Scatter-based scantables.

This gains about 0.2% on derf, 0.1% on hd and 0.4% on stdhd. I can put
this under an experimental flag if wanted, just trying to get my patch
queue in shape.

Change-Id: Ibe1a30fe0e0b07bec4802e0f3ff0ba22e505f576

diff --git a/configure b/configure
index dea79dd..8cd3b93 100755 (executable)
--- a/configure
+++ b/configure
@@ -250,6 +250,7 @@ EXPERIMENT_LIST="
     modelcoefprob
     loop_dering
     implicit_compoundinter_weight
+    scatterscan
 "
 CONFIG_LIST="
     external_build

diff --git a/vp9/common/vp9_entropy.c b/vp9/common/vp9_entropy.c
index 22c2905..a1dbd3e 100644 (file)
--- a/vp9/common/vp9_entropy.c
+++ b/vp9/common/vp9_entropy.c
@@ -63,6 +63,153 @@ DECLARE_ALIGNED(16, const uint8_t, vp9_pt_energy_class[MAX_ENTROPY_TOKENS]) = {
   0, 1, 2, 3, 3, 4, 4, 5, 5, 5, 5, 5
 };
 
+#if CONFIG_SCATTERSCAN
+DECLARE_ALIGNED(16, const int, vp9_default_zig_zag1d_4x4[16]) = {
+  0,  4,  1,  5,
+  8,  2, 12,  9,
+  3,  6, 13, 10,
+  7, 14, 11, 15,
+};
+
+DECLARE_ALIGNED(16, const int, vp9_col_scan_4x4[16]) = {
+  0,  4,  8,  1,
+  12,  5,  9,  2,
+  13,  6, 10,  3,
+  7, 14, 11, 15,
+};
+
+DECLARE_ALIGNED(16, const int, vp9_row_scan_4x4[16]) = {
+  0,  1,  4,  2,
+  5,  3,  6,  8,
+  9,  7, 12, 10,
+  13, 11, 14, 15,
+};
+
+DECLARE_ALIGNED(64, const int, vp9_default_zig_zag1d_8x8[64]) = {
+  0,  8,  1, 16,  9,  2, 17, 24,
+  10,  3, 18, 25, 32, 11,  4, 26,
+  33, 19, 40, 12, 34, 27,  5, 41,
+  20, 48, 13, 35, 42, 28, 21,  6,
+  49, 56, 36, 43, 29,  7, 14, 50,
+  57, 44, 22, 37, 15, 51, 58, 30,
+  45, 23, 52, 59, 38, 31, 60, 53,
+  46, 39, 61, 54, 47, 62, 55, 63,
+};
+
+DECLARE_ALIGNED(16, const int, vp9_col_scan_8x8[64]) = {
+  0,  8, 16,  1, 24,  9, 32, 17,
+  2, 40, 25, 10, 33, 18, 48,  3,
+  26, 41, 11, 56, 19, 34,  4, 49,
+  27, 42, 12, 35, 20, 57, 50, 28,
+  5, 43, 13, 36, 58, 51, 21, 44,
+  6, 29, 59, 37, 14, 52, 22,  7,
+  45, 60, 30, 15, 38, 53, 23, 46,
+  31, 61, 39, 54, 47, 62, 55, 63,
+};
+
+DECLARE_ALIGNED(16, const int, vp9_row_scan_8x8[64]) = {
+  0,  1,  2,  8,  9,  3, 16, 10,
+  4, 17, 11, 24,  5, 18, 25, 12,
+  19, 26, 32,  6, 13, 20, 33, 27,
+  7, 34, 40, 21, 28, 41, 14, 35,
+  48, 42, 29, 36, 49, 22, 43, 15,
+  56, 37, 50, 44, 30, 57, 23, 51,
+  58, 45, 38, 52, 31, 59, 53, 46,
+  60, 39, 61, 47, 54, 55, 62, 63,
+};
+
+DECLARE_ALIGNED(16, const int, vp9_default_zig_zag1d_16x16[256]) = {
+  0,  16,   1,  32,  17,   2,  48,  33,  18,   3,  64,  34,  49,  19,  65,  80,
+  50,   4,  35,  66,  20,  81,  96,  51,   5,  36,  82,  97,  67, 112,  21,  52,
+  98,  37,  83, 113,   6,  68, 128,  53,  22,  99, 114,  84,   7, 129,  38,  69,
+  100, 115, 144, 130,  85,  54,  23,   8, 145,  39,  70, 116, 101, 131, 160, 146,
+  55,  86,  24,  71, 132, 117, 161,  40,   9, 102, 147, 176, 162,  87,  56,  25,
+  133, 118, 177, 148,  72, 103,  41, 163,  10, 192, 178,  88,  57, 134, 149, 119,
+  26, 164,  73, 104, 193,  42, 179, 208,  11, 135,  89, 165, 120, 150,  58, 194,
+  180,  27,  74, 209, 105, 151, 136,  43,  90, 224, 166, 195, 181, 121, 210,  59,
+  12, 152, 106, 167, 196,  75, 137, 225, 211, 240, 182, 122,  91,  28, 197,  13,
+  226, 168, 183, 153,  44, 212, 138, 107, 241,  60,  29, 123, 198, 184, 227, 169,
+  242,  76, 213, 154,  45,  92,  14, 199, 139,  61, 228, 214, 170, 185, 243, 108,
+  77, 155,  30,  15, 200, 229, 124, 215, 244,  93,  46, 186, 171, 201, 109, 140,
+  230,  62, 216, 245,  31, 125,  78, 156, 231,  47, 187, 202, 217,  94, 246, 141,
+  63, 232, 172, 110, 247, 157,  79, 218, 203, 126, 233, 188, 248,  95, 173, 142,
+  219, 111, 249, 234, 158, 127, 189, 204, 250, 235, 143, 174, 220, 205, 159, 251,
+  190, 221, 175, 236, 237, 191, 206, 252, 222, 253, 207, 238, 223, 254, 239, 255,
+};
+
+DECLARE_ALIGNED(16, const int, vp9_col_scan_16x16[256]) = {
+  0,  16,  32,  48,   1,  64,  17,  80,  33,  96,  49,   2,  65, 112,  18,  81,
+  34, 128,  50,  97,   3,  66, 144,  19, 113,  35,  82, 160,  98,  51, 129,   4,
+  67, 176,  20, 114, 145,  83,  36,  99, 130,  52, 192,   5, 161,  68, 115,  21,
+  146,  84, 208, 177,  37, 131, 100,  53, 162, 224,  69,   6, 116, 193, 147,  85,
+  22, 240, 132,  38, 178, 101, 163,  54, 209, 117,  70,   7, 148, 194,  86, 179,
+  225,  23, 133,  39, 164,   8, 102, 210, 241,  55, 195, 118, 149,  71, 180,  24,
+  87, 226, 134, 165, 211,  40, 103,  56,  72, 150, 196, 242, 119,   9, 181, 227,
+  88, 166,  25, 135,  41, 104, 212,  57, 151, 197, 120,  73, 243, 182, 136, 167,
+  213,  89,  10, 228, 105, 152, 198,  26,  42, 121, 183, 244, 168,  58, 137, 229,
+  74, 214,  90, 153, 199, 184,  11, 106, 245,  27, 122, 230, 169,  43, 215,  59,
+  200, 138, 185, 246,  75,  12,  91, 154, 216, 231, 107,  28,  44, 201, 123, 170,
+  60, 247, 232,  76, 139,  13,  92, 217, 186, 248, 155, 108,  29, 124,  45, 202,
+  233, 171,  61,  14,  77, 140,  15, 249,  93,  30, 187, 156, 218,  46, 109, 125,
+  62, 172,  78, 203,  31, 141, 234,  94,  47, 188,  63, 157, 110, 250, 219,  79,
+  126, 204, 173, 142,  95, 189, 111, 235, 158, 220, 251, 127, 174, 143, 205, 236,
+  159, 190, 221, 252, 175, 206, 237, 191, 253, 222, 238, 207, 254, 223, 239, 255,
+};
+
+DECLARE_ALIGNED(16, const int, vp9_row_scan_16x16[256]) = {
+  0,   1,   2,  16,   3,  17,   4,  18,  32,   5,  33,  19,   6,  34,  48,  20,
+  49,   7,  35,  21,  50,  64,   8,  36,  65,  22,  51,  37,  80,   9,  66,  52,
+  23,  38,  81,  67,  10,  53,  24,  82,  68,  96,  39,  11,  54,  83,  97,  69,
+  25,  98,  84,  40, 112,  55,  12,  70,  99, 113,  85,  26,  41,  56, 114, 100,
+  13,  71, 128,  86,  27, 115, 101, 129,  42,  57,  72, 116,  14,  87, 130, 102,
+  144,  73, 131, 117,  28,  58,  15,  88,  43, 145, 103, 132, 146, 118,  74, 160,
+  89, 133, 104,  29,  59, 147, 119,  44, 161, 148,  90, 105, 134, 162, 120, 176,
+  75, 135, 149,  30,  60, 163, 177,  45, 121,  91, 106, 164, 178, 150, 192, 136,
+  165, 179,  31, 151, 193,  76, 122,  61, 137, 194, 107, 152, 180, 208,  46, 166,
+  167, 195,  92, 181, 138, 209, 123, 153, 224, 196,  77, 168, 210, 182, 240, 108,
+  197,  62, 154, 225, 183, 169, 211,  47, 139,  93, 184, 226, 212, 241, 198, 170,
+  124, 155, 199,  78, 213, 185, 109, 227, 200,  63, 228, 242, 140, 214, 171, 186,
+  156, 229, 243, 125,  94, 201, 244, 215, 216, 230, 141, 187, 202,  79, 172, 110,
+  157, 245, 217, 231,  95, 246, 232, 126, 203, 247, 233, 173, 218, 142, 111, 158,
+  188, 248, 127, 234, 219, 249, 189, 204, 143, 174, 159, 250, 235, 205, 220, 175,
+  190, 251, 221, 191, 206, 236, 207, 237, 252, 222, 253, 223, 238, 239, 254, 255,
+};
+
+DECLARE_ALIGNED(16, const int, vp9_default_zig_zag1d_32x32[1024]) = {
+  0,   32,    1,   64,   33,    2,   96,   65,   34,  128,    3,   97,   66,  160,  129,   35,   98,    4,   67,  130,  161,  192,   36,   99,  224,    5,  162,  193,   68,  131,   37,  100,
+  225,  194,  256,  163,   69,  132,    6,  226,  257,  288,  195,  101,  164,   38,  258,    7,  227,  289,  133,  320,   70,  196,  165,  290,  259,  228,   39,  321,  102,  352,    8,  197,
+  71,  134,  322,  291,  260,  353,  384,  229,  166,  103,   40,  354,  323,  292,  135,  385,  198,  261,   72,    9,  416,  167,  386,  355,  230,  324,  104,  293,   41,  417,  199,  136,
+  262,  387,  448,  325,  356,   10,   73,  418,  231,  168,  449,  294,  388,  105,  419,  263,   42,  200,  357,  450,  137,  480,   74,  326,  232,   11,  389,  169,  295,  420,  106,  451,
+  481,  358,  264,  327,  201,   43,  138,  512,  482,  390,  296,  233,  170,  421,   75,  452,  359,   12,  513,  265,  483,  328,  107,  202,  514,  544,  422,  391,  453,  139,   44,  234,
+  484,  297,  360,  171,   76,  515,  545,  266,  329,  454,   13,  423,  392,  203,  108,  546,  485,  576,  298,  235,  140,  361,  516,  330,  172,  547,   45,  424,  455,  267,  393,  577,
+  486,   77,  204,  517,  362,  548,  608,   14,  456,  299,  578,  109,  236,  425,  394,  487,  609,  331,  141,  579,  518,   46,  268,   15,  173,  549,  610,  640,  363,   78,  519,  488,
+  300,  205,   16,  457,  580,  426,  550,  395,  110,  237,  611,  641,  332,  672,  142,  642,  269,  458,   47,  581,  427,  489,  174,  364,  520,  612,  551,  673,   79,  206,  301,  643,
+  704,   17,  111,  490,  674,  238,  582,   48,  521,  613,  333,  396,  459,  143,  270,  552,  644,  705,  736,  365,   80,  675,  583,  175,  428,  706,  112,  302,  207,  614,  553,   49,
+  645,  522,  737,  397,  768,  144,  334,   18,  676,  491,  239,  615,  707,  584,   81,  460,  176,  271,  738,  429,  113,  800,  366,  208,  523,  708,  646,  554,  677,  769,   19,  145,
+  585,  739,  240,  303,   50,  461,  616,  398,  647,  335,  492,  177,   82,  770,  832,  555,  272,  430,  678,  209,  709,  114,  740,  801,  617,   51,  304,  679,  524,  367,  586,  241,
+  20,  146,  771,  864,   83,  802,  648,  493,  399,  273,  336,  710,  178,  462,  833,  587,  741,  115,  305,  711,  368,  525,  618,  803,  210,  896,  680,  834,  772,   52,  649,  147,
+  431,  494,  556,  242,  400,  865,  337,   21,  928,  179,  742,   84,  463,  274,  369,  804,  650,  557,  743,  960,  835,  619,  773,  306,  211,  526,  432,  992,  588,  712,  116,  243,
+  866,  495,  681,  558,  805,  589,  401,  897,   53,  338,  148,  682,  867,  464,  275,   22,  370,  433,  307,  620,  527,  836,  774,  651,  713,  744,   85,  180,  621,  465,  929,  775,
+  496,  898,  212,  339,  244,  402,  590,  117,  559,  714,  434,   23,  868,  930,  806,  683,  528,  652,  371,  961,  149,  837,   54,  899,  745,  276,  993,  497,  403,  622,  181,  776,
+  746,  529,  560,  435,   86,  684,  466,  308,  591,  653,  715,  807,  340,  869,  213,  962,  245,  838,  561,  931,  808,  592,  118,  498,  372,  623,  685,  994,  467,  654,  747,  900,
+  716,  277,  150,   55,   24,  404,  530,  839,  777,  655,  182,  963,  840,  686,  778,  309,  870,  341,   87,  499,  809,  624,  593,  436,  717,  932,  214,  246,  995,  718,  625,  373,
+  562,   25,  119,  901,  531,  468,  964,  748,  810,  278,  779,  500,  563,  656,  405,  687,  871,  872,  594,  151,  933,  749,  841,  310,  657,  626,  595,  437,  688,  183,  996,  965,
+  902,  811,  342,  750,  689,  719,  532,   56,  215,  469,  934,  374,  247,  720,  780,  564,  781,  842,  406,   26,  751,  903,  873,   57,  279,  627,  501,  658,  843,  997,  812,  904,
+  88,  813,  438,  752,  935,  936,  311,  596,  533,  690,  343,  966,  874,   89,  120,  470,  721,  875,  659,  782,  565,  998,  375,  844,  845,   27,  628,  967,  121,  905,  968,  152,
+  937,  814,  753,  502,  691,  783,  184,  153,  722,  407,   58,  815,  999,  660,  597,  723,  534,  906,  216,  439,  907,  248,  185,  876,  846,  692,  784,  629,   90,  969,  280,  754,
+  938,  939,  217,  847,  566,  471,  785,  816,  877, 1000,  249,  878,  661,  503,  312,  970,  755,  122,  817,  281,  344,  786,  598,  724,   28,   59,   29,  154,  535,  630,  376, 1001,
+  313,  908,  186,   91,  848,  849,  345,  909,  940,  879,  408,  818,  693, 1002,  971,  941,  567,  377,  218,  756,  910,  787,  440,  123,  880,  725,  662,  250,  819, 1003,  282,  972,
+  850,  599,  472,  409,  155,  441,  942,  757,  788,  694,  911,  881,  314,  631,  973,  504,  187, 1004,  346,  473,  851,  943,  820,  726,   60,  505,  219,  378,  912,  974,   30,   31,
+  536,  882, 1005,   92,  251,  663,  944,  913,  283,  695,  883,  568, 1006,  975,  410,  442,  945,  789,  852,  537, 1007,  124,  315,   61,  758,  821,  600,  914,  976,  569,  474,  347,
+  156, 1008,  915,   93,  977,  506,  946,  727,  379,  884,  188,  632,  601, 1009,  790,  853,  978,  947,  220,  411,  125,  633,  664,  759,  252,  443,  916,  538,  157,  822,   62,  570,
+  979,  284, 1010,  885,  948,  189,  475,   94,  316,  665,  696, 1011,  854,  791,  980,  221,  348,   63,  917,  602,  380,  507,  253,  126,  697,  823,  634,  285,  728,  949,  886,   95,
+  158,  539, 1012,  317,  412,  444,  760,  571,  190,  981,  729,  918,  127,  666,  349,  381,  476,  855,  761, 1013,  603,  222,  159,  698,  950,  508,  254,  792,  286,  635,  887,  793,
+  413,  191,  982,  445,  540,  318,  730,  667,  223,  824,  919, 1014,  350,  477,  572,  255,  825,  951,  762,  509,  604,  856,  382,  699,  287,  319,  636,  983,  794,  414,  541,  731,
+  857,  888,  351,  446,  573, 1015,  668,  889,  478,  826,  383,  763,  605,  920,  510,  637,  415,  700,  921,  858,  447,  952,  542,  795,  479,  953,  732,  890,  669,  574,  511,  984,
+  827,  985,  922, 1016,  764,  606,  543,  701,  859,  638, 1017,  575,  796,  954,  733,  891,  670,  607,  828,  986,  765,  923,  639, 1018,  702,  860,  955,  671,  892,  734,  797,  703,
+  987,  829, 1019,  766,  924,  735,  861,  956,  988,  893,  767,  798,  830, 1020,  925,  957,  799,  862,  831,  989,  894, 1021,  863,  926,  895,  958,  990, 1022,  927,  959,  991, 1023,
+};
+#else  // CONFIG_SCATTERSCAN
 DECLARE_ALIGNED(16, const int, vp9_default_zig_zag1d_4x4[16]) = {
   0,  1,  4,  8,
   5,  2,  3,  6,
@@ -220,6 +367,7 @@ DECLARE_ALIGNED(16, const int, vp9_default_zig_zag1d_32x32[1024]) = {
   951,  920,  889,  858,  827,  796,  765,  734,  703,  735,  766,  797,  828,  859,  890,  921,  952,  983, 1014, 1015,  984,  953,  922,  891,  860,  829,  798,  767,  799,  830,  861,  892,
   923,  954,  985, 1016, 1017,  986,  955,  924,  893,  862,  831,  863,  894,  925,  956,  987, 1018, 1019,  988,  957,  926,  895,  927,  958,  989, 1020, 1021,  990,  959,  991, 1022, 1023,
 };
+#endif  // CONFIG_SCATTERSCAN
 
 /* Array indices are identical to previously-existing CONTEXT_NODE indices */
 

diff --git a/vp9/decoder/vp9_dequantize.c b/vp9/decoder/vp9_dequantize.c
index 57eed17..9aebcdc 100644 (file)
--- a/vp9/decoder/vp9_dequantize.c
+++ b/vp9/decoder/vp9_dequantize.c
@@ -218,6 +218,7 @@ void vp9_dequant_idct_add_8x8_c(int16_t *input, const int16_t *dq,
     input[0] = 0;
 
     vp9_add_constant_residual_8x8(out, pred, pitch, dest, stride);
+#if !CONFIG_SCATTERSCAN
   } else if (eob <= 10) {
     input[1] *= dq[1];
     input[2] *= dq[1];
@@ -237,6 +238,7 @@ void vp9_dequant_idct_add_8x8_c(int16_t *input, const int16_t *dq,
     input[24] = 0;
 
     vp9_add_residual_8x8(output, pred, pitch, dest, stride);
+#endif
   } else {
     int i;
 
@@ -301,6 +303,7 @@ void vp9_dequant_idct_add_16x16_c(int16_t *input, const int16_t *dq,
     input[0] = 0;
 
     vp9_add_constant_residual_16x16(out, pred, pitch, dest, stride);
+#if !CONFIG_SCATTERSCAN
   } else if (eob <= 10) {
     input[0] *= dq[0];
 
@@ -323,6 +326,7 @@ void vp9_dequant_idct_add_16x16_c(int16_t *input, const int16_t *dq,
     input[48] = 0;
 
     vp9_add_residual_16x16(output, pred, pitch, dest, stride);
+#endif
   } else {
     int i;
 
@@ -352,6 +356,7 @@ void vp9_dequant_idct_add_32x32_c(int16_t *input, const int16_t *dq,
       vp9_short_idct1_32x32(input, output);
       vp9_add_constant_residual_32x32(output[0], pred, pitch, dest, stride);
       input[0] = 0;
+#if !CONFIG_SCATTERSCAN
     } else if (eob <= 10) {
       input[1] = input[1] * dq[1] / 2;
       input[2] = input[2] * dq[1] / 2;
@@ -372,6 +377,7 @@ void vp9_dequant_idct_add_32x32_c(int16_t *input, const int16_t *dq,
       input[96] = 0;
 
       vp9_add_residual_32x32(output, pred, pitch, dest, stride);
+#endif
     } else {
       int i;
       for (i = 1; i < 1024; i++)