4. 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
5. Per-Pixel Linked Lists with Direct3D 11 Element Element Element Element Link Link Link Link Nicolas Thibieroz European ISV Relations AMD
6. 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 Element Element Element Element Link Link Link Link
7. 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
9. 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
10. 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: structFragmentAndLinkBuffer_STRUCT { FragmentData_STRUCTFragmentData; // Fragment data uintuNext; // Link to next fragment }; RWStructuredBuffer <FragmentAndLinkBuffer_STRUCT> FLBuffer;
11. 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: RWByteAddressBufferStartOffsetBuffer;
12. 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)
13. Linked List Creation (2a) Start Offset Buffer -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 Viewport -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 0 Fragment and Link Buffer Counter = Fragment and Link Buffer Fragment and Link Buffer Fragment and Link Buffer
14. Linked List Creation (2b) Start Offset Buffer -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 Viewport -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 1 0 Fragment and Link Buffer Counter = Fragment and Link Buffer -1 Fragment and Link Buffer -1 Fragment and Link Buffer
15. Linked List Creation (2c) Start Offset Buffer -1 -1 -1 -1 -1 -1 -1 0 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 Viewport -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 1 2 3 Fragment and Link Buffer Counter = Fragment and Link Buffer -1 Fragment and Link Buffer -1 -1 Fragment and Link Buffer
16. Linked List Creation (2d) Start Offset Buffer -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 0 -1 -1 -1 -1 -1 -1 Viewport -1 -1 -1 -1 -1 -1 -1 -1 -1 1 2 -1 -1 -1 -1 -1 -1 -1 4 3 5 Fragment and Link Buffer Counter = -1 -1 -1 Fragment and Link Buffer 0 -1 Fragment and Link Buffer
17. Linked List Creation - Code float PS_StoreFragments(PS_INPUT input) : SV_Target { // Calculate fragment data (color, depth, etc.) FragmentData_STRUCTFragmentData = ComputeFragment(); // Retrieve current pixel count and increase counter uintuPixelCount = FLBuffer.IncrementCounter(); // Exchange offsets in StartOffsetBuffer uintvPos = uint(input.vPos); uintuStartOffsetAddress= 4 * ( (SCREEN_WIDTH*vPos.y) + vPos.x ); uintuOldStartOffset; 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; }
19. Rendering Pixels (1) “Start Offset” Buffer and “Fragment & Link” Buffer now bound as SRV Buffer<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
20. Rendering from Linked List Start Offset Buffer -1 -1 -1 -1 -1 -1 4 -1 -1 -1 -1 3 3 -1 -1 -1 -1 -1 -1 Render Target -1 -1 -1 -1 -1 -1 -1 -1 -1 1 2 -1 -1 -1 -1 -1 -1 -1 Fragment and Link Buffer -1 -1 -1 -1 Fragment and Link Buffer 0 0 -1 Fragment and Link Buffer
21. Rendering Pixels (2) float4 PS_RenderFragments(PS_INPUT input) : SV_Target { // Calculate UINT-aligned start offset buffer address uintvPos = uint(input.vPos); uintuStartOffsetAddress = SCREEN_WIDTH*vPos.y + vPos.x; // Fetch offset of first fragment for current pixel uintuOffset = 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); }
23. 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
24. 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: structFragmentAndLinkBuffer_STRUCT { uintuPixelColor; // Packed pixel color uintuDepth; // Pixel depth uintuNext; // 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
25. 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 { ... }
26. 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
27. Sorting and Blending 0.95 0.93 0.87 0.98 Background color Temp Array Render Target PS color 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 0.98 0.87 0.95 0.93 -1 0 12 34
28. 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 intnNumPixels=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); (...)
29. 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; }
31. 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; uintuCoverage : SV_COVERAGE; }
32. Linked List Structure Almost unchanged from previously Depth is now packed into 24 bits 8 Bits are used to store coverage structFragmentAndLinkBuffer_STRUCT { uintuPixelColor; // Packed pixel color uintuDepthAndCoverage; // Depth + coverage uintuNext; // Address of next link };
33. Sample Coverage Example Pixel Center Sample Third sample is covered uCoverage = 0x04 (0100 in binary) Element.uDepthAndCoverage = ( In.vPos.z*(2^24-1) << 8 ) | In.uCoverage;
34. 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
35. Rendering Samples (2) static uint2 SortedPixels[MAX_SORTED_PIXELS]; // Parse linked list for all pixels at this position // and store them into temp array for later sorting intnNumPixels=0; while (uOffset!=0xFFFFFFFF) { // Retrieve pixel at current offset Element=FLBufferSRV[uOffset]; // Retrieve pixel coverage from linked list element uintuCoverage=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; }
38. 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
39. 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
40. Indirect Illumination w/o Indirect Shadows Draw scene g-buffer Draw Reflective Shadowmap (RSM) RSM shows the part of the scene that recieves direct light from the light source Draw Indirect Light buffer at ½ res RSM texels are used as light sources on g-buffer pixels for indirect lighting Upsample Indirect Light (IL) Draw final image adding IL
41. 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
42. Step 2 RSM needs to allow reconstruction of World/Camera space position World/Camera 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
43. Step 3 Render a ½ 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
44. Computing IL at a G-buf Pixel 1 Sum up contribution of all VPLs in the kernel
45. Computing IL at a G-buf Pixel 2 RSM texel/VPL g-buffer pixel This term is very similar to terms used in radiosity form factor computations
46. Computing IL at a G-buf Pixel 3 A naive solution for smooth IL needs to consider four VPL kernels with centers at t0, t1, t2 and t3. stx : sub RSM texel x position [0.0, 1.0[ sty : sub RSM texel y position [0.0, 1.0[
47. 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 VPL kernel at t0 stx : sub texel x position [0.0, 1.0[ VPL kernel at t2 sty : sub texel y position [0.0, 1.0[ VPL kernel at t1 VPL kernel at t3
48. 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 !
50. Step 4 Indirect Light buffer is ½ res Perform a bilateral upsampling step SeePeter-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
53. Combined Image DEMO ~280 FPS on a HD5970 @ 1280x1024 for a 15x15 VPL kernel Deffered IL pass + bilateral upsampling costs ~2.5 ms
54. How to add Indirect Shadows Use a CS and the linked lists technique Insert blocker geomety of IL into 3D grid of lists – let‘s use the triangles of the blocker for now see backup for alternative data structure Look at a kernel of VPLs again 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 Subtract blocked indirect light from IL buffer Blurred version of low res blocked IL is used Blur is combined bilateral blurring/upsampling
55. Insert tris into 3D grid of triangle lists Rasterize dynamic blockers to 3D grid using a CS and atomics Scene
56. Insert tris into 3D grid of triangle lists (0,1,0) Rasterize dynamic blockers to 3D grid using a CS and atomics World space 3D grid of triangle lists around IL blockers laid out in a UAV Scene eol = End of list (0xffffffff)
62. Final Image DEMO ~70 FPS on a HD5970 @ 1280x1024 ~300 million rays per second for Indirect Shadows Ray casting costs ~9 ms
63. 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 (see backup) 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
64. Q&A Holger Gruen holger.gruen@AMD.com Nicolas Thibieroz nicolas.thibieroz@AMD.com Credits for the basic idea of how to implement PPLL under Direct3D 11 go to Jakub Klarowicz (Techland),Holger Gruen and Nicolas Thibieroz (AMD)
66. Computing IL at a g-buf pixel 1 Want to support low res RSMs Want to create smooth indirect light Goal is bi-linear filtering of four VPL-Kernels Otherwise results don‘t look smooth
67. Computing IL at a g-buf pixel 2 stx : sub texel x position [0.0, 1.0[ sty : sub texel y position [0.0, 1.0[
68. Computing IL at a g-buf pixel 3 For smooth IL one needs to consider four VPL kernels with centers at t0, t1, t2 and t3. stx : sub texel x position [0.0, 1.0[ sty : sub texel y position [0.0, 1.0[
69. Computing IL at a g-buf pixel 4 Center at t0 VPL kernel at t0 stx : sub texel x position [0.0, 1.0[ sty : sub texel y position [0.0, 1.0[
70. Computing IL at a g-buf pixel 4 Center at t1 VPL kernel at t0 stx : sub texel x position [0.0, 1.0[ sty : sub texel y position [0.0, 1.0[ VPL kernel at t1
71. Computing IL at a g-buf pixel 5 Center at t2 stx : sub texel x position [0.0, 1.0[ VPL kernel at t2 VPL kernel at t0 sty : sub texel y position [0.0, 1.0[ VPL kernel at t1
72. Computing IL at a g-buf pixel 6 Center at t3 VPL kernel at t0 stx : sub texel x position [0.0, 1.0[ VPL kernel at t2 sty : sub texel y position [0.0, 1.0[ VPL kernel at t1 VPL kernel at t3
73. Computing IL at a g-buf pixel 7 IndirectLight = (1.0f-sty) * ((1.0f-stx) * + stx * ) + (0.0f+sty) * ((1.0f-stx) * + stx * ) Evaluation of 4 big VPL kernels is slow VPL kernel at t0 stx : sub texel x position [0.0, 1.0[ VPL kernel at t2 sty : sub texel y position [0.0, 1.0[ VPL kernel at t1 VPL kernel at t3
74. Computing IL at a g-buf pixel 8 VPL kernel at t0 stx : sub texel x position [0.0, 1.0[ VPL kernel at t2 sty : sub texel y position [0.0, 1.0[ VPL kernel at t1 VPL kernel at t3
75. Computing IL at a g-buf pixel 9 B1 B2 B0 B3 B4 B5 B6 B7 B8 VPL kernel at t0 stx : sub texel x position [0.0, 1.0[ VPL kernel at t2 sty : sub texel y position [0.0, 1.0[ VPL kernel at t1 VPL kernel at t3
76. Computing IL at a g-buf pixel 9 IndirectLight = (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 Evaluation of 7 small and 1 bigger VPL kernels is fast stx : sub texel x position [0.0, 1.0[ sty : sub texel y position [0.0, 1.0[
77. Insert Tris into 2D Map of Lists of Tris Rasterize blockers of IL from view of light Light 2D buffer Scene
78. Insert Tris into 2D Map of Lists of Tris Rasterize blockers of IL from view of light using a GS and conservative rasterization Light Scene 2D buffer of lists of triangles written to by scattering PS eol = End of list (0xffffffff)
80. Linked List Creation (2) For every pixel: Calculate pixel data (color, depth etc.) Retrieve current pixel count from Fragment & Link UAV and increment counter uintuPixelCount =FragmentAndLinkBuffer.IncrementCounter(); Swap offsets in Start Offset UAV uintuOldStartOffset; StartOffsetBuffer.InterlockedExchange( PixelScreenSpacePositionLinearAddress, uPixelCount, uOldStartOffset); Add new entry to Fragment & Link UAV FragmentAndLinkBuffer_STRUCT Element; Element.FragmentData = FragmentData; Element.uNext = uOldStartOffset; FragmentAndLinkBuffer[uPixelCount] = Element;
81. Linked List Creation (3d) Start Offset Buffer -1 -1 -1 -1 -1 -1 -1 3 4 -1 -1 -1 -1 -1 -1 -1 -1 -1 Render Target -1 -1 -1 -1 -1 -1 -1 -1 -1 1 2 -1 -1 -1 -1 -1 -1 -1 Fragment and Link Buffer Fragment and Link Buffer 0 -1 -1 -1 -1 Fragment and Link Buffer Fragment and Link Buffer
82. PPLL Pros and Cons Pros: Allows storing of variable number of elements per pixel Only one atomic operation per pixel Using built-in HW UAV counter Good performance Cons: Traversal causes a lot of random access Special case for storing MSAA samples
83. Alternative Technique:Per Pixel Array There is a two pass method that allows the construction of per-pixel arrays First described in ’Sample Based Visibility for Soft Shadows using Alias-free Shadow Maps’(E.Sintorn, E. Eisemann, U. Assarsson, EG 2008) Uses a pre-fix sum to allocate multiple entries into a buffer Elements belonging to the same location will therefore be contiguous in memory Faster traversal (less random access) Not as fast as PPLL for scenes with ‘high’ depth complexity in our testing
84. Rendering SamplesAlternate Method Writes pixels straight into non-MSAA RT Only viable if access to samples is no longer required Samples must be resolved into pixels before writing This affects the manual blending process Samples with the same coverage are blended back-to-front ...then averaged (resolved) with other samples before written out to RT This method can prove slightly faster than per-sample rendering
85. Rendering FragmentsAlternate Method (2x MSAA) // Retrieve current sample colors from background texture float4 vCurrentColor0=BackgroundTexture.Load(int3(vPos.xy, 0)); float4 vCurrentColor1=BackgroundTexture.Load(int3(vPos.xy, 1)); // Rendering pixels using SRCALPHA-INVSRCALPHA blending for (int k=0; k<nNumPixels; k++) { // Retrieve next unblended furthermost pixel float4 vPixColor= UnpackFromUint(SortedPixels[k].x); // Retrieve sample coverage uintuCoverage=UnpackCoverage(SortedPixels[k].y); // Manual blending between current samples and previous ones if (uCoverage & (1<<0)) vCurrentColor0= lerp(vCurrentColor0.xyz, vPixColor.xyz, vPixColor.w); if (uCoverage & (1<<1)) vCurrentColor1= lerp(vCurrentColor1.xyz, vPixColor.xyz, vPixColor.w); } // Return resolved and manually-blended color return (vCurrentColor0+vCurrentColor1)/0.5; }
Notes de l'éditeur
A node is composed of Element and Link
Uses a Pixel Shader 5.0 to store fragments into linked list: because we need trianglesrasterized as normal through the rendering pipeline. Briefly describe UAVs.Uses atomic operations: only 1 such operation per pixel is used (although it has feedback)
“Fragment & Link” buffer only use Write access actually.FragmentData: typically fragment color, but also transparency, depth, blend mode id, etc.Must be large enough to store all fragments: allocation must take into account maximum number of elements to store. E.g. “average overdraw per pixel” can be used for allocation: width*height*averageoverdraw*sizeof(Element&Link)Fragment data struct should be kept minimal to reduce performance requirements (and improve performance in the process by performing less memory operations)
StartOffsetBuffer is 1D buffer despite being screen-sized; this is because atomic operations are not supported on 2D UAV textures. Initialized to magic value: thus by default the linked lists are empty since they point to -1 (end of list).
“Viewport” is used to illustrate incoming triangle (as no RT is bound)Here fragment data is just the pixel color for simplification.Start Offset Buffer is initialized to “magic value” (-1)
We’re going to exchange the offset in the StartOffsetBuffer with the current value of the counter (before incrementation).This is because the current value of the counter gives us the offset (address) of the fragment we’re storing.
Assuming a rasterization order of left to right for the green triangle
Assuming a rasterization order of left to right for the yellow triangle
Current pixel count is TOTAL pixel count, not pixel count per pixel location.Not using Append Buffers because they require the DXGI_FORMAT_R32_UNKNOWN format.Times 4 because startoffsetaddress is in bytes, not UINT (1 UINT = 4 bytes).
No “4x”to calculate offset address here becauseStartOffsetBufferSRV is declared with the uint type.
No point storing transparent fragments that are hidden by opaque geometry, i.e. that fail the depth test.
May also get away with packed color and depth into the same uint!: this would make the element size a nice 64 bits.Deferred rendering engines may wish to store the same properties in F&L structure as opaque pixels, but watch out for memory and performance impact of large structure!If using different blend modes per pixel then you would need to store blend mode as id into fragment data.Depth is stored as uint – we’ll explain why later on.
Default DX11 pixel pipeline executes pixel shader before depth test[earlydepthstencil] allows the depth test to be performed *before* pixel shader.Normally the HW will try to do this for you automatically but a PS writing to UAVs cannot benefit from this treatment – a bit like alpha testing.Allows performance savings and rendering correctness!: less fragments to store = good!
Sparse memory accesses = slow!: especially true as original rendering order can mean that fragments belonging to the same screen coordinates can be (very) far apart from each other.
Under-blending can be used insteadNo need for background input thenActual HW blending mode enabledBlending mode to use with under-blending is ONE, SRCALPHA
SortedPixels[].x storescolor, SortedPixels[].y stores depthUse (nNumPixels>=MAX_SORTED_PIXELS) test to prevent overflowActual sort algorithm used can be anything (e.g. insertion sort, etc.)
This happens after sorting.vCurrentColor.xyz= vFragColor.w * vFragColor.xyz + (1.0-vFragColor.w)* vCurrentColor.xyz;
The problem with OIT via PPLL and MSAA is that fragment sorting and blending must happens on a per-sample basis, which means a lot of storage and processing.Storing individual samples into Linked Lists requires a huge amount of memory: as most pixels are not edge pixels, they would get all their samples covered. Unfortunately schemes mixing edge and non-edge pixels proved difficult to implement and did not generate satisfactory results.SV_COVERAGE isDX11 only
Depth is now packed into 24 bits: this is why depth is stored as uint
Blue triangle covers one depth sample
Assuming a rasterization order of left to right for the yellow triangle
Traversal causes a lot of random access: Elements belonging to a same pixel location can be far apart in LL!
Not as fast: probably because multiple passes involved.
Only viable if access to samples is no longer required: (e.g. for HDR-correct tone mapping)Sorting is unaffected because stored samples share the same depth.Samples with the same coverage are blended back-to-front: this means blending happens on a per-sample basis as it normally would.
Code illustration for 2 samplesSortedPixels[] contains pixel color in .x and depth+coverage in .y