void pix_compare :: processYUV_MMX(imageStruct &image, imageStruct &right)
{
long datasize = image.xsize * image.ysize * image.csize;
datasize=datasize/sizeof(__m64)+(datasize%sizeof(__m64)!=0);
__m64*leftPix = (__m64*)image.data;
__m64*rightPix = (__m64*)right.data;
__m64 l, r, b;
__m64 mask = _mm_setr_pi8((unsigned char)0x00,
(unsigned char)0xFF,
(unsigned char)0x00,
(unsigned char)0xFF,
(unsigned char)0x00,
(unsigned char)0xFF,
(unsigned char)0x00,
(unsigned char)0xFF);
__m64 zeros = _mm_set1_pi8((unsigned char)0x00);
//format is U Y V Y
if (m_direction) {
while(datasize--){
l=leftPix[datasize];
r=rightPix[datasize];
b=_mm_subs_pu8(l, r);
b=_mm_and_si64(b, mask);
b=_mm_cmpeq_pi32(b, zeros);
r=_mm_and_si64(r, b);
l=_mm_andnot_si64(b, l);
leftPix[datasize]=_mm_or_si64(l, r);
}
} else {
while(datasize--){
l=leftPix[datasize];
r=rightPix[datasize];
b=_mm_subs_pu8(r, l);
b=_mm_and_si64(b, mask);
b=_mm_cmpeq_pi32(b, zeros);
r=_mm_and_si64(r, b);
l=_mm_andnot_si64(b, l);
leftPix[datasize]=_mm_or_si64(l, r);
}
}
_mm_empty();
}
void pix_background :: processRGBAMMX(imageStruct &image)
{
long i,pixsize;
pixsize = image.xsize * image.ysize * image.csize;
if(m_savedImage.xsize!=image.xsize ||
m_savedImage.ysize!=image.ysize ||
m_savedImage.format!=image.format)m_reset=1;
m_savedImage.xsize=image.xsize;
m_savedImage.ysize=image.ysize;
m_savedImage.setCsizeByFormat(image.format);
m_savedImage.reallocate();
if (m_reset){
memcpy(m_savedImage.data,image.data,pixsize);
}
m_reset=0;
i=pixsize/sizeof(__m64)+(pixsize%sizeof(__m64)!=0);
__m64*data =(__m64*)image.data;
__m64*saved=(__m64*)m_savedImage.data;
const __m64 thresh=_mm_set_pi8(m_Yrange, m_Urange, m_Vrange, m_Arange,
m_Yrange, m_Urange, m_Vrange, m_Arange);
const __m64 offset=_mm_set_pi8(1, 1, 1, 1, 1, 1, 1, 1);
__m64 newpix, oldpix, m1;
while(i--){
/* 7ops, 3memops */
/* i have the feeling that this is not faster at all!
* even if i have the 3memops + ONLY 1 _mm_subs_pu8()
* i am equally slow as the generic code;
* adding the other instruction does not change much
*/
newpix=*data;
oldpix=*saved++;
m1 = newpix;
m1 = _mm_subs_pu8 (m1, oldpix);
oldpix= _mm_subs_pu8 (oldpix, newpix);
m1 = _mm_or_si64 (m1, oldpix); // |oldpix-newpix|
m1 = _mm_adds_pu8 (m1, offset);
m1 = _mm_subs_pu8 (m1, thresh);
m1 = _mm_cmpeq_pi32 (m1, _mm_setzero_si64()); // |oldpix-newpix|>thresh
m1 = _mm_andnot_si64(m1, newpix);
*data++ = m1;
}
_mm_empty();
}
void pix_background :: processYUVMMX(imageStruct &image)
{
long pixsize;
pixsize = image.xsize * image.ysize * image.csize;
if(m_savedImage.xsize!=image.xsize ||
m_savedImage.ysize!=image.ysize ||
m_savedImage.format!=image.format)m_reset=1;
m_savedImage.xsize=image.xsize;
m_savedImage.ysize=image.ysize;
m_savedImage.setCsizeByFormat(image.format);
m_savedImage.reallocate();
if (m_reset){
memcpy(m_savedImage.data,image.data,pixsize);
// return;
}
m_reset=0;
int i=pixsize/sizeof(__m64)+(pixsize%sizeof(__m64)!=0);
__m64*data =(__m64*)image.data;
__m64*saved=(__m64*)m_savedImage.data;
const __m64 thresh=_mm_set_pi8(m_Urange, m_Yrange, m_Vrange, m_Yrange,
m_Urange, m_Yrange, m_Vrange, m_Yrange);
const __m64 offset=_mm_set_pi8(1, 1, 1, 1, 1, 1, 1, 1);
const __m64 black =_mm_set_pi8((unsigned char)0x00,
(unsigned char)0x80,
(unsigned char)0x00,
(unsigned char)0x80,
(unsigned char)0x00,
(unsigned char)0x80,
(unsigned char)0x00,
(unsigned char)0x80);
__m64 newpix, oldpix, m1;
while(i--){
newpix=*data;
oldpix=*saved++;
m1 = newpix;
m1 = _mm_subs_pu8 (m1, oldpix);
oldpix= _mm_subs_pu8 (oldpix, newpix);
m1 = _mm_or_si64 (m1, oldpix); // |oldpix-newpix|
m1 = _mm_adds_pu8 (m1, offset); // to make thresh=0 work correctly
m1 = _mm_subs_pu8 (m1, thresh); // m1>thresh -> saturation -> 0
m1 = _mm_cmpeq_pi32 (m1, _mm_setzero_si64()); // |oldpix-newpix|>thresh
oldpix= black;
oldpix= _mm_and_si64 (oldpix, m1);
m1 = _mm_andnot_si64 (m1, newpix);
m1 = _mm_or_si64 (m1, oldpix);
*data++ = m1;
}
_mm_empty();
}
__m64 test86(__m64 a, __m64 b) {
// CHECK: pcmpeqd
return _mm_cmpeq_pi32(a, b);
}
/* since sin_ps and cos_ps are almost identical, sincos_ps could replace both of them..
it is almost as fast, and gives you a free cosine with your sine */
void sincos_ps(v4sfu *xptr, v4sfu *sptr, v4sfu *cptr) {
__m128 x=*((__m128 *)xptr), *s=(__m128 *)sptr, *c=(__m128 *)cptr, xmm1, xmm2, xmm3 = _mm_setzero_ps(), sign_bit_sin, y;
#ifdef USE_SSE2
__m128i emm0, emm2, emm4;
#else
__m64 mm0, mm1, mm2, mm3, mm4, mm5;
#endif
sign_bit_sin = x;
/* take the absolute value */
x = _mm_and_ps(x, *(__m128*)_ps_inv_sign_mask);
/* extract the sign bit (upper one) */
sign_bit_sin = _mm_and_ps(sign_bit_sin, *(__m128*)_ps_sign_mask);
/* scale by 4/Pi */
y = _mm_mul_ps(x, *(__m128*)_ps_cephes_FOPI);
#ifdef USE_SSE2
/* store the integer part of y in emm2 */
emm2 = _mm_cvttps_epi32(y);
/* j=(j+1) & (~1) (see the cephes sources) */
emm2 = _mm_add_epi32(emm2, *(__m128i*)_pi32_1);
emm2 = _mm_and_si128(emm2, *(__m128i*)_pi32_inv1);
y = _mm_cvtepi32_ps(emm2);
emm4 = emm2;
/* get the swap sign flag for the sine */
emm0 = _mm_and_si128(emm2, *(__m128i*)_pi32_4);
emm0 = _mm_slli_epi32(emm0, 29);
__m128 swap_sign_bit_sin = _mm_castsi128_ps(emm0);
/* get the polynom selection mask for the sine*/
emm2 = _mm_and_si128(emm2, *(__m128i*)_pi32_2);
emm2 = _mm_cmpeq_epi32(emm2, _mm_setzero_si128());
__m128 poly_mask = _mm_castsi128_ps(emm2);
#else
/* store the integer part of y in mm2:mm3 */
xmm3 = _mm_movehl_ps(xmm3, y);
mm2 = _mm_cvttps_pi32(y);
mm3 = _mm_cvttps_pi32(xmm3);
/* j=(j+1) & (~1) (see the cephes sources) */
mm2 = _mm_add_pi32(mm2, *(__m64*)_pi32_1);
mm3 = _mm_add_pi32(mm3, *(__m64*)_pi32_1);
mm2 = _mm_and_si64(mm2, *(__m64*)_pi32_inv1);
mm3 = _mm_and_si64(mm3, *(__m64*)_pi32_inv1);
y = _mm_cvtpi32x2_ps(mm2, mm3);
mm4 = mm2;
mm5 = mm3;
/* get the swap sign flag for the sine */
mm0 = _mm_and_si64(mm2, *(__m64*)_pi32_4);
mm1 = _mm_and_si64(mm3, *(__m64*)_pi32_4);
mm0 = _mm_slli_pi32(mm0, 29);
mm1 = _mm_slli_pi32(mm1, 29);
__m128 swap_sign_bit_sin;
COPY_MM_TO_XMM(mm0, mm1, swap_sign_bit_sin);
/* get the polynom selection mask for the sine */
mm2 = _mm_and_si64(mm2, *(__m64*)_pi32_2);
mm3 = _mm_and_si64(mm3, *(__m64*)_pi32_2);
mm2 = _mm_cmpeq_pi32(mm2, _mm_setzero_si64());
mm3 = _mm_cmpeq_pi32(mm3, _mm_setzero_si64());
__m128 poly_mask;
COPY_MM_TO_XMM(mm2, mm3, poly_mask);
#endif
/* The magic pass: "******"
x = ((x - y * DP1) - y * DP2) - y * DP3; */
xmm1 = *(__m128*)_ps_minus_cephes_DP1;
xmm2 = *(__m128*)_ps_minus_cephes_DP2;
xmm3 = *(__m128*)_ps_minus_cephes_DP3;
xmm1 = _mm_mul_ps(y, xmm1);
xmm2 = _mm_mul_ps(y, xmm2);
xmm3 = _mm_mul_ps(y, xmm3);
x = _mm_add_ps(x, xmm1);
x = _mm_add_ps(x, xmm2);
x = _mm_add_ps(x, xmm3);
#ifdef USE_SSE2
emm4 = _mm_sub_epi32(emm4, *(__m128i*)_pi32_2);
emm4 = _mm_andnot_si128(emm4, *(__m128i*)_pi32_4);
emm4 = _mm_slli_epi32(emm4, 29);
__m128 sign_bit_cos = _mm_castsi128_ps(emm4);
#else
/* get the sign flag for the cosine */
mm4 = _mm_sub_pi32(mm4, *(__m64*)_pi32_2);
mm5 = _mm_sub_pi32(mm5, *(__m64*)_pi32_2);
mm4 = _mm_andnot_si64(mm4, *(__m64*)_pi32_4);
mm5 = _mm_andnot_si64(mm5, *(__m64*)_pi32_4);
mm4 = _mm_slli_pi32(mm4, 29);
mm5 = _mm_slli_pi32(mm5, 29);
__m128 sign_bit_cos;
COPY_MM_TO_XMM(mm4, mm5, sign_bit_cos);
_mm_empty(); /* good-bye mmx */
#endif
sign_bit_sin = _mm_xor_ps(sign_bit_sin, swap_sign_bit_sin);
/* Evaluate the first polynom (0 <= x <= Pi/4) */
__m128 z = _mm_mul_ps(x,x);
y = *(__m128*)_ps_coscof_p0;
y = _mm_mul_ps(y, z);
y = _mm_add_ps(y, *(__m128*)_ps_coscof_p1);
y = _mm_mul_ps(y, z);
y = _mm_add_ps(y, *(__m128*)_ps_coscof_p2);
y = _mm_mul_ps(y, z);
y = _mm_mul_ps(y, z);
__m128 tmp = _mm_mul_ps(z, *(__m128*)_ps_0p5);
y = _mm_sub_ps(y, tmp);
y = _mm_add_ps(y, *(__m128*)_ps_1);
/* Evaluate the second polynom (Pi/4 <= x <= 0) */
__m128 y2 = *(__m128*)_ps_sincof_p0;
y2 = _mm_mul_ps(y2, z);
y2 = _mm_add_ps(y2, *(__m128*)_ps_sincof_p1);
y2 = _mm_mul_ps(y2, z);
y2 = _mm_add_ps(y2, *(__m128*)_ps_sincof_p2);
y2 = _mm_mul_ps(y2, z);
y2 = _mm_mul_ps(y2, x);
y2 = _mm_add_ps(y2, x);
/* select the correct result from the two polynoms */
xmm3 = poly_mask;
__m128 ysin2 = _mm_and_ps(xmm3, y2);
__m128 ysin1 = _mm_andnot_ps(xmm3, y);
y2 = _mm_sub_ps(y2,ysin2);
y = _mm_sub_ps(y, ysin1);
xmm1 = _mm_add_ps(ysin1,ysin2);
xmm2 = _mm_add_ps(y,y2);
/* update the sign */
*s = _mm_xor_ps(xmm1, sign_bit_sin);
*c = _mm_xor_ps(xmm2, sign_bit_cos);
}
/* almost the same as sin_ps */
__m128 cos_ps(v4sfu *xPtr) { // any x
__m128 x=*((__m128 *)xPtr);
__m128 xmm1, xmm2 = _mm_setzero_ps(), xmm3, y;
#ifdef USE_SSE2
__m128i emm0, emm2;
#else
__m64 mm0, mm1, mm2, mm3;
#endif
/* take the absolute value */
x = _mm_and_ps(x, *(__m128*)_ps_inv_sign_mask);
/* scale by 4/Pi */
y = _mm_mul_ps(x, *(__m128*)_ps_cephes_FOPI);
#ifdef USE_SSE2
/* store the integer part of y in mm0 */
emm2 = _mm_cvttps_epi32(y);
/* j=(j+1) & (~1) (see the cephes sources) */
emm2 = _mm_add_epi32(emm2, *(__m128i*)_pi32_1);
emm2 = _mm_and_si128(emm2, *(__m128i*)_pi32_inv1);
y = _mm_cvtepi32_ps(emm2);
emm2 = _mm_sub_epi32(emm2, *(__m128i*)_pi32_2);
/* get the swap sign flag */
emm0 = _mm_andnot_si128(emm2, *(__m128i*)_pi32_4);
emm0 = _mm_slli_epi32(emm0, 29);
/* get the polynom selection mask */
emm2 = _mm_and_si128(emm2, *(__m128i*)_pi32_2);
emm2 = _mm_cmpeq_epi32(emm2, _mm_setzero_si128());
__m128 sign_bit = _mm_castsi128_ps(emm0);
__m128 poly_mask = _mm_castsi128_ps(emm2);
#else
/* store the integer part of y in mm0:mm1 */
xmm2 = _mm_movehl_ps(xmm2, y);
mm2 = _mm_cvttps_pi32(y);
mm3 = _mm_cvttps_pi32(xmm2);
/* j=(j+1) & (~1) (see the cephes sources) */
mm2 = _mm_add_pi32(mm2, *(__m64*)_pi32_1);
mm3 = _mm_add_pi32(mm3, *(__m64*)_pi32_1);
mm2 = _mm_and_si64(mm2, *(__m64*)_pi32_inv1);
mm3 = _mm_and_si64(mm3, *(__m64*)_pi32_inv1);
y = _mm_cvtpi32x2_ps(mm2, mm3);
mm2 = _mm_sub_pi32(mm2, *(__m64*)_pi32_2);
mm3 = _mm_sub_pi32(mm3, *(__m64*)_pi32_2);
/* get the swap sign flag in mm0:mm1 and the
polynom selection mask in mm2:mm3 */
mm0 = _mm_andnot_si64(mm2, *(__m64*)_pi32_4);
mm1 = _mm_andnot_si64(mm3, *(__m64*)_pi32_4);
mm0 = _mm_slli_pi32(mm0, 29);
mm1 = _mm_slli_pi32(mm1, 29);
mm2 = _mm_and_si64(mm2, *(__m64*)_pi32_2);
mm3 = _mm_and_si64(mm3, *(__m64*)_pi32_2);
mm2 = _mm_cmpeq_pi32(mm2, _mm_setzero_si64());
mm3 = _mm_cmpeq_pi32(mm3, _mm_setzero_si64());
__m128 sign_bit, poly_mask;
COPY_MM_TO_XMM(mm0, mm1, sign_bit);
COPY_MM_TO_XMM(mm2, mm3, poly_mask);
_mm_empty(); /* good-bye mmx */
#endif
/* The magic pass: "******"
x = ((x - y * DP1) - y * DP2) - y * DP3; */
xmm1 = *(__m128*)_ps_minus_cephes_DP1;
xmm2 = *(__m128*)_ps_minus_cephes_DP2;
xmm3 = *(__m128*)_ps_minus_cephes_DP3;
xmm1 = _mm_mul_ps(y, xmm1);
xmm2 = _mm_mul_ps(y, xmm2);
xmm3 = _mm_mul_ps(y, xmm3);
x = _mm_add_ps(x, xmm1);
x = _mm_add_ps(x, xmm2);
x = _mm_add_ps(x, xmm3);
/* Evaluate the first polynom (0 <= x <= Pi/4) */
y = *(__m128*)_ps_coscof_p0;
__m128 z = _mm_mul_ps(x,x);
y = _mm_mul_ps(y, z);
y = _mm_add_ps(y, *(__m128*)_ps_coscof_p1);
y = _mm_mul_ps(y, z);
y = _mm_add_ps(y, *(__m128*)_ps_coscof_p2);
y = _mm_mul_ps(y, z);
y = _mm_mul_ps(y, z);
__m128 tmp = _mm_mul_ps(z, *(__m128*)_ps_0p5);
y = _mm_sub_ps(y, tmp);
y = _mm_add_ps(y, *(__m128*)_ps_1);
/* Evaluate the second polynom (Pi/4 <= x <= 0) */
__m128 y2 = *(__m128*)_ps_sincof_p0;
y2 = _mm_mul_ps(y2, z);
y2 = _mm_add_ps(y2, *(__m128*)_ps_sincof_p1);
y2 = _mm_mul_ps(y2, z);
y2 = _mm_add_ps(y2, *(__m128*)_ps_sincof_p2);
y2 = _mm_mul_ps(y2, z);
y2 = _mm_mul_ps(y2, x);
y2 = _mm_add_ps(y2, x);
/* select the correct result from the two polynoms */
xmm3 = poly_mask;
y2 = _mm_and_ps(xmm3, y2); //, xmm3);
y = _mm_andnot_ps(xmm3, y);
y = _mm_add_ps(y,y2);
/* update the sign */
y = _mm_xor_ps(y, sign_bit);
return y;
}
__m64 test_mm_cmpeq_pi32(__m64 a, __m64 b) {
// CHECK-LABEL: test_mm_cmpeq_pi32
// CHECK: call x86_mmx @llvm.x86.mmx.pcmpeq.d
return _mm_cmpeq_pi32(a, b);
}
