void x264_quant_2x2_dc_altivec( int16_t dct[2][2], int mf, int bias )
{
vector bool short mskA;
vec_u32_t i_qbitsv;
vec_u16_t coefvA;
vec_u32_t multEvenvA, multOddvA;
vec_s16_t zerov, one;
vec_s16_t temp1v, temp2v;
vec_u16_t mfv;
vec_u16_t biasv;
vect_ushort_u mf_u;
mf_u.s[0]=mf;
mfv = vec_splat( mf_u.v, 0 );
vect_int_u qbits_u;
qbits_u.s[0]=16;
i_qbitsv = vec_splat(qbits_u.v, 0);
vect_ushort_u bias_u;
bias_u.s[0]=bias;
biasv = vec_splat(bias_u.v, 0);
zerov = vec_splat_s16(0);
one = vec_splat_s16(1);
QUANT_4_U_DC(0);
}
void x264_quant_4x4_altivec( int16_t dct[4][4], uint16_t mf[16], uint16_t bias[16] )
{
vector bool short mskA;
vec_u32_t i_qbitsv;
vec_u16_t coefvA;
vec_u32_t multEvenvA, multOddvA;
vec_u16_t mfvA;
vec_u16_t biasvA;
vec_s16_t zerov, one;
vector bool short mskB;
vec_u16_t coefvB;
vec_u32_t multEvenvB, multOddvB;
vec_u16_t mfvB;
vec_u16_t biasvB;
vec_s16_t temp1v, temp2v;
vect_int_u qbits_u;
qbits_u.s[0]=16;
i_qbitsv = vec_splat(qbits_u.v, 0);
zerov = vec_splat_s16(0);
one = vec_splat_s16(1);
QUANT_16_U( 0, 16 );
}
void x264_quant_8x8_altivec( int16_t dct[8][8], uint16_t mf[64], uint16_t bias[64] )
{
vector bool short mskA;
vec_u32_t i_qbitsv;
vec_u16_t coefvA;
vec_u32_t multEvenvA, multOddvA;
vec_u16_t mfvA;
vec_u16_t biasvA;
vec_s16_t zerov, one;
vector bool short mskB;
vec_u16_t coefvB;
vec_u32_t multEvenvB, multOddvB;
vec_u16_t mfvB;
vec_u16_t biasvB;
vec_s16_t temp1v, temp2v;
vect_int_u qbits_u;
qbits_u.s[0]=16;
i_qbitsv = vec_splat(qbits_u.v, 0);
zerov = vec_splat_s16(0);
one = vec_splat_s16(1);
int i;
for ( i=0; i<4; i++ ) {
QUANT_16_U( i*2*16, i*2*16+16 );
}
}
void ff_vp3_idct_put_altivec(uint8_t *dst, int stride, DCTELEM block[64])
{
vec_u8 t;
IDCT_START
// pixels are signed; so add 128*16 in addition to the normal 8
vec_s16 v2048 = vec_sl(vec_splat_s16(1), vec_splat_u16(11));
eight = vec_add(eight, v2048);
IDCT_1D(NOP, NOP)
TRANSPOSE8(b0, b1, b2, b3, b4, b5, b6, b7);
IDCT_1D(ADD8, SHIFT4)
#define PUT(a)\
t = vec_packsu(a, a);\
vec_ste((vec_u32)t, 0, (unsigned int *)dst);\
vec_ste((vec_u32)t, 4, (unsigned int *)dst);
PUT(b0) dst += stride;
PUT(b1) dst += stride;
PUT(b2) dst += stride;
PUT(b3) dst += stride;
PUT(b4) dst += stride;
PUT(b5) dst += stride;
PUT(b6) dst += stride;
PUT(b7)
}
uint32_t
dequant_mpeg_intra_altivec_c(int16_t * data,
const int16_t * coeff,
const uint32_t quant,
const uint32_t dcscalar,
const uint16_t * mpeg_quant_matrices)
{
register const uint16_t *intra_matrix = get_intra_matrix(mpeg_quant_matrices);
register const int16_t *coeff_ptr = coeff;
register int16_t *data_ptr = data;
register vec_sint16_t ox00;
register vec_sint16_t level;
register vec_sint16_t vec_2048;
register vec_uint16_t vintra;
register vec_uint32_t swap;
register vec_uint32_t even,odd;
register vec_uint32_t et,ot,t;
vec_uint32_t vquant;
vector bool short zero_less;
vector bool short overflow;
#ifdef DEBUG
if((long)data & 0xf)
fprintf(stderr, "xvidcore: error in dequant_mpeg_intra_altivec_c, incorrect align: %x\n", data);
#endif
/* Initialize */
ox00 = vec_splat_s16(0);
*((uint32_t*)&vquant) = quant;
vquant = vec_splat(vquant,0);
swap = vec_rl(vquant, vec_splat_u32(-16));
vec_2048 = (vec_sint16_t)vec_rl(vec_splat_u16(8),vec_splat_u16(8));
DEQUANT_MPEG_INTRA();
DEQUANT_MPEG_INTRA();
DEQUANT_MPEG_INTRA();
DEQUANT_MPEG_INTRA();
DEQUANT_MPEG_INTRA();
DEQUANT_MPEG_INTRA();
DEQUANT_MPEG_INTRA();
DEQUANT_MPEG_INTRA();
/* Process the first */
data[0] = coeff[0] * dcscalar;
if (data[0] < -2048) {
data[0] = -2048;
} else if (data[0] > 2047) {
data[0] = 2047;
}
return 0;
}
uint32_t
dequant_h263_inter_altivec_c(int16_t *data,
int16_t *coeff,
const uint32_t quant,
const uint16_t *mpeg_quant_matrices)
{
vector signed short acLevel;
vector signed short vec_2048;
vector unsigned short quant_m_2;
vector unsigned short quant_add;
vector unsigned short t;
register vector unsigned int even;
register vector unsigned int odd;
register vector unsigned int high;
register vector unsigned int low;
register vector unsigned char zerovec;
vector bool short equal_zero;
vector bool short less_zero;
vector bool short overflow;
#ifdef DEBUG
/* print alignment errors if this is on */
if(((unsigned)data) & 0x15)
fprintf(stderr, "dequant_h263_inter_altivec_c:incorrect align, data: %lx\n", (long)data);
#endif
/* initialize */
*((unsigned short*)&quant_m_2) = (unsigned short)(quant << 1);
quant_m_2 = vec_splat(quant_m_2,0);
*((unsigned short*)&quant_add) = (unsigned short)(quant & 1 ? quant : quant - 1);
quant_add = vec_splat(quant_add,0);
vec_2048 = vec_sl(vec_splat_s16(1), vec_splat_u16(11));
zerovec = vec_splat_u8(0);
/* dequant */
DEQUANT_H263_INTER_ALTIVEC();
DEQUANT_H263_INTER_ALTIVEC();
DEQUANT_H263_INTER_ALTIVEC();
DEQUANT_H263_INTER_ALTIVEC();
DEQUANT_H263_INTER_ALTIVEC();
DEQUANT_H263_INTER_ALTIVEC();
DEQUANT_H263_INTER_ALTIVEC();
DEQUANT_H263_INTER_ALTIVEC();
return 0;
}
/* *************************************************************************
NAME: test_mladd
USAGE:
test_madd();
returns: void
DESCRIPTION:
see how the combination multiply/add operation works
this will work on low order 16-bits
REFERENCES:
Ian Ollmann's Altivec Tutorial
LIMITATIONS:
GLOBAL VARIABLES:
accessed: none
modified: none
FUNCTIONS CALLED:
fprintf
vec_madd - multiply two short vectors and add to the sum a short
all in one operation
REVISION HISTORY:
STR Description of Revision Author
06-Mar-11 initial coding kaj
************************************************************************* */
void test_mladd(void)
{
vector unsigned short shortVector1 =
{ 0, 2, 4, 8, 16, 32, 64, 128 };
vector unsigned short addVector;
vector unsigned short coeffVector;
vector unsigned short resultVector;
vector short shortVector2 =
{ -128, -64, -32, -16, 0, 16, 32, 64};
vector short addVector2 =
{ -10, -10, -10, -10, 0, 10, 10, 10};
vector short coeffVector2;
vector short resultVector2;
short printshort[SHORT_ARRAYSIZE] __attribute__ ((aligned (16)));
coeffVector = vec_splat_u16(2);
addVector = vec_splat_u16(0);
/* print vectors performing mladd on */
fprintf(stderr,"-----------------------------------------------------------"
"\n\n");
printVecUShorts("vec_mladd unsigned input vector 1", shortVector1,
SHORT_ARRAYSIZE);
printVecUShorts("vec_mladd unsigned input vector to add", addVector,
SHORT_ARRAYSIZE);
printVecUShorts("vec_mladd unsigned coeffvector to multiply", coeffVector,
SHORT_ARRAYSIZE);
/* calculate */
resultVector = vec_mladd(shortVector1,coeffVector,addVector);
printVecUShorts("vec_mladd vector (Input*2+0)", resultVector,
SHORT_ARRAYSIZE);
/* signed shorts */
coeffVector2 = vec_splat_s16(2);
/* print signed short vectors performing mladd on */
fprintf(stderr,"----------------------------------------------------------"
"\n\n");
printVecShorts("vec_mladd signed input vector 1", shortVector2,
SHORT_ARRAYSIZE);
printVecShorts("vec_mladd signed input vector to add", addVector2,
SHORT_ARRAYSIZE);
printVecShorts("vec_mladd signed coeffvector to multiply", coeffVector2,
SHORT_ARRAYSIZE);
/* calculate */
resultVector2 = vec_mladd(shortVector2,coeffVector2,addVector2);
printVecShorts("vec_mladd vector (Input*2 + 10(increment pos & neg by 10)",
resultVector2,SHORT_ARRAYSIZE);
} /* test_mladd */
void x264_add8x8_idct_dc_altivec( uint8_t *p_dst, int16_t dct[4] )
{
vec_s16_t dcv;
vec_s16_t v32 = vec_sl( vec_splat_s16( 8 ), vec_splat_u16( 2 ) );
vec_u16_t v6 = vec_splat_u16( 6 );
vec_s16_t dctv = vec_vsx_ld( 0, dct );
dctv = vec_sra( vec_add( dctv, v32 ), v6 );
dcv = (vec_s16_t)vec_mergeh( (vec_s32_t)vec_splat( dctv, 0 ), (vec_s32_t)vec_splat( dctv, 1 ) );
dcv = (vec_s16_t)vec_mergeh( (vec_s32_t)dcv, (vec_s32_t)dcv );
idct8_dc_altivec( &p_dst[0], dcv );
dcv = (vec_s16_t)vec_mergeh( (vec_s32_t)vec_splat( dctv, 2 ), (vec_s32_t)vec_splat( dctv, 3 ) );
dcv = (vec_s16_t)vec_mergeh( (vec_s32_t)dcv, (vec_s32_t)dcv );
idct8_dc_altivec( &p_dst[4*FDEC_STRIDE+0], dcv );
}
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);
DECLARE_ALIGNED_16(signed int, ABCD[4]) =
{((8 - x) * (8 - y)),
(( x) * (8 - y)),
((8 - x) * ( y)),
(( x) * ( y))};
register int i;
vec_u8 fperm;
const vec_s32 vABCD = vec_ld(0, ABCD);
const vec_s16 vA = vec_splat((vec_s16)vABCD, 1);
const vec_s16 vB = vec_splat((vec_s16)vABCD, 3);
const vec_s16 vC = vec_splat((vec_s16)vABCD, 5);
const vec_s16 vD = vec_splat((vec_s16)vABCD, 7);
LOAD_ZERO;
const vec_s16 v32ss = vec_sl(vec_splat_s16(1),vec_splat_u16(5));
const vec_u16 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;
vec_u8 vsrcAuc, av_uninit(vsrcBuc), vsrcperm0, vsrcperm1;
vec_u8 vsrc0uc, vsrc1uc;
vec_s16 vsrc0ssH, vsrc1ssH;
vec_u8 vsrcCuc, vsrc2uc, vsrc3uc;
vec_s16 vsrc2ssH, vsrc3ssH, psum;
vec_u8 vdst, ppsum, vfdst, fsum;
POWERPC_PERF_START_COUNT(PREFIX_h264_chroma_mc8_num, 1);
if (((unsigned long)dst) % 16 == 0) {
fperm = (vec_u8){0x10, 0x11, 0x12, 0x13,
0x14, 0x15, 0x16, 0x17,
0x08, 0x09, 0x0A, 0x0B,
0x0C, 0x0D, 0x0E, 0x0F};
} else {
fperm = (vec_u8){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 = (vec_s16)vec_mergeh(zero_u8v,(vec_u8)vsrc0uc);
vsrc1ssH = (vec_s16)vec_mergeh(zero_u8v,(vec_u8)vsrc1uc);
if (ABCD[3]) {
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);
CHROMA_MC8_ALTIVEC_CORE
}
} else {
vec_u8 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);
CHROMA_MC8_ALTIVEC_CORE
}
}
} else {
/* 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);
DECLARE_ALIGNED_16(signed int, ABCD[4]) =
{((8 - x) * (8 - y)),
((x) * (8 - y)),
((8 - x) * (y)),
((x) * (y))};
register int i;
vec_u8_t fperm;
const vec_s32_t vABCD = vec_ld(0, ABCD);
const vec_s16_t vA = vec_splat((vec_s16_t)vABCD, 1);
const vec_s16_t vB = vec_splat((vec_s16_t)vABCD, 3);
const vec_s16_t vC = vec_splat((vec_s16_t)vABCD, 5);
const vec_s16_t vD = vec_splat((vec_s16_t)vABCD, 7);
LOAD_ZERO;
const vec_s16_t v32ss = vec_sl(vec_splat_s16(1),vec_splat_u16(5));
const vec_u16_t 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;
vec_u8_t vsrcAuc, vsrcBuc, vsrcperm0, vsrcperm1;
vec_u8_t vsrc0uc, vsrc1uc;
vec_s16_t vsrc0ssH, vsrc1ssH;
vec_u8_t vsrcCuc, vsrc2uc, vsrc3uc;
vec_s16_t vsrc2ssH, vsrc3ssH, psum;
vec_u8_t vdst, ppsum, vfdst, fsum;
POWERPC_PERF_START_COUNT(PREFIX_h264_chroma_mc8_num, 1);
if (((unsigned long)dst) % 16 == 0) {
fperm = (vec_u8_t)AVV(0x10, 0x11, 0x12, 0x13,
0x14, 0x15, 0x16, 0x17,
0x08, 0x09, 0x0A, 0x0B,
0x0C, 0x0D, 0x0E, 0x0F);
} else {
fperm = (vec_u8_t)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 = (vec_s16_t)vec_mergeh(zero_u8v,(vec_u8_t)vsrc0uc);
vsrc1ssH = (vec_s16_t)vec_mergeh(zero_u8v,(vec_u8_t)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 = (vec_s16_t)vec_mergeh(zero_u8v,(vec_u8_t)vsrc2uc);
vsrc3ssH = (vec_s16_t)vec_mergeh(zero_u8v,(vec_u8_t)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 = (vec_u8_t)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 {
vec_u8_t 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 = (vec_s16_t)vec_mergeh(zero_u8v,(vec_u8_t)vsrc2uc);
vsrc3ssH = (vec_s16_t)vec_mergeh(zero_u8v,(vec_u8_t)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 = (vec_u8_t)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);
}
/* AltiVec version of dct_unquantize_h263
this code assumes `block' is 16 bytes-aligned */
static void dct_unquantize_h263_altivec(MpegEncContext *s,
DCTELEM *block, int n, int qscale)
{
int i, level, qmul, qadd;
int nCoeffs;
assert(s->block_last_index[n]>=0);
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 always 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);
DECLARE_ALIGNED(16, short, qmul8) = qmul;
DECLARE_ALIGNED(16, short, qadd8) = 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_splat((vec_s16)vec_lde(0, &qmul8), 0);
qaddv = vec_splat((vec_s16)vec_lde(0, &qadd8), 0);
nqaddv = vec_sub(vczero, qaddv);
#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;
}
}
}
/* 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);
POWERPC_PERF_START_COUNT(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 = (const vector signed short)AVV(20);
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 = (const vector signed int)AVV(512);
const vector unsigned int v16ui = (const vector unsigned int)AVV(16);
register int align = ((((unsigned long)src) - 2) % 16);
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;
}
const vector signed short srcP0A = (vector signed short)vec_mergeh((vector unsigned char)vzero, srcP0);
const vector signed short srcP0B = (vector signed short)vec_mergel((vector unsigned char)vzero, srcP0);
const vector signed short srcP1A = (vector signed short)vec_mergeh((vector unsigned char)vzero, srcP1);
const vector signed short srcP1B = (vector signed short)vec_mergel((vector unsigned char)vzero, srcP1);
const vector signed short srcP2A = (vector signed short)vec_mergeh((vector unsigned char)vzero, srcP2);
const vector signed short srcP2B = (vector signed short)vec_mergel((vector unsigned char)vzero, srcP2);
const vector signed short srcP3A = (vector signed short)vec_mergeh((vector unsigned char)vzero, srcP3);
const vector signed short srcP3B = (vector signed short)vec_mergel((vector unsigned char)vzero, srcP3);
const vector signed short srcM1A = (vector signed short)vec_mergeh((vector unsigned char)vzero, srcM1);
const vector signed short srcM1B = (vector signed short)vec_mergel((vector unsigned char)vzero, srcM1);
const vector signed short srcM2A = (vector signed short)vec_mergeh((vector unsigned char)vzero, srcM2);
const vector signed short srcM2B = (vector signed short)vec_mergel((vector unsigned char)vzero, srcM2);
const vector signed short sum1A = vec_adds(srcP0A, srcP1A);
const vector signed short sum1B = vec_adds(srcP0B, srcP1B);
const vector signed short sum2A = vec_adds(srcM1A, srcP2A);
const vector signed short sum2B = vec_adds(srcM1B, srcP2B);
const vector signed short sum3A = vec_adds(srcM2A, srcP3A);
const vector signed short sum3B = vec_adds(srcM2B, srcP3B);
const vector signed short pp1A = vec_mladd(sum1A, v20ss, sum3A);
const vector signed short pp1B = vec_mladd(sum1B, v20ss, sum3B);
const vector signed short pp2A = vec_mladd(sum2A, v5ss, (vector signed short)vzero);
const vector signed short pp2B = vec_mladd(sum2B, v5ss, (vector signed short)vzero);
const vector signed short psumA = vec_sub(pp1A, pp2A);
const vector signed short 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 */
}
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);
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 - (tmpStride * 21);
vector signed short tmpM2ssA = vec_ld(0, tmpbis);
vector signed short tmpM2ssB = vec_ld(16, tmpbis);
tmpbis += tmpStride;
vector signed short tmpM1ssA = vec_ld(0, tmpbis);
vector signed short tmpM1ssB = vec_ld(16, tmpbis);
tmpbis += tmpStride;
vector signed short tmpP0ssA = vec_ld(0, tmpbis);
vector signed short tmpP0ssB = vec_ld(16, tmpbis);
tmpbis += tmpStride;
vector signed short tmpP1ssA = vec_ld(0, tmpbis);
vector signed short tmpP1ssB = vec_ld(16, tmpbis);
tmpbis += tmpStride;
vector signed short tmpP2ssA = vec_ld(0, tmpbis);
vector signed short 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);
tmpbis += tmpStride;
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);
tmpM2ssA = tmpM1ssA;
tmpM2ssB = tmpM1ssB;
tmpM1ssA = tmpP0ssA;
tmpM1ssB = tmpP0ssB;
tmpP0ssA = tmpP1ssA;
tmpP0ssB = tmpP1ssB;
tmpP1ssA = tmpP2ssA;
tmpP1ssB = tmpP2ssB;
tmpP2ssA = tmpP3ssA;
tmpP2ssB = tmpP3ssB;
const vector signed int pp1Ae = vec_mule(sum1A, v20ss);
const vector signed int pp1Ao = vec_mulo(sum1A, v20ss);
const vector signed int pp1Be = vec_mule(sum1B, v20ss);
const vector signed int pp1Bo = vec_mulo(sum1B, v20ss);
const vector signed int pp2Ae = vec_mule(sum2A, v5ss);
const vector signed int pp2Ao = vec_mulo(sum2A, v5ss);
const vector signed int pp2Be = vec_mule(sum2B, v5ss);
const vector signed int pp2Bo = vec_mulo(sum2B, v5ss);
const vector signed int pp3Ae = vec_sra((vector signed int)sum3A, v16ui);
const vector signed int pp3Ao = vec_mulo(sum3A, v1ss);
const vector signed int pp3Be = vec_sra((vector signed int)sum3B, v16ui);
const vector signed int pp3Bo = vec_mulo(sum3B, v1ss);
const vector signed int pp1cAe = vec_add(pp1Ae, v512si);
const vector signed int pp1cAo = vec_add(pp1Ao, v512si);
const vector signed int pp1cBe = vec_add(pp1Be, v512si);
const vector signed int pp1cBo = vec_add(pp1Bo, v512si);
const vector signed int pp32Ae = vec_sub(pp3Ae, pp2Ae);
const vector signed int pp32Ao = vec_sub(pp3Ao, pp2Ao);
const vector signed int pp32Be = vec_sub(pp3Be, pp2Be);
const vector signed int pp32Bo = vec_sub(pp3Bo, pp2Bo);
const vector signed int sumAe = vec_add(pp1cAe, pp32Ae);
const vector signed int sumAo = vec_add(pp1cAo, pp32Ao);
const vector signed int sumBe = vec_add(pp1cBe, pp32Be);
const vector signed int sumBo = vec_add(pp1cBo, pp32Bo);
const vector signed int ssumAe = vec_sra(sumAe, v10ui);
const vector signed int ssumAo = vec_sra(sumAo, v10ui);
const vector signed int ssumBe = vec_sra(sumBe, v10ui);
const vector signed int ssumBo = vec_sra(sumBo, v10ui);
const vector signed short ssume = vec_packs(ssumAe, ssumBe);
const vector signed short ssumo = vec_packs(ssumAo, ssumBo);
const vector unsigned char sumv = vec_packsu(ssume, ssumo);
const vector unsigned char sum = vec_perm(sumv, sumv, mperm);
const vector unsigned char dst1 = vec_ld(0, dst);
const vector unsigned char dst2 = vec_ld(16, dst);
const vector unsigned char vdst = vec_perm(dst1, dst2, vec_lvsl(0, dst));
vector unsigned char fsum;
OP_U8_ALTIVEC(fsum, sum, vdst);
const vector unsigned char rsum = vec_perm(fsum, fsum, dstperm);
const vector unsigned char fdst1 = vec_sel(dst1, rsum, dstmask);
const vector unsigned char 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);
}
/* 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);
POWERPC_PERF_START_COUNT(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 = (const vector signed short)AVV(20);
const vector unsigned short v5us = vec_splat_u16(5);
const vector signed short v5ss = vec_splat_s16(5);
const vector signed short v16ss = (const vector signed short)AVV(16);
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);
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);
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);
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);
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);
for (i = 0 ; i < 16 ; i++) {
const vector unsigned char srcP3a = vec_ld(0, srcbis);
const vector unsigned char srcP3b = vec_ld(16, srcbis);
const vector unsigned char srcP3 = vec_perm(srcP3a, srcP3b, perm);
const vector signed short srcP3ssA = (vector signed short)vec_mergeh((vector unsigned char)vzero, srcP3);
const vector signed short srcP3ssB = (vector signed short)vec_mergel((vector unsigned char)vzero, srcP3);
srcbis += srcStride;
const vector signed short sum1A = vec_adds(srcP0ssA, srcP1ssA);
const vector signed short sum1B = vec_adds(srcP0ssB, srcP1ssB);
const vector signed short sum2A = vec_adds(srcM1ssA, srcP2ssA);
const vector signed short sum2B = vec_adds(srcM1ssB, srcP2ssB);
const vector signed short sum3A = vec_adds(srcM2ssA, srcP3ssA);
const vector signed short 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;
const vector signed short pp1A = vec_mladd(sum1A, v20ss, v16ss);
const vector signed short pp1B = vec_mladd(sum1B, v20ss, v16ss);
const vector signed short pp2A = vec_mladd(sum2A, v5ss, (vector signed short)vzero);
const vector signed short pp2B = vec_mladd(sum2B, v5ss, (vector signed short)vzero);
const vector signed short pp3A = vec_add(sum3A, pp1A);
const vector signed short pp3B = vec_add(sum3B, pp1B);
const vector signed short psumA = vec_sub(pp3A, pp2A);
const vector signed short psumB = vec_sub(pp3B, pp2B);
const vector signed short sumA = vec_sra(psumA, v5us);
const vector signed short sumB = vec_sra(psumB, v5us);
const vector unsigned char sum = vec_packsu(sumA, sumB);
const vector unsigned char dst1 = vec_ld(0, dst);
const vector unsigned char dst2 = vec_ld(16, dst);
const vector unsigned char vdst = vec_perm(dst1, dst2, vec_lvsl(0, dst));
vector unsigned char fsum;
OP_U8_ALTIVEC(fsum, sum, vdst);
const vector unsigned char rsum = vec_perm(fsum, fsum, dstperm);
const vector unsigned char fdst1 = vec_sel(dst1, rsum, dstmask);
const vector unsigned char 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);
}
void
gimp_composite_dodge_rgba8_rgba8_rgba8_altivec (GimpCompositeContext *ctx)
{
const guchar *A = ctx->A;
const guchar *B = ctx->B;
guchar *D = ctx->D;
guint length = ctx->n_pixels;
vector unsigned char a,b,d;
vector unsigned char alpha_a,alpha_b,alpha;
vector signed short ox0001=vec_splat_s16(1);
union
{
vector signed short v;
vector unsigned short vu;
gushort u16[8];
} ah,al,bh,bl;
while (length >= 4)
{
a=LoadUnaligned(A);
b=LoadUnaligned(B);
alpha_a=vec_and(a, alphamask);
alpha_b=vec_and(b, alphamask);
alpha=vec_min(alpha_a, alpha_b);
ah.v=vec_unpackh((vector signed char)a);
ah.v=vec_sl(ah.v,ox0008);
al.v=vec_unpackl((vector signed char)a);
al.v=vec_sl(al.v,ox0008);
b=vec_nor(b,b);
bh.v=vec_unpackh((vector signed char)b);
bh.v=vec_and(bh.v,ox00ff);
bh.v=vec_add(bh.v,ox0001);
bl.v=vec_unpackl((vector signed char)b);
bl.v=vec_and(bl.v,ox00ff);
bl.v=vec_add(bl.v,ox0001);
ah.u16[0]=ah.u16[0]/bh.u16[0];
ah.u16[1]=ah.u16[1]/bh.u16[1];
ah.u16[2]=ah.u16[2]/bh.u16[2];
ah.u16[4]=ah.u16[4]/bh.u16[4];
ah.u16[5]=ah.u16[5]/bh.u16[5];
ah.u16[6]=ah.u16[6]/bh.u16[6];
al.u16[0]=al.u16[0]/bl.u16[0];
al.u16[1]=al.u16[1]/bl.u16[1];
al.u16[2]=al.u16[2]/bl.u16[2];
al.u16[4]=al.u16[4]/bl.u16[4];
al.u16[5]=al.u16[5]/bl.u16[5];
al.u16[6]=al.u16[6]/bl.u16[6];
d=vec_packs(ah.vu,al.vu);
d=vec_andc(d, alphamask);
d=vec_or(d, alpha);
StoreUnaligned(d, D);
A+=16;
B+=16;
D+=16;
length-=4;
}
length = length*4;
a=LoadUnalignedLess(A, length);
b=LoadUnalignedLess(B, length);
alpha_a=vec_and(a, alphamask);
alpha_b=vec_and(b, alphamask);
alpha=vec_min(alpha_a, alpha_b);
ah.v=vec_unpackh((vector signed char)a);
ah.v=vec_sl(ah.v,ox0008);
al.v=vec_unpackl((vector signed char)a);
al.v=vec_sl(al.v,ox0008);
b=vec_nor(b,b);
bh.v=vec_unpackh((vector signed char)b);
bh.v=vec_and(bh.v,ox00ff);
bh.v=vec_add(bh.v,ox0001);
bl.v=vec_unpackl((vector signed char)b);
bl.v=vec_and(bl.v,ox00ff);
bl.v=vec_add(bl.v,ox0001);
ah.u16[0]=ah.u16[0]/bh.u16[0];
ah.u16[1]=ah.u16[1]/bh.u16[1];
ah.u16[2]=ah.u16[2]/bh.u16[2];
ah.u16[4]=ah.u16[4]/bh.u16[4];
ah.u16[5]=ah.u16[5]/bh.u16[5];
ah.u16[6]=ah.u16[6]/bh.u16[6];
al.u16[0]=al.u16[0]/bl.u16[0];
al.u16[1]=al.u16[1]/bl.u16[1];
al.u16[2]=al.u16[2]/bl.u16[2];
al.u16[4]=al.u16[4]/bl.u16[4];
al.u16[5]=al.u16[5]/bl.u16[5];
al.u16[6]=al.u16[6]/bl.u16[6];
d=vec_packs(ah.vu,al.vu);
d=vec_andc(d, alphamask);
d=vec_or(d, alpha);
StoreUnalignedLess(d, D, length);
}
static void PREFIX_h264_qpel16_hv_lowpass_altivec(uint8_t * dst, int16_t * tmp, uint8_t * src, int dstStride, int tmpStride, int srcStride) {
register int i;
LOAD_ZERO;
const vec_u8 permM2 = vec_lvsl(-2, src);
const vec_u8 permM1 = vec_lvsl(-1, src);
const vec_u8 permP0 = vec_lvsl(+0, src);
const vec_u8 permP1 = vec_lvsl(+1, src);
const vec_u8 permP2 = vec_lvsl(+2, src);
const vec_u8 permP3 = vec_lvsl(+3, src);
const vec_s16 v20ss = vec_sl(vec_splat_s16(5),vec_splat_u16(2));
const vec_u32 v10ui = vec_splat_u32(10);
const vec_s16 v5ss = vec_splat_s16(5);
const vec_s16 v1ss = vec_splat_s16(1);
const vec_s32 v512si = vec_sl(vec_splat_s32(1),vec_splat_u32(9));
const vec_u32 v16ui = vec_sl(vec_splat_u32(1),vec_splat_u32(4));
register int align = ((((unsigned long)src) - 2) % 16);
vec_s16 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 vec_u8 mperm = (const vec_u8)
{0x00, 0x08, 0x01, 0x09, 0x02, 0x0A, 0x03, 0x0B,
0x04, 0x0C, 0x05, 0x0D, 0x06, 0x0E, 0x07, 0x0F};
int16_t *tmpbis = tmp;
vec_s16 tmpM1ssA, tmpM1ssB, tmpM2ssA, tmpM2ssB,
tmpP0ssA, tmpP0ssB, tmpP1ssA, tmpP1ssB,
tmpP2ssA, tmpP2ssB;
vec_s32 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;
vec_u8 fsum, sumv, sum;
vec_s16 ssume, ssumo;
src -= (2 * srcStride);
for (i = 0 ; i < 21 ; i ++) {
vec_u8 srcM2, srcM1, srcP0, srcP1, srcP2, srcP3;
vec_u8 srcR1 = vec_ld(-2, src);
vec_u8 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: {
vec_u8 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: {
vec_u8 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: {
vec_u8 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: {
vec_u8 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 = (vec_s16) vec_mergeh(zero_u8v, srcP0);
srcP0B = (vec_s16) vec_mergel(zero_u8v, srcP0);
srcP1A = (vec_s16) vec_mergeh(zero_u8v, srcP1);
srcP1B = (vec_s16) vec_mergel(zero_u8v, srcP1);
srcP2A = (vec_s16) vec_mergeh(zero_u8v, srcP2);
srcP2B = (vec_s16) vec_mergel(zero_u8v, srcP2);
srcP3A = (vec_s16) vec_mergeh(zero_u8v, srcP3);
srcP3B = (vec_s16) vec_mergel(zero_u8v, srcP3);
srcM1A = (vec_s16) vec_mergeh(zero_u8v, srcM1);
srcM1B = (vec_s16) vec_mergel(zero_u8v, srcM1);
srcM2A = (vec_s16) vec_mergeh(zero_u8v, srcM2);
srcM2B = (vec_s16) vec_mergel(zero_u8v, 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, zero_s16v);
pp2B = vec_mladd(sum2B, v5ss, zero_s16v);
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 vec_s16 tmpP3ssA = vec_ld(0, tmpbis);
const vec_s16 tmpP3ssB = vec_ld(16, tmpbis);
const vec_s16 sum1A = vec_adds(tmpP0ssA, tmpP1ssA);
const vec_s16 sum1B = vec_adds(tmpP0ssB, tmpP1ssB);
const vec_s16 sum2A = vec_adds(tmpM1ssA, tmpP2ssA);
const vec_s16 sum2B = vec_adds(tmpM1ssB, tmpP2ssB);
const vec_s16 sum3A = vec_adds(tmpM2ssA, tmpP3ssA);
const vec_s16 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((vec_s32)sum3A, v16ui);
pp3Ao = vec_mulo(sum3A, v1ss);
pp3Be = vec_sra((vec_s32)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);
ASSERT_ALIGNED(dst);
OP_U8_ALTIVEC(fsum, sum, vec_ld(0, dst));
vec_st(fsum, 0, dst);
dst += dstStride;
}
}
static av_always_inline
void put_vp8_epel_v_altivec_core(uint8_t *dst, ptrdiff_t dst_stride,
uint8_t *src, ptrdiff_t src_stride,
int h, int my, int w, int is6tap)
{
LOAD_V_SUBPEL_FILTER(my-1);
vec_u8 s0, s1, s2, s3, s4, s5, filt, align_vech, perm_vec, align_vecl;
vec_s16 s0f, s1f, s2f, s3f, s4f, s5f, f16h, f16l;
vec_s16 c64 = vec_sl(vec_splat_s16(1), vec_splat_u16(6));
vec_u16 c7 = vec_splat_u16(7);
// we want pixels 0-7 to be in the even positions and 8-15 in the odd,
// so combine this permute with the alignment permute vector
align_vech = vec_lvsl(0, src);
align_vecl = vec_sld(align_vech, align_vech, 8);
if (w ==16)
perm_vec = vec_mergeh(align_vech, align_vecl);
else
perm_vec = vec_mergeh(align_vech, align_vech);
if (is6tap)
s0 = load_with_perm_vec(-2*src_stride, src, perm_vec);
s1 = load_with_perm_vec(-1*src_stride, src, perm_vec);
s2 = load_with_perm_vec( 0*src_stride, src, perm_vec);
s3 = load_with_perm_vec( 1*src_stride, src, perm_vec);
if (is6tap)
s4 = load_with_perm_vec( 2*src_stride, src, perm_vec);
src += (2+is6tap)*src_stride;
while (h --> 0) {
if (is6tap)
s5 = load_with_perm_vec(0, src, perm_vec);
else
s4 = load_with_perm_vec(0, src, perm_vec);
FILTER_V(f16h, vec_mule);
if (w == 16) {
FILTER_V(f16l, vec_mulo);
filt = vec_packsu(f16h, f16l);
vec_st(filt, 0, dst);
} else {
filt = vec_packsu(f16h, f16h);
if (w == 4)
filt = (vec_u8)vec_splat((vec_u32)filt, 0);
else
vec_ste((vec_u32)filt, 4, (uint32_t*)dst);
vec_ste((vec_u32)filt, 0, (uint32_t*)dst);
}
if (is6tap)
s0 = s1;
s1 = s2;
s2 = s3;
s3 = s4;
if (is6tap)
s4 = s5;
dst += dst_stride;
src += src_stride;
}
}
/** Do inverse transform on 8x8 block
*/
static void vc1_inv_trans_8x8_altivec(DCTELEM block[64],
int sign, int rangered)
{
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_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);
const vector unsigned short rangered_shift = vec_splat_u16(1);
const vector signed short signed_bias = vec_sl(vec_splat_s16(4),
vec_splat_u16(4));
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);
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);
if (rangered)
{
if (!sign)
{
src0 = vec_sub(src0, signed_bias);
src1 = vec_sub(src1, signed_bias);
src2 = vec_sub(src2, signed_bias);
src3 = vec_sub(src3, signed_bias);
src4 = vec_sub(src4, signed_bias);
src5 = vec_sub(src5, signed_bias);
src6 = vec_sub(src6, signed_bias);
src7 = vec_sub(src7, signed_bias);
}
src0 = vec_sl(src0, rangered_shift);
src1 = vec_sl(src1, rangered_shift);
src2 = vec_sl(src2, rangered_shift);
src3 = vec_sl(src3, rangered_shift);
src4 = vec_sl(src4, rangered_shift);
src5 = vec_sl(src5, rangered_shift);
src6 = vec_sl(src6, rangered_shift);
src7 = vec_sl(src7, rangered_shift);
}
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);
}
void v_resample16_altivec(uint8_t *dst, int dst_width, const uint8_t *src,
int wrap, int16_t *filter)
{
int sum, i;
const uint8_t *s;
vector unsigned char *tv, tmp, dstv, zero;
vec_ss_t srchv[4], srclv[4], fv[4];
vector signed short zeros, sumhv, sumlv;
s = src;
for(i=0;i<4;i++) {
/*
The vec_madds later on does an implicit >>15 on the result.
Since FILTER_BITS is 8, and we have 15 bits of magnitude in
a signed short, we have just enough bits to pre-shift our
filter constants <<7 to compensate for vec_madds.
*/
fv[i].s[0] = filter[i] << (15-FILTER_BITS);
fv[i].v = vec_splat(fv[i].v, 0);
}
zero = vec_splat_u8(0);
zeros = vec_splat_s16(0);
/*
When we're resampling, we'd ideally like both our input buffers,
and output buffers to be 16-byte aligned, so we can do both aligned
reads and writes. Sadly we can't always have this at the moment, so
we opt for aligned writes, as unaligned writes have a huge overhead.
To do this, do enough scalar resamples to get dst 16-byte aligned.
*/
i = (-(int)dst) & 0xf;
while(i>0) {
sum = s[0 * wrap] * filter[0] +
s[1 * wrap] * filter[1] +
s[2 * wrap] * filter[2] +
s[3 * wrap] * filter[3];
sum = sum >> FILTER_BITS;
if (sum<0) sum = 0; else if (sum>255) sum=255;
dst[0] = sum;
dst++;
s++;
dst_width--;
i--;
}
/* Do our altivec resampling on 16 pixels at once. */
while(dst_width>=16) {
/* Read 16 (potentially unaligned) bytes from each of
4 lines into 4 vectors, and split them into shorts.
Interleave the multipy/accumulate for the resample
filter with the loads to hide the 3 cycle latency
the vec_madds have. */
tv = (vector unsigned char *) &s[0 * wrap];
tmp = vec_perm(tv[0], tv[1], vec_lvsl(0, &s[i * wrap]));
srchv[0].v = (vector signed short) vec_mergeh(zero, tmp);
srclv[0].v = (vector signed short) vec_mergel(zero, tmp);
sumhv = vec_madds(srchv[0].v, fv[0].v, zeros);
sumlv = vec_madds(srclv[0].v, fv[0].v, zeros);
tv = (vector unsigned char *) &s[1 * wrap];
tmp = vec_perm(tv[0], tv[1], vec_lvsl(0, &s[1 * wrap]));
srchv[1].v = (vector signed short) vec_mergeh(zero, tmp);
srclv[1].v = (vector signed short) vec_mergel(zero, tmp);
sumhv = vec_madds(srchv[1].v, fv[1].v, sumhv);
sumlv = vec_madds(srclv[1].v, fv[1].v, sumlv);
tv = (vector unsigned char *) &s[2 * wrap];
tmp = vec_perm(tv[0], tv[1], vec_lvsl(0, &s[2 * wrap]));
srchv[2].v = (vector signed short) vec_mergeh(zero, tmp);
srclv[2].v = (vector signed short) vec_mergel(zero, tmp);
sumhv = vec_madds(srchv[2].v, fv[2].v, sumhv);
sumlv = vec_madds(srclv[2].v, fv[2].v, sumlv);
tv = (vector unsigned char *) &s[3 * wrap];
tmp = vec_perm(tv[0], tv[1], vec_lvsl(0, &s[3 * wrap]));
srchv[3].v = (vector signed short) vec_mergeh(zero, tmp);
srclv[3].v = (vector signed short) vec_mergel(zero, tmp);
sumhv = vec_madds(srchv[3].v, fv[3].v, sumhv);
sumlv = vec_madds(srclv[3].v, fv[3].v, sumlv);
/* Pack the results into our destination vector,
and do an aligned write of that back to memory. */
dstv = vec_packsu(sumhv, sumlv) ;
vec_st(dstv, 0, (vector unsigned char *) dst);
dst+=16;
s+=16;
dst_width-=16;
}
/* If there are any leftover pixels, resample them
with the slow scalar method. */
while(dst_width>0) {
sum = s[0 * wrap] * filter[0] +
s[1 * wrap] * filter[1] +
s[2 * wrap] * filter[2] +
s[3 * wrap] * filter[3];
sum = sum >> FILTER_BITS;
if (sum<0) sum = 0; else if (sum>255) sum=255;
dst[0] = sum;
dst++;
s++;
dst_width--;
}
}
static void PREFIX_h264_qpel16_v_lowpass_altivec(uint8_t *dst,
const uint8_t *src,
int dstStride, int srcStride)
{
register int i;
LOAD_ZERO;
vec_u8 perm;
#if HAVE_BIGENDIAN
perm = vec_lvsl(0, src);
#endif
const vec_s16 v20ss = vec_sl(vec_splat_s16(5),vec_splat_u16(2));
const vec_u16 v5us = vec_splat_u16(5);
const vec_s16 v5ss = vec_splat_s16(5);
const vec_s16 v16ss = vec_sl(vec_splat_s16(1),vec_splat_u16(4));
const uint8_t *srcbis = src - (srcStride * 2);
const vec_u8 srcM2 = load_with_perm_vec(0, srcbis, perm);
srcbis += srcStride;
const vec_u8 srcM1 = load_with_perm_vec(0, srcbis, perm);
srcbis += srcStride;
const vec_u8 srcP0 = load_with_perm_vec(0, srcbis, perm);
srcbis += srcStride;
const vec_u8 srcP1 = load_with_perm_vec(0, srcbis, perm);
srcbis += srcStride;
const vec_u8 srcP2 = load_with_perm_vec(0, srcbis, perm);
srcbis += srcStride;
vec_s16 srcM2ssA = (vec_s16) VEC_MERGEH(zero_u8v, srcM2);
vec_s16 srcM2ssB = (vec_s16) VEC_MERGEL(zero_u8v, srcM2);
vec_s16 srcM1ssA = (vec_s16) VEC_MERGEH(zero_u8v, srcM1);
vec_s16 srcM1ssB = (vec_s16) VEC_MERGEL(zero_u8v, srcM1);
vec_s16 srcP0ssA = (vec_s16) VEC_MERGEH(zero_u8v, srcP0);
vec_s16 srcP0ssB = (vec_s16) VEC_MERGEL(zero_u8v, srcP0);
vec_s16 srcP1ssA = (vec_s16) VEC_MERGEH(zero_u8v, srcP1);
vec_s16 srcP1ssB = (vec_s16) VEC_MERGEL(zero_u8v, srcP1);
vec_s16 srcP2ssA = (vec_s16) VEC_MERGEH(zero_u8v, srcP2);
vec_s16 srcP2ssB = (vec_s16) VEC_MERGEL(zero_u8v, srcP2);
vec_s16 pp1A, pp1B, pp2A, pp2B, pp3A, pp3B,
psumA, psumB, sumA, sumB,
srcP3ssA, srcP3ssB,
sum1A, sum1B, sum2A, sum2B, sum3A, sum3B;
vec_u8 sum, fsum, srcP3;
for (i = 0 ; i < 16 ; i++) {
srcP3 = load_with_perm_vec(0, srcbis, perm);
srcbis += srcStride;
srcP3ssA = (vec_s16) VEC_MERGEH(zero_u8v, srcP3);
srcP3ssB = (vec_s16) VEC_MERGEL(zero_u8v, srcP3);
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, zero_s16v);
pp2B = vec_mladd(sum2B, v5ss, zero_s16v);
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);
ASSERT_ALIGNED(dst);
OP_U8_ALTIVEC(fsum, sum, vec_ld(0, dst));
vec_st(fsum, 0, dst);
dst += dstStride;
}
}
static void PREFIX_h264_qpel16_h_lowpass_altivec(uint8_t *dst,
const uint8_t *src,
int dstStride, int srcStride)
{
register int i;
LOAD_ZERO;
vec_u8 permM2, permM1, permP0, permP1, permP2, permP3;
const vec_s16 v5ss = vec_splat_s16(5);
const vec_u16 v5us = vec_splat_u16(5);
const vec_s16 v20ss = vec_sl(vec_splat_s16(5),vec_splat_u16(2));
const vec_s16 v16ss = vec_sl(vec_splat_s16(1),vec_splat_u16(4));
vec_u8 srcM2, srcM1, srcP0, srcP1, srcP2, srcP3;
register int align = ((((unsigned long)src) - 2) % 16);
vec_s16 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;
vec_u8 sum, fsum;
#if HAVE_BIGENDIAN
permM2 = vec_lvsl(-2, src);
permM1 = vec_lvsl(-1, src);
permP0 = vec_lvsl(+0, src);
permP1 = vec_lvsl(+1, src);
permP2 = vec_lvsl(+2, src);
permP3 = vec_lvsl(+3, src);
#endif /* HAVE_BIGENDIAN */
for (i = 0 ; i < 16 ; i ++) {
load_alignment(src, align, permM2, permM1, permP0, permP1, permP2, permP3);
srcP0A = (vec_s16) VEC_MERGEH(zero_u8v, srcP0);
srcP0B = (vec_s16) VEC_MERGEL(zero_u8v, srcP0);
srcP1A = (vec_s16) VEC_MERGEH(zero_u8v, srcP1);
srcP1B = (vec_s16) VEC_MERGEL(zero_u8v, srcP1);
srcP2A = (vec_s16) VEC_MERGEH(zero_u8v, srcP2);
srcP2B = (vec_s16) VEC_MERGEL(zero_u8v, srcP2);
srcP3A = (vec_s16) VEC_MERGEH(zero_u8v, srcP3);
srcP3B = (vec_s16) VEC_MERGEL(zero_u8v, srcP3);
srcM1A = (vec_s16) VEC_MERGEH(zero_u8v, srcM1);
srcM1B = (vec_s16) VEC_MERGEL(zero_u8v, srcM1);
srcM2A = (vec_s16) VEC_MERGEH(zero_u8v, srcM2);
srcM2B = (vec_s16) VEC_MERGEL(zero_u8v, 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, zero_s16v);
pp2B = vec_mladd(sum2B, v5ss, zero_s16v);
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);
ASSERT_ALIGNED(dst);
OP_U8_ALTIVEC(fsum, sum, vec_ld(0, dst));
vec_st(fsum, 0, dst);
src += srcStride;
dst += dstStride;
}
}
/* 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, 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;
}
}
}
uint32_t
dequant_h263_intra_altivec_c(int16_t *data,
const int16_t *coeff,
const uint32_t quant,
const uint32_t dcscalar,
const uint16_t *mpeg_quant_matrices)
{
vector signed short acLevel;
vector signed short vec_2048;
vector unsigned short quant_add;
vector unsigned short quant_m_2;
vector unsigned short t;
vector bool short equal_zero;
vector bool short less_zero;
vector bool short overflow;
register vector unsigned int even;
register vector unsigned int odd;
register vector unsigned int high;
register vector unsigned int low;
register vector unsigned char zerovec;
register int16_t *data_ptr;
register int16_t *coeff_ptr;
#ifdef DEBUG
if(((unsigned)data) & 0x15)
fprintf(stderr, "dequant_h263_intra_altivec_c:incorrect align, data: %lx\n", (long)data);
#endif
/* initialize */
*((unsigned short*)&quant_add) = (unsigned short)(quant & 1 ? quant : quant - 1);
quant_add = vec_splat(quant_add,0);
*((unsigned short*)&quant_m_2) = (unsigned short)(quant << 1);
quant_m_2 = vec_splat(quant_m_2,0);
vec_2048 = vec_sl(vec_splat_s16(1), vec_splat_u16(11));
zerovec = vec_splat_u8(0);
data_ptr = (int16_t*)data;
coeff_ptr = (int16_t*)coeff;
/* dequant */
DEQUANT_H263_INTRA_ALTIVEC();
DEQUANT_H263_INTRA_ALTIVEC();
DEQUANT_H263_INTRA_ALTIVEC();
DEQUANT_H263_INTRA_ALTIVEC();
DEQUANT_H263_INTRA_ALTIVEC();
DEQUANT_H263_INTRA_ALTIVEC();
DEQUANT_H263_INTRA_ALTIVEC();
DEQUANT_H263_INTRA_ALTIVEC();
/* data[0] is special */
data[0] = coeff[0] * dcscalar;
if(data[0] < -2048)
data[0] = -2048;
else if(data[0] > 2047)
data[0] = 2047;
return 0;
}
void iquant_intra_m1_altivec(IQUANT_INTRA_PDECL)
{
int i;
vector signed short vsrc;
uint16_t *qmat;
vector unsigned short vqmat;
vector unsigned short vmquant;
vector bool short eqzero, ltzero;
vector signed short val, t0;
vector signed short zero, one;
vector unsigned int four;
vector signed short min, max;
int offset, offset2;
int16_t dst0;
union {
vector unsigned short vu16;
unsigned short mquant;
vector signed int vs32;
struct {
signed int pad[3];
signed int sum;
} s;
} vu;
#ifdef ALTIVEC_DST
DataStreamControl dsc;
#endif
#ifdef ALTIVEC_VERIFY /* {{{ */
if (NOT_VECTOR_ALIGNED(wsp->intra_q_mat))
mjpeg_error_exit1("iquant_intra_m1: wsp->intra_q_mat %% 16 != 0, (%d)",
wsp->intra_q_mat);
if (NOT_VECTOR_ALIGNED(src))
mjpeg_error_exit1("iquant_intra_m1: src %% 16 != 0, (%d)", src);
if (NOT_VECTOR_ALIGNED(dst))
mjpeg_error_exit1("iquant_intra_m1: dst %% 16 != 0, (%d)", dst);
for (i = 0; i < 64; i++)
if (src[i] < -256 || src[i] > 255)
mjpeg_error_exit1("iquant_intra_m2: -256 > src[%i] > 255, (%d)",
i, src[i]);
#endif /* }}} */
AMBER_START;
dst0 = src[0] << (3 - dc_prec);
qmat = (uint16_t*)wsp->intra_q_mat;
#ifdef ALTIVEC_DST
dsc.control = DATA_STREAM_CONTROL(64/8,1,0);
vec_dst(src, dsc.control, 0);
vec_dst(qmat, dsc.control, 1);
#endif
/* vmquant = (vector unsigned short)(mquant); */
vu.mquant = (unsigned short)mquant;
vmquant = vec_splat(vu.vu16, 0);
zero = vec_splat_s16(0);
one = vec_splat_s16(1);
four = vec_splat_u32(4);
/* max = (2047); min = (-2048); {{{ */
vu8(max) = vec_splat_u8(0x7);
t0 = vec_splat_s16(-1); /* 0xffff */
vu8(max) = vec_mergeh(vu8(max), vu8(t0)); /* 0x07ff == 2047 */
min = vec_sub(t0, max);
/* }}} */
offset = 0;
#if 1
vsrc = vec_ld(offset, (signed short*)src);
vqmat = vec_ld(offset, (unsigned short*)qmat);
i = (64/8) - 1;
do {
/* intra_q[i] * mquant */
vu16(vqmat) = vec_mulo(vu8(vqmat), vu8(vmquant));
/* save sign */
ltzero = vec_cmplt(vsrc, zero);
eqzero = vec_cmpeq(vsrc, zero);
/* val = abs(src) */
t0 = vec_sub(zero, vsrc);
val = vec_max(t0, vsrc);
/* val = (src * quant) >> 4 */
vs32(t0) = vec_mule(val, vs16(vqmat));
vs32(val) = vec_mulo(val, vs16(vqmat));
vs32(t0) = vec_sra(vs32(t0), four);
vs16(t0) = vec_pack(vs32(t0), vs32(t0));
vs32(val) = vec_sra(vs32(val), four);
vs16(val) = vec_pack(vs32(val), vs32(val));
val = vec_mergeh(vs16(t0), vs16(val));
offset2 = offset;
offset += 8*sizeof(int16_t);
vsrc = vec_ld(offset, (signed short*)src);
vqmat = vec_ld(offset, (unsigned short*)qmat);
/* val = val - 1&~(val|val==0) */
t0 = vec_or(val, eqzero);
t0 = vec_andc(one, t0);
val = vec_sub(val, t0);
/* restore sign */
t0 = vec_sub(zero, val);
val = vec_sel(val, t0, ltzero);
/* val = (val > 2047) ? ((val < -2048) ? -2048 : val); */
val = vec_min(val, max);
val = vec_max(val, min);
vec_st(val, offset2, dst);
} while (--i);
/* intra_q[i] * mquant */
vu16(vqmat) = vec_mulo(vu8(vqmat), vu8(vmquant));
/* save sign */
ltzero = vec_cmplt(vsrc, zero);
eqzero = vec_cmpeq(vsrc, zero);
/* val = abs(src) */
t0 = vec_sub(zero, vsrc);
val = vec_max(t0, vsrc);
/* val = (src * quant) >> 4 */
vs32(t0) = vec_mule(val, vs16(vqmat));
vs32(val) = vec_mulo(val, vs16(vqmat));
vs32(t0) = vec_sra(vs32(t0), four);
vs16(t0) = vec_pack(vs32(t0), vs32(t0));
vs32(val) = vec_sra(vs32(val), four);
vs16(val) = vec_pack(vs32(val), vs32(val));
val = vec_mergeh(vs16(t0), vs16(val));
/* val = val - 1&~(val|val==0) */
t0 = vec_or(val, eqzero);
t0 = vec_andc(one, t0);
val = vec_sub(val, t0);
/* restore sign */
t0 = vec_sub(zero, val);
val = vec_sel(val, t0, ltzero);
/* val = (val > 2047) ? ((val < -2048) ? -2048 : val); */
val = vec_min(val, max);
val = vec_max(val, min);
vec_st(val, offset, dst);
#else
/* {{{ */
i = (64/8);
do {
vsrc = vec_ld(offset, (signed short*)src);
vqmat = vec_ld(offset, (unsigned short*)qmat);
/* intra_q[i] * mquant */
vu16(vqmat) = vec_mulo(vu8(vqmat), vu8(vmquant));
/* save sign */
ltzero = vec_cmplt(vsrc, zero);
eqzero = vec_cmpeq(vsrc, zero);
/* val = abs(src) */
t0 = vec_sub(zero, vsrc);
val = vec_max(t0, vsrc);
/* val = (src * quant) >> 4 */
vs32(t0) = vec_mule(val, vs16(vqmat));
vs32(val) = vec_mulo(val, vs16(vqmat));
vs32(t0) = vec_sra(vs32(t0), four);
vs16(t0) = vec_pack(vs32(t0), vs32(t0));
vs32(val) = vec_sra(vs32(val), four);
vs16(val) = vec_pack(vs32(val), vs32(val));
val = vec_mergeh(vs16(t0), vs16(val));
/* val = val - 1&~(val|val==0) */
t0 = vec_or(val, eqzero);
t0 = vec_andc(one, t0);
val = vec_sub(val, t0);
/* restore sign */
t0 = vec_sub(zero, val);
val = vec_sel(val, t0, ltzero);
/* val = (val > 2047) ? ((val < -2048) ? -2048 : val); */
val = vec_min(val, max);
val = vec_max(val, min);
vec_st(val, offset, dst);
offset += 8*sizeof(int16_t);
} while (--i);
/* }}} */
#endif
dst[0] = dst0;
AMBER_STOP;
}
uint32_t
dequant_mpeg_inter_altivec_c(int16_t * data,
const int16_t * coeff,
const uint32_t quant,
const uint16_t * mpeg_quant_matrices)
{
register uint32_t sum;
register const uint16_t *inter_matrix = get_inter_matrix(mpeg_quant_matrices);
register vec_sint16_t ox00;
register vec_sint16_t v2048;
register vec_sint16_t level;
register vec_uint16_t vinter;
register vec_uint32_t hi,lo;
register vec_uint32_t sw_hi,sw_lo;
register vec_uint32_t swap;
register vec_uint32_t t,v16;
vec_uint32_t vsum;
vec_uint32_t vquant;
vector bool short zero_eq;
vector bool short zero_less;
vector bool short overflow;
#ifdef DEBUG
if((long)data & 0xf)
fprintf(stderr, "xvidcore: error in dequant_mpeg_inter_altivec_c, incorrect align: %x\n", data);
#endif
/* Initialization */
ox00 = vec_splat_s16(0);
v16 = vec_splat_u32(-16);
v2048 = vec_rl(vec_splat_s16(8),vec_splat_u16(8));
vsum = (vec_uint32_t)ox00;
*((uint32_t*)&vquant) = quant;
vquant = vec_splat(vquant,0);
swap = vec_rl(vquant,v16);
DEQUANT_MPEG_INTER();
DEQUANT_MPEG_INTER();
DEQUANT_MPEG_INTER();
DEQUANT_MPEG_INTER();
DEQUANT_MPEG_INTER();
DEQUANT_MPEG_INTER();
DEQUANT_MPEG_INTER();
DEQUANT_MPEG_INTER();
sum = ((uint32_t*)&vsum)[0];
sum ^= ((uint32_t*)&vsum)[1];
sum ^= ((uint32_t*)&vsum)[2];
sum ^= ((uint32_t*)&vsum)[3];
/* mismatch control */
if((sum & 1) == 0) {
data -= 1;
*data ^= 1;
}
return 0;
}
DECLARE_ALIGNED(16, signed int, ABCD)[4] =
{
((8 - x) * (8 - y)),
(( x) * (8 - y)),
((8 - x) * ( y)),
(( x) * ( y))
};
register int i;
vec_u8 fperm;
const vec_s32 vABCD = vec_ld(0, ABCD);
const vec_s16 vA = vec_splat((vec_s16)vABCD, 1);
const vec_s16 vB = vec_splat((vec_s16)vABCD, 3);
const vec_s16 vC = vec_splat((vec_s16)vABCD, 5);
const vec_s16 vD = vec_splat((vec_s16)vABCD, 7);
LOAD_ZERO;
const vec_s16 v32ss = vec_sl(vec_splat_s16(1), vec_splat_u16(5));
const vec_u16 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;
vec_u8 vsrcAuc, av_uninit(vsrcBuc), vsrcperm0, vsrcperm1;
vec_u8 vsrc0uc, vsrc1uc;
vec_s16 vsrc0ssH, vsrc1ssH;
vec_u8 vsrcCuc, vsrc2uc, vsrc3uc;
vec_s16 vsrc2ssH, vsrc3ssH, psum;
vec_u8 vdst, ppsum, vfdst, fsum;
if (((unsigned long)dst) % 16 == 0)
{
fperm = (vec_u8)
{
/* 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);
POWERPC_PERF_START_COUNT(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 v20ss = (const vector signed short)AVV(20);
const vector unsigned short v5us = vec_splat_u16(5);
const vector signed short v5ss = vec_splat_s16(5);
const vector signed short v16ss = (const vector signed short)AVV(16);
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);
register int align = ((((unsigned long)src) - 2) % 16);
for (i = 0 ; i < 16 ; 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;
}
const vector signed short srcP0A = (vector signed short)vec_mergeh((vector unsigned char)vzero, srcP0);
const vector signed short srcP0B = (vector signed short)vec_mergel((vector unsigned char)vzero, srcP0);
const vector signed short srcP1A = (vector signed short)vec_mergeh((vector unsigned char)vzero, srcP1);
const vector signed short srcP1B = (vector signed short)vec_mergel((vector unsigned char)vzero, srcP1);
const vector signed short srcP2A = (vector signed short)vec_mergeh((vector unsigned char)vzero, srcP2);
const vector signed short srcP2B = (vector signed short)vec_mergel((vector unsigned char)vzero, srcP2);
const vector signed short srcP3A = (vector signed short)vec_mergeh((vector unsigned char)vzero, srcP3);
const vector signed short srcP3B = (vector signed short)vec_mergel((vector unsigned char)vzero, srcP3);
const vector signed short srcM1A = (vector signed short)vec_mergeh((vector unsigned char)vzero, srcM1);
const vector signed short srcM1B = (vector signed short)vec_mergel((vector unsigned char)vzero, srcM1);
const vector signed short srcM2A = (vector signed short)vec_mergeh((vector unsigned char)vzero, srcM2);
const vector signed short srcM2B = (vector signed short)vec_mergel((vector unsigned char)vzero, srcM2);
const vector signed short sum1A = vec_adds(srcP0A, srcP1A);
const vector signed short sum1B = vec_adds(srcP0B, srcP1B);
const vector signed short sum2A = vec_adds(srcM1A, srcP2A);
const vector signed short sum2B = vec_adds(srcM1B, srcP2B);
const vector signed short sum3A = vec_adds(srcM2A, srcP3A);
const vector signed short sum3B = vec_adds(srcM2B, srcP3B);
const vector signed short pp1A = vec_mladd(sum1A, v20ss, v16ss);
const vector signed short pp1B = vec_mladd(sum1B, v20ss, v16ss);
const vector signed short pp2A = vec_mladd(sum2A, v5ss, (vector signed short)vzero);
const vector signed short pp2B = vec_mladd(sum2B, v5ss, (vector signed short)vzero);
const vector signed short pp3A = vec_add(sum3A, pp1A);
const vector signed short pp3B = vec_add(sum3B, pp1B);
const vector signed short psumA = vec_sub(pp3A, pp2A);
const vector signed short psumB = vec_sub(pp3B, pp2B);
const vector signed short sumA = vec_sra(psumA, v5us);
const vector signed short sumB = vec_sra(psumB, v5us);
const vector unsigned char sum = vec_packsu(sumA, sumB);
const vector unsigned char dst1 = vec_ld(0, dst);
const vector unsigned char dst2 = vec_ld(16, dst);
const vector unsigned char vdst = vec_perm(dst1, dst2, vec_lvsl(0, dst));
vector unsigned char fsum;
OP_U8_ALTIVEC(fsum, sum, vdst);
const vector unsigned char rsum = vec_perm(fsum, fsum, dstperm);
const vector unsigned char fdst1 = vec_sel(dst1, rsum, dstmask);
const vector unsigned char 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);
}
static void PREFIX_h264_qpel16_v_lowpass_altivec(uint8_t * dst, uint8_t * src, int dstStride, int srcStride) {
register int i;
LOAD_ZERO;
const vec_u8 perm = vec_lvsl(0, src);
const vec_s16 v20ss = vec_sl(vec_splat_s16(5),vec_splat_u16(2));
const vec_u16 v5us = vec_splat_u16(5);
const vec_s16 v5ss = vec_splat_s16(5);
const vec_s16 v16ss = vec_sl(vec_splat_s16(1),vec_splat_u16(4));
uint8_t *srcbis = src - (srcStride * 2);
const vec_u8 srcM2a = vec_ld(0, srcbis);
const vec_u8 srcM2b = vec_ld(16, srcbis);
const vec_u8 srcM2 = vec_perm(srcM2a, srcM2b, perm);
//srcbis += srcStride;
const vec_u8 srcM1a = vec_ld(0, srcbis += srcStride);
const vec_u8 srcM1b = vec_ld(16, srcbis);
const vec_u8 srcM1 = vec_perm(srcM1a, srcM1b, perm);
//srcbis += srcStride;
const vec_u8 srcP0a = vec_ld(0, srcbis += srcStride);
const vec_u8 srcP0b = vec_ld(16, srcbis);
const vec_u8 srcP0 = vec_perm(srcP0a, srcP0b, perm);
//srcbis += srcStride;
const vec_u8 srcP1a = vec_ld(0, srcbis += srcStride);
const vec_u8 srcP1b = vec_ld(16, srcbis);
const vec_u8 srcP1 = vec_perm(srcP1a, srcP1b, perm);
//srcbis += srcStride;
const vec_u8 srcP2a = vec_ld(0, srcbis += srcStride);
const vec_u8 srcP2b = vec_ld(16, srcbis);
const vec_u8 srcP2 = vec_perm(srcP2a, srcP2b, perm);
//srcbis += srcStride;
vec_s16 srcM2ssA = (vec_s16) vec_mergeh(zero_u8v, srcM2);
vec_s16 srcM2ssB = (vec_s16) vec_mergel(zero_u8v, srcM2);
vec_s16 srcM1ssA = (vec_s16) vec_mergeh(zero_u8v, srcM1);
vec_s16 srcM1ssB = (vec_s16) vec_mergel(zero_u8v, srcM1);
vec_s16 srcP0ssA = (vec_s16) vec_mergeh(zero_u8v, srcP0);
vec_s16 srcP0ssB = (vec_s16) vec_mergel(zero_u8v, srcP0);
vec_s16 srcP1ssA = (vec_s16) vec_mergeh(zero_u8v, srcP1);
vec_s16 srcP1ssB = (vec_s16) vec_mergel(zero_u8v, srcP1);
vec_s16 srcP2ssA = (vec_s16) vec_mergeh(zero_u8v, srcP2);
vec_s16 srcP2ssB = (vec_s16) vec_mergel(zero_u8v, srcP2);
vec_s16 pp1A, pp1B, pp2A, pp2B, pp3A, pp3B,
psumA, psumB, sumA, sumB,
srcP3ssA, srcP3ssB,
sum1A, sum1B, sum2A, sum2B, sum3A, sum3B;
vec_u8 sum, fsum, srcP3a, srcP3b, srcP3;
for (i = 0 ; i < 16 ; i++) {
srcP3a = vec_ld(0, srcbis += srcStride);
srcP3b = vec_ld(16, srcbis);
srcP3 = vec_perm(srcP3a, srcP3b, perm);
srcP3ssA = (vec_s16) vec_mergeh(zero_u8v, srcP3);
srcP3ssB = (vec_s16) vec_mergel(zero_u8v, 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, zero_s16v);
pp2B = vec_mladd(sum2B, v5ss, zero_s16v);
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);
ASSERT_ALIGNED(dst);
OP_U8_ALTIVEC(fsum, sum, vec_ld(0, dst));
vec_st(fsum, 0, dst);
dst += dstStride;
}
}
/* 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);
POWERPC_PERF_START_COUNT(PREFIX_h264_chroma_mc8_num, 1);
signed int ABCD[4] __attribute__((aligned(16)));
register int i;
ABCD[0] = ((8 - x) * (8 - y));
ABCD[1] = ((x) * (8 - y));
ABCD[2] = ((8 - x) * (y));
ABCD[3] = ((x) * (y));
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 = (const vector signed short)AVV(32);
const vector unsigned short v6us = vec_splat_u16(6);
vector unsigned char fperm;
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);
}
register int loadSecond = (((unsigned long)src) % 16) <= 7 ? 0 : 1;
register int reallyBadAlign = (((unsigned long)src) % 16) == 15 ? 1 : 0;
vector unsigned char vsrcAuc;
vector unsigned char vsrcBuc;
vector unsigned char vsrcperm0;
vector unsigned char vsrcperm1;
vsrcAuc = vec_ld(0, src);
if (loadSecond)
vsrcBuc = vec_ld(16, src);
vsrcperm0 = vec_lvsl(0, src);
vsrcperm1 = vec_lvsl(1, src);
vector unsigned char vsrc0uc;
vector unsigned char vsrc1uc;
vsrc0uc = vec_perm(vsrcAuc, vsrcBuc, vsrcperm0);
if (reallyBadAlign)
vsrc1uc = vsrcBuc;
else
vsrc1uc = vec_perm(vsrcAuc, vsrcBuc, vsrcperm1);
vector signed short vsrc0ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero, (vector unsigned char)vsrc0uc);
vector signed short vsrc1ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero, (vector unsigned char)vsrc1uc);
if (!loadSecond) {// -> !reallyBadAlign
for (i = 0 ; i < h ; i++) {
vector unsigned char vsrcCuc;
vsrcCuc = vec_ld(stride + 0, src);
vector unsigned char vsrc2uc;
vector unsigned char vsrc3uc;
vsrc2uc = vec_perm(vsrcCuc, vsrcCuc, vsrcperm0);
vsrc3uc = vec_perm(vsrcCuc, vsrcCuc, vsrcperm1);
vector signed short vsrc2ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero, (vector unsigned char)vsrc2uc);
vector signed short vsrc3ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero, (vector unsigned char)vsrc3uc);
vector signed short psum;
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);
vector unsigned char vdst = vec_ld(0, dst);
vector unsigned char ppsum = (vector unsigned char)vec_packsu(psum, psum);
vector unsigned char vfdst = vec_perm(vdst, ppsum, fperm);
vector unsigned char fsum;
OP_U8_ALTIVEC(fsum, vfdst, vdst);
vec_st(fsum, 0, dst);
vsrc0ssH = vsrc2ssH;
vsrc1ssH = vsrc3ssH;
dst += stride;
src += stride;
}
} else {
for (i = 0 ; i < h ; i++) {
vector unsigned char vsrcCuc;
vector unsigned char vsrcDuc;
vsrcCuc = vec_ld(stride + 0, src);
vsrcDuc = vec_ld(stride + 16, src);
vector unsigned char vsrc2uc;
vector unsigned char vsrc3uc;
vsrc2uc = vec_perm(vsrcCuc, vsrcDuc, vsrcperm0);
if (reallyBadAlign)
vsrc3uc = vsrcDuc;
else
vsrc3uc = vec_perm(vsrcCuc, vsrcDuc, vsrcperm1);
vector signed short vsrc2ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero, (vector unsigned char)vsrc2uc);
vector signed short vsrc3ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero, (vector unsigned char)vsrc3uc);
vector signed short psum;
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);
vector unsigned char vdst = vec_ld(0, dst);
vector unsigned char ppsum = (vector unsigned char)vec_pack(psum, psum);
vector unsigned char vfdst = vec_perm(vdst, ppsum, fperm);
vector unsigned char fsum;
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);
}
static void PREFIX_h264_qpel16_h_lowpass_altivec(uint8_t * dst, uint8_t * src, int dstStride, int srcStride) {
register int i;
LOAD_ZERO;
const vec_u8 permM2 = vec_lvsl(-2, src);
const vec_u8 permM1 = vec_lvsl(-1, src);
const vec_u8 permP0 = vec_lvsl(+0, src);
const vec_u8 permP1 = vec_lvsl(+1, src);
const vec_u8 permP2 = vec_lvsl(+2, src);
const vec_u8 permP3 = vec_lvsl(+3, src);
const vec_s16 v5ss = vec_splat_s16(5);
const vec_u16 v5us = vec_splat_u16(5);
const vec_s16 v20ss = vec_sl(vec_splat_s16(5),vec_splat_u16(2));
const vec_s16 v16ss = vec_sl(vec_splat_s16(1),vec_splat_u16(4));
vec_u8 srcM2, srcM1, srcP0, srcP1, srcP2, srcP3;
register int align = ((((unsigned long)src) - 2) % 16);
vec_s16 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;
vec_u8 sum, fsum;
for (i = 0 ; i < 16 ; i ++) {
vec_u8 srcR1 = vec_ld(-2, src);
vec_u8 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: {
vec_u8 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: {
vec_u8 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: {
vec_u8 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: {
vec_u8 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 = (vec_s16) vec_mergeh(zero_u8v, srcP0);
srcP0B = (vec_s16) vec_mergel(zero_u8v, srcP0);
srcP1A = (vec_s16) vec_mergeh(zero_u8v, srcP1);
srcP1B = (vec_s16) vec_mergel(zero_u8v, srcP1);
srcP2A = (vec_s16) vec_mergeh(zero_u8v, srcP2);
srcP2B = (vec_s16) vec_mergel(zero_u8v, srcP2);
srcP3A = (vec_s16) vec_mergeh(zero_u8v, srcP3);
srcP3B = (vec_s16) vec_mergel(zero_u8v, srcP3);
srcM1A = (vec_s16) vec_mergeh(zero_u8v, srcM1);
srcM1B = (vec_s16) vec_mergel(zero_u8v, srcM1);
srcM2A = (vec_s16) vec_mergeh(zero_u8v, srcM2);
srcM2B = (vec_s16) vec_mergel(zero_u8v, 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, zero_s16v);
pp2B = vec_mladd(sum2B, v5ss, zero_s16v);
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);
ASSERT_ALIGNED(dst);
OP_U8_ALTIVEC(fsum, sum, vec_ld(0, dst));
vec_st(fsum, 0, dst);
src += srcStride;
dst += dstStride;
}
}
void YUV422_to_BGRA_altivec(const unsigned char *yuvdata,
size_t pixelnum, unsigned char *output)
{
const vector unsigned char *UYVY_ptr=reinterpret_cast<const vector unsigned char *>(yuvdata);
vector unsigned char *BGRA_ptr=reinterpret_cast<vector unsigned char *>(output);
vector unsigned int vShift;
vector signed short tempU, tempV, tempY, tempUV, out1, out2, out3, out4;
vector signed short v16, v128, a255, szero, one;
vector unsigned char zero;
vector signed short t0, t1, t2, tempGB1, tempGB2, tempRA1, tempRA2;
vector signed short vU_G, vV_G, vU_B, vU_R, y0, hiImage, loImage;
vector unsigned int uv_rEven, uv_rOdd, uv_rHi, uv_rLo,
uv_gUEven, uv_gVEven, uv_gUOdd, uv_gVOdd, uv_gHi, uv_gLo,
uv_bEven, uv_bOdd;
vector signed int tempUhi, tempUlo, tempVhi, tempVlo;
vector signed int yEven, yOdd;
vector unsigned int t0Even, t0Odd, t1Even, t1Odd, t2Even, t2Odd;
/* Load the equation constants. */
vector signed short vConst =
static_cast<vector signed short>(298, 519, 409, 16, 128, 255, -100, -210 );
vector unsigned char vPerm1 =
static_cast<vector unsigned char>( 0, 1, 16, 17, 8, 9, 24, 25,
2, 3, 18, 19, 10, 11, 26, 27 );
vector unsigned char vPerm2 =
static_cast<vector unsigned char>( 4, 5, 20, 21, 12, 13, 28, 29,
6, 7, 22, 23, 14, 15, 30, 31 );
vector unsigned char vPermY =
static_cast<vector unsigned char>( 2, 3, 6, 7, 10, 11, 14, 15,
18, 19, 22, 23, 26, 27, 30, 31 );
vector unsigned char vPermU =
static_cast<vector unsigned char>( 0, 1, 16, 17, 4, 5, 20, 21,
8, 9, 24, 25, 12, 13, 28, 29 );
vector unsigned char vPermV =
static_cast<vector unsigned char>( 2, 3, 18, 19, 6, 7, 22, 23,
10, 11, 26, 27, 14, 15, 30, 31 );
vector unsigned char vOutPerm1 =
static_cast<vector unsigned char>( 0, 1, 2, 3, 16, 17, 18, 19,
4, 5, 6, 7, 20, 21, 22, 23 );
vector unsigned char vOutPerm2 =
static_cast<vector unsigned char>( 8, 9, 10, 11, 24, 25, 26, 27,
12, 13, 14, 15, 28, 29, 30, 31 );
vector unsigned char uvPerm =
static_cast<vector unsigned char>( 0, 1, 4, 5, 8, 9, 12, 13,
16, 17, 20, 21, 24, 25, 28, 29 );
zero = vec_splat_u8(0);
szero = vec_splat_s16(0);
one = vec_splat_s16(1);
vShift = vec_splat_u32(8);
a255 = vec_splat( vConst, 5 ); // alpha channel = 255
vU_G = vec_splat( vConst, 6 ); // -100
vV_G = vec_splat( vConst, 7 ); // -210
vU_B = vec_splat( vConst, 1 ); // 519
vU_R = vec_splat( vConst, 2 ); // 409
y0 = vec_splat( vConst, 0 ); // 298
v16 = vec_splat( vConst, 3 ); // 16
v128 = vec_splat( vConst, 4 ); // 128
for ( unsigned int i = 0; i < (pixelnum/sizeof(vector unsigned char)); i++ ) {
// Load UYUV input vector
const vector unsigned char *vec1 = UYVY_ptr++;
//expand the UInt8's to short's
hiImage = static_cast<vector signed short>(vec_mergeh( zero, *vec1 ));
loImage = static_cast<vector signed short>(vec_mergel( zero, *vec1 ));
tempUV = static_cast<vector signed short>(vec_perm( hiImage, loImage, uvPerm ));
tempY = static_cast<vector signed short>(vec_perm( hiImage, loImage, vPermY ));
// subtract UV_OFFSET from UV's (should this be saturated?)
tempUV = static_cast<vector signed short>(vec_sub( tempUV, v128 ));
// subtract Y-OFFSET from Y's (should this be saturated?)
tempY = static_cast<vector signed short>(vec_sub( tempY, v16 ));
// expand to UUUU UUUU and VVVV VVVV
tempU = vec_perm(tempUV, tempUV, vPermU);
tempV = vec_perm(tempUV, tempUV, vPermV);
//below:
//
//error: cannot convert `vector int' to `vector unsigned int' in assignment
tempUhi = vec_mule( tempU, one );
// unsigned int = vec_mule( signed short, signed short )
// should be
// signed int = vec_mule( signed short, signed short )
tempUlo = vec_mulo( tempU, one );
tempVhi = vec_mule( tempV, one );
tempVlo = vec_mulo( tempV, one );
// uv_r = YUV2RGB_12*u + YUV2RGB_13*v
// uv_r = (-1)*u + 409*v (or "409*V - U")
uv_rEven = vec_mule( tempV, vU_R );
uv_rOdd = vec_mulo( tempV, vU_R );
uv_rHi = vec_sub( uv_rEven, tempUhi );
uv_rLo = vec_sub( uv_rOdd, tempUlo );
// uv_g = YUV2RGB_22*u + YUV2RGB_23*v
// uv_g = -100*u + (-210)*v
// multiply U by -100
uv_gUEven = vec_mule( tempU, vU_G );
uv_gUOdd = vec_mulo( tempU, vU_G );
// multiply V by -210
uv_gVEven = vec_mule( tempV, vV_G );
uv_gVOdd = vec_mulo( tempV, vV_G );
// add U & V products
uv_gHi = vec_add( uv_gUEven, uv_gVEven );
uv_gLo = vec_add( uv_gUOdd, uv_gVOdd );
// uv_b = YUV2RGB_32*u + YUV2RGB_33*v
// uv_b = 519*u + 0*v
uv_bEven = vec_mule( tempU, vU_B );
uv_bOdd = vec_mulo( tempU, vU_B );
// y = YUV2RGB_11 * tempY
// y = 298* (tempY - 16)
yEven = vec_mule( tempY, y0 );
yOdd = vec_mulo( tempY, y0 );
// add while int's
t0Even = vec_add( yEven, uv_bEven );
t0Odd = vec_add( yOdd, uv_bOdd );
t1Even = vec_add( yEven, uv_gHi );
t1Odd = vec_add( yOdd, uv_gLo );
t2Even = vec_add( yEven, uv_rHi );
t2Odd = vec_add( yOdd, uv_rLo );
// shift while int's
t0Even = vec_sra( t0Even, vShift );
t0Odd = vec_sra( t0Odd, vShift );
t1Even = vec_sra( t1Even, vShift );
t1Odd = vec_sra( t1Odd, vShift );
t2Even = vec_sra( t2Even, vShift );
t2Odd = vec_sra( t2Odd, vShift );
// pack down to shorts
t0 = vec_packs( t0Even, t0Odd );
t1 = vec_packs( t1Even, t1Odd );
t2 = vec_packs( t2Even, t2Odd );
// Permute to GBGBGBGB GBGBGBGB + re-interleave even & odd
tempGB1 = vec_perm( t1, t0, vPerm1 );
tempGB2 = vec_perm( t1, t0, vPerm2 );
// Permute to ARARARAR ARARARAR + re-interleave even & odd
tempRA1 = vec_perm( a255, t2, vPerm1 );
tempRA2 = vec_perm( a255, t2, vPerm2 );
// Permute to ARGB's
out1 = vec_perm( tempRA1, tempGB1, vOutPerm1 );
out2 = vec_perm( tempRA1, tempGB1, vOutPerm2 );
out3 = vec_perm( tempRA2, tempGB2, vOutPerm1 );
out4 = vec_perm( tempRA2, tempGB2, vOutPerm2 );
// pack down to char's
*BGRA_ptr = vec_packsu( out1, out2 );
BGRA_ptr++;
*BGRA_ptr = vec_packsu( out3, out4 );
BGRA_ptr++;
}
}
