[FFmpeg-devel] [PATCH v3] libswscale/ppc: VSX-optimize 9-16 bit yuv2planeX
Carl Eugen Hoyos
ceffmpeg at gmail.com
Wed Jan 9 23:26:25 EET 2019
2019-01-08 10:11 GMT+01:00, Lauri Kasanen <cand at gmx.com>:
> ./ffmpeg_g -f rawvideo -pix_fmt rgb24 -s hd1080 -i /dev/zero -pix_fmt
> yuv420p16be \
> -s 1920x1728 -f null -vframes 100 -v error -nostats -
>
> 9-14 bit funcs get about 6x speedup, 16-bit gets about 15x.
> Fate passes, each format tested with an image to video conversion.
>
> Only POWER8 includes 32-bit vector multiplies, so POWER7 is locked out
> of the 16-bit function. This includes the vec_mulo/mule functions too,
> not just vmuluwm.
>
> yuv420p9le
> 12341 UNITS in planarX, 130976 runs, 96 skips
> 73752 UNITS in planarX, 131066 runs, 6 skips
> yuv420p9be
> 12364 UNITS in planarX, 131025 runs, 47 skips
> 73001 UNITS in planarX, 131055 runs, 17 skips
> yuv420p10le
> 12386 UNITS in planarX, 131042 runs, 30 skips
> 72735 UNITS in planarX, 131062 runs, 10 skips
> yuv420p10be
> 12337 UNITS in planarX, 131045 runs, 27 skips
> 72734 UNITS in planarX, 131057 runs, 15 skips
> yuv420p12le
> 12236 UNITS in planarX, 131058 runs, 14 skips
> 73029 UNITS in planarX, 131062 runs, 10 skips
> yuv420p12be
> 12218 UNITS in planarX, 130973 runs, 99 skips
> 72402 UNITS in planarX, 131069 runs, 3 skips
> yuv420p14le
> 12168 UNITS in planarX, 131067 runs, 5 skips
> 72480 UNITS in planarX, 131069 runs, 3 skips
> yuv420p14be
> 12358 UNITS in planarX, 130948 runs, 124 skips
> 73772 UNITS in planarX, 131063 runs, 9 skips
> yuv420p16le
> 10439 UNITS in planarX, 130911 runs, 161 skips
> 157923 UNITS in planarX, 131068 runs, 4 skips
> yuv420p16be
> 10463 UNITS in planarX, 130874 runs, 198 skips
> 154405 UNITS in planarX, 131061 runs, 11 skips
>
> Signed-off-by: Lauri Kasanen <cand at gmx.com>
> ---
>
> v2: Separate macros so that yuv2plane1_16_vsx remains available for power7
> v3: Remove accidental tabs, switch to HAVE_POWER8 from configure + runtime
> check
>
> As far as I can tell, for HAVE_POWER8 to be defined, -march has to be at
> least
> power8, meaning with the current setup such a binary wouldn't run on POWER7.
> However using the configure define lets it be disabled in configure like
> Michael
> pointed out, and having the runtime check doesn't hurt any (it allows for
> future
> splits like on x86, where one binary can run on low cpu but use higher ISA
> if
> available).
>
> libswscale/ppc/swscale_ppc_template.c | 4 +-
> libswscale/ppc/swscale_vsx.c | 195
> +++++++++++++++++++++++++++++++++-
> 2 files changed, 193 insertions(+), 6 deletions(-)
>
> diff --git a/libswscale/ppc/swscale_ppc_template.c
> b/libswscale/ppc/swscale_ppc_template.c
> index 00e4b99..11decab 100644
> --- a/libswscale/ppc/swscale_ppc_template.c
> +++ b/libswscale/ppc/swscale_ppc_template.c
> @@ -21,7 +21,7 @@
> * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301
> USA
> */
>
> -static void FUNC(yuv2planeX_16)(const int16_t *filter, int filterSize,
> +static void FUNC(yuv2planeX_8_16)(const int16_t *filter, int filterSize,
> const int16_t **src, uint8_t *dest,
> const uint8_t *dither, int offset, int x)
> {
> @@ -88,7 +88,7 @@ static void FUNC(yuv2planeX)(const int16_t *filter, int
> filterSize,
> yuv2planeX_u(filter, filterSize, src, dest, dst_u, dither, offset, 0);
>
> for (i = dst_u; i < dstW - 15; i += 16)
> - FUNC(yuv2planeX_16)(filter, filterSize, src, dest + i, dither,
> + FUNC(yuv2planeX_8_16)(filter, filterSize, src, dest + i, dither,
> offset, i);
>
> yuv2planeX_u(filter, filterSize, src, dest, dstW, dither, offset, i);
> diff --git a/libswscale/ppc/swscale_vsx.c b/libswscale/ppc/swscale_vsx.c
> index 70da6ae..77680f8 100644
> --- a/libswscale/ppc/swscale_vsx.c
> +++ b/libswscale/ppc/swscale_vsx.c
> @@ -83,6 +83,8 @@
> #include "swscale_ppc_template.c"
> #undef FUNC
>
> +#undef vzero
> +
> #endif /* !HAVE_BIGENDIAN */
>
> static void yuv2plane1_8_u(const int16_t *src, uint8_t *dest, int dstW,
> @@ -180,6 +182,76 @@ static void yuv2plane1_nbps_vsx(const int16_t *src,
> uint16_t *dest, int dstW,
> yuv2plane1_nbps_u(src, dest, dstW, big_endian, output_bits, i);
> }
>
> +static void yuv2planeX_nbps_u(const int16_t *filter, int filterSize,
> + const int16_t **src, uint16_t *dest, int
> dstW,
> + int big_endian, int output_bits, int start)
> +{
> + int i;
> + int shift = 11 + 16 - output_bits;
> +
> + for (i = start; i < dstW; i++) {
> + int val = 1 << (shift - 1);
> + int j;
> +
> + for (j = 0; j < filterSize; j++)
> + val += src[j][i] * filter[j];
> +
> + output_pixel(&dest[i], val);
> + }
> +}
> +
> +static void yuv2planeX_nbps_vsx(const int16_t *filter, int filterSize,
> + const int16_t **src, uint16_t *dest, int
> dstW,
> + int big_endian, int output_bits)
> +{
> + const int dst_u = -(uintptr_t)dest & 7;
> + const int shift = 11 + 16 - output_bits;
> + const int add = (1 << (shift - 1));
> + const int clip = (1 << output_bits) - 1;
> + const uint16_t swap = big_endian ? 8 : 0;
> + const vector uint32_t vadd = (vector uint32_t) {add, add, add, add};
> + const vector uint32_t vshift = (vector uint32_t) {shift, shift, shift,
> shift};
> + const vector uint16_t vswap = (vector uint16_t) {swap, swap, swap,
> swap, swap, swap, swap, swap};
> + const vector uint16_t vlargest = (vector uint16_t) {clip, clip, clip,
> clip, clip, clip, clip, clip};
> + const vector int16_t vzero = vec_splat_s16(0);
> + const vector uint8_t vperm = (vector uint8_t) {0, 1, 8, 9, 2, 3, 10,
> 11, 4, 5, 12, 13, 6, 7, 14, 15};
> + vector int16_t vfilter[MAX_FILTER_SIZE], vin;
> + vector uint16_t v;
> + vector uint32_t vleft, vright, vtmp;
> + int i, j;
> +
> + for (i = 0; i < filterSize; i++) {
> + vfilter[i] = (vector int16_t) {filter[i], filter[i], filter[i],
> filter[i],
> + filter[i], filter[i], filter[i],
> filter[i]};
> + }
> +
> + yuv2planeX_nbps_u(filter, filterSize, src, dest, dst_u, big_endian,
> output_bits, 0);
> +
> + for (i = dst_u; i < dstW - 7; i += 8) {
> + vleft = vright = vadd;
> +
> + for (j = 0; j < filterSize; j++) {
> + vin = vec_vsx_ld(0, &src[j][i]);
> + vtmp = (vector uint32_t) vec_mule(vin, vfilter[j]);
> + vleft = vec_add(vleft, vtmp);
> + vtmp = (vector uint32_t) vec_mulo(vin, vfilter[j]);
> + vright = vec_add(vright, vtmp);
> + }
> +
> + vleft = vec_sra(vleft, vshift);
> + vright = vec_sra(vright, vshift);
> + v = vec_packsu(vleft, vright);
> + v = (vector uint16_t) vec_max((vector int16_t) v, vzero);
> + v = vec_min(v, vlargest);
> + v = vec_rl(v, vswap);
> + v = vec_perm(v, v, vperm);
> + vec_st(v, 0, &dest[i]);
> + }
> +
> + yuv2planeX_nbps_u(filter, filterSize, src, dest, dstW, big_endian,
> output_bits, i);
> +}
> +
> +
> #undef output_pixel
>
> #define output_pixel(pos, val, bias, signedness) \
> @@ -234,7 +306,97 @@ static void yuv2plane1_16_vsx(const int32_t *src,
> uint16_t *dest, int dstW,
> yuv2plane1_16_u(src, dest, dstW, big_endian, output_bits, i);
> }
>
> +#ifdef HAVE_POWER8
> +
> +static void yuv2planeX_16_u(const int16_t *filter, int filterSize,
> + const int32_t **src, uint16_t *dest, int dstW,
> + int big_endian, int output_bits, int start)
> +{
> + int i;
> + int shift = 15;
> +
> + for (i = start; i < dstW; i++) {
> + int val = 1 << (shift - 1);
> + int j;
> +
> + /* range of val is [0,0x7FFFFFFF], so 31 bits, but with
> lanczos/spline
> + * filters (or anything with negative coeffs, the range can be
> slightly
> + * wider in both directions. To account for this overflow, we
> subtract
> + * a constant so it always fits in the signed range (assuming a
> + * reasonable filterSize), and re-add that at the end. */
> + val -= 0x40000000;
> + for (j = 0; j < filterSize; j++)
> + val += src[j][i] * (unsigned)filter[j];
> +
> + output_pixel(&dest[i], val, 0x8000, int);
> + }
> +}
> +
> +static void yuv2planeX_16_vsx(const int16_t *filter, int filterSize,
> + const int32_t **src, uint16_t *dest, int
> dstW,
> + int big_endian, int output_bits)
> +{
> + const int dst_u = -(uintptr_t)dest & 7;
> + const int shift = 15;
> + const int bias = 0x8000;
> + const int add = (1 << (shift - 1)) - 0x40000000;
> + const uint16_t swap = big_endian ? 8 : 0;
> + const vector uint32_t vadd = (vector uint32_t) {add, add, add, add};
> + const vector uint32_t vshift = (vector uint32_t) {shift, shift, shift,
> shift};
> + const vector uint16_t vswap = (vector uint16_t) {swap, swap, swap,
> swap, swap, swap, swap, swap};
> + const vector uint16_t vbias = (vector uint16_t) {bias, bias, bias,
> bias, bias, bias, bias, bias};
> + vector int32_t vfilter[MAX_FILTER_SIZE];
> + vector uint16_t v;
> + vector uint32_t vleft, vright, vtmp;
> + vector int32_t vin32l, vin32r;
> + int i, j;
> +
> + for (i = 0; i < filterSize; i++) {
> + vfilter[i] = (vector int32_t) {filter[i], filter[i], filter[i],
> filter[i]};
> + }
> +
> + yuv2planeX_16_u(filter, filterSize, src, dest, dst_u, big_endian,
> output_bits, 0);
> +
> + for (i = dst_u; i < dstW - 7; i += 8) {
> + vleft = vright = vadd;
> +
> + for (j = 0; j < filterSize; j++) {
> + vin32l = vec_vsx_ld(0, &src[j][i]);
> + vin32r = vec_vsx_ld(0, &src[j][i + 4]);
> +
> +#ifdef __GNUC__
> + // GCC does not support vmuluwm yet. Bug open.
> + __asm__("vmuluwm %0, %1, %2" : "=v"(vtmp) : "v"(vin32l),
> "v"(vfilter[j]));
> + vleft = vec_add(vleft, vtmp);
> + __asm__("vmuluwm %0, %1, %2" : "=v"(vtmp) : "v"(vin32r),
> "v"(vfilter[j]));
> + vright = vec_add(vright, vtmp);
> +#else
> + // No idea which compilers this works in, untested. Copied from
> libsimdpp
> + vtmp = vec_vmuluwm(vin32l, vfilter[j]);
> + vleft = vec_add(vleft, vtmp);
> + vtmp = vec_vmuluwm(vin32r, vfilter[j]);
> + vright = vec_add(vright, vtmp);
> +#endif
Is there no xlc installed on your test system?
I suspect an earlier patch from you already
broke xlc compilation...
Carl Eugen
More information about the ffmpeg-devel
mailing list