COSCUP 2023 - Make Your Own Ray Tracing GPU with FPGA
0 likes607 views
This document outlines a session by Owen Wu at Coscup 2023 about building a ray tracing GPU using FPGA technology. It covers the fundamentals of hardware description languages, electronic design automation, and FPGA design, emphasizing the differences between hardware and software thinking. The session includes resources for beginners, an overview of ray tracing techniques, and an open-sourced project for creating a basic ray tracing GPU architecture.
COSCUP 2023 - Make Your Own Ray Tracing GPU with FPGA
- 1. Make Your Own Ray Tracing GPU with FPGA COSCUP 2023 Owen Wu fallingcat@gmail.com
- 2. Agenda ● Self-Intro ● Session Overview ● How to Start ○ HDL ○ EDA ○ FPGA ● Think Differently ● Ray Tracing ● HomebrewGPU projrect
- 3. Self-Intro
- 4. Self-Intro ● Game developer ○ Game engine development ○ PC ○ Console ○ Mobile ● GPU software engineer ○ Optimization from the perspective of hardware ○ AMD ○ Arm ● https://tinyurl.com/owenwu
- 6. Session Overview ● This session is for software engineer ● This session is NOT for hardware engineer ● Basic intro for the beginner ● Making a chip is very easy and cheap nowaday ○ 6 months from zero to a workable GPU ● Turn you algorithm into hardware, think differently ● Many open sourced projects on GitHub
- 7. How to Start
- 8. How to Start ● HDL (Haardware Description Language) ○ Language ● EDA (Electronic Design Automation) ○ Compiler ● FPGA (Field Programmable Gate Array) ○ Hardware
- 10. HDL ● VHDL ○ Ada ● Verilog ○ C ● Connect modules to design a whole chip ● Clock is the way to sync all modules ○ 100M Hz - generate 100M signals in one second ● You can think module as a function in C ● Every module can only do very limited works ○ Works need to be finished in one clock ● Books for begineer ○ Programming FPGAs: Getting Started with Verilog ○ Introduction to Verilog
- 12. EDA ● Convert HDL to bitstream file ● Upload bitstream file to FPGA to execute ● Many steps ○ IC Design ○ Synthesis ○ Verification ○ Physical Design ● You can think EDA as a compiler ● FPGA makers provide basic EDA for free ○ Xilinx Vivado ● You can also try EDA online ○ https://8bitworkshop.com/v3.10.1/?platform=verilog&file=cloc k_divider.v
- 14. FPGA ● Upload bitstream file to configure logic blocks ● FPGA development board integrate many components ○ VGA/HDMI output ○ Memory ○ LED ○ 7 segment disply ○ SD Card ○ SoC ● FPGA has different number of logic cells ○ Which decides how complex the design can be ● Elbert V2 for beginner ● Nexys A7 for more complex design
- 16. Parallel ● Software is serial ● Hardware is parallel ● Every modules work simultaneously ● Software optimizatios may not work with hardware ● Don’t use software thinking when designing hardware B() MUX A() R C
- 17. Latency v.s. Throughput ● CPU performance depends on latency ● Low latency means that the instruction can be completed quickly ● Instruction has order dependency ● GPU only care how long it takes to finish a frame ● Pixel doesn’t have order dependency ● Pixels can be executed simultaneously ● If the latency of one pixel is 32 clock ● The hardware executes 64 streams simultaneously ● The throughput will be 2 pixel per clock
- 18. Pipeline ● Hardware has many different modules ● All modules need to work simultaneously to get the best performance ● Use pipeline to split the tasks ● Every module process different pixel at the same time Surface (5 clk) Shadow (5 clk) Shading (5 clk) 15 clk/pixel Surface Shadow Shading pixel 1 pixel2 pixel 1 pixel3 pixel2 pixel 1 pixel 4 pixel 3 pixel 2 5 clk/pixel clk 0 clk 5 clk 10 clk 15
- 19. Ray Tracing
- 20. Ray Tracing ● Ray tracing is easy to implement ● Ray hit objects then reflect to camera ● Invert the ray ● Cast a ray from each pixel of the screen ● Find a closest hit of ray and objects ● Decide the color of the pixel ○ If there is a hit, the color of hit object ○ If there is no hit, the color of background ● Ray Tracing in One Weekend at GitHub
- 21. Ray Tracing - Reflection ● For the hit on a object ● If the object is reflective ● Cast a reflection ray from hit point ● Find the closet hit of the reflection ray ● Recursively cast reflection rays ● Blend the colosr of reflected objects into final shading
- 22. Ray Tracing - Shadow ● For the hit on a object ● Cast a new ray from hit point toward light ● Is there is any hit of the new ray ● If yes, the pixel is in shadow ● Otherwise the pixel is not in shadow
- 23. BVH(Bounding Volume Hierarchy) ● Accelerate the hit detection between ray and primitives ○ Quickly exclude the nodes which don’t have intersection ● Use AABB(Axis-aligned Bounding Box) to split the space ● Traverse the AABB until reach the leaf ● Detect the hits of ray between the primitive in leaf AABB AABB AABB AABB AABB Left Node Right Node
- 25. HomebrewGPU project ● Open sourced project ○ https://github.com/fallingcat/HomebrewGPU ● Implement a basic ray tracing GPU ○ Voxel based rendering ○ BVH acceleration ○ Shading/Reflection/Refraction/Shadow ● 6 months from zero to complete
- 27. Architecture
- 28. Architecture ● Thread Generator ○ Generate one thread per clock for each ray core ○ Each thread presents one pixel ○ The thread will go through ray core and output the final color ● BVH Structure ○ BVH structure stores the BVH tree structure data ○ Accepts the node or leaf query from ray core ○ Output the node or leaf data to ray core ● Primitive Unit ○ Primitive Unit stores the raw data of all primitives ○ Accepts the query from ray core and output primitive data ● Ray Core ○ Ray core process one thread to output the final color ○ Accepts the thread from thread generator or reflection/refraction ray ● Frame Buffer Writer ○ Cache the output of ray cores and write the pixel to frame buffer ○ Some threads with reflection/refraction take longer to get the final color ○ Wait util all threads in one cache set are finished then write the data to the frame buffer
- 29. Architecture ● Surface stage ○ Process the ray from camera and find the closest hit of the ray ○ Pass the hit information to next stage ● Shadow stage ○ Cast a ray from closest hit position to light source ○ It will pass the shadow information to next stage ● Shade stage ○ Use the closest hit information to decide if it's the final color or reflection/refraction will occur ○ Cast a reflection/refraction ray and pass the data back to surface stage ○ Recursively feed back to surface stage
- 31. Q & A ● Owen (fallingcat@gmail.com) ● https://github.com/fallingcat/Ho mebrewGPU
- 32. Thanks!