Graphics acceleration

An example of line-drawing by the ATI Radeon’s 2D graphics engine

Bresenham’s algorithm

• Recall this iterative algorithm for doing a ‘scanline conversion’ for a straight line

• It required five parameters:– The starting endpoint coordinates: (X0,Y0)– The ending endpoint coordinates: (X1,Y1)– The foreground color for the solid-color line

• It begins by initializing a decision-variable– errorTerm = 2*deltaY - deltaX;

Algorithm’s main loop

for (int y = Y0, x = X0; x <= X1; x++)

{

drawPixel( x, y, color );

if ( errorTerm >= 0 ) { errorTerm += 2*delY; }

else { y += 1; errorTerm += 2*(delY – delX); }

}

How much work for CPU?

• Example: To draw the longest visible line (in 1024x768 graphics mode) will require approximately 10,000 CPU instructions

• The loop gets executed once for each of the 1024 horizontal pixels, and each pass through that loop requires about ten CPU operations: moves, compares, branches, adds and subtracts, plus the function-calls

Is acceleration possible?

• The IBM 8514/A appeared in late 1980s

• It could do line-drawing (and some other common graphics operations) if just a few parameters were supplied

• So instead of requiring the CPU to do ten thousand operations, the CPU could do maybe ten operations, then let the 8514/A graphics engine do the rest of the work!

8514/A Block Diagram

VRAMmemory

ROM

RAMDAC

Display Monitor

PC Bus Interface

LUT DAC

Display processor

Drawing engine

CRT controller

PC Bus

Graphicsprocessor

CPU

ATI improved on IBM’s 8514/A

• Various OEM vendors soon introduced their own graphics accelerator designs

• Because IBM had not released details of its design, others had to create their own programming interfaces – all are different

• Early PC graphics software was therefore NOT portable between hardware platforms

How does X300 draw lines?

• To demonstrate the line-drawing ability of our classroom’s Radeon X300 graphics processors, we wrote ‘drawline.cpp’ demo

• We did not have access to ATI’s official Radeon programming manual, but we had several such manuals from other vendors, and we found ‘clues’ in source-code files for the Linux Radeon device-driver

Programming concepts

• Our demo-program must first verify that it is running on a Radeon-equipped machine

• It must determine how it can communicate with the Radeon’s graphics accelerator

• Normal VGA registers are at ‘standard’ I/O port-addresses, but the graphics engine is outside the scope of established standards

Peripheral Component Interconnect

• An industry committee (led by Intel) has established a standard mechanism that PC device-drivers can use to identify the peripheral devices that a workstation has, and their mechanisms for communication

• To simplify the Pre-Boot Execution code, modern PC’s provide ROM-BIOS routines that can be called to identify peripherals

PCI Configuration Space

ADDITIONALPCI

CONFIGURATIONDATA

Each peripheral device has a set of nonvolatile memory-locationswhich store information about that device using a standard layout

PCI CONFIGURATION HEADER

1024bytes

256 bytes

This device-information is accessed via I/O Port-Addresses 0x3C8-0x3CF

PCI Configuration Header

BASE-ADDRESSRESOURCE 0

BASE-ADDRESSRESOURCE 1

BASE-ADDRESSRESOURCE 2

BASE-ADDRESSRESOURCE 3

DEVICEID

VENDORID

VENDOR-ID = 0x1002: Advanced Technologies, Incorporated DEVICE-ID = 0x5B60: ATI Radeon X300 graphics processor BASE-ADDRESS for RESOURCE 1 is the 2D engine’s I/O port

Sixteen longword entries (256 bytes)

Our ‘findsvga.cpp’ utility will show you the PCI Configuration Space for anyperipheral devices of Class 0x030000 (i.e., VGA-compatible graphics cards)

Interface to PCI BIOS

• Our ‘dosio.c’ device-driver (and ‘int86.cpp’ companion code) allow us access to BIOS

• The PCI BIOS services are accessible (in the Pentium’s virtual-8086 mode) using function 0xB1 of software interrupt 0x1A

• There are several subfunctions – you can find documentation online – for example, Professor Ralf Brown’s Interrupt List

return_radeon_port_address();

• Our demo invokes these PCI ROM-BIOS subfunctions to discover which I/O Port our Radeon’s 2D graphics engine uses – Subfunction 1: Detect BIOS presence– Subfunction 3: Find Device in a Class– Subfunction A: Read Configuration Dword

• Configuration Dword at offset 0x14 holds I/O Port-Address for 2D graphics engine

The ATI I/O Port Interface

MM_INDEX MM_DATA

iobase + 0 iobase + 4

You output a register’s index to the iobase + 0 address

Then you have read or write access to that register at the iobase + 4 address

Many 2D engine registers!

• You can peruse the ‘radeon.h’ header-file to see names and register-index numbers for the Radeon 2D graphics accelerator

• You could also write a programming loop to input the contents from various offsets and thereby get some idea of which ones appear to hold ‘live’ values (i.e.,hundreds!)

• Only a small number used in line-drawing

Main Line-Drawing registers

• DP_GUI_MASTER_CNTL

• DP_BRUSH_FRGD_COLOR

• DP_BRUSH_BKGD_COLOR

• DP_WRITE_MSK

• DST_LINE_START

• DST_LINE_END

Others that affect drawing

• RB2D_DSTCACHE_MODE• MC_FB_LOCATION• DEFAULT_PITCH_OFFSET• DST_PITCH_OFFSET• SRC_PITCH_OFFSET• DP_DATATYPE• DEFAULT_SC_TOP_LEFT• DEFAULT_SC_BOTTOM_RIGHT

CPU/GPU synchronization

IntelPentium

CPU

ATIRadeon

GPU

When CPU off-loads the work of drawing lines (and doing other commonGraphical operations) tp the Graphics Processing Unit, then this frees up the CPU to execute other instructions – but it opens up the possibility thatthe CPU will send more drawing commands to the GPU, even before theGPU is finished doing earlier commands. Some mechanism is needed toprevent the GPU from becoming overwhelmed by work the CPU sends it.

Solution is a FIFO for pending commands, plus a Status Register

Engine has 64 FIFO slots

• Before the CPU initiates a new drawing command, it checks to see if there are enough free slots in the command FIFO for storing that command’s parameters

• The CPU can do ‘busy-waiting’ until the GPU reports that enough FIFO slots are ready to accept new command-arguments

• An alternative is ‘interrupt-driven’ drawing

Testing ‘drawline.cpp’

• We developed our ‘drawline.cpp’ demo on a Radeon 7000 graphics card, then tested it on a newer and faster Radeon 9250

• Our code worked fine

• Tonight we shall try it on the Radeon X300

• If these various models of the Radeon are fully compatible with one another, we can expect our demo to work fine on the X300

Hardware changes?

• But if any significant differences exist in the various Radeon design-generations, then we may discover that our ‘drawline’ fails to perform properly on an X300

• We would then have to explore the ways in which Radeon designs have changed, and try to devise ‘fixes’ for any flaws that we have found in our software application

In-class exercises

• Try running the ‘drawline.cpp’ application on our classroom or CS Lab workstation: maybe it works fine, maybe it doesn’t

• Look at the source-code files for the Linux ‘open-source’ ATI Radeon device-driver

• If our ‘drawline’ work ok, see if you can add code that programs the engine to fill rectangles or copy screen-areas; or, if ‘drawline’ fails, see if you can devise a ‘fix’