Function transform_8u

deps/opencv/modules/core/src/matmul.cpp:1373–1492

#endif

static void
transform_8u( const uchar* src, uchar* dst, const float* m, int len, int scn, int dcn )
{
#if CV_SSE2
    const int BITS = 10, SCALE = 1 << BITS;
    const float MAX_M = (float)(1 << (15 - BITS));

    if( USE_SSE2 && scn == 3 && dcn == 3 &&
        std::abs(m[0]) < MAX_M && std::abs(m[1]) < MAX_M && std::abs(m[2]) < MAX_M && std::abs(m[3]) < MAX_M*256 &&
        std::abs(m[4]) < MAX_M && std::abs(m[5]) < MAX_M && std::abs(m[6]) < MAX_M && std::abs(m[7]) < MAX_M*256 &&
        std::abs(m[8]) < MAX_M && std::abs(m[9]) < MAX_M && std::abs(m[10]) < MAX_M && std::abs(m[11]) < MAX_M*256 )
    {
        // faster fixed-point transformation
        short m00 = saturate_cast<short>(m[0]*SCALE), m01 = saturate_cast<short>(m[1]*SCALE),
            m02 = saturate_cast<short>(m[2]*SCALE), m10 = saturate_cast<short>(m[4]*SCALE),
            m11 = saturate_cast<short>(m[5]*SCALE), m12 = saturate_cast<short>(m[6]*SCALE),
            m20 = saturate_cast<short>(m[8]*SCALE), m21 = saturate_cast<short>(m[9]*SCALE),
            m22 = saturate_cast<short>(m[10]*SCALE);
        int m03 = saturate_cast<int>((m[3]+0.5f)*SCALE), m13 = saturate_cast<int>((m[7]+0.5f)*SCALE ),
            m23 = saturate_cast<int>((m[11]+0.5f)*SCALE);

        __m128i m0 = _mm_setr_epi16(0, m00, m01, m02, m00, m01, m02, 0);
        __m128i m1 = _mm_setr_epi16(0, m10, m11, m12, m10, m11, m12, 0);
        __m128i m2 = _mm_setr_epi16(0, m20, m21, m22, m20, m21, m22, 0);
        __m128i m3 = _mm_setr_epi32(m03, m13, m23, 0);
        int x = 0;

        for( ; x <= (len - 8)*3; x += 8*3 )
        {
            __m128i z = _mm_setzero_si128(), t0, t1, t2, r0, r1;
            __m128i v0 = _mm_loadl_epi64((const __m128i*)(src + x));
            __m128i v1 = _mm_loadl_epi64((const __m128i*)(src + x + 8));
            __m128i v2 = _mm_loadl_epi64((const __m128i*)(src + x + 16)), v3;
            v0 = _mm_unpacklo_epi8(v0, z); // b0 g0 r0 b1 g1 r1 b2 g2
            v1 = _mm_unpacklo_epi8(v1, z); // r2 b3 g3 r3 b4 g4 r4 b5
            v2 = _mm_unpacklo_epi8(v2, z); // g5 r5 b6 g6 r6 b7 g7 r7

            v3 = _mm_srli_si128(v2, 2); // ? b6 g6 r6 b7 g7 r7 0
            v2 = _mm_or_si128(_mm_slli_si128(v2, 10), _mm_srli_si128(v1, 6)); // ? b4 g4 r4 b5 g5 r5 ?
            v1 = _mm_or_si128(_mm_slli_si128(v1, 6), _mm_srli_si128(v0, 10)); // ? b2 g2 r2 b3 g3 r3 ?
            v0 = _mm_slli_si128(v0, 2); // 0 b0 g0 r0 b1 g1 r1 ?

            // process pixels 0 & 1
            t0 = _mm_madd_epi16(v0, m0); // a0 b0 a1 b1
            t1 = _mm_madd_epi16(v0, m1); // c0 d0 c1 d1
            t2 = _mm_madd_epi16(v0, m2); // e0 f0 e1 f1
            v0 = _mm_unpacklo_epi32(t0, t1); // a0 c0 b0 d0
            t0 = _mm_unpackhi_epi32(t0, t1); // a1 b1 c1 d1
            t1 = _mm_unpacklo_epi32(t2, z);  // e0 0 f0 0
            t2 = _mm_unpackhi_epi32(t2, z);  // e1 0 f1 0
            r0 = _mm_add_epi32(_mm_add_epi32(_mm_unpacklo_epi64(v0, t1), _mm_unpackhi_epi64(v0,t1)), m3); // B0 G0 R0 0
            r1 = _mm_add_epi32(_mm_add_epi32(_mm_unpacklo_epi64(t0, t2), _mm_unpackhi_epi64(t0,t2)), m3); // B1 G1 R1 0
            r0 = _mm_srai_epi32(r0, BITS);
            r1 = _mm_srai_epi32(r1, BITS);
            v0 = _mm_packus_epi16(_mm_packs_epi32(_mm_slli_si128(r0, 4), r1), z); // 0 B0 G0 R0 B1 G1 R1 0

            // process pixels 2 & 3
            t0 = _mm_madd_epi16(v1, m0); // a0 b0 a1 b1