[x265] [PATCH]Add: Auto AQ Mode
Nomis101
Nomis101 at web.de
Wed Apr 15 17:59:02 CEST 2020
What did happen with this auto-aq idea? Is it planned to pursue this idea further? I quite liked this idea and it gives me a nice reduction in file size.
Thanks and regards.
Am 27.02.20 um 09:59 schrieb Niranjan Bala:
> # HG changeset patch
> # User Niranjan <niranjan at multicorewareinc.com <mailto:niranjan at multicorewareinc.com>>
> # Date 1582622787 -19800
> # Tue Feb 25 14:56:27 2020 +0530
> # Node ID 2ec7d27dc576cd993fd0bfe4b116514e8f2ddc0c
> # Parent 1a47c3802fa38bdb4606e3d8203b73b4474d1ce6
> Add: Auto AQ Mode
>
> This patch does the following:
> 1. Automatically decides the AQ Mode for each frame, using its scene statistics,
> such as luma intensity and edge density.
> 2. Add option "--auto-aq" to enable auto detection of AQ Mode per frame.
>
> diff -r 1a47c3802fa3 -r 2ec7d27dc576 doc/reST/cli.rst
> --- a/doc/reST/cli.rst Mon Feb 24 12:51:41 2020 +0530
> +++ b/doc/reST/cli.rst Tue Feb 25 14:56:27 2020 +0530
> @@ -1721,6 +1721,12 @@
> 5. Same as AQ mode 3, but uses edge density instead of auto-variance.
> i.e, AQ with bias towards dark scenes which have high edge density.
>
> +.. option:: --auto-aq --no-auto-aq
> +
> + To enable and disable automatic AQ mode detection per frame.
> + This option adaptively sets the AQ mode for each frame between 2, 3, 4 and 5 based on the scene statistics.
> + Default: disabled.
> +
> .. option:: --aq-strength <float>
>
> Adjust the strength of the adaptive quantization offsets. Setting
> diff -r 1a47c3802fa3 -r 2ec7d27dc576 source/CMakeLists.txt
> --- a/source/CMakeLists.txt Mon Feb 24 12:51:41 2020 +0530
> +++ b/source/CMakeLists.txt Tue Feb 25 14:56:27 2020 +0530
> @@ -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 189)
> +set(X265_BUILD 190)
> configure_file("${PROJECT_SOURCE_DIR}/x265.def.in <http://x265.def.in>"
> "${PROJECT_BINARY_DIR}/x265.def")
> configure_file("${PROJECT_SOURCE_DIR}/x265_config.h.in <http://x265_config.h.in>"
> diff -r 1a47c3802fa3 -r 2ec7d27dc576 source/common/common.h
> --- a/source/common/common.h Mon Feb 24 12:51:41 2020 +0530
> +++ b/source/common/common.h Tue Feb 25 14:56:27 2020 +0530
> @@ -131,6 +131,7 @@
> typedef int64_t ssum2_t;
> #define SHIFT_TO_BITPLANE 9
> #define HISTOGRAM_BINS 1024
> +#define BRIGHTNESS_THRESHOLD 120 // The threshold above which a pixel is bright
> #else
> typedef uint8_t pixel;
> typedef uint16_t sum_t;
> @@ -139,6 +140,7 @@
> typedef int32_t ssum2_t; // Signed sum
> #define SHIFT_TO_BITPLANE 7
> #define HISTOGRAM_BINS 256
> +#define BRIGHTNESS_THRESHOLD 30 // The threshold above which a pixel is bright
> #endif // if HIGH_BIT_DEPTH
>
> #if X265_DEPTH < 10
> @@ -163,6 +165,8 @@
> #define MIN_QPSCALE 0.21249999999999999
> #define MAX_MAX_QPSCALE 615.46574234477100
>
> +#define FRAME_BRIGHTNESS_THRESHOLD 50.0 // Min % of pixels in a frame, that are above BRIGHTNESS_THRESHOLD for it to be considered a bright frame
> +#define FRAME_EDGE_THRESHOLD 10.0 // Min % of edge pixels in a frame, for it to be considered to have high edge density
>
> template<typename T>
> inline T x265_min(T a, T b) { return a < b ? a : b; }
> diff -r 1a47c3802fa3 -r 2ec7d27dc576 source/common/frame.cpp
> --- a/source/common/frame.cpp Mon Feb 24 12:51:41 2020 +0530
> +++ b/source/common/frame.cpp Tue Feb 25 14:56:27 2020 +0530
> @@ -63,6 +63,7 @@
> m_thetaPic = NULL;
> m_edgeBitPlane = NULL;
> m_edgeBitPic = NULL;
> + m_frameAq = X265_AQ_NONE;
> }
>
> bool Frame::create(x265_param *param, float* quantOffsets)
> @@ -103,7 +104,7 @@
> CHECKED_MALLOC_ZERO(m_classifyCount, uint32_t, size);
> }
>
> - if (param->rc.aqMode == X265_AQ_EDGE || param->rc.aqMode == X265_AQ_EDGE_BIASED || (param->rc.zonefileCount && param->rc.aqMode != 0))
> + if (param->rc.aqMode == X265_AQ_EDGE || param->rc.aqMode == X265_AQ_EDGE_BIASED || param->rc.bAutoAq || (param->rc.zonefileCount && param->rc.aqMode != 0))
> {
> uint32_t numCuInWidth = (param->sourceWidth + param->maxCUSize - 1) / param->maxCUSize;
> uint32_t numCuInHeight = (param->sourceHeight + param->maxCUSize - 1) / param->maxCUSize;
> diff -r 1a47c3802fa3 -r 2ec7d27dc576 source/common/frame.h
> --- a/source/common/frame.h Mon Feb 24 12:51:41 2020 +0530
> +++ b/source/common/frame.h Tue Feb 25 14:56:27 2020 +0530
> @@ -141,6 +141,9 @@
> pixel* m_edgeBitPlane;
> pixel* m_edgeBitPic;
>
> + /* AQ mode for each frame */
> + int m_frameAq;
> +
> Frame();
>
> bool create(x265_param *param, float* quantOffsets);
> diff -r 1a47c3802fa3 -r 2ec7d27dc576 source/common/lowres.cpp
> --- a/source/common/lowres.cpp Mon Feb 24 12:51:41 2020 +0530
> +++ b/source/common/lowres.cpp Tue Feb 25 14:56:27 2020 +0530
> @@ -190,6 +190,9 @@
> }
> }
>
> + if (param->rc.bAutoAq)
> + lowresEdgePlane = X265_MALLOC(pixel, lumaStride * (lines + (origPic->m_lumaMarginY * 2)));
> +
> return true;
>
> fail:
> @@ -235,6 +238,7 @@
> X265_FREE(edgeInclined);
> X265_FREE(qpAqMotionOffset);
> X265_FREE(blockVariance);
> + X265_FREE(lowresEdgePlane);
> if (maxAQDepth > 0)
> {
> for (uint32_t d = 0; d < 4; d++)
> diff -r 1a47c3802fa3 -r 2ec7d27dc576 source/common/lowres.h
> --- a/source/common/lowres.h Mon Feb 24 12:51:41 2020 +0530
> +++ b/source/common/lowres.h Tue Feb 25 14:56:27 2020 +0530
> @@ -44,6 +44,9 @@
> pixel* fpelLowerResPlane[3];
> pixel* lowerResPlane[4];
>
> + /* Edge Plane in Lowres */
> + pixel* lowresEdgePlane;
> +
> bool isWeighted;
> bool isLowres;
> bool isHMELowres;
> diff -r 1a47c3802fa3 -r 2ec7d27dc576 source/common/param.cpp
> --- a/source/common/param.cpp Mon Feb 24 12:51:41 2020 +0530
> +++ b/source/common/param.cpp Tue Feb 25 14:56:27 2020 +0530
> @@ -286,6 +286,7 @@
> param->rc.bEnableConstVbv = 0;
> param->bResetZoneConfig = 1;
> param->reconfigWindowSize = 0;
> + param->rc.bAutoAq = 0;
>
> /* Video Usability Information (VUI) */
> param->vui.aspectRatioIdc = 0;
> @@ -1229,6 +1230,7 @@
> OPT("multi-pass-opt-analysis") p->analysisMultiPassRefine = atobool(value);
> OPT("multi-pass-opt-distortion") p->analysisMultiPassDistortion = atobool(value);
> OPT("aq-motion") p->bAQMotion = atobool(value);
> + OPT("auto-aq") p->rc.bAutoAq = atobool(value);
> OPT("dynamic-rd") p->dynamicRd = atof(value);
> OPT("analysis-reuse-level")
> {
> @@ -1877,9 +1879,12 @@
> param->maxNumReferences, (param->limitReferences & X265_REF_LIMIT_CU) ? "on" : "off",
> (param->limitReferences & X265_REF_LIMIT_DEPTH) ? "on" : "off");
>
> - if (param->rc.aqMode)
> + if (param->rc.aqMode && !param->rc.bAutoAq)
> x265_log(param, X265_LOG_INFO, "AQ: mode / str / qg-size / cu-tree : %d / %0.1f / %d / %d\n", param->rc.aqMode,
> param->rc.aqStrength, param->rc.qgSize, param->rc.cuTree);
> + else if (param->rc.bAutoAq)
> + x265_log(param, X265_LOG_INFO, "AQ: mode / str / qg-size / cu-tree : auto / %0.1f / %d / %d\n", param->rc.aqStrength,
> + param->rc.qgSize, param->rc.cuTree);
>
> if (param->bLossless)
> x265_log(param, X265_LOG_INFO, "Rate Control : Lossless\n");
> @@ -2187,6 +2192,7 @@
> s += sprintf(s, " hist-threshold=%.2f", p->edgeTransitionThreshold);
> BOOL(p->bOptCUDeltaQP, "opt-cu-delta-qp");
> BOOL(p->bAQMotion, "aq-motion");
> + BOOL(p->rc.bAutoAq, "auto-aq");
> BOOL(p->bEmitHDR10SEI, "hdr10");
> BOOL(p->bHDR10Opt, "hdr10-opt");
> BOOL(p->bDhdr10opt, "dhdr10-opt");
> @@ -2468,6 +2474,7 @@
> dst->rc.bEnableConstVbv = src->rc.bEnableConstVbv;
> dst->rc.hevcAq = src->rc.hevcAq;
> dst->rc.qpAdaptationRange = src->rc.qpAdaptationRange;
> + dst->rc.bAutoAq = src->rc.bAutoAq;
>
> dst->vui.aspectRatioIdc = src->vui.aspectRatioIdc;
> dst->vui.sarWidth = src->vui.sarWidth;
> diff -r 1a47c3802fa3 -r 2ec7d27dc576 source/encoder/slicetype.cpp
> --- a/source/encoder/slicetype.cpp Mon Feb 24 12:51:41 2020 +0530
> +++ b/source/encoder/slicetype.cpp Tue Feb 25 14:56:27 2020 +0530
> @@ -473,9 +473,9 @@
> if (!(param->rc.bStatRead && param->rc.cuTree && IS_REFERENCED(curFrame)))
> {
> /* Calculate Qp offset for each 16x16 or 8x8 block in the frame */
> - if (param->rc.aqMode == X265_AQ_NONE || param->rc.aqStrength == 0)
> + if (curFrame->m_frameAq == X265_AQ_NONE || param->rc.aqStrength == 0)
> {
> - if (param->rc.aqMode && param->rc.aqStrength == 0)
> + if (curFrame->m_frameAq && param->rc.aqStrength == 0)
> {
> if (quantOffsets)
> {
> @@ -516,18 +516,18 @@
> double bias_strength = 0.f;
> double strength = 0.f;
>
> - if (param->rc.aqMode == X265_AQ_EDGE || param->rc.aqMode == X265_AQ_EDGE_BIASED)
> + if (curFrame->m_frameAq == X265_AQ_EDGE || curFrame->m_frameAq == X265_AQ_EDGE_BIASED)
> edgeFilter(curFrame, param);
>
> - if (param->rc.aqMode == X265_AQ_EDGE && !param->bHistBasedSceneCut && param->enableRecursionSkip >= EDGE_BASED_RSKIP)
> + if (curFrame->m_frameAq == X265_AQ_EDGE && !param->bHistBasedSceneCut && param->enableRecursionSkip >= EDGE_BASED_RSKIP)
> {
> pixel* src = curFrame->m_edgePic + curFrame->m_fencPic->m_lumaMarginY * curFrame->m_fencPic->m_stride + curFrame->m_fencPic->m_lumaMarginX;
> primitives.planecopy_pp_shr(src, curFrame->m_fencPic->m_stride, curFrame->m_edgeBitPic,
> curFrame->m_fencPic->m_stride, curFrame->m_fencPic->m_picWidth, curFrame->m_fencPic->m_picHeight, SHIFT_TO_BITPLANE);
> }
>
> - if (param->rc.aqMode == X265_AQ_AUTO_VARIANCE || param->rc.aqMode == X265_AQ_AUTO_VARIANCE_BIASED ||
> - param->rc.aqMode == X265_AQ_EDGE || param->rc.aqMode == X265_AQ_EDGE_BIASED)
> + if (curFrame->m_frameAq == X265_AQ_AUTO_VARIANCE || curFrame->m_frameAq == X265_AQ_AUTO_VARIANCE_BIASED ||
> + curFrame->m_frameAq == X265_AQ_EDGE || curFrame->m_frameAq == X265_AQ_EDGE_BIASED)
> {
> double bit_depth_correction = 1.f / (1 << (2 * (X265_DEPTH - 8)));
> for (int blockY = 0; blockY < maxRow; blockY += loopIncr)
> @@ -536,7 +536,7 @@
> {
> uint32_t energy, edgeDensity, avgAngle;
> energy = acEnergyCu(curFrame, blockX, blockY, param->internalCsp, param->rc.qgSize);
> - if (param->rc.aqMode == X265_AQ_EDGE || param->rc.aqMode == X265_AQ_EDGE_BIASED)
> + if (curFrame->m_frameAq == X265_AQ_EDGE || curFrame->m_frameAq == X265_AQ_EDGE_BIASED)
> {
> edgeDensity = edgeDensityCu(curFrame, avgAngle, blockX, blockY, param->rc.qgSize);
> if (edgeDensity)
> @@ -576,17 +576,17 @@
> {
> for (int blockX = 0; blockX < maxCol; blockX += loopIncr)
> {
> - if (param->rc.aqMode == X265_AQ_AUTO_VARIANCE_BIASED)
> + if (curFrame->m_frameAq == X265_AQ_AUTO_VARIANCE_BIASED)
> {
> qp_adj = curFrame->m_lowres.qpCuTreeOffset[blockXY];
> qp_adj = strength * (qp_adj - avg_adj) + bias_strength * (1.f - modeTwoConst / (qp_adj * qp_adj));
> }
> - else if (param->rc.aqMode == X265_AQ_AUTO_VARIANCE)
> + else if (curFrame->m_frameAq == X265_AQ_AUTO_VARIANCE)
> {
> qp_adj = curFrame->m_lowres.qpCuTreeOffset[blockXY];
> qp_adj = strength * (qp_adj - avg_adj);
> }
> - else if (param->rc.aqMode == X265_AQ_EDGE)
> + else if (curFrame->m_frameAq == X265_AQ_EDGE)
> {
> inclinedEdge = curFrame->m_lowres.edgeInclined[blockXY];
> qp_adj = curFrame->m_lowres.qpCuTreeOffset[blockXY];
> @@ -595,7 +595,7 @@
> else
> qp_adj = strength * (qp_adj - avg_adj);
> }
> - else if (param->rc.aqMode == X265_AQ_EDGE_BIASED)
> + else if (curFrame->m_frameAq == X265_AQ_EDGE_BIASED)
> {
> inclinedEdge = curFrame->m_lowres.edgeInclined[blockXY];
> qp_adj = curFrame->m_lowres.qpCuTreeOffset[blockXY];
> @@ -1387,6 +1387,44 @@
> }
> }
>
> +double computeBrightnessIntensity(pixel *inPlane, int width, int height, intptr_t stride)
> +{
> + pixel* rowStart = inPlane;
> + double count = 0;
> +
> + for (int i = 0; i < height; i++)
> + {
> + for (int j = 0; j < width; j++)
> + {
> + if (rowStart[j] > BRIGHTNESS_THRESHOLD)
> + count++;
> + }
> + rowStart += stride;
> + }
> +
> + /* Returns the brightness percentage of the input plane */
> + return (count / (width * height)) * 100;
> +}
> +
> +double computeEdgeIntensity(pixel *inPlane, int width, int height, intptr_t stride)
> +{
> + pixel* rowStart = inPlane;
> + double count = 0;
> +
> + for (int i = 0; i < height; i++)
> + {
> + for (int j = 0; j < width; j++)
> + {
> + if (rowStart[j] > 0)
> + count++;
> + }
> + rowStart += stride;
> + }
> +
> + /* Returns the edge percentage of the input plane */
> + return (count / (width * height)) * 100;
> +}
> +
> void PreLookaheadGroup::processTasks(int workerThreadID)
> {
> if (workerThreadID < 0)
> @@ -1401,6 +1439,36 @@
> ProfileScopeEvent(prelookahead);
> m_lock.release();
> preFrame->m_lowres.init(preFrame->m_fencPic, preFrame->m_poc);
> +
> + /* Auto AQ */
> + if (preFrame->m_param->rc.bAutoAq)
> + {
> + int heightL = preFrame->m_lowres.lines;
> + int widthL = preFrame->m_lowres.width;
> + pixel *lumaPlane = preFrame->m_lowres.fpelPlane[0];
> + intptr_t stride = preFrame->m_lowres.lumaStride;
> + double brightnessIntensity = 0, edgeIntensity = 0;
> +
> + /* Edge plane computation */
> + memset(preFrame->m_lowres.lowresEdgePlane, 0, stride * (heightL + (preFrame->m_fencPic->m_lumaMarginY * 2)) * sizeof(pixel));
> + pixel* lowresEdgePic = preFrame->m_lowres.lowresEdgePlane + preFrame->m_fencPic->m_lumaMarginY * stride + preFrame->m_fencPic->m_lumaMarginX;
> + computeEdge(lowresEdgePic, lumaPlane, NULL, stride, heightL, widthL, false);
> +
> + /*Frame edge percentage computation */
> + edgeIntensity = computeEdgeIntensity(lowresEdgePic, widthL, heightL, stride);
> +
> + /* Frame Brightness percentage computation */
> + brightnessIntensity = computeBrightnessIntensity(lumaPlane, widthL, heightL, stride);
> +
> + /* AQ mode switch */
> + if (edgeIntensity < FRAME_EDGE_THRESHOLD)
> + preFrame->m_frameAq = brightnessIntensity > FRAME_BRIGHTNESS_THRESHOLD? X265_AQ_AUTO_VARIANCE : X265_AQ_AUTO_VARIANCE_BIASED;
> + else
> + preFrame->m_frameAq = brightnessIntensity > FRAME_BRIGHTNESS_THRESHOLD? X265_AQ_EDGE : X265_AQ_EDGE_BIASED;
> + }
> + else
> + preFrame->m_frameAq = preFrame->m_param->rc.aqMode;
> +
> if (m_lookahead.m_bAdaptiveQuant)
> tld.calcAdaptiveQuantFrame(preFrame, m_lookahead.m_param);
> tld.lowresIntraEstimate(preFrame->m_lowres, m_lookahead.m_param->rc.qgSize);
> diff -r 1a47c3802fa3 -r 2ec7d27dc576 source/test/regression-tests.txt
> --- a/source/test/regression-tests.txt Mon Feb 24 12:51:41 2020 +0530
> +++ b/source/test/regression-tests.txt Tue Feb 25 14:56:27 2020 +0530
> @@ -171,6 +171,7 @@
> crowd_run_1080p50.yuv, --preset slow --ctu 32 --rskip 3 --rskip-edge-threshold 5 --hist-scenecut --hist-threshold 0.1
> crowd_run_1080p50.yuv, --preset slower --ctu 16 --rskip 3 --rskip-edge-threshold 5 --hist-scenecut --hist-threshold 0.1 --aq-mode 4
> sintel_trailer_2k_1920x1080_24.yuv, --preset medium --aq-mode 5
> +ducks_take_off_420_720p50.y4m, --preset slow --auto-aq --aq-strength 1.5 --aq-motion
>
> # Main12 intraCost overflow bug test
> 720p50_parkrun_ter.y4m,--preset medium
> diff -r 1a47c3802fa3 -r 2ec7d27dc576 source/x265.h
> --- a/source/x265.h Mon Feb 24 12:51:41 2020 +0530
> +++ b/source/x265.h Tue Feb 25 14:56:27 2020 +0530
> @@ -1477,6 +1477,9 @@
> /* internally enable if tune grain is set */
> int bEnableConstVbv;
>
> + /* automatically switch AQ mode for each frame */
> + int bAutoAq;
> +
> } rc;
>
> /*== Video Usability Information ==*/
> diff -r 1a47c3802fa3 -r 2ec7d27dc576 source/x265cli.h
> --- a/source/x265cli.h Mon Feb 24 12:51:41 2020 +0530
> +++ b/source/x265cli.h Tue Feb 25 14:56:27 2020 +0530
> @@ -172,6 +172,8 @@
> { "qp", required_argument, NULL, 'q' },
> { "aq-mode", required_argument, NULL, 0 },
> { "aq-strength", required_argument, NULL, 0 },
> + { "auto-aq", no_argument, NULL, 0 },
> + { "no-auto-aq", no_argument, NULL, 0 },
> { "rc-grain", no_argument, NULL, 0 },
> { "no-rc-grain", no_argument, NULL, 0 },
> { "ipratio", required_argument, NULL, 0 },
> @@ -594,6 +596,7 @@
> H0(" --aq-strength <float> Reduces blocking and blurring in flat and textured areas (0 to 3.0). Default %.2f\n", param->rc.aqStrength);
> H0(" --qp-adaptation-range <float> Delta QP range by QP adaptation based on a psycho-visual model (1.0 to 6.0). Default %.2f\n", param->rc.qpAdaptationRange);
> H0(" --[no-]aq-motion Block level QP adaptation based on the relative motion between the block and the frame. Default %s\n", OPT(param->bAQMotion));
> + H1(" --[no-]auto-aq Automatically decides the AQ Mode for each frame, using its scene statistics, such as luma intensity and edge density. Default %s\n", OPT(param->rc.bAutoAq));
> H0(" --qg-size <int> Specifies the size of the quantization group (64, 32, 16, 8). Default %d\n", param->rc.qgSize);
> H0(" --[no-]cutree Enable cutree for Adaptive Quantization. Default %s\n", OPT(param->rc.cuTree));
> H0(" --[no-]rc-grain Enable ratecontrol mode to handle grains specifically. turned on with tune grain. Default %s\n", OPT(param->rc.bEnableGrain));
> Thanks & Regards
> *Niranjan Kumar B*
> Video Codec Engineer
> Media & AI Analytics
> +91 958 511 1449
> <https://multicorewareinc.com/>
>
> _______________________________________________
> x265-devel mailing list
> x265-devel at videolan.org
> https://mailman.videolan.org/listinfo/x265-devel
>
More information about the x265-devel
mailing list