ninodimontalcino / advanced-denoising-mcrenderings

Fast and robust denosing of monte carlo renderings using color and feature information (AVX optimized)

ASL Project - Team 053

Main Paper: Robust Denoising using Feature and Color Information, Rouselle https://onlinelibrary.wiley.com/doi/full/10.1111/cgf.12219

Team Members

• Alexandre Binninger ([email protected])
• Félicité Lordon--de Bonniol du Trémont ([email protected])
• Alexandre Poirrier ([email protected])
• Nino Scherrer ([email protected])

Folder Structure

.
|── analysis                      # Scripts for automated performance evaluation and plotting
|── implementations               # Implemented Denoising Algorithms
│   ├──  basic.cpp                      # Vanilla Implementation of referenced denoising algorithm
│   ├──  <...>.cpp                      # Optimized Version (mention technique)
├── reference_implementation      # Joint Non-Local Means Implementation (Matlab)
├── renderings                    # Extracted Buffers of Monte Carlo Renderings
│   ├──  100spp                         # MC-Rendering using 100 samples per pixel
|   |    ├── _800x600                            # 800x600 (used for validation)
│   |    ├── <size>                              # <size> x <size> rendering
│   ├──  5000spp_GT                      # MC-Rendering using 5000 samples per pixel (Claimed Grountd-Truth)
├── src                           # Libraries
│   ├── ext                             # External Libraries    
│   |   ├── openexr                             # OpenEXR (used for Data Loading)
│   |   ├── zlib                                # zlib (data compression used by OpenEXR)
├── tests                         # Testing directory
|   ├──  test_generation              # Modified reference_implementation to create test data
|   ├──  test_data                    # Generated test data

Building

mkdir build
cd build
mkdir temp
git submodule update --init --recursive
cmake ..
make
./main

Tests

To compile tests:

mkdir build
cd build
cmake ..
make
cd ..
build/denoise_test

In this version, tests are for the FLT function (only the main computation, not the estimators for SURE). Test data could also be used to compare Matlab OpenEXR and the C++ openEXR implementation.

Autotuning script

Allows to automatically unroll loops to find best version.

Please place the script and run from build directory. If you want to unroll a file filename.c, create another file filename_unroll.c next to it, and replace loops to unroll with the unrolling tags (see below).

Usage:

python unroll.py ../filename.c n1,n2,n3,...

with n1, n2, ... the unrolling factors to test.

Note: best is to keep only one reference function in the register.hpp to get faster results.

Unrolling tags

Use // $unroll 8 to specify a portion of code to unroll loop (until // $end_unroll tag). The unrolling factor is specified in the starting tag.

Then use $i for varying indices, and $n refers to the unrolling factor.

Example:

// $unroll 2
scalar sum_r_$i;
for(int xp = R; xp < W - R; ++xp) {
    for(int yp = R + F_R; yp < H - R - F_R; yp+=$n) {
        sum_r_$i = 0.f;

        for (int k=-F_R; k<=F_R; k++){
            sum_r_$i += temp[xp * W + yp+k+$i];
        }
        temp2_r[xp * W + yp+$i] = sum_r_$i;
    }
}
// $end_unroll

unrolls as

// $unroll 2
scalar sum_r_0;
for(int xp = R; xp < W - R; ++xp) {
    for(int yp = R + F_R; yp < H - R - F_R; yp+=2) {
        sum_r_0 = 0.f;
        sum_r_1 = 0.f;

        for (int k=-F_R; k<=F_R; k++){
            sum_r_0 += temp[xp * W + yp+k+0];
            sum_r_1 += temp[xp * W + yp+k+1];
        }
        temp2_r[xp * W + yp+0] = sum_r_0;
        temp2_r[xp * W + yp+1] = sum_r_1;
    }
}
// $end_unroll

Instead of // $unroll n, use // $auto_unroll for automatically unroll multiple times and compare versions.