static int32_t scalarproduct_and_madd_int16_altivec(int16_t *v1, const int16_t *v2, const int16_t *v3, int order, int mul)
{
LOAD_ZERO;
vec_s16 *pv1 = (vec_s16*)v1;
vec_s16 *pv2 = (vec_s16*)v2;
vec_s16 *pv3 = (vec_s16*)v3;
register vec_s16 muls = {mul,mul,mul,mul,mul,mul,mul,mul};
register vec_s16 t0, t1, i0, i1;
register vec_s16 i2 = pv2[0], i3 = pv3[0];
register vec_s32 res = zero_s32v;
register vec_u8 align = vec_lvsl(0, v2);
int32_t ires;
order >>= 4;
do {
t0 = vec_perm(i2, pv2[1], align);
i2 = pv2[2];
t1 = vec_perm(pv2[1], i2, align);
i0 = pv1[0];
i1 = pv1[1];
res = vec_msum(t0, i0, res);
res = vec_msum(t1, i1, res);
t0 = vec_perm(i3, pv3[1], align);
i3 = pv3[2];
t1 = vec_perm(pv3[1], i3, align);
pv1[0] = vec_mladd(t0, muls, i0);
pv1[1] = vec_mladd(t1, muls, i1);
pv1 += 2;
pv2 += 2;
pv3 += 2;
} while(--order);
res = vec_splat(vec_sums(res, zero_s32v), 3);
vec_ste(res, 0, &ires);
return ires;
}
SIMD_INLINE v128_u16 Interpolate(v128_u16 s[2][2], v128_u16 k[2][2])
{
v128_u16 r = vec_mladd(s[0][0], k[0][0], K16_BILINEAR_ROUND_TERM);
r = vec_mladd(s[0][1], k[0][1], r);
r = vec_mladd(s[1][0], k[1][0], r);
r = vec_mladd(s[1][1], k[1][1], r);
return vec_sr(r, K16_BILINEAR_SHIFT);
}
/* *************************************************************************
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 */
template <int part> SIMD_INLINE v128_u16 Laplace(v128_u8 a[3][3])
{
return vec_sub(vec_mladd(K16_0008, UnpackU8<part>(a[1][1]), K16_0000),
vec_add(vec_add(vec_add(UnpackU8<part>(a[0][0]), UnpackU8<part>(a[0][1])),
vec_add(UnpackU8<part>(a[0][2]), UnpackU8<part>(a[1][0]))),
vec_add(vec_add(UnpackU8<part>(a[1][2]), UnpackU8<part>(a[2][0])),
vec_add(UnpackU8<part>(a[2][1]), UnpackU8<part>(a[2][2])))));
}
static force_inline vector unsigned int
pix_multiply (vector unsigned int p, vector unsigned int a)
{
vector unsigned short hi, lo, mod;
/* unpack to short */
hi = (vector unsigned short)
vec_mergeh ((vector unsigned char)AVV (0),
(vector unsigned char)p);
mod = (vector unsigned short)
vec_mergeh ((vector unsigned char)AVV (0),
(vector unsigned char)a);
hi = vec_mladd (hi, mod, (vector unsigned short)
AVV (0x0080, 0x0080, 0x0080, 0x0080,
0x0080, 0x0080, 0x0080, 0x0080));
hi = vec_adds (hi, vec_sr (hi, vec_splat_u16 (8)));
hi = vec_sr (hi, vec_splat_u16 (8));
/* unpack to short */
lo = (vector unsigned short)
vec_mergel ((vector unsigned char)AVV (0),
(vector unsigned char)p);
mod = (vector unsigned short)
vec_mergel ((vector unsigned char)AVV (0),
(vector unsigned char)a);
lo = vec_mladd (lo, mod, (vector unsigned short)
AVV (0x0080, 0x0080, 0x0080, 0x0080,
0x0080, 0x0080, 0x0080, 0x0080));
lo = vec_adds (lo, vec_sr (lo, vec_splat_u16 (8)));
lo = vec_sr (lo, vec_splat_u16 (8));
return (vector unsigned int)vec_packsu (hi, lo);
}
static int32_t scalarproduct_and_madd_int16_altivec(int16_t *v1,
const int16_t *v2,
const int16_t *v3,
int order, int mul)
{
LOAD_ZERO;
vec_s16 *pv1 = (vec_s16 *) v1;
register vec_s16 muls = { mul, mul, mul, mul, mul, mul, mul, mul };
register vec_s16 t0, t1, i0, i1, i4, i2, i3;
register vec_s32 res = zero_s32v;
#if HAVE_BIGENDIAN
register vec_u8 align = vec_lvsl(0, v2);
i2 = vec_ld(0, v2);
i3 = vec_ld(0, v3);
#endif
int32_t ires;
order >>= 4;
do {
GET_T(t0,t1,v2,i1,i2);
i0 = pv1[0];
i1 = pv1[1];
res = vec_msum(t0, i0, res);
res = vec_msum(t1, i1, res);
GET_T(t0,t1,v3,i4,i3);
pv1[0] = vec_mladd(t0, muls, i0);
pv1[1] = vec_mladd(t1, muls, i1);
pv1 += 2;
v2 += 16;
v3 += 16;
} while (--order);
res = vec_splat(vec_sums(res, zero_s32v), 3);
vec_ste(res, 0, &ires);
return ires;
}
void r_dimpatchD_ALTIVEC(const DCanvas *const cvs, argb_t color, int alpha, int x1, int y1, int w, int h)
{
int x, y, i;
argb_t *line;
int invAlpha = 256 - alpha;
int dpitch = cvs->pitch / sizeof(argb_t);
line = (argb_t *)cvs->buffer + y1 * dpitch;
int batches = w / 4;
int remainder = w & 3;
// AltiVec temporaries:
const vu16 zero = {0, 0, 0, 0, 0, 0, 0, 0};
const vu16 upper8mask = {0, 0xff, 0xff, 0xff, 0, 0xff, 0xff, 0xff};
const vu16 blendAlpha = {0, alpha, alpha, alpha, 0, alpha, alpha, alpha};
const vu16 blendInvAlpha = {0, invAlpha, invAlpha, invAlpha, 0, invAlpha, invAlpha, invAlpha};
const vu16 blendColor = {0, RPART(color), GPART(color), BPART(color), 0, RPART(color), GPART(color), BPART(color)};
const vu16 blendMult = vec_mladd(blendColor, blendAlpha, zero);
for (y = y1; y < y1 + h; y++)
{
// AltiVec optimize the bulk in batches of 4 colors:
for (i = 0, x = x1; i < batches; ++i, x += 4)
{
const vu32 input = {line[x + 0], line[x + 1], line[x + 2], line[x + 3]};
const vu32 output = (vu32)blend4vs1_altivec(input, blendMult, blendInvAlpha, upper8mask);
vec_ste(output, 0, &line[x]);
vec_ste(output, 4, &line[x]);
vec_ste(output, 8, &line[x]);
vec_ste(output, 12, &line[x]);
}
if (remainder)
{
// Pick up the remainder:
for (; x < x1 + w; x++)
{
line[x] = alphablend1a(line[x], color, alpha);
}
}
line += dpitch;
}
}
static void predict_16x16_p_altivec( uint8_t *src )
{
int H = 0, V = 0;
for( int i = 1; i <= 8; i++ )
{
H += i * ( src[7+i - FDEC_STRIDE ] - src[7-i - FDEC_STRIDE ] );
V += i * ( src[(7+i)*FDEC_STRIDE -1] - src[(7-i)*FDEC_STRIDE -1] );
}
int a = 16 * ( src[15*FDEC_STRIDE -1] + src[15 - FDEC_STRIDE] );
int b = ( 5 * H + 32 ) >> 6;
int c = ( 5 * V + 32 ) >> 6;
int i00 = a - b * 7 - c * 7 + 16;
vec_s16_u i00_u, b_u, c_u;
i00_u.s[0] = i00;
b_u.s[0] = b;
c_u.s[0] = c;
vec_u16_t val5_v = vec_splat_u16(5);
vec_s16_t i00_v, b_v, c_v;
i00_v = vec_splat(i00_u.v, 0);
b_v = vec_splat(b_u.v, 0);
c_v = vec_splat(c_u.v, 0);
vec_s16_t induc_v = (vec_s16_t) CV(0, 1, 2, 3, 4, 5, 6, 7);
vec_s16_t b8_v = vec_sl(b_v, vec_splat_u16(3));
vec_s16_t add_i0_b_0v = vec_mladd(induc_v, b_v, i00_v);
vec_s16_t add_i0_b_8v = vec_adds(b8_v, add_i0_b_0v);
for( int y = 0; y < 16; y++ )
{
vec_s16_t shift_0_v = vec_sra(add_i0_b_0v, val5_v);
vec_s16_t shift_8_v = vec_sra(add_i0_b_8v, val5_v);
vec_u8_t com_sat_v = vec_packsu(shift_0_v, shift_8_v);
vec_st( com_sat_v, 0, &src[0]);
src += FDEC_STRIDE;
add_i0_b_0v = vec_adds(add_i0_b_0v, c_v);
add_i0_b_8v = vec_adds(add_i0_b_8v, c_v);
}
}
static void predict_8x8c_p_altivec( uint8_t *src )
{
int H = 0, V = 0;
for( int i = 0; i < 4; i++ )
{
H += ( i + 1 ) * ( src[4+i - FDEC_STRIDE] - src[2 - i -FDEC_STRIDE] );
V += ( i + 1 ) * ( src[-1 +(i+4)*FDEC_STRIDE] - src[-1+(2-i)*FDEC_STRIDE] );
}
int a = 16 * ( src[-1+7*FDEC_STRIDE] + src[7 - FDEC_STRIDE] );
int b = ( 17 * H + 16 ) >> 5;
int c = ( 17 * V + 16 ) >> 5;
int i00 = a -3*b -3*c + 16;
vec_s16_u i00_u, b_u, c_u;
i00_u.s[0] = i00;
b_u.s[0] = b;
c_u.s[0] = c;
vec_u16_t val5_v = vec_splat_u16(5);
vec_s16_t i00_v, b_v, c_v;
i00_v = vec_splat(i00_u.v, 0);
b_v = vec_splat(b_u.v, 0);
c_v = vec_splat(c_u.v, 0);
vec_s16_t induc_v = (vec_s16_t) CV(0, 1, 2, 3, 4, 5, 6, 7);
vec_s16_t add_i0_b_0v = vec_mladd(induc_v, b_v, i00_v);
PREP_STORE8;
for( int i = 0; i < 8; ++i )
{
vec_s16_t shift_0_v = vec_sra(add_i0_b_0v, val5_v);
vec_u8_t com_sat_v = vec_packsu(shift_0_v, shift_0_v);
VEC_STORE8(com_sat_v, &src[0]);
src += FDEC_STRIDE;
add_i0_b_0v = vec_adds(add_i0_b_0v, c_v);
}
}
/* 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);
}
void
jsimd_idct_islow_altivec (void * dct_table_, JCOEFPTR coef_block,
JSAMPARRAY output_buf, JDIMENSION output_col)
{
short *dct_table = (short *)dct_table_;
int *outptr;
__vector short row0, row1, row2, row3, row4, row5, row6, row7,
col0, col1, col2, col3, col4, col5, col6, col7,
quant0, quant1, quant2, quant3, quant4, quant5, quant6, quant7,
tmp0, tmp1, tmp2, tmp3, z3, z4,
z34l, z34h, col71l, col71h, col26l, col26h, col53l, col53h,
row71l, row71h, row26l, row26h, row53l, row53h,
out0, out1, out2, out3, out4, out5, out6, out7;
__vector int tmp0l, tmp0h, tmp1l, tmp1h, tmp2l, tmp2h, tmp3l, tmp3h,
tmp10l, tmp10h, tmp11l, tmp11h, tmp12l, tmp12h, tmp13l, tmp13h,
z3l, z3h, z4l, z4h,
out0l, out0h, out1l, out1h, out2l, out2h, out3l, out3h, out4l, out4h,
out5l, out5h, out6l, out6h, out7l, out7h;
__vector signed char outb;
/* Constants */
__vector short pw_zero = { __8X(0) },
pw_f130_f054 = { __4X2(F_0_541 + F_0_765, F_0_541) },
pw_f054_mf130 = { __4X2(F_0_541, F_0_541 - F_1_847) },
pw_mf078_f117 = { __4X2(F_1_175 - F_1_961, F_1_175) },
pw_f117_f078 = { __4X2(F_1_175, F_1_175 - F_0_390) },
pw_mf060_mf089 = { __4X2(F_0_298 - F_0_899, -F_0_899) },
pw_mf089_f060 = { __4X2(-F_0_899, F_1_501 - F_0_899) },
pw_mf050_mf256 = { __4X2(F_2_053 - F_2_562, -F_2_562) },
pw_mf256_f050 = { __4X2(-F_2_562, F_3_072 - F_2_562) };
__vector unsigned short pass1_bits = { __8X(PASS1_BITS) };
__vector int pd_zero = { __4X(0) },
pd_descale_p1 = { __4X(1 << (DESCALE_P1 - 1)) },
pd_descale_p2 = { __4X(1 << (DESCALE_P2 - 1)) };
__vector unsigned int descale_p1 = { __4X(DESCALE_P1) },
descale_p2 = { __4X(DESCALE_P2) },
const_bits = { __4X(CONST_BITS) };
__vector signed char pb_centerjsamp = { __16X(CENTERJSAMPLE) };
/* Pass 1: process columns */
col0 = vec_ld(0, coef_block);
col1 = vec_ld(16, coef_block);
col2 = vec_ld(32, coef_block);
col3 = vec_ld(48, coef_block);
col4 = vec_ld(64, coef_block);
col5 = vec_ld(80, coef_block);
col6 = vec_ld(96, coef_block);
col7 = vec_ld(112, coef_block);
tmp1 = vec_or(col1, col2);
tmp2 = vec_or(col3, col4);
tmp1 = vec_or(tmp1, tmp2);
tmp3 = vec_or(col5, col6);
tmp3 = vec_or(tmp3, col7);
tmp1 = vec_or(tmp1, tmp3);
quant0 = vec_ld(0, dct_table);
col0 = vec_mladd(col0, quant0, pw_zero);
if (vec_all_eq(tmp1, pw_zero)) {
/* AC terms all zero */
col0 = vec_sl(col0, pass1_bits);
row0 = vec_splat(col0, 0);
row1 = vec_splat(col0, 1);
row2 = vec_splat(col0, 2);
row3 = vec_splat(col0, 3);
row4 = vec_splat(col0, 4);
row5 = vec_splat(col0, 5);
row6 = vec_splat(col0, 6);
row7 = vec_splat(col0, 7);
} else {
quant1 = vec_ld(16, dct_table);
quant2 = vec_ld(32, dct_table);
quant3 = vec_ld(48, dct_table);
quant4 = vec_ld(64, dct_table);
quant5 = vec_ld(80, dct_table);
quant6 = vec_ld(96, dct_table);
quant7 = vec_ld(112, dct_table);
col1 = vec_mladd(col1, quant1, pw_zero);
col2 = vec_mladd(col2, quant2, pw_zero);
col3 = vec_mladd(col3, quant3, pw_zero);
col4 = vec_mladd(col4, quant4, pw_zero);
col5 = vec_mladd(col5, quant5, pw_zero);
col6 = vec_mladd(col6, quant6, pw_zero);
col7 = vec_mladd(col7, quant7, pw_zero);
DO_IDCT(col, 1);
TRANSPOSE(out, row);
}
/* Pass 2: process rows */
DO_IDCT(row, 2);
TRANSPOSE(out, col);
outb = vec_packs(col0, col0);
outb = vec_add(outb, pb_centerjsamp);
outptr = (int *)(output_buf[0] + output_col);
vec_ste((__vector int)outb, 0, outptr);
vec_ste((__vector int)outb, 4, outptr);
outb = vec_packs(col1, col1);
outb = vec_add(outb, pb_centerjsamp);
outptr = (int *)(output_buf[1] + output_col);
vec_ste((__vector int)outb, 0, outptr);
vec_ste((__vector int)outb, 4, outptr);
outb = vec_packs(col2, col2);
outb = vec_add(outb, pb_centerjsamp);
outptr = (int *)(output_buf[2] + output_col);
vec_ste((__vector int)outb, 0, outptr);
vec_ste((__vector int)outb, 4, outptr);
outb = vec_packs(col3, col3);
outb = vec_add(outb, pb_centerjsamp);
outptr = (int *)(output_buf[3] + output_col);
vec_ste((__vector int)outb, 0, outptr);
vec_ste((__vector int)outb, 4, outptr);
outb = vec_packs(col4, col4);
outb = vec_add(outb, pb_centerjsamp);
outptr = (int *)(output_buf[4] + output_col);
vec_ste((__vector int)outb, 0, outptr);
vec_ste((__vector int)outb, 4, outptr);
outb = vec_packs(col5, col5);
outb = vec_add(outb, pb_centerjsamp);
outptr = (int *)(output_buf[5] + output_col);
vec_ste((__vector int)outb, 0, outptr);
vec_ste((__vector int)outb, 4, outptr);
outb = vec_packs(col6, col6);
outb = vec_add(outb, pb_centerjsamp);
outptr = (int *)(output_buf[6] + output_col);
vec_ste((__vector int)outb, 0, outptr);
vec_ste((__vector int)outb, 4, outptr);
outb = vec_packs(col7, col7);
outb = vec_add(outb, pb_centerjsamp);
outptr = (int *)(output_buf[7] + output_col);
vec_ste((__vector int)outb, 0, outptr);
vec_ste((__vector int)outb, 4, outptr);
}
/* 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;
}
}
}
void biweight_h264_WxH_vmx(uint8_t *dst, uint8_t *src, int stride, int log2_denom,
int weightd, int weights, int offset, int w, int h)
{
int y, dst_aligned, src_aligned;
//vec_u8 vsrc, vdst;
//vec_s16 vtemp, vweights, vweightd, voffset, v0, v1, v2, v3;
//vec_u16 vlog2_denom;
__vector4 vsrc, vdst;
__vector4 vtemp, vweights, vweightd, voffset, v0, v1, v2, v3;
__vector4 vlog2_denom;
DECLARE_ALIGNED(16, int32_t, temp)[4];
//LOAD_ZERO;
offset = ((offset + 1) | 1) << log2_denom;
temp[0] = log2_denom+1;
temp[1] = weights;
temp[2] = weightd;
temp[3] = offset;
vtemp = __lvx(temp, 0);
/*
vlog2_denom = (vec_u16)vec_splat(vtemp, 1);
vweights = vec_splat(vtemp, 3);
vweightd = vec_splat(vtemp, 5);
voffset = vec_splat(vtemp, 7);
*/
vlog2_denom = __vsplth(vtemp,1);
vweights = __vsplth(vtemp,3);
vweightd = __vsplth(vtemp,5);
voffset = __vsplth(vtemp,7);
dst_aligned = !((unsigned long)dst & 0xf);
src_aligned = !((unsigned long)src & 0xf);
for (y=0; y<h; y++) {
vdst = __lvx(dst, 0);
vsrc = __lvx(src, 0);
v0 = __vmrghb(__vzero(), vdst);
v1 = __vmrglb(__vzero(), vdst);
v2 = __vmrghb(__vzero(), vsrc);
v3 = __vmrglb(__vzero(), vsrc);
if (w == 8) {
if (src_aligned)
v3 = v2;
else
v2 = v3;
}
if (w == 16 || dst_aligned) {
v0 = vec_mladd(v0, vweightd, __vzero());
v2 = vec_mladd(v2, vweights, __vzero());
v0 = __vaddshs (v0, voffset);
v0 = __vaddshs(v0, v2);
v0 = __vsrah(v0, vlog2_denom);
}
if (w == 16 || !dst_aligned) {
v1 = vec_mladd(v1, vweightd, __vzero());
v3 = vec_mladd(v3, vweights, __vzero());
v1 = __vaddshs(v1, voffset);
v1 = __vaddshs(v1, v3);
v1 = __vsrah(v1, vlog2_denom);
}
vdst = __vpkshus(v0, v1);
__stvx(vdst, dst, 0);
dst += stride;
src += stride;
}
}
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 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);
}
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);
}
/* 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;
}
}
/* 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_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 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;
}
}
/* 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);
}
