Data Storage and Transfer in OpenGL®Data Storage and Transfer in OpenGL®

Kurt AkeleyNVIDIA CorporationNVIDIA U Presentation, 25 July 2003

NVIDIA CONFIDENTIAL

Outline

Data movement is the issueMemory typesDesign examples

Buffer objectsSuperbuffers (in process)

Programmable GPU memoryQ & A

Data MovementData Movement

NVIDIA CONFIDENTIAL

Semiconductor Scaling Rates

From: Digital Systems Engineering, Dally and Poulton

31.28Aggregate off-chip bandwidth

71.11750Pins per package

1.31.71Die-length wire delay / gate delay

1.00Device-length wire delay

1.31.71Capability (grids / gate delay)

(5)0.87150 pSGate Delay

1.751.491 BMoore’s Law (grids on a die)**

Years to Double (Half)

Yearly Factor

Current ValueParameter

** Ignores multi-layer metal, 8-layers in 2001

NVIDIA CONFIDENTIAL

Two Approaches

Move data faster (optimize speed)Point-to-point wiringAdvanced protocols (e.g. clock in data)Wide interfaces (256-bit GPUs)

Move data less (optimize locality)AlgorithmArchitecture (e.g. pipeline GPU)Cache data

NVIDIA CONFIDENTIAL

It’s All Cache!

Memory TypesMemory Types

NVIDIA CONFIDENTIAL

OpenGL’s Memory Types

Byte memoryClient pixel data

Element memoryFramebufferTexture object’s image

OthersDisplay list memory...

NVIDIA CONFIDENTIAL

Memory Properties

Server sideClient side *Typical location

Not allowedMay be allowedOut-of-order ops

Not allowedMay be allowedCPU mapping

PersistentPer-transactionFormatting

Component, indexByteBasic unit

Element memoryByte memoryProperty

Client pixel data Framebuffer

* We’ll see an exception soon

Buffer ObjectsBuffer Objects

NVIDIA CONFIDENTIAL

Classic Immediate Mode

ConvenientFlexible

Match client data structuresPer-vertex / per-primitive

SlowVector form slightly faster

float loc[2] = {2, 3};

glColor3f(1, 1, 1);

glBegin(GL_LINES);

glVertex2f(0, 1);

glVertex2fv(loc);

glEnd();

NVIDIA CONFIDENTIAL

Vertex Arrays

glVertexPointer(size, type, stride, *pointer);

size

type

stride

pointer

glDrawArrays(mode, first, count);

glEnableClientState(GL_VERTEX_ARRAY);

enable

NVIDIA CONFIDENTIAL

Vertex Array Problems

Client-side array is far from pipelineHard for application to use the right memory

On PC, AGP memory

NVIDIA CONFIDENTIAL

Solutions

NV_vertex_array_rangeMake vertex arrays work betterRelax coherence rules to enable pullingHelp client to position arrays in AGP memory

Array objectsExtend vertex arrays with element memory

NVIDIA CONFIDENTIAL

Good Qualities of Vertex Arrays

Easy to mix and match array dataShare vertex array with multiple color arraysDisable any array for debugging

Memory mappedClient can cherry-pick updatesExtra copies avoidedGreat for immediate mode rendering

Easy to mix usage modelsStatic array dataDynamic array data

NVIDIA CONFIDENTIAL

ARB Solution: Server-side Buffer Objects

Byte memory (mappable)Server side (efficient for rendering, sharable)

ApplicationByte

Memory+ State

Buffer Object

GL ServerGL Client

NVIDIA CONFIDENTIAL

Modifying Buffer Object Data

FunctionalglBufferSubDataARB(target, offset, size, *data);

glGetBufferSubDataARB(target, offset, size, *data);

Always safeMemory mapped

void *glMapBufferARB(target, access);

GLboolean *glUnmapBufferARB(target);

May result in data lossMay not be faster

NVIDIA CONFIDENTIAL

Mapping Rules

Be prepared for data lossMap for brief periods only

Map it, modify it, then unmap itDon’t render from a mapped buffer

Don’t pass a map pointer to the GL

NVIDIA CONFIDENTIAL

Store Vertex Arrays in Buffers

VertexPointer(size, type, stride, *pointer);

size

type

stride

pointer

bufferARRAY_BUFFER_ARB

offset

NVIDIA CONFIDENTIAL

Client and Server State

Application

Buffer Objects

GL ServerGL Client

ARRAY_BUFFER_ARB

size

type

stride

pointer

buffer

Buffer objects are server stateVertex arrays parameters are client state

size

type

stride

pointer

buffer

size

type

stride

pointer

buffer

size

type

stride

pointer

buffer

Byte Memory+ State

SuperbuffersSuperbuffers(In Process)(In Process)

NVIDIA CONFIDENTIAL

The OpenGL “Machine Tool”

Point, Line,Polygon

Rasterization

UnpackPixels

PackPixels

PixelOperations

ImageRasterization

TextureMemory

FragmentOperations

UnpackVertexes

VertexOperations

Framebuffer

Image

Geometry

NVIDIA CONFIDENTIAL

Problems

Framebuffer to texture transferCopy is too slowCurrent acceleration methods (pbuffers) are

Window-system specificAwkward to use

Framebuffer to geometry transferCopy is too slowThere are no acceleration methods

Render to texture

NVIDIA CONFIDENTIAL

Why Not Add Destructive Copy?

Virtue of simplicityIssues

Can’t be anticipatedForces 2x memory when not otherwise needed

NVIDIA CONFIDENTIAL

Superbuffer Solution

Invent framebuffer objectsMuch like texture objects

Invent element memory objects1D, 2D, or 3D images

Attach element memory objectsTo texture objects, as image levelsTo framebuffer objects, as color, depth, and stencil buffers

NVIDIA CONFIDENTIAL

Superbuffers

Texture Object

Memory Object

FramebufferObject

Buffer Object

?

Can element memory objects be attached to

buffer objects?

Can element memory objects be attached to

buffer objects?

GPU MemoryGPU Memory

NVIDIA CONFIDENTIAL

So Far So Good

Byte / element distinction makes senseTexture is element memory

Always accessed through functional interfaceFramebuffer is element memory

Rendered through functional interfaceCannot be read by vertex/fragment processor

Language arrays are “byte” memory

NVIDIA CONFIDENTIAL

Summary

NVIDIA is committed to OpenGL

OpenGL is evolvingBuffer objects will be in 1.5, due this summerSuperbuffers are underway, no target date

Byte / element memory distinction has been useful

NVIDIA CONFIDENTIAL

For More Information

www.opengl.orghttp://developer.nvidia.com/view.asp?IO=presentations