Ifrit-v2

GPU/CPU-Parallelized tile-based software rasterizer.

Successor to following repos:

• Ifrit
• Iris (TinyRenderer CPP)
• Iris (TinyRenderer C#)

Features

• Performance:

  • Multithreaded Rasterization
  • SIMD Vectorization
  • CUDA Acceleration (Incomplete)
    • Double Buffering / Overlapped Memory Transfer
    • Device Vector / Dynamic Array

• Rendering:

  • Homogeneous Space Clipping
  • Programmable VS/FS
  • Z Pre-Pass (CUDA-Only)

• Presentation:

  • Terminal Rendering (ASCII Characters/Color Image)

Performance

Test performed on 2048x2048 RGBA FP32 Image + 2048x2048 FP32 Depth Attachment. Time consumption in presentation stage (displaying texture via OpenGL) is ignored.

Frame Rate

Model	Triangles	CPU Single Thread*	CPU Multi-thread*	CUDA w/ Copy-back	CUDA w/o Copy-back**
Yomiya	70275	38 FPS	80 FPS	123 FPS^	400 FPS^
Stanford Bunny	208353	20 FPS	80 FPS	124 FPS^	320 FPS^
Khronos Sponza	786801	2 FPS	10 FPS	105 FPS	239 FPS
Intel Sponza	11241912	1 FPS	7 FPS	99 FPS	112 FPS

*. Under optimization

**. Might be influenced by other applications which utilize GPU

^. Result of previous commit.

Test Environment

• CPU: 12th Gen Intel(R) Core(TM) i9-12900H

  • Test with 16 threads + AVX2 Instructions

• GPU: NVIDIA GeForce RTX 3070 Ti Laptop GPU

• Shading: World-space normal

Dependencies

• Hardware Requirements:
  • SSE
  • AVX2
  • CUDA 12.4
• Presentation Dependencies:
  • Terminal (Windows Terminal)
  • OpenGL 3.3
  • GLFW 3.3
  • GLAD
• Compile Dependencies:
  • CMake 3.28
  • MSVC (Visual Studio 2022)
    • C++17 is required
    • C++20 is recommended for best performance
  • NVCC

Ongoing Plan

• Bug Fixing
  • Undeterministic Behaviors in CUDA Renderer Piepline
• CPU Pipeline Optimization
  • Performance: SIMD for tile-level pixel shading
  • Performance: Z Pre-Pass
  • Performance: Reduce fp division
• CUDA Pipeline Optimization
  • Performance: Batched Geometry Processing
  • Performance: Limited Varyings Or Exhaustive Local Mem Writes
  • Performance: Pixel Processing Bottleneck
  • Performance: Binner / Geometry Processing Low Throughput
  • Performance: Pixel Processing Low Cache Utilization Efficiency
• Blending
• Scanline Rasterizer (IMR)

References