redflash renderer / Raytracing Camp 7

レイトレ合宿7の自作レンダラーのプレゼン資料です。
OptiX6上でレイマーチングとポリゴンをハイブリッドにレンダリングできるパストレーサーを実装しました。
https://github.com/gam0022/redflash
https://sites.google.com/site/raytracingcamp7/

Transcript

1. SFEqBTISFOEFSFS
   2019/09/07 ϨΠτϨ߹॓7
   @gam0022 / Sho HOSODA
   

   View Slide

2. View Slide

3. Α͋͘Δػೳ͸͢΂࣮ͯ૷ʂ
   • Unidirectional Path Tracing
   ‣ implemented in NVIDIA® OptiX 6.0
   • Disney BRDF
   • Next Event Estimation (Direct Light Sampling)
   • Multiple Importance Sampling
   • Primitives
   ‣ Sphere
   ‣ Mesh
   ‣ Distance Function (Raymarching)
   • ACES Filmic Tone Mapping
   • Deep Learning Denoising
   3
   

   View Slide

4. Raymarching × Mesh
   • ϨΠϚʔνϯάͱϙϦΰϯϝογϡ͕

   ࠞࡏͨ͠γʔϯΛ

   ύετϨʔγϯάʹΑΔ

   Ұ؏ͨ͠ϨϯμϦϯά͕Ͱ͖ͨʂ
   • ϙϦΰϯͰ͸೉͍͠

   ෳࡶͳδΦϝτϦΛ

   ϨΠϚʔνϯάͰදݱ
   4
   

   View Slide

5. ϨΠϚʔνϯάͱ͸
   • ڑ཭ؔ਺ΛώϯτʹϨΠΛগͣͭ͠ਐΊΔ͜ͱͰަࠩ൑ఆ
   • ڑ཭ؔ਺ = ೚ҙͷ఺p͔Β෺ମද໘΁ͷ࠷୹ڑ཭Λฦؔ͢਺
   5
   float sdSphere( vec3 p, float s )
   {
   return length(p)-s;
   }
   float udBox( vec3 p, vec3 b )
   {
   return length(max(abs(p)-b,0.0));
   }
   http://iquilezles.org/www/articles/distfunctions/distfunctions.htm
   

   View Slide

6. NVIDIA® OptiXͱϨΠϚʔνϯάͷ౷߹
   • IntersectionProgram ͱ BoundingBoxProgram Λ࣮૷͢Ε͹OK
   ‣ ϨΠϚʔνϯάʹΑΔަࠩ൑ఆͱAABBͷఆٛΛCUDAͰ࣮૷
   • ৄࡉ͸ϒϩάهࣄͰ
   • https://gam0022.net/blog/2019/08/05/optix-raymarching-pathtracing/
   6
   

   View Slide

7. ϨΠϚʔνϯά͸ܭࢉྔ͕ଟ͍
   • IntersectionProgram ͷதͰڑ཭ؔ਺Λ਺ඦճ΋ධՁ͢Δඞཁ͕͋Δ
   • GPUͰ΋ϑϥΫλϧΛؾܰʹϨΠϚʔνϯά͢Δͱܹॏʹ…
   7
   ߴ଎ԽΛؤுͬͨʂ
   

   View Slide

8. ڑ཭ؔ਺ͷܰྔԽ
   • Mandelbox ͱ͍͏఻౷తͳϑϥΫλϧਤܗ
   • if෼ذ΋͋ͬͯॲཧ͕ॏ͍…
   • ݟͨ໨ͷྼԽΛ཈͑ͳ͕Βڑ཭ؔ਺ΛܰྔԽʂ
   8
   // http://blog.hvidtfeldts.net/index.php/2011/11/distance-estimated-3d-fractals-vi-the-mandelbox/
   float minRadius2 = 0.5;
   float fixedRadius2 = 1.0;
   float foldingLimit = 1.0;
   #define Iterations 8
   void sphereFold(inout vec3 z, inout float dz) {
   float r2 = dot(z,z);
   if (r2 < minRadius2) {
   // linear inner scaling
   float temp = (fixedRadius2 / minRadius2);
   z *= temp;
   dz *= temp;
   } else if (r2 < fixedRadius2) {
   // this is the actual sphere inversion
   float temp = fixedRadius2 / r2;
   z *= temp;
   dz *= temp;
   }
   }
   void boxFold(inout vec3 z, inout float dz) {
   z = clamp(z, -foldingLimit, foldingLimit) * 2.0 - z;
   }
   float dMbox(vec3 z) {
   vec3 offset = z;
   float dr = 1.0;
   for (int n = 0; n < Iterations; n++) {
   boxFold(z, dr); // Reflect
   sphereFold(z, dr); // Sphere Inversion
   z = kadoScale * z + offset; // Scale & Translate
   dr = dr * abs(kadoScale) + 1.0;
   }
   float r = length(z);
   return r / abs(dr);
   }
   float dMandel(float3 p, float scale, int n) {
   float4 q0 = float4 (p, 1.0);
   float4 q = q0;
   for (int i = 0; i < n; i++) {
   q.xyz = clamp(q.xyz, -1.0, 1.0) * 2.0 - q.xyz;
   q = q * scale / clamp(dot(q.xyz, q.xyz), 0.5, 1.0) + q0;
   }
   return length(q.xyz) / abs(q.w);
   }
   ܰྔԽ
   ※GLSL
   ※HLSL
   ҰൠԽ͞Εͨܗ͔ࣜΒύϥϝʔλ
   ൣғʹ੍ݶΛ͚ͭͯॲཧΛ؆ུԽ
   

   View Slide

9. ϨΠϚʔνϯάͷিಥ൑ఆͷਫ਼౓ͷLOD 1/2
   • Χϝϥʹ͍ۙ෦෼͸ detail ͕ཉ͍͠
   • ԕ͍෦෼͸ݫີʹিಥ൑ఆ͠ͳͯ͘ྑ͍
   ͭ·Γ-0%(Level of Detail) ͕͍ͨ͠ʂ
   • Mandelboxͷڑ཭ؔ਺͸ %JTUBODF&TUJNBUPSʢڑ཭ਪఆثʣ
   ‣ ༗ݶͷΠςϨʔγϣϯճ਺Ͱ͸ද໘ʹ઴ۙ͢Δ͕ɺڑ཭0ʹͳΒͳ͍
   ‣ ద౰ͳڑ཭ eps ͰিಥͱΈͳͯ͠ܭࢉΛଧͪ੾Δඞཁ͕͋Δ
   ‣ eps Λখ͘͢͞ΔͱɺΑΓࡉ͔͍ detail ·ͰՄࢹԽͰ͖Δ
   ‣ ԕܠ·Ͱ eps ͕খ͍͞ͱߴप೾੒෼͕ݱΕͯ݁Ռ͕Ԛ͘ͳΔʢMipMap OFFঢ়ଶʣ
   9
   ଎౓໘
   ඼࣭໘
   

   View Slide

10. ϨΠϚʔνϯάͷিಥ൑ఆͷਫ਼౓ͷLOD 2/2
    • Χϝϥ͔Βͷڑ཭ = ϨΠ͕ਐΜͩڑ཭ Λ༻͍ͯ eps Λಈతʹܾఆ
    • ϨΠϚʔνϯάͷিಥ൑ఆͷਫ਼౓ͷLOD͕Ͱ͖ͨʂ
    • ۙܠʙԕܠ͕ࠞࡏͨ͠ΧοτͰ΋៉ྷʹඳըͰ͖Δʂ
    10
    

    View Slide

11. 1ճͷlaunchͰͳΔ΂ͨ͘͘͞ΜαϯϓϦϯά͢Δઓུ
    • OptiXͰ͸ launch ͰγʔϯͷඳըΛ࣮ߦͰ͖Δ
    • sample_per_launchʢ1ճͷlaunch͝ͱͷαϯϓϦϯάճ਺ʣ

    Λ૿΍͢ͱɺCPUॲཧͷΦʔόʔϔουΛݮΒͤΔ
    ‣ υϩʔίʔϧ࡟ݮͷ࠷దԽͱಉ͡ݪཧ
    • ࠷ॳͷ4αϯϓϦϯάͰϚγϯͷੑೳΛϕϯνϚʔΫ
    ‣ ࣌ؒ੾ΕʹͳΒͳ͍࠷େͷ sample_per_launch Λܾఆ
    ‣ launch ίετΛ࠷খԽ
    11
    TBNQMF@QFS@MBVODI UPUBMTBNQMF
    
    
    
    RTX2070 60ඵͷܭଌ݁Ռ
    

    View Slide

12. Deep Learning Denoising
    • 10sppͷ݁ՌɻڻҟతͳσϊΠζੑೳ
    12
    Original Denoised
    

    View Slide

13. • Deep Learning Denoising ༻ʹ albedo ͱ normal Λੜ੒
    ‣ Bufferʢ= RenderTextureʣ͕߹ܭͰ5ຕ΋ඞཁʹͳͬͯ͠·ͬͨ
    • ΢Οϯυ΢ʹදࣔ͠ͳ͍ͳΒ RT_BUFFER_GPU_LOCAL ͰOK
    ‣ createBuffer ͷୈ1Ҿ਺ͰࢦఆͰ͖Δ
    ‣ An RT_BUFFER_INPUT_OUTPUT has separate copies on each device that are not synchronized
    • 60ඵؒͷαϯϓϦϯά਺͕ 88 → 100 (13.6%) ʹ૿͑ͨʂ
    RT_BUFFER_GPU_LOCAL ʹΑΔ࠷దԽ
    13
    normal albedo liner color
    gamma corrected

    + tonemapped
    denoised
    

    View Slide

14. ·ͱΊ
    ͜ΕΒͷ࠷దԽʹΑͬͯ…
    14
    ,ղ૾౓ग़ྗΛ࣮ݱʂ
    

    View Slide

15. ผࢹ఺͔ΒͷϨϯμϦϯά݁Ռ
    15
    

    View Slide

16. View Slide

17. View Slide

18. View Slide