falkTX
/
FFmpeg
mirror of https://github.com/falkTX/FFmpeg.git
1
0
Fork 0
Code Issues 0 Releases 338 Wiki Activity
Browse Source

Merge commit '022184a646896c6b27c6cca387039b030685266e' 

* commit '022184a646896c6b27c6cca387039b030685266e':
  ppc: dsputil: more K&R formatting cosmetics

Conflicts:
	libavcodec/ppc/gmc_altivec.c

Merged-by: Michael Niedermayer <michaelni@gmx.at>
tags/n2.3
Michael Niedermayer 12 years ago
parent
13405ecdfe 022184a646
commit
adf82fe4f7
4 changed files with 437 additions and 416 deletions
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. +191
    -193
      libavcodec/ppc/fdct_altivec.c
  2. +42
    -41
      libavcodec/ppc/gmc_altivec.c
  3. +154
    -141
      libavcodec/ppc/idct_altivec.c
  4. +50
    -41
      libavcodec/ppc/int_altivec.c

+ 191
- 193
libavcodec/ppc/fdct_altivec.c View File

@@ -22,39 +22,37 @@
#if HAVE_ALTIVEC_H
#include <altivec.h>
#endif

#include "libavutil/common.h"
#include "dsputil_altivec.h"

#define vs16(v) ((vector signed short)(v))
#define vs32(v) ((vector signed int)(v))
#define vu8(v) ((vector unsigned char)(v))
#define vu16(v) ((vector unsigned short)(v))
#define vu32(v) ((vector unsigned int)(v))


#define C1 0.98078525066375732421875000 /* cos(1*PI/16) */
#define C2 0.92387950420379638671875000 /* cos(2*PI/16) */
#define C3 0.83146959543228149414062500 /* cos(3*PI/16) */
#define C4 0.70710676908493041992187500 /* cos(4*PI/16) */
#define C5 0.55557024478912353515625000 /* cos(5*PI/16) */
#define C6 0.38268342614173889160156250 /* cos(6*PI/16) */
#define C7 0.19509032368659973144531250 /* cos(7*PI/16) */
#define SQRT_2 1.41421353816986083984375000 /* sqrt(2) */

#define vs16(v) ((vector signed short) (v))
#define vs32(v) ((vector signed int) (v))
#define vu8(v) ((vector unsigned char) (v))
#define vu16(v) ((vector unsigned short) (v))
#define vu32(v) ((vector unsigned int) (v))

#define C1 0.98078525066375732421875000 /* cos(1 * PI / 16) */
#define C2 0.92387950420379638671875000 /* cos(2 * PI / 16) */
#define C3 0.83146959543228149414062500 /* cos(3 * PI / 16) */
#define C4 0.70710676908493041992187500 /* cos(4 * PI / 16) */
#define C5 0.55557024478912353515625000 /* cos(5 * PI / 16) */
#define C6 0.38268342614173889160156250 /* cos(6 * PI / 16) */
#define C7 0.19509032368659973144531250 /* cos(7 * PI / 16) */
#define SQRT_2 1.41421353816986083984375000 /* sqrt(2) */

#define W0 -(2 * C2)
#define W1 (2 * C6)
#define W1 (2 * C6)
#define W2 (SQRT_2 * C6)
#define W3 (SQRT_2 * C3)
#define W4 (SQRT_2 * (-C1 + C3 + C5 - C7))
#define W5 (SQRT_2 * ( C1 + C3 - C5 + C7))
#define W6 (SQRT_2 * ( C1 + C3 + C5 - C7))
#define W7 (SQRT_2 * ( C1 + C3 - C5 - C7))
#define W8 (SQRT_2 * ( C7 - C3))
#define W5 (SQRT_2 * (C1 + C3 - C5 + C7))
#define W6 (SQRT_2 * (C1 + C3 + C5 - C7))
#define W7 (SQRT_2 * (C1 + C3 - C5 - C7))
#define W8 (SQRT_2 * (C7 - C3))
#define W9 (SQRT_2 * (-C1 - C3))
#define WA (SQRT_2 * (-C3 - C5))
#define WB (SQRT_2 * ( C5 - C3))

#define WB (SQRT_2 * (C5 - C3))

static vector float fdctconsts[3] = {
{ W0, W1, W2, W3 },
@@ -75,8 +73,7 @@ static vector float fdctconsts[3] = {
#define LD_WA vec_splat(cnsts2, 2)
#define LD_WB vec_splat(cnsts2, 3)


#define FDCTROW(b0,b1,b2,b3,b4,b5,b6,b7) /* {{{ */ \
#define FDCTROW(b0, b1, b2, b3, b4, b5, b6, b7) /* {{{ */ \
x0 = vec_add(b0, b7); /* x0 = b0 + b7; */ \
x7 = vec_sub(b0, b7); /* x7 = b0 - b7; */ \
x1 = vec_add(b1, b6); /* x1 = b1 + b6; */ \
@@ -91,41 +88,41 @@ static vector float fdctconsts[3] = {
b0 = vec_add(b7, b1); /* b0 = b7 + b1; */ \
b4 = vec_sub(b7, b1); /* b4 = b7 - b1; */ \
\
b2 = vec_sub(x0, x3); /* b2 = x0 - x3; */ \
b6 = vec_sub(x1, x2); /* b6 = x1 - x2; */ \
b5 = vec_add(b6, b2); /* b5 = b6 + b2; */ \
b2 = vec_sub(x0, x3); /* b2 = x0 - x3; */ \
b6 = vec_sub(x1, x2); /* b6 = x1 - x2; */ \
b5 = vec_add(b6, b2); /* b5 = b6 + b2; */ \
cnst = LD_W2; \
b5 = vec_madd(cnst, b5, mzero); /* b5 = b5 * W2; */ \
b5 = vec_madd(cnst, b5, mzero); /* b5 = b5 * W2; */ \
cnst = LD_W1; \
b2 = vec_madd(cnst, b2, b5); /* b2 = b5 + b2 * W1; */ \
b2 = vec_madd(cnst, b2, b5); /* b2 = b5 + b2 * W1; */ \
cnst = LD_W0; \
b6 = vec_madd(cnst, b6, b5); /* b6 = b5 + b6 * W0; */ \
b6 = vec_madd(cnst, b6, b5); /* b6 = b5 + b6 * W0; */ \
\
x0 = vec_add(x4, x7); /* x0 = x4 + x7; */ \
x1 = vec_add(x5, x6); /* x1 = x5 + x6; */ \
x2 = vec_add(x4, x6); /* x2 = x4 + x6; */ \
x3 = vec_add(x5, x7); /* x3 = x5 + x7; */ \
x8 = vec_add(x2, x3); /* x8 = x2 + x3; */ \
x0 = vec_add(x4, x7); /* x0 = x4 + x7; */ \
x1 = vec_add(x5, x6); /* x1 = x5 + x6; */ \
x2 = vec_add(x4, x6); /* x2 = x4 + x6; */ \
x3 = vec_add(x5, x7); /* x3 = x5 + x7; */ \
x8 = vec_add(x2, x3); /* x8 = x2 + x3; */ \
cnst = LD_W3; \
x8 = vec_madd(cnst, x8, mzero); /* x8 = x8 * W3; */ \
x8 = vec_madd(cnst, x8, mzero); /* x8 = x8 * W3; */ \
\
cnst = LD_W8; \
x0 = vec_madd(cnst, x0, mzero); /* x0 *= W8; */ \
x0 = vec_madd(cnst, x0, mzero); /* x0 *= W8; */ \
cnst = LD_W9; \
x1 = vec_madd(cnst, x1, mzero); /* x1 *= W9; */ \
x1 = vec_madd(cnst, x1, mzero); /* x1 *= W9; */ \
cnst = LD_WA; \
x2 = vec_madd(cnst, x2, x8); /* x2 = x2 * WA + x8; */ \
x2 = vec_madd(cnst, x2, x8); /* x2 = x2 * WA + x8; */ \
cnst = LD_WB; \
x3 = vec_madd(cnst, x3, x8); /* x3 = x3 * WB + x8; */ \
x3 = vec_madd(cnst, x3, x8); /* x3 = x3 * WB + x8; */ \
\
cnst = LD_W4; \
b7 = vec_madd(cnst, x4, x0); /* b7 = x4 * W4 + x0; */ \
b7 = vec_madd(cnst, x4, x0); /* b7 = x4 * W4 + x0; */ \
cnst = LD_W5; \
b5 = vec_madd(cnst, x5, x1); /* b5 = x5 * W5 + x1; */ \
b5 = vec_madd(cnst, x5, x1); /* b5 = x5 * W5 + x1; */ \
cnst = LD_W6; \
b3 = vec_madd(cnst, x6, x1); /* b3 = x6 * W6 + x1; */ \
b3 = vec_madd(cnst, x6, x1); /* b3 = x6 * W6 + x1; */ \
cnst = LD_W7; \
b1 = vec_madd(cnst, x7, x0); /* b1 = x7 * W7 + x0; */ \
b1 = vec_madd(cnst, x7, x0); /* b1 = x7 * W7 + x0; */ \
\
b7 = vec_add(b7, x2); /* b7 = b7 + x2; */ \
b5 = vec_add(b5, x3); /* b5 = b5 + x3; */ \
@@ -133,7 +130,7 @@ static vector float fdctconsts[3] = {
b1 = vec_add(b1, x3); /* b1 = b1 + x3; */ \
/* }}} */

#define FDCTCOL(b0,b1,b2,b3,b4,b5,b6,b7) /* {{{ */ \
#define FDCTCOL(b0, b1, b2, b3, b4, b5, b6, b7) /* {{{ */ \
x0 = vec_add(b0, b7); /* x0 = b0 + b7; */ \
x7 = vec_sub(b0, b7); /* x7 = b0 - b7; */ \
x1 = vec_add(b1, b6); /* x1 = b1 + b6; */ \
@@ -148,41 +145,41 @@ static vector float fdctconsts[3] = {
b0 = vec_add(b7, b1); /* b0 = b7 + b1; */ \
b4 = vec_sub(b7, b1); /* b4 = b7 - b1; */ \
\
b2 = vec_sub(x0, x3); /* b2 = x0 - x3; */ \
b6 = vec_sub(x1, x2); /* b6 = x1 - x2; */ \
b5 = vec_add(b6, b2); /* b5 = b6 + b2; */ \
b2 = vec_sub(x0, x3); /* b2 = x0 - x3; */ \
b6 = vec_sub(x1, x2); /* b6 = x1 - x2; */ \
b5 = vec_add(b6, b2); /* b5 = b6 + b2; */ \
cnst = LD_W2; \
b5 = vec_madd(cnst, b5, mzero); /* b5 = b5 * W2; */ \
b5 = vec_madd(cnst, b5, mzero); /* b5 = b5 * W2; */ \
cnst = LD_W1; \
b2 = vec_madd(cnst, b2, b5); /* b2 = b5 + b2 * W1; */ \
b2 = vec_madd(cnst, b2, b5); /* b2 = b5 + b2 * W1; */ \
cnst = LD_W0; \
b6 = vec_madd(cnst, b6, b5); /* b6 = b5 + b6 * W0; */ \
b6 = vec_madd(cnst, b6, b5); /* b6 = b5 + b6 * W0; */ \
\
x0 = vec_add(x4, x7); /* x0 = x4 + x7; */ \
x1 = vec_add(x5, x6); /* x1 = x5 + x6; */ \
x2 = vec_add(x4, x6); /* x2 = x4 + x6; */ \
x3 = vec_add(x5, x7); /* x3 = x5 + x7; */ \
x8 = vec_add(x2, x3); /* x8 = x2 + x3; */ \
x0 = vec_add(x4, x7); /* x0 = x4 + x7; */ \
x1 = vec_add(x5, x6); /* x1 = x5 + x6; */ \
x2 = vec_add(x4, x6); /* x2 = x4 + x6; */ \
x3 = vec_add(x5, x7); /* x3 = x5 + x7; */ \
x8 = vec_add(x2, x3); /* x8 = x2 + x3; */ \
cnst = LD_W3; \
x8 = vec_madd(cnst, x8, mzero); /* x8 = x8 * W3; */ \
x8 = vec_madd(cnst, x8, mzero); /* x8 = x8 * W3; */ \
\
cnst = LD_W8; \
x0 = vec_madd(cnst, x0, mzero); /* x0 *= W8; */ \
x0 = vec_madd(cnst, x0, mzero); /* x0 *= W8; */ \
cnst = LD_W9; \
x1 = vec_madd(cnst, x1, mzero); /* x1 *= W9; */ \
x1 = vec_madd(cnst, x1, mzero); /* x1 *= W9; */ \
cnst = LD_WA; \
x2 = vec_madd(cnst, x2, x8); /* x2 = x2 * WA + x8; */ \
x2 = vec_madd(cnst, x2, x8); /* x2 = x2 * WA + x8; */ \
cnst = LD_WB; \
x3 = vec_madd(cnst, x3, x8); /* x3 = x3 * WB + x8; */ \
x3 = vec_madd(cnst, x3, x8); /* x3 = x3 * WB + x8; */ \
\
cnst = LD_W4; \
b7 = vec_madd(cnst, x4, x0); /* b7 = x4 * W4 + x0; */ \
b7 = vec_madd(cnst, x4, x0); /* b7 = x4 * W4 + x0; */ \
cnst = LD_W5; \
b5 = vec_madd(cnst, x5, x1); /* b5 = x5 * W5 + x1; */ \
b5 = vec_madd(cnst, x5, x1); /* b5 = x5 * W5 + x1; */ \
cnst = LD_W6; \
b3 = vec_madd(cnst, x6, x1); /* b3 = x6 * W6 + x1; */ \
b3 = vec_madd(cnst, x6, x1); /* b3 = x6 * W6 + x1; */ \
cnst = LD_W7; \
b1 = vec_madd(cnst, x7, x0); /* b1 = x7 * W7 + x0; */ \
b1 = vec_madd(cnst, x7, x0); /* b1 = x7 * W7 + x0; */ \
\
b7 = vec_add(b7, x2); /* b7 += x2; */ \
b5 = vec_add(b5, x3); /* b5 += x3; */ \
@@ -190,10 +187,7 @@ static vector float fdctconsts[3] = {
b1 = vec_add(b1, x3); /* b1 += x3; */ \
/* }}} */



/* two dimensional discrete cosine transform */

void ff_fdct_altivec(int16_t *block)
{
vector signed short *bp;
@@ -205,56 +199,57 @@ void ff_fdct_altivec(int16_t *block)

/* setup constants {{{ */
/* mzero = -0.0 */
mzero = ((vector float)vec_splat_u32(-1));
mzero = ((vector float)vec_sl(vu32(mzero), vu32(mzero)));
cp = fdctconsts;
cnsts0 = vec_ld(0, cp); cp++;
cnsts1 = vec_ld(0, cp); cp++;
mzero = ((vector float) vec_splat_u32(-1));
mzero = ((vector float) vec_sl(vu32(mzero), vu32(mzero)));
cp = fdctconsts;
cnsts0 = vec_ld(0, cp);
cp++;
cnsts1 = vec_ld(0, cp);
cp++;
cnsts2 = vec_ld(0, cp);
/* }}} */


/* 8x8 matrix transpose (vector short[8]) {{{ */
#define MERGE_S16(hl,a,b) vec_merge##hl(vs16(a), vs16(b))
#define MERGE_S16(hl, a, b) vec_merge ## hl(vs16(a), vs16(b))

bp = (vector signed short*)block;
b00 = ((vector float)vec_ld(0, bp));
b40 = ((vector float)vec_ld(16*4, bp));
b01 = ((vector float)MERGE_S16(h, b00, b40));
b11 = ((vector float)MERGE_S16(l, b00, b40));
bp = (vector signed short *) block;
b00 = ((vector float) vec_ld(0, bp));
b40 = ((vector float) vec_ld(16 * 4, bp));
b01 = ((vector float) MERGE_S16(h, b00, b40));
b11 = ((vector float) MERGE_S16(l, b00, b40));
bp++;
b10 = ((vector float)vec_ld(0, bp));
b50 = ((vector float)vec_ld(16*4, bp));
b21 = ((vector float)MERGE_S16(h, b10, b50));
b31 = ((vector float)MERGE_S16(l, b10, b50));
b10 = ((vector float) vec_ld(0, bp));
b50 = ((vector float) vec_ld(16 * 4, bp));
b21 = ((vector float) MERGE_S16(h, b10, b50));
b31 = ((vector float) MERGE_S16(l, b10, b50));
bp++;
b20 = ((vector float)vec_ld(0, bp));
b60 = ((vector float)vec_ld(16*4, bp));
b41 = ((vector float)MERGE_S16(h, b20, b60));
b51 = ((vector float)MERGE_S16(l, b20, b60));
b20 = ((vector float) vec_ld(0, bp));
b60 = ((vector float) vec_ld(16 * 4, bp));
b41 = ((vector float) MERGE_S16(h, b20, b60));
b51 = ((vector float) MERGE_S16(l, b20, b60));
bp++;
b30 = ((vector float)vec_ld(0, bp));
b70 = ((vector float)vec_ld(16*4, bp));
b61 = ((vector float)MERGE_S16(h, b30, b70));
b71 = ((vector float)MERGE_S16(l, b30, b70));
x0 = ((vector float)MERGE_S16(h, b01, b41));
x1 = ((vector float)MERGE_S16(l, b01, b41));
x2 = ((vector float)MERGE_S16(h, b11, b51));
x3 = ((vector float)MERGE_S16(l, b11, b51));
x4 = ((vector float)MERGE_S16(h, b21, b61));
x5 = ((vector float)MERGE_S16(l, b21, b61));
x6 = ((vector float)MERGE_S16(h, b31, b71));
x7 = ((vector float)MERGE_S16(l, b31, b71));
b00 = ((vector float)MERGE_S16(h, x0, x4));
b10 = ((vector float)MERGE_S16(l, x0, x4));
b20 = ((vector float)MERGE_S16(h, x1, x5));
b30 = ((vector float)MERGE_S16(l, x1, x5));
b40 = ((vector float)MERGE_S16(h, x2, x6));
b50 = ((vector float)MERGE_S16(l, x2, x6));
b60 = ((vector float)MERGE_S16(h, x3, x7));
b70 = ((vector float)MERGE_S16(l, x3, x7));
b30 = ((vector float) vec_ld(0, bp));
b70 = ((vector float) vec_ld(16 * 4, bp));
b61 = ((vector float) MERGE_S16(h, b30, b70));
b71 = ((vector float) MERGE_S16(l, b30, b70));
x0 = ((vector float) MERGE_S16(h, b01, b41));
x1 = ((vector float) MERGE_S16(l, b01, b41));
x2 = ((vector float) MERGE_S16(h, b11, b51));
x3 = ((vector float) MERGE_S16(l, b11, b51));
x4 = ((vector float) MERGE_S16(h, b21, b61));
x5 = ((vector float) MERGE_S16(l, b21, b61));
x6 = ((vector float) MERGE_S16(h, b31, b71));
x7 = ((vector float) MERGE_S16(l, b31, b71));
b00 = ((vector float) MERGE_S16(h, x0, x4));
b10 = ((vector float) MERGE_S16(l, x0, x4));
b20 = ((vector float) MERGE_S16(h, x1, x5));
b30 = ((vector float) MERGE_S16(l, x1, x5));
b40 = ((vector float) MERGE_S16(h, x2, x6));
b50 = ((vector float) MERGE_S16(l, x2, x6));
b60 = ((vector float) MERGE_S16(h, x3, x7));
b70 = ((vector float) MERGE_S16(l, x3, x7));

#undef MERGE_S16
/* }}} */
@@ -264,32 +259,32 @@ void ff_fdct_altivec(int16_t *block)
* takes advantage of this. */

/* fdct rows {{{ */
x0 = ((vector float)vec_add(vs16(b00), vs16(b70)));
x7 = ((vector float)vec_sub(vs16(b00), vs16(b70)));
x1 = ((vector float)vec_add(vs16(b10), vs16(b60)));
x6 = ((vector float)vec_sub(vs16(b10), vs16(b60)));
x2 = ((vector float)vec_add(vs16(b20), vs16(b50)));
x5 = ((vector float)vec_sub(vs16(b20), vs16(b50)));
x3 = ((vector float)vec_add(vs16(b30), vs16(b40)));
x4 = ((vector float)vec_sub(vs16(b30), vs16(b40)));
b70 = ((vector float)vec_add(vs16(x0), vs16(x3)));
b10 = ((vector float)vec_add(vs16(x1), vs16(x2)));
b00 = ((vector float)vec_add(vs16(b70), vs16(b10)));
b40 = ((vector float)vec_sub(vs16(b70), vs16(b10)));
#define CTF0(n) \
b##n##1 = ((vector float)vec_unpackl(vs16(b##n##0))); \
b##n##0 = ((vector float)vec_unpackh(vs16(b##n##0))); \
b##n##1 = vec_ctf(vs32(b##n##1), 0); \
b##n##0 = vec_ctf(vs32(b##n##0), 0);
x0 = ((vector float) vec_add(vs16(b00), vs16(b70)));
x7 = ((vector float) vec_sub(vs16(b00), vs16(b70)));
x1 = ((vector float) vec_add(vs16(b10), vs16(b60)));
x6 = ((vector float) vec_sub(vs16(b10), vs16(b60)));
x2 = ((vector float) vec_add(vs16(b20), vs16(b50)));
x5 = ((vector float) vec_sub(vs16(b20), vs16(b50)));
x3 = ((vector float) vec_add(vs16(b30), vs16(b40)));
x4 = ((vector float) vec_sub(vs16(b30), vs16(b40)));
b70 = ((vector float) vec_add(vs16(x0), vs16(x3)));
b10 = ((vector float) vec_add(vs16(x1), vs16(x2)));
b00 = ((vector float) vec_add(vs16(b70), vs16(b10)));
b40 = ((vector float) vec_sub(vs16(b70), vs16(b10)));
#define CTF0(n) \
b ## n ## 1 = ((vector float) vec_unpackl(vs16(b ## n ## 0))); \
b ## n ## 0 = ((vector float) vec_unpackh(vs16(b ## n ## 0))); \
b ## n ## 1 = vec_ctf(vs32(b ## n ## 1), 0); \
b ## n ## 0 = vec_ctf(vs32(b ## n ## 0), 0);

CTF0(0);
CTF0(4);

b20 = ((vector float)vec_sub(vs16(x0), vs16(x3)));
b60 = ((vector float)vec_sub(vs16(x1), vs16(x2)));
b20 = ((vector float) vec_sub(vs16(x0), vs16(x3)));
b60 = ((vector float) vec_sub(vs16(x1), vs16(x2)));

CTF0(2);
CTF0(6);
@@ -300,20 +295,20 @@ void ff_fdct_altivec(int16_t *block)
x1 = vec_add(b61, b21);

cnst = LD_W2;
x0 = vec_madd(cnst, x0, mzero);
x1 = vec_madd(cnst, x1, mzero);
x0 = vec_madd(cnst, x0, mzero);
x1 = vec_madd(cnst, x1, mzero);
cnst = LD_W1;
b20 = vec_madd(cnst, b20, x0);
b21 = vec_madd(cnst, b21, x1);
b20 = vec_madd(cnst, b20, x0);
b21 = vec_madd(cnst, b21, x1);
cnst = LD_W0;
b60 = vec_madd(cnst, b60, x0);
b61 = vec_madd(cnst, b61, x1);
b60 = vec_madd(cnst, b60, x0);
b61 = vec_madd(cnst, b61, x1);

#define CTFX(x,b) \
b##0 = ((vector float)vec_unpackh(vs16(x))); \
b##1 = ((vector float)vec_unpackl(vs16(x))); \
b##0 = vec_ctf(vs32(b##0), 0); \
b##1 = vec_ctf(vs32(b##1), 0); \
#define CTFX(x, b) \
b ## 0 = ((vector float) vec_unpackh(vs16(x))); \
b ## 1 = ((vector float) vec_unpackl(vs16(x))); \
b ## 0 = vec_ctf(vs32(b ## 0), 0); \
b ## 1 = vec_ctf(vs32(b ## 1), 0); \

CTFX(x4, b7);
CTFX(x5, b5);
@@ -322,64 +317,62 @@ void ff_fdct_altivec(int16_t *block)

#undef CTFX


x0 = vec_add(b70, b10);
x1 = vec_add(b50, b30);
x2 = vec_add(b70, b30);
x3 = vec_add(b50, b10);
x8 = vec_add(x2, x3);
x0 = vec_add(b70, b10);
x1 = vec_add(b50, b30);
x2 = vec_add(b70, b30);
x3 = vec_add(b50, b10);
x8 = vec_add(x2, x3);
cnst = LD_W3;
x8 = vec_madd(cnst, x8, mzero);
x8 = vec_madd(cnst, x8, mzero);

cnst = LD_W8;
x0 = vec_madd(cnst, x0, mzero);
x0 = vec_madd(cnst, x0, mzero);
cnst = LD_W9;
x1 = vec_madd(cnst, x1, mzero);
x1 = vec_madd(cnst, x1, mzero);
cnst = LD_WA;
x2 = vec_madd(cnst, x2, x8);
x2 = vec_madd(cnst, x2, x8);
cnst = LD_WB;
x3 = vec_madd(cnst, x3, x8);
x3 = vec_madd(cnst, x3, x8);

cnst = LD_W4;
b70 = vec_madd(cnst, b70, x0);
b70 = vec_madd(cnst, b70, x0);
cnst = LD_W5;
b50 = vec_madd(cnst, b50, x1);
b50 = vec_madd(cnst, b50, x1);
cnst = LD_W6;
b30 = vec_madd(cnst, b30, x1);
b30 = vec_madd(cnst, b30, x1);
cnst = LD_W7;
b10 = vec_madd(cnst, b10, x0);
b10 = vec_madd(cnst, b10, x0);

b70 = vec_add(b70, x2);
b50 = vec_add(b50, x3);
b30 = vec_add(b30, x2);
b10 = vec_add(b10, x3);


x0 = vec_add(b71, b11);
x1 = vec_add(b51, b31);
x2 = vec_add(b71, b31);
x3 = vec_add(b51, b11);
x8 = vec_add(x2, x3);
x0 = vec_add(b71, b11);
x1 = vec_add(b51, b31);
x2 = vec_add(b71, b31);
x3 = vec_add(b51, b11);
x8 = vec_add(x2, x3);
cnst = LD_W3;
x8 = vec_madd(cnst, x8, mzero);
x8 = vec_madd(cnst, x8, mzero);

cnst = LD_W8;
x0 = vec_madd(cnst, x0, mzero);
x0 = vec_madd(cnst, x0, mzero);
cnst = LD_W9;
x1 = vec_madd(cnst, x1, mzero);
x1 = vec_madd(cnst, x1, mzero);
cnst = LD_WA;
x2 = vec_madd(cnst, x2, x8);
x2 = vec_madd(cnst, x2, x8);
cnst = LD_WB;
x3 = vec_madd(cnst, x3, x8);
x3 = vec_madd(cnst, x3, x8);

cnst = LD_W4;
b71 = vec_madd(cnst, b71, x0);
b71 = vec_madd(cnst, b71, x0);
cnst = LD_W5;
b51 = vec_madd(cnst, b51, x1);
b51 = vec_madd(cnst, b51, x1);
cnst = LD_W6;
b31 = vec_madd(cnst, b31, x1);
b31 = vec_madd(cnst, b31, x1);
cnst = LD_W7;
b11 = vec_madd(cnst, b11, x0);
b11 = vec_madd(cnst, b11, x0);

b71 = vec_add(b71, x2);
b51 = vec_add(b51, x3);
@@ -387,7 +380,6 @@ void ff_fdct_altivec(int16_t *block)
b11 = vec_add(b11, x3);
/* }}} */


/* 8x8 matrix transpose (vector float[8][2]) {{{ */
x0 = vec_mergel(b00, b20);
x1 = vec_mergeh(b00, b20);
@@ -430,28 +422,34 @@ void ff_fdct_altivec(int16_t *block)
b31 = vec_mergel(x4, x6);
/* }}} */


FDCTCOL(b00, b10, b20, b30, b40, b50, b60, b70);
FDCTCOL(b01, b11, b21, b31, b41, b51, b61, b71);


/* round, convert back to short {{{ */
#define CTS(n) \
b##n##0 = vec_round(b##n##0); \
b##n##1 = vec_round(b##n##1); \
b##n##0 = ((vector float)vec_cts(b##n##0, 0)); \
b##n##1 = ((vector float)vec_cts(b##n##1, 0)); \
b##n##0 = ((vector float)vec_pack(vs32(b##n##0), vs32(b##n##1))); \
vec_st(vs16(b##n##0), 0, bp);

bp = (vector signed short*)block;
CTS(0); bp++;
CTS(1); bp++;
CTS(2); bp++;
CTS(3); bp++;
CTS(4); bp++;
CTS(5); bp++;
CTS(6); bp++;
#define CTS(n) \
b ## n ## 0 = vec_round(b ## n ## 0); \
b ## n ## 1 = vec_round(b ## n ## 1); \
b ## n ## 0 = ((vector float) vec_cts(b ## n ## 0, 0)); \
b ## n ## 1 = ((vector float) vec_cts(b ## n ## 1, 0)); \
b ## n ## 0 = ((vector float) vec_pack(vs32(b ## n ## 0), \
vs32(b ## n ## 1))); \
vec_st(vs16(b ## n ## 0), 0, bp);

bp = (vector signed short *) block;
CTS(0);
bp++;
CTS(1);
bp++;
CTS(2);
bp++;
CTS(3);
bp++;
CTS(4);
bp++;
CTS(5);
bp++;
CTS(6);
bp++;
CTS(7);

#undef CTS


+ 42
- 41
libavcodec/ppc/gmc_altivec.c View File

@@ -27,32 +27,36 @@

/* AltiVec-enhanced gmc1. ATM this code assumes stride is a multiple of 8
* to preserve proper dst alignment. */
void ff_gmc1_altivec(uint8_t *dst /* align 8 */, uint8_t *src /* align1 */, int stride, int h, int x16, int y16, int rounder)
void ff_gmc1_altivec(uint8_t *dst /* align 8 */, uint8_t *src /* align1 */,
int stride, int h, int x16, int y16, int rounder)
{
const DECLARE_ALIGNED(16, unsigned short, rounder_a) = rounder;
const DECLARE_ALIGNED(16, unsigned short, ABCD)[8] =
{
(16-x16)*(16-y16), /* A */
( x16)*(16-y16), /* B */
(16-x16)*( y16), /* C */
( x16)*( y16), /* D */
0, 0, 0, 0 /* padding */
};
register const vector unsigned char vczero = (const vector unsigned char)vec_splat_u8(0);
register const vector unsigned short vcsr8 = (const vector unsigned short)vec_splat_u16(8);
register vector unsigned char dstv, dstv2, src_0, src_1, srcvA, srcvB, srcvC, srcvD;
register vector unsigned short Av, Bv, Cv, Dv, rounderV, tempA, tempB, tempC, tempD;
const DECLARE_ALIGNED(16, unsigned short, ABCD)[8] = {
(16 - x16) * (16 - y16), /* A */
(x16) * (16 - y16), /* B */
(16 - x16) * (y16), /* C */
(x16) * (y16), /* D */
0, 0, 0, 0 /* padding */
};
register const vector unsigned char vczero =
(const vector unsigned char) vec_splat_u8(0);
register const vector unsigned short vcsr8 =
(const vector unsigned short) vec_splat_u16(8);
register vector unsigned char dstv, dstv2, src_0, src_1,
srcvA, srcvB, srcvC, srcvD;
register vector unsigned short Av, Bv, Cv, Dv, rounderV,
tempA, tempB, tempC, tempD;
int i;
unsigned long dst_odd = (unsigned long)dst & 0x0000000F;
unsigned long src_really_odd = (unsigned long)src & 0x0000000F;
unsigned long dst_odd = (unsigned long) dst & 0x0000000F;
unsigned long src_really_odd = (unsigned long) src & 0x0000000F;

tempA = vec_ld(0, (const unsigned short*)ABCD);
Av = vec_splat(tempA, 0);
Bv = vec_splat(tempA, 1);
Cv = vec_splat(tempA, 2);
Dv = vec_splat(tempA, 3);
tempA = vec_ld(0, (const unsigned short *) ABCD);
Av = vec_splat(tempA, 0);
Bv = vec_splat(tempA, 1);
Cv = vec_splat(tempA, 2);
Dv = vec_splat(tempA, 3);

rounderV = vec_splat((vec_u16)vec_lde(0, &rounder_a), 0);
rounderV = vec_splat((vec_u16) vec_lde(0, &rounder_a), 0);

/* we'll be able to pick-up our 9 char elements at src from those
* 32 bytes we load the first batch here, as inside the loop we can
@@ -61,36 +65,34 @@ void ff_gmc1_altivec(uint8_t *dst /* align 8 */, uint8_t *src /* align1 */, int
src_1 = vec_ld(16, src);
srcvA = vec_perm(src_0, src_1, vec_lvsl(0, src));

if (src_really_odd != 0x0000000F) {
if (src_really_odd != 0x0000000F)
/* If (src & 0xF) == 0xF, then (src + 1) is properly aligned
* on the second vector. */
srcvB = vec_perm(src_0, src_1, vec_lvsl(1, src));
} else {
else
srcvB = src_1;
}
srcvA = vec_mergeh(vczero, srcvA);
srcvB = vec_mergeh(vczero, srcvB);

for(i=0; i<h; i++) {
dst_odd = (unsigned long)dst & 0x0000000F;
src_really_odd = (((unsigned long)src) + stride) & 0x0000000F;
for (i = 0; i < h; i++) {
dst_odd = (unsigned long) dst & 0x0000000F;
src_really_odd = (((unsigned long) src) + stride) & 0x0000000F;

dstv = vec_ld(0, dst);

/* We'll be able to pick-up our 9 char elements at src + stride from
* those 32 bytes then reuse the resulting 2 vectors srvcC and srcvD
* as the next srcvA and srcvB. */
src_0 = vec_ld(stride + 0, src);
src_0 = vec_ld(stride + 0, src);
src_1 = vec_ld(stride + 16, src);
srcvC = vec_perm(src_0, src_1, vec_lvsl(stride + 0, src));

if (src_really_odd != 0x0000000F) {
if (src_really_odd != 0x0000000F)
/* If (src & 0xF) == 0xF, then (src + 1) is properly aligned
* on the second vector. */
srcvD = vec_perm(src_0, src_1, vec_lvsl(stride + 1, src));
} else {
else
srcvD = src_1;
}

srcvC = vec_mergeh(vczero, srcvC);
srcvD = vec_mergeh(vczero, srcvD);
@@ -98,23 +100,22 @@ void ff_gmc1_altivec(uint8_t *dst /* align 8 */, uint8_t *src /* align1 */, int
/* OK, now we (finally) do the math :-)
* Those four instructions replace 32 int muls & 32 int adds.
* Isn't AltiVec nice? */
tempA = vec_mladd((vector unsigned short)srcvA, Av, rounderV);
tempB = vec_mladd((vector unsigned short)srcvB, Bv, tempA);
tempC = vec_mladd((vector unsigned short)srcvC, Cv, tempB);
tempD = vec_mladd((vector unsigned short)srcvD, Dv, tempC);
tempA = vec_mladd((vector unsigned short) srcvA, Av, rounderV);
tempB = vec_mladd((vector unsigned short) srcvB, Bv, tempA);
tempC = vec_mladd((vector unsigned short) srcvC, Cv, tempB);
tempD = vec_mladd((vector unsigned short) srcvD, Dv, tempC);

srcvA = srcvC;
srcvB = srcvD;

tempD = vec_sr(tempD, vcsr8);

dstv2 = vec_pack(tempD, (vector unsigned short)vczero);
dstv2 = vec_pack(tempD, (vector unsigned short) vczero);

if (dst_odd) {
dstv2 = vec_perm(dstv, dstv2, vcprm(0,1,s0,s1));
} else {
dstv2 = vec_perm(dstv, dstv2, vcprm(s0,s1,2,3));
}
if (dst_odd)
dstv2 = vec_perm(dstv, dstv2, vcprm(0, 1, s0, s1));
else
dstv2 = vec_perm(dstv, dstv2, vcprm(s0, s1, 2, 3));

vec_st(dstv2, 0, dst);



+ 154
- 141
libavcodec/ppc/idct_altivec.c View File

@@ -36,147 +36,153 @@
#if HAVE_ALTIVEC_H
#include <altivec.h>
#endif

#include "libavutil/ppc/types_altivec.h"
#include "dsputil_altivec.h"

#define IDCT_HALF \
/* 1st stage */ \
t1 = vec_mradds (a1, vx7, vx1 ); \
t8 = vec_mradds (a1, vx1, vec_subs (zero, vx7)); \
t7 = vec_mradds (a2, vx5, vx3); \
t3 = vec_mradds (ma2, vx3, vx5); \
t1 = vec_mradds(a1, vx7, vx1); \
t8 = vec_mradds(a1, vx1, vec_subs(zero, vx7)); \
t7 = vec_mradds(a2, vx5, vx3); \
t3 = vec_mradds(ma2, vx3, vx5); \
\
/* 2nd stage */ \
t5 = vec_adds (vx0, vx4); \
t0 = vec_subs (vx0, vx4); \
t2 = vec_mradds (a0, vx6, vx2); \
t4 = vec_mradds (a0, vx2, vec_subs (zero, vx6)); \
t6 = vec_adds (t8, t3); \
t3 = vec_subs (t8, t3); \
t8 = vec_subs (t1, t7); \
t1 = vec_adds (t1, t7); \
t5 = vec_adds(vx0, vx4); \
t0 = vec_subs(vx0, vx4); \
t2 = vec_mradds(a0, vx6, vx2); \
t4 = vec_mradds(a0, vx2, vec_subs(zero, vx6)); \
t6 = vec_adds(t8, t3); \
t3 = vec_subs(t8, t3); \
t8 = vec_subs(t1, t7); \
t1 = vec_adds(t1, t7); \
\
/* 3rd stage */ \
t7 = vec_adds (t5, t2); \
t2 = vec_subs (t5, t2); \
t5 = vec_adds (t0, t4); \
t0 = vec_subs (t0, t4); \
t4 = vec_subs (t8, t3); \
t3 = vec_adds (t8, t3); \
t7 = vec_adds(t5, t2); \
t2 = vec_subs(t5, t2); \
t5 = vec_adds(t0, t4); \
t0 = vec_subs(t0, t4); \
t4 = vec_subs(t8, t3); \
t3 = vec_adds(t8, t3); \
\
/* 4th stage */ \
vy0 = vec_adds (t7, t1); \
vy7 = vec_subs (t7, t1); \
vy1 = vec_mradds (c4, t3, t5); \
vy6 = vec_mradds (mc4, t3, t5); \
vy2 = vec_mradds (c4, t4, t0); \
vy5 = vec_mradds (mc4, t4, t0); \
vy3 = vec_adds (t2, t6); \
vy4 = vec_subs (t2, t6);


#define IDCT \
vec_s16 vx0, vx1, vx2, vx3, vx4, vx5, vx6, vx7; \
vec_s16 vy0, vy1, vy2, vy3, vy4, vy5, vy6, vy7; \
vec_s16 a0, a1, a2, ma2, c4, mc4, zero, bias; \
vec_s16 t0, t1, t2, t3, t4, t5, t6, t7, t8; \
vec_u16 shift; \
\
c4 = vec_splat (constants[0], 0); \
a0 = vec_splat (constants[0], 1); \
a1 = vec_splat (constants[0], 2); \
a2 = vec_splat (constants[0], 3); \
mc4 = vec_splat (constants[0], 4); \
ma2 = vec_splat (constants[0], 5); \
bias = (vec_s16)vec_splat ((vec_s32)constants[0], 3); \
\
zero = vec_splat_s16 (0); \
shift = vec_splat_u16 (4); \
\
vx0 = vec_mradds (vec_sl (block[0], shift), constants[1], zero); \
vx1 = vec_mradds (vec_sl (block[1], shift), constants[2], zero); \
vx2 = vec_mradds (vec_sl (block[2], shift), constants[3], zero); \
vx3 = vec_mradds (vec_sl (block[3], shift), constants[4], zero); \
vx4 = vec_mradds (vec_sl (block[4], shift), constants[1], zero); \
vx5 = vec_mradds (vec_sl (block[5], shift), constants[4], zero); \
vx6 = vec_mradds (vec_sl (block[6], shift), constants[3], zero); \
vx7 = vec_mradds (vec_sl (block[7], shift), constants[2], zero); \
\
IDCT_HALF \
\
vx0 = vec_mergeh (vy0, vy4); \
vx1 = vec_mergel (vy0, vy4); \
vx2 = vec_mergeh (vy1, vy5); \
vx3 = vec_mergel (vy1, vy5); \
vx4 = vec_mergeh (vy2, vy6); \
vx5 = vec_mergel (vy2, vy6); \
vx6 = vec_mergeh (vy3, vy7); \
vx7 = vec_mergel (vy3, vy7); \
\
vy0 = vec_mergeh (vx0, vx4); \
vy1 = vec_mergel (vx0, vx4); \
vy2 = vec_mergeh (vx1, vx5); \
vy3 = vec_mergel (vx1, vx5); \
vy4 = vec_mergeh (vx2, vx6); \
vy5 = vec_mergel (vx2, vx6); \
vy6 = vec_mergeh (vx3, vx7); \
vy7 = vec_mergel (vx3, vx7); \
\
vx0 = vec_adds (vec_mergeh (vy0, vy4), bias); \
vx1 = vec_mergel (vy0, vy4); \
vx2 = vec_mergeh (vy1, vy5); \
vx3 = vec_mergel (vy1, vy5); \
vx4 = vec_mergeh (vy2, vy6); \
vx5 = vec_mergel (vy2, vy6); \
vx6 = vec_mergeh (vy3, vy7); \
vx7 = vec_mergel (vy3, vy7); \
\
IDCT_HALF \
\
shift = vec_splat_u16 (6); \
vx0 = vec_sra (vy0, shift); \
vx1 = vec_sra (vy1, shift); \
vx2 = vec_sra (vy2, shift); \
vx3 = vec_sra (vy3, shift); \
vx4 = vec_sra (vy4, shift); \
vx5 = vec_sra (vy5, shift); \
vx6 = vec_sra (vy6, shift); \
vx7 = vec_sra (vy7, shift);

vy0 = vec_adds(t7, t1); \
vy7 = vec_subs(t7, t1); \
vy1 = vec_mradds(c4, t3, t5); \
vy6 = vec_mradds(mc4, t3, t5); \
vy2 = vec_mradds(c4, t4, t0); \
vy5 = vec_mradds(mc4, t4, t0); \
vy3 = vec_adds(t2, t6); \
vy4 = vec_subs(t2, t6);

#define IDCT \
vec_s16 vx0, vx1, vx2, vx3, vx4, vx5, vx6, vx7; \
vec_s16 vy0, vy1, vy2, vy3, vy4, vy5, vy6, vy7; \
vec_s16 a0, a1, a2, ma2, c4, mc4, zero, bias; \
vec_s16 t0, t1, t2, t3, t4, t5, t6, t7, t8; \
vec_u16 shift; \
\
c4 = vec_splat(constants[0], 0); \
a0 = vec_splat(constants[0], 1); \
a1 = vec_splat(constants[0], 2); \
a2 = vec_splat(constants[0], 3); \
mc4 = vec_splat(constants[0], 4); \
ma2 = vec_splat(constants[0], 5); \
bias = (vec_s16) vec_splat((vec_s32) constants[0], 3); \
\
zero = vec_splat_s16(0); \
shift = vec_splat_u16(4); \
\
vx0 = vec_mradds(vec_sl(block[0], shift), constants[1], zero); \
vx1 = vec_mradds(vec_sl(block[1], shift), constants[2], zero); \
vx2 = vec_mradds(vec_sl(block[2], shift), constants[3], zero); \
vx3 = vec_mradds(vec_sl(block[3], shift), constants[4], zero); \
vx4 = vec_mradds(vec_sl(block[4], shift), constants[1], zero); \
vx5 = vec_mradds(vec_sl(block[5], shift), constants[4], zero); \
vx6 = vec_mradds(vec_sl(block[6], shift), constants[3], zero); \
vx7 = vec_mradds(vec_sl(block[7], shift), constants[2], zero); \
\
IDCT_HALF \
\
vx0 = vec_mergeh(vy0, vy4); \
vx1 = vec_mergel(vy0, vy4); \
vx2 = vec_mergeh(vy1, vy5); \
vx3 = vec_mergel(vy1, vy5); \
vx4 = vec_mergeh(vy2, vy6); \
vx5 = vec_mergel(vy2, vy6); \
vx6 = vec_mergeh(vy3, vy7); \
vx7 = vec_mergel(vy3, vy7); \
\
vy0 = vec_mergeh(vx0, vx4); \
vy1 = vec_mergel(vx0, vx4); \
vy2 = vec_mergeh(vx1, vx5); \
vy3 = vec_mergel(vx1, vx5); \
vy4 = vec_mergeh(vx2, vx6); \
vy5 = vec_mergel(vx2, vx6); \
vy6 = vec_mergeh(vx3, vx7); \
vy7 = vec_mergel(vx3, vx7); \
\
vx0 = vec_adds(vec_mergeh(vy0, vy4), bias); \
vx1 = vec_mergel(vy0, vy4); \
vx2 = vec_mergeh(vy1, vy5); \
vx3 = vec_mergel(vy1, vy5); \
vx4 = vec_mergeh(vy2, vy6); \
vx5 = vec_mergel(vy2, vy6); \
vx6 = vec_mergeh(vy3, vy7); \
vx7 = vec_mergel(vy3, vy7); \
\
IDCT_HALF \
\
shift = vec_splat_u16(6); \
vx0 = vec_sra(vy0, shift); \
vx1 = vec_sra(vy1, shift); \
vx2 = vec_sra(vy2, shift); \
vx3 = vec_sra(vy3, shift); \
vx4 = vec_sra(vy4, shift); \
vx5 = vec_sra(vy5, shift); \
vx6 = vec_sra(vy6, shift); \
vx7 = vec_sra(vy7, shift);

static const vec_s16 constants[5] = {
{23170, 13573, 6518, 21895, -23170, -21895, 32, 31},
{16384, 22725, 21407, 19266, 16384, 19266, 21407, 22725},
{22725, 31521, 29692, 26722, 22725, 26722, 29692, 31521},
{21407, 29692, 27969, 25172, 21407, 25172, 27969, 29692},
{19266, 26722, 25172, 22654, 19266, 22654, 25172, 26722}
{ 23170, 13573, 6518, 21895, -23170, -21895, 32, 31 },
{ 16384, 22725, 21407, 19266, 16384, 19266, 21407, 22725 },
{ 22725, 31521, 29692, 26722, 22725, 26722, 29692, 31521 },
{ 21407, 29692, 27969, 25172, 21407, 25172, 27969, 29692 },
{ 19266, 26722, 25172, 22654, 19266, 22654, 25172, 26722 }
};

void ff_idct_put_altivec(uint8_t* dest, int stride, int16_t *blk)
void ff_idct_put_altivec(uint8_t *dest, int stride, int16_t *blk)
{
vec_s16 *block = (vec_s16*)blk;
vec_s16 *block = (vec_s16 *) blk;
vec_u8 tmp;

IDCT

#define COPY(dest,src) \
tmp = vec_packsu (src, src); \
vec_ste ((vec_u32)tmp, 0, (unsigned int *)dest); \
vec_ste ((vec_u32)tmp, 4, (unsigned int *)dest);

COPY (dest, vx0) dest += stride;
COPY (dest, vx1) dest += stride;
COPY (dest, vx2) dest += stride;
COPY (dest, vx3) dest += stride;
COPY (dest, vx4) dest += stride;
COPY (dest, vx5) dest += stride;
COPY (dest, vx6) dest += stride;
COPY (dest, vx7)
#define COPY(dest, src) \
tmp = vec_packsu(src, src); \
vec_ste((vec_u32) tmp, 0, (unsigned int *) dest); \
vec_ste((vec_u32) tmp, 4, (unsigned int *) dest);

COPY(dest, vx0)
dest += stride;
COPY(dest, vx1)
dest += stride;
COPY(dest, vx2)
dest += stride;
COPY(dest, vx3)
dest += stride;
COPY(dest, vx4)
dest += stride;
COPY(dest, vx5)
dest += stride;
COPY(dest, vx6)
dest += stride;
COPY(dest, vx7)
}

void ff_idct_add_altivec(uint8_t* dest, int stride, int16_t *blk)
void ff_idct_add_altivec(uint8_t *dest, int stride, int16_t *blk)
{
vec_s16 *block = (vec_s16*)blk;
vec_s16 *block = (vec_s16 *) blk;
vec_u8 tmp;
vec_s16 tmp2, tmp3;
vec_u8 perm0;
@@ -185,27 +191,34 @@ void ff_idct_add_altivec(uint8_t* dest, int stride, int16_t *blk)

IDCT

p0 = vec_lvsl (0, dest);
p1 = vec_lvsl (stride, dest);
p = vec_splat_u8 (-1);
perm0 = vec_mergeh (p, p0);
perm1 = vec_mergeh (p, p1);

#define ADD(dest,src,perm) \
/* *(uint64_t *)&tmp = *(uint64_t *)dest; */ \
tmp = vec_ld (0, dest); \
tmp2 = (vec_s16)vec_perm (tmp, (vec_u8)zero, perm); \
tmp3 = vec_adds (tmp2, src); \
tmp = vec_packsu (tmp3, tmp3); \
vec_ste ((vec_u32)tmp, 0, (unsigned int *)dest); \
vec_ste ((vec_u32)tmp, 4, (unsigned int *)dest);

ADD (dest, vx0, perm0) dest += stride;
ADD (dest, vx1, perm1) dest += stride;
ADD (dest, vx2, perm0) dest += stride;
ADD (dest, vx3, perm1) dest += stride;
ADD (dest, vx4, perm0) dest += stride;
ADD (dest, vx5, perm1) dest += stride;
ADD (dest, vx6, perm0) dest += stride;
ADD (dest, vx7, perm1)
p0 = vec_lvsl(0, dest);
p1 = vec_lvsl(stride, dest);
p = vec_splat_u8(-1);
perm0 = vec_mergeh(p, p0);
perm1 = vec_mergeh(p, p1);

#define ADD(dest, src, perm) \
/* *(uint64_t *) &tmp = *(uint64_t *) dest; */ \
tmp = vec_ld(0, dest); \
tmp2 = (vec_s16) vec_perm(tmp, (vec_u8) zero, perm); \
tmp3 = vec_adds(tmp2, src); \
tmp = vec_packsu(tmp3, tmp3); \
vec_ste((vec_u32) tmp, 0, (unsigned int *) dest); \
vec_ste((vec_u32) tmp, 4, (unsigned int *) dest);

ADD(dest, vx0, perm0)
dest += stride;
ADD(dest, vx1, perm1)
dest += stride;
ADD(dest, vx2, perm0)
dest += stride;
ADD(dest, vx3, perm1)
dest += stride;
ADD(dest, vx4, perm0)
dest += stride;
ADD(dest, vx5, perm1)
dest += stride;
ADD(dest, vx6, perm0)
dest += stride;
ADD(dest, vx7, perm1)
}

+ 50
- 41
libavcodec/ppc/int_altivec.c View File

@@ -31,27 +31,28 @@
#include "libavutil/attributes.h"
#include "libavutil/ppc/types_altivec.h"
#include "libavcodec/dsputil.h"

#include "dsputil_altivec.h"

static int ssd_int8_vs_int16_altivec(const int8_t *pix1, const int16_t *pix2,
int size) {
int size)
{
int i, size16;
vector signed char vpix1;
vector signed short vpix2, vdiff, vpix1l,vpix1h;
union { vector signed int vscore;
int32_t score[4];
} u;
vector signed short vpix2, vdiff, vpix1l, vpix1h;
union {
vector signed int vscore;
int32_t score[4];
} u;
u.vscore = vec_splat_s32(0);

// XXX lazy way, fix it later

#define vec_unaligned_load(b) \
vec_perm(vec_ld(0,b),vec_ld(15,b),vec_lvsl(0, b));
#define vec_unaligned_load(b) \
vec_perm(vec_ld(0, b), vec_ld(15, b), vec_lvsl(0, b));

size16 = size >> 4;
while(size16) {
// score += (pix1[i]-pix2[i])*(pix1[i]-pix2[i]);
while (size16) {
// score += (pix1[i] - pix2[i]) * (pix1[i] - pix2[i]);
// load pix1 and the first batch of pix2

vpix1 = vec_unaligned_load(pix1);
@@ -62,20 +63,20 @@ static int ssd_int8_vs_int16_altivec(const int8_t *pix1, const int16_t *pix2,
vdiff = vec_sub(vpix1h, vpix2);
vpix1l = vec_unpackl(vpix1);
// load another batch from pix2
vpix2 = vec_unaligned_load(pix2);
vpix2 = vec_unaligned_load(pix2);
u.vscore = vec_msum(vdiff, vdiff, u.vscore);
vdiff = vec_sub(vpix1l, vpix2);
vdiff = vec_sub(vpix1l, vpix2);
u.vscore = vec_msum(vdiff, vdiff, u.vscore);
pix1 += 16;
pix2 += 8;
pix1 += 16;
pix2 += 8;
size16--;
}
u.vscore = vec_sums(u.vscore, vec_splat_s32(0));

size %= 16;
for (i = 0; i < size; i++) {
u.score[3] += (pix1[i]-pix2[i])*(pix1[i]-pix2[i]);
}
for (i = 0; i < size; i++)
u.score[3] += (pix1[i] - pix2[i]) * (pix1[i] - pix2[i]);
return u.score[3];
}

@@ -88,56 +89,64 @@ static int32_t scalarproduct_int16_altivec(const int16_t *v1, const int16_t *v2,
register vec_s32 res = vec_splat_s32(0), t;
int32_t ires;

for(i = 0; i < order; i += 8){
for (i = 0; i < order; i += 8) {
vec1 = vec_unaligned_load(v1);
t = vec_msum(vec1, vec_ld(0, v2), zero_s32v);
res = vec_sums(t, res);
v1 += 8;
v2 += 8;
t = vec_msum(vec1, vec_ld(0, v2), zero_s32v);
res = vec_sums(t, res);
v1 += 8;
v2 += 8;
}
res = vec_splat(res, 3);
vec_ste(res, 0, &ires);

return ires;
}

static int32_t scalarproduct_and_madd_int16_altivec(int16_t *v1, const int16_t *v2, const int16_t *v3, int order, int mul)
static int32_t scalarproduct_and_madd_int16_altivec(int16_t *v1,
const int16_t *v2,
const int16_t *v3,
int order, int mul)
{
LOAD_ZERO;
vec_s16 *pv1 = (vec_s16*)v1;
register vec_s16 muls = {mul,mul,mul,mul,mul,mul,mul,mul};
vec_s16 *pv1 = (vec_s16 *) v1;
register vec_s16 muls = { mul, mul, mul, mul, mul, mul, mul, mul };
register vec_s16 t0, t1, i0, i1, i4;
register vec_s16 i2 = vec_ld(0, v2), i3 = vec_ld(0, v3);
register vec_s32 res = zero_s32v;
register vec_u8 align = vec_lvsl(0, v2);
int32_t ires;

order >>= 4;
do {
i1 = vec_ld(16, v2);
t0 = vec_perm(i2, i1, align);
i2 = vec_ld(32, v2);
t1 = vec_perm(i1, i2, align);
i0 = pv1[0];
i1 = pv1[1];
res = vec_msum(t0, i0, res);
res = vec_msum(t1, i1, res);
i4 = vec_ld(16, v3);
t0 = vec_perm(i3, i4, align);
i3 = vec_ld(32, v3);
t1 = vec_perm(i4, i3, align);
i1 = vec_ld(16, v2);
t0 = vec_perm(i2, i1, align);
i2 = vec_ld(32, v2);
t1 = vec_perm(i1, i2, align);
i0 = pv1[0];
i1 = pv1[1];
res = vec_msum(t0, i0, res);
res = vec_msum(t1, i1, res);
i4 = vec_ld(16, v3);
t0 = vec_perm(i3, i4, align);
i3 = vec_ld(32, v3);
t1 = vec_perm(i4, i3, align);
pv1[0] = vec_mladd(t0, muls, i0);
pv1[1] = vec_mladd(t1, muls, i1);
pv1 += 2;
v2 += 16;
v3 += 16;
} while(--order);
pv1 += 2;
v2 += 16;
v3 += 16;
} while (--order);
res = vec_splat(vec_sums(res, zero_s32v), 3);
vec_ste(res, 0, &ires);

return ires;
}

av_cold void ff_int_init_altivec(DSPContext *c, AVCodecContext *avctx)
{
c->ssd_int8_vs_int16 = ssd_int8_vs_int16_altivec;

c->scalarproduct_int16 = scalarproduct_int16_altivec;

c->scalarproduct_and_madd_int16 = scalarproduct_and_madd_int16_altivec;
}

Write Preview
Loading…
Cancel
Save