24
24
#include <stdlib.h>
26
27
#include "libavutil/cpu.h"
28
#include "libavutil/ppc/types_altivec.h"
29
#include "libavutil/ppc/util_altivec.h"
27
30
#include "libavcodec/dsputil.h"
28
31
#include "libavcodec/mpegvideo.h"
30
#include "util_altivec.h"
31
#include "types_altivec.h"
32
33
#include "dsputil_altivec.h"
34
// Swaps two variables (used for altivec registers)
37
__typeof__(a) swap_temp=a; \
42
// transposes a matrix consisting of four vectors with four elements each
43
#define TRANSPOSE4(a,b,c,d) \
45
__typeof__(a) _trans_ach = vec_mergeh(a, c); \
46
__typeof__(a) _trans_acl = vec_mergel(a, c); \
47
__typeof__(a) _trans_bdh = vec_mergeh(b, d); \
48
__typeof__(a) _trans_bdl = vec_mergel(b, d); \
50
a = vec_mergeh(_trans_ach, _trans_bdh); \
51
b = vec_mergel(_trans_ach, _trans_bdh); \
52
c = vec_mergeh(_trans_acl, _trans_bdl); \
53
d = vec_mergel(_trans_acl, _trans_bdl); \
57
// Loads a four-byte value (int or float) from the target address
58
// into every element in the target vector. Only works if the
59
// target address is four-byte aligned (which should be always).
60
#define LOAD4(vec, address) \
62
__typeof__(vec)* _load_addr = (__typeof__(vec)*)(address); \
63
vector unsigned char _perm_vec = vec_lvsl(0,(address)); \
64
vec = vec_ld(0, _load_addr); \
65
vec = vec_perm(vec, vec, _perm_vec); \
66
vec = vec_splat(vec, 0); \
70
#define FOUROF(a) {a,a,a,a}
72
static int dct_quantize_altivec(MpegEncContext* s,
74
int qscale, int* overflow)
77
vector float row0, row1, row2, row3, row4, row5, row6, row7;
78
vector float alt0, alt1, alt2, alt3, alt4, alt5, alt6, alt7;
79
const vector float zero = (const vector float)FOUROF(0.);
80
// used after quantize step
83
// Load the data into the row/alt vectors
85
vector signed short data0, data1, data2, data3, data4, data5, data6, data7;
87
data0 = vec_ld(0, data);
88
data1 = vec_ld(16, data);
89
data2 = vec_ld(32, data);
90
data3 = vec_ld(48, data);
91
data4 = vec_ld(64, data);
92
data5 = vec_ld(80, data);
93
data6 = vec_ld(96, data);
94
data7 = vec_ld(112, data);
96
// Transpose the data before we start
97
TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7);
99
// load the data into floating point vectors. We load
100
// the high half of each row into the main row vectors
101
// and the low half into the alt vectors.
102
row0 = vec_ctf(vec_unpackh(data0), 0);
103
alt0 = vec_ctf(vec_unpackl(data0), 0);
104
row1 = vec_ctf(vec_unpackh(data1), 0);
105
alt1 = vec_ctf(vec_unpackl(data1), 0);
106
row2 = vec_ctf(vec_unpackh(data2), 0);
107
alt2 = vec_ctf(vec_unpackl(data2), 0);
108
row3 = vec_ctf(vec_unpackh(data3), 0);
109
alt3 = vec_ctf(vec_unpackl(data3), 0);
110
row4 = vec_ctf(vec_unpackh(data4), 0);
111
alt4 = vec_ctf(vec_unpackl(data4), 0);
112
row5 = vec_ctf(vec_unpackh(data5), 0);
113
alt5 = vec_ctf(vec_unpackl(data5), 0);
114
row6 = vec_ctf(vec_unpackh(data6), 0);
115
alt6 = vec_ctf(vec_unpackl(data6), 0);
116
row7 = vec_ctf(vec_unpackh(data7), 0);
117
alt7 = vec_ctf(vec_unpackl(data7), 0);
120
// The following block could exist as a separate an altivec dct
121
// function. However, if we put it inline, the DCT data can remain
122
// in the vector local variables, as floats, which we'll use during the
125
const vector float vec_0_298631336 = (vector float)FOUROF(0.298631336f);
126
const vector float vec_0_390180644 = (vector float)FOUROF(-0.390180644f);
127
const vector float vec_0_541196100 = (vector float)FOUROF(0.541196100f);
128
const vector float vec_0_765366865 = (vector float)FOUROF(0.765366865f);
129
const vector float vec_0_899976223 = (vector float)FOUROF(-0.899976223f);
130
const vector float vec_1_175875602 = (vector float)FOUROF(1.175875602f);
131
const vector float vec_1_501321110 = (vector float)FOUROF(1.501321110f);
132
const vector float vec_1_847759065 = (vector float)FOUROF(-1.847759065f);
133
const vector float vec_1_961570560 = (vector float)FOUROF(-1.961570560f);
134
const vector float vec_2_053119869 = (vector float)FOUROF(2.053119869f);
135
const vector float vec_2_562915447 = (vector float)FOUROF(-2.562915447f);
136
const vector float vec_3_072711026 = (vector float)FOUROF(3.072711026f);
139
int whichPass, whichHalf;
141
for(whichPass = 1; whichPass<=2; whichPass++) {
142
for(whichHalf = 1; whichHalf<=2; whichHalf++) {
143
vector float tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
144
vector float tmp10, tmp11, tmp12, tmp13;
145
vector float z1, z2, z3, z4, z5;
147
tmp0 = vec_add(row0, row7); // tmp0 = dataptr[0] + dataptr[7];
148
tmp7 = vec_sub(row0, row7); // tmp7 = dataptr[0] - dataptr[7];
149
tmp3 = vec_add(row3, row4); // tmp3 = dataptr[3] + dataptr[4];
150
tmp4 = vec_sub(row3, row4); // tmp4 = dataptr[3] - dataptr[4];
151
tmp1 = vec_add(row1, row6); // tmp1 = dataptr[1] + dataptr[6];
152
tmp6 = vec_sub(row1, row6); // tmp6 = dataptr[1] - dataptr[6];
153
tmp2 = vec_add(row2, row5); // tmp2 = dataptr[2] + dataptr[5];
154
tmp5 = vec_sub(row2, row5); // tmp5 = dataptr[2] - dataptr[5];
156
tmp10 = vec_add(tmp0, tmp3); // tmp10 = tmp0 + tmp3;
157
tmp13 = vec_sub(tmp0, tmp3); // tmp13 = tmp0 - tmp3;
158
tmp11 = vec_add(tmp1, tmp2); // tmp11 = tmp1 + tmp2;
159
tmp12 = vec_sub(tmp1, tmp2); // tmp12 = tmp1 - tmp2;
162
// dataptr[0] = (DCTELEM) ((tmp10 + tmp11) << PASS1_BITS);
163
row0 = vec_add(tmp10, tmp11);
165
// dataptr[4] = (DCTELEM) ((tmp10 - tmp11) << PASS1_BITS);
166
row4 = vec_sub(tmp10, tmp11);
169
// z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
170
z1 = vec_madd(vec_add(tmp12, tmp13), vec_0_541196100, (vector float)zero);
172
// dataptr[2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
173
// CONST_BITS-PASS1_BITS);
174
row2 = vec_madd(tmp13, vec_0_765366865, z1);
176
// dataptr[6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
177
// CONST_BITS-PASS1_BITS);
178
row6 = vec_madd(tmp12, vec_1_847759065, z1);
180
z1 = vec_add(tmp4, tmp7); // z1 = tmp4 + tmp7;
181
z2 = vec_add(tmp5, tmp6); // z2 = tmp5 + tmp6;
182
z3 = vec_add(tmp4, tmp6); // z3 = tmp4 + tmp6;
183
z4 = vec_add(tmp5, tmp7); // z4 = tmp5 + tmp7;
185
// z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
186
z5 = vec_madd(vec_add(z3, z4), vec_1_175875602, (vector float)zero);
188
// z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
189
z3 = vec_madd(z3, vec_1_961570560, z5);
191
// z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
192
z4 = vec_madd(z4, vec_0_390180644, z5);
194
// The following adds are rolled into the multiplies above
195
// z3 = vec_add(z3, z5); // z3 += z5;
196
// z4 = vec_add(z4, z5); // z4 += z5;
198
// z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
199
// Wow! It's actually more efficient to roll this multiply
200
// into the adds below, even thought the multiply gets done twice!
201
// z2 = vec_madd(z2, vec_2_562915447, (vector float)zero);
203
// z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
204
// Same with this one...
205
// z1 = vec_madd(z1, vec_0_899976223, (vector float)zero);
207
// tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
208
// dataptr[7] = (DCTELEM) DESCALE(tmp4 + z1 + z3, CONST_BITS-PASS1_BITS);
209
row7 = vec_madd(tmp4, vec_0_298631336, vec_madd(z1, vec_0_899976223, z3));
211
// tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
212
// dataptr[5] = (DCTELEM) DESCALE(tmp5 + z2 + z4, CONST_BITS-PASS1_BITS);
213
row5 = vec_madd(tmp5, vec_2_053119869, vec_madd(z2, vec_2_562915447, z4));
215
// tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
216
// dataptr[3] = (DCTELEM) DESCALE(tmp6 + z2 + z3, CONST_BITS-PASS1_BITS);
217
row3 = vec_madd(tmp6, vec_3_072711026, vec_madd(z2, vec_2_562915447, z3));
219
// tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
220
// dataptr[1] = (DCTELEM) DESCALE(tmp7 + z1 + z4, CONST_BITS-PASS1_BITS);
221
row1 = vec_madd(z1, vec_0_899976223, vec_madd(tmp7, vec_1_501321110, z4));
223
// Swap the row values with the alts. If this is the first half,
224
// this sets up the low values to be acted on in the second half.
225
// If this is the second half, it puts the high values back in
226
// the row values where they are expected to be when we're done.
237
if (whichPass == 1) {
238
// transpose the data for the second pass
240
// First, block transpose the upper right with lower left.
246
// Now, transpose each block of four
247
TRANSPOSE4(row0, row1, row2, row3);
248
TRANSPOSE4(row4, row5, row6, row7);
249
TRANSPOSE4(alt0, alt1, alt2, alt3);
250
TRANSPOSE4(alt4, alt5, alt6, alt7);
255
// perform the quantize step, using the floating point data
256
// still in the row/alt registers
259
const vector signed int* qmat;
260
vector float bias, negBias;
263
vector signed int baseVector;
265
// We must cache element 0 in the intra case
266
// (it needs special handling).
267
baseVector = vec_cts(vec_splat(row0, 0), 0);
268
vec_ste(baseVector, 0, &oldBaseValue);
270
qmat = (vector signed int*)s->q_intra_matrix[qscale];
271
biasAddr = &s->intra_quant_bias;
273
qmat = (vector signed int*)s->q_inter_matrix[qscale];
274
biasAddr = &s->inter_quant_bias;
277
// Load the bias vector (We add 0.5 to the bias so that we're
278
// rounding when we convert to int, instead of flooring.)
280
vector signed int biasInt;
281
const vector float negOneFloat = (vector float)FOUROF(-1.0f);
282
LOAD4(biasInt, biasAddr);
283
bias = vec_ctf(biasInt, QUANT_BIAS_SHIFT);
284
negBias = vec_madd(bias, negOneFloat, zero);
288
vector float q0, q1, q2, q3, q4, q5, q6, q7;
290
q0 = vec_ctf(qmat[0], QMAT_SHIFT);
291
q1 = vec_ctf(qmat[2], QMAT_SHIFT);
292
q2 = vec_ctf(qmat[4], QMAT_SHIFT);
293
q3 = vec_ctf(qmat[6], QMAT_SHIFT);
294
q4 = vec_ctf(qmat[8], QMAT_SHIFT);
295
q5 = vec_ctf(qmat[10], QMAT_SHIFT);
296
q6 = vec_ctf(qmat[12], QMAT_SHIFT);
297
q7 = vec_ctf(qmat[14], QMAT_SHIFT);
299
row0 = vec_sel(vec_madd(row0, q0, negBias), vec_madd(row0, q0, bias),
300
vec_cmpgt(row0, zero));
301
row1 = vec_sel(vec_madd(row1, q1, negBias), vec_madd(row1, q1, bias),
302
vec_cmpgt(row1, zero));
303
row2 = vec_sel(vec_madd(row2, q2, negBias), vec_madd(row2, q2, bias),
304
vec_cmpgt(row2, zero));
305
row3 = vec_sel(vec_madd(row3, q3, negBias), vec_madd(row3, q3, bias),
306
vec_cmpgt(row3, zero));
307
row4 = vec_sel(vec_madd(row4, q4, negBias), vec_madd(row4, q4, bias),
308
vec_cmpgt(row4, zero));
309
row5 = vec_sel(vec_madd(row5, q5, negBias), vec_madd(row5, q5, bias),
310
vec_cmpgt(row5, zero));
311
row6 = vec_sel(vec_madd(row6, q6, negBias), vec_madd(row6, q6, bias),
312
vec_cmpgt(row6, zero));
313
row7 = vec_sel(vec_madd(row7, q7, negBias), vec_madd(row7, q7, bias),
314
vec_cmpgt(row7, zero));
316
q0 = vec_ctf(qmat[1], QMAT_SHIFT);
317
q1 = vec_ctf(qmat[3], QMAT_SHIFT);
318
q2 = vec_ctf(qmat[5], QMAT_SHIFT);
319
q3 = vec_ctf(qmat[7], QMAT_SHIFT);
320
q4 = vec_ctf(qmat[9], QMAT_SHIFT);
321
q5 = vec_ctf(qmat[11], QMAT_SHIFT);
322
q6 = vec_ctf(qmat[13], QMAT_SHIFT);
323
q7 = vec_ctf(qmat[15], QMAT_SHIFT);
325
alt0 = vec_sel(vec_madd(alt0, q0, negBias), vec_madd(alt0, q0, bias),
326
vec_cmpgt(alt0, zero));
327
alt1 = vec_sel(vec_madd(alt1, q1, negBias), vec_madd(alt1, q1, bias),
328
vec_cmpgt(alt1, zero));
329
alt2 = vec_sel(vec_madd(alt2, q2, negBias), vec_madd(alt2, q2, bias),
330
vec_cmpgt(alt2, zero));
331
alt3 = vec_sel(vec_madd(alt3, q3, negBias), vec_madd(alt3, q3, bias),
332
vec_cmpgt(alt3, zero));
333
alt4 = vec_sel(vec_madd(alt4, q4, negBias), vec_madd(alt4, q4, bias),
334
vec_cmpgt(alt4, zero));
335
alt5 = vec_sel(vec_madd(alt5, q5, negBias), vec_madd(alt5, q5, bias),
336
vec_cmpgt(alt5, zero));
337
alt6 = vec_sel(vec_madd(alt6, q6, negBias), vec_madd(alt6, q6, bias),
338
vec_cmpgt(alt6, zero));
339
alt7 = vec_sel(vec_madd(alt7, q7, negBias), vec_madd(alt7, q7, bias),
340
vec_cmpgt(alt7, zero));
346
// Store the data back into the original block
348
vector signed short data0, data1, data2, data3, data4, data5, data6, data7;
350
data0 = vec_pack(vec_cts(row0, 0), vec_cts(alt0, 0));
351
data1 = vec_pack(vec_cts(row1, 0), vec_cts(alt1, 0));
352
data2 = vec_pack(vec_cts(row2, 0), vec_cts(alt2, 0));
353
data3 = vec_pack(vec_cts(row3, 0), vec_cts(alt3, 0));
354
data4 = vec_pack(vec_cts(row4, 0), vec_cts(alt4, 0));
355
data5 = vec_pack(vec_cts(row5, 0), vec_cts(alt5, 0));
356
data6 = vec_pack(vec_cts(row6, 0), vec_cts(alt6, 0));
357
data7 = vec_pack(vec_cts(row7, 0), vec_cts(alt7, 0));
360
// Clamp for overflow
361
vector signed int max_q_int, min_q_int;
362
vector signed short max_q, min_q;
364
LOAD4(max_q_int, &s->max_qcoeff);
365
LOAD4(min_q_int, &s->min_qcoeff);
367
max_q = vec_pack(max_q_int, max_q_int);
368
min_q = vec_pack(min_q_int, min_q_int);
370
data0 = vec_max(vec_min(data0, max_q), min_q);
371
data1 = vec_max(vec_min(data1, max_q), min_q);
372
data2 = vec_max(vec_min(data2, max_q), min_q);
373
data4 = vec_max(vec_min(data4, max_q), min_q);
374
data5 = vec_max(vec_min(data5, max_q), min_q);
375
data6 = vec_max(vec_min(data6, max_q), min_q);
376
data7 = vec_max(vec_min(data7, max_q), min_q);
380
vector bool char zero_01, zero_23, zero_45, zero_67;
381
vector signed char scanIndexes_01, scanIndexes_23, scanIndexes_45, scanIndexes_67;
382
vector signed char negOne = vec_splat_s8(-1);
383
vector signed char* scanPtr =
384
(vector signed char*)(s->intra_scantable.inverse);
385
signed char lastNonZeroChar;
387
// Determine the largest non-zero index.
388
zero_01 = vec_pack(vec_cmpeq(data0, (vector signed short)zero),
389
vec_cmpeq(data1, (vector signed short)zero));
390
zero_23 = vec_pack(vec_cmpeq(data2, (vector signed short)zero),
391
vec_cmpeq(data3, (vector signed short)zero));
392
zero_45 = vec_pack(vec_cmpeq(data4, (vector signed short)zero),
393
vec_cmpeq(data5, (vector signed short)zero));
394
zero_67 = vec_pack(vec_cmpeq(data6, (vector signed short)zero),
395
vec_cmpeq(data7, (vector signed short)zero));
398
scanIndexes_01 = vec_sel(scanPtr[0], negOne, zero_01);
399
scanIndexes_23 = vec_sel(scanPtr[1], negOne, zero_23);
400
scanIndexes_45 = vec_sel(scanPtr[2], negOne, zero_45);
401
scanIndexes_67 = vec_sel(scanPtr[3], negOne, zero_67);
404
scanIndexes_01 = vec_max(scanIndexes_01, scanIndexes_23);
405
scanIndexes_45 = vec_max(scanIndexes_45, scanIndexes_67);
408
scanIndexes_01 = vec_max(scanIndexes_01, scanIndexes_45);
411
scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne),
412
vec_mergel(scanIndexes_01, negOne));
415
scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne),
416
vec_mergel(scanIndexes_01, negOne));
419
scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne),
420
vec_mergel(scanIndexes_01, negOne));
423
scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne),
424
vec_mergel(scanIndexes_01, negOne));
426
scanIndexes_01 = vec_splat(scanIndexes_01, 0);
429
vec_ste(scanIndexes_01, 0, &lastNonZeroChar);
431
lastNonZero = lastNonZeroChar;
433
// While the data is still in vectors we check for the transpose IDCT permute
434
// and handle it using the vector unit if we can. This is the permute used
435
// by the altivec idct, so it is common when using the altivec dct.
437
if ((lastNonZero > 0) && (s->dsp.idct_permutation_type == FF_TRANSPOSE_IDCT_PERM)) {
438
TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7);
441
vec_st(data0, 0, data);
442
vec_st(data1, 16, data);
443
vec_st(data2, 32, data);
444
vec_st(data3, 48, data);
445
vec_st(data4, 64, data);
446
vec_st(data5, 80, data);
447
vec_st(data6, 96, data);
448
vec_st(data7, 112, data);
452
// special handling of block[0]
456
oldBaseValue /= s->y_dc_scale;
458
oldBaseValue /= s->c_dc_scale;
461
// Divide by 8, rounding the result
462
data[0] = (oldBaseValue + 4) >> 3;
465
// We handled the transpose permutation above and we don't
466
// need to permute the "no" permutation case.
467
if ((lastNonZero > 0) &&
468
(s->dsp.idct_permutation_type != FF_TRANSPOSE_IDCT_PERM) &&
469
(s->dsp.idct_permutation_type != FF_NO_IDCT_PERM)) {
470
ff_block_permute(data, s->dsp.idct_permutation,
471
s->intra_scantable.scantable, lastNonZero);
477
35
/* AltiVec version of dct_unquantize_h263
478
36
this code assumes `block' is 16 bytes-aligned */
479
37
static void dct_unquantize_h263_altivec(MpegEncContext *s,