From: Hui Su Date: Tue, 17 Jul 2018 05:05:19 +0000 (-0700) Subject: ML based rectangular partition search pruning X-Git-Tag: v1.8.0~367^2 X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=2713dba538406c10dd9684f2df877aef014583d8;p=libvpx ML based rectangular partition search pruning Add a ML model to predict if rectangular partition search can be skipped without much coding loss. This model is enabled for speed 0 low bitdepth only. Impact on coding performance is minor: avg_psnr ovr_psnr ssim lowres -0.005% 0.005% 0.017% midres 0.100% 0.114% 0.134% hdres 0.048% 0.083% 0.074% jvet480p 0.035% 0.027% 0.044% jvet720p 0.094% 0.090% 0.174% Tested encoding speed over 20 midres and hdres clips, average speed gain is about 8%; maximum speed gain is 23%. Change-Id: I5d4029dec7134c53ac68ab6cf0c8077dc0b767ed --- diff --git a/vp9/encoder/vp9_context_tree.h b/vp9/encoder/vp9_context_tree.h index 2bcc26e94..c4600d4bc 100644 --- a/vp9/encoder/vp9_context_tree.h +++ b/vp9/encoder/vp9_context_tree.h @@ -56,6 +56,7 @@ typedef struct { // scope of refactoring. int rate; int64_t dist; + int64_t rdcost; #if CONFIG_VP9_TEMPORAL_DENOISING unsigned int newmv_sse; diff --git a/vp9/encoder/vp9_encodeframe.c b/vp9/encoder/vp9_encodeframe.c index 1b559fad3..8bba7d98a 100644 --- a/vp9/encoder/vp9_encodeframe.c +++ b/vp9/encoder/vp9_encodeframe.c @@ -3622,6 +3622,604 @@ static int ml_predict_breakout(const VP9_COMP *const cpi, BLOCK_SIZE bsize, #undef Q_CTX #undef RESOLUTION_CTX +#define FEATURES 17 +#define LABELS 4 +static const float rect_part_nn_weights_16_layer0[FEATURES * 32] = { + 1.262885f, -0.533345f, -0.161280f, 0.106098f, 0.194799f, 0.003600f, + 0.394783f, -0.053954f, 0.264474f, -0.016651f, 0.376765f, 0.221471f, + 0.489799f, 0.054924f, 0.018292f, 0.037633f, -0.053430f, 1.092426f, + 0.205791f, -0.055661f, -0.227335f, 0.301274f, -0.169917f, 0.100426f, + 0.254388f, 0.103465f, 0.189560f, 0.116479f, 1.647195f, -0.667044f, + 0.067795f, -0.044580f, 0.019428f, 0.072938f, -0.797569f, -0.077539f, + -0.225636f, 0.262883f, -1.048009f, 0.210118f, -0.416156f, -0.143741f, + -0.296985f, 0.205918f, -0.517383f, -0.118527f, -0.396606f, -0.113128f, + -0.279468f, 0.096141f, -0.342051f, -0.337036f, 0.143222f, -0.860280f, + 0.137169f, 0.339767f, -0.336076f, 0.071988f, 0.251557f, -0.004068f, + 0.170734f, 0.237283f, -0.332443f, 0.073643f, 0.375357f, 0.220407f, + 0.150708f, -0.176979f, 0.265786f, -0.105878f, -0.337465f, -0.000491f, + 0.234308f, -0.098973f, 0.129038f, -0.205936f, -0.034793f, -0.106981f, + 0.009974f, 0.037861f, -0.282874f, -0.354414f, 0.023021f, -0.266749f, + -0.041762f, -0.721725f, 0.182262f, -0.273945f, 0.123722f, -0.036749f, + -0.788645f, -0.081560f, -0.472226f, 0.004654f, -0.756766f, -0.132186f, + 1.085412f, -0.221324f, -0.072577f, -0.172834f, -0.104831f, -1.391641f, + -0.345893f, 0.194442f, -0.306583f, -0.041813f, -0.267635f, -0.218568f, + -0.178452f, 0.044421f, -0.128042f, -0.094797f, -0.253724f, 0.273931f, + 0.144843f, -0.401416f, -0.014354f, -0.348929f, 0.123550f, 0.494504f, + -0.007050f, -0.143830f, 0.111292f, 0.211057f, -1.579988f, 0.117744f, + -1.732487f, 0.009320f, -1.162696f, 0.176687f, -0.705609f, 0.524827f, + 0.089822f, 0.082976f, -0.023681f, 0.006120f, -0.907175f, -0.026273f, + 0.019027f, 0.027170f, -0.462563f, -0.535335f, 0.202231f, 0.709803f, + -0.112251f, -1.213869f, 0.225714f, 0.323785f, -0.518254f, -0.014235f, + -0.070790f, -0.369589f, 0.373399f, 0.002738f, 0.175113f, 0.084529f, + -0.101586f, -0.018978f, 0.773392f, -0.673230f, -0.549279f, 0.790196f, + 0.658609f, -0.826831f, -0.514211f, 0.575341f, -0.711311f, 0.276289f, + -0.435715f, 0.392986f, -0.079298f, -0.318719f, 0.188429f, -0.114366f, + 0.172527f, -0.261721f, -0.216761f, 0.163822f, -0.189374f, -0.391901f, + 0.142013f, -0.135046f, 0.144419f, 0.053887f, 0.074673f, -0.290791f, + -0.039560f, -0.103830f, -0.330263f, -0.042091f, 0.050646f, -0.057466f, + -0.069064f, -0.412864f, 0.071097f, 0.126693f, 0.175397f, -0.168485f, + 0.018129f, -0.419188f, -0.272024f, -0.436859f, -0.425711f, -0.024382f, + 0.248042f, -0.169090f, -0.346878f, -0.070926f, 0.292278f, -0.197610f, + -0.218286f, 0.290846f, 0.297843f, 0.247394f, -0.160736f, 0.110314f, + 0.276000f, -0.301676f, -0.232816f, -0.127576f, -0.174457f, -0.124503f, + 0.264880f, -0.332379f, 0.012659f, -0.197333f, 0.604700f, 0.801582f, + 0.758702f, 0.691880f, 0.440917f, 0.773548f, 0.064242f, 1.147508f, + -0.127543f, -0.189628f, -0.122994f, -0.226776f, -0.053531f, -0.187548f, + 0.226554f, -0.273451f, 0.011751f, 0.009133f, 0.185091f, 0.003031f, + 0.000525f, 0.221829f, 0.331550f, -0.202558f, -0.286550f, 0.100683f, + 0.268818f, 0.179971f, -0.050016f, 0.579665f, 0.015911f, 0.033068f, + 0.077768f, -0.017757f, -1.411251f, 0.051519f, -1.745767f, 0.011258f, + -1.947372f, 0.111396f, -1.112755f, -0.008989f, -0.006211f, -0.002098f, + -0.015236f, -0.095697f, -0.095820f, 0.044622f, -0.112096f, 0.060000f, + 0.138957f, -0.462708f, 0.590790f, -0.021405f, -0.283744f, -1.141749f, + 0.213121f, -0.332311f, -0.314090f, -0.789311f, 0.157605f, -0.438019f, + 0.642189f, -0.340764f, -0.996025f, 0.109871f, 0.106128f, -0.010505f, + -0.117233f, -0.223194f, 0.344105f, -0.308754f, 0.386020f, -0.305270f, + -0.538281f, -0.270720f, -0.101688f, 0.207580f, 0.237153f, -0.055730f, + 0.842779f, 0.393543f, 0.007886f, -0.318167f, 0.603768f, 0.388241f, + 0.421536f, 0.632080f, 0.423965f, 0.371472f, 0.456827f, 0.488134f, + 0.358997f, 0.032621f, -0.017104f, 0.032198f, 0.113266f, -0.312277f, + 0.178189f, 0.234180f, 0.134271f, -0.414889f, 0.774141f, -0.225043f, + 0.614052f, -0.279921f, 1.329141f, -0.140827f, 0.797267f, -0.171361f, + 0.066205f, 0.339976f, 0.015223f, 0.193725f, -0.245067f, -0.035578f, + -0.084043f, 0.086756f, 0.029478f, -0.845370f, 0.388613f, -1.215236f, + 0.304573f, -0.439884f, -0.293969f, -0.107988f, -0.267837f, -0.695339f, + -0.702099f, 0.359047f, 0.511730f, 1.429516f, 0.216959f, -0.313828f, + 0.068062f, -0.124917f, -0.648327f, -0.308411f, -0.378467f, -0.429288f, + -0.032415f, -0.357005f, 0.170068f, 0.161167f, -0.250280f, -0.320468f, + -0.408987f, -0.201496f, -0.155996f, 0.021067f, 0.141083f, -0.202733f, + -0.130953f, -0.278148f, -0.042051f, 0.070576f, 0.009982f, -0.044326f, + -0.346851f, -0.255397f, -0.346456f, 0.281781f, 0.001618f, 0.120648f, + 0.297140f, 0.198343f, 0.186104f, 0.183548f, -0.344482f, 0.182258f, + 0.291003f, -0.330228f, -0.048174f, 0.133694f, 0.264582f, 0.229671f, + -0.167251f, -0.316040f, 0.191829f, 0.153417f, -0.345158f, -0.212790f, + -0.878872f, -0.313099f, -0.028368f, 0.065869f, -0.695388f, 1.102812f, + -0.605539f, 0.400680f, -0.350120f, -0.432965f, 0.034553f, -0.693476f, + -0.045708f, 0.492409f, -0.043825f, -0.430522f, 0.071159f, -0.317376f, + -1.164842f, 0.112394f, 0.034137f, -0.611882f, 0.251020f, -0.245113f, + 0.286093f, -0.187883f, 0.340263f, -0.211592f, -0.065706f, -0.332148f, + 0.104026f, -0.003206f, 0.036397f, 0.206499f, 0.161962f, 0.037663f, + -0.313039f, -0.199837f, 0.117952f, -0.182145f, -0.343724f, 0.017625f, + 0.033427f, -0.288075f, -0.101873f, -0.083378f, 0.147870f, 0.049598f, + -0.241824f, 0.070494f, 0.140942f, -0.013795f, 0.020023f, -0.192213f, + -0.320505f, -0.193072f, 0.147260f, 0.311352f, 0.053486f, 0.183716f, + 0.142535f, 0.294333f, -0.054853f, 0.293314f, -0.025398f, 0.190815f, + -0.137574f, -0.191864f, -0.190950f, -0.205988f, -0.199046f, -0.017582f, + -0.149347f, 0.131040f, 0.006854f, -0.350732f, 0.113301f, -0.194371f, + -0.296885f, -0.249199f, -0.193946f, 0.116150f, -0.310411f, -0.325851f, + -0.053275f, -0.063419f, 0.204170f, -0.091940f, -0.146229f, 0.298173f, + 0.053349f, -0.368540f, 0.235629f, -0.317825f, -0.107304f, -0.114618f, + 0.058709f, -0.272070f, 0.076224f, 0.110668f, -0.193282f, -0.135440f, + -0.267950f, -0.102285f, 0.102699f, -0.159082f, 0.262721f, -0.263227f, + 0.094509f, -0.113405f, 0.069888f, -0.169665f, 0.070800f, 0.035432f, + 0.054243f, 0.264229f, 0.117416f, 0.091568f, -0.022069f, -0.069214f, + 0.124543f, 0.070413f, -0.039343f, 0.082823f, -0.838348f, 0.153727f, + -0.000947f, 0.270348f, -1.404952f, -0.159680f, -0.234320f, 0.061023f, + 0.271660f, -0.541834f, 0.570828f, -0.277254f, +}; + +static const float rect_part_nn_bias_16_layer0[32] = { + 0.045740f, 0.292685f, -0.754007f, -0.150412f, -0.006171f, 0.005915f, + 0.000167f, 0.322797f, -0.381793f, 0.349786f, 0.003878f, -0.307203f, + 0.000000f, 0.029122f, 0.000000f, 0.625494f, 0.302105f, -0.362807f, + -0.034002f, -0.573278f, 0.240021f, 0.083965f, 0.000000f, -0.018979f, + -0.147739f, -0.036990f, 0.000000f, 0.000000f, -0.026790f, -0.000036f, + -0.073448f, 0.398328f, +}; + +static const float rect_part_nn_weights_16_layer1[32 * LABELS] = { + 0.095090f, 0.831754f, 0.484433f, 0.472945f, 0.086165f, -0.442388f, + 0.176263f, -0.760247f, 0.419932f, -0.131377f, 0.075814f, 0.089844f, + -0.294718f, 0.299808f, -0.318435f, -0.623205f, -0.346703f, 0.494356f, + 0.949221f, 0.524653f, 0.044095f, 0.428540f, 0.402571f, -0.216920f, + 0.423915f, 1.023334f, -0.366449f, 0.395057f, 0.057576f, 0.094019f, + 0.247685f, -0.007200f, -0.420023f, -0.728965f, -0.063040f, -0.071321f, + 0.209298f, 0.486625f, -0.244375f, 0.263219f, -0.250463f, -0.260301f, + 0.068579f, 0.177644f, -0.155311f, -0.027606f, -0.101614f, 0.553046f, + -0.462729f, -0.237568f, -0.589316f, 0.045182f, 0.551759f, -0.196872f, + 0.183040f, 0.054341f, 0.252784f, -0.536486f, -0.024425f, 0.154942f, + -0.086636f, 0.360416f, 0.214773f, -0.170876f, -0.363522f, -0.464099f, + 0.145494f, -0.099329f, 0.343718f, 0.286427f, 0.085540f, -0.105182f, + 0.155543f, 0.290939f, -0.067069f, 0.228399f, 0.178247f, 0.113031f, + -0.067336f, 0.441062f, 0.132364f, -0.263403f, -0.263925f, -0.083613f, + -0.268577f, -0.204442f, 0.052526f, 0.334787f, -0.064285f, -0.197875f, + 0.296405f, 0.396440f, 0.033231f, 0.229087f, 0.118289f, 0.490894f, + -0.527582f, -0.897206f, -0.325708f, -0.433018f, -0.053989f, 0.223814f, + -0.352319f, 0.772440f, -0.108648f, -0.082859f, -0.342718f, 0.033022f, + -0.309199f, -0.560337f, 0.208476f, 0.520309f, -0.241035f, -0.560391f, + -1.268968f, -0.267567f, 0.129461f, -0.385547f, 0.080142f, 0.065785f, + -0.159324f, -0.580704f, -0.315150f, -0.224900f, -0.110807f, -0.230163f, + 0.307266f, 0.153446f, +}; + +static const float rect_part_nn_bias_16_layer1[LABELS] = { + -0.455437f, + 0.255310f, + 0.452974f, + -0.278733f, +}; + +static const NN_CONFIG rect_part_nnconfig_16 = { + FEATURES, // num_inputs + LABELS, // num_outputs + 1, // num_hidden_layers + { + 32, + }, // num_hidden_nodes + { + rect_part_nn_weights_16_layer0, + rect_part_nn_weights_16_layer1, + }, + { + rect_part_nn_bias_16_layer0, + rect_part_nn_bias_16_layer1, + }, +}; + +static const float rect_part_nn_weights_32_layer0[FEATURES * 32] = { + 0.735110f, -0.238477f, 0.101978f, 0.311671f, -0.123833f, 1.596506f, + -0.341982f, -0.480170f, -0.247587f, 0.613159f, -0.279899f, -0.740856f, + 0.499051f, 0.039041f, 0.056763f, 0.258874f, 0.470812f, -0.121635f, + -0.318852f, -0.098677f, -0.214714f, -0.159974f, -0.305400f, -0.344477f, + -0.260653f, -0.007737f, -0.053016f, -0.158079f, 0.151911f, -0.057685f, + -0.230948f, -0.165940f, -0.127591f, -0.192084f, 1.890390f, -0.315123f, + -0.714531f, -0.015355f, 0.186437f, 0.305504f, 0.035343f, -0.556783f, + 0.239364f, -0.297789f, 0.202735f, -0.707576f, 0.710250f, 0.223346f, + -0.291511f, 0.235778f, 0.455338f, -0.059402f, 0.084530f, -0.115117f, + -0.103696f, -0.192821f, 0.114579f, -0.223487f, 0.306864f, 0.021887f, + -0.028040f, 0.087866f, 0.038870f, -0.081742f, -0.056052f, -0.130837f, + 0.201058f, 0.293391f, 1.880344f, 0.339162f, 0.040928f, -0.503942f, + 0.476333f, 0.259272f, 0.629416f, 0.869369f, 0.622841f, 1.012843f, + 0.715795f, 1.958844f, -1.697462f, 0.071334f, 0.074189f, 0.014585f, + -0.002536f, 0.021900f, 0.151883f, 0.169501f, -0.333018f, -0.247512f, + -0.418575f, -0.473960f, -0.004501f, -0.280939f, -0.162188f, -0.355632f, + 0.136654f, -0.100967f, -0.350435f, -0.135386f, 0.037237f, 0.136982f, + -0.084157f, -0.073248f, 0.021792f, 0.077429f, -0.083042f, -3.169569f, + 0.016261f, -3.351328f, 0.021120f, -3.572247f, 0.023870f, -4.312754f, + 0.040973f, -0.038328f, -0.015052f, 0.017702f, 0.101427f, 0.115458f, + -0.304792f, 0.021826f, -0.157998f, 0.341022f, -0.013465f, 0.105076f, + -0.261465f, 0.318730f, 0.065701f, 0.314879f, -0.064785f, 0.282824f, + 0.100542f, 0.057260f, -0.003756f, -0.026214f, -0.264641f, 0.275545f, + -0.049201f, -0.283015f, -0.057363f, 0.183570f, 0.243161f, -0.255764f, + 0.099747f, -0.156157f, -0.262494f, 0.231521f, -0.262617f, -0.186096f, + 0.171720f, 0.018983f, -0.145545f, 0.197662f, -0.001502f, -0.267526f, + 0.001960f, 0.003260f, 0.045237f, -0.377174f, -0.042499f, -0.015278f, + -0.196779f, -0.262797f, -0.318427f, -0.126092f, -0.339723f, 0.205288f, + -0.544284f, -0.507896f, -0.316622f, -0.090312f, -0.250917f, -0.337263f, + -0.220199f, -0.296591f, -0.116816f, 0.052381f, 0.145681f, 0.016521f, + -0.093549f, -0.097822f, 0.023140f, -0.010346f, 0.036181f, 0.145826f, + -0.139123f, -0.462638f, -0.007315f, 0.156533f, -0.102787f, 0.143586f, + -0.092094f, -0.144220f, -0.168994f, -0.045833f, 0.021628f, -0.421794f, + -0.055857f, 0.217931f, -0.061937f, -0.028768f, -0.078250f, -0.426939f, + -0.223118f, -0.230080f, -0.194988f, -0.197673f, -0.020918f, 0.139945f, + 0.186951f, -0.071317f, -0.084007f, -0.138597f, 0.101950f, 0.093870f, + 0.153226f, 0.017799f, -0.088539f, -0.037796f, 0.340412f, 0.183305f, + 0.391880f, -1.127417f, 0.132762f, -0.228565f, 0.399035f, 0.017483f, + -0.041619f, 0.017849f, 0.092340f, 0.054204f, 0.681185f, 0.421034f, + 0.112520f, -0.040618f, -0.040148f, -0.360647f, 0.053555f, 0.192854f, + 0.076968f, -0.179224f, -0.081617f, -0.287661f, -0.191072f, -0.310227f, + -0.332226f, -0.039786f, -0.247795f, -0.232201f, -0.333533f, -0.077995f, + -0.471732f, 0.051829f, 0.090488f, 0.142465f, -0.120490f, -0.286151f, + -0.049117f, -0.251082f, 0.211884f, -0.223366f, 0.063565f, 0.229938f, + -0.059348f, -0.029573f, -0.064303f, -0.156148f, 0.086958f, -0.297613f, + -0.125107f, 0.062718f, 0.339137f, -0.218896f, -0.057290f, -0.236670f, + -0.143783f, -0.119429f, 0.242320f, -0.323464f, -0.178377f, 0.238275f, + -0.025042f, 0.074798f, 0.111329f, -0.299773f, -0.151748f, -0.261607f, + 0.215626f, 0.202243f, -0.121896f, -0.024283f, -0.293854f, -0.018232f, + -0.012629f, -0.199297f, -0.060595f, 0.432339f, -0.158735f, -0.028380f, + 0.326639f, 0.222546f, -0.218135f, -0.495955f, -0.015055f, -0.104206f, + -0.268823f, 0.116765f, 0.041769f, -0.187095f, 0.225090f, 0.198195f, + 0.001502f, -0.219212f, -0.244779f, -0.017690f, -0.033197f, -0.339813f, + -0.325453f, 0.002499f, -0.066113f, 0.043235f, 0.324275f, -0.630642f, + -1.440551f, 0.174527f, 0.124619f, -1.187345f, 1.372693f, -0.278393f, + -0.058673f, -0.286338f, 1.708757f, -0.325094f, -0.543172f, -0.229411f, + 0.169927f, 0.175064f, 0.198321f, 0.117351f, 0.220882f, 0.138078f, + -0.158000f, -0.286708f, 0.096046f, -0.321788f, 0.206949f, -0.014473f, + -0.321234f, 0.100033f, -0.108266f, 0.166824f, 0.032904f, -0.065760f, + -0.303896f, 0.180342f, -0.301145f, -0.352554f, 0.149089f, 0.013277f, + 0.256019f, -0.109770f, 1.832588f, -0.132568f, 1.527658f, -0.164252f, + -0.857880f, -0.242694f, -0.553797f, 0.334023f, -0.332759f, -0.166203f, + -0.223175f, 0.007953f, -0.175865f, -0.134590f, -0.023858f, -0.011983f, + 0.054403f, -0.147054f, -0.176901f, -0.166893f, -0.292662f, -0.010569f, + -0.041744f, -0.060398f, -0.237584f, 0.154246f, -0.083270f, -0.314016f, + -0.374736f, 0.100063f, 0.048401f, -0.061952f, -0.178816f, 0.157243f, + 0.221991f, -0.065035f, 0.098517f, -0.190704f, -0.210613f, -0.274884f, + -0.341442f, -0.205281f, 0.073644f, 0.130667f, 0.149194f, -0.018172f, + 1.796154f, -1.017806f, -0.169655f, 0.104239f, 0.344313f, 0.643042f, + 0.730177f, 0.270776f, 0.581631f, -1.090649f, 0.707472f, 1.411035f, + 0.268739f, 0.178860f, -0.062251f, -0.118611f, -0.215759f, 0.023485f, + -0.105320f, 0.036396f, -0.059604f, 0.090024f, 0.095224f, -0.053497f, + -0.084040f, 0.055836f, 0.111678f, 0.014886f, -0.178380f, 0.079662f, + -0.123580f, 0.057379f, -0.409844f, -0.305386f, -0.987808f, -0.291094f, + 0.063966f, 0.263709f, -0.337221f, 0.720093f, 0.105030f, 0.848950f, + 0.071835f, 0.228972f, 0.057705f, -2.154561f, -0.201303f, -0.058856f, + -0.020081f, 0.029375f, 0.234837f, -0.001063f, 0.042527f, 0.014567f, + -0.299420f, -0.289117f, 0.275219f, 0.263596f, -0.186026f, -0.111364f, + -0.118393f, -0.318778f, 0.010710f, -0.286836f, -0.070330f, -0.049497f, + 0.093162f, -0.298085f, 0.204761f, -0.206633f, -0.009057f, -0.235372f, + 0.185300f, -0.271814f, 0.281732f, 0.268149f, -0.018967f, 0.162748f, + -0.086694f, -0.063839f, -0.097473f, -0.280120f, 0.324688f, 0.157911f, + -0.064794f, -0.266017f, -0.305608f, -0.196854f, -0.185767f, 0.199455f, + 0.102264f, 0.070866f, 0.172045f, 0.266433f, -0.176167f, 0.251657f, + -0.239220f, 0.229667f, 0.156115f, -0.221345f, 0.270720f, 0.109367f, + 0.230352f, -0.384561f, -0.026329f, 0.005928f, -0.087685f, -0.097995f, + -0.153864f, 0.117211f, -0.226492f, -0.379832f, -0.201714f, 0.049707f, + -0.292120f, 0.114074f, -0.085307f, -0.485356f, -0.347405f, 0.089361f, + -0.419273f, -0.320764f, -0.107254f, -0.274615f, -0.292991f, 0.095602f, + -0.078789f, 0.138927f, 0.270813f, 0.205814f, 0.065003f, 0.169171f, + 0.056142f, -0.005792f, 0.059483f, 0.060149f, +}; + +static const float rect_part_nn_bias_32_layer0[32] = { + -1.749808f, 0.000000f, 0.239736f, -0.000424f, 0.431792f, -0.150833f, + 2.866760f, 0.000000f, 0.000000f, -0.281434f, 0.000000f, -0.150086f, + 0.000000f, -0.008346f, -0.204104f, -0.006581f, 0.000000f, -0.197006f, + 0.000000f, -0.735287f, -0.028345f, -1.180116f, -0.106524f, 0.000000f, + 0.075879f, -0.150966f, -2.438914f, 0.000000f, -0.011775f, -0.024204f, + -0.138235f, -0.123763f, +}; + +static const float rect_part_nn_weights_32_layer1[32 * LABELS] = { + 0.622235f, 0.264894f, -0.424216f, 0.103989f, 1.401192f, -0.063838f, + -5.216846f, 0.329234f, -0.293113f, 0.457519f, -0.271899f, 0.043771f, + -0.203823f, 0.573535f, -0.192703f, 0.054939f, 0.163019f, 0.124803f, + 0.160664f, 0.385406f, -0.091403f, 0.320204f, 0.101181f, -0.157792f, + -0.095555f, -0.255011f, 1.326614f, -0.138076f, -0.082434f, -0.342442f, + 0.184067f, -0.076395f, 0.050263f, 0.251065f, 0.291743f, 0.197838f, + -0.950922f, 0.280202f, 2.904905f, -0.219434f, 0.284386f, 0.375005f, + 0.193817f, -0.298663f, -0.255364f, -0.297545f, 0.030518f, -0.023892f, + -0.396120f, -0.253027f, 0.237235f, -0.550249f, -0.076817f, -0.201374f, + 0.292708f, 0.341936f, -0.532215f, 0.180634f, -0.943291f, -0.217179f, + 0.251611f, -0.306310f, 0.229054f, -0.350337f, -0.192707f, 0.146781f, + 0.409007f, 0.279088f, -0.307357f, 0.199059f, 2.780962f, 0.163723f, + -0.226445f, 0.242830f, 0.220356f, -0.057621f, 0.196677f, -0.179975f, + -0.314636f, 0.218271f, -0.278653f, -0.226286f, 0.034275f, -0.320149f, + 0.154779f, 0.074937f, -0.015650f, -0.281735f, -0.495227f, -0.075036f, + -0.871024f, -0.350643f, 0.343468f, 0.095665f, 0.447121f, -0.059040f, + 0.244757f, 0.223122f, 0.272544f, 0.129678f, -1.700183f, 0.254869f, + 2.528983f, 0.217362f, 0.327765f, -0.129369f, -0.003560f, -0.532537f, + 0.080216f, -0.739488f, -0.299813f, 0.185421f, 0.265994f, 0.152268f, + -0.401829f, -0.901380f, 0.347747f, -0.524845f, -0.201163f, 0.063585f, + -0.517479f, -0.077816f, -0.735739f, -0.161411f, -0.113607f, -0.306188f, + 0.190817f, -0.362567f, +}; + +static const float rect_part_nn_bias_32_layer1[LABELS] = { + -0.833530f, + 0.860502f, + 0.708645f, + -1.083700f, +}; + +static const NN_CONFIG rect_part_nnconfig_32 = { + FEATURES, // num_inputs + LABELS, // num_outputs + 1, // num_hidden_layers + { + 32, + }, // num_hidden_nodes + { + rect_part_nn_weights_32_layer0, + rect_part_nn_weights_32_layer1, + }, + { + rect_part_nn_bias_32_layer0, + rect_part_nn_bias_32_layer1, + }, +}; + +static const float rect_part_nn_weights_64_layer0[FEATURES * 32] = { + 0.029424f, -0.295893f, -0.313259f, -0.090484f, -0.104946f, 0.121361f, + 0.137971f, -0.137984f, -0.328158f, -0.137280f, -0.276995f, -0.153118f, + 0.187893f, 0.105787f, -0.236591f, -0.114325f, -0.000708f, 1.936191f, + 0.048491f, -0.026048f, -0.206916f, 0.830237f, -0.152354f, 0.074191f, + -0.153813f, 0.148942f, -0.103457f, 0.028252f, 1.758264f, -2.123016f, + 0.120182f, 0.049954f, 0.110450f, -0.199360f, 0.642198f, 0.040225f, + -0.140886f, 0.091833f, -0.122788f, 1.172115f, -0.833333f, -0.505218f, + 0.736050f, -0.109958f, -0.839030f, -0.399916f, 1.029718f, 0.408977f, + -0.836882f, 0.389683f, -1.134413f, -1.529672f, -0.146351f, 0.089298f, + 0.083772f, -0.697869f, 1.683311f, -0.882446f, 0.494428f, -0.122128f, + 0.659819f, -0.057178f, -0.915390f, -0.192412f, 0.046613f, 0.010697f, + 0.040782f, 0.110807f, -0.225332f, -0.327730f, -0.114825f, 0.063511f, + 0.050503f, 0.023602f, 0.006524f, -0.274547f, -0.607145f, -0.143812f, + -0.327689f, -0.333072f, -0.017138f, -0.183992f, -0.200622f, -0.262463f, + -0.132799f, -0.018155f, -0.534214f, -0.385994f, 0.116278f, -0.752879f, + -0.090734f, -0.249152f, 0.071716f, 0.029603f, -0.382456f, -0.122894f, + 1.349552f, -0.885192f, 0.257903f, -0.265945f, -0.045579f, 0.112247f, + -0.122810f, -0.258285f, -0.145427f, -0.127442f, 0.072778f, 0.072549f, + 0.182149f, 0.239403f, 0.167205f, -0.291616f, -0.281237f, 0.335735f, + 0.208511f, -0.239628f, -0.022236f, -0.177370f, 0.207808f, 0.023535f, + 0.137455f, 0.016406f, -0.138685f, 0.188732f, 0.205513f, 0.209787f, + 0.060592f, 0.239954f, -0.128341f, -0.291585f, 0.022141f, -0.311201f, + -0.010199f, -0.314224f, -0.351915f, -0.079775f, -0.260028f, -0.015953f, + 0.007404f, 0.051589f, 0.019771f, -2.337926f, 0.024596f, -2.512399f, + -0.023138f, -2.421380f, 0.016515f, -3.269775f, 0.026844f, -0.053660f, + -0.013213f, -0.029248f, 0.114357f, 0.259100f, -0.141749f, -0.106802f, + -0.117323f, -0.294698f, -0.316012f, -0.328013f, 0.016459f, 0.136175f, + 0.223327f, 0.322312f, -0.297297f, 0.118286f, -0.317197f, -0.116692f, + 0.262236f, -0.032443f, -0.392128f, -0.199989f, -0.383621f, 0.008347f, + -0.079302f, -0.005529f, 0.049261f, 0.145948f, -0.263592f, -0.317109f, + 0.260015f, -0.499341f, -0.171764f, -0.017815f, 0.149186f, 0.178294f, + -0.492198f, 0.016956f, 0.008067f, -0.057734f, -0.189979f, -0.131489f, + -0.163303f, 0.121378f, -0.172272f, 0.125891f, 0.120654f, 0.071314f, + 0.117423f, -0.242167f, 0.047170f, 0.234302f, -0.355370f, -0.336112f, + -0.255471f, -0.267792f, -0.135367f, -0.284411f, 0.254592f, 0.098749f, + 0.224989f, 0.258450f, -0.306878f, 0.153551f, -0.175806f, -0.244459f, + -0.274922f, 0.254346f, 0.110309f, 0.036054f, 0.095133f, -0.589646f, + 0.080543f, 0.154155f, 0.133797f, -0.401518f, 0.798127f, 0.066742f, + 1.449216f, 0.282498f, 1.210638f, -0.280643f, 0.572386f, -0.308133f, + -0.053143f, 0.008437f, 0.269565f, 0.347616f, 0.087180f, -0.771104f, + 0.200800f, 0.157578f, 0.474128f, -0.971488f, 0.193451f, 0.340339f, + -0.123425f, 0.560754f, -0.139621f, -0.281721f, -0.100162f, 0.250926f, + 0.281100f, 0.197680f, 0.138629f, 1.045823f, 0.339047f, 0.036698f, + -0.159210f, 0.727869f, -1.371850f, 0.116241f, -2.180194f, 0.214055f, + -0.213691f, 0.447957f, -1.129966f, 0.543598f, 0.147599f, 0.060034f, + -0.049415f, -0.095858f, 0.290599f, 0.059512f, 0.198343f, -0.211903f, + 0.158736f, -0.090220f, -0.221992f, 0.198320f, 0.028632f, -0.408238f, + -0.368266f, -0.218740f, -0.379023f, -0.173573f, -0.035179f, 0.240176f, + 0.237714f, -0.417132f, -0.184989f, 0.046818f, -0.016965f, -0.524012f, + -0.094848f, -0.225678f, 0.021766f, -0.028366f, 0.072343f, -0.039980f, + 0.023334f, -0.392397f, 0.164450f, -0.201650f, -0.519754f, -0.023352f, + -4.559466f, -0.115996f, 0.135844f, 0.152599f, -0.111570f, 1.870310f, + 0.003522f, 1.893098f, -0.134055f, 1.850787f, 0.085160f, -2.203354f, + 0.380799f, -0.074047f, 0.023760f, 0.077310f, 0.273381f, -1.163135f, + -0.024976f, 0.093252f, 0.011445f, -0.129009f, -2.200677f, -0.013703f, + -1.964109f, -0.027246f, -2.135679f, 0.049465f, -3.879032f, 0.195114f, + -0.018085f, 0.016755f, 0.036330f, 0.169138f, 0.003548f, -0.028565f, + -0.178196f, -0.020577f, -0.104330f, -0.270961f, -0.282822f, -0.228735f, + -0.292561f, 0.271648f, 0.129171f, 0.376168f, -0.265005f, -0.093002f, + -0.185514f, 0.025598f, 0.055265f, -0.212784f, -0.249005f, 0.051507f, + -0.267868f, 0.162227f, -0.237365f, 0.267479f, -0.051543f, -0.288800f, + -0.246119f, 0.216296f, 0.226888f, -0.123005f, 0.068040f, -0.096630f, + -0.100500f, 0.161640f, -0.349187f, -0.061229f, 0.042915f, 0.024949f, + -0.083086f, -0.407249f, -0.428306f, -0.381137f, -0.508822f, 0.354796f, + -0.612346f, -0.230076f, -0.734103f, -0.550571f, -0.318788f, -0.300091f, + -0.336045f, -0.494406f, -0.206900f, 0.079942f, 0.149065f, -0.533360f, + 0.940431f, -0.078860f, 1.418633f, -0.117527f, 1.349170f, 0.242658f, + 0.559328f, 0.258770f, -0.014508f, -0.204775f, -0.292631f, 0.498345f, + -0.274918f, 0.051670f, 0.157748f, -0.179721f, -0.183330f, -0.393550f, + -0.208848f, 0.060742f, -0.159654f, 0.047757f, -0.400256f, -0.084606f, + -0.080619f, -0.359664f, -0.078305f, -0.455653f, 0.227624f, -0.385606f, + -0.060326f, -0.209831f, -0.077008f, 0.148862f, 0.209908f, 0.047655f, + -0.342292f, -0.088375f, -0.115465f, 0.082700f, 0.036465f, -0.001792f, + -0.285730f, 0.114632f, 0.239254f, -0.348543f, 0.044916f, -0.299003f, + -0.244756f, -0.180802f, 0.314253f, -0.127788f, -0.221512f, 0.034787f, + -0.208388f, 0.349156f, 0.265975f, -0.068335f, 0.261372f, 0.146705f, + -0.098729f, 0.293699f, -0.111342f, 0.207402f, -0.038772f, 0.124135f, + -0.237450f, -0.191511f, -0.052240f, -0.237151f, 0.005013f, 0.139441f, + -0.153634f, -0.021596f, -0.036220f, -0.077873f, -0.085995f, -0.254555f, + -0.204382f, -0.082362f, 0.941796f, 0.253800f, -0.957468f, 0.095795f, + 0.122046f, -0.310364f, 0.087301f, 0.012704f, 0.193265f, -0.058303f, + 0.250452f, 0.835269f, 0.507383f, 0.109957f, -0.145028f, -0.114419f, + -0.225618f, 0.132387f, -0.063335f, -0.325776f, -0.346173f, -0.006653f, + -0.133534f, -0.085549f, -0.050177f, 0.173103f, 0.025421f, 0.105512f, + 0.258036f, 0.153116f, 0.290202f, -0.333699f, -0.072405f, -0.124069f, + -0.241933f, -0.313318f, 0.013623f, -0.237440f, -0.232228f, -0.170850f, + -0.039212f, 0.162468f, -0.330162f, -0.218462f, -0.287064f, -0.181673f, + -0.161059f, 0.024664f, -0.108642f, -0.231707f, 0.217994f, -1.128878f, + 0.093010f, 0.101513f, 0.055895f, -0.354538f, 0.844174f, 0.254335f, + 1.920298f, -0.230777f, 0.798144f, 0.206425f, 0.580655f, -0.177645f, + -0.412061f, 0.112629f, -0.476438f, 0.209436f, +}; + +static const float rect_part_nn_bias_64_layer0[32] = { + 0.000000f, 0.345406f, -0.499542f, -1.718246f, -0.147443f, -0.408843f, + -0.008997f, -0.107946f, 2.117510f, 0.000000f, -0.141830f, -0.049079f, + 0.000000f, -1.331136f, -1.417843f, -0.485054f, -0.100856f, -0.230750f, + -2.574372f, 2.310627f, -0.030363f, 0.000000f, -0.310119f, -1.314316f, + -0.108766f, -0.107918f, 0.000000f, 0.000000f, 0.093643f, 0.000000f, + 0.000000f, -0.902343f, +}; + +static const float rect_part_nn_weights_64_layer1[32 * LABELS] = { + 0.404567f, 1.168492f, 0.051714f, 0.827941f, 0.135334f, 0.456922f, + -0.370524f, 0.062865f, -3.076300f, -0.290613f, 0.280029f, -0.101778f, + 0.250216f, 0.347721f, 0.466400f, 0.030845f, 0.114570f, 0.089456f, + 1.519938f, -3.493788f, 0.264212f, -0.109125f, 0.306644f, 0.368206f, + -0.052168f, -0.229630f, -0.339932f, -0.080472f, 0.319845f, 0.143818f, + -0.172595f, 0.372777f, -0.082072f, -0.505781f, -0.288321f, -0.473028f, + -0.027567f, -0.034329f, -0.291965f, -0.063262f, 1.721741f, 0.118914f, + 0.183681f, 0.041611f, 0.266371f, 0.005896f, -0.484705f, 0.665535f, + -0.240945f, -0.017963f, -1.409440f, 2.031976f, 0.240327f, -0.116604f, + 0.273245f, -0.170570f, -0.085491f, -0.340315f, -0.209651f, -0.217460f, + -0.249373f, 0.009193f, 0.009467f, -0.272909f, 0.308472f, -0.551173f, + 0.168374f, -0.583229f, 0.140082f, -0.585715f, -0.010929f, 0.159779f, + 1.438104f, 0.293111f, -0.053339f, -0.101828f, -0.280573f, -0.211265f, + -0.323605f, -0.540908f, 0.101366f, -0.005288f, -1.517046f, 2.078767f, + 0.215597f, 0.144012f, 0.315888f, -0.251324f, 0.150482f, -0.137871f, + 0.235116f, -0.194202f, -0.153475f, -0.312384f, -0.375510f, 0.336488f, + -0.379837f, -1.004979f, -0.312587f, -0.406174f, 0.154290f, -0.539766f, + -0.230074f, 0.303564f, 0.719439f, -0.235108f, -0.204978f, 0.399229f, + 0.290222f, -0.278713f, -0.667069f, -0.420550f, 0.164893f, -0.459689f, + -1.035368f, 0.818909f, 0.275137f, -0.291006f, -0.061505f, 0.052737f, + -0.084871f, -0.348335f, 0.312544f, 0.120753f, -0.707222f, -0.010050f, + -0.137148f, -0.351765f, +}; + +static const float rect_part_nn_bias_64_layer1[LABELS] = { + -0.926768f, + 0.765832f, + 0.663683f, + -0.621865f, +}; + +static const NN_CONFIG rect_part_nnconfig_64 = { + FEATURES, // num_inputs + LABELS, // num_outputs + 1, // num_hidden_layers + { + 32, + }, // num_hidden_nodes + { + rect_part_nn_weights_64_layer0, + rect_part_nn_weights_64_layer1, + }, + { + rect_part_nn_bias_64_layer0, + rect_part_nn_bias_64_layer1, + }, +}; + +static void ml_prune_rect_partition(VP9_COMP *const cpi, MACROBLOCK *const x, + BLOCK_SIZE bsize, + const PC_TREE *const pc_tree, + int *allow_horz, int *allow_vert, + int64_t ref_rd, int mi_row, int mi_col) { + const NN_CONFIG *nn_config = NULL; + float score[LABELS] = { + 0.0f, + }; + int thresh = -1; + int i; + + if (ref_rd <= 0 || ref_rd > 1000000000) return; + + switch (bsize) { + case BLOCK_8X8: break; + case BLOCK_16X16: + nn_config = &rect_part_nnconfig_16; + thresh = cpi->sf.ml_prune_rect_partition_threhold[1]; + break; + case BLOCK_32X32: + nn_config = &rect_part_nnconfig_32; + thresh = cpi->sf.ml_prune_rect_partition_threhold[2]; + break; + case BLOCK_64X64: + nn_config = &rect_part_nnconfig_64; + thresh = cpi->sf.ml_prune_rect_partition_threhold[3]; + break; + default: assert(0 && "Unexpected block size."); return; + } + if (!nn_config || thresh < 0) return; + + // Feature extraction and model score calculation. + { + const int64_t none_rdcost = pc_tree->none.rdcost; + const VP9_COMMON *const cm = &cpi->common; + const int dc_q = vp9_dc_quant(cm->base_qindex, 0, cm->bit_depth); + int feature_index = 0; + unsigned int block_var = 0; + unsigned int sub_block_var[4] = { 0 }; + float features[FEATURES]; + + features[feature_index++] = + (float)(pc_tree->partitioning == PARTITION_NONE); + features[feature_index++] = logf((float)(dc_q * dc_q) / 256.0f + 1.0f); + + // Calculate source pixel variance. + { + struct buf_2d buf; + const BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_SPLIT); + const int bs = 4 * num_4x4_blocks_wide_lookup[bsize]; + const MACROBLOCKD *const xd = &x->e_mbd; + vp9_setup_src_planes(x, cpi->Source, mi_row, mi_col); + + (void)xd; +#if CONFIG_VP9_HIGHBITDEPTH + if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) { + block_var = vp9_high_get_sby_perpixel_variance(cpi, &x->plane[0].src, + bsize, xd->bd); + } else { + block_var = vp9_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize); + } +#else + block_var = vp9_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize); +#endif // CONFIG_VP9_HIGHBITDEPTH + + buf.stride = x->plane[0].src.stride; + for (i = 0; i < 4; ++i) { + const int x_idx = (i & 1) * bs / 2; + const int y_idx = (i >> 1) * bs / 2; + buf.buf = x->plane[0].src.buf + x_idx + y_idx * buf.stride; +#if CONFIG_VP9_HIGHBITDEPTH + if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) { + sub_block_var[i] = + vp9_high_get_sby_perpixel_variance(cpi, &buf, subsize, xd->bd); + } else { + sub_block_var[i] = vp9_get_sby_perpixel_variance(cpi, &buf, subsize); + } +#else + sub_block_var[i] = vp9_get_sby_perpixel_variance(cpi, &buf, subsize); +#endif // CONFIG_VP9_HIGHBITDEPTH + } + } + + features[feature_index++] = logf((float)block_var + 1.0f); + features[feature_index++] = logf((float)ref_rd + 1.0f); + features[feature_index++] = (none_rdcost > 0 && none_rdcost < 1000000000) + ? (float)pc_tree->none.skippable + : 0.0f; + + for (i = 0; i < 4; ++i) { + const int64_t this_rd = pc_tree->split[i]->none.rdcost; + const int rd_valid = this_rd > 0 && this_rd < 1000000000; + // Ratio between sub-block RD and whole block RD. + features[feature_index++] = + rd_valid ? ((float)this_rd / (float)ref_rd) : 1.0f; + // Sub-block skippable. + features[feature_index++] = + rd_valid ? ((float)pc_tree->split[i]->none.skippable) : 0.0f; + } + + { + const float denom = (float)(block_var + 1); + const float low_b = 0.1f; + const float high_b = 10.0f; + for (i = 0; i < 4; ++i) { + // Ratio between the quarter sub-block variance and the + // whole-block variance. + float var_ratio = (float)(sub_block_var[i] + 1) / denom; + if (var_ratio < low_b) var_ratio = low_b; + if (var_ratio > high_b) var_ratio = high_b; + features[feature_index++] = var_ratio; + } + } + assert(feature_index == FEATURES); + nn_predict(features, nn_config, score); + } + + // Make decisions based on the model score. + { + int max_score = -1000; + int horz = 0, vert = 0; + int int_score[LABELS]; + for (i = 0; i < LABELS; ++i) { + int_score[i] = (int)(100 * score[i]); + max_score = VPXMAX(int_score[i], max_score); + } + thresh = max_score - thresh; + for (i = 0; i < LABELS; ++i) { + if (int_score[i] >= thresh) { + if ((i >> 0) & 1) horz = 1; + if ((i >> 1) & 1) vert = 1; + } + } + *allow_horz = *allow_horz && horz; + *allow_vert = *allow_vert && vert; + } +} +#undef FEATURES + int get_rdmult_delta(VP9_COMP *cpi, BLOCK_SIZE bsize, int mi_row, int mi_col, int orig_rdmult) { TplDepFrame *tpl_frame = &cpi->tpl_stats[cpi->twopass.gf_group.index]; @@ -3687,7 +4285,7 @@ static void rd_pick_partition(VP9_COMP *cpi, ThreadData *td, ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE]; PARTITION_CONTEXT sl[8], sa[8]; TOKENEXTRA *tp_orig = *tp; - PICK_MODE_CONTEXT *ctx = &pc_tree->none; + PICK_MODE_CONTEXT *const ctx = &pc_tree->none; int i; const int pl = partition_plane_context(xd, mi_row, mi_col, bsize); BLOCK_SIZE subsize; @@ -3854,6 +4452,7 @@ static void rd_pick_partition(VP9_COMP *cpi, ThreadData *td, if (partition_none_allowed) { rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &this_rdc, bsize, ctx, best_rdc.rdcost); + ctx->rdcost = this_rdc.rdcost; if (this_rdc.rate != INT_MAX) { if (cpi->sf.prune_ref_frame_for_rect_partitions) { const int ref1 = ctx->mic.ref_frame[0]; @@ -3979,6 +4578,10 @@ static void rd_pick_partition(VP9_COMP *cpi, ThreadData *td, // PARTITION_SPLIT // TODO(jingning): use the motion vectors given by the above search as // the starting point of motion search in the following partition type check. + pc_tree->split[0]->none.rdcost = 0; + pc_tree->split[1]->none.rdcost = 0; + pc_tree->split[2]->none.rdcost = 0; + pc_tree->split[3]->none.rdcost = 0; if (do_split || must_split) { subsize = get_subsize(bsize, PARTITION_SPLIT); load_pred_mv(x, ctx); @@ -4087,6 +4690,21 @@ static void rd_pick_partition(VP9_COMP *cpi, ThreadData *td, if (used_frames) pc_tree->vertical[1].skip_ref_frame_mask = ~used_frames; } + { + int do_ml_rect_partition_pruning = + !frame_is_intra_only(cm) && !force_horz_split && !force_vert_split && + (partition_horz_allowed || partition_vert_allowed) && bsize > BLOCK_8X8; +#if CONFIG_VP9_HIGHBITDEPTH + if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) + do_ml_rect_partition_pruning = 0; +#endif + if (do_ml_rect_partition_pruning) { + ml_prune_rect_partition(cpi, x, bsize, pc_tree, &partition_horz_allowed, + &partition_vert_allowed, best_rdc.rdcost, mi_row, + mi_col); + } + } + // PARTITION_HORZ if (partition_horz_allowed && (do_rect || vp9_active_h_edge(cpi, mi_row, mi_step))) { @@ -4300,6 +4918,7 @@ static void encode_rd_sb_row(VP9_COMP *cpi, ThreadData *td, rd_auto_partition_range(cpi, tile_info, xd, mi_row, mi_col, &x->min_partition_size, &x->max_partition_size); } + td->pc_root->none.rdcost = 0; rd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col, BLOCK_64X64, &dummy_rdc, INT64_MAX, td->pc_root); } diff --git a/vp9/encoder/vp9_speed_features.c b/vp9/encoder/vp9_speed_features.c index dca47924e..118d0bd34 100644 --- a/vp9/encoder/vp9_speed_features.c +++ b/vp9/encoder/vp9_speed_features.c @@ -196,9 +196,14 @@ static void set_good_speed_feature_framesize_independent(VP9_COMP *cpi, sf->adaptive_rd_thresh_row_mt = 0; sf->allow_skip_recode = 1; sf->less_rectangular_check = 1; - sf->use_square_partition_only = !frame_is_boosted(cpi); + sf->use_square_partition_only = !boosted; sf->prune_ref_frame_for_rect_partitions = 1; + sf->ml_prune_rect_partition_threhold[0] = -1; + sf->ml_prune_rect_partition_threhold[1] = 350; + sf->ml_prune_rect_partition_threhold[2] = 325; + sf->ml_prune_rect_partition_threhold[3] = 250; + if (cpi->twopass.fr_content_type == FC_GRAPHICS_ANIMATION) { sf->exhaustive_searches_thresh = (1 << 22); } else { @@ -216,11 +221,17 @@ static void set_good_speed_feature_framesize_independent(VP9_COMP *cpi, if (speed >= 1) { sf->enable_tpl_model = 0; sf->prune_ref_frame_for_rect_partitions = 0; + + sf->ml_prune_rect_partition_threhold[0] = -1; + sf->ml_prune_rect_partition_threhold[1] = -1; + sf->ml_prune_rect_partition_threhold[2] = -1; + sf->ml_prune_rect_partition_threhold[3] = -1; + if (oxcf->pass == 2) { TWO_PASS *const twopass = &cpi->twopass; if ((twopass->fr_content_type == FC_GRAPHICS_ANIMATION) || vp9_internal_image_edge(cpi)) { - sf->use_square_partition_only = !frame_is_boosted(cpi); + sf->use_square_partition_only = !boosted; } else { sf->use_square_partition_only = !frame_is_intra_only(cm); } @@ -887,6 +898,10 @@ void vp9_set_speed_features_framesize_independent(VP9_COMP *cpi) { sf->limit_newmv_early_exit = 0; sf->bias_golden = 0; sf->base_mv_aggressive = 0; + sf->ml_prune_rect_partition_threhold[0] = -1; + sf->ml_prune_rect_partition_threhold[1] = -1; + sf->ml_prune_rect_partition_threhold[2] = -1; + sf->ml_prune_rect_partition_threhold[3] = -1; // Some speed-up features even for best quality as minimal impact on quality. sf->adaptive_rd_thresh = 1; diff --git a/vp9/encoder/vp9_speed_features.h b/vp9/encoder/vp9_speed_features.h index 331ca49d7..bc468c0f4 100644 --- a/vp9/encoder/vp9_speed_features.h +++ b/vp9/encoder/vp9_speed_features.h @@ -338,6 +338,9 @@ typedef struct SPEED_FEATURES { // Prune reference frames for rectangular partitions. int prune_ref_frame_for_rect_partitions; + // Threshold values used for ML based rectangular partition search pruning. + int ml_prune_rect_partition_threhold[4]; + // Sets min and max partition sizes for this 64x64 region based on the // same 64x64 in last encoded frame, and the left and above neighbor. AUTO_MIN_MAX_MODE auto_min_max_partition_size;