[SCM] libav/experimental: altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
siretart at users.alioth.debian.org
siretart at users.alioth.debian.org
Sun Jun 30 15:34:17 UTC 2013
The following commit has been merged in the experimental branch:
commit f2677d6b260dd9a16e19c57fa6009d28175a542c
Author: Brian Foley <bfoley at compsoc.nuigalway.ie>
Date: Tue Nov 19 23:22:29 2002 +0000
altivec optimizations patch by (Brian Foley <bfoley at compsoc dot nuigalway dot ie>)
Originally committed as revision 1240 to svn://svn.ffmpeg.org/ffmpeg/trunk
diff --git a/libavcodec/ppc/dsputil_altivec.c b/libavcodec/ppc/dsputil_altivec.c
index 8a50ccb..1e877d5 100644
--- a/libavcodec/ppc/dsputil_altivec.c
+++ b/libavcodec/ppc/dsputil_altivec.c
@@ -24,6 +24,216 @@
#include <sys/sysctl.h>
#endif
+int pix_abs16x16_x2_altivec(uint8_t *pix1, uint8_t *pix2, int line_size)
+{
+ int s, i;
+ vector unsigned char *tv, zero;
+ vector unsigned char pix1v, pix2v, pix2iv, avgv, t5;
+ vector unsigned int sad;
+ vector signed int sumdiffs;
+
+ s = 0;
+ zero = vec_splat_u8(0);
+ sad = vec_splat_u32(0);
+ for(i=0;i<16;i++) {
+ /*
+ Read unaligned pixels into our vectors. The vectors are as follows:
+ pix1v: pix1[0]-pix1[15]
+ pix2v: pix2[0]-pix2[15] pix2iv: pix2[1]-pix2[16]
+ */
+ tv = (vector unsigned char *) pix1;
+ pix1v = vec_perm(tv[0], tv[1], vec_lvsl(0, pix1));
+
+ tv = (vector unsigned char *) &pix2[0];
+ pix2v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[0]));
+
+ tv = (vector unsigned char *) &pix2[1];
+ pix2iv = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[1]));
+
+ /* Calculate the average vector */
+ avgv = vec_avg(pix2v, pix2iv);
+
+ /* Calculate a sum of abs differences vector */
+ t5 = vec_sub(vec_max(pix1v, avgv), vec_min(pix1v, avgv));
+
+ /* Add each 4 pixel group together and put 4 results into sad */
+ sad = vec_sum4s(t5, sad);
+
+ pix1 += line_size;
+ pix2 += line_size;
+ }
+ /* Sum up the four partial sums, and put the result into s */
+ sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
+ sumdiffs = vec_splat(sumdiffs, 3);
+ vec_ste(sumdiffs, 0, &s);
+
+ return s;
+}
+
+int pix_abs16x16_y2_altivec(uint8_t *pix1, uint8_t *pix2, int line_size)
+{
+ int s, i;
+ vector unsigned char *tv, zero;
+ vector unsigned char pix1v, pix2v, pix3v, avgv, t5;
+ vector unsigned int sad;
+ vector signed int sumdiffs;
+ uint8_t *pix3 = pix2 + line_size;
+
+ s = 0;
+ zero = vec_splat_u8(0);
+ sad = vec_splat_u32(0);
+
+ /*
+ Due to the fact that pix3 = pix2 + line_size, the pix3 of one
+ iteration becomes pix2 in the next iteration. We can use this
+ fact to avoid a potentially expensive unaligned read, each
+ time around the loop.
+ Read unaligned pixels into our vectors. The vectors are as follows:
+ pix2v: pix2[0]-pix2[15]
+ Split the pixel vectors into shorts
+ */
+ tv = (vector unsigned char *) &pix2[0];
+ pix2v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[0]));
+
+ for(i=0;i<16;i++) {
+ /*
+ Read unaligned pixels into our vectors. The vectors are as follows:
+ pix1v: pix1[0]-pix1[15]
+ pix3v: pix3[0]-pix3[15]
+ */
+ tv = (vector unsigned char *) pix1;
+ pix1v = vec_perm(tv[0], tv[1], vec_lvsl(0, pix1));
+
+ tv = (vector unsigned char *) &pix3[0];
+ pix3v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix3[0]));
+
+ /* Calculate the average vector */
+ avgv = vec_avg(pix2v, pix3v);
+
+ /* Calculate a sum of abs differences vector */
+ t5 = vec_sub(vec_max(pix1v, avgv), vec_min(pix1v, avgv));
+
+ /* Add each 4 pixel group together and put 4 results into sad */
+ sad = vec_sum4s(t5, sad);
+
+ pix1 += line_size;
+ pix2v = pix3v;
+ pix3 += line_size;
+
+ }
+
+ /* Sum up the four partial sums, and put the result into s */
+ sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
+ sumdiffs = vec_splat(sumdiffs, 3);
+ vec_ste(sumdiffs, 0, &s);
+ return s;
+}
+
+int pix_abs16x16_xy2_altivec(uint8_t *pix1, uint8_t *pix2, int line_size)
+{
+ int s, i;
+ uint8_t *pix3 = pix2 + line_size;
+ vector unsigned char *tv, avgv, t5, zero;
+ vector unsigned char pix1v, pix2v, pix3v, pix2iv, pix3iv;
+ vector unsigned short pix2lv, pix2hv, pix2ilv, pix2ihv;
+ vector unsigned short pix3lv, pix3hv, pix3ilv, pix3ihv;
+ vector unsigned short avghv, avglv, two, shift_mask;
+ vector unsigned short t1, t2, t3, t4;
+ vector unsigned int sad;
+ vector signed int sumdiffs;
+
+ shift_mask = (vector unsigned short) (0x3fff, 0x3fff, 0x3fff, 0x3fff,
+ 0x3fff, 0x3fff, 0x3fff, 0x3fff);
+ zero = vec_splat_u8(0);
+ two = vec_splat_u16(2);
+ sad = vec_splat_u32(0);
+
+ s = 0;
+
+ /*
+ Due to the fact that pix3 = pix2 + line_size, the pix3 of one
+ iteration becomes pix2 in the next iteration. We can use this
+ fact to avoid a potentially expensive unaligned read, as well
+ as some splitting, and vector addition each time around the loop.
+ Read unaligned pixels into our vectors. The vectors are as follows:
+ pix2v: pix2[0]-pix2[15] pix2iv: pix2[1]-pix2[16]
+ Split the pixel vectors into shorts
+ */
+ tv = (vector unsigned char *) &pix2[0];
+ pix2v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[0]));
+
+ tv = (vector unsigned char *) &pix2[1];
+ pix2iv = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[1]));
+
+ pix2hv = (vector unsigned short) vec_mergeh(zero, pix2v);
+ pix2lv = (vector unsigned short) vec_mergel(zero, pix2v);
+ pix2ihv = (vector unsigned short) vec_mergeh(zero, pix2iv);
+ pix2ilv = (vector unsigned short) vec_mergel(zero, pix2iv);
+ t1 = vec_add(pix2hv, pix2ihv);
+ t2 = vec_add(pix2lv, pix2ilv);
+
+ for(i=0;i<16;i++) {
+ /*
+ Read unaligned pixels into our vectors. The vectors are as follows:
+ pix1v: pix1[0]-pix1[15]
+ pix3v: pix3[0]-pix3[15] pix3iv: pix3[1]-pix3[16]
+ */
+ tv = (vector unsigned char *) pix1;
+ pix1v = vec_perm(tv[0], tv[1], vec_lvsl(0, pix1));
+
+ tv = (vector unsigned char *) &pix3[0];
+ pix3v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix3[0]));
+
+ tv = (vector unsigned char *) &pix3[1];
+ pix3iv = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix3[1]));
+
+ /*
+ Note that Altivec does have vec_avg, but this works on vector pairs
+ and rounds up. We could do avg(avg(a,b),avg(c,d)), but the rounding
+ would mean that, for example, avg(3,0,0,1) = 2, when it should be 1.
+ Instead, we have to split the pixel vectors into vectors of shorts,
+ and do the averaging by hand.
+ */
+
+ /* Split the pixel vectors into shorts */
+ pix3hv = (vector unsigned short) vec_mergeh(zero, pix3v);
+ pix3lv = (vector unsigned short) vec_mergel(zero, pix3v);
+ pix3ihv = (vector unsigned short) vec_mergeh(zero, pix3iv);
+ pix3ilv = (vector unsigned short) vec_mergel(zero, pix3iv);
+
+ /* Do the averaging on them */
+ t3 = vec_add(pix3hv, pix3ihv);
+ t4 = vec_add(pix3lv, pix3ilv);
+
+ avghv = vec_add(vec_add(t1, t3), two);
+ avghv= vec_and(vec_srl(avghv, two), shift_mask);
+
+ avglv = vec_add(vec_add(t2, t4), two);
+ avglv = vec_and(vec_srl(avglv, two), shift_mask);
+
+ /* Pack the shorts back into a result */
+ avgv = vec_pack(avghv, avglv);
+
+ /* Calculate a sum of abs differences vector */
+ t5 = vec_sub(vec_max(pix1v, avgv), vec_min(pix1v, avgv));
+
+ /* Add each 4 pixel group together and put 4 results into sad */
+ sad = vec_sum4s(t5, sad);
+
+ pix1 += line_size;
+ pix3 += line_size;
+ /* Transfer the calculated values for pix3 into pix2 */
+ t1 = t3;
+ t2 = t4;
+ }
+ /* Sum up the four partial sums, and put the result into s */
+ sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
+ sumdiffs = vec_splat(sumdiffs, 3);
+ vec_ste(sumdiffs, 0, &s);
+
+ return s;
+}
+
int pix_abs16x16_altivec(uint8_t *pix1, uint8_t *pix2, int line_size)
{
int i, s;
@@ -108,6 +318,49 @@ int pix_abs8x8_altivec(uint8_t *pix1, uint8_t *pix2, int line_size)
return s;
}
+int pix_norm1_altivec(uint8_t *pix, int line_size)
+{
+ int s, i;
+ vector unsigned char *tv, zero;
+ vector unsigned char pixv;
+ vector unsigned short pixlv, pixhv, zeros;
+ vector unsigned int sv;
+ vector signed int sum;
+ vector unsigned char perm_stoint_h = (vector unsigned char)
+ (16, 16, 0, 1, 16, 16, 2, 3, 16, 16, 4, 5, 16, 16, 6, 7);
+
+ vector unsigned char perm_stoint_l = (vector unsigned char)
+ (16, 16, 8, 9, 16, 16, 10, 11, 16, 16, 12, 13, 16, 16, 14, 15);
+
+ zero = vec_splat_u8(0);
+ zeros = vec_splat_u16(0);
+ sv = vec_splat_u32(0);
+
+ s = 0;
+ for (i = 0; i < 16; i++) {
+ /* Read in the potentially unaligned pixels */
+ tv = (vector unsigned char *) pix;
+ pixv = vec_perm(tv[0], tv[1], vec_lvsl(0, pix));
+
+ /* Split them into two vectors of shorts */
+ pixhv = (vector unsigned short) vec_mergeh(zero, pixv);
+ pixlv = (vector unsigned short) vec_mergel(zero, pixv);
+
+
+ /* Square the values and add them to our sum */
+ sv = vec_msum(pixhv, pixhv, sv);
+ sv = vec_msum(pixlv, pixlv, sv);
+
+ pix += line_size;
+ }
+ /* Sum up the four partial sums, and put the result into s */
+ sum = vec_sums((vector signed int) sv, (vector signed int) zero);
+ sum = vec_splat(sum, 3);
+ vec_ste(sum, 0, &s);
+
+ return s;
+}
+
int pix_sum_altivec(UINT8 * pix, int line_size)
{
diff --git a/libavcodec/ppc/dsputil_altivec.h b/libavcodec/ppc/dsputil_altivec.h
index bdf8f5f..d4d259d 100644
--- a/libavcodec/ppc/dsputil_altivec.h
+++ b/libavcodec/ppc/dsputil_altivec.h
@@ -16,9 +16,13 @@
* License along with this library; if not, write to the Free Software
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
*/
-
+
+extern int pix_abs16x16_x2_altivec(uint8_t *pix1, uint8_t *pix2, int line_size);
+extern int pix_abs16x16_y2_altivec(uint8_t *pix1, uint8_t *pix2, int line_size);
+extern int pix_abs16x16_xy2_altivec(uint8_t *pix1, uint8_t *pix2, int line_size);
extern int pix_abs16x16_altivec(uint8_t *pix1, uint8_t *pix2, int line_size);
extern int pix_abs8x8_altivec(uint8_t *pix1, uint8_t *pix2, int line_size);
+extern int pix_norm1_altivec(uint8_t *pix, int line_size);
extern int pix_sum_altivec(UINT8 * pix, int line_size);
extern void diff_pixels_altivec(DCTELEM* block, const UINT8* s1, const UINT8* s2, int stride);
extern void get_pixels_altivec(DCTELEM* block, const UINT8 * pixels, int line_size);
diff --git a/libavcodec/ppc/dsputil_ppc.c b/libavcodec/ppc/dsputil_ppc.c
index 4be3abb..a515cf3 100644
--- a/libavcodec/ppc/dsputil_ppc.c
+++ b/libavcodec/ppc/dsputil_ppc.c
@@ -32,8 +32,12 @@ void dsputil_init_ppc(DSPContext* c, unsigned mask)
#if HAVE_ALTIVEC
if (has_altivec()) {
// Altivec specific optimisations
+ c->pix_abs16x16_x2 = pix_abs16x16_x2_altivec;
+ c->pix_abs16x16_y2 = pix_abs16x16_y2_altivec;
+ c->pix_abs16x16_xy2 = pix_abs16x16_xy2_altivec;
c->pix_abs16x16 = pix_abs16x16_altivec;
c->pix_abs8x8 = pix_abs8x8_altivec;
+ c->pix_norm1 = pix_norm1_altivec;
c->pix_sum = pix_sum_altivec;
c->diff_pixels = diff_pixels_altivec;
c->get_pixels = get_pixels_altivec;
--
Libav/FFmpeg packaging
More information about the pkg-multimedia-commits
mailing list