if (EXTERNAL_SSE2(cpu_flags)) {
c->use_mmx_vfilter = 1;
c->yuv2planeX = yuv2yuvX_sse2;
#if HAVE_SSE3_EXTERNAL
if (EXTERNAL_SSE3(cpu_flags))
c->yuv2planeX = yuv2yuvX_sse3;
if (EXTERNAL_SSE3(cpu_flags))
c->yuv2planeX = yuv2yuvX_sse3;
#endif
#if HAVE_AVX2_EXTERNAL
if (EXTERNAL_AVX2_FAST(cpu_flags))
c->yuv2planeX = yuv2yuvX_avx2;
if (EXTERNAL_AVX2_FAST(cpu_flags))
c->yuv2planeX = yuv2yuvX_avx2;
#endif
#endif /* HAVE_MMXEXT_EXTERNAL */
}
#endif /* HAVE_SSE2_EXTERNAL */
#if HAVE_MMXEXT_INLINE
if (!(c->opts.flags & SWS_FULL_CHR_H_INT)) {
switch (c->opts.dst_format) {
libswscale/x86/yuv2yuvX.asm
+25 -19
View File
@@ -34,13 +34,13 @@ SECTION .text
;-----------------------------------------------------------------------------
%macro YUV2YUVX_FUNC 0
cglobal yuv2yuvX, 7, 7, 8, filter, filterSize, src, dest, dstW, dither, offset
cglobal yuv2yuvX, 7, 7, 6+2*cpuflag(sse3), filter, filterSize, src, dest, dstW, dither, offset
%if notcpuflag(sse3)
%define movr mova
%define movr movq
%define unroll 1
%else
%define movr movdqu
%define unroll 2
%define unroll 4
%endif
movsxdifnidn dstWq, dstWd
movsxdifnidn offsetq, offsetd
@@ -67,14 +67,14 @@ cglobal yuv2yuvX, 7, 7, 8, filter, filterSize, src, dest, dstW, dither, offset
mov srcq, [filterSizeq]
punpcklbw m3, m0
psllw m1, m1, 3
paddw m3, m3, m1
psraw m7, m3, 4
paddw m1, m3
psraw m1, 4
.outerloop:
mova m4, m7
mova m3, m7
mova m3, m1
%if cpuflag(sse3)
mova m6, m7
mova m1, m7
mova m4, m1
mova m6, m1
mova m7, m1
%endif
.loop:
%if cpuflag(avx2)
@@ -82,31 +82,37 @@ cglobal yuv2yuvX, 7, 7, 8, filter, filterSize, src, dest, dstW, dither, offset
%elif cpuflag(sse3)
movddup m0, [filterSizeq + 8]
%else
mova m0, [filterSizeq + 8]
movq m0, [filterSizeq + 8]
punpcklqdq m0, m0
%endif
%if cpuflag(sse3)
pmulhw m2, m0, [srcq + offsetq * 2]
pmulhw m5, m0, [srcq + offsetq * 2 + mmsize]
paddw m3, m3, m2
paddw m4, m4, m5
%if cpuflag(sse3)
pmulhw m2, m0, [srcq + offsetq * 2 + 2 * mmsize]
pmulhw m5, m0, [srcq + offsetq * 2 + 3 * mmsize]
paddw m6, m6, m2
paddw m1, m1, m5
paddw m7, m7, m5
%else
movu m2, [srcq + offsetq * 2]
pmulhw m2, m0
paddw m3, m2
%endif
add filterSizeq, 0x10
mov srcq, [filterSizeq]
test srcq, srcq
jnz .loop
psraw m3, m3, 3
%if cpuflag(sse3)
psraw m4, m4, 3
%if cpuflag(sse3)
psraw m6, m6, 3
psraw m1, m1, 3
%endif
psraw m7, m7, 3
packuswb m3, m3, m4
%if cpuflag(sse3)
packuswb m6, m6, m1
packuswb m6, m6, m7
%else
packuswb m3, m3
%endif
mov srcq, [filterq]
%if cpuflag(avx2)
@@ -117,14 +123,14 @@ cglobal yuv2yuvX, 7, 7, 8, filter, filterSize, src, dest, dstW, dither, offset
%if cpuflag(sse3)
movr [destq + offsetq + mmsize], m6
%endif
add offsetq, mmsize * unroll
add offsetq, mmsize / 2 * unroll
mov filterSizeq, filterq
cmp offsetq, dstWq
jb .outerloop
RET
%endmacro
INIT_MMX mmxext
INIT_XMM sse2
YUV2YUVX_FUNC
INIT_XMM sse3
YUV2YUVX_FUNC
tests/checkasm/sw_scale.c
+3 -3
View File
@@ -181,9 +181,9 @@ static void check_yuv2yuvX(int accurate, int bit_depth, int dst_pix_format)
static const int input_sizes[] = {8, 24, 128, 144, 256, 512};
const char *accurate_str = (accurate) ? "accurate" : "approximate";
declare_func_emms(AV_CPU_FLAG_MMX, void, const int16_t *filter,
int filterSize, const int16_t **src, uint8_t *dest,
int dstW, const uint8_t *dither, int offset);
declare_func(void, const int16_t *filter,
int filterSize, const int16_t **src, uint8_t *dest,
int dstW, const uint8_t *dither, int offset);
const int16_t **src;
LOCAL_ALIGNED_16(int16_t, src_pixels, [LARGEST_FILTER * LARGEST_INPUT_SIZE]);