oit and indirect illumination using dx11 linked lists - amd

OIT and Indirect Illumination using DX11 Linked Lists

Holger Gruen AMD ISV RelationsNicolas Thibieroz AMD ISV Relations

Agenda

Introduction Linked List Rendering Order Independent Transparency Indirect Illumination Q&A

Introduction Direct3D 11 HW opens the door to many

new rendering algorithms In particular per pixel linked lists allow

for a number of new techniquesOIT, Indirect Shadows, Ray Tracing of dynamic scenes, REYES surface dicing, custom AA, Irregular Z-buffering, custom blending, Advanced Depth of Field, etc.

This talk will walk you through:A DX11 implementation of per-pixel linked list and two effects that utilize this techique OIT Indirect Illumination

Per-Pixel Linked Lists with Direct3D 11

ElementLink

ElementLink

ElementLink

ElementLink

Nicolas ThibierozEuropean ISV RelationsAMD

Why Linked Lists? Data structure useful for programming

Very hard to implement efficiently with previous real-time graphics APIs

DX11 allows efficient creation and parsing of linked lists

Per-pixel linked lists A collection of linked lists enumerating all pixels

belonging to the same screen position

ElementLink

ElementLink

ElementLink

ElementLink

Two-step process

1) Linked List Creation Store incoming fragments into linked

lists

2) Rendering from Linked List Linked List traversal and processing of

stored fragments

Creating Per-Pixel Linked Lists

PS5.0 and UAVs Uses a Pixel Shader 5.0 to store

fragments into linked lists Not a Compute Shader 5.0!

Uses atomic operations Two UAV buffers required

- “Fragment & Link” buffer - “Start Offset” buffer

UAV = Unordered Access View

Fragment & Link Buffer The “Fragment & Link” buffer

contains data and link for all fragments to store

Must be large enough to store all fragments

Created with Counter support D3D11_BUFFER_UAV_FLAG_COUNTER flag in UAV view

Declaration:struct FragmentAndLinkBuffer_STRUCT{ FragmentData_STRUCT FragmentData; // Fragment data uint uNext; // Link to next fragment};RWStructuredBuffer <FragmentAndLinkBuffer_STRUCT> FLBuffer;

Start Offset Buffer The “Start Offset” buffer contains

the offset of the last fragment written at every pixel location

Screen-sized:(width * height * sizeof(UINT32) )

Initialized to magic value (e.g. -1) Magic value indicates no more fragments are

stored (i.e. end of the list) Declaration:

RWByteAddressBuffer StartOffsetBuffer;

Linked List Creation (1) No color Render Target bound!

No rendering yet, just storing in L.L. Depth buffer bound if needed

OIT will need it in a few slides UAVs bounds as input/output:

StartOffsetBuffer (R/W) FragmentAndLinkBuffer (W)

-1 -1 -1 -1 -1 -1

-1 -1 -1 -1 -1 -1

-1 -1 -1 -1 -1 -1Viewport

Start Offset Buffer

-1 -1 -1 -1 -1 -1

-1 -1 -1 -1 -1 -1

-1 -1 -1 -1 -1 -1

Fragment and Link Buffer

Linked List Creation (2a)



Fragment and Link Buffer Counter = 0

-1 -1 -1 -1 -1 -1

-1 -1 -1 -1 -1

-1 -1 -1 -1 -1 -1

Start Offset Buffer

-1 -1 -1 -1 -1 -1

-1 -1 -1 -1 -1 -1

-1 -1 -1 -1 -1 -1


Linked List Creation (2b)

-1Fragment and Link Buffer


-1

-1


Viewport

-1 -1

-1 -1 -1 -1 -1 -1

-1 0 -1 -1 -1 -1

-1 -1 -1 -1 -1 -1

Start Offset Buffer

-1 -1 -1 -1 -1 -1

-1 -1 -1 -1

-1 -1 -1 -1 -1 -1


Linked List Creation (2c)



-1


-1-1

Viewport

-1

-1 -1 -1 -1 -1 -1

-1 -1 -1 -1

-1 -1 -1 -1 -1 -1

Start Offset Buffer

-1 -1 -1 -1 -1 -1

-1 -1 -1 1 2 -1

-1 -1 -1 -1 -1 -1

Linked List Creation (2d)



-1 -1-1

0


0

5

-1

Viewport

Linked List Creation - Codefloat PS_StoreFragments(PS_INPUT input) : SV_Target{ // Calculate fragment data (color, depth, etc.) FragmentData_STRUCT FragmentData = ComputeFragment();

// Retrieve current pixel count and increase counter uint uPixelCount = FLBuffer.IncrementCounter();

// Exchange offsets in StartOffsetBuffer uint vPos = uint(input.vPos); uint uStartOffsetAddress= 4 * ( (SCREEN_WIDTH*vPos.y) + vPos.x ); uint uOldStartOffset; StartOffsetBuffer.InterlockedExchange(

uStartOffsetAddress, uPixelCount, uOldStartOffset);

// Add new fragment entry in Fragment & Link Buffer FragmentAndLinkBuffer_STRUCT Element; Element.FragmentData = FragmentData; Element.uNext = uOldStartOffset; FLBuffer[uPixelCount] = Element;}

Traversing Per-Pixel Linked Lists

Rendering Pixels (1) “Start Offset” Buffer and “Fragment &

Link” Buffer now bound as SRVBuffer<uint> StartOffsetBufferSRV;StructuredBuffer<FragmentAndLinkBuffer_STRUCT>

FLBufferSRV; Render a fullscreen quad For each pixel, parse the linked list and

retrieve fragments for this screen position Process list of fragments as required

Depends on algorithm e.g. sorting, finding maximum, etc.

SRV = Shader Resource View

Render Target

Start Offset Buffer

Rendering from Linked List




-1 -1 -1

4-1 -1 -1 -1 -1 -1

-1 -1 -1 -1

-1 -1 -1 -1 -1 -1

-1 -1 -1 -1 -1 -1

-1 -1 -1 1 2 -1

-1 -1 -1 -1 -1 -1

33

0

0

-1

-1

Rendering Pixels (2)float4 PS_RenderFragments(PS_INPUT input) : SV_Target{ // Calculate UINT-aligned start offset buffer address uint vPos = uint(input.vPos); uint uStartOffsetAddress = SCREEN_WIDTH*vPos.y + vPos.x; // Fetch offset of first fragment for current pixel uint uOffset = StartOffsetBufferSRV.Load(uStartOffsetAddress);

// Parse linked list for all fragments at this position float4 FinalColor=float4(0,0,0,0); while (uOffset!=0xFFFFFFFF) // 0xFFFFFFFF is magic

value { // Retrieve pixel at current offset Element=FLBufferSRV[uOffset]; // Process pixel as required ProcessPixel(Element, FinalColor); // Retrieve next offset uOffset = Element.uNext; } return (FinalColor);}

Order-Independent Transparency via Per-Pixel Linked Lists

Nicolas ThibierozEuropean ISV RelationsAMD

Description Straight application of the linked

list algorithm Stores transparent fragments into

PPLL Rendering phase sorts pixels in a

back-to-front order and blends them manually in a pixel shader Blend mode can be unique per-pixel!

Special case for MSAA support

Linked List Structure Optimize performance by reducing

amount of data to write to/read from UAV E.g. uint instead of float4 for color Example data structure for OIT:struct FragmentAndLinkBuffer_STRUCT{ uint uPixelColor; // Packed pixel color uint uDepth; // Pixel depth uint uNext; // Address of next link};

May also get away with packed color and depth into the same uint! (if same alpha) 16 bits color (565) + 16 bits depth Performance/memory/quality trade-off

Visible Fragments Only! Use [earlydepthstencil] in front of

Linked List creation pixel shader This ensures only transparent fragments

that pass the depth test are stored i.e. Visible fragments!

Allows performance savings and rendering correctness!

[earlydepthstencil]float PS_StoreFragments(PS_INPUT input) : SV_Target{ ...}

Sorting Pixels Sorting in place requires R/W

access to Linked List Sparse memory accesses = slow! Better way is to copy all pixels into

array of temp registers Then do the sorting

Temp array declaration means a hard limit on number of pixel per screen coordinates Required trade-off for performance

Sorting and Blending

Blend fragments back to front in PS Blending algorithm up to app Example: SRCALPHA-INVSRCALPHA Or unique per pixel! (stored in fragment data)

Background passed as input texture Actual HW blending mode disabled

0.87

00.98

-10.95

120.93

340.95 0.93 0.87 0.98

Temp ArrayBackground

color

PS color

Render Target

Storing Pixels for Sorting (...) static uint2 SortedPixels[MAX_SORTED_PIXELS]; // Parse linked list for all pixels at this position // and store them into temp array for later sorting int nNumPixels=0; while (uOffset!=0xFFFFFFFF) { // Retrieve pixel at current offset Element=FLBufferSRV[uOffset]; // Copy pixel data into temp array SortedPixels[nNumPixels++]=

uint2(Element.uPixelColor, Element.uDepth); // Retrieve next offset [flatten]uOffset = (nNumPixels>=MAX_SORTED_PIXELS) ? 0xFFFFFFFF : Element.uNext; }

// Sort pixels in-place SortPixelsInPlace(SortedPixels, nNumPixels); (...)

Pixel Blending in PS (...) // Retrieve current color from background texture float4 vCurrentColor=BackgroundTexture.Load(int3(vPos.xy, 0));

// Rendering pixels using SRCALPHA-INVSRCALPHA blending for (int k=0; k<nNumPixels; k++) { // Retrieve next unblended furthermost pixel float4 vPixColor= UnpackFromUint(SortedPixels[k].x);

// Manual blending between current fragment and previous one vCurrentColor.xyz= lerp(vCurrentColor.xyz, vPixColor.xyz, vPixColor.w); }

// Return manually-blended color return vCurrentColor;}

OIT via Per-Pixel Linked Lists with MSAA Support

Sample Coverage Storing individual samples into Linked

Lists requires a huge amount of memory ... and performance will suffer!

Solution is to store transparent pixels into PPLL as before

But including sample coverage too! Requires as many bits as MSAA mode

Declare SV_COVERAGE in PS structure struct PS_INPUT { float3 vNormal : NORMAL; float2 vTex : TEXCOORD; float4 vPos : SV_POSITION; uint uCoverage : SV_COVERAGE; }

Linked List Structure Almost unchanged from previously Depth is now packed into 24 bits 8 Bits are used to store coveragestruct FragmentAndLinkBuffer_STRUCT{ uint uPixelColor; // Packed pixel color uint uDepthAndCoverage; // Depth + coverage uint uNext; // Address of next link};

Pixel Center

Sample

Sample Coverage Example

Third sample is covered uCoverage = 0x04 (0100 in binary)

Element.uDepthAndCoverage = ( In.vPos.z*(2^24-1) << 8 ) | In.uCoverage;

Rendering Samples (1) Rendering phase needs to be able to write

individual samples Thus PS is run at sample frequency

Can be done by declaring SV_SAMPLEINDEX in input structure

Parse linked list and store pixels into temp array for later sorting Similar to non-MSAA case Difference is to only store sample if coverage

matches sample index being rasterized

static uint2 SortedPixels[MAX_SORTED_PIXELS];

// Parse linked list for all pixels at this position // and store them into temp array for later sorting int nNumPixels=0; while (uOffset!=0xFFFFFFFF) { // Retrieve pixel at current offset Element=FLBufferSRV[uOffset];

// Retrieve pixel coverage from linked list element uint uCoverage=UnpackCoverage(Element.uDepthAndCoverage); if ( uCoverage & (1<<In.uSampleIndex) ) { // Coverage matches current sample so copy pixel SortedPixels[nNumPixels++]=Element; } // Retrieve next offset [flatten]uOffset = (nNumPixels>=MAX_SORTED_PIXELS) ? 0xFFFFFFFF : Element.uNext; }

Rendering Samples (2)

OIT Linked List DemoDEMO

Holger GruenEuropean ISV Relations AMD

Direct3D 11 Indirect Illumination

Indirect Illumination Introduction 1

Real-time Indirect illumination is an active research topic

Numerous approaches existReflective Shadow Maps (RSM) [Dachsbacher/Stammiger05]Splatting Indirect Illumination [Dachsbacher/Stammiger2006]Multi-Res Splatting of Illumination [Wyman2009]Light propagation volumes [Kapalanyan2009]Approximating Dynamic Global Illumination in Image Space [Ritschel2009]

Only a few support indirect shadowsImperfect Shadow Maps [Ritschel/Grosch2008]Micro-Rendering for Scalable, Parallel Final Gathering(SSDO) [Ritschel2010] Cascaded light propagation volumes for real-time indirect illumination [Kapalanyan/Dachsbacher2010]

Most approaches somehow extend to multi-bounce lighting

Indirect Illumination Introduction 2

This section will coverAn efficient and simple DX9-compliant RSM based implementation for smooth one bounce indirect illumination

Indirect shadows are ignored here

A Direct3D 11 technique that traces rays to compute indirect shadows

Part of this technique could generally be used for ray-tracing dynamic scenes

Indirect Illumination w/o Indirect Shadows1. Draw scene g-buffer2. Draw Reflective Shadowmap (RSM)

1. RSM shows the part of the scene that recieves direct light from the light source

3. Draw Indirect Light buffer at ¼ res 1. RSM texels are used as light sources on g-

buffer pixels for indirect lighting 4. Upsample Indirect Light (IL)5. Draw final image adding IL

Step 1

G-Buffer needs to allow reconstruction of World/Camera space position World/Camera space normal Color/ Albedo

DXGI_FORMAT_R32G32B32A32_FLOAT positions may be required for precise ray queries for indirect shadows

Step 2

RSM needs to allow reconstruction of World space position World space normal Color/ Albedo

Only draw emitters of indirect light DXGI_FORMAT_R32G32B32A32_FLOAT

position may be required for ray precise queries for indirect shadows

Step 3 Render a ¼ res IL as a deferred op Transform g-buffer pix to RSM space

->Light Space->project to RSM texel space Use a kernel of RSM texels as light

sources RSM texels also called Virtual Point Light(VPL)

Kernel size depends on Desired speed Desired look of the effect RSM resolution

Computing IL at a G-buf Pixel 1

Sum up contribution of all VPLs in the kernel


g-buffer pixel

RSM texel/VPL

LNpN

pP

LPpL

pL

PPPPD

VPLVPL

pL

LPVPL AreaCol

PPDNsatDNsatonContributi

2

This term is very similar to terms used in radiosity form factor computations


stx : sub RSM texel x position [0.0, 1.0[sty : sub RSM texel y position [0.0, 1.0[

A naive solution for smooth IL needs to consider four VPL kernels with centers at t0, t1, t2 and t3.

Computing IL at a g-buf pixel 4

IndirectLight = (1.0f-sty) * ((1.0f-stx) * + stx * ) + (0.0f+sty) * ((1.0f-stx) * + stx * )

Evaluation of 4 big VPL kernels is slow

stx : sub texel x position [0.0, 1.0[sty : sub texel y position [0.0, 1.0[

VPL kernel at t0VPL kernel at t1

VPL kernel at t2VPL kernel at t3

Computing IL at a g-buf pixel 5

SmoothIndirectLight = (1.0f-sty)*(((1.0f-stx)*(B0+B3)+stx*(B2+B5))+B1)+

(0.0f+sty)*(((1.0f-stx)*(B6+B3)+stx*(B8+B5))+B7)+B4 stx : sub RSM texel x position of g-buf pix [0.0, 1.0[sty : sub RSM texel y position of g-buf pix [0.0, 1.0[

This trick is probably known to some of you already. See backup for a detailed explanation !

Indirect Light Buffer

Step 4 Indirect Light buffer is ¼ res Perform a bilateral upsampling

step See Peter-Pike Sloan, Naga K. Govindaraju, Derek

Nowrouzezahrai, John Snyder. "Image-Based Proxy Accumulation for Real-Time Soft Global Illumination". Pacific Graphics 2007

Result is a full resolution IL

Step 5 Combine

Direct Illumination Indirect Illumination Shadows (not mentioned)

Combined Image

~280 FPS on a HD5970 @ 1280x1024for a 15x15 VPL kernelDeffered IL pass + bilateral upsampling costs ~2.5 ms

DEMO

How to add Indirect Shadows1. Use a CS and the linked lists technique

Insert blockers of IL into 3D grid of lists – let‘s use triangles for now

see backup for alternative data structure

2. Look at a kernel of VPLs again3. Only accumulate light of VPLs that are

occluded by blocker tris Trace rays through 3d grid to detect occluded VPLs Render low res buffer only

4. Subtract blocked indirect light from IL buffer Subtract a blurred version of low res version

Blur is combined bilateral blurring/upsampling

Insert tris into 3D grid of triangle lists

Scene

Rasterize dynamic

blockers to 3D grid

using a CS and

atomics

Insert tris into 3D grid of triangle lists

World space 3D grid of triangle lists

around IL blockers laid out in a UAV

(0,1,0)

eol = End of list (0xffffffff)

Scene

Rasterize dynamic

blockers to 3D grid

using a CS and

atomics

3D Grid Demo


Emitter of greenlight

Blocker of green light

Expectedindirect shadow

Blocked Indirect Light

Subtracting Blocked IL

Final Image

~300 million rays per second for Indirect Shadows

~70 FPS on a HD5970 @ 1280x1024

DEMO

Ray casting costs ~9 ms

Future directions Speed up IL rendering

Render IL at even lower res Look into multi-res RSMs

Speed up ray-tracing Per pixel array of lists for depth buckets Other data structures

Raytrace other primitive types Splats, fuzzy ellipsoids etc. Proxy geometry or bounding volumes of blockers

Get rid of Interlocked*() ops Just mark grid cells as occupied Lower quality but could work on earlier hardware Need to solve self-occlusion issues

Q&A

Holger Gruen [email protected] Thibieroz

[email protected]

Credits for the basic idea of how to implement PPLL under Direct3D 11 go to Jakub Klarowicz (Techland), Holger Gruen and Nicolas Thibieroz (AMD)

mailto:[email protected]

mailto:[email protected]

oit and indirect illumination using dx11 linked lists - amd

Documents