How do fragments get generated by rasterizer in OpenGL

Question

I came across the description of rasterization and it basically says that when an object gets projected onto screen that what happens is that a scan takes place over all the pixels on the window/screen and decides if the pixel/fragment is within the triangle and hence determines that the pixel/fragment is inside the triangle and follows with further processing for the pixel/fragment such as colouring etc

Now since i am studying OpenGL and i do know that OpenGL probably has its own implementations of this process i was wondering whether this also takes place with OpenGL since of the "Scan-Conversion" process of vertices that i have read in OpenGL tutorial

Now another question related to this i have is that i know that the image/screen/window of pixels is an image or 2d array of pixels also known as the default framebuffer that is linear

So what i am wondering is if that is the case, how would projecting the 3 vertices of a triangle define which pixels are covered in side it?

Does the rasterizer draw the edges of a triangle first and then scans through each pixel or 2d array of pixels (also known as the default framebuffer) and sees if the points are between the lines using some mathematical method or some other simpler process happens?

Answer 1

and i do know that OpenGL probably has its own implementations of this process

OpenGL is just a specification document. What runs on a computer is an OpenGL implementation, most of the time as part of a GPU driver. The actual workload is carried out by a GPU…

this also takes place with OpenGL since of the "Scan-Conversion" process of vertices that i have read in OpenGL tutorial

most likely not. As a matter of fact last weekend I was attending a Khronos (the group that specifies OpenGL) event hosted by AMD and one of AMD's GPU engineers was lamenting that newbies have the scanline algorithm in mind with OpenGL, Direct3D, Mantel, Vulkan, etc., while GPUs do something entirely different.

2d array of pixels also known as the default framebuffer that is linear

actually the memory layout of pixels as used internally by the GPU is not linear (i.e. row-by-row) but follows a pattern that gives efficient localized access. For linear access GPUs have extremely efficient copy engines that allow for practically zero overhead conversion between the internal and linear format.

The exact layout used internally is a detail only the GPU engineers have insight into, though. But the fact that memory is not organized linearly but in a localized fashion is also one reason, that the traditional scanline algorithm is not used by GPUs.

So what i am wondering is if that is the case, how would projecting the 3 vertices of a triangle define which pixels are covered in side it?

Any method that satisfies the requirements of the OpenGL specification is allowed. The details are part of the OpenGL implementation, i.e. usually the combination of particular GPU model and driver version.

Answer 2

The scanline algorithm is what people did in software back in the 1990s, before modern GPUs. GPU developers figured out rather quickly that the algorithms you use for software rendering are vastly different from the algorithms you would implement in a VLSI implementation with billions of transistors. Algorithms optimized for hardware implementation tend to look fairly alien to anyone who comes from a software background anyway.

Another thing I'd like to clear up is that OpenGL doesn't say anything about "how" you render, it's just "what" you render. OpenGL implementations are free to do it however they please. We can find out "what" by reading the OpenGL standard, but "how" is buried in secrets kept by the GPU vendors.

Finally, before we start, the articles you linked are unrelated. They are about how ultrasonic scans work.

What do we know about scan conversion?

• Scan conversion has as input a number of primitives. For our purposes, let's assume that they're all triangles (which is increasingly true these days).

• Every triangle must be clipped by the clipping planes. This can add up to three additional sides to the triangle, in the worst case (turning it into a hexagon). This has to happen before perspective projection.

• Every primitive must go through perspective projection. This process takes each vertex with homogeneous coordinates (X, Y, Z, W) and converts it to (X/W, Y/W, Z/W).

• The framebuffer is usually organized hierarchically into tiles, not linearly like the way you do in software. Furthermore, the processing might be done at more than one hierarchical level. The reason why we use linear organization in software is because it takes extra cycles to compute memory addresses in a hierarchical layout. However, VLSI implementations do not suffer from this problem, they can simply wire up the bits in a register how they want to make an address from it.

So you can see that in software, tiles are "complicated and slow" but in hardware they're "easy and fast".

Some notes looking at the R5xx manual:

The R5xx series is positively ancient (2005) but the documentation is available online (search for "R5xx_Acceleration_v1.5.pdf"). It mentions two scan converters, so the pipeline looks something like this:

primitive output -> coarse scan converter -> quad scan converter -> fragment shader

The coarse scan converter appears to operate on larger tiles of configurable size (8x8 to 32x32), and has multiple selectable modes, an "intercept based" and a "bounding box based" mode.

The quad scan converter then takes the output of the coarse scan converter and outputs individual quads, which are groups of four samples. The depth values for each quad may be represented as four discrete values or as a plane equation. The plane equation allows the entire quad to be discarded quickly if the corresponding quad in the depth buffer also is specified as a plane equation. This is called "early Z" and it is a common optimization.

The fragment shader then works on one quad at a time. The quad might contain samples outside the triangle, which will then get discarded.

It's worth mentioning again that this is an old graphics card. Modern graphics cards are more complicated. For example, the R5xx doesn't even let you sample textures from the vertex shaders.

If you want an even more radically different picture, look up the PowerVR GPU implementations which use something called "tile-based deferred rendering". These modern and powerful GPUs are optimized for low cost and low power consumption, and they challenge a lot of your assumptions about how renderers work.

Answer 3

Quoting from GPU Gems: Parallel Prefix Sum (Scan) with CUDA, it describes how OpenGL does its scan conversion and compares it with CUDA which I think suffices as the answer of my question:

Prior to the introduction of CUDA, several researchers implemented scan using graphics APIs such as OpenGL and Direct3D (see Section 39.3.4 for more). To demonstrate the advantages CUDA has over these APIs for computations like scan, in this section we briefly describe the work-efficient OpenGL inclusive-scan implementation of Sengupta et al. (2006). Their implementation is a hybrid algorithm that performs a configurable number of reduce steps as shown in Algorithm 5. It then runs the double-buffered version of the sum scan algorithm previously shown in Algorithm 2 on the result of the reduce step. Finally it performs the down-sweep as shown in Algorithm 6.

Example 5. The Reduce Step of the OpenGL Scan Algorithm

1: for d = 1 to log2 n do 
2:     for all k = 1 to n/2 d  – 1 in parallel do 
3:          a[d][k] = a[d – 1][2k] + a[d – 1][2k + 1]]
Example 6. The Down-Sweep Step of the OpenGL Scan Algorithm

1: for d = log2 n – 1 down to 0 do 
2:     for all k = 0 to n/2 d  – 1 in parallel do 
3:          if i > 0 then 
4:             if k mod 2 U2260.GIF 0 then 
5:                  a[d][k] = a[d + 1][k/2]
6:             else 
7:                  a[d][i] = a[d + 1][k/2 – 1]

The OpenGL scan computation is implemented using pixel shaders, and each a[d] array is a two-dimensional texture on the GPU. Writing to these arrays is performed using render-to-texture in OpenGL. Thus, each loop iteration in Algorithm 5 and Algorithm 2 requires reading from one texture and writing to another.

The main advantages CUDA has over OpenGL are its on-chip shared memory, thread synchronization functionality, and scatter writes to memory, which are not exposed to OpenGL pixel shaders. CUDA divides the work of a large scan into many blocks, and each block is processed entirely on-chip by a single multiprocessor before any data is written to off-chip memory. In OpenGL, all memory updates are off-chip memory updates. Thus, the bandwidth used by the OpenGL implementation is much higher and therefore performance is lower, as shown previously in Figure 39-7.