Movatterモバイル変換

(unabridged slide deck)NVIDIA OpenGL in 2012: Version 4.3 is here! Mark Kilgard

Mark Kilgard Principal System Software Engineer OpenGL driver and API evolution Cg (“C for graphics”) shading language GPU-accelerated path rendering OpenGL Utility Toolkit (GLUT) implementer Author of OpenGL for the X Window System Co-author of Cg Tutorial Worked on OpenGL for 20+ years

Talk DetailsLocation: West Hall Meeting Room 503, Los Angeles Convention CenterDate: Wednesday, August 8, 2012Time: 11:50 AM - 12:50 PMMark Kilgard (Principal Software Engineer, NVIDIA)Abstract: Attend this session to get the most out of OpenGL on NVIDIA Quadro and GeForceGPUs. Learn about the new features in OpenGL 4.3, particularly Compute Shaders. Other topicsinclude bindless graphics; Linux improvements; and how to best use the modern OpenGLgraphics pipeline. Learn how your application can benefit from NVIDIA's leadership drivingOpenGL as a cross-platform, open industry standard.Topic Areas: Computer Graphics; Development Tools & Libraries; Visualization; Image andVideo ProcessingLevel: IntermediateWatch video replay: http://nvidia.fullviewmedia.com/siggraph2012/ondemand/SS104.html

Outline State of OpenGL & OpenGL’s importance to NVIDIA Compute Shaders explored Other stuff in OpenGL 4.3 Further NVIDIA OpenGL Work How to exploit OpenGL’s modern graphics pipeline

State of OpenGL &OpenGL’s importance to NVIDIA

OpenGL Standard is 20 Years and Strong

Think back to Computing in 1992Programming Languages ANSI C (C 89) was just 3 years old C++ still implemented as a front-end to C OpenGL in 1992 provided FORTRAN and Pascal bindingsOne year before NCSA Mosaic web browser first written Now WebGL standard in almost every browserWindows version Windows 3.1 ships! NT 3.1 still a year awayEntertainment Great video game graphics? Mortal Kombat? Top grossing movie (Aladdin) was animated Back when animated movies were still hand-drawn

Then and Now 2012 OpenGL 4.3: Real-time Global IlluminationOpenGL 1.0: Per-vertex lighting 1992 [Crassin]

Big News 4.3 OpenGL 4.3 announced Monday here at SIGGRAPH August 6, 2012 Moments later… NVIDIA beta OpenGL 4.3 driver on the web http://www.nvidia.com/content/devzone/opengl-driver-4.3.html OpenGL 4.3 brings substantial new features Compute Shaders! OpenGL Shading Language (GLSL) updates (multi-dimensional arrays, etc.) New texture functionality (stencil texturing, more queries)MarqueeFeature New buffer functionality (clear buffers, invalidate buffers, etc.) More Direct3D-isms (texture views, parity with DirectX compute shaders) OpenGL ES 3.0 compatibility

NVIDIA’s OpenGL Leverage GeForce Programmable Graphics (GLSL, Cg) Debugging with Tegra Parallel Nsight Quadro OptiX

Single 3D API for Every Platform Windows OS X Linux Android Solaris FreeBSD

OpenGL 3D Graphics API • cross-platform • most functional • peak performance • open standard • inter-operable • well specified & documented • 20 years of compatibility

OpenGL Spawns Closely Related Standards Congratulations: WebGL officially approved, February 2012 “The web is now 3D enabled”

Accelerating OpenGL Innovation 2004 2005 2006 2007 2008 2009 2010 2011 2012 DirectX 10.0 DirectX 10.1 DirectX 11 DirectX 9.0c OpenGL 3.1 OpenGL 3.3 +OpenGL 2.0 OpenGL 2.1 OpenGL 3.0 OpenGL 3.2 OpenGL 4.0 OpenGL 4.3 Now with OpenGL 4.1 compute • OpenGL has fast innovation + standardization shaders! - Pace is 7 new spec versions in four years - Actual implementations following specifications closely OpenGL 4.2 • OpenGL 4.3 is a superset of DirectX 11 functionality - While retaining backwards compatibility

OpenGL Today – DirectX 11 Superset Buffer and Event Interop First-class graphics + compute solution OpenGL 4.3 = graphics + compute shaders NVIDIA still has existing inter-op with CUDA / OpenCL Shaders can be saved to and loaded from binary blobs Ability to query a binary shader, and save it for reuse later Flow of content between desktop and mobile Brings ES 2.0 and 3.0 API and capabilities to desktop WebGL bridging desktop and mobile Cross platform Mac, Windows, Linux, Android, Solaris, FreeBSD Result of being an open standard

Increasing Functional Scope of OpenGL First-class Compute Shaders 4.3 Tessellation Features 4.0 Geometry Shaders 3.X Vertex and Fragment Shaders 2.X Fixed Function 1.XArguably, OpenGL 4.3is a 5.0 worthy feature-set!

Classic OpenGL State Machine From 1991-2007 * vertex & fragment processing got programmable 2001 & 2003 [source: GL 1.0 specification, 1992]

OpenGL 3.0 Conceptual Processing Flow (2008) uniform/ primitive topology, parameters transformed Legend vertex data Geometric primitive buffer objects Vertex Vertex assembly & vertices assembly primitive processing batch transformed processing programmable pixels operations type, vertex fragments vertex data attributes point, line, and polygon fixed-function filtered texels geometry operations Transform texture fragments buffer data vertex transform feedback fetches buffer objects feedback pixels in framebuffer object textures buffer objects stenciling, depth testing, primitive batch type, vertex blending, accumulation vertex indices, texture vertex attributes texture fetches buffer buffer data, objects Texture Fragment Raster unmap Framebuffer mapping fragment processing operations buffer texture Command Buffer fetches parser store pixel map buffer, pack get buffer buffer data objects image and bitmap texture fragments pixel image pixel image or unpack Pixel specification texture image buffer specification objects packing pixels to pack image OpenGL 3.0 rectangles, bitmaps Image Pixel Pixel primitive unpacking unpacked processing pixels processing copy pixels, copy texture image

Control point Patch Patch tessellation Patch evaluation (2010) assembly & processing transformed transformed generation transformed processing control points processing transformed patch patch, bivariatetessellation patch control patch domain points patch topology, evaluated patch vertextexture Legendfetches primitive topology, patch data transformed Geometric primitive Vertex Vertex vertex data programmable vertices assembly & operations assembly primitive processing pixels batch transformed processing fragments type, vertex fixed-function attributes point, line, operations filtered texels vertex data geometry and polygon buffer data vertex Transform texture fragments compute buffer transform feedback fetches objects feedback pixels in framebuffer object textures buffer primitive batch type, objects stenciling, depth testing, vertex vertex indices, texture blending, accumulation vertex attributes texture fetches buffer buffer data, objects Texture Fragment Raster unmap Framebuffer mapping fragment processing operations buffer pixel texture Command Buffer pack fetches parser store buffer map buffer, get buffer objects texture image and bitmap data pixel fragments image pixel image or unpack Pixel specification texture image packing specification buffer objects pixels to pack image OpenGL 4.0 rectangles, Image Pixel Pixel bitmaps primitive copy pixels, unpacking unpacked processing pixels processing copy texture image

Control point Patch Patch tessellation Patch evaluation (2012) assembly & processing transformed transformed generation transformed processing control points processing transformed patch patch, bivariatetessellation patch control patch domain points patch topology, evaluated patch vertextexture Legendfetches primitive topology, patch data transformed Geometric primitive Vertex Vertex vertex data programmable vertices assembly & operations assembly primitive processing pixels batch transformed processing fragments type, vertex fixed-function attributes point, line, operations filtered texels vertex data geometry and polygon buffer data vertex Transform texture fragments compute buffer transform feedback fetches objects feedback pixels in framebuffer object textures buffer primitive batch type, objects stenciling, depth testing, vertex vertex indices, texture blending, accumulation vertex attributes texture fetches buffer buffer data, objects Texture Fragment Raster unmap Framebuffer mapping fragment processing operations Command buffer Buffer texture fetches parser store pixel pack map buffer, get buffer buffer objects texture image and bitmap data Compute pixel fragments image pixel image or unpack Pixel processing specification texture image buffer packing specification objects pixels to pack image OpenGL 4.3 rectangles, Image Pixel Pixel bitmaps primitive copy pixels, unpacking unpacked processing pixels processing copy texture image

OpenGL 4.3 Processing Pipelines From Application From Application Vertex Puller Dispatch Indirect Dispatch Element Array Buffer b Buffer b Draw Indirect Buffer b Vertex Shader Image Load / Store t/b Compute Shader Tessellation Control Vertex Buffer Object b Shader Atomic Counter b Tessellation Primitive Generator Shader Storage b Tessellation Evaluation Shader Texture Fetch t/b Geometry Shader Uniform Block b Transform Feedback Transform Feedback Buffer b Legend Rasterization From Application Fixed Function Stage Fragment Shader Pixel Assembly Pixel Unpack Buffer b Programmable Stage b – Buffer Binding Raster Operations Pixel Operations Texture Image t t – Texture Binding Framebuffer Pixel Pack Pixel Pack Buffer b Arrows indicate data flow

OpenGL 4.3Compute Shadersexplored

Why Compute Shaders? particle physics Execute algorithmically general-purpose GLSL shaders Read and write uniforms and images Grid-oriented Single Program, Multiple Data (SPMD) fluid execution model with communication via shared variables behavior Process graphics data in context of the graphics pipeline Easier than interoperating with a compute API when processing ‘close to the pixel’ crowd Avoids involved “inter-op” APIs to connect OpenGL simulation objects to CUDA or OpenCL Complementary to OpenCL Gives full access to OpenGL objects (multisample buffers, ray etc.) tracing Same GLSL language used for graphic shaders In contrast to CUDA C/C++, not a full heterogonous (CPU/GPU) programming framework using full ANSI C Standard part of all OpenGL 4.3 implementations global Matches DirectX 11 functionality illumination

Compute Shader Particle System Demo Mike Bailey @ Oregon State University co-author of 2nd edition now available

OpenGL 4.3 Compute Shaders Single Program, Multiple Data (SPMD) Execution Model Mental model: “scores of threads jump into same function at once” Hierarchical thread structure Threads in Groups in Dispatches invocation work group dispatch (thread)

Single Program, Multiple Data Example Standard C Code, running single-threadedvoid SAXPY_CPU(int n, float alpha, float x[256], float y[256]){ if (n > 256) n = 256; for (int i = 0; i < n; i++) // loop over each element explicitly y[i] = alpha*x[i] + y[i];}#version 430layout(local_size_x=256) in; // spawn groups of 256 threads!buffer xBuffer { float x[]; }; buffer yBuffer { float y[]; };uniform float alpha;void main(){ int i = int(gl_GlobalInvocationID.x); if (i < x.length()) // derive size from buffer bound y[i] = alpha*x[i] + y[i];} OpenGL Compute Shader, running SPMD SAXPY = BLAS library's single-precision alpha times x plus y

Examples of Single-threaded Execution vs. SPMD Programming Systems Single-threaded Single Program, Multiple Data C/C++ CUDA C/C++ FORTRAN DirectCompute Pascal OpenCL OpenGL Compute Shaders CPU-centric, GPU-centric,hard to make multi-threaded & parallel naturally expresses parallelism

Per-Thread Local Variables Each thread can read/write variables “private” to its execution Each thread gets its own unique storage for each local variable work group local Compute Shader source code thread #1 v i int i; float v; no access to locals i++; of other threads v = 2*v + i; thread #2 v i

Special Per-thread Variables Work group can have a 1D, 2D or 3D “shape” Specified via Compute Shader input declarations Compute Shader syntax examples 1D, 256 threads: layout(local_size_x=256) in; 2D, 8x10 thread shape: layout(local_size_x=8,local_size_y=10) in; 3D, 4x4x4 thread shape: layout(local_size_x=4,local_size_y=4,local_size_z=4) in; Every thread in work group has its own invocation # Accessed through built-in variable gl_LocalInvocationID=(4,1,0) in uvec3 gl_LocalInvocationID; Think of every thread having a “who am I?” variable Using these variables, threads are expected to Index arrays Determine their flow control Compute thread-dependent computations 6x3 work group

Per-Work Group Shared Variables Any thread in a work group can read/write shared variables Typical idiom is to index by each thread’s invocation # Compute Shader a[0][0] a[0][1] a[0][2] a[0][3] source code shared float a[3][4]; a[1][0] a[1][1] a[1][2] a[1][3] unsigned int x = gl_LocalInvocationID.x a[2][0] a[2][1] a[2][2] a[2][3] unsigned int y = gl_LocalInvocationID.y no access to shared variables of a different work group a[y][x] = 2*ndx; a[y][x^1] += a[y][x]; a[0][0] a[0][1] a[0][2] a[0][3] memoryBarrierShared(); a[y][x^2] += a[y][x]; a[1][0] a[1][1] a[1][2] a[1][3] use shared memory barriers a[2][0] a[2][1] a[2][2] a[2][3] to synchronize access to shared variables

work groupsReading and WritingGlobal ResourcesIn addition to local andshared variables…Compute Shaders can alsoaccess global resources Read-only Textures Uniform buffer objects red Read-write green blue color Texture images x vertex 0 Uniform buffers y z position atomic Shader storage buffers red counters green Atomic counters image blue color Bindless buffers (within texture) x vertex 1 y Take care updating textures z z position shared read-write buffer object resources global OpenGL resources

Simple Compute Shader Let’s just copy from one 2D texture image to another… Pseudo-code: for each pixel in source image  pixels could be copied copy pixel to destination image fully in parallel How would we write this as a compute shader...

Simple Compute Shader Let’s just copy from one 2D texture image to another… #version 430 // use OpenGL 4.3’s GLSL with Compute Shaders #define TILE_WIDTH 16 #define TILE_HEIGHT 16 const ivec2 tileSize = ivec2(TILE_WIDTH,TILE_HEIGHT); layout(binding=0,rgba8) uniform image2D input_image; layout(binding=1,rgba8) uniform image2D output_image; layout(local_size_x=TILE_WIDTH,local_size_y=TILE_HEIGHT) in; void main() { const ivec2 tile_xy = ivec2(gl_WorkGroupID); const ivec2 thread_xy = ivec2(gl_LocalInvocationID); const ivec2 pixel_xy = tile_xy*tileSize + thread_xy; vec4 pixel = imageLoad(input_image, pixel_xy); imageStore(output_image, pixel_xy, pixel); }

Compiles into NV_compute_program5 Assembly!!NVcp5.0 # NV_compute_program5 assemblyGROUP_SIZE 16 16; # work group is 16x16 so 256 threadsPARAM c[2] = { program.local[0..1] }; # internal constantsTEMP R0, R1; # temporariesIMAGE images[] = { image[0..7] }; # input & output imagesMAD.S R1.xy,invocation.groupid,{16,16,0,0}.x,invocation.localid;MOV.S R0.x, c[0];LOADIM.U32 R0.x, R1, images[R0.x], 2D; # load from input imageMOV.S R1.z, c[1].x;UP4UB.F R0, R0.x; # unpack RGBA pixel into float4 vectorSTOREIM.F images[R1.z], R0, R1, 2D; # store to output imageEND

What is NV_compute_program5? NVIDIA has always provided assembly-level interfaces to GPU programmability in OpenGL NV_gpu_program5 is Shader Model 5.0 assembly And NV_gpu_program4 was for Shader Model 4.0 NV_tessellation_program5 is programmable tessellation extension NV_compute_program5 is further extension for Compute Shaders Advantages of assembly extensions Faster load-time for shaders Easier target for dynamic shader generation Allows other languages/tools, such as Cg, to target the underlying hardware Provides concrete underlying execution model You don’t have to guess if your GLSL compiles well or not

Launching a Compute Shader First write your compute shader Request GLSL 4.30 in your source code: #version 430 More on this later… Second compile your compute shader Same compilation process as standard GLSL graphics shaders… glCreateShader/glShaderSource with Compute Shader token GLuint compute_shader = glCreateShader(GL_COMPUTE_SHADER); glCreateProgram/glAttachShader/glLinkProgram (compute and graphics shaders cannot mix in the same program) Bind to your program object glUseProgram(compute_shader); Dispatch a grid of work groups dispatches a glDispatchCompute(4, 4, 3); 4x4x3 grid of work groups

Launching the Copy Compute ShaderSetup for copying from source to destination texture Create an input (source) texture object glTextureStorage2DEXT(input_texobj, GL_TEXTURE_2D, 1, GL_RGBA8, width, height); OpenGL 4.2 or glTextureSubImage2DEXT(input_texobj, GL_TEXTURE_2D, ARB_texture- /*level*/0, /*x,y*/0,0, width, height, _storage plus GL_RGBA, GL_UNSIGNED_BYTE, image); EXT_direct_state_access Create an empty output (destination) texture object glTextureStorage2DEXT(output_texobj, GL_TEXTURE_2D, 1, GL_RGBA8, width, height); Bind level zero of both textures to texture images 0 and 1 GLboolean is_not_layered = GL_FALSE; glBindImageTexture(/*image*/0, input_texobj, /*level*/0, OpenGL 4.2 or is_not_layered, /*layer*/0, GL_READ_ONLY, GL_RGBA8); ARB_shader- glBindImageTexture( /*image*/1, output_texobj, /*level*/0, _image- is_not_layered, /*layer*/0, GL_READ_WRITE, GL_RGBA8); _load_store Use the copy compute shader glUseProgram(compute_shader);Dispatch sufficient work group instances of the copy compute shaderglDispatchCompute((width + 15) / 16, (height + 15) / 16), 1); OpenGL 4.3

Copy Compute Shader Execution Input (source) image Output (destination) image

Copy Compute Shader Tiling gl_WorkGroupID=[x,y] [0,4] [1,4] [2,4] [3,4] [4,4] [0,4] [1,4] [2,4] [3,4] [4,4] [0,3] [1,3] [2,3] [3,3] [4,3] [0,3] [1,3] [2,3] [3,3] [4,3] [0,2] [1,2] [2,2] [3,2] [4,2] [0,2] [1,2] [2,2] [3,2] [4,2] [0,1] [1,1] [2,1] [3,1] [4,1] [0,1] [1,1] [2,1] [3,1] [4,1] [0,0] [1,0] [2,0] [3,0] [4,0] [0,0] [1,0] [2,0] [3,0] [4,0] Input (source) image 76x76 Output (destination) image 76x76

Next Example: General Convolution Discrete convolution: common image processing operation Building block for blurs, sharpening, edge detection, etc. Example: 5x5 convolution (N=5) of source (input) image s Generates destination (output) image d, given NxN matrix of weights w

Output Image after 5x5 Gaussian Blur sigma=2.0

Implementing a General Convolution Basic algorithm Tile-oriented: generate MxM pixel tiles So operating on a (M+2N)x(M+2N) region of the image Phase 1: Read all the pixels for a region from input image Phase 2: Perform weighted sum of pixels in [-N,N]x[-N,N] region around each output pixel Phase 3: Output the result pixel to output image

General Convolution: Preliminaries// Various kernel-wide constantsconst int tileWidth = 16, tileHeight = 16;const int filterWidth = 5, filterHeight = 5;const ivec2 tileSize = ivec2(tileWidth,tileHeight);const ivec2 filterOffset = ivec2(filterWidth/2,filterHeight/2);const ivec2 neighborhoodSize = tileSize + 2*filterOffset;// Declare the input and output images.layout(binding=0,rgba8) uniform image2D input_image;layout(binding=1,rgba8) uniform image2D output_image;uniform vec4 weight[filterHeight][filterWidth];uniform ivec4 imageBounds; // Bounds of the input image for pixel coordinate clamping.void retirePhase() { memoryBarrierShared(); barrier(); }ivec2 clampLocation(ivec2 xy) { // Clamp the image pixel location to the image boundary. return clamp(xy, imageBounds.xy, imageBounds.zw);}

General Convolution: Phase 1layout(local_size_x=TILE_WIDTH,local_size_y=TILE_HEIGHT) in;shared vec4 pixel[NEIGHBORHOOD_HEIGHT][NEIGHBORHOOD_WIDTH];void main() { const ivec2 tile_xy = ivec2(gl_WorkGroupID); const ivec2 thread_xy = ivec2(gl_LocalInvocationID); const ivec2 pixel_xy = tile_xy*tileSize + thread_xy; const uint x = thread_xy.x; const uint y = thread_xy.y; // Phase 1: Read the image's neighborhood into shared pixel arrays. for (int j=0; j<neighborhoodSize.y; j += tileHeight) { for (int i=0; i<neighborhoodSize.x; i += tileWidth) { if (x+i < neighborhoodSize.x && y+j < neighborhoodSize.y) { const ivec2 read_at = clampLocation(pixel_xy+ivec2(i,j)-filterOffset); pixel[y+j][x+i] = imageLoad(input_image, read_at); } } } retirePhase();

General Convolution: Phases 2 & 3 // Phase 2: Compute general convolution. vec4 result = vec4(0); for (int j=0; j<filterHeight; j++) { for (int i=0; i<filterWidth; i++) { result += pixel[y+j][x+i] * weight[j][i]; } } // Phase 3: Store result to output image. imageStore(output_image, pixel_xy, result);}

Separable Convolution Many important convolutions expressible in “separable” form More efficient to evaluate Allows two step process: 1) blur rows, then 2) blur columns Two sets of weights: column vector weights c and row vector weights r Practical example for demonstrating Compute Shader shared variables…

Example Separable Convolutions Original Original Original

Example Separable ConvolutionsGaussian filter, sigma=2.25 Sobel filter, horizontal Sobel filter, vertical

Example Separable Convolutions Weights0.026232 0.035279 0.038941 0.035279 0.026232 -1 0 1 -1 2 10.0352790.038941 0.047446 0.052371 0.052371 0.057807 0.047446 0.052371 0.035279 0.038941 -2 0 2 0 0 00.035279 0.047446 0.052371 0.047446 0.0352790.026232 0.035279 0.038941 0.035279 0.026232 -1 0 1 -1 2 1 = = =0.1619640.2178200.240432 1 1 0.161964 0.217820 .240432 0.217820 0.1619640.2178200.161964 2 -1 0 1 -1 0 1 2 1 1 5x5 Gaussian filter, sigma=2.25 Sobel filter, horizontal Sobel filter, vertical

GLSL Separable Filter Implementation<< assume preliminaries from earlier general convolution example>>layout(local_size_x=TILE_WIDTH,local_size_y=NEIGHBORHOOD_HEIGHT) in;shared vec4 pixel[NEIGHBORHOOD_HEIGHT][NEIGHBORHOOD_WIDTH]; // values read from input imageshared vec4 row[NEIGHBORHOOD_HEIGHT][TILE_WIDTH]; // weighted row sumsvoid main() // separable convolution{ const ivec2 tile_xy = ivec2(gl_WorkGroupID); const ivec2 thread_xy = ivec2(gl_LocalInvocationID); const ivec2 pixel_xy = tile_xy*tileSize + (thread_xy-ivec2(0,filterOffset.y)); const uint x = thread_xy.x; const uint y = thread_xy.y; // Phase 1: Read the image's neighborhood into shared pixel arrays. for (int i=0; i<NEIGHBORHOOD_WIDTH; i += TILE_WIDTH) { if (x+i < NEIGHBORHOOD_WIDTH) { const ivec2 read_at = clampLocation(pixel_xy+ivec2(i-filterOffset.x,0)); pixel[y][x+i] = imageLoad(input_image, read_at); } } retirePhase();

GLSL Separable Filter Implementation // Phase 2: Weighted sum the rows horizontally. row[y][x] = vec4(0); for (int i=0; i<filterWidth; i++) { row[y][x] += pixel[y][x+i] * rowWeight[i]; } retirePhase(); // Phase 3: Weighted sum the row sums vertically and write result to output image. // Does this thread correspond to a tile pixel? // Recall: There are more threads in the Y direction than tileHeight. if (y < tileHeight) { vec4 result = vec4(0); for (int i=0; i<filterHeight; i++) { result += row[y+i][x] * columnWeight[i]; } // Phase 4: Store result to output image. const ivec2 pixel_xy = tile_xy*tileSize + thread_xy; imageStore(output_image, pixel_xy, result); }}

Compute Shader Median Filter Simple idea “For each pixel, replace it with the median-valued pixel in its NxN neighborhood” Non-linear, good for image enhancement through noise reduction Expensive: naively, requires lots sorting to find median Very expensive when the neighborhood is large Reasonably efficient with Compute Shaders

Median Filter Example Noisy appearance in candy Original

Median Filter Example Noisy lost in blur But text is blurry too Gaussian 5x5 blur

Median Filter Example Noisy gone Text still sharp Median filter 5x5

Large Median Filters for Impressionistic Effect Original 7x7 Estimated Median Filter

OpenGL Evolves ModularlyEach core revision is specified as a set of extensions 4.3 Example: ARB_compute_shader Puts together all the functionality for compute shaders ARB_compute_shader Describe in its own text file ARB_ES3_compatibility May have dependencies on other extensions many more … Dependencies are stated explicitlyA core OpenGL revision (such as OpenGL 4.3) “bundles” a set ofagreed extensions—and mandates their mutual support Note: implementations can also “unbundle” ARB extensions for hardware unable to support the latest core revisionSo easiest to describe OpenGL 4.3 based on its bundledextensions…

OpenGL 4.3 debugging support ARB_debug_output OpenGL can present debug information back to developer ARB_debug_output2 Easier enabling of debug output ARB_debug_group Hierarchical grouping of debug tagging ARB_debug_label Label OpenGL objects for debugging

OpenGL 4.3 new texture functionality ARB_texture_view Provide different ways to interpret texture data without duplicating the texture Match DX11 functionality ARB_internalformat_query2 Find out actual supported limits for most texture parameters ARB_copy_image Direct copy of pixels between textures and render buffers ARB_texture_buffer_range Create texture buffer object corresponding to a sub-range of a buffer’s data store ARB_stencil_texturing Read stencil bits of a packed depth-stencil texture ARB_texture_storage_multisample Immutable storage objects for multisampled textures

OpenGL 4.3 new buffer functionalityARB_shader_storage_buffer_object Enables shader stages to read & write to very large buffers NVIDIA hardware allows every shader stage to read & write structs, arrays, scalars, etc.ARB_invalidate_subdata Invalidate all or some of the contents of textures and buffersARB_clear_buffer_object Clear a buffer object with a constant valueARB_vertex_attrib_binding Separate vertex attribute state from the data stores of each arrayARB_robust_buffer_access_behavior Shader read/write to an object only allowed to data owned by the application Applies to out of bounds accesses

OpenGL 4.3 new pipeline functionalityARB_compute_shader Introduces new shader stage Enables advanced processing algorithms that harness the parallelism of GPUsARB_multi_draw_indirect Draw many GPU generated objects with one callARB_program_interface_query Generic API to enumerate active variables and interface blocks for each stage Enumerate active variables in interfaces between separable program objectsARB_ES3_compatibility features not previously present in OpenGL Brings EAC and ETC2 texture compression formatsARB_framebuffer_no_attachments Render to an arbitrary sized framebuffer without actual populated pixels

GLSL 4.3 new functionality ARB_arrays_of_arrays Allows multi-dimensional arrays in GLSL. float f[4][3]; ARB_shader_image_size Query size of an image in a shader ARB_explicit_uniform_location Set location of a default-block uniform in the shader ARB_texture_query_levels Query number of mipmap levels accessible through a sampler uniform ARB_fragment_layer_viewport gl_Layer and gl_ViewportIndex now available to fragment shader

New KHR and ARB extensions Not part of core but important and standardized at same time as OpenGL 4.3… KHR_texture_compression_astc_ldr Adaptive Scalable Texture Compression (ASTC) 1-4 component, low bit rate < 1 bit/pixel – 8 bit/pixel ARB_robustness_isolation If application causes GPU reset, no other application will be affected For WebGL and other un-trusted 3D content sources

Getting at OpenGL 4.3 Easiest approach… Use OpenGL Extension Wrangler (GLEW) Release 1.9.0 already has OpenGL 4.3 support http://glew.sourceforge.net

Further NVIDIA OpenGL Work Linux enhancements Path Rendering for Resolution-independent 2D graphics Bindless Graphics Commitment to API Compatibility

OpenGL-relatedLinux Improvements  Support for X Resize, Rotate, and Reflect Extension  Also known as RandR  Version 1.2 and 1.3  OpenGL enables, by default, “Sync to Vertical Blank”  Locks your glXSwapBuffers to the monitor refresh rates  Matches Windows default now  Previously disabled by default

OpenGL-relatedLinux Improvements Expose additional full-scene antialiasing (FSAA) modes 16x multisample FSAA on all GeForce GPUs 2x2 supersampling of 4x multisampling Ultra high-quality FSAA modes for Quadro GPUs 32x multisample FSAA – 2x2 supersampling of 8x multisampling 64x multisample FSAA – 4x4 supersampling of 4x multisampling Coverage sample FSAA on GeForce 8 series and better 4 color/depth samples + 12 depth samples

Multisampling FSAA Patterns aliased 2x multisampling 4x multisampling 8x multisampling 1 sample/pixel 2 samples/pixel 4 samples/pixel 8 samples/pixel 64 bits/pixel 128 bits/pixel 256 bits/pixel 512 bits/pixel Assume: 32-bit RGBA + 24-bit Z + 8-bit Stencil = 64 bits/sample

Supersampling FSAA Patterns 2x2 supersampling 2x2 supersampling 4x4 supersampling of 4x multisampling of 8x multisampling of 16x multisampling 16 samples/pixel 32 samples/pixel 64 samples/pixel 1024 bits/pixel 2048 bits/pixel 4096 bits/pixel Quadro GPUs Assume: 32-bit RGBA + 24-bit Z + 8-bit Stencil = 64 bits/sample

Image Quality EvolvedNVIDIA Fast Approximated Anti-Alias (FXAA) Supported on Windows for several driver releases… Now enabled for Linux in 304.xx drivers

NVIDIA X Server Settings for LinuxControl Panel

GLX Protocol Network transparent OpenGL Run OpenGL app on one machine, display the X and 3D on a different machine 3D app X server GLX OpenGL Server GLX Client OpenGL network connection

OpenGL-relatedLinux Improvements Official GLX Protocol support for OpenGL extensions ARB_half_float_pixel EXT_point_parameters ARB_transpose_matrix EXT_stencil_two_side EXT_blend_equation_separate EXT_depth_bounds_test NV_copy_image EXT_framebuffer_blit NV_depth_buffer_float EXT_framebuffer_multisample NV_half_float EXT_packed_depth_stencil NV_occlusion_query NV_point_sprite NV_register_combiners2 NV_texture_barrier

OpenGL-relatedLinux Improvements Tentative GLX Protocol support for OpenGL extensionsARB_map_buffer_range EXT_vertex_attrib_64bitARB_shader_subroutine NV_conditional_renderARB_stencil_two_side NV_framebuffer_multisample_coverageEXT_transform_feedback2 NV_texture_barrier NV_transform_feedback2

Synchronizing X11-based OpenGL Streams New extension GL_EXT_x11_sync_object Bridges the X Synchronization Extension with OpenGL 3.2 “sync” objects (ARB_sync) Introduces new OpenGL command GLintptr sync_handle; GLsync glImportSyncEXT (GLenum external_sync_type, GLintptr external_sync, GLbitfield flags); external_sync_type must be GL_SYNC_X11_FENCE_EXT flags must be zero

Other Linux Updates GL_CLAMP behaves in conformant way now Long-standing work around for original Quake 3 Enabled 10-bit per component X desktop support GeForce 8 and better GPUs Support for 3D Vision Pro stereo now

What is 3D Vision Pro? For Professionals All of 3D Vision support, plus • Radio frequency (RF) glasses, Bidirectional • Query compass, accelerometer, battery • Many RF channels – no collision • Up to ~120 feet • No line of sight needed to emitter • NVAPI to control

NV_path_rendering An NVIDIA OpenGL extension GPU-accelerates resolution- independent 2D graphics Very fast! Supports PostScript-style renderingCome to my afternoon talk tolearn more “GPU-Accelerated 2D and Web Rendering” This room 2:40 PM - 3:40 PM

Pixel pipeline Vertex pipeline Path pipeline Application Path specificationPixel assembly Vertex assembly Transform path (unpack) Vertex operations transform feedback Primitive assemblyPixel operations Primitive operations Fill/Stroke Covering Pixel pack Rasterization read Texture Fragment operations back memory Fill/Stroke Application Raster operations Stenciling Framebuffer Display

NVIDIA’s Vision of Bindless Graphics Problem: Binding to different objects (textures, buffers) takes a lot of validation time in driver And applications are limited to a small palette of bound buffers and textures Approach of OpenGL, but also Direct3D Solution: Exposes GPU virtual addresses Let shaders and vertex puller access buffer and texture memory by its virtual address! Kepler GPUs support bindless texture

Prior to Bindless Graphics Traditional OpenGL GPU memory reads are “indirected” through bindings Limited number of texture units and vertex array attributes glBindTexture—for texture images and buffers glBindBuffer—for vertex arrays

Buffer-centric Evolution Data moves onto GPU, away from CPU Apps on CPUs just too slow at moving data otherwise Array Element Buffer glBegin, glDrawElements, etc. Object (VeBO) Texture Buffer Vertex Array Buffer Object Object (TexBO) Vertex Puller (VaBO) texel data Transform Feedback Buffer (XBO) Vertex Shading Pixel Unpack vertex data Buffer (PuBO) Texturing Geometry Shading Parameter Buffer glDrawPixels, glTexImage2D, etc. Object (PaBO) Pixel Pixel Pack Buffer Fragment Pipeline (PpBO) Uniform Buffer Shading glReadPixels, pixel data Object (UBO) etc. parameter data Framebuffer

Kepler – Bindless Textures Enormous increase in the number of unique textures available to shaders at run-time More different materials and richer texture detail in a scene texture #0 Shader code texture #1 Shader code texture #2 … texture #127 … Pre-Kepler texture binding model Kepler bindless textures over 1 million unique textures

Kepler – Bindless Textures Pre-Kepler texture binding model Kepler bindless textures CPU CPU Load texture A Load textures A, B, C Load texture B Draw() Load texture C Bind texture A to slot I GPU Bind texture B to slot J Read from texture A Draw() Read from texture B Read from texture C GPU Read from texture at slot I Read from texture at slot J CPU Bind texture C to slot K Bindless model reduces CPU Draw() overhead and improves GPU access GPU efficiency Read from texture at slot K

Bindless Textures Apropos for ray-tracing and advanced rendering where textures cannot be “bound” in advance Shader code

Bindless performance benefit Numbers obtained with a directed test

More Information on Bindless Texture Kepler has new NV_bindless_texture extension Texture companion to NV_vertex_buffer_unified_memory for bindless vertex arrays NV_shader_buffer_load for bindless shader buffer reads NV_shader_buffer_store (also NEW) for bindless shader buffer writes API specification publically available http://developer.download.nvidia.com/opengl/specs/GL_NV_bindless_texture.txt

API Usage to Initialize Bindless Texture Make a conventional OpenGL texture object With a 32-bit GLuint name Query a 64-bit texture handle from 32-bit texture name GLuint64 glGetTextureHandleNV(GLuint); Make handle resident in context’s GPU address space void glMakeTextureHandleResidentNV(GLuint64);

Writing GLSL for Bindless Textures Request GLSL to understand bindless textures #version 400 // or later #extension GL_NV_bindless_texture : require Declare a sampler in the normal way in sampler2D bindless_texture; Alternatively, access bindless samplers in big array: uniform Samplers { sampler2D lotsOfSamplers[256]; } Exciting: 256 samplers exceeds the available texture units!

Update Sampler Uniforms withBindless Texture Handle Get a location for a sampler or image uniform GLint loc = glGetUniformLocation(program, “bindless_texture”); GLint loc_array = glGetUniformLocation(program, “lotsOfSamplers”); Then set sampler to the bindless texture handle glProgramUniformHandleui64NV(program, location, 1, &bindless_handle);

NVIDIA’s Position on OpenGL Deprecation:Core vs. Compatibility Profiles OpenGL 3.1 introduced notion of Best advice for real developers “core” profile Simply use the “compatibility” Idea was remove stuff from core to profile make OpenGL “good-er” Easiest course of action Requesting the core profile requires Well-intentioned perhaps but… special context creation gymnastics Throws API backward Avoids the frustration of “they decided compatibility out the window to remove what??” Lots of useful functionality got Allows you to use existing OpenGL libraries and code as-is removed that is in fast hardware Examples: Polygon mode, line width, GL_QUADS No, your program won’t go faster for using the “core” profile Lots of easy-to-use, effective API It may go slower because of extra “is got labeled deprecated this allowed to work?” checks Immediate mode Display lists Nothing changes with OpenGL 4.3 NVIDIA still committed to compatibility without compromise

How to exploit OpenGL’smodern graphics pipeline Albrecht Dürer’s less-than-modern rendering approaches

Modern OpenGL Pipeline Ideas Case Study: CAD assemblies Geometry complexity less problematic than scene complexity Hardware can render billions of triangles, but doesn‘t like spoon feeding Many parts for individual pieces / geometry features (bevels, chamfers, joints...) Must remain addressable as individual to select, colorize, hide, transform... Need to lower CPU overhead so that GPU can plow through large chunks of work models courtesy of PTC

Concepts Minimize CPU/GPU interaction Allow GPU to update its own data Lower api usage when scene is changed little Avoid data redundancy Data stored once on GPU, referenced multiple time Update only once (less host to gpu transfers) Increase batching potential Further cuts api calls Less driver CPU work

Data Organization Large Scene Buffers Matrices, Bounding Boxes, data intended for culling and drawing Grouped Content Buffers Materials belonging to same shader, lights, view... When not using bindless, higher numbers might improve batching, but also more “work” to organize data. Draw Command Buffer Scratch buffer that is sometimes rebuilt Hosts data of all objects or references of “active objects” OpenGL 4.x allows most data to be stored and consumed on GPU (draw indirect)

OpenGL Technology GL 3.x uniform samplerBuffer matrixBuffer; // need helper functions Texture Buffers (TexBO, unsized mat4 getMatrix (samplerBuffer buf, int i){ return mat4( texelFetch (buf,(i*4)+0), 1D array of basic vector types) texelFetch (buf,(i*4)+1) ... Uniform Buffer Objects (UBO, } arbitrary types, size limitation uniform viewBuffer { 64kb) mat4 viewInvTM; mat4 viewProjTM; Texture Arrays (pack multiple float time; same-sized textures in one array) ... } GL 4.x // NEW 4.3 allows unsized arrays as last entry Shader Storage Buffer (SSBO, // and tighter array packing layout(std430) buffer matrixBuffer { unsized arbitrary buffer access) int willCostOnly16BytesNow[4]; (NEW 4.3) mat4 matrices[]; }

OpenGL Technology GL 4.x glGenTextures (2,tex); // create texture with complete mipchain ARB_texture_storage helps driver glTexStorage (GL_..,levels, GL_RGBA8, w, h); // subimage data in later to create immutable “complete” glTexSubImage (GL_.., 0,..., mipData[0]) texture at once // NEW 4.3 // create another texture that references the // same data and interprets a single mip ARB_texture_view (NEW 4.3) // slightly differently allows multiple views (internal- glTextureView (tex[1], GL_TEXTURE_2D, tex[2], format casting with same texel bit GL_R32UI, minlevel, numlevels, minlayer, numlayers); size) on same texture data // NEW 4.3 bind range of buffer ARB_texture_buffer_range (NEW glTexBufferRange (GL_..., GL_RGBA32F, buffer, offset, size); 4.3)

NVIDIA Technology // GLSL with true pointers NVIDIA Bindless Graphics uniform mat4* matrixBuffer; Exposes gpu resources directly // API glUniformui64NV (shd->matrixLocation, (pointers or objects) scene->matrixADDR); Reduces CPU cache thrashing greatly mat->diffuse = glGetTextureHandleNV (texobj); GL 3.x // later instead of glBindTexture NV_shader_buffer_load (SBL) for glUniformHandleui64NV (shd->diffuseLocation, arbitrary unsized cross buffer access mat->diffuse) // GLSL NV_vertex_buffer_unified_memory // can also store textures in resources, (VBUM) separates vertex data from // virtually no restrictions on # format uniform materialBuffer { sampler2D howManyTexturesIWant[LARGE]; GL 4.x } // virtual sparse texturing NV_bindless_texture allows sampler uniform usampler2D virtualTex; references anywhere ... sampler2D (packUint2x32 ( texelFetch (virtualTex, coord).xy));

Data Transfer // NEW 4.3 copy rectangles of textures Textures glCopyImageSubData ( srcName, srcTarget, srcLevel,srcX,srcY,srcZ, ARB_pixel_buffer_object (GL 2.x) dstName, dstTarget, dstLevel,dstX,dstY,dstZ, srcWidth, srcHeight, srcDepth); Now ARB_copy_image (NEW 4.3) Buffers // EXT_direct_state_access style usage shown // classic functions exists as well ARB_map_buffer_range (GL 3.x) for // range map and invalidate fast mapping void* data; ARB_invalidate_subdata (NEW 4.3) data = glMapNamedBufferRangeEXT (textBuffer, 0, sizeof(MyChar) * textLength, ARB_clear_buffer_object (NEW 4.3) GL_MAP_INVALIDATE_RANGE_BIT | GL_MAP_UNSYNCHRONIZED_BIT); allows memset() operations ARB_sync (GL 3.x) for efficient // NEW 4.3 clearing a buffer GLuint zero[1] = {0}; threaded streaming glClearNamedBufferDataEXT (visibleBuffer, GTC 2012 “Optimizing Texture Transfers” GL_R32UI, GL_RED, GL_UNSIGNED_INT, zero);

Drawing the Objects /* setup once, similar to glVertexAttribPointer but with relative offset last */ Classic: bind buffers/textures glVertexAttribFormat (ATTR_NORMAL, 3, and draw GL_FLOAT, GL_TRUE, offsetof(Vertex,normal)); glVertexAttribFormat (ATTR_POS, 3, NVIDIA Bindless Graphics allows GL_FLOAT, GL_FALSE, offsetof(Vertex,pos)); // bind to stream very fast switching glVertexAttribBinding (ATTR_NORMAL, 0); Group by bindings glVertexAttribBinding (ATTR_POS, 0); Enhanced: // switch single stream buffer glBindVertexBuffer (0, bufID, 0, sizeof(Vertex)); ARB_vertex_attrib_binding (NEW // NV_vertex_buffer_unified_memory 4.3): allows less buffer changes // enable once and set stride Similar to VBUM it separates glEnableClientState (GL_VERTEX...NV);... format from data glBindVertexBuffer (0, 0, 0, sizeof(Vertex)); Map multiple vertex attributes to // switch buffer one buffer glBufferAddressRangeNV (GL_VERTEX...,0,bufADDR, bufSize);

Drawing the Objects struct MyMaterial { Enhanced: vec4 diffuse; int shadeType; Grow buffers and dynamically index ... TexBO: can be large, but ugly to fetch }; uniform materialBuffer { UBO: fast, but size limited MyMaterial materials[128]; SSBO: large }; buffer transformBuffer { mat4 transforms[]; Pass assignment index as glUniform , }; glVertexAttribI (faster) ... gl_FragColor = materials[assign.x].diffuse; Go bindless beyond VBUM // bindless pointer datatypes struct Object { Bypassing binding completely MyMaterial* material; mat4* transform; };

Draw Call Reduction MultiDraw Render ranges from current VBO/IBO Single draw call for many distinct objects Reduces overhead for low complexity objects DrawIndirect DrawElementsIndirect Store drawcall information on GPU as well { GLuint count; (primitiveCount...) GLuint instanceCount; Let GPU create/modify such buffers to GLuint firstIndex; GLint baseVertex; generate frame‘s drawcall buffers GLuint baseInstance; } MultiDrawIndirect (NEW 4.3)

Drawing the Objects Combine drawcalls with MultiDraw 1111 00000000 2 3 44444 1 How to find object‘s transform, material... assignment? in vec4 oPos; GL 3.x sacrifice a vertex attribute in int objID; flat out ivec4 assigns; Inside main vertex buffer encode uniform isamplerBuffer assignBuffer; object index uniform samplerBuffer matrixBuffer; ... Fetch assign indices from assigns = texelFetch (assignBuffer, samplerBuffer objID); Matrix/material ... assignments independent of geometry data worldTM = getMatrix (matrixBuffer, assigns.x); GL 4.x MultiDrawIndirect exposes vec4 wPos = worldTM * oPos; ... “baseInstance” to get assignments

BaseInstance as Unique Object ID MultiDrawIndirect uses instanced drawing Can replicate a vertex attribute (material... assignment) Regular : VArray[ gl_VertexID + baseVertex ] Divisor != 0 : VArray[ gl_InstanceID / VAttribDivisor + baseInstance ] AssignBuffer (divisor) Combined Attributes VertexBuffer (regular) MultiDrawBuffer count = 9 count = 6 baseVertex = 0 baseVertex = 9 baseInstance = 1 baseInstance = 0

Recap Addressed: Can render many low complexity objects stored in same vbo Buffers with indexable content to lower overhead MultiDraw/Indirect for keeping objects independent baseInstance to provide unique index/assignments NVIDIA Bindless Graphics Instead of passing “index“ can pass pointer to buffer Buffer can store texture references Lowers CPU work even further for hot loop What remains: State and Shader switching (can’t do much about state, but...)

Shader Switching subroutine void shade_fn (); Could use indexed subroutines subroutine (shade_fn) vec4 metal() ... subroutine (shade_fn) vec4 wood () ... If shaders are similar in resource // content of array set by gl api call consumption (register usage) subroutine uniform shade_fn shadeFuncs[2]; might want to combine them (GL flat in ivec4 assigns; 4.x) void main(){ gl_FragColor = shadeFuncs[assigns.y](); } Initiliaze the subroutine array // bindless pointer casting and texture sampling once, then dynamically index vec4 metal() { MetalParams* metal = packPtr (assigns.zw); ... texture (metal->roughnessMap, uv); } NVIDIA Bindless Graphics vec4 wood() { pointers allow casting buffer WoodParams* metal = packPtr (assigns.zw); addresses ... }

Let the GPU do More Work So far CPU still responsible for most decision making and hot loop (Multi) DrawIndirect allows GPU to generate its own work Scene (objects, bboxs, materials..) described in buffers / indices / pointers Process scene data and build command lists for active objects E.g do culling, LOD picking, selection highlighting...

OpenGL Computing // API - COMPUTE glDispatchCompute (gx, gy, gz); // GLSL shared float s_mem[SOMESIZE]; ... s_mem[gl_LocalInvocationIndex] = ... ARB_compute_shader (NEW 4.3) // API - FBO Dispatch threads with shared memory glBindFrameBuffer (...); support (as in CUDA/CL) glFramebufferParameteri (..., GL_FRAMEBUFFER_DEFAULT_WIDTH, 2048); Access to ALL resources, textures, no glFramebufferParameteri (..., interop, all in GLSL GL_FRAMEBUFFER_DEFAULT_HEIGHT, 2048); glDrawArrays(...) NV_BINDLESS benefit from pointer acccess // GLSL ... imageStore(...ivec2(gl_FragCoord),.); ARB_framebuffer_no_attachments (NEW 4.3) // API - XFB glEnable (GL_RASTERIZER_DISCARD); Use rasterizer to spawn threads and glDrawArrays (GL_POINTS,0, count); SSBO/imageStores to record your results // GLSL buffer indirectBuffer { GL 3.x Transform Feedback (XFB) DrawIndirect commands[]; } Allows simple 1D Kernels ... commands[gl_VertexID].instanceCount = visible ? 1 : 0;

Culling Processing Matrix and bounding box (bbox) buffer, object buffer (which matrix to use with which bbox) XFB or “invisible” rendering to create output buffer Key: Single draw call for ALL active objects! No state changes Results “Readback” GPU to Host Can use XFB to pack into a single bit stream for all active objects 0,1,0,1,1,1,0,0,0 “Indirect” GPU to GPU Set DrawIndirect‘s instanceCount to 0 or 1

Culling Techniques Frustum (GL 3.x) XFB, VertexShader output 1 or 0 VertexAttributes are bbox index, matrix index, data fetched via TBO alternatively can feed bboxes directly, 2x vec3) HiZ Occlusion (GL 3.x) Depth-Pass (useful for fragment bound scenes anyway) Create mipmap pyramid, MAX depth XFB, VertexShader Compare object‘s closest clipspace bbox against z value of depth mip Projected size determines Mip level chosen by clipspace 2D area depth mip level  mip texel covers object

Culling Techniques Passing bbox fragments Raster Occlusion (GL 4.x) enable object Depth-Pass Raster “invisble” bounding boxes // GLSL fragment shader Geometry Shader to create the 3 sides // from ARB_shader_image_load_store depth buffer discards occluded fragments layout(early_fragment_tests) in; Fragment Shader does visible[objindex] = 1 buffer indirectBuffer { Temporal Coherence (vertex-bound) }; DrawIndirect commands[]; Render last visible flat in int objID; Test all bboxes against current depth void main(){ commands[objID].instanceCount = 1; Render newly added visible: (~last) & (visible) } Each object drawn only once // some other shader would have // cleared to 0 before

Realtime Global Illumination Octree-Based Sparse Voxelization for Real-Time Global Illumination Technique by Cyril Crassin et al. (GTC 2012, also presents here at SIGGRAPH) http://www.seas.upenn.edu/~pcozzi/OpenGLInsights/OpenGLInsights-SparseV Uses only OpenGL! Generates voxelization of the scene as well as tracing it for global effects (indirect lighting, glossy reflections)

Realtime Global Illumination Octree management NV_shader_buffer_load Pointers allow efficiently to manage access to the octree memory cells Casting is also possible to interpret the data for node-type easily Not limited to buffer bindings, can access many buffers at once pool Linear Memory Octree 8 9 6 7 4 5 2 3 1

Realtime Global Illumination Scene Voxelization ARB_atomic_counter to generate work queues (Draw/Dispatch)Indirect to construct tree asynchronously with glMemoryBarrier providing dependency information Attachment-less FBO used to rasterize triangles to voxels Material attributes (color, normal) contribute to a voxel cell (NV_shader_atomic_float) NV_shader_buffer_store, ARB_shader_image_load_store to write to voxels/octree cells

Don’t Forget the 20th Anniversary Party Date: August 8th 2012 ( today! ) Location: JW Marriott Los Angeles at LA Live Venue: Gold Ballroom – Salon 1

Other OpenGL-related NVIDIA Sessions at SIGGRAPHGPU-Accelerated 2D and Web Rendering Wednesday in West Hall 503 (this room), 2:40 PM - 3:40 PM Mark Kilgard, Principal Software Engineer, NVIDIAGPU Ray Tracing and OptiX Wednesday in West Hall 503, 3:50 PM - 4:50 PM David McAllister, OptiX Manager, NVIDIA Phillip Miller, Director, Workstation Software Product Management, NVIDIAVoxel Cone Tracing & Sparse Voxel Octree for Real-time Global Illumination Wednesday in NVIDIA Booth, 3:50 PM - 4:50 PM Cyril Crassin, Postdoctoral Research Scientist, NVIDIA ResearchOpenSubdiv: High Performance GPU Subdivision Surface Drawing Wednesday in NVIDIA Booth, 3:00 PM - 3:30 PM Thursday in NVIDIA Booth, 10:00 AM - 10:30 AM (2nd time) Pixar Animation Studios GPU Team, PixarnvFX : A New Scene & Material Effect Framework for OpenGL and DirectX Thursday in NVIDIA Booth, 2:00 PM - 2:30 PM Tristan Lorach, Developer Relations Senior Engineer, NVIDIA

Movatterモバイル変換

Change Language

SIGGRAPH 2012: NVIDIA OpenGL for 2012

Embed presentation

Recommended

More Related Content

What's hot

Similar to SIGGRAPH 2012: NVIDIA OpenGL for 2012

More from Mark Kilgard

Recently uploaded

SIGGRAPH 2012: NVIDIA OpenGL for 2012

Editor's Notes