ScaleRowDown2Box_Odd_SSSE3 for odd source width subsampling.

[GoogleCodeExporter]

ScaleRowDown2Box_SSSE3 does 32x2 to 16x1
ScaleRowDown2Box_AVX2 does 64x2 to 32x1
ScaleRowDown2Box_NEON does 32x2 to 16x1
where destination width is multiple of 16

Any variations support odd destination width, but source is a multiple of 2.

Suggest doing an Odd variation of Any that does odd source width.
ScaleRowDown2Box_Odd_SSSE3 already exist for I420Blend subsampling.

For reference this is SSSE3 gcc code:

void ScaleRowDown2Box_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
                            uint8* dst_ptr, int dst_width) {
  asm volatile (
    "pcmpeqb    %%xmm4,%%xmm4                  \n"
    "psrlw      $0xf,%%xmm4                    \n"
    "packuswb   %%xmm4,%%xmm4                  \n"
    "pxor       %%xmm5,%%xmm5                  \n"

    LABELALIGN
  "1:                                          \n"
    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
    MEMOPREG(movdqu,0x00,0,3,1,xmm2)           //  movdqu  (%0,%3,1),%%xmm2
    MEMOPREG(movdqu,0x10,0,3,1,xmm3)           //  movdqu  0x10(%0,%3,1),%%xmm3
    "lea       " MEMLEA(0x20,0) ",%0           \n"
    "pmaddubsw  %%xmm4,%%xmm0                  \n"
    "pmaddubsw  %%xmm4,%%xmm1                  \n"
    "pmaddubsw  %%xmm4,%%xmm2                  \n"
    "pmaddubsw  %%xmm4,%%xmm3                  \n"
    "paddw      %%xmm2,%%xmm0                  \n"
    "paddw      %%xmm3,%%xmm1                  \n"
    "psrlw      $0x1,%%xmm0                    \n"
    "psrlw      $0x1,%%xmm1                    \n"
    "pavgw      %%xmm5,%%xmm0                  \n"
    "pavgw      %%xmm5,%%xmm1                  \n"
    "packuswb   %%xmm1,%%xmm0                  \n"
    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
    "lea       " MEMLEA(0x10,1) ",%1           \n"
    "sub       $0x10,%2                        \n"
    "jg        1b                              \n"
  : "+r"(src_ptr),    // %0
    "+r"(dst_ptr),    // %1
    "+r"(dst_width)   // %2
  : "r"((intptr_t)(src_stride))   // %3
  : "memory", "cc", NACL_R14
    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
  );
}

// Fixed scale down for odd source width.  Used by I420Blend subsampling.
// Since dst_width is (width + 1) / 2, this function scales one less pixel
// and copies the last pixel.
#define SDODD(NAMEANY, SCALEROWDOWN_SIMD, SCALEROWDOWN_C, FACTOR, BPP, MASK)   \
    void NAMEANY(const uint8* src_ptr, ptrdiff_t src_stride,                   \
                 uint8* dst_ptr, int dst_width) {                              \
      int r = (int)((unsigned int)(dst_width - 1) % (MASK + 1));               \
      int n = dst_width - r;                                                   \
      if (n > 0) {                                                             \
        SCALEROWDOWN_SIMD(src_ptr, src_stride, dst_ptr, n);                    \
      }                                                                        \
      SCALEROWDOWN_C(src_ptr + (n * FACTOR) * BPP, src_stride,                 \
                     dst_ptr + n * BPP, r);                                    \
    }

SDANY(ScaleRowDown2Box_Any_AVX2, ScaleRowDown2Box_AVX2, ScaleRowDown2Box_C,
      2, 1, 31)
SDODD(ScaleRowDown2Box_Odd_AVX2, ScaleRowDown2Box_AVX2, ScaleRowDown2Box_Odd_C,
      2, 1, 31)
#endif

Original issue reported on code.google.com by fbarch...@google.com on 26 Jan 2016 at 2:01