00001 /* 00002 * Copyright (c) 2002 Dieter Shirley 00003 * 00004 * dct_unquantize_h263_altivec: 00005 * Copyright (c) 2003 Romain Dolbeau <romain@dolbeau.org> 00006 * 00007 * This file is part of FFmpeg. 00008 * 00009 * FFmpeg is free software; you can redistribute it and/or 00010 * modify it under the terms of the GNU Lesser General Public 00011 * License as published by the Free Software Foundation; either 00012 * version 2.1 of the License, or (at your option) any later version. 00013 * 00014 * FFmpeg is distributed in the hope that it will be useful, 00015 * but WITHOUT ANY WARRANTY; without even the implied warranty of 00016 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 00017 * Lesser General Public License for more details. 00018 * 00019 * You should have received a copy of the GNU Lesser General Public 00020 * License along with FFmpeg; if not, write to the Free Software 00021 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA 00022 */ 00023 00024 #include <stdlib.h> 00025 #include <stdio.h> 00026 #include "libavcodec/dsputil.h" 00027 #include "libavcodec/mpegvideo.h" 00028 00029 #include "gcc_fixes.h" 00030 00031 #include "dsputil_ppc.h" 00032 #include "util_altivec.h" 00033 // Swaps two variables (used for altivec registers) 00034 #define SWAP(a,b) \ 00035 do { \ 00036 __typeof__(a) swap_temp=a; \ 00037 a=b; \ 00038 b=swap_temp; \ 00039 } while (0) 00040 00041 // transposes a matrix consisting of four vectors with four elements each 00042 #define TRANSPOSE4(a,b,c,d) \ 00043 do { \ 00044 __typeof__(a) _trans_ach = vec_mergeh(a, c); \ 00045 __typeof__(a) _trans_acl = vec_mergel(a, c); \ 00046 __typeof__(a) _trans_bdh = vec_mergeh(b, d); \ 00047 __typeof__(a) _trans_bdl = vec_mergel(b, d); \ 00048 \ 00049 a = vec_mergeh(_trans_ach, _trans_bdh); \ 00050 b = vec_mergel(_trans_ach, _trans_bdh); \ 00051 c = vec_mergeh(_trans_acl, _trans_bdl); \ 00052 d = vec_mergel(_trans_acl, _trans_bdl); \ 00053 } while (0) 00054 00055 00056 // Loads a four-byte value (int or float) from the target address 00057 // into every element in the target vector. Only works if the 00058 // target address is four-byte aligned (which should be always). 00059 #define LOAD4(vec, address) \ 00060 { \ 00061 __typeof__(vec)* _load_addr = (__typeof__(vec)*)(address); \ 00062 vector unsigned char _perm_vec = vec_lvsl(0,(address)); \ 00063 vec = vec_ld(0, _load_addr); \ 00064 vec = vec_perm(vec, vec, _perm_vec); \ 00065 vec = vec_splat(vec, 0); \ 00066 } 00067 00068 00069 #define FOUROF(a) {a,a,a,a} 00070 00071 int dct_quantize_altivec(MpegEncContext* s, 00072 DCTELEM* data, int n, 00073 int qscale, int* overflow) 00074 { 00075 int lastNonZero; 00076 vector float row0, row1, row2, row3, row4, row5, row6, row7; 00077 vector float alt0, alt1, alt2, alt3, alt4, alt5, alt6, alt7; 00078 const vector float zero = (const vector float)FOUROF(0.); 00079 // used after quantize step 00080 int oldBaseValue = 0; 00081 00082 // Load the data into the row/alt vectors 00083 { 00084 vector signed short data0, data1, data2, data3, data4, data5, data6, data7; 00085 00086 data0 = vec_ld(0, data); 00087 data1 = vec_ld(16, data); 00088 data2 = vec_ld(32, data); 00089 data3 = vec_ld(48, data); 00090 data4 = vec_ld(64, data); 00091 data5 = vec_ld(80, data); 00092 data6 = vec_ld(96, data); 00093 data7 = vec_ld(112, data); 00094 00095 // Transpose the data before we start 00096 TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7); 00097 00098 // load the data into floating point vectors. We load 00099 // the high half of each row into the main row vectors 00100 // and the low half into the alt vectors. 00101 row0 = vec_ctf(vec_unpackh(data0), 0); 00102 alt0 = vec_ctf(vec_unpackl(data0), 0); 00103 row1 = vec_ctf(vec_unpackh(data1), 0); 00104 alt1 = vec_ctf(vec_unpackl(data1), 0); 00105 row2 = vec_ctf(vec_unpackh(data2), 0); 00106 alt2 = vec_ctf(vec_unpackl(data2), 0); 00107 row3 = vec_ctf(vec_unpackh(data3), 0); 00108 alt3 = vec_ctf(vec_unpackl(data3), 0); 00109 row4 = vec_ctf(vec_unpackh(data4), 0); 00110 alt4 = vec_ctf(vec_unpackl(data4), 0); 00111 row5 = vec_ctf(vec_unpackh(data5), 0); 00112 alt5 = vec_ctf(vec_unpackl(data5), 0); 00113 row6 = vec_ctf(vec_unpackh(data6), 0); 00114 alt6 = vec_ctf(vec_unpackl(data6), 0); 00115 row7 = vec_ctf(vec_unpackh(data7), 0); 00116 alt7 = vec_ctf(vec_unpackl(data7), 0); 00117 } 00118 00119 // The following block could exist as a separate an altivec dct 00120 // function. However, if we put it inline, the DCT data can remain 00121 // in the vector local variables, as floats, which we'll use during the 00122 // quantize step... 00123 { 00124 const vector float vec_0_298631336 = (vector float)FOUROF(0.298631336f); 00125 const vector float vec_0_390180644 = (vector float)FOUROF(-0.390180644f); 00126 const vector float vec_0_541196100 = (vector float)FOUROF(0.541196100f); 00127 const vector float vec_0_765366865 = (vector float)FOUROF(0.765366865f); 00128 const vector float vec_0_899976223 = (vector float)FOUROF(-0.899976223f); 00129 const vector float vec_1_175875602 = (vector float)FOUROF(1.175875602f); 00130 const vector float vec_1_501321110 = (vector float)FOUROF(1.501321110f); 00131 const vector float vec_1_847759065 = (vector float)FOUROF(-1.847759065f); 00132 const vector float vec_1_961570560 = (vector float)FOUROF(-1.961570560f); 00133 const vector float vec_2_053119869 = (vector float)FOUROF(2.053119869f); 00134 const vector float vec_2_562915447 = (vector float)FOUROF(-2.562915447f); 00135 const vector float vec_3_072711026 = (vector float)FOUROF(3.072711026f); 00136 00137 00138 int whichPass, whichHalf; 00139 00140 for(whichPass = 1; whichPass<=2; whichPass++) { 00141 for(whichHalf = 1; whichHalf<=2; whichHalf++) { 00142 vector float tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; 00143 vector float tmp10, tmp11, tmp12, tmp13; 00144 vector float z1, z2, z3, z4, z5; 00145 00146 tmp0 = vec_add(row0, row7); // tmp0 = dataptr[0] + dataptr[7]; 00147 tmp7 = vec_sub(row0, row7); // tmp7 = dataptr[0] - dataptr[7]; 00148 tmp3 = vec_add(row3, row4); // tmp3 = dataptr[3] + dataptr[4]; 00149 tmp4 = vec_sub(row3, row4); // tmp4 = dataptr[3] - dataptr[4]; 00150 tmp1 = vec_add(row1, row6); // tmp1 = dataptr[1] + dataptr[6]; 00151 tmp6 = vec_sub(row1, row6); // tmp6 = dataptr[1] - dataptr[6]; 00152 tmp2 = vec_add(row2, row5); // tmp2 = dataptr[2] + dataptr[5]; 00153 tmp5 = vec_sub(row2, row5); // tmp5 = dataptr[2] - dataptr[5]; 00154 00155 tmp10 = vec_add(tmp0, tmp3); // tmp10 = tmp0 + tmp3; 00156 tmp13 = vec_sub(tmp0, tmp3); // tmp13 = tmp0 - tmp3; 00157 tmp11 = vec_add(tmp1, tmp2); // tmp11 = tmp1 + tmp2; 00158 tmp12 = vec_sub(tmp1, tmp2); // tmp12 = tmp1 - tmp2; 00159 00160 00161 // dataptr[0] = (DCTELEM) ((tmp10 + tmp11) << PASS1_BITS); 00162 row0 = vec_add(tmp10, tmp11); 00163 00164 // dataptr[4] = (DCTELEM) ((tmp10 - tmp11) << PASS1_BITS); 00165 row4 = vec_sub(tmp10, tmp11); 00166 00167 00168 // z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100); 00169 z1 = vec_madd(vec_add(tmp12, tmp13), vec_0_541196100, (vector float)zero); 00170 00171 // dataptr[2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865), 00172 // CONST_BITS-PASS1_BITS); 00173 row2 = vec_madd(tmp13, vec_0_765366865, z1); 00174 00175 // dataptr[6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065), 00176 // CONST_BITS-PASS1_BITS); 00177 row6 = vec_madd(tmp12, vec_1_847759065, z1); 00178 00179 z1 = vec_add(tmp4, tmp7); // z1 = tmp4 + tmp7; 00180 z2 = vec_add(tmp5, tmp6); // z2 = tmp5 + tmp6; 00181 z3 = vec_add(tmp4, tmp6); // z3 = tmp4 + tmp6; 00182 z4 = vec_add(tmp5, tmp7); // z4 = tmp5 + tmp7; 00183 00184 // z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */ 00185 z5 = vec_madd(vec_add(z3, z4), vec_1_175875602, (vector float)zero); 00186 00187 // z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */ 00188 z3 = vec_madd(z3, vec_1_961570560, z5); 00189 00190 // z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */ 00191 z4 = vec_madd(z4, vec_0_390180644, z5); 00192 00193 // The following adds are rolled into the multiplies above 00194 // z3 = vec_add(z3, z5); // z3 += z5; 00195 // z4 = vec_add(z4, z5); // z4 += z5; 00196 00197 // z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */ 00198 // Wow! It's actually more efficient to roll this multiply 00199 // into the adds below, even thought the multiply gets done twice! 00200 // z2 = vec_madd(z2, vec_2_562915447, (vector float)zero); 00201 00202 // z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */ 00203 // Same with this one... 00204 // z1 = vec_madd(z1, vec_0_899976223, (vector float)zero); 00205 00206 // tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */ 00207 // dataptr[7] = (DCTELEM) DESCALE(tmp4 + z1 + z3, CONST_BITS-PASS1_BITS); 00208 row7 = vec_madd(tmp4, vec_0_298631336, vec_madd(z1, vec_0_899976223, z3)); 00209 00210 // tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */ 00211 // dataptr[5] = (DCTELEM) DESCALE(tmp5 + z2 + z4, CONST_BITS-PASS1_BITS); 00212 row5 = vec_madd(tmp5, vec_2_053119869, vec_madd(z2, vec_2_562915447, z4)); 00213 00214 // tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */ 00215 // dataptr[3] = (DCTELEM) DESCALE(tmp6 + z2 + z3, CONST_BITS-PASS1_BITS); 00216 row3 = vec_madd(tmp6, vec_3_072711026, vec_madd(z2, vec_2_562915447, z3)); 00217 00218 // tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */ 00219 // dataptr[1] = (DCTELEM) DESCALE(tmp7 + z1 + z4, CONST_BITS-PASS1_BITS); 00220 row1 = vec_madd(z1, vec_0_899976223, vec_madd(tmp7, vec_1_501321110, z4)); 00221 00222 // Swap the row values with the alts. If this is the first half, 00223 // this sets up the low values to be acted on in the second half. 00224 // If this is the second half, it puts the high values back in 00225 // the row values where they are expected to be when we're done. 00226 SWAP(row0, alt0); 00227 SWAP(row1, alt1); 00228 SWAP(row2, alt2); 00229 SWAP(row3, alt3); 00230 SWAP(row4, alt4); 00231 SWAP(row5, alt5); 00232 SWAP(row6, alt6); 00233 SWAP(row7, alt7); 00234 } 00235 00236 if (whichPass == 1) { 00237 // transpose the data for the second pass 00238 00239 // First, block transpose the upper right with lower left. 00240 SWAP(row4, alt0); 00241 SWAP(row5, alt1); 00242 SWAP(row6, alt2); 00243 SWAP(row7, alt3); 00244 00245 // Now, transpose each block of four 00246 TRANSPOSE4(row0, row1, row2, row3); 00247 TRANSPOSE4(row4, row5, row6, row7); 00248 TRANSPOSE4(alt0, alt1, alt2, alt3); 00249 TRANSPOSE4(alt4, alt5, alt6, alt7); 00250 } 00251 } 00252 } 00253 00254 // perform the quantize step, using the floating point data 00255 // still in the row/alt registers 00256 { 00257 const int* biasAddr; 00258 const vector signed int* qmat; 00259 vector float bias, negBias; 00260 00261 if (s->mb_intra) { 00262 vector signed int baseVector; 00263 00264 // We must cache element 0 in the intra case 00265 // (it needs special handling). 00266 baseVector = vec_cts(vec_splat(row0, 0), 0); 00267 vec_ste(baseVector, 0, &oldBaseValue); 00268 00269 qmat = (vector signed int*)s->q_intra_matrix[qscale]; 00270 biasAddr = &(s->intra_quant_bias); 00271 } else { 00272 qmat = (vector signed int*)s->q_inter_matrix[qscale]; 00273 biasAddr = &(s->inter_quant_bias); 00274 } 00275 00276 // Load the bias vector (We add 0.5 to the bias so that we're 00277 // rounding when we convert to int, instead of flooring.) 00278 { 00279 vector signed int biasInt; 00280 const vector float negOneFloat = (vector float)FOUROF(-1.0f); 00281 LOAD4(biasInt, biasAddr); 00282 bias = vec_ctf(biasInt, QUANT_BIAS_SHIFT); 00283 negBias = vec_madd(bias, negOneFloat, zero); 00284 } 00285 00286 { 00287 vector float q0, q1, q2, q3, q4, q5, q6, q7; 00288 00289 q0 = vec_ctf(qmat[0], QMAT_SHIFT); 00290 q1 = vec_ctf(qmat[2], QMAT_SHIFT); 00291 q2 = vec_ctf(qmat[4], QMAT_SHIFT); 00292 q3 = vec_ctf(qmat[6], QMAT_SHIFT); 00293 q4 = vec_ctf(qmat[8], QMAT_SHIFT); 00294 q5 = vec_ctf(qmat[10], QMAT_SHIFT); 00295 q6 = vec_ctf(qmat[12], QMAT_SHIFT); 00296 q7 = vec_ctf(qmat[14], QMAT_SHIFT); 00297 00298 row0 = vec_sel(vec_madd(row0, q0, negBias), vec_madd(row0, q0, bias), 00299 vec_cmpgt(row0, zero)); 00300 row1 = vec_sel(vec_madd(row1, q1, negBias), vec_madd(row1, q1, bias), 00301 vec_cmpgt(row1, zero)); 00302 row2 = vec_sel(vec_madd(row2, q2, negBias), vec_madd(row2, q2, bias), 00303 vec_cmpgt(row2, zero)); 00304 row3 = vec_sel(vec_madd(row3, q3, negBias), vec_madd(row3, q3, bias), 00305 vec_cmpgt(row3, zero)); 00306 row4 = vec_sel(vec_madd(row4, q4, negBias), vec_madd(row4, q4, bias), 00307 vec_cmpgt(row4, zero)); 00308 row5 = vec_sel(vec_madd(row5, q5, negBias), vec_madd(row5, q5, bias), 00309 vec_cmpgt(row5, zero)); 00310 row6 = vec_sel(vec_madd(row6, q6, negBias), vec_madd(row6, q6, bias), 00311 vec_cmpgt(row6, zero)); 00312 row7 = vec_sel(vec_madd(row7, q7, negBias), vec_madd(row7, q7, bias), 00313 vec_cmpgt(row7, zero)); 00314 00315 q0 = vec_ctf(qmat[1], QMAT_SHIFT); 00316 q1 = vec_ctf(qmat[3], QMAT_SHIFT); 00317 q2 = vec_ctf(qmat[5], QMAT_SHIFT); 00318 q3 = vec_ctf(qmat[7], QMAT_SHIFT); 00319 q4 = vec_ctf(qmat[9], QMAT_SHIFT); 00320 q5 = vec_ctf(qmat[11], QMAT_SHIFT); 00321 q6 = vec_ctf(qmat[13], QMAT_SHIFT); 00322 q7 = vec_ctf(qmat[15], QMAT_SHIFT); 00323 00324 alt0 = vec_sel(vec_madd(alt0, q0, negBias), vec_madd(alt0, q0, bias), 00325 vec_cmpgt(alt0, zero)); 00326 alt1 = vec_sel(vec_madd(alt1, q1, negBias), vec_madd(alt1, q1, bias), 00327 vec_cmpgt(alt1, zero)); 00328 alt2 = vec_sel(vec_madd(alt2, q2, negBias), vec_madd(alt2, q2, bias), 00329 vec_cmpgt(alt2, zero)); 00330 alt3 = vec_sel(vec_madd(alt3, q3, negBias), vec_madd(alt3, q3, bias), 00331 vec_cmpgt(alt3, zero)); 00332 alt4 = vec_sel(vec_madd(alt4, q4, negBias), vec_madd(alt4, q4, bias), 00333 vec_cmpgt(alt4, zero)); 00334 alt5 = vec_sel(vec_madd(alt5, q5, negBias), vec_madd(alt5, q5, bias), 00335 vec_cmpgt(alt5, zero)); 00336 alt6 = vec_sel(vec_madd(alt6, q6, negBias), vec_madd(alt6, q6, bias), 00337 vec_cmpgt(alt6, zero)); 00338 alt7 = vec_sel(vec_madd(alt7, q7, negBias), vec_madd(alt7, q7, bias), 00339 vec_cmpgt(alt7, zero)); 00340 } 00341 00342 00343 } 00344 00345 // Store the data back into the original block 00346 { 00347 vector signed short data0, data1, data2, data3, data4, data5, data6, data7; 00348 00349 data0 = vec_pack(vec_cts(row0, 0), vec_cts(alt0, 0)); 00350 data1 = vec_pack(vec_cts(row1, 0), vec_cts(alt1, 0)); 00351 data2 = vec_pack(vec_cts(row2, 0), vec_cts(alt2, 0)); 00352 data3 = vec_pack(vec_cts(row3, 0), vec_cts(alt3, 0)); 00353 data4 = vec_pack(vec_cts(row4, 0), vec_cts(alt4, 0)); 00354 data5 = vec_pack(vec_cts(row5, 0), vec_cts(alt5, 0)); 00355 data6 = vec_pack(vec_cts(row6, 0), vec_cts(alt6, 0)); 00356 data7 = vec_pack(vec_cts(row7, 0), vec_cts(alt7, 0)); 00357 00358 { 00359 // Clamp for overflow 00360 vector signed int max_q_int, min_q_int; 00361 vector signed short max_q, min_q; 00362 00363 LOAD4(max_q_int, &(s->max_qcoeff)); 00364 LOAD4(min_q_int, &(s->min_qcoeff)); 00365 00366 max_q = vec_pack(max_q_int, max_q_int); 00367 min_q = vec_pack(min_q_int, min_q_int); 00368 00369 data0 = vec_max(vec_min(data0, max_q), min_q); 00370 data1 = vec_max(vec_min(data1, max_q), min_q); 00371 data2 = vec_max(vec_min(data2, max_q), min_q); 00372 data4 = vec_max(vec_min(data4, max_q), min_q); 00373 data5 = vec_max(vec_min(data5, max_q), min_q); 00374 data6 = vec_max(vec_min(data6, max_q), min_q); 00375 data7 = vec_max(vec_min(data7, max_q), min_q); 00376 } 00377 00378 { 00379 vector bool char zero_01, zero_23, zero_45, zero_67; 00380 vector signed char scanIndexes_01, scanIndexes_23, scanIndexes_45, scanIndexes_67; 00381 vector signed char negOne = vec_splat_s8(-1); 00382 vector signed char* scanPtr = 00383 (vector signed char*)(s->intra_scantable.inverse); 00384 signed char lastNonZeroChar; 00385 00386 // Determine the largest non-zero index. 00387 zero_01 = vec_pack(vec_cmpeq(data0, (vector signed short)zero), 00388 vec_cmpeq(data1, (vector signed short)zero)); 00389 zero_23 = vec_pack(vec_cmpeq(data2, (vector signed short)zero), 00390 vec_cmpeq(data3, (vector signed short)zero)); 00391 zero_45 = vec_pack(vec_cmpeq(data4, (vector signed short)zero), 00392 vec_cmpeq(data5, (vector signed short)zero)); 00393 zero_67 = vec_pack(vec_cmpeq(data6, (vector signed short)zero), 00394 vec_cmpeq(data7, (vector signed short)zero)); 00395 00396 // 64 biggest values 00397 scanIndexes_01 = vec_sel(scanPtr[0], negOne, zero_01); 00398 scanIndexes_23 = vec_sel(scanPtr[1], negOne, zero_23); 00399 scanIndexes_45 = vec_sel(scanPtr[2], negOne, zero_45); 00400 scanIndexes_67 = vec_sel(scanPtr[3], negOne, zero_67); 00401 00402 // 32 largest values 00403 scanIndexes_01 = vec_max(scanIndexes_01, scanIndexes_23); 00404 scanIndexes_45 = vec_max(scanIndexes_45, scanIndexes_67); 00405 00406 // 16 largest values 00407 scanIndexes_01 = vec_max(scanIndexes_01, scanIndexes_45); 00408 00409 // 8 largest values 00410 scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne), 00411 vec_mergel(scanIndexes_01, negOne)); 00412 00413 // 4 largest values 00414 scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne), 00415 vec_mergel(scanIndexes_01, negOne)); 00416 00417 // 2 largest values 00418 scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne), 00419 vec_mergel(scanIndexes_01, negOne)); 00420 00421 // largest value 00422 scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne), 00423 vec_mergel(scanIndexes_01, negOne)); 00424 00425 scanIndexes_01 = vec_splat(scanIndexes_01, 0); 00426 00427 00428 vec_ste(scanIndexes_01, 0, &lastNonZeroChar); 00429 00430 lastNonZero = lastNonZeroChar; 00431 00432 // While the data is still in vectors we check for the transpose IDCT permute 00433 // and handle it using the vector unit if we can. This is the permute used 00434 // by the altivec idct, so it is common when using the altivec dct. 00435 00436 if ((lastNonZero > 0) && (s->dsp.idct_permutation_type == FF_TRANSPOSE_IDCT_PERM)) { 00437 TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7); 00438 } 00439 00440 vec_st(data0, 0, data); 00441 vec_st(data1, 16, data); 00442 vec_st(data2, 32, data); 00443 vec_st(data3, 48, data); 00444 vec_st(data4, 64, data); 00445 vec_st(data5, 80, data); 00446 vec_st(data6, 96, data); 00447 vec_st(data7, 112, data); 00448 } 00449 } 00450 00451 // special handling of block[0] 00452 if (s->mb_intra) { 00453 if (!s->h263_aic) { 00454 if (n < 4) 00455 oldBaseValue /= s->y_dc_scale; 00456 else 00457 oldBaseValue /= s->c_dc_scale; 00458 } 00459 00460 // Divide by 8, rounding the result 00461 data[0] = (oldBaseValue + 4) >> 3; 00462 } 00463 00464 // We handled the transpose permutation above and we don't 00465 // need to permute the "no" permutation case. 00466 if ((lastNonZero > 0) && 00467 (s->dsp.idct_permutation_type != FF_TRANSPOSE_IDCT_PERM) && 00468 (s->dsp.idct_permutation_type != FF_NO_IDCT_PERM)) { 00469 ff_block_permute(data, s->dsp.idct_permutation, 00470 s->intra_scantable.scantable, lastNonZero); 00471 } 00472 00473 return lastNonZero; 00474 } 00475 00476 /* AltiVec version of dct_unquantize_h263 00477 this code assumes `block' is 16 bytes-aligned */ 00478 void dct_unquantize_h263_altivec(MpegEncContext *s, 00479 DCTELEM *block, int n, int qscale) 00480 { 00481 POWERPC_PERF_DECLARE(altivec_dct_unquantize_h263_num, 1); 00482 int i, level, qmul, qadd; 00483 int nCoeffs; 00484 00485 assert(s->block_last_index[n]>=0); 00486 00487 POWERPC_PERF_START_COUNT(altivec_dct_unquantize_h263_num, 1); 00488 00489 qadd = (qscale - 1) | 1; 00490 qmul = qscale << 1; 00491 00492 if (s->mb_intra) { 00493 if (!s->h263_aic) { 00494 if (n < 4) 00495 block[0] = block[0] * s->y_dc_scale; 00496 else 00497 block[0] = block[0] * s->c_dc_scale; 00498 }else 00499 qadd = 0; 00500 i = 1; 00501 nCoeffs= 63; //does not always use zigzag table 00502 } else { 00503 i = 0; 00504 nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ]; 00505 } 00506 00507 { 00508 register const vector signed short vczero = (const vector signed short)vec_splat_s16(0); 00509 DECLARE_ALIGNED_16(short, qmul8[]) = 00510 { 00511 qmul, qmul, qmul, qmul, 00512 qmul, qmul, qmul, qmul 00513 }; 00514 DECLARE_ALIGNED_16(short, qadd8[]) = 00515 { 00516 qadd, qadd, qadd, qadd, 00517 qadd, qadd, qadd, qadd 00518 }; 00519 DECLARE_ALIGNED_16(short, nqadd8[]) = 00520 { 00521 -qadd, -qadd, -qadd, -qadd, 00522 -qadd, -qadd, -qadd, -qadd 00523 }; 00524 register vector signed short blockv, qmulv, qaddv, nqaddv, temp1; 00525 register vector bool short blockv_null, blockv_neg; 00526 register short backup_0 = block[0]; 00527 register int j = 0; 00528 00529 qmulv = vec_ld(0, qmul8); 00530 qaddv = vec_ld(0, qadd8); 00531 nqaddv = vec_ld(0, nqadd8); 00532 00533 #if 0 // block *is* 16 bytes-aligned, it seems. 00534 // first make sure block[j] is 16 bytes-aligned 00535 for(j = 0; (j <= nCoeffs) && ((((unsigned long)block) + (j << 1)) & 0x0000000F) ; j++) { 00536 level = block[j]; 00537 if (level) { 00538 if (level < 0) { 00539 level = level * qmul - qadd; 00540 } else { 00541 level = level * qmul + qadd; 00542 } 00543 block[j] = level; 00544 } 00545 } 00546 #endif 00547 00548 // vectorize all the 16 bytes-aligned blocks 00549 // of 8 elements 00550 for(; (j + 7) <= nCoeffs ; j+=8) { 00551 blockv = vec_ld(j << 1, block); 00552 blockv_neg = vec_cmplt(blockv, vczero); 00553 blockv_null = vec_cmpeq(blockv, vczero); 00554 // choose between +qadd or -qadd as the third operand 00555 temp1 = vec_sel(qaddv, nqaddv, blockv_neg); 00556 // multiply & add (block{i,i+7} * qmul [+-] qadd) 00557 temp1 = vec_mladd(blockv, qmulv, temp1); 00558 // put 0 where block[{i,i+7} used to have 0 00559 blockv = vec_sel(temp1, blockv, blockv_null); 00560 vec_st(blockv, j << 1, block); 00561 } 00562 00563 // if nCoeffs isn't a multiple of 8, finish the job 00564 // using good old scalar units. 00565 // (we could do it using a truncated vector, 00566 // but I'm not sure it's worth the hassle) 00567 for(; j <= nCoeffs ; j++) { 00568 level = block[j]; 00569 if (level) { 00570 if (level < 0) { 00571 level = level * qmul - qadd; 00572 } else { 00573 level = level * qmul + qadd; 00574 } 00575 block[j] = level; 00576 } 00577 } 00578 00579 if (i == 1) { 00580 // cheat. this avoid special-casing the first iteration 00581 block[0] = backup_0; 00582 } 00583 } 00584 POWERPC_PERF_STOP_COUNT(altivec_dct_unquantize_h263_num, nCoeffs == 63); 00585 } 00586 00587 00588 void idct_put_altivec(uint8_t *dest, int line_size, int16_t *block); 00589 void idct_add_altivec(uint8_t *dest, int line_size, int16_t *block); 00590 00591 void MPV_common_init_altivec(MpegEncContext *s) 00592 { 00593 if ((mm_flags & FF_MM_ALTIVEC) == 0) return; 00594 00595 if (s->avctx->lowres==0) { 00596 if ((s->avctx->idct_algo == FF_IDCT_AUTO) || 00597 (s->avctx->idct_algo == FF_IDCT_ALTIVEC)) { 00598 s->dsp.idct_put = idct_put_altivec; 00599 s->dsp.idct_add = idct_add_altivec; 00600 s->dsp.idct_permutation_type = FF_TRANSPOSE_IDCT_PERM; 00601 } 00602 } 00603 00604 // Test to make sure that the dct required alignments are met. 00605 if ((((long)(s->q_intra_matrix) & 0x0f) != 0) || 00606 (((long)(s->q_inter_matrix) & 0x0f) != 0)) { 00607 av_log(s->avctx, AV_LOG_INFO, "Internal Error: q-matrix blocks must be 16-byte aligned " 00608 "to use AltiVec DCT. Reverting to non-AltiVec version.\n"); 00609 return; 00610 } 00611 00612 if (((long)(s->intra_scantable.inverse) & 0x0f) != 0) { 00613 av_log(s->avctx, AV_LOG_INFO, "Internal Error: scan table blocks must be 16-byte aligned " 00614 "to use AltiVec DCT. Reverting to non-AltiVec version.\n"); 00615 return; 00616 } 00617 00618 00619 if ((s->avctx->dct_algo == FF_DCT_AUTO) || 00620 (s->avctx->dct_algo == FF_DCT_ALTIVEC)) { 00621 #if 0 /* seems to cause trouble under some circumstances */ 00622 s->dct_quantize = dct_quantize_altivec; 00623 #endif 00624 s->dct_unquantize_h263_intra = dct_unquantize_h263_altivec; 00625 s->dct_unquantize_h263_inter = dct_unquantize_h263_altivec; 00626 } 00627 }