Bilinear filter with SSE4.1 intrinsics

I am trying to figure out a reasonably fast bilinear filtering function just for one filtered sample at a time now as an exercise in getting used to using intrinsics - up to SSE41 is fine.

So far I have the following:

inline __m128i DivideBy255_8xUint16(const __m128i value)
{
    //  Blinn 16bit divide by 255 trick but across 8 packed 16bit values
    const __m128i plus128 = _mm_add_epi16(value, _mm_set1_epi16(128));
    const __m128i plus128ThenDivideBy256 = _mm_srli_epi16(plus128, 8);          //  TODO:   Should this be an arithmetic or logical shift or does it matter?
    const __m128i partial = _mm_add_epi16(plus128, plus128ThenDivideBy256);
    const __m128i result = _mm_srli_epi16(partial, 8);                          //  TODO:   Should this be an arithmetic or logical shift or does it matter?


    return result;
}


inline uint32_t BilinearSSE41(const uint8_t* data, uint32_t pitch, uint32_t width, uint32_t height, float u, float v)
{
    //  TODO:   There are probably intrinsics I haven't found yet to avoid using these?
    //  0x80 is high bit set which means zero out that component
    const __m128i unpack_fraction_u_mask = _mm_set_epi8(0x80, 0, 0x80, 0, 0x80, 0, 0x80, 0, 0x80, 0, 0x80, 0, 0x80, 0, 0x80, 0);
    const __m128i unpack_fraction_v_mask = _mm_set_epi8(0x80, 1, 0x80, 1, 0x80, 1, 0x80, 1, 0x80, 1, 0x80, 1, 0x80, 1, 0x80, 1);
    const __m128i unpack_two_texels_mask = _mm_set_epi8(0x80, 7, 0x80, 6, 0x80, 5, 0x80, 4, 0x80, 3, 0x80, 2, 0x80, 1, 0x80, 0);


    //  TODO:   Potentially wasting two channels of operations for now
    const __m128i size = _mm_set_epi32(0, 0, height - 1, width - 1);
    const __m128 uv = _mm_set_ps(0.0f, 0.0f, v, u);

    const __m128 floor_uv_f = _mm_floor_ps(uv);
    const __m128 fraction_uv_f = _mm_sub_ps(uv, floor_uv_f);
    const __m128 fraction255_uv_f = _mm_mul_ps(fraction_uv_f, _mm_set_ps1(255.0f));
    const __m128i fraction255_uv_i = _mm_cvttps_epi32(fraction255_uv_f);    //  TODO:   Did this get rounded correctly?

    const __m128i fraction255_u_i = _mm_shuffle_epi8(fraction255_uv_i, unpack_fraction_u_mask); //  Splat fraction_u*255 across all 16 bit words
    const __m128i fraction255_v_i = _mm_shuffle_epi8(fraction255_uv_i, unpack_fraction_v_mask); //  Splat fraction_v*255 across all 16 bit words

    const __m128i inverse_fraction255_u_i = _mm_sub_epi16(_mm_set1_epi16(255), fraction255_u_i);
    const __m128i inverse_fraction255_v_i = _mm_sub_epi16(_mm_set1_epi16(255), fraction255_v_i);

    const __m128i floor_uv_i = _mm_cvttps_epi32(floor_uv_f);
    const __m128i clipped_floor_uv_i = _mm_min_epu32(floor_uv_i, size); //  TODO:   I haven't clamped this probably if uv was less than zero yet...


    //  TODO:   Calculating the addresses in the SSE register set would maybe be better

    int u0 = _mm_extract_epi32(floor_uv_i, 0);
    int v0 = _mm_extract_epi32(floor_uv_i, 1);


    const uint8_t* row = data + (u0<<2) + pitch*v0;


    const __m128i row0_packed = _mm_loadl_epi64((const __m128i*)data);
    const __m128i row0 = _mm_shuffle_epi8(row0_packed, unpack_two_texels_mask);

    const __m128i row1_packed = _mm_loadl_epi64((const __m128i*)(data + pitch));
    const __m128i row1 = _mm_shuffle_epi8(row1_packed, unpack_two_texels_mask);


    //  Compute (row0*fraction)/255 + row1*(255 - fraction)/255 - probably slight precision loss across addition!
    const __m128i vlerp0 = DivideBy255_8xUint16(_mm_mullo_epi16(row0, fraction255_v_i));
    const __m128i vlerp1 = DivideBy255_8xUint16(_mm_mullo_epi16(row1, inverse_fraction255_v_i));
    const __m128i vlerp = _mm_adds_epi16(vlerp0, vlerp1);

    const __m128i hlerp0 = DivideBy255_8xUint16(_mm_mullo_epi16(vlerp, fraction255_u_i));
    const __m128i hlerp1 = DivideBy255_8xUint16(_mm_srli_si128(_mm_mullo_epi16(vlerp, inverse_fraction255_u_i), 16 - 2*4));
    const __m128i hlerp = _mm_adds_epi16(hlerp0, hlerp1);


    //  Pack down to 8bit from 16bit components and return 32bit ARGB result
    return _mm_extract_epi32(_mm_packus_epi16(hlerp, hlerp), 0);
}

The code assumes the image data is ARGB8 and has an extra column and row to handle edge cases without having to branch.

I am after advice on what instructions I can use to bring down the size of this gangly mess and of course how 开发者_C百科it can be improved to run faster!

Thanks :)

Nothing specific to say about your code. But I wrote my own Bilinear scaling code using SSE2. See the StackOverflow question Help me improve some more SSE2 code for more details.

In my code I calculate the horizontal and vertical fractions and indexes first rather than per pixel. I think this is faster.

My code under core2 cpus seems to be memory limited rather than cpu so not doing the precalc might be faster.

Noticed your comment "TODO: Should this be an arithmetic or logical shift or does it matter?"

Arithmetic shift is for signed integers. Logical shift is for unsigned integers.

    0x80000000 >> 4 is 0xf8000000 // Arithmetic shift
    0x80000000 >> 4 is 0x08000000 // Logical shift