CING APTI BUNDLE WITH M EMORY ORY U SA GE P REDICT ION : WITH - PowerPoint PPT Presentation

A DAP TIVE R AY AY - BU LE T RACIN CING APTI BUNDLE WITH M EMORY ORY U SA GE P REDICT ION : WITH SAGE DICTIO E FFI FFICIENT G LOBAL AL I LLUMIN ION IN IN L ARGE GE S CENES MINATIO Yusuke ke To Toku kuyoshi (Squar are Enix Co., Ltd.) .) Ta Taka kashi Se Seki kine (Squar are Enix Co., Ltd.) .) Tiag ago d da Silva (Square Enix Co. o., Ltd., Univ. of of Tok okyo) Tak akas ashi Kan anai (Univ. of of Tok okyo)

L IGHT IGHT M AP APS FO FOR L ARG ARGE S CENES ES 45.3 M texel light maps Scene with 4.9 km in diameter (3.7 M triangles) Computation time: 1396 secs (2000 sample directions) (GPU: NVIDIA GeForce GTX 580 1.5GB memory)

I NTRODU ION DUCTIO 1. Intr Introduc uction 2. Ada daptive T Til iling for R or Ray ay-bu bundl dles 3. Expe perimental Results & Future Work ork

R AY AY - BUNDL DLE T RACI RACING  Set of parallel rays for a sample direction [Sbert96]  Implemented with GPU rasterization [Szirmay-Kalos98, Hachisuka05]  Benefits: HW acceleration, tessellation etc.  Multi-fragment problem is identical to OIT  Per Per-pixel li l linke ked-lis ist [Yang10]

L IM ITED M EMORY RY C APACI ACITY OF OF GPU GPU S IMIT Ray ay-bu bundl dle trac racing g is weak ak in lar arge ge scenes Light lea eaking er error Me Memory ov overflow w of of the li lists  Uniformly distributed rays  Memory usage is unknown before rendering  Inhomogeneous light map density  Excessive memory has to be allocated  High-resolution ray-bundle buffer is required

U NIFO FORM T IL ING [T HIBIEROZ 11] ILIN  Proposed for real-time linked-list OIT  Split a render target into smaller tiled regions  Each tile is rendered separately  Unsuitable for off-line rendering  Overflow is still unpredictable  Scene-dependent parameter tuning split plit 1 rende der r targe get 8x8 8 rende der r targe gets

O UR UR C ONT ONS NTRIBUTIONS  Memory usage prediction for linked-list ray-bundles  Adaptive tile subdivision using the above prediction  Reduce the risk of memory overflow & light leaking error  Avoid over-splitting  Less parameter tuning Uniform rm tili ling Our adapt ptive tili ling

A DAPTI PTIVE T IL ING FO FOR R AY AY - BUNDLES LES ILIN 1. Intr Introduc uction 2. Ada daptive T Til iling for R or Ray ay-bu bundl dles 3. Expe perimental Results & Future Work ork

A DAPTI PTIVE T IL ING ILIN  Based on adaptive shadow mapping [Fernando01]  Quadtree-based tile subdivision  According to a low-resolution scene analysis  Analysis for memory usage prediction is also added  The overflow risk is reduced dramatically  It is not completely eliminated, however Our contribu bution ions

I MPORTAN CE & F & F RAG RAGMENT C OUNT NT A NALY LYSIS ANCE  Render two mipmaps from the ray-bundle direction  Pixels as quadtree nodes (resolution: 2 n ) render Importance mipmap Fragment count mipmap (re required ray ray de density) (memory usa sage pe per r ray ray)

R ECU RSIVE T IL ILE S UBDIV ISION CURS DIVIS  Start from the top mip level (root of the quadtree)  A tile is subdivided when overflow is predicted Subdi bdivision c con ondi dition for each tile Requ quire ired r d ray-bu bundl dle pix ixel l count Estima imated u d upper r bound computed with computed with importance mipmap fragm gment nt c count unt m mipm pmap

E XPER ENTAL R ES ULTS ERIMEN ESULT & & C ONCLUSI ONS SIONS 1. Intr Introduc uction 2. Ada daptive T Til iling for R or Ray ay-bu bundl dles 3. Expe perimental Results & Future Work ork

N O T IL ING ILIN 2000 sample directions Ray-bundle resolution: 1024 2 Node buffer size: 5M nodes Analysis resolution: 1024 2 100 100 secs MSE: SE: 2. 2.04 0435e-2 ov overflow ratio: 0% Grou Ground tru ruth GPU: NVIDIA GeForce GTX 580 with 1.5GB memory

35 35 X 35 U 35 U NIFO FORM T IL ING ILIN 2000 sample directions Ray-bundle resolution: 1024 2 Node buffer size: 5M nodes Analysis resolution: 1024 2 1381 1381 secs MSE: SE: 3. 3.33 3344e-3 overflow ratio: 6.96% ov Grou Ground tru ruth GPU: NVIDIA GeForce GTX 580 with 1.5GB memory

O UR UR A DAPTI PTIVE T IL ING ILIN (172.7 tiles / direction) 2000 sample directions Ray-bundle resolution: 1024 2 Node buffer size: 5M nodes Analysis resolution: 1024 2 1396 1396 secs MSE: SE: 2. 2.53 5349e-4 overflow w ratio: 1.27e 7e-2% Grou Ground tru ruth GPU: NVIDIA GeForce GTX 580 with 1.5GB memory

C OMPU TION T IM IMES PER ER S AMPLE LE D IRE CTION MPUTA TATI RECT 2% o overhead head 7.9 13.3 2.4 7.5 12.5 Analysi sis Ren ender ering 0.3 0.5 0.3 0.2 0.3 Mipm pmappi pping 0.3 0.2 0.3 0.4 0.4 Tile S Subdivisi sion 0.7 0.8 0.6 0.8 0.7 GPU GPU-CPU D Data Copy Copy 291.6 405.7 69.9 269.8 418.9 Ray ay-bu bundl dle Cre Creation 180.3 274.7 63.8 217.4 286.9 Light M Map Update te (ms) GPU: NVIDIA GeForce GTX 580 with 1.5GB memory

C ONCLUSI ONS SIONS  Adaptive tiling for linked-list ray-bundles  A tiles is subdivided when overflow is predicted  The risk of memory overflow is reduced dramatically  Less parameter tuning  Memory usage prediction  Using the fragment count mipmap  Demonstrated baking light maps of large scenes  With a limited memory capacity

F UT URE W ORK RK UTUR  Improving the analysis accuracy  Supersampling  Conservative rasterization [Hasselgren05]  Ray-bundle warping  Rectilinear texture warping [Rosen12] Warpi rping  Real-time linked-list OIT  For an arbitrary node buffer size

T HA HANK Y OU OU