[x265] [PATCH] Fix MV Wrap-around
chen
chenm003 at 163.com
Fri Mar 29 03:24:04 CET 2019
Could you please make a new compile time option to handle 8K? similar as high bit depth, because 64-bits operators much expensive in most platform.
At 2019-03-29 08:23:14, tamura at pegasys-inc.com wrote:
># HG changeset patch
># User Masaharu Tamura <tamura at pegasys-inc.com>
># Date 1553818841 -32400
># Fri Mar 29 09:20:41 2019 +0900
># Node ID 0018ca1ca42f1c91422b9a78091967cc45bb944c
># Parent 3e5032228123c1898d408a30e45ac15eb687ffb6
>Fix MV Wrap-around
>
>Fixed that wrap-around from MV structure overflow occurred around 8K pixels or over.
>
>diff -r 3e5032228123 -r 0018ca1ca42f source/CMakeLists.txt
>--- a/source/CMakeLists.txt Fri Mar 22 13:09:49 2019 +0530
>+++ b/source/CMakeLists.txt Fri Mar 29 09:20:41 2019 +0900
>@@ -29,7 +29,7 @@
> option(STATIC_LINK_CRT "Statically link C runtime for release builds" OFF)
> mark_as_advanced(FPROFILE_USE FPROFILE_GENERATE NATIVE_BUILD)
> # X265_BUILD must be incremented each time the public API is changed
>-set(X265_BUILD 172)
>+set(X265_BUILD 173)
> configure_file("${PROJECT_SOURCE_DIR}/x265.def.in"
> "${PROJECT_BINARY_DIR}/x265.def")
> configure_file("${PROJECT_SOURCE_DIR}/x265_config.h.in"
>diff -r 3e5032228123 -r 0018ca1ca42f source/common/cudata.cpp
>--- a/source/common/cudata.cpp Fri Mar 22 13:09:49 2019 +0530
>+++ b/source/common/cudata.cpp Fri Mar 29 09:20:41 2019 +0900
>@@ -106,7 +106,7 @@
> int mvx = x265_clip3(-32768, 32767, (scale * mv.x + 127 + (scale * mv.x < 0)) >> 8);
> int mvy = x265_clip3(-32768, 32767, (scale * mv.y + 127 + (scale * mv.y < 0)) >> 8);
>
>- return MV((int16_t)mvx, (int16_t)mvy);
>+ return MV((int32_t)mvx, (int32_t)mvy);
> }
>
> }
>@@ -1917,11 +1917,11 @@
> const uint32_t mvshift = 2;
> uint32_t offset = 8;
>
>- int16_t xmax = (int16_t)((m_slice->m_sps->picWidthInLumaSamples + offset - m_cuPelX - 1) << mvshift);
>- int16_t xmin = -(int16_t)((m_encData->m_param->maxCUSize + offset + m_cuPelX - 1) << mvshift);
>+ int32_t xmax = (int32_t)((m_slice->m_sps->picWidthInLumaSamples + offset - m_cuPelX - 1) << mvshift);
>+ int32_t xmin = -(int32_t)((m_encData->m_param->maxCUSize + offset + m_cuPelX - 1) << mvshift);
>
>- int16_t ymax = (int16_t)((m_slice->m_sps->picHeightInLumaSamples + offset - m_cuPelY - 1) << mvshift);
>- int16_t ymin = -(int16_t)((m_encData->m_param->maxCUSize + offset + m_cuPelY - 1) << mvshift);
>+ int32_t ymax = (int32_t)((m_slice->m_sps->picHeightInLumaSamples + offset - m_cuPelY - 1) << mvshift);
>+ int32_t ymin = -(int32_t)((m_encData->m_param->maxCUSize + offset + m_cuPelY - 1) << mvshift);
>
> outMV.x = X265_MIN(xmax, X265_MAX(xmin, outMV.x));
> outMV.y = X265_MIN(ymax, X265_MAX(ymin, outMV.y));
>diff -r 3e5032228123 -r 0018ca1ca42f source/common/mv.h
>--- a/source/common/mv.h Fri Mar 22 13:09:49 2019 +0530
>+++ b/source/common/mv.h Fri Mar 29 09:20:41 2019 +0900
>@@ -39,16 +39,16 @@
> public:
>
> union {
>- struct { int16_t x, y; };
>+ struct { int32_t x, y; };
>
>- int32_t word;
>+ int64_t word;
> };
>
> MV() {}
>- MV(int32_t w) : word(w) {}
>- MV(int16_t _x, int16_t _y) : x(_x), y(_y) {}
>+ MV(int64_t w) : word(w) {}
>+ MV(int32_t _x, int32_t _y) : x(_x), y(_y) {}
>
>- MV& operator =(uint32_t w) { word = w; return *this; }
>+ MV& operator =(uint64_t w) { word = w; return *this; }
>
> MV& operator +=(const MV& other) { x += other.x; y += other.y; return *this; }
>
>@@ -67,7 +67,7 @@
>
> MV operator >>(int i) const { return MV(x >> i, y >> i); }
>
>- MV operator *(int16_t i) const { return MV(x * i, y * i); }
>+ MV operator *(int32_t i) const { return MV(x * i, y * i); }
>
> MV operator -(const MV& other) const { return MV(x - other.x, y - other.y); }
>
>@@ -87,7 +87,7 @@
>
> bool inline notZero() const { return this->word != 0; }
>
>- bool inline isSubpel() const { return (this->word & 0x00030003) != 0; }
>+ bool inline isSubpel() const { return (this->word & 0x0000000300000003ll) != 0; }
>
> MV mvmin(const MV& m) const { return MV(x > m.x ? m.x : x, y > m.y ? m.y : y); }
>
>diff -r 3e5032228123 -r 0018ca1ca42f source/encoder/encoder.cpp
>--- a/source/encoder/encoder.cpp Fri Mar 22 13:09:49 2019 +0530
>+++ b/source/encoder/encoder.cpp Fri Mar 29 09:20:41 2019 +0900
>@@ -3631,8 +3631,8 @@
> (analysis->interData)->refIdx[i][count + pu] = refIdx[i][d];
> if (m_param->scaleFactor)
> {
>- mv[i][d].x *= (int16_t)m_param->scaleFactor;
>- mv[i][d].y *= (int16_t)m_param->scaleFactor;
>+ mv[i][d].x *= (int32_t)m_param->scaleFactor;
>+ mv[i][d].y *= (int32_t)m_param->scaleFactor;
> }
> memcpy(&(analysis->interData)->mv[i][count + pu], &mv[i][d], sizeof(MV));
> }
>@@ -4014,8 +4014,8 @@
> {
> (analysis->interData)->mvpIdx[i][count + pu] = mvpIdx[i][d];
> (analysis->interData)->refIdx[i][count + pu] = refIdx[i][d];
>- mvCopy[i].x = mv[i][d].x * (int16_t)m_param->scaleFactor;
>- mvCopy[i].y = mv[i][d].y * (int16_t)m_param->scaleFactor;
>+ mvCopy[i].x = mv[i][d].x * (int32_t)m_param->scaleFactor;
>+ mvCopy[i].y = mv[i][d].y * (int32_t)m_param->scaleFactor;
> memcpy(&(analysis->interData)->mv[i][count + pu], &mvCopy[i], sizeof(MV));
> }
> }
>diff -r 3e5032228123 -r 0018ca1ca42f source/encoder/frameencoder.cpp
>--- a/source/encoder/frameencoder.cpp Fri Mar 22 13:09:49 2019 +0530
>+++ b/source/encoder/frameencoder.cpp Fri Mar 29 09:20:41 2019 +0900
>@@ -1406,8 +1406,8 @@
> }
>
> // Initialize restrict on MV range in slices
>- tld.analysis.m_sliceMinY = -(int16_t)(rowInSlice * m_param->maxCUSize * 4) + 3 * 4;
>- tld.analysis.m_sliceMaxY = (int16_t)((endRowInSlicePlus1 - 1 - row) * (m_param->maxCUSize * 4) - 4 * 4);
>+ tld.analysis.m_sliceMinY = -(int32_t)(rowInSlice * m_param->maxCUSize * 4) + 3 * 4;
>+ tld.analysis.m_sliceMaxY = (int32_t)((endRowInSlicePlus1 - 1 - row) * (m_param->maxCUSize * 4) - 4 * 4);
>
> // Handle single row slice
> if (tld.analysis.m_sliceMaxY < tld.analysis.m_sliceMinY)
>diff -r 3e5032228123 -r 0018ca1ca42f source/encoder/motion.cpp
>--- a/source/encoder/motion.cpp Fri Mar 22 13:09:49 2019 +0530
>+++ b/source/encoder/motion.cpp Fri Mar 29 09:20:41 2019 +0900
>@@ -382,10 +382,10 @@
> 4 * 5
> 7
> */
>- const int16_t top = omv.y - dist;
>- const int16_t bottom = omv.y + dist;
>- const int16_t left = omv.x - dist;
>- const int16_t right = omv.x + dist;
>+ const int32_t top = omv.y - dist;
>+ const int32_t bottom = omv.y + dist;
>+ const int32_t left = omv.x - dist;
>+ const int32_t right = omv.x + dist;
>
> if (top >= mvmin.y && left >= mvmin.x && right <= mvmax.x && bottom <= mvmax.y)
> {
>@@ -430,14 +430,14 @@
> Points 2, 4, 5, 7 are dist
> Points 1, 3, 6, 8 are dist>>1
> */
>- const int16_t top = omv.y - dist;
>- const int16_t bottom = omv.y + dist;
>- const int16_t left = omv.x - dist;
>- const int16_t right = omv.x + dist;
>- const int16_t top2 = omv.y - (dist >> 1);
>- const int16_t bottom2 = omv.y + (dist >> 1);
>- const int16_t left2 = omv.x - (dist >> 1);
>- const int16_t right2 = omv.x + (dist >> 1);
>+ const int32_t top = omv.y - dist;
>+ const int32_t bottom = omv.y + dist;
>+ const int32_t left = omv.x - dist;
>+ const int32_t right = omv.x + dist;
>+ const int32_t top2 = omv.y - (dist >> 1);
>+ const int32_t bottom2 = omv.y + (dist >> 1);
>+ const int32_t left2 = omv.x - (dist >> 1);
>+ const int32_t right2 = omv.x + (dist >> 1);
> saved = bcost;
>
> if (top >= mvmin.y && left >= mvmin.x &&
>@@ -502,10 +502,10 @@
>
> for (int16_t dist = 16; dist <= (int16_t)merange; dist <<= 1)
> {
>- const int16_t top = omv.y - dist;
>- const int16_t bottom = omv.y + dist;
>- const int16_t left = omv.x - dist;
>- const int16_t right = omv.x + dist;
>+ const int32_t top = omv.y - dist;
>+ const int32_t bottom = omv.y + dist;
>+ const int32_t left = omv.x - dist;
>+ const int32_t right = omv.x + dist;
>
> saved = bcost;
> if (top >= mvmin.y && left >= mvmin.x &&
>@@ -530,10 +530,10 @@
>
> for (int16_t index = 1; index < 4; index++)
> {
>- int16_t posYT = top + ((dist >> 2) * index);
>- int16_t posYB = bottom - ((dist >> 2) * index);
>- int16_t posXL = omv.x - ((dist >> 2) * index);
>- int16_t posXR = omv.x + ((dist >> 2) * index);
>+ int32_t posYT = top + ((dist >> 2) * index);
>+ int32_t posYB = bottom - ((dist >> 2) * index);
>+ int32_t posXL = omv.x - ((dist >> 2) * index);
>+ int32_t posXR = omv.x + ((dist >> 2) * index);
>
> COST_MV_PT_DIST_X4(posXL, posYT, 0, dist,
> posXR, posYT, 0, dist,
>@@ -561,10 +561,10 @@
> }
> for (int16_t index = 1; index < 4; index++)
> {
>- int16_t posYT = top + ((dist >> 2) * index);
>- int16_t posYB = bottom - ((dist >> 2) * index);
>- int16_t posXL = omv.x - ((dist >> 2) * index);
>- int16_t posXR = omv.x + ((dist >> 2) * index);
>+ int32_t posYT = top + ((dist >> 2) * index);
>+ int32_t posYB = bottom - ((dist >> 2) * index);
>+ int32_t posXL = omv.x - ((dist >> 2) * index);
>+ int32_t posXR = omv.x + ((dist >> 2) * index);
>
> if (posYT >= mvmin.y) // check top
> {
>@@ -1235,10 +1235,10 @@
> case X265_SEA:
> {
> // Successive Elimination Algorithm
>- const int16_t minX = X265_MAX(omv.x - (int16_t)merange, mvmin.x);
>- const int16_t minY = X265_MAX(omv.y - (int16_t)merange, mvmin.y);
>- const int16_t maxX = X265_MIN(omv.x + (int16_t)merange, mvmax.x);
>- const int16_t maxY = X265_MIN(omv.y + (int16_t)merange, mvmax.y);
>+ const int32_t minX = X265_MAX(omv.x - (int32_t)merange, mvmin.x);
>+ const int32_t minY = X265_MAX(omv.y - (int32_t)merange, mvmin.y);
>+ const int32_t maxX = X265_MIN(omv.x + (int32_t)merange, mvmax.x);
>+ const int32_t maxY = X265_MIN(omv.y + (int32_t)merange, mvmax.y);
> const uint16_t *p_cost_mvx = m_cost_mvx - qmvp.x;
> const uint16_t *p_cost_mvy = m_cost_mvy - qmvp.y;
> int16_t* meScratchBuffer = NULL;
>diff -r 3e5032228123 -r 0018ca1ca42f source/encoder/search.cpp
>--- a/source/encoder/search.cpp Fri Mar 22 13:09:49 2019 +0530
>+++ b/source/encoder/search.cpp Fri Mar 29 09:20:41 2019 +0900
>@@ -2633,7 +2633,7 @@
>
> void Search::setSearchRange(const CUData& cu, const MV& mvp, int merange, MV& mvmin, MV& mvmax) const
> {
>- MV dist((int16_t)merange << 2, (int16_t)merange << 2);
>+ MV dist((int32_t)merange << 2, (int32_t)merange << 2);
> mvmin = mvp - dist;
> mvmax = mvp + dist;
>
>@@ -2670,8 +2670,8 @@
> mvmax >>= 2;
>
> /* conditional clipping for frame parallelism */
>- mvmin.y = X265_MIN(mvmin.y, (int16_t)m_refLagPixels);
>- mvmax.y = X265_MIN(mvmax.y, (int16_t)m_refLagPixels);
>+ mvmin.y = X265_MIN(mvmin.y, (int32_t)m_refLagPixels);
>+ mvmax.y = X265_MIN(mvmax.y, (int32_t)m_refLagPixels);
>
> /* conditional clipping for negative mv range */
> mvmax.y = X265_MAX(mvmax.y, mvmin.y);
>diff -r 3e5032228123 -r 0018ca1ca42f source/encoder/search.h
>--- a/source/encoder/search.h Fri Mar 22 13:09:49 2019 +0530
>+++ b/source/encoder/search.h Fri Mar 29 09:20:41 2019 +0900
>@@ -283,8 +283,8 @@
> int32_t m_maxTUDepth;
> uint16_t m_limitTU;
>
>- int16_t m_sliceMaxY;
>- int16_t m_sliceMinY;
>+ int32_t m_sliceMaxY;
>+ int32_t m_sliceMinY;
>
> #if DETAILED_CU_STATS
> /* Accumulate CU statistics separately for each frame encoder */
>diff -r 3e5032228123 -r 0018ca1ca42f source/encoder/slicetype.cpp
>--- a/source/encoder/slicetype.cpp Fri Mar 22 13:09:49 2019 +0530
>+++ b/source/encoder/slicetype.cpp Fri Mar 29 09:20:41 2019 +0900
>@@ -2860,10 +2860,10 @@
>
> // TODO: restrict to slices boundaries
> // establish search bounds that don't cross extended frame boundaries
>- mvmin.x = (int16_t)(-cuX * cuSize - 8);
>- mvmin.y = (int16_t)(-cuY * cuSize - 8);
>- mvmax.x = (int16_t)((widthInCU - cuX - 1) * cuSize + 8);
>- mvmax.y = (int16_t)((heightInCU - cuY - 1) * cuSize + 8);
>+ mvmin.x = (int32_t)(-cuX * cuSize - 8);
>+ mvmin.y = (int32_t)(-cuY * cuSize - 8);
>+ mvmax.x = (int32_t)((widthInCU - cuX - 1) * cuSize + 8);
>+ mvmax.y = (int32_t)((heightInCU - cuY - 1) * cuSize + 8);
>
> for (int i = 0; i < 1 + bBidir; i++)
> {
>diff -r 3e5032228123 -r 0018ca1ca42f source/encoder/weightPrediction.cpp
>--- a/source/encoder/weightPrediction.cpp Fri Mar 22 13:09:49 2019 +0530
>+++ b/source/encoder/weightPrediction.cpp Fri Mar 29 09:20:41 2019 +0900
>@@ -69,15 +69,15 @@
> for (int y = 0; y < ref.lines; y += cuSize)
> {
> intptr_t pixoff = y * stride;
>- mvmin.y = (int16_t)((-y - 8) * mvshift);
>- mvmax.y = (int16_t)((ref.lines - y - 1 + 8) * mvshift);
>+ mvmin.y = (int32_t)((-y - 8) * mvshift);
>+ mvmax.y = (int32_t)((ref.lines - y - 1 + 8) * mvshift);
>
> for (int x = 0; x < ref.width; x += cuSize, pixoff += cuSize, cu++)
> {
> ALIGN_VAR_16(pixel, buf8x8[8 * 8]);
> intptr_t bstride = 8;
>- mvmin.x = (int16_t)((-x - 8) * mvshift);
>- mvmax.x = (int16_t)((ref.width - x - 1 + 8) * mvshift);
>+ mvmin.x = (int32_t)((-x - 8) * mvshift);
>+ mvmax.x = (int32_t)((ref.width - x - 1 + 8) * mvshift);
>
> /* clip MV to available pixels */
> MV mv = mvs[cu];
>@@ -113,8 +113,8 @@
> * into the lowres structures */
> int cu = y * cache.lowresWidthInCU;
> intptr_t pixoff = y * stride;
>- mvmin.y = (int16_t)((-y - 8) * mvshift);
>- mvmax.y = (int16_t)((height - y - 1 + 8) * mvshift);
>+ mvmin.y = (int32_t)((-y - 8) * mvshift);
>+ mvmax.y = (int32_t)((height - y - 1 + 8) * mvshift);
>
> for (int x = 0; x < width; x += bw, cu++, pixoff += bw)
> {
>@@ -126,8 +126,8 @@
> mv.y >>= cache.vshift;
>
> /* clip MV to available pixels */
>- mvmin.x = (int16_t)((-x - 8) * mvshift);
>- mvmax.x = (int16_t)((width - x - 1 + 8) * mvshift);
>+ mvmin.x = (int32_t)((-x - 8) * mvshift);
>+ mvmax.x = (int32_t)((width - x - 1 + 8) * mvshift);
> mv = mv.clipped(mvmin, mvmax);
>
> intptr_t fpeloffset = (mv.y >> 2) * stride + (mv.x >> 2);
>diff -r 3e5032228123 -r 0018ca1ca42f source/x265.h
>--- a/source/x265.h Fri Mar 22 13:09:49 2019 +0530
>+++ b/source/x265.h Fri Mar 29 09:20:41 2019 +0900
>@@ -147,9 +147,9 @@
> typedef struct x265_analysis_MV
> {
> union{
>- struct { int16_t x, y; };
>+ struct { int32_t x, y; };
>
>- int32_t word;
>+ int64_t word;
> };
> }x265_analysis_MV;
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.videolan.org/pipermail/x265-devel/attachments/20190329/27f191ae/attachment-0001.html>
More information about the x265-devel
mailing list