int32_t dot_product(int16_t *x,
int16_t *y,
uint32_t N, //must be a multiple of 8
uint8_t output_shift)
{
uint32_t n;
#if defined(__x86_64__) || defined(__i386__)
__m128i *x128,*y128,mmtmp1,mmtmp2,mmtmp3,mmcumul,mmcumul_re,mmcumul_im;
__m64 mmtmp7;
__m128i minus_i = _mm_set_epi16(-1,1,-1,1,-1,1,-1,1);
int32_t result;
x128 = (__m128i*) x;
y128 = (__m128i*) y;
mmcumul_re = _mm_setzero_si128();
mmcumul_im = _mm_setzero_si128();
for (n=0; n<(N>>2); n++) {
//printf("n=%d, x128=%p, y128=%p\n",n,x128,y128);
// print_shorts("x",&x128[0]);
// print_shorts("y",&y128[0]);
// this computes Re(z) = Re(x)*Re(y) + Im(x)*Im(y)
mmtmp1 = _mm_madd_epi16(x128[0],y128[0]);
// print_ints("re",&mmtmp1);
// mmtmp1 contains real part of 4 consecutive outputs (32-bit)
// shift and accumulate results
mmtmp1 = _mm_srai_epi32(mmtmp1,output_shift);
mmcumul_re = _mm_add_epi32(mmcumul_re,mmtmp1);
// print_ints("re",&mmcumul_re);
// this computes Im(z) = Re(x)*Im(y) - Re(y)*Im(x)
mmtmp2 = _mm_shufflelo_epi16(y128[0],_MM_SHUFFLE(2,3,0,1));
// print_shorts("y",&mmtmp2);
mmtmp2 = _mm_shufflehi_epi16(mmtmp2,_MM_SHUFFLE(2,3,0,1));
// print_shorts("y",&mmtmp2);
mmtmp2 = _mm_sign_epi16(mmtmp2,minus_i);
// print_shorts("y",&mmtmp2);
mmtmp3 = _mm_madd_epi16(x128[0],mmtmp2);
// print_ints("im",&mmtmp3);
// mmtmp3 contains imag part of 4 consecutive outputs (32-bit)
// shift and accumulate results
mmtmp3 = _mm_srai_epi32(mmtmp3,output_shift);
mmcumul_im = _mm_add_epi32(mmcumul_im,mmtmp3);
// print_ints("im",&mmcumul_im);
x128++;
y128++;
}
// this gives Re Re Im Im
mmcumul = _mm_hadd_epi32(mmcumul_re,mmcumul_im);
// print_ints("cumul1",&mmcumul);
// this gives Re Im Re Im
mmcumul = _mm_hadd_epi32(mmcumul,mmcumul);
// print_ints("cumul2",&mmcumul);
//mmcumul = _mm_srai_epi32(mmcumul,output_shift);
// extract the lower half
mmtmp7 = _mm_movepi64_pi64(mmcumul);
// print_ints("mmtmp7",&mmtmp7);
// pack the result
mmtmp7 = _mm_packs_pi32(mmtmp7,mmtmp7);
// print_shorts("mmtmp7",&mmtmp7);
// convert back to integer
result = _mm_cvtsi64_si32(mmtmp7);
_mm_empty();
_m_empty();
return(result);
#elif defined(__arm__)
int16x4_t *x_128=(int16x4_t*)x;
int16x4_t *y_128=(int16x4_t*)y;
int32x4_t tmp_re,tmp_im;
int32x4_t tmp_re1,tmp_im1;
int32x4_t re_cumul,im_cumul;
int32x2_t re_cumul2,im_cumul2;
int32x4_t shift = vdupq_n_s32(-output_shift);
int32x2x2_t result2;
int16_t conjug[4]__attribute__((aligned(16))) = {-1,1,-1,1} ;
re_cumul = vdupq_n_s32(0);
im_cumul = vdupq_n_s32(0);
for (n=0; n<(N>>2); n++) {
tmp_re = vmull_s16(*x_128++, *y_128++);
//tmp_re = [Re(x[0])Re(y[0]) Im(x[0])Im(y[0]) Re(x[1])Re(y[1]) Im(x[1])Im(y[1])]
tmp_re1 = vmull_s16(*x_128++, *y_128++);
//tmp_re1 = [Re(x1[1])Re(x2[1]) Im(x1[1])Im(x2[1]) Re(x1[1])Re(x2[2]) Im(x1[1])Im(x2[2])]
tmp_re = vcombine_s32(vpadd_s32(vget_low_s32(tmp_re),vget_high_s32(tmp_re)),
vpadd_s32(vget_low_s32(tmp_re1),vget_high_s32(tmp_re1)));
//tmp_re = [Re(ch[0])Re(rx[0])+Im(ch[0])Im(ch[0]) Re(ch[1])Re(rx[1])+Im(ch[1])Im(ch[1]) Re(ch[2])Re(rx[2])+Im(ch[2]) Im(ch[2]) Re(ch[3])Re(rx[3])+Im(ch[3])Im(ch[3])]
tmp_im = vmull_s16(vrev32_s16(vmul_s16(*x_128++,*(int16x4_t*)conjug)),*y_128++);
//tmp_im = [-Im(ch[0])Re(rx[0]) Re(ch[0])Im(rx[0]) -Im(ch[1])Re(rx[1]) Re(ch[1])Im(rx[1])]
tmp_im1 = vmull_s16(vrev32_s16(vmul_s16(*x_128++,*(int16x4_t*)conjug)),*y_128++);
//tmp_im1 = [-Im(ch[2])Re(rx[2]) Re(ch[2])Im(rx[2]) -Im(ch[3])Re(rx[3]) Re(ch[3])Im(rx[3])]
tmp_im = vcombine_s32(vpadd_s32(vget_low_s32(tmp_im),vget_high_s32(tmp_im)),
vpadd_s32(vget_low_s32(tmp_im1),vget_high_s32(tmp_im1)));
//tmp_im = [-Im(ch[0])Re(rx[0])+Re(ch[0])Im(rx[0]) -Im(ch[1])Re(rx[1])+Re(ch[1])Im(rx[1]) -Im(ch[2])Re(rx[2])+Re(ch[2])Im(rx[2]) -Im(ch[3])Re(rx[3])+Re(ch[3])Im(rx[3])]
re_cumul = vqaddq_s32(re_cumul,vqshlq_s32(tmp_re,shift));
im_cumul = vqaddq_s32(im_cumul,vqshlq_s32(tmp_im,shift));
}
re_cumul2 = vpadd_s32(vget_low_s32(re_cumul),vget_high_s32(re_cumul));
im_cumul2 = vpadd_s32(vget_low_s32(im_cumul),vget_high_s32(im_cumul));
re_cumul2 = vpadd_s32(re_cumul2,re_cumul2);
im_cumul2 = vpadd_s32(im_cumul2,im_cumul2);
result2 = vzip_s32(re_cumul2,im_cumul2);
return(vget_lane_s32(result2.val[0],0));
#endif
}
void weighted_merge_planar_mmx(BYTE *p1, const BYTE *p2, int p1_pitch, int p2_pitch, int width, int height, int weight, int invweight) {
__m64 round_mask = _mm_set1_pi32(0x4000);
__m64 zero = _mm_setzero_si64();
__m64 mask = _mm_set_pi16(weight, invweight, weight, invweight);
int wMod8 = (width/8) * 8;
for (int y = 0; y < height; y++) {
for (int x = 0; x < wMod8; x += 8) {
__m64 px1 = *(reinterpret_cast<const __m64*>(p1+x)); //y7y6 y5y4 y3y2 y1y0
__m64 px2 = *(reinterpret_cast<const __m64*>(p2+x)); //Y7Y6 Y5Y4 Y3Y2 Y1Y0
__m64 p0123 = _mm_unpacklo_pi8(px1, px2); //Y3y3 Y2y2 Y1y1 Y0y0
__m64 p4567 = _mm_unpackhi_pi8(px1, px2); //Y7y7 Y6y6 Y5y5 Y4y4
__m64 p01 = _mm_unpacklo_pi8(p0123, zero); //00Y1 00y1 00Y0 00y0
__m64 p23 = _mm_unpackhi_pi8(p0123, zero); //00Y3 00y3 00Y2 00y2
__m64 p45 = _mm_unpacklo_pi8(p4567, zero); //00Y5 00y5 00Y4 00y4
__m64 p67 = _mm_unpackhi_pi8(p4567, zero); //00Y7 00y7 00Y6 00y6
p01 = _mm_madd_pi16(p01, mask);
p23 = _mm_madd_pi16(p23, mask);
p45 = _mm_madd_pi16(p45, mask);
p67 = _mm_madd_pi16(p67, mask);
p01 = _mm_add_pi32(p01, round_mask);
p23 = _mm_add_pi32(p23, round_mask);
p45 = _mm_add_pi32(p45, round_mask);
p67 = _mm_add_pi32(p67, round_mask);
p01 = _mm_srli_pi32(p01, 15);
p23 = _mm_srli_pi32(p23, 15);
p45 = _mm_srli_pi32(p45, 15);
p67 = _mm_srli_pi32(p67, 15);
p0123 = _mm_packs_pi32(p01, p23);
p4567 = _mm_packs_pi32(p45, p67);
__m64 result = _mm_packs_pu16(p0123, p4567);
*reinterpret_cast<__m64*>(p1+x) = result;
}
for (int x = wMod8; x < width; x++) {
p1[x] = (p1[x]*invweight + p2[x]*weight + 16384) >> 15;
}
p1 += p1_pitch;
p2 += p2_pitch;
}
_mm_empty();
}
static void weighted_merge_luma_yuy2_mmx(BYTE *src, const BYTE *luma, int pitch, int luma_pitch,int width, int height, int weight, int invweight)
{
__m64 round_mask = _mm_set1_pi32(0x4000);
__m64 mask = _mm_set_pi16(weight, invweight, weight, invweight);
__m64 luma_mask = _mm_set1_pi16(0x00FF);
#pragma warning(push)
#pragma warning(disable: 4309)
__m64 chroma_mask = _mm_set1_pi16(0xFF00);
#pragma warning(pop)
int wMod8 = (width/8) * 8;
for (int y = 0; y < height; y++) {
for (int x = 0; x < wMod8; x += 8) {
__m64 px1 = *reinterpret_cast<const __m64*>(src+x); //V1 Y3 U1 Y2 V0 Y1 U0 Y0
__m64 px2 = *reinterpret_cast<const __m64*>(luma+x); //v1 y3 u1 y2 v0 y1 u0 y0
__m64 src_lo = _mm_unpacklo_pi16(px1, px2); //v0 y1 V0 Y1 u0 y0 U0 Y0
__m64 src_hi = _mm_unpackhi_pi16(px1, px2);
src_lo = _mm_and_si64(src_lo, luma_mask); //00 v0 00 V0 00 u0 00 U0
src_hi = _mm_and_si64(src_hi, luma_mask);
src_lo = _mm_madd_pi16(src_lo, mask);
src_hi = _mm_madd_pi16(src_hi, mask);
src_lo = _mm_add_pi32(src_lo, round_mask);
src_hi = _mm_add_pi32(src_hi, round_mask);
src_lo = _mm_srli_pi32(src_lo, 15);
src_hi = _mm_srli_pi32(src_hi, 15);
__m64 result_luma = _mm_packs_pi32(src_lo, src_hi);
__m64 result_chroma = _mm_and_si64(px1, chroma_mask);
__m64 result = _mm_or_si64(result_chroma, result_luma);
*reinterpret_cast<__m64*>(src+x) = result;
}
for (int x = wMod8; x < width; x+=2) {
src[x] = (luma[x] * weight + src[x] * invweight + 16384) >> 15;
}
src += pitch;
luma += luma_pitch;
}
_mm_empty();
}
__m64 test76(__m64 a, __m64 b) {
// CHECK: packssdw
return _mm_packs_pi32(a, b);
}
////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
// Convert YUY2 to RGB24.
VOID Yuy2ToRgb24_mmx(PBYTE pbDstX, INT iDstXStride, PBYTE pbSrcX, INT iSrcXStride, UINT uWidth, INT iHeight)
{
UINT x;
INT y;
INT iDstXDif;
INT iSrcXDif;
INT yy, bu, guv, rv;
M64 y0, y1, u0, v0, uv_temp1, uv_temp2, mz;
M64 r0, g0, b0, r1, g1, b1;
M64 rgb0, rgb1, rgb2, rgb3;
M64 bu0, gu0, gv0, rv0, bu1, rv1, guv0, guv1;
if (iHeight < 0)
{
iHeight = -iHeight;
pbSrcX += (iHeight - 1) * iSrcXStride;
iSrcXStride = -iSrcXStride;
}
iDstXDif = iDstXStride - (uWidth * 3);
iSrcXDif = iSrcXStride - (uWidth * 2);
mz = _mm_setzero_si64();
for (y = iHeight; y; y--)
{
for (x = uWidth / 8; x; x--)
{
y0 = ((PM64) pbSrcX)[0];
y1 = ((PM64) pbSrcX)[1];
u0 = y0;
v0 = y1;
y0 = _mm_and_si64(y0, g_mWord00FF);
y1 = _mm_and_si64(y1, g_mWord00FF);
u0 = _mm_srli_pi16(u0, 8);
v0 = _mm_srli_pi16(v0, 8);
uv_temp1 = _mm_srli_pi32(u0, 16);
u0 = _mm_slli_pi32(u0, 16);
u0 = _mm_srli_pi32(u0, 16);
uv_temp2 = _mm_srli_pi32(v0, 16);
v0 = _mm_slli_pi32(v0, 16);
v0 = _mm_srli_pi32(v0, 16);
u0 = _mm_packs_pi32(u0, v0);
v0 = _mm_packs_pi32(uv_temp1, uv_temp2);
// Calculate coefficient.
u0 = _mm_subs_pi16(u0, g_mSub80);
v0 = _mm_subs_pi16(v0, g_mSub80);
gu0 = _mm_mullo_pi16(u0, g_mUGMul);
gv0 = _mm_mullo_pi16(v0, g_mVGMul);
bu0 = _mm_mullo_pi16(u0, g_mUBMul);
rv0 = _mm_mullo_pi16(v0, g_mVRMul);
guv0 = _mm_adds_pi16(gu0, gv0);
guv1 = _mm_unpackhi_pi16(guv0, guv0); // guv3 guv3 guv2 guv2
guv0 = _mm_unpacklo_pi16(guv0, guv0); // guv1 guv1 guv0 guv0
bu1 = _mm_unpackhi_pi16(bu0, bu0); // bu3 bu3 bu2 bu2
bu0 = _mm_unpacklo_pi16(bu0, bu0); // bu1 bu1 bu0 bu0
rv1 = _mm_unpackhi_pi16(rv0, rv0); // rv3 rv3 rv2 rv2
rv0 = _mm_unpacklo_pi16(rv0, rv0); // rv1 rv1 rv0 rv0
// Process for row 0.
y1 = _mm_subs_pi16(y1, g_mSub10);
y0 = _mm_subs_pi16(y0, g_mSub10);
y1 = _mm_mullo_pi16(y1, g_mYYMul);
y0 = _mm_mullo_pi16(y0, g_mYYMul);
g1 = _mm_subs_pi16(y1, guv1); // g7 g6 g5 g4
g0 = _mm_subs_pi16(y0, guv0); // g3 g2 g1 g0
g1 = _mm_srai_pi16(g1, SCALEBITS);
g0 = _mm_srai_pi16(g0, SCALEBITS);
g0 = _mm_packs_pu16(g0, g1); // g7 g6 ...g1 g0
b1 = _mm_adds_pi16(y1, bu1);
b0 = _mm_adds_pi16(y0, bu0);
b1 = _mm_srai_pi16(b1, SCALEBITS);
b0 = _mm_srai_pi16(b0, SCALEBITS);
b0 = _mm_packs_pu16(b0, b1);
r1 = _mm_adds_pi16(y1, rv1);
r0 = _mm_adds_pi16(y0, rv0);
r1 = _mm_srai_pi16(r1, SCALEBITS);
r0 = _mm_srai_pi16(r0, SCALEBITS);
r0 = _mm_packs_pu16(r0, r1);
r1 = _mm_unpackhi_pi8(b0, r0); // r7 b7 r6 b6 r5 b5 r4 b4
r0 = _mm_unpacklo_pi8(b0, r0); // r3 b3 r2 b2 r1 b1 r0 b0
g1 = _mm_unpackhi_pi8(g0, mz); // 0 g7 0 g6 0 g5 0 g4
g0 = _mm_unpacklo_pi8(g0, mz); // 0 g3 0 g2 0 g1 0 g0
rgb0 = _mm_unpacklo_pi8(r0, g0); // 0 r1 g1 b1 0 r0 g0 b0
rgb1 = _mm_unpackhi_pi8(r0, g0); // 0 r3 g3 b3 0 r2 g2 b2
rgb2 = _mm_unpacklo_pi8(r1, g1); // 0 r5 g5 b5 0 r4 g4 b4
rgb3 = _mm_unpackhi_pi8(r1, g1); // 0 r7 g7 b7 0 r6 g6 b6
// Write out row 0.
*((PDWORD) (pbDstX + 0)) = _mm_cvtsi64_si32(rgb0); rgb0 = _mm_srli_si64(rgb0, 32);
*((PDWORD) (pbDstX + 3)) = _mm_cvtsi64_si32(rgb0);
*((PDWORD) (pbDstX + 6)) = _mm_cvtsi64_si32(rgb1); rgb1 = _mm_srli_si64(rgb1, 32);
*((PDWORD) (pbDstX + 9)) = _mm_cvtsi64_si32(rgb1);
*((PDWORD) (pbDstX + 12)) = _mm_cvtsi64_si32(rgb2); rgb2 = _mm_srli_si64(rgb2, 32);
*((PDWORD) (pbDstX + 15)) = _mm_cvtsi64_si32(rgb2);
*((PDWORD) (pbDstX + 18)) = _mm_cvtsi64_si32(rgb3); rgb3 = _mm_srli_si64(rgb3, 32);
*((PDWORD) (pbDstX + 21)) = _mm_cvtsi64_si32(rgb3);
pbDstX += 24;
pbSrcX += 16;
}
for (x = (uWidth & 7) / 2; x; x--)
{
bu = g_iBUTab[pbSrcX[1]];
guv = g_iGUTab[pbSrcX[1]] + g_iGVTab[pbSrcX[3]];
rv = g_iRVTab[pbSrcX[3]];
yy = g_iYYTab[pbSrcX[0]];
pbDstX[0] = _Clip((yy + bu) >> SCALEBITS_OUT);
pbDstX[1] = _Clip((yy - guv) >> SCALEBITS_OUT);
pbDstX[2] = _Clip((yy + rv) >> SCALEBITS_OUT);
yy = g_iYYTab[pbSrcX[2]];
pbDstX[3] = _Clip((yy + bu) >> SCALEBITS_OUT);
pbDstX[4] = _Clip((yy - guv) >> SCALEBITS_OUT);
pbDstX[5] = _Clip((yy + rv) >> SCALEBITS_OUT);
pbDstX += 6;
pbSrcX += 4;
}
pbDstX += iDstXDif;
pbSrcX += iSrcXDif;
}
_mm_empty();
}
__m64 test_mm_packs_pi32(__m64 a, __m64 b) {
// CHECK-LABEL: test_mm_packs_pi32
// CHECK: call x86_mmx @llvm.x86.mmx.packssdw
return _mm_packs_pi32(a, b);
}
