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 r_dimpatchD_MMX(const DCanvas *const cvs, argb_t color, int alpha, int x1, int y1, int w, int h)
{
int x, y, i;
argb_t *line;
int invAlpha = 256 - alpha;
int dpitch = cvs->pitch / sizeof(DWORD);
line = (argb_t *)cvs->buffer + y1 * dpitch;
int batches = w / 2;
int remainder = w & 1;
// MMX temporaries:
const __m64 upper8mask = _mm_set_pi16(0, 0xff, 0xff, 0xff);
const __m64 blendAlpha = _mm_set_pi16(0, alpha, alpha, alpha);
const __m64 blendInvAlpha = _mm_set_pi16(0, invAlpha, invAlpha, invAlpha);
const __m64 blendColor = _mm_set_pi16(0, RPART(color), GPART(color), BPART(color));
const __m64 blendMult = _mm_mullo_pi16(blendColor, blendAlpha);
for (y = y1; y < y1 + h; y++)
{
// MMX optimize the bulk in batches of 2 colors:
for (i = 0, x = x1; i < batches; ++i, x += 2)
{
#if 1
const __m64 input = _mm_setr_pi32(line[x + 0], line[x + 1]);
#else
// NOTE(jsd): No guarantee of 64-bit alignment; cannot use.
const __m64 input = *((__m64 *)line[x]);
#endif
const __m64 output = blend2vs1_mmx(input, blendMult, blendInvAlpha, upper8mask);
#if 1
line[x+0] = _mm_cvtsi64_si32(_mm_srli_si64(output, 32*0));
line[x+1] = _mm_cvtsi64_si32(_mm_srli_si64(output, 32*1));
#else
// NOTE(jsd): No guarantee of 64-bit alignment; cannot use.
*((__m64 *)line[x]) = output;
#endif
}
if (remainder)
{
// Pick up the remainder:
for (; x < x1 + w; x++)
{
line[x] = alphablend1a(line[x], color, alpha);
}
}
line += dpitch;
}
// Required to reset FP:
_mm_empty();
}
void weak_horizontal_chroma_MMX(unsigned char pix[], const int xstride, const unsigned char alpha , const unsigned char beta, const unsigned char tc0)
{
__m64 mp1 = _mm_set_pi16( 0,0,pix[-2*xstride + 1], pix[-2*xstride]);
__m64 mp0 = _mm_set_pi16( 0,0,pix[-1*xstride + 1], pix[-1*xstride]);
__m64 mq0 = _mm_set_pi16( 0,0,pix[1], pix[0]);
__m64 mq1 = _mm_set_pi16( 0,0,pix[xstride + 1], pix[xstride]);
__m64 mdiff_p0_q0 = _mm_sub_pi16(mq0,mp0); //abs(q0 - p0)
__m64 mdiff_p1_p0 = _mm_sub_pi16(mp0,mp1); //abs(p1 - p0)
__m64 mdiff_q1_q0 = _mm_sub_pi16(mq0, mq1); //abs(q1 - q0)
//To calculate the mask
__m64 malpha = _mm_set_pi16(0,0,alpha,alpha);
__m64 malphab = _mm_set_pi16(0,0,-alpha,-alpha);
__m64 mbeta = _mm_set_pi16(0,0,beta,beta);
__m64 mbetab = _mm_set_pi16(0,0,-beta,-beta);
__m64 mask0 = _mm_and_si64( _mm_cmpgt_pi16(malpha, mdiff_p0_q0), _mm_cmpgt_pi16(mdiff_p0_q0,malphab));
__m64 mask1 = _mm_and_si64( _mm_cmpgt_pi16(mbeta, mdiff_p1_p0), _mm_cmpgt_pi16(mdiff_p1_p0,mbetab));
__m64 mask2 = _mm_and_si64( _mm_cmpgt_pi16(mbeta, mdiff_q1_q0), _mm_cmpgt_pi16(mdiff_q1_q0,mbetab));
__m64 first_mask = _mm_and_si64 (_mm_and_si64 (mask0,mask1),mask2);
__m64 mdiff_q0_p0 = _mm_sub_pi16(mq0,mp0); //(q0 - p0)
__m64 mlshift = _mm_set_pi16(0,0,0,2);
__m64 minter_1 = _mm_sll_pi16(mdiff_q0_p0, mlshift);//inter_1 = (q0 - p0 ) << 2;
__m64 minter_2 = _mm_sub_pi16(mp1, mq1);//(p1 - q1)
__m64 madd4 = _mm_set_pi16(4,4,4,4);
__m64 minter_3 = _mm_add_pi16(minter_2, madd4);//inter_2 = (p1 - q1) + 4;
__m64 minter_4 = _mm_add_pi16(minter_3,minter_1); //(inter_1 + inter_2)
__m64 mrshift3 = _mm_set_pi16(0,0,0,3);
__m64 minter5 = _mm_sra_pi16(minter_4, mrshift3);
//Clip3
__m64 m_tc0 = _mm_set_pi16(0,0,tc0,tc0);
__m64 m_tcb0 = _mm_set_pi16(0,0,-tc0,-tc0);
__m64 mres_c3 = _mm_min_pi16(_mm_max_pi16(minter5,m_tcb0),m_tc0); //CLIP3(-tc0, tc0, addp2 - p1 );
__m64 merror2 = _mm_and_si64 (mres_c3,first_mask);
__m64 result_p0 = _mm_add_pi16(merror2,mp0); //_mm_shuffle_pi16(_mm_add_pi16(merror2,mq1), 0x1B);
__m64 result_q0 = _mm_sub_pi16(mq0, merror2);//_mm_shuffle_pi16(_mm_sub_pi16(mq1, merror2), 0x1B);
__m64 mrshift = _mm_set_pi16(0,0,0,1);
*((unsigned short* )(&pix[-xstride])) = _mm_cvtsi64_si32(_mm_packs_pu16(result_p0,mrshift));
*((unsigned short* )(&pix[0])) = _mm_cvtsi64_si32(_mm_packs_pu16(result_q0,mrshift));
empty();
}
unsigned int mmx_hash_bucket_data(unsigned char *key, int size, int NoOfItems)
{
char *p, *end;
__m64 v1, v2, s;
int val;
if (size < 8) return(fnv_data2bucket(key, size, NoOfItems));
p=key;
end=key+size;
_mm_empty(); // emms
v1=_mm_set1_pi32(FNV_INIT_VAL);
while ((end-p) > 7)
{
v2=_mm_setr_pi32(*p,*(p+4));
v1=_mm_add_pi16(v1, v2);
v1=_mm_slli_pi32(v1, 3);
p+=8;
}
val=_mm_cvtsi64_si32(v1);
_mm_empty(); // emms
if (val < 0) val=1-val;
val =val % NoOfItems;
return(val);
}
unsigned int interpolvline_mmx_3(unsigned char* image, int PicWidthInPix){
__m64 mm_A = _mm_set_pi16(image[-2 * PicWidthInPix + 3],image[-2 * PicWidthInPix + 2],image[-2 * PicWidthInPix + 1],image[-2 * PicWidthInPix]);
__m64 mm_B = _mm_set_pi16(image[-1 * PicWidthInPix + 3],image[-1 * PicWidthInPix + 2],image[-1 * PicWidthInPix + 1],image[-1 * PicWidthInPix]);
__m64 mm_C = _mm_set_pi16(image[3],image[2],image[1],image[0]);
__m64 mm_D = _mm_set_pi16(image[1 * PicWidthInPix + 3],image[1 * PicWidthInPix + 2],image[1 * PicWidthInPix + 1],image[1 * PicWidthInPix]);
__m64 mm_E = _mm_set_pi16(image[2 * PicWidthInPix + 3],image[2 * PicWidthInPix + 2],image[2 * PicWidthInPix + 1],image[2 * PicWidthInPix]);
__m64 mm_F = _mm_set_pi16(image[3 * PicWidthInPix + 3],image[3 * PicWidthInPix + 2],image[3 * PicWidthInPix + 1],image[3 * PicWidthInPix]);
__m64 mm_AF = _mm_add_pi16(mm_A,mm_F);//A + F
__m64 mm_inter0 = _mm_add_pi16(mm_AF,_mm_set_pi16(16,16,16,16));//A + F + 16
__m64 mm_BE = _mm_add_pi16(mm_B,mm_E);//B + E
__m64 mm_CD = _mm_add_pi16(mm_C,mm_D);//C + D
__m64 mm_CDS = _mm_slli_pi16(mm_CD,2);//(C + D) << 2
__m64 mm_inter1 = _mm_sub_pi16(mm_CDS,mm_BE);//((C + D) << 2)-(B + E)
__m64 mm_5 = _mm_set_pi16(5,5,5,5);
__m64 mm_inter_3 = _mm_mullo_pi16(mm_inter1, mm_5);//(((C + D) << 2)-(B + E))*5
__m64 mm_result = _mm_add_pi16(mm_inter_3,mm_inter0);//A + F + 16 + (((C + D) << 2)-(B + E))*5
__m64 mm_zero = _mm_setzero_si64();
__m64 mm_clip = _mm_max_pi16(mm_result,mm_zero);//Clip with 0
__m64 mm_ret = _mm_srai_pi16(mm_clip,5);
__m64 mm_clip1 = _mm_min_pi16(mm_ret,_mm_set_pi16(255,255,255,255)); //Clip with 255
__m64 test = _mm_avg_pu8(mm_clip1,mm_D);//(ptr_img[0] + ptr_rf[0] + 1) >> 1
__m64 test1 =_mm_packs_pu16(test,mm_zero);
unsigned int ret = _mm_cvtsi64_si32(test1);
empty();
return ret;
}
unsigned int interpolhline_mmx_2(unsigned char* image){
__m64 mm_A = _mm_set_pi16(image[1],image[0],image[-1],image[-2]);
__m64 mm_B = _mm_set_pi16(image[2],image[1],image[0],image[-1]);
__m64 mm_C = _mm_set_pi16(image[3],image[2],image[1],image[0]);
__m64 mm_D = _mm_set_pi16(image[4],image[3],image[2],image[1]);
__m64 mm_E = _mm_set_pi16(image[5],image[4],image[3],image[2]);
__m64 mm_F = _mm_set_pi16(image[6],image[5],image[4],image[3]);
__m64 mm_AF = _mm_add_pi16(mm_A,mm_F);//A + F
__m64 mm_inter0 = _mm_add_pi16(mm_AF,_mm_set_pi16(16,16,16,16));//A + F + 16
__m64 mm_BE = _mm_add_pi16(mm_B,mm_E);//B + E
__m64 mm_CD = _mm_add_pi16(mm_C,mm_D);//C + D
__m64 mm_CDS = _mm_slli_pi16(mm_CD,2);//(C + D) << 2
__m64 mm_inter1 = _mm_sub_pi16(mm_CDS,mm_BE);//((C + D) << 2)-(B + E)
__m64 mm_5 = _mm_set_pi16(5,5,5,5);
__m64 mm_inter_3 = _mm_mullo_pi16(mm_inter1, mm_5);//(((C + D) << 2)-(B + E))*5
__m64 mm_result = _mm_add_pi16(mm_inter_3,mm_inter0);//A + F + 16 + (((C + D) << 2)-(B + E))*5
__m64 mm_zero = _mm_setzero_si64();
__m64 mm_clip = _mm_max_pi16(mm_result,mm_zero);//Clip with 0
__m64 mm_ret = _mm_srai_pi16(mm_clip,5);
__m64 mm_clip1 = _mm_min_pi16(mm_ret,_mm_set_pi16(255,255,255,255)); //Clip with 255
__m64 result =_mm_packs_pu16(mm_clip1,mm_zero);
unsigned int ret = _mm_cvtsi64_si32(result);
empty();
return ret;
}
unsigned int interpolvline64_2_mmx(__m64* temp){
__m64 res;
__m64 ptr = _mm_setzero_si64();
short A = _mm_extract_pi16(temp[0],0);
short B = _mm_extract_pi16(temp[1],0);
short C = _mm_extract_pi16(temp[2],0);
short D = _mm_extract_pi16(temp[3],0);
short E = _mm_extract_pi16(temp[4],0);
short F = _mm_extract_pi16(temp[5],0);
unsigned int result = A + F - 5 * (short)(B + E) + 20 * (short)(C + D) + 512;
ptr = _mm_insert_pi16(ptr,CLIP255_16(result >> 10),0);
A = _mm_extract_pi16(temp[0],1);
B = _mm_extract_pi16(temp[1],1);
C = _mm_extract_pi16(temp[2],1);
D = _mm_extract_pi16(temp[3],1);
E = _mm_extract_pi16(temp[4],1);
F = _mm_extract_pi16(temp[5],1);
result = A + F - 5 * (short)(B + E) + 20 * (short)(C + D) + 512;
ptr = _mm_insert_pi16(ptr,CLIP255_16(result >> 10),1);
A = _mm_extract_pi16(temp[0],2);
B = _mm_extract_pi16(temp[1],2);
C = _mm_extract_pi16(temp[2],2);
D = _mm_extract_pi16(temp[3],2);
E = _mm_extract_pi16(temp[4],2);
F = _mm_extract_pi16(temp[5],2);
result = A + F - 5 * (short)(B + E) + 20 * (short)(C + D) + 512;
ptr = _mm_insert_pi16(ptr,CLIP255_16(result >> 10),2);
A = _mm_extract_pi16(temp[0],3);
B = _mm_extract_pi16(temp[1],3);
C = _mm_extract_pi16(temp[2],3);
D = _mm_extract_pi16(temp[3],3);
E = _mm_extract_pi16(temp[4],3);
F = _mm_extract_pi16(temp[5],3);
result = A + F - 5 * (short)(B + E) + 20 * (short)(C + D) + 512;
ptr = _mm_insert_pi16(ptr,CLIP255_16(result >> 10),3);
res = _mm_set_pi8(0,0,0,0,_mm_extract_pi16(ptr,3),_mm_extract_pi16(ptr,2),_mm_extract_pi16(ptr,1),_mm_extract_pi16(ptr,0));
result = _mm_cvtsi64_si32(res);
empty();
return result;
}
int32_t od_mc_compute_satd8_4x4_sse2(const unsigned char *src, int systride,
const unsigned char *ref, int rystride) {
int32_t satd;
__m64 sums;
__m64 a;
__m64 b;
__m64 c;
__m64 d;
a = od_load_convert_subtract_x4(src + 0*systride, ref + 0*rystride);
b = od_load_convert_subtract_x4(src + 1*systride, ref + 1*rystride);
c = od_load_convert_subtract_x4(src + 2*systride, ref + 2*rystride);
d = od_load_convert_subtract_x4(src + 3*systride, ref + 3*rystride);
/*Vertical 1D transform.*/
od_mc_butterfly_2x2_16x4(&a, &b, &c, &d);
od_mc_butterfly_2x2_16x4(&a, &b, &c, &d);
od_transpose16x4(&a, &b, &c, &d);
/*Horizontal 1D transform.*/
od_mc_butterfly_2x2_16x4(&a, &b, &c, &d);
/*Use the fact that (abs(a+b)+abs(a-b))/2=max(abs(a),abs(b)) to merge the
final butterfly stage with the calculating the absolute values and the
first stage of accumulation.
Calculates (abs(a+b)+abs(a-b))/2-0x7FFF.
An offset must be added to the final sum before rounding to account for
subtracting 0x7FFF.*/
a = _mm_sub_pi16(_mm_max_pi16(a, b), _mm_adds_pi16(_mm_add_pi16(a, b),
_mm_set1_pi16(0x7FFF)));
c = _mm_sub_pi16(_mm_max_pi16(c, d), _mm_adds_pi16(_mm_add_pi16(c, d),
_mm_set1_pi16(0x7FFF)));
/*Take the sum of all the absolute values.*/
sums = _mm_add_pi16(a, c);
/*Sum the elements of the vector.*/
sums = _mm_add_pi16(sums, _mm_shuffle_pi16(sums, _MM_SHUFFLE(0, 1, 2, 3)));
sums = _mm_add_pi16(sums, _mm_shuffle_pi16(sums, _MM_SHUFFLE(2, 3, 0, 1)));
sums = _mm_unpacklo_pi16(sums, _mm_setzero_si64());
satd = _mm_cvtsi64_si32(sums);
/*Subtract the offset (8) and round.*/
satd = (satd + 1 - 8) >> 1;
#if defined(OD_CHECKASM)
{
int32_t c_satd;
c_satd = od_mc_compute_satd8_4x4_c(src, systride, ref, rystride);
if (satd != c_satd) {
fprintf(stderr, "od_mc_compute_satd %ix%i check failed: %i!=%i\n",
4, 4, satd, c_satd);
}
}
#endif
return satd;
}
////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
// Convert YV12 to RGB24.
VOID Yv12ToRgb24_mmx(PBYTE pbDstX, INT iDstXStride,
PBYTE pbSrcY, PBYTE pbSrcU, PBYTE pbSrcV, INT iSrcYStride, INT iSrcUvStride,
UINT uWidth, INT iHeight)
{
UINT x;
INT y;
INT iDstXDif;
INT iSrcYDif;
INT iSrcUvDif;
INT yy, bu, guv, rv;
M64 y0, y1, u0, v0, 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;
pbDstX += (iHeight - 1) * iDstXStride;
iDstXStride = -iDstXStride;
}
iDstXDif = iDstXStride - (uWidth * 3);
iSrcYDif = iSrcYStride - uWidth;
iSrcUvDif = iSrcUvStride - (uWidth / 2);
mz = _mm_setzero_si64();
for (y = iHeight / 2; y; y--)
{
for (x = uWidth / 8; x; x--)
{
// Calculate coefficient.
u0 = _mm_cvtsi32_si64(*((PDWORD) pbSrcU)); // [ | u3 u2 u1 u0]
v0 = _mm_cvtsi32_si64(*((PDWORD) pbSrcV)); // [ | v3 v2 v1 v0]
u0 = _mm_unpacklo_pi8(u0, mz); // u3 u2 u1 u0
v0 = _mm_unpacklo_pi8(v0, mz); // v3 v2 v1 v0
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.
y0 = *((PM64) pbSrcY); // [YYYY | YYYY]; row 0
y1 = _mm_unpackhi_pi8(y0, mz); // y7 y6 y5 y4
y0 = _mm_unpacklo_pi8(y0, mz); // y3 y2 y1 y0
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);
// Process for row 1.
y0 = *((PM64) (pbSrcY + iSrcYStride)); // [YYYY | YYYY]; row 1
y1 = _mm_unpackhi_pi8(y0, mz); // y7 y6 y5 y4
y0 = _mm_unpacklo_pi8(y0, mz); // y3 y2 y1 y0
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 1.
*((PDWORD) (pbDstX + iDstXStride + 0)) = _mm_cvtsi64_si32(rgb0); rgb0 = _mm_srli_si64(rgb0, 32);
*((PDWORD) (pbDstX + iDstXStride + 3)) = _mm_cvtsi64_si32(rgb0);
*((PDWORD) (pbDstX + iDstXStride + 6)) = _mm_cvtsi64_si32(rgb1); rgb1 = _mm_srli_si64(rgb1, 32);
*((PDWORD) (pbDstX + iDstXStride + 9)) = _mm_cvtsi64_si32(rgb1);
*((PDWORD) (pbDstX + iDstXStride + 12)) = _mm_cvtsi64_si32(rgb2); rgb2 = _mm_srli_si64(rgb2, 32);
*((PDWORD) (pbDstX + iDstXStride + 15)) = _mm_cvtsi64_si32(rgb2);
*((PDWORD) (pbDstX + iDstXStride + 18)) = _mm_cvtsi64_si32(rgb3); rgb3 = _mm_srli_si64(rgb3, 32);
*((PDWORD) (pbDstX + iDstXStride + 21)) = _mm_cvtsi64_si32(rgb3);
pbDstX += 24;
pbSrcY += 8;
pbSrcU += 4;
pbSrcV += 4;
}
for (x = (uWidth & 7) / 2; x; x--)
{
bu = g_iBUTab[pbSrcU[0]];
guv = g_iGUTab[pbSrcU[0]] + g_iGVTab[pbSrcV[0]];
rv = g_iRVTab[pbSrcV[0]];
yy = g_iYYTab[pbSrcY[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[pbSrcY[1]];
pbDstX[3] = _Clip((yy + bu) >> SCALEBITS_OUT);
pbDstX[4] = _Clip((yy - guv) >> SCALEBITS_OUT);
pbDstX[5] = _Clip((yy + rv) >> SCALEBITS_OUT);
yy = g_iYYTab[pbSrcY[iSrcYStride]];
pbDstX[iDstXStride + 0] = _Clip((yy + bu) >> SCALEBITS_OUT);
pbDstX[iDstXStride + 1] = _Clip((yy - guv) >> SCALEBITS_OUT);
pbDstX[iDstXStride + 2] = _Clip((yy + rv) >> SCALEBITS_OUT);
yy = g_iYYTab[pbSrcY[iSrcYStride + 1]];
pbDstX[iDstXStride + 3] = _Clip((yy + bu) >> SCALEBITS_OUT);
pbDstX[iDstXStride + 4] = _Clip((yy - guv) >> SCALEBITS_OUT);
pbDstX[iDstXStride + 5] = _Clip((yy + rv) >> SCALEBITS_OUT);
pbDstX += 6;
pbSrcY += 2;
pbSrcU++;
pbSrcV++;
}
pbDstX += iDstXDif + iDstXStride;
pbSrcY += iSrcYDif + iSrcYStride;
pbSrcU += iSrcUvDif;
pbSrcV += iSrcUvDif;
}
_mm_empty();
}
void rtv_lucent4cols_MMX(byte *source, argb_t *dest, int bga, int fga)
{
// SSE2 temporaries:
const __m64 upper8mask = _mm_set_pi16(0, 0xff, 0xff, 0xff);
const __m64 fgAlpha = _mm_set_pi16(0, fga, fga, fga);
const __m64 bgAlpha = _mm_set_pi16(0, bga, bga, bga);
#if 1
const __m64 bgColors01 = _mm_setr_pi32(dest[0], dest[1]);
#else
const __m64 bgColors01 = *((__m64 *)&dest[0]);
#endif
const __m64 fgColors01 = _mm_setr_pi32(
rt_mapcolor<argb_t>(dcol.colormap, source[0]),
rt_mapcolor<argb_t>(dcol.colormap, source[1])
);
const __m64 finalColors01 = _mm_packs_pu16(
_mm_srli_pi16(
_mm_adds_pi16(
_mm_mullo_pi16(_mm_and_si64(_mm_unpacklo_pi8(bgColors01, bgColors01), upper8mask), bgAlpha),
_mm_mullo_pi16(_mm_and_si64(_mm_unpacklo_pi8(fgColors01, fgColors01), upper8mask), fgAlpha)
),
8
),
_mm_srli_pi16(
_mm_adds_pi16(
_mm_mullo_pi16(_mm_and_si64(_mm_unpackhi_pi8(bgColors01, bgColors01), upper8mask), bgAlpha),
_mm_mullo_pi16(_mm_and_si64(_mm_unpackhi_pi8(fgColors01, fgColors01), upper8mask), fgAlpha)
),
8
)
);
#if 1
const __m64 bgColors23 = _mm_setr_pi32(dest[2], dest[3]);
#else
// NOTE(jsd): No guarantee of 64-bit alignment; cannot use.
const __m64 bgColors23 = *((__m64 *)&dest[2]);
#endif
const __m64 fgColors23 = _mm_setr_pi32(
rt_mapcolor<argb_t>(dcol.colormap, source[2]),
rt_mapcolor<argb_t>(dcol.colormap, source[3])
);
const __m64 finalColors23 = _mm_packs_pu16(
_mm_srli_pi16(
_mm_adds_pi16(
_mm_mullo_pi16(_mm_and_si64(_mm_unpacklo_pi8(bgColors23, bgColors23), upper8mask), bgAlpha),
_mm_mullo_pi16(_mm_and_si64(_mm_unpacklo_pi8(fgColors23, fgColors23), upper8mask), fgAlpha)
),
8
),
_mm_srli_pi16(
_mm_adds_pi16(
_mm_mullo_pi16(_mm_and_si64(_mm_unpackhi_pi8(bgColors23, bgColors23), upper8mask), bgAlpha),
_mm_mullo_pi16(_mm_and_si64(_mm_unpackhi_pi8(fgColors23, fgColors23), upper8mask), fgAlpha)
),
8
)
);
#if 1
dest[0] = _mm_cvtsi64_si32(_mm_srli_si64(finalColors01, 32*0));
dest[1] = _mm_cvtsi64_si32(_mm_srli_si64(finalColors01, 32*1));
dest[2] = _mm_cvtsi64_si32(_mm_srli_si64(finalColors23, 32*0));
dest[3] = _mm_cvtsi64_si32(_mm_srli_si64(finalColors23, 32*1));
#else
// NOTE(jsd): No guarantee of 64-bit alignment; cannot use.
*((__m64 *)&dest[0]) = finalColors01;
*((__m64 *)&dest[2]) = finalColors23;
#endif
// Required to reset FP:
_mm_empty();
}
////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
// 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();
}
////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
// Convert YUY2 to YV12.
VOID Yuy2ToYv12_mmx(PBYTE pbDstY, PBYTE pbDstU, PBYTE pbDstV, INT iDstYStride, INT iDstUvStride,
PBYTE pbSrcX, INT iSrcXStride, UINT uWidth, INT iHeight)
{
UINT x;
INT y;
INT iSrcXDif;
INT iDstYDif;
INT iDstUvDif;
M64 m0, m1, m2, m3, m4, m5, m6, m7;
if (iHeight < 0)
{
iHeight = -iHeight;
pbSrcX += (iHeight - 1) * iSrcXStride;
iSrcXStride = -iSrcXStride;
}
iSrcXDif = iSrcXStride - (uWidth * 2);
iDstYDif = iDstYStride - uWidth;
iDstUvDif = iDstUvStride - (uWidth / 2);
m7 = g_mWord00FF;
for (y = iHeight / 2; y; y--)
{
for (x = uWidth / 8; x; x--)
{
m0 = ((PM64) pbSrcX)[0];
m1 = ((PM64) pbSrcX)[1];
m2 = ((PM64) (pbSrcX + iSrcXStride))[0];
m3 = ((PM64) (pbSrcX + iSrcXStride))[1];
m4 = m0;
m5 = m2;
m4 = _mm_srli_pi16(m4, 8);
m5 = _mm_srli_pi16(m5, 8);
m4 = _mm_and_si64(m4, m7);
m5 = _mm_and_si64(m5, m7);
m4 = _mm_add_pi16(m4, m5);
m5 = m1;
m6 = m3;
m5 = _mm_srli_pi16(m5, 8);
m6 = _mm_srli_pi16(m6, 8);
m5 = _mm_and_si64(m5, m7);
m6 = _mm_and_si64(m6, m7);
m5 = _mm_add_pi16(m5, m6);
m4 = _mm_add_pi16(m4, g_mWord0001);
m5 = _mm_add_pi16(m5, g_mWord0001);
m4 = _mm_srli_pi16(m4, 1);
m5 = _mm_srli_pi16(m5, 1);
m0 = _mm_and_si64(m0, m7);
m1 = _mm_and_si64(m1, m7);
m2 = _mm_and_si64(m2, m7);
m3 = _mm_and_si64(m3, m7);
m0 = _mm_packs_pu16(m0, m1);
m2 = _mm_packs_pu16(m2, m3);
((PM64) pbDstY)[0] = m0;
((PM64) (pbDstY + iDstYStride))[0] = m2;
m4 = _mm_packs_pu16(m4, m5);
m5 = m4;
m4 = _mm_srli_si64(m4, 8);
m5 = _mm_and_si64(m5, m7);
m4 = _mm_and_si64(m4, m7);
m5 = _mm_packs_pu16(m5, m5);
m4 = _mm_packs_pu16(m4, m4);
((PDWORD) pbDstU)[0] = _mm_cvtsi64_si32(m5);
((PDWORD) pbDstV)[0] = _mm_cvtsi64_si32(m4);
pbSrcX += 16;
pbDstY += 8;
pbDstU += 4;
pbDstV += 4;
}
for (x = (uWidth & 7) / 2; x; x--)
{
pbDstY[0] = pbSrcX[0];
pbDstU[0] = (pbSrcX[1] + pbSrcX[iSrcXStride + 1] + 1) / 2;
pbDstY[1] = pbSrcX[2];
pbDstV[0] = (pbSrcX[3] + pbSrcX[iSrcXStride + 3] + 1) / 2;
pbDstY[iDstYStride + 0] = pbSrcX[iSrcXStride + 0];
pbDstY[iDstYStride + 1] = pbSrcX[iSrcXStride + 2];
pbSrcX += 4;
pbDstY += 2;
pbDstU++;
pbDstV++;
}
pbSrcX += iSrcXDif + iSrcXStride;
pbDstY += iDstYDif + iDstYStride;
pbDstU += iDstUvDif;
pbDstV += iDstUvDif;
}
_mm_empty();
}
//mbl test with Pocket_PC
void weak_horizontal_luma_MMX(unsigned char pix[], const int xstride, const unsigned char alpha, const unsigned char beta, const unsigned char tc0){
__m64 mp2 = _mm_set_pi16(pix[-3*xstride + 3], pix[-3*xstride + 2], pix[-3*xstride + 1], pix[-3*xstride]);
__m64 mp1 = _mm_set_pi16(pix[-2*xstride + 3], pix[-2*xstride + 2], pix[-2*xstride + 1], pix[-2*xstride]);
__m64 mp0 = _mm_set_pi16(pix[-1*xstride + 3], pix[-1*xstride + 2], pix[-1*xstride + 1], pix[-1*xstride]);
__m64 mq0 = _mm_set_pi16(pix[3], pix[2], pix[1], pix[0]);
__m64 mq1 = _mm_set_pi16(pix[xstride + 3], pix[xstride + 2], pix[xstride + 1], pix[xstride]);
__m64 mq2 = _mm_set_pi16(pix[2*xstride + 3], pix[2*xstride + 2], pix[2*xstride + 1], pix[2*xstride]);
__m64 mrshift = _mm_set_pi16(0,0,0,1);
__m64 maddp0_q0 = _mm_avg_pu8(mp0,mq0); //addp0_q0 = (p0 + q0 + 1) >> 1;
__m64 maddp2 = _mm_add_pi16(maddp0_q0,mp2); //addp2 = (p2 + addp0_q0);
__m64 maddq2 = _mm_add_pi16(maddp0_q0,mq2); //addp2 = (p2 + addp0_q0);
__m64 maddp2_s = _mm_srl_pi16(maddp2,mrshift); //addp2 = (p2 + addp0_q0) >> 1;
__m64 maddq2_s = _mm_srl_pi16(maddq2,mrshift); //addp2 = (p2 + addp0_q0) >> 1;
__m64 mp1_c = _mm_sub_pi16(maddp2_s, mp1); //addp2 - p1
__m64 mq1_c = _mm_sub_pi16(maddq2_s, mq1); // addq2 - q1
//To calculate the mask
__m64 malpha = _mm_set_pi16(alpha,alpha,alpha,alpha);
__m64 malphab = _mm_set_pi16(-alpha,-alpha,-alpha,-alpha);
__m64 mbeta = _mm_set_pi16(beta,beta,beta,beta);
__m64 mbetab = _mm_set_pi16(-beta,-beta,-beta,-beta);
__m64 mdiff_p0_q0 = _mm_sub_pi16(mq0,mp0); //abs(q0 - p0)
__m64 mdiff_p1_p0 = _mm_sub_pi16(mp0,mp1); //abs(p1 - p0)
__m64 mdiff_q1_q0 = _mm_sub_pi16(mq0, mq1); //abs(q1 - q0)
__m64 mdiff_p2_p0 = _mm_sub_pi16(mp2,mp0); //abs(p2 - p0 ))
__m64 mdiff_q2_q0 = _mm_sub_pi16(mq2,mq0); //abs(q2 - q0)
__m64 mask0 = _mm_and_si64( _mm_cmpgt_pi16(malpha, mdiff_p0_q0), _mm_cmpgt_pi16(mdiff_p0_q0,malphab));
__m64 mask1 = _mm_and_si64( _mm_cmpgt_pi16(mbeta, mdiff_p1_p0), _mm_cmpgt_pi16(mdiff_p1_p0,mbetab));
__m64 mask2 = _mm_and_si64( _mm_cmpgt_pi16(mbeta, mdiff_q1_q0), _mm_cmpgt_pi16(mdiff_q1_q0,mbetab));
__m64 mask3 = _mm_and_si64( _mm_cmpgt_pi16(mbeta, mdiff_p2_p0), _mm_cmpgt_pi16(mdiff_p2_p0,mbetab));
__m64 mask4 = _mm_and_si64( _mm_cmpgt_pi16(mbeta, mdiff_q2_q0), _mm_cmpgt_pi16(mdiff_q2_q0,mbetab));
__m64 first_mask = _mm_and_si64 (_mm_and_si64 (mask0,mask1),mask2); //(abs(q0 - p0) < alpha) && (abs(p1 - p0) < beta) && (abs(q1 - q0) < beta)
__m64 second_mask = _mm_and_si64 (first_mask,mask3);
__m64 third_mask = _mm_and_si64 (first_mask,mask4);
__m64 mdiff_q0_p0 = _mm_sub_pi16(mq0,mp0); //(q0 - p0)
__m64 mlshift = _mm_set_pi16(0,0,0,2);
__m64 minter_1 = _mm_sll_pi16(mdiff_q0_p0, mlshift);//inter_1 = (q0 - p0 ) << 2;
__m64 minter_2 = _mm_sub_pi16(mp1, mq1);//(p1 - q1)
__m64 madd4 = _mm_set_pi16(4,4,4,4);
__m64 minter_3 = _mm_add_pi16(minter_2, madd4);//inter_2 = (p1 - q1) + 4;
__m64 minter_4 = _mm_add_pi16(minter_3,minter_1); //(inter_1 + inter_2)
__m64 mrshift3 = _mm_set_pi16(0,0,0,3);
__m64 minter5 = _mm_sra_pi16(minter_4, mrshift3);
//Clip3
__m64 m_tc0 = _mm_set_pi16(tc0,tc0,tc0,tc0);
__m64 m_tcb0 = _mm_set_pi16(-tc0,-tc0,-tc0,-tc0);
__m64 mres_c1 = _mm_min_pi16(_mm_max_pi16(mp1_c,m_tcb0),m_tc0); //CLIP3(-tc0, tc0, addp2 - p1 );
__m64 mres_c2 = _mm_min_pi16(_mm_max_pi16(mq1_c,m_tcb0),m_tc0); //CLIP3(-tc0, tc0, addq2 - q1 );
__m64 merror0 = _mm_and_si64 (mres_c1,second_mask);
__m64 merror1 = _mm_and_si64 (mres_c2,third_mask);
__m64 m_1 = _mm_set_pi16(1,1,1,1);
__m64 m_and1 = _mm_and_si64 (mask3, m_1); //tc++; if abs( p2 - p0 ) < beta
__m64 m_and2 = _mm_and_si64 (mask4, m_1); //tc++; if abs( q2 - q0 ) < beta
__m64 m_tc = _mm_add_pi16(m_and2,_mm_add_pi16(m_tc0,m_and1));
__m64 m_tcn =_mm_sub_pi16(_mm_sub_pi16(m_tcb0,m_and1),m_and2);
__m64 mres_c3 = _mm_min_pi16(_mm_max_pi16(minter5,m_tcn),m_tc); //CLIP3(-tc0, tc0, addp2 - p1 );
__m64 merror2 = _mm_and_si64 (mres_c3,first_mask);
__m64 result_p1 = _mm_add_pi16(merror0,mp1); //_mm_shuffle_pi16(_mm_add_pi16(merror0,mp1), 0x1B);
__m64 result_q1 = _mm_add_pi16(merror1,mq1); //_mm_shuffle_pi16(_mm_add_pi16(merror1,mq1), 0x1B);
__m64 result_p0 = _mm_add_pi16(merror2,mp0); //_mm_shuffle_pi16(_mm_add_pi16(merror2,mq1), 0x1B);
__m64 result_q0 = _mm_sub_pi16(mq0, merror2);//_mm_shuffle_pi16(_mm_sub_pi16(mq1, merror2), 0x1B);
*((unsigned int* )(&pix[-2*xstride])) = _mm_cvtsi64_si32(_mm_packs_pu16(result_p1,mrshift));
*((unsigned int* )(&pix[-xstride])) = _mm_cvtsi64_si32(_mm_packs_pu16(result_p0,mrshift));
*((unsigned int* )(&pix[0])) = _mm_cvtsi64_si32(_mm_packs_pu16(result_q0,mrshift));
*((unsigned int* )(&pix[xstride])) = _mm_cvtsi64_si32(_mm_packs_pu16(result_q1,mrshift));
empty();
}
unsigned int interpolvline64_3_mmx(__m64* temp){
__m64 res,res1;
__m64 ptr = _mm_setzero_si64();
__m64 mm_16 = _mm_set_pi16(16,16,16,16);
short A = _mm_extract_pi16(temp[0],0);
short B = _mm_extract_pi16(temp[1],0);
short C = _mm_extract_pi16(temp[2],0);
short D = _mm_extract_pi16(temp[3],0);
short E = _mm_extract_pi16(temp[4],0);
short F = _mm_extract_pi16(temp[5],0);
unsigned int result = A + F - 5 * (short)(B + E) + 20 * (short)(C + D) + 512;
ptr = _mm_insert_pi16(ptr,CLIP255_16(result >> 10),0);
A = _mm_extract_pi16(temp[0],1);
B = _mm_extract_pi16(temp[1],1);
C = _mm_extract_pi16(temp[2],1);
D = _mm_extract_pi16(temp[3],1);
E = _mm_extract_pi16(temp[4],1);
F = _mm_extract_pi16(temp[5],1);
result = A + F - 5 * (short)(B + E) + 20 * (short)(C + D) + 512;
ptr = _mm_insert_pi16(ptr,CLIP255_16(result >> 10),1);
A = _mm_extract_pi16(temp[0],2);
B = _mm_extract_pi16(temp[1],2);
C = _mm_extract_pi16(temp[2],2);
D = _mm_extract_pi16(temp[3],2);
E = _mm_extract_pi16(temp[4],2);
F = _mm_extract_pi16(temp[5],2);
result = A + F - 5 * (short)(B + E) + 20 * (short)(C + D) + 512;
ptr = _mm_insert_pi16(ptr,CLIP255_16(result >> 10),2);
A = _mm_extract_pi16(temp[0],3);
B = _mm_extract_pi16(temp[1],3);
C = _mm_extract_pi16(temp[2],3);
D = _mm_extract_pi16(temp[3],3);
E = _mm_extract_pi16(temp[4],3);
F = _mm_extract_pi16(temp[5],3);
result = A + F - 5 * (short)(B + E) + 20 * (short)(C + D) + 512;
ptr = _mm_insert_pi16(ptr,CLIP255_16(result >> 10),3);
res = _mm_add_pi16(temp[3],mm_16);
res1 = _mm_srai_pi16(res,5);
res1 = _mm_max_pi16(res1,_mm_set_pi16(0,0,0,0));
res1 = _mm_min_pi16(res1,_mm_set_pi16(255,255,255,255)); //Clip
res = _mm_set_pi8(0,0,0,0,_mm_extract_pi16(ptr,3),_mm_extract_pi16(ptr,2),_mm_extract_pi16(ptr,1),_mm_extract_pi16(ptr,0));
res1 =_mm_set_pi8(0,0,0,0,_mm_extract_pi16(res1,3),_mm_extract_pi16(res1,2),_mm_extract_pi16(res1,1),_mm_extract_pi16(res1,0));
res = _mm_avg_pu8(res,res1);//(ptr_img[0] + ptr_rf[0] + 1) >> 1
result = _mm_cvtsi64_si32(res);
empty();
return result;
}
void
mlib_m_ImageMaximum_U8_3(
mlib_s32 *res32,
const mlib_image *img)
{
/* src address */
__m64 *sp, *sl;
/* src data */
__m64 sd;
/* max values */
__m64 max0, max1, max2, max3;
/* edge mask */
mlib_s32 emask;
/* loop variables */
mlib_s32 n1;
/* height of image */
mlib_s32 height = mlib_ImageGetHeight(img);
/* elements to next row */
mlib_s32 slb = mlib_ImageGetStride(img);
mlib_s32 width = mlib_ImageGetWidth(img) * 3;
mlib_u8 *dend;
if (slb == width) {
width *= height;
height = 1;
}
sp = sl = (__m64 *) mlib_ImageGetData(img);
max1 = _mm_set1_pi8(MLIB_U8_MIN);
max2 = _mm_set1_pi8(MLIB_U8_MIN);
max3 = _mm_set1_pi8(MLIB_U8_MIN);
for (; height > 0; height--) {
n1 = width;
dend = (mlib_u8 *)sp + width;
for (; n1 > 23; n1 -= 24) {
sd = (*sp++);
MLIB_M_IMAGE_MAXIMUM_U8(max1, max1, sd);
sd = (*sp++);
MLIB_M_IMAGE_MAXIMUM_U8(max2, max2, sd);
sd = (*sp++);
MLIB_M_IMAGE_MAXIMUM_U8(max3, max3, sd);
}
if (n1 > 0) {
emask = (n1 > 7) ? 0xFF : (0xFF << (8 - n1));
sd = (*sp++);
MLIB_M_IMAGE_MAXIMUM_U8_M32(max1, max1, sd, emask);
n1 = ((mlib_u8 *)dend - (mlib_u8 *)sp);
if (n1 > 0) {
emask = (n1 > 7) ? 0xFF : (0xFF << (8 - n1));
sd = (*sp++);
MLIB_M_IMAGE_MAXIMUM_U8_M32(max2, max2, sd,
emask);
n1 = ((mlib_u8 *)dend - (mlib_u8 *)sp);
if (n1 > 0) {
emask = (0xFF << (8 - n1));
sd = *sp;
MLIB_M_IMAGE_MAXIMUM_U8_M32(max3, max3,
sd, emask);
}
}
}
sp = sl = (__m64 *) ((mlib_u8 *)sl + slb);
}
MLIB_M_IMAGE_MAXIMUM_U8_M64(max0, max1, _mm_srli_si64(max2, 8),
mmx_write_64(0x00ffffffffffffffll));
MLIB_M_IMAGE_MAXIMUM_U8_M64(max0, max0, _mm_slli_si64(max2, 16),
mmx_write_64(0x0000000000ff0000ll));
MLIB_M_IMAGE_MAXIMUM_U8_M64(max0, max0, _mm_srli_si64(max3, 16),
mmx_write_64(0x0000ffffffffffffll));
MLIB_M_IMAGE_MAXIMUM_U8_M64(max0, max0, _mm_slli_si64(max3, 8),
mmx_write_64(0x0000000000ffff00ll));
MLIB_M_IMAGE_MAXIMUM_U8_M64(max0, max0, _mm_srli_si64(max0, 24),
mmx_write_64(0x000000ffff000000ll));
MLIB_M_IMAGE_MAXIMUM_U8_M64(max0, max0, _mm_srli_si64(max0, 24),
mmx_write_64(0x0000000000ffffffll));
res32[0] = _mm_cvtsi64_si32(_mm_and_si64(max0,
mmx_write_64(0x00000000000000ffll)));
res32[1] =
_mm_cvtsi64_si32(_mm_and_si64(_mm_srli_si64(max0, 8),
mmx_write_64(0x00000000000000ffll)));
res32[2] =
_mm_cvtsi64_si32(_mm_and_si64(_mm_srli_si64(max0, 16),
mmx_write_64(0x00000000000000ffll)));
_mm_empty();
}
int test_mm_cvtsi64_si32(__m64 a) {
// CHECK-LABEL: test_mm_cvtsi64_si32
// CHECK: extractelement <2 x i32>
return _mm_cvtsi64_si32(a);
}