[x265] [PATCH 2 of 3] implementation for minimum CU size
Deepthi Nandakumar
deepthi at multicorewareinc.com
Wed Feb 18 10:12:44 CET 2015
I agree. Santhoshini, except for the whitespace fixes and set_globals
addition, no changes should be necessary. The max depth is set in SPS
already.
On Tue, Feb 17, 2015 at 11:14 AM, Steve Borho <steve at borho.org> wrote:
> On 02/16, santhoshini at multicorewareinc.com wrote:
> > # HG changeset patch
> > # User Santhoshini Sekar<santhoshini at multicorewareinc.com>
> > # Date 1424077388 -19800
> > # Mon Feb 16 14:33:08 2015 +0530
> > # Node ID 24e3fc60cb4183fff50921a332f0fd0d98cc0f56
> > # Parent 59466f1455ce52920eb642cf06db4b4de8d8ff10
> > implementation for minimum CU size
> >
> > diff -r 59466f1455ce -r 24e3fc60cb41 source/common/cudata.cpp
> > --- a/source/common/cudata.cpp Mon Feb 16 14:28:19 2015 +0530
> > +++ b/source/common/cudata.cpp Mon Feb 16 14:33:08 2015 +0530
> > @@ -2066,14 +2066,14 @@
> >
> > #define CU_SET_FLAG(bitfield, flag, value) (bitfield) = ((bitfield) &
> (~(flag))) | ((~((value) - 1)) & (flag))
> >
> > -void CUData::calcCTUGeoms(uint32_t ctuWidth, uint32_t ctuHeight,
> uint32_t maxCUSize, CUGeom cuDataArray[CUGeom::MAX_GEOMS])
> > +void CUData::calcCTUGeoms(uint32_t ctuWidth, uint32_t ctuHeight,
> uint32_t maxCUSize, uint32_t minCUSize, CUGeom
> cuDataArray[CUGeom::MAX_GEOMS])
> > {
> > // Initialize the coding blocks inside the CTB
> > - for (uint32_t log2CUSize = g_log2Size[maxCUSize], rangeCUIdx = 0;
> log2CUSize >= MIN_LOG2_CU_SIZE; log2CUSize--)
> > + for (uint32_t log2CUSize = g_log2Size[maxCUSize], rangeCUIdx = 0;
> log2CUSize >= g_log2Size[minCUSize]; log2CUSize--)
> > {
> > uint32_t blockSize = 1 << log2CUSize;
> > uint32_t sbWidth = 1 << (g_log2Size[maxCUSize] - log2CUSize);
> > - int32_t lastLevelFlag = log2CUSize == MIN_LOG2_CU_SIZE;
> > + int32_t lastLevelFlag = log2CUSize == g_log2Size[minCUSize];
> > for (uint32_t sbY = 0; sbY < sbWidth; sbY++)
> > {
> > for (uint32_t sbX = 0; sbX < sbWidth; sbX++)
> > diff -r 59466f1455ce -r 24e3fc60cb41 source/common/cudata.h
> > --- a/source/common/cudata.h Mon Feb 16 14:28:19 2015 +0530
> > +++ b/source/common/cudata.h Mon Feb 16 14:33:08 2015 +0530
> > @@ -158,7 +158,7 @@
> > CUData();
> >
> > void initialize(const CUDataMemPool& dataPool, uint32_t depth,
> int csp, int instance);
> > - static void calcCTUGeoms(uint32_t ctuWidth, uint32_t ctuHeight,
> uint32_t maxCUSize, CUGeom cuDataArray[CUGeom::MAX_GEOMS]);
> > + static void calcCTUGeoms(uint32_t ctuWidth, uint32_t ctuHeight,
> uint32_t maxCUSize, uint32_t minCUSize, CUGeom
> cuDataArray[CUGeom::MAX_GEOMS]);
> >
> > void initCTU(const Frame& frame, uint32_t cuAddr, int qp);
> > void initSubCU(const CUData& ctu, const CUGeom& cuGeom);
> > diff -r 59466f1455ce -r 24e3fc60cb41 source/common/param.cpp
> > --- a/source/common/param.cpp Mon Feb 16 14:28:19 2015 +0530
> > +++ b/source/common/param.cpp Mon Feb 16 14:33:08 2015 +0530
> > @@ -127,6 +127,7 @@
> >
> > /* CU definitions */
> > param->maxCUSize = 64;
> > + param->minCUSize = 8;
> > param->tuQTMaxInterDepth = 1;
> > param->tuQTMaxIntraDepth = 1;
> > param->maxTUSize = 32;
> > @@ -978,6 +979,8 @@
> > "x265 was compiled for 8bit encodes, only 8bit internal depth
> supported");
> > #endif
> >
> > + CHECK(param->minCUSize != 64 && param->minCUSize != 32 &&
> param->minCUSize != 16 && param->minCUSize != 8,
> > + "minimim CU size must be 8, 16, 32, or 64");
> > CHECK(param->rc.qp < -6 * (param->internalBitDepth - 8) ||
> param->rc.qp > QP_MAX_SPEC,
> > "QP exceeds supported range (-QpBDOffsety to 51)");
> > CHECK(param->fpsNum == 0 || param->fpsDenom == 0,
> > @@ -1166,14 +1169,20 @@
> > }
> > else
> > {
> > + if (param->minCUSize > param->maxCUSize)
> > + {
> > + x265_log(param, X265_LOG_WARNING, "Min CU size should be less
> than or equal to max CU size, setting min CU size = %d\n",
> param->maxCUSize);
> > + param->minCUSize = param->maxCUSize;
> > + }
>
> w/s here is borked
>
> > uint32_t maxLog2CUSize = (uint32_t)g_log2Size[param->maxCUSize];
> > + uint32_t minLog2CUSize = (uint32_t)g_log2Size[param->minCUSize];
> >
> > // set max CU width & height
> > g_maxCUSize = param->maxCUSize;
> > g_maxLog2CUSize = maxLog2CUSize;
> >
> > // compute actual CU depth with respect to config depth and max
> transform size
> > - g_maxCUDepth = maxLog2CUSize - MIN_LOG2_CU_SIZE;
> > + g_maxCUDepth = maxLog2CUSize - minLog2CUSize;
> > g_unitSizeDepth = maxLog2CUSize - LOG2_UNIT_SIZE;
>
> If g_maxCUDepth is going to be based on minLog2CUSize, then we must
> validate this in x265_set_globals().
>
> We have a design decision to make here. We could leave g_maxCUDepth
> unmodified here and always signal the min CU size as 8x8 but voluntarily
> never split 16x16 (or 32x32) CUs. We will have to signal no-splits at
> the lowest remaining level since they will not be implied, this is a
> compression cost. Or we can signal the intended min CU size and enforce
> this same limit for all encodes in the same process: a loss of
> flexibility but a slight gain in compression efficiency.
>
> I'd love to hear dissenting opinions, but based on what I know most
> use-cases that use multiple encodes in a single process will use the
> same base preset for all of them. In that case, it is better to err on
> the side of compression efficiency.
>
> > // initialize partition order
> > @@ -1195,7 +1204,7 @@
> > if (param->interlaceMode)
> > x265_log(param, X265_LOG_INFO, "Interlaced field inputs
> : %s\n", x265_interlace_names[param->interlaceMode]);
> >
> > - x265_log(param, X265_LOG_INFO, "Coding QT: max CU size, min CU size
> : %d / %d\n", param->maxCUSize, 8);
> > + x265_log(param, X265_LOG_INFO, "Coding QT: max CU size, min CU size
> : %d / %d\n", param->maxCUSize, param->minCUSize);
> >
> > x265_log(param, X265_LOG_INFO, "Residual QT: max TU size, max depth
> : %d / %d inter / %d intra\n",
> > param->maxTUSize, param->tuQTMaxInterDepth,
> param->tuQTMaxIntraDepth);
> > diff -r 59466f1455ce -r 24e3fc60cb41 source/encoder/analysis.cpp
> > --- a/source/encoder/analysis.cpp Mon Feb 16 14:28:19 2015 +0530
> > +++ b/source/encoder/analysis.cpp Mon Feb 16 14:33:08 2015 +0530
> > @@ -260,7 +260,7 @@
> > checkIntra(md.pred[PRED_INTRA], cuGeom, SIZE_2Nx2N, NULL, NULL);
> > checkBestMode(md.pred[PRED_INTRA], depth);
> >
> > - if (depth == g_maxCUDepth)
> > + if (cuGeom.log2CUSize == 3 &&
> m_slice->m_sps->quadtreeTULog2MinSize < 3)
> > {
> > md.pred[PRED_INTRA_NxN].cu.initSubCU(parentCTU, cuGeom);
> > checkIntra(md.pred[PRED_INTRA_NxN], cuGeom, SIZE_NxN, NULL,
> NULL);
> > @@ -472,7 +472,7 @@
> > {
> > case 0:
> > slave->checkIntra(md.pred[PRED_INTRA], *m_curGeom,
> SIZE_2Nx2N, NULL, NULL);
> > - if (m_curGeom->depth == g_maxCUDepth &&
> m_curGeom->log2CUSize > m_slice->m_sps->quadtreeTULog2MinSize)
> > + if (m_curGeom->log2CUSize == 3 &&
> m_slice->m_sps->quadtreeTULog2MinSize < 3)
> > slave->checkIntra(md.pred[PRED_INTRA_NxN], *m_curGeom,
> SIZE_NxN, NULL, NULL);
> > break;
> >
> > @@ -556,7 +556,7 @@
> > if (bTryIntra)
> > {
> > md.pred[PRED_INTRA].cu.initSubCU(parentCTU, cuGeom);
> > - if (depth == g_maxCUDepth && cuGeom.log2CUSize >
> m_slice->m_sps->quadtreeTULog2MinSize)
> > + if (cuGeom.log2CUSize == 3 &&
> m_slice->m_sps->quadtreeTULog2MinSize < 3)
> > md.pred[PRED_INTRA_NxN].cu.initSubCU(parentCTU, cuGeom);
> > }
> >
> > @@ -704,7 +704,7 @@
> > if (bTryIntra)
> > {
> > checkBestMode(md.pred[PRED_INTRA], depth);
> > - if (depth == g_maxCUDepth && cuGeom.log2CUSize >
> m_slice->m_sps->quadtreeTULog2MinSize)
> > + if (cuGeom.log2CUSize == 3 &&
> m_slice->m_sps->quadtreeTULog2MinSize < 3)
> > checkBestMode(md.pred[PRED_INTRA_NxN], depth);
> > }
> > }
> > @@ -1187,7 +1187,7 @@
> > checkIntra(md.pred[PRED_INTRA], cuGeom, SIZE_2Nx2N,
> NULL, NULL);
> > checkBestMode(md.pred[PRED_INTRA], depth);
> >
> > - if (depth == g_maxCUDepth && cuGeom.log2CUSize >
> m_slice->m_sps->quadtreeTULog2MinSize)
> > + if (cuGeom.log2CUSize == 3 &&
> m_slice->m_sps->quadtreeTULog2MinSize < 3)
> > {
> > md.pred[PRED_INTRA_NxN].cu.initSubCU(parentCTU,
> cuGeom);
> > checkIntra(md.pred[PRED_INTRA_NxN], cuGeom,
> SIZE_NxN, NULL, NULL);
> > diff -r 59466f1455ce -r 24e3fc60cb41 source/encoder/encoder.cpp
> > --- a/source/encoder/encoder.cpp Mon Feb 16 14:28:19 2015 +0530
> > +++ b/source/encoder/encoder.cpp Mon Feb 16 14:33:08 2015 +0530
> > @@ -961,7 +961,7 @@
> > finalLog.cuInterDistribution[depth][m] +=
> enclog.cuInterDistribution[depth][m];
> > }
> >
> > - if (depth == g_maxCUDepth)
> > + if (m_frameEncoder[i]->m_cuGeoms->log2CUSize == 3 &&
> m_sps.quadtreeTULog2MinSize < 3)
> > finalLog.cntIntraNxN += enclog.cntIntraNxN;
> > if (sliceType != I_SLICE)
> > {
> > @@ -1054,14 +1054,14 @@
> > cuIntraDistribution[1],
> cuIntraDistribution[2]);
> > if (sliceType != I_SLICE)
> > {
> > - if (depth == g_maxCUDepth)
> > + if (g_log2Size[cuSize] == 3 &&
> m_sps.quadtreeTULog2MinSize < 3)
> > len += sprintf(stats + len, " %dx%d "X265_LL
> "%%", cuSize / 2, cuSize / 2, cntIntraNxN);
> > }
> >
> > len += sprintf(stats + len, ")");
> > if (sliceType == I_SLICE)
> > {
> > - if (depth == g_maxCUDepth)
> > + if (g_log2Size[cuSize] == 3 &&
> m_sps.quadtreeTULog2MinSize < 3)
> > len += sprintf(stats + len, " %dx%d: "X265_LL
> "%%", cuSize / 2, cuSize / 2, cntIntraNxN);
> > }
> > }
> > @@ -1708,10 +1708,10 @@
> > m_conformanceWindow.leftOffset = 0;
> >
> > /* set pad size if width is not multiple of the minimum CU size */
> > - if (p->sourceWidth & (MIN_CU_SIZE - 1))
> > + if (p->sourceWidth & (p->minCUSize - 1))
> > {
> > - uint32_t rem = p->sourceWidth & (MIN_CU_SIZE - 1);
> > - uint32_t padsize = MIN_CU_SIZE - rem;
> > + uint32_t rem = p->sourceWidth & (p->minCUSize - 1);
> > + uint32_t padsize = p->minCUSize - rem;
> > p->sourceWidth += padsize;
> >
> > m_conformanceWindow.bEnabled = true;
> > @@ -1719,10 +1719,10 @@
> > }
> >
> > /* set pad size if height is not multiple of the minimum CU size */
> > - if (p->sourceHeight & (MIN_CU_SIZE - 1))
> > + if (p->sourceHeight & (p->minCUSize - 1))
> > {
> > - uint32_t rem = p->sourceHeight & (MIN_CU_SIZE - 1);
> > - uint32_t padsize = MIN_CU_SIZE - rem;
> > + uint32_t rem = p->sourceHeight & (p->minCUSize - 1);
> > + uint32_t padsize = p->minCUSize - rem;
> > p->sourceHeight += padsize;
> >
> > m_conformanceWindow.bEnabled = true;
> > diff -r 59466f1455ce -r 24e3fc60cb41 source/encoder/frameencoder.cpp
> > --- a/source/encoder/frameencoder.cpp Mon Feb 16 14:28:19 2015 +0530
> > +++ b/source/encoder/frameencoder.cpp Mon Feb 16 14:33:08 2015 +0530
> > @@ -143,6 +143,7 @@
> > {
> > /* Geoms only vary between CTUs in the presence of picture edges */
> > int maxCUSize = m_param->maxCUSize;
> > + int minCUSize = m_param->minCUSize;
> > int heightRem = m_param->sourceHeight & (maxCUSize - 1);
> > int widthRem = m_param->sourceWidth & (maxCUSize - 1);
> > int allocGeoms = 1; // body
> > @@ -157,7 +158,7 @@
> > return false;
> >
> > // body
> > - CUData::calcCTUGeoms(maxCUSize, maxCUSize, maxCUSize, m_cuGeoms);
> > + CUData::calcCTUGeoms(maxCUSize, maxCUSize, maxCUSize, minCUSize,
> m_cuGeoms);
> > memset(m_ctuGeomMap, 0, sizeof(uint32_t) * m_numRows * m_numCols);
> > if (allocGeoms == 1)
> > return true;
> > @@ -166,7 +167,7 @@
> > if (widthRem)
> > {
> > // right
> > - CUData::calcCTUGeoms(widthRem, maxCUSize, maxCUSize, m_cuGeoms
> + countGeoms * CUGeom::MAX_GEOMS);
> > + CUData::calcCTUGeoms(widthRem, maxCUSize, maxCUSize, minCUSize,
> m_cuGeoms + countGeoms * CUGeom::MAX_GEOMS);
> > for (uint32_t i = 0; i < m_numRows; i++)
> > {
> > uint32_t ctuAddr = m_numCols * (i + 1) - 1;
> > @@ -177,7 +178,7 @@
> > if (heightRem)
> > {
> > // bottom
> > - CUData::calcCTUGeoms(maxCUSize, heightRem, maxCUSize, m_cuGeoms
> + countGeoms * CUGeom::MAX_GEOMS);
> > + CUData::calcCTUGeoms(maxCUSize, heightRem, maxCUSize,
> minCUSize, m_cuGeoms + countGeoms * CUGeom::MAX_GEOMS);
> > for (uint32_t i = 0; i < m_numCols; i++)
> > {
> > uint32_t ctuAddr = m_numCols * (m_numRows - 1) + i;
> > @@ -188,7 +189,7 @@
> > if (widthRem)
> > {
> > // corner
> > - CUData::calcCTUGeoms(widthRem, heightRem, maxCUSize,
> m_cuGeoms + countGeoms * CUGeom::MAX_GEOMS);
> > + CUData::calcCTUGeoms(widthRem, heightRem, maxCUSize,
> minCUSize, m_cuGeoms + countGeoms * CUGeom::MAX_GEOMS);
> >
> > uint32_t ctuAddr = m_numCols * m_numRows - 1;
> > m_ctuGeomMap[ctuAddr] = countGeoms * CUGeom::MAX_GEOMS;
> > @@ -1038,7 +1039,7 @@
> > else if (ctu.m_partSize[absPartIdx] != SIZE_2Nx2N)
> > {
> > /* TODO: log intra modes at absPartIdx +0 to +3 */
> > - X265_CHECK(depth == g_maxCUDepth, "Intra NxN found at
> improbable depth\n");
> > + X265_CHECK(ctu.m_log2CUSize[absPartIdx] == 3 &&
> ctu.m_slice->m_sps->quadtreeTULog2MinSize < 3, "Intra NxN found at
> improbable depth\n");
> > log->cntIntraNxN++;
> > log->cntIntra[depth]--;
> > }
> > @@ -1086,7 +1087,7 @@
> >
> > if (ctu.m_partSize[absPartIdx] != SIZE_2Nx2N)
> > {
> > - X265_CHECK(depth == g_maxCUDepth, "Intra NxN found
> at improbable depth\n");
> > + X265_CHECK(ctu.m_log2CUSize[absPartIdx] == 3 &&
> ctu.m_slice->m_sps->quadtreeTULog2MinSize < 3, "Intra NxN found at
> improbable depth\n");
> > log->cntIntraNxN++;
> > /* TODO: log intra modes at absPartIdx +0 to +3 */
> > }
> > diff -r 59466f1455ce -r 24e3fc60cb41 source/x265.h
> > --- a/source/x265.h Mon Feb 16 14:28:19 2015 +0530
> > +++ b/source/x265.h Mon Feb 16 14:33:08 2015 +0530
> > @@ -579,6 +579,10 @@
> > * frame parallelism will become because of the increase in rows.
> default 64 */
> > uint32_t maxCUSize;
> >
> > + /* Miniumum CU width and height in pixels. The size must be 64,
> 32, 16, or 8.
> > + * default 8 */
> > + uint32_t minCUSize;
> > +
> > /* Enable rectangular motion prediction partitions (vertical and
> > * horizontal), available at all CU depths from 64x64 to 8x8.
> Default is
> > * disabled */
> > _______________________________________________
> > 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
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.videolan.org/pipermail/x265-devel/attachments/20150218/0621e37d/attachment-0001.html>
More information about the x265-devel
mailing list