I'm trying to implement one complex algorithm using GPU. The only problem is HW limitations and maximum available feature level is 9_3.
Algorithm is basically "stereo matching"-like algorithm for two images. Because of mentioned limitations all calculations has to be performed in Vertex/Pixel shaders only (there is no computation API available). Vertex shaders are rather useless here so I considered them as pass-through vertex shaders.
Let me shortly describe the algorithm:
Take two images and calculate cost volume maps (basically conterting RGB to Grayscale -> translate right image by D and subtract it from the left image). This step is repeated around 20 times for different D which generates Texture3D.
Problem here: I cannot simply create one Pixel Shader which calculates
those 20 repetitions in one go because of size limitation of Pixel
Shader (max. 512 arithmetics), so I'm forced to call Draw() in a loop
in C++ which unnecessary involves CPU while all operations are done on
the same two images - it seems to me like I have one bottleneck here. I know that there are multiple render targets but: there are max. 8 targets (I need 20+), if I want to generate 8 results in one pixel shader I exceed it's size limit (512 arithmetic for my HW).
Then I need to calculate for each of calculated textures box filter with windows where r > 9.
Another problem here: Because window is so big I need to split box filtering into two Pixel Shaders (vertical and horizontal direction separately) because loops unrolling stage results with very long code. Manual implementation of those loops won't help cuz still it would create to big pixel shader. So another bottleneck here - CPU needs to be involved to pass results from temp texture (result of V pass) to the second pass (H pass).
Then in next step some arithmetic operations are applied for each pair of results from 1st step and 2nd step.
I haven't reach yet here with my development so no idea what kind of bottlenecks are waiting for me here.
Then minimal D (value of parameter from 1st step) is taken for each pixel based on pixel value from step 3.
... same as in step 3.
Here basically is VERY simple graph showing my current implementation (excluding steps 3 and 4).
Red dots/circles/whatever are temporary buffers (textures) where partial results are stored and at every red dot CPU is getting involved.
Question 1: Isn't it possible somehow to let GPU know how to perform each branch form up to the bottom without involving CPU and leading to bottleneck? I.e. to program sequence of graphics pipelines in one go and then let the GPU do it's job.
One additional question about render-to-texture thing: Does all textures resides in GPU memory all the time even between Draw() method calls and Pixel/Vertex shaders switching? Or there is any transfer from GPU to CPU happening... Cuz this may be another issue here which leads to bottleneck.
Any help would be appreciated!
Thank you in advance.
Best regards,
Lukasz
Writing computational algorithms in pixel shaders can be very difficult. Writing such algorithms for 9_3 target can be impossible. Too much restrictions. But, well, I think I know how to workaround your problems.
1. Shader repetition
First of all, it is unclear, what do you call "bottleneck" here. Yes, theoretically, draw calls in for loop is a performance loss. But does it bottleneck? Does your application really looses performance here? How much? Only profilers (CPU and GPU) can answer. But to run it, you must first complete your algorithm (stages 3 and 4). So, I'd better stick with current solution, and started to implement whole algorithm, then profile and than fix performance issues.
But, if you feel ready to tweaks... Common "repetition" technology is instancing. You can create one more vertex buffer (called instance buffer), which will contains parameters not for each vertex, but for one draw instance. Then you do all the stuff with one DrawInstanced() call.
For you first stage, instance buffer can contain your D value and index of target Texture3D layer. You can pass-through them from vertex shader.
As always, you have a tradeof here: simplicity of code to (probably) performance.
2. Multi-pass rendering
CPU needs to be involved to pass results from temp texture (result of
V pass) to the second pass (H pass)
Typically, you do chaining like this, so no CPU involved:
// Pass 1: from pTexture0 to pTexture1
// ...set up pipeline state for Pass1 here...
pContext->PSSetShaderResources(slot, 1, pTexture0); // source
pContext->OMSetRenderTargets(1, pTexture1, 0); // target
pContext->Draw(...);
// Pass 2: from pTexture1 to pTexture2
// ...set up pipeline state for Pass1 here...
pContext->PSSetShaderResources(slot, 1, pTexture1); // previous target is now source
pContext->OMSetRenderTargets(1, pTexture2, 0);
pContext->Draw(...);
// Pass 3: ...
Note, that pTexture1 must have both D3D11_BIND_SHADER_RESOURCE and D3D11_BIND_RENDER_TARGET flags. You can have multiple input textures and multiple render targets. Just make sure, that every next pass knows what previous pass outputs.
And if previous pass uses more resources than current, don't forget to unbind unneeded, to prevent hard-to-find errors:
pContext->PSSetShaderResources(2, 1, 0);
pContext->PSSetShaderResources(3, 1, 0);
pContext->PSSetShaderResources(4, 1, 0);
// Only 0 and 1 texture slots will be used
3. Resource data location
Does all textures resides in GPU memory all the time even between
Draw() method calls and Pixel/Vertex shaders switching?
We can never know that. Driver chooses appropriate location for resources. But if you have resources created with DEFAULT usage and 0 CPU access flag, you can be almost sure it will always be in video memory.
Hope it helps. Happy coding!
Related
What i'm doing is GPGPU on WebGL and I don't know the access pattern which I'd be talking about applies to general graphics and gaming programs. In our code, frequently, we come across data which needs to be summarized or reduced per output texel. A very simple example is matrix multiplication during which, for every output texel, your return a value which is a dot product of a row of one input and a column of the other input.
This has been the sore point of our performance because of not so much the computation but multiplied data access. So I've been trying to find a pattern of reads or data layouts which would expedite this operation and I have been completely unsuccessful.
I will be describing some assumptions and some schemes below. The sample code for all these are under https://github.com/jeffsaremi/webgl-experiments
Unfortunately due to size I wasn't able to use the 'snippet' feature of StackOverflow. NOTE: All examples write to console not the html page.
Base matmul implementation: Example: [2,3]x[3,4]->[2,4] . This produces in a simplistic form 2 textures of (w:3,h:2) and (w:4,h:3). For each output texel I will be reading along the X axis of the left texture but going along the Y axis of the right texture. (see webgl-matmul.html)
Assuming that GPU accesses data similar to CPU -- that is block by block -- if I read along the width of the texture I should be hitting the cache pretty often.
For this, I'd layout both textures in a way that I'd be doing dot products of corresponding rows (along texture width) only. Example: [2,3]x[4,3]->[2,4] . Note that the data for the right texture is now transposed so that for each output texel I'd be doing a dot product of one row from the left and one row from the right. (see webgl-matmul-shared-alongX.html)
To ensure that the above assumption is indeed working, I created a negative test also. In this test I'd be reading along the Y axis of both left and right textures which should have the worst performance ever. Data is pre-transposed so that the results make sense. Example: [3,2]x[3,4]->[2,4]. (see webgl-matmul-shared-alongY.html).
So I ran these -- and I hope you could do as well to see -- and I found no evidence to support existence or non-existence of such caching behavior. You need to run each example a few times to get consistent results for comparison.
Then I came along this paper http://fileadmin.cs.lth.se/cs/Personal/Michael_Doggett/pubs/doggett12-tc.pdf which in short claims that the GPU caches data in blocks (or tiles as I call them).
Based on this promising lead I created a version of matmul (or dot product) which uses blocks of 2x2 to do its calculation. Prior to using this of course I had to rearrange my inputs into such layout. The cost of that re-arrangement is not included in my comparison. Let's say I could do that once and run my matmul many times after. Even this scheme did not contribute anything to the performance if not taking something away. (see webgl-dotprod-tiled.html).
A this point I am completely out of ideas and any hints would be appreciated.
thanks
I am trying to move from OpenGL to Metal for my iOS apps. In my OpenGL code I use glColorMask (if I want to write only to selected channels, for example only to alpha channel of a texture) in many places.
In Metal, for render pipeline (though vertex and fragment shader) seems like MTLColorWriteMask is the equivalent of glColorMask. I can setup it up while creating a MTLRenderPipelineState through the MTLRenderPipelineDescriptor.
But I could not find a similar option for compute pipeline (through kernel function). I always need to write all the channels (red, green, blue and alpha) every time I write to an output texture. What if I want to preserve the alpha (or any other channel) and only want to modify the color channels? I can create a copy of the output texture and use it as one of the inputs and read alpha from it to preserve the values but that is expensive.
Computer memory architectures don't like writing only some bytes of data. A write to 1 out of 4 bytes usually involves reading those four bytes into the cache, modifying one of them in the cache, and then writing those four bytes back out into memory. Well, most computers read/write a lot more than 4 bytes at a time, but you get the idea.
This happens with framebuffers too. If you do a partial write mask, the hardware is still going to be doing the equivalent of a read/modify/write on that texture. It's just not changing all of the bytes its reads.
So you can do the same thing from your compute shader. Read the 4-vector value, modify the channels you want, and then write it back out. As long as the read and write are from the same shader invocation, there should be no synchronization problems (assuming that no other invocations are trying to read/write to that same location, but if that were the case, you'd have problems anyway).
I am in the middle of rendering different textures on multiple meshes of a model, but I do not have much clues about the procedures. Someone suggested for each mesh, create its own descriptor sets and call vkCmdBindDescriptorSets() and vkCmdDrawIndexed() for rendering like this:
// Pipeline with descriptor set layout that matches the shared descriptor sets
vkCmdBindPipeline(...pipelines.mesh...);
...
// Mesh A
vkCmdBindDescriptorSets(...&meshA.descriptorSet... );
vkCmdDrawIndexed(...);
// Mesh B
vkCmdBindDescriptorSets(...&meshB.descriptorSet... );
vkCmdDrawIndexed(...);
However, the above approach is quite different from the chopper sample and vulkan's samples that makes me have no idea where to start the change. I really appreciate any help to guide me to a correct direction.
Cheers
You have a conceptual object which is made of multiple meshes which have different texturing needs. The general ways to deal with this are:
Change descriptor sets between parts of the object. Painful, but it works on all Vulkan-capable hardware.
Employ array textures. Each individual mesh fetches its data from a particular layer in the array texture. Of course, this restricts you to having each sub-mesh use textures of the same size. But it works on all Vulkan-capable hardware (up to 128 array elements, minimum). The array layer for a particular mesh can be provided as a push-constant, or a base instance if that's available.
Note that if you manage to be able to do it by base instance, then you can render the entire object with a multi-draw indirect command. Though it's not clear that a short multi-draw indirect would be faster than just baking a short sequence of drawing commands into a command buffer.
Employ sampler arrays, as Sascha Willems suggests. Presumably, the array index for the sub-mesh is provided as a push-constant or a multi-draw's draw index. The problem is that, regardless of how that array index is provided, it will have to be a dynamically uniform expression. And Vulkan implementations are not required to allow you to index a sampler array with a dynamically uniform expression. The base requirement is just a constant expression.
This limits you to hardware that supports the shaderSampledImageArrayDynamicIndexing feature. So you have to ask for that, and if it's not available, then you've got to work around that with #1 or #2. Or just don't run on that hardware. But the last one means that you can't run on any mobile hardware, since most of them don't support this feature as of yet.
Note that I am not saying you shouldn't use this method. I just want you to be aware that there are costs. There's a lot of hardware out there that can't do this. So you need to plan for that.
The person that suggested the above code fragment was me I guess ;)
This is only one way of doing it. You don't necessarily have to create one descriptor set per mesh or per texture. If your mesh e.g. uses 4 different textures, you could bind all of them at once to different binding points and select them in the shader.
And if you a take a look at NVIDIA's chopper sample, they do it pretty much the same way only with some more abstraction.
The example also sets up descriptor sets for the textures used :
VkDescriptorSet *textureDescriptors = m_renderer->getTextureDescriptorSets();
binds them a few lines later :
VkDescriptorSet sets[3] = { sceneDescriptor, textureDescriptors[0], m_transform_descriptor_set };
vkCmdBindDescriptorSets(m_draw_command[inCommandIndex], VK_PIPELINE_BIND_POINT_GRAPHICS, layout, 0, 3, sets, 0, NULL);
and then renders the mesh with the bound descriptor sets :
vkCmdDrawIndexedIndirect(m_draw_command[inCommandIndex], sceneIndirectBuffer, 0, inCount, sizeof(VkDrawIndexedIndirectCommand));
vkCmdDraw(m_draw_command[inCommandIndex], 1, 1, 0, 0);
If you take a look at initDescriptorSets you can see that they also create separate descriptor sets for the cubemap, the terrain, etc.
The LunarG examples should work similar, though if I'm not mistaken they never use more than one texture?
I'm looking at getting a program written for DirectX11 to play nice on DirectX10. To do that, I need to compile the shaders for model 4, not 5. Right now the only problem with that is that the geometry shaders use instancing which is unsupported by 4. The general model is
[instance(NUM_INSTANCES)]
void Gs(..., in uint instanceId : SV_GSInstanceID) { }
I can't seem to find many documents on why this exists, because my thought is: can't I just replace this with a loop from instanceId=0 to instanceId=NUM_INSTANCES-1?
The answer seems to be no, as it doesn't seems to output correctly, but besides my exact problem - can you help me understand why the concept of instancing exists. Is there some implication on the entire pipeline that instancing has beyond simply calling the main function twice with a different index?
With regards to why my replacement did not work:
Geometry shaders are annotated with [maxvertexcount(N)]. I had incorrectly assumed this was the vertex input count, and ignored it. In fact, input is determined by the type of primitive coming in, and so this was about the output. Before, if N was my output over I instances, each instance output N vertices. But now that I want to use a loop, a single instance outputs N*I vertices. As such, the answer was to do as I suggested, and also use [maxvertexcount(N*NUM_INSTANCES)].
To more broadly answer my question on why instances may be useful in a world that already has loops, I can only guess
Loops are not truly supported in shaders, it turns out - graphics card cores do not have a concept of control flow. When loops are written in shaders, the loop is unrolled (see [unroll]). This has limitations, makes compilation slower, and makes the shader blob bigger.
Instances can be parallelized - one GPU core can run one instance of a shader while another runs the next instance of the same shader with the same input.
If I am generating 0-12 triangles in a compute shader, is there a way I can stream them to a buffer that will then be used for rendering to screen?
My current strategy is:
create a buffer of float3 of size threads * 12, so can store the maximum possible number of triangles;
write to the buffer using an index that depends on the thread position in the grid, so there are no race conditions.
If I want to render from this though, I would need to skip the empty memory. It sounds ugly, but probably there is no other way currently. I know CUDA geometry shaders can have variable length output, but I wonder if/how games on iOS can generate variable-length data on GPU.
UPDATE 1:
As soon as I wrote the question, I thought about the possibility of using a second buffer that would point out how many triangles are available for each block. The vertex shader would then process all vertices of all triangles of that block.
This will not solve the problem of the unused memory though and as I have a big number of threads, the total memory wasted would be considerable.
What you're looking for is the Metal equivalent of D3D's "AppendStructuredBuffer". You want a type that can have structures added to it atomically.
I'm not familiar with Metal, but it does support Atomic operations such as 'Add' which is all you really need to roll your own Append Buffer. Initialise the counter to 0 and have each thread add '1' to the counter and use the original value as the index to write to in your buffer.