Skip to content

F
ffmpeg-mt
Commit c435c416 authored by diego's avatar diego
Browse files
Options

cosmetics: Make libavcodec/ppc/dsputil_altivec.c conform to style guidelines. 

This includes indentation changes, comment reformatting, consistent brace                                                                                                             
placement and some prettyprinting.


git-svn-id: file:///var/local/repositories/ffmpeg/trunk@14318 9553f0bf-9b14-0410-a0b8-cfaf0461ba5b
parent 33297bf0
Show whitespace changes
Inline Side-by-side
Showing with 437 additions and 516 deletions
libavcodec/ppc/dsputil_altivec.c
View file @ c435c416
......@@ -39,12 +39,10 @@ int sad16_x2_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h
s = 0;
sad = (vector unsigned int)vec_splat_u32(0);
for(i=0;i<h;i++) {
/*
Read unaligned pixels into our vectors. The vectors are as follows:
for (i = 0; i < h; 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]
*/
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));
......@@ -88,24 +86,20 @@ int sad16_y2_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h
s = 0;
sad = (vector unsigned int)vec_splat_u32(0);
/*
Due to the fact that pix3 = pix2 + line_size, the pix3 of one
/* 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
*/
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<h;i++) {
/*
Read unaligned pixels into our vectors. The vectors are as follows:
for (i = 0; i < h; i++) {
/* Read unaligned pixels into our vectors. The vectors are as follows:
pix1v: pix1[0]-pix1[15]
pix3v: pix3[0]-pix3[15]
*/
pix3v: pix3[0]-pix3[15] */
tv = (vector unsigned char *) pix1;
pix1v = vec_perm(tv[0], tv[1], vec_lvsl(0, pix1));
......@@ -154,15 +148,13 @@ int sad16_xy2_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int
s = 0;
/*
Due to the fact that pix3 = pix2 + line_size, the pix3 of one
/* 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
*/
Split the pixel vectors into shorts */
tv = (vector unsigned char *) &pix2[0];
pix2v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[0]));
......@@ -176,12 +168,10 @@ int sad16_xy2_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int
t1 = vec_add(pix2hv, pix2ihv);
t2 = vec_add(pix2lv, pix2ilv);
for(i=0;i<h;i++) {
/*
Read unaligned pixels into our vectors. The vectors are as follows:
for (i = 0; i < h; 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]
*/
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));
......@@ -191,13 +181,11 @@ int sad16_xy2_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int
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
/* 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.
*/
and do the averaging by hand. */
/* Split the pixel vectors into shorts */
pix3hv = (vector unsigned short) vec_mergeh(zero, pix3v);
......@@ -248,7 +236,7 @@ int sad16_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
sad = (vector unsigned int)vec_splat_u32(0);
for(i=0;i<h;i++) {
for (i = 0; i < h; i++) {
/* Read potentially unaligned pixels into t1 and t2 */
perm1 = vec_lvsl(0, pix1);
pix1v = (vector unsigned char *) pix1;
......@@ -291,7 +279,7 @@ int sad8_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
permclear = (vector unsigned char)AVV(255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0);
for(i=0;i<h;i++) {
for (i = 0; i < h; i++) {
/* Read potentially unaligned pixels into t1 and t2
Since we're reading 16 pixels, and actually only want 8,
mask out the last 8 pixels. The 0s don't change the sum. */
......@@ -373,7 +361,7 @@ int sse8_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
permclear = (vector unsigned char)AVV(255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0);
for(i=0;i<h;i++) {
for (i = 0; i < h; i++) {
/* Read potentially unaligned pixels into t1 and t2
Since we're reading 16 pixels, and actually only want 8,
mask out the last 8 pixels. The 0s don't change the sum. */
......@@ -384,10 +372,8 @@ int sse8_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
t1 = vec_and(vec_perm(pix1v[0], pix1v[1], perm1), permclear);
t2 = vec_and(vec_perm(pix2v[0], pix2v[1], perm2), permclear);
/*
Since we want to use unsigned chars, we can take advantage
of the fact that abs(a-b)^2 = (a-b)^2.
*/
/* Since we want to use unsigned chars, we can take advantage
of the fact that abs(a-b)^2 = (a-b)^2. */
/* Calculate abs differences vector */
t3 = vec_max(t1, t2);
......@@ -426,7 +412,7 @@ int sse16_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
sum = (vector unsigned int)vec_splat_u32(0);
for(i=0;i<h;i++) {
for (i = 0; i < h; i++) {
/* Read potentially unaligned pixels into t1 and t2 */
perm1 = vec_lvsl(0, pix1);
pix1v = (vector unsigned char *) pix1;
......@@ -435,10 +421,8 @@ int sse16_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
t1 = vec_perm(pix1v[0], pix1v[1], perm1);
t2 = vec_perm(pix2v[0], pix2v[1], perm2);
/*
Since we want to use unsigned chars, we can take advantage
of the fact that abs(a-b)^2 = (a-b)^2.
*/
/* Since we want to use unsigned chars, we can take advantage
of the fact that abs(a-b)^2 = (a-b)^2. */
/* Calculate abs differences vector */
t3 = vec_max(t1, t2);
......@@ -500,8 +484,7 @@ void get_pixels_altivec(DCTELEM *restrict block, const uint8_t *pixels, int line
const vector unsigned char zero = (const vector unsigned char)vec_splat_u8(0);
vector signed short shorts;
for(i=0;i<8;i++)
{
for (i = 0; i < 8; i++) {
// Read potentially unaligned pixels.
// We're reading 16 pixels, and actually only want 8,
// but we simply ignore the extras.
......@@ -527,8 +510,7 @@ void diff_pixels_altivec(DCTELEM *restrict block, const uint8_t *s1,
const vector unsigned char zero = (const vector unsigned char)vec_splat_u8(0);
vector signed short shorts1, shorts2;
for(i=0;i<4;i++)
{
for (i = 0; i < 4; i++) {
// Read potentially unaligned pixels
// We're reading 16 pixels, and actually only want 8,
// but we simply ignore the extras.
......@@ -596,16 +578,14 @@ void add_bytes_altivec(uint8_t *dst, uint8_t *src, int w) {
register vector unsigned char vdst, vsrc;
/* dst and src are 16 bytes-aligned (guaranteed) */
for(i = 0 ; (i + 15) < w ; i+=16)
{
for (i = 0 ; (i + 15) < w ; i+=16) {
vdst = vec_ld(i, (unsigned char*)dst);
vsrc = vec_ld(i, (unsigned char*)src);
vdst = vec_add(vsrc, vdst);
vec_st(vdst, i, (unsigned char*)dst);
}
/* if w is not a multiple of 16 */
for (; (i < w) ; i++)
{
for (; (i < w) ; i++) {
dst[i] = src[i];
}
}
......@@ -632,7 +612,7 @@ POWERPC_PERF_START_COUNT(altivec_put_pixels16_num, 1);
// -funroll-loops w/ this is bad - 74 cycles again.
// all this is on a 7450, tuning for the 7450
#if 0
for(i=0; i<h; i++) {
for (i = 0; i < h; i++) {
pixelsv1 = vec_ld(0, (unsigned char*)pixels);
pixelsv2 = vec_ld(16, (unsigned char*)pixels);
vec_st(vec_perm(pixelsv1, pixelsv2, perm),
......@@ -641,7 +621,7 @@ POWERPC_PERF_START_COUNT(altivec_put_pixels16_num, 1);
block +=line_size;
}
#else
for(i=0; i<h; i+=4) {
for (i = 0; i < h; i += 4) {
pixelsv1 = vec_ld(0, (unsigned char*)pixels);
pixelsv2 = vec_ld(15, (unsigned char*)pixels);
pixelsv1B = vec_ld(line_size, (unsigned char*)pixels);
......@@ -676,7 +656,7 @@ POWERPC_PERF_DECLARE(altivec_avg_pixels16_num, 1);
POWERPC_PERF_START_COUNT(altivec_avg_pixels16_num, 1);
for(i=0; i<h; i++) {
for (i = 0; i < h; i++) {
pixelsv1 = vec_ld(0, (unsigned char*)pixels);
pixelsv2 = vec_ld(16, (unsigned char*)pixels);
blockv = vec_ld(0, block);
......@@ -700,10 +680,8 @@ POWERPC_PERF_DECLARE(altivec_avg_pixels8_num, 1);
POWERPC_PERF_START_COUNT(altivec_avg_pixels8_num, 1);
for (i = 0; i < h; i++) {
/*
block is 8 bytes-aligned, so we're either in the
left block (16 bytes-aligned) or in the right block (not)
*/
/* block is 8 bytes-aligned, so we're either in the
left block (16 bytes-aligned) or in the right block (not) */
int rightside = ((unsigned long)block & 0x0000000F);
blockv = vec_ld(0, block);
......@@ -711,12 +689,9 @@ POWERPC_PERF_START_COUNT(altivec_avg_pixels8_num, 1);
pixelsv2 = vec_ld(16, (unsigned char*)pixels);
pixelsv = vec_perm(pixelsv1, pixelsv2, vec_lvsl(0, pixels));
if (rightside)
{
if (rightside) {
pixelsv = vec_perm(blockv, pixelsv, vcprm(0,1,s0,s1));
}
else
{
} else {
pixelsv = vec_perm(blockv, pixelsv, vcprm(s0,s1,2,3));
}
......@@ -736,25 +711,18 @@ void put_pixels8_xy2_altivec(uint8_t *block, const uint8_t *pixels, int line_siz
{
POWERPC_PERF_DECLARE(altivec_put_pixels8_xy2_num, 1);
register int i;
register vector unsigned char
pixelsv1, pixelsv2,
pixelsavg;
register vector unsigned char
blockv, temp1, temp2;
register vector unsigned short
pixelssum1, pixelssum2, temp3;
register vector unsigned char pixelsv1, pixelsv2, pixelsavg;
register vector unsigned char blockv, temp1, temp2;
register vector unsigned short pixelssum1, pixelssum2, temp3;
register const vector unsigned char vczero = (const vector unsigned char)vec_splat_u8(0);
register const vector unsigned short vctwo = (const vector unsigned short)vec_splat_u16(2);
temp1 = vec_ld(0, pixels);
temp2 = vec_ld(16, pixels);
pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(0, pixels));
if ((((unsigned long)pixels) & 0x0000000F) == 0x0000000F)
{
if ((((unsigned long)pixels) & 0x0000000F) == 0x0000000F) {
pixelsv2 = temp2;
}
else
{
} else {
pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
}
pixelsv1 = vec_mergeh(vczero, pixelsv1);
......@@ -771,12 +739,9 @@ POWERPC_PERF_START_COUNT(altivec_put_pixels8_xy2_num, 1);
temp1 = vec_ld(line_size, pixels);
temp2 = vec_ld(line_size + 16, pixels);
pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels));
if (((((unsigned long)pixels) + line_size) & 0x0000000F) == 0x0000000F)
{
if (((((unsigned long)pixels) + line_size) & 0x0000000F) == 0x0000000F) {
pixelsv2 = temp2;
}
else
{
} else {
pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
}
......@@ -789,12 +754,9 @@ POWERPC_PERF_START_COUNT(altivec_put_pixels8_xy2_num, 1);
pixelssum1 = vec_add(pixelssum2, vctwo);
pixelsavg = vec_packsu(temp3, (vector unsigned short) vczero);
if (rightside)
{
if (rightside) {
blockv = vec_perm(blockv, pixelsavg, vcprm(0, 1, s0, s1));
}
else
{
} else {
blockv = vec_perm(blockv, pixelsavg, vcprm(s0, s1, 2, 3));
}
......@@ -812,13 +774,9 @@ void put_no_rnd_pixels8_xy2_altivec(uint8_t *block, const uint8_t *pixels, int l
{
POWERPC_PERF_DECLARE(altivec_put_no_rnd_pixels8_xy2_num, 1);
register int i;
register vector unsigned char
pixelsv1, pixelsv2,
pixelsavg;
register vector unsigned char
blockv, temp1, temp2;
register vector unsigned short
pixelssum1, pixelssum2, temp3;
register vector unsigned char pixelsv1, pixelsv2, pixelsavg;
register vector unsigned char blockv, temp1, temp2;
register vector unsigned short pixelssum1, pixelssum2, temp3;
register const vector unsigned char vczero = (const vector unsigned char)vec_splat_u8(0);
register const vector unsigned short vcone = (const vector unsigned short)vec_splat_u16(1);
register const vector unsigned short vctwo = (const vector unsigned short)vec_splat_u16(2);
......@@ -826,12 +784,9 @@ POWERPC_PERF_DECLARE(altivec_put_no_rnd_pixels8_xy2_num, 1);
temp1 = vec_ld(0, pixels);
temp2 = vec_ld(16, pixels);
pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(0, pixels));
if ((((unsigned long)pixels) & 0x0000000F) == 0x0000000F)
{
if ((((unsigned long)pixels) & 0x0000000F) == 0x0000000F) {
pixelsv2 = temp2;
}
else
{
} else {
pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
}
pixelsv1 = vec_mergeh(vczero, pixelsv1);
......@@ -848,12 +803,9 @@ POWERPC_PERF_START_COUNT(altivec_put_no_rnd_pixels8_xy2_num, 1);
temp1 = vec_ld(line_size, pixels);
temp2 = vec_ld(line_size + 16, pixels);
pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels));
if (((((unsigned long)pixels) + line_size) & 0x0000000F) == 0x0000000F)
{
if (((((unsigned long)pixels) + line_size) & 0x0000000F) == 0x0000000F) {
pixelsv2 = temp2;
}
else
{
} else {
pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
}
......@@ -866,12 +818,9 @@ POWERPC_PERF_START_COUNT(altivec_put_no_rnd_pixels8_xy2_num, 1);
pixelssum1 = vec_add(pixelssum2, vcone);
pixelsavg = vec_packsu(temp3, (vector unsigned short) vczero);
if (rightside)
{
if (rightside) {
blockv = vec_perm(blockv, pixelsavg, vcprm(0, 1, s0, s1));
}
else
{
} else {
blockv = vec_perm(blockv, pixelsavg, vcprm(s0, s1, 2, 3));
}
......@@ -889,13 +838,10 @@ void put_pixels16_xy2_altivec(uint8_t * block, const uint8_t * pixels, int line_
{
POWERPC_PERF_DECLARE(altivec_put_pixels16_xy2_num, 1);
register int i;
register vector unsigned char
pixelsv1, pixelsv2, pixelsv3, pixelsv4;
register vector unsigned char
blockv, temp1, temp2;
register vector unsigned short
pixelssum1, pixelssum2, temp3,
pixelssum3, pixelssum4, temp4;
register vector unsigned char pixelsv1, pixelsv2, pixelsv3, pixelsv4;
register vector unsigned char blockv, temp1, temp2;
register vector unsigned short temp3, temp4,
pixelssum1, pixelssum2, pixelssum3, pixelssum4;
register const vector unsigned char vczero = (const vector unsigned char)vec_splat_u8(0);
register const vector unsigned short vctwo = (const vector unsigned short)vec_splat_u16(2);
......@@ -904,12 +850,9 @@ POWERPC_PERF_START_COUNT(altivec_put_pixels16_xy2_num, 1);
temp1 = vec_ld(0, pixels);
temp2 = vec_ld(16, pixels);
pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(0, pixels));
if ((((unsigned long)pixels) & 0x0000000F) == 0x0000000F)
{
if ((((unsigned long)pixels) & 0x0000000F) == 0x0000000F) {
pixelsv2 = temp2;
}
else
{
} else {
pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
}
pixelsv3 = vec_mergel(vczero, pixelsv1);
......@@ -929,12 +872,9 @@ POWERPC_PERF_START_COUNT(altivec_put_pixels16_xy2_num, 1);
temp1 = vec_ld(line_size, pixels);
temp2 = vec_ld(line_size + 16, pixels);
pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels));
if (((((unsigned long)pixels) + line_size) & 0x0000000F) == 0x0000000F)
{
if (((((unsigned long)pixels) + line_size) & 0x0000000F) == 0x0000000F) {
pixelsv2 = temp2;
}
else
{
} else {
pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
}
......@@ -971,13 +911,10 @@ void put_no_rnd_pixels16_xy2_altivec(uint8_t * block, const uint8_t * pixels, in
{
POWERPC_PERF_DECLARE(altivec_put_no_rnd_pixels16_xy2_num, 1);
register int i;
register vector unsigned char
pixelsv1, pixelsv2, pixelsv3, pixelsv4;
register vector unsigned char
blockv, temp1, temp2;
register vector unsigned short
pixelssum1, pixelssum2, temp3,
pixelssum3, pixelssum4, temp4;
register vector unsigned char pixelsv1, pixelsv2, pixelsv3, pixelsv4;
register vector unsigned char blockv, temp1, temp2;
register vector unsigned short temp3, temp4,
pixelssum1, pixelssum2, pixelssum3, pixelssum4;
register const vector unsigned char vczero = (const vector unsigned char)vec_splat_u8(0);
register const vector unsigned short vcone = (const vector unsigned short)vec_splat_u16(1);
register const vector unsigned short vctwo = (const vector unsigned short)vec_splat_u16(2);
......@@ -987,12 +924,9 @@ POWERPC_PERF_START_COUNT(altivec_put_no_rnd_pixels16_xy2_num, 1);
temp1 = vec_ld(0, pixels);
temp2 = vec_ld(16, pixels);
pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(0, pixels));
if ((((unsigned long)pixels) & 0x0000000F) == 0x0000000F)
{
if ((((unsigned long)pixels) & 0x0000000F) == 0x0000000F) {
pixelsv2 = temp2;
}
else
{
} else {
pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
}
pixelsv3 = vec_mergel(vczero, pixelsv1);
......@@ -1012,12 +946,9 @@ POWERPC_PERF_START_COUNT(altivec_put_no_rnd_pixels16_xy2_num, 1);
temp1 = vec_ld(line_size, pixels);
temp2 = vec_ld(line_size + 16, pixels);
pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels));
if (((((unsigned long)pixels) + line_size) & 0x0000000F) == 0x0000000F)
{
if (((((unsigned long)pixels) + line_size) & 0x0000000F) == 0x0000000F) {
pixelsv2 = temp2;
}
else
{
} else {
pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
}
......@@ -1088,11 +1019,9 @@ POWERPC_PERF_START_COUNT(altivec_hadamard8_diff8x8_num, 1);
dstO = vec_perm(dst1, dst2, vec_lvsl(stride * i, dst)); \
/* promote the unsigned chars to signed shorts */ \
/* we're in the 8x8 function, we only care for the first 8 */ \
srcV = \
(vector signed short)vec_mergeh((vector signed char)vzero, \
srcV = (vector signed short)vec_mergeh((vector signed char)vzero, \
(vector signed char)srcO); \
dstV = \
(vector signed short)vec_mergeh((vector signed char)vzero, \
dstV = (vector signed short)vec_mergeh((vector signed char)vzero, \
(vector signed char)dstO); \
/* subtractions inside the first butterfly */ \
but0 = vec_sub(srcV, dstV); \
......@@ -1159,25 +1088,22 @@ POWERPC_PERF_STOP_COUNT(altivec_hadamard8_diff8x8_num, 1);
}
/*
16x8 works with 16 elements ; it allows to avoid replicating
loads, and give the compiler more rooms for scheduling.
It's only used from inside hadamard8_diff16_altivec.
Unfortunately, it seems gcc-3.3 is a bit dumb, and
the compiled code has a LOT of spill code, it seems
gcc (unlike xlc) cannot keep everything in registers
by itself. The following code include hand-made
registers allocation. It's not clean, but on
a 7450 the resulting code is much faster (best case
fall from 700+ cycles to 550).
xlc doesn't add spill code, but it doesn't know how to
schedule for the 7450, and its code isn't much faster than
gcc-3.3 on the 7450 (but uses 25% less instructions...)
On the 970, the hand-made RA is still a win (around 690
vs. around 780), but xlc goes to around 660 on the
regular C code...
16x8 works with 16 elements; it allows to avoid replicating loads, and
give the compiler more rooms for scheduling. It's only used from
inside hadamard8_diff16_altivec.
Unfortunately, it seems gcc-3.3 is a bit dumb, and the compiled code has a LOT
of spill code, it seems gcc (unlike xlc) cannot keep everything in registers
by itself. The following code include hand-made registers allocation. It's not
clean, but on a 7450 the resulting code is much faster (best case fall from
700+ cycles to 550).
xlc doesn't add spill code, but it doesn't know how to schedule for the 7450,
and its code isn't much faster than gcc-3.3 on the 7450 (but uses 25% less
instructions...)
On the 970, the hand-made RA is still a win (around 690 vs. around 780), but
xlc goes to around 660 on the regular C code...
*/
static int hadamard8_diff16x8_altivec(/*MpegEncContext*/ void *s, uint8_t *dst, uint8_t *src, int stride, int h) {
......@@ -1255,17 +1181,13 @@ static int hadamard8_diff16x8_altivec(/*MpegEncContext*/ void *s, uint8_t *dst,
dst2 = vec_ld((stride * i) + 16, dst); \
dstO = vec_perm(dst1, dst2, vec_lvsl(stride * i, dst)); \
/* promote the unsigned chars to signed shorts */ \
srcV = \
(vector signed short)vec_mergeh((vector signed char)vzero, \
srcV = (vector signed short)vec_mergeh((vector signed char)vzero, \
(vector signed char)srcO); \
dstV = \
(vector signed short)vec_mergeh((vector signed char)vzero, \
dstV = (vector signed short)vec_mergeh((vector signed char)vzero, \
(vector signed char)dstO); \
srcW = \
(vector signed short)vec_mergel((vector signed char)vzero, \
srcW = (vector signed short)vec_mergel((vector signed char)vzero, \
(vector signed char)srcO); \
dstW = \
(vector signed short)vec_mergel((vector signed char)vzero, \
dstW = (vector signed short)vec_mergel((vector signed char)vzero, \
(vector signed char)dstO); \
/* subtractions inside the first butterfly */ \
but0 = vec_sub(srcV, dstV); \
......@@ -1401,7 +1323,7 @@ static void vorbis_inverse_coupling_altivec(float *mag, float *ang,
vector bool int t0, t1;
const vector unsigned int v_31 = //XXX
vec_add(vec_add(vec_splat_u32(15),vec_splat_u32(15)),vec_splat_u32(1));
for(i=0; i<blocksize; i+=4) {
for (i = 0; i < blocksize; i += 4) {
m = vec_ld(0, mag+i);
a = vec_ld(0, ang+i);
t0 = vec_cmple(m, (vector float)vec_splat_u32(0));
......@@ -1452,8 +1374,7 @@ POWERPC_PERF_START_COUNT(altivec_avg_pixels8_xy2_num, 1);
temp1 = vec_ld(line_size, pixels);
temp2 = vec_ld(line_size + 16, pixels);
pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels));
if (((((unsigned long)pixels) + line_size) & 0x0000000F) == 0x0000000F)
{
if (((((unsigned long)pixels) + line_size) & 0x0000000F) == 0x0000000F) {
pixelsv2 = temp2;
} else {
pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment