2 * Copyright (c) 2002 Dieter Shirley
4 * This library is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU Lesser General Public
6 * License as published by the Free Software Foundation; either
7 * version 2 of the License, or (at your option) any later version.
9 * This library is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
12 * Lesser General Public License for more details.
14 * You should have received a copy of the GNU Lesser General Public
15 * License along with this library; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
21 #include "../dsputil.h"
22 #include "../mpegvideo.h"
24 // Swaps two variables (used for altivec registers)
27 __typeof__(a) swap_temp=a; \
32 // transposes a matrix consisting of four vectors with four elements each
33 #define TRANSPOSE4(a,b,c,d) \
35 __typeof__(a) _trans_ach = vec_mergeh(a, c); \
36 __typeof__(a) _trans_acl = vec_mergel(a, c); \
37 __typeof__(a) _trans_bdh = vec_mergeh(b, d); \
38 __typeof__(a) _trans_bdl = vec_mergel(b, d); \
40 a = vec_mergeh(_trans_ach, _trans_bdh); \
41 b = vec_mergel(_trans_ach, _trans_bdh); \
42 c = vec_mergeh(_trans_acl, _trans_bdl); \
43 d = vec_mergel(_trans_acl, _trans_bdl); \
46 #define TRANSPOSE8(a,b,c,d,e,f,g,h) \
48 __typeof__(a) _A1, _B1, _C1, _D1, _E1, _F1, _G1, _H1; \
49 __typeof__(a) _A2, _B2, _C2, _D2, _E2, _F2, _G2, _H2; \
51 _A1 = vec_mergeh (a, e); \
52 _B1 = vec_mergel (a, e); \
53 _C1 = vec_mergeh (b, f); \
54 _D1 = vec_mergel (b, f); \
55 _E1 = vec_mergeh (c, g); \
56 _F1 = vec_mergel (c, g); \
57 _G1 = vec_mergeh (d, h); \
58 _H1 = vec_mergel (d, h); \
60 _A2 = vec_mergeh (_A1, _E1); \
61 _B2 = vec_mergel (_A1, _E1); \
62 _C2 = vec_mergeh (_B1, _F1); \
63 _D2 = vec_mergel (_B1, _F1); \
64 _E2 = vec_mergeh (_C1, _G1); \
65 _F2 = vec_mergel (_C1, _G1); \
66 _G2 = vec_mergeh (_D1, _H1); \
67 _H2 = vec_mergel (_D1, _H1); \
69 a = vec_mergeh (_A2, _E2); \
70 b = vec_mergel (_A2, _E2); \
71 c = vec_mergeh (_B2, _F2); \
72 d = vec_mergel (_B2, _F2); \
73 e = vec_mergeh (_C2, _G2); \
74 f = vec_mergel (_C2, _G2); \
75 g = vec_mergeh (_D2, _H2); \
76 h = vec_mergel (_D2, _H2); \
80 // Loads a four-byte value (int or float) from the target address
81 // into every element in the target vector. Only works if the
82 // target address is four-byte aligned (which should be always).
83 #define LOAD4(vec, address) \
85 __typeof__(vec)* _load_addr = (__typeof__(vec)*)(address); \
86 vector unsigned char _perm_vec = vec_lvsl(0,(address)); \
87 vec = vec_ld(0, _load_addr); \
88 vec = vec_perm(vec, vec, _perm_vec); \
89 vec = vec_splat(vec, 0); \
92 int dct_quantize_altivec(MpegEncContext* s,
94 int qscale, int* overflow)
97 vector float row0, row1, row2, row3, row4, row5, row6, row7;
98 vector float alt0, alt1, alt2, alt3, alt4, alt5, alt6, alt7;
99 const vector float zero = (const vector float)(0.0f);
101 // Load the data into the row/alt vectors
103 vector signed short data0, data1, data2, data3, data4, data5, data6, data7;
105 data0 = vec_ld(0, data);
106 data1 = vec_ld(16, data);
107 data2 = vec_ld(32, data);
108 data3 = vec_ld(48, data);
109 data4 = vec_ld(64, data);
110 data5 = vec_ld(80, data);
111 data6 = vec_ld(96, data);
112 data7 = vec_ld(112, data);
114 // Transpose the data before we start
115 TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7);
117 // load the data into floating point vectors. We load
118 // the high half of each row into the main row vectors
119 // and the low half into the alt vectors.
120 row0 = vec_ctf(vec_unpackh(data0), 0);
121 alt0 = vec_ctf(vec_unpackl(data0), 0);
122 row1 = vec_ctf(vec_unpackh(data1), 0);
123 alt1 = vec_ctf(vec_unpackl(data1), 0);
124 row2 = vec_ctf(vec_unpackh(data2), 0);
125 alt2 = vec_ctf(vec_unpackl(data2), 0);
126 row3 = vec_ctf(vec_unpackh(data3), 0);
127 alt3 = vec_ctf(vec_unpackl(data3), 0);
128 row4 = vec_ctf(vec_unpackh(data4), 0);
129 alt4 = vec_ctf(vec_unpackl(data4), 0);
130 row5 = vec_ctf(vec_unpackh(data5), 0);
131 alt5 = vec_ctf(vec_unpackl(data5), 0);
132 row6 = vec_ctf(vec_unpackh(data6), 0);
133 alt6 = vec_ctf(vec_unpackl(data6), 0);
134 row7 = vec_ctf(vec_unpackh(data7), 0);
135 alt7 = vec_ctf(vec_unpackl(data7), 0);
138 // The following block could exist as a separate an altivec dct
139 // function. However, if we put it inline, the DCT data can remain
140 // in the vector local variables, as floats, which we'll use during the
143 const vector float vec_0_298631336 = (vector float)(0.298631336f);
144 const vector float vec_0_390180644 = (vector float)(-0.390180644f);
145 const vector float vec_0_541196100 = (vector float)(0.541196100f);
146 const vector float vec_0_765366865 = (vector float)(0.765366865f);
147 const vector float vec_0_899976223 = (vector float)(-0.899976223f);
148 const vector float vec_1_175875602 = (vector float)(1.175875602f);
149 const vector float vec_1_501321110 = (vector float)(1.501321110f);
150 const vector float vec_1_847759065 = (vector float)(-1.847759065f);
151 const vector float vec_1_961570560 = (vector float)(-1.961570560f);
152 const vector float vec_2_053119869 = (vector float)(2.053119869f);
153 const vector float vec_2_562915447 = (vector float)(-2.562915447f);
154 const vector float vec_3_072711026 = (vector float)(3.072711026f);
157 int whichPass, whichHalf;
159 for(whichPass = 1; whichPass<=2; whichPass++)
161 for(whichHalf = 1; whichHalf<=2; whichHalf++)
163 vector float tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
164 vector float tmp10, tmp11, tmp12, tmp13;
165 vector float z1, z2, z3, z4, z5;
167 tmp0 = vec_add(row0, row7); // tmp0 = dataptr[0] + dataptr[7];
168 tmp7 = vec_sub(row0, row7); // tmp7 = dataptr[0] - dataptr[7];
169 tmp3 = vec_add(row3, row4); // tmp3 = dataptr[3] + dataptr[4];
170 tmp4 = vec_sub(row3, row4); // tmp4 = dataptr[3] - dataptr[4];
171 tmp1 = vec_add(row1, row6); // tmp1 = dataptr[1] + dataptr[6];
172 tmp6 = vec_sub(row1, row6); // tmp6 = dataptr[1] - dataptr[6];
173 tmp2 = vec_add(row2, row5); // tmp2 = dataptr[2] + dataptr[5];
174 tmp5 = vec_sub(row2, row5); // tmp5 = dataptr[2] - dataptr[5];
176 tmp10 = vec_add(tmp0, tmp3); // tmp10 = tmp0 + tmp3;
177 tmp13 = vec_sub(tmp0, tmp3); // tmp13 = tmp0 - tmp3;
178 tmp11 = vec_add(tmp1, tmp2); // tmp11 = tmp1 + tmp2;
179 tmp12 = vec_sub(tmp1, tmp2); // tmp12 = tmp1 - tmp2;
182 // dataptr[0] = (DCTELEM) ((tmp10 + tmp11) << PASS1_BITS);
183 row0 = vec_add(tmp10, tmp11);
185 // dataptr[4] = (DCTELEM) ((tmp10 - tmp11) << PASS1_BITS);
186 row4 = vec_sub(tmp10, tmp11);
189 // z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
190 z1 = vec_madd(vec_add(tmp12, tmp13), vec_0_541196100, (vector float)zero);
192 // dataptr[2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
193 // CONST_BITS-PASS1_BITS);
194 row2 = vec_madd(tmp13, vec_0_765366865, z1);
196 // dataptr[6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
197 // CONST_BITS-PASS1_BITS);
198 row6 = vec_madd(tmp12, vec_1_847759065, z1);
200 z1 = vec_add(tmp4, tmp7); // z1 = tmp4 + tmp7;
201 z2 = vec_add(tmp5, tmp6); // z2 = tmp5 + tmp6;
202 z3 = vec_add(tmp4, tmp6); // z3 = tmp4 + tmp6;
203 z4 = vec_add(tmp5, tmp7); // z4 = tmp5 + tmp7;
205 // z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
206 z5 = vec_madd(vec_add(z3, z4), vec_1_175875602, (vector float)zero);
208 // z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
209 z3 = vec_madd(z3, vec_1_961570560, z5);
211 // z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
212 z4 = vec_madd(z4, vec_0_390180644, z5);
214 // The following adds are rolled into the multiplies above
215 // z3 = vec_add(z3, z5); // z3 += z5;
216 // z4 = vec_add(z4, z5); // z4 += z5;
218 // z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
219 // Wow! It's actually more effecient to roll this multiply
220 // into the adds below, even thought the multiply gets done twice!
221 // z2 = vec_madd(z2, vec_2_562915447, (vector float)zero);
223 // z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
224 // Same with this one...
225 // z1 = vec_madd(z1, vec_0_899976223, (vector float)zero);
227 // tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
228 // dataptr[7] = (DCTELEM) DESCALE(tmp4 + z1 + z3, CONST_BITS-PASS1_BITS);
229 row7 = vec_madd(tmp4, vec_0_298631336, vec_madd(z1, vec_0_899976223, z3));
231 // tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
232 // dataptr[5] = (DCTELEM) DESCALE(tmp5 + z2 + z4, CONST_BITS-PASS1_BITS);
233 row5 = vec_madd(tmp5, vec_2_053119869, vec_madd(z2, vec_2_562915447, z4));
235 // tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
236 // dataptr[3] = (DCTELEM) DESCALE(tmp6 + z2 + z3, CONST_BITS-PASS1_BITS);
237 row3 = vec_madd(tmp6, vec_3_072711026, vec_madd(z2, vec_2_562915447, z3));
239 // tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
240 // dataptr[1] = (DCTELEM) DESCALE(tmp7 + z1 + z4, CONST_BITS-PASS1_BITS);
241 row1 = vec_madd(z1, vec_0_899976223, vec_madd(tmp7, vec_1_501321110, z4));
243 // Swap the row values with the alts. If this is the first half,
244 // this sets up the low values to be acted on in the second half.
245 // If this is the second half, it puts the high values back in
246 // the row values where they are expected to be when we're done.
259 // transpose the data for the second pass
261 // First, block transpose the upper right with lower left.
267 // Now, transpose each block of four
268 TRANSPOSE4(row0, row1, row2, row3);
269 TRANSPOSE4(row4, row5, row6, row7);
270 TRANSPOSE4(alt0, alt1, alt2, alt3);
271 TRANSPOSE4(alt4, alt5, alt6, alt7);
276 // used after quantise step
277 int oldBaseValue = 0;
279 // perform the quantise step, using the floating point data
280 // still in the row/alt registers
283 const vector signed int* qmat;
284 vector float bias, negBias;
288 vector signed int baseVector;
290 // We must cache element 0 in the intra case
291 // (it needs special handling).
292 baseVector = vec_cts(vec_splat(row0, 0), 0);
293 vec_ste(baseVector, 0, &oldBaseValue);
295 qmat = (vector signed int*)s->q_intra_matrix[qscale];
296 biasAddr = &(s->intra_quant_bias);
300 qmat = (vector signed int*)s->q_inter_matrix[qscale];
301 biasAddr = &(s->inter_quant_bias);
304 // Load the bias vector (We add 0.5 to the bias so that we're
305 // rounding when we convert to int, instead of flooring.)
307 vector signed int biasInt;
308 const vector float negOneFloat = (vector float)(-1.0f);
309 LOAD4(biasInt, biasAddr);
310 bias = vec_ctf(biasInt, QUANT_BIAS_SHIFT);
311 negBias = vec_madd(bias, negOneFloat, zero);
315 vector float q0, q1, q2, q3, q4, q5, q6, q7;
317 q0 = vec_ctf(qmat[0], QMAT_SHIFT);
318 q1 = vec_ctf(qmat[2], QMAT_SHIFT);
319 q2 = vec_ctf(qmat[4], QMAT_SHIFT);
320 q3 = vec_ctf(qmat[6], QMAT_SHIFT);
321 q4 = vec_ctf(qmat[8], QMAT_SHIFT);
322 q5 = vec_ctf(qmat[10], QMAT_SHIFT);
323 q6 = vec_ctf(qmat[12], QMAT_SHIFT);
324 q7 = vec_ctf(qmat[14], QMAT_SHIFT);
326 row0 = vec_sel(vec_madd(row0, q0, negBias), vec_madd(row0, q0, bias),
327 vec_cmpgt(row0, zero));
328 row1 = vec_sel(vec_madd(row1, q1, negBias), vec_madd(row1, q1, bias),
329 vec_cmpgt(row1, zero));
330 row2 = vec_sel(vec_madd(row2, q2, negBias), vec_madd(row2, q2, bias),
331 vec_cmpgt(row2, zero));
332 row3 = vec_sel(vec_madd(row3, q3, negBias), vec_madd(row3, q3, bias),
333 vec_cmpgt(row3, zero));
334 row4 = vec_sel(vec_madd(row4, q4, negBias), vec_madd(row4, q4, bias),
335 vec_cmpgt(row4, zero));
336 row5 = vec_sel(vec_madd(row5, q5, negBias), vec_madd(row5, q5, bias),
337 vec_cmpgt(row5, zero));
338 row6 = vec_sel(vec_madd(row6, q6, negBias), vec_madd(row6, q6, bias),
339 vec_cmpgt(row6, zero));
340 row7 = vec_sel(vec_madd(row7, q7, negBias), vec_madd(row7, q7, bias),
341 vec_cmpgt(row7, zero));
343 q0 = vec_ctf(qmat[1], QMAT_SHIFT);
344 q1 = vec_ctf(qmat[3], QMAT_SHIFT);
345 q2 = vec_ctf(qmat[5], QMAT_SHIFT);
346 q3 = vec_ctf(qmat[7], QMAT_SHIFT);
347 q4 = vec_ctf(qmat[9], QMAT_SHIFT);
348 q5 = vec_ctf(qmat[11], QMAT_SHIFT);
349 q6 = vec_ctf(qmat[13], QMAT_SHIFT);
350 q7 = vec_ctf(qmat[15], QMAT_SHIFT);
352 alt0 = vec_sel(vec_madd(alt0, q0, negBias), vec_madd(alt0, q0, bias),
353 vec_cmpgt(alt0, zero));
354 alt1 = vec_sel(vec_madd(alt1, q1, negBias), vec_madd(alt1, q1, bias),
355 vec_cmpgt(alt1, zero));
356 alt2 = vec_sel(vec_madd(alt2, q2, negBias), vec_madd(alt2, q2, bias),
357 vec_cmpgt(alt2, zero));
358 alt3 = vec_sel(vec_madd(alt3, q3, negBias), vec_madd(alt3, q3, bias),
359 vec_cmpgt(alt3, zero));
360 alt4 = vec_sel(vec_madd(alt4, q4, negBias), vec_madd(alt4, q4, bias),
361 vec_cmpgt(alt4, zero));
362 alt5 = vec_sel(vec_madd(alt5, q5, negBias), vec_madd(alt5, q5, bias),
363 vec_cmpgt(alt5, zero));
364 alt6 = vec_sel(vec_madd(alt6, q6, negBias), vec_madd(alt6, q6, bias),
365 vec_cmpgt(alt6, zero));
366 alt7 = vec_sel(vec_madd(alt7, q7, negBias), vec_madd(alt7, q7, bias),
367 vec_cmpgt(alt7, zero));
373 // Store the data back into the original block
375 vector signed short data0, data1, data2, data3, data4, data5, data6, data7;
377 data0 = vec_pack(vec_cts(row0, 0), vec_cts(alt0, 0));
378 data1 = vec_pack(vec_cts(row1, 0), vec_cts(alt1, 0));
379 data2 = vec_pack(vec_cts(row2, 0), vec_cts(alt2, 0));
380 data3 = vec_pack(vec_cts(row3, 0), vec_cts(alt3, 0));
381 data4 = vec_pack(vec_cts(row4, 0), vec_cts(alt4, 0));
382 data5 = vec_pack(vec_cts(row5, 0), vec_cts(alt5, 0));
383 data6 = vec_pack(vec_cts(row6, 0), vec_cts(alt6, 0));
384 data7 = vec_pack(vec_cts(row7, 0), vec_cts(alt7, 0));
387 // Clamp for overflow
388 vector signed int max_q_int, min_q_int;
389 vector signed short max_q, min_q;
391 LOAD4(max_q_int, &(s->max_qcoeff));
392 LOAD4(min_q_int, &(s->min_qcoeff));
394 max_q = vec_pack(max_q_int, max_q_int);
395 min_q = vec_pack(min_q_int, min_q_int);
397 data0 = vec_max(vec_min(data0, max_q), min_q);
398 data1 = vec_max(vec_min(data1, max_q), min_q);
399 data2 = vec_max(vec_min(data2, max_q), min_q);
400 data4 = vec_max(vec_min(data4, max_q), min_q);
401 data5 = vec_max(vec_min(data5, max_q), min_q);
402 data6 = vec_max(vec_min(data6, max_q), min_q);
403 data7 = vec_max(vec_min(data7, max_q), min_q);
406 vector bool char zero_01, zero_23, zero_45, zero_67;
407 vector signed char scanIndices_01, scanIndices_23, scanIndices_45, scanIndices_67;
408 vector signed char negOne = vec_splat_s8(-1);
409 vector signed char* scanPtr =
410 (vector signed char*)(s->intra_scantable.inverse);
412 // Determine the largest non-zero index.
413 zero_01 = vec_pack(vec_cmpeq(data0, (vector short)zero),
414 vec_cmpeq(data1, (vector short)zero));
415 zero_23 = vec_pack(vec_cmpeq(data2, (vector short)zero),
416 vec_cmpeq(data3, (vector short)zero));
417 zero_45 = vec_pack(vec_cmpeq(data4, (vector short)zero),
418 vec_cmpeq(data5, (vector short)zero));
419 zero_67 = vec_pack(vec_cmpeq(data6, (vector short)zero),
420 vec_cmpeq(data7, (vector short)zero));
423 scanIndices_01 = vec_sel(scanPtr[0], negOne, zero_01);
424 scanIndices_23 = vec_sel(scanPtr[1], negOne, zero_23);
425 scanIndices_45 = vec_sel(scanPtr[2], negOne, zero_45);
426 scanIndices_67 = vec_sel(scanPtr[3], negOne, zero_67);
429 scanIndices_01 = vec_max(scanIndices_01, scanIndices_23);
430 scanIndices_45 = vec_max(scanIndices_45, scanIndices_67);
433 scanIndices_01 = vec_max(scanIndices_01, scanIndices_45);
436 scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
437 vec_mergel(scanIndices_01, negOne));
440 scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
441 vec_mergel(scanIndices_01, negOne));
444 scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
445 vec_mergel(scanIndices_01, negOne));
448 scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
449 vec_mergel(scanIndices_01, negOne));
451 scanIndices_01 = vec_splat(scanIndices_01, 0);
453 signed char lastNonZeroChar;
455 vec_ste(scanIndices_01, 0, &lastNonZeroChar);
457 lastNonZero = lastNonZeroChar;
459 // While the data is still in vectors we check for the transpose IDCT permute
460 // and handle it using the vector unit if we can. This is the permute used
461 // by the altivec idct, so it is common when using the altivec dct.
463 if ((lastNonZero > 0) && (s->idct_permutation_type == FF_TRANSPOSE_IDCT_PERM))
465 TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7);
468 vec_st(data0, 0, data);
469 vec_st(data1, 16, data);
470 vec_st(data2, 32, data);
471 vec_st(data3, 48, data);
472 vec_st(data4, 64, data);
473 vec_st(data5, 80, data);
474 vec_st(data6, 96, data);
475 vec_st(data7, 112, data);
478 // special handling of block[0]
484 oldBaseValue /= s->y_dc_scale;
486 oldBaseValue /= s->c_dc_scale;
489 // Divide by 8, rounding the result
490 data[0] = (oldBaseValue + 4) >> 3;
493 // We handled the tranpose permutation above and we don't
494 // need to permute the "no" permutation case.
495 if ((lastNonZero > 0) &&
496 (s->idct_permutation_type != FF_TRANSPOSE_IDCT_PERM) &&
497 (s->idct_permutation_type != FF_NO_IDCT_PERM))
499 ff_block_permute(data, s->idct_permutation,
500 s->intra_scantable.scantable, lastNonZero);