upsampling.c 11 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222
  1. // Copyright 2011 Google Inc. All Rights Reserved.
  2. //
  3. // Use of this source code is governed by a BSD-style license
  4. // that can be found in the COPYING file in the root of the source
  5. // tree. An additional intellectual property rights grant can be found
  6. // in the file PATENTS. All contributing project authors may
  7. // be found in the AUTHORS file in the root of the source tree.
  8. // -----------------------------------------------------------------------------
  9. //
  10. // YUV to RGB upsampling functions.
  11. //
  12. // Author: somnath@google.com (Somnath Banerjee)
  13. #include "./dsp.h"
  14. #include "./yuv.h"
  15. #include <assert.h>
  16. //------------------------------------------------------------------------------
  17. // Fancy upsampler
  18. #ifdef FANCY_UPSAMPLING
  19. // Fancy upsampling functions to convert YUV to RGB
  20. WebPUpsampleLinePairFunc WebPUpsamplers[MODE_LAST];
  21. // Given samples laid out in a square as:
  22. // [a b]
  23. // [c d]
  24. // we interpolate u/v as:
  25. // ([9*a + 3*b + 3*c + d 3*a + 9*b + 3*c + d] + [8 8]) / 16
  26. // ([3*a + b + 9*c + 3*d a + 3*b + 3*c + 9*d] [8 8]) / 16
  27. // We process u and v together stashed into 32bit (16bit each).
  28. #define LOAD_UV(u, v) ((u) | ((v) << 16))
  29. #define UPSAMPLE_FUNC(FUNC_NAME, FUNC, XSTEP) \
  30. static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y, \
  31. const uint8_t* top_u, const uint8_t* top_v, \
  32. const uint8_t* cur_u, const uint8_t* cur_v, \
  33. uint8_t* top_dst, uint8_t* bottom_dst, int len) { \
  34. int x; \
  35. const int last_pixel_pair = (len - 1) >> 1; \
  36. uint32_t tl_uv = LOAD_UV(top_u[0], top_v[0]); /* top-left sample */ \
  37. uint32_t l_uv = LOAD_UV(cur_u[0], cur_v[0]); /* left-sample */ \
  38. assert(top_y != NULL); \
  39. { \
  40. const uint32_t uv0 = (3 * tl_uv + l_uv + 0x00020002u) >> 2; \
  41. FUNC(top_y[0], uv0 & 0xff, (uv0 >> 16), top_dst); \
  42. } \
  43. if (bottom_y != NULL) { \
  44. const uint32_t uv0 = (3 * l_uv + tl_uv + 0x00020002u) >> 2; \
  45. FUNC(bottom_y[0], uv0 & 0xff, (uv0 >> 16), bottom_dst); \
  46. } \
  47. for (x = 1; x <= last_pixel_pair; ++x) { \
  48. const uint32_t t_uv = LOAD_UV(top_u[x], top_v[x]); /* top sample */ \
  49. const uint32_t uv = LOAD_UV(cur_u[x], cur_v[x]); /* sample */ \
  50. /* precompute invariant values associated with first and second diagonals*/\
  51. const uint32_t avg = tl_uv + t_uv + l_uv + uv + 0x00080008u; \
  52. const uint32_t diag_12 = (avg + 2 * (t_uv + l_uv)) >> 3; \
  53. const uint32_t diag_03 = (avg + 2 * (tl_uv + uv)) >> 3; \
  54. { \
  55. const uint32_t uv0 = (diag_12 + tl_uv) >> 1; \
  56. const uint32_t uv1 = (diag_03 + t_uv) >> 1; \
  57. FUNC(top_y[2 * x - 1], uv0 & 0xff, (uv0 >> 16), \
  58. top_dst + (2 * x - 1) * XSTEP); \
  59. FUNC(top_y[2 * x - 0], uv1 & 0xff, (uv1 >> 16), \
  60. top_dst + (2 * x - 0) * XSTEP); \
  61. } \
  62. if (bottom_y != NULL) { \
  63. const uint32_t uv0 = (diag_03 + l_uv) >> 1; \
  64. const uint32_t uv1 = (diag_12 + uv) >> 1; \
  65. FUNC(bottom_y[2 * x - 1], uv0 & 0xff, (uv0 >> 16), \
  66. bottom_dst + (2 * x - 1) * XSTEP); \
  67. FUNC(bottom_y[2 * x + 0], uv1 & 0xff, (uv1 >> 16), \
  68. bottom_dst + (2 * x + 0) * XSTEP); \
  69. } \
  70. tl_uv = t_uv; \
  71. l_uv = uv; \
  72. } \
  73. if (!(len & 1)) { \
  74. { \
  75. const uint32_t uv0 = (3 * tl_uv + l_uv + 0x00020002u) >> 2; \
  76. FUNC(top_y[len - 1], uv0 & 0xff, (uv0 >> 16), \
  77. top_dst + (len - 1) * XSTEP); \
  78. } \
  79. if (bottom_y != NULL) { \
  80. const uint32_t uv0 = (3 * l_uv + tl_uv + 0x00020002u) >> 2; \
  81. FUNC(bottom_y[len - 1], uv0 & 0xff, (uv0 >> 16), \
  82. bottom_dst + (len - 1) * XSTEP); \
  83. } \
  84. } \
  85. }
  86. // All variants implemented.
  87. UPSAMPLE_FUNC(UpsampleRgbLinePair, VP8YuvToRgb, 3)
  88. UPSAMPLE_FUNC(UpsampleBgrLinePair, VP8YuvToBgr, 3)
  89. UPSAMPLE_FUNC(UpsampleRgbaLinePair, VP8YuvToRgba, 4)
  90. UPSAMPLE_FUNC(UpsampleBgraLinePair, VP8YuvToBgra, 4)
  91. UPSAMPLE_FUNC(UpsampleArgbLinePair, VP8YuvToArgb, 4)
  92. UPSAMPLE_FUNC(UpsampleRgba4444LinePair, VP8YuvToRgba4444, 2)
  93. UPSAMPLE_FUNC(UpsampleRgb565LinePair, VP8YuvToRgb565, 2)
  94. #undef LOAD_UV
  95. #undef UPSAMPLE_FUNC
  96. #endif // FANCY_UPSAMPLING
  97. //------------------------------------------------------------------------------
  98. #if !defined(FANCY_UPSAMPLING)
  99. #define DUAL_SAMPLE_FUNC(FUNC_NAME, FUNC) \
  100. static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bot_y, \
  101. const uint8_t* top_u, const uint8_t* top_v, \
  102. const uint8_t* bot_u, const uint8_t* bot_v, \
  103. uint8_t* top_dst, uint8_t* bot_dst, int len) { \
  104. const int half_len = len >> 1; \
  105. int x; \
  106. assert(top_dst != NULL); \
  107. { \
  108. for (x = 0; x < half_len; ++x) { \
  109. FUNC(top_y[2 * x + 0], top_u[x], top_v[x], top_dst + 8 * x + 0); \
  110. FUNC(top_y[2 * x + 1], top_u[x], top_v[x], top_dst + 8 * x + 4); \
  111. } \
  112. if (len & 1) FUNC(top_y[2 * x + 0], top_u[x], top_v[x], top_dst + 8 * x); \
  113. } \
  114. if (bot_dst != NULL) { \
  115. for (x = 0; x < half_len; ++x) { \
  116. FUNC(bot_y[2 * x + 0], bot_u[x], bot_v[x], bot_dst + 8 * x + 0); \
  117. FUNC(bot_y[2 * x + 1], bot_u[x], bot_v[x], bot_dst + 8 * x + 4); \
  118. } \
  119. if (len & 1) FUNC(bot_y[2 * x + 0], bot_u[x], bot_v[x], bot_dst + 8 * x); \
  120. } \
  121. }
  122. DUAL_SAMPLE_FUNC(DualLineSamplerBGRA, VP8YuvToBgra)
  123. DUAL_SAMPLE_FUNC(DualLineSamplerARGB, VP8YuvToArgb)
  124. #undef DUAL_SAMPLE_FUNC
  125. #endif // !FANCY_UPSAMPLING
  126. WebPUpsampleLinePairFunc WebPGetLinePairConverter(int alpha_is_last) {
  127. WebPInitUpsamplers();
  128. VP8YUVInit();
  129. #ifdef FANCY_UPSAMPLING
  130. return WebPUpsamplers[alpha_is_last ? MODE_BGRA : MODE_ARGB];
  131. #else
  132. return (alpha_is_last ? DualLineSamplerBGRA : DualLineSamplerARGB);
  133. #endif
  134. }
  135. //------------------------------------------------------------------------------
  136. // YUV444 converter
  137. #define YUV444_FUNC(FUNC_NAME, FUNC, XSTEP) \
  138. static void FUNC_NAME(const uint8_t* y, const uint8_t* u, const uint8_t* v, \
  139. uint8_t* dst, int len) { \
  140. int i; \
  141. for (i = 0; i < len; ++i) FUNC(y[i], u[i], v[i], &dst[i * XSTEP]); \
  142. }
  143. YUV444_FUNC(Yuv444ToRgb, VP8YuvToRgb, 3)
  144. YUV444_FUNC(Yuv444ToBgr, VP8YuvToBgr, 3)
  145. YUV444_FUNC(Yuv444ToRgba, VP8YuvToRgba, 4)
  146. YUV444_FUNC(Yuv444ToBgra, VP8YuvToBgra, 4)
  147. YUV444_FUNC(Yuv444ToArgb, VP8YuvToArgb, 4)
  148. YUV444_FUNC(Yuv444ToRgba4444, VP8YuvToRgba4444, 2)
  149. YUV444_FUNC(Yuv444ToRgb565, VP8YuvToRgb565, 2)
  150. #undef YUV444_FUNC
  151. const WebPYUV444Converter WebPYUV444Converters[MODE_LAST] = {
  152. Yuv444ToRgb, // MODE_RGB
  153. Yuv444ToRgba, // MODE_RGBA
  154. Yuv444ToBgr, // MODE_BGR
  155. Yuv444ToBgra, // MODE_BGRA
  156. Yuv444ToArgb, // MODE_ARGB
  157. Yuv444ToRgba4444, // MODE_RGBA_4444
  158. Yuv444ToRgb565, // MODE_RGB_565
  159. Yuv444ToRgba, // MODE_rgbA
  160. Yuv444ToBgra, // MODE_bgrA
  161. Yuv444ToArgb, // MODE_Argb
  162. Yuv444ToRgba4444 // MODE_rgbA_4444
  163. };
  164. //------------------------------------------------------------------------------
  165. // Main calls
  166. extern void WebPInitUpsamplersSSE2(void);
  167. extern void WebPInitUpsamplersNEON(void);
  168. void WebPInitUpsamplers(void) {
  169. #ifdef FANCY_UPSAMPLING
  170. WebPUpsamplers[MODE_RGB] = UpsampleRgbLinePair;
  171. WebPUpsamplers[MODE_RGBA] = UpsampleRgbaLinePair;
  172. WebPUpsamplers[MODE_BGR] = UpsampleBgrLinePair;
  173. WebPUpsamplers[MODE_BGRA] = UpsampleBgraLinePair;
  174. WebPUpsamplers[MODE_ARGB] = UpsampleArgbLinePair;
  175. WebPUpsamplers[MODE_RGBA_4444] = UpsampleRgba4444LinePair;
  176. WebPUpsamplers[MODE_RGB_565] = UpsampleRgb565LinePair;
  177. WebPUpsamplers[MODE_rgbA] = UpsampleRgbaLinePair;
  178. WebPUpsamplers[MODE_bgrA] = UpsampleBgraLinePair;
  179. WebPUpsamplers[MODE_Argb] = UpsampleArgbLinePair;
  180. WebPUpsamplers[MODE_rgbA_4444] = UpsampleRgba4444LinePair;
  181. // If defined, use CPUInfo() to overwrite some pointers with faster versions.
  182. if (VP8GetCPUInfo != NULL) {
  183. #if defined(WEBP_USE_SSE2)
  184. if (VP8GetCPUInfo(kSSE2)) {
  185. WebPInitUpsamplersSSE2();
  186. }
  187. #endif
  188. #if defined(WEBP_USE_NEON)
  189. if (VP8GetCPUInfo(kNEON)) {
  190. WebPInitUpsamplersNEON();
  191. }
  192. #endif
  193. }
  194. #endif // FANCY_UPSAMPLING
  195. }
  196. //------------------------------------------------------------------------------