[x265] [PATCH 1 of 6] asm: sse4 version of costCoeffGroupNxN in codeCoeffNxN

Deepthi Nandakumar deepthi at multicorewareinc.com
Fri Jun 5 09:10:22 CEST 2015


Thanks, Min.

Sumalatha - can you please work on adding testbench support for the new
codeCoeff primitive that has been added?

On Fri, Jun 5, 2015 at 12:43 AM, Min Chen <chenm003 at 163.com> wrote:

> # HG changeset patch
> # User Min Chen <chenm003 at 163.com>
> # Date 1433445185 25200
> # Node ID 24f347c00df01352fa6860e05b376846d8d8cc74
> # Parent  093618ce0b26ea4703b5928f618d2895cf6daf32
> asm: sse4 version of costCoeffGroupNxN in codeCoeffNxN
> ---
>  source/common/constants.cpp          |    5 +-
>  source/common/constants.h            |    2 +-
>  source/common/contexts.h             |    3 +-
>  source/common/dct.cpp                |   57 +++++++++++
>  source/common/primitives.h           |    4 +
>  source/common/x86/asm-primitives.cpp |    2 +
>  source/common/x86/pixel-util.h       |    3 +
>  source/common/x86/pixel-util8.asm    |  176
> ++++++++++++++++++++++++++++++++++
>  source/encoder/entropy.cpp           |  113 ++++++++++------------
>  9 files changed, 297 insertions(+), 68 deletions(-)
>
> diff -r 093618ce0b26 -r 24f347c00df0 source/common/constants.cpp
> --- a/source/common/constants.cpp       Tue Jun 02 17:21:24 2015 +0800
> +++ b/source/common/constants.cpp       Thu Jun 04 12:13:05 2015 -0700
> @@ -324,11 +324,12 @@
>        4,  12, 20, 28,  5, 13, 21, 29,  6, 14, 22, 30,  7, 15, 23, 31, 36,
> 44, 52, 60, 37, 45, 53, 61, 38, 46, 54, 62, 39, 47, 55, 63 }
>  };
>
> -ALIGN_VAR_16(const uint16_t, g_scan4x4[NUM_SCAN_TYPE][4 * 4]) =
> +ALIGN_VAR_16(const uint16_t, g_scan4x4[NUM_SCAN_TYPE + 1][4 * 4]) =
>  {
>      { 0,  4,  1,  8,  5,  2, 12,  9,  6,  3, 13, 10,  7, 14, 11, 15 },
>      { 0,  1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12, 13, 14, 15 },
> -    { 0,  4,  8, 12,  1,  5,  9, 13,  2,  6, 10, 14,  3,  7, 11, 15 }
> +    { 0,  4,  8, 12,  1,  5,  9, 13,  2,  6, 10, 14,  3,  7, 11, 15 },
> +    { 0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0 }
>  };
>
>  const uint16_t g_scan16x16[16 * 16] =
> diff -r 093618ce0b26 -r 24f347c00df0 source/common/constants.h
> --- a/source/common/constants.h Tue Jun 02 17:21:24 2015 +0800
> +++ b/source/common/constants.h Thu Jun 04 12:13:05 2015 -0700
> @@ -83,7 +83,7 @@
>  extern const uint16_t* const g_scanOrder[NUM_SCAN_TYPE][NUM_SCAN_SIZE];
>  extern const uint16_t* const g_scanOrderCG[NUM_SCAN_TYPE][NUM_SCAN_SIZE];
>  extern const uint16_t g_scan8x8diag[8 * 8];
> -extern const uint16_t g_scan4x4[NUM_SCAN_TYPE][4 * 4];
> +extern const uint16_t g_scan4x4[NUM_SCAN_TYPE + 1][4 * 4];  // +1 for
> safe buffer area for codeCoeffNxN assembly optimize, there have up to 15
> bytes beyond bound read
>
>  extern const uint8_t g_lastCoeffTable[32];
>  extern const uint8_t g_goRiceRange[5]; // maximum value coded with Rice
> codes
> diff -r 093618ce0b26 -r 24f347c00df0 source/common/contexts.h
> --- a/source/common/contexts.h  Tue Jun 02 17:21:24 2015 +0800
> +++ b/source/common/contexts.h  Thu Jun 04 12:13:05 2015 -0700
> @@ -102,11 +102,12 @@
>  #define OFF_TQUANT_BYPASS_FLAG_CTX (OFF_TRANSFORMSKIP_FLAG_CTX + 2 *
> NUM_TRANSFORMSKIP_FLAG_CTX)
>  #define MAX_OFF_CTX_MOD            (OFF_TQUANT_BYPASS_FLAG_CTX +
>  NUM_TQUANT_BYPASS_FLAG_CTX)
>
> +extern "C" const uint32_t g_entropyStateBits[128];
> +
>  namespace x265 {
>  // private namespace
>
>  extern const uint32_t g_entropyBits[128];
> -extern const uint32_t g_entropyStateBits[128];
>  extern const uint8_t g_nextState[128][2];
>
>  #define sbacGetMps(S)            ((S) & 1)
> diff -r 093618ce0b26 -r 24f347c00df0 source/common/dct.cpp
> --- a/source/common/dct.cpp     Tue Jun 02 17:21:24 2015 +0800
> +++ b/source/common/dct.cpp     Thu Jun 04 12:13:05 2015 -0700
> @@ -29,6 +29,7 @@
>
>  #include "common.h"
>  #include "primitives.h"
> +#include "contexts.h"   // costCoeffNxN_c
>
>  using namespace x265;
>
> @@ -817,6 +818,61 @@
>      return ((lastNZPosInCG << 16) | firstNZPosInCG);
>  }
>
> +
> +uint32_t costCoeffNxN_c(const uint16_t *scan, const coeff_t *coeff,
> intptr_t trSize, uint16_t *absCoeff, const uint8_t *tabSigCtx, uint32_t
> scanFlagMask, uint8_t *baseCtx, int offset, int scanPosSigOff, int
> subPosBase)
> +{
> +    ALIGN_VAR_32(uint16_t, tmpCoeff[SCAN_SET_SIZE]);
> +    uint32_t numNonZero = (scanPosSigOff < (SCAN_SET_SIZE - 1) ? 1 : 0);
> +    uint32_t sum = 0;
> +
> +    // correct offset to match assembly
> +    absCoeff -= numNonZero;
> +
> +    for (int i = 0; i < MLS_CG_SIZE; i++)
> +    {
> +        tmpCoeff[i * MLS_CG_SIZE + 0] = (uint16_t)abs(coeff[i * trSize +
> 0]);
> +        tmpCoeff[i * MLS_CG_SIZE + 1] = (uint16_t)abs(coeff[i * trSize +
> 1]);
> +        tmpCoeff[i * MLS_CG_SIZE + 2] = (uint16_t)abs(coeff[i * trSize +
> 2]);
> +        tmpCoeff[i * MLS_CG_SIZE + 3] = (uint16_t)abs(coeff[i * trSize +
> 3]);
> +    }
> +
> +    do
> +    {
> +        uint32_t blkPos, sig, ctxSig;
> +        blkPos = scan[scanPosSigOff];
> +        const uint32_t posZeroMask = (subPosBase + scanPosSigOff) ? ~0 :
> 0;
> +        sig     = scanFlagMask & 1;
> +        scanFlagMask >>= 1;
> +        X265_CHECK((uint32_t)(tmpCoeff[blkPos] != 0) == sig, "sign bit
> mistake\n");
> +        if ((scanPosSigOff != 0) || (subPosBase == 0) || numNonZero)
> +        {
> +            const uint32_t cnt = tabSigCtx[blkPos] + offset;
> +            ctxSig = cnt & posZeroMask;
> +
> +            //X265_CHECK(ctxSig == Quant::getSigCtxInc(patternSigCtx,
> log2TrSize, trSize, codingParameters.scan[subPosBase + scanPosSigOff],
> bIsLuma, codingParameters.firstSignificanceMapContext), "sigCtx
> mistake!\n");;
> +            //encodeBin(sig, baseCtx[ctxSig]);
> +            const uint32_t mstate = baseCtx[ctxSig];
> +            const uint32_t mps = mstate & 1;
> +            const uint32_t stateBits = g_entropyStateBits[mstate ^ sig];
> +            uint32_t nextState = (stateBits >> 24) + mps;
> +            if ((mstate ^ sig) == 1)
> +                nextState = sig;
> +            X265_CHECK(sbacNext(mstate, sig) == nextState, "nextState
> check failure\n");
> +            X265_CHECK(sbacGetEntropyBits(mstate, sig) == (stateBits &
> 0xFFFFFF), "entropyBits check failure\n");
> +            baseCtx[ctxSig] = (uint8_t)nextState;
> +            sum += stateBits;
> +        }
> +        assert(numNonZero <= 15);
> +        assert(blkPos <= 15);
> +        absCoeff[numNonZero] = tmpCoeff[blkPos];
> +        numNonZero += sig;
> +        scanPosSigOff--;
> +    }
> +    while(scanPosSigOff >= 0);
> +
> +    return (sum & 0xFFFFFF);
> +}
> +
>  }  // closing - anonymous file-static namespace
>
>  namespace x265 {
> @@ -851,5 +907,6 @@
>
>      p.scanPosLast = scanPosLast_c;
>      p.findPosFirstLast = findPosFirstLast_c;
> +    p.costCoeffNxN = costCoeffNxN_c;
>  }
>  }
> diff -r 093618ce0b26 -r 24f347c00df0 source/common/primitives.h
> --- a/source/common/primitives.h        Tue Jun 02 17:21:24 2015 +0800
> +++ b/source/common/primitives.h        Thu Jun 04 12:13:05 2015 -0700
> @@ -186,6 +186,8 @@
>  typedef int (*scanPosLast_t)(const uint16_t *scan, const coeff_t *coeff,
> uint16_t *coeffSign, uint16_t *coeffFlag, uint8_t *coeffNum, int numSig,
> const uint16_t* scanCG4x4, const int trSize);
>  typedef uint32_t (*findPosFirstLast_t)(const int16_t *dstCoeff, const
> intptr_t trSize, const uint16_t scanTbl[16]);
>
> +typedef uint32_t (*costCoeffNxN_t)(const uint16_t *scan, const coeff_t
> *coeff, intptr_t trSize, uint16_t *absCoeff, const uint8_t *tabSigCtx,
> uint32_t scanFlagMask, uint8_t *baseCtx, int offset, int scanPosSigOff, int
> subPosBase);
> +
>  /* Function pointers to optimized encoder primitives. Each pointer can
> reference
>   * either an assembly routine, a SIMD intrinsic primitive, or a C
> function */
>  struct EncoderPrimitives
> @@ -310,6 +312,8 @@
>      scanPosLast_t         scanPosLast;
>      findPosFirstLast_t    findPosFirstLast;
>
> +    costCoeffNxN_t        costCoeffNxN;
> +
>      /* There is one set of chroma primitives per color space. An encoder
> will
>       * have just a single color space and thus it will only ever use one
> entry
>       * in this array. However we always fill all entries in the array in
> case
> diff -r 093618ce0b26 -r 24f347c00df0 source/common/x86/asm-primitives.cpp
> --- a/source/common/x86/asm-primitives.cpp      Tue Jun 02 17:21:24 2015
> +0800
> +++ b/source/common/x86/asm-primitives.cpp      Thu Jun 04 12:13:05 2015
> -0700
> @@ -2048,6 +2048,9 @@
>
>          ALL_LUMA_CU(psy_cost_pp, psyCost_pp, sse4);
>          ALL_LUMA_CU(psy_cost_ss, psyCost_ss, sse4);
> +
> +        // TODO: it is passed smoke test, but we need testbench, so
> temporary disable
> +        //p.costCoeffNxN = x265_costCoeffNxN_sse4;
>  #endif
>      }
>      if (cpuMask & X265_CPU_AVX)
> diff -r 093618ce0b26 -r 24f347c00df0 source/common/x86/pixel-util.h
> --- a/source/common/x86/pixel-util.h    Tue Jun 02 17:21:24 2015 +0800
> +++ b/source/common/x86/pixel-util.h    Thu Jun 04 12:13:05 2015 -0700
> @@ -82,6 +82,9 @@
>  int x265_scanPosLast_avx2_bmi2(const uint16_t *scan, const coeff_t
> *coeff, uint16_t *coeffSign, uint16_t *coeffFlag, uint8_t *coeffNum, int
> numSig, const uint16_t* scanCG4x4, const int trSize);
>  uint32_t x265_findPosFirstLast_ssse3(const int16_t *dstCoeff, const
> intptr_t trSize, const uint16_t scanTbl[16]);
>
> +uint32_t x265_costCoeffNxN_sse4(const uint16_t *scan, const coeff_t
> *coeff, intptr_t trSize, uint16_t *absCoeff, const uint8_t *tabSigCtx,
> uint32_t scanFlagMask, uint8_t *baseCtx, int offset, int scanPosSigOff, int
> subPosBase);
> +
> +
>  #define SETUP_CHROMA_PIXELSUB_PS_FUNC(W, H, cpu) \
>      void x265_pixel_sub_ps_ ## W ## x ## H ## cpu(int16_t*  dest,
> intptr_t destride, const pixel* src0, const pixel* src1, intptr_t
> srcstride0, intptr_t srcstride1); \
>      void x265_pixel_add_ps_ ## W ## x ## H ## cpu(pixel* dest, intptr_t
> destride, const pixel* src0, const int16_t*  src1, intptr_t srcStride0,
> intptr_t srcStride1);
> diff -r 093618ce0b26 -r 24f347c00df0 source/common/x86/pixel-util8.asm
> --- a/source/common/x86/pixel-util8.asm Tue Jun 02 17:21:24 2015 +0800
> +++ b/source/common/x86/pixel-util8.asm Thu Jun 04 12:13:05 2015 -0700
> @@ -71,6 +71,7 @@
>  cextern pb_64
>  cextern hmul_16p
>  cextern trans8_shuf
> +cextern_naked g_entropyStateBits
>
>
>  ;-----------------------------------------------------------------------------
>  ; void getResidual(pixel *fenc, pixel *pred, int16_t *residual, intptr_t
> stride)
> @@ -6362,3 +6363,178 @@
>      add         [r1 + 4 * 4], r6d
>      RET
>  %endif ; ARCH_X86_64
> +
> +
> +; uint32_t costCoeffNxN(uint16_t *scan, coeff_t *coeff, intptr_t trSize,
> uint16_t *absCoeff, uint8_t *tabSigCtx, uint16_t scanFlagMask, uint8_t
> *baseCtx, int offset, int subPosBase)
> +;for (int i = 0; i < MLS_CG_SIZE; i++)
> +;{
> +;    tmpCoeff[i * MLS_CG_SIZE + 0] = (uint16_t)abs(coeff[blkPosBase + i *
> trSize + 0]);
> +;    tmpCoeff[i * MLS_CG_SIZE + 1] = (uint16_t)abs(coeff[blkPosBase + i *
> trSize + 1]);
> +;    tmpCoeff[i * MLS_CG_SIZE + 2] = (uint16_t)abs(coeff[blkPosBase + i *
> trSize + 2]);
> +;    tmpCoeff[i * MLS_CG_SIZE + 3] = (uint16_t)abs(coeff[blkPosBase + i *
> trSize + 3]);
> +;}
> +;do
> +;{
> +;    uint32_t blkPos, sig, ctxSig;
> +;    blkPos = g_scan4x4[codingParameters.scanType][scanPosSigOff];
> +;    const uint32_t posZeroMask = (subPosBase + scanPosSigOff) ? ~0 : 0;
> +;    sig     = scanFlagMask & 1;
> +;    scanFlagMask >>= 1;
> +;    if (scanPosSigOff + (subSet == 0) + numNonZero)
> +;    {
> +;        const uint32_t cnt = tabSigCtx[blkPos] + offset + posOffset;
> +;        ctxSig = cnt & posZeroMask;
> +;
> +;        const uint32_t mstate = baseCtx[ctxSig];
> +;        const uint32_t mps = mstate & 1;
> +;        const uint32_t stateBits = g_entropyStateBits[mstate ^ sig];
> +;        uint32_t nextState = (stateBits >> 24) + mps;
> +;        if ((mstate ^ sig) == 1)
> +;            nextState = sig;
> +;        baseCtx[ctxSig] = (uint8_t)nextState;
> +;        sum += stateBits;
> +;    }
> +;    absCoeff[numNonZero] = tmpCoeff[blkPos];
> +;    numNonZero += sig;
> +;    scanPosSigOff--;
> +;}
> +;while(scanPosSigOff >= 0);
> +; sum &= 0xFFFFFF
> +
> +%if ARCH_X86_64
> +; uint32_t costCoeffNxN(uint16_t *scan, coeff_t *coeff, intptr_t trSize,
> uint16_t *absCoeff, uint8_t *tabSigCtx, uint16_t scanFlagMask, uint8_t
> *baseCtx, int offset, int scanPosSigOff, int subPosBase)
> +INIT_XMM sse4
> +cglobal costCoeffNxN, 6,11,5
> +    add         r2d, r2d
> +
> +    ; abs(coeff)
> +    movh        m1, [r1]
> +    movhps      m1, [r1 + r2]
> +    movh        m2, [r1 + r2 * 2]
> +    lea         r2, [r2 * 3]
> +    movhps      m2, [r1 + r2]
> +    pabsw       m1, m1
> +    pabsw       m2, m2
> +    ; r[1-2] free here
> +
> +    ; WARNING: beyond-bound read here!
> +    ; loading scan table
> +    mov         r2d, r8m
> +    xor         r2d, 15
> +    movu        m0, [r0 + r2 * 2]
> +    movu        m3, [r0 + r2 * 2 + mmsize]
> +    packuswb    m0, m3
> +    pxor        m0, [pb_15]
> +    xchg        r2d, r8m
> +    ; r[0-1] free here
> +
> +    ; reorder coeff
> +    mova        m3, [deinterleave_shuf]
> +    pshufb      m1, m3
> +    pshufb      m2, m3
> +    punpcklqdq  m3, m1, m2
> +    punpckhqdq  m1, m2
> +    pshufb      m3, m0
> +    pshufb      m1, m0
> +    punpcklbw   m2, m3, m1
> +    punpckhbw   m3, m1
> +    ; r[0-1], m[1] free here
> +
> +    ; loading tabSigCtx (+offset)
> +    mova        m1, [r4]
> +    pshufb      m1, m0
> +    movd        m4, r7m
> +    pxor        m5, m5
> +    pshufb      m4, m5
> +    paddb       m1, m4
> +
> +    ; register mapping
> +    ; m0 - Zigzag
> +    ; m1 - sigCtx
> +    ; {m3,m2} - abs(coeff)
> +    ; r0 - g_entropyStateBits
> +    ; r1 - baseCtx
> +    ; r2 - scanPosSigOff
> +    ; r3 - absCoeff
> +    ; r4 - nonZero
> +    ; r5 - scanFlagMask
> +    ; r6 - sum
> +    lea         r0, [g_entropyStateBits]
> +    mov         r1, r6mp
> +    xor         r6d, r6d
> +    xor         r4d, r4d
> +    xor         r8d, r8d
> +
> +    test        r2d, r2d
> +    jz         .idx_zero
> +
> +.loop:
> +;   {
> +;        const uint32_t cnt = tabSigCtx[blkPos] + offset + posOffset;
> +;        ctxSig = cnt & posZeroMask;
> +;        const uint32_t mstate = baseCtx[ctxSig];
> +;        const uint32_t mps = mstate & 1;
> +;        const uint32_t stateBits = g_entropyStateBits[mstate ^ sig];
> +;        uint32_t nextState = (stateBits >> 24) + mps;
> +;        if ((mstate ^ sig) == 1)
> +;            nextState = sig;
> +;        baseCtx[ctxSig] = (uint8_t)nextState;
> +;        sum += stateBits;
> +;    }
> +;    absCoeff[numNonZero] = tmpCoeff[blkPos];
> +;    numNonZero += sig;
> +;    scanPosSigOff--;
> +
> +    pextrw      [r3 + r4 * 2], m2, 0            ; absCoeff[numNonZero] =
> tmpCoeff[blkPos]
> +    shr         r5d, 1
> +    setc        r8b                             ; r8 = sig
> +    add         r4d, r8d                        ; numNonZero += sig
> +    palignr     m4, m3, m2, 2
> +    psrldq      m3, 2
> +    mova        m2, m4
> +    movd        r7d, m1                         ; r7 = ctxSig
> +    movzx       r7d, r7b
> +    psrldq      m1, 1
> +    movzx       r9d, byte [r1 + r7]             ; mstate = baseCtx[ctxSig]
> +    mov         r10d, r9d
> +    and         r10d, 1                         ; mps = mstate & 1
> +    xor         r9d, r8d                        ; r9 = mstate ^ sig
> +    add         r6d, [r0 + r9 * 4]              ; sum +=
> g_entropyStateBits[mstate ^ sig]
> +    add         r10b, byte [r0 + r9 * 4 + 3]    ; nextState = (stateBits
> >> 24) + mps
> +    cmp         r9b, 1
> +    cmove       r10d, r8d
> +    mov    byte [r1 + r7], r10b
> +
> +    dec         r2d
> +    jg         .loop
> +
> +.idx_zero:
> +    pextrw      [r3 + r4 * 2], m2, 0            ; absCoeff[numNonZero] =
> tmpCoeff[blkPos]
> +    add         r4b, r8m
> +    xor         r2d, r2d
> +    cmp    word r9m, 0
> +    sete        r2b
> +    add         r4b, r2b
> +    jz         .exit
> +
> +    dec         r2b
> +    movd        r3d, m1
> +    and         r2d, r3d
> +
> +    movzx       r3d, byte [r1 + r2]             ; mstate = baseCtx[ctxSig]
> +    mov         r4d, r5d
> +    xor         r5d, r3d                        ; r0 = mstate ^ sig
> +    and         r3d, 1                          ; mps = mstate & 1
> +    add         r6d, [r0 + r5 * 4]              ; sum +=
> g_entropyStateBits[mstate ^ sig]
> +    add         r3b, [r0 + r5 * 4 + 3]          ; nextState = (stateBits
> >> 24) + mps
> +    cmp         r5b, 1
> +    cmove       r3d, r4d
> +    mov    byte [r1 + r2], r3b
> +
> +.exit:
> +%ifnidn eax,r6d
> +    mov         eax, r6d
> +%endif
> +    and         eax, 0xFFFFFF
> +    RET
> +%endif ; ARCH_X86_64
> diff -r 093618ce0b26 -r 24f347c00df0 source/encoder/entropy.cpp
> --- a/source/encoder/entropy.cpp        Tue Jun 02 17:21:24 2015 +0800
> +++ b/source/encoder/entropy.cpp        Thu Jun 04 12:13:05 2015 -0700
> @@ -1517,12 +1517,12 @@
>      uint8_t * const baseCtx = bIsLuma ? &m_contextState[OFF_SIG_FLAG_CTX]
> : &m_contextState[OFF_SIG_FLAG_CTX + NUM_SIG_FLAG_CTX_LUMA];
>      uint32_t c1 = 1;
>      int scanPosSigOff = scanPosLast - (lastScanSet << MLS_CG_SIZE) - 1;
> -    int absCoeff[1 << MLS_CG_SIZE];
> +    ALIGN_VAR_32(uint16_t, absCoeff[1 << MLS_CG_SIZE]);
>      uint32_t numNonZero = 1;
>      unsigned long lastNZPosInCG;
>      unsigned long firstNZPosInCG;
>
> -    absCoeff[0] = int(abs(coeff[posLast]));
> +    absCoeff[0] = (uint16_t)abs(coeff[posLast]);
>
>      for (int subSet = lastScanSet; subSet >= 0; subSet--)
>      {
> @@ -1600,19 +1600,20 @@
>
>              const int offset =
> codingParameters.firstSignificanceMapContext;
>              ALIGN_VAR_32(uint16_t, tmpCoeff[SCAN_SET_SIZE]);
> -            // TODO: accelerate by PABSW
>              const uint32_t blkPosBase  =
> codingParameters.scan[subPosBase];
> -            for (int i = 0; i < MLS_CG_SIZE; i++)
> -            {
> -                tmpCoeff[i * MLS_CG_SIZE + 0] =
> (uint16_t)abs(coeff[blkPosBase + i * trSize + 0]);
> -                tmpCoeff[i * MLS_CG_SIZE + 1] =
> (uint16_t)abs(coeff[blkPosBase + i * trSize + 1]);
> -                tmpCoeff[i * MLS_CG_SIZE + 2] =
> (uint16_t)abs(coeff[blkPosBase + i * trSize + 2]);
> -                tmpCoeff[i * MLS_CG_SIZE + 3] =
> (uint16_t)abs(coeff[blkPosBase + i * trSize + 3]);
> -            }
>
>              X265_CHECK(scanPosSigOff >= 0, "scanPosSigOff check
> failure\n");
>              if (m_bitIf)
>              {
> +                // TODO: accelerate by PABSW
> +                for (int i = 0; i < MLS_CG_SIZE; i++)
> +                {
> +                    tmpCoeff[i * MLS_CG_SIZE + 0] =
> (uint16_t)abs(coeff[blkPosBase + i * trSize + 0]);
> +                    tmpCoeff[i * MLS_CG_SIZE + 1] =
> (uint16_t)abs(coeff[blkPosBase + i * trSize + 1]);
> +                    tmpCoeff[i * MLS_CG_SIZE + 2] =
> (uint16_t)abs(coeff[blkPosBase + i * trSize + 2]);
> +                    tmpCoeff[i * MLS_CG_SIZE + 3] =
> (uint16_t)abs(coeff[blkPosBase + i * trSize + 3]);
> +                }
> +
>                  if (log2TrSize == 2)
>                  {
>                      do
> @@ -1667,6 +1668,15 @@
>                  uint32_t sum = 0;
>                  if (log2TrSize == 2)
>                  {
> +                    // TODO: accelerate by PABSW
> +                    for (int i = 0; i < MLS_CG_SIZE; i++)
> +                    {
> +                        tmpCoeff[i * MLS_CG_SIZE + 0] =
> (uint16_t)abs(coeff[blkPosBase + i * trSize + 0]);
> +                        tmpCoeff[i * MLS_CG_SIZE + 1] =
> (uint16_t)abs(coeff[blkPosBase + i * trSize + 1]);
> +                        tmpCoeff[i * MLS_CG_SIZE + 2] =
> (uint16_t)abs(coeff[blkPosBase + i * trSize + 2]);
> +                        tmpCoeff[i * MLS_CG_SIZE + 3] =
> (uint16_t)abs(coeff[blkPosBase + i * trSize + 3]);
> +                    }
> +
>                      do
>                      {
>                          uint32_t blkPos, sig, ctxSig;
> @@ -1681,7 +1691,7 @@
>                              const uint32_t mstate = baseCtx[ctxSig];
>                              const uint32_t mps = mstate & 1;
>                              const uint32_t stateBits =
> g_entropyStateBits[mstate ^ sig];
> -                            uint32_t nextState = (stateBits >> 23) + mps;
> +                            uint32_t nextState = (stateBits >> 24) + mps;
>                              if ((mstate ^ sig) == 1)
>                                  nextState = sig;
>                              X265_CHECK(sbacNext(mstate, sig) ==
> nextState, "nextState check failure\n");
> @@ -1698,39 +1708,13 @@
>                  else
>                  {
>                      X265_CHECK((log2TrSize > 2), "log2TrSize must be more
> than 2 in this path!\n");
> +                    const uint8_t *tabSigCtx =
> table_cnt[(uint32_t)patternSigCtx];
>
> -                    const uint8_t *tabSigCtx =
> table_cnt[(uint32_t)patternSigCtx];
> -                    do
> -                    {
> -                        uint32_t blkPos, sig, ctxSig;
> -                        blkPos =
> g_scan4x4[codingParameters.scanType][scanPosSigOff];
> -                        const uint32_t posZeroMask = (subPosBase +
> scanPosSigOff) ? ~0 : 0;
> -                        sig     = scanFlagMask & 1;
> -                        scanFlagMask >>= 1;
> -                        X265_CHECK((uint32_t)(tmpCoeff[blkPos] != 0) ==
> sig, "sign bit mistake\n");
> -                        if (scanPosSigOff != 0 || subSet == 0 ||
> numNonZero)
> -                        {
> -                            const uint32_t cnt = tabSigCtx[blkPos] +
> offset;
> -                            ctxSig = (cnt + posOffset) & posZeroMask;
> +                    sum =
> primitives.costCoeffNxN(g_scan4x4[codingParameters.scanType],
> &coeff[blkPosBase], (intptr_t)trSize, absCoeff + numNonZero, tabSigCtx,
> scanFlagMask, baseCtx, offset + posOffset, scanPosSigOff, subPosBase);
>
> -                            X265_CHECK(ctxSig ==
> Quant::getSigCtxInc(patternSigCtx, log2TrSize, trSize,
> codingParameters.scan[subPosBase + scanPosSigOff], bIsLuma,
> codingParameters.firstSignificanceMapContext), "sigCtx mistake!\n");;
> -                            //encodeBin(sig, baseCtx[ctxSig]);
> -                            const uint32_t mstate = baseCtx[ctxSig];
> -                            const uint32_t mps = mstate & 1;
> -                            const uint32_t stateBits =
> g_entropyStateBits[mstate ^ sig];
> -                            uint32_t nextState = (stateBits >> 23) + mps;
> -                            if ((mstate ^ sig) == 1)
> -                                nextState = sig;
> -                            X265_CHECK(sbacNext(mstate, sig) ==
> nextState, "nextState check failure\n");
> -                            X265_CHECK(sbacGetEntropyBits(mstate, sig) ==
> (stateBits & 0xFFFFFF), "entropyBits check failure\n");
> -                            baseCtx[ctxSig] = (uint8_t)nextState;
> -                            sum += stateBits;
> -                        }
> -                        absCoeff[numNonZero] = tmpCoeff[blkPos];
> -                        numNonZero += sig;
> -                        scanPosSigOff--;
> -                    }
> -                    while(scanPosSigOff >= 0);
> +#if CHECKED_BUILD || _DEBUG
> +                    numNonZero = coeffNum[subSet];
> +#endif
>                  } // end of non 4x4 path
>                  sum &= 0xFFFFFF;
>
> @@ -2271,28 +2255,6 @@
>      0x0050c, 0x29bab, 0x004c1, 0x2a674, 0x004a7, 0x2aa5e, 0x0046f,
> 0x2b32f, 0x0041f, 0x2c0ad, 0x003e7, 0x2ca8d, 0x003ba, 0x2d323, 0x0010c,
> 0x3bfbb
>  };
>
> -// [8 24] --> [stateMPS BitCost], [stateLPS BitCost]
> -const uint32_t g_entropyStateBits[128] =
> -{
> -    // Corrected table, most notably for last state
> -    0x01007b23, 0x000085f9, 0x020074a0, 0x00008cbc, 0x03006ee4,
> 0x01009354, 0x040067f4, 0x02009c1b,
> -    0x050060b0, 0x0200a62a, 0x06005a9c, 0x0400af5b, 0x0700548d,
> 0x0400b955, 0x08004f56, 0x0500c2a9,
> -    0x09004a87, 0x0600cbf7, 0x0a0045d6, 0x0700d5c3, 0x0b004144,
> 0x0800e01b, 0x0c003d88, 0x0900e937,
> -    0x0d0039e0, 0x0900f2cd, 0x0e003663, 0x0b00fc9e, 0x0f003347,
> 0x0b010600, 0x10003050, 0x0c010f95,
> -    0x11002d4d, 0x0d011a02, 0x12002ad3, 0x0d012333, 0x1300286e,
> 0x0f012cad, 0x14002604, 0x0f0136df,
> -    0x15002425, 0x10013f48, 0x160021f4, 0x100149c4, 0x1700203e,
> 0x1201527b, 0x18001e4d, 0x12015d00,
> -    0x19001c99, 0x130166de, 0x1a001b18, 0x13017017, 0x1b0019a5,
> 0x15017988, 0x1c001841, 0x15018327,
> -    0x1d0016df, 0x16018d50, 0x1e0015d9, 0x16019547, 0x1f00147c,
> 0x1701a083, 0x2000138e, 0x1801a8a3,
> -    0x21001251, 0x1801b418, 0x22001166, 0x1901bd27, 0x23001068,
> 0x1a01c77b, 0x24000f7f, 0x1a01d18e,
> -    0x25000eda, 0x1b01d91a, 0x26000e19, 0x1b01e254, 0x27000d4f,
> 0x1c01ec9a, 0x28000c90, 0x1d01f6e0,
> -    0x29000c01, 0x1d01fef8, 0x2a000b5f, 0x1e0208b1, 0x2b000ab6,
> 0x1e021362, 0x2c000a15, 0x1e021e46,
> -    0x2d000988, 0x1f02285d, 0x2e000934, 0x20022ea8, 0x2f0008a8,
> 0x200239b2, 0x3000081d, 0x21024577,
> -    0x310007c9, 0x21024ce6, 0x32000763, 0x21025663, 0x33000710,
> 0x22025e8f, 0x340006a0, 0x22026a26,
> -    0x35000672, 0x23026f23, 0x360005e8, 0x23027ef8, 0x370005ba,
> 0x230284b5, 0x3800055e, 0x24029057,
> -    0x3900050c, 0x24029bab, 0x3a0004c1, 0x2402a674, 0x3b0004a7,
> 0x2502aa5e, 0x3c00046f, 0x2502b32f,
> -    0x3d00041f, 0x2502c0ad, 0x3e0003e7, 0x2602ca8d, 0x3e0003ba,
> 0x2602d323, 0x3f00010c, 0x3f03bfbb,
> -};
> -
>  const uint8_t g_nextState[128][2] =
>  {
>      { 2, 1 }, { 0, 3 }, { 4, 0 }, { 1, 5 }, { 6, 2 }, { 3, 7 }, { 8, 4 },
> { 5, 9 },
> @@ -2314,3 +2276,26 @@
>  };
>
>  }
> +
> +// [8 24] --> [stateMPS BitCost], [stateLPS BitCost]
> +extern "C" const uint32_t g_entropyStateBits[128] =
> +{
> +    // Corrected table, most notably for last state
> +    0x02007B23, 0x000085F9, 0x040074A0, 0x00008CBC, 0x06006EE4,
> 0x02009354, 0x080067F4, 0x04009C1B,
> +    0x0A0060B0, 0x0400A62A, 0x0C005A9C, 0x0800AF5B, 0x0E00548D,
> 0x0800B955, 0x10004F56, 0x0A00C2A9,
> +    0x12004A87, 0x0C00CBF7, 0x140045D6, 0x0E00D5C3, 0x16004144,
> 0x1000E01B, 0x18003D88, 0x1200E937,
> +    0x1A0039E0, 0x1200F2CD, 0x1C003663, 0x1600FC9E, 0x1E003347,
> 0x16010600, 0x20003050, 0x18010F95,
> +    0x22002D4D, 0x1A011A02, 0x24002AD3, 0x1A012333, 0x2600286E,
> 0x1E012CAD, 0x28002604, 0x1E0136DF,
> +    0x2A002425, 0x20013F48, 0x2C0021F4, 0x200149C4, 0x2E00203E,
> 0x2401527B, 0x30001E4D, 0x24015D00,
> +    0x32001C99, 0x260166DE, 0x34001B18, 0x26017017, 0x360019A5,
> 0x2A017988, 0x38001841, 0x2A018327,
> +    0x3A0016DF, 0x2C018D50, 0x3C0015D9, 0x2C019547, 0x3E00147C,
> 0x2E01A083, 0x4000138E, 0x3001A8A3,
> +    0x42001251, 0x3001B418, 0x44001166, 0x3201BD27, 0x46001068,
> 0x3401C77B, 0x48000F7F, 0x3401D18E,
> +    0x4A000EDA, 0x3601D91A, 0x4C000E19, 0x3601E254, 0x4E000D4F,
> 0x3801EC9A, 0x50000C90, 0x3A01F6E0,
> +    0x52000C01, 0x3A01FEF8, 0x54000B5F, 0x3C0208B1, 0x56000AB6,
> 0x3C021362, 0x58000A15, 0x3C021E46,
> +    0x5A000988, 0x3E02285D, 0x5C000934, 0x40022EA8, 0x5E0008A8,
> 0x400239B2, 0x6000081D, 0x42024577,
> +    0x620007C9, 0x42024CE6, 0x64000763, 0x42025663, 0x66000710,
> 0x44025E8F, 0x680006A0, 0x44026A26,
> +    0x6A000672, 0x46026F23, 0x6C0005E8, 0x46027EF8, 0x6E0005BA,
> 0x460284B5, 0x7000055E, 0x48029057,
> +    0x7200050C, 0x48029BAB, 0x740004C1, 0x4802A674, 0x760004A7,
> 0x4A02AA5E, 0x7800046F, 0x4A02B32F,
> +    0x7A00041F, 0x4A02C0AD, 0x7C0003E7, 0x4C02CA8D, 0x7C0003BA,
> 0x4C02D323, 0x7E00010C, 0x7E03BFBB,
> +};
> +
>
> _______________________________________________
> 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/20150605/b31fced8/attachment-0001.html>


More information about the x265-devel mailing list