#include "libavcodec/me_cmp.h"
#if HAVE_ALTIVEC
+
+#if HAVE_BIGENDIAN
+#define GET_PERM(per1, per2, pix) {\
+ per1 = vec_lvsl(0, pix);\
+ per2 = vec_add(per1, vec_splat_u8(1));\
+}
+#define LOAD_PIX(v, iv, pix, per1, per2) {\
+ vector unsigned char pix2l = vec_ld(0, pix);\
+ vector unsigned char pix2r = vec_ld(16, pix);\
+ v = vec_perm(pix2l, pix2r, per1);\
+ iv = vec_perm(pix2l, pix2r, per2);\
+}
+#else
+#define GET_PERM(per1, per2, pix) {}
+#define LOAD_PIX(v, iv, pix, per1, per2) {\
+ v = vec_vsx_ld(0, pix);\
+ iv = vec_vsx_ld(1, pix);\
+}
+#endif
static int sad16_x2_altivec(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
- int line_size, int h)
+ ptrdiff_t stride, int h)
{
- int i, s = 0;
+ int i;
+ int __attribute__((aligned(16))) s = 0;
const vector unsigned char zero =
(const vector unsigned char) vec_splat_u8(0);
- vector unsigned char perm1 = vec_lvsl(0, pix2);
- vector unsigned char perm2 = vec_add(perm1, vec_splat_u8(1));
vector unsigned int sad = (vector unsigned int) vec_splat_u32(0);
vector signed int sumdiffs;
+ vector unsigned char perm1, perm2, pix2v, pix2iv;
+ GET_PERM(perm1, perm2, pix2);
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] */
vector unsigned char pix1v = vec_ld(0, pix1);
- vector unsigned char pix2l = vec_ld(0, pix2);
- vector unsigned char pix2r = vec_ld(16, pix2);
- vector unsigned char pix2v = vec_perm(pix2l, pix2r, perm1);
- vector unsigned char pix2iv = vec_perm(pix2l, pix2r, perm2);
+ LOAD_PIX(pix2v, pix2iv, pix2, perm1, perm2);
/* Calculate the average vector. */
vector unsigned char avgv = vec_avg(pix2v, pix2iv);
/* Add each 4 pixel group together and put 4 results into sad. */
sad = vec_sum4s(t5, sad);
- pix1 += line_size;
- pix2 += line_size;
+ pix1 += stride;
+ pix2 += stride;
}
/* Sum up the four partial sums, and put the result into s. */
sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
}
static int sad16_y2_altivec(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
- int line_size, int h)
+ ptrdiff_t stride, int h)
{
- int i, s = 0;
+ int i;
+ int __attribute__((aligned(16))) s = 0;
const vector unsigned char zero =
(const vector unsigned char) vec_splat_u8(0);
- vector unsigned char perm = vec_lvsl(0, pix2);
vector unsigned char pix1v, pix3v, avgv, t5;
vector unsigned int sad = (vector unsigned int) vec_splat_u32(0);
vector signed int sumdiffs;
- uint8_t *pix3 = pix2 + line_size;
- /* Due to the fact that pix3 = pix2 + line_size, the pix3 of one
+ uint8_t *pix3 = pix2 + stride;
+
+ /* Due to the fact that pix3 = pix2 + stride, 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. */
- vector unsigned char pix2l = vec_ld(0, pix2);
- vector unsigned char pix2r = vec_ld(15, pix2);
- vector unsigned char pix2v = vec_perm(pix2l, pix2r, perm);
+ vector unsigned char pix2v = VEC_LD(0, pix2);
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] */
pix1v = vec_ld(0, pix1);
-
- pix2l = vec_ld(0, pix3);
- pix2r = vec_ld(15, pix3);
- pix3v = vec_perm(pix2l, pix2r, perm);
+ pix3v = VEC_LD(0, pix3);
/* Calculate the average vector. */
avgv = vec_avg(pix2v, pix3v);
/* Add each 4 pixel group together and put 4 results into sad. */
sad = vec_sum4s(t5, sad);
- pix1 += line_size;
+ pix1 += stride;
pix2v = pix3v;
- pix3 += line_size;
+ pix3 += stride;
}
/* Sum up the four partial sums, and put the result into s. */
}
static int sad16_xy2_altivec(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
- int line_size, int h)
+ ptrdiff_t stride, int h)
{
- int i, s = 0;
- uint8_t *pix3 = pix2 + line_size;
+ int i;
+ int __attribute__((aligned(16))) s = 0;
+ uint8_t *pix3 = pix2 + stride;
const vector unsigned char zero =
(const vector unsigned char) vec_splat_u8(0);
const vector unsigned short two =
(const vector unsigned short) vec_splat_u16(2);
vector unsigned char avgv, t5;
- vector unsigned char perm1 = vec_lvsl(0, pix2);
- vector unsigned char perm2 = vec_add(perm1, vec_splat_u8(1));
vector unsigned char pix1v, pix3v, pix3iv;
vector unsigned short pix3lv, pix3hv, pix3ilv, pix3ihv;
vector unsigned short avghv, avglv;
vector unsigned int sad = (vector unsigned int) vec_splat_u32(0);
vector signed int sumdiffs;
+ vector unsigned char perm1, perm2, pix2v, pix2iv;
+ GET_PERM(perm1, perm2, pix2);
- /* Due to the fact that pix3 = pix2 + line_size, the pix3 of one
+ /* Due to the fact that pix3 = pix2 + stride, 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. */
- vector unsigned char pix2l = vec_ld(0, pix2);
- vector unsigned char pix2r = vec_ld(16, pix2);
- vector unsigned char pix2v = vec_perm(pix2l, pix2r, perm1);
- vector unsigned char pix2iv = vec_perm(pix2l, pix2r, perm2);
-
+ LOAD_PIX(pix2v, pix2iv, pix2, perm1, perm2);
vector unsigned short pix2hv =
- (vector unsigned short) vec_mergeh(zero, pix2v);
+ (vector unsigned short) VEC_MERGEH(zero, pix2v);
vector unsigned short pix2lv =
- (vector unsigned short) vec_mergel(zero, pix2v);
+ (vector unsigned short) VEC_MERGEL(zero, pix2v);
vector unsigned short pix2ihv =
- (vector unsigned short) vec_mergeh(zero, pix2iv);
+ (vector unsigned short) VEC_MERGEH(zero, pix2iv);
vector unsigned short pix2ilv =
- (vector unsigned short) vec_mergel(zero, pix2iv);
+ (vector unsigned short) VEC_MERGEL(zero, pix2iv);
+
vector unsigned short t1 = vec_add(pix2hv, pix2ihv);
vector unsigned short t2 = vec_add(pix2lv, pix2ilv);
vector unsigned short t3, t4;
* pix1v: pix1[0] - pix1[15]
* pix3v: pix3[0] - pix3[15] pix3iv: pix3[1] - pix3[16] */
pix1v = vec_ld(0, pix1);
-
- pix2l = vec_ld(0, pix3);
- pix2r = vec_ld(16, pix3);
- pix3v = vec_perm(pix2l, pix2r, perm1);
- pix3iv = vec_perm(pix2l, pix2r, perm2);
+ LOAD_PIX(pix3v, pix3iv, pix3, perm1, perm2);
/* 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
* 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);
+ 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);
/* Add each 4 pixel group together and put 4 results into sad. */
sad = vec_sum4s(t5, sad);
- pix1 += line_size;
- pix3 += line_size;
+ pix1 += stride;
+ pix3 += stride;
/* Transfer the calculated values for pix3 into pix2. */
t1 = t3;
t2 = t4;
}
static int sad16_altivec(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
- int line_size, int h)
+ ptrdiff_t stride, int h)
{
- int i, s;
+ int i;
+ int __attribute__((aligned(16))) s;
const vector unsigned int zero =
(const vector unsigned int) vec_splat_u32(0);
- vector unsigned char perm = vec_lvsl(0, pix2);
vector unsigned int sad = (vector unsigned int) vec_splat_u32(0);
vector signed int sumdiffs;
for (i = 0; i < h; i++) {
/* Read potentially unaligned pixels into t1 and t2. */
- vector unsigned char pix2l = vec_ld(0, pix2);
- vector unsigned char pix2r = vec_ld(15, pix2);
- vector unsigned char t1 = vec_ld(0, pix1);
- vector unsigned char t2 = vec_perm(pix2l, pix2r, perm);
+ vector unsigned char t1 =vec_ld(0, pix1);
+ vector unsigned char t2 = VEC_LD(0, pix2);
/* Calculate a sum of abs differences vector. */
vector unsigned char t3 = vec_max(t1, t2);
/* Add each 4 pixel group together and put 4 results into sad. */
sad = vec_sum4s(t5, sad);
- pix1 += line_size;
- pix2 += line_size;
+ pix1 += stride;
+ pix2 += stride;
}
/* Sum up the four partial sums, and put the result into s. */
}
static int sad8_altivec(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
- int line_size, int h)
+ ptrdiff_t stride, int h)
{
- int i, s;
+ int i;
+ int __attribute__((aligned(16))) s;
const vector unsigned int zero =
(const vector unsigned int) vec_splat_u32(0);
const vector unsigned char permclear =
(vector unsigned char)
{ 255, 255, 255, 255, 255, 255, 255, 255, 0, 0, 0, 0, 0, 0, 0, 0 };
- vector unsigned char perm1 = vec_lvsl(0, pix1);
- vector unsigned char perm2 = vec_lvsl(0, pix2);
vector unsigned int sad = (vector unsigned int) vec_splat_u32(0);
vector signed int sumdiffs;
/* 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. */
- vector unsigned char pix1l = vec_ld(0, pix1);
- vector unsigned char pix1r = vec_ld(7, pix1);
- vector unsigned char pix2l = vec_ld(0, pix2);
- vector unsigned char pix2r = vec_ld(7, pix2);
- vector unsigned char t1 = vec_and(vec_perm(pix1l, pix1r, perm1),
- permclear);
- vector unsigned char t2 = vec_and(vec_perm(pix2l, pix2r, perm2),
- permclear);
+ vector unsigned char pix1l = VEC_LD(0, pix1);
+ vector unsigned char pix2l = VEC_LD(0, pix2);
+ vector unsigned char t1 = vec_and(pix1l, permclear);
+ vector unsigned char t2 = vec_and(pix2l, permclear);
/* Calculate a sum of abs differences vector. */
vector unsigned char t3 = vec_max(t1, t2);
/* Add each 4 pixel group together and put 4 results into sad. */
sad = vec_sum4s(t5, sad);
- pix1 += line_size;
- pix2 += line_size;
+ pix1 += stride;
+ pix2 += stride;
}
/* Sum up the four partial sums, and put the result into s. */
/* Sum of Squared Errors for an 8x8 block, AltiVec-enhanced.
* It's the sad8_altivec code above w/ squaring added. */
static int sse8_altivec(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
- int line_size, int h)
+ ptrdiff_t stride, int h)
{
- int i, s;
+ int i;
+ int __attribute__((aligned(16))) s;
const vector unsigned int zero =
(const vector unsigned int) vec_splat_u32(0);
const vector unsigned char permclear =
(vector unsigned char)
{ 255, 255, 255, 255, 255, 255, 255, 255, 0, 0, 0, 0, 0, 0, 0, 0 };
- vector unsigned char perm1 = vec_lvsl(0, pix1);
- vector unsigned char perm2 = vec_lvsl(0, pix2);
vector unsigned int sum = (vector unsigned int) vec_splat_u32(0);
vector signed int sumsqr;
/* 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. */
- vector unsigned char pix1l = vec_ld(0, pix1);
- vector unsigned char pix1r = vec_ld(7, pix1);
- vector unsigned char pix2l = vec_ld(0, pix2);
- vector unsigned char pix2r = vec_ld(7, pix2);
- vector unsigned char t1 = vec_and(vec_perm(pix1l, pix1r, perm1),
- permclear);
- vector unsigned char t2 = vec_and(vec_perm(pix2l, pix2r, perm2),
- permclear);
+ vector unsigned char t1 = vec_and(VEC_LD(0, pix1), permclear);
+ vector unsigned char t2 = vec_and(VEC_LD(0, pix2), permclear);
/* Since we want to use unsigned chars, we can take advantage
* of the fact that abs(a - b) ^ 2 = (a - b) ^ 2. */
/* Square the values and add them to our sum. */
sum = vec_msum(t5, t5, sum);
- pix1 += line_size;
- pix2 += line_size;
+ pix1 += stride;
+ pix2 += stride;
}
/* Sum up the four partial sums, and put the result into s. */
/* Sum of Squared Errors for a 16x16 block, AltiVec-enhanced.
* It's the sad16_altivec code above w/ squaring added. */
static int sse16_altivec(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
- int line_size, int h)
+ ptrdiff_t stride, int h)
{
- int i, s;
+ int i;
+ int __attribute__((aligned(16))) s;
const vector unsigned int zero =
(const vector unsigned int) vec_splat_u32(0);
- vector unsigned char perm = vec_lvsl(0, pix2);
vector unsigned int sum = (vector unsigned int) vec_splat_u32(0);
vector signed int sumsqr;
for (i = 0; i < h; i++) {
/* Read potentially unaligned pixels into t1 and t2. */
- vector unsigned char pix2l = vec_ld(0, pix2);
- vector unsigned char pix2r = vec_ld(15, pix2);
vector unsigned char t1 = vec_ld(0, pix1);
- vector unsigned char t2 = vec_perm(pix2l, pix2r, perm);
+ vector unsigned char t2 = VEC_LD(0, pix2);
/* Since we want to use unsigned chars, we can take advantage
* of the fact that abs(a - b) ^ 2 = (a - b) ^ 2. */
/* Square the values and add them to our sum. */
sum = vec_msum(t5, t5, sum);
- pix1 += line_size;
- pix2 += line_size;
+ pix1 += stride;
+ pix2 += stride;
}
/* Sum up the four partial sums, and put the result into s. */
sumsqr = vec_sums((vector signed int) sum, (vector signed int) zero);
sumsqr = vec_splat(sumsqr, 3);
- vec_ste(sumsqr, 0, &s);
+ vec_ste(sumsqr, 0, &s);
return s;
}
static int hadamard8_diff8x8_altivec(MpegEncContext *s, uint8_t *dst,
- uint8_t *src, int stride, int h)
+ uint8_t *src, ptrdiff_t stride, int h)
{
- int sum;
+ int __attribute__((aligned(16))) sum;
register const vector unsigned char vzero =
(const vector unsigned char) vec_splat_u8(0);
register vector signed short temp0, temp1, temp2, temp3, temp4,
{ 0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0F,
0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07 };
+
#define ONEITERBUTTERFLY(i, res) \
{ \
- register vector unsigned char src1 = vec_ld(stride * i, src); \
- register vector unsigned char src2 = vec_ld(stride * i + 15, src); \
- register vector unsigned char srcO = \
- vec_perm(src1, src2, vec_lvsl(stride * i, src)); \
- register vector unsigned char dst1 = vec_ld(stride * i, dst); \
- register vector unsigned char dst2 = vec_ld(stride * i + 15, dst); \
- register vector unsigned char dstO = \
- vec_perm(dst1, dst2, vec_lvsl(stride * i, dst)); \
+ register vector unsigned char srcO = unaligned_load(stride * i, src); \
+ register vector unsigned char dstO = unaligned_load(stride * i, dst);\
\
/* Promote the unsigned chars to signed shorts. */ \
/* We're in the 8x8 function, we only care for the first 8. */ \
register vector signed short srcV = \
- (vector signed short) vec_mergeh((vector signed char) vzero, \
+ (vector signed short) VEC_MERGEH((vector signed char) vzero, \
(vector signed char) srcO); \
register vector signed short dstV = \
- (vector signed short) vec_mergeh((vector signed char) vzero, \
+ (vector signed short) VEC_MERGEH((vector signed char) vzero, \
(vector signed char) dstO); \
\
/* subtractions inside the first butterfly */ \
register vector signed short op3 = vec_perm(but2, but2, perm3); \
res = vec_mladd(but2, vprod3, op3); \
}
+
ONEITERBUTTERFLY(0, temp0);
ONEITERBUTTERFLY(1, temp1);
ONEITERBUTTERFLY(2, temp2);
vsum = vec_sum4s(vec_abs(line7C), vsum);
vsum = vec_sums(vsum, (vector signed int) vzero);
vsum = vec_splat(vsum, 3);
+
vec_ste(vsum, 0, &sum);
}
return sum;
* but xlc goes to around 660 on the regular C code...
*/
static int hadamard8_diff16x8_altivec(MpegEncContext *s, uint8_t *dst,
- uint8_t *src, int stride, int h)
+ uint8_t *src, ptrdiff_t stride, int h)
{
- int sum;
+ int __attribute__((aligned(16))) sum;
register vector signed short
temp0 __asm__ ("v0"),
temp1 __asm__ ("v1"),
#define ONEITERBUTTERFLY(i, res1, res2) \
{ \
- register vector unsigned char src1 __asm__ ("v22") = \
- vec_ld(stride * i, src); \
- register vector unsigned char src2 __asm__ ("v23") = \
- vec_ld(stride * i + 16, src); \
register vector unsigned char srcO __asm__ ("v22") = \
- vec_perm(src1, src2, vec_lvsl(stride * i, src)); \
- register vector unsigned char dst1 __asm__ ("v24") = \
- vec_ld(stride * i, dst); \
- register vector unsigned char dst2 __asm__ ("v25") = \
- vec_ld(stride * i + 16, dst); \
+ unaligned_load(stride * i, src); \
register vector unsigned char dstO __asm__ ("v23") = \
- vec_perm(dst1, dst2, vec_lvsl(stride * i, dst)); \
+ unaligned_load(stride * i, dst);\
\
/* Promote the unsigned chars to signed shorts. */ \
register vector signed short srcV __asm__ ("v24") = \
- (vector signed short) vec_mergeh((vector signed char) vzero, \
+ (vector signed short) VEC_MERGEH((vector signed char) vzero, \
(vector signed char) srcO); \
register vector signed short dstV __asm__ ("v25") = \
- (vector signed short) vec_mergeh((vector signed char) vzero, \
+ (vector signed short) VEC_MERGEH((vector signed char) vzero, \
(vector signed char) dstO); \
register vector signed short srcW __asm__ ("v26") = \
- (vector signed short) vec_mergel((vector signed char) vzero, \
+ (vector signed short) VEC_MERGEL((vector signed char) vzero, \
(vector signed char) srcO); \
register vector signed short dstW __asm__ ("v27") = \
- (vector signed short) vec_mergel((vector signed char) vzero, \
+ (vector signed short) VEC_MERGEL((vector signed char) vzero, \
(vector signed char) dstO); \
\
/* subtractions inside the first butterfly */ \
res1 = vec_mladd(but2, vprod3, op3); \
res2 = vec_mladd(but2S, vprod3, op3S); \
}
+
ONEITERBUTTERFLY(0, temp0, temp0S);
ONEITERBUTTERFLY(1, temp1, temp1S);
ONEITERBUTTERFLY(2, temp2, temp2S);
vsum = vec_sum4s(vec_abs(line7CS), vsum);
vsum = vec_sums(vsum, (vector signed int) vzero);
vsum = vec_splat(vsum, 3);
+
vec_ste(vsum, 0, &sum);
}
return sum;
}
static int hadamard8_diff16_altivec(MpegEncContext *s, uint8_t *dst,
- uint8_t *src, int stride, int h)
+ uint8_t *src, ptrdiff_t stride, int h)
{
int score = hadamard8_diff16x8_altivec(s, dst, src, stride, 8);