jidctflt.c 8.5 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240
  1. /*
  2. * jidctflt.c
  3. *
  4. * This file was part of the Independent JPEG Group's software:
  5. * Copyright (C) 1994-1998, Thomas G. Lane.
  6. * Modified 2010 by Guido Vollbeding.
  7. * libjpeg-turbo Modifications:
  8. * Copyright (C) 2014, D. R. Commander.
  9. * For conditions of distribution and use, see the accompanying README.ijg
  10. * file.
  11. *
  12. * This file contains a floating-point implementation of the
  13. * inverse DCT (Discrete Cosine Transform). In the IJG code, this routine
  14. * must also perform dequantization of the input coefficients.
  15. *
  16. * This implementation should be more accurate than either of the integer
  17. * IDCT implementations. However, it may not give the same results on all
  18. * machines because of differences in roundoff behavior. Speed will depend
  19. * on the hardware's floating point capacity.
  20. *
  21. * A 2-D IDCT can be done by 1-D IDCT on each column followed by 1-D IDCT
  22. * on each row (or vice versa, but it's more convenient to emit a row at
  23. * a time). Direct algorithms are also available, but they are much more
  24. * complex and seem not to be any faster when reduced to code.
  25. *
  26. * This implementation is based on Arai, Agui, and Nakajima's algorithm for
  27. * scaled DCT. Their original paper (Trans. IEICE E-71(11):1095) is in
  28. * Japanese, but the algorithm is described in the Pennebaker & Mitchell
  29. * JPEG textbook (see REFERENCES section in file README.ijg). The following
  30. * code is based directly on figure 4-8 in P&M.
  31. * While an 8-point DCT cannot be done in less than 11 multiplies, it is
  32. * possible to arrange the computation so that many of the multiplies are
  33. * simple scalings of the final outputs. These multiplies can then be
  34. * folded into the multiplications or divisions by the JPEG quantization
  35. * table entries. The AA&N method leaves only 5 multiplies and 29 adds
  36. * to be done in the DCT itself.
  37. * The primary disadvantage of this method is that with a fixed-point
  38. * implementation, accuracy is lost due to imprecise representation of the
  39. * scaled quantization values. However, that problem does not arise if
  40. * we use floating point arithmetic.
  41. */
  42. #define JPEG_INTERNALS
  43. #include "jinclude.h"
  44. #include "jpeglib.h"
  45. #include "jdct.h" /* Private declarations for DCT subsystem */
  46. #ifdef DCT_FLOAT_SUPPORTED
  47. /*
  48. * This module is specialized to the case DCTSIZE = 8.
  49. */
  50. #if DCTSIZE != 8
  51. Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
  52. #endif
  53. /* Dequantize a coefficient by multiplying it by the multiplier-table
  54. * entry; produce a float result.
  55. */
  56. #define DEQUANTIZE(coef, quantval) (((FAST_FLOAT)(coef)) * (quantval))
  57. /*
  58. * Perform dequantization and inverse DCT on one block of coefficients.
  59. */
  60. GLOBAL(void)
  61. jpeg_idct_float(j_decompress_ptr cinfo, jpeg_component_info *compptr,
  62. JCOEFPTR coef_block, JSAMPARRAY output_buf,
  63. JDIMENSION output_col)
  64. {
  65. FAST_FLOAT tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
  66. FAST_FLOAT tmp10, tmp11, tmp12, tmp13;
  67. FAST_FLOAT z5, z10, z11, z12, z13;
  68. JCOEFPTR inptr;
  69. FLOAT_MULT_TYPE *quantptr;
  70. FAST_FLOAT *wsptr;
  71. JSAMPROW outptr;
  72. JSAMPLE *range_limit = cinfo->sample_range_limit;
  73. int ctr;
  74. FAST_FLOAT workspace[DCTSIZE2]; /* buffers data between passes */
  75. #define _0_125 ((FLOAT_MULT_TYPE)0.125)
  76. /* Pass 1: process columns from input, store into work array. */
  77. inptr = coef_block;
  78. quantptr = (FLOAT_MULT_TYPE *)compptr->dct_table;
  79. wsptr = workspace;
  80. for (ctr = DCTSIZE; ctr > 0; ctr--) {
  81. /* Due to quantization, we will usually find that many of the input
  82. * coefficients are zero, especially the AC terms. We can exploit this
  83. * by short-circuiting the IDCT calculation for any column in which all
  84. * the AC terms are zero. In that case each output is equal to the
  85. * DC coefficient (with scale factor as needed).
  86. * With typical images and quantization tables, half or more of the
  87. * column DCT calculations can be simplified this way.
  88. */
  89. if (inptr[DCTSIZE * 1] == 0 && inptr[DCTSIZE * 2] == 0 &&
  90. inptr[DCTSIZE * 3] == 0 && inptr[DCTSIZE * 4] == 0 &&
  91. inptr[DCTSIZE * 5] == 0 && inptr[DCTSIZE * 6] == 0 &&
  92. inptr[DCTSIZE * 7] == 0) {
  93. /* AC terms all zero */
  94. FAST_FLOAT dcval = DEQUANTIZE(inptr[DCTSIZE * 0],
  95. quantptr[DCTSIZE * 0] * _0_125);
  96. wsptr[DCTSIZE * 0] = dcval;
  97. wsptr[DCTSIZE * 1] = dcval;
  98. wsptr[DCTSIZE * 2] = dcval;
  99. wsptr[DCTSIZE * 3] = dcval;
  100. wsptr[DCTSIZE * 4] = dcval;
  101. wsptr[DCTSIZE * 5] = dcval;
  102. wsptr[DCTSIZE * 6] = dcval;
  103. wsptr[DCTSIZE * 7] = dcval;
  104. inptr++; /* advance pointers to next column */
  105. quantptr++;
  106. wsptr++;
  107. continue;
  108. }
  109. /* Even part */
  110. tmp0 = DEQUANTIZE(inptr[DCTSIZE * 0], quantptr[DCTSIZE * 0] * _0_125);
  111. tmp1 = DEQUANTIZE(inptr[DCTSIZE * 2], quantptr[DCTSIZE * 2] * _0_125);
  112. tmp2 = DEQUANTIZE(inptr[DCTSIZE * 4], quantptr[DCTSIZE * 4] * _0_125);
  113. tmp3 = DEQUANTIZE(inptr[DCTSIZE * 6], quantptr[DCTSIZE * 6] * _0_125);
  114. tmp10 = tmp0 + tmp2; /* phase 3 */
  115. tmp11 = tmp0 - tmp2;
  116. tmp13 = tmp1 + tmp3; /* phases 5-3 */
  117. tmp12 = (tmp1 - tmp3) * ((FAST_FLOAT)1.414213562) - tmp13; /* 2*c4 */
  118. tmp0 = tmp10 + tmp13; /* phase 2 */
  119. tmp3 = tmp10 - tmp13;
  120. tmp1 = tmp11 + tmp12;
  121. tmp2 = tmp11 - tmp12;
  122. /* Odd part */
  123. tmp4 = DEQUANTIZE(inptr[DCTSIZE * 1], quantptr[DCTSIZE * 1] * _0_125);
  124. tmp5 = DEQUANTIZE(inptr[DCTSIZE * 3], quantptr[DCTSIZE * 3] * _0_125);
  125. tmp6 = DEQUANTIZE(inptr[DCTSIZE * 5], quantptr[DCTSIZE * 5] * _0_125);
  126. tmp7 = DEQUANTIZE(inptr[DCTSIZE * 7], quantptr[DCTSIZE * 7] * _0_125);
  127. z13 = tmp6 + tmp5; /* phase 6 */
  128. z10 = tmp6 - tmp5;
  129. z11 = tmp4 + tmp7;
  130. z12 = tmp4 - tmp7;
  131. tmp7 = z11 + z13; /* phase 5 */
  132. tmp11 = (z11 - z13) * ((FAST_FLOAT)1.414213562); /* 2*c4 */
  133. z5 = (z10 + z12) * ((FAST_FLOAT)1.847759065); /* 2*c2 */
  134. tmp10 = z5 - z12 * ((FAST_FLOAT)1.082392200); /* 2*(c2-c6) */
  135. tmp12 = z5 - z10 * ((FAST_FLOAT)2.613125930); /* 2*(c2+c6) */
  136. tmp6 = tmp12 - tmp7; /* phase 2 */
  137. tmp5 = tmp11 - tmp6;
  138. tmp4 = tmp10 - tmp5;
  139. wsptr[DCTSIZE * 0] = tmp0 + tmp7;
  140. wsptr[DCTSIZE * 7] = tmp0 - tmp7;
  141. wsptr[DCTSIZE * 1] = tmp1 + tmp6;
  142. wsptr[DCTSIZE * 6] = tmp1 - tmp6;
  143. wsptr[DCTSIZE * 2] = tmp2 + tmp5;
  144. wsptr[DCTSIZE * 5] = tmp2 - tmp5;
  145. wsptr[DCTSIZE * 3] = tmp3 + tmp4;
  146. wsptr[DCTSIZE * 4] = tmp3 - tmp4;
  147. inptr++; /* advance pointers to next column */
  148. quantptr++;
  149. wsptr++;
  150. }
  151. /* Pass 2: process rows from work array, store into output array. */
  152. wsptr = workspace;
  153. for (ctr = 0; ctr < DCTSIZE; ctr++) {
  154. outptr = output_buf[ctr] + output_col;
  155. /* Rows of zeroes can be exploited in the same way as we did with columns.
  156. * However, the column calculation has created many nonzero AC terms, so
  157. * the simplification applies less often (typically 5% to 10% of the time).
  158. * And testing floats for zero is relatively expensive, so we don't bother.
  159. */
  160. /* Even part */
  161. /* Apply signed->unsigned and prepare float->int conversion */
  162. z5 = wsptr[0] + ((FAST_FLOAT)CENTERJSAMPLE + (FAST_FLOAT)0.5);
  163. tmp10 = z5 + wsptr[4];
  164. tmp11 = z5 - wsptr[4];
  165. tmp13 = wsptr[2] + wsptr[6];
  166. tmp12 = (wsptr[2] - wsptr[6]) * ((FAST_FLOAT)1.414213562) - tmp13;
  167. tmp0 = tmp10 + tmp13;
  168. tmp3 = tmp10 - tmp13;
  169. tmp1 = tmp11 + tmp12;
  170. tmp2 = tmp11 - tmp12;
  171. /* Odd part */
  172. z13 = wsptr[5] + wsptr[3];
  173. z10 = wsptr[5] - wsptr[3];
  174. z11 = wsptr[1] + wsptr[7];
  175. z12 = wsptr[1] - wsptr[7];
  176. tmp7 = z11 + z13;
  177. tmp11 = (z11 - z13) * ((FAST_FLOAT)1.414213562);
  178. z5 = (z10 + z12) * ((FAST_FLOAT)1.847759065); /* 2*c2 */
  179. tmp10 = z5 - z12 * ((FAST_FLOAT)1.082392200); /* 2*(c2-c6) */
  180. tmp12 = z5 - z10 * ((FAST_FLOAT)2.613125930); /* 2*(c2+c6) */
  181. tmp6 = tmp12 - tmp7;
  182. tmp5 = tmp11 - tmp6;
  183. tmp4 = tmp10 - tmp5;
  184. /* Final output stage: float->int conversion and range-limit */
  185. outptr[0] = range_limit[((int)(tmp0 + tmp7)) & RANGE_MASK];
  186. outptr[7] = range_limit[((int)(tmp0 - tmp7)) & RANGE_MASK];
  187. outptr[1] = range_limit[((int)(tmp1 + tmp6)) & RANGE_MASK];
  188. outptr[6] = range_limit[((int)(tmp1 - tmp6)) & RANGE_MASK];
  189. outptr[2] = range_limit[((int)(tmp2 + tmp5)) & RANGE_MASK];
  190. outptr[5] = range_limit[((int)(tmp2 - tmp5)) & RANGE_MASK];
  191. outptr[3] = range_limit[((int)(tmp3 + tmp4)) & RANGE_MASK];
  192. outptr[4] = range_limit[((int)(tmp3 - tmp4)) & RANGE_MASK];
  193. wsptr += DCTSIZE; /* advance pointer to next row */
  194. }
  195. }
  196. #endif /* DCT_FLOAT_SUPPORTED */