jfdctfst.c 7.8 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233
  1. /*
  2. * jfdctfst.c
  3. *
  4. * Copyright (C) 1994-1996, Thomas G. Lane.
  5. * Modified 2003-2015 by Guido Vollbeding.
  6. * This file is part of the Independent JPEG Group's software.
  7. * For conditions of distribution and use, see the accompanying README file.
  8. *
  9. * This file contains a fast, not so accurate integer implementation of the
  10. * forward DCT (Discrete Cosine Transform).
  11. *
  12. * A 2-D DCT can be done by 1-D DCT on each row followed by 1-D DCT
  13. * on each column. Direct algorithms are also available, but they are
  14. * much more complex and seem not to be any faster when reduced to code.
  15. *
  16. * This implementation is based on Arai, Agui, and Nakajima's algorithm for
  17. * scaled DCT. Their original paper (Trans. IEICE E-71(11):1095) is in
  18. * Japanese, but the algorithm is described in the Pennebaker & Mitchell
  19. * JPEG textbook (see REFERENCES section in file README). The following code
  20. * is based directly on figure 4-8 in P&M.
  21. * While an 8-point DCT cannot be done in less than 11 multiplies, it is
  22. * possible to arrange the computation so that many of the multiplies are
  23. * simple scalings of the final outputs. These multiplies can then be
  24. * folded into the multiplications or divisions by the JPEG quantization
  25. * table entries. The AA&N method leaves only 5 multiplies and 29 adds
  26. * to be done in the DCT itself.
  27. * The primary disadvantage of this method is that with fixed-point math,
  28. * accuracy is lost due to imprecise representation of the scaled
  29. * quantization values. The smaller the quantization table entry, the less
  30. * precise the scaled value, so this implementation does worse with high-
  31. * quality-setting files than with low-quality ones.
  32. */
  33. #define JPEG_INTERNALS
  34. #include "jinclude.h"
  35. #include "jpeglib.h"
  36. #include "jdct.h" /* Private declarations for DCT subsystem */
  37. #ifdef DCT_IFAST_SUPPORTED
  38. /*
  39. * This module is specialized to the case DCTSIZE = 8.
  40. */
  41. #if DCTSIZE != 8
  42. Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
  43. #endif
  44. /* Scaling decisions are generally the same as in the LL&M algorithm;
  45. * see jfdctint.c for more details. However, we choose to descale
  46. * (right shift) multiplication products as soon as they are formed,
  47. * rather than carrying additional fractional bits into subsequent additions.
  48. * This compromises accuracy slightly, but it lets us save a few shifts.
  49. * More importantly, 16-bit arithmetic is then adequate (for 8-bit samples)
  50. * everywhere except in the multiplications proper; this saves a good deal
  51. * of work on 16-bit-int machines.
  52. *
  53. * Again to save a few shifts, the intermediate results between pass 1 and
  54. * pass 2 are not upscaled, but are represented only to integral precision.
  55. *
  56. * A final compromise is to represent the multiplicative constants to only
  57. * 8 fractional bits, rather than 13. This saves some shifting work on some
  58. * machines, and may also reduce the cost of multiplication (since there
  59. * are fewer one-bits in the constants).
  60. */
  61. #define CONST_BITS 8
  62. /* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
  63. * causing a lot of useless floating-point operations at run time.
  64. * To get around this we use the following pre-calculated constants.
  65. * If you change CONST_BITS you may want to add appropriate values.
  66. * (With a reasonable C compiler, you can just rely on the FIX() macro...)
  67. */
  68. #if CONST_BITS == 8
  69. #define FIX_0_382683433 ((INT32) 98) /* FIX(0.382683433) */
  70. #define FIX_0_541196100 ((INT32) 139) /* FIX(0.541196100) */
  71. #define FIX_0_707106781 ((INT32) 181) /* FIX(0.707106781) */
  72. #define FIX_1_306562965 ((INT32) 334) /* FIX(1.306562965) */
  73. #else
  74. #define FIX_0_382683433 FIX(0.382683433)
  75. #define FIX_0_541196100 FIX(0.541196100)
  76. #define FIX_0_707106781 FIX(0.707106781)
  77. #define FIX_1_306562965 FIX(1.306562965)
  78. #endif
  79. /* We can gain a little more speed, with a further compromise in accuracy,
  80. * by omitting the addition in a descaling shift. This yields an incorrectly
  81. * rounded result half the time...
  82. */
  83. #ifndef USE_ACCURATE_ROUNDING
  84. #undef DESCALE
  85. #define DESCALE(x,n) RIGHT_SHIFT(x, n)
  86. #endif
  87. /* Multiply a DCTELEM variable by an INT32 constant, and immediately
  88. * descale to yield a DCTELEM result.
  89. */
  90. #define MULTIPLY(var,const) ((DCTELEM) DESCALE((var) * (const), CONST_BITS))
  91. /*
  92. * Perform the forward DCT on one block of samples.
  93. *
  94. * cK represents cos(K*pi/16).
  95. */
  96. GLOBAL(void)
  97. jpeg_fdct_ifast (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
  98. {
  99. DCTELEM tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
  100. DCTELEM tmp10, tmp11, tmp12, tmp13;
  101. DCTELEM z1, z2, z3, z4, z5, z11, z13;
  102. DCTELEM *dataptr;
  103. JSAMPROW elemptr;
  104. int ctr;
  105. SHIFT_TEMPS
  106. /* Pass 1: process rows. */
  107. dataptr = data;
  108. for (ctr = 0; ctr < DCTSIZE; ctr++) {
  109. elemptr = sample_data[ctr] + start_col;
  110. /* Load data into workspace */
  111. tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[7]);
  112. tmp7 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[7]);
  113. tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[6]);
  114. tmp6 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[6]);
  115. tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[5]);
  116. tmp5 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[5]);
  117. tmp3 = GETJSAMPLE(elemptr[3]) + GETJSAMPLE(elemptr[4]);
  118. tmp4 = GETJSAMPLE(elemptr[3]) - GETJSAMPLE(elemptr[4]);
  119. /* Even part */
  120. tmp10 = tmp0 + tmp3; /* phase 2 */
  121. tmp13 = tmp0 - tmp3;
  122. tmp11 = tmp1 + tmp2;
  123. tmp12 = tmp1 - tmp2;
  124. /* Apply unsigned->signed conversion. */
  125. dataptr[0] = tmp10 + tmp11 - 8 * CENTERJSAMPLE; /* phase 3 */
  126. dataptr[4] = tmp10 - tmp11;
  127. z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */
  128. dataptr[2] = tmp13 + z1; /* phase 5 */
  129. dataptr[6] = tmp13 - z1;
  130. /* Odd part */
  131. tmp10 = tmp4 + tmp5; /* phase 2 */
  132. tmp11 = tmp5 + tmp6;
  133. tmp12 = tmp6 + tmp7;
  134. /* The rotator is modified from fig 4-8 to avoid extra negations. */
  135. z5 = MULTIPLY(tmp10 - tmp12, FIX_0_382683433); /* c6 */
  136. z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5; /* c2-c6 */
  137. z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5; /* c2+c6 */
  138. z3 = MULTIPLY(tmp11, FIX_0_707106781); /* c4 */
  139. z11 = tmp7 + z3; /* phase 5 */
  140. z13 = tmp7 - z3;
  141. dataptr[5] = z13 + z2; /* phase 6 */
  142. dataptr[3] = z13 - z2;
  143. dataptr[1] = z11 + z4;
  144. dataptr[7] = z11 - z4;
  145. dataptr += DCTSIZE; /* advance pointer to next row */
  146. }
  147. /* Pass 2: process columns. */
  148. dataptr = data;
  149. for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
  150. tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
  151. tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
  152. tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
  153. tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
  154. tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
  155. tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
  156. tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
  157. tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
  158. /* Even part */
  159. tmp10 = tmp0 + tmp3; /* phase 2 */
  160. tmp13 = tmp0 - tmp3;
  161. tmp11 = tmp1 + tmp2;
  162. tmp12 = tmp1 - tmp2;
  163. dataptr[DCTSIZE*0] = tmp10 + tmp11; /* phase 3 */
  164. dataptr[DCTSIZE*4] = tmp10 - tmp11;
  165. z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */
  166. dataptr[DCTSIZE*2] = tmp13 + z1; /* phase 5 */
  167. dataptr[DCTSIZE*6] = tmp13 - z1;
  168. /* Odd part */
  169. tmp10 = tmp4 + tmp5; /* phase 2 */
  170. tmp11 = tmp5 + tmp6;
  171. tmp12 = tmp6 + tmp7;
  172. /* The rotator is modified from fig 4-8 to avoid extra negations. */
  173. z5 = MULTIPLY(tmp10 - tmp12, FIX_0_382683433); /* c6 */
  174. z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5; /* c2-c6 */
  175. z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5; /* c2+c6 */
  176. z3 = MULTIPLY(tmp11, FIX_0_707106781); /* c4 */
  177. z11 = tmp7 + z3; /* phase 5 */
  178. z13 = tmp7 - z3;
  179. dataptr[DCTSIZE*5] = z13 + z2; /* phase 6 */
  180. dataptr[DCTSIZE*3] = z13 - z2;
  181. dataptr[DCTSIZE*1] = z11 + z4;
  182. dataptr[DCTSIZE*7] = z11 - z4;
  183. dataptr++; /* advance pointer to next column */
  184. }
  185. }
  186. #endif /* DCT_IFAST_SUPPORTED */