/* Routine optimized for unshuffling a buffer for a type size of 16 bytes. */
static void
unshuffle16(uint8_t* dest, uint8_t* orig, size_t size)
{
size_t i, j, k;
size_t neblock, numof16belem;
__m128i xmm1[16], xmm2[16];
neblock = size / 16;
numof16belem = neblock / 16;
for (i = 0, k = 0; i < numof16belem; i++, k += 16) {
/* Load the first 128 bytes in 16 XMM registrers */
for (j = 0; j < 16; j++) {
xmm1[j] = ((__m128i *)orig)[j*numof16belem+i];
}
/* Shuffle bytes */
for (j = 0; j < 8; j++) {
/* Compute the low 32 bytes */
xmm2[j] = _mm_unpacklo_epi8(xmm1[j*2], xmm1[j*2+1]);
/* Compute the hi 32 bytes */
xmm2[8+j] = _mm_unpackhi_epi8(xmm1[j*2], xmm1[j*2+1]);
}
/* Shuffle 2-byte words */
for (j = 0; j < 8; j++) {
/* Compute the low 32 bytes */
xmm1[j] = _mm_unpacklo_epi16(xmm2[j*2], xmm2[j*2+1]);
/* Compute the hi 32 bytes */
xmm1[8+j] = _mm_unpackhi_epi16(xmm2[j*2], xmm2[j*2+1]);
}
/* Shuffle 4-byte dwords */
for (j = 0; j < 8; j++) {
/* Compute the low 32 bytes */
xmm2[j] = _mm_unpacklo_epi32(xmm1[j*2], xmm1[j*2+1]);
/* Compute the hi 32 bytes */
xmm2[8+j] = _mm_unpackhi_epi32(xmm1[j*2], xmm1[j*2+1]);
}
/* Shuffle 8-byte qwords */
for (j = 0; j < 8; j++) {
/* Compute the low 32 bytes */
xmm1[j] = _mm_unpacklo_epi64(xmm2[j*2], xmm2[j*2+1]);
/* Compute the hi 32 bytes */
xmm1[8+j] = _mm_unpackhi_epi64(xmm2[j*2], xmm2[j*2+1]);
}
/* Store the result vectors in proper order */
((__m128i *)dest)[k+0] = xmm1[0];
((__m128i *)dest)[k+1] = xmm1[8];
((__m128i *)dest)[k+2] = xmm1[4];
((__m128i *)dest)[k+3] = xmm1[12];
((__m128i *)dest)[k+4] = xmm1[2];
((__m128i *)dest)[k+5] = xmm1[10];
((__m128i *)dest)[k+6] = xmm1[6];
((__m128i *)dest)[k+7] = xmm1[14];
((__m128i *)dest)[k+8] = xmm1[1];
((__m128i *)dest)[k+9] = xmm1[9];
((__m128i *)dest)[k+10] = xmm1[5];
((__m128i *)dest)[k+11] = xmm1[13];
((__m128i *)dest)[k+12] = xmm1[3];
((__m128i *)dest)[k+13] = xmm1[11];
((__m128i *)dest)[k+14] = xmm1[7];
((__m128i *)dest)[k+15] = xmm1[15];
}
}
void EmitColorIndices_Intrinsics( const byte *colorBlock, const byte *minColor, const byte *maxColor, byte *&outData )
{
ALIGN16( byte color0[16] );
ALIGN16( byte color1[16] );
ALIGN16( byte color2[16] );
ALIGN16( byte color3[16] );
ALIGN16( byte result[16] );
// mov esi, maxColor
// mov edi, minColor
__m128i t0, t1, t2, t3, t4, t5, t6, t7;
t7 = _mm_setzero_si128();
//t7 = _mm_xor_si128(t7, t7);
_mm_store_si128 ( (__m128i*) &result, t7 );
//t0 = _mm_load_si128 ( (__m128i*) maxColor );
t0 = _mm_cvtsi32_si128( *(int*)maxColor);
// Bitwise AND
__m128i tt = _mm_load_si128 ( (__m128i*) SIMD_SSE2_byte_colorMask );
t0 = _mm_and_si128(t0, tt);
t0 = _mm_unpacklo_epi8(t0, t7);
t4 = _mm_shufflelo_epi16( t0, R_SHUFFLE_D( 0, 3, 2, 3 ));
t5 = _mm_shufflelo_epi16( t0, R_SHUFFLE_D( 3, 1, 3, 3 ));
t4 = _mm_srli_epi16(t4, 5);
t5 = _mm_srli_epi16(t5, 6);
// Bitwise Logical OR
t0 = _mm_or_si128(t0, t4);
t0 = _mm_or_si128(t0, t5); // t0 contains color0 in 565
//t1 = _mm_load_si128 ( (__m128i*) minColor );
t1 = _mm_cvtsi32_si128( *(int*)minColor);
t1 = _mm_and_si128(t1, tt);
t1 = _mm_unpacklo_epi8(t1, t7);
t4 = _mm_shufflelo_epi16( t1, R_SHUFFLE_D( 0, 3, 2, 3 ));
t5 = _mm_shufflelo_epi16( t1, R_SHUFFLE_D( 3, 1, 3, 3 ));
t4 = _mm_srli_epi16(t4, 5);
t5 = _mm_srli_epi16(t5, 6);
t1 = _mm_or_si128(t1, t4);
t1 = _mm_or_si128(t1, t5); // t1 contains color1 in 565
t2 = t0;
t2 = _mm_packus_epi16(t2, t7);
t2 = _mm_shuffle_epi32( t2, R_SHUFFLE_D( 0, 1, 0, 1 ));
_mm_store_si128 ( (__m128i*) &color0, t2 );
t6 = t0;
t6 = _mm_add_epi16(t6, t0);
t6 = _mm_add_epi16(t6, t1);
// Multiply Packed Signed Integers and Store High Result
__m128i tw3 = _mm_load_si128 ( (__m128i*) SIMD_SSE2_word_div_by_3 );
t6 = _mm_mulhi_epi16(t6, tw3);
t6 = _mm_packus_epi16(t6, t7);
t6 = _mm_shuffle_epi32( t6, R_SHUFFLE_D( 0, 1, 0, 1 ));
_mm_store_si128 ( (__m128i*) &color2, t6 );
t3 = t1;
t3 = _mm_packus_epi16(t3, t7);
t3 = _mm_shuffle_epi32( t3, R_SHUFFLE_D( 0, 1, 0, 1 ));
_mm_store_si128 ( (__m128i*) &color1, t3 );
t1 = _mm_add_epi16(t1, t1);
t0 = _mm_add_epi16(t0, t1);
t0 = _mm_mulhi_epi16(t0, tw3);
t0 = _mm_packus_epi16(t0, t7);
t0 = _mm_shuffle_epi32( t0, R_SHUFFLE_D( 0, 1, 0, 1 ));
_mm_store_si128 ( (__m128i*) &color3, t0 );
__m128i w0 = _mm_load_si128 ( (__m128i*) SIMD_SSE2_word_0);
__m128i w1 = _mm_load_si128 ( (__m128i*) SIMD_SSE2_word_1);
__m128i w2 = _mm_load_si128 ( (__m128i*) SIMD_SSE2_word_2);
// mov eax, 32
// mov esi, colorBlock
int x = 32;
//const byte *c = colorBlock;
while (x >= 0)
{
t3 = _mm_loadl_epi64( (__m128i*) (colorBlock+x+0));
t3 = _mm_shuffle_epi32( t3, R_SHUFFLE_D( 0, 2, 1, 3 ));
t5 = _mm_loadl_epi64( (__m128i*) (colorBlock+x+8));
t5 = _mm_shuffle_epi32( t5, R_SHUFFLE_D( 0, 2, 1, 3 ));
t0 = t3;
t6 = t5;
// Compute Sum of Absolute Difference
__m128i c0 = _mm_load_si128 ( (__m128i*) color0 );
t0 = _mm_sad_epu8(t0, c0);
t6 = _mm_sad_epu8(t6, c0);
// Pack with Signed Saturation
t0 = _mm_packs_epi32 (t0, t6);
t1 = t3;
t6 = t5;
__m128i c1 = _mm_load_si128 ( (__m128i*) color1 );
t1 = _mm_sad_epu8(t1, c1);
t6 = _mm_sad_epu8(t6, c1);
t1 = _mm_packs_epi32 (t1, t6);
t2 = t3;
t6 = t5;
__m128i c2 = _mm_load_si128 ( (__m128i*) color2 );
t2 = _mm_sad_epu8(t2, c2);
t6 = _mm_sad_epu8(t6, c2);
t2 = _mm_packs_epi32 (t2, t6);
__m128i c3 = _mm_load_si128 ( (__m128i*) color3 );
t3 = _mm_sad_epu8(t3, c3);
t5 = _mm_sad_epu8(t5, c3);
t3 = _mm_packs_epi32 (t3, t5);
t4 = _mm_loadl_epi64( (__m128i*) (colorBlock+x+16));
t4 = _mm_shuffle_epi32( t4, R_SHUFFLE_D( 0, 2, 1, 3 ));
t5 = _mm_loadl_epi64( (__m128i*) (colorBlock+x+24));
t5 = _mm_shuffle_epi32( t5, R_SHUFFLE_D( 0, 2, 1, 3 ));
t6 = t4;
t7 = t5;
t6 = _mm_sad_epu8(t6, c0);
t7 = _mm_sad_epu8(t7, c0);
t6 = _mm_packs_epi32 (t6, t7);
t0 = _mm_packs_epi32 (t0, t6); // d0
t6 = t4;
t7 = t5;
t6 = _mm_sad_epu8(t6, c1);
t7 = _mm_sad_epu8(t7, c1);
t6 = _mm_packs_epi32 (t6, t7);
t1 = _mm_packs_epi32 (t1, t6); // d1
t6 = t4;
t7 = t5;
t6 = _mm_sad_epu8(t6, c2);
t7 = _mm_sad_epu8(t7, c2);
t6 = _mm_packs_epi32 (t6, t7);
t2 = _mm_packs_epi32 (t2, t6); // d2
t4 = _mm_sad_epu8(t4, c3);
t5 = _mm_sad_epu8(t5, c3);
t4 = _mm_packs_epi32 (t4, t5);
t3 = _mm_packs_epi32 (t3, t4); // d3
t7 = _mm_load_si128 ( (__m128i*) result );
t7 = _mm_slli_epi32( t7, 16);
t4 = t0;
t5 = t1;
// Compare Packed Signed Integers for Greater Than
t0 = _mm_cmpgt_epi16(t0, t3); // b0
t1 = _mm_cmpgt_epi16(t1, t2); // b1
t4 = _mm_cmpgt_epi16(t4, t2); // b2
t5 = _mm_cmpgt_epi16(t5, t3); // b3
t2 = _mm_cmpgt_epi16(t2, t3); // b4
t4 = _mm_and_si128(t4, t1); // x0
t5 = _mm_and_si128(t5, t0); // x1
t2 = _mm_and_si128(t2, t0); // x2
t4 = _mm_or_si128(t4, t5);
t2 = _mm_and_si128(t2, w1);
t4 = _mm_and_si128(t4, w2);
t2 = _mm_or_si128(t2, t4);
t5 = _mm_shuffle_epi32( t2, R_SHUFFLE_D( 2, 3, 0, 1 ));
// Unpack Low Data
t2 = _mm_unpacklo_epi16 ( t2, w0);
t5 = _mm_unpacklo_epi16 ( t5, w0);
//t5 = _mm_slli_si128 ( t5, 8);
t5 = _mm_slli_epi32( t5, 8);
t7 = _mm_or_si128(t7, t5);
t7 = _mm_or_si128(t7, t2);
_mm_store_si128 ( (__m128i*) &result, t7 );
x -=32;
}
t4 = _mm_shuffle_epi32( t7, R_SHUFFLE_D( 1, 2, 3, 0 ));
t5 = _mm_shuffle_epi32( t7, R_SHUFFLE_D( 2, 3, 0, 1 ));
t6 = _mm_shuffle_epi32( t7, R_SHUFFLE_D( 3, 0, 1, 2 ));
t4 = _mm_slli_epi32 ( t4, 2);
t5 = _mm_slli_epi32 ( t5, 4);
t6 = _mm_slli_epi32 ( t6, 6);
t7 = _mm_or_si128(t7, t4);
t7 = _mm_or_si128(t7, t5);
t7 = _mm_or_si128(t7, t6);
//_mm_store_si128 ( (__m128i*) outData, t7 );
int r = _mm_cvtsi128_si32 (t7);
memcpy(outData, &r, 4); // Anything better ?
outData += 4;
}
// Does one or two inverse transforms.
static void ITransform(const uint8_t* ref, const int16_t* in, uint8_t* dst,
int do_two) {
// This implementation makes use of 16-bit fixed point versions of two
// multiply constants:
// K1 = sqrt(2) * cos (pi/8) ~= 85627 / 2^16
// K2 = sqrt(2) * sin (pi/8) ~= 35468 / 2^16
//
// To be able to use signed 16-bit integers, we use the following trick to
// have constants within range:
// - Associated constants are obtained by subtracting the 16-bit fixed point
// version of one:
// k = K - (1 << 16) => K = k + (1 << 16)
// K1 = 85267 => k1 = 20091
// K2 = 35468 => k2 = -30068
// - The multiplication of a variable by a constant become the sum of the
// variable and the multiplication of that variable by the associated
// constant:
// (x * K) >> 16 = (x * (k + (1 << 16))) >> 16 = ((x * k ) >> 16) + x
const __m128i k1 = _mm_set1_epi16(20091);
const __m128i k2 = _mm_set1_epi16(-30068);
__m128i T0, T1, T2, T3;
// Load and concatenate the transform coefficients (we'll do two inverse
// transforms in parallel). In the case of only one inverse transform, the
// second half of the vectors will just contain random value we'll never
// use nor store.
__m128i in0, in1, in2, in3;
{
in0 = _mm_loadl_epi64((const __m128i*)&in[0]);
in1 = _mm_loadl_epi64((const __m128i*)&in[4]);
in2 = _mm_loadl_epi64((const __m128i*)&in[8]);
in3 = _mm_loadl_epi64((const __m128i*)&in[12]);
// a00 a10 a20 a30 x x x x
// a01 a11 a21 a31 x x x x
// a02 a12 a22 a32 x x x x
// a03 a13 a23 a33 x x x x
if (do_two) {
const __m128i inB0 = _mm_loadl_epi64((const __m128i*)&in[16]);
const __m128i inB1 = _mm_loadl_epi64((const __m128i*)&in[20]);
const __m128i inB2 = _mm_loadl_epi64((const __m128i*)&in[24]);
const __m128i inB3 = _mm_loadl_epi64((const __m128i*)&in[28]);
in0 = _mm_unpacklo_epi64(in0, inB0);
in1 = _mm_unpacklo_epi64(in1, inB1);
in2 = _mm_unpacklo_epi64(in2, inB2);
in3 = _mm_unpacklo_epi64(in3, inB3);
// a00 a10 a20 a30 b00 b10 b20 b30
// a01 a11 a21 a31 b01 b11 b21 b31
// a02 a12 a22 a32 b02 b12 b22 b32
// a03 a13 a23 a33 b03 b13 b23 b33
}
}
// Vertical pass and subsequent transpose.
{
// First pass, c and d calculations are longer because of the "trick"
// multiplications.
const __m128i a = _mm_add_epi16(in0, in2);
const __m128i b = _mm_sub_epi16(in0, in2);
// c = MUL(in1, K2) - MUL(in3, K1) = MUL(in1, k2) - MUL(in3, k1) + in1 - in3
const __m128i c1 = _mm_mulhi_epi16(in1, k2);
const __m128i c2 = _mm_mulhi_epi16(in3, k1);
const __m128i c3 = _mm_sub_epi16(in1, in3);
const __m128i c4 = _mm_sub_epi16(c1, c2);
const __m128i c = _mm_add_epi16(c3, c4);
// d = MUL(in1, K1) + MUL(in3, K2) = MUL(in1, k1) + MUL(in3, k2) + in1 + in3
const __m128i d1 = _mm_mulhi_epi16(in1, k1);
const __m128i d2 = _mm_mulhi_epi16(in3, k2);
const __m128i d3 = _mm_add_epi16(in1, in3);
const __m128i d4 = _mm_add_epi16(d1, d2);
const __m128i d = _mm_add_epi16(d3, d4);
// Second pass.
const __m128i tmp0 = _mm_add_epi16(a, d);
const __m128i tmp1 = _mm_add_epi16(b, c);
const __m128i tmp2 = _mm_sub_epi16(b, c);
const __m128i tmp3 = _mm_sub_epi16(a, d);
// Transpose the two 4x4.
// a00 a01 a02 a03 b00 b01 b02 b03
// a10 a11 a12 a13 b10 b11 b12 b13
// a20 a21 a22 a23 b20 b21 b22 b23
// a30 a31 a32 a33 b30 b31 b32 b33
const __m128i transpose0_0 = _mm_unpacklo_epi16(tmp0, tmp1);
const __m128i transpose0_1 = _mm_unpacklo_epi16(tmp2, tmp3);
const __m128i transpose0_2 = _mm_unpackhi_epi16(tmp0, tmp1);
const __m128i transpose0_3 = _mm_unpackhi_epi16(tmp2, tmp3);
// a00 a10 a01 a11 a02 a12 a03 a13
// a20 a30 a21 a31 a22 a32 a23 a33
// b00 b10 b01 b11 b02 b12 b03 b13
// b20 b30 b21 b31 b22 b32 b23 b33
const __m128i transpose1_0 = _mm_unpacklo_epi32(transpose0_0, transpose0_1);
const __m128i transpose1_1 = _mm_unpacklo_epi32(transpose0_2, transpose0_3);
const __m128i transpose1_2 = _mm_unpackhi_epi32(transpose0_0, transpose0_1);
const __m128i transpose1_3 = _mm_unpackhi_epi32(transpose0_2, transpose0_3);
// a00 a10 a20 a30 a01 a11 a21 a31
// b00 b10 b20 b30 b01 b11 b21 b31
// a02 a12 a22 a32 a03 a13 a23 a33
// b02 b12 a22 b32 b03 b13 b23 b33
T0 = _mm_unpacklo_epi64(transpose1_0, transpose1_1);
T1 = _mm_unpackhi_epi64(transpose1_0, transpose1_1);
T2 = _mm_unpacklo_epi64(transpose1_2, transpose1_3);
T3 = _mm_unpackhi_epi64(transpose1_2, transpose1_3);
// a00 a10 a20 a30 b00 b10 b20 b30
// a01 a11 a21 a31 b01 b11 b21 b31
// a02 a12 a22 a32 b02 b12 b22 b32
// a03 a13 a23 a33 b03 b13 b23 b33
}
// Horizontal pass and subsequent transpose.
{
// First pass, c and d calculations are longer because of the "trick"
// multiplications.
const __m128i four = _mm_set1_epi16(4);
const __m128i dc = _mm_add_epi16(T0, four);
const __m128i a = _mm_add_epi16(dc, T2);
const __m128i b = _mm_sub_epi16(dc, T2);
// c = MUL(T1, K2) - MUL(T3, K1) = MUL(T1, k2) - MUL(T3, k1) + T1 - T3
const __m128i c1 = _mm_mulhi_epi16(T1, k2);
const __m128i c2 = _mm_mulhi_epi16(T3, k1);
const __m128i c3 = _mm_sub_epi16(T1, T3);
const __m128i c4 = _mm_sub_epi16(c1, c2);
const __m128i c = _mm_add_epi16(c3, c4);
// d = MUL(T1, K1) + MUL(T3, K2) = MUL(T1, k1) + MUL(T3, k2) + T1 + T3
const __m128i d1 = _mm_mulhi_epi16(T1, k1);
const __m128i d2 = _mm_mulhi_epi16(T3, k2);
const __m128i d3 = _mm_add_epi16(T1, T3);
const __m128i d4 = _mm_add_epi16(d1, d2);
const __m128i d = _mm_add_epi16(d3, d4);
// Second pass.
const __m128i tmp0 = _mm_add_epi16(a, d);
const __m128i tmp1 = _mm_add_epi16(b, c);
const __m128i tmp2 = _mm_sub_epi16(b, c);
const __m128i tmp3 = _mm_sub_epi16(a, d);
const __m128i shifted0 = _mm_srai_epi16(tmp0, 3);
const __m128i shifted1 = _mm_srai_epi16(tmp1, 3);
const __m128i shifted2 = _mm_srai_epi16(tmp2, 3);
const __m128i shifted3 = _mm_srai_epi16(tmp3, 3);
// Transpose the two 4x4.
// a00 a01 a02 a03 b00 b01 b02 b03
// a10 a11 a12 a13 b10 b11 b12 b13
// a20 a21 a22 a23 b20 b21 b22 b23
// a30 a31 a32 a33 b30 b31 b32 b33
const __m128i transpose0_0 = _mm_unpacklo_epi16(shifted0, shifted1);
const __m128i transpose0_1 = _mm_unpacklo_epi16(shifted2, shifted3);
const __m128i transpose0_2 = _mm_unpackhi_epi16(shifted0, shifted1);
const __m128i transpose0_3 = _mm_unpackhi_epi16(shifted2, shifted3);
// a00 a10 a01 a11 a02 a12 a03 a13
// a20 a30 a21 a31 a22 a32 a23 a33
// b00 b10 b01 b11 b02 b12 b03 b13
// b20 b30 b21 b31 b22 b32 b23 b33
const __m128i transpose1_0 = _mm_unpacklo_epi32(transpose0_0, transpose0_1);
const __m128i transpose1_1 = _mm_unpacklo_epi32(transpose0_2, transpose0_3);
const __m128i transpose1_2 = _mm_unpackhi_epi32(transpose0_0, transpose0_1);
const __m128i transpose1_3 = _mm_unpackhi_epi32(transpose0_2, transpose0_3);
// a00 a10 a20 a30 a01 a11 a21 a31
// b00 b10 b20 b30 b01 b11 b21 b31
// a02 a12 a22 a32 a03 a13 a23 a33
// b02 b12 a22 b32 b03 b13 b23 b33
T0 = _mm_unpacklo_epi64(transpose1_0, transpose1_1);
T1 = _mm_unpackhi_epi64(transpose1_0, transpose1_1);
T2 = _mm_unpacklo_epi64(transpose1_2, transpose1_3);
T3 = _mm_unpackhi_epi64(transpose1_2, transpose1_3);
// a00 a10 a20 a30 b00 b10 b20 b30
// a01 a11 a21 a31 b01 b11 b21 b31
// a02 a12 a22 a32 b02 b12 b22 b32
// a03 a13 a23 a33 b03 b13 b23 b33
}
// Add inverse transform to 'ref' and store.
{
const __m128i zero = _mm_setzero_si128();
// Load the reference(s).
__m128i ref0, ref1, ref2, ref3;
if (do_two) {
// Load eight bytes/pixels per line.
ref0 = _mm_loadl_epi64((const __m128i*)&ref[0 * BPS]);
ref1 = _mm_loadl_epi64((const __m128i*)&ref[1 * BPS]);
ref2 = _mm_loadl_epi64((const __m128i*)&ref[2 * BPS]);
ref3 = _mm_loadl_epi64((const __m128i*)&ref[3 * BPS]);
} else {
// Load four bytes/pixels per line.
ref0 = _mm_cvtsi32_si128(*(const int*)&ref[0 * BPS]);
ref1 = _mm_cvtsi32_si128(*(const int*)&ref[1 * BPS]);
ref2 = _mm_cvtsi32_si128(*(const int*)&ref[2 * BPS]);
ref3 = _mm_cvtsi32_si128(*(const int*)&ref[3 * BPS]);
}
// Convert to 16b.
ref0 = _mm_unpacklo_epi8(ref0, zero);
ref1 = _mm_unpacklo_epi8(ref1, zero);
ref2 = _mm_unpacklo_epi8(ref2, zero);
ref3 = _mm_unpacklo_epi8(ref3, zero);
// Add the inverse transform(s).
ref0 = _mm_add_epi16(ref0, T0);
ref1 = _mm_add_epi16(ref1, T1);
ref2 = _mm_add_epi16(ref2, T2);
ref3 = _mm_add_epi16(ref3, T3);
// Unsigned saturate to 8b.
ref0 = _mm_packus_epi16(ref0, ref0);
ref1 = _mm_packus_epi16(ref1, ref1);
ref2 = _mm_packus_epi16(ref2, ref2);
ref3 = _mm_packus_epi16(ref3, ref3);
// Store the results.
if (do_two) {
// Store eight bytes/pixels per line.
_mm_storel_epi64((__m128i*)&dst[0 * BPS], ref0);
_mm_storel_epi64((__m128i*)&dst[1 * BPS], ref1);
_mm_storel_epi64((__m128i*)&dst[2 * BPS], ref2);
_mm_storel_epi64((__m128i*)&dst[3 * BPS], ref3);
} else {
// Store four bytes/pixels per line.
*((int32_t *)&dst[0 * BPS]) = _mm_cvtsi128_si32(ref0);
*((int32_t *)&dst[1 * BPS]) = _mm_cvtsi128_si32(ref1);
*((int32_t *)&dst[2 * BPS]) = _mm_cvtsi128_si32(ref2);
*((int32_t *)&dst[3 * BPS]) = _mm_cvtsi128_si32(ref3);
}
}
}
mlib_status
__mlib_VectorConvert_S32_S16_Mod(
mlib_s32 *z,
const mlib_s16 *x,
mlib_s32 n)
{
if (n < 1)
return (MLIB_FAILURE);
mlib_s32 i, ax, az, nstep, n1, n2, n3;
mlib_s16 *px = (mlib_s16 *)x;
mlib_s32 *pz = (mlib_s32 *)z;
__m128i xbuf, zlo, zhi, zero;
zero = _mm_setzero_si128();
ax = (mlib_addr)x & 15;
az = (mlib_addr)z & 15;
nstep = 16 / sizeof (mlib_s16);
n1 = ((16 - ax) & 15) / sizeof (mlib_s16);
n2 = (n - n1) / nstep;
n3 = n - n1 - n2 * nstep;
if (n2 < 1) {
for (i = 0; i < n; i++) {
*pz++ = *px++;
}
} else {
for (i = 0; i < n1; i++) {
*pz++ = *px++;
}
if ((ax * 2 & 15) == az) {
for (i = 0; i < n2; i++) {
xbuf = _mm_load_si128((__m128i *)px);
zlo = _mm_unpacklo_epi16(zero, xbuf);
zhi = _mm_unpackhi_epi16(zero, xbuf);
zlo = _mm_srai_epi32(zlo, 16);
zhi = _mm_srai_epi32(zhi, 16);
_mm_store_si128((__m128i *)pz, zlo);
_mm_store_si128((__m128i *)pz + 1, zhi);
px += nstep;
pz += nstep;
}
} else {
for (i = 0; i < n2; i++) {
xbuf = _mm_load_si128((__m128i *)px);
zlo = _mm_unpacklo_epi16(zero, xbuf);
zhi = _mm_unpackhi_epi16(zero, xbuf);
zlo = _mm_srai_epi32(zlo, 16);
zhi = _mm_srai_epi32(zhi, 16);
_mm_storeu_si128((__m128i *)pz, zlo);
_mm_storeu_si128((__m128i *)pz + 1, zhi);
px += nstep;
pz += nstep;
}
}
for (i = 0; i < n3; i++) {
*pz++ = *px++;
}
}
return (MLIB_SUCCESS);
}
FourColorVals(const uint16_t *p) {
assert(((uintptr_t)p & 7) == 0);//assert aligned
vec = _mm_unpacklo_epi16(_mm_loadl_epi64((__m128i*)p),_mm_setzero_si128());
}
static void GF_FUNC_ALIGN VS_CC
proc_8bit_sse2(convolution_t *ch, uint8_t *buff, int bstride, int width,
int height, int stride, uint8_t *dstp, const uint8_t *srcp)
{
uint8_t *p0 = buff + 16;
uint8_t *p1 = p0 + bstride;
uint8_t *p2 = p1 + bstride;
uint8_t *p3 = p2 + bstride;
uint8_t *p4 = p3 + bstride;
uint8_t *orig = p0, *end = p4;
line_copy8(p0, srcp + 2 * stride , width, 2);
line_copy8(p1, srcp + stride, width, 2);
line_copy8(p2, srcp, width, 2);
srcp += stride;
line_copy8(p3, srcp, width, 2);
__m128i zero = _mm_setzero_si128();
__m128 rdiv = _mm_set1_ps((float)ch->rdiv);
__m128 bias = _mm_set1_ps((float)ch->bias);
__m128i matrix[25];
for (int i = 0; i < 25; i++) {
matrix[i] = _mm_unpacklo_epi16(_mm_set1_epi16((int16_t)ch->m[i]), zero);
}
for (int y = 0; y < height; y++) {
srcp += stride * (y < height - 2 ? 1 : -1);
line_copy8(p4, srcp, width, 2);
uint8_t *array[] = {
p0 - 2, p0 - 1, p0, p0 + 1, p0 + 2,
p1 - 2, p1 - 1, p1, p1 + 1, p1 + 2,
p2 - 2, p2 - 1, p2, p2 + 1, p2 + 2,
p3 - 2, p3 - 1, p3, p3 + 1, p3 + 2,
p4 - 2, p4 - 1, p4, p4 + 1, p4 + 2
};
for (int x = 0; x < width; x += 16) {
__m128i sum[4] = { zero, zero, zero, zero };
for (int i = 0; i < 25; i++) {
__m128i xmm0, xmm1, xmm2;
xmm0 = _mm_loadu_si128((__m128i *)(array[i] + x));
xmm2 = _mm_unpackhi_epi8(xmm0, zero);
xmm0 = _mm_unpacklo_epi8(xmm0, zero);
xmm1 = _mm_unpackhi_epi16(xmm0, zero);
xmm0 = _mm_unpacklo_epi16(xmm0, zero);
sum[0] = _mm_add_epi32(sum[0], _mm_madd_epi16(xmm0, matrix[i]));
sum[1] = _mm_add_epi32(sum[1], _mm_madd_epi16(xmm1, matrix[i]));
xmm1 = _mm_unpackhi_epi16(xmm2, zero);
xmm0 = _mm_unpacklo_epi16(xmm2, zero);
sum[2] = _mm_add_epi32(sum[2], _mm_madd_epi16(xmm0, matrix[i]));
sum[3] = _mm_add_epi32(sum[3], _mm_madd_epi16(xmm1, matrix[i]));
}
for (int i = 0; i < 4; i++) {
__m128 sumfp = _mm_cvtepi32_ps(sum[i]);
sumfp = _mm_mul_ps(sumfp, rdiv);
sumfp = _mm_add_ps(sumfp, bias);
if (!ch->saturate) {
sumfp = mm_abs_ps(sumfp);
}
sum[i] = _mm_cvttps_epi32(sumfp);
}
sum[0] = _mm_packs_epi32(sum[0], sum[1]);
sum[1] = _mm_packs_epi32(sum[2], sum[3]);
sum[0] = _mm_packus_epi16(sum[0], sum[1]);
_mm_store_si128((__m128i *)(dstp + x), sum[0]);
}
dstp += stride;
p0 = p1;
p1 = p2;
p2 = p3;
p3 = p4;
p4 = (p4 == end) ? orig : p4 + bstride;
}
}
int
global_sse2_word(int queryLength,
unsigned short *profile,
const unsigned char *dbSeq,
int dbLength,
unsigned short gapOpen,
unsigned short gapExtend,
unsigned short ceiling,
struct f_struct *f_str)
{
int i, j;
int score;
int scale;
int temp;
int distance;
int offset;
int position;
int cmp;
int iter;
__m128i *pvH;
__m128i *pvE;
__m128i vE, vF, vH;
__m128i vHNext;
__m128i vFPrev;
__m128i vGapOpen;
__m128i vGapExtend;
__m128i vCeiling;
__m128i vScale;
__m128i vScaleAmt;
__m128i vScaleTmp;
__m128i vTemp;
__m128i vNull;
__m128i *pvScore;
scale = 0;
iter = (queryLength + 7) / 8;
offset = (queryLength - 1) % iter;
position = 7 - (queryLength - 1) / iter;
pvH = (__m128i *)f_str->workspace;
pvE = pvH + iter;
/* Load gap opening penalty to all elements of a constant */
vGapOpen = _mm_setzero_si128(); /* transfered from Apple Devel smith_waterman_sse2.c fix */
vGapOpen = _mm_insert_epi16 (vGapOpen, gapOpen, 0);
vGapOpen = _mm_shufflelo_epi16 (vGapOpen, 0);
vGapOpen = _mm_shuffle_epi32 (vGapOpen, 0);
/* Load gap extension penalty to all elements of a constant */
vGapExtend = _mm_setzero_si128(); /* transfered from Apple Devel smith_waterman_sse2.c fix */
vGapExtend = _mm_insert_epi16 (vGapExtend, gapExtend, 0);
vGapExtend = _mm_shufflelo_epi16 (vGapExtend, 0);
vGapExtend = _mm_shuffle_epi32 (vGapExtend, 0);
/* Generate the ceiling before scaling */
vTemp = _mm_setzero_si128(); /* transfered from Apple Devel smith_waterman_sse2.c fix */
vTemp = _mm_insert_epi16 (vTemp, ceiling, 0);
vTemp = _mm_shufflelo_epi16 (vTemp, 0);
vTemp = _mm_shuffle_epi32 (vTemp, 0);
vCeiling = _mm_cmpeq_epi16 (vTemp, vTemp);
vCeiling = _mm_srli_epi16 (vCeiling, 1);
vCeiling = _mm_subs_epi16 (vCeiling, vTemp);
vCeiling = _mm_subs_epi16 (vCeiling, vGapOpen);
vNull = _mm_cmpeq_epi16 (vTemp, vTemp);
vNull = _mm_slli_epi16 (vNull, 15);
vScaleAmt = _mm_xor_si128 (vNull, vNull);
/* Zero out the storage vector */
vTemp = _mm_adds_epi16 (vNull, vGapOpen);
for (i = 0; i < iter; i++) {
_mm_store_si128 (pvH + i, vTemp);
_mm_store_si128 (pvE + i, vNull);
}
/* initialize F */
vF = vNull;
vFPrev = vNull;
/* load and scale H for the next round */
vTemp = _mm_srli_si128 (vGapOpen, 14);
vH = _mm_load_si128 (pvH + iter - 1);
vH = _mm_adds_epi16 (vH, vTemp);
for (i = 0; i < dbLength; ++i) {
/* fetch first data asap. */
pvScore = (__m128i *) profile + dbSeq[i] * iter;
vF = vNull;
vH = _mm_max_epi16 (vH, vFPrev);
for (j = 0; j < iter; j++) {
/* correct H from the previous columns F */
vHNext = _mm_load_si128 (pvH + j);
vHNext = _mm_max_epi16 (vHNext, vFPrev);
/* load and correct E value */
vE = _mm_load_si128 (pvE + j);
vTemp = _mm_subs_epi16 (vHNext, vGapOpen);
vE = _mm_max_epi16 (vE, vTemp);
_mm_store_si128 (pvE + j, vE);
/* add score to vH */
vH = _mm_adds_epi16 (vH, *pvScore++);
/* get max from vH, vE and vF */
vH = _mm_max_epi16 (vH, vE);
vH = _mm_max_epi16 (vH, vF);
_mm_store_si128 (pvH + j, vH);
/* update vF value */
vH = _mm_subs_epi16 (vH, vGapOpen);
vF = _mm_max_epi16 (vF, vH);
/* load the next h values */
vH = vHNext;
}
/* check if we need to scale before the next round */
vTemp = _mm_cmpgt_epi16 (vF, vCeiling);
cmp = _mm_movemask_epi8 (vTemp);
/* broadcast F values */
vF = _mm_xor_si128 (vF, vNull);
vTemp = _mm_slli_si128 (vF, 2);
vTemp = _mm_subs_epu16 (vTemp, vScaleAmt);
vF = max_epu16 (vF, vTemp);
vTemp = _mm_slli_si128 (vF, 4);
vScaleTmp = _mm_slli_si128 (vScaleAmt, 2);
vScaleTmp = _mm_adds_epu16 (vScaleTmp, vScaleAmt);
vTemp = _mm_subs_epu16 (vTemp, vScaleTmp);
vF = max_epu16 (vF, vTemp);
vTemp = _mm_slli_si128 (vScaleTmp, 4);
vScaleTmp = _mm_adds_epu16 (vScaleTmp, vTemp);
vTemp = _mm_slli_si128 (vF, 8);
vTemp = _mm_subs_epu16 (vTemp, vScaleTmp);
vF = max_epu16 (vF, vTemp);
/* scale if necessary */
if (cmp != 0x0000) {
__m128i vScale1;
__m128i vScale2;
vScale = _mm_slli_si128 (vF, 2);
vScale = _mm_subs_epu16 (vScale, vGapOpen);
vScale = _mm_subs_epu16 (vScale, vScaleAmt);
vTemp = _mm_slli_si128 (vScale, 2);
vTemp = _mm_subs_epu16 (vScale, vTemp);
vScaleAmt = _mm_adds_epu16 (vScaleAmt, vTemp);
vTemp = _mm_slli_si128 (vScale, 2);
vTemp = _mm_subs_epu16 (vTemp, vScale);
vScaleAmt = _mm_subs_epu16 (vScaleAmt, vTemp);
/* rescale the previous F */
vF = _mm_subs_epu16 (vF, vScale);
/* check if we can continue in signed 16-bits */
vTemp = _mm_xor_si128 (vF, vNull);
vTemp = _mm_cmpgt_epi16 (vTemp, vCeiling);
cmp = _mm_movemask_epi8 (vTemp);
if (cmp != 0x0000) {
return OVERFLOW_SCORE;
}
vTemp = _mm_adds_epi16 (vCeiling, vCeiling);
vScale1 = _mm_subs_epu16 (vScale, vTemp);
vScale2 = _mm_subs_epu16 (vScale, vScale1);
/* scale all the vectors */
for (j = 0; j < iter; j++) {
/* load H and E */
vH = _mm_load_si128 (pvH + j);
vE = _mm_load_si128 (pvE + j);
/* get max from vH, vE and vF */
vH = _mm_subs_epi16 (vH, vScale1);
vH = _mm_subs_epi16 (vH, vScale2);
vE = _mm_subs_epi16 (vE, vScale1);
vE = _mm_subs_epi16 (vE, vScale2);
/* save the H and E */
_mm_store_si128 (pvH + j, vH);
_mm_store_si128 (pvE + j, vE);
}
vScale = vScaleAmt;
for (j = 0; j < position; ++j) {
vScale = _mm_slli_si128 (vScale, 2);
}
/* calculate the final scaling amount */
vTemp = _mm_xor_si128 (vTemp, vTemp);
vScale1 = _mm_unpacklo_epi16 (vScale, vTemp);
vScale2 = _mm_unpackhi_epi16 (vScale, vTemp);
vScale = _mm_add_epi32 (vScale1, vScale2);
vTemp = _mm_srli_si128 (vScale, 8);
vScale = _mm_add_epi32 (vScale, vTemp);
vTemp = _mm_srli_si128 (vScale, 4);
vScale = _mm_add_epi32 (vScale, vTemp);
scale = (int) (unsigned short) _mm_extract_epi16 (vScale, 0);
temp = (int) (unsigned short) _mm_extract_epi16 (vScale, 1);
scale = scale + (temp << 16);
}
/* scale the F value for the next round */
vFPrev = _mm_slli_si128 (vF, 2);
vFPrev = _mm_subs_epu16 (vFPrev, vScaleAmt);
vFPrev = _mm_xor_si128 (vFPrev, vNull);
/* load and scale H for the next round */
vH = _mm_load_si128 (pvH + iter - 1);
vH = _mm_xor_si128 (vH, vNull);
vH = _mm_slli_si128 (vH, 2);
vH = _mm_subs_epu16 (vH, vScaleAmt);
vH = _mm_insert_epi16 (vH, gapOpen, 0);
vH = _mm_xor_si128 (vH, vNull);
}
vH = _mm_load_si128 (pvH + offset);
vH = _mm_max_epi16 (vH, vFPrev);
for (j = 0; j < position; ++j) {
vH = _mm_slli_si128 (vH, 2);
}
score = (int) (signed short) _mm_extract_epi16 (vH, 7);
score = score + SHORT_BIAS;
/* return largest score */
distance = (queryLength + dbLength) * gapExtend;
score = score - (gapOpen * 2) - distance + scale;
return score;
}
static LW_FORCEINLINE void fill_rgb_buffer_sse41( BYTE *rgb_buffer, BYTE *lw48_ptr )
{
static const USHORT LW_ALIGN(16) PW_32768[8] = { 32768, 32768, 32768, 32768, 32768, 32768, 32768, 32768 };
static const short LW_ALIGN(16) PW_28672[8] = { 28672, 28672, 28672, 28672, 28672, 28672, 28672, 28672 };
static const short LW_ALIGN(16) PW_9539[8] = { 9539, 9539, 9539, 9539, 9539, 9539, 9539, 9539 };
static const short LW_ALIGN(16) PW_13074[8] = { 13074, 13074, 13074, 13074, 13074, 13074, 13074, 13074 };
static const short LW_ALIGN(16) PW_16531[8] = { 16531, 16531, 16531, 16531, 16531, 16531, 16531, 16531 };
static const short LW_ALIGN(16) PW_M3203_M6808[8] = { -3203, -6808, -3203, -6808, -3203, -6808, -3203, -6808 };
static const int LW_ALIGN(16) PD_1_20[4] = { (1<<20), (1<<20), (1<<20), (1<<20) };
static const char LW_ALIGN(16) LW48_SHUFFLE[3][16] = {
{ 0, 1, 6, 7, 12, 13, 2, 3, 8, 9, 14, 15, 4, 5, 10, 11 },
{ 2, 3, 8, 9, 14, 15, 4, 5, 10, 11, 0, 1, 6, 7, 12, 13 },
{ 4, 5, 10, 11, 0, 1, 6, 7, 12, 13, 2, 3, 8, 9, 14, 15 }
};
__m128i x0, x1, x2, x3, x4, x5, x6, x7;
x5 = _mm_loadu_si128((__m128i *)(lw48_ptr + 0));
x6 = _mm_loadu_si128((__m128i *)(lw48_ptr + 16));
x7 = _mm_loadu_si128((__m128i *)(lw48_ptr + 32));
x0 = _mm_blend_epi16(x5, x6, 0x80+0x10+0x02);
x0 = _mm_blend_epi16(x0, x7, 0x20+0x04);
x1 = _mm_blend_epi16(x5, x6, 0x20+0x04);
x1 = _mm_blend_epi16(x1, x7, 0x40+0x08+0x01);
x2 = _mm_blend_epi16(x5, x6, 0x40+0x08+0x01);
x2 = _mm_blend_epi16(x2, x7, 0x80+0x10+0x02);
x0 = _mm_shuffle_epi8(x0, _mm_load_si128((__m128i*)LW48_SHUFFLE[0])); /* Y */
x1 = _mm_shuffle_epi8(x1, _mm_load_si128((__m128i*)LW48_SHUFFLE[1])); /* Cb */
x2 = _mm_shuffle_epi8(x2, _mm_load_si128((__m128i*)LW48_SHUFFLE[2])); /* Cr */
x0 = _mm_sub_epi16(x0, _mm_load_si128((__m128i*)PW_32768));
x1 = _mm_sub_epi16(x1, _mm_load_si128((__m128i*)PW_32768));
x2 = _mm_sub_epi16(x2, _mm_load_si128((__m128i*)PW_32768));
/* y_tmp = ((y - 4096) * 9539) */
/* = ((y - 32768) + (32768 - 4096)) * 9539 */
/* = ((y - 32768) * 9539 + 28672 * 9539 */
x3 = _mm_unpacklo_epi16(x0, _mm_load_si128((__m128i*)PW_28672));
x4 = _mm_unpackhi_epi16(x0, _mm_load_si128((__m128i*)PW_28672));
x3 = _mm_madd_epi16(x3, _mm_load_si128((__m128i*)PW_9539));
x4 = _mm_madd_epi16(x4, _mm_load_si128((__m128i*)PW_9539));
/* G = ((y_tmp + ((cb-32768) * -3203) + ((cr-32768) * -6808)) + (1<<20)) >> 21 */
x5 = _mm_unpacklo_epi16(x1, x2);
x6 = _mm_unpackhi_epi16(x1, x2);
x5 = _mm_madd_epi16(x5, _mm_load_si128((__m128i*)PW_M3203_M6808));
x6 = _mm_madd_epi16(x6, _mm_load_si128((__m128i*)PW_M3203_M6808));
x5 = _mm_add_epi32(x5, x3);
x6 = _mm_add_epi32(x6, x4);
x5 = _mm_add_epi32(x5, _mm_load_si128((__m128i*)PD_1_20));
x6 = _mm_add_epi32(x6, _mm_load_si128((__m128i*)PD_1_20));
x5 = _mm_srai_epi32(x5, 21);
x6 = _mm_srai_epi32(x6, 21);
x5 = _mm_packs_epi32(x5, x6);
_mm_store_si128((__m128i*)(rgb_buffer + 16), x5);
/* R = ((y_tmp + ((cr-32768) * 13074) + (1<<20)) >> 21 */
x0 = _mm_mullo_epi16(x2, _mm_load_si128((__m128i*)PW_13074));
x7 = _mm_mulhi_epi16(x2, _mm_load_si128((__m128i*)PW_13074));
x6 = _mm_unpacklo_epi16(x0, x7);
x7 = _mm_unpackhi_epi16(x0, x7);
x6 = _mm_add_epi32(x6, x3);
x7 = _mm_add_epi32(x7, x4);
x6 = _mm_add_epi32(x6, _mm_load_si128((__m128i*)PD_1_20));
x7 = _mm_add_epi32(x7, _mm_load_si128((__m128i*)PD_1_20));
x6 = _mm_srai_epi32(x6, 21);
x7 = _mm_srai_epi32(x7, 21);
x6 = _mm_packs_epi32(x6, x7);
_mm_store_si128((__m128i*)(rgb_buffer + 32), x6);
/* B = ((y_tmp + ((cb-32768) * 16531) + (1<<20)) >> 21 */
x2 = _mm_mullo_epi16(x1, _mm_load_si128((__m128i*)PW_16531));
x7 = _mm_mulhi_epi16(x1, _mm_load_si128((__m128i*)PW_16531));
x0 = _mm_unpacklo_epi16(x2, x7);
x7 = _mm_unpackhi_epi16(x2, x7);
x0 = _mm_add_epi32(x0, x3);
x7 = _mm_add_epi32(x7, x4);
x0 = _mm_add_epi32(x0, _mm_load_si128((__m128i*)PD_1_20));
x7 = _mm_add_epi32(x7, _mm_load_si128((__m128i*)PD_1_20));
x0 = _mm_srai_epi32(x0, 21);
x7 = _mm_srai_epi32(x7, 21);
x7 = _mm_packs_epi32(x0, x7);
_mm_store_si128((__m128i*)(rgb_buffer + 0), x7);
}
void tuned_ConvertULY4ToRGB(uint8_t *pDstBegin, uint8_t *pDstEnd, const uint8_t *pYBegin, const uint8_t *pUBegin, const uint8_t *pVBegin, size_t cbWidth, ssize_t scbStride)
{
const int shift = 13;
__m128i xy2rgb = _mm_set2_epi16_shift((-16 * C::Y2RGB + 0.5) / 0xff, C::Y2RGB, shift);
__m128i vu2r = _mm_set2_epi16_shift(C::V2R, 0, shift);
__m128i vu2g = _mm_set2_epi16_shift(C::V2G, C::U2G, shift);
__m128i vu2b = _mm_set2_epi16_shift(0, C::U2B, shift);
auto y = pYBegin;
auto u = pUBegin;
auto v = pVBegin;
for (auto p = pDstBegin; p != pDstEnd; p += scbStride)
{
auto pp = p;
for (; pp <= p + cbWidth - 16; pp += T::BYPP * 4)
{
__m128i yy = _mm_cvtsi32_si128(*(const int *)y);
__m128i uu = _mm_cvtsi32_si128(*(const int *)u);
__m128i vv = _mm_cvtsi32_si128(*(const int *)v);
__m128i xy = _mm_unpacklo_epi8(_mm_unpacklo_epi8(yy, _mm_setone_si128()), _mm_setzero_si128()); // 00 ff 00 Y3 00 ff 00 Y2 00 ff 00 Y1 00 ff 00 Y0
__m128i vu = _mm_unpacklo_epi8(_mm_unpacklo_epi8(uu, vv), _mm_setzero_si128()); // 00 V3 00 U3 00 V2 00 U2 00 V1 00 U1 00 V0 00 U0
vu = _mm_sub_epi16(vu, _mm_set1_epi16(128));
__m128i rgbtmp = _mm_madd_epi16(xy, xy2rgb);
auto xyuv2rgb = [rgbtmp, vu, shift](__m128i vu2rgb) -> __m128i {
__m128i rgb = _mm_add_epi32(rgbtmp, _mm_madd_epi16(vu, vu2rgb));
rgb = _mm_srai_epi32(rgb, shift);
rgb = _mm_packs_epi32(rgb, rgb);
rgb = _mm_packus_epi16(rgb, rgb);
return rgb;
};
__m128i rr = xyuv2rgb(vu2r);
__m128i gg = xyuv2rgb(vu2g);
__m128i bb = xyuv2rgb(vu2b);
if (std::is_same<T, CBGRAColorOrder>::value)
{
__m128i bgrx = _mm_unpacklo_epi16(_mm_unpacklo_epi8(bb, gg), _mm_unpacklo_epi8(rr, _mm_setone_si128()));
_mm_storeu_si128((__m128i *)pp, bgrx);
}
#ifdef __SSSE3__
else if (std::is_same<T, CBGRColorOrder>::value)
{
__m128i bgrx = _mm_unpacklo_epi16(_mm_unpacklo_epi8(bb, gg), _mm_unpacklo_epi8(rr, rr));
__m128i bgr = _mm_shuffle_epi8(bgrx, _mm_set_epi8(-1, -1, -1, -1, 14, 13, 12, 10, 9, 8, 6, 5, 4, 2, 1, 0));
_mm_storeu_si128((__m128i *)pp, bgr);
}
#endif
else if (std::is_same<T, CARGBColorOrder>::value)
{
__m128i xrgb = _mm_unpacklo_epi16(_mm_unpacklo_epi8(rr, rr), _mm_unpacklo_epi8(gg, bb));
_mm_storeu_si128((__m128i *)pp, xrgb);
}
#ifdef __SSSE3__
else if (std::is_same<T, CRGBColorOrder>::value)
{
__m128i xrgb = _mm_unpacklo_epi16(_mm_unpacklo_epi8(_mm_setone_si128(), rr), _mm_unpacklo_epi8(gg, bb));
__m128i rgb = _mm_shuffle_epi8(xrgb, _mm_set_epi8(-1, -1, -1, -1, 15, 14, 13, 11, 10, 9, 7, 6, 5, 3, 2, 1));
_mm_storeu_si128((__m128i *)pp, rgb);
}
#endif
y += 4;
u += 4;
v += 4;
}
for (; pp < p + cbWidth; pp += T::BYPP)
{
__m128i xy = _mm_cvtsi32_si128(*y | 0x00ff0000);
__m128i uu = _mm_cvtsi32_si128(*u);
__m128i vv = _mm_cvtsi32_si128(*v);
__m128i vu = _mm_unpacklo_epi8(_mm_unpacklo_epi8(uu, vv), _mm_setzero_si128()); // 00 V3 00 U3 00 V2 00 U2 00 V1 00 U1 00 V0 00 U0
vu = _mm_sub_epi16(vu, _mm_set1_epi16(128));
__m128i rgbtmp = _mm_madd_epi16(xy, xy2rgb);
auto xyuv2rgb = [rgbtmp, vu, shift](__m128i vu2rgb) -> __m128i {
__m128i rgb = _mm_add_epi32(rgbtmp, _mm_madd_epi16(vu, vu2rgb));
rgb = _mm_srai_epi32(rgb, shift);
rgb = _mm_packs_epi32(rgb, rgb);
rgb = _mm_packus_epi16(rgb, rgb);
return rgb;
};
__m128i rr = xyuv2rgb(vu2r);
__m128i gg = xyuv2rgb(vu2g);
__m128i bb = xyuv2rgb(vu2b);
if (std::is_same<T, CBGRAColorOrder>::value)
{
__m128i bgrx = _mm_unpacklo_epi16(_mm_unpacklo_epi8(bb, gg), _mm_unpacklo_epi8(rr, _mm_setone_si128()));
*(uint32_t *)pp = _mm_cvtsi128_si32(bgrx);
}
else if (std::is_same<T, CARGBColorOrder>::value)
{
__m128i xrgb = _mm_unpacklo_epi16(_mm_unpacklo_epi8(rr, rr), _mm_unpacklo_epi8(gg, bb));
*(uint32_t *)pp = _mm_cvtsi128_si32(xrgb);
}
else if (std::is_same<T, CBGRColorOrder>::value || std::is_same<T, CRGBColorOrder>::value)
{
*(pp + T::B) = (uint8_t)_mm_cvtsi128_si32(bb);
*(pp + T::G) = (uint8_t)_mm_cvtsi128_si32(gg);
*(pp + T::R) = (uint8_t)_mm_cvtsi128_si32(rr);
}
y += 1;
u += 1;
v += 1;
}
}
}
void vp9_short_fdct4x4_sse2(int16_t *input, int16_t *output, int pitch) {
// The 2D transform is done with two passes which are actually pretty
// similar. In the first one, we transform the columns and transpose
// the results. In the second one, we transform the rows. To achieve that,
// as the first pass results are transposed, we tranpose the columns (that
// is the transposed rows) and transpose the results (so that it goes back
// in normal/row positions).
const int stride = pitch >> 1;
int pass;
// Constants
// When we use them, in one case, they are all the same. In all others
// it's a pair of them that we need to repeat four times. This is done
// by constructing the 32 bit constant corresponding to that pair.
const __m128i k__cospi_p16_p16 = _mm_set1_epi16(cospi_16_64);
const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
const __m128i k__nonzero_bias_a = _mm_setr_epi16(0, 1, 1, 1, 1, 1, 1, 1);
const __m128i k__nonzero_bias_b = _mm_setr_epi16(1, 0, 0, 0, 0, 0, 0, 0);
const __m128i kOne = _mm_set1_epi16(1);
__m128i in0, in1, in2, in3;
// Load inputs.
{
in0 = _mm_loadl_epi64((const __m128i *)(input + 0 * stride));
in1 = _mm_loadl_epi64((const __m128i *)(input + 1 * stride));
in2 = _mm_loadl_epi64((const __m128i *)(input + 2 * stride));
in3 = _mm_loadl_epi64((const __m128i *)(input + 3 * stride));
// x = x << 4
in0 = _mm_slli_epi16(in0, 4);
in1 = _mm_slli_epi16(in1, 4);
in2 = _mm_slli_epi16(in2, 4);
in3 = _mm_slli_epi16(in3, 4);
// if (i == 0 && input[0]) input[0] += 1;
{
// The mask will only contain wether the first value is zero, all
// other comparison will fail as something shifted by 4 (above << 4)
// can never be equal to one. To increment in the non-zero case, we
// add the mask and one for the first element:
// - if zero, mask = -1, v = v - 1 + 1 = v
// - if non-zero, mask = 0, v = v + 0 + 1 = v + 1
__m128i mask = _mm_cmpeq_epi16(in0, k__nonzero_bias_a);
in0 = _mm_add_epi16(in0, mask);
in0 = _mm_add_epi16(in0, k__nonzero_bias_b);
}
}
// Do the two transform/transpose passes
for (pass = 0; pass < 2; ++pass) {
// Transform 1/2: Add/substract
const __m128i r0 = _mm_add_epi16(in0, in3);
const __m128i r1 = _mm_add_epi16(in1, in2);
const __m128i r2 = _mm_sub_epi16(in1, in2);
const __m128i r3 = _mm_sub_epi16(in0, in3);
// Transform 1/2: Interleave to do the multiply by constants which gets us
// into 32 bits.
const __m128i t0 = _mm_unpacklo_epi16(r0, r1);
const __m128i t2 = _mm_unpacklo_epi16(r2, r3);
const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16);
const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16);
const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08);
const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24);
const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
// Combine and transpose
const __m128i res0 = _mm_packs_epi32(w0, w2);
const __m128i res1 = _mm_packs_epi32(w4, w6);
// 00 01 02 03 20 21 22 23
// 10 11 12 13 30 31 32 33
const __m128i tr0_0 = _mm_unpacklo_epi16(res0, res1);
const __m128i tr0_1 = _mm_unpackhi_epi16(res0, res1);
// 00 10 01 11 02 12 03 13
// 20 30 21 31 22 32 23 33
in0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
in2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
// 00 10 20 30 01 11 21 31 in0 contains 0 followed by 1
// 02 12 22 32 03 13 23 33 in2 contains 2 followed by 3
if (0 == pass) {
// Extract values in the high part for second pass as transform code
// only uses the first four values.
in1 = _mm_unpackhi_epi64(in0, in0);
in3 = _mm_unpackhi_epi64(in2, in2);
} else {
// Post-condition output and store it (v + 1) >> 2, taking advantage
// of the fact 1/3 are stored just after 0/2.
__m128i out01 = _mm_add_epi16(in0, kOne);
__m128i out23 = _mm_add_epi16(in2, kOne);
out01 = _mm_srai_epi16(out01, 2);
out23 = _mm_srai_epi16(out23, 2);
_mm_storeu_si128((__m128i *)(output + 0 * 4), out01);
_mm_storeu_si128((__m128i *)(output + 2 * 4), out23);
}
}
}
void vp9_short_fdct16x16_sse2(int16_t *input, int16_t *output, int pitch) {
// The 2D transform is done with two passes which are actually pretty
// similar. In the first one, we transform the columns and transpose
// the results. In the second one, we transform the rows. To achieve that,
// as the first pass results are transposed, we tranpose the columns (that
// is the transposed rows) and transpose the results (so that it goes back
// in normal/row positions).
const int stride = pitch >> 1;
int pass;
// We need an intermediate buffer between passes.
int16_t intermediate[256];
int16_t *in = input;
int16_t *out = intermediate;
// Constants
// When we use them, in one case, they are all the same. In all others
// it's a pair of them that we need to repeat four times. This is done
// by constructing the 32 bit constant corresponding to that pair.
const __m128i k__cospi_p16_p16 = _mm_set1_epi16(cospi_16_64);
const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
const __m128i k__cospi_m24_m08 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64);
const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64);
const __m128i k__cospi_p30_p02 = pair_set_epi16(cospi_30_64, cospi_2_64);
const __m128i k__cospi_p14_p18 = pair_set_epi16(cospi_14_64, cospi_18_64);
const __m128i k__cospi_m02_p30 = pair_set_epi16(-cospi_2_64, cospi_30_64);
const __m128i k__cospi_m18_p14 = pair_set_epi16(-cospi_18_64, cospi_14_64);
const __m128i k__cospi_p22_p10 = pair_set_epi16(cospi_22_64, cospi_10_64);
const __m128i k__cospi_p06_p26 = pair_set_epi16(cospi_6_64, cospi_26_64);
const __m128i k__cospi_m10_p22 = pair_set_epi16(-cospi_10_64, cospi_22_64);
const __m128i k__cospi_m26_p06 = pair_set_epi16(-cospi_26_64, cospi_6_64);
const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
const __m128i kOne = _mm_set1_epi16(1);
// Do the two transform/transpose passes
for (pass = 0; pass < 2; ++pass) {
// We process eight columns (transposed rows in second pass) at a time.
int column_start;
for (column_start = 0; column_start < 16; column_start += 8) {
__m128i in00, in01, in02, in03, in04, in05, in06, in07;
__m128i in08, in09, in10, in11, in12, in13, in14, in15;
__m128i input0, input1, input2, input3, input4, input5, input6, input7;
__m128i step1_0, step1_1, step1_2, step1_3;
__m128i step1_4, step1_5, step1_6, step1_7;
__m128i step2_1, step2_2, step2_3, step2_4, step2_5, step2_6;
__m128i step3_0, step3_1, step3_2, step3_3;
__m128i step3_4, step3_5, step3_6, step3_7;
__m128i res00, res01, res02, res03, res04, res05, res06, res07;
__m128i res08, res09, res10, res11, res12, res13, res14, res15;
// Load and pre-condition input.
if (0 == pass) {
in00 = _mm_loadu_si128((const __m128i *)(in + 0 * stride));
in01 = _mm_loadu_si128((const __m128i *)(in + 1 * stride));
in02 = _mm_loadu_si128((const __m128i *)(in + 2 * stride));
in03 = _mm_loadu_si128((const __m128i *)(in + 3 * stride));
in04 = _mm_loadu_si128((const __m128i *)(in + 4 * stride));
in05 = _mm_loadu_si128((const __m128i *)(in + 5 * stride));
in06 = _mm_loadu_si128((const __m128i *)(in + 6 * stride));
in07 = _mm_loadu_si128((const __m128i *)(in + 7 * stride));
in08 = _mm_loadu_si128((const __m128i *)(in + 8 * stride));
in09 = _mm_loadu_si128((const __m128i *)(in + 9 * stride));
in10 = _mm_loadu_si128((const __m128i *)(in + 10 * stride));
in11 = _mm_loadu_si128((const __m128i *)(in + 11 * stride));
in12 = _mm_loadu_si128((const __m128i *)(in + 12 * stride));
in13 = _mm_loadu_si128((const __m128i *)(in + 13 * stride));
in14 = _mm_loadu_si128((const __m128i *)(in + 14 * stride));
in15 = _mm_loadu_si128((const __m128i *)(in + 15 * stride));
// x = x << 2
in00 = _mm_slli_epi16(in00, 2);
in01 = _mm_slli_epi16(in01, 2);
in02 = _mm_slli_epi16(in02, 2);
in03 = _mm_slli_epi16(in03, 2);
in04 = _mm_slli_epi16(in04, 2);
in05 = _mm_slli_epi16(in05, 2);
in06 = _mm_slli_epi16(in06, 2);
in07 = _mm_slli_epi16(in07, 2);
in08 = _mm_slli_epi16(in08, 2);
in09 = _mm_slli_epi16(in09, 2);
in10 = _mm_slli_epi16(in10, 2);
in11 = _mm_slli_epi16(in11, 2);
in12 = _mm_slli_epi16(in12, 2);
in13 = _mm_slli_epi16(in13, 2);
in14 = _mm_slli_epi16(in14, 2);
in15 = _mm_slli_epi16(in15, 2);
} else {
in00 = _mm_loadu_si128((const __m128i *)(in + 0 * 16));
in01 = _mm_loadu_si128((const __m128i *)(in + 1 * 16));
in02 = _mm_loadu_si128((const __m128i *)(in + 2 * 16));
in03 = _mm_loadu_si128((const __m128i *)(in + 3 * 16));
in04 = _mm_loadu_si128((const __m128i *)(in + 4 * 16));
in05 = _mm_loadu_si128((const __m128i *)(in + 5 * 16));
in06 = _mm_loadu_si128((const __m128i *)(in + 6 * 16));
in07 = _mm_loadu_si128((const __m128i *)(in + 7 * 16));
in08 = _mm_loadu_si128((const __m128i *)(in + 8 * 16));
in09 = _mm_loadu_si128((const __m128i *)(in + 9 * 16));
in10 = _mm_loadu_si128((const __m128i *)(in + 10 * 16));
in11 = _mm_loadu_si128((const __m128i *)(in + 11 * 16));
in12 = _mm_loadu_si128((const __m128i *)(in + 12 * 16));
in13 = _mm_loadu_si128((const __m128i *)(in + 13 * 16));
in14 = _mm_loadu_si128((const __m128i *)(in + 14 * 16));
in15 = _mm_loadu_si128((const __m128i *)(in + 15 * 16));
// x = (x + 1) >> 2
in00 = _mm_add_epi16(in00, kOne);
in01 = _mm_add_epi16(in01, kOne);
in02 = _mm_add_epi16(in02, kOne);
in03 = _mm_add_epi16(in03, kOne);
in04 = _mm_add_epi16(in04, kOne);
in05 = _mm_add_epi16(in05, kOne);
in06 = _mm_add_epi16(in06, kOne);
in07 = _mm_add_epi16(in07, kOne);
in08 = _mm_add_epi16(in08, kOne);
in09 = _mm_add_epi16(in09, kOne);
in10 = _mm_add_epi16(in10, kOne);
in11 = _mm_add_epi16(in11, kOne);
in12 = _mm_add_epi16(in12, kOne);
in13 = _mm_add_epi16(in13, kOne);
in14 = _mm_add_epi16(in14, kOne);
in15 = _mm_add_epi16(in15, kOne);
in00 = _mm_srai_epi16(in00, 2);
in01 = _mm_srai_epi16(in01, 2);
in02 = _mm_srai_epi16(in02, 2);
in03 = _mm_srai_epi16(in03, 2);
in04 = _mm_srai_epi16(in04, 2);
in05 = _mm_srai_epi16(in05, 2);
in06 = _mm_srai_epi16(in06, 2);
in07 = _mm_srai_epi16(in07, 2);
in08 = _mm_srai_epi16(in08, 2);
in09 = _mm_srai_epi16(in09, 2);
in10 = _mm_srai_epi16(in10, 2);
in11 = _mm_srai_epi16(in11, 2);
in12 = _mm_srai_epi16(in12, 2);
in13 = _mm_srai_epi16(in13, 2);
in14 = _mm_srai_epi16(in14, 2);
in15 = _mm_srai_epi16(in15, 2);
}
in += 8;
// Calculate input for the first 8 results.
{
input0 = _mm_add_epi16(in00, in15);
input1 = _mm_add_epi16(in01, in14);
input2 = _mm_add_epi16(in02, in13);
input3 = _mm_add_epi16(in03, in12);
input4 = _mm_add_epi16(in04, in11);
input5 = _mm_add_epi16(in05, in10);
input6 = _mm_add_epi16(in06, in09);
input7 = _mm_add_epi16(in07, in08);
}
// Calculate input for the next 8 results.
{
step1_0 = _mm_sub_epi16(in07, in08);
step1_1 = _mm_sub_epi16(in06, in09);
step1_2 = _mm_sub_epi16(in05, in10);
step1_3 = _mm_sub_epi16(in04, in11);
step1_4 = _mm_sub_epi16(in03, in12);
step1_5 = _mm_sub_epi16(in02, in13);
step1_6 = _mm_sub_epi16(in01, in14);
step1_7 = _mm_sub_epi16(in00, in15);
}
// Work on the first eight values; fdct8_1d(input, even_results);
{
// Add/substract
const __m128i q0 = _mm_add_epi16(input0, input7);
const __m128i q1 = _mm_add_epi16(input1, input6);
const __m128i q2 = _mm_add_epi16(input2, input5);
const __m128i q3 = _mm_add_epi16(input3, input4);
const __m128i q4 = _mm_sub_epi16(input3, input4);
const __m128i q5 = _mm_sub_epi16(input2, input5);
const __m128i q6 = _mm_sub_epi16(input1, input6);
const __m128i q7 = _mm_sub_epi16(input0, input7);
// Work on first four results
{
// Add/substract
const __m128i r0 = _mm_add_epi16(q0, q3);
const __m128i r1 = _mm_add_epi16(q1, q2);
const __m128i r2 = _mm_sub_epi16(q1, q2);
const __m128i r3 = _mm_sub_epi16(q0, q3);
// Interleave to do the multiply by constants which gets us
// into 32 bits.
const __m128i t0 = _mm_unpacklo_epi16(r0, r1);
const __m128i t1 = _mm_unpackhi_epi16(r0, r1);
const __m128i t2 = _mm_unpacklo_epi16(r2, r3);
const __m128i t3 = _mm_unpackhi_epi16(r2, r3);
const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16);
const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_p16);
const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16);
const __m128i u3 = _mm_madd_epi16(t1, k__cospi_p16_m16);
const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08);
const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p24_p08);
const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24);
const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m08_p24);
// dct_const_round_shift
const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
// Combine
res00 = _mm_packs_epi32(w0, w1);
res08 = _mm_packs_epi32(w2, w3);
res04 = _mm_packs_epi32(w4, w5);
res12 = _mm_packs_epi32(w6, w7);
}
// Work on next four results
{
// Interleave to do the multiply by constants which gets us
// into 32 bits.
const __m128i d0 = _mm_unpacklo_epi16(q6, q5);
const __m128i d1 = _mm_unpackhi_epi16(q6, q5);
const __m128i e0 = _mm_madd_epi16(d0, k__cospi_p16_m16);
const __m128i e1 = _mm_madd_epi16(d1, k__cospi_p16_m16);
const __m128i e2 = _mm_madd_epi16(d0, k__cospi_p16_p16);
const __m128i e3 = _mm_madd_epi16(d1, k__cospi_p16_p16);
// dct_const_round_shift
const __m128i f0 = _mm_add_epi32(e0, k__DCT_CONST_ROUNDING);
const __m128i f1 = _mm_add_epi32(e1, k__DCT_CONST_ROUNDING);
const __m128i f2 = _mm_add_epi32(e2, k__DCT_CONST_ROUNDING);
const __m128i f3 = _mm_add_epi32(e3, k__DCT_CONST_ROUNDING);
const __m128i s0 = _mm_srai_epi32(f0, DCT_CONST_BITS);
const __m128i s1 = _mm_srai_epi32(f1, DCT_CONST_BITS);
const __m128i s2 = _mm_srai_epi32(f2, DCT_CONST_BITS);
const __m128i s3 = _mm_srai_epi32(f3, DCT_CONST_BITS);
// Combine
const __m128i r0 = _mm_packs_epi32(s0, s1);
const __m128i r1 = _mm_packs_epi32(s2, s3);
// Add/substract
const __m128i x0 = _mm_add_epi16(q4, r0);
const __m128i x1 = _mm_sub_epi16(q4, r0);
const __m128i x2 = _mm_sub_epi16(q7, r1);
const __m128i x3 = _mm_add_epi16(q7, r1);
// Interleave to do the multiply by constants which gets us
// into 32 bits.
const __m128i t0 = _mm_unpacklo_epi16(x0, x3);
const __m128i t1 = _mm_unpackhi_epi16(x0, x3);
const __m128i t2 = _mm_unpacklo_epi16(x1, x2);
const __m128i t3 = _mm_unpackhi_epi16(x1, x2);
const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p28_p04);
const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p28_p04);
const __m128i u2 = _mm_madd_epi16(t0, k__cospi_m04_p28);
const __m128i u3 = _mm_madd_epi16(t1, k__cospi_m04_p28);
const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p12_p20);
const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p12_p20);
const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m20_p12);
const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m20_p12);
// dct_const_round_shift
const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
// Combine
res02 = _mm_packs_epi32(w0, w1);
res14 = _mm_packs_epi32(w2, w3);
res10 = _mm_packs_epi32(w4, w5);
res06 = _mm_packs_epi32(w6, w7);
}
}
// Work on the next eight values; step1 -> odd_results
{
// step 2
{
const __m128i t0 = _mm_unpacklo_epi16(step1_5, step1_2);
const __m128i t1 = _mm_unpackhi_epi16(step1_5, step1_2);
const __m128i t2 = _mm_unpacklo_epi16(step1_4, step1_3);
const __m128i t3 = _mm_unpackhi_epi16(step1_4, step1_3);
const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_m16);
const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_m16);
const __m128i u2 = _mm_madd_epi16(t2, k__cospi_p16_m16);
const __m128i u3 = _mm_madd_epi16(t3, k__cospi_p16_m16);
// dct_const_round_shift
const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
// Combine
step2_2 = _mm_packs_epi32(w0, w1);
step2_3 = _mm_packs_epi32(w2, w3);
}
{
const __m128i t0 = _mm_unpacklo_epi16(step1_5, step1_2);
const __m128i t1 = _mm_unpackhi_epi16(step1_5, step1_2);
const __m128i t2 = _mm_unpacklo_epi16(step1_4, step1_3);
const __m128i t3 = _mm_unpackhi_epi16(step1_4, step1_3);
const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16);
const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_p16);
const __m128i u2 = _mm_madd_epi16(t2, k__cospi_p16_p16);
const __m128i u3 = _mm_madd_epi16(t3, k__cospi_p16_p16);
// dct_const_round_shift
const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
// Combine
step2_5 = _mm_packs_epi32(w0, w1);
step2_4 = _mm_packs_epi32(w2, w3);
}
// step 3
{
step3_0 = _mm_add_epi16(step1_0, step2_3);
step3_1 = _mm_add_epi16(step1_1, step2_2);
step3_2 = _mm_sub_epi16(step1_1, step2_2);
step3_3 = _mm_sub_epi16(step1_0, step2_3);
step3_4 = _mm_sub_epi16(step1_7, step2_4);
step3_5 = _mm_sub_epi16(step1_6, step2_5);
step3_6 = _mm_add_epi16(step1_6, step2_5);
step3_7 = _mm_add_epi16(step1_7, step2_4);
}
// step 4
{
const __m128i t0 = _mm_unpacklo_epi16(step3_1, step3_6);
const __m128i t1 = _mm_unpackhi_epi16(step3_1, step3_6);
const __m128i t2 = _mm_unpacklo_epi16(step3_2, step3_5);
const __m128i t3 = _mm_unpackhi_epi16(step3_2, step3_5);
const __m128i u0 = _mm_madd_epi16(t0, k__cospi_m08_p24);
const __m128i u1 = _mm_madd_epi16(t1, k__cospi_m08_p24);
const __m128i u2 = _mm_madd_epi16(t2, k__cospi_m24_m08);
const __m128i u3 = _mm_madd_epi16(t3, k__cospi_m24_m08);
// dct_const_round_shift
const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
// Combine
step2_1 = _mm_packs_epi32(w0, w1);
step2_2 = _mm_packs_epi32(w2, w3);
}
{
const __m128i t0 = _mm_unpacklo_epi16(step3_1, step3_6);
const __m128i t1 = _mm_unpackhi_epi16(step3_1, step3_6);
const __m128i t2 = _mm_unpacklo_epi16(step3_2, step3_5);
const __m128i t3 = _mm_unpackhi_epi16(step3_2, step3_5);
const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p24_p08);
const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p24_p08);
const __m128i u2 = _mm_madd_epi16(t2, k__cospi_m08_p24);
const __m128i u3 = _mm_madd_epi16(t3, k__cospi_m08_p24);
// dct_const_round_shift
const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
// Combine
step2_6 = _mm_packs_epi32(w0, w1);
step2_5 = _mm_packs_epi32(w2, w3);
}
// step 5
{
step1_0 = _mm_add_epi16(step3_0, step2_1);
step1_1 = _mm_sub_epi16(step3_0, step2_1);
step1_2 = _mm_sub_epi16(step3_3, step2_2);
step1_3 = _mm_add_epi16(step3_3, step2_2);
step1_4 = _mm_add_epi16(step3_4, step2_5);
step1_5 = _mm_sub_epi16(step3_4, step2_5);
step1_6 = _mm_sub_epi16(step3_7, step2_6);
step1_7 = _mm_add_epi16(step3_7, step2_6);
}
// step 6
{
const __m128i t0 = _mm_unpacklo_epi16(step1_0, step1_7);
const __m128i t1 = _mm_unpackhi_epi16(step1_0, step1_7);
const __m128i t2 = _mm_unpacklo_epi16(step1_1, step1_6);
const __m128i t3 = _mm_unpackhi_epi16(step1_1, step1_6);
const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p30_p02);
const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p30_p02);
const __m128i u2 = _mm_madd_epi16(t2, k__cospi_p14_p18);
const __m128i u3 = _mm_madd_epi16(t3, k__cospi_p14_p18);
// dct_const_round_shift
const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
// Combine
res01 = _mm_packs_epi32(w0, w1);
res09 = _mm_packs_epi32(w2, w3);
}
{
const __m128i t0 = _mm_unpacklo_epi16(step1_2, step1_5);
const __m128i t1 = _mm_unpackhi_epi16(step1_2, step1_5);
const __m128i t2 = _mm_unpacklo_epi16(step1_3, step1_4);
const __m128i t3 = _mm_unpackhi_epi16(step1_3, step1_4);
const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p22_p10);
const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p22_p10);
const __m128i u2 = _mm_madd_epi16(t2, k__cospi_p06_p26);
const __m128i u3 = _mm_madd_epi16(t3, k__cospi_p06_p26);
// dct_const_round_shift
const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
// Combine
res05 = _mm_packs_epi32(w0, w1);
res13 = _mm_packs_epi32(w2, w3);
}
{
const __m128i t0 = _mm_unpacklo_epi16(step1_2, step1_5);
const __m128i t1 = _mm_unpackhi_epi16(step1_2, step1_5);
const __m128i t2 = _mm_unpacklo_epi16(step1_3, step1_4);
const __m128i t3 = _mm_unpackhi_epi16(step1_3, step1_4);
const __m128i u0 = _mm_madd_epi16(t0, k__cospi_m10_p22);
const __m128i u1 = _mm_madd_epi16(t1, k__cospi_m10_p22);
const __m128i u2 = _mm_madd_epi16(t2, k__cospi_m26_p06);
const __m128i u3 = _mm_madd_epi16(t3, k__cospi_m26_p06);
// dct_const_round_shift
const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
// Combine
res11 = _mm_packs_epi32(w0, w1);
res03 = _mm_packs_epi32(w2, w3);
}
{
const __m128i t0 = _mm_unpacklo_epi16(step1_0, step1_7);
const __m128i t1 = _mm_unpackhi_epi16(step1_0, step1_7);
const __m128i t2 = _mm_unpacklo_epi16(step1_1, step1_6);
const __m128i t3 = _mm_unpackhi_epi16(step1_1, step1_6);
const __m128i u0 = _mm_madd_epi16(t0, k__cospi_m02_p30);
const __m128i u1 = _mm_madd_epi16(t1, k__cospi_m02_p30);
const __m128i u2 = _mm_madd_epi16(t2, k__cospi_m18_p14);
const __m128i u3 = _mm_madd_epi16(t3, k__cospi_m18_p14);
// dct_const_round_shift
const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
// Combine
res15 = _mm_packs_epi32(w0, w1);
res07 = _mm_packs_epi32(w2, w3);
}
}
// Transpose the results, do it as two 8x8 transposes.
{
// 00 01 02 03 04 05 06 07
// 10 11 12 13 14 15 16 17
// 20 21 22 23 24 25 26 27
// 30 31 32 33 34 35 36 37
// 40 41 42 43 44 45 46 47
// 50 51 52 53 54 55 56 57
// 60 61 62 63 64 65 66 67
// 70 71 72 73 74 75 76 77
const __m128i tr0_0 = _mm_unpacklo_epi16(res00, res01);
const __m128i tr0_1 = _mm_unpacklo_epi16(res02, res03);
const __m128i tr0_2 = _mm_unpackhi_epi16(res00, res01);
const __m128i tr0_3 = _mm_unpackhi_epi16(res02, res03);
const __m128i tr0_4 = _mm_unpacklo_epi16(res04, res05);
const __m128i tr0_5 = _mm_unpacklo_epi16(res06, res07);
const __m128i tr0_6 = _mm_unpackhi_epi16(res04, res05);
const __m128i tr0_7 = _mm_unpackhi_epi16(res06, res07);
// 00 10 01 11 02 12 03 13
// 20 30 21 31 22 32 23 33
// 04 14 05 15 06 16 07 17
// 24 34 25 35 26 36 27 37
// 40 50 41 51 42 52 43 53
// 60 70 61 71 62 72 63 73
// 54 54 55 55 56 56 57 57
// 64 74 65 75 66 76 67 77
const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3);
const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3);
const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
// 00 10 20 30 01 11 21 31
// 40 50 60 70 41 51 61 71
// 02 12 22 32 03 13 23 33
// 42 52 62 72 43 53 63 73
// 04 14 24 34 05 15 21 36
// 44 54 64 74 45 55 61 76
// 06 16 26 36 07 17 27 37
// 46 56 66 76 47 57 67 77
const __m128i tr2_0 = _mm_unpacklo_epi64(tr1_0, tr1_4);
const __m128i tr2_1 = _mm_unpackhi_epi64(tr1_0, tr1_4);
const __m128i tr2_2 = _mm_unpacklo_epi64(tr1_2, tr1_6);
const __m128i tr2_3 = _mm_unpackhi_epi64(tr1_2, tr1_6);
const __m128i tr2_4 = _mm_unpacklo_epi64(tr1_1, tr1_5);
const __m128i tr2_5 = _mm_unpackhi_epi64(tr1_1, tr1_5);
const __m128i tr2_6 = _mm_unpacklo_epi64(tr1_3, tr1_7);
const __m128i tr2_7 = _mm_unpackhi_epi64(tr1_3, tr1_7);
// 00 10 20 30 40 50 60 70
// 01 11 21 31 41 51 61 71
// 02 12 22 32 42 52 62 72
// 03 13 23 33 43 53 63 73
// 04 14 24 34 44 54 64 74
// 05 15 25 35 45 55 65 75
// 06 16 26 36 46 56 66 76
// 07 17 27 37 47 57 67 77
_mm_storeu_si128((__m128i *)(out + 0 * 16), tr2_0);
_mm_storeu_si128((__m128i *)(out + 1 * 16), tr2_1);
_mm_storeu_si128((__m128i *)(out + 2 * 16), tr2_2);
_mm_storeu_si128((__m128i *)(out + 3 * 16), tr2_3);
_mm_storeu_si128((__m128i *)(out + 4 * 16), tr2_4);
_mm_storeu_si128((__m128i *)(out + 5 * 16), tr2_5);
_mm_storeu_si128((__m128i *)(out + 6 * 16), tr2_6);
_mm_storeu_si128((__m128i *)(out + 7 * 16), tr2_7);
}
{
// 00 01 02 03 04 05 06 07
// 10 11 12 13 14 15 16 17
// 20 21 22 23 24 25 26 27
// 30 31 32 33 34 35 36 37
// 40 41 42 43 44 45 46 47
// 50 51 52 53 54 55 56 57
// 60 61 62 63 64 65 66 67
// 70 71 72 73 74 75 76 77
const __m128i tr0_0 = _mm_unpacklo_epi16(res08, res09);
const __m128i tr0_1 = _mm_unpacklo_epi16(res10, res11);
const __m128i tr0_2 = _mm_unpackhi_epi16(res08, res09);
const __m128i tr0_3 = _mm_unpackhi_epi16(res10, res11);
const __m128i tr0_4 = _mm_unpacklo_epi16(res12, res13);
const __m128i tr0_5 = _mm_unpacklo_epi16(res14, res15);
const __m128i tr0_6 = _mm_unpackhi_epi16(res12, res13);
const __m128i tr0_7 = _mm_unpackhi_epi16(res14, res15);
// 00 10 01 11 02 12 03 13
// 20 30 21 31 22 32 23 33
// 04 14 05 15 06 16 07 17
// 24 34 25 35 26 36 27 37
// 40 50 41 51 42 52 43 53
// 60 70 61 71 62 72 63 73
// 54 54 55 55 56 56 57 57
// 64 74 65 75 66 76 67 77
const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3);
const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3);
const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
// 00 10 20 30 01 11 21 31
// 40 50 60 70 41 51 61 71
// 02 12 22 32 03 13 23 33
// 42 52 62 72 43 53 63 73
// 04 14 24 34 05 15 21 36
// 44 54 64 74 45 55 61 76
// 06 16 26 36 07 17 27 37
// 46 56 66 76 47 57 67 77
const __m128i tr2_0 = _mm_unpacklo_epi64(tr1_0, tr1_4);
const __m128i tr2_1 = _mm_unpackhi_epi64(tr1_0, tr1_4);
const __m128i tr2_2 = _mm_unpacklo_epi64(tr1_2, tr1_6);
const __m128i tr2_3 = _mm_unpackhi_epi64(tr1_2, tr1_6);
const __m128i tr2_4 = _mm_unpacklo_epi64(tr1_1, tr1_5);
const __m128i tr2_5 = _mm_unpackhi_epi64(tr1_1, tr1_5);
const __m128i tr2_6 = _mm_unpacklo_epi64(tr1_3, tr1_7);
const __m128i tr2_7 = _mm_unpackhi_epi64(tr1_3, tr1_7);
// 00 10 20 30 40 50 60 70
// 01 11 21 31 41 51 61 71
// 02 12 22 32 42 52 62 72
// 03 13 23 33 43 53 63 73
// 04 14 24 34 44 54 64 74
// 05 15 25 35 45 55 65 75
// 06 16 26 36 46 56 66 76
// 07 17 27 37 47 57 67 77
// Store results
_mm_storeu_si128((__m128i *)(out + 8 + 0 * 16), tr2_0);
_mm_storeu_si128((__m128i *)(out + 8 + 1 * 16), tr2_1);
_mm_storeu_si128((__m128i *)(out + 8 + 2 * 16), tr2_2);
_mm_storeu_si128((__m128i *)(out + 8 + 3 * 16), tr2_3);
_mm_storeu_si128((__m128i *)(out + 8 + 4 * 16), tr2_4);
_mm_storeu_si128((__m128i *)(out + 8 + 5 * 16), tr2_5);
_mm_storeu_si128((__m128i *)(out + 8 + 6 * 16), tr2_6);
_mm_storeu_si128((__m128i *)(out + 8 + 7 * 16), tr2_7);
}
out += 8*16;
}
// Setup in/out for next pass.
in = intermediate;
out = output;
}
}
void vp9_short_fdct8x8_sse2(int16_t *input, int16_t *output, int pitch) {
const int stride = pitch >> 1;
int pass;
// Constants
// When we use them, in one case, they are all the same. In all others
// it's a pair of them that we need to repeat four times. This is done
// by constructing the 32 bit constant corresponding to that pair.
const __m128i k__cospi_p16_p16 = _mm_set1_epi16(cospi_16_64);
const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64);
const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64);
const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
// Load input
__m128i in0 = _mm_loadu_si128((const __m128i *)(input + 0 * stride));
__m128i in1 = _mm_loadu_si128((const __m128i *)(input + 1 * stride));
__m128i in2 = _mm_loadu_si128((const __m128i *)(input + 2 * stride));
__m128i in3 = _mm_loadu_si128((const __m128i *)(input + 3 * stride));
__m128i in4 = _mm_loadu_si128((const __m128i *)(input + 4 * stride));
__m128i in5 = _mm_loadu_si128((const __m128i *)(input + 5 * stride));
__m128i in6 = _mm_loadu_si128((const __m128i *)(input + 6 * stride));
__m128i in7 = _mm_loadu_si128((const __m128i *)(input + 7 * stride));
// Pre-condition input (shift by two)
in0 = _mm_slli_epi16(in0, 2);
in1 = _mm_slli_epi16(in1, 2);
in2 = _mm_slli_epi16(in2, 2);
in3 = _mm_slli_epi16(in3, 2);
in4 = _mm_slli_epi16(in4, 2);
in5 = _mm_slli_epi16(in5, 2);
in6 = _mm_slli_epi16(in6, 2);
in7 = _mm_slli_epi16(in7, 2);
// We do two passes, first the columns, then the rows. The results of the
// first pass are transposed so that the same column code can be reused. The
// results of the second pass are also transposed so that the rows (processed
// as columns) are put back in row positions.
for (pass = 0; pass < 2; pass++) {
// To store results of each pass before the transpose.
__m128i res0, res1, res2, res3, res4, res5, res6, res7;
// Add/substract
const __m128i q0 = _mm_add_epi16(in0, in7);
const __m128i q1 = _mm_add_epi16(in1, in6);
const __m128i q2 = _mm_add_epi16(in2, in5);
const __m128i q3 = _mm_add_epi16(in3, in4);
const __m128i q4 = _mm_sub_epi16(in3, in4);
const __m128i q5 = _mm_sub_epi16(in2, in5);
const __m128i q6 = _mm_sub_epi16(in1, in6);
const __m128i q7 = _mm_sub_epi16(in0, in7);
// Work on first four results
{
// Add/substract
const __m128i r0 = _mm_add_epi16(q0, q3);
const __m128i r1 = _mm_add_epi16(q1, q2);
const __m128i r2 = _mm_sub_epi16(q1, q2);
const __m128i r3 = _mm_sub_epi16(q0, q3);
// Interleave to do the multiply by constants which gets us into 32bits
const __m128i t0 = _mm_unpacklo_epi16(r0, r1);
const __m128i t1 = _mm_unpackhi_epi16(r0, r1);
const __m128i t2 = _mm_unpacklo_epi16(r2, r3);
const __m128i t3 = _mm_unpackhi_epi16(r2, r3);
const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16);
const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_p16);
const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16);
const __m128i u3 = _mm_madd_epi16(t1, k__cospi_p16_m16);
const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08);
const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p24_p08);
const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24);
const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m08_p24);
// dct_const_round_shift
const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
// Combine
res0 = _mm_packs_epi32(w0, w1);
res4 = _mm_packs_epi32(w2, w3);
res2 = _mm_packs_epi32(w4, w5);
res6 = _mm_packs_epi32(w6, w7);
}
// Work on next four results
{
// Interleave to do the multiply by constants which gets us into 32bits
const __m128i d0 = _mm_unpacklo_epi16(q6, q5);
const __m128i d1 = _mm_unpackhi_epi16(q6, q5);
const __m128i e0 = _mm_madd_epi16(d0, k__cospi_p16_m16);
const __m128i e1 = _mm_madd_epi16(d1, k__cospi_p16_m16);
const __m128i e2 = _mm_madd_epi16(d0, k__cospi_p16_p16);
const __m128i e3 = _mm_madd_epi16(d1, k__cospi_p16_p16);
// dct_const_round_shift
const __m128i f0 = _mm_add_epi32(e0, k__DCT_CONST_ROUNDING);
const __m128i f1 = _mm_add_epi32(e1, k__DCT_CONST_ROUNDING);
const __m128i f2 = _mm_add_epi32(e2, k__DCT_CONST_ROUNDING);
const __m128i f3 = _mm_add_epi32(e3, k__DCT_CONST_ROUNDING);
const __m128i s0 = _mm_srai_epi32(f0, DCT_CONST_BITS);
const __m128i s1 = _mm_srai_epi32(f1, DCT_CONST_BITS);
const __m128i s2 = _mm_srai_epi32(f2, DCT_CONST_BITS);
const __m128i s3 = _mm_srai_epi32(f3, DCT_CONST_BITS);
// Combine
const __m128i r0 = _mm_packs_epi32(s0, s1);
const __m128i r1 = _mm_packs_epi32(s2, s3);
// Add/substract
const __m128i x0 = _mm_add_epi16(q4, r0);
const __m128i x1 = _mm_sub_epi16(q4, r0);
const __m128i x2 = _mm_sub_epi16(q7, r1);
const __m128i x3 = _mm_add_epi16(q7, r1);
// Interleave to do the multiply by constants which gets us into 32bits
const __m128i t0 = _mm_unpacklo_epi16(x0, x3);
const __m128i t1 = _mm_unpackhi_epi16(x0, x3);
const __m128i t2 = _mm_unpacklo_epi16(x1, x2);
const __m128i t3 = _mm_unpackhi_epi16(x1, x2);
const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p28_p04);
const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p28_p04);
const __m128i u2 = _mm_madd_epi16(t0, k__cospi_m04_p28);
const __m128i u3 = _mm_madd_epi16(t1, k__cospi_m04_p28);
const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p12_p20);
const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p12_p20);
const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m20_p12);
const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m20_p12);
// dct_const_round_shift
const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
// Combine
res1 = _mm_packs_epi32(w0, w1);
res7 = _mm_packs_epi32(w2, w3);
res5 = _mm_packs_epi32(w4, w5);
res3 = _mm_packs_epi32(w6, w7);
}
// Transpose the 8x8.
{
// 00 01 02 03 04 05 06 07
// 10 11 12 13 14 15 16 17
// 20 21 22 23 24 25 26 27
// 30 31 32 33 34 35 36 37
// 40 41 42 43 44 45 46 47
// 50 51 52 53 54 55 56 57
// 60 61 62 63 64 65 66 67
// 70 71 72 73 74 75 76 77
const __m128i tr0_0 = _mm_unpacklo_epi16(res0, res1);
const __m128i tr0_1 = _mm_unpacklo_epi16(res2, res3);
const __m128i tr0_2 = _mm_unpackhi_epi16(res0, res1);
const __m128i tr0_3 = _mm_unpackhi_epi16(res2, res3);
const __m128i tr0_4 = _mm_unpacklo_epi16(res4, res5);
const __m128i tr0_5 = _mm_unpacklo_epi16(res6, res7);
const __m128i tr0_6 = _mm_unpackhi_epi16(res4, res5);
const __m128i tr0_7 = _mm_unpackhi_epi16(res6, res7);
// 00 10 01 11 02 12 03 13
// 20 30 21 31 22 32 23 33
// 04 14 05 15 06 16 07 17
// 24 34 25 35 26 36 27 37
// 40 50 41 51 42 52 43 53
// 60 70 61 71 62 72 63 73
// 54 54 55 55 56 56 57 57
// 64 74 65 75 66 76 67 77
const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3);
const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3);
const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
// 00 10 20 30 01 11 21 31
// 40 50 60 70 41 51 61 71
// 02 12 22 32 03 13 23 33
// 42 52 62 72 43 53 63 73
// 04 14 24 34 05 15 21 36
// 44 54 64 74 45 55 61 76
// 06 16 26 36 07 17 27 37
// 46 56 66 76 47 57 67 77
in0 = _mm_unpacklo_epi64(tr1_0, tr1_4);
in1 = _mm_unpackhi_epi64(tr1_0, tr1_4);
in2 = _mm_unpacklo_epi64(tr1_2, tr1_6);
in3 = _mm_unpackhi_epi64(tr1_2, tr1_6);
in4 = _mm_unpacklo_epi64(tr1_1, tr1_5);
in5 = _mm_unpackhi_epi64(tr1_1, tr1_5);
in6 = _mm_unpacklo_epi64(tr1_3, tr1_7);
in7 = _mm_unpackhi_epi64(tr1_3, tr1_7);
// 00 10 20 30 40 50 60 70
// 01 11 21 31 41 51 61 71
// 02 12 22 32 42 52 62 72
// 03 13 23 33 43 53 63 73
// 04 14 24 34 44 54 64 74
// 05 15 25 35 45 55 65 75
// 06 16 26 36 46 56 66 76
// 07 17 27 37 47 57 67 77
}
}
// Post-condition output and store it
{
// Post-condition (division by two)
// division of two 16 bits signed numbers using shifts
// n / 2 = (n - (n >> 15)) >> 1
const __m128i sign_in0 = _mm_srai_epi16(in0, 15);
const __m128i sign_in1 = _mm_srai_epi16(in1, 15);
const __m128i sign_in2 = _mm_srai_epi16(in2, 15);
const __m128i sign_in3 = _mm_srai_epi16(in3, 15);
const __m128i sign_in4 = _mm_srai_epi16(in4, 15);
const __m128i sign_in5 = _mm_srai_epi16(in5, 15);
const __m128i sign_in6 = _mm_srai_epi16(in6, 15);
const __m128i sign_in7 = _mm_srai_epi16(in7, 15);
in0 = _mm_sub_epi16(in0, sign_in0);
in1 = _mm_sub_epi16(in1, sign_in1);
in2 = _mm_sub_epi16(in2, sign_in2);
in3 = _mm_sub_epi16(in3, sign_in3);
in4 = _mm_sub_epi16(in4, sign_in4);
in5 = _mm_sub_epi16(in5, sign_in5);
in6 = _mm_sub_epi16(in6, sign_in6);
in7 = _mm_sub_epi16(in7, sign_in7);
in0 = _mm_srai_epi16(in0, 1);
in1 = _mm_srai_epi16(in1, 1);
in2 = _mm_srai_epi16(in2, 1);
in3 = _mm_srai_epi16(in3, 1);
in4 = _mm_srai_epi16(in4, 1);
in5 = _mm_srai_epi16(in5, 1);
in6 = _mm_srai_epi16(in6, 1);
in7 = _mm_srai_epi16(in7, 1);
// store results
_mm_storeu_si128((__m128i *)(output + 0 * 8), in0);
_mm_storeu_si128((__m128i *)(output + 1 * 8), in1);
_mm_storeu_si128((__m128i *)(output + 2 * 8), in2);
_mm_storeu_si128((__m128i *)(output + 3 * 8), in3);
_mm_storeu_si128((__m128i *)(output + 4 * 8), in4);
_mm_storeu_si128((__m128i *)(output + 5 * 8), in5);
_mm_storeu_si128((__m128i *)(output + 6 * 8), in6);
_mm_storeu_si128((__m128i *)(output + 7 * 8), in7);
}
}
/*
********************************************************************************
*
* @brief This function performs a 4x4 inverse hadamard transform on the 4x4 DC coefficients
* of a 16x16 intra prediction macroblock, and then performs scaling.
* prediction buffer
*
* @par Description:
* The DC coefficients pass through a 2-stage inverse hadamard transform.
* This inverse transformed content is scaled to based on Qp value.
*
* @param[in] pi2_src
* input 4x4 block of DC coefficients
*
* @param[out] pi2_out
* output 4x4 block
*
* @param[in] pu2_iscal_mat
* pointer to scaling list
*
* @param[in] pu2_weigh_mat
* pointer to weight matrix
*
* @param[in] u4_qp_div_6
* Floor (qp/6)
*
* @param[in] pi4_tmp
* temporary buffer of size 1*16
*
* @returns none
*
* @remarks none
*
*******************************************************************************
*/
void ih264_ihadamard_scaling_4x4_ssse3(WORD16* pi2_src,
WORD16* pi2_out,
const UWORD16 *pu2_iscal_mat,
const UWORD16 *pu2_weigh_mat,
UWORD32 u4_qp_div_6,
WORD32* pi4_tmp)
{
int val = 0xFFFF;
__m128i src_r0_r1, src_r2_r3, sign_reg, zero_8x16b = _mm_setzero_si128();
__m128i src_r0, src_r1, src_r2, src_r3;
__m128i temp0, temp1, temp2, temp3;
__m128i add_rshift = _mm_set1_epi32((1 << (5 - u4_qp_div_6)));
__m128i mult_val = _mm_set1_epi32(pu2_iscal_mat[0] * pu2_weigh_mat[0]);
__m128i mask = _mm_set1_epi32(val);
UNUSED (pi4_tmp);
mult_val = _mm_and_si128(mult_val, mask);
src_r0_r1 = _mm_loadu_si128((__m128i *) (pi2_src)); //a00 a01 a02 a03 a10 a11 a12 a13 -- the source matrix 0th,1st row
src_r2_r3 = _mm_loadu_si128((__m128i *) (pi2_src + 8)); //a20 a21 a22 a23 a30 a31 a32 a33 -- the source matrix 2nd,3rd row
sign_reg = _mm_cmpgt_epi16(zero_8x16b, src_r0_r1);
src_r0 = _mm_unpacklo_epi16(src_r0_r1, sign_reg);
src_r1 = _mm_unpackhi_epi16(src_r0_r1, sign_reg);
sign_reg = _mm_cmpgt_epi16(zero_8x16b, src_r2_r3);
src_r2 = _mm_unpacklo_epi16(src_r2_r3, sign_reg);
src_r3 = _mm_unpackhi_epi16(src_r2_r3, sign_reg);
/* Perform Inverse transform */
/*-------------------------------------------------------------*/
/* IDCT [ Horizontal transformation ] */
/*-------------------------------------------------------------*/
// Matrix transpose
/*
* a0 a1 a2 a3
* b0 b1 b2 b3
* c0 c1 c2 c3
* d0 d1 d2 d3
*/
temp0 = _mm_unpacklo_epi32(src_r0, src_r1); //a0 b0 a1 b1
temp2 = _mm_unpacklo_epi32(src_r2, src_r3); //c0 d0 c1 d1
temp1 = _mm_unpackhi_epi32(src_r0, src_r1); //a2 b2 a3 b3
temp3 = _mm_unpackhi_epi32(src_r2, src_r3); //c2 d2 c3 d3
src_r0 = _mm_unpacklo_epi64(temp0, temp2); //a0 b0 c0 d0
src_r1 = _mm_unpackhi_epi64(temp0, temp2); //a1 b1 c1 d1
src_r2 = _mm_unpacklo_epi64(temp1, temp3); //a2 b2 c2 d2
src_r3 = _mm_unpackhi_epi64(temp1, temp3); //a3 b3 c3 d3
temp0 = _mm_add_epi32(src_r0, src_r3);
temp1 = _mm_add_epi32(src_r1, src_r2);
temp2 = _mm_sub_epi32(src_r1, src_r2);
temp3 = _mm_sub_epi32(src_r0, src_r3);
src_r0 = _mm_add_epi32(temp0, temp1);
src_r1 = _mm_add_epi32(temp2, temp3);
src_r2 = _mm_sub_epi32(temp0, temp1);
src_r3 = _mm_sub_epi32(temp3, temp2);
/*-------------------------------------------------------------*/
/* IDCT [ Vertical transformation ] */
/*-------------------------------------------------------------*/
// Matrix transpose
/*
* a0 b0 c0 d0
* a1 b1 c1 d1
* a2 b2 c2 d2
* a3 b3 c3 d3
*/
temp0 = _mm_unpacklo_epi32(src_r0, src_r1); //a0 a1 b0 b1
temp2 = _mm_unpacklo_epi32(src_r2, src_r3); //a2 a3 b2 b3
temp1 = _mm_unpackhi_epi32(src_r0, src_r1); //c0 c1 d0 d1
temp3 = _mm_unpackhi_epi32(src_r2, src_r3); //c2 c3 d2 d3
src_r0 = _mm_unpacklo_epi64(temp0, temp2); //a0 a1 a2 a3
src_r1 = _mm_unpackhi_epi64(temp0, temp2); //b0 b1 b2 b3
src_r2 = _mm_unpacklo_epi64(temp1, temp3); //c0 c1 c2 c3
src_r3 = _mm_unpackhi_epi64(temp1, temp3); //d0 d1 d2 d3
temp0 = _mm_add_epi32(src_r0, src_r3);
temp1 = _mm_add_epi32(src_r1, src_r2);
temp2 = _mm_sub_epi32(src_r1, src_r2);
temp3 = _mm_sub_epi32(src_r0, src_r3);
src_r0 = _mm_add_epi32(temp0, temp1);
src_r1 = _mm_add_epi32(temp2, temp3);
src_r2 = _mm_sub_epi32(temp0, temp1);
src_r3 = _mm_sub_epi32(temp3, temp2);
src_r0 = _mm_and_si128(src_r0, mask);
src_r1 = _mm_and_si128(src_r1, mask);
src_r2 = _mm_and_si128(src_r2, mask);
src_r3 = _mm_and_si128(src_r3, mask);
src_r0 = _mm_madd_epi16(src_r0, mult_val);
src_r1 = _mm_madd_epi16(src_r1, mult_val);
src_r2 = _mm_madd_epi16(src_r2, mult_val);
src_r3 = _mm_madd_epi16(src_r3, mult_val);
//Scaling
if(u4_qp_div_6 >= 6)
{
src_r0 = _mm_slli_epi32(src_r0, u4_qp_div_6 - 6);
src_r1 = _mm_slli_epi32(src_r1, u4_qp_div_6 - 6);
src_r2 = _mm_slli_epi32(src_r2, u4_qp_div_6 - 6);
src_r3 = _mm_slli_epi32(src_r3, u4_qp_div_6 - 6);
}
else
{
temp0 = _mm_add_epi32(src_r0, add_rshift);
temp1 = _mm_add_epi32(src_r1, add_rshift);
temp2 = _mm_add_epi32(src_r2, add_rshift);
temp3 = _mm_add_epi32(src_r3, add_rshift);
src_r0 = _mm_srai_epi32(temp0, 6 - u4_qp_div_6);
src_r1 = _mm_srai_epi32(temp1, 6 - u4_qp_div_6);
src_r2 = _mm_srai_epi32(temp2, 6 - u4_qp_div_6);
src_r3 = _mm_srai_epi32(temp3, 6 - u4_qp_div_6);
}
src_r0_r1 = _mm_packs_epi32(src_r0, src_r1);
src_r2_r3 = _mm_packs_epi32(src_r2, src_r3);
_mm_storeu_si128((__m128i *) (&pi2_out[0]), src_r0_r1);
_mm_storeu_si128((__m128i *) (&pi2_out[8]), src_r2_r3);
}
/* motion templates */
CV_IMPL void
cvUpdateMotionHistory( const void* silhouette, void* mhimg,
double timestamp, double mhi_duration )
{
CvMat silhstub, *silh = cvGetMat(silhouette, &silhstub);
CvMat mhistub, *mhi = cvGetMat(mhimg, &mhistub);
if( !CV_IS_MASK_ARR( silh ))
CV_Error( CV_StsBadMask, "" );
if( CV_MAT_TYPE( mhi->type ) != CV_32FC1 )
CV_Error( CV_StsUnsupportedFormat, "" );
if( !CV_ARE_SIZES_EQ( mhi, silh ))
CV_Error( CV_StsUnmatchedSizes, "" );
CvSize size = cvGetMatSize( mhi );
if( CV_IS_MAT_CONT( mhi->type & silh->type ))
{
size.width *= size.height;
size.height = 1;
}
float ts = (float)timestamp;
float delbound = (float)(timestamp - mhi_duration);
int x, y;
#if CV_SSE2
volatile bool useSIMD = cv::checkHardwareSupport(CV_CPU_SSE2);
#endif
for( y = 0; y < size.height; y++ )
{
const uchar* silhData = silh->data.ptr + silh->step*y;
float* mhiData = (float*)(mhi->data.ptr + mhi->step*y);
x = 0;
#if CV_SSE2
if( useSIMD )
{
__m128 ts4 = _mm_set1_ps(ts), db4 = _mm_set1_ps(delbound);
for( ; x <= size.width - 8; x += 8 )
{
__m128i z = _mm_setzero_si128();
__m128i s = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i*)(silhData + x)), z);
__m128 s0 = _mm_cvtepi32_ps(_mm_unpacklo_epi16(s, z)), s1 = _mm_cvtepi32_ps(_mm_unpackhi_epi16(s, z));
__m128 v0 = _mm_loadu_ps(mhiData + x), v1 = _mm_loadu_ps(mhiData + x + 4);
__m128 fz = _mm_setzero_ps();
v0 = _mm_and_ps(v0, _mm_cmpge_ps(v0, db4));
v1 = _mm_and_ps(v1, _mm_cmpge_ps(v1, db4));
__m128 m0 = _mm_and_ps(_mm_xor_ps(v0, ts4), _mm_cmpneq_ps(s0, fz));
__m128 m1 = _mm_and_ps(_mm_xor_ps(v1, ts4), _mm_cmpneq_ps(s1, fz));
v0 = _mm_xor_ps(v0, m0);
v1 = _mm_xor_ps(v1, m1);
_mm_storeu_ps(mhiData + x, v0);
_mm_storeu_ps(mhiData + x + 4, v1);
}
}
#endif
for( ; x < size.width; x++ )
{
float val = mhiData[x];
val = silhData[x] ? ts : val < delbound ? 0 : val;
mhiData[x] = val;
}
}
}
/// CURRENTLY SAME CODE AS SCALAR !!
/// REPLACE HERE WITH SSE intrinsics
static void partialButterflyInverse16_simd(short *src, short *dst, int shift)
{
int add = 1<<(shift-1);
//we cast the original 16X16 matrix to an SIMD vector type
__m128i *g_aiT16_vec = (__m128i *)g_aiT16;
//We cast the input source (which is basically random numbers(see the main function for details)) to an SIMD vector type
//We also cast the output to an SIMD vector type
__m128i *in_vec = (__m128i *) src;
__m128i *out_vec = (__m128i *) dst;
//we declare an 8X8 array and cast it to an SIMD vector type
short gt[8][8] __attribute__ ((aligned (16)));
__m128i *gt_vec = (__m128i *)gt;
//we declare an 16X16 array and cast it to an SIMD vector type
short random[16][16] __attribute__ ((aligned (16)));
__m128i *random_vec = (__m128i *)random;
trans_g_aiT16(g_aiT16_vec,gt_vec);
tranpose8x8(in_vec,2, random_vec,0);
tranpose8x8(in_vec,3, random_vec,8);
tranpose8x8(in_vec,0, random_vec,16);
tranpose8x8(in_vec,1, random_vec,24);
for (int j=0; j<16; j++)
{
/* Utilizing symmetry properties to the maximum to minimize the number of multiplications */
__m128i I0 = _mm_load_si128 (&random_vec[j]);
__m128i II0 = _mm_load_si128 (&random_vec[j+16]);
// for (int k=0; k<8; k++)
//here we are loading up the transposed values in the initial matrix
//multiplying it with the input numbers to produce intermediate 32-bit integers
// we then sum up adjacent pairs of 32-bit integers and store them in the destination register
__m128i I1 = _mm_load_si128 (>_vec[0]);
__m128i I2 = _mm_madd_epi16 (I1, I0);
__m128i I3 = _mm_load_si128 (>_vec[1]);
__m128i I4 = _mm_madd_epi16 (I3, I0);
__m128i I5 = _mm_load_si128 (>_vec[2]);
__m128i I6 = _mm_madd_epi16 (I5, I0);
__m128i I7 = _mm_load_si128 (>_vec[3]);
__m128i I8 = _mm_madd_epi16 (I7, I0);
__m128i I9 = _mm_load_si128 (>_vec[4]);
__m128i I10 = _mm_madd_epi16 (I9, I0);
__m128i I11 = _mm_load_si128 (>_vec[5]);
__m128i I12 = _mm_madd_epi16 (I11, I0);
__m128i I13 = _mm_load_si128 (>_vec[6]);
__m128i I14 = _mm_madd_epi16 (I13, I0);
__m128i I15 = _mm_load_si128 (>_vec[7]);
__m128i I16 = _mm_madd_epi16 (I15, I0);
//horizontally add the partial results obtained from thee previous step
__m128i A1 =_mm_hadd_epi32 (I2, I4);
__m128i A2 =_mm_hadd_epi32 (I6, I8);
__m128i R1 =_mm_hadd_epi32 (A1, A2);
__m128i A3 =_mm_hadd_epi32 (I10, I12);
__m128i A4 =_mm_hadd_epi32 (I14, I16);
__m128i R2 =_mm_hadd_epi32 (A3, A4);
// O[k] = T[0]+T[1]+T[2]+T[3];
// for (int k=0; k<4; k++)
// {
//load the original matrix values, multiply it with the random values
//store the low bits to I2 and the hi bits to I3
I1 = _mm_load_si128 (>_vec[8]);
I2 = _mm_mullo_epi16 (I1, II0);
I3 = _mm_mulhi_epi16 (I1, II0);
__m128i lowI23 = _mm_unpacklo_epi16(I2,I3);
__m128i hiI23 = _mm_unpackhi_epi16(I2,I3);
__m128i temp1 = _mm_add_epi32(lowI23,hiI23);
__m128i temp5 = _mm_hsub_epi32 (lowI23, hiI23);
I4 = _mm_load_si128 (>_vec[9]);
I5 = _mm_mullo_epi16 (I4, II0);
I6 = _mm_mulhi_epi16 (I4, II0);
__m128i lowI56 = _mm_unpacklo_epi16(I5,I6);
__m128i hiI56 = _mm_unpackhi_epi16(I5,I6);
__m128i temp2 = _mm_add_epi32(lowI56,hiI56);
__m128i temp6 = _mm_hsub_epi32 (lowI56, hiI56);
I7 = _mm_load_si128 (>_vec[10]);
I8 = _mm_mullo_epi16 (I7, II0);
I9 = _mm_mulhi_epi16 (I7, II0);
__m128i lowI89 = _mm_unpacklo_epi16(I8,I9);
__m128i hiI89 = _mm_unpackhi_epi16(I8,I9);
__m128i temp3 = _mm_add_epi32(lowI89,hiI89);
__m128i temp7 = _mm_hsub_epi32 (lowI89, hiI89);
I10 = _mm_load_si128 (>_vec[11]);
I11 = _mm_mullo_epi16 (I10, II0);
I12 = _mm_mulhi_epi16 (I10, II0);
__m128i lowI1112 = _mm_unpacklo_epi16(I11,I12);
__m128i hiI1112 = _mm_unpackhi_epi16(I11,I12);
__m128i temp4 = _mm_add_epi32(lowI1112,hiI1112);
__m128i temp8 = _mm_hsub_epi32 (lowI1112, hiI1112);
__m128i A5 =_mm_hadd_epi32 (temp1, temp2);
__m128i A6 =_mm_hadd_epi32 (temp3, temp4);
__m128i R3 =_mm_hadd_epi32 (A5, A6);
__m128i A7 =_mm_hadd_epi32 (temp8, temp7);
__m128i A8 =_mm_hadd_epi32 (temp6, temp5);
__m128i R4 =_mm_hadd_epi32 (A7, A8);
///////////////////////////
__m128i add_reg = _mm_set1_epi32(add);
__m128i sum_vec0 = _mm_add_epi32(R3,R1);
sum_vec0 = _mm_add_epi32(sum_vec0,add_reg);
sum_vec0 = _mm_srai_epi32(sum_vec0, shift); // shift right
__m128i sum_vec1 = _mm_add_epi32(R4,R2);
sum_vec1 = _mm_add_epi32(sum_vec1,add_reg);
sum_vec1 = _mm_srai_epi32(sum_vec1, shift); // shift right
__m128i finalres0 = _mm_packs_epi32(sum_vec0, sum_vec1); // shrink packed 32bit to packed 16 bit and saturate
_mm_store_si128 (&out_vec[2*j], finalres0);
__m128i sum_vec2 = _mm_sub_epi32(R4, R2);
sum_vec2 = _mm_add_epi32(sum_vec2,add_reg);
sum_vec2 = _mm_srai_epi32(sum_vec2, shift); // shift right
__m128i sum_vec3 = _mm_sub_epi32(R3, R1);
sum_vec3 = _mm_add_epi32(sum_vec3,add_reg);
sum_vec3 = _mm_srai_epi32(sum_vec3, shift); // shift right
I5 = _mm_unpackhi_epi32(sum_vec2, sum_vec3);
I6 = _mm_unpacklo_epi32(sum_vec2, sum_vec3);
I7 = _mm_unpackhi_epi32(I5, I6);
I8 = _mm_unpacklo_epi32(I5, I6);
I9 = _mm_unpacklo_epi32(I7, I8);
I10 = _mm_unpackhi_epi32(I7, I8);
sum_vec3 = _mm_packs_epi32(I9, I10); // shrink packed 32bit to packed 16 bit and saturate
_mm_store_si128 (&out_vec[2*j+1], sum_vec3);
}
}
/*__forceinline*/ bool Cmp_ClutBuffer_SavedClut<u16>(u16* saved_clut, u32 csa, u32 clutsize)
{
assert((clutsize&31) == 0);
#ifdef ZEROGS_SSE2
__m128i zero_128 = _mm_setzero_si128();
#endif
u16* clut = (u16*)GetClutBufferAddress<u32>(csa); // Keep aligned version for sse2
// which side to cmp
u32 clutsize_right;
u32 clutsize_left;
if (csa < 16) {
clutsize_right = min(clutsize, (16-csa)*32);
clutsize_left = clutsize - clutsize_right;
} else {
clutsize_right = 0;
clutsize_left = clutsize;
}
while (clutsize_right > 0)
{
#ifdef ZEROGS_SSE2
// only lower 16 bits of dword are valid
__m128i clut_0 = _mm_load_si128((__m128i*)clut);
__m128i clut_1 = _mm_load_si128((__m128i*)clut+1);
__m128i clut_2 = _mm_load_si128((__m128i*)clut+2);
__m128i clut_3 = _mm_load_si128((__m128i*)clut+3);
// value must converted to 32 bits
__m128i saved_clut_0 = _mm_load_si128((__m128i*)saved_clut);
__m128i saved_clut_1 = _mm_load_si128((__m128i*)saved_clut+1);
__m128i result = _mm_cmpeq_epi16(_mm_unpacklo_epi16(saved_clut_0, zero_128), clut_0);
__m128i result_tmp = _mm_cmpeq_epi16(_mm_unpackhi_epi16(saved_clut_0, zero_128), clut_1);
result = _mm_and_si128(result, result_tmp);
result_tmp = _mm_cmpeq_epi16(_mm_unpacklo_epi16(saved_clut_1, zero_128), clut_2);
result = _mm_and_si128(result, result_tmp);
result_tmp = _mm_cmpeq_epi16(_mm_unpackhi_epi16(saved_clut_1, zero_128), clut_3);
result = _mm_and_si128(result, result_tmp);
u32 result_int = _mm_movemask_epi8(result);
// only lower 16bits must be checked
if ((result_int&0x3333) != 0x3333)
return true;
#else
for (int i = 0; i < 16; ++i)
if (saved_clut[i] != clut[2*i]) return true;
#endif
saved_clut += 16;
clut += 32;
clutsize_right -= 32;
}
if(csa < 16) {
// go back to the base before processing left clut column
clut = (u16*)GetClutBufferAddress<u32>(0); // Keep aligned version for sse2
}
while (clutsize_left > 0)
{
#ifdef ZEROGS_SSE2
// only higher 16 bits of dword are valid
__m128i clut_0 = _mm_load_si128((__m128i*)clut);
__m128i clut_1 = _mm_load_si128((__m128i*)clut+1);
__m128i clut_2 = _mm_load_si128((__m128i*)clut+2);
__m128i clut_3 = _mm_load_si128((__m128i*)clut+3);
// value must converted to 32 bits (with 0 in lower 16 bits)
__m128i saved_clut_0 = _mm_load_si128((__m128i*)saved_clut);
__m128i saved_clut_1 = _mm_load_si128((__m128i*)saved_clut+1);
__m128i result = _mm_cmpeq_epi16(_mm_unpacklo_epi16(zero_128, saved_clut_0), clut_0);
__m128i result_tmp = _mm_cmpeq_epi16(_mm_unpackhi_epi16(zero_128, saved_clut_0), clut_1);
result = _mm_and_si128(result, result_tmp);
result_tmp = _mm_cmpeq_epi16(_mm_unpacklo_epi16(zero_128, saved_clut_1), clut_2);
result = _mm_and_si128(result, result_tmp);
result_tmp = _mm_cmpeq_epi16(_mm_unpackhi_epi16(zero_128, saved_clut_1), clut_3);
result = _mm_and_si128(result, result_tmp);
u32 result_int = _mm_movemask_epi8(result);
// only higher 16bits must be checked
if ((result_int&0xCCCC) != 0xCCCC)
return true;
#else
// Note +1 because we change higher 16 bits
for (int i = 0; i < 16; ++i)
if (saved_clut[i] != clut[2*i+1]) return true;
#endif
saved_clut += 16;
clut += 32;
clutsize_left -= 32;
}
return false;
}
static void GF_FUNC_ALIGN VS_CC
proc_16bit_sse2(convolution_t *ch, uint8_t *buff, int bstride, int width,
int height, int stride, uint8_t *d, const uint8_t *s)
{
const uint16_t *srcp = (uint16_t *)s;
uint16_t *dstp = (uint16_t *)d;
stride /= 2;
bstride /= 2;
uint16_t *p0 = (uint16_t *)buff + 8;
uint16_t *p1 = p0 + bstride;
uint16_t *p2 = p1 + bstride;
uint16_t *p3 = p2 + bstride;
uint16_t *p4 = p3 + bstride;
uint16_t *orig = p0, *end = p4;
line_copy16(p0, srcp + 2 * stride, width, 2);
line_copy16(p1, srcp + stride, width, 2);
line_copy16(p2, srcp, width, 2);
srcp += stride;
line_copy16(p3, srcp, width, 2);
__m128i zero = _mm_setzero_si128();
__m128 rdiv = _mm_set1_ps((float)ch->rdiv);
__m128 bias = _mm_set1_ps((float)ch->bias);
__m128i max = _mm_set1_epi32(0xFFFF);
__m128 matrix[25];
for (int i = 0; i < 25; i++) {
matrix[i] = _mm_set1_ps((float)ch->m[i]);
}
for (int y = 0; y < height; y++) {
srcp += stride * (y < height - 2 ? 1 : -1);
line_copy16(p4, srcp, width, 2);
uint16_t *array[] = {
p0 - 2, p0 - 1, p0, p0 + 1, p0 + 2,
p1 - 2, p1 - 1, p1, p1 + 1, p1 + 2,
p2 - 2, p2 - 1, p2, p2 + 1, p2 + 2,
p3 - 2, p3 - 1, p3, p3 + 1, p3 + 2,
p4 - 2, p4 - 1, p4, p4 + 1, p4 + 2
};
for (int x = 0; x < width; x += 8) {
__m128 sum[2] = {(__m128)zero, (__m128)zero};
for (int i = 0; i < 25; i++) {
__m128i xmm0 = _mm_loadu_si128((__m128i *)(array[i] + x));
__m128 xmm1 = _mm_cvtepi32_ps(_mm_unpacklo_epi16(xmm0, zero));
__m128 xmm2 = _mm_cvtepi32_ps(_mm_unpackhi_epi16(xmm0, zero));
xmm1 = _mm_mul_ps(xmm1, matrix[i]);
xmm2 = _mm_mul_ps(xmm2, matrix[i]);
sum[0] = _mm_add_ps(sum[0], xmm1);
sum[1] = _mm_add_ps(sum[1], xmm2);
}
__m128i sumi[2];
for (int i = 0; i < 2; i++) {
sum[i] = _mm_mul_ps(sum[i], rdiv);
sum[i] = _mm_add_ps(sum[i], bias);
if (!ch->saturate) {
sum[i] = mm_abs_ps(sum[i]);
}
sumi[i] = _mm_cvtps_epi32(sum[i]);
sumi[i] = mm_min_epi32(sumi[i], max);
__m128i mask = _mm_cmpgt_epi32(sumi[i], zero);
sumi[i] = _mm_and_si128(sumi[i], mask);
}
sumi[0] = mm_cast_epi32(sumi[0], sumi[1]);
_mm_store_si128((__m128i *)(dstp + x), sumi[0]);
}
dstp += stride;
p0 = p1;
p1 = p2;
p2 = p3;
p3 = p4;
p4 = (p4 == end) ? orig : p4 + bstride;
}
}
static void FTransformSSE2(const uint8_t* src, const uint8_t* ref,
int16_t* out) {
const __m128i zero = _mm_setzero_si128();
const __m128i seven = _mm_set1_epi16(7);
const __m128i k7500 = _mm_set1_epi32(7500);
const __m128i k14500 = _mm_set1_epi32(14500);
const __m128i k51000 = _mm_set1_epi32(51000);
const __m128i k12000_plus_one = _mm_set1_epi32(12000 + (1 << 16));
const __m128i k5352_2217 = _mm_set_epi16(5352, 2217, 5352, 2217,
5352, 2217, 5352, 2217);
const __m128i k2217_5352 = _mm_set_epi16(2217, -5352, 2217, -5352,
2217, -5352, 2217, -5352);
__m128i v01, v32;
// Difference between src and ref and initial transpose.
{
// Load src and convert to 16b.
const __m128i src0 = _mm_loadl_epi64((__m128i*)&src[0 * BPS]);
const __m128i src1 = _mm_loadl_epi64((__m128i*)&src[1 * BPS]);
const __m128i src2 = _mm_loadl_epi64((__m128i*)&src[2 * BPS]);
const __m128i src3 = _mm_loadl_epi64((__m128i*)&src[3 * BPS]);
const __m128i src_0 = _mm_unpacklo_epi8(src0, zero);
const __m128i src_1 = _mm_unpacklo_epi8(src1, zero);
const __m128i src_2 = _mm_unpacklo_epi8(src2, zero);
const __m128i src_3 = _mm_unpacklo_epi8(src3, zero);
// Load ref and convert to 16b.
const __m128i ref0 = _mm_loadl_epi64((__m128i*)&ref[0 * BPS]);
const __m128i ref1 = _mm_loadl_epi64((__m128i*)&ref[1 * BPS]);
const __m128i ref2 = _mm_loadl_epi64((__m128i*)&ref[2 * BPS]);
const __m128i ref3 = _mm_loadl_epi64((__m128i*)&ref[3 * BPS]);
const __m128i ref_0 = _mm_unpacklo_epi8(ref0, zero);
const __m128i ref_1 = _mm_unpacklo_epi8(ref1, zero);
const __m128i ref_2 = _mm_unpacklo_epi8(ref2, zero);
const __m128i ref_3 = _mm_unpacklo_epi8(ref3, zero);
// Compute difference.
const __m128i diff0 = _mm_sub_epi16(src_0, ref_0);
const __m128i diff1 = _mm_sub_epi16(src_1, ref_1);
const __m128i diff2 = _mm_sub_epi16(src_2, ref_2);
const __m128i diff3 = _mm_sub_epi16(src_3, ref_3);
// Transpose.
// 00 01 02 03 0 0 0 0
// 10 11 12 13 0 0 0 0
// 20 21 22 23 0 0 0 0
// 30 31 32 33 0 0 0 0
const __m128i transpose0_0 = _mm_unpacklo_epi16(diff0, diff1);
const __m128i transpose0_1 = _mm_unpacklo_epi16(diff2, diff3);
// 00 10 01 11 02 12 03 13
// 20 30 21 31 22 32 23 33
const __m128i v23 = _mm_unpackhi_epi32(transpose0_0, transpose0_1);
v01 = _mm_unpacklo_epi32(transpose0_0, transpose0_1);
v32 = _mm_shuffle_epi32(v23, _MM_SHUFFLE(1, 0, 3, 2));
// a02 a12 a22 a32 a03 a13 a23 a33
// a00 a10 a20 a30 a01 a11 a21 a31
// a03 a13 a23 a33 a02 a12 a22 a32
}
// First pass and subsequent transpose.
{
// Same operations are done on the (0,3) and (1,2) pairs.
// b0 = (a0 + a3) << 3
// b1 = (a1 + a2) << 3
// b3 = (a0 - a3) << 3
// b2 = (a1 - a2) << 3
const __m128i a01 = _mm_add_epi16(v01, v32);
const __m128i a32 = _mm_sub_epi16(v01, v32);
const __m128i b01 = _mm_slli_epi16(a01, 3);
const __m128i b32 = _mm_slli_epi16(a32, 3);
const __m128i b11 = _mm_unpackhi_epi64(b01, b01);
const __m128i b22 = _mm_unpackhi_epi64(b32, b32);
// e0 = b0 + b1
// e2 = b0 - b1
const __m128i e0 = _mm_add_epi16(b01, b11);
const __m128i e2 = _mm_sub_epi16(b01, b11);
const __m128i e02 = _mm_unpacklo_epi64(e0, e2);
// e1 = (b3 * 5352 + b2 * 2217 + 14500) >> 12
// e3 = (b3 * 2217 - b2 * 5352 + 7500) >> 12
const __m128i b23 = _mm_unpacklo_epi16(b22, b32);
const __m128i c1 = _mm_madd_epi16(b23, k5352_2217);
const __m128i c3 = _mm_madd_epi16(b23, k2217_5352);
const __m128i d1 = _mm_add_epi32(c1, k14500);
const __m128i d3 = _mm_add_epi32(c3, k7500);
const __m128i e1 = _mm_srai_epi32(d1, 12);
const __m128i e3 = _mm_srai_epi32(d3, 12);
const __m128i e13 = _mm_packs_epi32(e1, e3);
// Transpose.
// 00 01 02 03 20 21 22 23
// 10 11 12 13 30 31 32 33
const __m128i transpose0_0 = _mm_unpacklo_epi16(e02, e13);
const __m128i transpose0_1 = _mm_unpackhi_epi16(e02, e13);
// 00 10 01 11 02 12 03 13
// 20 30 21 31 22 32 23 33
const __m128i v23 = _mm_unpackhi_epi32(transpose0_0, transpose0_1);
v01 = _mm_unpacklo_epi32(transpose0_0, transpose0_1);
v32 = _mm_shuffle_epi32(v23, _MM_SHUFFLE(1, 0, 3, 2));
// 02 12 22 32 03 13 23 33
// 00 10 20 30 01 11 21 31
// 03 13 23 33 02 12 22 32
}
// Second pass
{
// Same operations are done on the (0,3) and (1,2) pairs.
// a0 = v0 + v3
// a1 = v1 + v2
// a3 = v0 - v3
// a2 = v1 - v2
const __m128i a01 = _mm_add_epi16(v01, v32);
const __m128i a32 = _mm_sub_epi16(v01, v32);
const __m128i a11 = _mm_unpackhi_epi64(a01, a01);
const __m128i a22 = _mm_unpackhi_epi64(a32, a32);
// d0 = (a0 + a1 + 7) >> 4;
// d2 = (a0 - a1 + 7) >> 4;
const __m128i b0 = _mm_add_epi16(a01, a11);
const __m128i b2 = _mm_sub_epi16(a01, a11);
const __m128i c0 = _mm_add_epi16(b0, seven);
const __m128i c2 = _mm_add_epi16(b2, seven);
const __m128i d0 = _mm_srai_epi16(c0, 4);
const __m128i d2 = _mm_srai_epi16(c2, 4);
// f1 = ((b3 * 5352 + b2 * 2217 + 12000) >> 16)
// f3 = ((b3 * 2217 - b2 * 5352 + 51000) >> 16)
const __m128i b23 = _mm_unpacklo_epi16(a22, a32);
const __m128i c1 = _mm_madd_epi16(b23, k5352_2217);
const __m128i c3 = _mm_madd_epi16(b23, k2217_5352);
const __m128i d1 = _mm_add_epi32(c1, k12000_plus_one);
const __m128i d3 = _mm_add_epi32(c3, k51000);
const __m128i e1 = _mm_srai_epi32(d1, 16);
const __m128i e3 = _mm_srai_epi32(d3, 16);
const __m128i f1 = _mm_packs_epi32(e1, e1);
const __m128i f3 = _mm_packs_epi32(e3, e3);
// f1 = f1 + (a3 != 0);
// The compare will return (0xffff, 0) for (==0, !=0). To turn that into the
// desired (0, 1), we add one earlier through k12000_plus_one.
const __m128i g1 = _mm_add_epi16(f1, _mm_cmpeq_epi16(a32, zero));
_mm_storel_epi64((__m128i*)&out[ 0], d0);
_mm_storel_epi64((__m128i*)&out[ 4], g1);
_mm_storel_epi64((__m128i*)&out[ 8], d2);
_mm_storel_epi64((__m128i*)&out[12], f3);
}
}
static void aom_filter_block1d4_v4_ssse3(
const uint8_t *src_ptr, ptrdiff_t src_pitch, uint8_t *output_ptr,
ptrdiff_t out_pitch, uint32_t output_height, const int16_t *filter) {
__m128i filtersReg;
__m128i addFilterReg32;
__m128i srcReg2, srcReg3, srcReg23, srcReg4, srcReg34, srcReg5, srcReg45,
srcReg6, srcReg56;
__m128i srcReg23_34_lo, srcReg45_56_lo;
__m128i srcReg2345_3456_lo, srcReg2345_3456_hi;
__m128i resReglo, resReghi;
__m128i firstFilters;
unsigned int i;
ptrdiff_t src_stride, dst_stride;
addFilterReg32 = _mm_set1_epi16(32);
filtersReg = _mm_loadu_si128((const __m128i *)filter);
// converting the 16 bit (short) to 8 bit (byte) and have the
// same data in both lanes of 128 bit register.
filtersReg = _mm_srai_epi16(filtersReg, 1);
filtersReg = _mm_packs_epi16(filtersReg, filtersReg);
firstFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi32(0x5040302u));
// multiple the size of the source and destination stride by two
src_stride = src_pitch << 1;
dst_stride = out_pitch << 1;
srcReg2 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 2));
srcReg3 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 3));
srcReg23 = _mm_unpacklo_epi32(srcReg2, srcReg3);
srcReg4 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 4));
// have consecutive loads on the same 256 register
srcReg34 = _mm_unpacklo_epi32(srcReg3, srcReg4);
srcReg23_34_lo = _mm_unpacklo_epi8(srcReg23, srcReg34);
for (i = output_height; i > 1; i -= 2) {
srcReg5 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 5));
srcReg45 = _mm_unpacklo_epi32(srcReg4, srcReg5);
srcReg6 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 6));
srcReg56 = _mm_unpacklo_epi32(srcReg5, srcReg6);
// merge every two consecutive registers
srcReg45_56_lo = _mm_unpacklo_epi8(srcReg45, srcReg56);
srcReg2345_3456_lo = _mm_unpacklo_epi16(srcReg23_34_lo, srcReg45_56_lo);
srcReg2345_3456_hi = _mm_unpackhi_epi16(srcReg23_34_lo, srcReg45_56_lo);
// multiply 2 adjacent elements with the filter and add the result
resReglo = _mm_maddubs_epi16(srcReg2345_3456_lo, firstFilters);
resReghi = _mm_maddubs_epi16(srcReg2345_3456_hi, firstFilters);
resReglo = _mm_hadds_epi16(resReglo, _mm_setzero_si128());
resReghi = _mm_hadds_epi16(resReghi, _mm_setzero_si128());
// shift by 6 bit each 16 bit
resReglo = _mm_adds_epi16(resReglo, addFilterReg32);
resReghi = _mm_adds_epi16(resReghi, addFilterReg32);
resReglo = _mm_srai_epi16(resReglo, 6);
resReghi = _mm_srai_epi16(resReghi, 6);
// shrink to 8 bit each 16 bits, the first lane contain the first
// convolve result and the second lane contain the second convolve
// result
resReglo = _mm_packus_epi16(resReglo, resReglo);
resReghi = _mm_packus_epi16(resReghi, resReghi);
src_ptr += src_stride;
*((uint32_t *)(output_ptr)) = _mm_cvtsi128_si32(resReglo);
*((uint32_t *)(output_ptr + out_pitch)) = _mm_cvtsi128_si32(resReghi);
output_ptr += dst_stride;
// save part of the registers for next strides
srcReg23_34_lo = srcReg45_56_lo;
srcReg4 = srcReg6;
}
}
// Hadamard transform
// Returns the difference between the weighted sum of the absolute value of
// transformed coefficients.
static int TTransformSSE2(const uint8_t* inA, const uint8_t* inB,
const uint16_t* const w) {
int32_t sum[4];
__m128i tmp_0, tmp_1, tmp_2, tmp_3;
const __m128i zero = _mm_setzero_si128();
const __m128i one = _mm_set1_epi16(1);
const __m128i three = _mm_set1_epi16(3);
// Load, combine and tranpose inputs.
{
const __m128i inA_0 = _mm_loadl_epi64((__m128i*)&inA[BPS * 0]);
const __m128i inA_1 = _mm_loadl_epi64((__m128i*)&inA[BPS * 1]);
const __m128i inA_2 = _mm_loadl_epi64((__m128i*)&inA[BPS * 2]);
const __m128i inA_3 = _mm_loadl_epi64((__m128i*)&inA[BPS * 3]);
const __m128i inB_0 = _mm_loadl_epi64((__m128i*)&inB[BPS * 0]);
const __m128i inB_1 = _mm_loadl_epi64((__m128i*)&inB[BPS * 1]);
const __m128i inB_2 = _mm_loadl_epi64((__m128i*)&inB[BPS * 2]);
const __m128i inB_3 = _mm_loadl_epi64((__m128i*)&inB[BPS * 3]);
// Combine inA and inB (we'll do two transforms in parallel).
const __m128i inAB_0 = _mm_unpacklo_epi8(inA_0, inB_0);
const __m128i inAB_1 = _mm_unpacklo_epi8(inA_1, inB_1);
const __m128i inAB_2 = _mm_unpacklo_epi8(inA_2, inB_2);
const __m128i inAB_3 = _mm_unpacklo_epi8(inA_3, inB_3);
// a00 b00 a01 b01 a02 b03 a03 b03 0 0 0 0 0 0 0 0
// a10 b10 a11 b11 a12 b12 a13 b13 0 0 0 0 0 0 0 0
// a20 b20 a21 b21 a22 b22 a23 b23 0 0 0 0 0 0 0 0
// a30 b30 a31 b31 a32 b32 a33 b33 0 0 0 0 0 0 0 0
// Transpose the two 4x4, discarding the filling zeroes.
const __m128i transpose0_0 = _mm_unpacklo_epi8(inAB_0, inAB_2);
const __m128i transpose0_1 = _mm_unpacklo_epi8(inAB_1, inAB_3);
// a00 a20 b00 b20 a01 a21 b01 b21 a02 a22 b02 b22 a03 a23 b03 b23
// a10 a30 b10 b30 a11 a31 b11 b31 a12 a32 b12 b32 a13 a33 b13 b33
const __m128i transpose1_0 = _mm_unpacklo_epi8(transpose0_0, transpose0_1);
const __m128i transpose1_1 = _mm_unpackhi_epi8(transpose0_0, transpose0_1);
// a00 a10 a20 a30 b00 b10 b20 b30 a01 a11 a21 a31 b01 b11 b21 b31
// a02 a12 a22 a32 b02 b12 b22 b32 a03 a13 a23 a33 b03 b13 b23 b33
// Convert to 16b.
tmp_0 = _mm_unpacklo_epi8(transpose1_0, zero);
tmp_1 = _mm_unpackhi_epi8(transpose1_0, zero);
tmp_2 = _mm_unpacklo_epi8(transpose1_1, zero);
tmp_3 = _mm_unpackhi_epi8(transpose1_1, zero);
// a00 a10 a20 a30 b00 b10 b20 b30
// a01 a11 a21 a31 b01 b11 b21 b31
// a02 a12 a22 a32 b02 b12 b22 b32
// a03 a13 a23 a33 b03 b13 b23 b33
}
// Horizontal pass and subsequent transpose.
{
// Calculate a and b (two 4x4 at once).
const __m128i a0 = _mm_slli_epi16(_mm_add_epi16(tmp_0, tmp_2), 2);
const __m128i a1 = _mm_slli_epi16(_mm_add_epi16(tmp_1, tmp_3), 2);
const __m128i a2 = _mm_slli_epi16(_mm_sub_epi16(tmp_1, tmp_3), 2);
const __m128i a3 = _mm_slli_epi16(_mm_sub_epi16(tmp_0, tmp_2), 2);
// b0_extra = (a0 != 0);
const __m128i b0_extra = _mm_andnot_si128(_mm_cmpeq_epi16 (a0, zero), one);
const __m128i b0_base = _mm_add_epi16(a0, a1);
const __m128i b1 = _mm_add_epi16(a3, a2);
const __m128i b2 = _mm_sub_epi16(a3, a2);
const __m128i b3 = _mm_sub_epi16(a0, a1);
const __m128i b0 = _mm_add_epi16(b0_base, b0_extra);
// a00 a01 a02 a03 b00 b01 b02 b03
// a10 a11 a12 a13 b10 b11 b12 b13
// a20 a21 a22 a23 b20 b21 b22 b23
// a30 a31 a32 a33 b30 b31 b32 b33
// Transpose the two 4x4.
const __m128i transpose0_0 = _mm_unpacklo_epi16(b0, b1);
const __m128i transpose0_1 = _mm_unpacklo_epi16(b2, b3);
const __m128i transpose0_2 = _mm_unpackhi_epi16(b0, b1);
const __m128i transpose0_3 = _mm_unpackhi_epi16(b2, b3);
// a00 a10 a01 a11 a02 a12 a03 a13
// a20 a30 a21 a31 a22 a32 a23 a33
// b00 b10 b01 b11 b02 b12 b03 b13
// b20 b30 b21 b31 b22 b32 b23 b33
const __m128i transpose1_0 = _mm_unpacklo_epi32(transpose0_0, transpose0_1);
const __m128i transpose1_1 = _mm_unpacklo_epi32(transpose0_2, transpose0_3);
const __m128i transpose1_2 = _mm_unpackhi_epi32(transpose0_0, transpose0_1);
const __m128i transpose1_3 = _mm_unpackhi_epi32(transpose0_2, transpose0_3);
// a00 a10 a20 a30 a01 a11 a21 a31
// b00 b10 b20 b30 b01 b11 b21 b31
// a02 a12 a22 a32 a03 a13 a23 a33
// b02 b12 a22 b32 b03 b13 b23 b33
tmp_0 = _mm_unpacklo_epi64(transpose1_0, transpose1_1);
tmp_1 = _mm_unpackhi_epi64(transpose1_0, transpose1_1);
tmp_2 = _mm_unpacklo_epi64(transpose1_2, transpose1_3);
tmp_3 = _mm_unpackhi_epi64(transpose1_2, transpose1_3);
// a00 a10 a20 a30 b00 b10 b20 b30
// a01 a11 a21 a31 b01 b11 b21 b31
// a02 a12 a22 a32 b02 b12 b22 b32
// a03 a13 a23 a33 b03 b13 b23 b33
}
// Vertical pass and difference of weighted sums.
{
// Load all inputs.
// TODO(cduvivier): Make variable declarations and allocations aligned so
// we can use _mm_load_si128 instead of _mm_loadu_si128.
const __m128i w_0 = _mm_loadu_si128((__m128i*)&w[0]);
const __m128i w_8 = _mm_loadu_si128((__m128i*)&w[8]);
// Calculate a and b (two 4x4 at once).
const __m128i a0 = _mm_add_epi16(tmp_0, tmp_2);
const __m128i a1 = _mm_add_epi16(tmp_1, tmp_3);
const __m128i a2 = _mm_sub_epi16(tmp_1, tmp_3);
const __m128i a3 = _mm_sub_epi16(tmp_0, tmp_2);
const __m128i b0 = _mm_add_epi16(a0, a1);
const __m128i b1 = _mm_add_epi16(a3, a2);
const __m128i b2 = _mm_sub_epi16(a3, a2);
const __m128i b3 = _mm_sub_epi16(a0, a1);
// Separate the transforms of inA and inB.
__m128i A_b0 = _mm_unpacklo_epi64(b0, b1);
__m128i A_b2 = _mm_unpacklo_epi64(b2, b3);
__m128i B_b0 = _mm_unpackhi_epi64(b0, b1);
__m128i B_b2 = _mm_unpackhi_epi64(b2, b3);
{
// sign(b) = b >> 15 (0x0000 if positive, 0xffff if negative)
const __m128i sign_A_b0 = _mm_srai_epi16(A_b0, 15);
const __m128i sign_A_b2 = _mm_srai_epi16(A_b2, 15);
const __m128i sign_B_b0 = _mm_srai_epi16(B_b0, 15);
const __m128i sign_B_b2 = _mm_srai_epi16(B_b2, 15);
// b = abs(b) = (b ^ sign) - sign
A_b0 = _mm_xor_si128(A_b0, sign_A_b0);
A_b2 = _mm_xor_si128(A_b2, sign_A_b2);
B_b0 = _mm_xor_si128(B_b0, sign_B_b0);
B_b2 = _mm_xor_si128(B_b2, sign_B_b2);
A_b0 = _mm_sub_epi16(A_b0, sign_A_b0);
A_b2 = _mm_sub_epi16(A_b2, sign_A_b2);
B_b0 = _mm_sub_epi16(B_b0, sign_B_b0);
B_b2 = _mm_sub_epi16(B_b2, sign_B_b2);
}
// b = abs(b) + 3
A_b0 = _mm_add_epi16(A_b0, three);
A_b2 = _mm_add_epi16(A_b2, three);
B_b0 = _mm_add_epi16(B_b0, three);
B_b2 = _mm_add_epi16(B_b2, three);
// abs((b + (b<0) + 3) >> 3) = (abs(b) + 3) >> 3
// b = (abs(b) + 3) >> 3
A_b0 = _mm_srai_epi16(A_b0, 3);
A_b2 = _mm_srai_epi16(A_b2, 3);
B_b0 = _mm_srai_epi16(B_b0, 3);
B_b2 = _mm_srai_epi16(B_b2, 3);
// weighted sums
A_b0 = _mm_madd_epi16(A_b0, w_0);
A_b2 = _mm_madd_epi16(A_b2, w_8);
B_b0 = _mm_madd_epi16(B_b0, w_0);
B_b2 = _mm_madd_epi16(B_b2, w_8);
A_b0 = _mm_add_epi32(A_b0, A_b2);
B_b0 = _mm_add_epi32(B_b0, B_b2);
// difference of weighted sums
A_b0 = _mm_sub_epi32(A_b0, B_b0);
_mm_storeu_si128((__m128i*)&sum[0], A_b0);
}
return sum[0] + sum[1] + sum[2] + sum[3];
}
SIMD_INLINE __m128i MulDiv16(__m128i dividend, __m128i divisor, const __m128 & KF_255_DIV_6)
{
const __m128i quotientLo = MulDiv32(_mm_unpacklo_epi16(dividend, K_ZERO), _mm_unpacklo_epi16(divisor, K_ZERO), KF_255_DIV_6);
const __m128i quotientHi = MulDiv32(_mm_unpackhi_epi16(dividend, K_ZERO), _mm_unpackhi_epi16(divisor, K_ZERO), KF_255_DIV_6);
return _mm_packs_epi32(quotientLo, quotientHi);
}
// Simple quantization
static int QuantizeBlockSSE2(int16_t in[16], int16_t out[16],
int n, const VP8Matrix* const mtx) {
const __m128i max_coeff_2047 = _mm_set1_epi16(2047);
const __m128i zero = _mm_set1_epi16(0);
__m128i sign0, sign8;
__m128i coeff0, coeff8;
__m128i out0, out8;
__m128i packed_out;
// Load all inputs.
// TODO(cduvivier): Make variable declarations and allocations aligned so that
// we can use _mm_load_si128 instead of _mm_loadu_si128.
__m128i in0 = _mm_loadu_si128((__m128i*)&in[0]);
__m128i in8 = _mm_loadu_si128((__m128i*)&in[8]);
const __m128i sharpen0 = _mm_loadu_si128((__m128i*)&mtx->sharpen_[0]);
const __m128i sharpen8 = _mm_loadu_si128((__m128i*)&mtx->sharpen_[8]);
const __m128i iq0 = _mm_loadu_si128((__m128i*)&mtx->iq_[0]);
const __m128i iq8 = _mm_loadu_si128((__m128i*)&mtx->iq_[8]);
const __m128i bias0 = _mm_loadu_si128((__m128i*)&mtx->bias_[0]);
const __m128i bias8 = _mm_loadu_si128((__m128i*)&mtx->bias_[8]);
const __m128i q0 = _mm_loadu_si128((__m128i*)&mtx->q_[0]);
const __m128i q8 = _mm_loadu_si128((__m128i*)&mtx->q_[8]);
const __m128i zthresh0 = _mm_loadu_si128((__m128i*)&mtx->zthresh_[0]);
const __m128i zthresh8 = _mm_loadu_si128((__m128i*)&mtx->zthresh_[8]);
// sign(in) = in >> 15 (0x0000 if positive, 0xffff if negative)
sign0 = _mm_srai_epi16(in0, 15);
sign8 = _mm_srai_epi16(in8, 15);
// coeff = abs(in) = (in ^ sign) - sign
coeff0 = _mm_xor_si128(in0, sign0);
coeff8 = _mm_xor_si128(in8, sign8);
coeff0 = _mm_sub_epi16(coeff0, sign0);
coeff8 = _mm_sub_epi16(coeff8, sign8);
// coeff = abs(in) + sharpen
coeff0 = _mm_add_epi16(coeff0, sharpen0);
coeff8 = _mm_add_epi16(coeff8, sharpen8);
// if (coeff > 2047) coeff = 2047
coeff0 = _mm_min_epi16(coeff0, max_coeff_2047);
coeff8 = _mm_min_epi16(coeff8, max_coeff_2047);
// out = (coeff * iQ + B) >> QFIX;
{
// doing calculations with 32b precision (QFIX=17)
// out = (coeff * iQ)
__m128i coeff_iQ0H = _mm_mulhi_epu16(coeff0, iq0);
__m128i coeff_iQ0L = _mm_mullo_epi16(coeff0, iq0);
__m128i coeff_iQ8H = _mm_mulhi_epu16(coeff8, iq8);
__m128i coeff_iQ8L = _mm_mullo_epi16(coeff8, iq8);
__m128i out_00 = _mm_unpacklo_epi16(coeff_iQ0L, coeff_iQ0H);
__m128i out_04 = _mm_unpackhi_epi16(coeff_iQ0L, coeff_iQ0H);
__m128i out_08 = _mm_unpacklo_epi16(coeff_iQ8L, coeff_iQ8H);
__m128i out_12 = _mm_unpackhi_epi16(coeff_iQ8L, coeff_iQ8H);
// expand bias from 16b to 32b
__m128i bias_00 = _mm_unpacklo_epi16(bias0, zero);
__m128i bias_04 = _mm_unpackhi_epi16(bias0, zero);
__m128i bias_08 = _mm_unpacklo_epi16(bias8, zero);
__m128i bias_12 = _mm_unpackhi_epi16(bias8, zero);
// out = (coeff * iQ + B)
out_00 = _mm_add_epi32(out_00, bias_00);
out_04 = _mm_add_epi32(out_04, bias_04);
out_08 = _mm_add_epi32(out_08, bias_08);
out_12 = _mm_add_epi32(out_12, bias_12);
// out = (coeff * iQ + B) >> QFIX;
out_00 = _mm_srai_epi32(out_00, QFIX);
out_04 = _mm_srai_epi32(out_04, QFIX);
out_08 = _mm_srai_epi32(out_08, QFIX);
out_12 = _mm_srai_epi32(out_12, QFIX);
// pack result as 16b
out0 = _mm_packs_epi32(out_00, out_04);
out8 = _mm_packs_epi32(out_08, out_12);
}
// get sign back (if (sign[j]) out_n = -out_n)
out0 = _mm_xor_si128(out0, sign0);
out8 = _mm_xor_si128(out8, sign8);
out0 = _mm_sub_epi16(out0, sign0);
out8 = _mm_sub_epi16(out8, sign8);
// in = out * Q
in0 = _mm_mullo_epi16(out0, q0);
in8 = _mm_mullo_epi16(out8, q8);
// if (coeff <= mtx->zthresh_) {in=0; out=0;}
{
__m128i cmp0 = _mm_cmpgt_epi16(coeff0, zthresh0);
__m128i cmp8 = _mm_cmpgt_epi16(coeff8, zthresh8);
in0 = _mm_and_si128(in0, cmp0);
in8 = _mm_and_si128(in8, cmp8);
_mm_storeu_si128((__m128i*)&in[0], in0);
_mm_storeu_si128((__m128i*)&in[8], in8);
out0 = _mm_and_si128(out0, cmp0);
out8 = _mm_and_si128(out8, cmp8);
}
// zigzag the output before storing it.
//
// The zigzag pattern can almost be reproduced with a small sequence of
// shuffles. After it, we only need to swap the 7th (ending up in third
// position instead of twelfth) and 8th values.
{
__m128i outZ0, outZ8;
outZ0 = _mm_shufflehi_epi16(out0, _MM_SHUFFLE(2, 1, 3, 0));
outZ0 = _mm_shuffle_epi32 (outZ0, _MM_SHUFFLE(3, 1, 2, 0));
outZ0 = _mm_shufflehi_epi16(outZ0, _MM_SHUFFLE(3, 1, 0, 2));
outZ8 = _mm_shufflelo_epi16(out8, _MM_SHUFFLE(3, 0, 2, 1));
outZ8 = _mm_shuffle_epi32 (outZ8, _MM_SHUFFLE(3, 1, 2, 0));
outZ8 = _mm_shufflelo_epi16(outZ8, _MM_SHUFFLE(1, 3, 2, 0));
_mm_storeu_si128((__m128i*)&out[0], outZ0);
_mm_storeu_si128((__m128i*)&out[8], outZ8);
packed_out = _mm_packs_epi16(outZ0, outZ8);
}
{
const int16_t outZ_12 = out[12];
const int16_t outZ_3 = out[3];
out[3] = outZ_12;
out[12] = outZ_3;
}
// detect if all 'out' values are zeroes or not
{
int32_t tmp[4];
_mm_storeu_si128((__m128i*)tmp, packed_out);
if (n) {
tmp[0] &= ~0xff;
}
return (tmp[3] || tmp[2] || tmp[1] || tmp[0]);
}
}
void aom_highbd_comp_avg_upsampled_pred_sse2(uint16_t *comp_pred,
const uint8_t *pred8, int width,
int height, const uint8_t *ref8,
const int ref_stride) {
const __m128i one = _mm_set1_epi16(1);
const int stride = ref_stride << 3;
int i, j;
uint16_t *pred = CONVERT_TO_SHORTPTR(pred8);
uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
if (width >= 8) {
// read 8 points at one time
for (i = 0; i < height; i++) {
for (j = 0; j < width; j += 8) {
__m128i s0 = _mm_cvtsi32_si128(*(const uint32_t *)ref);
__m128i s1 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 8));
__m128i s2 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 16));
__m128i s3 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 24));
__m128i s4 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 32));
__m128i s5 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 40));
__m128i s6 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 48));
__m128i s7 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 56));
__m128i p0 = _mm_loadu_si128((const __m128i *)pred);
__m128i t0, t1, t2, t3;
t0 = _mm_unpacklo_epi16(s0, s1);
t1 = _mm_unpacklo_epi16(s2, s3);
t2 = _mm_unpacklo_epi16(s4, s5);
t3 = _mm_unpacklo_epi16(s6, s7);
t0 = _mm_unpacklo_epi32(t0, t1);
t2 = _mm_unpacklo_epi32(t2, t3);
t0 = _mm_unpacklo_epi64(t0, t2);
p0 = _mm_adds_epu16(t0, p0);
p0 = _mm_adds_epu16(p0, one);
p0 = _mm_srli_epi16(p0, 1);
_mm_storeu_si128((__m128i *)(comp_pred), p0);
comp_pred += 8;
pred += 8;
ref += 8 * 8;
}
ref += stride - (width << 3);
}
} else {
// read 4 points at one time
for (i = 0; i < height; i++) {
for (j = 0; j < width; j += 4) {
__m128i s0 = _mm_cvtsi32_si128(*(const uint32_t *)ref);
__m128i s1 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 8));
__m128i s2 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 16));
__m128i s3 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 24));
__m128i p0 = _mm_loadl_epi64((const __m128i *)pred);
__m128i t0, t1;
t0 = _mm_unpacklo_epi16(s0, s1);
t1 = _mm_unpacklo_epi16(s2, s3);
t0 = _mm_unpacklo_epi32(t0, t1);
p0 = _mm_adds_epu16(t0, p0);
p0 = _mm_adds_epu16(p0, one);
p0 = _mm_srli_epi16(p0, 1);
_mm_storel_epi64((__m128i *)(comp_pred), p0);
comp_pred += 4;
pred += 4;
ref += 4 * 8;
}
ref += stride - (width << 3);
}
}
}
rfx_dwt_2d_decode_block_horiz_sse2(INT16* l, INT16* h, INT16* dst, int subband_width)
{
int y, n;
INT16* l_ptr = l;
INT16* h_ptr = h;
INT16* dst_ptr = dst;
int first;
int last;
__m128i l_n;
__m128i h_n;
__m128i h_n_m;
__m128i tmp_n;
__m128i dst_n;
__m128i dst_n_p;
__m128i dst1;
__m128i dst2;
for (y = 0; y < subband_width; y++)
{
/* Even coefficients */
for (n = 0; n < subband_width; n += 8)
{
/* dst[2n] = l[n] - ((h[n-1] + h[n] + 1) >> 1); */
l_n = _mm_load_si128((__m128i*) l_ptr);
h_n = _mm_load_si128((__m128i*) h_ptr);
h_n_m = _mm_loadu_si128((__m128i*) (h_ptr - 1));
if (n == 0)
{
first = _mm_extract_epi16(h_n_m, 1);
h_n_m = _mm_insert_epi16(h_n_m, first, 0);
}
tmp_n = _mm_add_epi16(h_n, h_n_m);
tmp_n = _mm_add_epi16(tmp_n, _mm_set1_epi16(1));
tmp_n = _mm_srai_epi16(tmp_n, 1);
dst_n = _mm_sub_epi16(l_n, tmp_n);
_mm_store_si128((__m128i*) l_ptr, dst_n);
l_ptr += 8;
h_ptr += 8;
}
l_ptr -= subband_width;
h_ptr -= subband_width;
/* Odd coefficients */
for (n = 0; n < subband_width; n += 8)
{
/* dst[2n + 1] = (h[n] << 1) + ((dst[2n] + dst[2n + 2]) >> 1); */
h_n = _mm_load_si128((__m128i*) h_ptr);
h_n = _mm_slli_epi16(h_n, 1);
dst_n = _mm_load_si128((__m128i*) (l_ptr));
dst_n_p = _mm_loadu_si128((__m128i*) (l_ptr + 1));
if (n == subband_width - 8)
{
last = _mm_extract_epi16(dst_n_p, 6);
dst_n_p = _mm_insert_epi16(dst_n_p, last, 7);
}
tmp_n = _mm_add_epi16(dst_n_p, dst_n);
tmp_n = _mm_srai_epi16(tmp_n, 1);
tmp_n = _mm_add_epi16(tmp_n, h_n);
dst1 = _mm_unpacklo_epi16(dst_n, tmp_n);
dst2 = _mm_unpackhi_epi16(dst_n, tmp_n);
_mm_store_si128((__m128i*) dst_ptr, dst1);
_mm_store_si128((__m128i*) (dst_ptr + 8), dst2);
l_ptr += 8;
h_ptr += 8;
dst_ptr += 16;
}
}
}
static WEBP_INLINE int DoQuantizeBlock_SSE41(int16_t in[16], int16_t out[16],
const uint16_t* const sharpen,
const VP8Matrix* const mtx) {
const __m128i max_coeff_2047 = _mm_set1_epi16(MAX_LEVEL);
const __m128i zero = _mm_setzero_si128();
__m128i out0, out8;
__m128i packed_out;
// Load all inputs.
__m128i in0 = _mm_loadu_si128((__m128i*)&in[0]);
__m128i in8 = _mm_loadu_si128((__m128i*)&in[8]);
const __m128i iq0 = _mm_loadu_si128((const __m128i*)&mtx->iq_[0]);
const __m128i iq8 = _mm_loadu_si128((const __m128i*)&mtx->iq_[8]);
const __m128i q0 = _mm_loadu_si128((const __m128i*)&mtx->q_[0]);
const __m128i q8 = _mm_loadu_si128((const __m128i*)&mtx->q_[8]);
// coeff = abs(in)
__m128i coeff0 = _mm_abs_epi16(in0);
__m128i coeff8 = _mm_abs_epi16(in8);
// coeff = abs(in) + sharpen
if (sharpen != NULL) {
const __m128i sharpen0 = _mm_loadu_si128((const __m128i*)&sharpen[0]);
const __m128i sharpen8 = _mm_loadu_si128((const __m128i*)&sharpen[8]);
coeff0 = _mm_add_epi16(coeff0, sharpen0);
coeff8 = _mm_add_epi16(coeff8, sharpen8);
}
// out = (coeff * iQ + B) >> QFIX
{
// doing calculations with 32b precision (QFIX=17)
// out = (coeff * iQ)
const __m128i coeff_iQ0H = _mm_mulhi_epu16(coeff0, iq0);
const __m128i coeff_iQ0L = _mm_mullo_epi16(coeff0, iq0);
const __m128i coeff_iQ8H = _mm_mulhi_epu16(coeff8, iq8);
const __m128i coeff_iQ8L = _mm_mullo_epi16(coeff8, iq8);
__m128i out_00 = _mm_unpacklo_epi16(coeff_iQ0L, coeff_iQ0H);
__m128i out_04 = _mm_unpackhi_epi16(coeff_iQ0L, coeff_iQ0H);
__m128i out_08 = _mm_unpacklo_epi16(coeff_iQ8L, coeff_iQ8H);
__m128i out_12 = _mm_unpackhi_epi16(coeff_iQ8L, coeff_iQ8H);
// out = (coeff * iQ + B)
const __m128i bias_00 = _mm_loadu_si128((const __m128i*)&mtx->bias_[0]);
const __m128i bias_04 = _mm_loadu_si128((const __m128i*)&mtx->bias_[4]);
const __m128i bias_08 = _mm_loadu_si128((const __m128i*)&mtx->bias_[8]);
const __m128i bias_12 = _mm_loadu_si128((const __m128i*)&mtx->bias_[12]);
out_00 = _mm_add_epi32(out_00, bias_00);
out_04 = _mm_add_epi32(out_04, bias_04);
out_08 = _mm_add_epi32(out_08, bias_08);
out_12 = _mm_add_epi32(out_12, bias_12);
// out = QUANTDIV(coeff, iQ, B, QFIX)
out_00 = _mm_srai_epi32(out_00, QFIX);
out_04 = _mm_srai_epi32(out_04, QFIX);
out_08 = _mm_srai_epi32(out_08, QFIX);
out_12 = _mm_srai_epi32(out_12, QFIX);
// pack result as 16b
out0 = _mm_packs_epi32(out_00, out_04);
out8 = _mm_packs_epi32(out_08, out_12);
// if (coeff > 2047) coeff = 2047
out0 = _mm_min_epi16(out0, max_coeff_2047);
out8 = _mm_min_epi16(out8, max_coeff_2047);
}
// put sign back
out0 = _mm_sign_epi16(out0, in0);
out8 = _mm_sign_epi16(out8, in8);
// in = out * Q
in0 = _mm_mullo_epi16(out0, q0);
in8 = _mm_mullo_epi16(out8, q8);
_mm_storeu_si128((__m128i*)&in[0], in0);
_mm_storeu_si128((__m128i*)&in[8], in8);
// zigzag the output before storing it. The re-ordering is:
// 0 1 2 3 4 5 6 7 | 8 9 10 11 12 13 14 15
// -> 0 1 4[8]5 2 3 6 | 9 12 13 10 [7]11 14 15
// There's only two misplaced entries ([8] and [7]) that are crossing the
// reg's boundaries.
// We use pshufb instead of pshuflo/pshufhi.
{
const __m128i kCst_lo = PSHUFB_CST(0, 1, 4, -1, 5, 2, 3, 6);
const __m128i kCst_7 = PSHUFB_CST(-1, -1, -1, -1, 7, -1, -1, -1);
const __m128i tmp_lo = _mm_shuffle_epi8(out0, kCst_lo);
const __m128i tmp_7 = _mm_shuffle_epi8(out0, kCst_7); // extract #7
const __m128i kCst_hi = PSHUFB_CST(1, 4, 5, 2, -1, 3, 6, 7);
const __m128i kCst_8 = PSHUFB_CST(-1, -1, -1, 0, -1, -1, -1, -1);
const __m128i tmp_hi = _mm_shuffle_epi8(out8, kCst_hi);
const __m128i tmp_8 = _mm_shuffle_epi8(out8, kCst_8); // extract #8
const __m128i out_z0 = _mm_or_si128(tmp_lo, tmp_8);
const __m128i out_z8 = _mm_or_si128(tmp_hi, tmp_7);
_mm_storeu_si128((__m128i*)&out[0], out_z0);
_mm_storeu_si128((__m128i*)&out[8], out_z8);
packed_out = _mm_packs_epi16(out_z0, out_z8);
}
// detect if all 'out' values are zeroes or not
return (_mm_movemask_epi8(_mm_cmpeq_epi8(packed_out, zero)) != 0xffff);
}
static void tranpose8x8(__m128i *input,int i_indx, __m128i *Transposed,int t_indx)
{
__m128i a;
__m128i b;
__m128i c;
__m128i d;
__m128i e;
__m128i f;
__m128i g;
__m128i h;
__m128i temp1;
__m128i temp2;
__m128i temp3;
__m128i temp4;
__m128i temp5;
__m128i temp6;
__m128i temp7;
__m128i temp8;
__m128i temp9;
__m128i temp10;
__m128i temp11;
__m128i temp12;
__m128i temp13;
__m128i temp14;
__m128i temp15;
__m128i temp16;
__m128i T0;
__m128i T1;
__m128i T2;
__m128i T3;
__m128i T4;
__m128i T5;
__m128i T6;
__m128i T7;
a = _mm_load_si128(&input[i_indx]);
b = _mm_load_si128(&input[i_indx+4 ]);
c = _mm_load_si128(&input[i_indx+8 ]);
d = _mm_load_si128(&input[i_indx+12]);
e = _mm_load_si128(&input[i_indx+16]);
f = _mm_load_si128(&input[i_indx+20]);
g = _mm_load_si128(&input[i_indx+24]);
h = _mm_load_si128(&input[i_indx+28]);
temp1 = _mm_unpacklo_epi16(a, b); //a03b03
temp2 = _mm_unpacklo_epi16(c, d);
temp3 = _mm_unpacklo_epi16(e, f);
temp4 = _mm_unpacklo_epi16(g, h);
temp5 = _mm_unpackhi_epi16(a, b);
temp6 = _mm_unpackhi_epi16(c, d);
temp7 = _mm_unpackhi_epi16(e, f);
temp8 = _mm_unpackhi_epi16(g, h);
temp9 = _mm_unpacklo_epi32(temp1, temp2); //a01b01c01d01
temp10 = _mm_unpackhi_epi32(temp1, temp2);
temp11 = _mm_unpacklo_epi32(temp3, temp4);
temp12 = _mm_unpackhi_epi32(temp3, temp4);
temp13 = _mm_unpacklo_epi32(temp5, temp6);
temp14 = _mm_unpackhi_epi32(temp5, temp6);
temp15 = _mm_unpacklo_epi32(temp7, temp8);
temp16 = _mm_unpackhi_epi32(temp7, temp8);
T0 = _mm_unpacklo_epi64(temp9, temp11); //a0b0c0d0e0f0g0h0
T1 = _mm_unpackhi_epi64(temp9, temp11);
T2 = _mm_unpacklo_epi64(temp10, temp12);
T3 = _mm_unpackhi_epi64(temp10, temp12);
T4 = _mm_unpacklo_epi64(temp13, temp15);
T5 = _mm_unpackhi_epi64(temp13, temp15);
T6 = _mm_unpacklo_epi64(temp14, temp16);
T7 = _mm_unpackhi_epi64(temp14, temp16);
_mm_store_si128(&Transposed[t_indx], T0); //store transposed 8X8 matrix
_mm_store_si128(&Transposed[t_indx+1], T1);
_mm_store_si128(&Transposed[t_indx+2], T2);
_mm_store_si128(&Transposed[t_indx+3], T3);
_mm_store_si128(&Transposed[t_indx+4], T4);
_mm_store_si128(&Transposed[t_indx+5], T5);
_mm_store_si128(&Transposed[t_indx+6], T6);
_mm_store_si128(&Transposed[t_indx+7], T7);
}
/* Input image must be large enough to have valid pixels for the offset (dx,dy).
I.e., with (dx,dy)=(-10,8), x-value up to -10 and y-values up to (h-1)+8 will be accessed.
The integral image will be access with (x,y) in [-1,w)x[-1,h).
Note also that we use 32bit for the integral image even though the values may overflow
that range. However, the modulo-arithmetic used when computing the block sums later
will be still correct when the block size is not too large.
*/
static void buildIntegralImage_SSE(uint32_t* integral_image, int integral_stride,
const uint8_t* current_image, int current_image_stride,
const uint8_t* compare_image, int compare_image_stride,
int w,int hStart, int hEnd,
int dx,int dy)
{
const __m128i zero = _mm_set1_epi8(0);
memset(integral_image -1 -integral_stride, 0, (w+1)*sizeof(uint32_t));
for (int y=hStart;y<hEnd;y++) {
const uint8_t* p1 = current_image + y *current_image_stride;
const uint8_t* p2 = compare_image + (y+dy)*compare_image_stride + dx;
uint32_t* out = integral_image + y*integral_stride-1;
__m128i prevadd = _mm_set1_epi32(0);
const int pixels_step = 16;
*out++ = 0;
for (int x=0 ; x<w ; x+=pixels_step)
{
__m128i pa, pb;
__m128i pla, plb;
__m128i ldiff, lldiff, lhdiff;
__m128i ltmp,htmp;
__m128i ladd,hadd;
__m128i pha,phb;
__m128i hdiff,hldiff,hhdiff;
__m128i l2tmp,h2tmp;
pa = _mm_loadu_si128((__m128i*)p1);
pb = _mm_loadu_si128((__m128i*)p2);
pla = _mm_unpacklo_epi8(pa,zero);
plb = _mm_unpacklo_epi8(pb,zero);
ldiff = _mm_sub_epi16(pla,plb);
ldiff = _mm_mullo_epi16(ldiff,ldiff);
lldiff = _mm_unpacklo_epi16(ldiff,zero);
lhdiff = _mm_unpackhi_epi16(ldiff,zero);
ltmp = _mm_slli_si128(lldiff, 4);
lldiff = _mm_add_epi32(lldiff, ltmp);
ltmp = _mm_slli_si128(lldiff, 8);
lldiff = _mm_add_epi32(lldiff, ltmp);
lldiff = _mm_add_epi32(lldiff, prevadd);
ladd = _mm_shuffle_epi32(lldiff, 0xff);
htmp = _mm_slli_si128(lhdiff, 4);
lhdiff = _mm_add_epi32(lhdiff, htmp);
htmp = _mm_slli_si128(lhdiff, 8);
lhdiff = _mm_add_epi32(lhdiff, htmp);
lhdiff = _mm_add_epi32(lhdiff, ladd);
prevadd = _mm_shuffle_epi32(lhdiff, 0xff);
_mm_store_si128((__m128i*)(out), lldiff);
_mm_store_si128((__m128i*)(out+4),lhdiff);
pha = _mm_unpackhi_epi8(pa,zero);
phb = _mm_unpackhi_epi8(pb,zero);
hdiff = _mm_sub_epi16(pha,phb);
hdiff = _mm_mullo_epi16(hdiff,hdiff);
hldiff = _mm_unpacklo_epi16(hdiff,zero);
hhdiff = _mm_unpackhi_epi16(hdiff,zero);
l2tmp = _mm_slli_si128(hldiff, 4);
hldiff = _mm_add_epi32(hldiff, l2tmp);
l2tmp = _mm_slli_si128(hldiff, 8);
hldiff = _mm_add_epi32(hldiff, l2tmp);
hldiff = _mm_add_epi32(hldiff, prevadd);
hadd = _mm_shuffle_epi32(hldiff, 0xff);
h2tmp = _mm_slli_si128(hhdiff, 4);
hhdiff = _mm_add_epi32(hhdiff, h2tmp);
h2tmp = _mm_slli_si128(hhdiff, 8);
hhdiff = _mm_add_epi32(hhdiff, h2tmp);
hhdiff = _mm_add_epi32(hhdiff, hadd);
prevadd = _mm_shuffle_epi32(hhdiff, 0xff);
_mm_store_si128((__m128i*)(out+8), hldiff);
_mm_store_si128((__m128i*)(out+12),hhdiff);
out+=pixels_step;
p1 +=pixels_step;
p2 +=pixels_step;
}
if (y>0) {
out = integral_image + y*integral_stride;
for (int x=0 ; x<w ; x+=pixels_step) {
*((__m128i*)out) = _mm_add_epi32(*(__m128i*)(out-integral_stride),
*(__m128i*)(out));
*((__m128i*)(out+4)) = _mm_add_epi32(*(__m128i*)(out+4-integral_stride),
*(__m128i*)(out+4));
*((__m128i*)(out+8)) = _mm_add_epi32(*(__m128i*)(out+8-integral_stride),
*(__m128i*)(out+8));
*((__m128i*)(out+12)) = _mm_add_epi32(*(__m128i*)(out+12-integral_stride),
*(__m128i*)(out+12));
out += 4*4;
}
}
}
}
static void trans_g_aiT16(__m128i *input, __m128i *Transposed)
{
__m128i a;
__m128i b;
__m128i c;
__m128i d;
__m128i e;
__m128i f;
__m128i g;
__m128i h;
__m128i temp1;
__m128i temp2;
__m128i temp3;
__m128i temp4;
__m128i temp5;
__m128i temp6;
__m128i temp7;
__m128i temp8;
__m128i temp9;
__m128i temp10;
__m128i temp11;
__m128i temp12;
__m128i temp13;
__m128i temp14;
__m128i temp15;
__m128i temp16;
__m128i T0;
__m128i T1;
__m128i T2;
__m128i T3;
__m128i T4;
__m128i T5;
__m128i T6;
__m128i T7;
a = _mm_load_si128(&input[2]);
b = _mm_load_si128(&input[6]);
c = _mm_load_si128(&input[10]);
d = _mm_load_si128(&input[14]);
e = _mm_load_si128(&input[18]);
f = _mm_load_si128(&input[22]);
g = _mm_load_si128(&input[26]);
h = _mm_load_si128(&input[30]);
//store 128 bits of integer data into the memory address given
_mm_store_si128(&Transposed[0], a); //store transposed 8X8 matrix
_mm_store_si128(&Transposed[1], b);
_mm_store_si128(&Transposed[2], c);
_mm_store_si128(&Transposed[3], d);
_mm_store_si128(&Transposed[4], e);
_mm_store_si128(&Transposed[5], f);
_mm_store_si128(&Transposed[6], g);
_mm_store_si128(&Transposed[7], h);
//load matrix input[0][0],[2][0]...
a = _mm_load_si128(&input[0]);
b = _mm_load_si128(&input[4]);
c = _mm_load_si128(&input[8]);
d = _mm_load_si128(&input[12]);
e = _mm_load_si128(&input[16]);
f = _mm_load_si128(&input[20]);
g = _mm_load_si128(&input[24]);
h = _mm_load_si128(&input[28]);
temp1 = _mm_unpacklo_epi16(a, b);
temp2 = _mm_unpacklo_epi16(c, d);
temp3 = _mm_unpacklo_epi16(e, f);
temp4 = _mm_unpacklo_epi16(g, h);
temp5 = _mm_unpackhi_epi16(a, b);
temp6 = _mm_unpackhi_epi16(c, d);
temp7 = _mm_unpackhi_epi16(e, f);
temp8 = _mm_unpackhi_epi16(g, h);
temp9 = _mm_unpacklo_epi32(temp1, temp2);
temp10 = _mm_unpackhi_epi32(temp1, temp2);
temp11 = _mm_unpacklo_epi32(temp3, temp4);
temp12 = _mm_unpackhi_epi32(temp3, temp4);
temp13 = _mm_unpacklo_epi32(temp5, temp6);
temp14 = _mm_unpackhi_epi32(temp5, temp6);
temp15 = _mm_unpacklo_epi32(temp7, temp8);
temp16 = _mm_unpackhi_epi32(temp7, temp8);
T0 = _mm_unpacklo_epi64(temp9, temp11);
T1 = _mm_unpackhi_epi64(temp9, temp11);
T2 = _mm_unpacklo_epi64(temp10, temp12);
T3 = _mm_unpackhi_epi64(temp10, temp12);
_mm_store_si128(&Transposed[8], T0); //store transposed 8X8 matrix
_mm_store_si128(&Transposed[9], T1);
_mm_store_si128(&Transposed[10], T2);
_mm_store_si128(&Transposed[11], T3);
}
static void FTransform(const uint8_t* src, const uint8_t* ref, int16_t* out) {
const __m128i zero = _mm_setzero_si128();
const __m128i seven = _mm_set1_epi16(7);
const __m128i k937 = _mm_set1_epi32(937);
const __m128i k1812 = _mm_set1_epi32(1812);
const __m128i k51000 = _mm_set1_epi32(51000);
const __m128i k12000_plus_one = _mm_set1_epi32(12000 + (1 << 16));
const __m128i k5352_2217 = _mm_set_epi16(5352, 2217, 5352, 2217,
5352, 2217, 5352, 2217);
const __m128i k2217_5352 = _mm_set_epi16(2217, -5352, 2217, -5352,
2217, -5352, 2217, -5352);
const __m128i k88p = _mm_set_epi16(8, 8, 8, 8, 8, 8, 8, 8);
const __m128i k88m = _mm_set_epi16(-8, 8, -8, 8, -8, 8, -8, 8);
const __m128i k5352_2217p = _mm_set_epi16(2217, 5352, 2217, 5352,
2217, 5352, 2217, 5352);
const __m128i k5352_2217m = _mm_set_epi16(-5352, 2217, -5352, 2217,
-5352, 2217, -5352, 2217);
__m128i v01, v32;
// Difference between src and ref and initial transpose.
{
// Load src and convert to 16b.
const __m128i src0 = _mm_loadl_epi64((const __m128i*)&src[0 * BPS]);
const __m128i src1 = _mm_loadl_epi64((const __m128i*)&src[1 * BPS]);
const __m128i src2 = _mm_loadl_epi64((const __m128i*)&src[2 * BPS]);
const __m128i src3 = _mm_loadl_epi64((const __m128i*)&src[3 * BPS]);
const __m128i src_0 = _mm_unpacklo_epi8(src0, zero);
const __m128i src_1 = _mm_unpacklo_epi8(src1, zero);
const __m128i src_2 = _mm_unpacklo_epi8(src2, zero);
const __m128i src_3 = _mm_unpacklo_epi8(src3, zero);
// Load ref and convert to 16b.
const __m128i ref0 = _mm_loadl_epi64((const __m128i*)&ref[0 * BPS]);
const __m128i ref1 = _mm_loadl_epi64((const __m128i*)&ref[1 * BPS]);
const __m128i ref2 = _mm_loadl_epi64((const __m128i*)&ref[2 * BPS]);
const __m128i ref3 = _mm_loadl_epi64((const __m128i*)&ref[3 * BPS]);
const __m128i ref_0 = _mm_unpacklo_epi8(ref0, zero);
const __m128i ref_1 = _mm_unpacklo_epi8(ref1, zero);
const __m128i ref_2 = _mm_unpacklo_epi8(ref2, zero);
const __m128i ref_3 = _mm_unpacklo_epi8(ref3, zero);
// Compute difference. -> 00 01 02 03 00 00 00 00
const __m128i diff0 = _mm_sub_epi16(src_0, ref_0);
const __m128i diff1 = _mm_sub_epi16(src_1, ref_1);
const __m128i diff2 = _mm_sub_epi16(src_2, ref_2);
const __m128i diff3 = _mm_sub_epi16(src_3, ref_3);
// Unpack and shuffle
// 00 01 02 03 0 0 0 0
// 10 11 12 13 0 0 0 0
// 20 21 22 23 0 0 0 0
// 30 31 32 33 0 0 0 0
const __m128i shuf01 = _mm_unpacklo_epi32(diff0, diff1);
const __m128i shuf23 = _mm_unpacklo_epi32(diff2, diff3);
// 00 01 10 11 02 03 12 13
// 20 21 30 31 22 23 32 33
const __m128i shuf01_p =
_mm_shufflehi_epi16(shuf01, _MM_SHUFFLE(2, 3, 0, 1));
const __m128i shuf23_p =
_mm_shufflehi_epi16(shuf23, _MM_SHUFFLE(2, 3, 0, 1));
// 00 01 10 11 03 02 13 12
// 20 21 30 31 23 22 33 32
const __m128i s01 = _mm_unpacklo_epi64(shuf01_p, shuf23_p);
const __m128i s32 = _mm_unpackhi_epi64(shuf01_p, shuf23_p);
// 00 01 10 11 20 21 30 31
// 03 02 13 12 23 22 33 32
const __m128i a01 = _mm_add_epi16(s01, s32);
const __m128i a32 = _mm_sub_epi16(s01, s32);
// [d0 + d3 | d1 + d2 | ...] = [a0 a1 | a0' a1' | ... ]
// [d0 - d3 | d1 - d2 | ...] = [a3 a2 | a3' a2' | ... ]
const __m128i tmp0 = _mm_madd_epi16(a01, k88p); // [ (a0 + a1) << 3, ... ]
const __m128i tmp2 = _mm_madd_epi16(a01, k88m); // [ (a0 - a1) << 3, ... ]
const __m128i tmp1_1 = _mm_madd_epi16(a32, k5352_2217p);
const __m128i tmp3_1 = _mm_madd_epi16(a32, k5352_2217m);
const __m128i tmp1_2 = _mm_add_epi32(tmp1_1, k1812);
const __m128i tmp3_2 = _mm_add_epi32(tmp3_1, k937);
const __m128i tmp1 = _mm_srai_epi32(tmp1_2, 9);
const __m128i tmp3 = _mm_srai_epi32(tmp3_2, 9);
const __m128i s03 = _mm_packs_epi32(tmp0, tmp2);
const __m128i s12 = _mm_packs_epi32(tmp1, tmp3);
const __m128i s_lo = _mm_unpacklo_epi16(s03, s12); // 0 1 0 1 0 1...
const __m128i s_hi = _mm_unpackhi_epi16(s03, s12); // 2 3 2 3 2 3
const __m128i v23 = _mm_unpackhi_epi32(s_lo, s_hi);
v01 = _mm_unpacklo_epi32(s_lo, s_hi);
v32 = _mm_shuffle_epi32(v23, _MM_SHUFFLE(1, 0, 3, 2)); // 3 2 3 2 3 2..
}
// Second pass
{
// Same operations are done on the (0,3) and (1,2) pairs.
// a0 = v0 + v3
// a1 = v1 + v2
// a3 = v0 - v3
// a2 = v1 - v2
const __m128i a01 = _mm_add_epi16(v01, v32);
const __m128i a32 = _mm_sub_epi16(v01, v32);
const __m128i a11 = _mm_unpackhi_epi64(a01, a01);
const __m128i a22 = _mm_unpackhi_epi64(a32, a32);
const __m128i a01_plus_7 = _mm_add_epi16(a01, seven);
// d0 = (a0 + a1 + 7) >> 4;
// d2 = (a0 - a1 + 7) >> 4;
const __m128i c0 = _mm_add_epi16(a01_plus_7, a11);
const __m128i c2 = _mm_sub_epi16(a01_plus_7, a11);
const __m128i d0 = _mm_srai_epi16(c0, 4);
const __m128i d2 = _mm_srai_epi16(c2, 4);
// f1 = ((b3 * 5352 + b2 * 2217 + 12000) >> 16)
// f3 = ((b3 * 2217 - b2 * 5352 + 51000) >> 16)
const __m128i b23 = _mm_unpacklo_epi16(a22, a32);
const __m128i c1 = _mm_madd_epi16(b23, k5352_2217);
const __m128i c3 = _mm_madd_epi16(b23, k2217_5352);
const __m128i d1 = _mm_add_epi32(c1, k12000_plus_one);
const __m128i d3 = _mm_add_epi32(c3, k51000);
const __m128i e1 = _mm_srai_epi32(d1, 16);
const __m128i e3 = _mm_srai_epi32(d3, 16);
const __m128i f1 = _mm_packs_epi32(e1, e1);
const __m128i f3 = _mm_packs_epi32(e3, e3);
// f1 = f1 + (a3 != 0);
// The compare will return (0xffff, 0) for (==0, !=0). To turn that into the
// desired (0, 1), we add one earlier through k12000_plus_one.
// -> f1 = f1 + 1 - (a3 == 0)
const __m128i g1 = _mm_add_epi16(f1, _mm_cmpeq_epi16(a32, zero));
const __m128i d0_g1 = _mm_unpacklo_epi64(d0, g1);
const __m128i d2_f3 = _mm_unpacklo_epi64(d2, f3);
_mm_storeu_si128((__m128i*)&out[0], d0_g1);
_mm_storeu_si128((__m128i*)&out[8], d2_f3);
}
}
__m64 interpolvline128_3(__m128i* temp){
__m128i xmm6;
__m64 ret;
__m128i xmm7 = _mm_setzero_si128();
__m128i xmm0 = _mm_load_si128(temp++);
__m128i xmm1 = _mm_load_si128(temp++);
__m128i xmm2 = _mm_load_si128(temp++);
__m128i xmm3 = _mm_load_si128(temp++);
__m128i xmm4 = _mm_load_si128(temp++);
__m128i xmm5 = _mm_load_si128(temp);
xmm1 = _mm_add_epi16(xmm1,xmm4);
xmm0 = _mm_add_epi16(xmm0,xmm5);
xmm6 = _mm_set_epi32(0xFFFBFFFB,0xFFFBFFFB,0xFFFBFFFB,0xFFFBFFFB);
xmm4 = _mm_mullo_epi16(xmm1, xmm6);
xmm5 = _mm_mulhi_epi16(xmm1, xmm6);
xmm1 = _mm_unpacklo_epi16(xmm4, xmm5);
xmm6 = _mm_unpackhi_epi16(xmm4, xmm5);
xmm7 = _mm_set_epi32(0x00140014,0x00140014,0x00140014,0x00140014);
xmm5 = _mm_add_epi16(xmm2,xmm3);
xmm4 = _mm_mullo_epi16(xmm5, xmm7);
xmm5 = _mm_mulhi_epi16(xmm5, xmm7);
xmm7 = _mm_unpacklo_epi16(xmm4, xmm5);
xmm4 = _mm_unpackhi_epi16(xmm4, xmm5);
xmm7 = _mm_add_epi32(xmm7,xmm1);
xmm4 = _mm_add_epi32(xmm4,xmm6);
xmm6 = _mm_set_epi32(0x00010001,0x00010001,0x00010001,0x00010001);
xmm6 = _mm_mulhi_epi16(xmm0, xmm6);
xmm1 = _mm_unpacklo_epi16(xmm0, xmm6);
xmm6 = _mm_unpackhi_epi16(xmm0, xmm6);
xmm7 = _mm_add_epi32(xmm7,xmm1);
xmm4 = _mm_add_epi32(xmm4,xmm6);
xmm1 = _mm_set_epi32(0x00000200,0x00000200,0x00000200,0x00000200);
xmm7 = _mm_add_epi32(xmm7,xmm1);
xmm4 = _mm_add_epi32(xmm4,xmm1);
xmm5 = _mm_setzero_si128();
xmm7 = _mm_srli_epi32(xmm7, 10);
xmm7 = _mm_max_epi16(xmm7, xmm5); // preventing negative values
xmm7 = _mm_slli_epi32(xmm7,16);
xmm7 = _mm_srli_epi32(xmm7,16);
xmm4 = _mm_srli_epi32(xmm4, 10);
xmm4 = _mm_max_epi16(xmm4, xmm5); // preventing negative values
xmm4 = _mm_slli_epi32(xmm4,16);
xmm4 = _mm_srli_epi32(xmm4,16);
xmm6 = _mm_packs_epi32(xmm7, xmm4);
xmm1 = _mm_set_epi32(0x00100010,0x00100010,0x00100010,0x00100010);
xmm2 = _mm_add_epi16(xmm2,xmm1);
xmm2 = _mm_max_epi16(xmm2, xmm5); // preventing negative values
xmm2 = _mm_srli_epi16(xmm2,5);
xmm3 = _mm_add_epi16(xmm3,xmm1);
xmm3 = _mm_max_epi16(xmm3, xmm5); // preventing negative values
xmm3 = _mm_srli_epi16(xmm3,5);
xmm2 = _mm_packus_epi16(xmm2,xmm5);
xmm3 = _mm_packus_epi16(xmm3,xmm5);
xmm6 = _mm_packus_epi16(xmm6,xmm5);
xmm7 = _mm_unpacklo_epi8(xmm2,xmm6);
xmm4 = _mm_unpacklo_epi8(xmm6,xmm3);
xmm6 = _mm_avg_epu8(xmm4,xmm7);
xmm6 = _mm_srli_epi16(xmm6,8);
xmm6 = _mm_packus_epi16(xmm6,xmm5);
ret = _mm_movepi64_pi64(xmm6);
_mm_empty();
return(ret);
}
