static int dct_quantize_altivec(MpegEncContext* s,
DCTELEM* data, int n,
int qscale, int* overflow)
{
int lastNonZero;
vector float row0, row1, row2, row3, row4, row5, row6, row7;
vector float alt0, alt1, alt2, alt3, alt4, alt5, alt6, alt7;
const vector float zero = (const vector float)FOUROF(0.);
// used after quantize step
int oldBaseValue = 0;
// Load the data into the row/alt vectors
{
vector signed short data0, data1, data2, data3, data4, data5, data6, data7;
data0 = vec_ld(0, data);
data1 = vec_ld(16, data);
data2 = vec_ld(32, data);
data3 = vec_ld(48, data);
data4 = vec_ld(64, data);
data5 = vec_ld(80, data);
data6 = vec_ld(96, data);
data7 = vec_ld(112, data);
// Transpose the data before we start
TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7);
// load the data into floating point vectors. We load
// the high half of each row into the main row vectors
// and the low half into the alt vectors.
row0 = vec_ctf(vec_unpackh(data0), 0);
alt0 = vec_ctf(vec_unpackl(data0), 0);
row1 = vec_ctf(vec_unpackh(data1), 0);
alt1 = vec_ctf(vec_unpackl(data1), 0);
row2 = vec_ctf(vec_unpackh(data2), 0);
alt2 = vec_ctf(vec_unpackl(data2), 0);
row3 = vec_ctf(vec_unpackh(data3), 0);
alt3 = vec_ctf(vec_unpackl(data3), 0);
row4 = vec_ctf(vec_unpackh(data4), 0);
alt4 = vec_ctf(vec_unpackl(data4), 0);
row5 = vec_ctf(vec_unpackh(data5), 0);
alt5 = vec_ctf(vec_unpackl(data5), 0);
row6 = vec_ctf(vec_unpackh(data6), 0);
alt6 = vec_ctf(vec_unpackl(data6), 0);
row7 = vec_ctf(vec_unpackh(data7), 0);
alt7 = vec_ctf(vec_unpackl(data7), 0);
}
// The following block could exist as a separate an altivec dct
// function. However, if we put it inline, the DCT data can remain
// in the vector local variables, as floats, which we'll use during the
// quantize step...
{
const vector float vec_0_298631336 = (vector float)FOUROF(0.298631336f);
const vector float vec_0_390180644 = (vector float)FOUROF(-0.390180644f);
const vector float vec_0_541196100 = (vector float)FOUROF(0.541196100f);
const vector float vec_0_765366865 = (vector float)FOUROF(0.765366865f);
const vector float vec_0_899976223 = (vector float)FOUROF(-0.899976223f);
const vector float vec_1_175875602 = (vector float)FOUROF(1.175875602f);
const vector float vec_1_501321110 = (vector float)FOUROF(1.501321110f);
const vector float vec_1_847759065 = (vector float)FOUROF(-1.847759065f);
const vector float vec_1_961570560 = (vector float)FOUROF(-1.961570560f);
const vector float vec_2_053119869 = (vector float)FOUROF(2.053119869f);
const vector float vec_2_562915447 = (vector float)FOUROF(-2.562915447f);
const vector float vec_3_072711026 = (vector float)FOUROF(3.072711026f);
int whichPass, whichHalf;
for(whichPass = 1; whichPass<=2; whichPass++) {
for(whichHalf = 1; whichHalf<=2; whichHalf++) {
vector float tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
vector float tmp10, tmp11, tmp12, tmp13;
vector float z1, z2, z3, z4, z5;
tmp0 = vec_add(row0, row7); // tmp0 = dataptr[0] + dataptr[7];
tmp7 = vec_sub(row0, row7); // tmp7 = dataptr[0] - dataptr[7];
tmp3 = vec_add(row3, row4); // tmp3 = dataptr[3] + dataptr[4];
tmp4 = vec_sub(row3, row4); // tmp4 = dataptr[3] - dataptr[4];
tmp1 = vec_add(row1, row6); // tmp1 = dataptr[1] + dataptr[6];
tmp6 = vec_sub(row1, row6); // tmp6 = dataptr[1] - dataptr[6];
tmp2 = vec_add(row2, row5); // tmp2 = dataptr[2] + dataptr[5];
tmp5 = vec_sub(row2, row5); // tmp5 = dataptr[2] - dataptr[5];
tmp10 = vec_add(tmp0, tmp3); // tmp10 = tmp0 + tmp3;
tmp13 = vec_sub(tmp0, tmp3); // tmp13 = tmp0 - tmp3;
tmp11 = vec_add(tmp1, tmp2); // tmp11 = tmp1 + tmp2;
tmp12 = vec_sub(tmp1, tmp2); // tmp12 = tmp1 - tmp2;
// dataptr[0] = (DCTELEM) ((tmp10 + tmp11) << PASS1_BITS);
row0 = vec_add(tmp10, tmp11);
// dataptr[4] = (DCTELEM) ((tmp10 - tmp11) << PASS1_BITS);
row4 = vec_sub(tmp10, tmp11);
// z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
z1 = vec_madd(vec_add(tmp12, tmp13), vec_0_541196100, (vector float)zero);
// dataptr[2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
// CONST_BITS-PASS1_BITS);
row2 = vec_madd(tmp13, vec_0_765366865, z1);
// dataptr[6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
// CONST_BITS-PASS1_BITS);
row6 = vec_madd(tmp12, vec_1_847759065, z1);
z1 = vec_add(tmp4, tmp7); // z1 = tmp4 + tmp7;
z2 = vec_add(tmp5, tmp6); // z2 = tmp5 + tmp6;
z3 = vec_add(tmp4, tmp6); // z3 = tmp4 + tmp6;
z4 = vec_add(tmp5, tmp7); // z4 = tmp5 + tmp7;
// z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
z5 = vec_madd(vec_add(z3, z4), vec_1_175875602, (vector float)zero);
// z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
z3 = vec_madd(z3, vec_1_961570560, z5);
// z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
z4 = vec_madd(z4, vec_0_390180644, z5);
// The following adds are rolled into the multiplies above
// z3 = vec_add(z3, z5); // z3 += z5;
// z4 = vec_add(z4, z5); // z4 += z5;
// z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
// Wow! It's actually more efficient to roll this multiply
// into the adds below, even thought the multiply gets done twice!
// z2 = vec_madd(z2, vec_2_562915447, (vector float)zero);
// z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
// Same with this one...
// z1 = vec_madd(z1, vec_0_899976223, (vector float)zero);
// tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
// dataptr[7] = (DCTELEM) DESCALE(tmp4 + z1 + z3, CONST_BITS-PASS1_BITS);
row7 = vec_madd(tmp4, vec_0_298631336, vec_madd(z1, vec_0_899976223, z3));
// tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
// dataptr[5] = (DCTELEM) DESCALE(tmp5 + z2 + z4, CONST_BITS-PASS1_BITS);
row5 = vec_madd(tmp5, vec_2_053119869, vec_madd(z2, vec_2_562915447, z4));
// tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
// dataptr[3] = (DCTELEM) DESCALE(tmp6 + z2 + z3, CONST_BITS-PASS1_BITS);
row3 = vec_madd(tmp6, vec_3_072711026, vec_madd(z2, vec_2_562915447, z3));
// tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
// dataptr[1] = (DCTELEM) DESCALE(tmp7 + z1 + z4, CONST_BITS-PASS1_BITS);
row1 = vec_madd(z1, vec_0_899976223, vec_madd(tmp7, vec_1_501321110, z4));
// Swap the row values with the alts. If this is the first half,
// this sets up the low values to be acted on in the second half.
// If this is the second half, it puts the high values back in
// the row values where they are expected to be when we're done.
SWAP(row0, alt0);
SWAP(row1, alt1);
SWAP(row2, alt2);
SWAP(row3, alt3);
SWAP(row4, alt4);
SWAP(row5, alt5);
SWAP(row6, alt6);
SWAP(row7, alt7);
}
if (whichPass == 1) {
// transpose the data for the second pass
// First, block transpose the upper right with lower left.
SWAP(row4, alt0);
SWAP(row5, alt1);
SWAP(row6, alt2);
SWAP(row7, alt3);
// Now, transpose each block of four
TRANSPOSE4(row0, row1, row2, row3);
TRANSPOSE4(row4, row5, row6, row7);
TRANSPOSE4(alt0, alt1, alt2, alt3);
TRANSPOSE4(alt4, alt5, alt6, alt7);
}
}
}
// perform the quantize step, using the floating point data
// still in the row/alt registers
{
const int* biasAddr;
const vector signed int* qmat;
vector float bias, negBias;
if (s->mb_intra) {
vector signed int baseVector;
// We must cache element 0 in the intra case
// (it needs special handling).
baseVector = vec_cts(vec_splat(row0, 0), 0);
vec_ste(baseVector, 0, &oldBaseValue);
qmat = (vector signed int*)s->q_intra_matrix[qscale];
biasAddr = &(s->intra_quant_bias);
} else {
qmat = (vector signed int*)s->q_inter_matrix[qscale];
biasAddr = &(s->inter_quant_bias);
}
// Load the bias vector (We add 0.5 to the bias so that we're
// rounding when we convert to int, instead of flooring.)
{
vector signed int biasInt;
const vector float negOneFloat = (vector float)FOUROF(-1.0f);
LOAD4(biasInt, biasAddr);
bias = vec_ctf(biasInt, QUANT_BIAS_SHIFT);
negBias = vec_madd(bias, negOneFloat, zero);
}
{
vector float q0, q1, q2, q3, q4, q5, q6, q7;
q0 = vec_ctf(qmat[0], QMAT_SHIFT);
q1 = vec_ctf(qmat[2], QMAT_SHIFT);
q2 = vec_ctf(qmat[4], QMAT_SHIFT);
q3 = vec_ctf(qmat[6], QMAT_SHIFT);
q4 = vec_ctf(qmat[8], QMAT_SHIFT);
q5 = vec_ctf(qmat[10], QMAT_SHIFT);
q6 = vec_ctf(qmat[12], QMAT_SHIFT);
q7 = vec_ctf(qmat[14], QMAT_SHIFT);
row0 = vec_sel(vec_madd(row0, q0, negBias), vec_madd(row0, q0, bias),
vec_cmpgt(row0, zero));
row1 = vec_sel(vec_madd(row1, q1, negBias), vec_madd(row1, q1, bias),
vec_cmpgt(row1, zero));
row2 = vec_sel(vec_madd(row2, q2, negBias), vec_madd(row2, q2, bias),
vec_cmpgt(row2, zero));
row3 = vec_sel(vec_madd(row3, q3, negBias), vec_madd(row3, q3, bias),
vec_cmpgt(row3, zero));
row4 = vec_sel(vec_madd(row4, q4, negBias), vec_madd(row4, q4, bias),
vec_cmpgt(row4, zero));
row5 = vec_sel(vec_madd(row5, q5, negBias), vec_madd(row5, q5, bias),
vec_cmpgt(row5, zero));
row6 = vec_sel(vec_madd(row6, q6, negBias), vec_madd(row6, q6, bias),
vec_cmpgt(row6, zero));
row7 = vec_sel(vec_madd(row7, q7, negBias), vec_madd(row7, q7, bias),
vec_cmpgt(row7, zero));
q0 = vec_ctf(qmat[1], QMAT_SHIFT);
q1 = vec_ctf(qmat[3], QMAT_SHIFT);
q2 = vec_ctf(qmat[5], QMAT_SHIFT);
q3 = vec_ctf(qmat[7], QMAT_SHIFT);
q4 = vec_ctf(qmat[9], QMAT_SHIFT);
q5 = vec_ctf(qmat[11], QMAT_SHIFT);
q6 = vec_ctf(qmat[13], QMAT_SHIFT);
q7 = vec_ctf(qmat[15], QMAT_SHIFT);
alt0 = vec_sel(vec_madd(alt0, q0, negBias), vec_madd(alt0, q0, bias),
vec_cmpgt(alt0, zero));
alt1 = vec_sel(vec_madd(alt1, q1, negBias), vec_madd(alt1, q1, bias),
vec_cmpgt(alt1, zero));
alt2 = vec_sel(vec_madd(alt2, q2, negBias), vec_madd(alt2, q2, bias),
vec_cmpgt(alt2, zero));
alt3 = vec_sel(vec_madd(alt3, q3, negBias), vec_madd(alt3, q3, bias),
vec_cmpgt(alt3, zero));
alt4 = vec_sel(vec_madd(alt4, q4, negBias), vec_madd(alt4, q4, bias),
vec_cmpgt(alt4, zero));
alt5 = vec_sel(vec_madd(alt5, q5, negBias), vec_madd(alt5, q5, bias),
vec_cmpgt(alt5, zero));
alt6 = vec_sel(vec_madd(alt6, q6, negBias), vec_madd(alt6, q6, bias),
vec_cmpgt(alt6, zero));
alt7 = vec_sel(vec_madd(alt7, q7, negBias), vec_madd(alt7, q7, bias),
vec_cmpgt(alt7, zero));
}
}
// Store the data back into the original block
{
vector signed short data0, data1, data2, data3, data4, data5, data6, data7;
data0 = vec_pack(vec_cts(row0, 0), vec_cts(alt0, 0));
data1 = vec_pack(vec_cts(row1, 0), vec_cts(alt1, 0));
data2 = vec_pack(vec_cts(row2, 0), vec_cts(alt2, 0));
data3 = vec_pack(vec_cts(row3, 0), vec_cts(alt3, 0));
data4 = vec_pack(vec_cts(row4, 0), vec_cts(alt4, 0));
data5 = vec_pack(vec_cts(row5, 0), vec_cts(alt5, 0));
data6 = vec_pack(vec_cts(row6, 0), vec_cts(alt6, 0));
data7 = vec_pack(vec_cts(row7, 0), vec_cts(alt7, 0));
{
// Clamp for overflow
vector signed int max_q_int, min_q_int;
vector signed short max_q, min_q;
LOAD4(max_q_int, &(s->max_qcoeff));
LOAD4(min_q_int, &(s->min_qcoeff));
max_q = vec_pack(max_q_int, max_q_int);
min_q = vec_pack(min_q_int, min_q_int);
data0 = vec_max(vec_min(data0, max_q), min_q);
data1 = vec_max(vec_min(data1, max_q), min_q);
data2 = vec_max(vec_min(data2, max_q), min_q);
data4 = vec_max(vec_min(data4, max_q), min_q);
data5 = vec_max(vec_min(data5, max_q), min_q);
data6 = vec_max(vec_min(data6, max_q), min_q);
data7 = vec_max(vec_min(data7, max_q), min_q);
}
{
vector bool char zero_01, zero_23, zero_45, zero_67;
vector signed char scanIndexes_01, scanIndexes_23, scanIndexes_45, scanIndexes_67;
vector signed char negOne = vec_splat_s8(-1);
vector signed char* scanPtr =
(vector signed char*)(s->intra_scantable.inverse);
signed char lastNonZeroChar;
// Determine the largest non-zero index.
zero_01 = vec_pack(vec_cmpeq(data0, (vector signed short)zero),
vec_cmpeq(data1, (vector signed short)zero));
zero_23 = vec_pack(vec_cmpeq(data2, (vector signed short)zero),
vec_cmpeq(data3, (vector signed short)zero));
zero_45 = vec_pack(vec_cmpeq(data4, (vector signed short)zero),
vec_cmpeq(data5, (vector signed short)zero));
zero_67 = vec_pack(vec_cmpeq(data6, (vector signed short)zero),
vec_cmpeq(data7, (vector signed short)zero));
// 64 biggest values
scanIndexes_01 = vec_sel(scanPtr[0], negOne, zero_01);
scanIndexes_23 = vec_sel(scanPtr[1], negOne, zero_23);
scanIndexes_45 = vec_sel(scanPtr[2], negOne, zero_45);
scanIndexes_67 = vec_sel(scanPtr[3], negOne, zero_67);
// 32 largest values
scanIndexes_01 = vec_max(scanIndexes_01, scanIndexes_23);
scanIndexes_45 = vec_max(scanIndexes_45, scanIndexes_67);
// 16 largest values
scanIndexes_01 = vec_max(scanIndexes_01, scanIndexes_45);
// 8 largest values
scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne),
vec_mergel(scanIndexes_01, negOne));
// 4 largest values
scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne),
vec_mergel(scanIndexes_01, negOne));
// 2 largest values
scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne),
vec_mergel(scanIndexes_01, negOne));
// largest value
scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne),
vec_mergel(scanIndexes_01, negOne));
scanIndexes_01 = vec_splat(scanIndexes_01, 0);
vec_ste(scanIndexes_01, 0, &lastNonZeroChar);
lastNonZero = lastNonZeroChar;
// While the data is still in vectors we check for the transpose IDCT permute
// and handle it using the vector unit if we can. This is the permute used
// by the altivec idct, so it is common when using the altivec dct.
if ((lastNonZero > 0) && (s->dsp.idct_permutation_type == FF_TRANSPOSE_IDCT_PERM)) {
TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7);
}
vec_st(data0, 0, data);
vec_st(data1, 16, data);
vec_st(data2, 32, data);
vec_st(data3, 48, data);
vec_st(data4, 64, data);
vec_st(data5, 80, data);
vec_st(data6, 96, data);
vec_st(data7, 112, data);
}
}
// special handling of block[0]
if (s->mb_intra) {
if (!s->h263_aic) {
if (n < 4)
oldBaseValue /= s->y_dc_scale;
else
oldBaseValue /= s->c_dc_scale;
}
// Divide by 8, rounding the result
data[0] = (oldBaseValue + 4) >> 3;
}
// We handled the transpose permutation above and we don't
// need to permute the "no" permutation case.
if ((lastNonZero > 0) &&
(s->dsp.idct_permutation_type != FF_TRANSPOSE_IDCT_PERM) &&
(s->dsp.idct_permutation_type != FF_NO_IDCT_PERM)) {
ff_block_permute(data, s->dsp.idct_permutation,
s->intra_scantable.scantable, lastNonZero);
}
return lastNonZero;
}
void dct64_altivec(real *a,real *b,real *c)
{
real __attribute__ ((aligned(16))) b1[0x20];
real __attribute__ ((aligned(16))) b2[0x20];
real *out0 = a;
real *out1 = b;
real *samples = c;
const vector float vczero = (const vector float)FOUROF(0.);
const vector unsigned char reverse = (const vector unsigned char)vcprm(3,2,1,0);
if (((unsigned long)b1 & 0x0000000F) ||
((unsigned long)b2 & 0x0000000F))
{
printf("MISALIGNED:\t%p\t%p\t%p\t%p\t%p\n",
b1, b2, a, b, samples);
}
#ifdef ALTIVEC_USE_REFERENCE_C_CODE
{
register real *costab = mp3lib_pnts[0];
b1[0x00] = samples[0x00] + samples[0x1F];
b1[0x01] = samples[0x01] + samples[0x1E];
b1[0x02] = samples[0x02] + samples[0x1D];
b1[0x03] = samples[0x03] + samples[0x1C];
b1[0x04] = samples[0x04] + samples[0x1B];
b1[0x05] = samples[0x05] + samples[0x1A];
b1[0x06] = samples[0x06] + samples[0x19];
b1[0x07] = samples[0x07] + samples[0x18];
b1[0x08] = samples[0x08] + samples[0x17];
b1[0x09] = samples[0x09] + samples[0x16];
b1[0x0A] = samples[0x0A] + samples[0x15];
b1[0x0B] = samples[0x0B] + samples[0x14];
b1[0x0C] = samples[0x0C] + samples[0x13];
b1[0x0D] = samples[0x0D] + samples[0x12];
b1[0x0E] = samples[0x0E] + samples[0x11];
b1[0x0F] = samples[0x0F] + samples[0x10];
b1[0x10] = (samples[0x0F] - samples[0x10]) * costab[0xF];
b1[0x11] = (samples[0x0E] - samples[0x11]) * costab[0xE];
b1[0x12] = (samples[0x0D] - samples[0x12]) * costab[0xD];
b1[0x13] = (samples[0x0C] - samples[0x13]) * costab[0xC];
b1[0x14] = (samples[0x0B] - samples[0x14]) * costab[0xB];
b1[0x15] = (samples[0x0A] - samples[0x15]) * costab[0xA];
b1[0x16] = (samples[0x09] - samples[0x16]) * costab[0x9];
b1[0x17] = (samples[0x08] - samples[0x17]) * costab[0x8];
b1[0x18] = (samples[0x07] - samples[0x18]) * costab[0x7];
b1[0x19] = (samples[0x06] - samples[0x19]) * costab[0x6];
b1[0x1A] = (samples[0x05] - samples[0x1A]) * costab[0x5];
b1[0x1B] = (samples[0x04] - samples[0x1B]) * costab[0x4];
b1[0x1C] = (samples[0x03] - samples[0x1C]) * costab[0x3];
b1[0x1D] = (samples[0x02] - samples[0x1D]) * costab[0x2];
b1[0x1E] = (samples[0x01] - samples[0x1E]) * costab[0x1];
b1[0x1F] = (samples[0x00] - samples[0x1F]) * costab[0x0];
}
{
register real *costab = mp3lib_pnts[1];
b2[0x00] = b1[0x00] + b1[0x0F];
b2[0x01] = b1[0x01] + b1[0x0E];
b2[0x02] = b1[0x02] + b1[0x0D];
b2[0x03] = b1[0x03] + b1[0x0C];
b2[0x04] = b1[0x04] + b1[0x0B];
b2[0x05] = b1[0x05] + b1[0x0A];
b2[0x06] = b1[0x06] + b1[0x09];
b2[0x07] = b1[0x07] + b1[0x08];
b2[0x08] = (b1[0x07] - b1[0x08]) * costab[7];
b2[0x09] = (b1[0x06] - b1[0x09]) * costab[6];
b2[0x0A] = (b1[0x05] - b1[0x0A]) * costab[5];
b2[0x0B] = (b1[0x04] - b1[0x0B]) * costab[4];
b2[0x0C] = (b1[0x03] - b1[0x0C]) * costab[3];
b2[0x0D] = (b1[0x02] - b1[0x0D]) * costab[2];
b2[0x0E] = (b1[0x01] - b1[0x0E]) * costab[1];
b2[0x0F] = (b1[0x00] - b1[0x0F]) * costab[0];
b2[0x10] = b1[0x10] + b1[0x1F];
b2[0x11] = b1[0x11] + b1[0x1E];
b2[0x12] = b1[0x12] + b1[0x1D];
b2[0x13] = b1[0x13] + b1[0x1C];
b2[0x14] = b1[0x14] + b1[0x1B];
b2[0x15] = b1[0x15] + b1[0x1A];
b2[0x16] = b1[0x16] + b1[0x19];
b2[0x17] = b1[0x17] + b1[0x18];
b2[0x18] = (b1[0x18] - b1[0x17]) * costab[7];
b2[0x19] = (b1[0x19] - b1[0x16]) * costab[6];
b2[0x1A] = (b1[0x1A] - b1[0x15]) * costab[5];
b2[0x1B] = (b1[0x1B] - b1[0x14]) * costab[4];
b2[0x1C] = (b1[0x1C] - b1[0x13]) * costab[3];
b2[0x1D] = (b1[0x1D] - b1[0x12]) * costab[2];
b2[0x1E] = (b1[0x1E] - b1[0x11]) * costab[1];
b2[0x1F] = (b1[0x1F] - b1[0x10]) * costab[0];
}
{
register real *costab = mp3lib_pnts[2];
b1[0x00] = b2[0x00] + b2[0x07];
b1[0x01] = b2[0x01] + b2[0x06];
b1[0x02] = b2[0x02] + b2[0x05];
b1[0x03] = b2[0x03] + b2[0x04];
b1[0x04] = (b2[0x03] - b2[0x04]) * costab[3];
b1[0x05] = (b2[0x02] - b2[0x05]) * costab[2];
b1[0x06] = (b2[0x01] - b2[0x06]) * costab[1];
b1[0x07] = (b2[0x00] - b2[0x07]) * costab[0];
b1[0x08] = b2[0x08] + b2[0x0F];
b1[0x09] = b2[0x09] + b2[0x0E];
b1[0x0A] = b2[0x0A] + b2[0x0D];
b1[0x0B] = b2[0x0B] + b2[0x0C];
b1[0x0C] = (b2[0x0C] - b2[0x0B]) * costab[3];
b1[0x0D] = (b2[0x0D] - b2[0x0A]) * costab[2];
b1[0x0E] = (b2[0x0E] - b2[0x09]) * costab[1];
b1[0x0F] = (b2[0x0F] - b2[0x08]) * costab[0];
b1[0x10] = b2[0x10] + b2[0x17];
b1[0x11] = b2[0x11] + b2[0x16];
b1[0x12] = b2[0x12] + b2[0x15];
b1[0x13] = b2[0x13] + b2[0x14];
b1[0x14] = (b2[0x13] - b2[0x14]) * costab[3];
b1[0x15] = (b2[0x12] - b2[0x15]) * costab[2];
b1[0x16] = (b2[0x11] - b2[0x16]) * costab[1];
b1[0x17] = (b2[0x10] - b2[0x17]) * costab[0];
b1[0x18] = b2[0x18] + b2[0x1F];
b1[0x19] = b2[0x19] + b2[0x1E];
b1[0x1A] = b2[0x1A] + b2[0x1D];
b1[0x1B] = b2[0x1B] + b2[0x1C];
b1[0x1C] = (b2[0x1C] - b2[0x1B]) * costab[3];
b1[0x1D] = (b2[0x1D] - b2[0x1A]) * costab[2];
b1[0x1E] = (b2[0x1E] - b2[0x19]) * costab[1];
b1[0x1F] = (b2[0x1F] - b2[0x18]) * costab[0];
}
#else /* ALTIVEC_USE_REFERENCE_C_CODE */
// How does it work ?
// the first three passes are reproducted in the three block below
// all computations are done on a 4 elements vector
// 'reverse' is a special perumtation vector used to reverse
// the order of the elements inside a vector.
// note that all loads/stores to b1 (b2) between passes 1 and 2 (2 and 3)
// have been removed, all elements are stored inside b1vX (b2vX)
{
register vector float
b1v0, b1v1, b1v2, b1v3,
b1v4, b1v5, b1v6, b1v7;
register vector float
temp1, temp2;
{
register real *costab = mp3lib_pnts[0];
register vector float
samplesv1, samplesv2, samplesv3, samplesv4,
samplesv5, samplesv6, samplesv7, samplesv8,
samplesv9;
register vector unsigned char samples_perm = vec_lvsl(0, samples);
register vector float costabv1, costabv2, costabv3, costabv4, costabv5;
register vector unsigned char costab_perm = vec_lvsl(0, costab);
samplesv1 = vec_ld(0, samples);
samplesv2 = vec_ld(16, samples);
samplesv1 = vec_perm(samplesv1, samplesv2, samples_perm);
samplesv3 = vec_ld(32, samples);
samplesv2 = vec_perm(samplesv2, samplesv3, samples_perm);
samplesv4 = vec_ld(48, samples);
samplesv3 = vec_perm(samplesv3, samplesv4, samples_perm);
samplesv5 = vec_ld(64, samples);
samplesv4 = vec_perm(samplesv4, samplesv5, samples_perm);
samplesv6 = vec_ld(80, samples);
samplesv5 = vec_perm(samplesv5, samplesv6, samples_perm);
samplesv7 = vec_ld(96, samples);
samplesv6 = vec_perm(samplesv6, samplesv7, samples_perm);
samplesv8 = vec_ld(112, samples);
samplesv7 = vec_perm(samplesv7, samplesv8, samples_perm);
samplesv9 = vec_ld(128, samples);
samplesv8 = vec_perm(samplesv8, samplesv9, samples_perm);
temp1 = vec_add(samplesv1,
vec_perm(samplesv8, samplesv8, reverse));
//vec_st(temp1, 0, b1);
b1v0 = temp1;
temp1 = vec_add(samplesv2,
vec_perm(samplesv7, samplesv7, reverse));
//vec_st(temp1, 16, b1);
b1v1 = temp1;
temp1 = vec_add(samplesv3,
vec_perm(samplesv6, samplesv6, reverse));
//vec_st(temp1, 32, b1);
b1v2 = temp1;
temp1 = vec_add(samplesv4,
vec_perm(samplesv5, samplesv5, reverse));
//vec_st(temp1, 48, b1);
b1v3 = temp1;
costabv1 = vec_ld(0, costab);
costabv2 = vec_ld(16, costab);
costabv1 = vec_perm(costabv1, costabv2, costab_perm);
costabv3 = vec_ld(32, costab);
costabv2 = vec_perm(costabv2, costabv3, costab_perm);
costabv4 = vec_ld(48, costab);
costabv3 = vec_perm(costabv3, costabv4, costab_perm);
costabv5 = vec_ld(64, costab);
costabv4 = vec_perm(costabv4, costabv5, costab_perm);
temp1 = vec_sub(vec_perm(samplesv4, samplesv4, reverse),
samplesv5);
temp2 = vec_madd(temp1,
vec_perm(costabv4, costabv4, reverse),
vczero);
//vec_st(temp2, 64, b1);
b1v4 = temp2;
temp1 = vec_sub(vec_perm(samplesv3, samplesv3, reverse),
samplesv6);
temp2 = vec_madd(temp1,
vec_perm(costabv3, costabv3, reverse),
vczero);
//vec_st(temp2, 80, b1);
b1v5 = temp2;
temp1 = vec_sub(vec_perm(samplesv2, samplesv2, reverse),
samplesv7);
temp2 = vec_madd(temp1,
vec_perm(costabv2, costabv2, reverse),
vczero);
//vec_st(temp2, 96, b1);
b1v6 = temp2;
temp1 = vec_sub(vec_perm(samplesv1, samplesv1, reverse),
samplesv8);
temp2 = vec_madd(temp1,
vec_perm(costabv1, costabv1, reverse),
vczero);
//vec_st(temp2, 112, b1);
b1v7 = temp2;
}
{
register vector float
b2v0, b2v1, b2v2, b2v3,
b2v4, b2v5, b2v6, b2v7;
{
register real *costab = mp3lib_pnts[1];
register vector float costabv1r, costabv2r, costabv1, costabv2, costabv3;
register vector unsigned char costab_perm = vec_lvsl(0, costab);
costabv1 = vec_ld(0, costab);
costabv2 = vec_ld(16, costab);
costabv1 = vec_perm(costabv1, costabv2, costab_perm);
costabv3 = vec_ld(32, costab);
costabv2 = vec_perm(costabv2, costabv3 , costab_perm);
costabv1r = vec_perm(costabv1, costabv1, reverse);
costabv2r = vec_perm(costabv2, costabv2, reverse);
temp1 = vec_add(b1v0, vec_perm(b1v3, b1v3, reverse));
//vec_st(temp1, 0, b2);
b2v0 = temp1;
temp1 = vec_add(b1v1, vec_perm(b1v2, b1v2, reverse));
//vec_st(temp1, 16, b2);
b2v1 = temp1;
temp2 = vec_sub(vec_perm(b1v1, b1v1, reverse), b1v2);
temp1 = vec_madd(temp2, costabv2r, vczero);
//vec_st(temp1, 32, b2);
b2v2 = temp1;
temp2 = vec_sub(vec_perm(b1v0, b1v0, reverse), b1v3);
temp1 = vec_madd(temp2, costabv1r, vczero);
//vec_st(temp1, 48, b2);
b2v3 = temp1;
temp1 = vec_add(b1v4, vec_perm(b1v7, b1v7, reverse));
//vec_st(temp1, 64, b2);
b2v4 = temp1;
temp1 = vec_add(b1v5, vec_perm(b1v6, b1v6, reverse));
//vec_st(temp1, 80, b2);
b2v5 = temp1;
temp2 = vec_sub(b1v6, vec_perm(b1v5, b1v5, reverse));
temp1 = vec_madd(temp2, costabv2r, vczero);
//vec_st(temp1, 96, b2);
b2v6 = temp1;
temp2 = vec_sub(b1v7, vec_perm(b1v4, b1v4, reverse));
temp1 = vec_madd(temp2, costabv1r, vczero);
//vec_st(temp1, 112, b2);
b2v7 = temp1;
}
{
register real *costab = mp3lib_pnts[2];
vector float costabv1r, costabv1, costabv2;
vector unsigned char costab_perm = vec_lvsl(0, costab);
costabv1 = vec_ld(0, costab);
costabv2 = vec_ld(16, costab);
costabv1 = vec_perm(costabv1, costabv2, costab_perm);
costabv1r = vec_perm(costabv1, costabv1, reverse);
temp1 = vec_add(b2v0, vec_perm(b2v1, b2v1, reverse));
vec_st(temp1, 0, b1);
temp2 = vec_sub(vec_perm(b2v0, b2v0, reverse), b2v1);
temp1 = vec_madd(temp2, costabv1r, vczero);
vec_st(temp1, 16, b1);
temp1 = vec_add(b2v2, vec_perm(b2v3, b2v3, reverse));
vec_st(temp1, 32, b1);
temp2 = vec_sub(b2v3, vec_perm(b2v2, b2v2, reverse));
temp1 = vec_madd(temp2, costabv1r, vczero);
vec_st(temp1, 48, b1);
temp1 = vec_add(b2v4, vec_perm(b2v5, b2v5, reverse));
vec_st(temp1, 64, b1);
temp2 = vec_sub(vec_perm(b2v4, b2v4, reverse), b2v5);
temp1 = vec_madd(temp2, costabv1r, vczero);
vec_st(temp1, 80, b1);
temp1 = vec_add(b2v6, vec_perm(b2v7, b2v7, reverse));
vec_st(temp1, 96, b1);
temp2 = vec_sub(b2v7, vec_perm(b2v6, b2v6, reverse));
temp1 = vec_madd(temp2, costabv1r, vczero);
vec_st(temp1, 112, b1);
}
}
}
#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
{
register real const cos0 = mp3lib_pnts[3][0];
register real const cos1 = mp3lib_pnts[3][1];
b2[0x00] = b1[0x00] + b1[0x03];
b2[0x01] = b1[0x01] + b1[0x02];
b2[0x02] = (b1[0x01] - b1[0x02]) * cos1;
b2[0x03] = (b1[0x00] - b1[0x03]) * cos0;
b2[0x04] = b1[0x04] + b1[0x07];
b2[0x05] = b1[0x05] + b1[0x06];
b2[0x06] = (b1[0x06] - b1[0x05]) * cos1;
b2[0x07] = (b1[0x07] - b1[0x04]) * cos0;
b2[0x08] = b1[0x08] + b1[0x0B];
b2[0x09] = b1[0x09] + b1[0x0A];
b2[0x0A] = (b1[0x09] - b1[0x0A]) * cos1;
b2[0x0B] = (b1[0x08] - b1[0x0B]) * cos0;
b2[0x0C] = b1[0x0C] + b1[0x0F];
b2[0x0D] = b1[0x0D] + b1[0x0E];
b2[0x0E] = (b1[0x0E] - b1[0x0D]) * cos1;
b2[0x0F] = (b1[0x0F] - b1[0x0C]) * cos0;
b2[0x10] = b1[0x10] + b1[0x13];
b2[0x11] = b1[0x11] + b1[0x12];
b2[0x12] = (b1[0x11] - b1[0x12]) * cos1;
b2[0x13] = (b1[0x10] - b1[0x13]) * cos0;
b2[0x14] = b1[0x14] + b1[0x17];
b2[0x15] = b1[0x15] + b1[0x16];
b2[0x16] = (b1[0x16] - b1[0x15]) * cos1;
b2[0x17] = (b1[0x17] - b1[0x14]) * cos0;
b2[0x18] = b1[0x18] + b1[0x1B];
b2[0x19] = b1[0x19] + b1[0x1A];
b2[0x1A] = (b1[0x19] - b1[0x1A]) * cos1;
b2[0x1B] = (b1[0x18] - b1[0x1B]) * cos0;
b2[0x1C] = b1[0x1C] + b1[0x1F];
b2[0x1D] = b1[0x1D] + b1[0x1E];
b2[0x1E] = (b1[0x1E] - b1[0x1D]) * cos1;
b2[0x1F] = (b1[0x1F] - b1[0x1C]) * cos0;
}
{
register real const cos0 = mp3lib_pnts[4][0];
b1[0x00] = b2[0x00] + b2[0x01];
b1[0x01] = (b2[0x00] - b2[0x01]) * cos0;
b1[0x02] = b2[0x02] + b2[0x03];
b1[0x03] = (b2[0x03] - b2[0x02]) * cos0;
b1[0x02] += b1[0x03];
b1[0x04] = b2[0x04] + b2[0x05];
b1[0x05] = (b2[0x04] - b2[0x05]) * cos0;
b1[0x06] = b2[0x06] + b2[0x07];
b1[0x07] = (b2[0x07] - b2[0x06]) * cos0;
b1[0x06] += b1[0x07];
b1[0x04] += b1[0x06];
b1[0x06] += b1[0x05];
b1[0x05] += b1[0x07];
b1[0x08] = b2[0x08] + b2[0x09];
b1[0x09] = (b2[0x08] - b2[0x09]) * cos0;
b1[0x0A] = b2[0x0A] + b2[0x0B];
b1[0x0B] = (b2[0x0B] - b2[0x0A]) * cos0;
b1[0x0A] += b1[0x0B];
b1[0x0C] = b2[0x0C] + b2[0x0D];
b1[0x0D] = (b2[0x0C] - b2[0x0D]) * cos0;
b1[0x0E] = b2[0x0E] + b2[0x0F];
b1[0x0F] = (b2[0x0F] - b2[0x0E]) * cos0;
b1[0x0E] += b1[0x0F];
b1[0x0C] += b1[0x0E];
b1[0x0E] += b1[0x0D];
b1[0x0D] += b1[0x0F];
b1[0x10] = b2[0x10] + b2[0x11];
b1[0x11] = (b2[0x10] - b2[0x11]) * cos0;
b1[0x12] = b2[0x12] + b2[0x13];
b1[0x13] = (b2[0x13] - b2[0x12]) * cos0;
b1[0x12] += b1[0x13];
b1[0x14] = b2[0x14] + b2[0x15];
b1[0x15] = (b2[0x14] - b2[0x15]) * cos0;
b1[0x16] = b2[0x16] + b2[0x17];
b1[0x17] = (b2[0x17] - b2[0x16]) * cos0;
b1[0x16] += b1[0x17];
b1[0x14] += b1[0x16];
b1[0x16] += b1[0x15];
b1[0x15] += b1[0x17];
b1[0x18] = b2[0x18] + b2[0x19];
b1[0x19] = (b2[0x18] - b2[0x19]) * cos0;
b1[0x1A] = b2[0x1A] + b2[0x1B];
b1[0x1B] = (b2[0x1B] - b2[0x1A]) * cos0;
b1[0x1A] += b1[0x1B];
b1[0x1C] = b2[0x1C] + b2[0x1D];
b1[0x1D] = (b2[0x1C] - b2[0x1D]) * cos0;
b1[0x1E] = b2[0x1E] + b2[0x1F];
b1[0x1F] = (b2[0x1F] - b2[0x1E]) * cos0;
b1[0x1E] += b1[0x1F];
b1[0x1C] += b1[0x1E];
b1[0x1E] += b1[0x1D];
b1[0x1D] += b1[0x1F];
}
out0[0x10*16] = b1[0x00];
out0[0x10*12] = b1[0x04];
out0[0x10* 8] = b1[0x02];
out0[0x10* 4] = b1[0x06];
out0[0x10* 0] = b1[0x01];
out1[0x10* 0] = b1[0x01];
out1[0x10* 4] = b1[0x05];
out1[0x10* 8] = b1[0x03];
out1[0x10*12] = b1[0x07];
b1[0x08] += b1[0x0C];
out0[0x10*14] = b1[0x08];
b1[0x0C] += b1[0x0a];
out0[0x10*10] = b1[0x0C];
b1[0x0A] += b1[0x0E];
out0[0x10* 6] = b1[0x0A];
b1[0x0E] += b1[0x09];
out0[0x10* 2] = b1[0x0E];
b1[0x09] += b1[0x0D];
out1[0x10* 2] = b1[0x09];
b1[0x0D] += b1[0x0B];
out1[0x10* 6] = b1[0x0D];
b1[0x0B] += b1[0x0F];
out1[0x10*10] = b1[0x0B];
out1[0x10*14] = b1[0x0F];
b1[0x18] += b1[0x1C];
out0[0x10*15] = b1[0x10] + b1[0x18];
out0[0x10*13] = b1[0x18] + b1[0x14];
b1[0x1C] += b1[0x1a];
out0[0x10*11] = b1[0x14] + b1[0x1C];
out0[0x10* 9] = b1[0x1C] + b1[0x12];
b1[0x1A] += b1[0x1E];
out0[0x10* 7] = b1[0x12] + b1[0x1A];
out0[0x10* 5] = b1[0x1A] + b1[0x16];
b1[0x1E] += b1[0x19];
out0[0x10* 3] = b1[0x16] + b1[0x1E];
out0[0x10* 1] = b1[0x1E] + b1[0x11];
b1[0x19] += b1[0x1D];
out1[0x10* 1] = b1[0x11] + b1[0x19];
out1[0x10* 3] = b1[0x19] + b1[0x15];
b1[0x1D] += b1[0x1B];
out1[0x10* 5] = b1[0x15] + b1[0x1D];
out1[0x10* 7] = b1[0x1D] + b1[0x13];
b1[0x1B] += b1[0x1F];
out1[0x10* 9] = b1[0x13] + b1[0x1B];
out1[0x10*11] = b1[0x1B] + b1[0x17];
out1[0x10*13] = b1[0x17] + b1[0x1F];
out1[0x10*15] = b1[0x1F];
}
