MCSTF optimization

doc/reST/cli.rst

@@ -1358,15 +1358,33 @@ Temporal / motion search options
 
 	Motion-compensated spatio-temporal filtering (MCSTF) improves the compression
 	efficiency of videos that contain a high level of noise. It introduces a
-	temporal filter before encoding and this filter is applied only to the I- and P-frames.
+	temporal filter before encoding, and this filter is applied unconditionally
+	to every I- and P-frame regardless of how noisy the source content actually is.
 	It utilizes previously generated motion vectors across different video content
 	resolutions to find the best temporal correspondence for low-pass filtering. Here,
 	motion estimation is applied between the central picture and each future or past
 	picture, thereby generating multiple motion-compensated predictions, which are then
 	combined by using adaptive filtering to produce a final noise-reduced picture.
+	To filter frames selectively based on their estimated noise level instead, use
+	:option:`--selective-mcstf`.
 	Default: disabled
 
-	Note : MCSTF should be enabled only with frame threads 1
+.. option:: --mcstf-ref-range
+
+	Number of reference frames to consider on each side (past and future) of the
+	current frame during Motion-compensated spatio-temporal filtering. A higher value
+	includes more temporal neighbors for filtering, which can improve noise reduction
+	at the cost of increased memory usage and processing time. The valid range is 1 to 4.
+	Default: 2
+
+.. option:: --selective-mcstf, --no-selective-mcstf
+
+	Gate MCSTF with a per-GOP noise estimate: skip the filter for GOPs whose
+	starting I/IDR frame (or scenecut) is estimated as clean (low noise), and
+	apply it only where it helps. Implicitly enables :option:`--mcstf` if it
+	was not already turned on. When :option:`--csv-log-level` is 2 or higher,
+	adds a Frame Noise and IsMCSTFEnabled column to the per-frame CSV log.
+	Default: disabled
 
 Spatial/intra options
 =====================

source/CMakeLists.txt

@@ -32,7 +32,7 @@ option(NATIVE_BUILD "Target the build CPU" OFF)
 option(STATIC_LINK_CRT "Statically link C and C++ runtimes 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 216)
+set(X265_BUILD 217)
 configure_file("${PROJECT_SOURCE_DIR}/x265.def.in"
                "${PROJECT_BINARY_DIR}/x265.def")
 configure_file("${PROJECT_SOURCE_DIR}/x265_config.h.in"

source/common/CMakeLists.txt

@@ -22,9 +22,10 @@ if(ENABLE_ASSEMBLY AND X86)
     set(SSE3  vec/dct-sse3.cpp)
     set(SSSE3 vec/dct-ssse3.cpp)
     set(SSE41 vec/dct-sse41.cpp)
+    set(AVX2  vec/temporalfilter-avx2.cpp)
 
     if(MSVC)
-        set(PRIMITIVES ${SSE3} ${SSSE3} ${SSE41})
+        set(PRIMITIVES ${SSE3} ${SSSE3} ${SSE41} ${AVX2})
         set(WARNDISABLE "/wd4100") # unreferenced formal parameter
         if(INTEL_CXX)
             add_definitions(/Qwd111) # statement is unreachable
@@ -39,6 +40,7 @@ if(ENABLE_ASSEMBLY AND X86)
             # x64 implies SSE4, so only add /arch:SSE2 if building for Win32
             set_source_files_properties(${SSE3} ${SSSE3} ${SSE41} PROPERTIES COMPILE_FLAGS "${WARNDISABLE} /arch:SSE2")
         endif()
+        set_source_files_properties(${AVX2} PROPERTIES COMPILE_FLAGS "${WARNDISABLE} /arch:AVX2")
     endif()
     if(GCC)
         if(CLANG)
@@ -48,10 +50,11 @@ if(ENABLE_ASSEMBLY AND X86)
             set(WARNDISABLE "-Wno-unused-parameter")
         endif()
         if(INTEL_CXX OR CLANG OR (NOT CC_VERSION VERSION_LESS 4.3))
-            set(PRIMITIVES ${SSE3} ${SSSE3} ${SSE41})
+            set(PRIMITIVES ${SSE3} ${SSSE3} ${SSE41} ${AVX2})
             set_source_files_properties(${SSE3}  PROPERTIES COMPILE_FLAGS "${WARNDISABLE} -msse3")
             set_source_files_properties(${SSSE3} PROPERTIES COMPILE_FLAGS "${WARNDISABLE} -mssse3")
             set_source_files_properties(${SSE41} PROPERTIES COMPILE_FLAGS "${WARNDISABLE} -msse4.1")
+            set_source_files_properties(${AVX2}  PROPERTIES COMPILE_FLAGS "${WARNDISABLE} -mavx2")
         endif()
     endif()
     set(VEC_PRIMITIVES vec/vec-primitives.cpp ${PRIMITIVES})

source/common/loongarch64/asm-primitives.cpp

@@ -936,7 +936,6 @@ void setupAssemblyPrimitives(EncoderPrimitives &p, int cpuMask)
         // mc
         p.frameInitLowres    = PFX(frame_init_lowres_core_lsx);
         p.frameInitLowerRes  = PFX(frame_init_lowres_core_lsx);
-        p.frameSubSampleLuma = PFX(frame_subsample_luma_lsx);
 
         p.fix8Unpack = PFX(cutree_fix8_unpack_lsx);
         p.fix8Pack   = PFX(cutree_fix8_pack_lsx);
@@ -1333,7 +1332,6 @@ void setupAssemblyPrimitives(EncoderPrimitives &p, int cpuMask)
         // mc
         p.frameInitLowres    = PFX(frame_init_lowres_core_lasx);
         p.frameInitLowerRes  = PFX(frame_init_lowres_core_lasx);
-        p.frameSubSampleLuma = PFX(frame_subsample_luma_lasx);
 
         p.fix8Unpack = PFX(cutree_fix8_unpack_lasx);
         p.fix8Pack   = PFX(cutree_fix8_pack_lasx);

source/common/loongarch64/mc-a2.S

@@ -278,123 +278,6 @@ function x265_frame_init_lowres_core_lasx
     blt             zero,   t0,     .LOOP_HEIGHT_FRAME_INIT_LOWRES_CORE_LASX
 endfunc
 
-.macro CALCULATE_FRAMESUBSAMPLELUMA_LSX_8 offset0
-    // load inRow[offset0 ~ offset0+15], inRowBelow[offset0 ~ offset0+15]
-    vldx            vr1,    a0,     \offset0
-    vldx            vr3,    t0,     \offset0
-    vaddwev.h.bu    vr4,    vr1,    vr3
-    vaddwod.h.bu    vr5,    vr1,    vr3
-    vsrari.h        vr4,    vr4,    1
-    vsrari.h        vr5,    vr5,    1
-    vadd.h          vr6,    vr4,    vr5
-    vsrarni.b.h     vr6,    vr6,    1
-.endm
-
-/*-----------------------------------------------------------------------------
-void x265_frame_subsample_luma_lsx(const pixel* src0, pixel* dst0,
-            intptr_t src_stride, intptr_t dst_stride, int width, int height)
------------------------------------------------------------------------------*/
-function x265_frame_subsample_luma_lsx
-    andi            t4,     a4,     7       // width % 8
-    srli.d          a4,     a4,     3       // width / 8
-
-.LOOP_HEIGHT_FRAME_SUBSAMPLE_LUMA_LSX:
-    add.d           t0,     a0,     a2      // *inRowBelow
-    addi.d          t1,     zero,   0
-    addi.d          t2,     zero,   0
-    addi.d          t8,     zero,   0       // index
-
-.LOOP_WIDTH_FRAME_SUBSAMPLE_LUMA_LSX:
-    bge             t8,     a4,     .HANDLING_MANTISSA_FRAMESUBSAMPLELUMA_LSX_7
-    CALCULATE_FRAMESUBSAMPLELUMA_LSX_8  t1
-    add.d           t3,     a1,     t2
-    vstelm.d        vr6,    t3,     0,      0
-
-    // prepare for next inner-loop
-    addi.d          t1,     t1,     16
-    addi.d          t2,     t2,     8
-    addi.d          t8,     t8,     1       // index++
-    b               .LOOP_WIDTH_FRAME_SUBSAMPLE_LUMA_LSX
-
-.HANDLING_MANTISSA_FRAMESUBSAMPLELUMA_LSX_7:
-    bge             zero,   t4,     .PREPARE_FOR_NEXT_OUTERLOOP_FRAMESUBSAMPLELUMA_LSX
-    CALCULATE_FRAMESUBSAMPLELUMA_LSX_8  t1              // calculate mantissa
-    addi.d          t8,     zero,   0
-    add.d           t3,     a1,     t2
-.irp   offset0, 0, 1, 2, 3, 4, 5, 6
-    vstelm.b        vr6,    t3,     0,      \offset0
-    addi.d          t8,     t8,     1
-    addi.d          t3,     t3,     1
-    bge             t8,     t4,     .PREPARE_FOR_NEXT_OUTERLOOP_FRAMESUBSAMPLELUMA_LSX
-.endr
-
-.PREPARE_FOR_NEXT_OUTERLOOP_FRAMESUBSAMPLELUMA_LSX:
-    // prepare for next outer-loop
-    add.d           a0,     a0,     a2
-    add.d           a0,     a0,     a2
-    add.d           a1,     a1,     a3
-    addi.d          a5,     a5,     -1
-    blt             zero,   a5,     .LOOP_HEIGHT_FRAME_SUBSAMPLE_LUMA_LSX
-endfunc
-
-.macro CALCULATE_FRAMESUBSAMPLELUMA_LASX_16 offset0
-    // load inRow[offset0 ~ offset0+31], inRowBelow[offset0 ~ offset0+31]
-    xvldx           xr1,    a0,     \offset0
-    xvldx           xr3,    t0,     \offset0
-    xvaddwev.h.bu   xr4,    xr1,    xr3
-    xvaddwod.h.bu   xr5,    xr1,    xr3
-    xvsrari.h       xr4,    xr4,    1
-    xvsrari.h       xr5,    xr5,    1
-    xvadd.h         xr6,    xr4,    xr5
-    xvsrarni.b.h    xr6,    xr6,    1
-    xvpermi.d       xr6,    xr6,    0b00001000
-.endm
-
-/*-----------------------------------------------------------------------------
-void x265_frame_subsample_luma_lasx(const pixel* src0, pixel* dst0,
-            intptr_t src_stride, intptr_t dst_stride, int width, int height)
------------------------------------------------------------------------------*/
-function x265_frame_subsample_luma_lasx
-    andi            t4,     a4,     15      // width % 16
-    srli.d          a4,     a4,     4       // width / 16
-
-.LOOP_HEIGHT_FRAME_SUBSAMPLE_LUMA_LASX:
-    add.d           t0,     a0,     a2      // *inRowBelow
-    addi.d          t1,     zero,   0
-    addi.d          t2,     zero,   0
-    addi.d          t8,     zero,   0       // index
-
-.LOOP_WIDTH_FRAME_SUBSAMPLE_LUMA_LASX:
-    bge             t8,     a4,     .HANDLING_MANTISSA_FRAMESUBSAMPLELUMA_LASX_15
-    CALCULATE_FRAMESUBSAMPLELUMA_LASX_16  t1
-    vstx            vr6,    a1,     t2
-    addi.d          t1,     t1,     32
-    addi.d          t2,     t2,     16
-    addi.d          t8,     t8,     1       // index++
-    b               .LOOP_WIDTH_FRAME_SUBSAMPLE_LUMA_LASX
-
-.HANDLING_MANTISSA_FRAMESUBSAMPLELUMA_LASX_15:
-    bge             zero,   t4,     .PREPARE_FOR_NEXT_OUTERLOOP_FRAMESUBSAMPLELUMA_LASX
-    CALCULATE_FRAMESUBSAMPLELUMA_LASX_16    t1              // calculate mantissa
-    addi.d          t8,     zero,   0
-    add.d           t3,     a1,     t2
-.irp   offset0, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14
-    vstelm.b        vr6,    t3,     0,      \offset0
-    addi.d          t8,     t8,     1
-    addi.d          t3,     t3,     1
-    bge             t8,     t4,     .PREPARE_FOR_NEXT_OUTERLOOP_FRAMESUBSAMPLELUMA_LASX
-.endr
-
-.PREPARE_FOR_NEXT_OUTERLOOP_FRAMESUBSAMPLELUMA_LASX:
-    // prepare for next loop
-    add.d           a0,     a0,     a2
-    add.d           a0,     a0,     a2
-    add.d           a1,     a1,     a3
-    addi.d          a5,     a5,     -1
-    blt             zero,   a5,     .LOOP_HEIGHT_FRAME_SUBSAMPLE_LUMA_LASX
-endfunc
-
-
 .macro HANDLING_MANTISSA_MC_A2_3  size, dst0, dst1, dst2, count, dst_adress, \
                         dst_stride, index0, index1, index2, terminate_lable
     addi.d          \count,         \count,         -1

source/common/loongarch64/pixel.h

@@ -372,8 +372,6 @@ void x265_addAvg_64x64_lasx(const int16_t* src0, const int16_t* src1, pixel* dst
 
 void x265_frame_init_lowres_core_lsx(const pixel *src0, pixel *dst0, pixel *dsth, pixel *dstv, pixel *dstc, intptr_t src_stride, intptr_t dst_stride, int width, int height);
 void x265_frame_init_lowres_core_lasx(const pixel *src0, pixel *dst0, pixel *dsth, pixel *dstv, pixel *dstc, intptr_t src_stride, intptr_t dst_stride, int width, int height);
-void x265_frame_subsample_luma_lsx(const pixel *src0, pixel *dst0, intptr_t src_stride, intptr_t dst_stride, int width, int height);
-void x265_frame_subsample_luma_lasx(const pixel *src0, pixel *dst0, intptr_t src_stride, intptr_t dst_stride, int width, int height);
 void x265_cutree_fix8_unpack_lsx(double *dst, uint16_t *src, int count);
 void x265_cutree_fix8_unpack_lasx(double *dst, uint16_t *src, int count);
 void x265_cutree_fix8_pack_lsx(uint16_t *dst, double *src, int count);

source/common/lowres.cpp

@@ -26,6 +26,12 @@
 #include "lowres.h"
 #include "mv.h"
 
+namespace X265_NS {
+    // forward declaration - defined in pixel.cpp, no SIMD override for MCSTF path
+    void frameInitLowresCoreMCSTF(const pixel* src0, pixel* dst0, pixel* dsth, pixel* dstv, pixel* dstc,
+        intptr_t src_stride, intptr_t dst_stride, int width, int height);
+}
+
 using namespace X265_NS;
 
 /*
@@ -194,11 +200,6 @@ bool Lowres::create(x265_param* param, PicYuv *origPic, uint32_t qgSize)
         }
     }
 
-    for (int i = 0; i < 4; i++)
-    {
-        CHECKED_MALLOC(lowresMcstfMvs[0][i], MV, cuCount);
-    }
-
     for (int i = 0; i < bframes + 2; i++)
     {
         CHECKED_MALLOC(lowresMvs[0][i], MV, cuCount);
@@ -287,10 +288,6 @@ void Lowres::destroy(x265_param* param)
         }
     }
 
-    for (int i = 0; i < 4; i++)
-    {
-        X265_FREE(lowresMcstfMvs[0][i]);
-    }
     X265_FREE(qpAqOffset);
     X265_FREE(invQscaleFactor);
     X265_FREE(qpCuTreeOffset);
@@ -344,11 +341,13 @@ void Lowres::destroy(x265_param* param)
     }
 }
 // (re) initialize lowres state
-void Lowres::init(PicYuv *origPic, int poc)
+void Lowres::init(PicYuv* origPic, int poc, bool bEnableTemporalFilter)
 {
     bLastMiniGopBFrame = false;
     bKeyframe = false; // Not a keyframe unless identified by lookahead
     bIsFadeEnd = false;
+    filterThisGOP = false;
+    noiseScore = -1;
     frameNum = poc;
     leadingBframes = 0;
     indB = 0;
@@ -368,21 +367,17 @@ void Lowres::init(PicYuv *origPic, int poc)
         lowresMvs[1][i][0].x = 0x7FFF;
     }
 
-    for (int i = 0; i < 4; i++)
-    {
-        lowresMcstfMvs[0][i][0].x = 0x7FFF;
-    }
-
     for (int i = 0; i < bframes + 2; i++)
         intraMbs[i] = 0;
     if (origPic->m_param->rc.vbvBufferSize)
         for (int i = 0; i < X265_LOOKAHEAD_MAX + 1; i++)
             plannedType[i] = X265_TYPE_AUTO;
 
     /* downscale and generate 4 hpel planes for lookahead */
-    primitives.frameInitLowres(origPic->m_picOrg[0],
-                               lowresPlane[0], lowresPlane[1], lowresPlane[2], lowresPlane[3],
-                               origPic->m_stride, lumaStride, width, lines);
+    if (bEnableTemporalFilter)
+        frameInitLowresCoreMCSTF(origPic->m_picOrg[0], lowresPlane[0], lowresPlane[1], lowresPlane[2], lowresPlane[3], origPic->m_stride, lumaStride, width, lines);
+    else
+        primitives.frameInitLowres(origPic->m_picOrg[0], lowresPlane[0], lowresPlane[1], lowresPlane[2], lowresPlane[3], origPic->m_stride, lumaStride, width, lines);
 
     /* extend hpel planes for motion search */
     extendPicBorder(lowresPlane[0], lumaStride, width, lines, origPic->m_lumaMarginX, origPic->m_lumaMarginY);
@@ -392,9 +387,11 @@ void Lowres::init(PicYuv *origPic, int poc)
 
     if (origPic->m_param->bEnableHME || origPic->m_param->bEnableTemporalFilter)
     {
-        primitives.frameInitLowerRes(lowresPlane[0],
-            lowerResPlane[0], lowerResPlane[1], lowerResPlane[2], lowerResPlane[3],
-            lumaStride, lumaStride/2, (width / 2), (lines / 2));
+        if (bEnableTemporalFilter)
+            frameInitLowresCoreMCSTF(lowresPlane[0], lowerResPlane[0], lowerResPlane[1], lowerResPlane[2], lowerResPlane[3], lumaStride, lumaStride / 2, (width / 2), (lines / 2));
+        else
+            primitives.frameInitLowerRes(lowresPlane[0], lowerResPlane[0], lowerResPlane[1], lowerResPlane[2], lowerResPlane[3], lumaStride, lumaStride / 2, (width / 2), (lines / 2));
+
         extendPicBorder(lowerResPlane[0], lumaStride/2, width/2, lines/2, origPic->m_lumaMarginX/2, origPic->m_lumaMarginY/2);
         extendPicBorder(lowerResPlane[1], lumaStride/2, width/2, lines/2, origPic->m_lumaMarginX/2, origPic->m_lumaMarginY/2);
         extendPicBorder(lowerResPlane[2], lumaStride/2, width/2, lines/2, origPic->m_lumaMarginX/2, origPic->m_lumaMarginY/2);

source/common/lowres.h

@@ -177,6 +177,8 @@ struct Lowres : public ReferencePlanes
     bool   bKeyframe;
     bool   bLastMiniGopBFrame;
     bool   bIsFadeEnd;
+    bool    filterThisGOP;  // Set by lookahead noise gate: true = GOP start was noisy, apply MCSTF
+    int32_t noiseScore;     // Raw noise score from estimateNoise() at GOP start; -1 for non-GOP-start frames
 
     double ipCostRatio;
 
@@ -192,7 +194,6 @@ struct Lowres : public ReferencePlanes
     uint16_t* lowresCosts[X265_BFRAME_MAX + 2][X265_BFRAME_MAX + 2];
     int32_t*  lowresMvCosts[2][X265_BFRAME_MAX + 2];
     MV*       lowresMvs[2][X265_BFRAME_MAX + 2];
-    MV*       lowresMcstfMvs[2][4];
     uint32_t  maxBlocksInRow;
     uint32_t  maxBlocksInCol;
     uint32_t  maxBlocksInRowFullRes;
@@ -253,7 +254,7 @@ struct Lowres : public ReferencePlanes
 
     bool create(x265_param* param, PicYuv *origPic, uint32_t qgSize);
     void destroy(x265_param* param);
-    void init(PicYuv *origPic, int poc);
+    void init(PicYuv* origPic, int poc, bool bEnableTemporalFilter);
 };
 }