static void
TEST (void)
{
__m64_union u, s1, s2;
__m64_union e;
int i, tmp;
s1.as_m64 = _mm_set_pi8 (30, 2, 3, 4, 10, 20, 30, 90);
s2.as_m64 = _mm_set_pi8 (88, 44, 3, 22, 11, 98, 76, 100);
u.as_m64 = test (s1.as_m64, s2.as_m64);
for (i = 0; i < 8; i++)
{
tmp = s1.as_char[i] - s2.as_char[i];
if (tmp > 255)
tmp = -1;
if (tmp < 0)
tmp = 0;
e.as_char[i] = tmp;
}
if (u.as_m64 != e.as_m64)
abort ();
}
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();
}
static void TEST (void)
{
__m64_union u, e, s1, s2;
int i;
s1.as_m64 = _mm_set_pi8 (1, 2, 3, 4, 5, 6, 7, 8);
s2.as_m64 = _mm_set_pi8 (8, 7, 6, 5, 4, 3, 2, 1);
u.as_m64 = test (s1.as_m64, s2.as_m64);
for (i = 0; i < 8; i++)
e.as_char[i] =
((unsigned char) s1.as_char[i] > (unsigned char) s2.as_char[i]) ?
s1.as_char[i] : s2.as_char[i];
if (u.as_m64 != e.as_m64)
abort ();
}
void pix_movement :: processGrayMMX(imageStruct &image)
{
// assume that the pix_size does not change !
bool doclear=(image.xsize*image.ysize != buffer.xsize*buffer.ysize);
buffer.xsize = image.xsize;
buffer.ysize = image.ysize;
buffer.reallocate();
if(doclear) {
buffer.setWhite();
}
buffer2.xsize = image.xsize;
buffer2.ysize = image.ysize;
buffer2.reallocate();
int pixsize = image.ysize * image.xsize / sizeof(__m64);
unsigned char thresh=threshold;
__m64*rp = (__m64*)image.data; // read pointer
__m64*wp = (__m64*)buffer.data; // write pointer to the copy
__m64*wp2= (__m64*)buffer2.data; // write pointer to the diff-image
__m64 m1, m2, grey;
__m64 thresh8=_mm_set_pi8(thresh,thresh,thresh,thresh,
thresh,thresh,thresh,thresh);
// there is still one problem with the threshold: is the cmpgt only for signed ?
while(pixsize--) {
grey = rp[pixsize]; // image.data
m2 = wp[pixsize]; // buffer.data
//m0 =_mm_cmpgt_pi8(grey, m2); // (grey>m2)
//m1 =_mm_subs_pu8 (grey, m2); // (grey-m2)
//m2 =_mm_subs_pu8 (m2, grey); // (m2-grey)
//m1 =_mm_and_si64 (m1, m0); // (m2-grey)&(grey>m2) ((??))
//m0 =_mm_andnot_si64(m0, m2); // !(grey>m2)&(grey-m2) ((??))
//m2 =_mm_or_si64 (m2, m0); // [(a-b)&(a>b)]|[(b-a)&!(a>b)]=abs(a-b)
// this is better: use saturated arithmetic!
m1 =_mm_subs_pu8 (grey, m2); // (grey-m2)
m2 =_mm_subs_pu8 (m2, grey); // (m2-grey)
wp[pixsize]=grey; // buffer.data
m2 = _mm_or_si64 (m2, m1); // |grey-m2|
m2 =_mm_subs_pu8 (m2, thresh8);
m2 =_mm_cmpgt_pi8(m2, _mm_setzero_si64());
wp2[pixsize]=m2; // output.data
}
_mm_empty();
image.data = buffer2.data;
}
__m64 test_mm_set_pi8(char a, char b, char c, char d, char e, char f, char g, char h) {
// CHECK-LABEL: test_mm_set_pi8
// CHECK: insertelement <8 x i8>
// CHECK: insertelement <8 x i8>
// CHECK: insertelement <8 x i8>
// CHECK: insertelement <8 x i8>
// CHECK: insertelement <8 x i8>
// CHECK: insertelement <8 x i8>
// CHECK: insertelement <8 x i8>
// CHECK: insertelement <8 x i8>
return _mm_set_pi8(a, b, c, d, e, f, g, h);
}
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;
}
void pix_background :: processGrayMMX(imageStruct &image){
int i;
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);
}
m_reset=0;
if(m_Yrange==0)return;
__m64*npixes=(__m64*)image.data;
__m64*opixes=(__m64*)m_savedImage.data;
__m64 newpix, oldpix, m1;
unsigned char thresh=m_Yrange-1;
__m64 thresh8=_mm_set_pi8(thresh,thresh,thresh,thresh,
thresh,thresh,thresh,thresh);
i=pixsize/sizeof(__m64)+(pixsize%sizeof(__m64)!=0);
while(i--){
newpix=npixes[i];
oldpix=opixes[i];
m1 = _mm_subs_pu8 (newpix, oldpix);
oldpix= _mm_subs_pu8 (oldpix, newpix);
m1 = _mm_or_si64 (m1, oldpix); // |oldpix-newpix|
m1 = _mm_subs_pu8 (m1, thresh8);
m1 = _mm_cmpgt_pi8(m1, _mm_setzero_si64()); // |oldpix-newpix|>thresh8
npixes[i] = _mm_and_si64(m1, newpix);
}
_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();
}
void uyvy_to_yuv422(int width, int height, int shift_picture_down, const uint8_t *input, uint8_t *output)
{
__m64 chroma_mask = _mm_set_pi8(255, 0, 255, 0, 255, 0, 255, 0);
__m64 luma_mask = _mm_set_pi8(0, 255, 0, 255, 0, 255, 0, 255);
const uint8_t *orig_input = input;
uint8_t *y_comp = output;
uint8_t *u_comp = output + width * height;
uint8_t *v_comp = u_comp + (int)((width * height)/2); // 4:2:2
int i, j;
// When preparing video for PAL DV50 encoding, the video must be shifted
// down by one line to change the field order to be bottom-field-first
int start_line = 0;
if (shift_picture_down) {
memset(y_comp, 0x10, width); // write one line of black Y
y_comp += width;
memset(u_comp, 0x80, width/2); // write one line of black U,V
u_comp += width/2;
memset(v_comp, 0x80, width/2); // write one line of black U,V
v_comp += width/2;
start_line = 1;
}
/* Do the y component */
for (j = start_line; j < height; j++)
{
// Consume 16 bytes of UYVY data per iteration (8 pixels worth)
for (i = 0; i < width*2; i += 16)
{
//__m64 m1 = _mm_and_si64 (*(__m64 *)input, luma_mask);
//__m64 m2 = _mm_and_si64 (*(__m64 *)(input+8), luma_mask);
//__m64 m2 = _mm_set_pi8 (0, 0, 0, 0, 0, 0, 0, 0);
//*(__m64 *)y_comp = _mm_packs_pu16 (m2, m1);
__m64 m0 = *(__m64 *)input;
__m64 m2 = _mm_srli_si64(m0, 8);
__m64 m3 = _mm_slli_si64(m0, 8);
m3 = _mm_and_si64 (m3, chroma_mask);
m2 = _mm_and_si64 (m2, luma_mask);
m2 = _mm_or_si64 (m2, m3);
m2= _mm_and_si64 (m2, luma_mask);
m0 = m2;
__m64 m1 = *(__m64 *)(input+8);
m2 = _mm_srli_si64(m1, 8);
m3 = _mm_slli_si64(m1, 8);
m3 = _mm_and_si64 (m3, chroma_mask);
m2 = _mm_and_si64 (m2, luma_mask);
m2 = _mm_or_si64 (m2, m3);
m2= _mm_and_si64 (m2, luma_mask);
m1 = m2;
*(__m64 *)y_comp = _mm_packs_pu16 (m0, m1);
y_comp += 8;
input += 16;
}
}
/* Do the chroma components */
input = orig_input;
for (j = start_line; j < height; j++)
{
/* Process every line for yuv 4:2:2 */
for (i = 0; i < width*2; i += 16)
{
__m64 m1 = _mm_unpacklo_pi8 (*(__m64 *)input, *(__m64 *)(input+8));
__m64 m2 = _mm_unpackhi_pi8 (*(__m64 *)input, *(__m64 *)(input+8));
__m64 m3 = _mm_unpacklo_pi8 (m1, m2);
__m64 m4 = _mm_unpackhi_pi8 (m1, m2);
//*(__m64 *)u_comp = _mm_unpacklo_pi8 (m1, m2);
//*(__m64 *)v_comp = _mm_unpackhi_pi8 (m1, m2);
memcpy (u_comp, &m3, 4);
memcpy (v_comp, &m4, 4);
u_comp += 4;
v_comp += 4;
input += 16;
}
}
_mm_empty(); // Clear aliased fp register state
}
void test_char_to_float(void)
{
__m64 narrow = _mm_set_pi8(0, 1, 2, 3, 252, 253, 254, 255);
__m64 zero = _mm_setzero_si64();
__m64 loshorts, hishorts;
__m128 lofloats, hifloats;
float lofloatarray[FLOAT_ARRAYSIZE] __attribute__ ((aligned (16)));
float hifloatarray[FLOAT_ARRAYSIZE] __attribute__ ((aligned (16)));
int16_t shortarray[SHORT_ARRAYSIZE] __attribute__ ((aligned (16)));
int i;
/* interleave zero with narrow and return halves: essentially widening */
/* elements from unsigned chars to unsigned shorts */
loshorts = _mm_unpacklo_pi8(narrow, zero);
hishorts = _mm_unpackhi_pi8(narrow, zero);
/* now turn the 4 shorts in loshorts into floats and store them in lofloats */
/* likewise hishorts into hifloats. */
/* bug in _mm_cvtpi16_ps ? */
lofloats = _mm_cvtpu16_ps1(loshorts);
hifloats = _mm_cvtpu16_ps1(hishorts);
_mm_store_ps(lofloatarray, lofloats);
_mm_store_ps(hifloatarray, hifloats);
/* we used SSE1 instructions that used __m64: add an _mm_empty */
_mm_empty();
/* now store loshorts in shortarray and lofloats into lofloatarray */
/* and print them. */
memcpy(shortarray, &loshorts, sizeof(shortarray));
fprintf(stderr, "loshorts ");
for(i= 0; i < SHORT_ARRAYSIZE; i++)
{
fprintf(stderr, "%d ", shortarray[i]);
}
fprintf(stderr, "\n");
fprintf(stderr, "lofloats ");
for(i= 0; i < FLOAT_ARRAYSIZE; i++)
{
fprintf(stderr, "%f ", lofloatarray[i]);
}
fprintf(stderr, "\n");
/* now store hishorts in shortarray and hifloats into lofloatarray */
/* and print them. */
memcpy(shortarray, &hishorts, sizeof(shortarray));
fprintf(stderr, "hishorts ");
for(i= 0; i < SHORT_ARRAYSIZE; i++)
{
fprintf(stderr, "%d ", shortarray[i]);
}
fprintf(stderr, "\n");
fprintf(stderr, "hifloats ");
for(i= 0; i < FLOAT_ARRAYSIZE; i++)
{
fprintf(stderr, "%f ", hifloatarray[i]);
}
fprintf(stderr, "\n");
}
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;
}
