0002-arm-asm-primitives.patch 24 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354
  1. Fixes "arm assembly fail to compile on 1.8"
  2. Downloaded from upstream bug report:
  3. https://bitbucket.org/multicoreware/x265/issues/406
  4. Signed-off-by: Bernd Kuhls <bernd.kuhls@t-online.de>
  5. --- ./source/common/arm/asm-primitives.cpp.orig 2018-05-21 02:33:10.000000000 -0600
  6. +++ ./source/common/arm/asm-primitives.cpp 2018-05-28 20:38:37.302378303 -0600
  7. @@ -48,77 +48,77 @@ void setupAssemblyPrimitives(EncoderPrim
  8. p.ssim_4x4x2_core = PFX(ssim_4x4x2_core_neon);
  9. // addAvg
  10. - p.pu[LUMA_4x4].addAvg = PFX(addAvg_4x4_neon);
  11. - p.pu[LUMA_4x8].addAvg = PFX(addAvg_4x8_neon);
  12. - p.pu[LUMA_4x16].addAvg = PFX(addAvg_4x16_neon);
  13. - p.pu[LUMA_8x4].addAvg = PFX(addAvg_8x4_neon);
  14. - p.pu[LUMA_8x8].addAvg = PFX(addAvg_8x8_neon);
  15. - p.pu[LUMA_8x16].addAvg = PFX(addAvg_8x16_neon);
  16. - p.pu[LUMA_8x32].addAvg = PFX(addAvg_8x32_neon);
  17. - p.pu[LUMA_12x16].addAvg = PFX(addAvg_12x16_neon);
  18. - p.pu[LUMA_16x4].addAvg = PFX(addAvg_16x4_neon);
  19. - p.pu[LUMA_16x8].addAvg = PFX(addAvg_16x8_neon);
  20. - p.pu[LUMA_16x12].addAvg = PFX(addAvg_16x12_neon);
  21. - p.pu[LUMA_16x16].addAvg = PFX(addAvg_16x16_neon);
  22. - p.pu[LUMA_16x32].addAvg = PFX(addAvg_16x32_neon);
  23. - p.pu[LUMA_16x64].addAvg = PFX(addAvg_16x64_neon);
  24. - p.pu[LUMA_24x32].addAvg = PFX(addAvg_24x32_neon);
  25. - p.pu[LUMA_32x8].addAvg = PFX(addAvg_32x8_neon);
  26. - p.pu[LUMA_32x16].addAvg = PFX(addAvg_32x16_neon);
  27. - p.pu[LUMA_32x24].addAvg = PFX(addAvg_32x24_neon);
  28. - p.pu[LUMA_32x32].addAvg = PFX(addAvg_32x32_neon);
  29. - p.pu[LUMA_32x64].addAvg = PFX(addAvg_32x64_neon);
  30. - p.pu[LUMA_48x64].addAvg = PFX(addAvg_48x64_neon);
  31. - p.pu[LUMA_64x16].addAvg = PFX(addAvg_64x16_neon);
  32. - p.pu[LUMA_64x32].addAvg = PFX(addAvg_64x32_neon);
  33. - p.pu[LUMA_64x48].addAvg = PFX(addAvg_64x48_neon);
  34. - p.pu[LUMA_64x64].addAvg = PFX(addAvg_64x64_neon);
  35. + p.pu[LUMA_4x4].addAvg[ALIGNED] = PFX(addAvg_4x4_neon);
  36. + p.pu[LUMA_4x8].addAvg[ALIGNED] = PFX(addAvg_4x8_neon);
  37. + p.pu[LUMA_4x16].addAvg[ALIGNED] = PFX(addAvg_4x16_neon);
  38. + p.pu[LUMA_8x4].addAvg[ALIGNED] = PFX(addAvg_8x4_neon);
  39. + p.pu[LUMA_8x8].addAvg[ALIGNED] = PFX(addAvg_8x8_neon);
  40. + p.pu[LUMA_8x16].addAvg[ALIGNED] = PFX(addAvg_8x16_neon);
  41. + p.pu[LUMA_8x32].addAvg[ALIGNED] = PFX(addAvg_8x32_neon);
  42. + p.pu[LUMA_12x16].addAvg[ALIGNED] = PFX(addAvg_12x16_neon);
  43. + p.pu[LUMA_16x4].addAvg[ALIGNED] = PFX(addAvg_16x4_neon);
  44. + p.pu[LUMA_16x8].addAvg[ALIGNED] = PFX(addAvg_16x8_neon);
  45. + p.pu[LUMA_16x12].addAvg[ALIGNED] = PFX(addAvg_16x12_neon);
  46. + p.pu[LUMA_16x16].addAvg[ALIGNED] = PFX(addAvg_16x16_neon);
  47. + p.pu[LUMA_16x32].addAvg[ALIGNED] = PFX(addAvg_16x32_neon);
  48. + p.pu[LUMA_16x64].addAvg[ALIGNED] = PFX(addAvg_16x64_neon);
  49. + p.pu[LUMA_24x32].addAvg[ALIGNED] = PFX(addAvg_24x32_neon);
  50. + p.pu[LUMA_32x8].addAvg[ALIGNED] = PFX(addAvg_32x8_neon);
  51. + p.pu[LUMA_32x16].addAvg[ALIGNED] = PFX(addAvg_32x16_neon);
  52. + p.pu[LUMA_32x24].addAvg[ALIGNED] = PFX(addAvg_32x24_neon);
  53. + p.pu[LUMA_32x32].addAvg[ALIGNED] = PFX(addAvg_32x32_neon);
  54. + p.pu[LUMA_32x64].addAvg[ALIGNED] = PFX(addAvg_32x64_neon);
  55. + p.pu[LUMA_48x64].addAvg[ALIGNED] = PFX(addAvg_48x64_neon);
  56. + p.pu[LUMA_64x16].addAvg[ALIGNED] = PFX(addAvg_64x16_neon);
  57. + p.pu[LUMA_64x32].addAvg[ALIGNED] = PFX(addAvg_64x32_neon);
  58. + p.pu[LUMA_64x48].addAvg[ALIGNED] = PFX(addAvg_64x48_neon);
  59. + p.pu[LUMA_64x64].addAvg[ALIGNED] = PFX(addAvg_64x64_neon);
  60. // chroma addAvg
  61. - p.chroma[X265_CSP_I420].pu[CHROMA_420_4x2].addAvg = PFX(addAvg_4x2_neon);
  62. - p.chroma[X265_CSP_I420].pu[CHROMA_420_4x4].addAvg = PFX(addAvg_4x4_neon);
  63. - p.chroma[X265_CSP_I420].pu[CHROMA_420_4x8].addAvg = PFX(addAvg_4x8_neon);
  64. - p.chroma[X265_CSP_I420].pu[CHROMA_420_4x16].addAvg = PFX(addAvg_4x16_neon);
  65. - p.chroma[X265_CSP_I420].pu[CHROMA_420_6x8].addAvg = PFX(addAvg_6x8_neon);
  66. - p.chroma[X265_CSP_I420].pu[CHROMA_420_8x2].addAvg = PFX(addAvg_8x2_neon);
  67. - p.chroma[X265_CSP_I420].pu[CHROMA_420_8x4].addAvg = PFX(addAvg_8x4_neon);
  68. - p.chroma[X265_CSP_I420].pu[CHROMA_420_8x6].addAvg = PFX(addAvg_8x6_neon);
  69. - p.chroma[X265_CSP_I420].pu[CHROMA_420_8x8].addAvg = PFX(addAvg_8x8_neon);
  70. - p.chroma[X265_CSP_I420].pu[CHROMA_420_8x16].addAvg = PFX(addAvg_8x16_neon);
  71. - p.chroma[X265_CSP_I420].pu[CHROMA_420_8x32].addAvg = PFX(addAvg_8x32_neon);
  72. - p.chroma[X265_CSP_I420].pu[CHROMA_420_12x16].addAvg = PFX(addAvg_12x16_neon);
  73. - p.chroma[X265_CSP_I420].pu[CHROMA_420_16x4].addAvg = PFX(addAvg_16x4_neon);
  74. - p.chroma[X265_CSP_I420].pu[CHROMA_420_16x8].addAvg = PFX(addAvg_16x8_neon);
  75. - p.chroma[X265_CSP_I420].pu[CHROMA_420_16x12].addAvg = PFX(addAvg_16x12_neon);
  76. - p.chroma[X265_CSP_I420].pu[CHROMA_420_16x16].addAvg = PFX(addAvg_16x16_neon);
  77. - p.chroma[X265_CSP_I420].pu[CHROMA_420_16x32].addAvg = PFX(addAvg_16x32_neon);
  78. - p.chroma[X265_CSP_I420].pu[CHROMA_420_24x32].addAvg = PFX(addAvg_24x32_neon);
  79. - p.chroma[X265_CSP_I420].pu[CHROMA_420_32x8].addAvg = PFX(addAvg_32x8_neon);
  80. - p.chroma[X265_CSP_I420].pu[CHROMA_420_32x16].addAvg = PFX(addAvg_32x16_neon);
  81. - p.chroma[X265_CSP_I420].pu[CHROMA_420_32x24].addAvg = PFX(addAvg_32x24_neon);
  82. - p.chroma[X265_CSP_I420].pu[CHROMA_420_32x32].addAvg = PFX(addAvg_32x32_neon);
  83. -
  84. - p.chroma[X265_CSP_I422].pu[CHROMA_422_4x8].addAvg = PFX(addAvg_4x8_neon);
  85. - p.chroma[X265_CSP_I422].pu[CHROMA_422_4x16].addAvg = PFX(addAvg_4x16_neon);
  86. - p.chroma[X265_CSP_I422].pu[CHROMA_422_4x32].addAvg = PFX(addAvg_4x32_neon);
  87. - p.chroma[X265_CSP_I422].pu[CHROMA_422_6x16].addAvg = PFX(addAvg_6x16_neon);
  88. - p.chroma[X265_CSP_I422].pu[CHROMA_422_8x4].addAvg = PFX(addAvg_8x4_neon);
  89. - p.chroma[X265_CSP_I422].pu[CHROMA_422_8x8].addAvg = PFX(addAvg_8x8_neon);
  90. - p.chroma[X265_CSP_I422].pu[CHROMA_422_8x12].addAvg = PFX(addAvg_8x12_neon);
  91. - p.chroma[X265_CSP_I422].pu[CHROMA_422_8x16].addAvg = PFX(addAvg_8x16_neon);
  92. - p.chroma[X265_CSP_I422].pu[CHROMA_422_8x32].addAvg = PFX(addAvg_8x32_neon);
  93. - p.chroma[X265_CSP_I422].pu[CHROMA_422_8x64].addAvg = PFX(addAvg_8x64_neon);
  94. - p.chroma[X265_CSP_I422].pu[CHROMA_422_12x32].addAvg = PFX(addAvg_12x32_neon);
  95. - p.chroma[X265_CSP_I422].pu[CHROMA_422_16x8].addAvg = PFX(addAvg_16x8_neon);
  96. - p.chroma[X265_CSP_I422].pu[CHROMA_422_16x16].addAvg = PFX(addAvg_16x16_neon);
  97. - p.chroma[X265_CSP_I422].pu[CHROMA_422_16x24].addAvg = PFX(addAvg_16x24_neon);
  98. - p.chroma[X265_CSP_I422].pu[CHROMA_422_16x32].addAvg = PFX(addAvg_16x32_neon);
  99. - p.chroma[X265_CSP_I422].pu[CHROMA_422_16x64].addAvg = PFX(addAvg_16x64_neon);
  100. - p.chroma[X265_CSP_I422].pu[CHROMA_422_24x64].addAvg = PFX(addAvg_24x64_neon);
  101. - p.chroma[X265_CSP_I422].pu[CHROMA_422_32x16].addAvg = PFX(addAvg_32x16_neon);
  102. - p.chroma[X265_CSP_I422].pu[CHROMA_422_32x32].addAvg = PFX(addAvg_32x32_neon);
  103. - p.chroma[X265_CSP_I422].pu[CHROMA_422_32x48].addAvg = PFX(addAvg_32x48_neon);
  104. - p.chroma[X265_CSP_I422].pu[CHROMA_422_32x64].addAvg = PFX(addAvg_32x64_neon);
  105. + p.chroma[X265_CSP_I420].pu[CHROMA_420_4x2].addAvg[ALIGNED] = PFX(addAvg_4x2_neon);
  106. + p.chroma[X265_CSP_I420].pu[CHROMA_420_4x4].addAvg[ALIGNED] = PFX(addAvg_4x4_neon);
  107. + p.chroma[X265_CSP_I420].pu[CHROMA_420_4x8].addAvg[ALIGNED] = PFX(addAvg_4x8_neon);
  108. + p.chroma[X265_CSP_I420].pu[CHROMA_420_4x16].addAvg[ALIGNED] = PFX(addAvg_4x16_neon);
  109. + p.chroma[X265_CSP_I420].pu[CHROMA_420_6x8].addAvg[ALIGNED] = PFX(addAvg_6x8_neon);
  110. + p.chroma[X265_CSP_I420].pu[CHROMA_420_8x2].addAvg[ALIGNED] = PFX(addAvg_8x2_neon);
  111. + p.chroma[X265_CSP_I420].pu[CHROMA_420_8x4].addAvg[ALIGNED] = PFX(addAvg_8x4_neon);
  112. + p.chroma[X265_CSP_I420].pu[CHROMA_420_8x6].addAvg[ALIGNED] = PFX(addAvg_8x6_neon);
  113. + p.chroma[X265_CSP_I420].pu[CHROMA_420_8x8].addAvg[ALIGNED] = PFX(addAvg_8x8_neon);
  114. + p.chroma[X265_CSP_I420].pu[CHROMA_420_8x16].addAvg[ALIGNED] = PFX(addAvg_8x16_neon);
  115. + p.chroma[X265_CSP_I420].pu[CHROMA_420_8x32].addAvg[ALIGNED] = PFX(addAvg_8x32_neon);
  116. + p.chroma[X265_CSP_I420].pu[CHROMA_420_12x16].addAvg[ALIGNED] = PFX(addAvg_12x16_neon);
  117. + p.chroma[X265_CSP_I420].pu[CHROMA_420_16x4].addAvg[ALIGNED] = PFX(addAvg_16x4_neon);
  118. + p.chroma[X265_CSP_I420].pu[CHROMA_420_16x8].addAvg[ALIGNED] = PFX(addAvg_16x8_neon);
  119. + p.chroma[X265_CSP_I420].pu[CHROMA_420_16x12].addAvg[ALIGNED] = PFX(addAvg_16x12_neon);
  120. + p.chroma[X265_CSP_I420].pu[CHROMA_420_16x16].addAvg[ALIGNED] = PFX(addAvg_16x16_neon);
  121. + p.chroma[X265_CSP_I420].pu[CHROMA_420_16x32].addAvg[ALIGNED] = PFX(addAvg_16x32_neon);
  122. + p.chroma[X265_CSP_I420].pu[CHROMA_420_24x32].addAvg[ALIGNED] = PFX(addAvg_24x32_neon);
  123. + p.chroma[X265_CSP_I420].pu[CHROMA_420_32x8].addAvg[ALIGNED] = PFX(addAvg_32x8_neon);
  124. + p.chroma[X265_CSP_I420].pu[CHROMA_420_32x16].addAvg[ALIGNED] = PFX(addAvg_32x16_neon);
  125. + p.chroma[X265_CSP_I420].pu[CHROMA_420_32x24].addAvg[ALIGNED] = PFX(addAvg_32x24_neon);
  126. + p.chroma[X265_CSP_I420].pu[CHROMA_420_32x32].addAvg[ALIGNED] = PFX(addAvg_32x32_neon);
  127. +
  128. + p.chroma[X265_CSP_I422].pu[CHROMA_422_4x8].addAvg[ALIGNED] = PFX(addAvg_4x8_neon);
  129. + p.chroma[X265_CSP_I422].pu[CHROMA_422_4x16].addAvg[ALIGNED] = PFX(addAvg_4x16_neon);
  130. + p.chroma[X265_CSP_I422].pu[CHROMA_422_4x32].addAvg[ALIGNED] = PFX(addAvg_4x32_neon);
  131. + p.chroma[X265_CSP_I422].pu[CHROMA_422_6x16].addAvg[ALIGNED] = PFX(addAvg_6x16_neon);
  132. + p.chroma[X265_CSP_I422].pu[CHROMA_422_8x4].addAvg[ALIGNED] = PFX(addAvg_8x4_neon);
  133. + p.chroma[X265_CSP_I422].pu[CHROMA_422_8x8].addAvg[ALIGNED] = PFX(addAvg_8x8_neon);
  134. + p.chroma[X265_CSP_I422].pu[CHROMA_422_8x12].addAvg[ALIGNED] = PFX(addAvg_8x12_neon);
  135. + p.chroma[X265_CSP_I422].pu[CHROMA_422_8x16].addAvg[ALIGNED] = PFX(addAvg_8x16_neon);
  136. + p.chroma[X265_CSP_I422].pu[CHROMA_422_8x32].addAvg[ALIGNED] = PFX(addAvg_8x32_neon);
  137. + p.chroma[X265_CSP_I422].pu[CHROMA_422_8x64].addAvg[ALIGNED] = PFX(addAvg_8x64_neon);
  138. + p.chroma[X265_CSP_I422].pu[CHROMA_422_12x32].addAvg[ALIGNED] = PFX(addAvg_12x32_neon);
  139. + p.chroma[X265_CSP_I422].pu[CHROMA_422_16x8].addAvg[ALIGNED] = PFX(addAvg_16x8_neon);
  140. + p.chroma[X265_CSP_I422].pu[CHROMA_422_16x16].addAvg[ALIGNED] = PFX(addAvg_16x16_neon);
  141. + p.chroma[X265_CSP_I422].pu[CHROMA_422_16x24].addAvg[ALIGNED] = PFX(addAvg_16x24_neon);
  142. + p.chroma[X265_CSP_I422].pu[CHROMA_422_16x32].addAvg[ALIGNED] = PFX(addAvg_16x32_neon);
  143. + p.chroma[X265_CSP_I422].pu[CHROMA_422_16x64].addAvg[ALIGNED] = PFX(addAvg_16x64_neon);
  144. + p.chroma[X265_CSP_I422].pu[CHROMA_422_24x64].addAvg[ALIGNED] = PFX(addAvg_24x64_neon);
  145. + p.chroma[X265_CSP_I422].pu[CHROMA_422_32x16].addAvg[ALIGNED] = PFX(addAvg_32x16_neon);
  146. + p.chroma[X265_CSP_I422].pu[CHROMA_422_32x32].addAvg[ALIGNED] = PFX(addAvg_32x32_neon);
  147. + p.chroma[X265_CSP_I422].pu[CHROMA_422_32x48].addAvg[ALIGNED] = PFX(addAvg_32x48_neon);
  148. + p.chroma[X265_CSP_I422].pu[CHROMA_422_32x64].addAvg[ALIGNED] = PFX(addAvg_32x64_neon);
  149. // quant
  150. p.quant = PFX(quant_neon);
  151. @@ -402,7 +402,7 @@ void setupAssemblyPrimitives(EncoderPrim
  152. p.scale2D_64to32 = PFX(scale2D_64to32_neon);
  153. // scale1D_128to64
  154. - p.scale1D_128to64 = PFX(scale1D_128to64_neon);
  155. + p.scale1D_128to64[ALIGNED] = PFX(scale1D_128to64_neon);
  156. // copy_count
  157. p.cu[BLOCK_4x4].copy_cnt = PFX(copy_cnt_4_neon);
  158. @@ -411,37 +411,37 @@ void setupAssemblyPrimitives(EncoderPrim
  159. p.cu[BLOCK_32x32].copy_cnt = PFX(copy_cnt_32_neon);
  160. // filterPixelToShort
  161. - p.pu[LUMA_4x4].convert_p2s = PFX(filterPixelToShort_4x4_neon);
  162. - p.pu[LUMA_4x8].convert_p2s = PFX(filterPixelToShort_4x8_neon);
  163. - p.pu[LUMA_4x16].convert_p2s = PFX(filterPixelToShort_4x16_neon);
  164. - p.pu[LUMA_8x4].convert_p2s = PFX(filterPixelToShort_8x4_neon);
  165. - p.pu[LUMA_8x8].convert_p2s = PFX(filterPixelToShort_8x8_neon);
  166. - p.pu[LUMA_8x16].convert_p2s = PFX(filterPixelToShort_8x16_neon);
  167. - p.pu[LUMA_8x32].convert_p2s = PFX(filterPixelToShort_8x32_neon);
  168. - p.pu[LUMA_12x16].convert_p2s = PFX(filterPixelToShort_12x16_neon);
  169. - p.pu[LUMA_16x4].convert_p2s = PFX(filterPixelToShort_16x4_neon);
  170. - p.pu[LUMA_16x8].convert_p2s = PFX(filterPixelToShort_16x8_neon);
  171. - p.pu[LUMA_16x12].convert_p2s = PFX(filterPixelToShort_16x12_neon);
  172. - p.pu[LUMA_16x16].convert_p2s = PFX(filterPixelToShort_16x16_neon);
  173. - p.pu[LUMA_16x32].convert_p2s = PFX(filterPixelToShort_16x32_neon);
  174. - p.pu[LUMA_16x64].convert_p2s = PFX(filterPixelToShort_16x64_neon);
  175. - p.pu[LUMA_24x32].convert_p2s = PFX(filterPixelToShort_24x32_neon);
  176. - p.pu[LUMA_32x8].convert_p2s = PFX(filterPixelToShort_32x8_neon);
  177. - p.pu[LUMA_32x16].convert_p2s = PFX(filterPixelToShort_32x16_neon);
  178. - p.pu[LUMA_32x24].convert_p2s = PFX(filterPixelToShort_32x24_neon);
  179. - p.pu[LUMA_32x32].convert_p2s = PFX(filterPixelToShort_32x32_neon);
  180. - p.pu[LUMA_32x64].convert_p2s = PFX(filterPixelToShort_32x64_neon);
  181. - p.pu[LUMA_48x64].convert_p2s = PFX(filterPixelToShort_48x64_neon);
  182. - p.pu[LUMA_64x16].convert_p2s = PFX(filterPixelToShort_64x16_neon);
  183. - p.pu[LUMA_64x32].convert_p2s = PFX(filterPixelToShort_64x32_neon);
  184. - p.pu[LUMA_64x48].convert_p2s = PFX(filterPixelToShort_64x48_neon);
  185. - p.pu[LUMA_64x64].convert_p2s = PFX(filterPixelToShort_64x64_neon);
  186. + p.pu[LUMA_4x4].convert_p2s[ALIGNED] = PFX(filterPixelToShort_4x4_neon);
  187. + p.pu[LUMA_4x8].convert_p2s[ALIGNED] = PFX(filterPixelToShort_4x8_neon);
  188. + p.pu[LUMA_4x16].convert_p2s[ALIGNED] = PFX(filterPixelToShort_4x16_neon);
  189. + p.pu[LUMA_8x4].convert_p2s[ALIGNED] = PFX(filterPixelToShort_8x4_neon);
  190. + p.pu[LUMA_8x8].convert_p2s[ALIGNED] = PFX(filterPixelToShort_8x8_neon);
  191. + p.pu[LUMA_8x16].convert_p2s[ALIGNED] = PFX(filterPixelToShort_8x16_neon);
  192. + p.pu[LUMA_8x32].convert_p2s[ALIGNED] = PFX(filterPixelToShort_8x32_neon);
  193. + p.pu[LUMA_12x16].convert_p2s[ALIGNED] = PFX(filterPixelToShort_12x16_neon);
  194. + p.pu[LUMA_16x4].convert_p2s[ALIGNED] = PFX(filterPixelToShort_16x4_neon);
  195. + p.pu[LUMA_16x8].convert_p2s[ALIGNED] = PFX(filterPixelToShort_16x8_neon);
  196. + p.pu[LUMA_16x12].convert_p2s[ALIGNED] = PFX(filterPixelToShort_16x12_neon);
  197. + p.pu[LUMA_16x16].convert_p2s[ALIGNED] = PFX(filterPixelToShort_16x16_neon);
  198. + p.pu[LUMA_16x32].convert_p2s[ALIGNED] = PFX(filterPixelToShort_16x32_neon);
  199. + p.pu[LUMA_16x64].convert_p2s[ALIGNED] = PFX(filterPixelToShort_16x64_neon);
  200. + p.pu[LUMA_24x32].convert_p2s[ALIGNED] = PFX(filterPixelToShort_24x32_neon);
  201. + p.pu[LUMA_32x8].convert_p2s[ALIGNED] = PFX(filterPixelToShort_32x8_neon);
  202. + p.pu[LUMA_32x16].convert_p2s[ALIGNED] = PFX(filterPixelToShort_32x16_neon);
  203. + p.pu[LUMA_32x24].convert_p2s[ALIGNED] = PFX(filterPixelToShort_32x24_neon);
  204. + p.pu[LUMA_32x32].convert_p2s[ALIGNED] = PFX(filterPixelToShort_32x32_neon);
  205. + p.pu[LUMA_32x64].convert_p2s[ALIGNED] = PFX(filterPixelToShort_32x64_neon);
  206. + p.pu[LUMA_48x64].convert_p2s[ALIGNED] = PFX(filterPixelToShort_48x64_neon);
  207. + p.pu[LUMA_64x16].convert_p2s[ALIGNED] = PFX(filterPixelToShort_64x16_neon);
  208. + p.pu[LUMA_64x32].convert_p2s[ALIGNED] = PFX(filterPixelToShort_64x32_neon);
  209. + p.pu[LUMA_64x48].convert_p2s[ALIGNED] = PFX(filterPixelToShort_64x48_neon);
  210. + p.pu[LUMA_64x64].convert_p2s[ALIGNED] = PFX(filterPixelToShort_64x64_neon);
  211. // Block_fill
  212. - p.cu[BLOCK_4x4].blockfill_s = PFX(blockfill_s_4x4_neon);
  213. - p.cu[BLOCK_8x8].blockfill_s = PFX(blockfill_s_8x8_neon);
  214. - p.cu[BLOCK_16x16].blockfill_s = PFX(blockfill_s_16x16_neon);
  215. - p.cu[BLOCK_32x32].blockfill_s = PFX(blockfill_s_32x32_neon);
  216. + p.cu[BLOCK_4x4].blockfill_s[ALIGNED] = PFX(blockfill_s_4x4_neon);
  217. + p.cu[BLOCK_8x8].blockfill_s[ALIGNED] = PFX(blockfill_s_8x8_neon);
  218. + p.cu[BLOCK_16x16].blockfill_s[ALIGNED] = PFX(blockfill_s_16x16_neon);
  219. + p.cu[BLOCK_32x32].blockfill_s[ALIGNED] = PFX(blockfill_s_32x32_neon);
  220. // Blockcopy_ss
  221. p.cu[BLOCK_4x4].copy_ss = PFX(blockcopy_ss_4x4_neon);
  222. @@ -495,21 +495,21 @@ void setupAssemblyPrimitives(EncoderPrim
  223. p.chroma[X265_CSP_I422].cu[BLOCK_422_32x64].copy_sp = PFX(blockcopy_sp_32x64_neon);
  224. // pixel_add_ps
  225. - p.cu[BLOCK_4x4].add_ps = PFX(pixel_add_ps_4x4_neon);
  226. - p.cu[BLOCK_8x8].add_ps = PFX(pixel_add_ps_8x8_neon);
  227. - p.cu[BLOCK_16x16].add_ps = PFX(pixel_add_ps_16x16_neon);
  228. - p.cu[BLOCK_32x32].add_ps = PFX(pixel_add_ps_32x32_neon);
  229. - p.cu[BLOCK_64x64].add_ps = PFX(pixel_add_ps_64x64_neon);
  230. + p.cu[BLOCK_4x4].add_ps[ALIGNED] = PFX(pixel_add_ps_4x4_neon);
  231. + p.cu[BLOCK_8x8].add_ps[ALIGNED] = PFX(pixel_add_ps_8x8_neon);
  232. + p.cu[BLOCK_16x16].add_ps[ALIGNED] = PFX(pixel_add_ps_16x16_neon);
  233. + p.cu[BLOCK_32x32].add_ps[ALIGNED] = PFX(pixel_add_ps_32x32_neon);
  234. + p.cu[BLOCK_64x64].add_ps[ALIGNED] = PFX(pixel_add_ps_64x64_neon);
  235. // chroma add_ps
  236. - p.chroma[X265_CSP_I420].cu[BLOCK_420_4x4].add_ps = PFX(pixel_add_ps_4x4_neon);
  237. - p.chroma[X265_CSP_I420].cu[BLOCK_420_8x8].add_ps = PFX(pixel_add_ps_8x8_neon);
  238. - p.chroma[X265_CSP_I420].cu[BLOCK_420_16x16].add_ps = PFX(pixel_add_ps_16x16_neon);
  239. - p.chroma[X265_CSP_I420].cu[BLOCK_420_32x32].add_ps = PFX(pixel_add_ps_32x32_neon);
  240. - p.chroma[X265_CSP_I422].cu[BLOCK_422_4x8].add_ps = PFX(pixel_add_ps_4x8_neon);
  241. - p.chroma[X265_CSP_I422].cu[BLOCK_422_8x16].add_ps = PFX(pixel_add_ps_8x16_neon);
  242. - p.chroma[X265_CSP_I422].cu[BLOCK_422_16x32].add_ps = PFX(pixel_add_ps_16x32_neon);
  243. - p.chroma[X265_CSP_I422].cu[BLOCK_422_32x64].add_ps = PFX(pixel_add_ps_32x64_neon);
  244. + p.chroma[X265_CSP_I420].cu[BLOCK_420_4x4].add_ps[ALIGNED] = PFX(pixel_add_ps_4x4_neon);
  245. + p.chroma[X265_CSP_I420].cu[BLOCK_420_8x8].add_ps[ALIGNED] = PFX(pixel_add_ps_8x8_neon);
  246. + p.chroma[X265_CSP_I420].cu[BLOCK_420_16x16].add_ps[ALIGNED] = PFX(pixel_add_ps_16x16_neon);
  247. + p.chroma[X265_CSP_I420].cu[BLOCK_420_32x32].add_ps[ALIGNED] = PFX(pixel_add_ps_32x32_neon);
  248. + p.chroma[X265_CSP_I422].cu[BLOCK_422_4x8].add_ps[ALIGNED] = PFX(pixel_add_ps_4x8_neon);
  249. + p.chroma[X265_CSP_I422].cu[BLOCK_422_8x16].add_ps[ALIGNED] = PFX(pixel_add_ps_8x16_neon);
  250. + p.chroma[X265_CSP_I422].cu[BLOCK_422_16x32].add_ps[ALIGNED] = PFX(pixel_add_ps_16x32_neon);
  251. + p.chroma[X265_CSP_I422].cu[BLOCK_422_32x64].add_ps[ALIGNED] = PFX(pixel_add_ps_32x64_neon);
  252. // cpy2Dto1D_shr
  253. p.cu[BLOCK_4x4].cpy2Dto1D_shr = PFX(cpy2Dto1D_shr_4x4_neon);
  254. @@ -518,10 +518,10 @@ void setupAssemblyPrimitives(EncoderPrim
  255. p.cu[BLOCK_32x32].cpy2Dto1D_shr = PFX(cpy2Dto1D_shr_32x32_neon);
  256. // ssd_s
  257. - p.cu[BLOCK_4x4].ssd_s = PFX(pixel_ssd_s_4x4_neon);
  258. - p.cu[BLOCK_8x8].ssd_s = PFX(pixel_ssd_s_8x8_neon);
  259. - p.cu[BLOCK_16x16].ssd_s = PFX(pixel_ssd_s_16x16_neon);
  260. - p.cu[BLOCK_32x32].ssd_s = PFX(pixel_ssd_s_32x32_neon);
  261. + p.cu[BLOCK_4x4].ssd_s[ALIGNED] = PFX(pixel_ssd_s_4x4_neon);
  262. + p.cu[BLOCK_8x8].ssd_s[ALIGNED] = PFX(pixel_ssd_s_8x8_neon);
  263. + p.cu[BLOCK_16x16].ssd_s[ALIGNED] = PFX(pixel_ssd_s_16x16_neon);
  264. + p.cu[BLOCK_32x32].ssd_s[ALIGNED] = PFX(pixel_ssd_s_32x32_neon);
  265. // sse_ss
  266. p.cu[BLOCK_4x4].sse_ss = PFX(pixel_sse_ss_4x4_neon);
  267. @@ -548,10 +548,10 @@ void setupAssemblyPrimitives(EncoderPrim
  268. p.chroma[X265_CSP_I422].cu[BLOCK_422_32x64].sub_ps = PFX(pixel_sub_ps_32x64_neon);
  269. // calc_Residual
  270. - p.cu[BLOCK_4x4].calcresidual = PFX(getResidual4_neon);
  271. - p.cu[BLOCK_8x8].calcresidual = PFX(getResidual8_neon);
  272. - p.cu[BLOCK_16x16].calcresidual = PFX(getResidual16_neon);
  273. - p.cu[BLOCK_32x32].calcresidual = PFX(getResidual32_neon);
  274. + p.cu[BLOCK_4x4].calcresidual[ALIGNED] = PFX(getResidual4_neon);
  275. + p.cu[BLOCK_8x8].calcresidual[ALIGNED] = PFX(getResidual8_neon);
  276. + p.cu[BLOCK_16x16].calcresidual[ALIGNED] = PFX(getResidual16_neon);
  277. + p.cu[BLOCK_32x32].calcresidual[ALIGNED] = PFX(getResidual32_neon);
  278. // sse_pp
  279. p.cu[BLOCK_4x4].sse_pp = PFX(pixel_sse_pp_4x4_neon);
  280. @@ -722,31 +722,31 @@ void setupAssemblyPrimitives(EncoderPrim
  281. p.pu[LUMA_64x64].sad_x4 = PFX(sad_x4_64x64_neon);
  282. // pixel_avg_pp
  283. - p.pu[LUMA_4x4].pixelavg_pp = PFX(pixel_avg_pp_4x4_neon);
  284. - p.pu[LUMA_4x8].pixelavg_pp = PFX(pixel_avg_pp_4x8_neon);
  285. - p.pu[LUMA_4x16].pixelavg_pp = PFX(pixel_avg_pp_4x16_neon);
  286. - p.pu[LUMA_8x4].pixelavg_pp = PFX(pixel_avg_pp_8x4_neon);
  287. - p.pu[LUMA_8x8].pixelavg_pp = PFX(pixel_avg_pp_8x8_neon);
  288. - p.pu[LUMA_8x16].pixelavg_pp = PFX(pixel_avg_pp_8x16_neon);
  289. - p.pu[LUMA_8x32].pixelavg_pp = PFX(pixel_avg_pp_8x32_neon);
  290. - p.pu[LUMA_12x16].pixelavg_pp = PFX(pixel_avg_pp_12x16_neon);
  291. - p.pu[LUMA_16x4].pixelavg_pp = PFX(pixel_avg_pp_16x4_neon);
  292. - p.pu[LUMA_16x8].pixelavg_pp = PFX(pixel_avg_pp_16x8_neon);
  293. - p.pu[LUMA_16x12].pixelavg_pp = PFX(pixel_avg_pp_16x12_neon);
  294. - p.pu[LUMA_16x16].pixelavg_pp = PFX(pixel_avg_pp_16x16_neon);
  295. - p.pu[LUMA_16x32].pixelavg_pp = PFX(pixel_avg_pp_16x32_neon);
  296. - p.pu[LUMA_16x64].pixelavg_pp = PFX(pixel_avg_pp_16x64_neon);
  297. - p.pu[LUMA_24x32].pixelavg_pp = PFX(pixel_avg_pp_24x32_neon);
  298. - p.pu[LUMA_32x8].pixelavg_pp = PFX(pixel_avg_pp_32x8_neon);
  299. - p.pu[LUMA_32x16].pixelavg_pp = PFX(pixel_avg_pp_32x16_neon);
  300. - p.pu[LUMA_32x24].pixelavg_pp = PFX(pixel_avg_pp_32x24_neon);
  301. - p.pu[LUMA_32x32].pixelavg_pp = PFX(pixel_avg_pp_32x32_neon);
  302. - p.pu[LUMA_32x64].pixelavg_pp = PFX(pixel_avg_pp_32x64_neon);
  303. - p.pu[LUMA_48x64].pixelavg_pp = PFX(pixel_avg_pp_48x64_neon);
  304. - p.pu[LUMA_64x16].pixelavg_pp = PFX(pixel_avg_pp_64x16_neon);
  305. - p.pu[LUMA_64x32].pixelavg_pp = PFX(pixel_avg_pp_64x32_neon);
  306. - p.pu[LUMA_64x48].pixelavg_pp = PFX(pixel_avg_pp_64x48_neon);
  307. - p.pu[LUMA_64x64].pixelavg_pp = PFX(pixel_avg_pp_64x64_neon);
  308. + p.pu[LUMA_4x4].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_4x4_neon);
  309. + p.pu[LUMA_4x8].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_4x8_neon);
  310. + p.pu[LUMA_4x16].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_4x16_neon);
  311. + p.pu[LUMA_8x4].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_8x4_neon);
  312. + p.pu[LUMA_8x8].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_8x8_neon);
  313. + p.pu[LUMA_8x16].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_8x16_neon);
  314. + p.pu[LUMA_8x32].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_8x32_neon);
  315. + p.pu[LUMA_12x16].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_12x16_neon);
  316. + p.pu[LUMA_16x4].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_16x4_neon);
  317. + p.pu[LUMA_16x8].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_16x8_neon);
  318. + p.pu[LUMA_16x12].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_16x12_neon);
  319. + p.pu[LUMA_16x16].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_16x16_neon);
  320. + p.pu[LUMA_16x32].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_16x32_neon);
  321. + p.pu[LUMA_16x64].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_16x64_neon);
  322. + p.pu[LUMA_24x32].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_24x32_neon);
  323. + p.pu[LUMA_32x8].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_32x8_neon);
  324. + p.pu[LUMA_32x16].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_32x16_neon);
  325. + p.pu[LUMA_32x24].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_32x24_neon);
  326. + p.pu[LUMA_32x32].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_32x32_neon);
  327. + p.pu[LUMA_32x64].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_32x64_neon);
  328. + p.pu[LUMA_48x64].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_48x64_neon);
  329. + p.pu[LUMA_64x16].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_64x16_neon);
  330. + p.pu[LUMA_64x32].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_64x32_neon);
  331. + p.pu[LUMA_64x48].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_64x48_neon);
  332. + p.pu[LUMA_64x64].pixelavg_pp[ALIGNED] = PFX(pixel_avg_pp_64x64_neon);
  333. // planecopy
  334. p.planecopy_cp = PFX(pixel_planecopy_cp_neon);