Libav 0.7.1
|
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 Libav. 00008 * 00009 * Libav 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 * Libav 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 Libav; 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 "libavutil/cpu.h" 00027 #include "libavcodec/dsputil.h" 00028 #include "libavcodec/mpegvideo.h" 00029 00030 #include "util_altivec.h" 00031 #include "types_altivec.h" 00032 #include "dsputil_altivec.h" 00033 00034 // Swaps two variables (used for altivec registers) 00035 #define SWAP(a,b) \ 00036 do { \ 00037 __typeof__(a) swap_temp=a; \ 00038 a=b; \ 00039 b=swap_temp; \ 00040 } while (0) 00041 00042 // transposes a matrix consisting of four vectors with four elements each 00043 #define TRANSPOSE4(a,b,c,d) \ 00044 do { \ 00045 __typeof__(a) _trans_ach = vec_mergeh(a, c); \ 00046 __typeof__(a) _trans_acl = vec_mergel(a, c); \ 00047 __typeof__(a) _trans_bdh = vec_mergeh(b, d); \ 00048 __typeof__(a) _trans_bdl = vec_mergel(b, d); \ 00049 \ 00050 a = vec_mergeh(_trans_ach, _trans_bdh); \ 00051 b = vec_mergel(_trans_ach, _trans_bdh); \ 00052 c = vec_mergeh(_trans_acl, _trans_bdl); \ 00053 d = vec_mergel(_trans_acl, _trans_bdl); \ 00054 } while (0) 00055 00056 00057 // Loads a four-byte value (int or float) from the target address 00058 // into every element in the target vector. Only works if the 00059 // target address is four-byte aligned (which should be always). 00060 #define LOAD4(vec, address) \ 00061 { \ 00062 __typeof__(vec)* _load_addr = (__typeof__(vec)*)(address); \ 00063 vector unsigned char _perm_vec = vec_lvsl(0,(address)); \ 00064 vec = vec_ld(0, _load_addr); \ 00065 vec = vec_perm(vec, vec, _perm_vec); \ 00066 vec = vec_splat(vec, 0); \ 00067 } 00068 00069 00070 #define FOUROF(a) {a,a,a,a} 00071 00072 static int dct_quantize_altivec(MpegEncContext* s, 00073 DCTELEM* data, int n, 00074 int qscale, int* overflow) 00075 { 00076 int lastNonZero; 00077 vector float row0, row1, row2, row3, row4, row5, row6, row7; 00078 vector float alt0, alt1, alt2, alt3, alt4, alt5, alt6, alt7; 00079 const vector float zero = (const vector float)FOUROF(0.); 00080 // used after quantize step 00081 int oldBaseValue = 0; 00082 00083 // Load the data into the row/alt vectors 00084 { 00085 vector signed short data0, data1, data2, data3, data4, data5, data6, data7; 00086 00087 data0 = vec_ld(0, data); 00088 data1 = vec_ld(16, data); 00089 data2 = vec_ld(32, data); 00090 data3 = vec_ld(48, data); 00091 data4 = vec_ld(64, data); 00092 data5 = vec_ld(80, data); 00093 data6 = vec_ld(96, data); 00094 data7 = vec_ld(112, data); 00095 00096 // Transpose the data before we start 00097 TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7); 00098 00099 // load the data into floating point vectors. We load 00100 // the high half of each row into the main row vectors 00101 // and the low half into the alt vectors. 00102 row0 = vec_ctf(vec_unpackh(data0), 0); 00103 alt0 = vec_ctf(vec_unpackl(data0), 0); 00104 row1 = vec_ctf(vec_unpackh(data1), 0); 00105 alt1 = vec_ctf(vec_unpackl(data1), 0); 00106 row2 = vec_ctf(vec_unpackh(data2), 0); 00107 alt2 = vec_ctf(vec_unpackl(data2), 0); 00108 row3 = vec_ctf(vec_unpackh(data3), 0); 00109 alt3 = vec_ctf(vec_unpackl(data3), 0); 00110 row4 = vec_ctf(vec_unpackh(data4), 0); 00111 alt4 = vec_ctf(vec_unpackl(data4), 0); 00112 row5 = vec_ctf(vec_unpackh(data5), 0); 00113 alt5 = vec_ctf(vec_unpackl(data5), 0); 00114 row6 = vec_ctf(vec_unpackh(data6), 0); 00115 alt6 = vec_ctf(vec_unpackl(data6), 0); 00116 row7 = vec_ctf(vec_unpackh(data7), 0); 00117 alt7 = vec_ctf(vec_unpackl(data7), 0); 00118 } 00119 00120 // The following block could exist as a separate an altivec dct 00121 // function. However, if we put it inline, the DCT data can remain 00122 // in the vector local variables, as floats, which we'll use during the 00123 // quantize step... 00124 { 00125 const vector float vec_0_298631336 = (vector float)FOUROF(0.298631336f); 00126 const vector float vec_0_390180644 = (vector float)FOUROF(-0.390180644f); 00127 const vector float vec_0_541196100 = (vector float)FOUROF(0.541196100f); 00128 const vector float vec_0_765366865 = (vector float)FOUROF(0.765366865f); 00129 const vector float vec_0_899976223 = (vector float)FOUROF(-0.899976223f); 00130 const vector float vec_1_175875602 = (vector float)FOUROF(1.175875602f); 00131 const vector float vec_1_501321110 = (vector float)FOUROF(1.501321110f); 00132 const vector float vec_1_847759065 = (vector float)FOUROF(-1.847759065f); 00133 const vector float vec_1_961570560 = (vector float)FOUROF(-1.961570560f); 00134 const vector float vec_2_053119869 = (vector float)FOUROF(2.053119869f); 00135 const vector float vec_2_562915447 = (vector float)FOUROF(-2.562915447f); 00136 const vector float vec_3_072711026 = (vector float)FOUROF(3.072711026f); 00137 00138 00139 int whichPass, whichHalf; 00140 00141 for(whichPass = 1; whichPass<=2; whichPass++) { 00142 for(whichHalf = 1; whichHalf<=2; whichHalf++) { 00143 vector float tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; 00144 vector float tmp10, tmp11, tmp12, tmp13; 00145 vector float z1, z2, z3, z4, z5; 00146 00147 tmp0 = vec_add(row0, row7); // tmp0 = dataptr[0] + dataptr[7]; 00148 tmp7 = vec_sub(row0, row7); // tmp7 = dataptr[0] - dataptr[7]; 00149 tmp3 = vec_add(row3, row4); // tmp3 = dataptr[3] + dataptr[4]; 00150 tmp4 = vec_sub(row3, row4); // tmp4 = dataptr[3] - dataptr[4]; 00151 tmp1 = vec_add(row1, row6); // tmp1 = dataptr[1] + dataptr[6]; 00152 tmp6 = vec_sub(row1, row6); // tmp6 = dataptr[1] - dataptr[6]; 00153 tmp2 = vec_add(row2, row5); // tmp2 = dataptr[2] + dataptr[5]; 00154 tmp5 = vec_sub(row2, row5); // tmp5 = dataptr[2] - dataptr[5]; 00155 00156 tmp10 = vec_add(tmp0, tmp3); // tmp10 = tmp0 + tmp3; 00157 tmp13 = vec_sub(tmp0, tmp3); // tmp13 = tmp0 - tmp3; 00158 tmp11 = vec_add(tmp1, tmp2); // tmp11 = tmp1 + tmp2; 00159 tmp12 = vec_sub(tmp1, tmp2); // tmp12 = tmp1 - tmp2; 00160 00161 00162 // dataptr[0] = (DCTELEM) ((tmp10 + tmp11) << PASS1_BITS); 00163 row0 = vec_add(tmp10, tmp11); 00164 00165 // dataptr[4] = (DCTELEM) ((tmp10 - tmp11) << PASS1_BITS); 00166 row4 = vec_sub(tmp10, tmp11); 00167 00168 00169 // z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100); 00170 z1 = vec_madd(vec_add(tmp12, tmp13), vec_0_541196100, (vector float)zero); 00171 00172 // dataptr[2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865), 00173 // CONST_BITS-PASS1_BITS); 00174 row2 = vec_madd(tmp13, vec_0_765366865, z1); 00175 00176 // dataptr[6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065), 00177 // CONST_BITS-PASS1_BITS); 00178 row6 = vec_madd(tmp12, vec_1_847759065, z1); 00179 00180 z1 = vec_add(tmp4, tmp7); // z1 = tmp4 + tmp7; 00181 z2 = vec_add(tmp5, tmp6); // z2 = tmp5 + tmp6; 00182 z3 = vec_add(tmp4, tmp6); // z3 = tmp4 + tmp6; 00183 z4 = vec_add(tmp5, tmp7); // z4 = tmp5 + tmp7; 00184 00185 // z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */ 00186 z5 = vec_madd(vec_add(z3, z4), vec_1_175875602, (vector float)zero); 00187 00188 // z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */ 00189 z3 = vec_madd(z3, vec_1_961570560, z5); 00190 00191 // z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */ 00192 z4 = vec_madd(z4, vec_0_390180644, z5); 00193 00194 // The following adds are rolled into the multiplies above 00195 // z3 = vec_add(z3, z5); // z3 += z5; 00196 // z4 = vec_add(z4, z5); // z4 += z5; 00197 00198 // z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */ 00199 // Wow! It's actually more efficient to roll this multiply 00200 // into the adds below, even thought the multiply gets done twice! 00201 // z2 = vec_madd(z2, vec_2_562915447, (vector float)zero); 00202 00203 // z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */ 00204 // Same with this one... 00205 // z1 = vec_madd(z1, vec_0_899976223, (vector float)zero); 00206 00207 // tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */ 00208 // dataptr[7] = (DCTELEM) DESCALE(tmp4 + z1 + z3, CONST_BITS-PASS1_BITS); 00209 row7 = vec_madd(tmp4, vec_0_298631336, vec_madd(z1, vec_0_899976223, z3)); 00210 00211 // tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */ 00212 // dataptr[5] = (DCTELEM) DESCALE(tmp5 + z2 + z4, CONST_BITS-PASS1_BITS); 00213 row5 = vec_madd(tmp5, vec_2_053119869, vec_madd(z2, vec_2_562915447, z4)); 00214 00215 // tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */ 00216 // dataptr[3] = (DCTELEM) DESCALE(tmp6 + z2 + z3, CONST_BITS-PASS1_BITS); 00217 row3 = vec_madd(tmp6, vec_3_072711026, vec_madd(z2, vec_2_562915447, z3)); 00218 00219 // tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */ 00220 // dataptr[1] = (DCTELEM) DESCALE(tmp7 + z1 + z4, CONST_BITS-PASS1_BITS); 00221 row1 = vec_madd(z1, vec_0_899976223, vec_madd(tmp7, vec_1_501321110, z4)); 00222 00223 // Swap the row values with the alts. If this is the first half, 00224 // this sets up the low values to be acted on in the second half. 00225 // If this is the second half, it puts the high values back in 00226 // the row values where they are expected to be when we're done. 00227 SWAP(row0, alt0); 00228 SWAP(row1, alt1); 00229 SWAP(row2, alt2); 00230 SWAP(row3, alt3); 00231 SWAP(row4, alt4); 00232 SWAP(row5, alt5); 00233 SWAP(row6, alt6); 00234 SWAP(row7, alt7); 00235 } 00236 00237 if (whichPass == 1) { 00238 // transpose the data for the second pass 00239 00240 // First, block transpose the upper right with lower left. 00241 SWAP(row4, alt0); 00242 SWAP(row5, alt1); 00243 SWAP(row6, alt2); 00244 SWAP(row7, alt3); 00245 00246 // Now, transpose each block of four 00247 TRANSPOSE4(row0, row1, row2, row3); 00248 TRANSPOSE4(row4, row5, row6, row7); 00249 TRANSPOSE4(alt0, alt1, alt2, alt3); 00250 TRANSPOSE4(alt4, alt5, alt6, alt7); 00251 } 00252 } 00253 } 00254 00255 // perform the quantize step, using the floating point data 00256 // still in the row/alt registers 00257 { 00258 const int* biasAddr; 00259 const vector signed int* qmat; 00260 vector float bias, negBias; 00261 00262 if (s->mb_intra) { 00263 vector signed int baseVector; 00264 00265 // We must cache element 0 in the intra case 00266 // (it needs special handling). 00267 baseVector = vec_cts(vec_splat(row0, 0), 0); 00268 vec_ste(baseVector, 0, &oldBaseValue); 00269 00270 qmat = (vector signed int*)s->q_intra_matrix[qscale]; 00271 biasAddr = &(s->intra_quant_bias); 00272 } else { 00273 qmat = (vector signed int*)s->q_inter_matrix[qscale]; 00274 biasAddr = &(s->inter_quant_bias); 00275 } 00276 00277 // Load the bias vector (We add 0.5 to the bias so that we're 00278 // rounding when we convert to int, instead of flooring.) 00279 { 00280 vector signed int biasInt; 00281 const vector float negOneFloat = (vector float)FOUROF(-1.0f); 00282 LOAD4(biasInt, biasAddr); 00283 bias = vec_ctf(biasInt, QUANT_BIAS_SHIFT); 00284 negBias = vec_madd(bias, negOneFloat, zero); 00285 } 00286 00287 { 00288 vector float q0, q1, q2, q3, q4, q5, q6, q7; 00289 00290 q0 = vec_ctf(qmat[0], QMAT_SHIFT); 00291 q1 = vec_ctf(qmat[2], QMAT_SHIFT); 00292 q2 = vec_ctf(qmat[4], QMAT_SHIFT); 00293 q3 = vec_ctf(qmat[6], QMAT_SHIFT); 00294 q4 = vec_ctf(qmat[8], QMAT_SHIFT); 00295 q5 = vec_ctf(qmat[10], QMAT_SHIFT); 00296 q6 = vec_ctf(qmat[12], QMAT_SHIFT); 00297 q7 = vec_ctf(qmat[14], QMAT_SHIFT); 00298 00299 row0 = vec_sel(vec_madd(row0, q0, negBias), vec_madd(row0, q0, bias), 00300 vec_cmpgt(row0, zero)); 00301 row1 = vec_sel(vec_madd(row1, q1, negBias), vec_madd(row1, q1, bias), 00302 vec_cmpgt(row1, zero)); 00303 row2 = vec_sel(vec_madd(row2, q2, negBias), vec_madd(row2, q2, bias), 00304 vec_cmpgt(row2, zero)); 00305 row3 = vec_sel(vec_madd(row3, q3, negBias), vec_madd(row3, q3, bias), 00306 vec_cmpgt(row3, zero)); 00307 row4 = vec_sel(vec_madd(row4, q4, negBias), vec_madd(row4, q4, bias), 00308 vec_cmpgt(row4, zero)); 00309 row5 = vec_sel(vec_madd(row5, q5, negBias), vec_madd(row5, q5, bias), 00310 vec_cmpgt(row5, zero)); 00311 row6 = vec_sel(vec_madd(row6, q6, negBias), vec_madd(row6, q6, bias), 00312 vec_cmpgt(row6, zero)); 00313 row7 = vec_sel(vec_madd(row7, q7, negBias), vec_madd(row7, q7, bias), 00314 vec_cmpgt(row7, zero)); 00315 00316 q0 = vec_ctf(qmat[1], QMAT_SHIFT); 00317 q1 = vec_ctf(qmat[3], QMAT_SHIFT); 00318 q2 = vec_ctf(qmat[5], QMAT_SHIFT); 00319 q3 = vec_ctf(qmat[7], QMAT_SHIFT); 00320 q4 = vec_ctf(qmat[9], QMAT_SHIFT); 00321 q5 = vec_ctf(qmat[11], QMAT_SHIFT); 00322 q6 = vec_ctf(qmat[13], QMAT_SHIFT); 00323 q7 = vec_ctf(qmat[15], QMAT_SHIFT); 00324 00325 alt0 = vec_sel(vec_madd(alt0, q0, negBias), vec_madd(alt0, q0, bias), 00326 vec_cmpgt(alt0, zero)); 00327 alt1 = vec_sel(vec_madd(alt1, q1, negBias), vec_madd(alt1, q1, bias), 00328 vec_cmpgt(alt1, zero)); 00329 alt2 = vec_sel(vec_madd(alt2, q2, negBias), vec_madd(alt2, q2, bias), 00330 vec_cmpgt(alt2, zero)); 00331 alt3 = vec_sel(vec_madd(alt3, q3, negBias), vec_madd(alt3, q3, bias), 00332 vec_cmpgt(alt3, zero)); 00333 alt4 = vec_sel(vec_madd(alt4, q4, negBias), vec_madd(alt4, q4, bias), 00334 vec_cmpgt(alt4, zero)); 00335 alt5 = vec_sel(vec_madd(alt5, q5, negBias), vec_madd(alt5, q5, bias), 00336 vec_cmpgt(alt5, zero)); 00337 alt6 = vec_sel(vec_madd(alt6, q6, negBias), vec_madd(alt6, q6, bias), 00338 vec_cmpgt(alt6, zero)); 00339 alt7 = vec_sel(vec_madd(alt7, q7, negBias), vec_madd(alt7, q7, bias), 00340 vec_cmpgt(alt7, zero)); 00341 } 00342 00343 00344 } 00345 00346 // Store the data back into the original block 00347 { 00348 vector signed short data0, data1, data2, data3, data4, data5, data6, data7; 00349 00350 data0 = vec_pack(vec_cts(row0, 0), vec_cts(alt0, 0)); 00351 data1 = vec_pack(vec_cts(row1, 0), vec_cts(alt1, 0)); 00352 data2 = vec_pack(vec_cts(row2, 0), vec_cts(alt2, 0)); 00353 data3 = vec_pack(vec_cts(row3, 0), vec_cts(alt3, 0)); 00354 data4 = vec_pack(vec_cts(row4, 0), vec_cts(alt4, 0)); 00355 data5 = vec_pack(vec_cts(row5, 0), vec_cts(alt5, 0)); 00356 data6 = vec_pack(vec_cts(row6, 0), vec_cts(alt6, 0)); 00357 data7 = vec_pack(vec_cts(row7, 0), vec_cts(alt7, 0)); 00358 00359 { 00360 // Clamp for overflow 00361 vector signed int max_q_int, min_q_int; 00362 vector signed short max_q, min_q; 00363 00364 LOAD4(max_q_int, &(s->max_qcoeff)); 00365 LOAD4(min_q_int, &(s->min_qcoeff)); 00366 00367 max_q = vec_pack(max_q_int, max_q_int); 00368 min_q = vec_pack(min_q_int, min_q_int); 00369 00370 data0 = vec_max(vec_min(data0, max_q), min_q); 00371 data1 = vec_max(vec_min(data1, max_q), min_q); 00372 data2 = vec_max(vec_min(data2, max_q), min_q); 00373 data4 = vec_max(vec_min(data4, max_q), min_q); 00374 data5 = vec_max(vec_min(data5, max_q), min_q); 00375 data6 = vec_max(vec_min(data6, max_q), min_q); 00376 data7 = vec_max(vec_min(data7, max_q), min_q); 00377 } 00378 00379 { 00380 vector bool char zero_01, zero_23, zero_45, zero_67; 00381 vector signed char scanIndexes_01, scanIndexes_23, scanIndexes_45, scanIndexes_67; 00382 vector signed char negOne = vec_splat_s8(-1); 00383 vector signed char* scanPtr = 00384 (vector signed char*)(s->intra_scantable.inverse); 00385 signed char lastNonZeroChar; 00386 00387 // Determine the largest non-zero index. 00388 zero_01 = vec_pack(vec_cmpeq(data0, (vector signed short)zero), 00389 vec_cmpeq(data1, (vector signed short)zero)); 00390 zero_23 = vec_pack(vec_cmpeq(data2, (vector signed short)zero), 00391 vec_cmpeq(data3, (vector signed short)zero)); 00392 zero_45 = vec_pack(vec_cmpeq(data4, (vector signed short)zero), 00393 vec_cmpeq(data5, (vector signed short)zero)); 00394 zero_67 = vec_pack(vec_cmpeq(data6, (vector signed short)zero), 00395 vec_cmpeq(data7, (vector signed short)zero)); 00396 00397 // 64 biggest values 00398 scanIndexes_01 = vec_sel(scanPtr[0], negOne, zero_01); 00399 scanIndexes_23 = vec_sel(scanPtr[1], negOne, zero_23); 00400 scanIndexes_45 = vec_sel(scanPtr[2], negOne, zero_45); 00401 scanIndexes_67 = vec_sel(scanPtr[3], negOne, zero_67); 00402 00403 // 32 largest values 00404 scanIndexes_01 = vec_max(scanIndexes_01, scanIndexes_23); 00405 scanIndexes_45 = vec_max(scanIndexes_45, scanIndexes_67); 00406 00407 // 16 largest values 00408 scanIndexes_01 = vec_max(scanIndexes_01, scanIndexes_45); 00409 00410 // 8 largest values 00411 scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne), 00412 vec_mergel(scanIndexes_01, negOne)); 00413 00414 // 4 largest values 00415 scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne), 00416 vec_mergel(scanIndexes_01, negOne)); 00417 00418 // 2 largest values 00419 scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne), 00420 vec_mergel(scanIndexes_01, negOne)); 00421 00422 // largest value 00423 scanIndexes_01 = vec_max(vec_mergeh(scanIndexes_01, negOne), 00424 vec_mergel(scanIndexes_01, negOne)); 00425 00426 scanIndexes_01 = vec_splat(scanIndexes_01, 0); 00427 00428 00429 vec_ste(scanIndexes_01, 0, &lastNonZeroChar); 00430 00431 lastNonZero = lastNonZeroChar; 00432 00433 // While the data is still in vectors we check for the transpose IDCT permute 00434 // and handle it using the vector unit if we can. This is the permute used 00435 // by the altivec idct, so it is common when using the altivec dct. 00436 00437 if ((lastNonZero > 0) && (s->dsp.idct_permutation_type == FF_TRANSPOSE_IDCT_PERM)) { 00438 TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7); 00439 } 00440 00441 vec_st(data0, 0, data); 00442 vec_st(data1, 16, data); 00443 vec_st(data2, 32, data); 00444 vec_st(data3, 48, data); 00445 vec_st(data4, 64, data); 00446 vec_st(data5, 80, data); 00447 vec_st(data6, 96, data); 00448 vec_st(data7, 112, data); 00449 } 00450 } 00451 00452 // special handling of block[0] 00453 if (s->mb_intra) { 00454 if (!s->h263_aic) { 00455 if (n < 4) 00456 oldBaseValue /= s->y_dc_scale; 00457 else 00458 oldBaseValue /= s->c_dc_scale; 00459 } 00460 00461 // Divide by 8, rounding the result 00462 data[0] = (oldBaseValue + 4) >> 3; 00463 } 00464 00465 // We handled the transpose permutation above and we don't 00466 // need to permute the "no" permutation case. 00467 if ((lastNonZero > 0) && 00468 (s->dsp.idct_permutation_type != FF_TRANSPOSE_IDCT_PERM) && 00469 (s->dsp.idct_permutation_type != FF_NO_IDCT_PERM)) { 00470 ff_block_permute(data, s->dsp.idct_permutation, 00471 s->intra_scantable.scantable, lastNonZero); 00472 } 00473 00474 return lastNonZero; 00475 } 00476 00477 /* AltiVec version of dct_unquantize_h263 00478 this code assumes `block' is 16 bytes-aligned */ 00479 static void dct_unquantize_h263_altivec(MpegEncContext *s, 00480 DCTELEM *block, int n, int qscale) 00481 { 00482 int i, level, qmul, qadd; 00483 int nCoeffs; 00484 00485 assert(s->block_last_index[n]>=0); 00486 00487 qadd = (qscale - 1) | 1; 00488 qmul = qscale << 1; 00489 00490 if (s->mb_intra) { 00491 if (!s->h263_aic) { 00492 if (n < 4) 00493 block[0] = block[0] * s->y_dc_scale; 00494 else 00495 block[0] = block[0] * s->c_dc_scale; 00496 }else 00497 qadd = 0; 00498 i = 1; 00499 nCoeffs= 63; //does not always use zigzag table 00500 } else { 00501 i = 0; 00502 nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ]; 00503 } 00504 00505 { 00506 register const vector signed short vczero = (const vector signed short)vec_splat_s16(0); 00507 DECLARE_ALIGNED(16, short, qmul8) = qmul; 00508 DECLARE_ALIGNED(16, short, qadd8) = qadd; 00509 register vector signed short blockv, qmulv, qaddv, nqaddv, temp1; 00510 register vector bool short blockv_null, blockv_neg; 00511 register short backup_0 = block[0]; 00512 register int j = 0; 00513 00514 qmulv = vec_splat((vec_s16)vec_lde(0, &qmul8), 0); 00515 qaddv = vec_splat((vec_s16)vec_lde(0, &qadd8), 0); 00516 nqaddv = vec_sub(vczero, qaddv); 00517 00518 #if 0 // block *is* 16 bytes-aligned, it seems. 00519 // first make sure block[j] is 16 bytes-aligned 00520 for(j = 0; (j <= nCoeffs) && ((((unsigned long)block) + (j << 1)) & 0x0000000F) ; j++) { 00521 level = block[j]; 00522 if (level) { 00523 if (level < 0) { 00524 level = level * qmul - qadd; 00525 } else { 00526 level = level * qmul + qadd; 00527 } 00528 block[j] = level; 00529 } 00530 } 00531 #endif 00532 00533 // vectorize all the 16 bytes-aligned blocks 00534 // of 8 elements 00535 for(; (j + 7) <= nCoeffs ; j+=8) { 00536 blockv = vec_ld(j << 1, block); 00537 blockv_neg = vec_cmplt(blockv, vczero); 00538 blockv_null = vec_cmpeq(blockv, vczero); 00539 // choose between +qadd or -qadd as the third operand 00540 temp1 = vec_sel(qaddv, nqaddv, blockv_neg); 00541 // multiply & add (block{i,i+7} * qmul [+-] qadd) 00542 temp1 = vec_mladd(blockv, qmulv, temp1); 00543 // put 0 where block[{i,i+7} used to have 0 00544 blockv = vec_sel(temp1, blockv, blockv_null); 00545 vec_st(blockv, j << 1, block); 00546 } 00547 00548 // if nCoeffs isn't a multiple of 8, finish the job 00549 // using good old scalar units. 00550 // (we could do it using a truncated vector, 00551 // but I'm not sure it's worth the hassle) 00552 for(; j <= nCoeffs ; j++) { 00553 level = block[j]; 00554 if (level) { 00555 if (level < 0) { 00556 level = level * qmul - qadd; 00557 } else { 00558 level = level * qmul + qadd; 00559 } 00560 block[j] = level; 00561 } 00562 } 00563 00564 if (i == 1) { 00565 // cheat. this avoid special-casing the first iteration 00566 block[0] = backup_0; 00567 } 00568 } 00569 } 00570 00571 00572 void MPV_common_init_altivec(MpegEncContext *s) 00573 { 00574 if (!(av_get_cpu_flags() & AV_CPU_FLAG_ALTIVEC)) return; 00575 00576 if (s->avctx->lowres==0) { 00577 if ((s->avctx->idct_algo == FF_IDCT_AUTO) || 00578 (s->avctx->idct_algo == FF_IDCT_ALTIVEC)) { 00579 s->dsp.idct_put = idct_put_altivec; 00580 s->dsp.idct_add = idct_add_altivec; 00581 s->dsp.idct_permutation_type = FF_TRANSPOSE_IDCT_PERM; 00582 } 00583 } 00584 00585 // Test to make sure that the dct required alignments are met. 00586 if ((((long)(s->q_intra_matrix) & 0x0f) != 0) || 00587 (((long)(s->q_inter_matrix) & 0x0f) != 0)) { 00588 av_log(s->avctx, AV_LOG_INFO, "Internal Error: q-matrix blocks must be 16-byte aligned " 00589 "to use AltiVec DCT. Reverting to non-AltiVec version.\n"); 00590 return; 00591 } 00592 00593 if (((long)(s->intra_scantable.inverse) & 0x0f) != 0) { 00594 av_log(s->avctx, AV_LOG_INFO, "Internal Error: scan table blocks must be 16-byte aligned " 00595 "to use AltiVec DCT. Reverting to non-AltiVec version.\n"); 00596 return; 00597 } 00598 00599 00600 if ((s->avctx->dct_algo == FF_DCT_AUTO) || 00601 (s->avctx->dct_algo == FF_DCT_ALTIVEC)) { 00602 #if 0 /* seems to cause trouble under some circumstances */ 00603 s->dct_quantize = dct_quantize_altivec; 00604 #endif 00605 s->dct_unquantize_h263_intra = dct_unquantize_h263_altivec; 00606 s->dct_unquantize_h263_inter = dct_unquantize_h263_altivec; 00607 } 00608 }