1 /* quantize.c, quantization / inverse quantization */
3 /* Copyright (C) 1996, MPEG Software Simulation Group. All Rights Reserved. */
6 * Disclaimer of Warranty
8 * These software programs are available to the user without any license fee or
9 * royalty on an "as is" basis. The MPEG Software Simulation Group disclaims
10 * any and all warranties, whether express, implied, or statuary, including any
11 * implied warranties or merchantability or of fitness for a particular
12 * purpose. In no event shall the copyright-holder be liable for any
13 * incidental, punitive, or consequential damages of any kind whatsoever
14 * arising from the use of these programs.
16 * This disclaimer of warranty extends to the user of these programs and user's
17 * customers, employees, agents, transferees, successors, and assigns.
19 * The MPEG Software Simulation Group does not represent or warrant that the
20 * programs furnished hereunder are free of infringement of any third-party
23 * Commercial implementations of MPEG-1 and MPEG-2 video, including shareware,
24 * are subject to royalty fees to patent holders. Many of these patents are
25 * general enough such that they are unavoidable regardless of implementation
36 #include "cpu_accel.h"
38 #include "fastintfns.h"
41 /* Global function pointers for SIMD-dependent functions */
42 int (*pquant_non_intra)(pict_data_s *picture, int16_t *src, int16_t *dst,
43 int mquant, int *nonsat_mquant);
44 int (*pquant_weight_coeff_sum)(int16_t *blk, uint16_t*i_quant_mat );
46 /* Local functions pointers for SIMD-dependent functions */
48 static void (*piquant_non_intra_m1)(int16_t *src, int16_t *dst, uint16_t *quant_mat);
51 static int quant_weight_coeff_sum( int16_t *blk, uint16_t * i_quant_mat );
52 static void iquant_non_intra_m1(int16_t *src, int16_t *dst, uint16_t *quant_mat);
56 Initialise quantization routines.
57 Currently just setting up MMX routines if available...
60 void init_quantizer_hv()
65 if( (flags & ACCEL_X86_MMX) != 0 ) /* MMX CPU */
67 if(verbose) fprintf( stderr, "SETTING " );
68 if( (flags & ACCEL_X86_3DNOW) != 0 )
70 if(verbose) fprintf( stderr, "3DNOW and ");
71 pquant_non_intra = quant_non_intra_hv_3dnow;
74 * else if ( (flags & ACCEL_X86_MMXEXT) != 0 )
76 * if(verbose) fprintf( stderr, "SSE and ");
77 * pquant_non_intra = quant_non_intra_hv_sse;
82 pquant_non_intra = quant_non_intra_hv;
85 if ( (flags & ACCEL_X86_MMXEXT) != 0 )
87 if(verbose) fprintf( stderr, "EXTENDED MMX");
88 pquant_weight_coeff_sum = quant_weight_coeff_sum_mmx;
89 piquant_non_intra_m1 = iquant_non_intra_m1_sse;
93 if(verbose) fprintf( stderr, "MMX");
94 pquant_weight_coeff_sum = quant_weight_coeff_sum_mmx;
95 piquant_non_intra_m1 = iquant_non_intra_m1_mmx;
97 if(verbose) fprintf( stderr, " for QUANTIZER!\n");
102 pquant_non_intra = quant_non_intra_hv;
103 pquant_weight_coeff_sum = quant_weight_coeff_sum;
104 piquant_non_intra_m1 = iquant_non_intra_m1;
110 * Computes the next quantisation up. Used to avoid saturation
111 * in macroblock coefficients - common in MPEG-1 - which causes
114 * NOTE: Does no range checking...
119 int next_larger_quant_hv( pict_data_s *picture, int quant )
121 if( picture->q_scale_type )
123 if( map_non_linear_mquant_hv[quant]+1 > 31 )
126 return non_linear_mquant_table_hv[map_non_linear_mquant_hv[quant]+1];
139 * Quantisation for intra blocks using Test Model 5 quantization
141 * this quantizer has a bias of 1/8 stepsize towards zero
142 * (except for the DC coefficient)
144 PRECONDITION: src dst point to *disinct* memory buffers...
145 of block_count *adjact* int16_t[64] arrays...
147 * RETURN: 1 If non-zero coefficients left after quantisaiont 0 otherwise
151 pict_data_s *picture,
162 int clipvalue = dctsatlim;
163 uint16_t *quant_mat = intra_q_tbl[mquant] /* intra_q */;
166 /* Inspired by suggestion by Juan. Quantize a little harder if we clip...
174 for( comp = 0; comp<block_count && !clipping; ++comp )
177 d = 8>>picture->dc_prec; /* intra_dc_mult */
178 pbuf[0] = (x>=0) ? (x+(d>>1))/d : -((-x+(d>>1))/d); /* round(x/d) */
181 for (i=1; i<64 ; i++)
185 /* RJ: save one divide operation */
186 y = ((abs(x) << 5)+ ((3 * quant_mat[i]) >> 2)) / (quant_mat[i] << 1);
190 mquant = next_larger_quant_hv( picture, mquant );
191 quant_mat = intra_q_tbl[mquant];
195 pbuf[i] = intsamesign(x,y);
202 *nonsat_mquant = mquant;
207 * Quantisation matrix weighted Coefficient sum fixed-point
208 * integer with low 16 bits fractional...
209 * To be used for rate control as a measure of dct block
214 int quant_weight_coeff_sum( int16_t *blk, uint16_t * i_quant_mat )
218 for( i = 0; i < 64; i+=2 )
220 sum += abs((int)blk[i]) * (i_quant_mat[i]) + abs((int)blk[i+1]) * (i_quant_mat[i+1]);
223 /* In case you're wondering typical average coeff_sum's for a rather
224 noisy video are around 20.0. */
230 * Quantisation for non-intra blocks using Test Model 5 quantization
232 * this quantizer has a bias of 1/8 stepsize towards zero
233 * (except for the DC coefficient)
235 * A.Stevens 2000: The above comment is nonsense. Only the intra quantiser does
236 * this. This one just truncates with a modest bias of 1/(4*quant_matrix_scale)
239 * PRECONDITION: src dst point to *disinct* memory buffers...
240 * of block_count *adjacent* int16_t[64] arrays...
242 * RETURN: A bit-mask of block_count bits indicating non-zero blocks (a 1).
244 * TODO: A candidate for use of efficient abs and "intsamesign". If only gcc understood
245 * PPro conditional moves...
248 int quant_non_intra_hv(
249 pict_data_s *picture,
250 int16_t *src, int16_t *dst,
258 int clipvalue = dctsatlim;
261 uint16_t *quant_mat = inter_q_tbl[mquant]/* inter_q */;
263 coeff_count = 64*block_count;
266 for (i=0; i<coeff_count; ++i)
271 nzflag = (nzflag<<1) | !!flags;
275 /* RJ: save one divide operation */
277 x = abs( ((int)src[i]) ) /*(src[i] >= 0 ? src[i] : -src[i])*/ ;
278 d = (int)quant_mat[(i&63)];
279 /* A.Stevens 2000: Given the math of non-intra frame
280 quantisation / inverse quantisation I always though the
281 funny little foudning factor was bogus. It seems to be
282 the encoder needs less bits if you simply divide!
285 y = (x<<4) / (d) /* (32*x + (d>>1))/(d*2*mquant)*/ ;
294 int new_mquant = next_larger_quant_hv( picture, mquant );
295 if( new_mquant != mquant )
298 quant_mat = inter_q_tbl[mquant];
309 dst[i] = intsamesign(src[i], y) /* (src[i] >= 0 ? y : -y) */;
312 nzflag = (nzflag<<1) | !!flags;
314 *nonsat_mquant = mquant;
318 /* MPEG-1 inverse quantization */
319 static void iquant1_intra(int16_t *src, int16_t *dst, int dc_prec, int mquant)
322 uint16_t *quant_mat = intra_q;
324 dst[0] = src[0] << (3-dc_prec);
327 val = (int)(src[i]*quant_mat[i]*mquant)/16;
329 /* mismatch control */
330 if ((val&1)==0 && val!=0)
331 val+= (val>0) ? -1 : 1;
334 dst[i] = (val>2047) ? 2047 : ((val<-2048) ? -2048 : val);
339 /* MPEG-2 inverse quantization */
340 void iquant_intra(int16_t *src, int16_t *dst, int dc_prec, int mquant)
345 iquant1_intra(src,dst,dc_prec, mquant);
348 sum = dst[0] = src[0] << (3-dc_prec);
351 val = (int)(src[i]*intra_q[i]*mquant)/16;
352 sum+= dst[i] = (val>2047) ? 2047 : ((val<-2048) ? -2048 : val);
355 /* mismatch control */
362 static void iquant_non_intra_m1(int16_t *src, int16_t *dst, uint16_t *quant_mat)
366 #ifndef ORIGINAL_CODE
373 val = (int)((2*val+(val>0 ? 1 : -1))*quant_mat[i])/32;
375 /* mismatch control */
376 if ((val&1)==0 && val!=0)
377 val+= (val>0) ? -1 : 1;
381 dst[i] = (val>2047) ? 2047 : ((val<-2048) ? -2048 : val);
390 val = ((val+val+1)*quant_mat[i]) >> 5;
391 /* mismatch control */
392 val -= (~(val&1))&(val!=0);
393 val = fastmin(val, 2047); /* Saturation */
395 dst[i] = intsamesign(src[i],val);
405 void iquant_non_intra(int16_t *src, int16_t *dst, int mquant )
411 (*piquant_non_intra_m1)(src,dst,inter_q_tbl[mquant]);
422 val = (int)((2*val+(val>0 ? 1 : -1))*inter_q[i]*mquant)/32;
423 sum+= dst[i] = (val>2047) ? 2047 : ((val<-2048) ? -2048 : val);
426 quant_mat = inter_q_tbl[mquant];
433 val = (int)((val+val+1)*quant_mat[i])>>5;
434 val = intmin( val, 2047);
437 dst[i] = intsamesign(src[i],val);
441 /* mismatch control */
447 void iquantize( pict_data_s *picture )
450 int16_t (*qblocks)[64] = picture->qblocks;
451 for (k=0; k<mb_per_pict; k++)
453 if (picture->mbinfo[k].mb_type & MB_INTRA)
454 for (j=0; j<block_count; j++)
455 iquant_intra(qblocks[k*block_count+j],
456 qblocks[k*block_count+j],
458 cur_picture.mbinfo[k].mquant);
460 for (j=0;j<block_count;j++)
461 iquant_non_intra(qblocks[k*block_count+j],
462 qblocks[k*block_count+j],
463 cur_picture.mbinfo[k].mquant);