Skip to content

README_Ishaan.md

Latest commit

 

History

History
69 lines (55 loc) · 2.77 KB

README_Ishaan.md

File metadata and controls

69 lines (55 loc) · 2.77 KB

CIS565: Project 1: CUDA Raytracer

Fall 2013 - Ishaan Singh

Due Thursday, 09/19/2013

Features:

  • Ray tracing per bounce. Compaction not completed yet.

  • Specular Reflection (make has reflection between 0.1 and 1)

  • Soft Shadows and Area Lights

  • Depth of Field

  • Jittered Antialiasing

  • MSAA: Multi Sample Anti Aliasing

  • Refraction (make has refraction between 0.1 and 1)

  • Camera Movement

    w-a-s-d-q-z move camera in world space

    [-]-o-p pan camera around

    up - down for moving the focal plan in and out

    left - right for increasing and decreasing aperture size

Images:

http://github.com/ishaan13/Project1-RayTracer/blob/master/renders/2K_10_noCompaction.bmp http://github.com/ishaan13/Project1-RayTracer/blob/master/renders/2K_10_noCompaction_jitter.bmp http://github.com/ishaan13/Project1-RayTracer/blob/master/renders/2K_10_noCompaction_jitter_DOF.bmp

Videos:

https://www.dropbox.com/s/htm8t0xxahi0z1l/RayTracerGPU.avi https://www.dropbox.com/s/xnmadcdf24i5zr4/RayTracerGPUDOF.avi

Performance Evaluation:

Peformance Boosts

  • Ray-Box intersection code has less branching
  • Bounce Parallelization

Performance increase strategy:

  • Compaction should increase speed massively because of decrease in dead threads
  • Put light source geometry into shared memory
  • Less global memory access in code. Use more register-variables

Performance Analysis:

Varying Number of threads per block and looking at time per iteration for a maximum of 10 bounces: on NVidia GT640M (384 cores, 1GB)

  • 4 x 4 : 395ms
  • 8 x 8 : 120ms
  • 12 x 12 : 139ms
  • 16 x 16 : 125ms
  • 24 x 24 : 229ms
  • 32 x 32 : 156ms

So, the interesting thing is that at multiples of a half-warp size per dimension, there's a significant increase in performance as all the local maxima's. It would be curious to understand why. The reason being that even the 24 x 24 is a multiple of a 32 warp size.