diff options
Diffstat (limited to 'contrib/ffmpeg/libavcodec/ppc')
18 files changed, 6579 insertions, 0 deletions
diff --git a/contrib/ffmpeg/libavcodec/ppc/dsputil_altivec.c b/contrib/ffmpeg/libavcodec/ppc/dsputil_altivec.c new file mode 100644 index 000000000..6f48893a4 --- /dev/null +++ b/contrib/ffmpeg/libavcodec/ppc/dsputil_altivec.c @@ -0,0 +1,1591 @@ +/* + * Copyright (c) 2002 Brian Foley + * Copyright (c) 2002 Dieter Shirley + * Copyright (c) 2003-2004 Romain Dolbeau <romain@dolbeau.org> + * + * This file is part of FFmpeg. + * + * FFmpeg 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 2.1 of the License, or (at your option) any later version. + * + * FFmpeg 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 FFmpeg; 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 "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) +{ + int i; + int s __attribute__((aligned(16))); + 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; + vector signed int sumdiffs; + + 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: + pix1v: pix1[0]-pix1[15] + pix2v: pix2[0]-pix2[15] pix2iv: pix2[1]-pix2[16] + */ + tv = (vector unsigned char *) pix1; + pix1v = vec_perm(tv[0], tv[1], vec_lvsl(0, pix1)); + + tv = (vector unsigned char *) &pix2[0]; + pix2v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[0])); + + tv = (vector unsigned char *) &pix2[1]; + pix2iv = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[1])); + + /* Calculate the average vector */ + avgv = vec_avg(pix2v, pix2iv); + + /* Calculate a sum of abs differences vector */ + t5 = vec_sub(vec_max(pix1v, avgv), vec_min(pix1v, avgv)); + + /* Add each 4 pixel group together and put 4 results into sad */ + sad = vec_sum4s(t5, sad); + + pix1 += line_size; + pix2 += line_size; + } + /* Sum up the four partial sums, and put the result into s */ + sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero); + sumdiffs = vec_splat(sumdiffs, 3); + vec_ste(sumdiffs, 0, &s); + + return s; +} + +int 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); + vector unsigned char *tv; + vector unsigned char pix1v, pix2v, pix3v, avgv, t5; + vector unsigned int sad; + vector signed int sumdiffs; + uint8_t *pix3 = pix2 + line_size; + + s = 0; + sad = (vector unsigned int)vec_splat_u32(0); + + /* + Due to the fact that pix3 = pix2 + line_size, the pix3 of one + iteration becomes pix2 in the next iteration. We can use this + fact to avoid a potentially expensive unaligned read, each + time around the loop. + Read unaligned pixels into our vectors. The vectors are as follows: + pix2v: pix2[0]-pix2[15] + Split the pixel vectors into shorts + */ + tv = (vector unsigned char *) &pix2[0]; + pix2v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[0])); + + for(i=0;i<h;i++) { + /* + Read unaligned pixels into our vectors. The vectors are as follows: + pix1v: pix1[0]-pix1[15] + pix3v: pix3[0]-pix3[15] + */ + tv = (vector unsigned char *) pix1; + pix1v = vec_perm(tv[0], tv[1], vec_lvsl(0, pix1)); + + tv = (vector unsigned char *) &pix3[0]; + pix3v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix3[0])); + + /* Calculate the average vector */ + avgv = vec_avg(pix2v, pix3v); + + /* Calculate a sum of abs differences vector */ + t5 = vec_sub(vec_max(pix1v, avgv), vec_min(pix1v, avgv)); + + /* Add each 4 pixel group together and put 4 results into sad */ + sad = vec_sum4s(t5, sad); + + pix1 += line_size; + pix2v = pix3v; + pix3 += line_size; + + } + + /* Sum up the four partial sums, and put the result into s */ + sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero); + sumdiffs = vec_splat(sumdiffs, 3); + vec_ste(sumdiffs, 0, &s); + return s; +} + +int sad16_xy2_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h) +{ + int i; + int s __attribute__((aligned(16))); + 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); + vector unsigned char *tv, avgv, t5; + vector unsigned char pix1v, pix2v, pix3v, pix2iv, pix3iv; + vector unsigned short pix2lv, pix2hv, pix2ilv, pix2ihv; + vector unsigned short pix3lv, pix3hv, pix3ilv, pix3ihv; + vector unsigned short avghv, avglv; + vector unsigned short t1, t2, t3, t4; + vector unsigned int sad; + vector signed int sumdiffs; + + sad = (vector unsigned int)vec_splat_u32(0); + + s = 0; + + /* + Due to the fact that pix3 = pix2 + line_size, the pix3 of one + iteration becomes pix2 in the next iteration. We can use this + fact to avoid a potentially expensive unaligned read, as well + as some splitting, and vector addition each time around the loop. + Read unaligned pixels into our vectors. The vectors are as follows: + pix2v: pix2[0]-pix2[15] pix2iv: pix2[1]-pix2[16] + Split the pixel vectors into shorts + */ + tv = (vector unsigned char *) &pix2[0]; + pix2v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[0])); + + tv = (vector unsigned char *) &pix2[1]; + pix2iv = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[1])); + + pix2hv = (vector unsigned short) vec_mergeh(zero, pix2v); + pix2lv = (vector unsigned short) vec_mergel(zero, pix2v); + pix2ihv = (vector unsigned short) vec_mergeh(zero, pix2iv); + pix2ilv = (vector unsigned short) vec_mergel(zero, pix2iv); + t1 = vec_add(pix2hv, pix2ihv); + t2 = vec_add(pix2lv, pix2ilv); + + for(i=0;i<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] + */ + tv = (vector unsigned char *) pix1; + pix1v = vec_perm(tv[0], tv[1], vec_lvsl(0, pix1)); + + tv = (vector unsigned char *) &pix3[0]; + pix3v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix3[0])); + + tv = (vector unsigned char *) &pix3[1]; + pix3iv = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix3[1])); + + /* + Note that Altivec does have vec_avg, but this works on vector pairs + and rounds up. We could do avg(avg(a,b),avg(c,d)), but the rounding + would mean that, for example, avg(3,0,0,1) = 2, when it should be 1. + Instead, we have to split the pixel vectors into vectors of shorts, + and do the averaging by hand. + */ + + /* Split the pixel vectors into shorts */ + pix3hv = (vector unsigned short) vec_mergeh(zero, pix3v); + pix3lv = (vector unsigned short) vec_mergel(zero, pix3v); + pix3ihv = (vector unsigned short) vec_mergeh(zero, pix3iv); + pix3ilv = (vector unsigned short) vec_mergel(zero, pix3iv); + + /* Do the averaging on them */ + t3 = vec_add(pix3hv, pix3ihv); + t4 = vec_add(pix3lv, pix3ilv); + + avghv = vec_sr(vec_add(vec_add(t1, t3), two), two); + avglv = vec_sr(vec_add(vec_add(t2, t4), two), two); + + /* Pack the shorts back into a result */ + avgv = vec_pack(avghv, avglv); + + /* Calculate a sum of abs differences vector */ + t5 = vec_sub(vec_max(pix1v, avgv), vec_min(pix1v, avgv)); + + /* Add each 4 pixel group together and put 4 results into sad */ + sad = vec_sum4s(t5, sad); + + pix1 += line_size; + pix3 += line_size; + /* Transfer the calculated values for pix3 into pix2 */ + t1 = t3; + t2 = t4; + } + /* Sum up the four partial sums, and put the result into s */ + sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero); + sumdiffs = vec_splat(sumdiffs, 3); + vec_ste(sumdiffs, 0, &s); + + return s; +} + +int 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; + 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++) { + /* Read potentially unaligned pixels into t1 and t2 */ + perm1 = vec_lvsl(0, pix1); + pix1v = (vector unsigned char *) 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); + + /* Calculate a sum of abs differences vector */ + t3 = vec_max(t1, t2); + t4 = vec_min(t1, t2); + t5 = vec_sub(t3, t4); + + /* Add each 4 pixel group together and put 4 results into sad */ + sad = vec_sum4s(t5, sad); + + pix1 += line_size; + pix2 += line_size; + } + + /* Sum up the four partial sums, and put the result into s */ + sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero); + sumdiffs = vec_splat(sumdiffs, 3); + vec_ste(sumdiffs, 0, &s); + + return s; +} + +int 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); + vector unsigned char perm1, perm2, permclear, *pix1v, *pix2v; + vector unsigned char t1, t2, t3,t4, t5; + vector unsigned int sad; + vector signed int sumdiffs; + + 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); + + 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. */ + perm1 = vec_lvsl(0, pix1); + pix1v = (vector unsigned char *) pix1; + perm2 = vec_lvsl(0, pix2); + pix2v = (vector unsigned char *) pix2; + t1 = vec_and(vec_perm(pix1v[0], pix1v[1], perm1), permclear); + t2 = vec_and(vec_perm(pix2v[0], pix2v[1], perm2), permclear); + + /* Calculate a sum of abs differences vector */ + t3 = vec_max(t1, t2); + t4 = vec_min(t1, t2); + t5 = vec_sub(t3, t4); + + /* Add each 4 pixel group together and put 4 results into sad */ + sad = vec_sum4s(t5, sad); + + pix1 += line_size; + pix2 += line_size; + } + + /* Sum up the four partial sums, and put the result into s */ + sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero); + sumdiffs = vec_splat(sumdiffs, 3); + vec_ste(sumdiffs, 0, &s); + + return s; +} + +int pix_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); + vector unsigned char *tv; + vector unsigned char pixv; + vector unsigned int sv; + vector signed int sum; + + sv = (vector unsigned int)vec_splat_u32(0); + + s = 0; + for (i = 0; i < 16; i++) { + /* Read in the potentially unaligned pixels */ + tv = (vector unsigned char *) pix; + pixv = vec_perm(tv[0], tv[1], vec_lvsl(0, pix)); + + /* Square the values, and add them to our sum */ + sv = vec_msum(pixv, pixv, sv); + + pix += line_size; + } + /* Sum up the four partial sums, and put the result into s */ + sum = vec_sums((vector signed int) sv, (vector signed int) zero); + sum = vec_splat(sum, 3); + vec_ste(sum, 0, &s); + + return s; +} + +/** + * Sum of Squared Errors for a 8x8 block. + * 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) +{ + 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, permclear, *pix1v, *pix2v; + vector unsigned char t1, t2, t3,t4, t5; + vector unsigned int sum; + vector signed int sumsqr; + + 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); + + + 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. */ + perm1 = vec_lvsl(0, pix1); + pix1v = (vector unsigned char *) pix1; + perm2 = vec_lvsl(0, pix2); + pix2v = (vector unsigned char *) pix2; + 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. + */ + + /* Calculate abs differences vector */ + t3 = vec_max(t1, t2); + t4 = vec_min(t1, t2); + t5 = vec_sub(t3, t4); + + /* Square the values and add them to our sum */ + sum = vec_msum(t5, t5, sum); + + pix1 += line_size; + pix2 += line_size; + } + + /* 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); + + return s; +} + +/** + * Sum of Squared Errors for a 16x16 block. + * 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) +{ + 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; + vector unsigned char t1, t2, t3,t4, t5; + vector unsigned int sum; + vector signed int sumsqr; + + sum = (vector unsigned int)vec_splat_u32(0); + + for(i=0;i<h;i++) { + /* Read potentially unaligned pixels into t1 and t2 */ + perm1 = vec_lvsl(0, pix1); + pix1v = (vector unsigned char *) 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); + + /* + 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); + t4 = vec_min(t1, t2); + t5 = vec_sub(t3, t4); + + /* Square the values and add them to our sum */ + sum = vec_msum(t5, t5, sum); + + pix1 += line_size; + pix2 += line_size; + } + + /* 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); + + return s; +} + +int pix_sum_altivec(uint8_t * pix, int line_size) +{ + 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))); + + sad = (vector unsigned int)vec_splat_u32(0); + + for (i = 0; i < 16; i++) { + /* Read the potentially unaligned 16 pixels into t1 */ + perm = vec_lvsl(0, pix); + pixv = (vector unsigned char *) pix; + t1 = vec_perm(pixv[0], pixv[1], perm); + + /* Add each 4 pixel group together and put 4 results into sad */ + sad = vec_sum4s(t1, sad); + + pix += line_size; + } + + /* Sum up the four partial sums, and put the result into s */ + sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero); + sumdiffs = vec_splat(sumdiffs, 3); + vec_ste(sumdiffs, 0, &s); + + return s; +} + +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); + vector signed short shorts; + + 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. + perm = vec_lvsl(0, pixels); + pixv = (vector unsigned char *) pixels; + bytes = vec_perm(pixv[0], pixv[1], perm); + + // convert the bytes into shorts + shorts = (vector signed short)vec_mergeh(zero, bytes); + + // save the data to the block, we assume the block is 16-byte aligned + vec_st(shorts, i*16, (vector signed short*)block); + + pixels += line_size; + } +} + +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); + vector signed short shorts1, shorts2; + + 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. + perm = vec_lvsl(0, s1); + pixv = (vector unsigned char *) s1; + bytes = vec_perm(pixv[0], pixv[1], perm); + + // convert the bytes into shorts + shorts1 = (vector signed short)vec_mergeh(zero, bytes); + + // Do the same for the second block of pixels + perm = vec_lvsl(0, s2); + pixv = (vector unsigned char *) s2; + bytes = vec_perm(pixv[0], pixv[1], perm); + + // convert the bytes into shorts + shorts2 = (vector signed short)vec_mergeh(zero, bytes); + + // Do the subtraction + shorts1 = vec_sub(shorts1, shorts2); + + // save the data to the block, we assume the block is 16-byte aligned + vec_st(shorts1, 0, (vector signed short*)block); + + s1 += stride; + s2 += stride; + block += 8; + + + // The code below is a copy of the code above... This is a manual + // unroll. + + // Read potentially unaligned pixels + // We're reading 16 pixels, and actually only want 8, + // but we simply ignore the extras. + perm = vec_lvsl(0, s1); + pixv = (vector unsigned char *) s1; + bytes = vec_perm(pixv[0], pixv[1], perm); + + // convert the bytes into shorts + shorts1 = (vector signed short)vec_mergeh(zero, bytes); + + // Do the same for the second block of pixels + perm = vec_lvsl(0, s2); + pixv = (vector unsigned char *) s2; + bytes = vec_perm(pixv[0], pixv[1], perm); + + // convert the bytes into shorts + shorts2 = (vector signed short)vec_mergeh(zero, bytes); + + // Do the subtraction + shorts1 = vec_sub(shorts1, shorts2); + + // save the data to the block, we assume the block is 16-byte aligned + vec_st(shorts1, 0, (vector signed short*)block); + + s1 += stride; + s2 += stride; + block += 8; + } +} + +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+=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]; + } +} + +/* 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); + register vector unsigned char pixelsv1, pixelsv2; + register vector unsigned char pixelsv1B, pixelsv2B; + register vector unsigned char pixelsv1C, pixelsv2C; + register vector unsigned char pixelsv1D, pixelsv2D; + + register vector unsigned char perm = vec_lvsl(0, pixels); + int i; + register int line_size_2 = line_size << 1; + 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; + } +#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; + } +#endif +POWERPC_PERF_STOP_COUNT(altivec_put_pixels16_num, 1); +} + +/* 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); + 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; + } + +POWERPC_PERF_STOP_COUNT(altivec_avg_pixels16_num, 1); +} + +/* next one assumes that ((line_size % 8) == 0) */ +void avg_pixels8_altivec(uint8_t * block, const uint8_t * pixels, int line_size, int h) +{ +POWERPC_PERF_DECLARE(altivec_avg_pixels8_num, 1); + 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; + } + +POWERPC_PERF_STOP_COUNT(altivec_avg_pixels8_num, 1); +} + +/* 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) +{ +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 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); +} + +/* 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) +{ +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 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); +} + +/* 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) +{ +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 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); +} + +/* 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) +{ +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 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); +} + +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); \ + } + ONEITERBUTTERFLY(0, temp0); + ONEITERBUTTERFLY(1, temp1); + ONEITERBUTTERFLY(2, temp2); + ONEITERBUTTERFLY(3, temp3); + ONEITERBUTTERFLY(4, temp4); + 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); + register vector signed short line2 = vec_add(temp2, temp3); + register vector signed short line3 = vec_sub(temp2, temp3); + register vector signed short line4 = vec_add(temp4, temp5); + register vector signed short line5 = vec_sub(temp4, temp5); + register vector signed short line6 = vec_add(temp6, temp7); + register vector signed short line7 = vec_sub(temp6, temp7); + + register vector signed short line0B = vec_add(line0, line2); + register vector signed short line2B = vec_sub(line0, line2); + register vector signed short line1B = vec_add(line1, line3); + register vector signed short line3B = vec_sub(line1, line3); + register vector signed short line4B = vec_add(line4, line6); + register vector signed short line6B = vec_sub(line4, line6); + register vector signed short line5B = vec_add(line5, line7); + register vector signed short line7B = vec_sub(line5, line7); + + register vector signed short line0C = vec_add(line0B, line4B); + register vector signed short line4C = vec_sub(line0B, line4B); + register vector signed short line1C = vec_add(line1B, line5B); + register vector signed short line5C = vec_sub(line1B, line5B); + register vector signed short line2C = vec_add(line2B, line6B); + register vector signed short line6C = vec_sub(line2B, line6B); + register vector signed short line3C = vec_add(line3B, line7B); + register vector signed short line7C = vec_sub(line3B, line7B); + + vsum = vec_sum4s(vec_abs(line0C), vec_splat_s32(0)); + vsum = vec_sum4s(vec_abs(line1C), vsum); + vsum = vec_sum4s(vec_abs(line2C), vsum); + vsum = vec_sum4s(vec_abs(line3C), vsum); + vsum = vec_sum4s(vec_abs(line4C), vsum); + vsum = vec_sum4s(vec_abs(line5C), vsum); + vsum = vec_sum4s(vec_abs(line6C), vsum); + 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); + } +POWERPC_PERF_STOP_COUNT(altivec_hadamard8_diff8x8_num, 1); + 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... +*/ + +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); \ + } + ONEITERBUTTERFLY(0, temp0, temp0S); + ONEITERBUTTERFLY(1, temp1, temp1S); + ONEITERBUTTERFLY(2, temp2, temp2S); + ONEITERBUTTERFLY(3, temp3, temp3S); + ONEITERBUTTERFLY(4, temp4, temp4S); + 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, + line1BS,line3BS,line4BS,line6BS,line5BS, + line7BS,line0CS,line4CS,line1CS,line5CS, + line2CS,line6CS,line3CS,line7CS; + + register vector signed short line0 = vec_add(temp0, temp1); + register vector signed short line1 = vec_sub(temp0, temp1); + register vector signed short line2 = vec_add(temp2, temp3); + register vector signed short line3 = vec_sub(temp2, temp3); + register vector signed short line4 = vec_add(temp4, temp5); + register vector signed short line5 = vec_sub(temp4, temp5); + register vector signed short line6 = vec_add(temp6, temp7); + register vector signed short line7 = vec_sub(temp6, temp7); + + register vector signed short line0B = vec_add(line0, line2); + register vector signed short line2B = vec_sub(line0, line2); + register vector signed short line1B = vec_add(line1, line3); + register vector signed short line3B = vec_sub(line1, line3); + register vector signed short line4B = vec_add(line4, line6); + register vector signed short line6B = vec_sub(line4, line6); + register vector signed short line5B = vec_add(line5, line7); + register vector signed short line7B = vec_sub(line5, line7); + + register vector signed short line0C = vec_add(line0B, line4B); + register vector signed short line4C = vec_sub(line0B, line4B); + register vector signed short line1C = vec_add(line1B, line5B); + register vector signed short line5C = vec_sub(line1B, line5B); + register vector signed short line2C = vec_add(line2B, line6B); + register vector signed short line6C = vec_sub(line2B, line6B); + register vector signed short line3C = vec_add(line3B, line7B); + register vector signed short line7C = vec_sub(line3B, line7B); + + vsum = vec_sum4s(vec_abs(line0C), vec_splat_s32(0)); + vsum = vec_sum4s(vec_abs(line1C), vsum); + vsum = vec_sum4s(vec_abs(line2C), vsum); + vsum = vec_sum4s(vec_abs(line3C), vsum); + vsum = vec_sum4s(vec_abs(line4C), vsum); + vsum = vec_sum4s(vec_abs(line5C), vsum); + vsum = vec_sum4s(vec_abs(line6C), vsum); + vsum = vec_sum4s(vec_abs(line7C), vsum); + + line0S = vec_add(temp0S, temp1S); + line1S = vec_sub(temp0S, temp1S); + line2S = vec_add(temp2S, temp3S); + line3S = vec_sub(temp2S, temp3S); + line4S = vec_add(temp4S, temp5S); + line5S = vec_sub(temp4S, temp5S); + line6S = vec_add(temp6S, temp7S); + line7S = vec_sub(temp6S, temp7S); + + line0BS = vec_add(line0S, line2S); + line2BS = vec_sub(line0S, line2S); + line1BS = vec_add(line1S, line3S); + line3BS = vec_sub(line1S, line3S); + line4BS = vec_add(line4S, line6S); + line6BS = vec_sub(line4S, line6S); + line5BS = vec_add(line5S, line7S); + line7BS = vec_sub(line5S, line7S); + + line0CS = vec_add(line0BS, line4BS); + line4CS = vec_sub(line0BS, line4BS); + line1CS = vec_add(line1BS, line5BS); + line5CS = vec_sub(line1BS, line5BS); + line2CS = vec_add(line2BS, line6BS); + line6CS = vec_sub(line2BS, line6BS); + line3CS = vec_add(line3BS, line7BS); + line7CS = vec_sub(line3BS, line7BS); + + vsum = vec_sum4s(vec_abs(line0CS), vsum); + vsum = vec_sum4s(vec_abs(line1CS), vsum); + vsum = vec_sum4s(vec_abs(line2CS), vsum); + vsum = vec_sum4s(vec_abs(line3CS), vsum); + vsum = vec_sum4s(vec_abs(line4CS), vsum); + vsum = vec_sum4s(vec_abs(line5CS), vsum); + vsum = vec_sum4s(vec_abs(line6CS), vsum); + 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; +} + +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; +} + +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; + } + } +#endif /* CONFIG_DARWIN */ + return 0; +#endif /* __AMIGAOS4__ */ +} + +static void vorbis_inverse_coupling_altivec(float *mag, float *ang, + int blocksize) +{ + int i; + vector float m, a; + 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) { + m = vec_ld(0, mag+i); + a = vec_ld(0, ang+i); + t0 = vec_cmple(m, (vector float)vec_splat_u32(0)); + t1 = vec_cmple(a, (vector float)vec_splat_u32(0)); + a = vec_xor(a, (vector float) vec_sl((vector unsigned int)t0, v_31)); + t0 = (vector bool int)vec_and(a, t1); + t1 = (vector bool int)vec_andc(a, t1); + a = vec_sub(m, (vector float)t1); + m = vec_add(m, (vector float)t0); + vec_stl(a, 0, ang+i); + vec_stl(m, 0, mag+i); + } +} + +/* 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) +{ +POWERPC_PERF_DECLARE(altivec_avg_pixels8_xy2_num, 1); + 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; + } + +POWERPC_PERF_STOP_COUNT(altivec_avg_pixels8_xy2_num, 1); +} + +void dsputil_init_altivec(DSPContext* c, AVCodecContext *avctx) +{ + c->pix_abs[0][1] = sad16_x2_altivec; + c->pix_abs[0][2] = sad16_y2_altivec; + c->pix_abs[0][3] = sad16_xy2_altivec; + c->pix_abs[0][0] = sad16_altivec; + c->pix_abs[1][0] = sad8_altivec; + c->sad[0]= sad16_altivec; + c->sad[1]= sad8_altivec; + c->pix_norm1 = pix_norm1_altivec; + c->sse[1]= sse8_altivec; + c->sse[0]= sse16_altivec; + c->pix_sum = pix_sum_altivec; + c->diff_pixels = diff_pixels_altivec; + c->get_pixels = get_pixels_altivec; + c->add_bytes= add_bytes_altivec; + 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->avg_pixels_tab[0][0] = avg_pixels16_altivec; + c->avg_pixels_tab[1][0] = avg_pixels8_altivec; + c->avg_pixels_tab[1][3] = avg_pixels8_xy2_altivec; + c->put_pixels_tab[1][3] = put_pixels8_xy2_altivec; + c->put_no_rnd_pixels_tab[1][3] = put_no_rnd_pixels8_xy2_altivec; + c->put_pixels_tab[0][3] = put_pixels16_xy2_altivec; + c->put_no_rnd_pixels_tab[0][3] = put_no_rnd_pixels16_xy2_altivec; + + c->hadamard8_diff[0] = hadamard8_diff16_altivec; + c->hadamard8_diff[1] = hadamard8_diff8x8_altivec; +#ifdef CONFIG_VORBIS_DECODER + c->vorbis_inverse_coupling = vorbis_inverse_coupling_altivec; +#endif +} diff --git a/contrib/ffmpeg/libavcodec/ppc/dsputil_altivec.h b/contrib/ffmpeg/libavcodec/ppc/dsputil_altivec.h new file mode 100644 index 000000000..560d778bb --- /dev/null +++ b/contrib/ffmpeg/libavcodec/ppc/dsputil_altivec.h @@ -0,0 +1,106 @@ +/* + * Copyright (c) 2002 Brian Foley + * Copyright (c) 2002 Dieter Shirley + * Copyright (c) 2003-2004 Romain Dolbeau <romain@dolbeau.org> + * + * This file is part of FFmpeg. + * + * FFmpeg 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 2.1 of the License, or (at your option) any later version. + * + * FFmpeg 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 FFmpeg; if not, write to the Free Software + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA + */ + +#ifndef _DSPUTIL_ALTIVEC_ +#define _DSPUTIL_ALTIVEC_ + +#include "dsputil_ppc.h" + +#ifdef HAVE_ALTIVEC + +extern int has_altivec(void); + +void put_pixels16_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h); + +void avg_pixels16_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h); + +// used to build registers permutation vectors (vcprm) +// the 's' are for words in the _s_econd vector +#define WORD_0 0x00,0x01,0x02,0x03 +#define WORD_1 0x04,0x05,0x06,0x07 +#define WORD_2 0x08,0x09,0x0a,0x0b +#define WORD_3 0x0c,0x0d,0x0e,0x0f +#define WORD_s0 0x10,0x11,0x12,0x13 +#define WORD_s1 0x14,0x15,0x16,0x17 +#define WORD_s2 0x18,0x19,0x1a,0x1b +#define WORD_s3 0x1c,0x1d,0x1e,0x1f + +#ifdef CONFIG_DARWIN +#define vcprm(a,b,c,d) (const vector unsigned char)(WORD_ ## a, WORD_ ## b, WORD_ ## c, WORD_ ## d) +#else +#define vcprm(a,b,c,d) (const vector unsigned char){WORD_ ## a, WORD_ ## b, WORD_ ## c, WORD_ ## d} +#endif + +// vcprmle is used to keep the same index as in the SSE version. +// it's the same as vcprm, with the index inversed +// ('le' is Little Endian) +#define vcprmle(a,b,c,d) vcprm(d,c,b,a) + +// used to build inverse/identity vectors (vcii) +// n is _n_egative, p is _p_ositive +#define FLOAT_n -1. +#define FLOAT_p 1. + + +#ifdef CONFIG_DARWIN +#define vcii(a,b,c,d) (const vector float)(FLOAT_ ## a, FLOAT_ ## b, FLOAT_ ## c, FLOAT_ ## d) +#else +#define vcii(a,b,c,d) (const vector float){FLOAT_ ## a, FLOAT_ ## b, FLOAT_ ## c, FLOAT_ ## d} +#endif + +// Transpose 8x8 matrix of 16-bit elements (in-place) +#define TRANSPOSE8(a,b,c,d,e,f,g,h) \ +do { \ + vector signed short A1, B1, C1, D1, E1, F1, G1, H1; \ + vector signed short A2, B2, C2, D2, E2, F2, G2, H2; \ + \ + A1 = vec_mergeh (a, e); \ + B1 = vec_mergel (a, e); \ + C1 = vec_mergeh (b, f); \ + D1 = vec_mergel (b, f); \ + E1 = vec_mergeh (c, g); \ + F1 = vec_mergel (c, g); \ + G1 = vec_mergeh (d, h); \ + H1 = vec_mergel (d, h); \ + \ + A2 = vec_mergeh (A1, E1); \ + B2 = vec_mergel (A1, E1); \ + C2 = vec_mergeh (B1, F1); \ + D2 = vec_mergel (B1, F1); \ + E2 = vec_mergeh (C1, G1); \ + F2 = vec_mergel (C1, G1); \ + G2 = vec_mergeh (D1, H1); \ + H2 = vec_mergel (D1, H1); \ + \ + a = vec_mergeh (A2, E2); \ + b = vec_mergel (A2, E2); \ + c = vec_mergeh (B2, F2); \ + d = vec_mergel (B2, F2); \ + e = vec_mergeh (C2, G2); \ + f = vec_mergel (C2, G2); \ + g = vec_mergeh (D2, H2); \ + h = vec_mergel (D2, H2); \ +} while (0) + +#endif /* HAVE_ALTIVEC */ + +#endif /* _DSPUTIL_ALTIVEC_ */ diff --git a/contrib/ffmpeg/libavcodec/ppc/dsputil_ppc.c b/contrib/ffmpeg/libavcodec/ppc/dsputil_ppc.c new file mode 100644 index 000000000..9169eaef0 --- /dev/null +++ b/contrib/ffmpeg/libavcodec/ppc/dsputil_ppc.c @@ -0,0 +1,325 @@ +/* + * Copyright (c) 2002 Brian Foley + * Copyright (c) 2002 Dieter Shirley + * Copyright (c) 2003-2004 Romain Dolbeau <romain@dolbeau.org> + * + * This file is part of FFmpeg. + * + * FFmpeg 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 2.1 of the License, or (at your option) any later version. + * + * FFmpeg 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 FFmpeg; if not, write to the Free Software + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA + */ + +#include "../dsputil.h" + +#include "dsputil_ppc.h" + +#ifdef HAVE_ALTIVEC +#include "dsputil_altivec.h" + +extern void fdct_altivec(int16_t *block); +extern void gmc1_altivec(uint8_t *dst, uint8_t *src, int stride, int h, + int x16, int y16, int rounder); +extern void idct_put_altivec(uint8_t *dest, int line_size, int16_t *block); +extern void idct_add_altivec(uint8_t *dest, int line_size, int16_t *block); + +void dsputil_h264_init_ppc(DSPContext* c, AVCodecContext *avctx); + +void dsputil_init_altivec(DSPContext* c, AVCodecContext *avctx); +void vc1dsp_init_altivec(DSPContext* c, AVCodecContext *avctx); +void snow_init_altivec(DSPContext* c, AVCodecContext *avctx); +void float_init_altivec(DSPContext* c, AVCodecContext *avctx); + +#endif + +int mm_flags = 0; + +int mm_support(void) +{ + int result = 0; +#ifdef HAVE_ALTIVEC + if (has_altivec()) { + result |= MM_ALTIVEC; + } +#endif /* result */ + return result; +} + +#ifdef POWERPC_PERFORMANCE_REPORT +unsigned long long perfdata[POWERPC_NUM_PMC_ENABLED][powerpc_perf_total][powerpc_data_total]; +/* list below must match enum in dsputil_ppc.h */ +static unsigned char* perfname[] = { + "ff_fft_calc_altivec", + "gmc1_altivec", + "dct_unquantize_h263_altivec", + "fdct_altivec", + "idct_add_altivec", + "idct_put_altivec", + "put_pixels16_altivec", + "avg_pixels16_altivec", + "avg_pixels8_altivec", + "put_pixels8_xy2_altivec", + "put_no_rnd_pixels8_xy2_altivec", + "put_pixels16_xy2_altivec", + "put_no_rnd_pixels16_xy2_altivec", + "hadamard8_diff8x8_altivec", + "hadamard8_diff16_altivec", + "avg_pixels8_xy2_altivec", + "clear_blocks_dcbz32_ppc", + "clear_blocks_dcbz128_ppc", + "put_h264_chroma_mc8_altivec", + "avg_h264_chroma_mc8_altivec", + "put_h264_qpel16_h_lowpass_altivec", + "avg_h264_qpel16_h_lowpass_altivec", + "put_h264_qpel16_v_lowpass_altivec", + "avg_h264_qpel16_v_lowpass_altivec", + "put_h264_qpel16_hv_lowpass_altivec", + "avg_h264_qpel16_hv_lowpass_altivec", + "" +}; +#include <stdio.h> +#endif + +#ifdef POWERPC_PERFORMANCE_REPORT +void powerpc_display_perf_report(void) +{ + int i, j; + av_log(NULL, AV_LOG_INFO, "PowerPC performance report\n Values are from the PMC registers, and represent whatever the registers are set to record.\n"); + for(i = 0 ; i < powerpc_perf_total ; i++) + { + for (j = 0; j < POWERPC_NUM_PMC_ENABLED ; j++) + { + if (perfdata[j][i][powerpc_data_num] != (unsigned long long)0) + av_log(NULL, AV_LOG_INFO, + " Function \"%s\" (pmc%d):\n\tmin: %"PRIu64"\n\tmax: %"PRIu64"\n\tavg: %1.2lf (%"PRIu64")\n", + perfname[i], + j+1, + perfdata[j][i][powerpc_data_min], + perfdata[j][i][powerpc_data_max], + (double)perfdata[j][i][powerpc_data_sum] / + (double)perfdata[j][i][powerpc_data_num], + perfdata[j][i][powerpc_data_num]); + } + } +} +#endif /* POWERPC_PERFORMANCE_REPORT */ + +/* ***** WARNING ***** WARNING ***** WARNING ***** */ +/* + clear_blocks_dcbz32_ppc will not work properly + on PowerPC processors with a cache line size + not equal to 32 bytes. + Fortunately all processor used by Apple up to + at least the 7450 (aka second generation G4) + use 32 bytes cache line. + This is due to the use of the 'dcbz' instruction. + It simply clear to zero a single cache line, + so you need to know the cache line size to use it ! + It's absurd, but it's fast... + + update 24/06/2003 : Apple released yesterday the G5, + with a PPC970. cache line size : 128 bytes. Oups. + The semantic of dcbz was changed, it always clear + 32 bytes. so the function below will work, but will + be slow. So I fixed check_dcbz_effect to use dcbzl, + which is defined to clear a cache line (as dcbz before). + So we still can distinguish, and use dcbz (32 bytes) + or dcbzl (one cache line) as required. + + see <http://developer.apple.com/technotes/tn/tn2087.html> + and <http://developer.apple.com/technotes/tn/tn2086.html> +*/ +void clear_blocks_dcbz32_ppc(DCTELEM *blocks) +{ +POWERPC_PERF_DECLARE(powerpc_clear_blocks_dcbz32, 1); + register int misal = ((unsigned long)blocks & 0x00000010); + register int i = 0; +POWERPC_PERF_START_COUNT(powerpc_clear_blocks_dcbz32, 1); +#if 1 + if (misal) { + ((unsigned long*)blocks)[0] = 0L; + ((unsigned long*)blocks)[1] = 0L; + ((unsigned long*)blocks)[2] = 0L; + ((unsigned long*)blocks)[3] = 0L; + i += 16; + } + for ( ; i < sizeof(DCTELEM)*6*64-31 ; i += 32) { +#ifndef __MWERKS__ + asm volatile("dcbz %0,%1" : : "b" (blocks), "r" (i) : "memory"); +#else + __dcbz( blocks, i ); +#endif + } + if (misal) { + ((unsigned long*)blocks)[188] = 0L; + ((unsigned long*)blocks)[189] = 0L; + ((unsigned long*)blocks)[190] = 0L; + ((unsigned long*)blocks)[191] = 0L; + i += 16; + } +#else + memset(blocks, 0, sizeof(DCTELEM)*6*64); +#endif +POWERPC_PERF_STOP_COUNT(powerpc_clear_blocks_dcbz32, 1); +} + +/* same as above, when dcbzl clear a whole 128B cache line + i.e. the PPC970 aka G5 */ +#ifdef HAVE_DCBZL +void clear_blocks_dcbz128_ppc(DCTELEM *blocks) +{ +POWERPC_PERF_DECLARE(powerpc_clear_blocks_dcbz128, 1); + register int misal = ((unsigned long)blocks & 0x0000007f); + register int i = 0; +POWERPC_PERF_START_COUNT(powerpc_clear_blocks_dcbz128, 1); +#if 1 + if (misal) { + // we could probably also optimize this case, + // but there's not much point as the machines + // aren't available yet (2003-06-26) + memset(blocks, 0, sizeof(DCTELEM)*6*64); + } + else + for ( ; i < sizeof(DCTELEM)*6*64 ; i += 128) { + asm volatile("dcbzl %0,%1" : : "b" (blocks), "r" (i) : "memory"); + } +#else + memset(blocks, 0, sizeof(DCTELEM)*6*64); +#endif +POWERPC_PERF_STOP_COUNT(powerpc_clear_blocks_dcbz128, 1); +} +#else +void clear_blocks_dcbz128_ppc(DCTELEM *blocks) +{ + memset(blocks, 0, sizeof(DCTELEM)*6*64); +} +#endif + +#ifdef HAVE_DCBZL +/* check dcbz report how many bytes are set to 0 by dcbz */ +/* update 24/06/2003 : replace dcbz by dcbzl to get + the intended effect (Apple "fixed" dcbz) + unfortunately this cannot be used unless the assembler + knows about dcbzl ... */ +long check_dcbzl_effect(void) +{ + register char *fakedata = (char*)av_malloc(1024); + register char *fakedata_middle; + register long zero = 0; + register long i = 0; + long count = 0; + + if (!fakedata) + { + return 0L; + } + + fakedata_middle = (fakedata + 512); + + memset(fakedata, 0xFF, 1024); + + /* below the constraint "b" seems to mean "Address base register" + in gcc-3.3 / RS/6000 speaks. seems to avoid using r0, so.... */ + asm volatile("dcbzl %0, %1" : : "b" (fakedata_middle), "r" (zero)); + + for (i = 0; i < 1024 ; i ++) + { + if (fakedata[i] == (char)0) + count++; + } + + av_free(fakedata); + + return count; +} +#else +long check_dcbzl_effect(void) +{ + return 0; +} +#endif + +static void prefetch_ppc(void *mem, int stride, int h) +{ + register const uint8_t *p = mem; + do { + asm volatile ("dcbt 0,%0" : : "r" (p)); + p+= stride; + } while(--h); +} + +void dsputil_init_ppc(DSPContext* c, AVCodecContext *avctx) +{ + // Common optimizations whether Altivec is available or not + c->prefetch = prefetch_ppc; + switch (check_dcbzl_effect()) { + case 32: + c->clear_blocks = clear_blocks_dcbz32_ppc; + break; + case 128: + c->clear_blocks = clear_blocks_dcbz128_ppc; + break; + default: + break; + } + +#ifdef HAVE_ALTIVEC + if(ENABLE_H264_DECODER) dsputil_h264_init_ppc(c, avctx); + + if (has_altivec()) { + mm_flags |= MM_ALTIVEC; + + dsputil_init_altivec(c, avctx); + if(ENABLE_SNOW_DECODER) snow_init_altivec(c, avctx); + if(ENABLE_VC1_DECODER || ENABLE_WMV3_DECODER) + vc1dsp_init_altivec(c, avctx); + float_init_altivec(c, avctx); + c->gmc1 = gmc1_altivec; + +#ifdef CONFIG_ENCODERS + if (avctx->dct_algo == FF_DCT_AUTO || + avctx->dct_algo == FF_DCT_ALTIVEC) + { + c->fdct = fdct_altivec; + } +#endif //CONFIG_ENCODERS + + if (avctx->lowres==0) + { + if ((avctx->idct_algo == FF_IDCT_AUTO) || + (avctx->idct_algo == FF_IDCT_ALTIVEC)) + { + c->idct_put = idct_put_altivec; + c->idct_add = idct_add_altivec; + c->idct_permutation_type = FF_TRANSPOSE_IDCT_PERM; + } + } + +#ifdef POWERPC_PERFORMANCE_REPORT + { + int i, j; + for (i = 0 ; i < powerpc_perf_total ; i++) + { + for (j = 0; j < POWERPC_NUM_PMC_ENABLED ; j++) + { + perfdata[j][i][powerpc_data_min] = 0xFFFFFFFFFFFFFFFFULL; + perfdata[j][i][powerpc_data_max] = 0x0000000000000000ULL; + perfdata[j][i][powerpc_data_sum] = 0x0000000000000000ULL; + perfdata[j][i][powerpc_data_num] = 0x0000000000000000ULL; + } + } + } +#endif /* POWERPC_PERFORMANCE_REPORT */ + } +#endif /* HAVE_ALTIVEC */ +} diff --git a/contrib/ffmpeg/libavcodec/ppc/dsputil_ppc.h b/contrib/ffmpeg/libavcodec/ppc/dsputil_ppc.h new file mode 100644 index 000000000..ab2b05780 --- /dev/null +++ b/contrib/ffmpeg/libavcodec/ppc/dsputil_ppc.h @@ -0,0 +1,155 @@ +/* + * Copyright (c) 2003-2004 Romain Dolbeau <romain@dolbeau.org> + * + * This file is part of FFmpeg. + * + * FFmpeg 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 2.1 of the License, or (at your option) any later version. + * + * FFmpeg 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 FFmpeg; if not, write to the Free Software + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA + */ + +#ifndef _DSPUTIL_PPC_ +#define _DSPUTIL_PPC_ + +#ifdef POWERPC_PERFORMANCE_REPORT +void powerpc_display_perf_report(void); +/* the 604* have 2, the G3* have 4, the G4s have 6, + and the G5 are completely different (they MUST use + POWERPC_MODE_64BITS, and let's hope all future 64 bis PPC + will use the same PMCs... */ +#define POWERPC_NUM_PMC_ENABLED 6 +/* if you add to the enum below, also add to the perfname array + in dsputil_ppc.c */ +enum powerpc_perf_index { + altivec_fft_num = 0, + altivec_gmc1_num, + altivec_dct_unquantize_h263_num, + altivec_fdct, + altivec_idct_add_num, + altivec_idct_put_num, + altivec_put_pixels16_num, + altivec_avg_pixels16_num, + altivec_avg_pixels8_num, + altivec_put_pixels8_xy2_num, + altivec_put_no_rnd_pixels8_xy2_num, + altivec_put_pixels16_xy2_num, + altivec_put_no_rnd_pixels16_xy2_num, + altivec_hadamard8_diff8x8_num, + altivec_hadamard8_diff16_num, + altivec_avg_pixels8_xy2_num, + powerpc_clear_blocks_dcbz32, + powerpc_clear_blocks_dcbz128, + altivec_put_h264_chroma_mc8_num, + altivec_avg_h264_chroma_mc8_num, + altivec_put_h264_qpel16_h_lowpass_num, + altivec_avg_h264_qpel16_h_lowpass_num, + altivec_put_h264_qpel16_v_lowpass_num, + altivec_avg_h264_qpel16_v_lowpass_num, + altivec_put_h264_qpel16_hv_lowpass_num, + altivec_avg_h264_qpel16_hv_lowpass_num, + powerpc_perf_total +}; +enum powerpc_data_index { + powerpc_data_min = 0, + powerpc_data_max, + powerpc_data_sum, + powerpc_data_num, + powerpc_data_total +}; +extern unsigned long long perfdata[POWERPC_NUM_PMC_ENABLED][powerpc_perf_total][powerpc_data_total]; + +#ifndef POWERPC_MODE_64BITS +#define POWERP_PMC_DATATYPE unsigned long +#define POWERPC_GET_PMC1(a) asm volatile("mfspr %0, 937" : "=r" (a)) +#define POWERPC_GET_PMC2(a) asm volatile("mfspr %0, 938" : "=r" (a)) +#if (POWERPC_NUM_PMC_ENABLED > 2) +#define POWERPC_GET_PMC3(a) asm volatile("mfspr %0, 941" : "=r" (a)) +#define POWERPC_GET_PMC4(a) asm volatile("mfspr %0, 942" : "=r" (a)) +#else +#define POWERPC_GET_PMC3(a) do {} while (0) +#define POWERPC_GET_PMC4(a) do {} while (0) +#endif +#if (POWERPC_NUM_PMC_ENABLED > 4) +#define POWERPC_GET_PMC5(a) asm volatile("mfspr %0, 929" : "=r" (a)) +#define POWERPC_GET_PMC6(a) asm volatile("mfspr %0, 930" : "=r" (a)) +#else +#define POWERPC_GET_PMC5(a) do {} while (0) +#define POWERPC_GET_PMC6(a) do {} while (0) +#endif +#else /* POWERPC_MODE_64BITS */ +#define POWERP_PMC_DATATYPE unsigned long long +#define POWERPC_GET_PMC1(a) asm volatile("mfspr %0, 771" : "=r" (a)) +#define POWERPC_GET_PMC2(a) asm volatile("mfspr %0, 772" : "=r" (a)) +#if (POWERPC_NUM_PMC_ENABLED > 2) +#define POWERPC_GET_PMC3(a) asm volatile("mfspr %0, 773" : "=r" (a)) +#define POWERPC_GET_PMC4(a) asm volatile("mfspr %0, 774" : "=r" (a)) +#else +#define POWERPC_GET_PMC3(a) do {} while (0) +#define POWERPC_GET_PMC4(a) do {} while (0) +#endif +#if (POWERPC_NUM_PMC_ENABLED > 4) +#define POWERPC_GET_PMC5(a) asm volatile("mfspr %0, 775" : "=r" (a)) +#define POWERPC_GET_PMC6(a) asm volatile("mfspr %0, 776" : "=r" (a)) +#else +#define POWERPC_GET_PMC5(a) do {} while (0) +#define POWERPC_GET_PMC6(a) do {} while (0) +#endif +#endif /* POWERPC_MODE_64BITS */ +#define POWERPC_PERF_DECLARE(a, cond) \ + POWERP_PMC_DATATYPE \ + pmc_start[POWERPC_NUM_PMC_ENABLED], \ + pmc_stop[POWERPC_NUM_PMC_ENABLED], \ + pmc_loop_index; +#define POWERPC_PERF_START_COUNT(a, cond) do { \ + POWERPC_GET_PMC6(pmc_start[5]); \ + POWERPC_GET_PMC5(pmc_start[4]); \ + POWERPC_GET_PMC4(pmc_start[3]); \ + POWERPC_GET_PMC3(pmc_start[2]); \ + POWERPC_GET_PMC2(pmc_start[1]); \ + POWERPC_GET_PMC1(pmc_start[0]); \ + } while (0) +#define POWERPC_PERF_STOP_COUNT(a, cond) do { \ + POWERPC_GET_PMC1(pmc_stop[0]); \ + POWERPC_GET_PMC2(pmc_stop[1]); \ + POWERPC_GET_PMC3(pmc_stop[2]); \ + POWERPC_GET_PMC4(pmc_stop[3]); \ + POWERPC_GET_PMC5(pmc_stop[4]); \ + POWERPC_GET_PMC6(pmc_stop[5]); \ + if (cond) \ + { \ + for(pmc_loop_index = 0; \ + pmc_loop_index < POWERPC_NUM_PMC_ENABLED; \ + pmc_loop_index++) \ + { \ + if (pmc_stop[pmc_loop_index] >= pmc_start[pmc_loop_index]) \ + { \ + POWERP_PMC_DATATYPE diff = \ + pmc_stop[pmc_loop_index] - pmc_start[pmc_loop_index]; \ + if (diff < perfdata[pmc_loop_index][a][powerpc_data_min]) \ + perfdata[pmc_loop_index][a][powerpc_data_min] = diff; \ + if (diff > perfdata[pmc_loop_index][a][powerpc_data_max]) \ + perfdata[pmc_loop_index][a][powerpc_data_max] = diff; \ + perfdata[pmc_loop_index][a][powerpc_data_sum] += diff; \ + perfdata[pmc_loop_index][a][powerpc_data_num] ++; \ + } \ + } \ + } \ +} while (0) +#else /* POWERPC_PERFORMANCE_REPORT */ +// those are needed to avoid empty statements. +#define POWERPC_PERF_DECLARE(a, cond) int altivec_placeholder __attribute__ ((unused)) +#define POWERPC_PERF_START_COUNT(a, cond) do {} while (0) +#define POWERPC_PERF_STOP_COUNT(a, cond) do {} while (0) +#endif /* POWERPC_PERFORMANCE_REPORT */ + +#endif /* _DSPUTIL_PPC_ */ diff --git a/contrib/ffmpeg/libavcodec/ppc/fdct_altivec.c b/contrib/ffmpeg/libavcodec/ppc/fdct_altivec.c new file mode 100644 index 000000000..2418c32bb --- /dev/null +++ b/contrib/ffmpeg/libavcodec/ppc/fdct_altivec.c @@ -0,0 +1,493 @@ +/* ffmpeg/libavcodec/ppc/fdct_altivec.c, this file is part of the + * AltiVec optimized library for the FFMPEG Multimedia System + * Copyright (C) 2003 James Klicman <james@klicman.org> + * + * This file is part of FFmpeg. + * + * FFmpeg 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 2.1 of the License, or (at your option) any later version. + * + * FFmpeg 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 FFmpeg; if not, write to the Free Software + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA + */ + + +#include "common.h" +#include "../dsputil.h" +#include "dsputil_altivec.h" +#include "gcc_fixes.h" + + +#define vs16(v) ((vector signed short)(v)) +#define vs32(v) ((vector signed int)(v)) +#define vu8(v) ((vector unsigned char)(v)) +#define vu16(v) ((vector unsigned short)(v)) +#define vu32(v) ((vector unsigned int)(v)) + + +#define C1 0.98078525066375732421875000 /* cos(1*PI/16) */ +#define C2 0.92387950420379638671875000 /* cos(2*PI/16) */ +#define C3 0.83146959543228149414062500 /* cos(3*PI/16) */ +#define C4 0.70710676908493041992187500 /* cos(4*PI/16) */ +#define C5 0.55557024478912353515625000 /* cos(5*PI/16) */ +#define C6 0.38268342614173889160156250 /* cos(6*PI/16) */ +#define C7 0.19509032368659973144531250 /* cos(7*PI/16) */ +#define SQRT_2 1.41421353816986083984375000 /* sqrt(2) */ + + +#define W0 -(2 * C2) +#define W1 (2 * C6) +#define W2 (SQRT_2 * C6) +#define W3 (SQRT_2 * C3) +#define W4 (SQRT_2 * (-C1 + C3 + C5 - C7)) +#define W5 (SQRT_2 * ( C1 + C3 - C5 + C7)) +#define W6 (SQRT_2 * ( C1 + C3 + C5 - C7)) +#define W7 (SQRT_2 * ( C1 + C3 - C5 - C7)) +#define W8 (SQRT_2 * ( C7 - C3)) +#define W9 (SQRT_2 * (-C1 - C3)) +#define WA (SQRT_2 * (-C3 - C5)) +#define WB (SQRT_2 * ( C5 - C3)) + + +static vector float fdctconsts[3] = { + (vector float)AVV( W0, W1, W2, W3 ), + (vector float)AVV( W4, W5, W6, W7 ), + (vector float)AVV( W8, W9, WA, WB ) +}; + +#define LD_W0 vec_splat(cnsts0, 0) +#define LD_W1 vec_splat(cnsts0, 1) +#define LD_W2 vec_splat(cnsts0, 2) +#define LD_W3 vec_splat(cnsts0, 3) +#define LD_W4 vec_splat(cnsts1, 0) +#define LD_W5 vec_splat(cnsts1, 1) +#define LD_W6 vec_splat(cnsts1, 2) +#define LD_W7 vec_splat(cnsts1, 3) +#define LD_W8 vec_splat(cnsts2, 0) +#define LD_W9 vec_splat(cnsts2, 1) +#define LD_WA vec_splat(cnsts2, 2) +#define LD_WB vec_splat(cnsts2, 3) + + +#define FDCTROW(b0,b1,b2,b3,b4,b5,b6,b7) /* {{{ */ \ + x0 = vec_add(b0, b7); /* x0 = b0 + b7; */ \ + x7 = vec_sub(b0, b7); /* x7 = b0 - b7; */ \ + x1 = vec_add(b1, b6); /* x1 = b1 + b6; */ \ + x6 = vec_sub(b1, b6); /* x6 = b1 - b6; */ \ + x2 = vec_add(b2, b5); /* x2 = b2 + b5; */ \ + x5 = vec_sub(b2, b5); /* x5 = b2 - b5; */ \ + x3 = vec_add(b3, b4); /* x3 = b3 + b4; */ \ + x4 = vec_sub(b3, b4); /* x4 = b3 - b4; */ \ + \ + b7 = vec_add(x0, x3); /* b7 = x0 + x3; */ \ + b1 = vec_add(x1, x2); /* b1 = x1 + x2; */ \ + b0 = vec_add(b7, b1); /* b0 = b7 + b1; */ \ + b4 = vec_sub(b7, b1); /* b4 = b7 - b1; */ \ + \ + b2 = vec_sub(x0, x3); /* b2 = x0 - x3; */ \ + b6 = vec_sub(x1, x2); /* b6 = x1 - x2; */ \ + b5 = vec_add(b6, b2); /* b5 = b6 + b2; */ \ + cnst = LD_W2; \ + b5 = vec_madd(cnst, b5, mzero); /* b5 = b5 * W2; */ \ + cnst = LD_W1; \ + b2 = vec_madd(cnst, b2, b5); /* b2 = b5 + b2 * W1; */ \ + cnst = LD_W0; \ + b6 = vec_madd(cnst, b6, b5); /* b6 = b5 + b6 * W0; */ \ + \ + x0 = vec_add(x4, x7); /* x0 = x4 + x7; */ \ + x1 = vec_add(x5, x6); /* x1 = x5 + x6; */ \ + x2 = vec_add(x4, x6); /* x2 = x4 + x6; */ \ + x3 = vec_add(x5, x7); /* x3 = x5 + x7; */ \ + x8 = vec_add(x2, x3); /* x8 = x2 + x3; */ \ + cnst = LD_W3; \ + x8 = vec_madd(cnst, x8, mzero); /* x8 = x8 * W3; */ \ + \ + cnst = LD_W8; \ + x0 = vec_madd(cnst, x0, mzero); /* x0 *= W8; */ \ + cnst = LD_W9; \ + x1 = vec_madd(cnst, x1, mzero); /* x1 *= W9; */ \ + cnst = LD_WA; \ + x2 = vec_madd(cnst, x2, x8); /* x2 = x2 * WA + x8; */ \ + cnst = LD_WB; \ + x3 = vec_madd(cnst, x3, x8); /* x3 = x3 * WB + x8; */ \ + \ + cnst = LD_W4; \ + b7 = vec_madd(cnst, x4, x0); /* b7 = x4 * W4 + x0; */ \ + cnst = LD_W5; \ + b5 = vec_madd(cnst, x5, x1); /* b5 = x5 * W5 + x1; */ \ + cnst = LD_W6; \ + b3 = vec_madd(cnst, x6, x1); /* b3 = x6 * W6 + x1; */ \ + cnst = LD_W7; \ + b1 = vec_madd(cnst, x7, x0); /* b1 = x7 * W7 + x0; */ \ + \ + b7 = vec_add(b7, x2); /* b7 = b7 + x2; */ \ + b5 = vec_add(b5, x3); /* b5 = b5 + x3; */ \ + b3 = vec_add(b3, x2); /* b3 = b3 + x2; */ \ + b1 = vec_add(b1, x3); /* b1 = b1 + x3; */ \ + /* }}} */ + +#define FDCTCOL(b0,b1,b2,b3,b4,b5,b6,b7) /* {{{ */ \ + x0 = vec_add(b0, b7); /* x0 = b0 + b7; */ \ + x7 = vec_sub(b0, b7); /* x7 = b0 - b7; */ \ + x1 = vec_add(b1, b6); /* x1 = b1 + b6; */ \ + x6 = vec_sub(b1, b6); /* x6 = b1 - b6; */ \ + x2 = vec_add(b2, b5); /* x2 = b2 + b5; */ \ + x5 = vec_sub(b2, b5); /* x5 = b2 - b5; */ \ + x3 = vec_add(b3, b4); /* x3 = b3 + b4; */ \ + x4 = vec_sub(b3, b4); /* x4 = b3 - b4; */ \ + \ + b7 = vec_add(x0, x3); /* b7 = x0 + x3; */ \ + b1 = vec_add(x1, x2); /* b1 = x1 + x2; */ \ + b0 = vec_add(b7, b1); /* b0 = b7 + b1; */ \ + b4 = vec_sub(b7, b1); /* b4 = b7 - b1; */ \ + \ + b2 = vec_sub(x0, x3); /* b2 = x0 - x3; */ \ + b6 = vec_sub(x1, x2); /* b6 = x1 - x2; */ \ + b5 = vec_add(b6, b2); /* b5 = b6 + b2; */ \ + cnst = LD_W2; \ + b5 = vec_madd(cnst, b5, mzero); /* b5 = b5 * W2; */ \ + cnst = LD_W1; \ + b2 = vec_madd(cnst, b2, b5); /* b2 = b5 + b2 * W1; */ \ + cnst = LD_W0; \ + b6 = vec_madd(cnst, b6, b5); /* b6 = b5 + b6 * W0; */ \ + \ + x0 = vec_add(x4, x7); /* x0 = x4 + x7; */ \ + x1 = vec_add(x5, x6); /* x1 = x5 + x6; */ \ + x2 = vec_add(x4, x6); /* x2 = x4 + x6; */ \ + x3 = vec_add(x5, x7); /* x3 = x5 + x7; */ \ + x8 = vec_add(x2, x3); /* x8 = x2 + x3; */ \ + cnst = LD_W3; \ + x8 = vec_madd(cnst, x8, mzero); /* x8 = x8 * W3; */ \ + \ + cnst = LD_W8; \ + x0 = vec_madd(cnst, x0, mzero); /* x0 *= W8; */ \ + cnst = LD_W9; \ + x1 = vec_madd(cnst, x1, mzero); /* x1 *= W9; */ \ + cnst = LD_WA; \ + x2 = vec_madd(cnst, x2, x8); /* x2 = x2 * WA + x8; */ \ + cnst = LD_WB; \ + x3 = vec_madd(cnst, x3, x8); /* x3 = x3 * WB + x8; */ \ + \ + cnst = LD_W4; \ + b7 = vec_madd(cnst, x4, x0); /* b7 = x4 * W4 + x0; */ \ + cnst = LD_W5; \ + b5 = vec_madd(cnst, x5, x1); /* b5 = x5 * W5 + x1; */ \ + cnst = LD_W6; \ + b3 = vec_madd(cnst, x6, x1); /* b3 = x6 * W6 + x1; */ \ + cnst = LD_W7; \ + b1 = vec_madd(cnst, x7, x0); /* b1 = x7 * W7 + x0; */ \ + \ + b7 = vec_add(b7, x2); /* b7 += x2; */ \ + b5 = vec_add(b5, x3); /* b5 += x3; */ \ + b3 = vec_add(b3, x2); /* b3 += x2; */ \ + b1 = vec_add(b1, x3); /* b1 += x3; */ \ + /* }}} */ + + + +/* two dimensional discrete cosine transform */ + +void fdct_altivec(int16_t *block) +{ +POWERPC_PERF_DECLARE(altivec_fdct, 1); + vector signed short *bp; + vector float *cp; + vector float b00, b10, b20, b30, b40, b50, b60, b70; + vector float b01, b11, b21, b31, b41, b51, b61, b71; + vector float mzero, cnst, cnsts0, cnsts1, cnsts2; + vector float x0, x1, x2, x3, x4, x5, x6, x7, x8; + + POWERPC_PERF_START_COUNT(altivec_fdct, 1); + + + /* setup constants {{{ */ + /* mzero = -0.0 */ + mzero = ((vector float)vec_splat_u32(-1)); + mzero = ((vector float)vec_sl(vu32(mzero), vu32(mzero))); + cp = fdctconsts; + cnsts0 = vec_ld(0, cp); cp++; + cnsts1 = vec_ld(0, cp); cp++; + cnsts2 = vec_ld(0, cp); + /* }}} */ + + + /* 8x8 matrix transpose (vector short[8]) {{{ */ +#define MERGE_S16(hl,a,b) vec_merge##hl(vs16(a), vs16(b)) + + bp = (vector signed short*)block; + b00 = ((vector float)vec_ld(0, bp)); + b40 = ((vector float)vec_ld(16*4, bp)); + b01 = ((vector float)MERGE_S16(h, b00, b40)); + b11 = ((vector float)MERGE_S16(l, b00, b40)); + bp++; + b10 = ((vector float)vec_ld(0, bp)); + b50 = ((vector float)vec_ld(16*4, bp)); + b21 = ((vector float)MERGE_S16(h, b10, b50)); + b31 = ((vector float)MERGE_S16(l, b10, b50)); + bp++; + b20 = ((vector float)vec_ld(0, bp)); + b60 = ((vector float)vec_ld(16*4, bp)); + b41 = ((vector float)MERGE_S16(h, b20, b60)); + b51 = ((vector float)MERGE_S16(l, b20, b60)); + bp++; + b30 = ((vector float)vec_ld(0, bp)); + b70 = ((vector float)vec_ld(16*4, bp)); + b61 = ((vector float)MERGE_S16(h, b30, b70)); + b71 = ((vector float)MERGE_S16(l, b30, b70)); + + x0 = ((vector float)MERGE_S16(h, b01, b41)); + x1 = ((vector float)MERGE_S16(l, b01, b41)); + x2 = ((vector float)MERGE_S16(h, b11, b51)); + x3 = ((vector float)MERGE_S16(l, b11, b51)); + x4 = ((vector float)MERGE_S16(h, b21, b61)); + x5 = ((vector float)MERGE_S16(l, b21, b61)); + x6 = ((vector float)MERGE_S16(h, b31, b71)); + x7 = ((vector float)MERGE_S16(l, b31, b71)); + + b00 = ((vector float)MERGE_S16(h, x0, x4)); + b10 = ((vector float)MERGE_S16(l, x0, x4)); + b20 = ((vector float)MERGE_S16(h, x1, x5)); + b30 = ((vector float)MERGE_S16(l, x1, x5)); + b40 = ((vector float)MERGE_S16(h, x2, x6)); + b50 = ((vector float)MERGE_S16(l, x2, x6)); + b60 = ((vector float)MERGE_S16(h, x3, x7)); + b70 = ((vector float)MERGE_S16(l, x3, x7)); + +#undef MERGE_S16 + /* }}} */ + + +/* Some of the initial calculations can be done as vector short before + * conversion to vector float. The following code section takes advantage + * of this. + */ +#if 1 + /* fdct rows {{{ */ + x0 = ((vector float)vec_add(vs16(b00), vs16(b70))); + x7 = ((vector float)vec_sub(vs16(b00), vs16(b70))); + x1 = ((vector float)vec_add(vs16(b10), vs16(b60))); + x6 = ((vector float)vec_sub(vs16(b10), vs16(b60))); + x2 = ((vector float)vec_add(vs16(b20), vs16(b50))); + x5 = ((vector float)vec_sub(vs16(b20), vs16(b50))); + x3 = ((vector float)vec_add(vs16(b30), vs16(b40))); + x4 = ((vector float)vec_sub(vs16(b30), vs16(b40))); + + b70 = ((vector float)vec_add(vs16(x0), vs16(x3))); + b10 = ((vector float)vec_add(vs16(x1), vs16(x2))); + + b00 = ((vector float)vec_add(vs16(b70), vs16(b10))); + b40 = ((vector float)vec_sub(vs16(b70), vs16(b10))); + +#define CTF0(n) \ + b##n##1 = ((vector float)vec_unpackl(vs16(b##n##0))); \ + b##n##0 = ((vector float)vec_unpackh(vs16(b##n##0))); \ + b##n##1 = vec_ctf(vs32(b##n##1), 0); \ + b##n##0 = vec_ctf(vs32(b##n##0), 0); + + CTF0(0); + CTF0(4); + + b20 = ((vector float)vec_sub(vs16(x0), vs16(x3))); + b60 = ((vector float)vec_sub(vs16(x1), vs16(x2))); + + CTF0(2); + CTF0(6); + +#undef CTF0 + + x0 = vec_add(b60, b20); + x1 = vec_add(b61, b21); + + cnst = LD_W2; + x0 = vec_madd(cnst, x0, mzero); + x1 = vec_madd(cnst, x1, mzero); + cnst = LD_W1; + b20 = vec_madd(cnst, b20, x0); + b21 = vec_madd(cnst, b21, x1); + cnst = LD_W0; + b60 = vec_madd(cnst, b60, x0); + b61 = vec_madd(cnst, b61, x1); + +#define CTFX(x,b) \ + b##0 = ((vector float)vec_unpackh(vs16(x))); \ + b##1 = ((vector float)vec_unpackl(vs16(x))); \ + b##0 = vec_ctf(vs32(b##0), 0); \ + b##1 = vec_ctf(vs32(b##1), 0); \ + + CTFX(x4, b7); + CTFX(x5, b5); + CTFX(x6, b3); + CTFX(x7, b1); + +#undef CTFX + + + x0 = vec_add(b70, b10); + x1 = vec_add(b50, b30); + x2 = vec_add(b70, b30); + x3 = vec_add(b50, b10); + x8 = vec_add(x2, x3); + cnst = LD_W3; + x8 = vec_madd(cnst, x8, mzero); + + cnst = LD_W8; + x0 = vec_madd(cnst, x0, mzero); + cnst = LD_W9; + x1 = vec_madd(cnst, x1, mzero); + cnst = LD_WA; + x2 = vec_madd(cnst, x2, x8); + cnst = LD_WB; + x3 = vec_madd(cnst, x3, x8); + + cnst = LD_W4; + b70 = vec_madd(cnst, b70, x0); + cnst = LD_W5; + b50 = vec_madd(cnst, b50, x1); + cnst = LD_W6; + b30 = vec_madd(cnst, b30, x1); + cnst = LD_W7; + b10 = vec_madd(cnst, b10, x0); + + b70 = vec_add(b70, x2); + b50 = vec_add(b50, x3); + b30 = vec_add(b30, x2); + b10 = vec_add(b10, x3); + + + x0 = vec_add(b71, b11); + x1 = vec_add(b51, b31); + x2 = vec_add(b71, b31); + x3 = vec_add(b51, b11); + x8 = vec_add(x2, x3); + cnst = LD_W3; + x8 = vec_madd(cnst, x8, mzero); + + cnst = LD_W8; + x0 = vec_madd(cnst, x0, mzero); + cnst = LD_W9; + x1 = vec_madd(cnst, x1, mzero); + cnst = LD_WA; + x2 = vec_madd(cnst, x2, x8); + cnst = LD_WB; + x3 = vec_madd(cnst, x3, x8); + + cnst = LD_W4; + b71 = vec_madd(cnst, b71, x0); + cnst = LD_W5; + b51 = vec_madd(cnst, b51, x1); + cnst = LD_W6; + b31 = vec_madd(cnst, b31, x1); + cnst = LD_W7; + b11 = vec_madd(cnst, b11, x0); + + b71 = vec_add(b71, x2); + b51 = vec_add(b51, x3); + b31 = vec_add(b31, x2); + b11 = vec_add(b11, x3); + /* }}} */ +#else + /* convert to float {{{ */ +#define CTF(n) \ + vs32(b##n##1) = vec_unpackl(vs16(b##n##0)); \ + vs32(b##n##0) = vec_unpackh(vs16(b##n##0)); \ + b##n##1 = vec_ctf(vs32(b##n##1), 0); \ + b##n##0 = vec_ctf(vs32(b##n##0), 0); \ + + CTF(0); + CTF(1); + CTF(2); + CTF(3); + CTF(4); + CTF(5); + CTF(6); + CTF(7); + +#undef CTF + /* }}} */ + + FDCTROW(b00, b10, b20, b30, b40, b50, b60, b70); + FDCTROW(b01, b11, b21, b31, b41, b51, b61, b71); +#endif + + + /* 8x8 matrix transpose (vector float[8][2]) {{{ */ + x0 = vec_mergel(b00, b20); + x1 = vec_mergeh(b00, b20); + x2 = vec_mergel(b10, b30); + x3 = vec_mergeh(b10, b30); + + b00 = vec_mergeh(x1, x3); + b10 = vec_mergel(x1, x3); + b20 = vec_mergeh(x0, x2); + b30 = vec_mergel(x0, x2); + + x4 = vec_mergel(b41, b61); + x5 = vec_mergeh(b41, b61); + x6 = vec_mergel(b51, b71); + x7 = vec_mergeh(b51, b71); + + b41 = vec_mergeh(x5, x7); + b51 = vec_mergel(x5, x7); + b61 = vec_mergeh(x4, x6); + b71 = vec_mergel(x4, x6); + + x0 = vec_mergel(b01, b21); + x1 = vec_mergeh(b01, b21); + x2 = vec_mergel(b11, b31); + x3 = vec_mergeh(b11, b31); + + x4 = vec_mergel(b40, b60); + x5 = vec_mergeh(b40, b60); + x6 = vec_mergel(b50, b70); + x7 = vec_mergeh(b50, b70); + + b40 = vec_mergeh(x1, x3); + b50 = vec_mergel(x1, x3); + b60 = vec_mergeh(x0, x2); + b70 = vec_mergel(x0, x2); + + b01 = vec_mergeh(x5, x7); + b11 = vec_mergel(x5, x7); + b21 = vec_mergeh(x4, x6); + b31 = vec_mergel(x4, x6); + /* }}} */ + + + FDCTCOL(b00, b10, b20, b30, b40, b50, b60, b70); + FDCTCOL(b01, b11, b21, b31, b41, b51, b61, b71); + + + /* round, convert back to short {{{ */ +#define CTS(n) \ + b##n##0 = vec_round(b##n##0); \ + b##n##1 = vec_round(b##n##1); \ + b##n##0 = ((vector float)vec_cts(b##n##0, 0)); \ + b##n##1 = ((vector float)vec_cts(b##n##1, 0)); \ + b##n##0 = ((vector float)vec_pack(vs32(b##n##0), vs32(b##n##1))); \ + vec_st(vs16(b##n##0), 0, bp); + + bp = (vector signed short*)block; + CTS(0); bp++; + CTS(1); bp++; + CTS(2); bp++; + CTS(3); bp++; + CTS(4); bp++; + CTS(5); bp++; + CTS(6); bp++; + CTS(7); + +#undef CTS + /* }}} */ + +POWERPC_PERF_STOP_COUNT(altivec_fdct, 1); +} + +/* vim:set foldmethod=marker foldlevel=0: */ diff --git a/contrib/ffmpeg/libavcodec/ppc/fft_altivec.c b/contrib/ffmpeg/libavcodec/ppc/fft_altivec.c new file mode 100644 index 000000000..384a774ff --- /dev/null +++ b/contrib/ffmpeg/libavcodec/ppc/fft_altivec.c @@ -0,0 +1,166 @@ +/* + * FFT/IFFT transforms + * AltiVec-enabled + * Copyright (c) 2003 Romain Dolbeau <romain@dolbeau.org> + * Based on code Copyright (c) 2002 Fabrice Bellard. + * + * This file is part of FFmpeg. + * + * FFmpeg 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 2.1 of the License, or (at your option) any later version. + * + * FFmpeg 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 FFmpeg; 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 "dsputil_altivec.h" + +/* + those three macros are from libavcodec/fft.c + and are required for the reference C code +*/ +/* butter fly op */ +#define BF(pre, pim, qre, qim, pre1, pim1, qre1, qim1) \ +{\ + FFTSample ax, ay, bx, by;\ + bx=pre1;\ + by=pim1;\ + ax=qre1;\ + ay=qim1;\ + pre = (bx + ax);\ + pim = (by + ay);\ + qre = (bx - ax);\ + qim = (by - ay);\ +} +#define MUL16(a,b) ((a) * (b)) +#define CMUL(pre, pim, are, aim, bre, bim) \ +{\ + pre = (MUL16(are, bre) - MUL16(aim, bim));\ + pim = (MUL16(are, bim) + MUL16(bre, aim));\ +} + + +/** + * Do a complex FFT with the parameters defined in ff_fft_init(). The + * input data must be permuted before with s->revtab table. No + * 1.0/sqrt(n) normalization is done. + * AltiVec-enabled + * This code assumes that the 'z' pointer is 16 bytes-aligned + * It also assumes all FFTComplex are 8 bytes-aligned pair of float + * The code is exactly the same as the SSE version, except + * that successive MUL + ADD/SUB have been merged into + * fused multiply-add ('vec_madd' in altivec) + */ +void ff_fft_calc_altivec(FFTContext *s, FFTComplex *z) +{ +POWERPC_PERF_DECLARE(altivec_fft_num, s->nbits >= 6); + register const vector float vczero = (const vector float)vec_splat_u32(0.); + + int ln = s->nbits; + int j, np, np2; + int nblocks, nloops; + register FFTComplex *p, *q; + FFTComplex *cptr, *cptr1; + int k; + +POWERPC_PERF_START_COUNT(altivec_fft_num, s->nbits >= 6); + + np = 1 << ln; + + { + vector float *r, a, b, a1, c1, c2; + + r = (vector float *)&z[0]; + + c1 = vcii(p,p,n,n); + + if (s->inverse) + { + c2 = vcii(p,p,n,p); + } + else + { + c2 = vcii(p,p,p,n); + } + + j = (np >> 2); + do { + a = vec_ld(0, r); + a1 = vec_ld(sizeof(vector float), r); + + b = vec_perm(a,a,vcprmle(1,0,3,2)); + a = vec_madd(a,c1,b); + /* do the pass 0 butterfly */ + + b = vec_perm(a1,a1,vcprmle(1,0,3,2)); + b = vec_madd(a1,c1,b); + /* do the pass 0 butterfly */ + + /* multiply third by -i */ + b = vec_perm(b,b,vcprmle(2,3,1,0)); + + /* do the pass 1 butterfly */ + vec_st(vec_madd(b,c2,a), 0, r); + vec_st(vec_nmsub(b,c2,a), sizeof(vector float), r); + + r += 2; + } while (--j != 0); + } + /* pass 2 .. ln-1 */ + + nblocks = np >> 3; + nloops = 1 << 2; + np2 = np >> 1; + + cptr1 = s->exptab1; + do { + p = z; + q = z + nloops; + j = nblocks; + do { + cptr = cptr1; + k = nloops >> 1; + do { + vector float a,b,c,t1; + + a = vec_ld(0, (float*)p); + b = vec_ld(0, (float*)q); + + /* complex mul */ + c = vec_ld(0, (float*)cptr); + /* cre*re cim*re */ + t1 = vec_madd(c, vec_perm(b,b,vcprmle(2,2,0,0)),vczero); + c = vec_ld(sizeof(vector float), (float*)cptr); + /* -cim*im cre*im */ + b = vec_madd(c, vec_perm(b,b,vcprmle(3,3,1,1)),t1); + + /* butterfly */ + vec_st(vec_add(a,b), 0, (float*)p); + vec_st(vec_sub(a,b), 0, (float*)q); + + p += 2; + q += 2; + cptr += 4; + } while (--k); + + p += nloops; + q += nloops; + } while (--j); + cptr1 += nloops * 2; + nblocks = nblocks >> 1; + nloops = nloops << 1; + } while (nblocks != 0); + +POWERPC_PERF_STOP_COUNT(altivec_fft_num, s->nbits >= 6); +} diff --git a/contrib/ffmpeg/libavcodec/ppc/float_altivec.c b/contrib/ffmpeg/libavcodec/ppc/float_altivec.c new file mode 100644 index 000000000..c6e43dec2 --- /dev/null +++ b/contrib/ffmpeg/libavcodec/ppc/float_altivec.c @@ -0,0 +1,194 @@ +/* + * Copyright (c) 2006 Luca Barbato <lu_zero@gentoo.org> + * + * This file is part of FFmpeg. + * + * FFmpeg 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 2.1 of the License, or (at your option) any later version. + * + * FFmpeg 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 FFmpeg; 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 "dsputil_altivec.h" + +static void vector_fmul_altivec(float *dst, const float *src, int len) +{ + int i; + vector float d0, d1, s, zero = (vector float)vec_splat_u32(0); + for(i=0; i<len-7; i+=8) { + d0 = vec_ld(0, dst+i); + s = vec_ld(0, src+i); + d1 = vec_ld(16, dst+i); + d0 = vec_madd(d0, s, zero); + d1 = vec_madd(d1, vec_ld(16,src+i), zero); + vec_st(d0, 0, dst+i); + vec_st(d1, 16, dst+i); + } +} + +static void vector_fmul_reverse_altivec(float *dst, const float *src0, + const float *src1, int len) +{ + int i; + vector float d, s0, s1, h0, l0, + s2, s3, zero = (vector float)vec_splat_u32(0); + src1 += len-4; + for(i=0; i<len-7; i+=8) { + s1 = vec_ld(0, src1-i); // [a,b,c,d] + s0 = vec_ld(0, src0+i); + l0 = vec_mergel(s1, s1); // [c,c,d,d] + s3 = vec_ld(-16, src1-i); + h0 = vec_mergeh(s1, s1); // [a,a,b,b] + s2 = vec_ld(16, src0+i); + s1 = vec_mergeh(vec_mergel(l0,h0), // [d,b,d,b] + vec_mergeh(l0,h0)); // [c,a,c,a] + // [d,c,b,a] + l0 = vec_mergel(s3, s3); + d = vec_madd(s0, s1, zero); + h0 = vec_mergeh(s3, s3); + vec_st(d, 0, dst+i); + s3 = vec_mergeh(vec_mergel(l0,h0), + vec_mergeh(l0,h0)); + d = vec_madd(s2, s3, zero); + vec_st(d, 16, dst+i); + } +} + +static void vector_fmul_add_add_altivec(float *dst, const float *src0, + const float *src1, const float *src2, + int src3, int len, int step) +{ + int i; + vector float d, s0, s1, s2, t0, t1, edges; + vector unsigned char align = vec_lvsr(0,dst), + mask = vec_lvsl(0, dst); + + t0 = vec_ld(0, dst); +#if 0 //FIXME: there is still something wrong + if (step == 2) { + int y; + vector float d0, d1, s3, t2; + vector unsigned int sel = + vec_mergeh(vec_splat_u32(-1), vec_splat_u32(0)); + t1 = vec_ld(16, dst); + for (i=0,y=0; i<len-3; i+=4,y+=8) { + + s0 = vec_ld(0,src0+i); + s1 = vec_ld(0,src1+i); + s2 = vec_ld(0,src2+i); + +// t0 = vec_ld(0, dst+y); //[x x x|a] +// t1 = vec_ld(16, dst+y); //[b c d|e] + t2 = vec_ld(31, dst+y); //[f g h|x] + + d = vec_madd(s0,s1,s2); // [A B C D] + + // [A A B B] + + // [C C D D] + + d0 = vec_perm(t0, t1, mask); // [a b c d] + + d0 = vec_sel(vec_mergeh(d, d), d0, sel); // [A b B d] + + edges = vec_perm(t1, t0, mask); + + t0 = vec_perm(edges, d0, align); // [x x x|A] + + t1 = vec_perm(d0, edges, align); // [b B d|e] + + vec_stl(t0, 0, dst+y); + + d1 = vec_perm(t1, t2, mask); // [e f g h] + + d1 = vec_sel(vec_mergel(d, d), d1, sel); // [C f D h] + + edges = vec_perm(t2, t1, mask); + + t1 = vec_perm(edges, d1, align); // [b B d|C] + + t2 = vec_perm(d1, edges, align); // [f D h|x] + + vec_stl(t1, 16, dst+y); + + t0 = t1; + + vec_stl(t2, 31, dst+y); + + t1 = t2; + } + } else + #endif + if (step == 1 && src3 == 0) + for (i=0; i<len-3; i+=4) { + t1 = vec_ld(15, dst+i); + s0 = vec_ld(0, src0+i); + s1 = vec_ld(0, src1+i); + s2 = vec_ld(0, src2+i); + edges = vec_perm(t1 ,t0, mask); + d = vec_madd(s0,s1,s2); + t1 = vec_perm(d, edges, align); + t0 = vec_perm(edges, d, align); + vec_st(t1, 15, dst+i); + vec_st(t0, 0, dst+i); + t0 = t1; + } + else + ff_vector_fmul_add_add_c(dst, src0, src1, src2, src3, len, step); +} + +void float_to_int16_altivec(int16_t *dst, const float *src, int len) +{ + int i; + vector float s0, s1; + vector signed int t0, t1; + vector signed short d0, d1, d; + vector unsigned char align; + if(((long)dst)&15) //FIXME + for(i=0; i<len-7; i+=8) { + s0 = vec_ld(0, src+i); + s1 = vec_ld(16, src+i); + t0 = vec_cts(s0, 0); + d0 = vec_ld(0, dst+i); + t1 = vec_cts(s1, 0); + d1 = vec_ld(15, dst+i); + d = vec_packs(t0,t1); + d1 = vec_perm(d1, d0, vec_lvsl(0,dst+i)); + align = vec_lvsr(0, dst+i); + d0 = vec_perm(d1, d, align); + d1 = vec_perm(d, d1, align); + vec_st(d0, 0, dst+i); + vec_st(d1,15, dst+i); + } + else + for(i=0; i<len-7; i+=8) { + s0 = vec_ld(0, src+i); + s1 = vec_ld(16, src+i); + t0 = vec_cts(s0, 0); + t1 = vec_cts(s1, 0); + d = vec_packs(t0,t1); + vec_st(d, 0, dst+i); + } +} + +void float_init_altivec(DSPContext* c, AVCodecContext *avctx) +{ + c->vector_fmul = vector_fmul_altivec; + c->vector_fmul_reverse = vector_fmul_reverse_altivec; + c->vector_fmul_add_add = vector_fmul_add_add_altivec; + if(!(avctx->flags & CODEC_FLAG_BITEXACT)) + c->float_to_int16 = float_to_int16_altivec; +} diff --git a/contrib/ffmpeg/libavcodec/ppc/gcc_fixes.h b/contrib/ffmpeg/libavcodec/ppc/gcc_fixes.h new file mode 100644 index 000000000..5a4a55188 --- /dev/null +++ b/contrib/ffmpeg/libavcodec/ppc/gcc_fixes.h @@ -0,0 +1,119 @@ +/* + * gcc fixes for altivec. + * Used to workaround broken gcc (FSF gcc-3 pre gcc-3.3) + * and to stay somewhat compatible with Darwin. + * + * This file is part of FFmpeg. + * + * FFmpeg 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 2.1 of the License, or (at your option) any later version. + * + * FFmpeg 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 FFmpeg; if not, write to the Free Software + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA + */ + +#ifndef _GCC_FIXES_ +#define _GCC_FIXES_ + +#ifdef HAVE_ALTIVEC_H +#include <altivec.h> +#endif + +#ifdef CONFIG_DARWIN +# ifndef __MWERKS__ +# define AVV(x...) (x) +# else +# define AVV +# endif +#define REG_v(a) asm ( #a ) +#else + +#define AVV(x...) {x} + +#if (__GNUC__ < 4) +# define REG_v(a) +#else +# define REG_v(a) asm ( #a ) +#endif + +#if (__GNUC__ * 100 + __GNUC_MINOR__ < 303) + +/* This code was provided to me by Bartosch Pixa + * as a separate header file (broken_mergel.h). + * thanks to lu_zero for the workaround. + * + * See this mail for more information: + * http://gcc.gnu.org/ml/gcc/2003-04/msg00967.html + */ + +static inline vector signed char ff_vmrglb (vector signed char const A, + vector signed char const B) +{ + static const vector unsigned char lowbyte = { + 0x08, 0x18, 0x09, 0x19, 0x0a, 0x1a, 0x0b, 0x1b, + 0x0c, 0x1c, 0x0d, 0x1d, 0x0e, 0x1e, 0x0f, 0x1f + }; + return vec_perm (A, B, lowbyte); +} + +static inline vector signed short ff_vmrglh (vector signed short const A, + vector signed short const B) +{ + static const vector unsigned char lowhalf = { + 0x08, 0x09, 0x18, 0x19, 0x0a, 0x0b, 0x1a, 0x1b, + 0x0c, 0x0d, 0x1c, 0x1d, 0x0e, 0x0f, 0x1e, 0x1f + }; + return vec_perm (A, B, lowhalf); +} + +static inline vector signed int ff_vmrglw (vector signed int const A, + vector signed int const B) +{ + static const vector unsigned char lowword = { + 0x08, 0x09, 0x0a, 0x0b, 0x18, 0x19, 0x1a, 0x1b, + 0x0c, 0x0d, 0x0e, 0x0f, 0x1c, 0x1d, 0x1e, 0x1f + }; + return vec_perm (A, B, lowword); +} +/*#define ff_vmrglb ff_vmrglb +#define ff_vmrglh ff_vmrglh +#define ff_vmrglw ff_vmrglw +*/ +#undef vec_mergel + +#define vec_mergel(a1, a2) \ +__ch (__bin_args_eq (vector signed char, (a1), vector signed char, (a2)), \ + ((vector signed char) ff_vmrglb ((vector signed char) (a1), (vector signed char) (a2))), \ +__ch (__bin_args_eq (vector unsigned char, (a1), vector unsigned char, (a2)), \ + ((vector unsigned char) ff_vmrglb ((vector signed char) (a1), (vector signed char) (a2))), \ +__ch (__bin_args_eq (vector signed short, (a1), vector signed short, (a2)), \ + ((vector signed short) ff_vmrglh ((vector signed short) (a1), (vector signed short) (a2))), \ +__ch (__bin_args_eq (vector unsigned short, (a1), vector unsigned short, (a2)), \ + ((vector unsigned short) ff_vmrglh ((vector signed short) (a1), (vector signed short) (a2))), \ +__ch (__bin_args_eq (vector float, (a1), vector float, (a2)), \ + ((vector float) ff_vmrglw ((vector signed int) (a1), (vector signed int) (a2))), \ +__ch (__bin_args_eq (vector signed int, (a1), vector signed int, (a2)), \ + ((vector signed int) ff_vmrglw ((vector signed int) (a1), (vector signed int) (a2))), \ +__ch (__bin_args_eq (vector unsigned int, (a1), vector unsigned int, (a2)), \ + ((vector unsigned int) ff_vmrglw ((vector signed int) (a1), (vector signed int) (a2))), \ + __altivec_link_error_invalid_argument ()))))))) + +#endif + +#endif /* CONFIG_DARWIN */ + +#ifndef __MWERKS__ +#define const_vector const vector +#else +#define const_vector vector +#endif + +#endif /* _GCC_FIXES_ */ diff --git a/contrib/ffmpeg/libavcodec/ppc/gmc_altivec.c b/contrib/ffmpeg/libavcodec/ppc/gmc_altivec.c new file mode 100644 index 000000000..42c936bb3 --- /dev/null +++ b/contrib/ffmpeg/libavcodec/ppc/gmc_altivec.c @@ -0,0 +1,146 @@ +/* + * GMC (Global Motion Compensation) + * AltiVec-enabled + * Copyright (c) 2003 Romain Dolbeau <romain@dolbeau.org> + * + * This file is part of FFmpeg. + * + * FFmpeg 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 2.1 of the License, or (at your option) any later version. + * + * FFmpeg 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 FFmpeg; 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 "dsputil_altivec.h" + +/* + altivec-enhanced gmc1. ATM this code assume stride is a multiple of 8, + to preserve proper dst alignement. +*/ +#define GMC1_PERF_COND (h==8) +void gmc1_altivec(uint8_t *dst /* align 8 */, uint8_t *src /* align1 */, int stride, int h, int x16, int y16, int rounder) +{ +POWERPC_PERF_DECLARE(altivec_gmc1_num, GMC1_PERF_COND); + const unsigned short __attribute__ ((aligned(16))) rounder_a[8] = + {rounder, rounder, rounder, rounder, + rounder, rounder, rounder, rounder}; + const unsigned short __attribute__ ((aligned(16))) ABCD[8] = + { + (16-x16)*(16-y16), /* A */ + ( x16)*(16-y16), /* B */ + (16-x16)*( y16), /* C */ + ( x16)*( y16), /* D */ + 0, 0, 0, 0 /* padding */ + }; + register const_vector unsigned char vczero = (const_vector unsigned char)vec_splat_u8(0); + register const_vector unsigned short vcsr8 = (const_vector unsigned short)vec_splat_u16(8); + register vector unsigned char dstv, dstv2, src_0, src_1, srcvA, srcvB, srcvC, srcvD; + register vector unsigned short Av, Bv, Cv, Dv, rounderV, tempA, tempB, tempC, tempD; + int i; + unsigned long dst_odd = (unsigned long)dst & 0x0000000F; + unsigned long src_really_odd = (unsigned long)src & 0x0000000F; + + +POWERPC_PERF_START_COUNT(altivec_gmc1_num, GMC1_PERF_COND); + + tempA = vec_ld(0, (unsigned short*)ABCD); + Av = vec_splat(tempA, 0); + Bv = vec_splat(tempA, 1); + Cv = vec_splat(tempA, 2); + Dv = vec_splat(tempA, 3); + + rounderV = vec_ld(0, (unsigned short*)rounder_a); + + // we'll be able to pick-up our 9 char elements + // at src from those 32 bytes + // we load the first batch here, as inside the loop + // we can re-use 'src+stride' from one iteration + // as the 'src' of the next. + src_0 = vec_ld(0, src); + src_1 = vec_ld(16, src); + srcvA = vec_perm(src_0, src_1, vec_lvsl(0, src)); + + if (src_really_odd != 0x0000000F) + { // if src & 0xF == 0xF, then (src+1) is properly aligned on the second vector. + srcvB = vec_perm(src_0, src_1, vec_lvsl(1, src)); + } + else + { + srcvB = src_1; + } + srcvA = vec_mergeh(vczero, srcvA); + srcvB = vec_mergeh(vczero, srcvB); + + for(i=0; i<h; i++) + { + dst_odd = (unsigned long)dst & 0x0000000F; + src_really_odd = (((unsigned long)src) + stride) & 0x0000000F; + + dstv = vec_ld(0, dst); + + // we we'll be able to pick-up our 9 char elements + // at src + stride from those 32 bytes + // then reuse the resulting 2 vectors srvcC and srcvD + // as the next srcvA and srcvB + src_0 = vec_ld(stride + 0, src); + src_1 = vec_ld(stride + 16, src); + srcvC = vec_perm(src_0, src_1, vec_lvsl(stride + 0, src)); + + if (src_really_odd != 0x0000000F) + { // if src & 0xF == 0xF, then (src+1) is properly aligned on the second vector. + srcvD = vec_perm(src_0, src_1, vec_lvsl(stride + 1, src)); + } + else + { + srcvD = src_1; + } + + srcvC = vec_mergeh(vczero, srcvC); + srcvD = vec_mergeh(vczero, srcvD); + + + // OK, now we (finally) do the math :-) + // those four instructions replaces 32 int muls & 32 int adds. + // isn't AltiVec nice ? + tempA = vec_mladd((vector unsigned short)srcvA, Av, rounderV); + tempB = vec_mladd((vector unsigned short)srcvB, Bv, tempA); + tempC = vec_mladd((vector unsigned short)srcvC, Cv, tempB); + tempD = vec_mladd((vector unsigned short)srcvD, Dv, tempC); + + srcvA = srcvC; + srcvB = srcvD; + + tempD = vec_sr(tempD, vcsr8); + + dstv2 = vec_pack(tempD, (vector unsigned short)vczero); + + if (dst_odd) + { + dstv2 = vec_perm(dstv, dstv2, vcprm(0,1,s0,s1)); + } + else + { + dstv2 = vec_perm(dstv, dstv2, vcprm(s0,s1,2,3)); + } + + vec_st(dstv2, 0, dst); + + dst += stride; + src += stride; + } + +POWERPC_PERF_STOP_COUNT(altivec_gmc1_num, GMC1_PERF_COND); +} diff --git a/contrib/ffmpeg/libavcodec/ppc/h264_altivec.c b/contrib/ffmpeg/libavcodec/ppc/h264_altivec.c new file mode 100644 index 000000000..4aa366f97 --- /dev/null +++ b/contrib/ffmpeg/libavcodec/ppc/h264_altivec.c @@ -0,0 +1,438 @@ +/* + * Copyright (c) 2004 Romain Dolbeau <romain@dolbeau.org> + * + * This file is part of FFmpeg. + * + * FFmpeg 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 2.1 of the License, or (at your option) any later version. + * + * FFmpeg 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 FFmpeg; 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 "dsputil_altivec.h" + +#define PUT_OP_U8_ALTIVEC(d, s, dst) d = s +#define AVG_OP_U8_ALTIVEC(d, s, dst) d = vec_avg(dst, s) + +#define OP_U8_ALTIVEC PUT_OP_U8_ALTIVEC +#define PREFIX_h264_chroma_mc8_altivec put_h264_chroma_mc8_altivec +#define PREFIX_h264_chroma_mc8_num altivec_put_h264_chroma_mc8_num +#define PREFIX_h264_qpel16_h_lowpass_altivec put_h264_qpel16_h_lowpass_altivec +#define PREFIX_h264_qpel16_h_lowpass_num altivec_put_h264_qpel16_h_lowpass_num +#define PREFIX_h264_qpel16_v_lowpass_altivec put_h264_qpel16_v_lowpass_altivec +#define PREFIX_h264_qpel16_v_lowpass_num altivec_put_h264_qpel16_v_lowpass_num +#define PREFIX_h264_qpel16_hv_lowpass_altivec put_h264_qpel16_hv_lowpass_altivec +#define PREFIX_h264_qpel16_hv_lowpass_num altivec_put_h264_qpel16_hv_lowpass_num +#include "h264_template_altivec.c" +#undef OP_U8_ALTIVEC +#undef PREFIX_h264_chroma_mc8_altivec +#undef PREFIX_h264_chroma_mc8_num +#undef PREFIX_h264_qpel16_h_lowpass_altivec +#undef PREFIX_h264_qpel16_h_lowpass_num +#undef PREFIX_h264_qpel16_v_lowpass_altivec +#undef PREFIX_h264_qpel16_v_lowpass_num +#undef PREFIX_h264_qpel16_hv_lowpass_altivec +#undef PREFIX_h264_qpel16_hv_lowpass_num + +#define OP_U8_ALTIVEC AVG_OP_U8_ALTIVEC +#define PREFIX_h264_chroma_mc8_altivec avg_h264_chroma_mc8_altivec +#define PREFIX_h264_chroma_mc8_num altivec_avg_h264_chroma_mc8_num +#define PREFIX_h264_qpel16_h_lowpass_altivec avg_h264_qpel16_h_lowpass_altivec +#define PREFIX_h264_qpel16_h_lowpass_num altivec_avg_h264_qpel16_h_lowpass_num +#define PREFIX_h264_qpel16_v_lowpass_altivec avg_h264_qpel16_v_lowpass_altivec +#define PREFIX_h264_qpel16_v_lowpass_num altivec_avg_h264_qpel16_v_lowpass_num +#define PREFIX_h264_qpel16_hv_lowpass_altivec avg_h264_qpel16_hv_lowpass_altivec +#define PREFIX_h264_qpel16_hv_lowpass_num altivec_avg_h264_qpel16_hv_lowpass_num +#include "h264_template_altivec.c" +#undef OP_U8_ALTIVEC +#undef PREFIX_h264_chroma_mc8_altivec +#undef PREFIX_h264_chroma_mc8_num +#undef PREFIX_h264_qpel16_h_lowpass_altivec +#undef PREFIX_h264_qpel16_h_lowpass_num +#undef PREFIX_h264_qpel16_v_lowpass_altivec +#undef PREFIX_h264_qpel16_v_lowpass_num +#undef PREFIX_h264_qpel16_hv_lowpass_altivec +#undef PREFIX_h264_qpel16_hv_lowpass_num + +#define H264_MC(OPNAME, SIZE, CODETYPE) \ +static void OPNAME ## h264_qpel ## SIZE ## _mc00_ ## CODETYPE (uint8_t *dst, uint8_t *src, int stride){\ + OPNAME ## pixels ## SIZE ## _ ## CODETYPE(dst, src, stride, SIZE);\ +}\ +\ +static void OPNAME ## h264_qpel ## SIZE ## _mc10_ ## CODETYPE(uint8_t *dst, uint8_t *src, int stride){ \ + DECLARE_ALIGNED_16(uint8_t, half[SIZE*SIZE]);\ + put_h264_qpel ## SIZE ## _h_lowpass_ ## CODETYPE(half, src, SIZE, stride);\ + OPNAME ## pixels ## SIZE ## _l2_ ## CODETYPE(dst, src, half, stride, stride, SIZE);\ +}\ +\ +static void OPNAME ## h264_qpel ## SIZE ## _mc20_ ## CODETYPE(uint8_t *dst, uint8_t *src, int stride){\ + OPNAME ## h264_qpel ## SIZE ## _h_lowpass_ ## CODETYPE(dst, src, stride, stride);\ +}\ +\ +static void OPNAME ## h264_qpel ## SIZE ## _mc30_ ## CODETYPE(uint8_t *dst, uint8_t *src, int stride){\ + DECLARE_ALIGNED_16(uint8_t, half[SIZE*SIZE]);\ + put_h264_qpel ## SIZE ## _h_lowpass_ ## CODETYPE(half, src, SIZE, stride);\ + OPNAME ## pixels ## SIZE ## _l2_ ## CODETYPE(dst, src+1, half, stride, stride, SIZE);\ +}\ +\ +static void OPNAME ## h264_qpel ## SIZE ## _mc01_ ## CODETYPE(uint8_t *dst, uint8_t *src, int stride){\ + DECLARE_ALIGNED_16(uint8_t, half[SIZE*SIZE]);\ + put_h264_qpel ## SIZE ## _v_lowpass_ ## CODETYPE(half, src, SIZE, stride);\ + OPNAME ## pixels ## SIZE ## _l2_ ## CODETYPE(dst, src, half, stride, stride, SIZE);\ +}\ +\ +static void OPNAME ## h264_qpel ## SIZE ## _mc02_ ## CODETYPE(uint8_t *dst, uint8_t *src, int stride){\ + OPNAME ## h264_qpel ## SIZE ## _v_lowpass_ ## CODETYPE(dst, src, stride, stride);\ +}\ +\ +static void OPNAME ## h264_qpel ## SIZE ## _mc03_ ## CODETYPE(uint8_t *dst, uint8_t *src, int stride){\ + DECLARE_ALIGNED_16(uint8_t, half[SIZE*SIZE]);\ + put_h264_qpel ## SIZE ## _v_lowpass_ ## CODETYPE(half, src, SIZE, stride);\ + OPNAME ## pixels ## SIZE ## _l2_ ## CODETYPE(dst, src+stride, half, stride, stride, SIZE);\ +}\ +\ +static void OPNAME ## h264_qpel ## SIZE ## _mc11_ ## CODETYPE(uint8_t *dst, uint8_t *src, int stride){\ + DECLARE_ALIGNED_16(uint8_t, halfH[SIZE*SIZE]);\ + DECLARE_ALIGNED_16(uint8_t, halfV[SIZE*SIZE]);\ + put_h264_qpel ## SIZE ## _h_lowpass_ ## CODETYPE(halfH, src, SIZE, stride);\ + put_h264_qpel ## SIZE ## _v_lowpass_ ## CODETYPE(halfV, src, SIZE, stride);\ + OPNAME ## pixels ## SIZE ## _l2_ ## CODETYPE(dst, halfH, halfV, stride, SIZE, SIZE);\ +}\ +\ +static void OPNAME ## h264_qpel ## SIZE ## _mc31_ ## CODETYPE(uint8_t *dst, uint8_t *src, int stride){\ + DECLARE_ALIGNED_16(uint8_t, halfH[SIZE*SIZE]);\ + DECLARE_ALIGNED_16(uint8_t, halfV[SIZE*SIZE]);\ + put_h264_qpel ## SIZE ## _h_lowpass_ ## CODETYPE(halfH, src, SIZE, stride);\ + put_h264_qpel ## SIZE ## _v_lowpass_ ## CODETYPE(halfV, src+1, SIZE, stride);\ + OPNAME ## pixels ## SIZE ## _l2_ ## CODETYPE(dst, halfH, halfV, stride, SIZE, SIZE);\ +}\ +\ +static void OPNAME ## h264_qpel ## SIZE ## _mc13_ ## CODETYPE(uint8_t *dst, uint8_t *src, int stride){\ + DECLARE_ALIGNED_16(uint8_t, halfH[SIZE*SIZE]);\ + DECLARE_ALIGNED_16(uint8_t, halfV[SIZE*SIZE]);\ + put_h264_qpel ## SIZE ## _h_lowpass_ ## CODETYPE(halfH, src + stride, SIZE, stride);\ + put_h264_qpel ## SIZE ## _v_lowpass_ ## CODETYPE(halfV, src, SIZE, stride);\ + OPNAME ## pixels ## SIZE ## _l2_ ## CODETYPE(dst, halfH, halfV, stride, SIZE, SIZE);\ +}\ +\ +static void OPNAME ## h264_qpel ## SIZE ## _mc33_ ## CODETYPE(uint8_t *dst, uint8_t *src, int stride){\ + DECLARE_ALIGNED_16(uint8_t, halfH[SIZE*SIZE]);\ + DECLARE_ALIGNED_16(uint8_t, halfV[SIZE*SIZE]);\ + put_h264_qpel ## SIZE ## _h_lowpass_ ## CODETYPE(halfH, src + stride, SIZE, stride);\ + put_h264_qpel ## SIZE ## _v_lowpass_ ## CODETYPE(halfV, src+1, SIZE, stride);\ + OPNAME ## pixels ## SIZE ## _l2_ ## CODETYPE(dst, halfH, halfV, stride, SIZE, SIZE);\ +}\ +\ +static void OPNAME ## h264_qpel ## SIZE ## _mc22_ ## CODETYPE(uint8_t *dst, uint8_t *src, int stride){\ + DECLARE_ALIGNED_16(int16_t, tmp[SIZE*(SIZE+8)]);\ + OPNAME ## h264_qpel ## SIZE ## _hv_lowpass_ ## CODETYPE(dst, tmp, src, stride, SIZE, stride);\ +}\ +\ +static void OPNAME ## h264_qpel ## SIZE ## _mc21_ ## CODETYPE(uint8_t *dst, uint8_t *src, int stride){\ + DECLARE_ALIGNED_16(uint8_t, halfH[SIZE*SIZE]);\ + DECLARE_ALIGNED_16(uint8_t, halfHV[SIZE*SIZE]);\ + DECLARE_ALIGNED_16(int16_t, tmp[SIZE*(SIZE+8)]);\ + put_h264_qpel ## SIZE ## _h_lowpass_ ## CODETYPE(halfH, src, SIZE, stride);\ + put_h264_qpel ## SIZE ## _hv_lowpass_ ## CODETYPE(halfHV, tmp, src, SIZE, SIZE, stride);\ + OPNAME ## pixels ## SIZE ## _l2_ ## CODETYPE(dst, halfH, halfHV, stride, SIZE, SIZE);\ +}\ +\ +static void OPNAME ## h264_qpel ## SIZE ## _mc23_ ## CODETYPE(uint8_t *dst, uint8_t *src, int stride){\ + DECLARE_ALIGNED_16(uint8_t, halfH[SIZE*SIZE]);\ + DECLARE_ALIGNED_16(uint8_t, halfHV[SIZE*SIZE]);\ + DECLARE_ALIGNED_16(int16_t, tmp[SIZE*(SIZE+8)]);\ + put_h264_qpel ## SIZE ## _h_lowpass_ ## CODETYPE(halfH, src + stride, SIZE, stride);\ + put_h264_qpel ## SIZE ## _hv_lowpass_ ## CODETYPE(halfHV, tmp, src, SIZE, SIZE, stride);\ + OPNAME ## pixels ## SIZE ## _l2_ ## CODETYPE(dst, halfH, halfHV, stride, SIZE, SIZE);\ +}\ +\ +static void OPNAME ## h264_qpel ## SIZE ## _mc12_ ## CODETYPE(uint8_t *dst, uint8_t *src, int stride){\ + DECLARE_ALIGNED_16(uint8_t, halfV[SIZE*SIZE]);\ + DECLARE_ALIGNED_16(uint8_t, halfHV[SIZE*SIZE]);\ + DECLARE_ALIGNED_16(int16_t, tmp[SIZE*(SIZE+8)]);\ + put_h264_qpel ## SIZE ## _v_lowpass_ ## CODETYPE(halfV, src, SIZE, stride);\ + put_h264_qpel ## SIZE ## _hv_lowpass_ ## CODETYPE(halfHV, tmp, src, SIZE, SIZE, stride);\ + OPNAME ## pixels ## SIZE ## _l2_ ## CODETYPE(dst, halfV, halfHV, stride, SIZE, SIZE);\ +}\ +\ +static void OPNAME ## h264_qpel ## SIZE ## _mc32_ ## CODETYPE(uint8_t *dst, uint8_t *src, int stride){\ + DECLARE_ALIGNED_16(uint8_t, halfV[SIZE*SIZE]);\ + DECLARE_ALIGNED_16(uint8_t, halfHV[SIZE*SIZE]);\ + DECLARE_ALIGNED_16(int16_t, tmp[SIZE*(SIZE+8)]);\ + put_h264_qpel ## SIZE ## _v_lowpass_ ## CODETYPE(halfV, src+1, SIZE, stride);\ + put_h264_qpel ## SIZE ## _hv_lowpass_ ## CODETYPE(halfHV, tmp, src, SIZE, SIZE, stride);\ + OPNAME ## pixels ## SIZE ## _l2_ ## CODETYPE(dst, halfV, halfHV, stride, SIZE, SIZE);\ +}\ + +/* this code assume that stride % 16 == 0 */ +void put_no_rnd_h264_chroma_mc8_altivec(uint8_t * dst, uint8_t * src, int stride, int h, int x, int y) { + signed int ABCD[4] __attribute__((aligned(16))) = + {((8 - x) * (8 - y)), + ((x) * (8 - y)), + ((8 - x) * (y)), + ((x) * (y))}; + register int i; + vector unsigned char fperm; + const vector signed int vABCD = vec_ld(0, ABCD); + const vector signed short vA = vec_splat((vector signed short)vABCD, 1); + const vector signed short vB = vec_splat((vector signed short)vABCD, 3); + const vector signed short vC = vec_splat((vector signed short)vABCD, 5); + const vector signed short vD = vec_splat((vector signed short)vABCD, 7); + const vector signed int vzero = vec_splat_s32(0); + const vector signed short v28ss = vec_sub(vec_sl(vec_splat_s16(1),vec_splat_u16(5)),vec_splat_s16(4)); + const vector unsigned short v6us = vec_splat_u16(6); + register int loadSecond = (((unsigned long)src) % 16) <= 7 ? 0 : 1; + register int reallyBadAlign = (((unsigned long)src) % 16) == 15 ? 1 : 0; + + vector unsigned char vsrcAuc, vsrcBuc, vsrcperm0, vsrcperm1; + vector unsigned char vsrc0uc, vsrc1uc; + vector signed short vsrc0ssH, vsrc1ssH; + vector unsigned char vsrcCuc, vsrc2uc, vsrc3uc; + vector signed short vsrc2ssH, vsrc3ssH, psum; + vector unsigned char vdst, ppsum, vfdst, fsum; + + if (((unsigned long)dst) % 16 == 0) { + fperm = (vector unsigned char)AVV(0x10, 0x11, 0x12, 0x13, + 0x14, 0x15, 0x16, 0x17, + 0x08, 0x09, 0x0A, 0x0B, + 0x0C, 0x0D, 0x0E, 0x0F); + } else { + fperm = (vector unsigned char)AVV(0x00, 0x01, 0x02, 0x03, + 0x04, 0x05, 0x06, 0x07, + 0x18, 0x19, 0x1A, 0x1B, + 0x1C, 0x1D, 0x1E, 0x1F); + } + + vsrcAuc = vec_ld(0, src); + + if (loadSecond) + vsrcBuc = vec_ld(16, src); + vsrcperm0 = vec_lvsl(0, src); + vsrcperm1 = vec_lvsl(1, src); + + vsrc0uc = vec_perm(vsrcAuc, vsrcBuc, vsrcperm0); + if (reallyBadAlign) + vsrc1uc = vsrcBuc; + else + vsrc1uc = vec_perm(vsrcAuc, vsrcBuc, vsrcperm1); + + vsrc0ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero, + (vector unsigned char)vsrc0uc); + vsrc1ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero, + (vector unsigned char)vsrc1uc); + + if (!loadSecond) {// -> !reallyBadAlign + for (i = 0 ; i < h ; i++) { + + + vsrcCuc = vec_ld(stride + 0, src); + + vsrc2uc = vec_perm(vsrcCuc, vsrcCuc, vsrcperm0); + vsrc3uc = vec_perm(vsrcCuc, vsrcCuc, vsrcperm1); + + vsrc2ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero, + (vector unsigned char)vsrc2uc); + vsrc3ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero, + (vector unsigned char)vsrc3uc); + + psum = vec_mladd(vA, vsrc0ssH, vec_splat_s16(0)); + psum = vec_mladd(vB, vsrc1ssH, psum); + psum = vec_mladd(vC, vsrc2ssH, psum); + psum = vec_mladd(vD, vsrc3ssH, psum); + psum = vec_add(v28ss, psum); + psum = vec_sra(psum, v6us); + + vdst = vec_ld(0, dst); + ppsum = (vector unsigned char)vec_packsu(psum, psum); + fsum = vec_perm(vdst, ppsum, fperm); + + vec_st(fsum, 0, dst); + + vsrc0ssH = vsrc2ssH; + vsrc1ssH = vsrc3ssH; + + dst += stride; + src += stride; + } + } else { + vector unsigned char vsrcDuc; + for (i = 0 ; i < h ; i++) { + vsrcCuc = vec_ld(stride + 0, src); + vsrcDuc = vec_ld(stride + 16, src); + + vsrc2uc = vec_perm(vsrcCuc, vsrcDuc, vsrcperm0); + if (reallyBadAlign) + vsrc3uc = vsrcDuc; + else + vsrc3uc = vec_perm(vsrcCuc, vsrcDuc, vsrcperm1); + + vsrc2ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero, + (vector unsigned char)vsrc2uc); + vsrc3ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero, + (vector unsigned char)vsrc3uc); + + psum = vec_mladd(vA, vsrc0ssH, vec_splat_s16(0)); + psum = vec_mladd(vB, vsrc1ssH, psum); + psum = vec_mladd(vC, vsrc2ssH, psum); + psum = vec_mladd(vD, vsrc3ssH, psum); + psum = vec_add(v28ss, psum); + psum = vec_sr(psum, v6us); + + vdst = vec_ld(0, dst); + ppsum = (vector unsigned char)vec_pack(psum, psum); + fsum = vec_perm(vdst, ppsum, fperm); + + vec_st(fsum, 0, dst); + + vsrc0ssH = vsrc2ssH; + vsrc1ssH = vsrc3ssH; + + dst += stride; + src += stride; + } + } +} + +static inline void put_pixels16_l2_altivec( uint8_t * dst, const uint8_t * src1, + const uint8_t * src2, int dst_stride, + int src_stride1, int h) +{ + int i; + vector unsigned char a, b, d, tmp1, tmp2, mask, mask_, edges, align; + + mask_ = vec_lvsl(0, src2); + + for (i = 0; i < h; i++) { + + tmp1 = vec_ld(i * src_stride1, src1); + mask = vec_lvsl(i * src_stride1, src1); + tmp2 = vec_ld(i * src_stride1 + 15, src1); + + a = vec_perm(tmp1, tmp2, mask); + + tmp1 = vec_ld(i * 16, src2); + tmp2 = vec_ld(i * 16 + 15, src2); + + b = vec_perm(tmp1, tmp2, mask_); + + tmp1 = vec_ld(0, dst); + mask = vec_lvsl(0, dst); + tmp2 = vec_ld(15, dst); + + d = vec_avg(a, b); + + edges = vec_perm(tmp2, tmp1, mask); + + align = vec_lvsr(0, dst); + + tmp2 = vec_perm(d, edges, align); + tmp1 = vec_perm(edges, d, align); + + vec_st(tmp2, 15, dst); + vec_st(tmp1, 0 , dst); + + dst += dst_stride; + } +} + +static inline void avg_pixels16_l2_altivec( uint8_t * dst, const uint8_t * src1, + const uint8_t * src2, int dst_stride, + int src_stride1, int h) +{ + int i; + vector unsigned char a, b, d, tmp1, tmp2, mask, mask_, edges, align; + + mask_ = vec_lvsl(0, src2); + + for (i = 0; i < h; i++) { + + tmp1 = vec_ld(i * src_stride1, src1); + mask = vec_lvsl(i * src_stride1, src1); + tmp2 = vec_ld(i * src_stride1 + 15, src1); + + a = vec_perm(tmp1, tmp2, mask); + + tmp1 = vec_ld(i * 16, src2); + tmp2 = vec_ld(i * 16 + 15, src2); + + b = vec_perm(tmp1, tmp2, mask_); + + tmp1 = vec_ld(0, dst); + mask = vec_lvsl(0, dst); + tmp2 = vec_ld(15, dst); + + d = vec_avg(vec_perm(tmp1, tmp2, mask), vec_avg(a, b)); + + edges = vec_perm(tmp2, tmp1, mask); + + align = vec_lvsr(0, dst); + + tmp2 = vec_perm(d, edges, align); + tmp1 = vec_perm(edges, d, align); + + vec_st(tmp2, 15, dst); + vec_st(tmp1, 0 , dst); + + dst += dst_stride; + } +} + +/* Implemented but could be faster +#define put_pixels16_l2_altivec(d,s1,s2,ds,s1s,h) put_pixels16_l2(d,s1,s2,ds,s1s,16,h) +#define avg_pixels16_l2_altivec(d,s1,s2,ds,s1s,h) avg_pixels16_l2(d,s1,s2,ds,s1s,16,h) + */ + + H264_MC(put_, 16, altivec) + H264_MC(avg_, 16, altivec) + +void dsputil_h264_init_ppc(DSPContext* c, AVCodecContext *avctx) { + +#ifdef HAVE_ALTIVEC + if (has_altivec()) { + c->put_h264_chroma_pixels_tab[0] = put_h264_chroma_mc8_altivec; + c->put_no_rnd_h264_chroma_pixels_tab[0] = put_no_rnd_h264_chroma_mc8_altivec; + c->avg_h264_chroma_pixels_tab[0] = avg_h264_chroma_mc8_altivec; + +#define dspfunc(PFX, IDX, NUM) \ + c->PFX ## _pixels_tab[IDX][ 0] = PFX ## NUM ## _mc00_altivec; \ + c->PFX ## _pixels_tab[IDX][ 1] = PFX ## NUM ## _mc10_altivec; \ + c->PFX ## _pixels_tab[IDX][ 2] = PFX ## NUM ## _mc20_altivec; \ + c->PFX ## _pixels_tab[IDX][ 3] = PFX ## NUM ## _mc30_altivec; \ + c->PFX ## _pixels_tab[IDX][ 4] = PFX ## NUM ## _mc01_altivec; \ + c->PFX ## _pixels_tab[IDX][ 5] = PFX ## NUM ## _mc11_altivec; \ + c->PFX ## _pixels_tab[IDX][ 6] = PFX ## NUM ## _mc21_altivec; \ + c->PFX ## _pixels_tab[IDX][ 7] = PFX ## NUM ## _mc31_altivec; \ + c->PFX ## _pixels_tab[IDX][ 8] = PFX ## NUM ## _mc02_altivec; \ + c->PFX ## _pixels_tab[IDX][ 9] = PFX ## NUM ## _mc12_altivec; \ + c->PFX ## _pixels_tab[IDX][10] = PFX ## NUM ## _mc22_altivec; \ + c->PFX ## _pixels_tab[IDX][11] = PFX ## NUM ## _mc32_altivec; \ + c->PFX ## _pixels_tab[IDX][12] = PFX ## NUM ## _mc03_altivec; \ + c->PFX ## _pixels_tab[IDX][13] = PFX ## NUM ## _mc13_altivec; \ + c->PFX ## _pixels_tab[IDX][14] = PFX ## NUM ## _mc23_altivec; \ + c->PFX ## _pixels_tab[IDX][15] = PFX ## NUM ## _mc33_altivec + + dspfunc(put_h264_qpel, 0, 16); + dspfunc(avg_h264_qpel, 0, 16); +#undef dspfunc + + } else +#endif /* HAVE_ALTIVEC */ + { + // Non-AltiVec PPC optimisations + + // ... pending ... + } +} diff --git a/contrib/ffmpeg/libavcodec/ppc/h264_template_altivec.c b/contrib/ffmpeg/libavcodec/ppc/h264_template_altivec.c new file mode 100644 index 000000000..e8ad67f2f --- /dev/null +++ b/contrib/ffmpeg/libavcodec/ppc/h264_template_altivec.c @@ -0,0 +1,719 @@ +/* + * Copyright (c) 2004 Romain Dolbeau <romain@dolbeau.org> + * + * This file is part of FFmpeg. + * + * FFmpeg 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 2.1 of the License, or (at your option) any later version. + * + * FFmpeg 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 FFmpeg; if not, write to the Free Software + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA + */ + +/* this code assume that stride % 16 == 0 */ +void PREFIX_h264_chroma_mc8_altivec(uint8_t * dst, uint8_t * src, int stride, int h, int x, int y) { + POWERPC_PERF_DECLARE(PREFIX_h264_chroma_mc8_num, 1); + signed int ABCD[4] __attribute__((aligned(16))) = + {((8 - x) * (8 - y)), + ((x) * (8 - y)), + ((8 - x) * (y)), + ((x) * (y))}; + register int i; + vector unsigned char fperm; + const vector signed int vABCD = vec_ld(0, ABCD); + const vector signed short vA = vec_splat((vector signed short)vABCD, 1); + const vector signed short vB = vec_splat((vector signed short)vABCD, 3); + const vector signed short vC = vec_splat((vector signed short)vABCD, 5); + const vector signed short vD = vec_splat((vector signed short)vABCD, 7); + const vector signed int vzero = vec_splat_s32(0); + const vector signed short v32ss = vec_sl(vec_splat_s16(1),vec_splat_u16(5)); + const vector unsigned short v6us = vec_splat_u16(6); + register int loadSecond = (((unsigned long)src) % 16) <= 7 ? 0 : 1; + register int reallyBadAlign = (((unsigned long)src) % 16) == 15 ? 1 : 0; + + vector unsigned char vsrcAuc, vsrcBuc, vsrcperm0, vsrcperm1; + vector unsigned char vsrc0uc, vsrc1uc; + vector signed short vsrc0ssH, vsrc1ssH; + vector unsigned char vsrcCuc, vsrc2uc, vsrc3uc; + vector signed short vsrc2ssH, vsrc3ssH, psum; + vector unsigned char vdst, ppsum, vfdst, fsum; + + POWERPC_PERF_START_COUNT(PREFIX_h264_chroma_mc8_num, 1); + + if (((unsigned long)dst) % 16 == 0) { + fperm = (vector unsigned char)AVV(0x10, 0x11, 0x12, 0x13, + 0x14, 0x15, 0x16, 0x17, + 0x08, 0x09, 0x0A, 0x0B, + 0x0C, 0x0D, 0x0E, 0x0F); + } else { + fperm = (vector unsigned char)AVV(0x00, 0x01, 0x02, 0x03, + 0x04, 0x05, 0x06, 0x07, + 0x18, 0x19, 0x1A, 0x1B, + 0x1C, 0x1D, 0x1E, 0x1F); + } + + vsrcAuc = vec_ld(0, src); + + if (loadSecond) + vsrcBuc = vec_ld(16, src); + vsrcperm0 = vec_lvsl(0, src); + vsrcperm1 = vec_lvsl(1, src); + + vsrc0uc = vec_perm(vsrcAuc, vsrcBuc, vsrcperm0); + if (reallyBadAlign) + vsrc1uc = vsrcBuc; + else + vsrc1uc = vec_perm(vsrcAuc, vsrcBuc, vsrcperm1); + + vsrc0ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero, + (vector unsigned char)vsrc0uc); + vsrc1ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero, + (vector unsigned char)vsrc1uc); + + if (!loadSecond) {// -> !reallyBadAlign + for (i = 0 ; i < h ; i++) { + + + vsrcCuc = vec_ld(stride + 0, src); + + vsrc2uc = vec_perm(vsrcCuc, vsrcCuc, vsrcperm0); + vsrc3uc = vec_perm(vsrcCuc, vsrcCuc, vsrcperm1); + + vsrc2ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero, + (vector unsigned char)vsrc2uc); + vsrc3ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero, + (vector unsigned char)vsrc3uc); + + psum = vec_mladd(vA, vsrc0ssH, vec_splat_s16(0)); + psum = vec_mladd(vB, vsrc1ssH, psum); + psum = vec_mladd(vC, vsrc2ssH, psum); + psum = vec_mladd(vD, vsrc3ssH, psum); + psum = vec_add(v32ss, psum); + psum = vec_sra(psum, v6us); + + vdst = vec_ld(0, dst); + ppsum = (vector unsigned char)vec_packsu(psum, psum); + vfdst = vec_perm(vdst, ppsum, fperm); + + OP_U8_ALTIVEC(fsum, vfdst, vdst); + + vec_st(fsum, 0, dst); + + vsrc0ssH = vsrc2ssH; + vsrc1ssH = vsrc3ssH; + + dst += stride; + src += stride; + } + } else { + vector unsigned char vsrcDuc; + for (i = 0 ; i < h ; i++) { + vsrcCuc = vec_ld(stride + 0, src); + vsrcDuc = vec_ld(stride + 16, src); + + vsrc2uc = vec_perm(vsrcCuc, vsrcDuc, vsrcperm0); + if (reallyBadAlign) + vsrc3uc = vsrcDuc; + else + vsrc3uc = vec_perm(vsrcCuc, vsrcDuc, vsrcperm1); + + vsrc2ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero, + (vector unsigned char)vsrc2uc); + vsrc3ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero, + (vector unsigned char)vsrc3uc); + + psum = vec_mladd(vA, vsrc0ssH, vec_splat_s16(0)); + psum = vec_mladd(vB, vsrc1ssH, psum); + psum = vec_mladd(vC, vsrc2ssH, psum); + psum = vec_mladd(vD, vsrc3ssH, psum); + psum = vec_add(v32ss, psum); + psum = vec_sr(psum, v6us); + + vdst = vec_ld(0, dst); + ppsum = (vector unsigned char)vec_pack(psum, psum); + vfdst = vec_perm(vdst, ppsum, fperm); + + OP_U8_ALTIVEC(fsum, vfdst, vdst); + + vec_st(fsum, 0, dst); + + vsrc0ssH = vsrc2ssH; + vsrc1ssH = vsrc3ssH; + + dst += stride; + src += stride; + } + } + POWERPC_PERF_STOP_COUNT(PREFIX_h264_chroma_mc8_num, 1); +} + +/* this code assume stride % 16 == 0 */ +static void PREFIX_h264_qpel16_h_lowpass_altivec(uint8_t * dst, uint8_t * src, int dstStride, int srcStride) { + POWERPC_PERF_DECLARE(PREFIX_h264_qpel16_h_lowpass_num, 1); + register int i; + + const vector signed int vzero = vec_splat_s32(0); + const vector unsigned char permM2 = vec_lvsl(-2, src); + const vector unsigned char permM1 = vec_lvsl(-1, src); + const vector unsigned char permP0 = vec_lvsl(+0, src); + const vector unsigned char permP1 = vec_lvsl(+1, src); + const vector unsigned char permP2 = vec_lvsl(+2, src); + const vector unsigned char permP3 = vec_lvsl(+3, src); + const vector signed short v5ss = vec_splat_s16(5); + const vector unsigned short v5us = vec_splat_u16(5); + const vector signed short v20ss = vec_sl(vec_splat_s16(5),vec_splat_u16(2)); + const vector signed short v16ss = vec_sl(vec_splat_s16(1),vec_splat_u16(4)); + const vector unsigned char dstperm = vec_lvsr(0, dst); + const vector unsigned char neg1 = + (const vector unsigned char) vec_splat_s8(-1); + + const vector unsigned char dstmask = + vec_perm((const vector unsigned char)vzero, + neg1, dstperm); + + vector unsigned char srcM2, srcM1, srcP0, srcP1, srcP2, srcP3; + + register int align = ((((unsigned long)src) - 2) % 16); + + vector signed short srcP0A, srcP0B, srcP1A, srcP1B, + srcP2A, srcP2B, srcP3A, srcP3B, + srcM1A, srcM1B, srcM2A, srcM2B, + sum1A, sum1B, sum2A, sum2B, sum3A, sum3B, + pp1A, pp1B, pp2A, pp2B, pp3A, pp3B, + psumA, psumB, sumA, sumB; + + vector unsigned char sum, dst1, dst2, vdst, fsum, + rsum, fdst1, fdst2; + + POWERPC_PERF_START_COUNT(PREFIX_h264_qpel16_h_lowpass_num, 1); + + for (i = 0 ; i < 16 ; i ++) { + vector unsigned char srcR1 = vec_ld(-2, src); + vector unsigned char srcR2 = vec_ld(14, src); + + switch (align) { + default: { + srcM2 = vec_perm(srcR1, srcR2, permM2); + srcM1 = vec_perm(srcR1, srcR2, permM1); + srcP0 = vec_perm(srcR1, srcR2, permP0); + srcP1 = vec_perm(srcR1, srcR2, permP1); + srcP2 = vec_perm(srcR1, srcR2, permP2); + srcP3 = vec_perm(srcR1, srcR2, permP3); + } break; + case 11: { + srcM2 = vec_perm(srcR1, srcR2, permM2); + srcM1 = vec_perm(srcR1, srcR2, permM1); + srcP0 = vec_perm(srcR1, srcR2, permP0); + srcP1 = vec_perm(srcR1, srcR2, permP1); + srcP2 = vec_perm(srcR1, srcR2, permP2); + srcP3 = srcR2; + } break; + case 12: { + vector unsigned char srcR3 = vec_ld(30, src); + srcM2 = vec_perm(srcR1, srcR2, permM2); + srcM1 = vec_perm(srcR1, srcR2, permM1); + srcP0 = vec_perm(srcR1, srcR2, permP0); + srcP1 = vec_perm(srcR1, srcR2, permP1); + srcP2 = srcR2; + srcP3 = vec_perm(srcR2, srcR3, permP3); + } break; + case 13: { + vector unsigned char srcR3 = vec_ld(30, src); + srcM2 = vec_perm(srcR1, srcR2, permM2); + srcM1 = vec_perm(srcR1, srcR2, permM1); + srcP0 = vec_perm(srcR1, srcR2, permP0); + srcP1 = srcR2; + srcP2 = vec_perm(srcR2, srcR3, permP2); + srcP3 = vec_perm(srcR2, srcR3, permP3); + } break; + case 14: { + vector unsigned char srcR3 = vec_ld(30, src); + srcM2 = vec_perm(srcR1, srcR2, permM2); + srcM1 = vec_perm(srcR1, srcR2, permM1); + srcP0 = srcR2; + srcP1 = vec_perm(srcR2, srcR3, permP1); + srcP2 = vec_perm(srcR2, srcR3, permP2); + srcP3 = vec_perm(srcR2, srcR3, permP3); + } break; + case 15: { + vector unsigned char srcR3 = vec_ld(30, src); + srcM2 = vec_perm(srcR1, srcR2, permM2); + srcM1 = srcR2; + srcP0 = vec_perm(srcR2, srcR3, permP0); + srcP1 = vec_perm(srcR2, srcR3, permP1); + srcP2 = vec_perm(srcR2, srcR3, permP2); + srcP3 = vec_perm(srcR2, srcR3, permP3); + } break; + } + + srcP0A = (vector signed short) + vec_mergeh((vector unsigned char)vzero, srcP0); + srcP0B = (vector signed short) + vec_mergel((vector unsigned char)vzero, srcP0); + srcP1A = (vector signed short) + vec_mergeh((vector unsigned char)vzero, srcP1); + srcP1B = (vector signed short) + vec_mergel((vector unsigned char)vzero, srcP1); + + srcP2A = (vector signed short) + vec_mergeh((vector unsigned char)vzero, srcP2); + srcP2B = (vector signed short) + vec_mergel((vector unsigned char)vzero, srcP2); + srcP3A = (vector signed short) + vec_mergeh((vector unsigned char)vzero, srcP3); + srcP3B = (vector signed short) + vec_mergel((vector unsigned char)vzero, srcP3); + + srcM1A = (vector signed short) + vec_mergeh((vector unsigned char)vzero, srcM1); + srcM1B = (vector signed short) + vec_mergel((vector unsigned char)vzero, srcM1); + srcM2A = (vector signed short) + vec_mergeh((vector unsigned char)vzero, srcM2); + srcM2B = (vector signed short) + vec_mergel((vector unsigned char)vzero, srcM2); + + sum1A = vec_adds(srcP0A, srcP1A); + sum1B = vec_adds(srcP0B, srcP1B); + sum2A = vec_adds(srcM1A, srcP2A); + sum2B = vec_adds(srcM1B, srcP2B); + sum3A = vec_adds(srcM2A, srcP3A); + sum3B = vec_adds(srcM2B, srcP3B); + + pp1A = vec_mladd(sum1A, v20ss, v16ss); + pp1B = vec_mladd(sum1B, v20ss, v16ss); + + pp2A = vec_mladd(sum2A, v5ss, (vector signed short)vzero); + pp2B = vec_mladd(sum2B, v5ss, (vector signed short)vzero); + + pp3A = vec_add(sum3A, pp1A); + pp3B = vec_add(sum3B, pp1B); + + psumA = vec_sub(pp3A, pp2A); + psumB = vec_sub(pp3B, pp2B); + + sumA = vec_sra(psumA, v5us); + sumB = vec_sra(psumB, v5us); + + sum = vec_packsu(sumA, sumB); + + dst1 = vec_ld(0, dst); + dst2 = vec_ld(16, dst); + vdst = vec_perm(dst1, dst2, vec_lvsl(0, dst)); + + OP_U8_ALTIVEC(fsum, sum, vdst); + + rsum = vec_perm(fsum, fsum, dstperm); + fdst1 = vec_sel(dst1, rsum, dstmask); + fdst2 = vec_sel(rsum, dst2, dstmask); + + vec_st(fdst1, 0, dst); + vec_st(fdst2, 16, dst); + + src += srcStride; + dst += dstStride; + } +POWERPC_PERF_STOP_COUNT(PREFIX_h264_qpel16_h_lowpass_num, 1); +} + +/* this code assume stride % 16 == 0 */ +static void PREFIX_h264_qpel16_v_lowpass_altivec(uint8_t * dst, uint8_t * src, int dstStride, int srcStride) { + POWERPC_PERF_DECLARE(PREFIX_h264_qpel16_v_lowpass_num, 1); + + register int i; + + const vector signed int vzero = vec_splat_s32(0); + const vector unsigned char perm = vec_lvsl(0, src); + const vector signed short v20ss = vec_sl(vec_splat_s16(5),vec_splat_u16(2)); + const vector unsigned short v5us = vec_splat_u16(5); + const vector signed short v5ss = vec_splat_s16(5); + const vector signed short v16ss = vec_sl(vec_splat_s16(1),vec_splat_u16(4)); + const vector unsigned char dstperm = vec_lvsr(0, dst); + const vector unsigned char neg1 = (const vector unsigned char)vec_splat_s8(-1); + const vector unsigned char dstmask = vec_perm((const vector unsigned char)vzero, neg1, dstperm); + + uint8_t *srcbis = src - (srcStride * 2); + + const vector unsigned char srcM2a = vec_ld(0, srcbis); + const vector unsigned char srcM2b = vec_ld(16, srcbis); + const vector unsigned char srcM2 = vec_perm(srcM2a, srcM2b, perm); +// srcbis += srcStride; + const vector unsigned char srcM1a = vec_ld(0, srcbis += srcStride); + const vector unsigned char srcM1b = vec_ld(16, srcbis); + const vector unsigned char srcM1 = vec_perm(srcM1a, srcM1b, perm); +// srcbis += srcStride; + const vector unsigned char srcP0a = vec_ld(0, srcbis += srcStride); + const vector unsigned char srcP0b = vec_ld(16, srcbis); + const vector unsigned char srcP0 = vec_perm(srcP0a, srcP0b, perm); +// srcbis += srcStride; + const vector unsigned char srcP1a = vec_ld(0, srcbis += srcStride); + const vector unsigned char srcP1b = vec_ld(16, srcbis); + const vector unsigned char srcP1 = vec_perm(srcP1a, srcP1b, perm); +// srcbis += srcStride; + const vector unsigned char srcP2a = vec_ld(0, srcbis += srcStride); + const vector unsigned char srcP2b = vec_ld(16, srcbis); + const vector unsigned char srcP2 = vec_perm(srcP2a, srcP2b, perm); +// srcbis += srcStride; + + vector signed short srcM2ssA = (vector signed short) + vec_mergeh((vector unsigned char)vzero, srcM2); + vector signed short srcM2ssB = (vector signed short) + vec_mergel((vector unsigned char)vzero, srcM2); + vector signed short srcM1ssA = (vector signed short) + vec_mergeh((vector unsigned char)vzero, srcM1); + vector signed short srcM1ssB = (vector signed short) + vec_mergel((vector unsigned char)vzero, srcM1); + vector signed short srcP0ssA = (vector signed short) + vec_mergeh((vector unsigned char)vzero, srcP0); + vector signed short srcP0ssB = (vector signed short) + vec_mergel((vector unsigned char)vzero, srcP0); + vector signed short srcP1ssA = (vector signed short) + vec_mergeh((vector unsigned char)vzero, srcP1); + vector signed short srcP1ssB = (vector signed short) + vec_mergel((vector unsigned char)vzero, srcP1); + vector signed short srcP2ssA = (vector signed short) + vec_mergeh((vector unsigned char)vzero, srcP2); + vector signed short srcP2ssB = (vector signed short) + vec_mergel((vector unsigned char)vzero, srcP2); + + vector signed short pp1A, pp1B, pp2A, pp2B, pp3A, pp3B, + psumA, psumB, sumA, sumB, + srcP3ssA, srcP3ssB, + sum1A, sum1B, sum2A, sum2B, sum3A, sum3B; + + vector unsigned char sum, dst1, dst2, vdst, fsum, rsum, fdst1, fdst2, + srcP3a, srcP3b, srcP3; + + POWERPC_PERF_START_COUNT(PREFIX_h264_qpel16_v_lowpass_num, 1); + + for (i = 0 ; i < 16 ; i++) { + srcP3a = vec_ld(0, srcbis += srcStride); + srcP3b = vec_ld(16, srcbis); + srcP3 = vec_perm(srcP3a, srcP3b, perm); + srcP3ssA = (vector signed short) + vec_mergeh((vector unsigned char)vzero, srcP3); + srcP3ssB = (vector signed short) + vec_mergel((vector unsigned char)vzero, srcP3); +// srcbis += srcStride; + + sum1A = vec_adds(srcP0ssA, srcP1ssA); + sum1B = vec_adds(srcP0ssB, srcP1ssB); + sum2A = vec_adds(srcM1ssA, srcP2ssA); + sum2B = vec_adds(srcM1ssB, srcP2ssB); + sum3A = vec_adds(srcM2ssA, srcP3ssA); + sum3B = vec_adds(srcM2ssB, srcP3ssB); + + srcM2ssA = srcM1ssA; + srcM2ssB = srcM1ssB; + srcM1ssA = srcP0ssA; + srcM1ssB = srcP0ssB; + srcP0ssA = srcP1ssA; + srcP0ssB = srcP1ssB; + srcP1ssA = srcP2ssA; + srcP1ssB = srcP2ssB; + srcP2ssA = srcP3ssA; + srcP2ssB = srcP3ssB; + + pp1A = vec_mladd(sum1A, v20ss, v16ss); + pp1B = vec_mladd(sum1B, v20ss, v16ss); + + pp2A = vec_mladd(sum2A, v5ss, (vector signed short)vzero); + pp2B = vec_mladd(sum2B, v5ss, (vector signed short)vzero); + + pp3A = vec_add(sum3A, pp1A); + pp3B = vec_add(sum3B, pp1B); + + psumA = vec_sub(pp3A, pp2A); + psumB = vec_sub(pp3B, pp2B); + + sumA = vec_sra(psumA, v5us); + sumB = vec_sra(psumB, v5us); + + sum = vec_packsu(sumA, sumB); + + dst1 = vec_ld(0, dst); + dst2 = vec_ld(16, dst); + vdst = vec_perm(dst1, dst2, vec_lvsl(0, dst)); + + OP_U8_ALTIVEC(fsum, sum, vdst); + + rsum = vec_perm(fsum, fsum, dstperm); + fdst1 = vec_sel(dst1, rsum, dstmask); + fdst2 = vec_sel(rsum, dst2, dstmask); + + vec_st(fdst1, 0, dst); + vec_st(fdst2, 16, dst); + + dst += dstStride; + } + POWERPC_PERF_STOP_COUNT(PREFIX_h264_qpel16_v_lowpass_num, 1); +} + +/* this code assume stride % 16 == 0 *and* tmp is properly aligned */ +static void PREFIX_h264_qpel16_hv_lowpass_altivec(uint8_t * dst, int16_t * tmp, uint8_t * src, int dstStride, int tmpStride, int srcStride) { + POWERPC_PERF_DECLARE(PREFIX_h264_qpel16_hv_lowpass_num, 1); + register int i; + const vector signed int vzero = vec_splat_s32(0); + const vector unsigned char permM2 = vec_lvsl(-2, src); + const vector unsigned char permM1 = vec_lvsl(-1, src); + const vector unsigned char permP0 = vec_lvsl(+0, src); + const vector unsigned char permP1 = vec_lvsl(+1, src); + const vector unsigned char permP2 = vec_lvsl(+2, src); + const vector unsigned char permP3 = vec_lvsl(+3, src); + const vector signed short v20ss = vec_sl(vec_splat_s16(5),vec_splat_u16(2)); + const vector unsigned int v10ui = vec_splat_u32(10); + const vector signed short v5ss = vec_splat_s16(5); + const vector signed short v1ss = vec_splat_s16(1); + const vector signed int v512si = vec_sl(vec_splat_s32(1),vec_splat_u32(9)); + const vector unsigned int v16ui = vec_sl(vec_splat_u32(1),vec_splat_u32(4)); + + register int align = ((((unsigned long)src) - 2) % 16); + + const vector unsigned char neg1 = (const vector unsigned char) + vec_splat_s8(-1); + + vector signed short srcP0A, srcP0B, srcP1A, srcP1B, + srcP2A, srcP2B, srcP3A, srcP3B, + srcM1A, srcM1B, srcM2A, srcM2B, + sum1A, sum1B, sum2A, sum2B, sum3A, sum3B, + pp1A, pp1B, pp2A, pp2B, psumA, psumB; + + const vector unsigned char dstperm = vec_lvsr(0, dst); + + const vector unsigned char dstmask = vec_perm((const vector unsigned char)vzero, neg1, dstperm); + + const vector unsigned char mperm = (const vector unsigned char) + AVV(0x00, 0x08, 0x01, 0x09, 0x02, 0x0A, 0x03, 0x0B, + 0x04, 0x0C, 0x05, 0x0D, 0x06, 0x0E, 0x07, 0x0F); + int16_t *tmpbis = tmp; + + vector signed short tmpM1ssA, tmpM1ssB, tmpM2ssA, tmpM2ssB, + tmpP0ssA, tmpP0ssB, tmpP1ssA, tmpP1ssB, + tmpP2ssA, tmpP2ssB; + + vector signed int pp1Ae, pp1Ao, pp1Be, pp1Bo, pp2Ae, pp2Ao, pp2Be, pp2Bo, + pp3Ae, pp3Ao, pp3Be, pp3Bo, pp1cAe, pp1cAo, pp1cBe, pp1cBo, + pp32Ae, pp32Ao, pp32Be, pp32Bo, sumAe, sumAo, sumBe, sumBo, + ssumAe, ssumAo, ssumBe, ssumBo; + vector unsigned char fsum, sumv, sum, dst1, dst2, vdst, + rsum, fdst1, fdst2; + vector signed short ssume, ssumo; + + POWERPC_PERF_START_COUNT(PREFIX_h264_qpel16_hv_lowpass_num, 1); + src -= (2 * srcStride); + for (i = 0 ; i < 21 ; i ++) { + vector unsigned char srcM2, srcM1, srcP0, srcP1, srcP2, srcP3; + vector unsigned char srcR1 = vec_ld(-2, src); + vector unsigned char srcR2 = vec_ld(14, src); + + switch (align) { + default: { + srcM2 = vec_perm(srcR1, srcR2, permM2); + srcM1 = vec_perm(srcR1, srcR2, permM1); + srcP0 = vec_perm(srcR1, srcR2, permP0); + srcP1 = vec_perm(srcR1, srcR2, permP1); + srcP2 = vec_perm(srcR1, srcR2, permP2); + srcP3 = vec_perm(srcR1, srcR2, permP3); + } break; + case 11: { + srcM2 = vec_perm(srcR1, srcR2, permM2); + srcM1 = vec_perm(srcR1, srcR2, permM1); + srcP0 = vec_perm(srcR1, srcR2, permP0); + srcP1 = vec_perm(srcR1, srcR2, permP1); + srcP2 = vec_perm(srcR1, srcR2, permP2); + srcP3 = srcR2; + } break; + case 12: { + vector unsigned char srcR3 = vec_ld(30, src); + srcM2 = vec_perm(srcR1, srcR2, permM2); + srcM1 = vec_perm(srcR1, srcR2, permM1); + srcP0 = vec_perm(srcR1, srcR2, permP0); + srcP1 = vec_perm(srcR1, srcR2, permP1); + srcP2 = srcR2; + srcP3 = vec_perm(srcR2, srcR3, permP3); + } break; + case 13: { + vector unsigned char srcR3 = vec_ld(30, src); + srcM2 = vec_perm(srcR1, srcR2, permM2); + srcM1 = vec_perm(srcR1, srcR2, permM1); + srcP0 = vec_perm(srcR1, srcR2, permP0); + srcP1 = srcR2; + srcP2 = vec_perm(srcR2, srcR3, permP2); + srcP3 = vec_perm(srcR2, srcR3, permP3); + } break; + case 14: { + vector unsigned char srcR3 = vec_ld(30, src); + srcM2 = vec_perm(srcR1, srcR2, permM2); + srcM1 = vec_perm(srcR1, srcR2, permM1); + srcP0 = srcR2; + srcP1 = vec_perm(srcR2, srcR3, permP1); + srcP2 = vec_perm(srcR2, srcR3, permP2); + srcP3 = vec_perm(srcR2, srcR3, permP3); + } break; + case 15: { + vector unsigned char srcR3 = vec_ld(30, src); + srcM2 = vec_perm(srcR1, srcR2, permM2); + srcM1 = srcR2; + srcP0 = vec_perm(srcR2, srcR3, permP0); + srcP1 = vec_perm(srcR2, srcR3, permP1); + srcP2 = vec_perm(srcR2, srcR3, permP2); + srcP3 = vec_perm(srcR2, srcR3, permP3); + } break; + } + + srcP0A = (vector signed short) + vec_mergeh((vector unsigned char)vzero, srcP0); + srcP0B = (vector signed short) + vec_mergel((vector unsigned char)vzero, srcP0); + srcP1A = (vector signed short) + vec_mergeh((vector unsigned char)vzero, srcP1); + srcP1B = (vector signed short) + vec_mergel((vector unsigned char)vzero, srcP1); + + srcP2A = (vector signed short) + vec_mergeh((vector unsigned char)vzero, srcP2); + srcP2B = (vector signed short) + vec_mergel((vector unsigned char)vzero, srcP2); + srcP3A = (vector signed short) + vec_mergeh((vector unsigned char)vzero, srcP3); + srcP3B = (vector signed short) + vec_mergel((vector unsigned char)vzero, srcP3); + + srcM1A = (vector signed short) + vec_mergeh((vector unsigned char)vzero, srcM1); + srcM1B = (vector signed short) + vec_mergel((vector unsigned char)vzero, srcM1); + srcM2A = (vector signed short) + vec_mergeh((vector unsigned char)vzero, srcM2); + srcM2B = (vector signed short) + vec_mergel((vector unsigned char)vzero, srcM2); + + sum1A = vec_adds(srcP0A, srcP1A); + sum1B = vec_adds(srcP0B, srcP1B); + sum2A = vec_adds(srcM1A, srcP2A); + sum2B = vec_adds(srcM1B, srcP2B); + sum3A = vec_adds(srcM2A, srcP3A); + sum3B = vec_adds(srcM2B, srcP3B); + + pp1A = vec_mladd(sum1A, v20ss, sum3A); + pp1B = vec_mladd(sum1B, v20ss, sum3B); + + pp2A = vec_mladd(sum2A, v5ss, (vector signed short)vzero); + pp2B = vec_mladd(sum2B, v5ss, (vector signed short)vzero); + + psumA = vec_sub(pp1A, pp2A); + psumB = vec_sub(pp1B, pp2B); + + vec_st(psumA, 0, tmp); + vec_st(psumB, 16, tmp); + + src += srcStride; + tmp += tmpStride; /* int16_t*, and stride is 16, so it's OK here */ + } + + tmpM2ssA = vec_ld(0, tmpbis); + tmpM2ssB = vec_ld(16, tmpbis); + tmpbis += tmpStride; + tmpM1ssA = vec_ld(0, tmpbis); + tmpM1ssB = vec_ld(16, tmpbis); + tmpbis += tmpStride; + tmpP0ssA = vec_ld(0, tmpbis); + tmpP0ssB = vec_ld(16, tmpbis); + tmpbis += tmpStride; + tmpP1ssA = vec_ld(0, tmpbis); + tmpP1ssB = vec_ld(16, tmpbis); + tmpbis += tmpStride; + tmpP2ssA = vec_ld(0, tmpbis); + tmpP2ssB = vec_ld(16, tmpbis); + tmpbis += tmpStride; + + for (i = 0 ; i < 16 ; i++) { + const vector signed short tmpP3ssA = vec_ld(0, tmpbis); + const vector signed short tmpP3ssB = vec_ld(16, tmpbis); + + const vector signed short sum1A = vec_adds(tmpP0ssA, tmpP1ssA); + const vector signed short sum1B = vec_adds(tmpP0ssB, tmpP1ssB); + const vector signed short sum2A = vec_adds(tmpM1ssA, tmpP2ssA); + const vector signed short sum2B = vec_adds(tmpM1ssB, tmpP2ssB); + const vector signed short sum3A = vec_adds(tmpM2ssA, tmpP3ssA); + const vector signed short sum3B = vec_adds(tmpM2ssB, tmpP3ssB); + + tmpbis += tmpStride; + + tmpM2ssA = tmpM1ssA; + tmpM2ssB = tmpM1ssB; + tmpM1ssA = tmpP0ssA; + tmpM1ssB = tmpP0ssB; + tmpP0ssA = tmpP1ssA; + tmpP0ssB = tmpP1ssB; + tmpP1ssA = tmpP2ssA; + tmpP1ssB = tmpP2ssB; + tmpP2ssA = tmpP3ssA; + tmpP2ssB = tmpP3ssB; + + pp1Ae = vec_mule(sum1A, v20ss); + pp1Ao = vec_mulo(sum1A, v20ss); + pp1Be = vec_mule(sum1B, v20ss); + pp1Bo = vec_mulo(sum1B, v20ss); + + pp2Ae = vec_mule(sum2A, v5ss); + pp2Ao = vec_mulo(sum2A, v5ss); + pp2Be = vec_mule(sum2B, v5ss); + pp2Bo = vec_mulo(sum2B, v5ss); + + pp3Ae = vec_sra((vector signed int)sum3A, v16ui); + pp3Ao = vec_mulo(sum3A, v1ss); + pp3Be = vec_sra((vector signed int)sum3B, v16ui); + pp3Bo = vec_mulo(sum3B, v1ss); + + pp1cAe = vec_add(pp1Ae, v512si); + pp1cAo = vec_add(pp1Ao, v512si); + pp1cBe = vec_add(pp1Be, v512si); + pp1cBo = vec_add(pp1Bo, v512si); + + pp32Ae = vec_sub(pp3Ae, pp2Ae); + pp32Ao = vec_sub(pp3Ao, pp2Ao); + pp32Be = vec_sub(pp3Be, pp2Be); + pp32Bo = vec_sub(pp3Bo, pp2Bo); + + sumAe = vec_add(pp1cAe, pp32Ae); + sumAo = vec_add(pp1cAo, pp32Ao); + sumBe = vec_add(pp1cBe, pp32Be); + sumBo = vec_add(pp1cBo, pp32Bo); + + ssumAe = vec_sra(sumAe, v10ui); + ssumAo = vec_sra(sumAo, v10ui); + ssumBe = vec_sra(sumBe, v10ui); + ssumBo = vec_sra(sumBo, v10ui); + + ssume = vec_packs(ssumAe, ssumBe); + ssumo = vec_packs(ssumAo, ssumBo); + + sumv = vec_packsu(ssume, ssumo); + sum = vec_perm(sumv, sumv, mperm); + + dst1 = vec_ld(0, dst); + dst2 = vec_ld(16, dst); + vdst = vec_perm(dst1, dst2, vec_lvsl(0, dst)); + + OP_U8_ALTIVEC(fsum, sum, vdst); + + rsum = vec_perm(fsum, fsum, dstperm); + fdst1 = vec_sel(dst1, rsum, dstmask); + fdst2 = vec_sel(rsum, dst2, dstmask); + + vec_st(fdst1, 0, dst); + vec_st(fdst2, 16, dst); + + dst += dstStride; + } + POWERPC_PERF_STOP_COUNT(PREFIX_h264_qpel16_hv_lowpass_num, 1); +} diff --git a/contrib/ffmpeg/libavcodec/ppc/idct_altivec.c b/contrib/ffmpeg/libavcodec/ppc/idct_altivec.c new file mode 100644 index 000000000..cee46fc25 --- /dev/null +++ b/contrib/ffmpeg/libavcodec/ppc/idct_altivec.c @@ -0,0 +1,237 @@ +/* + * Copyright (c) 2001 Michel Lespinasse + * + * This file is part of FFmpeg. + * + * FFmpeg 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 2.1 of the License, or (at your option) any later version. + * + * FFmpeg 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 FFmpeg; if not, write to the Free Software + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA + * + */ + +/* + * NOTE: This code is based on GPL code from the libmpeg2 project. The + * author, Michel Lespinasses, has given explicit permission to release + * under LGPL as part of ffmpeg. + * + */ + +/* + * FFMpeg integration by Dieter Shirley + * + * This file is a direct copy of the altivec idct module from the libmpeg2 + * project. I've deleted all of the libmpeg2 specific code, renamed the functions and + * re-ordered the function parameters. The only change to the IDCT function + * itself was to factor out the partial transposition, and to perform a full + * transpose at the end of the function. + */ + + +#include <stdlib.h> /* malloc(), free() */ +#include <string.h> +#include "../dsputil.h" + +#include "gcc_fixes.h" + +#include "dsputil_altivec.h" + +#define vector_s16_t vector signed short +#define const_vector_s16_t const_vector signed short +#define vector_u16_t vector unsigned short +#define vector_s8_t vector signed char +#define vector_u8_t vector unsigned char +#define vector_s32_t vector signed int +#define vector_u32_t vector unsigned int + +#define IDCT_HALF \ + /* 1st stage */ \ + t1 = vec_mradds (a1, vx7, vx1 ); \ + t8 = vec_mradds (a1, vx1, vec_subs (zero, vx7)); \ + t7 = vec_mradds (a2, vx5, vx3); \ + t3 = vec_mradds (ma2, vx3, vx5); \ + \ + /* 2nd stage */ \ + t5 = vec_adds (vx0, vx4); \ + t0 = vec_subs (vx0, vx4); \ + t2 = vec_mradds (a0, vx6, vx2); \ + t4 = vec_mradds (a0, vx2, vec_subs (zero, vx6)); \ + t6 = vec_adds (t8, t3); \ + t3 = vec_subs (t8, t3); \ + t8 = vec_subs (t1, t7); \ + t1 = vec_adds (t1, t7); \ + \ + /* 3rd stage */ \ + t7 = vec_adds (t5, t2); \ + t2 = vec_subs (t5, t2); \ + t5 = vec_adds (t0, t4); \ + t0 = vec_subs (t0, t4); \ + t4 = vec_subs (t8, t3); \ + t3 = vec_adds (t8, t3); \ + \ + /* 4th stage */ \ + vy0 = vec_adds (t7, t1); \ + vy7 = vec_subs (t7, t1); \ + vy1 = vec_mradds (c4, t3, t5); \ + vy6 = vec_mradds (mc4, t3, t5); \ + vy2 = vec_mradds (c4, t4, t0); \ + vy5 = vec_mradds (mc4, t4, t0); \ + vy3 = vec_adds (t2, t6); \ + vy4 = vec_subs (t2, t6); + + +#define IDCT \ + vector_s16_t vx0, vx1, vx2, vx3, vx4, vx5, vx6, vx7; \ + vector_s16_t vy0, vy1, vy2, vy3, vy4, vy5, vy6, vy7; \ + vector_s16_t a0, a1, a2, ma2, c4, mc4, zero, bias; \ + vector_s16_t t0, t1, t2, t3, t4, t5, t6, t7, t8; \ + vector_u16_t shift; \ + \ + c4 = vec_splat (constants[0], 0); \ + a0 = vec_splat (constants[0], 1); \ + a1 = vec_splat (constants[0], 2); \ + a2 = vec_splat (constants[0], 3); \ + mc4 = vec_splat (constants[0], 4); \ + ma2 = vec_splat (constants[0], 5); \ + bias = (vector_s16_t)vec_splat ((vector_s32_t)constants[0], 3); \ + \ + zero = vec_splat_s16 (0); \ + shift = vec_splat_u16 (4); \ + \ + vx0 = vec_mradds (vec_sl (block[0], shift), constants[1], zero); \ + vx1 = vec_mradds (vec_sl (block[1], shift), constants[2], zero); \ + vx2 = vec_mradds (vec_sl (block[2], shift), constants[3], zero); \ + vx3 = vec_mradds (vec_sl (block[3], shift), constants[4], zero); \ + vx4 = vec_mradds (vec_sl (block[4], shift), constants[1], zero); \ + vx5 = vec_mradds (vec_sl (block[5], shift), constants[4], zero); \ + vx6 = vec_mradds (vec_sl (block[6], shift), constants[3], zero); \ + vx7 = vec_mradds (vec_sl (block[7], shift), constants[2], zero); \ + \ + IDCT_HALF \ + \ + vx0 = vec_mergeh (vy0, vy4); \ + vx1 = vec_mergel (vy0, vy4); \ + vx2 = vec_mergeh (vy1, vy5); \ + vx3 = vec_mergel (vy1, vy5); \ + vx4 = vec_mergeh (vy2, vy6); \ + vx5 = vec_mergel (vy2, vy6); \ + vx6 = vec_mergeh (vy3, vy7); \ + vx7 = vec_mergel (vy3, vy7); \ + \ + vy0 = vec_mergeh (vx0, vx4); \ + vy1 = vec_mergel (vx0, vx4); \ + vy2 = vec_mergeh (vx1, vx5); \ + vy3 = vec_mergel (vx1, vx5); \ + vy4 = vec_mergeh (vx2, vx6); \ + vy5 = vec_mergel (vx2, vx6); \ + vy6 = vec_mergeh (vx3, vx7); \ + vy7 = vec_mergel (vx3, vx7); \ + \ + vx0 = vec_adds (vec_mergeh (vy0, vy4), bias); \ + vx1 = vec_mergel (vy0, vy4); \ + vx2 = vec_mergeh (vy1, vy5); \ + vx3 = vec_mergel (vy1, vy5); \ + vx4 = vec_mergeh (vy2, vy6); \ + vx5 = vec_mergel (vy2, vy6); \ + vx6 = vec_mergeh (vy3, vy7); \ + vx7 = vec_mergel (vy3, vy7); \ + \ + IDCT_HALF \ + \ + shift = vec_splat_u16 (6); \ + vx0 = vec_sra (vy0, shift); \ + vx1 = vec_sra (vy1, shift); \ + vx2 = vec_sra (vy2, shift); \ + vx3 = vec_sra (vy3, shift); \ + vx4 = vec_sra (vy4, shift); \ + vx5 = vec_sra (vy5, shift); \ + vx6 = vec_sra (vy6, shift); \ + vx7 = vec_sra (vy7, shift); + + +static const_vector_s16_t constants[5] = { + (vector_s16_t) AVV(23170, 13573, 6518, 21895, -23170, -21895, 32, 31), + (vector_s16_t) AVV(16384, 22725, 21407, 19266, 16384, 19266, 21407, 22725), + (vector_s16_t) AVV(22725, 31521, 29692, 26722, 22725, 26722, 29692, 31521), + (vector_s16_t) AVV(21407, 29692, 27969, 25172, 21407, 25172, 27969, 29692), + (vector_s16_t) AVV(19266, 26722, 25172, 22654, 19266, 22654, 25172, 26722) +}; + +void idct_put_altivec(uint8_t* dest, int stride, vector_s16_t* block) +{ +POWERPC_PERF_DECLARE(altivec_idct_put_num, 1); + vector_u8_t tmp; + +#ifdef POWERPC_PERFORMANCE_REPORT +POWERPC_PERF_START_COUNT(altivec_idct_put_num, 1); +#endif + IDCT + +#define COPY(dest,src) \ + tmp = vec_packsu (src, src); \ + vec_ste ((vector_u32_t)tmp, 0, (unsigned int *)dest); \ + vec_ste ((vector_u32_t)tmp, 4, (unsigned int *)dest); + + COPY (dest, vx0) dest += stride; + COPY (dest, vx1) dest += stride; + COPY (dest, vx2) dest += stride; + COPY (dest, vx3) dest += stride; + COPY (dest, vx4) dest += stride; + COPY (dest, vx5) dest += stride; + COPY (dest, vx6) dest += stride; + COPY (dest, vx7) + +POWERPC_PERF_STOP_COUNT(altivec_idct_put_num, 1); +} + +void idct_add_altivec(uint8_t* dest, int stride, vector_s16_t* block) +{ +POWERPC_PERF_DECLARE(altivec_idct_add_num, 1); + vector_u8_t tmp; + vector_s16_t tmp2, tmp3; + vector_u8_t perm0; + vector_u8_t perm1; + vector_u8_t p0, p1, p; + +#ifdef POWERPC_PERFORMANCE_REPORT +POWERPC_PERF_START_COUNT(altivec_idct_add_num, 1); +#endif + + IDCT + + p0 = vec_lvsl (0, dest); + p1 = vec_lvsl (stride, dest); + p = vec_splat_u8 (-1); + perm0 = vec_mergeh (p, p0); + perm1 = vec_mergeh (p, p1); + +#define ADD(dest,src,perm) \ + /* *(uint64_t *)&tmp = *(uint64_t *)dest; */ \ + tmp = vec_ld (0, dest); \ + tmp2 = (vector_s16_t)vec_perm (tmp, (vector_u8_t)zero, perm); \ + tmp3 = vec_adds (tmp2, src); \ + tmp = vec_packsu (tmp3, tmp3); \ + vec_ste ((vector_u32_t)tmp, 0, (unsigned int *)dest); \ + vec_ste ((vector_u32_t)tmp, 4, (unsigned int *)dest); + + ADD (dest, vx0, perm0) dest += stride; + ADD (dest, vx1, perm1) dest += stride; + ADD (dest, vx2, perm0) dest += stride; + ADD (dest, vx3, perm1) dest += stride; + ADD (dest, vx4, perm0) dest += stride; + ADD (dest, vx5, perm1) dest += stride; + ADD (dest, vx6, perm0) dest += stride; + ADD (dest, vx7, perm1) + +POWERPC_PERF_STOP_COUNT(altivec_idct_add_num, 1); +} + diff --git a/contrib/ffmpeg/libavcodec/ppc/mathops.h b/contrib/ffmpeg/libavcodec/ppc/mathops.h new file mode 100644 index 000000000..6af23f246 --- /dev/null +++ b/contrib/ffmpeg/libavcodec/ppc/mathops.h @@ -0,0 +1,33 @@ +/* + * simple math operations + * Copyright (c) 2001, 2002 Fabrice Bellard. + * Copyright (c) 2006 Michael Niedermayer <michaelni@gmx.at> et al + * + * This file is part of FFmpeg. + * + * FFmpeg 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 2.1 of the License, or (at your option) any later version. + * + * FFmpeg 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 FFmpeg; if not, write to the Free Software + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA + */ + +#if defined(ARCH_POWERPC_405) +/* signed 16x16 -> 32 multiply add accumulate */ +# define MAC16(rt, ra, rb) \ + asm ("maclhw %0, %2, %3" : "=r" (rt) : "0" (rt), "r" (ra), "r" (rb)); + +/* signed 16x16 -> 32 multiply */ +# define MUL16(ra, rb) \ + ({ int __rt; + asm ("mullhw %0, %1, %2" : "=r" (__rt) : "r" (ra), "r" (rb)); + __rt; }) +#endif diff --git a/contrib/ffmpeg/libavcodec/ppc/mpegvideo_altivec.c b/contrib/ffmpeg/libavcodec/ppc/mpegvideo_altivec.c new file mode 100644 index 000000000..3822cb20e --- /dev/null +++ b/contrib/ffmpeg/libavcodec/ppc/mpegvideo_altivec.c @@ -0,0 +1,603 @@ +/* + * Copyright (c) 2002 Dieter Shirley + * + * dct_unquantize_h263_altivec: + * Copyright (c) 2003 Romain Dolbeau <romain@dolbeau.org> + * + * This file is part of FFmpeg. + * + * FFmpeg 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 2.1 of the License, or (at your option) any later version. + * + * FFmpeg 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 FFmpeg; if not, write to the Free Software + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA + */ + +#include <stdlib.h> +#include <stdio.h> +#include "../dsputil.h" +#include "../mpegvideo.h" + +#include "gcc_fixes.h" + +#include "dsputil_altivec.h" + +// Swaps two variables (used for altivec registers) +#define SWAP(a,b) \ +do { \ + __typeof__(a) swap_temp=a; \ + a=b; \ + b=swap_temp; \ +} while (0) + +// transposes a matrix consisting of four vectors with four elements each +#define TRANSPOSE4(a,b,c,d) \ +do { \ + __typeof__(a) _trans_ach = vec_mergeh(a, c); \ + __typeof__(a) _trans_acl = vec_mergel(a, c); \ + __typeof__(a) _trans_bdh = vec_mergeh(b, d); \ + __typeof__(a) _trans_bdl = vec_mergel(b, d); \ + \ + a = vec_mergeh(_trans_ach, _trans_bdh); \ + b = vec_mergel(_trans_ach, _trans_bdh); \ + c = vec_mergeh(_trans_acl, _trans_bdl); \ + d = vec_mergel(_trans_acl, _trans_bdl); \ +} while (0) + + +// Loads a four-byte value (int or float) from the target address +// into every element in the target vector. Only works if the +// target address is four-byte aligned (which should be always). +#define LOAD4(vec, address) \ +{ \ + __typeof__(vec)* _load_addr = (__typeof__(vec)*)(address); \ + vector unsigned char _perm_vec = vec_lvsl(0,(address)); \ + vec = vec_ld(0, _load_addr); \ + vec = vec_perm(vec, vec, _perm_vec); \ + vec = vec_splat(vec, 0); \ +} + + +#ifdef CONFIG_DARWIN +#define FOUROF(a) (a) +#else +// slower, for dumb non-apple GCC +#define FOUROF(a) {a,a,a,a} +#endif +int dct_quantize_altivec(MpegEncContext* s, + DCTELEM* data, int n, + int qscale, int* overflow) +{ + int lastNonZero; + vector float row0, row1, row2, row3, row4, row5, row6, row7; + vector float alt0, alt1, alt2, alt3, alt4, alt5, alt6, alt7; + const_vector float zero = (const_vector float)FOUROF(0.); + // used after quantise step + int oldBaseValue = 0; + + // Load the data into the row/alt vectors + { + vector signed short data0, data1, data2, data3, data4, data5, data6, data7; + + data0 = vec_ld(0, data); + data1 = vec_ld(16, data); + data2 = vec_ld(32, data); + data3 = vec_ld(48, data); + data4 = vec_ld(64, data); + data5 = vec_ld(80, data); + data6 = vec_ld(96, data); + data7 = vec_ld(112, data); + + // Transpose the data before we start + TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7); + + // load the data into floating point vectors. We load + // the high half of each row into the main row vectors + // and the low half into the alt vectors. + row0 = vec_ctf(vec_unpackh(data0), 0); + alt0 = vec_ctf(vec_unpackl(data0), 0); + row1 = vec_ctf(vec_unpackh(data1), 0); + alt1 = vec_ctf(vec_unpackl(data1), 0); + row2 = vec_ctf(vec_unpackh(data2), 0); + alt2 = vec_ctf(vec_unpackl(data2), 0); + row3 = vec_ctf(vec_unpackh(data3), 0); + alt3 = vec_ctf(vec_unpackl(data3), 0); + row4 = vec_ctf(vec_unpackh(data4), 0); + alt4 = vec_ctf(vec_unpackl(data4), 0); + row5 = vec_ctf(vec_unpackh(data5), 0); + alt5 = vec_ctf(vec_unpackl(data5), 0); + row6 = vec_ctf(vec_unpackh(data6), 0); + alt6 = vec_ctf(vec_unpackl(data6), 0); + row7 = vec_ctf(vec_unpackh(data7), 0); + alt7 = vec_ctf(vec_unpackl(data7), 0); + } + + // The following block could exist as a separate an altivec dct + // function. However, if we put it inline, the DCT data can remain + // in the vector local variables, as floats, which we'll use during the + // quantize step... + { + const vector float vec_0_298631336 = (vector float)FOUROF(0.298631336f); + const vector float vec_0_390180644 = (vector float)FOUROF(-0.390180644f); + const vector float vec_0_541196100 = (vector float)FOUROF(0.541196100f); + const vector float vec_0_765366865 = (vector float)FOUROF(0.765366865f); + const vector float vec_0_899976223 = (vector float)FOUROF(-0.899976223f); + const vector float vec_1_175875602 = (vector float)FOUROF(1.175875602f); + const vector float vec_1_501321110 = (vector float)FOUROF(1.501321110f); + const vector float vec_1_847759065 = (vector float)FOUROF(-1.847759065f); + const vector float vec_1_961570560 = (vector float)FOUROF(-1.961570560f); + const vector float vec_2_053119869 = (vector float)FOUROF(2.053119869f); + const vector float vec_2_562915447 = (vector float)FOUROF(-2.562915447f); + const vector float vec_3_072711026 = (vector float)FOUROF(3.072711026f); + + + int whichPass, whichHalf; + + for(whichPass = 1; whichPass<=2; whichPass++) + { + for(whichHalf = 1; whichHalf<=2; whichHalf++) + { + vector float tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; + vector float tmp10, tmp11, tmp12, tmp13; + vector float z1, z2, z3, z4, z5; + + tmp0 = vec_add(row0, row7); // tmp0 = dataptr[0] + dataptr[7]; + tmp7 = vec_sub(row0, row7); // tmp7 = dataptr[0] - dataptr[7]; + tmp3 = vec_add(row3, row4); // tmp3 = dataptr[3] + dataptr[4]; + tmp4 = vec_sub(row3, row4); // tmp4 = dataptr[3] - dataptr[4]; + tmp1 = vec_add(row1, row6); // tmp1 = dataptr[1] + dataptr[6]; + tmp6 = vec_sub(row1, row6); // tmp6 = dataptr[1] - dataptr[6]; + tmp2 = vec_add(row2, row5); // tmp2 = dataptr[2] + dataptr[5]; + tmp5 = vec_sub(row2, row5); // tmp5 = dataptr[2] - dataptr[5]; + + tmp10 = vec_add(tmp0, tmp3); // tmp10 = tmp0 + tmp3; + tmp13 = vec_sub(tmp0, tmp3); // tmp13 = tmp0 - tmp3; + tmp11 = vec_add(tmp1, tmp2); // tmp11 = tmp1 + tmp2; + tmp12 = vec_sub(tmp1, tmp2); // tmp12 = tmp1 - tmp2; + + + // dataptr[0] = (DCTELEM) ((tmp10 + tmp11) << PASS1_BITS); + row0 = vec_add(tmp10, tmp11); + + // dataptr[4] = (DCTELEM) ((tmp10 - tmp11) << PASS1_BITS); + row4 = vec_sub(tmp10, tmp11); + + + // z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100); + z1 = vec_madd(vec_add(tmp12, tmp13), vec_0_541196100, (vector float)zero); + + // dataptr[2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865), + // CONST_BITS-PASS1_BITS); + row2 = vec_madd(tmp13, vec_0_765366865, z1); + + // dataptr[6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065), + // CONST_BITS-PASS1_BITS); + row6 = vec_madd(tmp12, vec_1_847759065, z1); + + z1 = vec_add(tmp4, tmp7); // z1 = tmp4 + tmp7; + z2 = vec_add(tmp5, tmp6); // z2 = tmp5 + tmp6; + z3 = vec_add(tmp4, tmp6); // z3 = tmp4 + tmp6; + z4 = vec_add(tmp5, tmp7); // z4 = tmp5 + tmp7; + + // z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */ + z5 = vec_madd(vec_add(z3, z4), vec_1_175875602, (vector float)zero); + + // z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */ + z3 = vec_madd(z3, vec_1_961570560, z5); + + // z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */ + z4 = vec_madd(z4, vec_0_390180644, z5); + + // The following adds are rolled into the multiplies above + // z3 = vec_add(z3, z5); // z3 += z5; + // z4 = vec_add(z4, z5); // z4 += z5; + + // z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */ + // Wow! It's actually more effecient to roll this multiply + // into the adds below, even thought the multiply gets done twice! + // z2 = vec_madd(z2, vec_2_562915447, (vector float)zero); + + // z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */ + // Same with this one... + // z1 = vec_madd(z1, vec_0_899976223, (vector float)zero); + + // tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */ + // dataptr[7] = (DCTELEM) DESCALE(tmp4 + z1 + z3, CONST_BITS-PASS1_BITS); + row7 = vec_madd(tmp4, vec_0_298631336, vec_madd(z1, vec_0_899976223, z3)); + + // tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */ + // dataptr[5] = (DCTELEM) DESCALE(tmp5 + z2 + z4, CONST_BITS-PASS1_BITS); + row5 = vec_madd(tmp5, vec_2_053119869, vec_madd(z2, vec_2_562915447, z4)); + + // tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */ + // dataptr[3] = (DCTELEM) DESCALE(tmp6 + z2 + z3, CONST_BITS-PASS1_BITS); + row3 = vec_madd(tmp6, vec_3_072711026, vec_madd(z2, vec_2_562915447, z3)); + + // tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */ + // dataptr[1] = (DCTELEM) DESCALE(tmp7 + z1 + z4, CONST_BITS-PASS1_BITS); + row1 = vec_madd(z1, vec_0_899976223, vec_madd(tmp7, vec_1_501321110, z4)); + + // Swap the row values with the alts. If this is the first half, + // this sets up the low values to be acted on in the second half. + // If this is the second half, it puts the high values back in + // the row values where they are expected to be when we're done. + SWAP(row0, alt0); + SWAP(row1, alt1); + SWAP(row2, alt2); + SWAP(row3, alt3); + SWAP(row4, alt4); + SWAP(row5, alt5); + SWAP(row6, alt6); + SWAP(row7, alt7); + } + + if (whichPass == 1) + { + // transpose the data for the second pass + + // First, block transpose the upper right with lower left. + SWAP(row4, alt0); + SWAP(row5, alt1); + SWAP(row6, alt2); + SWAP(row7, alt3); + + // Now, transpose each block of four + TRANSPOSE4(row0, row1, row2, row3); + TRANSPOSE4(row4, row5, row6, row7); + TRANSPOSE4(alt0, alt1, alt2, alt3); + TRANSPOSE4(alt4, alt5, alt6, alt7); + } + } + } + + // perform the quantise step, using the floating point data + // still in the row/alt registers + { + const int* biasAddr; + const vector signed int* qmat; + vector float bias, negBias; + + if (s->mb_intra) + { + vector signed int baseVector; + + // We must cache element 0 in the intra case + // (it needs special handling). + baseVector = vec_cts(vec_splat(row0, 0), 0); + vec_ste(baseVector, 0, &oldBaseValue); + + qmat = (vector signed int*)s->q_intra_matrix[qscale]; + biasAddr = &(s->intra_quant_bias); + } + else + { + qmat = (vector signed int*)s->q_inter_matrix[qscale]; + biasAddr = &(s->inter_quant_bias); + } + + // Load the bias vector (We add 0.5 to the bias so that we're + // rounding when we convert to int, instead of flooring.) + { + vector signed int biasInt; + const vector float negOneFloat = (vector float)FOUROF(-1.0f); + LOAD4(biasInt, biasAddr); + bias = vec_ctf(biasInt, QUANT_BIAS_SHIFT); + negBias = vec_madd(bias, negOneFloat, zero); + } + + { + vector float q0, q1, q2, q3, q4, q5, q6, q7; + + q0 = vec_ctf(qmat[0], QMAT_SHIFT); + q1 = vec_ctf(qmat[2], QMAT_SHIFT); + q2 = vec_ctf(qmat[4], QMAT_SHIFT); + q3 = vec_ctf(qmat[6], QMAT_SHIFT); + q4 = vec_ctf(qmat[8], QMAT_SHIFT); + q5 = vec_ctf(qmat[10], QMAT_SHIFT); + q6 = vec_ctf(qmat[12], QMAT_SHIFT); + q7 = vec_ctf(qmat[14], QMAT_SHIFT); + + row0 = vec_sel(vec_madd(row0, q0, negBias), vec_madd(row0, q0, bias), + vec_cmpgt(row0, zero)); + row1 = vec_sel(vec_madd(row1, q1, negBias), vec_madd(row1, q1, bias), + vec_cmpgt(row1, zero)); + row2 = vec_sel(vec_madd(row2, q2, negBias), vec_madd(row2, q2, bias), + vec_cmpgt(row2, zero)); + row3 = vec_sel(vec_madd(row3, q3, negBias), vec_madd(row3, q3, bias), + vec_cmpgt(row3, zero)); + row4 = vec_sel(vec_madd(row4, q4, negBias), vec_madd(row4, q4, bias), + vec_cmpgt(row4, zero)); + row5 = vec_sel(vec_madd(row5, q5, negBias), vec_madd(row5, q5, bias), + vec_cmpgt(row5, zero)); + row6 = vec_sel(vec_madd(row6, q6, negBias), vec_madd(row6, q6, bias), + vec_cmpgt(row6, zero)); + row7 = vec_sel(vec_madd(row7, q7, negBias), vec_madd(row7, q7, bias), + vec_cmpgt(row7, zero)); + + q0 = vec_ctf(qmat[1], QMAT_SHIFT); + q1 = vec_ctf(qmat[3], QMAT_SHIFT); + q2 = vec_ctf(qmat[5], QMAT_SHIFT); + q3 = vec_ctf(qmat[7], QMAT_SHIFT); + q4 = vec_ctf(qmat[9], QMAT_SHIFT); + q5 = vec_ctf(qmat[11], QMAT_SHIFT); + q6 = vec_ctf(qmat[13], QMAT_SHIFT); + q7 = vec_ctf(qmat[15], QMAT_SHIFT); + + alt0 = vec_sel(vec_madd(alt0, q0, negBias), vec_madd(alt0, q0, bias), + vec_cmpgt(alt0, zero)); + alt1 = vec_sel(vec_madd(alt1, q1, negBias), vec_madd(alt1, q1, bias), + vec_cmpgt(alt1, zero)); + alt2 = vec_sel(vec_madd(alt2, q2, negBias), vec_madd(alt2, q2, bias), + vec_cmpgt(alt2, zero)); + alt3 = vec_sel(vec_madd(alt3, q3, negBias), vec_madd(alt3, q3, bias), + vec_cmpgt(alt3, zero)); + alt4 = vec_sel(vec_madd(alt4, q4, negBias), vec_madd(alt4, q4, bias), + vec_cmpgt(alt4, zero)); + alt5 = vec_sel(vec_madd(alt5, q5, negBias), vec_madd(alt5, q5, bias), + vec_cmpgt(alt5, zero)); + alt6 = vec_sel(vec_madd(alt6, q6, negBias), vec_madd(alt6, q6, bias), + vec_cmpgt(alt6, zero)); + alt7 = vec_sel(vec_madd(alt7, q7, negBias), vec_madd(alt7, q7, bias), + vec_cmpgt(alt7, zero)); + } + + + } + + // Store the data back into the original block + { + vector signed short data0, data1, data2, data3, data4, data5, data6, data7; + + data0 = vec_pack(vec_cts(row0, 0), vec_cts(alt0, 0)); + data1 = vec_pack(vec_cts(row1, 0), vec_cts(alt1, 0)); + data2 = vec_pack(vec_cts(row2, 0), vec_cts(alt2, 0)); + data3 = vec_pack(vec_cts(row3, 0), vec_cts(alt3, 0)); + data4 = vec_pack(vec_cts(row4, 0), vec_cts(alt4, 0)); + data5 = vec_pack(vec_cts(row5, 0), vec_cts(alt5, 0)); + data6 = vec_pack(vec_cts(row6, 0), vec_cts(alt6, 0)); + data7 = vec_pack(vec_cts(row7, 0), vec_cts(alt7, 0)); + + { + // Clamp for overflow + vector signed int max_q_int, min_q_int; + vector signed short max_q, min_q; + + LOAD4(max_q_int, &(s->max_qcoeff)); + LOAD4(min_q_int, &(s->min_qcoeff)); + + max_q = vec_pack(max_q_int, max_q_int); + min_q = vec_pack(min_q_int, min_q_int); + + data0 = vec_max(vec_min(data0, max_q), min_q); + data1 = vec_max(vec_min(data1, max_q), min_q); + data2 = vec_max(vec_min(data2, max_q), min_q); + data4 = vec_max(vec_min(data4, max_q), min_q); + data5 = vec_max(vec_min(data5, max_q), min_q); + data6 = vec_max(vec_min(data6, max_q), min_q); + data7 = vec_max(vec_min(data7, max_q), min_q); + } + + { + vector bool char zero_01, zero_23, zero_45, zero_67; + vector signed char scanIndices_01, scanIndices_23, scanIndices_45, scanIndices_67; + vector signed char negOne = vec_splat_s8(-1); + vector signed char* scanPtr = + (vector signed char*)(s->intra_scantable.inverse); + signed char lastNonZeroChar; + + // Determine the largest non-zero index. + zero_01 = vec_pack(vec_cmpeq(data0, (vector signed short)zero), + vec_cmpeq(data1, (vector signed short)zero)); + zero_23 = vec_pack(vec_cmpeq(data2, (vector signed short)zero), + vec_cmpeq(data3, (vector signed short)zero)); + zero_45 = vec_pack(vec_cmpeq(data4, (vector signed short)zero), + vec_cmpeq(data5, (vector signed short)zero)); + zero_67 = vec_pack(vec_cmpeq(data6, (vector signed short)zero), + vec_cmpeq(data7, (vector signed short)zero)); + + // 64 biggest values + scanIndices_01 = vec_sel(scanPtr[0], negOne, zero_01); + scanIndices_23 = vec_sel(scanPtr[1], negOne, zero_23); + scanIndices_45 = vec_sel(scanPtr[2], negOne, zero_45); + scanIndices_67 = vec_sel(scanPtr[3], negOne, zero_67); + + // 32 largest values + scanIndices_01 = vec_max(scanIndices_01, scanIndices_23); + scanIndices_45 = vec_max(scanIndices_45, scanIndices_67); + + // 16 largest values + scanIndices_01 = vec_max(scanIndices_01, scanIndices_45); + + // 8 largest values + scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne), + vec_mergel(scanIndices_01, negOne)); + + // 4 largest values + scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne), + vec_mergel(scanIndices_01, negOne)); + + // 2 largest values + scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne), + vec_mergel(scanIndices_01, negOne)); + + // largest value + scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne), + vec_mergel(scanIndices_01, negOne)); + + scanIndices_01 = vec_splat(scanIndices_01, 0); + + + vec_ste(scanIndices_01, 0, &lastNonZeroChar); + + lastNonZero = lastNonZeroChar; + + // While the data is still in vectors we check for the transpose IDCT permute + // and handle it using the vector unit if we can. This is the permute used + // by the altivec idct, so it is common when using the altivec dct. + + if ((lastNonZero > 0) && (s->dsp.idct_permutation_type == FF_TRANSPOSE_IDCT_PERM)) + { + TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7); + } + + vec_st(data0, 0, data); + vec_st(data1, 16, data); + vec_st(data2, 32, data); + vec_st(data3, 48, data); + vec_st(data4, 64, data); + vec_st(data5, 80, data); + vec_st(data6, 96, data); + vec_st(data7, 112, data); + } + } + + // special handling of block[0] + if (s->mb_intra) + { + if (!s->h263_aic) + { + if (n < 4) + oldBaseValue /= s->y_dc_scale; + else + oldBaseValue /= s->c_dc_scale; + } + + // Divide by 8, rounding the result + data[0] = (oldBaseValue + 4) >> 3; + } + + // We handled the tranpose permutation above and we don't + // need to permute the "no" permutation case. + if ((lastNonZero > 0) && + (s->dsp.idct_permutation_type != FF_TRANSPOSE_IDCT_PERM) && + (s->dsp.idct_permutation_type != FF_NO_IDCT_PERM)) + { + ff_block_permute(data, s->dsp.idct_permutation, + s->intra_scantable.scantable, lastNonZero); + } + + return lastNonZero; +} +#undef FOUROF + +/* + AltiVec version of dct_unquantize_h263 + this code assumes `block' is 16 bytes-aligned +*/ +void dct_unquantize_h263_altivec(MpegEncContext *s, + DCTELEM *block, int n, int qscale) +{ +POWERPC_PERF_DECLARE(altivec_dct_unquantize_h263_num, 1); + int i, level, qmul, qadd; + int nCoeffs; + + assert(s->block_last_index[n]>=0); + +POWERPC_PERF_START_COUNT(altivec_dct_unquantize_h263_num, 1); + + qadd = (qscale - 1) | 1; + qmul = qscale << 1; + + if (s->mb_intra) { + if (!s->h263_aic) { + if (n < 4) + block[0] = block[0] * s->y_dc_scale; + else + block[0] = block[0] * s->c_dc_scale; + }else + qadd = 0; + i = 1; + nCoeffs= 63; //does not allways use zigzag table + } else { + i = 0; + nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ]; + } + + { + register const_vector signed short vczero = (const_vector signed short)vec_splat_s16(0); + short __attribute__ ((aligned(16))) qmul8[] = + { + qmul, qmul, qmul, qmul, + qmul, qmul, qmul, qmul + }; + short __attribute__ ((aligned(16))) qadd8[] = + { + qadd, qadd, qadd, qadd, + qadd, qadd, qadd, qadd + }; + short __attribute__ ((aligned(16))) nqadd8[] = + { + -qadd, -qadd, -qadd, -qadd, + -qadd, -qadd, -qadd, -qadd + }; + register vector signed short blockv, qmulv, qaddv, nqaddv, temp1; + register vector bool short blockv_null, blockv_neg; + register short backup_0 = block[0]; + register int j = 0; + + qmulv = vec_ld(0, qmul8); + qaddv = vec_ld(0, qadd8); + nqaddv = vec_ld(0, nqadd8); + +#if 0 // block *is* 16 bytes-aligned, it seems. + // first make sure block[j] is 16 bytes-aligned + for(j = 0; (j <= nCoeffs) && ((((unsigned long)block) + (j << 1)) & 0x0000000F) ; j++) { + level = block[j]; + if (level) { + if (level < 0) { + level = level * qmul - qadd; + } else { + level = level * qmul + qadd; + } + block[j] = level; + } + } +#endif + + // vectorize all the 16 bytes-aligned blocks + // of 8 elements + for(; (j + 7) <= nCoeffs ; j+=8) + { + blockv = vec_ld(j << 1, block); + blockv_neg = vec_cmplt(blockv, vczero); + blockv_null = vec_cmpeq(blockv, vczero); + // choose between +qadd or -qadd as the third operand + temp1 = vec_sel(qaddv, nqaddv, blockv_neg); + // multiply & add (block{i,i+7} * qmul [+-] qadd) + temp1 = vec_mladd(blockv, qmulv, temp1); + // put 0 where block[{i,i+7} used to have 0 + blockv = vec_sel(temp1, blockv, blockv_null); + vec_st(blockv, j << 1, block); + } + + // if nCoeffs isn't a multiple of 8, finish the job + // using good old scalar units. + // (we could do it using a truncated vector, + // but I'm not sure it's worth the hassle) + for(; j <= nCoeffs ; j++) { + level = block[j]; + if (level) { + if (level < 0) { + level = level * qmul - qadd; + } else { + level = level * qmul + qadd; + } + block[j] = level; + } + } + + if (i == 1) + { // cheat. this avoid special-casing the first iteration + block[0] = backup_0; + } + } +POWERPC_PERF_STOP_COUNT(altivec_dct_unquantize_h263_num, nCoeffs == 63); +} diff --git a/contrib/ffmpeg/libavcodec/ppc/mpegvideo_ppc.c b/contrib/ffmpeg/libavcodec/ppc/mpegvideo_ppc.c new file mode 100644 index 000000000..c5e822f77 --- /dev/null +++ b/contrib/ffmpeg/libavcodec/ppc/mpegvideo_ppc.c @@ -0,0 +1,87 @@ +/* + * Copyright (c) 2002 Dieter Shirley + * + * This file is part of FFmpeg. + * + * FFmpeg 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 2.1 of the License, or (at your option) any later version. + * + * FFmpeg 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 FFmpeg; if not, write to the Free Software + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA + */ + +#include "../dsputil.h" +#include "../mpegvideo.h" +#include <time.h> + +#ifdef HAVE_ALTIVEC +#include "dsputil_altivec.h" +#endif + +extern int dct_quantize_altivec(MpegEncContext *s, + DCTELEM *block, int n, + int qscale, int *overflow); +extern void dct_unquantize_h263_altivec(MpegEncContext *s, + DCTELEM *block, int n, int qscale); + +extern void idct_put_altivec(uint8_t *dest, int line_size, int16_t *block); +extern void idct_add_altivec(uint8_t *dest, int line_size, int16_t *block); + + +void MPV_common_init_ppc(MpegEncContext *s) +{ +#ifdef HAVE_ALTIVEC + if (has_altivec()) + { + if (s->avctx->lowres==0) + { + if ((s->avctx->idct_algo == FF_IDCT_AUTO) || + (s->avctx->idct_algo == FF_IDCT_ALTIVEC)) + { + s->dsp.idct_put = idct_put_altivec; + s->dsp.idct_add = idct_add_altivec; + s->dsp.idct_permutation_type = FF_TRANSPOSE_IDCT_PERM; + } + } + + // Test to make sure that the dct required alignments are met. + if ((((long)(s->q_intra_matrix) & 0x0f) != 0) || + (((long)(s->q_inter_matrix) & 0x0f) != 0)) + { + av_log(s->avctx, AV_LOG_INFO, "Internal Error: q-matrix blocks must be 16-byte aligned " + "to use Altivec DCT. Reverting to non-altivec version.\n"); + return; + } + + if (((long)(s->intra_scantable.inverse) & 0x0f) != 0) + { + av_log(s->avctx, AV_LOG_INFO, "Internal Error: scan table blocks must be 16-byte aligned " + "to use Altivec DCT. Reverting to non-altivec version.\n"); + return; + } + + + if ((s->avctx->dct_algo == FF_DCT_AUTO) || + (s->avctx->dct_algo == FF_DCT_ALTIVEC)) + { +#if 0 /* seems to cause trouble under some circumstances */ + s->dct_quantize = dct_quantize_altivec; +#endif + s->dct_unquantize_h263_intra = dct_unquantize_h263_altivec; + s->dct_unquantize_h263_inter = dct_unquantize_h263_altivec; + } + } else +#endif + { + /* Non-AltiVec PPC optimisations here */ + } +} + diff --git a/contrib/ffmpeg/libavcodec/ppc/snow_altivec.c b/contrib/ffmpeg/libavcodec/ppc/snow_altivec.c new file mode 100644 index 000000000..b15672ffe --- /dev/null +++ b/contrib/ffmpeg/libavcodec/ppc/snow_altivec.c @@ -0,0 +1,788 @@ +/* + * Altivec optimized snow DSP utils + * Copyright (c) 2006 Luca Barbato <lu_zero@gentoo.org> + * + * This file is part of FFmpeg. + * + * FFmpeg 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 2.1 of the License, or (at your option) any later version. + * + * FFmpeg 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 FFmpeg; 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 "dsputil_altivec.h" +#include "../snow.h" + +#undef NDEBUG +#include <assert.h> + + + +//FIXME remove this replication +#define slice_buffer_get_line(slice_buf, line_num) ((slice_buf)->line[line_num] ? (slice_buf)->line[line_num] : slice_buffer_load_line((slice_buf), (line_num))) + +static DWTELEM * slice_buffer_load_line(slice_buffer * buf, int line) +{ + int offset; + DWTELEM * buffer; + +// av_log(NULL, AV_LOG_DEBUG, "Cache hit: %d\n", line); + + assert(buf->data_stack_top >= 0); +// assert(!buf->line[line]); + if (buf->line[line]) + return buf->line[line]; + + offset = buf->line_width * line; + buffer = buf->data_stack[buf->data_stack_top]; + buf->data_stack_top--; + buf->line[line] = buffer; + +// av_log(NULL, AV_LOG_DEBUG, "slice_buffer_load_line: line: %d remaining: %d\n", line, buf->data_stack_top + 1); + + return buffer; +} + + +//altivec code + +void ff_snow_horizontal_compose97i_altivec(DWTELEM *b, int width) +{ + const int w2= (width+1)>>1; + DECLARE_ALIGNED_16(DWTELEM, temp[(width>>1)]); + const int w_l= (width>>1); + const int w_r= w2 - 1; + int i; + vector signed int t1, t2, x, y, tmp1, tmp2; + vector signed int *vbuf, *vtmp; + vector unsigned char align; + + + + { // Lift 0 + DWTELEM * const ref = b + w2 - 1; + DWTELEM b_0 = b[0]; + vbuf = (vector signed int *)b; + + tmp1 = vec_ld (0, ref); + align = vec_lvsl (0, ref); + tmp2 = vec_ld (15, ref); + t1= vec_perm(tmp1, tmp2, align); + + i = 0; + + for (i=0; i<w_l-15; i+=16) { +#if 0 + b[i+0] = b[i+0] - ((3 * (ref[i+0] + ref[i+1]) + 4) >> 3); + b[i+1] = b[i+1] - ((3 * (ref[i+1] + ref[i+2]) + 4) >> 3); + b[i+2] = b[i+2] - ((3 * (ref[i+2] + ref[i+3]) + 4) >> 3); + b[i+3] = b[i+3] - ((3 * (ref[i+3] + ref[i+4]) + 4) >> 3); +#else + + tmp1 = vec_ld (0, ref+4+i); + tmp2 = vec_ld (15, ref+4+i); + + t2 = vec_perm(tmp1, tmp2, align); + + y = vec_add(t1,vec_sld(t1,t2,4)); + y = vec_add(vec_add(y,y),y); + + tmp1 = vec_ld (0, ref+8+i); + + y = vec_add(y, vec_splat_s32(4)); + y = vec_sra(y, vec_splat_u32(3)); + + tmp2 = vec_ld (15, ref+8+i); + + *vbuf = vec_sub(*vbuf, y); + + t1=t2; + + vbuf++; + + t2 = vec_perm(tmp1, tmp2, align); + + y = vec_add(t1,vec_sld(t1,t2,4)); + y = vec_add(vec_add(y,y),y); + + tmp1 = vec_ld (0, ref+12+i); + + y = vec_add(y, vec_splat_s32(4)); + y = vec_sra(y, vec_splat_u32(3)); + + tmp2 = vec_ld (15, ref+12+i); + + *vbuf = vec_sub(*vbuf, y); + + t1=t2; + + vbuf++; + + t2 = vec_perm(tmp1, tmp2, align); + + y = vec_add(t1,vec_sld(t1,t2,4)); + y = vec_add(vec_add(y,y),y); + + tmp1 = vec_ld (0, ref+16+i); + + y = vec_add(y, vec_splat_s32(4)); + y = vec_sra(y, vec_splat_u32(3)); + + tmp2 = vec_ld (15, ref+16+i); + + *vbuf = vec_sub(*vbuf, y); + + t1=t2; + + t2 = vec_perm(tmp1, tmp2, align); + + y = vec_add(t1,vec_sld(t1,t2,4)); + y = vec_add(vec_add(y,y),y); + + vbuf++; + + y = vec_add(y, vec_splat_s32(4)); + y = vec_sra(y, vec_splat_u32(3)); + *vbuf = vec_sub(*vbuf, y); + + t1=t2; + + vbuf++; + +#endif + } + + snow_horizontal_compose_lift_lead_out(i, b, b, ref, width, w_l, 0, W_DM, W_DO, W_DS); + b[0] = b_0 - ((W_DM * 2 * ref[1]+W_DO)>>W_DS); + } + + { // Lift 1 + DWTELEM * const dst = b+w2; + + i = 0; + for(; (((long)&dst[i]) & 0xF) && i<w_r; i++){ + dst[i] = dst[i] - (b[i] + b[i + 1]); + } + + align = vec_lvsl(0, b+i); + tmp1 = vec_ld(0, b+i); + vbuf = (vector signed int*) (dst + i); + tmp2 = vec_ld(15, b+i); + + t1 = vec_perm(tmp1, tmp2, align); + + for (; i<w_r-3; i+=4) { + +#if 0 + dst[i] = dst[i] - (b[i] + b[i + 1]); + dst[i+1] = dst[i+1] - (b[i+1] + b[i + 2]); + dst[i+2] = dst[i+2] - (b[i+2] + b[i + 3]); + dst[i+3] = dst[i+3] - (b[i+3] + b[i + 4]); +#else + + tmp1 = vec_ld(0, b+4+i); + tmp2 = vec_ld(15, b+4+i); + + t2 = vec_perm(tmp1, tmp2, align); + + y = vec_add(t1, vec_sld(t1,t2,4)); + *vbuf = vec_sub (*vbuf, y); + + vbuf++; + + t1 = t2; + +#endif + + } + + snow_horizontal_compose_lift_lead_out(i, dst, dst, b, width, w_r, 1, W_CM, W_CO, W_CS); + } + + { // Lift 2 + DWTELEM * const ref = b+w2 - 1; + DWTELEM b_0 = b[0]; + vbuf= (vector signed int *) b; + + tmp1 = vec_ld (0, ref); + align = vec_lvsl (0, ref); + tmp2 = vec_ld (15, ref); + t1= vec_perm(tmp1, tmp2, align); + + i = 0; + for (; i<w_l-15; i+=16) { +#if 0 + b[i] = b[i] - (((8 -(ref[i] + ref[i+1])) - (b[i] <<2)) >> 4); + b[i+1] = b[i+1] - (((8 -(ref[i+1] + ref[i+2])) - (b[i+1]<<2)) >> 4); + b[i+2] = b[i+2] - (((8 -(ref[i+2] + ref[i+3])) - (b[i+2]<<2)) >> 4); + b[i+3] = b[i+3] - (((8 -(ref[i+3] + ref[i+4])) - (b[i+3]<<2)) >> 4); +#else + tmp1 = vec_ld (0, ref+4+i); + tmp2 = vec_ld (15, ref+4+i); + + t2 = vec_perm(tmp1, tmp2, align); + + y = vec_add(t1,vec_sld(t1,t2,4)); + y = vec_sub(vec_splat_s32(8),y); + + tmp1 = vec_ld (0, ref+8+i); + + x = vec_sl(*vbuf,vec_splat_u32(2)); + y = vec_sra(vec_sub(y,x),vec_splat_u32(4)); + + tmp2 = vec_ld (15, ref+8+i); + + *vbuf = vec_sub( *vbuf, y); + + t1 = t2; + + vbuf++; + + t2 = vec_perm(tmp1, tmp2, align); + + y = vec_add(t1,vec_sld(t1,t2,4)); + y = vec_sub(vec_splat_s32(8),y); + + tmp1 = vec_ld (0, ref+12+i); + + x = vec_sl(*vbuf,vec_splat_u32(2)); + y = vec_sra(vec_sub(y,x),vec_splat_u32(4)); + + tmp2 = vec_ld (15, ref+12+i); + + *vbuf = vec_sub( *vbuf, y); + + t1 = t2; + + vbuf++; + + t2 = vec_perm(tmp1, tmp2, align); + + y = vec_add(t1,vec_sld(t1,t2,4)); + y = vec_sub(vec_splat_s32(8),y); + + tmp1 = vec_ld (0, ref+16+i); + + x = vec_sl(*vbuf,vec_splat_u32(2)); + y = vec_sra(vec_sub(y,x),vec_splat_u32(4)); + + tmp2 = vec_ld (15, ref+16+i); + + *vbuf = vec_sub( *vbuf, y); + + t1 = t2; + + vbuf++; + + t2 = vec_perm(tmp1, tmp2, align); + + y = vec_add(t1,vec_sld(t1,t2,4)); + y = vec_sub(vec_splat_s32(8),y); + + t1 = t2; + + x = vec_sl(*vbuf,vec_splat_u32(2)); + y = vec_sra(vec_sub(y,x),vec_splat_u32(4)); + *vbuf = vec_sub( *vbuf, y); + + vbuf++; + +#endif + } + + snow_horizontal_compose_liftS_lead_out(i, b, b, ref, width, w_l); + b[0] = b_0 - (((-2 * ref[1] + W_BO) - 4 * b_0) >> W_BS); + } + + { // Lift 3 + DWTELEM * const src = b+w2; + + vbuf = (vector signed int *)b; + vtmp = (vector signed int *)temp; + + i = 0; + align = vec_lvsl(0, src); + + for (; i<w_r-3; i+=4) { +#if 0 + temp[i] = src[i] - ((-3*(b[i] + b[i+1]))>>1); + temp[i+1] = src[i+1] - ((-3*(b[i+1] + b[i+2]))>>1); + temp[i+2] = src[i+2] - ((-3*(b[i+2] + b[i+3]))>>1); + temp[i+3] = src[i+3] - ((-3*(b[i+3] + b[i+4]))>>1); +#else + tmp1 = vec_ld(0,src+i); + t1 = vec_add(vbuf[0],vec_sld(vbuf[0],vbuf[1],4)); + tmp2 = vec_ld(15,src+i); + t1 = vec_sub(vec_splat_s32(0),t1); //bad! + t1 = vec_add(t1,vec_add(t1,t1)); + t2 = vec_perm(tmp1 ,tmp2 ,align); + t1 = vec_sra(t1,vec_splat_u32(1)); + vbuf++; + *vtmp = vec_sub(t2,t1); + vtmp++; + +#endif + + } + + snow_horizontal_compose_lift_lead_out(i, temp, src, b, width, w_r, 1, -3, 0, 1); + } + + { + //Interleave + int a; + vector signed int *t = (vector signed int *)temp, + *v = (vector signed int *)b; + + snow_interleave_line_header(&i, width, b, temp); + + for (; (i & 0xE) != 0xE; i-=2){ + b[i+1] = temp[i>>1]; + b[i] = b[i>>1]; + } + for (i-=14; i>=0; i-=16){ + a=i/4; + + v[a+3]=vec_mergel(v[(a>>1)+1],t[(a>>1)+1]); + v[a+2]=vec_mergeh(v[(a>>1)+1],t[(a>>1)+1]); + v[a+1]=vec_mergel(v[a>>1],t[a>>1]); + v[a]=vec_mergeh(v[a>>1],t[a>>1]); + + } + + } +} + +void ff_snow_vertical_compose97i_altivec(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, DWTELEM *b3, DWTELEM *b4, DWTELEM *b5, int width) +{ + int i, w4 = width/4; + vector signed int *v0, *v1,*v2,*v3,*v4,*v5; + vector signed int t1, t2; + + v0=(vector signed int *)b0; + v1=(vector signed int *)b1; + v2=(vector signed int *)b2; + v3=(vector signed int *)b3; + v4=(vector signed int *)b4; + v5=(vector signed int *)b5; + + for (i=0; i< w4;i++) + { + + #if 0 + b4[i] -= (3*(b3[i] + b5[i])+4)>>3; + b3[i] -= ((b2[i] + b4[i])); + b2[i] += ((b1[i] + b3[i])+4*b2[i]+8)>>4; + b1[i] += (3*(b0[i] + b2[i]))>>1; + #else + t1 = vec_add(v3[i], v5[i]); + t2 = vec_add(t1, vec_add(t1,t1)); + t1 = vec_add(t2, vec_splat_s32(4)); + v4[i] = vec_sub(v4[i], vec_sra(t1,vec_splat_u32(3))); + + v3[i] = vec_sub(v3[i], vec_add(v2[i], v4[i])); + + t1 = vec_add(vec_splat_s32(8), vec_add(v1[i], v3[i])); + t2 = vec_sl(v2[i], vec_splat_u32(2)); + v2[i] = vec_add(v2[i], vec_sra(vec_add(t1,t2),vec_splat_u32(4))); + t1 = vec_add(v0[i], v2[i]); + t2 = vec_add(t1, vec_add(t1,t1)); + v1[i] = vec_add(v1[i], vec_sra(t2,vec_splat_u32(1))); + + #endif + } + + for(i*=4; i < width; i++) + { + b4[i] -= (W_DM*(b3[i] + b5[i])+W_DO)>>W_DS; + b3[i] -= (W_CM*(b2[i] + b4[i])+W_CO)>>W_CS; + b2[i] += (W_BM*(b1[i] + b3[i])+4*b2[i]+W_BO)>>W_BS; + b1[i] += (W_AM*(b0[i] + b2[i])+W_AO)>>W_AS; + } +} + +#define LOAD_BLOCKS \ + tmp1 = vec_ld(0, &block[3][y*src_stride]);\ + align = vec_lvsl(0, &block[3][y*src_stride]);\ + tmp2 = vec_ld(15, &block[3][y*src_stride]);\ +\ + b3 = vec_perm(tmp1,tmp2,align);\ +\ + tmp1 = vec_ld(0, &block[2][y*src_stride]);\ + align = vec_lvsl(0, &block[2][y*src_stride]);\ + tmp2 = vec_ld(15, &block[2][y*src_stride]);\ +\ + b2 = vec_perm(tmp1,tmp2,align);\ +\ + tmp1 = vec_ld(0, &block[1][y*src_stride]);\ + align = vec_lvsl(0, &block[1][y*src_stride]);\ + tmp2 = vec_ld(15, &block[1][y*src_stride]);\ +\ + b1 = vec_perm(tmp1,tmp2,align);\ +\ + tmp1 = vec_ld(0, &block[0][y*src_stride]);\ + align = vec_lvsl(0, &block[0][y*src_stride]);\ + tmp2 = vec_ld(15, &block[0][y*src_stride]);\ +\ + b0 = vec_perm(tmp1,tmp2,align); + +#define LOAD_OBMCS \ + tmp1 = vec_ld(0, obmc1);\ + align = vec_lvsl(0, obmc1);\ + tmp2 = vec_ld(15, obmc1);\ +\ + ob1 = vec_perm(tmp1,tmp2,align);\ +\ + tmp1 = vec_ld(0, obmc2);\ + align = vec_lvsl(0, obmc2);\ + tmp2 = vec_ld(15, obmc2);\ +\ + ob2 = vec_perm(tmp1,tmp2,align);\ +\ + tmp1 = vec_ld(0, obmc3);\ + align = vec_lvsl(0, obmc3);\ + tmp2 = vec_ld(15, obmc3);\ +\ + ob3 = vec_perm(tmp1,tmp2,align);\ +\ + tmp1 = vec_ld(0, obmc4);\ + align = vec_lvsl(0, obmc4);\ + tmp2 = vec_ld(15, obmc4);\ +\ + ob4 = vec_perm(tmp1,tmp2,align); + +/* interleave logic + * h1 <- [ a,b,a,b, a,b,a,b, a,b,a,b, a,b,a,b ] + * h2 <- [ c,d,c,d, c,d,c,d, c,d,c,d, c,d,c,d ] + * h <- [ a,b,c,d, a,b,c,d, a,b,c,d, a,b,c,d ] + */ + +#define STEPS_0_1\ + h1 = (vector unsigned short)\ + vec_mergeh(ob1, ob2);\ +\ + h2 = (vector unsigned short)\ + vec_mergeh(ob3, ob4);\ +\ + ih = (vector unsigned char)\ + vec_mergeh(h1,h2);\ +\ + l1 = (vector unsigned short) vec_mergeh(b3, b2);\ +\ + ih1 = (vector unsigned char) vec_mergel(h1, h2);\ +\ + l2 = (vector unsigned short) vec_mergeh(b1, b0);\ +\ + il = (vector unsigned char) vec_mergeh(l1, l2);\ +\ + v[0] = (vector signed int) vec_msum(ih, il, vec_splat_u32(0));\ +\ + il1 = (vector unsigned char) vec_mergel(l1, l2);\ +\ + v[1] = (vector signed int) vec_msum(ih1, il1, vec_splat_u32(0)); + +#define FINAL_STEP_SCALAR\ + for(x=0; x<b_w; x++)\ + if(add){\ + vbuf[x] += dst[x + src_x];\ + vbuf[x] = (vbuf[x] + (1<<(FRAC_BITS-1))) >> FRAC_BITS;\ + if(vbuf[x]&(~255)) vbuf[x]= ~(vbuf[x]>>31);\ + dst8[x + y*src_stride] = vbuf[x];\ + }else{\ + dst[x + src_x] -= vbuf[x];\ + } + +static void inner_add_yblock_bw_8_obmc_16_altivec(uint8_t *obmc, + const int obmc_stride, + uint8_t * * block, int b_w, + int b_h, int src_x, int src_y, + int src_stride, slice_buffer * sb, + int add, uint8_t * dst8) +{ + int y, x; + DWTELEM * dst; + vector unsigned short h1, h2, l1, l2; + vector unsigned char ih, il, ih1, il1, tmp1, tmp2, align; + vector unsigned char b0,b1,b2,b3; + vector unsigned char ob1,ob2,ob3,ob4; + + DECLARE_ALIGNED_16(int, vbuf[16]); + vector signed int *v = (vector signed int *)vbuf, *d; + + for(y=0; y<b_h; y++){ + //FIXME ugly missue of obmc_stride + + uint8_t *obmc1= obmc + y*obmc_stride; + uint8_t *obmc2= obmc1+ (obmc_stride>>1); + uint8_t *obmc3= obmc1+ obmc_stride*(obmc_stride>>1); + uint8_t *obmc4= obmc3+ (obmc_stride>>1); + + dst = slice_buffer_get_line(sb, src_y + y); + d = (vector signed int *)(dst + src_x); + +//FIXME i could avoid some loads! + + // load blocks + LOAD_BLOCKS + + // load obmcs + LOAD_OBMCS + + // steps 0 1 + STEPS_0_1 + + FINAL_STEP_SCALAR + + } + +} + +#define STEPS_2_3\ + h1 = (vector unsigned short) vec_mergel(ob1, ob2);\ +\ + h2 = (vector unsigned short) vec_mergel(ob3, ob4);\ +\ + ih = (vector unsigned char) vec_mergeh(h1,h2);\ +\ + l1 = (vector unsigned short) vec_mergel(b3, b2);\ +\ + l2 = (vector unsigned short) vec_mergel(b1, b0);\ +\ + ih1 = (vector unsigned char) vec_mergel(h1,h2);\ +\ + il = (vector unsigned char) vec_mergeh(l1,l2);\ +\ + v[2] = (vector signed int) vec_msum(ih, il, vec_splat_u32(0));\ +\ + il1 = (vector unsigned char) vec_mergel(l1,l2);\ +\ + v[3] = (vector signed int) vec_msum(ih1, il1, vec_splat_u32(0)); + + +static void inner_add_yblock_bw_16_obmc_32_altivec(uint8_t *obmc, + const int obmc_stride, + uint8_t * * block, int b_w, + int b_h, int src_x, int src_y, + int src_stride, slice_buffer * sb, + int add, uint8_t * dst8) +{ + int y, x; + DWTELEM * dst; + vector unsigned short h1, h2, l1, l2; + vector unsigned char ih, il, ih1, il1, tmp1, tmp2, align; + vector unsigned char b0,b1,b2,b3; + vector unsigned char ob1,ob2,ob3,ob4; + DECLARE_ALIGNED_16(int, vbuf[b_w]); + vector signed int *v = (vector signed int *)vbuf, *d; + + for(y=0; y<b_h; y++){ + //FIXME ugly missue of obmc_stride + + uint8_t *obmc1= obmc + y*obmc_stride; + uint8_t *obmc2= obmc1+ (obmc_stride>>1); + uint8_t *obmc3= obmc1+ obmc_stride*(obmc_stride>>1); + uint8_t *obmc4= obmc3+ (obmc_stride>>1); + + dst = slice_buffer_get_line(sb, src_y + y); + d = (vector signed int *)(dst + src_x); + + // load blocks + LOAD_BLOCKS + + // load obmcs + LOAD_OBMCS + + // steps 0 1 2 3 + STEPS_0_1 + + STEPS_2_3 + + FINAL_STEP_SCALAR + + } +} + +#define FINAL_STEP_VEC \ +\ + if(add)\ + {\ + for(x=0; x<b_w/4; x++)\ + {\ + v[x] = vec_add(v[x], d[x]);\ + v[x] = vec_sra(vec_add(v[x],\ + vec_sl( vec_splat_s32(1),\ + vec_splat_u32(7))),\ + vec_splat_u32(8));\ +\ + mask = (vector bool int) vec_sl((vector signed int)\ + vec_cmpeq(v[x],v[x]),vec_splat_u32(8));\ + mask = (vector bool int) vec_and(v[x],vec_nor(mask,mask));\ +\ + mask = (vector bool int)\ + vec_cmpeq((vector signed int)mask,\ + (vector signed int)vec_splat_u32(0));\ +\ + vs = vec_sra(v[x],vec_splat_u32(8));\ + vs = vec_sra(v[x],vec_splat_u32(8));\ + vs = vec_sra(v[x],vec_splat_u32(15));\ +\ + vs = vec_nor(vs,vs);\ +\ + v[x]= vec_sel(v[x],vs,mask);\ + }\ +\ + for(x=0; x<b_w; x++)\ + dst8[x + y*src_stride] = vbuf[x];\ +\ + }\ + else\ + for(x=0; x<b_w/4; x++)\ + d[x] = vec_sub(d[x], v[x]); + +static void inner_add_yblock_a_bw_8_obmc_16_altivec(uint8_t *obmc, + const int obmc_stride, + uint8_t * * block, int b_w, + int b_h, int src_x, int src_y, + int src_stride, slice_buffer * sb, + int add, uint8_t * dst8) +{ + int y, x; + DWTELEM * dst; + vector bool int mask; + vector signed int vs; + vector unsigned short h1, h2, l1, l2; + vector unsigned char ih, il, ih1, il1, tmp1, tmp2, align; + vector unsigned char b0,b1,b2,b3; + vector unsigned char ob1,ob2,ob3,ob4; + + DECLARE_ALIGNED_16(int, vbuf[16]); + vector signed int *v = (vector signed int *)vbuf, *d; + + for(y=0; y<b_h; y++){ + //FIXME ugly missue of obmc_stride + + uint8_t *obmc1= obmc + y*obmc_stride; + uint8_t *obmc2= obmc1+ (obmc_stride>>1); + uint8_t *obmc3= obmc1+ obmc_stride*(obmc_stride>>1); + uint8_t *obmc4= obmc3+ (obmc_stride>>1); + + dst = slice_buffer_get_line(sb, src_y + y); + d = (vector signed int *)(dst + src_x); + +//FIXME i could avoid some loads! + + // load blocks + LOAD_BLOCKS + + // load obmcs + LOAD_OBMCS + + // steps 0 1 + STEPS_0_1 + + FINAL_STEP_VEC + + } + +} + +static void inner_add_yblock_a_bw_16_obmc_32_altivec(uint8_t *obmc, + const int obmc_stride, + uint8_t * * block, int b_w, + int b_h, int src_x, int src_y, + int src_stride, slice_buffer * sb, + int add, uint8_t * dst8) +{ + int y, x; + DWTELEM * dst; + vector bool int mask; + vector signed int vs; + vector unsigned short h1, h2, l1, l2; + vector unsigned char ih, il, ih1, il1, tmp1, tmp2, align; + vector unsigned char b0,b1,b2,b3; + vector unsigned char ob1,ob2,ob3,ob4; + DECLARE_ALIGNED_16(int, vbuf[b_w]); + vector signed int *v = (vector signed int *)vbuf, *d; + + for(y=0; y<b_h; y++){ + //FIXME ugly missue of obmc_stride + + uint8_t *obmc1= obmc + y*obmc_stride; + uint8_t *obmc2= obmc1+ (obmc_stride>>1); + uint8_t *obmc3= obmc1+ obmc_stride*(obmc_stride>>1); + uint8_t *obmc4= obmc3+ (obmc_stride>>1); + + dst = slice_buffer_get_line(sb, src_y + y); + d = (vector signed int *)(dst + src_x); + + // load blocks + LOAD_BLOCKS + + // load obmcs + LOAD_OBMCS + + // steps 0 1 2 3 + STEPS_0_1 + + STEPS_2_3 + + FINAL_STEP_VEC + + } +} + + +void ff_snow_inner_add_yblock_altivec(uint8_t *obmc, const int obmc_stride, + uint8_t * * block, int b_w, int b_h, + int src_x, int src_y, int src_stride, + slice_buffer * sb, int add, + uint8_t * dst8) +{ + if (src_x&15) { + if (b_w == 16) + inner_add_yblock_bw_16_obmc_32_altivec(obmc, obmc_stride, block, + b_w, b_h, src_x, src_y, + src_stride, sb, add, dst8); + else if (b_w == 8) + inner_add_yblock_bw_8_obmc_16_altivec(obmc, obmc_stride, block, + b_w, b_h, src_x, src_y, + src_stride, sb, add, dst8); + else + ff_snow_inner_add_yblock(obmc, obmc_stride, block, b_w, b_h, src_x, + src_y, src_stride, sb, add, dst8); + } else { + if (b_w == 16) + inner_add_yblock_a_bw_16_obmc_32_altivec(obmc, obmc_stride, block, + b_w, b_h, src_x, src_y, + src_stride, sb, add, dst8); + else if (b_w == 8) + inner_add_yblock_a_bw_8_obmc_16_altivec(obmc, obmc_stride, block, + b_w, b_h, src_x, src_y, + src_stride, sb, add, dst8); + else + ff_snow_inner_add_yblock(obmc, obmc_stride, block, b_w, b_h, src_x, + src_y, src_stride, sb, add, dst8); + } +} + + +void snow_init_altivec(DSPContext* c, AVCodecContext *avctx) +{ + c->horizontal_compose97i = ff_snow_horizontal_compose97i_altivec; + c->vertical_compose97i = ff_snow_vertical_compose97i_altivec; + c->inner_add_yblock = ff_snow_inner_add_yblock_altivec; +} diff --git a/contrib/ffmpeg/libavcodec/ppc/types_altivec.h b/contrib/ffmpeg/libavcodec/ppc/types_altivec.h new file mode 100644 index 000000000..f29026e04 --- /dev/null +++ b/contrib/ffmpeg/libavcodec/ppc/types_altivec.h @@ -0,0 +1,41 @@ +/* + * Copyright (c) 2006 Guillaume Poirier <gpoirier@mplayerhq.hu> + * + * This file is part of FFmpeg. + * + * FFmpeg 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 2.1 of the License, or (at your option) any later version. + * + * FFmpeg 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 FFmpeg; if not, write to the Free Software + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA + */ + +/*********************************************************************** + * Vector types + **********************************************************************/ +#define vec_u8_t vector unsigned char +#define vec_s8_t vector signed char +#define vec_u16_t vector unsigned short +#define vec_s16_t vector signed short +#define vec_u32_t vector unsigned int +#define vec_s32_t vector signed int + +/*********************************************************************** + * Null vector + **********************************************************************/ +#define LOAD_ZERO const vec_u8_t zerov = vec_splat_u8( 0 ) + +#define zero_u8v (vec_u8_t) zerov +#define zero_s8v (vec_s8_t) zerov +#define zero_u16v (vec_u16_t) zerov +#define zero_s16v (vec_s16_t) zerov +#define zero_u32v (vec_u32_t) zerov +#define zero_s32v (vec_s32_t) zerov diff --git a/contrib/ffmpeg/libavcodec/ppc/vc1dsp_altivec.c b/contrib/ffmpeg/libavcodec/ppc/vc1dsp_altivec.c new file mode 100644 index 000000000..114c9d41f --- /dev/null +++ b/contrib/ffmpeg/libavcodec/ppc/vc1dsp_altivec.c @@ -0,0 +1,338 @@ +/* + * VC-1 and WMV3 decoder - DSP functions AltiVec-optimized + * Copyright (c) 2006 Konstantin Shishkov + * + * This file is part of FFmpeg. + * + * FFmpeg 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 2.1 of the License, or (at your option) any later version. + * + * FFmpeg 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 FFmpeg; 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 "dsputil_altivec.h" + +// main steps of 8x8 transform +#define STEP8(s0, s1, s2, s3, s4, s5, s6, s7, vec_rnd) \ +do { \ + t0 = vec_sl(vec_add(s0, s4), vec_2); \ + t0 = vec_add(vec_sl(t0, vec_1), t0); \ + t0 = vec_add(t0, vec_rnd); \ + t1 = vec_sl(vec_sub(s0, s4), vec_2); \ + t1 = vec_add(vec_sl(t1, vec_1), t1); \ + t1 = vec_add(t1, vec_rnd); \ + t2 = vec_add(vec_sl(s6, vec_2), vec_sl(s6, vec_1)); \ + t2 = vec_add(t2, vec_sl(s2, vec_4)); \ + t3 = vec_add(vec_sl(s2, vec_2), vec_sl(s2, vec_1)); \ + t3 = vec_sub(t3, vec_sl(s6, vec_4)); \ + t4 = vec_add(t0, t2); \ + t5 = vec_add(t1, t3); \ + t6 = vec_sub(t1, t3); \ + t7 = vec_sub(t0, t2); \ +\ + t0 = vec_sl(vec_add(s1, s3), vec_4); \ + t0 = vec_add(t0, vec_sl(s5, vec_3)); \ + t0 = vec_add(t0, vec_sl(s7, vec_2)); \ + t0 = vec_add(t0, vec_sub(s5, s3)); \ +\ + t1 = vec_sl(vec_sub(s1, s5), vec_4); \ + t1 = vec_sub(t1, vec_sl(s7, vec_3)); \ + t1 = vec_sub(t1, vec_sl(s3, vec_2)); \ + t1 = vec_sub(t1, vec_add(s1, s7)); \ +\ + t2 = vec_sl(vec_sub(s7, s3), vec_4); \ + t2 = vec_add(t2, vec_sl(s1, vec_3)); \ + t2 = vec_add(t2, vec_sl(s5, vec_2)); \ + t2 = vec_add(t2, vec_sub(s1, s7)); \ +\ + t3 = vec_sl(vec_sub(s5, s7), vec_4); \ + t3 = vec_sub(t3, vec_sl(s3, vec_3)); \ + t3 = vec_add(t3, vec_sl(s1, vec_2)); \ + t3 = vec_sub(t3, vec_add(s3, s5)); \ +\ + s0 = vec_add(t4, t0); \ + s1 = vec_add(t5, t1); \ + s2 = vec_add(t6, t2); \ + s3 = vec_add(t7, t3); \ + s4 = vec_sub(t7, t3); \ + s5 = vec_sub(t6, t2); \ + s6 = vec_sub(t5, t1); \ + s7 = vec_sub(t4, t0); \ +}while(0) + +#define SHIFT_HOR8(s0, s1, s2, s3, s4, s5, s6, s7) \ +do { \ + s0 = vec_sra(s0, vec_3); \ + s1 = vec_sra(s1, vec_3); \ + s2 = vec_sra(s2, vec_3); \ + s3 = vec_sra(s3, vec_3); \ + s4 = vec_sra(s4, vec_3); \ + s5 = vec_sra(s5, vec_3); \ + s6 = vec_sra(s6, vec_3); \ + s7 = vec_sra(s7, vec_3); \ +}while(0) + +#define SHIFT_VERT8(s0, s1, s2, s3, s4, s5, s6, s7) \ +do { \ + s0 = vec_sra(s0, vec_7); \ + s1 = vec_sra(s1, vec_7); \ + s2 = vec_sra(s2, vec_7); \ + s3 = vec_sra(s3, vec_7); \ + s4 = vec_sra(vec_add(s4, vec_1s), vec_7); \ + s5 = vec_sra(vec_add(s5, vec_1s), vec_7); \ + s6 = vec_sra(vec_add(s6, vec_1s), vec_7); \ + s7 = vec_sra(vec_add(s7, vec_1s), vec_7); \ +}while(0) + +/* main steps of 4x4 transform */ +#define STEP4(s0, s1, s2, s3, vec_rnd) \ +do { \ + t1 = vec_add(vec_sl(s0, vec_4), s0); \ + t1 = vec_add(t1, vec_rnd); \ + t2 = vec_add(vec_sl(s2, vec_4), s2); \ + t0 = vec_add(t1, t2); \ + t1 = vec_sub(t1, t2); \ + t3 = vec_sl(vec_sub(s3, s1), vec_1); \ + t3 = vec_add(t3, vec_sl(t3, vec_2)); \ + t2 = vec_add(t3, vec_sl(s1, vec_5)); \ + t3 = vec_add(t3, vec_sl(s3, vec_3)); \ + t3 = vec_add(t3, vec_sl(s3, vec_2)); \ + s0 = vec_add(t0, t2); \ + s1 = vec_sub(t1, t3); \ + s2 = vec_add(t1, t3); \ + s3 = vec_sub(t0, t2); \ +}while (0) + +#define SHIFT_HOR4(s0, s1, s2, s3) \ + s0 = vec_sra(s0, vec_3); \ + s1 = vec_sra(s1, vec_3); \ + s2 = vec_sra(s2, vec_3); \ + s3 = vec_sra(s3, vec_3); + +#define SHIFT_VERT4(s0, s1, s2, s3) \ + s0 = vec_sra(s0, vec_7); \ + s1 = vec_sra(s1, vec_7); \ + s2 = vec_sra(s2, vec_7); \ + s3 = vec_sra(s3, vec_7); + +/** Do inverse transform on 8x8 block +*/ +static void vc1_inv_trans_8x8_altivec(DCTELEM block[64]) +{ + vector signed short src0, src1, src2, src3, src4, src5, src6, src7; + vector signed int s0, s1, s2, s3, s4, s5, s6, s7; + vector signed int s8, s9, sA, sB, sC, sD, sE, sF; + vector signed int t0, t1, t2, t3, t4, t5, t6, t7; + const vector signed int vec_64 = vec_sl(vec_splat_s32(4), vec_splat_u32(4)); + const vector unsigned int vec_7 = vec_splat_u32(7); + const vector unsigned int vec_5 = vec_splat_u32(5); + const vector unsigned int vec_4 = vec_splat_u32(4); + const vector signed int vec_4s = vec_splat_s32(4); + const vector unsigned int vec_3 = vec_splat_u32(3); + const vector unsigned int vec_2 = vec_splat_u32(2); + const vector signed int vec_1s = vec_splat_s32(1); + const vector unsigned int vec_1 = vec_splat_u32(1); + + + src0 = vec_ld( 0, block); + src1 = vec_ld( 16, block); + src2 = vec_ld( 32, block); + src3 = vec_ld( 48, block); + src4 = vec_ld( 64, block); + src5 = vec_ld( 80, block); + src6 = vec_ld( 96, block); + src7 = vec_ld(112, block); + + TRANSPOSE8(src0, src1, src2, src3, src4, src5, src6, src7); + s0 = vec_unpackl(src0); + s1 = vec_unpackl(src1); + s2 = vec_unpackl(src2); + s3 = vec_unpackl(src3); + s4 = vec_unpackl(src4); + s5 = vec_unpackl(src5); + s6 = vec_unpackl(src6); + s7 = vec_unpackl(src7); + s8 = vec_unpackh(src0); + s9 = vec_unpackh(src1); + sA = vec_unpackh(src2); + sB = vec_unpackh(src3); + sC = vec_unpackh(src4); + sD = vec_unpackh(src5); + sE = vec_unpackh(src6); + sF = vec_unpackh(src7); + STEP8(s0, s1, s2, s3, s4, s5, s6, s7, vec_4s); + SHIFT_HOR8(s0, s1, s2, s3, s4, s5, s6, s7); + STEP8(s8, s9, sA, sB, sC, sD, sE, sF, vec_4s); + SHIFT_HOR8(s8, s9, sA, sB, sC, sD, sE, sF); + src0 = vec_pack(s8, s0); + src1 = vec_pack(s9, s1); + src2 = vec_pack(sA, s2); + src3 = vec_pack(sB, s3); + src4 = vec_pack(sC, s4); + src5 = vec_pack(sD, s5); + src6 = vec_pack(sE, s6); + src7 = vec_pack(sF, s7); + TRANSPOSE8(src0, src1, src2, src3, src4, src5, src6, src7); + + s0 = vec_unpackl(src0); + s1 = vec_unpackl(src1); + s2 = vec_unpackl(src2); + s3 = vec_unpackl(src3); + s4 = vec_unpackl(src4); + s5 = vec_unpackl(src5); + s6 = vec_unpackl(src6); + s7 = vec_unpackl(src7); + s8 = vec_unpackh(src0); + s9 = vec_unpackh(src1); + sA = vec_unpackh(src2); + sB = vec_unpackh(src3); + sC = vec_unpackh(src4); + sD = vec_unpackh(src5); + sE = vec_unpackh(src6); + sF = vec_unpackh(src7); + STEP8(s0, s1, s2, s3, s4, s5, s6, s7, vec_64); + SHIFT_VERT8(s0, s1, s2, s3, s4, s5, s6, s7); + STEP8(s8, s9, sA, sB, sC, sD, sE, sF, vec_64); + SHIFT_VERT8(s8, s9, sA, sB, sC, sD, sE, sF); + src0 = vec_pack(s8, s0); + src1 = vec_pack(s9, s1); + src2 = vec_pack(sA, s2); + src3 = vec_pack(sB, s3); + src4 = vec_pack(sC, s4); + src5 = vec_pack(sD, s5); + src6 = vec_pack(sE, s6); + src7 = vec_pack(sF, s7); + + vec_st(src0, 0, block); + vec_st(src1, 16, block); + vec_st(src2, 32, block); + vec_st(src3, 48, block); + vec_st(src4, 64, block); + vec_st(src5, 80, block); + vec_st(src6, 96, block); + vec_st(src7,112, block); +} + +/** Do inverse transform on 8x4 part of block +*/ +static void vc1_inv_trans_8x4_altivec(DCTELEM block[64], int n) +{ + vector signed short src0, src1, src2, src3, src4, src5, src6, src7; + vector signed int s0, s1, s2, s3, s4, s5, s6, s7; + vector signed int s8, s9, sA, sB, sC, sD, sE, sF; + vector signed int t0, t1, t2, t3, t4, t5, t6, t7; + const vector signed int vec_64 = vec_sl(vec_splat_s32(4), vec_splat_u32(4)); + const vector unsigned int vec_7 = vec_splat_u32(7); + const vector unsigned int vec_5 = vec_splat_u32(5); + const vector unsigned int vec_4 = vec_splat_u32(4); + const vector signed int vec_4s = vec_splat_s32(4); + const vector unsigned int vec_3 = vec_splat_u32(3); + const vector unsigned int vec_2 = vec_splat_u32(2); + const vector unsigned int vec_1 = vec_splat_u32(1); + + src0 = vec_ld( 0, block); + src1 = vec_ld( 16, block); + src2 = vec_ld( 32, block); + src3 = vec_ld( 48, block); + src4 = vec_ld( 64, block); + src5 = vec_ld( 80, block); + src6 = vec_ld( 96, block); + src7 = vec_ld(112, block); + + TRANSPOSE8(src0, src1, src2, src3, src4, src5, src6, src7); + s0 = vec_unpackl(src0); + s1 = vec_unpackl(src1); + s2 = vec_unpackl(src2); + s3 = vec_unpackl(src3); + s4 = vec_unpackl(src4); + s5 = vec_unpackl(src5); + s6 = vec_unpackl(src6); + s7 = vec_unpackl(src7); + s8 = vec_unpackh(src0); + s9 = vec_unpackh(src1); + sA = vec_unpackh(src2); + sB = vec_unpackh(src3); + sC = vec_unpackh(src4); + sD = vec_unpackh(src5); + sE = vec_unpackh(src6); + sF = vec_unpackh(src7); + STEP8(s0, s1, s2, s3, s4, s5, s6, s7, vec_4s); + SHIFT_HOR8(s0, s1, s2, s3, s4, s5, s6, s7); + STEP8(s8, s9, sA, sB, sC, sD, sE, sF, vec_4s); + SHIFT_HOR8(s8, s9, sA, sB, sC, sD, sE, sF); + src0 = vec_pack(s8, s0); + src1 = vec_pack(s9, s1); + src2 = vec_pack(sA, s2); + src3 = vec_pack(sB, s3); + src4 = vec_pack(sC, s4); + src5 = vec_pack(sD, s5); + src6 = vec_pack(sE, s6); + src7 = vec_pack(sF, s7); + TRANSPOSE8(src0, src1, src2, src3, src4, src5, src6, src7); + + if(!n){ // upper half of block + s0 = vec_unpackh(src0); + s1 = vec_unpackh(src1); + s2 = vec_unpackh(src2); + s3 = vec_unpackh(src3); + s8 = vec_unpackl(src0); + s9 = vec_unpackl(src1); + sA = vec_unpackl(src2); + sB = vec_unpackl(src3); + STEP4(s0, s1, s2, s3, vec_64); + SHIFT_VERT4(s0, s1, s2, s3); + STEP4(s8, s9, sA, sB, vec_64); + SHIFT_VERT4(s8, s9, sA, sB); + src0 = vec_pack(s0, s8); + src1 = vec_pack(s1, s9); + src2 = vec_pack(s2, sA); + src3 = vec_pack(s3, sB); + + vec_st(src0, 0, block); + vec_st(src1, 16, block); + vec_st(src2, 32, block); + vec_st(src3, 48, block); + } else { //lower half of block + s0 = vec_unpackh(src4); + s1 = vec_unpackh(src5); + s2 = vec_unpackh(src6); + s3 = vec_unpackh(src7); + s8 = vec_unpackl(src4); + s9 = vec_unpackl(src5); + sA = vec_unpackl(src6); + sB = vec_unpackl(src7); + STEP4(s0, s1, s2, s3, vec_64); + SHIFT_VERT4(s0, s1, s2, s3); + STEP4(s8, s9, sA, sB, vec_64); + SHIFT_VERT4(s8, s9, sA, sB); + src4 = vec_pack(s0, s8); + src5 = vec_pack(s1, s9); + src6 = vec_pack(s2, sA); + src7 = vec_pack(s3, sB); + + vec_st(src4, 64, block); + vec_st(src5, 80, block); + vec_st(src6, 96, block); + vec_st(src7,112, block); + } +} + + +void vc1dsp_init_altivec(DSPContext* dsp, AVCodecContext *avctx) { + dsp->vc1_inv_trans_8x8 = vc1_inv_trans_8x8_altivec; + dsp->vc1_inv_trans_8x4 = vc1_inv_trans_8x4_altivec; +} |