NVIDIA's OpenGL Functionality

Room A5 | Monday, September, 20th, 16:00 - 17:20 San Jose Convention Center, San Jose, California NVIDIA’s OpenGL Functionality

Mark J. Kilgard Principal System Software Engineer OpenGL driver Cg (“C for graphics”) shading language OpenGL Utility Toolkit (GLUT) implementer Author of OpenGL for the X Window System Co-author of Cg Tutorial

Outline OpenGL’s importance to NVIDIA OpenGL 3.3 and 4.0 OpenGL 4.1 Loose ends: deprecation, Cg, further extensions

OpenGL Leverage SceniX CompleX OptiX Parallel Nsight Cg

Example of Hybrid Rendering with OptiX OptiX (Ray tracing) OpenGL (Rasterization)

Parallel Nsight Provides OpenGL Profiling Configure Application Trace Settings

Parallel Nsight Provides OpenGL Profiling Magnified trace options shows specific OpenGL (and Cg) tracing options

Parallel Nsight Provides OpenGL Profiling

Parallel Nsight Provides OpenGL Profiling Trace of mix of OpenGL and CUDA shows glFinish & OpenGL draw calls

OpenGL In Every NVIDIA Business

OpenGL on Quadro World class OpenGL 4 drivers 18 years of uninterrupted API compatibility Workstation application certifications Workstation application profiles Display list optimizations Fast antialiased lines Largest memory configurations: 6 gigabytes GPU affinity Enhanced interop with CUDA and multi-GPU OpenGL Advanced multi-GPU rendering Overlays Genlock Unified Back Buffer for less framebuffer memory usage Cross-platform Windows XP, Vista, Win7, Linux, Mac, FreeBSD, Solaris SLI Mosaic Advanced multi-monitor support First class, in-house Linux support

NVIDIA 3D Vision Pro Introduced at SIGGRAPH 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

OpenGL and Tesla Tesla for high performance Compute Error Correcting Code (ECC) memory 10x double-precision Scalable and rack-mountable Quadro for professional graphics and Compute Use interop to mix OpenGL and Compute Tesla M2050 / M2070

OpenGL on Tegra Tegra ultra low power mobile web processor

OpenGL on GeForce Lifestyle Adobe Creative Suite 5 (CS5) Photoshop, Premiere, Encore Internet Accelerated browsing Games Cross-platform gaming Mac, Linux, and Windows

NVIDIA OpenGL Initiatives API improvements Direct State Access (DSA) Bindless Graphics Framebuffer Memory info Texture Barrier First class Direct3D-isms WebGL OpenGL ES Quadro Quadro SDI Capture Quadro SDI Output Quad Buffered Stereo Many more! Data interoperability and control DirectX interop CUDA interop OpenCL interop VDPAU interop Multi-GPU OpenGL interop GPU Affinity

OpenGL Interoperability Multi-GPU

DirectX / OpenGL Interoperability Application OpenGL driver DirectX driver GPU Display DirectX interop is a GL API extension ( WGL_NVX_dx_interop ) Premise: create resources in DirectX, get access to them within OpenGL Read/write support for DirectX 9 textures, render targets and vertex buffers Implicit synchronization is done between Direct3D and OpenGL Soon: DirectX 10/11 support Texture / Vertex buffer object data

DirectX / OpenGL Interoperability // create Direct3D device and resources the regular way direct3D-> CreateDevice (..., &dxDevice); dxDevice-> CreateRenderTarget (width, height, D3DFMT_A8R8G8B8 , D3DMULTISAMPLE_4_SAMPLES , 0, FALSE, &dxColorBuffer, NULL); dxDevice-> CreateDepthStencilSurface (width, height, D3DFMT_D24S8 , D3DMULTISAMPLE_4_SAMPLES , 0, FALSE, &dxDepthBuffer, NULL);

DirectX / OpenGL Interoperability // Register DirectX device for OpenGL interop HANDLE gl_handleD3D = wglDXOpenDeviceNVX (dxDevice); // Register DirectX render targets as GL texture objects GLuint names[2]; HANDLE handles[2]; handles[0] = wglDXRegisterObjectNVX (gl_handleD3D, dxColorBuffer, names[0], GL_TEXTURE_2D_MULTISAMPLE, WGL_ACCESS_READ_WRITE_NVX ); handles[1] = wglDXRegisterObjectNVX (gl_handleD3D, dxDepthBuffer, names[0], GL_TEXTURE_2D_MULTISAMPLE, WGL_ACCESS_READ_WRITE_NVX ); // Now textures can be used as normal OpenGL textures

DirectX / OpenGL Interoperability // Rendering example: DirectX and OpenGL rendering to the // same render target direct3d_render_pass(); // D3D renders to the render targets as usual // Lock the render targets for GL access wglDXLockObjectsNVX (handleD3D, 2, handles); opengl_render_pass(); // OpenGL renders using the textures as render // targets (e.g., attached to an FBO) wglDXUnlockObjectsNVX (handleD3D, 2, handles); direct3d_swap_buffers(); // Direct3D presents the results on the screen

CUDA / OpenCL Interop example Application OpenGL driver CUDA driver Quadro or GeForce Display Tesla or GeForce Interop fast Interop APIs are extensions to OpenGL and CUDA Multi-card interop 2x faster on Quadro / Tesla. Transfer between cards, minimal CPU involvement GeForce requires CPU copy

Multi-GPU OpenGL Interop Application OpenGL driver OpenGL driver Quadro Display Quadro Interop fast Interop using NV_copy_image OpenGL extension Transfer directly between cards, minimal CPU involvement Quadro only GPU affinity Off-screen buffer

NVIDIA OpenGL 4 DirectX 11 superset All of Fermi exposed!

OpenGL Version Progression OpenGL 2.0 2.1 OpenGL 3.0 3.1 3.2 3.3 OpenGL 4.0 4.1 OpenGL 1.0 1.1 1.2.1 1.3 1.4 1.5 Silicon Graphics heritage DirectX 9 era GPUs DirectX 10 era GPUs DirectX 11 era GPUs (2008-2010) (1992-2003) (2004-2007) (2010-)

2010: Big year for OpenGL Three core updates in a single year! OpenGL 4.0 & 3.3 Released March 2010 at Game Developers Conf. (GDC) OpenGL 4.1 Released July 2010 at SIGGRAPH Cross-vendor supports for modern GPU features OpenGL 4.1 more functional than DirectX 11

OpenGL Versions for NVIDIA GPUs OpenGL 4.1+ Fermi-based GPUs GeForce 4xx, Quadro DirectX 11-class GPU functionality OpenGL 3.3 GeForce 8, 9, and 200 series GPUs DirectX 10-class GPU functionality OpenGL 2.1 GeForce 5, 6, and 7 series GPUs DirectX 9-class GPU functionality

Well Worth Reviewing OpenGL 3.3 + 4.0 Additions OpenGL 4.1 is the latest But don’t miss out on OpenGL 4.0’s goodness Honestly, 4.0 is the bigger upgrade Of course, NVIDIA is shipping OpenGL 4.1 A few of the “big ticket” items in 4.0 Texture cube map arrays Double-precision values within shaders Programmable tessellation

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

OpenGL 3.2 Conceptual Processing Flow (pre-2010) Vertex processing Pixel processing Texture mapping Geometric primitive assembly & processing Image primitive processing Transform feedback Pixel unpacking Pixel packing Vertex assembly pixels in framebuffer object textures texture buffer objects texture image specification image rectangles, bitmaps primitive topology, transformed vertex data vertex texture fetches pixel pack buffer objects pixel unpack buffer objects vertex buffer objects transform feedback buffer objects buffer data, unmap buffer geometry texture fetches primitive batch type, vertex indices, vertex attributes primitive batch type, vertex data fragment texture fetches pixel image or texture image specification map buffer, get buffer data transformed vertex attributes image and bitmap fragments point, line, and polygon fragments pixels to pack unpacked pixels pixels fragments filtered texels buffer data vertices Legend programmable operations fixed-function operations copy pixels, copy texture image Buffer store uniform/ parameters buffer objects Fragment processing stenciling, depth testing, blending, accumulation OpenGL 3.2 Raster operations Framebuffer Command parser

Vertex processing Pixel processing Texture mapping Geometric primitive assembly & processing Image primitive processing Transform feedback Pixel unpacking Pixel packing Vertex assembly pixels in framebuffer object textures texture buffer objects texture image specification image rectangles, bitmaps primitive topology, transformed vertex data vertex texture fetches pixel pack buffer objects pixel unpack buffer objects vertex buffer objects transform feedback buffer objects buffer data, unmap buffer geometry texture fetches primitive batch type, vertex indices, vertex attributes primitive batch type, vertex data fragment texture fetches pixel image or texture image specification map buffer, get buffer data transformed vertex attributes image and bitmap fragments point, line, and polygon fragments pixels to pack unpacked pixels pixels fragments filtered texels buffer data vertices Legend programmable operations fixed-function operations copy pixels, copy texture image Buffer store Fragment processing stenciling, depth testing, blending, accumulation Control point processing Patch assembly & processing Patch tessellation generation Patch evaluation processing patch control points transformed control points transformed patch transformed patch, bivariate domain transformed patch tessellation texture fetches patch topology, evaluated patch vertex patch data OpenGL 4.1 Raster operations Framebuffer Command parser

OpenGL 3.3 Laundry List (1) Sampler objects to decouple sampler (i.e. glTexParameter state) from texture image state (i.e. glTexImage * state) Mimics Direct3D’s decoupling of sampler and texture image state Mostly useful to mimic Direct3D’s model of operation glBindSampler(GLuint texunit,GLuint sampler) binds a named sampler object Zero for a sampler name means use the texture object’s sampler state Alternate occlusion query target ( GL_ANY_SAMPLES_PASSED ) returning a Boolean result instead of the samples passed

OpenGL 3.3 Laundry List (2) New GLSL functions ( floatBitsToInt , intBitsToFloat , etc.) to convert to and from floating-values to integer bit encodings New extended blend functions involving two output colors from the fragment shader (instead of one) New GL_SRC_COLOR 1 , etc. blend function modes Explicit mapping in your shader of GLSL vertex shader inputs to vertex attributes Example: layout(location = 3) in vec4 normal; Avoids otherwise needed calls to glBindAttribLocation in your applications Explicit mappings in your shader of GLSL fragment shader output colors to bound color buffers Example: layout(location=2) out vec4 color Avoids otherwise needed calls to glBindFragDataLocation in your applications

OpenGL 3.3 Laundry List (3) New unsigned 10-10-10-2 integer internal texture format ( GL_RGB10_A2UI ) Like the fixed-point GL_RGB10_A2 format, but just with unsigned integer components The ability to swizzle the texture components returned by a texture fetch Mimics an ability of the PlayStation 3 architecture R G B A 1 0 X Y Z W texture fetch result delivered to shader fetched texture’s RGBA glTexParameteri(GL_TEXTURE_SWIZZLE_R, GL_BLUE) glTexParameteri(GL_TEXTURE_SWIZZLE_G, GL_BLUE) glTexParameteri(GL_TEXTURE_SWIZZLE_B, GL_ONE) glTexParameteri(GL_TEXTURE_SWIZZLE_A, GL_GREEN)

OpenGL 3.3 Laundry List (4) Timer query objects Allows timing of OpenGL command sequences without timing the idle time Allows accurate instrumentation of GPU bottlenecks Instanced arrays Mimics the behavior of Direct3D’s geometry instancing OpenGL’s support for immediate mode and inexpensive primitive batches makes geometry instancing needless Primary use would be emulating Direct3D’s model New command glVertexAttribDivisor New 10-10-10-2 vertex array format types Signed and unsigned versions ( GL_INT_2_10_10_10_REV and GL_UNSIGNED_INT_2_10_10_10_REV ) Good for packed vertex normals

OpenGL 3.3 Support & Older GPUs Pre-Fermi (DirectX 10 era) GPUs support OpenGL 3.3 Such pre-Fermi GPUs will never natively support OpenGL 4 (FYI: NVIDIA drivers do provide slooow emulation support for newer architectures; useful for experiments, but that’s about all) Only NVIDIA’s Fermi (and on) support OpenGL 4 However many ARB extensions introduced by OpenGL 4 are supported by older DirectX 10 (and sometimes 9) GPUs Examples: OpenGL 4.1’s ARB_viewport_array is exposed on DirectX 10 era NVIDIA GPUs NVIDIA’s policy is to expose as much of OpenGL functionality as a given GPU generation can possibly expose

OpenGL 4.0 Features Categories Texturing Framebuffer Transform feedback Programmability Indexing Numerics Subroutines Fragment processing Tessellation Fairly broad categories

OpenGL 4.0 Texturing Features 3-component buffer texture formats RGB formats for floats ( GL_RGB32F ), signed ( GL_RGB32I ) & unsigned integers ( GL_RGB32UI ) GLSL functions ( textureLOD ) to return vector of automatic level-of-detail parameters GLSL function ( textureGather ) returns any single component of 2x2 footprint in RGBA vector Including depth comparisons Arbitrary varying offsets Independent offsets for each of 4 returned texels Cube map texture arrays ( GL_TEXTURE_CUBE_MAP_ARRAY ) Texture array where every 6 slices form a cube map

Cube Map Texture Arrays Extends 1D and 2D texture array concept to omni-directional cube map texture images accessed by 4D ( s , t , r , i ) coordinate set where ( s , t , r ) is an un-normalized direction vector And i is a texture array index “slice” Uses: Bidirectional Radiance Distribution Functions (BRDFs), spherical harmonics, time varying environment maps slice i = 0 slice i = 1 slice i = 2

OpenGL 4.0 Framebuffer Features Per-color buffer blending controls Each shader color output can specify its own Blend equation Separate blend functions for RGB and alpha New blend commands ( glBlendEquationi , glBlendFunci , etc.) take a buffer index parameter Better blending control for multiple render targets Per-sample shading Fragment shader executes per-sample, instead of per-pixel Easy API: First glMinSampleShadingARB(1.0) says per-sample shade 100% of samples Then enable: glEnable(GL_SAMPLE_SHADING)

OpenGL 4.0 Transform Feedback Features Transform feedback objects to encapsulated transform feedback related state ( glBindTransformFeedback , etc.) Allows the ability to pause ( glPauseTransformFeedback ) and resume ( glResumeTransformFeedback ) transform feedback operations Draw primitives captured by transform feedback ( glDrawTransformFeedback ) without having to explicitly query back the number of captured primitives Multiple separate vertex streams can be written during transform feedback Semi-related functionality : New commands glDrawArraysInstanced and glDrawElementsInstancedBaseVertex can source their “draw arrays” and “draw elements” parameters from a GPU buffer object primitive processing rasterization pixel processing framebuffer transform feedback buffer data framebuffer buffer data

OpenGL Shading Language Update OpenGL Shading Language (GLSL) version now synchronized with OpenGL version number So GLSL 4.00 corresponds to OpenGL 4.0 And GLSL 3.30 corresponds to OpenGL 3.3 Previously, confusingly, GLSL version number didn’t match OpenGL core version number Example: OpenGL 3.2 depended on GLSL 1.50 Remember to request the version of GLSL you require Example: #version 400 compatibility Best to always request compatibility or risk forced shader re-writes due to backward compatibility and deprecation issues

OpenGL 4.0 Programmable Features: Indexing Indexing into arrays of samplers with varying indices Example syntax: sampler2D sampler_array[3] (Different from a “texture array” which is one texture object that stores multiple texture image array slices) Old GLSL requirement: index must be uniform for all uses Now : Shader Model 5.0 in GLSL 4.x & NV_gpu_program5 allow arbitrary varying expressions for indices Similarly, now allows indexing into arrays of uniform blocks

OpenGL 4.0 Programmable Features: Numeric Data Type Control Implicit signed and unsigned integer conversions “ precise ” type qualifier to avoid optimization-induced invariance Fused floating-point multiply-add (MAD) operation ( fma ) Splitting a floating-point value into significand and exponent ( frexp ) Constructing a floating-point value from a significand and exponent ( ldexp ) Integer bit-field manipulation ( bitfieldExtract , findLSB , etc.) Packing and unpacking small fixed-point values into larger scalar values ( packUnorm2x16 , unpackUnorm4x8 , etc.) NOTE: Recall 3.3 added raw cast of IEEE floating-point values to and from integers ( floatBitsToInt , intBitsToFloat , etc.)

Double-precision data types Scalar double Vector dvec4 , etc Matrix dmat3 (3x3), dmat4x2 , etc. Doubles can reside in buffer objects Double-precision operations Multiply, add, multiply-add (MAD) Relational operators, including vector comparisons Absolute value Minimum, maximum, clamping, etc. Packing and unpacking No support for double-precision angle, trigonometric, or logarithmic functions OpenGL 4.0 Programmable Features: Double-Precision Numeric

Double Precision in OpenGL Shaders Double-precision magnified 970,000x Single-precision magnified 970,000x Mandelbrot set interactively visualized

Mix of Double- and Single-Precision fp32 fp32 fp32 fp32 fp32 fp32 fp32 fp32 fp32 fp32 fp32 fp32

Mandelbrot Shaders Compared #version 400 compatibility in vec2 position; out vec4 color; uniform float max_iterations; uniform sampler1D lut; uniform mat2x3 matrix; float mandel( vec2 c) { vec2 z = vec2 (0.0); float iterations = 0.0; while(iterations < max_iterations) { vec2 z2 = z*z; if (z2.x + z2.y > 4.0) break; z = vec2 (z2.x - z2.y, 2.0 * z.x * z.y) + c; iterations++; } return iterations; } void main() { vec2 pos = vec2 (dot(matrix[0], vec3 (position,1)), dot(matrix[1], vec3 (position,1))); float iterations = mandel(pos); // False-color pixel based on iteration // count in look-up table float s = iterations / max_iterations; color = texture(lut, s); } #version 400 compatibility in vec2 position; out vec4 color; uniform float max_iterations; uniform sampler1D lut; uniform dmat2x3 matrix; float mandel( dvec2 c) { dvec2 z = dvec2 (0.0); float iterations = 0.0; while(iterations < max_iterations) { dvec2 z2 = z*z; if (z2.x + z2.y > 4.0) break; z = dvec2 (z2.x - z2.y, 2.0 * z.x * z.y) + c; iterations++; } return iterations; } void main() { dvec2 pos = dvec2 (dot(matrix[0], dvec3 (position,1)), dot(matrix[1], dvec3 (position,1))); float iterations = mandel(pos); // False-color pixel based on iteration // count in look-up table float s = iterations / max_iterations; color = texture(lut, s); } Double-precision dmat2x3 dvec2 dvec3 Single-precision mat2x3 vec2 vec3

OpenGL 4.0 Programmable Features: Subroutines GLSL shaders support subroutine functions Typed subroutines variables can be bound to subroutine functions Allows for “indirect” subroutine calls Query glGetSubroutineIndex obtains index names for subroutine function names Command glUniformSubrountinesiv updates a program’s subroutine uniform variables with subroutine index names Subroutine uniforms are per-stage Unlike data value uniforms which apply to all stages of a program object Use of subroutines can avoid/minimize shader recompilation costs by dynamically switching subroutines

OpenGL 4.0 Programmable Features: Fragment Shaders Mask of covered samples ( gl_SampleMask[ ] ) of a fragment is read-able Fragment shader can “know” which samples within its pixel footprint are covered Fragment attribute interpolation control By offset relative to the pixel center ( interpolateAtOffset ) By a varying sample index ( interpolateAtSample ) Or at the centroid of the covered samples ( interpolateAtCentroid ) pixel center relative offset dx dy sample index 0 1 2 3 centroid centroid position of 3 covered samples pixel footprints interpolation modes

OpenGL 4.0 Programmable Features: Tessellation Shaders Higher-level geometric primitives: Patches Quadrilateral patches Triangular patches Iso-line patches Two new shaders stages Existing vertex shader stage re-tasked to transform control points Tessellation control shader Executes a gang of cooperating threads per patch One thread per output control point Computes level-of-detail (LOD) parameters, per-patch parameters, and/or performs patch basis change Hard-wired tessellation primitive generator emits topology for the patch Based on LOD parameters and patch type Tessellation evaluation shader Evaluates position of tessellated vertex within patch And performs conventional vertex processing (lighting, etc.)

Programmable Tessellation Data Flow vertex shader clipping, setup, & rasterization fragment shader assembled primitive fragments index to vertex attribute puller indices primitive topology & in-band vertex attributes vertex attributes vertex shader geometry shader clipping, setup, & rasterization fragment shader OpenGL 4.0 added tessellation shaders assembled primitive fragments index to vertex attribute puller indices primitive topology & in-band vertex attributes vertex attributes tessellation control (hull) shader tessellation evaluation (domain) shader tessellation generator level-of-detail parameters patch control points & parameters fine primitive topology (u, v) coordinates primitive stream control points control point (vertex) shader vertex attributes rectangular & triangluar patch OpenGL 3.2 programmable fixed-function Legend geometry shader primitive stream

“DirectX 11 Done Right” What it means Architecture designed for Tessellation & Compute Tessellation = feed-forward operation No loop-back or cycling data through memory All programmable graphics pipeline stages can operate in parallel in a Fermi GPU vertex shaders, tessellation control & evaluation shaders, geometry shaders, and fragment shaders Plus in parallel with all fixed-function stages

Tessellated Triangle and Quad Notice fractional tessellation scheme in wire-frame pattern

Iso-line Tessellation Intended for hair, particular animation and combing Scientific visualization applications too

Bezier Triangle Patch Bezier triangle’s tessellation Bezier triangle’s control point hull

Bezier Triangle Control Points b 300 b 210 b 120 b 030 b 021 b 012 b 003 b 102 b 201 b 111 P(u,v) = u 3 b 300 + 3 u 2 v b 210 + 3 uv 2 b 120 + v 3 b 030 + 3 u 2 w b 201 + 6 uvw b 111 + 3 v 2 w b 021 + 3 uw 2 b 102 + 3 uw 2 b 012 + w 3 b 003 where w=1-u-v

“Precise” Evaluation Math for Bezier Triangle Example precise vec4 www = w*w*w * vec4 (b300,1), uuu = u*u*u * vec4 (b030,1), vvv = v*v*v * vec4 (b003,1), wwu3 = 3*w*w*u * vec4 (b210,1), wuu3 = 3*w*u*u * vec4 (b120,1), wwv3 = 3*w*w*v * vec4 (b201,1), uuv3 = 3*u*u*v * vec4 (b021,1), wvv3 = 3*w*v*v * vec4 (b102,1), uvv3 = 3*u*v*v * vec4 (b012,1), wuv6 = 6*w*u*v * vec4 (b111,1); precise vec4 wuE = wwu3 + wuu3, uvE = uvv3 + uuv3, wvE = wwv3 + wvv3, E = wuE + uvE + wvE; precise vec4 C = www + uuu + vvv; precise vec4 p = C + E; Notice evaluation needs homogenous coordinates, hence vec4(x,1)

Bi-cubic Patch Mesh b 00 b 10 b 20 b 30 b 01 b 11 b 21 b 31 b 02 b 12 b 22 b 32 b 03 b 13 b 23 b 33

Bi-cubic Quadrilateral Patch Bi-cubic patch’s tessellation Bi-cubic patch’s control point hull

Surfaces are Determined by Their Control Points Moving control points displaces the evaluated surfaces

Utah Teapot: Bi-cubic Patch Mesh

Bi-cubic Patches Used in Production Rendering Production renderers for film and video rely on surfaces

Amplification of Standard Triangle Meshes Original faceted mesh Tessellation results in triangle amplification with curvature Results of Phong shading of curved

Art Pipeline for Tessellation … Generate LODs Displaced Surface GPU Control Cage Smooth Surface Polygon Mesh

Displacement Mapping Displacement Mapping - Sample displacement value from a texture - Displace vertex along its normal Watertight Normals - cross product of a pair of tangent, bi-tangent vectors - discontinuities occur at shared corners and edges - define corner and edge ownership

Tessellation Shader Example: Starting Simple Control point (vertex) shader transforms vertex positions to eye-space Simple 4x3 affine matrix transformation Tessellation control shader accepts a 3 control points Forms a standard triangle Passes each control point on to tessellation valuation Computes a level-of-detail scalar for each edge based on the scaled window-space length of each edge So the tessellation of the generated triangle patch adapts to the screen-space size of the triangle Tessellation evaluation shader runs at each tessellated vertex of triangle patch Gets 3 control points + (u,v,w) barycentric triangle weights Weights control points to make a new position Computes a surface normal The normal is constant for the entire triangle in this case, but generally the normal would vary Outputs barycentric weights as RGB color (to visualize)

Simple Example Rendered Original single triangle Adaptive tessellation automatically tessellates distant, small triangle less and close, large triangle more Visualization of barycentric weights used by tessellation evaluation shader

GLSL Control Point (Vertex) Shader #version 400 compatibility // Multiply XYZ vertex by 4x3 modelview // matrix (assumed non-projective) vec3 transform(precise vec3 v) { vec3 r; // Remember: GLSL is oddly column-major // so [col][row] for ( int i=0; i<3; i++) { r[i] = v[0]* gl_ModelViewMatrix [0][i] + v[1]* gl_ModelViewMatrix [1][i] + v[2]* gl_ModelViewMatrix [2][i] + gl_ModelViewMatrix [3][i]; } return r; } uniform vec2 wh; // scaled width & height of screen out vec3 eye_space_pos; out vec2 scaled_window_space_pos; out vec3 eye_normal; void main( void ) { eye_space_pos = transform(gl_Vertex.xyz); vec4 ndc_pos = gl_ModelViewProjectionMatrix * vec4 ( gl_Vertex .xyz,1); scaled_window_space_pos = wh*(ndc_pos.xy / ndc_pos.w); // Pass along two sets of texture coordinates... gl_TexCoord [0] = gl_MultiTexCoord0 ; gl_TexCoord [1] = gl_MultiTexCoord1 ; mat3 nm = gl_NormalMatrix ; vec3 n = gl_Normal ; eye_normal = normalize (nm * n); }

GLSL Tessellation Control Shader #version 400 compatibility layout ( vertices =3) out ; // thread ID #define TID gl_InvocationID out float sharable_len[]; in vec3 eye_space_pos[]; in vec2 scaled_window_space_pos[]; out vec3 eye_space_pos2[]; void main( void ) { sharable_len[TID] = distance (scaled_window_space_pos[TID], scaled_window_space_pos[(TID+1)%3]); barrier (); float len0 = sharable_len[0], len1 = sharable_len[1], len2 = sharable_len[2]; eye_space_pos2[TID] = eye_space_pos[TID]; // Limit level-of-detail output to thread 0 . if (TID == 0) { // Outer LOD gl_TessLevelOuter [0]=len1; //V1-to-V2 edge gl_TessLevelOuter [1]=len2; //V2-to-V0 edge gl_TessLevelOuter [2]=len0; //V0-to-V1 edge // Inner LOD gl_TessLevelInner[0] = max (len0, max (len1, len2)); } }

GLSL Tessellation Evaluation Shader #version 400 compatibility layout ( triangles ) in ; layout ( ccw ) in ; layout ( fractional_odd_spacing ) in ; in float sharable_len[]; in vec3 eye_space_pos2[]; vec3 lerp3( in vec3 attrs[3], vec3 uvw) { return attrs[0] * uvw[0] + attrs[1] * uvw[1] + attrs[2] * uvw[2]; } vec4 applyPerspective( mat4 affine_matrix, vec3 v) { vec4 r; r[0] = affine_matrix[0][0] * v[0]; r[1] = affine_matrix[1][1] * v[1]; r[2] = affine_matrix[2][2] * v[2] + affine_matrix[3][2]; r[3] = -v[2]; return r; } void main( void ) { vec3 barycentric_weights = gl_TessCoord ; vec3 v[3]; for ( int i = 0; i < 3; i++) { v[i] = eye_space_pos2[i]; } vec3 p = lerp3(v, barycentric_weights); gl_Position = applyPerspective( gl_ProjectionMatrix , p); vec3 dpdv = eye_space_pos2[1] - eye_space_pos2[0]; vec3 dpdw = eye_space_pos2[2] - eye_space_pos2[0]; vec3 normal = cross (dpdv, dpdw); // Show barycentric weights as color. vec4 color = vec4 (barycentric_weights, 1.0); gl_FrontColor = color; gl_TexCoord [0] = vec4 ( normalize (normal), length ( normalize (normal))); gl_TexCoord [1] = vec4 (p, 1); }

More Complicated Example: Adaptive PN Triangles Idea : Send a triangle mesh with positions & normals Have the tessellation unit tessellated a curved triangle that matches the per-triangle positions & normals Nice because existing triangle meshes can get extra triangle detail Details Same control point (vertex) shader as before Tessellation control shader Input: 3 control points + normals Output: 10 control points for a Bezier triangle patch Output: 6 controls points for quadratic normal interpolation Output: adaptive level-of-detail (LOD) based on edge lengths in scaled window space Tessellation evaluation shader Does cubic interpolation of Bezier triangle to compute position Does quadratic interpolation to compute surface normal

Adaptive PN Triangles Tessellation Rendered Visualization of barycentric weights used by tessellation evaluation shader Original monkey head faceted mesh Tessellated & Phong shaded Wire frame of adaptive tessellation

Adaptive PN Triangle Tessellation Control Shader #version 400 compatibility // interior control points are per-patch layout ( vertices =3) out ; #define TID gl_InvocationID // thread ID #define P eye_space_pos #define P2 eye_space_pos2 #define N eye_normal in vec3 eye_space_pos[]; in vec2 scaled_window_space_pos[]; in vec3 eye_normal[]; out vec3 eye_space_pos2[]; out vec3 eye_space_normal[]; out vec3 ccw_cp[]; out vec3 cw_cp[]; out vec3 mid_normal[]; patch out vec3 b111; void main(void) { // TIDp1 = counter-clockwise (plus 1) // control point index from TID int TIDp1 = TID<2 ? TID+1 : 0; // ^^ Means TIDp1 = (TID+1)%3; // TIDm1 = clockwise (minus 1) control point // index from TID int TIDm1 = TID>0 ? TID-1 : 2; // ^^ Means TIDp1 = (TID+2)%3; // Compute triangle LOD parameters. // Each of 3 threads computes each of 3 edge lengths. gl_TessLevelOuter [TID] = distance (scaled_window_space_pos[TIDm1], scaled_window_space_pos[TID]); barrier (); gl_TessLevelInner [0] = max ( gl_TessLevelOuter [0], max ( gl_TessLevelOuter [1], gl_TessLevelOuter [2])); // ccw_cp[TID] is the control point immediate // counter-clockwise from P[TID] // cwc_cp[TID] is the control point immediate // clockwise from P[TID] ccw_cp[TID] = (2*P[TID] + P[TIDp1] – dot (P[TIDp1]-P[TID],N[TID])*N[TID])/3.0; cw_cp[TID] = (2*P[TID] + P[TIDm1] – dot (P[TIDm1]-P[TID],N[TID])*N[TID])/3.0; P2[TID] = ccw_cp[TID] + cw_cp[TID]; // (b210+b120+b021+b012+b102+b201)/6.0 vec3 E = (P2[0] + P2[1] + P2[2])/6.0, V = (P[0] + P[1] + P[3])/3.0; b111 = E + (E-V)/2.0; P2[TID] = P[TID]; eye_space_normal[TID] = N[TID]; float v = 2* dot (P[TIDp1]-P[TID],N[TID]+N[TIDp1]) / dot (P[TIDp1]-P[TID],P[TIDp1]-P[TID]); vec3 h = N[TID] + N[TIDp1] - v*(P[TIDp1]-P[TID]); mid_normal[TID] = normalize (h); }

PN Triangle Control Shader Compute window-space edge distances Thread parallel Scaled edge length = exterior (edge) LOD Maximum scaled edge length = interior LOD Compute Bezier triangle edge control points Thread parallel Compute Bezier triangle central control point Pass through eye-space position & normal Thread parallel Compute quadratic normal edge control points Thread parallel 1. 2. 3. 4. 5.

#version 400 compatibility layout ( triangles ) in ; layout ( ccw ) in ; layout ( fractional_even_spacing ) in ; // Assumes matrix in gluPerspective form // Intended to be cheaper than full 4x4 // transform by projection matrix // Compiles to just 5 MAD ops precise vec4 applyPerspective( precise mat4 affine_matrix, precise vec4 v) { precise vec4 r; r[0] = affine_matrix[0][0] * v[0]; r[1] = affine_matrix[1][1] * v[1]; r[2] = affine_matrix[2][2] * v[2] + affine_matrix[3][2]*v[3]; r[3] = -v[2]; return r; } // 10 input control points for // Bezier triangle position in precise vec3 eye_space_pos2[]; in precise vec3 ccw_cp[]; in precise vec3 cw_cp[]; patch in precise vec3 b111; // 6 input control points for // quadratic normal interpolation in precise vec3 eye_space_normal[]; in vec3 mid_normal[]; Adaptive PN Triangle Tessellation Evaluation Shader (1 of 3)

void main( void ) { // Evaluate position by weighting triangle // vertex positions with the generated vertex's // barycentric weights. vec3 barycentric_weights = gl_TessCoord ; precise float u = barycentric_weights.x, v = barycentric_weights.y, w = barycentric_weights.z; // w should be 1- u - v vec3 triangle_vertex[3]; for ( int i = 0; i < 3; i++) { triangle_vertex[i] = eye_space_pos2[i]; } // 10 position control points of a Bezier triangle precise vec3 b300 = eye_space_pos2[0], b030 = eye_space_pos2[1], b003 = eye_space_pos2[2], b210 = ccw_cp[0], b201 = cw_cp[0], b021 = ccw_cp[1], b120 = cw_cp[1], b102 = ccw_cp[2], b012 = cw_cp[2]; // Weight the position control points with a // (cubic) Bezier triangle basis precise vec4 www = w*w*w * vec4 (b300,1), uuu = u*u*u * vec4 (b030,1), vvv = v*v*v * vec4 (b003,1), wwu3 = 3*w*w*u * vec4 (b210,1), wuu3 = 3*w*u*u * vec4 (b120,1), wwv3 = 3*w*w*v * vec4 (b201,1), uuv3 = 3*u*u*v * vec4 (b021,1), wvv3 = 3*w*v*v * vec4 (b102,1), uvv3 = 3*u*v*v * vec4 (b012,1), wuv6 = 6*w*u*v * vec4 (b111,1), wuE = wwu3 + wuu3, uvE = uvv3 + uuv3, wvE = wwv3 + wvv3, E = wuE + uvE + wvE, C = www + uuu + vvv, p = C + E, clip_space_p = applyPerspective( gl_ProjectionMatrix , p); gl_Position = clip_space_p; Adaptive PN Triangle Tessellation Evaluation Shader (2 of 3)

// Weight the normal control points with a quadratic basis vec3 n200 = eye_space_normal[0], n020 = eye_space_normal[1], n002 = eye_space_normal[2], n110 = mid_normal[0], n011 = mid_normal[1], n101 = mid_normal[2], normal = n200*w*w + n020*u*u + n002*v*v + n110*w*u + n011*u*v + n101*w*v; // Visualize barycentric weights as color. vec4 color = vec4 (barycentric_weights, 1.0); gl_FrontColor = color; // Output the normalized surface normal gl_TexCoord [0] = vec4 (normalize(normal),1); // Output the (perspective-divided) eye-space view vector gl_TexCoord [1] = vec4 (p.xyz/p.w, 1); } Adaptive PN Triangle Tessellation Evaluation Shader (3 of 3)

PN Triangle Evaluation Shader Bezier triangle position evaluation Cubic evaluation using 10 position control points ~80 Multiply-Add operations Apply perspective transformation Multiply position by spare frustum matrix 5 Multiply-Add operations Bezier triangle normal evaluation Quadratic evaluation using 6 normal control points 3×12 Multiply-Add operations Output vectors Barycentric coordinates encoded as color Normalized surface normal and eye direction vectors

Careful About Cracks at Patch Seams Patch seams need to be computed very carefully Otherwise you get cracks Animation makes cracks obvious! Be sure of the following Your LOD parameters at the seam of two patches are computed bit-for-bit identical The control points that influence the evaluation along a patch edge are bit-for-bit identical and use symmetric evaluation Influence means has “non-zero weight” A+B+C ≠ C+B+A in floating-point precise keyword can help this artificially injected cracks

What is in OpenGL 4.1? ARB_separate_shader_objects Ability to bind programs individually to programmable stages ARB_viewport_array Multiple viewports for the same rendering surface, or one per surface ARB_ES2_compatibility Pulling missing functionality from OpenGL ES 2.0 into OpenGL ARB_get_program_binary Query and load a binary blob for program objects ARB_vertex_attrib_64_bit Provides 64-bit floating-point component vertex shader inputs ARB_shader_precision Documents precision requirements for several floating-point operations

Separate Shader Objects Previous, GLSL forced linking of shaders into a monolithic program object Assembly shaders never had this restriction Neither does Direct3D Separate shader objects allow mix-and-match shaders in GLSL bump mapping fragment shader red velvet fragment shader wobbly torus vertex shader smooth torus vertex shader Avoids combinatorial explosion of shader combinations ARB_separate_shader_objects

Viewport Array Geometry shader can “steer” output primitives to 1 of 16 viewports into the framebuffer surface GLSL geometry shader writes to gl_ViewportIndex Value from 0 to 15 Separate scissor rectangle per viewport too 1 primitive in 6 primitives out each projected to different cube face & directed to different viewport geometry shader ARB_viewport_array

ES Compatibility OpenGL ES 2.0 introduced stuff not in mainstream OpenGL OpenGL 4.1 synchronizes mainstream OpenGL with ES 2.0 Functionality Floating-point versions of glDepthRange and glClearDepth glDepthRangef & glClearDepthf Query for shader range & precision: glGetShaderPrecisionFormat Because ES 2.0 devices might have less precision and range than single-precision Ability to “release” the shader compiler when all shader compilation is done: glReleaseShaderCompiler glShaderBinary allows loading of pre-compiled shader binaries ARB_ES2_compatibility

Query for Binary Shader Blobs Adds commands to query and re-specify a program object returned as a binary shader blob glGetProgramBinary returns a binary blob for a compiled & linked program object glProgramBinary re-specifies a program object with the opaque data from a binary blob from glGetProgramBinary glProgramParameteri can query if a binary program is actually retrievable Rationale : Faster shader loader, can skip some of the expense of GLSL compiler ARB_get_program_binary

64-bit Vertex Attributes New vertex attribute commands for passing in double-precision attributes glVertexAttribL [1-4] d {v} For immediate mode 64-bit double precision attributes glVertexAttribLPointer For specifying vertex arrays glVertexArrayVertexAttribLOffsetEXT included if EXT_direct_state_access is supported In order to make sense, this requires 64-bit shader math So needs ARB_gpu_shader_fp64 or OpenGL 4.0 NOTE: Existing immediate mode calls in OpenGL such as glVertex3d simply marshal inputs to single-precision ARB_vertex_attrib_64_bit

Shader Precision Adds precision requirements to GLSL’s execution environment Specifies relative error in terms of “units of last place” (ULP) various GLSL math operation should have So consistent with Direct3D’s requirements Actually a NOP for NVIDIA GPUs NVIDIA already meets these requirements Request with #extension GL_ARB_shader_precision : enable ARB_shader_precision

New ARB only Extensions ARB_robustness Robust features to grant greater stability when using untrusted shaders and code ARB_debug_output Callback mechanism to receive errors and warning messages from the GL ARB_cl_event Create OpenGL sync objects linked to OpenCL event objects

Technology behind OpenGL 4.0 NVIDIA ARB_tessellation_shader ARB_gpu_shader5 ARB_shader_subroutine ARB_timer_query ARB_transform_feedback2 ARB_transform_feedback3 ARB_gpu_shader_fp64 ARB_sample_shading ARB_texture_buffer_object_rgb32 ARB_draw_indirect ARB_texture_swizzle ARB_texture_query_LOD ARB_instanced_arrays NVIDIA and AMD ARB_texture_gather ARB_sampler_objects AMD ARB_vertex_type_2101010_rev ARB_shader_bit_encoding ARB_blend_func_extended ARB_draw_buffers_blend ARB_texture_cube_map_array ARB_occlusion_query2 TransGaming ARB_texture_rgb10_a2ui Intel ARB_explicit_attrib_location

Technology behind OpenGL 4.1 NVIDIA ARB_separate_shader_objects ARB_ES2_compatibility ARB_vertex_attrib_64 bit ARB_texture_filter_anisotropic ARB_robustness Create_context_robust_access AMD ARB_viewport_array ARB_debug_output ARB_shader_stencil_export Apple ARB_get_program_binary Intel ARB_shader_precision Khronos ARB_cl_event

Accelerating OpenGL Innovation OpenGL increased pace of innovation Six new spec versions in two years Actual implementations following specifications closely OpenGL 4 is a superset of DirectX 11 functionality While retaining backwards compatibility DirectX 10.1 2004 2006 2008 OpenGL 2.0 OpenGL 2.1 OpenGL 3.0 OpenGL 3.1 2009 DirectX 9.0c DirectX 10.0 DirectX 11 2005 2007 OpenGL 3.2 OpenGL 3.3 + OpenGL 4.0 2011 NOW! OpenGL 4.1

Additional OpenGL Topics Deprecation Cg 3.0 Additional OpenGL extensions

NVIDIA’s Position on OpenGL Deprecation: Core vs. Compatibility Profiles OpenGL 3.1 introduced notion of “core” profile Idea was remove stuff from core to make OpenGL “good-er” Well-intentioned perhaps but… Throws API backward compatibility out the window Lots of useful functionality got removed that is in fast hardware Examples: Polygon mode, line width Lots of easy-to-use, effective API got labeled deprecated Immediate mode Display lists Best advice for real developers Simply use the “compatibility” profile Easiest course of action Requesting the core profile requires special context creation gymnastics Avoids the frustration of “they decided to remove what??” Allows you to use existing OpenGL libraries and code easily No, your program won’t go faster for using the “core” profile It may go slower because of extra “is this allowed to work?” checks

What got ( er, didn’t get ) deprecated? Feature Hardware accelerated YES Line stipple YES Quadrilaterals (GL_QUAD*) and polygons YES Regular (non-sprite) points YES Two-sided lighting YES, key cases optimized to be faster than shaders Fixed-function vertex processing YES Edge flags YES Immediate mode YES Line widths > 1

What got ( er, didn’t get ) deprecated? Feature Hardware accelerated YES GL_DEPTH_TEXTURE_MODE YES ALPHA, LUMINANCE, LUMINANCE_ALPHA, INTENSITY texture formats Driver accelerated for speed glBitmap YES glPixelZoom YES glDrawPixels CPU vector instructions + GPU Pixel transfer modes YES Polygon stipple YES Separate polygon draw modes: glPolygonMode

What got ( er, didn’t get ) deprecated? Feature Hardware accelerated YES Auxiliary color buffers YES glCopyPixels YES, fp16 Accumulation buffer YES Alpha test YES Fixed-function fragment processing YES Automatic mipmap generation YES Texture borders YES Texture GL_CLAMP wrap mode

What got ( er, didn’t get ) deprecated? Feature Hardware accelerated Best advice : Overwhelming majority of deprecated features are fully hardware-accelerated by NVIDIA GPUs and their drivers so take advantage of them—they aren’t going away Think of these features as what you lack in Direct3D YES Color material No, but could be by Fermi Evaluators No, but selection could be Selection and feedback modes Highly optimized by NVIDIA, fastest way to send geometry Display lists Ignored, always fastest + nicest Hints No Attribute stacks

Cg 3.0 Update (1) Cg 3.0 includes Shader Model 5.0 support New gp5vp , gp5gp , and gp5fp profiles correspond to Shader Model 5.0 vertex, geometry, and fragment profiles Compiles to assembly text for the NV_gpu_program5 assembly-level shading language Cg 3.0 also introduces programmable tessellation profiles gp5tcp = NV_gpu_program5 tessellation control program gp5tep = NV_gpu_program5 tessellation evaluation program Cg 3.0 also support DirectX 11 tessellation profiles So Cg provides one common language to target either OpenGL or Direct3D programmable tessellation

Cg 3.0 Update (2) OpenGL Shading Language 4.10 profiles Compile Cg to assembly, GLSL, or HLSL Support for up to 32 texture units Previously 16 Unlike GLSL, CgFX provides a powerful effect system compatible with DirectX 9’s FX Download Cg 3.0 now from http://developer.nvidia.com/object/cg_download.html

Cg Off-line Compiler Compiles GLSL to Assembly GLSL is “opaque” about the assembly generated Hard to know the quality of the code generated with the driver Cg Toolkit 3.0 is a useful development tool even for GLSL programmers Cg Toolkit’s cgc command-line compiler actually accepts both the Cg (cgc’s default) and GLSL (with –oglsl flag) languages Uses the same compiler technology used to compile GLSL within the driver Example command line for compiling earlier tessellation control and evaluation GLSL shaders cgc -profile gp5tcp -oglsl -po InputPatchSize=3 -po PATCH_3 filename.glsl (assumes 3 input control points and 3 outputs points) cgc -profile gp5tep -oglsl -po PATCH_3 filename.glsl (assumes 3 input control points)

More NVIDIA OpenGL Features to Explore (1) Bindless Graphics NV_vertex_buffer_unified_memory (VBUM), NV_shader_buffer_load (SBL), and NV_shader_buffer_store (SBS) Supports mapping OpenGL buffer objects to GPU virtual address space Permits faster vertex array state changes by avoiding expensive glBindBuffer commands Direct State Access (DSA) Avoid error-prone OpenGL API’s reliance on “selectors” such as glMatrixMode and glActiveTexture Instead use new selector-free OpenGL commands Example: glMatrixMultfEXT and glTextureParameteriEXT Results in more read-able, less error-prone OpenGL code

Video Decode and Presentation API for Unix (VDPAU) interoperability Use NV_vdpau_interop extension to access VDPAU video and output surfaces for texturing and rendering Allows OpenGL to process and display contents of video streams decoded with VDPAU NV_gpu_program5 Assembly Extension Cg compile generates assembly for this extension (the gp5 * profiles) Allows you to off-line compile high-level shading languages to low-level assembly Better visibility into low-level execution environment and shading Instruction Set of Fermi More NVIDIA OpenGL Features to Explore (2)

More NVIDIA OpenGL Features to Explore (3) New low- and high-dynamic range Block Pixel Texture Compression (BPTC) formats ARB_texture_compression_bptc Bit-for-bit compatible with DirectX 11 block texture compression formats The BC6H and BC7 formats Random-access reads and writes to memory from shaders EXT_shader_image_load_store Includes atomic read-modify-write operations within shaders Introduces a glMemoryBarrierEXT operation to manage memory coherency

OpenGL 3D Graphics API cross-platform most functional peak performance open standard inter-operable well specified & documented 18+ years of compatibility

OpenGL 4 – DirectX 11 Superset Interop with a complete compute solution OpenGL is for graphics – CUDA / OpenCL is for compute 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 All of OpenGL ES 2.0 capabilities available on desktop EGL on Desktop in the works WebGL bridging desktop and mobile Cross platform Mac, Windows, Linux, Android, Solaris, FreeBSD Result of being an open standard Buffer and Event Interop

Other OpenGL-related Sessions at GTC 2158: Pipeline with Quadro Digital Video Pipeline with OpenGL Monday, 13:00, Room C 2113: WebGL: Bringing 3D to the Web Tuesday, 15:00, Room A8 2227: OpenGL 4.0 Tessellation for Professional Applications Tuesday, 15:00, Room A1 2056: Next-Generation Rendering with CgFX Tuesday, 16:00, Room C 2285: Disney Studios’ GPU-Accelerated Animatic Lighting Process with Soft Shadows and Depth of Field Wednesday, 17:00, Room L 2016: VDPAU: PureVideo on Unix Thursday, 15:00, Room A1

NVIDIA's OpenGL Functionality

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to NVIDIA's OpenGL Functionality (20)

More from Mark Kilgard (20)

Recently uploaded (20)

NVIDIA's OpenGL Functionality