[x265] [PATCH] inline mvcost() to reduce address operators
Steve Borho
steve at borho.org
Wed May 6 19:04:23 CEST 2015
On 05/06, chen wrote:
>
>
> At 2015-05-06 08:38:24,"Steve Borho" <steve at borho.org> wrote:
> >On 05/05, Min Chen wrote:
> >> # HG changeset patch
> >> # User Min Chen <chenm003 at 163.com>
> >> # Date 1430862259 25200
> >> # Node ID 50ce2c0ddfbb743b45f678ee2e6b796762ad868f
> >> # Parent f32e6464225afa02983af1b1905f50cdccae5244
> >> inline mvcost() to reduce address operators
> >
> >I'm skeptical that this is a good idea. have you measured the difference
> >in performance with encoders built with profile-guided optimizations?
>
> I found this idea from vtune assembly report (preset ultrafast), it show bottleneck in signed address extendsion because offset is signed integer.
> In the ICL, when we use keyword 'restrict' can avoid part of these reduce operators.
>
> after this patch, I got 10% improve in ME module or call ~2% in total encode.
I see, so this is bypassing the int->short->int signed int conversions.
In that case it's probably ok to explode the code this way.
> >
> >> source/encoder/motion.cpp | 48 ++++++++++++++++++++++++++++++--------------
> >> 1 files changed, 33 insertions(+), 15 deletions(-)
> >>
> >> diff -r f32e6464225a -r 50ce2c0ddfbb source/encoder/motion.cpp
> >> --- a/source/encoder/motion.cpp Mon May 04 15:15:42 2015 -0500
> >> +++ b/source/encoder/motion.cpp Tue May 05 14:44:19 2015 -0700
> >> @@ -234,9 +234,14 @@
> >> pix_base + (m1x) + (m1y) * stride, \
> >> pix_base + (m2x) + (m2y) * stride, \
> >> stride, costs); \
> >> - (costs)[0] += mvcost((bmv + MV(m0x, m0y)) << 2); \
> >> - (costs)[1] += mvcost((bmv + MV(m1x, m1y)) << 2); \
> >> - (costs)[2] += mvcost((bmv + MV(m2x, m2y)) << 2); \
> >> + const uint16_t *base_mvx = &m_cost_mvx[(bmv.x + (m0x)) << 2]; \
> >> + const uint16_t *base_mvy = &m_cost_mvy[(bmv.y + (m0y)) << 2]; \
> >> + X265_CHECK(mvcost((bmv + MV(m0x, m0y)) << 2) == (base_mvx[((m0x) - (m0x)) << 2] + base_mvy[((m0y) - (m0y)) << 2]), "mvcost() check failure\n"); \
> >> + X265_CHECK(mvcost((bmv + MV(m1x, m1y)) << 2) == (base_mvx[((m1x) - (m0x)) << 2] + base_mvy[((m1y) - (m0y)) << 2]), "mvcost() check failure\n"); \
> >> + X265_CHECK(mvcost((bmv + MV(m2x, m2y)) << 2) == (base_mvx[((m2x) - (m0x)) << 2] + base_mvy[((m2y) - (m0y)) << 2]), "mvcost() check failure\n"); \
> >> + (costs)[0] += (base_mvx[((m0x) - (m0x)) << 2] + base_mvy[((m0y) - (m0y)) << 2]); \
> >> + (costs)[1] += (base_mvx[((m1x) - (m0x)) << 2] + base_mvy[((m1y) - (m0y)) << 2]); \
> >> + (costs)[2] += (base_mvx[((m2x) - (m0x)) << 2] + base_mvy[((m2y) - (m0y)) << 2]); \
> >> }
> >>
> >> #define COST_MV_PT_DIST_X4(m0x, m0y, p0, d0, m1x, m1y, p1, d1, m2x, m2y, p2, d2, m3x, m3y, p3, d3) \
> >> @@ -247,10 +252,10 @@
> >> fref + (m2x) + (m2y) * stride, \
> >> fref + (m3x) + (m3y) * stride, \
> >> stride, costs); \
> >> - costs[0] += mvcost(MV(m0x, m0y) << 2); \
> >> - costs[1] += mvcost(MV(m1x, m1y) << 2); \
> >> - costs[2] += mvcost(MV(m2x, m2y) << 2); \
> >> - costs[3] += mvcost(MV(m3x, m3y) << 2); \
> >> + (costs)[0] += mvcost(MV(m0x, m0y) << 2); \
> >> + (costs)[1] += mvcost(MV(m1x, m1y) << 2); \
> >> + (costs)[2] += mvcost(MV(m2x, m2y) << 2); \
> >> + (costs)[3] += mvcost(MV(m3x, m3y) << 2); \
why was this one spared? you only added parents around the macro
argument.
> >> COPY4_IF_LT(bcost, costs[0], bmv, MV(m0x, m0y), bPointNr, p0, bDistance, d0); \
> >> COPY4_IF_LT(bcost, costs[1], bmv, MV(m1x, m1y), bPointNr, p1, bDistance, d1); \
> >> COPY4_IF_LT(bcost, costs[2], bmv, MV(m2x, m2y), bPointNr, p2, bDistance, d2); \
> >> @@ -266,10 +271,16 @@
> >> pix_base + (m2x) + (m2y) * stride, \
> >> pix_base + (m3x) + (m3y) * stride, \
> >> stride, costs); \
> >> - costs[0] += mvcost((omv + MV(m0x, m0y)) << 2); \
> >> - costs[1] += mvcost((omv + MV(m1x, m1y)) << 2); \
> >> - costs[2] += mvcost((omv + MV(m2x, m2y)) << 2); \
> >> - costs[3] += mvcost((omv + MV(m3x, m3y)) << 2); \
> >> + const uint16_t *base_mvx = &m_cost_mvx[(omv.x << 2)]; \
> >> + const uint16_t *base_mvy = &m_cost_mvy[(omv.y << 2)]; \
> >> + X265_CHECK(mvcost((omv + MV(m0x, m0y)) << 2) == (base_mvx[(m0x) << 2] + base_mvy[(m0y) << 2]), "mvcost() check failure\n"); \
> >> + X265_CHECK(mvcost((omv + MV(m1x, m1y)) << 2) == (base_mvx[(m1x) << 2] + base_mvy[(m1y) << 2]), "mvcost() check failure\n"); \
> >> + X265_CHECK(mvcost((omv + MV(m2x, m2y)) << 2) == (base_mvx[(m2x) << 2] + base_mvy[(m2y) << 2]), "mvcost() check failure\n"); \
> >> + X265_CHECK(mvcost((omv + MV(m3x, m3y)) << 2) == (base_mvx[(m3x) << 2] + base_mvy[(m3y) << 2]), "mvcost() check failure\n"); \
> >> + costs[0] += (base_mvx[(m0x) << 2] + base_mvy[(m0y) << 2]); \
> >> + costs[1] += (base_mvx[(m1x) << 2] + base_mvy[(m1y) << 2]); \
> >> + costs[2] += (base_mvx[(m2x) << 2] + base_mvy[(m2y) << 2]); \
> >> + costs[3] += (base_mvx[(m3x) << 2] + base_mvy[(m3y) << 2]); \
> >> COPY2_IF_LT(bcost, costs[0], bmv, omv + MV(m0x, m0y)); \
> >> COPY2_IF_LT(bcost, costs[1], bmv, omv + MV(m1x, m1y)); \
> >> COPY2_IF_LT(bcost, costs[2], bmv, omv + MV(m2x, m2y)); \
> >> @@ -285,10 +296,17 @@
> >> pix_base + (m2x) + (m2y) * stride, \
> >> pix_base + (m3x) + (m3y) * stride, \
> >> stride, costs); \
> >> - (costs)[0] += mvcost((bmv + MV(m0x, m0y)) << 2); \
> >> - (costs)[1] += mvcost((bmv + MV(m1x, m1y)) << 2); \
> >> - (costs)[2] += mvcost((bmv + MV(m2x, m2y)) << 2); \
> >> - (costs)[3] += mvcost((bmv + MV(m3x, m3y)) << 2); \
> >> + /* TODO: use restrict keyword in ICL */ \
use restrict where?
> >> + const uint16_t *base_mvx = &m_cost_mvx[(bmv.x << 2)]; \
> >> + const uint16_t *base_mvy = &m_cost_mvy[(bmv.y << 2)]; \
> >> + X265_CHECK(mvcost((bmv + MV(m0x, m0y)) << 2) == (base_mvx[(m0x) << 2] + base_mvy[(m0y) << 2]), "mvcost() check failure\n"); \
> >> + X265_CHECK(mvcost((bmv + MV(m1x, m1y)) << 2) == (base_mvx[(m1x) << 2] + base_mvy[(m1y) << 2]), "mvcost() check failure\n"); \
> >> + X265_CHECK(mvcost((bmv + MV(m2x, m2y)) << 2) == (base_mvx[(m2x) << 2] + base_mvy[(m2y) << 2]), "mvcost() check failure\n"); \
> >> + X265_CHECK(mvcost((bmv + MV(m3x, m3y)) << 2) == (base_mvx[(m3x) << 2] + base_mvy[(m3y) << 2]), "mvcost() check failure\n"); \
> >> + (costs)[0] += (base_mvx[(m0x) << 2] + base_mvy[(m0y) << 2]); \
> >> + (costs)[1] += (base_mvx[(m1x) << 2] + base_mvy[(m1y) << 2]); \
> >> + (costs)[2] += (base_mvx[(m2x) << 2] + base_mvy[(m2y) << 2]); \
> >> + (costs)[3] += (base_mvx[(m3x) << 2] + base_mvy[(m3y) << 2]); \
> >> }
> >>
> >> #define DIA1_ITER(mx, my) \
> >>
> >> _______________________________________________
> >> x265-devel mailing list
> >> x265-devel at videolan.org
> >> https://mailman.videolan.org/listinfo/x265-devel
> >
> >--
> >Steve Borho
> >_______________________________________________
> >x265-devel mailing list
> >x265-devel at videolan.org
> >https://mailman.videolan.org/listinfo/x265-devel
> _______________________________________________
> x265-devel mailing list
> x265-devel at videolan.org
> https://mailman.videolan.org/listinfo/x265-devel
--
Steve Borho
More information about the x265-devel
mailing list