/* next one assumes that ((line_size % 8) == 0) */
static void avg_pixels8_altivec(uint8_t * block, const uint8_t * pixels, ptrdiff_t line_size, int h)
{
register vector unsigned char pixelsv1, pixelsv2, pixelsv, blockv;
int i;
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, pixels);
pixelsv2 = vec_ld(16, pixels);
pixelsv = vec_perm(pixelsv1, pixelsv2, vec_lvsl(0, pixels));
if (rightside) {
pixelsv = vec_perm(blockv, pixelsv, vcprm(0,1,s0,s1));
} else {
pixelsv = vec_perm(blockv, pixelsv, vcprm(s0,s1,2,3));
}
blockv = vec_avg(blockv, pixelsv);
vec_st(blockv, 0, block);
pixels += line_size;
block += line_size;
}
}
/* next one assumes that ((line_size % 8) == 0) */
static void put_no_rnd_pixels8_xy2_altivec(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h)
{
register int i;
register vector unsigned char pixelsv1, pixelsv2, pixelsavg;
register vector unsigned char blockv, temp1, temp2;
register vector unsigned short pixelssum1, pixelssum2, temp3;
register const vector unsigned char vczero = (const vector unsigned char)vec_splat_u8(0);
register const vector unsigned short vcone = (const vector unsigned short)vec_splat_u16(1);
register const vector unsigned short vctwo = (const vector unsigned short)vec_splat_u16(2);
temp1 = vec_ld(0, pixels);
temp2 = vec_ld(16, pixels);
pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(0, pixels));
if ((((unsigned long)pixels) & 0x0000000F) == 0x0000000F) {
pixelsv2 = temp2;
} else {
pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
}
pixelsv1 = vec_mergeh(vczero, pixelsv1);
pixelsv2 = vec_mergeh(vczero, pixelsv2);
pixelssum1 = vec_add((vector unsigned short)pixelsv1,
(vector unsigned short)pixelsv2);
pixelssum1 = vec_add(pixelssum1, vcone);
for (i = 0; i < h ; i++) {
int rightside = ((unsigned long)block & 0x0000000F);
blockv = vec_ld(0, block);
temp1 = vec_ld(line_size, pixels);
temp2 = vec_ld(line_size + 16, pixels);
pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels));
if (((((unsigned long)pixels) + line_size) & 0x0000000F) == 0x0000000F) {
pixelsv2 = temp2;
} else {
pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
}
pixelsv1 = vec_mergeh(vczero, pixelsv1);
pixelsv2 = vec_mergeh(vczero, pixelsv2);
pixelssum2 = vec_add((vector unsigned short)pixelsv1,
(vector unsigned short)pixelsv2);
temp3 = vec_add(pixelssum1, pixelssum2);
temp3 = vec_sra(temp3, vctwo);
pixelssum1 = vec_add(pixelssum2, vcone);
pixelsavg = vec_packsu(temp3, (vector unsigned short) vczero);
if (rightside) {
blockv = vec_perm(blockv, pixelsavg, vcprm(0, 1, s0, s1));
} else {
blockv = vec_perm(blockv, pixelsavg, vcprm(s0, s1, 2, 3));
}
vec_st(blockv, 0, block);
block += line_size;
pixels += line_size;
}
}
static void vector_fmul_window_altivec(float *dst, const float *src0, const float *src1, const float *win, int len)
{
vector float zero, t0, t1, s0, s1, wi, wj;
const vector unsigned char reverse = vcprm(3,2,1,0);
int i,j;
dst += len;
win += len;
src0+= len;
zero = (vector float)vec_splat_u32(0);
for(i=-len*4, j=len*4-16; i<0; i+=16, j-=16) {
s0 = vec_ld(i, src0);
s1 = vec_ld(j, src1);
wi = vec_ld(i, win);
wj = vec_ld(j, win);
s1 = vec_perm(s1, s1, reverse);
wj = vec_perm(wj, wj, reverse);
t0 = vec_madd(s0, wj, zero);
t0 = vec_nmsub(s1, wi, t0);
t1 = vec_madd(s0, wi, zero);
t1 = vec_madd(s1, wj, t1);
t1 = vec_perm(t1, t1, reverse);
vec_st(t0, i, dst);
vec_st(t1, j, dst);
}
}
static void ff_imdct_calc_altivec(FFTContext *s, FFTSample *output, const FFTSample *input)
{
int k;
int n = 1 << s->mdct_bits;
int n4 = n >> 2;
int n16 = n >> 4;
vec_u32 sign = {1U<<31,1U<<31,1U<<31,1U<<31};
vec_u32 *p0 = (vec_u32*)(output+n4);
vec_u32 *p1 = (vec_u32*)(output+n4*3);
ff_imdct_half_altivec(s, output+n4, input);
for (k = 0; k < n16; k++) {
vec_u32 a = p0[k] ^ sign;
vec_u32 b = p1[-k-1];
p0[-k-1] = vec_perm(a, a, vcprm(3,2,1,0));
p1[k] = vec_perm(b, b, vcprm(3,2,1,0));
}
}
/* AltiVec-enhanced gmc1. ATM this code assumes stride is a multiple of 8
* to preserve proper dst alignment. */
void ff_gmc1_altivec(uint8_t *dst /* align 8 */, uint8_t *src /* align1 */,
int stride, int h, int x16, int y16, int rounder)
{
int i;
const DECLARE_ALIGNED(16, unsigned short, rounder_a) = rounder;
const DECLARE_ALIGNED(16, unsigned short, ABCD)[8] = {
(16 - x16) * (16 - y16), /* A */
(x16) * (16 - y16), /* B */
(16 - x16) * (y16), /* C */
(x16) * (y16), /* D */
0, 0, 0, 0 /* padding */
};
register const vector unsigned char vczero =
(const vector unsigned char) vec_splat_u8(0);
register const vector unsigned short vcsr8 =
(const vector unsigned short) vec_splat_u16(8);
register vector unsigned char dstv, dstv2, srcvB, srcvC, srcvD;
register vector unsigned short tempB, tempC, tempD;
unsigned long dst_odd = (unsigned long) dst & 0x0000000F;
unsigned long src_really_odd = (unsigned long) src & 0x0000000F;
register vector unsigned short tempA =
vec_ld(0, (const unsigned short *) ABCD);
register vector unsigned short Av = vec_splat(tempA, 0);
register vector unsigned short Bv = vec_splat(tempA, 1);
register vector unsigned short Cv = vec_splat(tempA, 2);
register vector unsigned short Dv = vec_splat(tempA, 3);
register vector unsigned short rounderV =
vec_splat((vec_u16) vec_lde(0, &rounder_a), 0);
/* we'll be able to pick-up our 9 char elements at src from those
* 32 bytes we load the first batch here, as inside the loop we can
* reuse 'src + stride' from one iteration as the 'src' of the next. */
register vector unsigned char src_0 = vec_ld(0, src);
register vector unsigned char src_1 = vec_ld(16, src);
register vector unsigned char 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'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 replace 32 int muls & 32 int adds.
* Isn't AltiVec nice? */
tempA = vec_mladd((vector unsigned short) srcvA, Av, rounderV);
tempB = vec_mladd((vector unsigned short) srcvB, Bv, tempA);
tempC = vec_mladd((vector unsigned short) srcvC, Cv, tempB);
tempD = vec_mladd((vector unsigned short) srcvD, Dv, tempC);
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;
}
}
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 DECLARE_ALIGNED_16(unsigned short, rounder_a[8]) =
{rounder, rounder, rounder, rounder,
rounder, rounder, rounder, rounder};
const DECLARE_ALIGNED_16(unsigned short, ABCD[8]) =
{
(16-x16)*(16-y16), /* A */
( x16)*(16-y16), /* B */
(16-x16)*( y16), /* C */
( x16)*( y16), /* D */
0, 0, 0, 0 /* padding */
};
register const_vector unsigned char vczero = (const_vector unsigned char)vec_splat_u8(0);
register const_vector unsigned short vcsr8 = (const_vector unsigned short)vec_splat_u16(8);
register vector unsigned char dstv, dstv2, src_0, src_1, srcvA, srcvB, srcvC, srcvD;
register vector unsigned short Av, Bv, Cv, Dv, rounderV, tempA, tempB, tempC, tempD;
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);
}
static void ff_imdct_half_altivec(FFTContext *s, FFTSample *output, const FFTSample *input)
{
int j, k;
int n = 1 << s->mdct_bits;
int n4 = n >> 2;
int n8 = n >> 3;
int n32 = n >> 5;
const uint16_t *revtabj = s->revtab;
const uint16_t *revtabk = s->revtab+n4;
const vec_f *tcos = (const vec_f*)(s->tcos+n8);
const vec_f *tsin = (const vec_f*)(s->tsin+n8);
const vec_f *pin = (const vec_f*)(input+n4);
vec_f *pout = (vec_f*)(output+n4);
/* pre rotation */
k = n32-1;
do {
vec_f cos,sin,cos0,sin0,cos1,sin1,re,im,r0,i0,r1,i1,a,b,c,d;
#define CMULA(p,o0,o1,o2,o3)\
a = pin[ k*2+p]; /* { z[k].re, z[k].im, z[k+1].re, z[k+1].im } */\
b = pin[-k*2-p-1]; /* { z[-k-2].re, z[-k-2].im, z[-k-1].re, z[-k-1].im } */\
re = vec_perm(a, b, vcprm(0,2,s0,s2)); /* { z[k].re, z[k+1].re, z[-k-2].re, z[-k-1].re } */\
im = vec_perm(a, b, vcprm(s3,s1,3,1)); /* { z[-k-1].im, z[-k-2].im, z[k+1].im, z[k].im } */\
cos = vec_perm(cos0, cos1, vcprm(o0,o1,s##o2,s##o3)); /* { cos[k], cos[k+1], cos[-k-2], cos[-k-1] } */\
sin = vec_perm(sin0, sin1, vcprm(o0,o1,s##o2,s##o3));\
r##p = im*cos - re*sin;\
i##p = re*cos + im*sin;
#define STORE2(v,dst)\
j = dst;\
vec_ste(v, 0, output+j*2);\
vec_ste(v, 4, output+j*2);
#define STORE8(p)\
a = vec_perm(r##p, i##p, vcprm(0,s0,0,s0));\
b = vec_perm(r##p, i##p, vcprm(1,s1,1,s1));\
c = vec_perm(r##p, i##p, vcprm(2,s2,2,s2));\
d = vec_perm(r##p, i##p, vcprm(3,s3,3,s3));\
STORE2(a, revtabk[ p*2-4]);\
STORE2(b, revtabk[ p*2-3]);\
STORE2(c, revtabj[-p*2+2]);\
STORE2(d, revtabj[-p*2+3]);
cos0 = tcos[k];
sin0 = tsin[k];
cos1 = tcos[-k-1];
sin1 = tsin[-k-1];
CMULA(0, 0,1,2,3);
CMULA(1, 2,3,0,1);
STORE8(0);
STORE8(1);
revtabj += 4;
revtabk -= 4;
k--;
} while(k >= 0);
ff_fft_calc_altivec(s, (FFTComplex*)output);
/* post rotation + reordering */
j = -n32;
k = n32-1;
do {
vec_f cos,sin,re,im,a,b,c,d;
#define CMULB(d0,d1,o)\
re = pout[o*2];\
im = pout[o*2+1];\
cos = tcos[o];\
sin = tsin[o];\
d0 = im*sin - re*cos;\
d1 = re*sin + im*cos;
CMULB(a,b,j);
CMULB(c,d,k);
pout[2*j] = vec_perm(a, d, vcprm(0,s3,1,s2));
pout[2*j+1] = vec_perm(a, d, vcprm(2,s1,3,s0));
pout[2*k] = vec_perm(c, b, vcprm(0,s3,1,s2));
pout[2*k+1] = vec_perm(c, b, vcprm(2,s1,3,s0));
j++;
k--;
} while(k >= 0);
}
void dct64_altivec(real *a,real *b,real *c)
{
real __attribute__ ((aligned(16))) b1[0x20];
real __attribute__ ((aligned(16))) b2[0x20];
real *out0 = a;
real *out1 = b;
real *samples = c;
const vector float vczero = (const vector float)FOUROF(0.);
const vector unsigned char reverse = (const vector unsigned char)vcprm(3,2,1,0);
if (((unsigned long)b1 & 0x0000000F) ||
((unsigned long)b2 & 0x0000000F))
{
printf("MISALIGNED:\t%p\t%p\t%p\t%p\t%p\n",
b1, b2, a, b, samples);
}
#ifdef ALTIVEC_USE_REFERENCE_C_CODE
{
register real *costab = mp3lib_pnts[0];
b1[0x00] = samples[0x00] + samples[0x1F];
b1[0x01] = samples[0x01] + samples[0x1E];
b1[0x02] = samples[0x02] + samples[0x1D];
b1[0x03] = samples[0x03] + samples[0x1C];
b1[0x04] = samples[0x04] + samples[0x1B];
b1[0x05] = samples[0x05] + samples[0x1A];
b1[0x06] = samples[0x06] + samples[0x19];
b1[0x07] = samples[0x07] + samples[0x18];
b1[0x08] = samples[0x08] + samples[0x17];
b1[0x09] = samples[0x09] + samples[0x16];
b1[0x0A] = samples[0x0A] + samples[0x15];
b1[0x0B] = samples[0x0B] + samples[0x14];
b1[0x0C] = samples[0x0C] + samples[0x13];
b1[0x0D] = samples[0x0D] + samples[0x12];
b1[0x0E] = samples[0x0E] + samples[0x11];
b1[0x0F] = samples[0x0F] + samples[0x10];
b1[0x10] = (samples[0x0F] - samples[0x10]) * costab[0xF];
b1[0x11] = (samples[0x0E] - samples[0x11]) * costab[0xE];
b1[0x12] = (samples[0x0D] - samples[0x12]) * costab[0xD];
b1[0x13] = (samples[0x0C] - samples[0x13]) * costab[0xC];
b1[0x14] = (samples[0x0B] - samples[0x14]) * costab[0xB];
b1[0x15] = (samples[0x0A] - samples[0x15]) * costab[0xA];
b1[0x16] = (samples[0x09] - samples[0x16]) * costab[0x9];
b1[0x17] = (samples[0x08] - samples[0x17]) * costab[0x8];
b1[0x18] = (samples[0x07] - samples[0x18]) * costab[0x7];
b1[0x19] = (samples[0x06] - samples[0x19]) * costab[0x6];
b1[0x1A] = (samples[0x05] - samples[0x1A]) * costab[0x5];
b1[0x1B] = (samples[0x04] - samples[0x1B]) * costab[0x4];
b1[0x1C] = (samples[0x03] - samples[0x1C]) * costab[0x3];
b1[0x1D] = (samples[0x02] - samples[0x1D]) * costab[0x2];
b1[0x1E] = (samples[0x01] - samples[0x1E]) * costab[0x1];
b1[0x1F] = (samples[0x00] - samples[0x1F]) * costab[0x0];
}
{
register real *costab = mp3lib_pnts[1];
b2[0x00] = b1[0x00] + b1[0x0F];
b2[0x01] = b1[0x01] + b1[0x0E];
b2[0x02] = b1[0x02] + b1[0x0D];
b2[0x03] = b1[0x03] + b1[0x0C];
b2[0x04] = b1[0x04] + b1[0x0B];
b2[0x05] = b1[0x05] + b1[0x0A];
b2[0x06] = b1[0x06] + b1[0x09];
b2[0x07] = b1[0x07] + b1[0x08];
b2[0x08] = (b1[0x07] - b1[0x08]) * costab[7];
b2[0x09] = (b1[0x06] - b1[0x09]) * costab[6];
b2[0x0A] = (b1[0x05] - b1[0x0A]) * costab[5];
b2[0x0B] = (b1[0x04] - b1[0x0B]) * costab[4];
b2[0x0C] = (b1[0x03] - b1[0x0C]) * costab[3];
b2[0x0D] = (b1[0x02] - b1[0x0D]) * costab[2];
b2[0x0E] = (b1[0x01] - b1[0x0E]) * costab[1];
b2[0x0F] = (b1[0x00] - b1[0x0F]) * costab[0];
b2[0x10] = b1[0x10] + b1[0x1F];
b2[0x11] = b1[0x11] + b1[0x1E];
b2[0x12] = b1[0x12] + b1[0x1D];
b2[0x13] = b1[0x13] + b1[0x1C];
b2[0x14] = b1[0x14] + b1[0x1B];
b2[0x15] = b1[0x15] + b1[0x1A];
b2[0x16] = b1[0x16] + b1[0x19];
b2[0x17] = b1[0x17] + b1[0x18];
b2[0x18] = (b1[0x18] - b1[0x17]) * costab[7];
b2[0x19] = (b1[0x19] - b1[0x16]) * costab[6];
b2[0x1A] = (b1[0x1A] - b1[0x15]) * costab[5];
b2[0x1B] = (b1[0x1B] - b1[0x14]) * costab[4];
b2[0x1C] = (b1[0x1C] - b1[0x13]) * costab[3];
b2[0x1D] = (b1[0x1D] - b1[0x12]) * costab[2];
b2[0x1E] = (b1[0x1E] - b1[0x11]) * costab[1];
b2[0x1F] = (b1[0x1F] - b1[0x10]) * costab[0];
}
{
register real *costab = mp3lib_pnts[2];
b1[0x00] = b2[0x00] + b2[0x07];
b1[0x01] = b2[0x01] + b2[0x06];
b1[0x02] = b2[0x02] + b2[0x05];
b1[0x03] = b2[0x03] + b2[0x04];
b1[0x04] = (b2[0x03] - b2[0x04]) * costab[3];
b1[0x05] = (b2[0x02] - b2[0x05]) * costab[2];
b1[0x06] = (b2[0x01] - b2[0x06]) * costab[1];
b1[0x07] = (b2[0x00] - b2[0x07]) * costab[0];
b1[0x08] = b2[0x08] + b2[0x0F];
b1[0x09] = b2[0x09] + b2[0x0E];
b1[0x0A] = b2[0x0A] + b2[0x0D];
b1[0x0B] = b2[0x0B] + b2[0x0C];
b1[0x0C] = (b2[0x0C] - b2[0x0B]) * costab[3];
b1[0x0D] = (b2[0x0D] - b2[0x0A]) * costab[2];
b1[0x0E] = (b2[0x0E] - b2[0x09]) * costab[1];
b1[0x0F] = (b2[0x0F] - b2[0x08]) * costab[0];
b1[0x10] = b2[0x10] + b2[0x17];
b1[0x11] = b2[0x11] + b2[0x16];
b1[0x12] = b2[0x12] + b2[0x15];
b1[0x13] = b2[0x13] + b2[0x14];
b1[0x14] = (b2[0x13] - b2[0x14]) * costab[3];
b1[0x15] = (b2[0x12] - b2[0x15]) * costab[2];
b1[0x16] = (b2[0x11] - b2[0x16]) * costab[1];
b1[0x17] = (b2[0x10] - b2[0x17]) * costab[0];
b1[0x18] = b2[0x18] + b2[0x1F];
b1[0x19] = b2[0x19] + b2[0x1E];
b1[0x1A] = b2[0x1A] + b2[0x1D];
b1[0x1B] = b2[0x1B] + b2[0x1C];
b1[0x1C] = (b2[0x1C] - b2[0x1B]) * costab[3];
b1[0x1D] = (b2[0x1D] - b2[0x1A]) * costab[2];
b1[0x1E] = (b2[0x1E] - b2[0x19]) * costab[1];
b1[0x1F] = (b2[0x1F] - b2[0x18]) * costab[0];
}
#else /* ALTIVEC_USE_REFERENCE_C_CODE */
// How does it work ?
// the first three passes are reproducted in the three block below
// all computations are done on a 4 elements vector
// 'reverse' is a special perumtation vector used to reverse
// the order of the elements inside a vector.
// note that all loads/stores to b1 (b2) between passes 1 and 2 (2 and 3)
// have been removed, all elements are stored inside b1vX (b2vX)
{
register vector float
b1v0, b1v1, b1v2, b1v3,
b1v4, b1v5, b1v6, b1v7;
register vector float
temp1, temp2;
{
register real *costab = mp3lib_pnts[0];
register vector float
samplesv1, samplesv2, samplesv3, samplesv4,
samplesv5, samplesv6, samplesv7, samplesv8,
samplesv9;
register vector unsigned char samples_perm = vec_lvsl(0, samples);
register vector float costabv1, costabv2, costabv3, costabv4, costabv5;
register vector unsigned char costab_perm = vec_lvsl(0, costab);
samplesv1 = vec_ld(0, samples);
samplesv2 = vec_ld(16, samples);
samplesv1 = vec_perm(samplesv1, samplesv2, samples_perm);
samplesv3 = vec_ld(32, samples);
samplesv2 = vec_perm(samplesv2, samplesv3, samples_perm);
samplesv4 = vec_ld(48, samples);
samplesv3 = vec_perm(samplesv3, samplesv4, samples_perm);
samplesv5 = vec_ld(64, samples);
samplesv4 = vec_perm(samplesv4, samplesv5, samples_perm);
samplesv6 = vec_ld(80, samples);
samplesv5 = vec_perm(samplesv5, samplesv6, samples_perm);
samplesv7 = vec_ld(96, samples);
samplesv6 = vec_perm(samplesv6, samplesv7, samples_perm);
samplesv8 = vec_ld(112, samples);
samplesv7 = vec_perm(samplesv7, samplesv8, samples_perm);
samplesv9 = vec_ld(128, samples);
samplesv8 = vec_perm(samplesv8, samplesv9, samples_perm);
temp1 = vec_add(samplesv1,
vec_perm(samplesv8, samplesv8, reverse));
//vec_st(temp1, 0, b1);
b1v0 = temp1;
temp1 = vec_add(samplesv2,
vec_perm(samplesv7, samplesv7, reverse));
//vec_st(temp1, 16, b1);
b1v1 = temp1;
temp1 = vec_add(samplesv3,
vec_perm(samplesv6, samplesv6, reverse));
//vec_st(temp1, 32, b1);
b1v2 = temp1;
temp1 = vec_add(samplesv4,
vec_perm(samplesv5, samplesv5, reverse));
//vec_st(temp1, 48, b1);
b1v3 = temp1;
costabv1 = vec_ld(0, costab);
costabv2 = vec_ld(16, costab);
costabv1 = vec_perm(costabv1, costabv2, costab_perm);
costabv3 = vec_ld(32, costab);
costabv2 = vec_perm(costabv2, costabv3, costab_perm);
costabv4 = vec_ld(48, costab);
costabv3 = vec_perm(costabv3, costabv4, costab_perm);
costabv5 = vec_ld(64, costab);
costabv4 = vec_perm(costabv4, costabv5, costab_perm);
temp1 = vec_sub(vec_perm(samplesv4, samplesv4, reverse),
samplesv5);
temp2 = vec_madd(temp1,
vec_perm(costabv4, costabv4, reverse),
vczero);
//vec_st(temp2, 64, b1);
b1v4 = temp2;
temp1 = vec_sub(vec_perm(samplesv3, samplesv3, reverse),
samplesv6);
temp2 = vec_madd(temp1,
vec_perm(costabv3, costabv3, reverse),
vczero);
//vec_st(temp2, 80, b1);
b1v5 = temp2;
temp1 = vec_sub(vec_perm(samplesv2, samplesv2, reverse),
samplesv7);
temp2 = vec_madd(temp1,
vec_perm(costabv2, costabv2, reverse),
vczero);
//vec_st(temp2, 96, b1);
b1v6 = temp2;
temp1 = vec_sub(vec_perm(samplesv1, samplesv1, reverse),
samplesv8);
temp2 = vec_madd(temp1,
vec_perm(costabv1, costabv1, reverse),
vczero);
//vec_st(temp2, 112, b1);
b1v7 = temp2;
}
{
register vector float
b2v0, b2v1, b2v2, b2v3,
b2v4, b2v5, b2v6, b2v7;
{
register real *costab = mp3lib_pnts[1];
register vector float costabv1r, costabv2r, costabv1, costabv2, costabv3;
register vector unsigned char costab_perm = vec_lvsl(0, costab);
costabv1 = vec_ld(0, costab);
costabv2 = vec_ld(16, costab);
costabv1 = vec_perm(costabv1, costabv2, costab_perm);
costabv3 = vec_ld(32, costab);
costabv2 = vec_perm(costabv2, costabv3 , costab_perm);
costabv1r = vec_perm(costabv1, costabv1, reverse);
costabv2r = vec_perm(costabv2, costabv2, reverse);
temp1 = vec_add(b1v0, vec_perm(b1v3, b1v3, reverse));
//vec_st(temp1, 0, b2);
b2v0 = temp1;
temp1 = vec_add(b1v1, vec_perm(b1v2, b1v2, reverse));
//vec_st(temp1, 16, b2);
b2v1 = temp1;
temp2 = vec_sub(vec_perm(b1v1, b1v1, reverse), b1v2);
temp1 = vec_madd(temp2, costabv2r, vczero);
//vec_st(temp1, 32, b2);
b2v2 = temp1;
temp2 = vec_sub(vec_perm(b1v0, b1v0, reverse), b1v3);
temp1 = vec_madd(temp2, costabv1r, vczero);
//vec_st(temp1, 48, b2);
b2v3 = temp1;
temp1 = vec_add(b1v4, vec_perm(b1v7, b1v7, reverse));
//vec_st(temp1, 64, b2);
b2v4 = temp1;
temp1 = vec_add(b1v5, vec_perm(b1v6, b1v6, reverse));
//vec_st(temp1, 80, b2);
b2v5 = temp1;
temp2 = vec_sub(b1v6, vec_perm(b1v5, b1v5, reverse));
temp1 = vec_madd(temp2, costabv2r, vczero);
//vec_st(temp1, 96, b2);
b2v6 = temp1;
temp2 = vec_sub(b1v7, vec_perm(b1v4, b1v4, reverse));
temp1 = vec_madd(temp2, costabv1r, vczero);
//vec_st(temp1, 112, b2);
b2v7 = temp1;
}
{
register real *costab = mp3lib_pnts[2];
vector float costabv1r, costabv1, costabv2;
vector unsigned char costab_perm = vec_lvsl(0, costab);
costabv1 = vec_ld(0, costab);
costabv2 = vec_ld(16, costab);
costabv1 = vec_perm(costabv1, costabv2, costab_perm);
costabv1r = vec_perm(costabv1, costabv1, reverse);
temp1 = vec_add(b2v0, vec_perm(b2v1, b2v1, reverse));
vec_st(temp1, 0, b1);
temp2 = vec_sub(vec_perm(b2v0, b2v0, reverse), b2v1);
temp1 = vec_madd(temp2, costabv1r, vczero);
vec_st(temp1, 16, b1);
temp1 = vec_add(b2v2, vec_perm(b2v3, b2v3, reverse));
vec_st(temp1, 32, b1);
temp2 = vec_sub(b2v3, vec_perm(b2v2, b2v2, reverse));
temp1 = vec_madd(temp2, costabv1r, vczero);
vec_st(temp1, 48, b1);
temp1 = vec_add(b2v4, vec_perm(b2v5, b2v5, reverse));
vec_st(temp1, 64, b1);
temp2 = vec_sub(vec_perm(b2v4, b2v4, reverse), b2v5);
temp1 = vec_madd(temp2, costabv1r, vczero);
vec_st(temp1, 80, b1);
temp1 = vec_add(b2v6, vec_perm(b2v7, b2v7, reverse));
vec_st(temp1, 96, b1);
temp2 = vec_sub(b2v7, vec_perm(b2v6, b2v6, reverse));
temp1 = vec_madd(temp2, costabv1r, vczero);
vec_st(temp1, 112, b1);
}
}
}
#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
{
register real const cos0 = mp3lib_pnts[3][0];
register real const cos1 = mp3lib_pnts[3][1];
b2[0x00] = b1[0x00] + b1[0x03];
b2[0x01] = b1[0x01] + b1[0x02];
b2[0x02] = (b1[0x01] - b1[0x02]) * cos1;
b2[0x03] = (b1[0x00] - b1[0x03]) * cos0;
b2[0x04] = b1[0x04] + b1[0x07];
b2[0x05] = b1[0x05] + b1[0x06];
b2[0x06] = (b1[0x06] - b1[0x05]) * cos1;
b2[0x07] = (b1[0x07] - b1[0x04]) * cos0;
b2[0x08] = b1[0x08] + b1[0x0B];
b2[0x09] = b1[0x09] + b1[0x0A];
b2[0x0A] = (b1[0x09] - b1[0x0A]) * cos1;
b2[0x0B] = (b1[0x08] - b1[0x0B]) * cos0;
b2[0x0C] = b1[0x0C] + b1[0x0F];
b2[0x0D] = b1[0x0D] + b1[0x0E];
b2[0x0E] = (b1[0x0E] - b1[0x0D]) * cos1;
b2[0x0F] = (b1[0x0F] - b1[0x0C]) * cos0;
b2[0x10] = b1[0x10] + b1[0x13];
b2[0x11] = b1[0x11] + b1[0x12];
b2[0x12] = (b1[0x11] - b1[0x12]) * cos1;
b2[0x13] = (b1[0x10] - b1[0x13]) * cos0;
b2[0x14] = b1[0x14] + b1[0x17];
b2[0x15] = b1[0x15] + b1[0x16];
b2[0x16] = (b1[0x16] - b1[0x15]) * cos1;
b2[0x17] = (b1[0x17] - b1[0x14]) * cos0;
b2[0x18] = b1[0x18] + b1[0x1B];
b2[0x19] = b1[0x19] + b1[0x1A];
b2[0x1A] = (b1[0x19] - b1[0x1A]) * cos1;
b2[0x1B] = (b1[0x18] - b1[0x1B]) * cos0;
b2[0x1C] = b1[0x1C] + b1[0x1F];
b2[0x1D] = b1[0x1D] + b1[0x1E];
b2[0x1E] = (b1[0x1E] - b1[0x1D]) * cos1;
b2[0x1F] = (b1[0x1F] - b1[0x1C]) * cos0;
}
{
register real const cos0 = mp3lib_pnts[4][0];
b1[0x00] = b2[0x00] + b2[0x01];
b1[0x01] = (b2[0x00] - b2[0x01]) * cos0;
b1[0x02] = b2[0x02] + b2[0x03];
b1[0x03] = (b2[0x03] - b2[0x02]) * cos0;
b1[0x02] += b1[0x03];
b1[0x04] = b2[0x04] + b2[0x05];
b1[0x05] = (b2[0x04] - b2[0x05]) * cos0;
b1[0x06] = b2[0x06] + b2[0x07];
b1[0x07] = (b2[0x07] - b2[0x06]) * cos0;
b1[0x06] += b1[0x07];
b1[0x04] += b1[0x06];
b1[0x06] += b1[0x05];
b1[0x05] += b1[0x07];
b1[0x08] = b2[0x08] + b2[0x09];
b1[0x09] = (b2[0x08] - b2[0x09]) * cos0;
b1[0x0A] = b2[0x0A] + b2[0x0B];
b1[0x0B] = (b2[0x0B] - b2[0x0A]) * cos0;
b1[0x0A] += b1[0x0B];
b1[0x0C] = b2[0x0C] + b2[0x0D];
b1[0x0D] = (b2[0x0C] - b2[0x0D]) * cos0;
b1[0x0E] = b2[0x0E] + b2[0x0F];
b1[0x0F] = (b2[0x0F] - b2[0x0E]) * cos0;
b1[0x0E] += b1[0x0F];
b1[0x0C] += b1[0x0E];
b1[0x0E] += b1[0x0D];
b1[0x0D] += b1[0x0F];
b1[0x10] = b2[0x10] + b2[0x11];
b1[0x11] = (b2[0x10] - b2[0x11]) * cos0;
b1[0x12] = b2[0x12] + b2[0x13];
b1[0x13] = (b2[0x13] - b2[0x12]) * cos0;
b1[0x12] += b1[0x13];
b1[0x14] = b2[0x14] + b2[0x15];
b1[0x15] = (b2[0x14] - b2[0x15]) * cos0;
b1[0x16] = b2[0x16] + b2[0x17];
b1[0x17] = (b2[0x17] - b2[0x16]) * cos0;
b1[0x16] += b1[0x17];
b1[0x14] += b1[0x16];
b1[0x16] += b1[0x15];
b1[0x15] += b1[0x17];
b1[0x18] = b2[0x18] + b2[0x19];
b1[0x19] = (b2[0x18] - b2[0x19]) * cos0;
b1[0x1A] = b2[0x1A] + b2[0x1B];
b1[0x1B] = (b2[0x1B] - b2[0x1A]) * cos0;
b1[0x1A] += b1[0x1B];
b1[0x1C] = b2[0x1C] + b2[0x1D];
b1[0x1D] = (b2[0x1C] - b2[0x1D]) * cos0;
b1[0x1E] = b2[0x1E] + b2[0x1F];
b1[0x1F] = (b2[0x1F] - b2[0x1E]) * cos0;
b1[0x1E] += b1[0x1F];
b1[0x1C] += b1[0x1E];
b1[0x1E] += b1[0x1D];
b1[0x1D] += b1[0x1F];
}
out0[0x10*16] = b1[0x00];
out0[0x10*12] = b1[0x04];
out0[0x10* 8] = b1[0x02];
out0[0x10* 4] = b1[0x06];
out0[0x10* 0] = b1[0x01];
out1[0x10* 0] = b1[0x01];
out1[0x10* 4] = b1[0x05];
out1[0x10* 8] = b1[0x03];
out1[0x10*12] = b1[0x07];
b1[0x08] += b1[0x0C];
out0[0x10*14] = b1[0x08];
b1[0x0C] += b1[0x0a];
out0[0x10*10] = b1[0x0C];
b1[0x0A] += b1[0x0E];
out0[0x10* 6] = b1[0x0A];
b1[0x0E] += b1[0x09];
out0[0x10* 2] = b1[0x0E];
b1[0x09] += b1[0x0D];
out1[0x10* 2] = b1[0x09];
b1[0x0D] += b1[0x0B];
out1[0x10* 6] = b1[0x0D];
b1[0x0B] += b1[0x0F];
out1[0x10*10] = b1[0x0B];
out1[0x10*14] = b1[0x0F];
b1[0x18] += b1[0x1C];
out0[0x10*15] = b1[0x10] + b1[0x18];
out0[0x10*13] = b1[0x18] + b1[0x14];
b1[0x1C] += b1[0x1a];
out0[0x10*11] = b1[0x14] + b1[0x1C];
out0[0x10* 9] = b1[0x1C] + b1[0x12];
b1[0x1A] += b1[0x1E];
out0[0x10* 7] = b1[0x12] + b1[0x1A];
out0[0x10* 5] = b1[0x1A] + b1[0x16];
b1[0x1E] += b1[0x19];
out0[0x10* 3] = b1[0x16] + b1[0x1E];
out0[0x10* 1] = b1[0x1E] + b1[0x11];
b1[0x19] += b1[0x1D];
out1[0x10* 1] = b1[0x11] + b1[0x19];
out1[0x10* 3] = b1[0x19] + b1[0x15];
b1[0x1D] += b1[0x1B];
out1[0x10* 5] = b1[0x15] + b1[0x1D];
out1[0x10* 7] = b1[0x1D] + b1[0x13];
b1[0x1B] += b1[0x1F];
out1[0x10* 9] = b1[0x13] + b1[0x1B];
out1[0x10*11] = b1[0x1B] + b1[0x17];
out1[0x10*13] = b1[0x17] + b1[0x1F];
out1[0x10*15] = b1[0x1F];
}
