* Copyright (c) 2002 Dieter Shirley
* Copyright (c) 2003-2004 Romain Dolbeau <romain@dolbeau.org>
*
- * This library is free software; you can redistribute it and/or
+ * This file is part of Libav.
+ *
+ * Libav is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
- * version 2 of the License, or (at your option) any later version.
+ * version 2.1 of the License, or (at your option) any later version.
*
- * This library is distributed in the hope that it will be useful,
+ * Libav is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
- * License along with this library; if not, write to the Free Software
+ * License along with Libav; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
-#include "../dsputil.h"
-
-#include "gcc_fixes.h"
-
+#include "config.h"
+#if HAVE_ALTIVEC_H
+#include <altivec.h>
+#endif
+#include "libavcodec/dsputil.h"
+#include "util_altivec.h"
+#include "types_altivec.h"
#include "dsputil_altivec.h"
-#ifdef CONFIG_DARWIN
-#include <sys/sysctl.h>
-#else /* CONFIG_DARWIN */
-#ifdef __AMIGAOS4__
-#include <exec/exec.h>
-#include <interfaces/exec.h>
-#include <proto/exec.h>
-#else /* __AMIGAOS4__ */
-#include <signal.h>
-#include <setjmp.h>
-
-static sigjmp_buf jmpbuf;
-static volatile sig_atomic_t canjump = 0;
-
-static void sigill_handler (int sig)
-{
- if (!canjump) {
- signal (sig, SIG_DFL);
- raise (sig);
- }
-
- canjump = 0;
- siglongjmp (jmpbuf, 1);
-}
-#endif /* CONFIG_DARWIN */
-#endif /* __AMIGAOS4__ */
-
-int sad16_x2_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
+static int sad16_x2_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
{
int i;
- int s __attribute__((aligned(16)));
- const_vector unsigned char zero = (const_vector unsigned char)vec_splat_u8(0);
+ int s;
+ const vector unsigned char zero = (const vector unsigned char)vec_splat_u8(0);
vector unsigned char *tv;
vector unsigned char pix1v, pix2v, pix2iv, avgv, t5;
vector unsigned int sad;
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));
return s;
}
-int sad16_y2_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
+static int sad16_y2_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
{
int i;
- int s __attribute__((aligned(16)));
- const_vector unsigned char zero = (const_vector unsigned char)vec_splat_u8(0);
+ int s;
+ const vector unsigned char zero = (const vector unsigned char)vec_splat_u8(0);
vector unsigned char *tv;
vector unsigned char pix1v, pix2v, pix3v, avgv, t5;
vector unsigned int sad;
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));
return s;
}
-int sad16_xy2_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
+static int sad16_xy2_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
{
int i;
- int s __attribute__((aligned(16)));
+ int s;
uint8_t *pix3 = pix2 + line_size;
- 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);
+ 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 *tv, avgv, t5;
vector unsigned char pix1v, pix2v, pix3v, pix2iv, pix3iv;
vector unsigned short pix2lv, pix2hv, pix2ilv, pix2ihv;
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]));
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);
+ 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<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));
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.
- */
+ /* 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);
+ 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);
return s;
}
-int sad16_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
+static int sad16_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
{
int i;
- int s __attribute__((aligned(16)));
- const_vector unsigned int zero = (const_vector unsigned int)vec_splat_u32(0);
- vector unsigned char perm1, perm2, *pix1v, *pix2v;
+ int s;
+ const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);
+ vector unsigned char perm1, perm2, pix1v_low, pix1v_high, pix2v_low, pix2v_high;
vector unsigned char t1, t2, t3,t4, t5;
vector unsigned int sad;
vector signed int sumdiffs;
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;
+ pix1v_high = vec_ld( 0, pix1);
+ pix1v_low = vec_ld(15, pix1);
perm2 = vec_lvsl(0, pix2);
- pix2v = (vector unsigned char *) pix2;
- t1 = vec_perm(pix1v[0], pix1v[1], perm1);
- t2 = vec_perm(pix2v[0], pix2v[1], perm2);
+ pix2v_high = vec_ld( 0, pix2);
+ pix2v_low = vec_ld(15, pix2);
+ t1 = vec_perm(pix1v_high, pix1v_low, perm1);
+ t2 = vec_perm(pix2v_high, pix2v_low, perm2);
/* Calculate a sum of abs differences vector */
t3 = vec_max(t1, t2);
return s;
}
-int sad8_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
+static int sad8_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
{
int i;
- int s __attribute__((aligned(16)));
- const_vector unsigned int zero = (const_vector unsigned int)vec_splat_u32(0);
+ int s;
+ const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);
vector unsigned char perm1, perm2, permclear, *pix1v, *pix2v;
vector unsigned char t1, t2, t3,t4, t5;
vector unsigned int sad;
sad = (vector unsigned int)vec_splat_u32(0);
- permclear = (vector unsigned char)AVV(255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0);
+ permclear = (vector unsigned char){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. */
return s;
}
-int pix_norm1_altivec(uint8_t *pix, int line_size)
+static int pix_norm1_altivec(uint8_t *pix, int line_size)
{
int i;
- int s __attribute__((aligned(16)));
- const_vector unsigned int zero = (const_vector unsigned int)vec_splat_u32(0);
+ int s;
+ const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);
vector unsigned char *tv;
vector unsigned char pixv;
vector unsigned int sv;
* AltiVec-enhanced.
* It's the sad8_altivec code above w/ squaring added.
*/
-int sse8_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
+static int sse8_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
{
int i;
- int s __attribute__((aligned(16)));
- const_vector unsigned int zero = (const_vector unsigned int)vec_splat_u32(0);
+ int s;
+ const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);
vector unsigned char perm1, perm2, permclear, *pix1v, *pix2v;
vector unsigned char t1, t2, t3,t4, t5;
vector unsigned int sum;
sum = (vector unsigned int)vec_splat_u32(0);
- permclear = (vector unsigned char)AVV(255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0);
+ permclear = (vector unsigned char){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. */
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);
* AltiVec-enhanced.
* It's the sad16_altivec code above w/ squaring added.
*/
-int sse16_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
+static int sse16_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
{
int i;
- int s __attribute__((aligned(16)));
- const_vector unsigned int zero = (const_vector unsigned int)vec_splat_u32(0);
+ int s;
+ const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);
vector unsigned char perm1, perm2, *pix1v, *pix2v;
vector unsigned char t1, t2, t3,t4, t5;
vector unsigned int sum;
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;
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);
return s;
}
-int pix_sum_altivec(uint8_t * pix, int line_size)
+static int pix_sum_altivec(uint8_t * pix, int line_size)
{
- const_vector unsigned int zero = (const_vector unsigned int)vec_splat_u32(0);
+ const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);
vector unsigned char perm, *pixv;
vector unsigned char t1;
vector unsigned int sad;
vector signed int sumdiffs;
int i;
- int s __attribute__((aligned(16)));
+ int s;
sad = (vector unsigned int)vec_splat_u32(0);
return s;
}
-void get_pixels_altivec(DCTELEM *restrict block, const uint8_t *pixels, int line_size)
+static void get_pixels_altivec(DCTELEM *restrict block, const uint8_t *pixels, int line_size)
{
int i;
vector unsigned char perm, bytes, *pixv;
- const_vector unsigned char zero = (const_vector unsigned char)vec_splat_u8(0);
+ 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.
}
}
-void diff_pixels_altivec(DCTELEM *restrict block, const uint8_t *s1,
+static void diff_pixels_altivec(DCTELEM *restrict block, const uint8_t *s1,
const uint8_t *s2, int stride)
{
int i;
vector unsigned char perm, bytes, *pixv;
- const_vector unsigned char zero = (const_vector unsigned char)vec_splat_u8(0);
+ 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.
}
}
-void add_bytes_altivec(uint8_t *dst, uint8_t *src, int w) {
-#ifdef ALTIVEC_USE_REFERENCE_C_CODE
- int i;
- for(i=0; i+7<w; i++){
- dst[i+0] += src[i+0];
- dst[i+1] += src[i+1];
- dst[i+2] += src[i+2];
- dst[i+3] += src[i+3];
- dst[i+4] += src[i+4];
- dst[i+5] += src[i+5];
- dst[i+6] += src[i+6];
- dst[i+7] += src[i+7];
- }
- for(; i<w; i++)
- dst[i+0] += src[i+0];
-#else /* ALTIVEC_USE_REFERENCE_C_CODE */
+
+static void clear_block_altivec(DCTELEM *block) {
+ LOAD_ZERO;
+ vec_st(zero_s16v, 0, block);
+ vec_st(zero_s16v, 16, block);
+ vec_st(zero_s16v, 32, block);
+ vec_st(zero_s16v, 48, block);
+ vec_st(zero_s16v, 64, block);
+ vec_st(zero_s16v, 80, block);
+ vec_st(zero_s16v, 96, block);
+ vec_st(zero_s16v, 112, block);
+}
+
+
+static void add_bytes_altivec(uint8_t *dst, uint8_t *src, int w) {
register int i;
register vector unsigned char vdst, vsrc;
/* dst and src are 16 bytes-aligned (guaranteed) */
- for(i = 0 ; (i + 15) < w ; i++)
- {
- vdst = vec_ld(i << 4, (unsigned char*)dst);
- vsrc = vec_ld(i << 4, (unsigned char*)src);
- vdst = vec_add(vsrc, vdst);
- vec_st(vdst, i << 4, (unsigned char*)dst);
+ 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++)
- {
- dst[i] = src[i];
+ for (; (i < w) ; i++) {
+ dst[i] = src[i];
}
-#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
}
/* next one assumes that ((line_size % 16) == 0) */
void put_pixels16_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
{
-POWERPC_PERF_DECLARE(altivec_put_pixels16_num, 1);
-#ifdef ALTIVEC_USE_REFERENCE_C_CODE
- int i;
-
-POWERPC_PERF_START_COUNT(altivec_put_pixels16_num, 1);
-
- for(i=0; i<h; i++) {
- *((uint32_t*)(block)) = LD32(pixels);
- *((uint32_t*)(block+4)) = LD32(pixels+4);
- *((uint32_t*)(block+8)) = LD32(pixels+8);
- *((uint32_t*)(block+12)) = LD32(pixels+12);
- pixels+=line_size;
- block +=line_size;
- }
-
-POWERPC_PERF_STOP_COUNT(altivec_put_pixels16_num, 1);
-
-#else /* ALTIVEC_USE_REFERENCE_C_CODE */
register vector unsigned char pixelsv1, pixelsv2;
register vector unsigned char pixelsv1B, pixelsv2B;
register vector unsigned char pixelsv1C, pixelsv2C;
register int line_size_3 = line_size + line_size_2;
register int line_size_4 = line_size << 2;
-POWERPC_PERF_START_COUNT(altivec_put_pixels16_num, 1);
// hand-unrolling the loop by 4 gains about 15%
// mininum execution time goes from 74 to 60 cycles
// it's faster than -funroll-loops, but using
// -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++) {
- pixelsv1 = vec_ld(0, (unsigned char*)pixels);
- pixelsv2 = vec_ld(16, (unsigned char*)pixels);
- vec_st(vec_perm(pixelsv1, pixelsv2, perm),
- 0, (unsigned char*)block);
- pixels+=line_size;
- block +=line_size;
+ for (i = 0; i < h; i++) {
+ pixelsv1 = vec_ld(0, pixels);
+ pixelsv2 = vec_ld(16, pixels);
+ vec_st(vec_perm(pixelsv1, pixelsv2, perm),
+ 0, block);
+ pixels+=line_size;
+ block +=line_size;
}
#else
- 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);
- pixelsv2B = vec_ld(15 + line_size, (unsigned char*)pixels);
- pixelsv1C = vec_ld(line_size_2, (unsigned char*)pixels);
- pixelsv2C = vec_ld(15 + line_size_2, (unsigned char*)pixels);
- pixelsv1D = vec_ld(line_size_3, (unsigned char*)pixels);
- pixelsv2D = vec_ld(15 + line_size_3, (unsigned char*)pixels);
- vec_st(vec_perm(pixelsv1, pixelsv2, perm),
- 0, (unsigned char*)block);
- vec_st(vec_perm(pixelsv1B, pixelsv2B, perm),
- line_size, (unsigned char*)block);
- vec_st(vec_perm(pixelsv1C, pixelsv2C, perm),
- line_size_2, (unsigned char*)block);
- vec_st(vec_perm(pixelsv1D, pixelsv2D, perm),
- line_size_3, (unsigned char*)block);
- pixels+=line_size_4;
- block +=line_size_4;
+ for (i = 0; i < h; i += 4) {
+ pixelsv1 = vec_ld( 0, pixels);
+ pixelsv2 = vec_ld(15, pixels);
+ pixelsv1B = vec_ld(line_size, pixels);
+ pixelsv2B = vec_ld(15 + line_size, pixels);
+ pixelsv1C = vec_ld(line_size_2, pixels);
+ pixelsv2C = vec_ld(15 + line_size_2, pixels);
+ pixelsv1D = vec_ld(line_size_3, pixels);
+ pixelsv2D = vec_ld(15 + line_size_3, pixels);
+ vec_st(vec_perm(pixelsv1, pixelsv2, perm),
+ 0, (unsigned char*)block);
+ vec_st(vec_perm(pixelsv1B, pixelsv2B, perm),
+ line_size, (unsigned char*)block);
+ vec_st(vec_perm(pixelsv1C, pixelsv2C, perm),
+ line_size_2, (unsigned char*)block);
+ vec_st(vec_perm(pixelsv1D, pixelsv2D, perm),
+ line_size_3, (unsigned char*)block);
+ pixels+=line_size_4;
+ block +=line_size_4;
}
#endif
-POWERPC_PERF_STOP_COUNT(altivec_put_pixels16_num, 1);
-
-#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
}
/* next one assumes that ((line_size % 16) == 0) */
#define op_avg(a,b) a = ( ((a)|(b)) - ((((a)^(b))&0xFEFEFEFEUL)>>1) )
void avg_pixels16_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
{
-POWERPC_PERF_DECLARE(altivec_avg_pixels16_num, 1);
-#ifdef ALTIVEC_USE_REFERENCE_C_CODE
- int i;
-
-POWERPC_PERF_START_COUNT(altivec_avg_pixels16_num, 1);
-
- for(i=0; i<h; i++) {
- op_avg(*((uint32_t*)(block)),LD32(pixels));
- op_avg(*((uint32_t*)(block+4)),LD32(pixels+4));
- op_avg(*((uint32_t*)(block+8)),LD32(pixels+8));
- op_avg(*((uint32_t*)(block+12)),LD32(pixels+12));
- pixels+=line_size;
- block +=line_size;
- }
-
-POWERPC_PERF_STOP_COUNT(altivec_avg_pixels16_num, 1);
-
-#else /* ALTIVEC_USE_REFERENCE_C_CODE */
register vector unsigned char pixelsv1, pixelsv2, pixelsv, blockv;
register vector unsigned char perm = vec_lvsl(0, pixels);
int i;
-POWERPC_PERF_START_COUNT(altivec_avg_pixels16_num, 1);
-
- 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);
- pixelsv = vec_perm(pixelsv1, pixelsv2, perm);
- blockv = vec_avg(blockv,pixelsv);
- vec_st(blockv, 0, (unsigned char*)block);
- pixels+=line_size;
- block +=line_size;
+ for (i = 0; i < h; i++) {
+ pixelsv1 = vec_ld( 0, pixels);
+ pixelsv2 = vec_ld(16,pixels);
+ blockv = vec_ld(0, block);
+ pixelsv = vec_perm(pixelsv1, pixelsv2, perm);
+ blockv = vec_avg(blockv,pixelsv);
+ vec_st(blockv, 0, (unsigned char*)block);
+ pixels+=line_size;
+ block +=line_size;
}
-
-POWERPC_PERF_STOP_COUNT(altivec_avg_pixels16_num, 1);
-
-#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
}
/* next one assumes that ((line_size % 8) == 0) */
-void avg_pixels8_altivec(uint8_t * block, const uint8_t * pixels, int line_size, int h)
+static void avg_pixels8_altivec(uint8_t * block, const uint8_t * pixels, int line_size, int h)
{
-POWERPC_PERF_DECLARE(altivec_avg_pixels8_num, 1);
-#ifdef ALTIVEC_USE_REFERENCE_C_CODE
- int i;
-POWERPC_PERF_START_COUNT(altivec_avg_pixels8_num, 1);
- for (i = 0; i < h; i++) {
- *((uint32_t *) (block)) =
- (((*((uint32_t *) (block))) |
- ((((const struct unaligned_32 *) (pixels))->l))) -
- ((((*((uint32_t *) (block))) ^
- ((((const struct unaligned_32 *) (pixels))->
- l))) & 0xFEFEFEFEUL) >> 1));
- *((uint32_t *) (block + 4)) =
- (((*((uint32_t *) (block + 4))) |
- ((((const struct unaligned_32 *) (pixels + 4))->l))) -
- ((((*((uint32_t *) (block + 4))) ^
- ((((const struct unaligned_32 *) (pixels +
- 4))->
- l))) & 0xFEFEFEFEUL) >> 1));
- pixels += line_size;
- block += line_size;
- }
-POWERPC_PERF_STOP_COUNT(altivec_avg_pixels8_num, 1);
-
-#else /* ALTIVEC_USE_REFERENCE_C_CODE */
register vector unsigned char pixelsv1, pixelsv2, pixelsv, blockv;
int i;
-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)
- */
- int rightside = ((unsigned long)block & 0x0000000F);
-
- blockv = vec_ld(0, block);
- pixelsv1 = vec_ld(0, (unsigned char*)pixels);
- pixelsv2 = vec_ld(16, (unsigned char*)pixels);
- pixelsv = vec_perm(pixelsv1, pixelsv2, vec_lvsl(0, pixels));
-
- if (rightside)
- {
- pixelsv = vec_perm(blockv, pixelsv, vcprm(0,1,s0,s1));
- }
- else
- {
- pixelsv = vec_perm(blockv, pixelsv, vcprm(s0,s1,2,3));
- }
-
- blockv = vec_avg(blockv, pixelsv);
-
- vec_st(blockv, 0, block);
-
- pixels += line_size;
- block += line_size;
- }
+ /* 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);
+ pixelsv1 = vec_ld( 0, pixels);
+ pixelsv2 = vec_ld(16, pixels);
+ pixelsv = vec_perm(pixelsv1, pixelsv2, vec_lvsl(0, pixels));
+
+ if (rightside) {
+ pixelsv = vec_perm(blockv, pixelsv, vcprm(0,1,s0,s1));
+ } else {
+ pixelsv = vec_perm(blockv, pixelsv, vcprm(s0,s1,2,3));
+ }
-POWERPC_PERF_STOP_COUNT(altivec_avg_pixels8_num, 1);
+ blockv = vec_avg(blockv, pixelsv);
-#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
+ vec_st(blockv, 0, block);
+
+ pixels += line_size;
+ block += line_size;
+ }
}
/* next one assumes that ((line_size % 8) == 0) */
-void put_pixels8_xy2_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
+static void put_pixels8_xy2_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
{
-POWERPC_PERF_DECLARE(altivec_put_pixels8_xy2_num, 1);
-#ifdef ALTIVEC_USE_REFERENCE_C_CODE
- int j;
-POWERPC_PERF_START_COUNT(altivec_put_pixels8_xy2_num, 1);
- for (j = 0; j < 2; j++) {
- int i;
- const uint32_t a = (((const struct unaligned_32 *) (pixels))->l);
- const uint32_t b =
- (((const struct unaligned_32 *) (pixels + 1))->l);
- uint32_t l0 =
- (a & 0x03030303UL) + (b & 0x03030303UL) + 0x02020202UL;
- uint32_t h0 =
- ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
- uint32_t l1, h1;
- pixels += line_size;
- for (i = 0; i < h; i += 2) {
- uint32_t a = (((const struct unaligned_32 *) (pixels))->l);
- uint32_t b = (((const struct unaligned_32 *) (pixels + 1))->l);
- l1 = (a & 0x03030303UL) + (b & 0x03030303UL);
- h1 = ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
- *((uint32_t *) block) =
- h0 + h1 + (((l0 + l1) >> 2) & 0x0F0F0F0FUL);
- pixels += line_size;
+ 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 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) {
+ pixelsv2 = temp2;
+ } else {
+ pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
+ }
+ pixelsv1 = vec_mergeh(vczero, pixelsv1);
+ pixelsv2 = vec_mergeh(vczero, pixelsv2);
+ pixelssum1 = vec_add((vector unsigned short)pixelsv1,
+ (vector unsigned short)pixelsv2);
+ pixelssum1 = vec_add(pixelssum1, vctwo);
+
+ for (i = 0; i < h ; i++) {
+ int rightside = ((unsigned long)block & 0x0000000F);
+ blockv = vec_ld(0, block);
+
+ 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) {
+ pixelsv2 = temp2;
+ } else {
+ pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
+ }
+
+ pixelsv1 = vec_mergeh(vczero, pixelsv1);
+ pixelsv2 = vec_mergeh(vczero, pixelsv2);
+ pixelssum2 = vec_add((vector unsigned short)pixelsv1,
+ (vector unsigned short)pixelsv2);
+ temp3 = vec_add(pixelssum1, pixelssum2);
+ temp3 = vec_sra(temp3, vctwo);
+ pixelssum1 = vec_add(pixelssum2, vctwo);
+ pixelsavg = vec_packsu(temp3, (vector unsigned short) vczero);
+
+ if (rightside) {
+ blockv = vec_perm(blockv, pixelsavg, vcprm(0, 1, s0, s1));
+ } else {
+ blockv = vec_perm(blockv, pixelsavg, vcprm(s0, s1, 2, 3));
+ }
+
+ vec_st(blockv, 0, block);
+
block += line_size;
- a = (((const struct unaligned_32 *) (pixels))->l);
- b = (((const struct unaligned_32 *) (pixels + 1))->l);
- l0 = (a & 0x03030303UL) + (b & 0x03030303UL) + 0x02020202UL;
- h0 = ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
- *((uint32_t *) block) =
- h0 + h1 + (((l0 + l1) >> 2) & 0x0F0F0F0FUL);
pixels += line_size;
- block += line_size;
- } pixels += 4 - line_size * (h + 1);
- block += 4 - line_size * h;
}
-
-POWERPC_PERF_STOP_COUNT(altivec_put_pixels8_xy2_num, 1);
-
-#else /* ALTIVEC_USE_REFERENCE_C_CODE */
- 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 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)
- {
- pixelsv2 = temp2;
- }
- else
- {
- pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
- }
- pixelsv1 = vec_mergeh(vczero, pixelsv1);
- pixelsv2 = vec_mergeh(vczero, pixelsv2);
- pixelssum1 = vec_add((vector unsigned short)pixelsv1,
- (vector unsigned short)pixelsv2);
- pixelssum1 = vec_add(pixelssum1, vctwo);
-
-POWERPC_PERF_START_COUNT(altivec_put_pixels8_xy2_num, 1);
- for (i = 0; i < h ; i++) {
- int rightside = ((unsigned long)block & 0x0000000F);
- blockv = vec_ld(0, block);
-
- 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)
- {
- pixelsv2 = temp2;
- }
- else
- {
- pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
- }
-
- pixelsv1 = vec_mergeh(vczero, pixelsv1);
- pixelsv2 = vec_mergeh(vczero, pixelsv2);
- pixelssum2 = vec_add((vector unsigned short)pixelsv1,
- (vector unsigned short)pixelsv2);
- temp3 = vec_add(pixelssum1, pixelssum2);
- temp3 = vec_sra(temp3, vctwo);
- pixelssum1 = vec_add(pixelssum2, vctwo);
- pixelsavg = vec_packsu(temp3, (vector unsigned short) vczero);
-
- if (rightside)
- {
- blockv = vec_perm(blockv, pixelsavg, vcprm(0, 1, s0, s1));
- }
- else
- {
- blockv = vec_perm(blockv, pixelsavg, vcprm(s0, s1, 2, 3));
- }
-
- vec_st(blockv, 0, block);
-
- block += line_size;
- pixels += line_size;
- }
-
-POWERPC_PERF_STOP_COUNT(altivec_put_pixels8_xy2_num, 1);
-#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
}
/* next one assumes that ((line_size % 8) == 0) */
-void put_no_rnd_pixels8_xy2_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
+static void put_no_rnd_pixels8_xy2_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
{
-POWERPC_PERF_DECLARE(altivec_put_no_rnd_pixels8_xy2_num, 1);
-#ifdef ALTIVEC_USE_REFERENCE_C_CODE
- int j;
-POWERPC_PERF_START_COUNT(altivec_put_no_rnd_pixels8_xy2_num, 1);
- for (j = 0; j < 2; j++) {
- int i;
- const uint32_t a = (((const struct unaligned_32 *) (pixels))->l);
- const uint32_t b =
- (((const struct unaligned_32 *) (pixels + 1))->l);
- uint32_t l0 =
- (a & 0x03030303UL) + (b & 0x03030303UL) + 0x01010101UL;
- uint32_t h0 =
- ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
- uint32_t l1, h1;
- pixels += line_size;
- for (i = 0; i < h; i += 2) {
- uint32_t a = (((const struct unaligned_32 *) (pixels))->l);
- uint32_t b = (((const struct unaligned_32 *) (pixels + 1))->l);
- l1 = (a & 0x03030303UL) + (b & 0x03030303UL);
- h1 = ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
- *((uint32_t *) block) =
- h0 + h1 + (((l0 + l1) >> 2) & 0x0F0F0F0FUL);
- pixels += line_size;
+ 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 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);
+
+ 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) {
+ pixelsv2 = temp2;
+ } else {
+ pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
+ }
+ pixelsv1 = vec_mergeh(vczero, pixelsv1);
+ pixelsv2 = vec_mergeh(vczero, pixelsv2);
+ pixelssum1 = vec_add((vector unsigned short)pixelsv1,
+ (vector unsigned short)pixelsv2);
+ pixelssum1 = vec_add(pixelssum1, vcone);
+
+ for (i = 0; i < h ; i++) {
+ int rightside = ((unsigned long)block & 0x0000000F);
+ blockv = vec_ld(0, block);
+
+ 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) {
+ pixelsv2 = temp2;
+ } else {
+ pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
+ }
+
+ pixelsv1 = vec_mergeh(vczero, pixelsv1);
+ pixelsv2 = vec_mergeh(vczero, pixelsv2);
+ pixelssum2 = vec_add((vector unsigned short)pixelsv1,
+ (vector unsigned short)pixelsv2);
+ temp3 = vec_add(pixelssum1, pixelssum2);
+ temp3 = vec_sra(temp3, vctwo);
+ pixelssum1 = vec_add(pixelssum2, vcone);
+ pixelsavg = vec_packsu(temp3, (vector unsigned short) vczero);
+
+ if (rightside) {
+ blockv = vec_perm(blockv, pixelsavg, vcprm(0, 1, s0, s1));
+ } else {
+ blockv = vec_perm(blockv, pixelsavg, vcprm(s0, s1, 2, 3));
+ }
+
+ vec_st(blockv, 0, block);
+
block += line_size;
- a = (((const struct unaligned_32 *) (pixels))->l);
- b = (((const struct unaligned_32 *) (pixels + 1))->l);
- l0 = (a & 0x03030303UL) + (b & 0x03030303UL) + 0x01010101UL;
- h0 = ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
- *((uint32_t *) block) =
- h0 + h1 + (((l0 + l1) >> 2) & 0x0F0F0F0FUL);
pixels += line_size;
- block += line_size;
- } pixels += 4 - line_size * (h + 1);
- block += 4 - line_size * h;
}
-
-POWERPC_PERF_STOP_COUNT(altivec_put_no_rnd_pixels8_xy2_num, 1);
-
-#else /* ALTIVEC_USE_REFERENCE_C_CODE */
- 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 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);
-
- 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)
- {
- pixelsv2 = temp2;
- }
- else
- {
- pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
- }
- pixelsv1 = vec_mergeh(vczero, pixelsv1);
- pixelsv2 = vec_mergeh(vczero, pixelsv2);
- pixelssum1 = vec_add((vector unsigned short)pixelsv1,
- (vector unsigned short)pixelsv2);
- pixelssum1 = vec_add(pixelssum1, vcone);
-
-POWERPC_PERF_START_COUNT(altivec_put_no_rnd_pixels8_xy2_num, 1);
- for (i = 0; i < h ; i++) {
- int rightside = ((unsigned long)block & 0x0000000F);
- blockv = vec_ld(0, block);
-
- 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)
- {
- pixelsv2 = temp2;
- }
- else
- {
- pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
- }
-
- pixelsv1 = vec_mergeh(vczero, pixelsv1);
- pixelsv2 = vec_mergeh(vczero, pixelsv2);
- pixelssum2 = vec_add((vector unsigned short)pixelsv1,
- (vector unsigned short)pixelsv2);
- temp3 = vec_add(pixelssum1, pixelssum2);
- temp3 = vec_sra(temp3, vctwo);
- pixelssum1 = vec_add(pixelssum2, vcone);
- pixelsavg = vec_packsu(temp3, (vector unsigned short) vczero);
-
- if (rightside)
- {
- blockv = vec_perm(blockv, pixelsavg, vcprm(0, 1, s0, s1));
- }
- else
- {
- blockv = vec_perm(blockv, pixelsavg, vcprm(s0, s1, 2, 3));
- }
-
- vec_st(blockv, 0, block);
-
- block += line_size;
- pixels += line_size;
- }
-
-POWERPC_PERF_STOP_COUNT(altivec_put_no_rnd_pixels8_xy2_num, 1);
-#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
}
/* next one assumes that ((line_size % 16) == 0) */
-void put_pixels16_xy2_altivec(uint8_t * block, const uint8_t * pixels, int line_size, int h)
+static void put_pixels16_xy2_altivec(uint8_t * block, const uint8_t * pixels, int line_size, int h)
{
-POWERPC_PERF_DECLARE(altivec_put_pixels16_xy2_num, 1);
-#ifdef ALTIVEC_USE_REFERENCE_C_CODE
- int j;
-POWERPC_PERF_START_COUNT(altivec_put_pixels16_xy2_num, 1);
- for (j = 0; j < 4; j++) {
- int i;
- const uint32_t a = (((const struct unaligned_32 *) (pixels))->l);
- const uint32_t b =
- (((const struct unaligned_32 *) (pixels + 1))->l);
- uint32_t l0 =
- (a & 0x03030303UL) + (b & 0x03030303UL) + 0x02020202UL;
- uint32_t h0 =
- ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
- uint32_t l1, h1;
- pixels += line_size;
- for (i = 0; i < h; i += 2) {
- uint32_t a = (((const struct unaligned_32 *) (pixels))->l);
- uint32_t b = (((const struct unaligned_32 *) (pixels + 1))->l);
- l1 = (a & 0x03030303UL) + (b & 0x03030303UL);
- h1 = ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
- *((uint32_t *) block) =
- h0 + h1 + (((l0 + l1) >> 2) & 0x0F0F0F0FUL);
- pixels += line_size;
+ register int i;
+ 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);
+
+ 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) {
+ pixelsv2 = temp2;
+ } else {
+ pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
+ }
+ pixelsv3 = vec_mergel(vczero, pixelsv1);
+ pixelsv4 = vec_mergel(vczero, pixelsv2);
+ pixelsv1 = vec_mergeh(vczero, pixelsv1);
+ pixelsv2 = vec_mergeh(vczero, pixelsv2);
+ pixelssum3 = vec_add((vector unsigned short)pixelsv3,
+ (vector unsigned short)pixelsv4);
+ pixelssum3 = vec_add(pixelssum3, vctwo);
+ pixelssum1 = vec_add((vector unsigned short)pixelsv1,
+ (vector unsigned short)pixelsv2);
+ pixelssum1 = vec_add(pixelssum1, vctwo);
+
+ for (i = 0; i < h ; i++) {
+ blockv = vec_ld(0, block);
+
+ 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) {
+ pixelsv2 = temp2;
+ } else {
+ pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
+ }
+
+ pixelsv3 = vec_mergel(vczero, pixelsv1);
+ pixelsv4 = vec_mergel(vczero, pixelsv2);
+ pixelsv1 = vec_mergeh(vczero, pixelsv1);
+ pixelsv2 = vec_mergeh(vczero, pixelsv2);
+
+ pixelssum4 = vec_add((vector unsigned short)pixelsv3,
+ (vector unsigned short)pixelsv4);
+ pixelssum2 = vec_add((vector unsigned short)pixelsv1,
+ (vector unsigned short)pixelsv2);
+ temp4 = vec_add(pixelssum3, pixelssum4);
+ temp4 = vec_sra(temp4, vctwo);
+ temp3 = vec_add(pixelssum1, pixelssum2);
+ temp3 = vec_sra(temp3, vctwo);
+
+ pixelssum3 = vec_add(pixelssum4, vctwo);
+ pixelssum1 = vec_add(pixelssum2, vctwo);
+
+ blockv = vec_packsu(temp3, temp4);
+
+ vec_st(blockv, 0, block);
+
block += line_size;
- a = (((const struct unaligned_32 *) (pixels))->l);
- b = (((const struct unaligned_32 *) (pixels + 1))->l);
- l0 = (a & 0x03030303UL) + (b & 0x03030303UL) + 0x02020202UL;
- h0 = ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
- *((uint32_t *) block) =
- h0 + h1 + (((l0 + l1) >> 2) & 0x0F0F0F0FUL);
pixels += line_size;
- block += line_size;
- } pixels += 4 - line_size * (h + 1);
- block += 4 - line_size * h;
}
-
-POWERPC_PERF_STOP_COUNT(altivec_put_pixels16_xy2_num, 1);
-
-#else /* ALTIVEC_USE_REFERENCE_C_CODE */
- 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 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);
-
-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)
- {
- pixelsv2 = temp2;
- }
- else
- {
- pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
- }
- pixelsv3 = vec_mergel(vczero, pixelsv1);
- pixelsv4 = vec_mergel(vczero, pixelsv2);
- pixelsv1 = vec_mergeh(vczero, pixelsv1);
- pixelsv2 = vec_mergeh(vczero, pixelsv2);
- pixelssum3 = vec_add((vector unsigned short)pixelsv3,
- (vector unsigned short)pixelsv4);
- pixelssum3 = vec_add(pixelssum3, vctwo);
- pixelssum1 = vec_add((vector unsigned short)pixelsv1,
- (vector unsigned short)pixelsv2);
- pixelssum1 = vec_add(pixelssum1, vctwo);
-
- for (i = 0; i < h ; i++) {
- blockv = vec_ld(0, block);
-
- 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)
- {
- pixelsv2 = temp2;
- }
- else
- {
- pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
- }
-
- pixelsv3 = vec_mergel(vczero, pixelsv1);
- pixelsv4 = vec_mergel(vczero, pixelsv2);
- pixelsv1 = vec_mergeh(vczero, pixelsv1);
- pixelsv2 = vec_mergeh(vczero, pixelsv2);
-
- pixelssum4 = vec_add((vector unsigned short)pixelsv3,
- (vector unsigned short)pixelsv4);
- pixelssum2 = vec_add((vector unsigned short)pixelsv1,
- (vector unsigned short)pixelsv2);
- temp4 = vec_add(pixelssum3, pixelssum4);
- temp4 = vec_sra(temp4, vctwo);
- temp3 = vec_add(pixelssum1, pixelssum2);
- temp3 = vec_sra(temp3, vctwo);
-
- pixelssum3 = vec_add(pixelssum4, vctwo);
- pixelssum1 = vec_add(pixelssum2, vctwo);
-
- blockv = vec_packsu(temp3, temp4);
-
- vec_st(blockv, 0, block);
-
- block += line_size;
- pixels += line_size;
- }
-
-POWERPC_PERF_STOP_COUNT(altivec_put_pixels16_xy2_num, 1);
-#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
}
/* next one assumes that ((line_size % 16) == 0) */
-void put_no_rnd_pixels16_xy2_altivec(uint8_t * block, const uint8_t * pixels, int line_size, int h)
+static void put_no_rnd_pixels16_xy2_altivec(uint8_t * block, const uint8_t * pixels, int line_size, int h)
{
-POWERPC_PERF_DECLARE(altivec_put_no_rnd_pixels16_xy2_num, 1);
-#ifdef ALTIVEC_USE_REFERENCE_C_CODE
- int j;
-POWERPC_PERF_START_COUNT(altivec_put_no_rnd_pixels16_xy2_num, 1);
- for (j = 0; j < 4; j++) {
- int i;
- const uint32_t a = (((const struct unaligned_32 *) (pixels))->l);
- const uint32_t b =
- (((const struct unaligned_32 *) (pixels + 1))->l);
- uint32_t l0 =
- (a & 0x03030303UL) + (b & 0x03030303UL) + 0x01010101UL;
- uint32_t h0 =
- ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
- uint32_t l1, h1;
- pixels += line_size;
- for (i = 0; i < h; i += 2) {
- uint32_t a = (((const struct unaligned_32 *) (pixels))->l);
- uint32_t b = (((const struct unaligned_32 *) (pixels + 1))->l);
- l1 = (a & 0x03030303UL) + (b & 0x03030303UL);
- h1 = ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
- *((uint32_t *) block) =
- h0 + h1 + (((l0 + l1) >> 2) & 0x0F0F0F0FUL);
- pixels += line_size;
+ register int i;
+ 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);
+
+ 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) {
+ pixelsv2 = temp2;
+ } else {
+ pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
+ }
+ pixelsv3 = vec_mergel(vczero, pixelsv1);
+ pixelsv4 = vec_mergel(vczero, pixelsv2);
+ pixelsv1 = vec_mergeh(vczero, pixelsv1);
+ pixelsv2 = vec_mergeh(vczero, pixelsv2);
+ pixelssum3 = vec_add((vector unsigned short)pixelsv3,
+ (vector unsigned short)pixelsv4);
+ pixelssum3 = vec_add(pixelssum3, vcone);
+ pixelssum1 = vec_add((vector unsigned short)pixelsv1,
+ (vector unsigned short)pixelsv2);
+ pixelssum1 = vec_add(pixelssum1, vcone);
+
+ for (i = 0; i < h ; i++) {
+ blockv = vec_ld(0, block);
+
+ 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) {
+ pixelsv2 = temp2;
+ } else {
+ pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
+ }
+
+ pixelsv3 = vec_mergel(vczero, pixelsv1);
+ pixelsv4 = vec_mergel(vczero, pixelsv2);
+ pixelsv1 = vec_mergeh(vczero, pixelsv1);
+ pixelsv2 = vec_mergeh(vczero, pixelsv2);
+
+ pixelssum4 = vec_add((vector unsigned short)pixelsv3,
+ (vector unsigned short)pixelsv4);
+ pixelssum2 = vec_add((vector unsigned short)pixelsv1,
+ (vector unsigned short)pixelsv2);
+ temp4 = vec_add(pixelssum3, pixelssum4);
+ temp4 = vec_sra(temp4, vctwo);
+ temp3 = vec_add(pixelssum1, pixelssum2);
+ temp3 = vec_sra(temp3, vctwo);
+
+ pixelssum3 = vec_add(pixelssum4, vcone);
+ pixelssum1 = vec_add(pixelssum2, vcone);
+
+ blockv = vec_packsu(temp3, temp4);
+
+ vec_st(blockv, 0, block);
+
block += line_size;
- a = (((const struct unaligned_32 *) (pixels))->l);
- b = (((const struct unaligned_32 *) (pixels + 1))->l);
- l0 = (a & 0x03030303UL) + (b & 0x03030303UL) + 0x01010101UL;
- h0 = ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
- *((uint32_t *) block) =
- h0 + h1 + (((l0 + l1) >> 2) & 0x0F0F0F0FUL);
pixels += line_size;
- block += line_size;
- } pixels += 4 - line_size * (h + 1);
- block += 4 - line_size * h;
}
-
-POWERPC_PERF_STOP_COUNT(altivec_put_no_rnd_pixels16_xy2_num, 1);
-
-#else /* ALTIVEC_USE_REFERENCE_C_CODE */
- 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 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);
-
-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)
- {
- pixelsv2 = temp2;
- }
- else
- {
- pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
- }
- pixelsv3 = vec_mergel(vczero, pixelsv1);
- pixelsv4 = vec_mergel(vczero, pixelsv2);
- pixelsv1 = vec_mergeh(vczero, pixelsv1);
- pixelsv2 = vec_mergeh(vczero, pixelsv2);
- pixelssum3 = vec_add((vector unsigned short)pixelsv3,
- (vector unsigned short)pixelsv4);
- pixelssum3 = vec_add(pixelssum3, vcone);
- pixelssum1 = vec_add((vector unsigned short)pixelsv1,
- (vector unsigned short)pixelsv2);
- pixelssum1 = vec_add(pixelssum1, vcone);
-
- for (i = 0; i < h ; i++) {
- blockv = vec_ld(0, block);
-
- 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)
- {
- pixelsv2 = temp2;
- }
- else
- {
- pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
- }
-
- pixelsv3 = vec_mergel(vczero, pixelsv1);
- pixelsv4 = vec_mergel(vczero, pixelsv2);
- pixelsv1 = vec_mergeh(vczero, pixelsv1);
- pixelsv2 = vec_mergeh(vczero, pixelsv2);
-
- pixelssum4 = vec_add((vector unsigned short)pixelsv3,
- (vector unsigned short)pixelsv4);
- pixelssum2 = vec_add((vector unsigned short)pixelsv1,
- (vector unsigned short)pixelsv2);
- temp4 = vec_add(pixelssum3, pixelssum4);
- temp4 = vec_sra(temp4, vctwo);
- temp3 = vec_add(pixelssum1, pixelssum2);
- temp3 = vec_sra(temp3, vctwo);
-
- pixelssum3 = vec_add(pixelssum4, vcone);
- pixelssum1 = vec_add(pixelssum2, vcone);
-
- blockv = vec_packsu(temp3, temp4);
-
- vec_st(blockv, 0, block);
-
- block += line_size;
- pixels += line_size;
- }
-
-POWERPC_PERF_STOP_COUNT(altivec_put_no_rnd_pixels16_xy2_num, 1);
-#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
}
-int hadamard8_diff8x8_altivec(/*MpegEncContext*/ void *s, uint8_t *dst, uint8_t *src, int stride, int h){
-POWERPC_PERF_DECLARE(altivec_hadamard8_diff8x8_num, 1);
- int sum;
- register const_vector unsigned char vzero = (const_vector unsigned char)vec_splat_u8(0);
- register vector signed short temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7;
-POWERPC_PERF_START_COUNT(altivec_hadamard8_diff8x8_num, 1);
- {
- register const_vector signed short vprod1 = (const_vector signed short)AVV( 1,-1, 1,-1, 1,-1, 1,-1);
- register const_vector signed short vprod2 = (const_vector signed short)AVV( 1, 1,-1,-1, 1, 1,-1,-1);
- register const_vector signed short vprod3 = (const_vector signed short)AVV( 1, 1, 1, 1,-1,-1,-1,-1);
- register const_vector unsigned char perm1 = (const_vector unsigned char)
- AVV(0x02, 0x03, 0x00, 0x01,
- 0x06, 0x07, 0x04, 0x05,
- 0x0A, 0x0B, 0x08, 0x09,
- 0x0E, 0x0F, 0x0C, 0x0D);
- register const_vector unsigned char perm2 = (const_vector unsigned char)
- AVV(0x04, 0x05, 0x06, 0x07,
- 0x00, 0x01, 0x02, 0x03,
- 0x0C, 0x0D, 0x0E, 0x0F,
- 0x08, 0x09, 0x0A, 0x0B);
- register const_vector unsigned char perm3 = (const_vector unsigned char)
- AVV(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, src2, srcO; \
- register vector unsigned char dst1, dst2, dstO; \
- register vector signed short srcV, dstV; \
- register vector signed short but0, but1, but2, op1, op2, op3; \
- src1 = vec_ld(stride * i, src); \
- if ((((stride * i) + (unsigned long)src) & 0x0000000F) > 8) \
- src2 = vec_ld((stride * i) + 16, src); \
- srcO = vec_perm(src1, src2, vec_lvsl(stride * i, src)); \
- dst1 = vec_ld(stride * i, dst); \
- if ((((stride * i) + (unsigned long)dst) & 0x0000000F) > 8) \
- dst2 = vec_ld((stride * i) + 16, dst); \
- 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, \
- (vector signed char)srcO); \
- dstV = \
- (vector signed short)vec_mergeh((vector signed char)vzero, \
- (vector signed char)dstO); \
- /* substractions inside the first butterfly */ \
- but0 = vec_sub(srcV, dstV); \
- op1 = vec_perm(but0, but0, perm1); \
- but1 = vec_mladd(but0, vprod1, op1); \
- op2 = vec_perm(but1, but1, perm2); \
- but2 = vec_mladd(but1, vprod2, op2); \
- op3 = vec_perm(but2, but2, perm3); \
- res = vec_mladd(but2, vprod3, op3); \
+static int hadamard8_diff8x8_altivec(/*MpegEncContext*/ void *s, uint8_t *dst, uint8_t *src, int stride, int h){
+ int sum;
+ register const vector unsigned char vzero =
+ (const vector unsigned char)vec_splat_u8(0);
+ register vector signed short temp0, temp1, temp2, temp3, temp4,
+ temp5, temp6, temp7;
+ {
+ register const vector signed short vprod1 =(const vector signed short)
+ { 1,-1, 1,-1, 1,-1, 1,-1 };
+ register const vector signed short vprod2 =(const vector signed short)
+ { 1, 1,-1,-1, 1, 1,-1,-1 };
+ register const vector signed short vprod3 =(const vector signed short)
+ { 1, 1, 1, 1,-1,-1,-1,-1 };
+ register const vector unsigned char perm1 = (const vector unsigned char)
+ {0x02, 0x03, 0x00, 0x01, 0x06, 0x07, 0x04, 0x05,
+ 0x0A, 0x0B, 0x08, 0x09, 0x0E, 0x0F, 0x0C, 0x0D};
+ register const vector unsigned char perm2 = (const vector unsigned char)
+ {0x04, 0x05, 0x06, 0x07, 0x00, 0x01, 0x02, 0x03,
+ 0x0C, 0x0D, 0x0E, 0x0F, 0x08, 0x09, 0x0A, 0x0B};
+ register const vector unsigned char perm3 = (const vector unsigned char)
+ {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, src2, srcO; \
+ register vector unsigned char dst1, dst2, dstO; \
+ register vector signed short srcV, dstV; \
+ register vector signed short but0, but1, but2, op1, op2, op3; \
+ src1 = vec_ld(stride * i, src); \
+ src2 = vec_ld((stride * i) + 15, src); \
+ srcO = vec_perm(src1, src2, vec_lvsl(stride * i, src)); \
+ dst1 = vec_ld(stride * i, dst); \
+ dst2 = vec_ld((stride * i) + 15, dst); \
+ 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, \
+ (vector signed char)srcO); \
+ dstV = (vector signed short)vec_mergeh((vector signed char)vzero, \
+ (vector signed char)dstO); \
+ /* subtractions inside the first butterfly */ \
+ but0 = vec_sub(srcV, dstV); \
+ op1 = vec_perm(but0, but0, perm1); \
+ but1 = vec_mladd(but0, vprod1, op1); \
+ op2 = vec_perm(but1, but1, perm2); \
+ but2 = vec_mladd(but1, vprod2, op2); \
+ op3 = vec_perm(but2, but2, perm3); \
+ res = vec_mladd(but2, vprod3, op3); \
}
ONEITERBUTTERFLY(0, temp0);
ONEITERBUTTERFLY(1, temp1);
ONEITERBUTTERFLY(5, temp5);
ONEITERBUTTERFLY(6, temp6);
ONEITERBUTTERFLY(7, temp7);
- }
+ }
#undef ONEITERBUTTERFLY
- {
+ {
register vector signed int vsum;
register vector signed short line0 = vec_add(temp0, temp1);
register vector signed short line1 = vec_sub(temp0, temp1);
vsum = vec_sums(vsum, (vector signed int)vzero);
vsum = vec_splat(vsum, 3);
vec_ste(vsum, 0, &sum);
- }
-POWERPC_PERF_STOP_COUNT(altivec_hadamard8_diff8x8_num, 1);
- return sum;
+ }
+ return sum;
}
/*
- 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 (arount 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) {
- int sum;
- register vector signed short
- temp0 REG_v(v0),
- temp1 REG_v(v1),
- temp2 REG_v(v2),
- temp3 REG_v(v3),
- temp4 REG_v(v4),
- temp5 REG_v(v5),
- temp6 REG_v(v6),
- temp7 REG_v(v7);
- register vector signed short
- temp0S REG_v(v8),
- temp1S REG_v(v9),
- temp2S REG_v(v10),
- temp3S REG_v(v11),
- temp4S REG_v(v12),
- temp5S REG_v(v13),
- temp6S REG_v(v14),
- temp7S REG_v(v15);
- register const_vector unsigned char vzero REG_v(v31)= (const_vector unsigned char)vec_splat_u8(0);
- {
- register const_vector signed short vprod1 REG_v(v16)= (const_vector signed short)AVV( 1,-1, 1,-1, 1,-1, 1,-1);
- register const_vector signed short vprod2 REG_v(v17)= (const_vector signed short)AVV( 1, 1,-1,-1, 1, 1,-1,-1);
- register const_vector signed short vprod3 REG_v(v18)= (const_vector signed short)AVV( 1, 1, 1, 1,-1,-1,-1,-1);
- register const_vector unsigned char perm1 REG_v(v19)= (const_vector unsigned char)
- AVV(0x02, 0x03, 0x00, 0x01,
- 0x06, 0x07, 0x04, 0x05,
- 0x0A, 0x0B, 0x08, 0x09,
- 0x0E, 0x0F, 0x0C, 0x0D);
- register const_vector unsigned char perm2 REG_v(v20)= (const_vector unsigned char)
- AVV(0x04, 0x05, 0x06, 0x07,
- 0x00, 0x01, 0x02, 0x03,
- 0x0C, 0x0D, 0x0E, 0x0F,
- 0x08, 0x09, 0x0A, 0x0B);
- register const_vector unsigned char perm3 REG_v(v21)= (const_vector unsigned char)
- AVV(0x08, 0x09, 0x0A, 0x0B,
- 0x0C, 0x0D, 0x0E, 0x0F,
- 0x00, 0x01, 0x02, 0x03,
- 0x04, 0x05, 0x06, 0x07);
-
-#define ONEITERBUTTERFLY(i, res1, res2) \
- { \
- register vector unsigned char src1 REG_v(v22), \
- src2 REG_v(v23), \
- dst1 REG_v(v24), \
- dst2 REG_v(v25), \
- srcO REG_v(v22), \
- dstO REG_v(v23); \
- \
- register vector signed short srcV REG_v(v24), \
- dstV REG_v(v25), \
- srcW REG_v(v26), \
- dstW REG_v(v27), \
- but0 REG_v(v28), \
- but0S REG_v(v29), \
- op1 REG_v(v30), \
- but1 REG_v(v22), \
- op1S REG_v(v23), \
- but1S REG_v(v24), \
- op2 REG_v(v25), \
- but2 REG_v(v26), \
- op2S REG_v(v27), \
- but2S REG_v(v28), \
- op3 REG_v(v29), \
- op3S REG_v(v30); \
- \
- src1 = vec_ld(stride * i, src); \
- src2 = vec_ld((stride * i) + 16, src); \
- srcO = vec_perm(src1, src2, vec_lvsl(stride * i, src)); \
- dst1 = vec_ld(stride * i, 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, \
- (vector signed char)srcO); \
- dstV = \
- (vector signed short)vec_mergeh((vector signed char)vzero, \
- (vector signed char)dstO); \
- srcW = \
- (vector signed short)vec_mergel((vector signed char)vzero, \
- (vector signed char)srcO); \
- dstW = \
- (vector signed short)vec_mergel((vector signed char)vzero, \
- (vector signed char)dstO); \
- /* substractions inside the first butterfly */ \
- but0 = vec_sub(srcV, dstV); \
- but0S = vec_sub(srcW, dstW); \
- op1 = vec_perm(but0, but0, perm1); \
- but1 = vec_mladd(but0, vprod1, op1); \
- op1S = vec_perm(but0S, but0S, perm1); \
- but1S = vec_mladd(but0S, vprod1, op1S); \
- op2 = vec_perm(but1, but1, perm2); \
- but2 = vec_mladd(but1, vprod2, op2); \
- op2S = vec_perm(but1S, but1S, perm2); \
- but2S = vec_mladd(but1S, vprod2, op2S); \
- op3 = vec_perm(but2, but2, perm3); \
- res1 = vec_mladd(but2, vprod3, op3); \
- op3S = vec_perm(but2S, but2S, perm3); \
- res2 = vec_mladd(but2S, vprod3, op3S); \
+ int sum;
+ register vector signed short
+ temp0 __asm__ ("v0"),
+ temp1 __asm__ ("v1"),
+ temp2 __asm__ ("v2"),
+ temp3 __asm__ ("v3"),
+ temp4 __asm__ ("v4"),
+ temp5 __asm__ ("v5"),
+ temp6 __asm__ ("v6"),
+ temp7 __asm__ ("v7");
+ register vector signed short
+ temp0S __asm__ ("v8"),
+ temp1S __asm__ ("v9"),
+ temp2S __asm__ ("v10"),
+ temp3S __asm__ ("v11"),
+ temp4S __asm__ ("v12"),
+ temp5S __asm__ ("v13"),
+ temp6S __asm__ ("v14"),
+ temp7S __asm__ ("v15");
+ register const vector unsigned char vzero __asm__ ("v31") =
+ (const vector unsigned char)vec_splat_u8(0);
+ {
+ register const vector signed short vprod1 __asm__ ("v16") =
+ (const vector signed short){ 1,-1, 1,-1, 1,-1, 1,-1 };
+ register const vector signed short vprod2 __asm__ ("v17") =
+ (const vector signed short){ 1, 1,-1,-1, 1, 1,-1,-1 };
+ register const vector signed short vprod3 __asm__ ("v18") =
+ (const vector signed short){ 1, 1, 1, 1,-1,-1,-1,-1 };
+ register const vector unsigned char perm1 __asm__ ("v19") =
+ (const vector unsigned char)
+ {0x02, 0x03, 0x00, 0x01, 0x06, 0x07, 0x04, 0x05,
+ 0x0A, 0x0B, 0x08, 0x09, 0x0E, 0x0F, 0x0C, 0x0D};
+ register const vector unsigned char perm2 __asm__ ("v20") =
+ (const vector unsigned char)
+ {0x04, 0x05, 0x06, 0x07, 0x00, 0x01, 0x02, 0x03,
+ 0x0C, 0x0D, 0x0E, 0x0F, 0x08, 0x09, 0x0A, 0x0B};
+ register const vector unsigned char perm3 __asm__ ("v21") =
+ (const vector unsigned char)
+ {0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0F,
+ 0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07};
+
+#define ONEITERBUTTERFLY(i, res1, res2) \
+ { \
+ register vector unsigned char src1 __asm__ ("v22"), \
+ src2 __asm__ ("v23"), \
+ dst1 __asm__ ("v24"), \
+ dst2 __asm__ ("v25"), \
+ srcO __asm__ ("v22"), \
+ dstO __asm__ ("v23"); \
+ \
+ register vector signed short srcV __asm__ ("v24"), \
+ dstV __asm__ ("v25"), \
+ srcW __asm__ ("v26"), \
+ dstW __asm__ ("v27"), \
+ but0 __asm__ ("v28"), \
+ but0S __asm__ ("v29"), \
+ op1 __asm__ ("v30"), \
+ but1 __asm__ ("v22"), \
+ op1S __asm__ ("v23"), \
+ but1S __asm__ ("v24"), \
+ op2 __asm__ ("v25"), \
+ but2 __asm__ ("v26"), \
+ op2S __asm__ ("v27"), \
+ but2S __asm__ ("v28"), \
+ op3 __asm__ ("v29"), \
+ op3S __asm__ ("v30"); \
+ \
+ src1 = vec_ld(stride * i, src); \
+ src2 = vec_ld((stride * i) + 16, src); \
+ srcO = vec_perm(src1, src2, vec_lvsl(stride * i, src)); \
+ dst1 = vec_ld(stride * i, 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, \
+ (vector signed char)srcO); \
+ dstV = (vector signed short)vec_mergeh((vector signed char)vzero, \
+ (vector signed char)dstO); \
+ srcW = (vector signed short)vec_mergel((vector signed char)vzero, \
+ (vector signed char)srcO); \
+ dstW = (vector signed short)vec_mergel((vector signed char)vzero, \
+ (vector signed char)dstO); \
+ /* subtractions inside the first butterfly */ \
+ but0 = vec_sub(srcV, dstV); \
+ but0S = vec_sub(srcW, dstW); \
+ op1 = vec_perm(but0, but0, perm1); \
+ but1 = vec_mladd(but0, vprod1, op1); \
+ op1S = vec_perm(but0S, but0S, perm1); \
+ but1S = vec_mladd(but0S, vprod1, op1S); \
+ op2 = vec_perm(but1, but1, perm2); \
+ but2 = vec_mladd(but1, vprod2, op2); \
+ op2S = vec_perm(but1S, but1S, perm2); \
+ but2S = vec_mladd(but1S, vprod2, op2S); \
+ op3 = vec_perm(but2, but2, perm3); \
+ res1 = vec_mladd(but2, vprod3, op3); \
+ op3S = vec_perm(but2S, but2S, perm3); \
+ res2 = vec_mladd(but2S, vprod3, op3S); \
}
ONEITERBUTTERFLY(0, temp0, temp0S);
ONEITERBUTTERFLY(1, temp1, temp1S);
ONEITERBUTTERFLY(5, temp5, temp5S);
ONEITERBUTTERFLY(6, temp6, temp6S);
ONEITERBUTTERFLY(7, temp7, temp7S);
- }
+ }
#undef ONEITERBUTTERFLY
- {
+ {
register vector signed int vsum;
register vector signed short line0S, line1S, line2S, line3S, line4S,
line5S, line6S, line7S, line0BS,line2BS,
vsum = vec_sums(vsum, (vector signed int)vzero);
vsum = vec_splat(vsum, 3);
vec_ste(vsum, 0, &sum);
- }
- return sum;
-}
-
-int hadamard8_diff16_altivec(/*MpegEncContext*/ void *s, uint8_t *dst, uint8_t *src, int stride, int h){
-POWERPC_PERF_DECLARE(altivec_hadamard8_diff16_num, 1);
- int score;
-POWERPC_PERF_START_COUNT(altivec_hadamard8_diff16_num, 1);
- score = hadamard8_diff16x8_altivec(s, dst, src, stride, 8);
- if (h==16) {
- dst += 8*stride;
- src += 8*stride;
- score += hadamard8_diff16x8_altivec(s, dst, src, stride, 8);
- }
-POWERPC_PERF_STOP_COUNT(altivec_hadamard8_diff16_num, 1);
- return score;
+ }
+ return sum;
}
-int has_altivec(void)
-{
-#ifdef __AMIGAOS4__
- ULONG result = 0;
- extern struct ExecIFace *IExec;
-
- IExec->GetCPUInfoTags(GCIT_VectorUnit, &result, TAG_DONE);
- if (result == VECTORTYPE_ALTIVEC) return 1;
- return 0;
-#else /* __AMIGAOS4__ */
-
-#ifdef CONFIG_DARWIN
- int sels[2] = {CTL_HW, HW_VECTORUNIT};
- int has_vu = 0;
- size_t len = sizeof(has_vu);
- int err;
-
- err = sysctl(sels, 2, &has_vu, &len, NULL, 0);
-
- if (err == 0) return (has_vu != 0);
-#else /* CONFIG_DARWIN */
-/* no Darwin, do it the brute-force way */
-/* this is borrowed from the libmpeg2 library */
- {
- signal (SIGILL, sigill_handler);
- if (sigsetjmp (jmpbuf, 1)) {
- signal (SIGILL, SIG_DFL);
- } else {
- canjump = 1;
-
- asm volatile ("mtspr 256, %0\n\t"
- "vand %%v0, %%v0, %%v0"
- :
- : "r" (-1));
-
- signal (SIGILL, SIG_DFL);
- return 1;
- }
+static int hadamard8_diff16_altivec(/*MpegEncContext*/ void *s, uint8_t *dst, uint8_t *src, int stride, int h){
+ int score;
+ score = hadamard8_diff16x8_altivec(s, dst, src, stride, 8);
+ if (h==16) {
+ dst += 8*stride;
+ src += 8*stride;
+ score += hadamard8_diff16x8_altivec(s, dst, src, stride, 8);
}
-#endif /* CONFIG_DARWIN */
- return 0;
-#endif /* __AMIGAOS4__ */
+ return score;
}
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));
}
/* next one assumes that ((line_size % 8) == 0) */
-void avg_pixels8_xy2_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
+static void avg_pixels8_xy2_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
{
-POWERPC_PERF_DECLARE(altivec_avg_pixels8_xy2_num, 1);
-#ifdef ALTIVEC_USE_REFERENCE_C_CODE
-
- int j;
-POWERPC_PERF_START_COUNT(altivec_avg_pixels8_xy2_num, 1);
- for (j = 0; j < 2; j++) {
- int i;
- const uint32_t a = (((const struct unaligned_32 *) (pixels))->l);
- const uint32_t b = (((const struct unaligned_32 *) (pixels + 1))->l);
- uint32_t l0 = (a & 0x03030303UL) + (b & 0x03030303UL) + 0x02020202UL;
- uint32_t h0 = ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
- uint32_t l1, h1;
- pixels += line_size;
- for (i = 0; i < h; i += 2) {
- uint32_t a = (((const struct unaligned_32 *) (pixels))->l);
- uint32_t b = (((const struct unaligned_32 *) (pixels + 1))->l);
- l1 = (a & 0x03030303UL) + (b & 0x03030303UL);
- h1 = ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
- *((uint32_t *) block) = rnd_avg32(*((uint32_t *) block), h0 + h1 + (((l0 + l1) >> 2) & 0x0F0F0F0FUL));
- pixels += line_size;
- block += line_size;
- a = (((const struct unaligned_32 *) (pixels))->l);
- b = (((const struct unaligned_32 *) (pixels + 1))->l);
- l0 = (a & 0x03030303UL) + (b & 0x03030303UL) + 0x02020202UL;
- h0 = ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
- *((uint32_t *) block) = rnd_avg32(*((uint32_t *) block), h0 + h1 + (((l0 + l1) >> 2) & 0x0F0F0F0FUL));
- pixels += line_size;
- block += line_size;
- } pixels += 4 - line_size * (h + 1);
- block += 4 - line_size * h;
- }
-POWERPC_PERF_STOP_COUNT(altivec_avg_pixels8_xy2_num, 1);
-#else /* ALTIVEC_USE_REFERENCE_C_CODE */
- register int i;
- register vector unsigned char
- pixelsv1, pixelsv2,
- pixelsavg;
- register vector unsigned char
- blockv, temp1, temp2, blocktemp;
- 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)
- {
- pixelsv2 = temp2;
- }
- else
- {
- pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
- }
- pixelsv1 = vec_mergeh(vczero, pixelsv1);
- pixelsv2 = vec_mergeh(vczero, pixelsv2);
- pixelssum1 = vec_add((vector unsigned short)pixelsv1,
- (vector unsigned short)pixelsv2);
- pixelssum1 = vec_add(pixelssum1, vctwo);
-
-POWERPC_PERF_START_COUNT(altivec_avg_pixels8_xy2_num, 1);
- for (i = 0; i < h ; i++) {
- int rightside = ((unsigned long)block & 0x0000000F);
- blockv = vec_ld(0, block);
-
- 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)
- {
- pixelsv2 = temp2;
- }
- else
- {
- pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
- }
-
- pixelsv1 = vec_mergeh(vczero, pixelsv1);
- pixelsv2 = vec_mergeh(vczero, pixelsv2);
- pixelssum2 = vec_add((vector unsigned short)pixelsv1,
- (vector unsigned short)pixelsv2);
- temp3 = vec_add(pixelssum1, pixelssum2);
- temp3 = vec_sra(temp3, vctwo);
- pixelssum1 = vec_add(pixelssum2, vctwo);
- pixelsavg = vec_packsu(temp3, (vector unsigned short) vczero);
-
- if (rightside)
- {
- blocktemp = vec_perm(blockv, pixelsavg, vcprm(0, 1, s0, s1));
- }
- else
- {
- blocktemp = vec_perm(blockv, pixelsavg, vcprm(s0, s1, 2, 3));
- }
-
- blockv = vec_avg(blocktemp, blockv);
- vec_st(blockv, 0, block);
-
- block += line_size;
- pixels += line_size;
- }
+ register int i;
+ register vector unsigned char pixelsv1, pixelsv2, pixelsavg;
+ register vector unsigned char blockv, temp1, temp2, blocktemp;
+ 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) {
+ pixelsv2 = temp2;
+ } else {
+ pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
+ }
+ pixelsv1 = vec_mergeh(vczero, pixelsv1);
+ pixelsv2 = vec_mergeh(vczero, pixelsv2);
+ pixelssum1 = vec_add((vector unsigned short)pixelsv1,
+ (vector unsigned short)pixelsv2);
+ pixelssum1 = vec_add(pixelssum1, vctwo);
+
+ for (i = 0; i < h ; i++) {
+ int rightside = ((unsigned long)block & 0x0000000F);
+ blockv = vec_ld(0, block);
+
+ 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) {
+ pixelsv2 = temp2;
+ } else {
+ pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
+ }
+
+ pixelsv1 = vec_mergeh(vczero, pixelsv1);
+ pixelsv2 = vec_mergeh(vczero, pixelsv2);
+ pixelssum2 = vec_add((vector unsigned short)pixelsv1,
+ (vector unsigned short)pixelsv2);
+ temp3 = vec_add(pixelssum1, pixelssum2);
+ temp3 = vec_sra(temp3, vctwo);
+ pixelssum1 = vec_add(pixelssum2, vctwo);
+ pixelsavg = vec_packsu(temp3, (vector unsigned short) vczero);
+
+ if (rightside) {
+ blocktemp = vec_perm(blockv, pixelsavg, vcprm(0, 1, s0, s1));
+ } else {
+ blocktemp = vec_perm(blockv, pixelsavg, vcprm(s0, s1, 2, 3));
+ }
+
+ blockv = vec_avg(blocktemp, blockv);
+ vec_st(blockv, 0, block);
-POWERPC_PERF_STOP_COUNT(altivec_avg_pixels8_xy2_num, 1);
-#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
+ block += line_size;
+ pixels += line_size;
+ }
}
void dsputil_init_altivec(DSPContext* c, AVCodecContext *avctx)
c->pix_sum = pix_sum_altivec;
c->diff_pixels = diff_pixels_altivec;
c->get_pixels = get_pixels_altivec;
-// next one disabled as it's untested.
-#if 0
+ c->clear_block = clear_block_altivec;
c->add_bytes= add_bytes_altivec;
-#endif /* 0 */
c->put_pixels_tab[0][0] = put_pixels16_altivec;
/* the two functions do the same thing, so use the same code */
c->put_no_rnd_pixels_tab[0][0] = put_pixels16_altivec;
c->hadamard8_diff[0] = hadamard8_diff16_altivec;
c->hadamard8_diff[1] = hadamard8_diff8x8_altivec;
- c->vorbis_inverse_coupling = vorbis_inverse_coupling_altivec;
+ if (CONFIG_VORBIS_DECODER)
+ c->vorbis_inverse_coupling = vorbis_inverse_coupling_altivec;
}