• Advertisement
Sign in to follow this  

DX11 D3DCompile() way too slow

This topic is 2557 days old which is more than the 365 day threshold we allow for new replies. Please post a new topic.

If you intended to correct an error in the post then please contact us.

Recommended Posts

So I am doing some implicit function evaluation with perlin noise with DirectCompute (on DX11 hw, cs_5 profile), and the whole thing takes about 2 minutes to complete. I was starting to get a bit dissapointed with the speed of my GPU, until I found out that 99% of that time wasn't spent in the shader, but in... D3DCompile()

My code is only about 5kb of source, and thru experimentation I have found out that any loops have the biggest influence on compile time: it is as if compile time is exponential to a version of the code with all loops unrolled and function calls inlined. Optimizing the loop counts where I could, I could get the compile time down to 1 minute.

tagging all the loops with [loop] had no effect, and neither had any of the D3D10_SHADER_SKIP_OPTIMIZATION or D3D10_SHADER_OPTIMIZATION_LEVEL0 flags. The only thing that had an impact was [fastopt], which reduced compilation time about 4x down to 15 seconds.

Thing is, I can't precompile this code to a blob, because the whole point of my program is to interactively change the code and see the result instantly (or at least fast, failing that). It is a bit ridiculous if the user has to wait 16 seconds to see the new result, of which only 1 second was actual computation.

How do I speed this up? I don't mind a lower optimisation level as clearly that will only ever be a win for me.

I have written optimizing compilers myself, and I know of no compiler that would take longer than a second on such small amount of code (on modern hw). Certainly optimizing hlsl is trivial compared to C/C++, because there's no aliasing problems etc. What the hell can it possibly be spending all that time on?

Oh, and I have searched google and this forum for solutions, but sofar haven't found any (e.g. http://www.gamedev.n..._1#entry4725295 and http://www.gamedev.n..._1#entry4574352)

Share this post


Link to post
Share on other sites
Advertisement
I'd consider posting your shader source if you're not trying to keep it secret, maybe there is some way it can be made simpler or cleaner to help the compiler figure it out faster.

Share this post


Link to post
Share on other sites
Try the [fastopt] attribute on your loops. This should short circuit the loop simulator in the compiler which would otherwise be exponential for the number of embedded loops.

Docs for this are here

Share this post


Link to post
Share on other sites

I'd consider posting your shader source if you're not trying to keep it secret, maybe there is some way it can be made simpler or cleaner to help the compiler figure it out faster.


I don't mind posting it, but like I said, I already experimented with commenting out code to see what effect it has on compilation speed, and it is purely due to for-loops (which I can't remove) and the amount of code in general. I can probably optimize the code more, but I want compilation to be a lot faster, not just a bit.

I guess I am hoping for some secret compiler flags that tone down whatever crazy stuff it is doing?



cbuffer consts
{
float4 sb;
float4 grad[16];
};

float4 Col(float3 c, float inside) { return float4(c, inside); }
bool Solid(float4 c) { return c.w > 0.999f; }

float rand(float2 co)
{
return frac(sin(dot(co.xy, float2(12.9898, 78.233))) * 43758.5453);
}

int randint(float2 co) { return int(rand(co) * 256); }

int Hash( float3 P )
{
return randint(P.xy) ^ randint(P.yz) ^ randint(P.zx);
}

float Snoise3D( float3 P )
{
const float F3 = 0.333333333333;
const float G3 = 0.166666666667;

float s = dot( P, F3 );
float3 Pi = floor( P + s );
float t = dot( Pi, G3 );

float3 P0 = Pi - t;
float3 Pf0 = P - P0;

float3 simplex[4];
float3 T = Pf0.xzy >= Pf0.yxz;
simplex[0] = 0;
simplex[1] = T.xzy > T.yxz;
simplex[2] = T.yxz <= T.xzy;
simplex[3] = 1;

float n = 0;

[loop][fastopt]
for (int i = 0; i<4; i++)
{
float3 Pf = Pf0 - simplex + G3 * i;
int h = Hash( Pi + simplex );
float d = saturate( 0.6f - dot( Pf, Pf ) );
d *= d;
n += d * d * dot((float3)grad[ h & 15 ], Pf);
}

return 32.0f * n;
}

float Turbulence3D( float3 p )
{
float res = 0;
float fact = 1;
float scale = 1;
float weight = 0;
[loop][fastopt]
for (int i = 0; i<4; i++)
{
res += fact * Snoise3D( p * scale );
weight += fact;
fact /= 2.5;
scale *= 2.5;
}
return res / weight;
}

float4 Fun(float3 p)
{
p = (p - sb.y) / sb.x;

return Col(0.5, Turbulence3D(p) < -0.3);
}

struct BufferStruct2
{
float4 a;
float4 b;
};

RWStructuredBuffer<BufferStruct2> g_OutBuff2 : register( u1 );

[numthreads(64, 1, 1)]
void main2( uint3 threadIDInGroup : SV_GroupThreadID,
uint3 groupID : SV_GroupID,
uint groupIndex : SV_GroupIndex,
uint3 dispatchThreadID : SV_DispatchThreadID )
{
BufferStruct2 e = g_OutBuff2[dispatchThreadID.x];
float3 a = (float3)e.a;
float3 b = (float3)e.b;

int subdivs = 10;
[loop][fastopt]
for (int i = 0; i < subdivs; i++)
{
float3 mid = (a + B) / 2;
float4 col = Fun(mid);
if (Solid(col)) a = mid;
else b = mid;
}

float3 pos = (a + B) / 2;


float3 col = float3(0, 0, 0);
int n = 0;
const float off = 0.6f;
[loop][fastopt]
for (float x = -off; x < off + 0.1f; x += off)
[loop][fastopt]
for (float y = -off; y < off + 0.1f; y += off)
[loop][fastopt]
for (float z = -off; z < off + 0.1f; z += off)
{
float4 acol = Fun(float3(x, y, z) + pos);
col += (float3)(acol * acol.w);
n += acol.w;
}
if (n > 1) col /= n;

e.a = float4(pos, 0);
e.b = float4(col, 0);
g_OutBuff2[dispatchThreadID.x] = e;
}

Share this post


Link to post
Share on other sites
I don't know about any "secret compiler flags", but if [fastopt] doesn't work for you then you can always just specify a lower optimization level. I *think* the default is D3D10_SHADER_OPTIMIZATION_LEVEL1, and you can specify 0, 2, or 3. You can also just disable optimizations altogether.

Share this post


Link to post
Share on other sites

Try the [fastopt] attribute on your loops. This should short circuit the loop simulator in the compiler which would otherwise be exponential for the number of embedded loops.

Docs for this are here

Thanks, but if you see my original post, I am already using that. It helps some, but not enough.

Share this post


Link to post
Share on other sites

I don't know about any "secret compiler flags", but if [fastopt] doesn't work for you then you can always just specify a lower optimization level. I *think* the default is D3D10_SHADER_OPTIMIZATION_LEVEL1, and you can specify 0, 2, or 3. You can also just disable optimizations altogether.


Thanks, but as I said in my original post, I tried both D3D10_SHADER_OPTIMIZATION_LEVEL0 and D3D10_SHADER_SKIP_OPTIMIZATION, and neither appear to have any effect on compilation time.

Share this post


Link to post
Share on other sites
On my PC (Core i7 920) pasting that code into ATI's GPU Shader Analyzer and compiling as cs_5_0 (and changing 'B' to 'b' to make it compile) gives me a compile time of about 5 seconds, regardless of settings. If it's taking you 15 seconds a CPU upgrade might help, or you're compiling it more than once...

This drops down to almost instant if I take out the code in the middle of those triple nested loops at the bottom, so that's clearly the slow bit to compile. Unfortunately playing with that code I was unable to noticeably speed up compilation time, without removing code.

You might want to consider some sort of cache of compilation results, but that's not easy if you want to ignore changes to the source that won't affect the compiled result like adding whitespace.

Another option is using extra threads to compile the code in the background after each change, to minimize the delay that the user sees.

Share this post


Link to post
Share on other sites

On my PC (Core i7 920) pasting that code into ATI's GPU Shader Analyzer and compiling as cs_5_0 (and changing 'B' to 'b' to make it compile) gives me a compile time of about 5 seconds, regardless of settings. If it's taking you 15 seconds a CPU upgrade might help, or you're compiling it more than once...

This drops down to almost instant if I take out the code in the middle of those triple nested loops at the bottom, so that's clearly the slow bit to compile. Unfortunately playing with that code I was unable to noticeably speed up compilation time, without removing code.

You might want to consider some sort of cache of compilation results, but that's not easy if you want to ignore changes to the source that won't affect the compiled result like adding whitespace.

Another option is using extra threads to compile the code in the background after each change, to minimize the delay that the user sees.


Thanks for trying that out. Yes, the version I sent was already optimized somewhat from the 15 second version, and takes 5.8 seconds to compile on mine.

Yeah, a cache won't work, as any part of the code may change from run to run, so it will never hit the cache. If DirectCompute had some form of "linking", I could compile parts of the code that never change separately, but I don't think that's possible either.

Seeing as how none of the compiler flags affect compilation speed, it is clearly ignoring my request not to try to optimize those loops.

Share this post


Link to post
Share on other sites
Some of the loop analysis can't be turned off regardless of the flags you specify. Anyway, compiler performance is being worked on for a future release as it has major issues especially for compute shaders. At the moment you won't be able to eak out better performance through anything other than HLSL code change to your algorithm which may or may not be possible.

Share this post


Link to post
Share on other sites

Some of the loop analysis can't be turned off regardless of the flags you specify. Anyway, compiler performance is being worked on for a future release as it has major issues especially for compute shaders. At the moment you won't be able to eak out better performance through anything other than HLSL code change to your algorithm which may or may not be possible.

That is good to hear. Making the compiler obey flags of cheap optimization (just basic constant folding & inlining) would be fantastic.

Rather than reducing the amount of code, my project was planned to involve extending this code significantly. I guess I will have to implement it some other way, as this clearly will never get me fast turn arounds.

Does anyone have experience with OpenCL having a fast minimal optimization mode? I'd prefer to avoid OpenCL if I can, it appears pretty messy compared to DirectCompute. I can't use CUDA-C, since I'm on an ATI chip. Either that, or it is back to CPU code for me just for turn-around's sake :(

Share this post


Link to post
Share on other sites

[quote name='DieterVW' timestamp='1295651691' post='4762706']
Some of the loop analysis can't be turned off regardless of the flags you specify. Anyway, compiler performance is being worked on for a future release as it has major issues especially for compute shaders. At the moment you won't be able to eak out better performance through anything other than HLSL code change to your algorithm which may or may not be possible.

That is good to hear. Making the compiler obey flags of cheap optimization (just basic constant folding & inlining) would be fantastic.

Rather than reducing the amount of code, my project was planned to involve extending this code significantly. I guess I will have to implement it some other way, as this clearly will never get me fast turn arounds.

Does anyone have experience with OpenCL having a fast minimal optimization mode? I'd prefer to avoid OpenCL if I can, it appears pretty messy compared to DirectCompute. I can't use CUDA-C, since I'm on an ATI chip. Either that, or it is back to CPU code for me just for turn-around's sake :(
[/quote]
I've never tried it, but could your program benefit from the dynamic linkage portion of D3D11? If your program can be chopped up into sections, then you could utilize the dynamic linkage to simply replace portions of the program when they are edited. This would require defining interfaces in your program for certain sections of it, but it could eliminate the expensive recompilation portion that is eating up so much time...

Share this post


Link to post
Share on other sites

I've never tried it, but could your program benefit from the dynamic linkage portion of D3D11? If your program can be chopped up into sections, then you could utilize the dynamic linkage to simply replace portions of the program when they are edited. This would require defining interfaces in your program for certain sections of it, but it could eliminate the expensive recompilation portion that is eating up so much time...


The DynamicShaderLinkage11 uses #include to compile the interfaces/classes and the code that uses it in a single D3DX11CompileFromFile call, which probably wouldn't speed things up much. So the question is, can I link a class instance defined in one separately compiled blob to my main shader in another blob?

I can set a ID3D11ClassInstance thru CSSetShader. I obtain such a pointer using GetClassInstance() on a ID3D11ClassLinkage, which can then find any instances in code passed thru CreateComputeShader with that ID3D11ClassLinkage as an argument.

However I don't think I can call CreateComputeShader / D3DX11CompileFromFile on a file that only contains interfaces and classes, because they assume and entrypoint? I could add a dummy entrypoint, but something tells me ID3D11ClassLinkage was not intended to work way (referring to a shader that's not bound to the pipeline at all)... or am I missing something?

And even if that works, the sample documentation says:

"The Direct3D 11 runtime efficiently links each of the selected methods at source level, inlining and optimizing the shader code as much as possible to provide an optimal shader for the GPU to execute."

Which makes sense, because for PS use no-one would use interfaces if they were slower than the equivalent inlined code. So does that mean that potentially part of all that slowness would return at CSSetShader() time?

Share this post


Link to post
Share on other sites
the linking will only help you if the code can be written and compiled once. If you're letting the user write new code directly then this won't work. Otherwise the functionality could be split across a large number of class which can be bound in any arbitrary order from the runtime to meet the currently requested functionality. The shader will have to have all the code available when it compiles but you will have the ability to change how the shader operates from the runtime side. Perhaps it wold help to know more about your tool/application.

Share this post


Link to post
Share on other sites
There is one way I've spotted that you could effectively split some of that shader into multiple pieces that you can 'link together - convert some of the functions to texture lookups. For example you could try creating a 2D texture which you sample to calculate the results of randint(x, y).

Share this post


Link to post
Share on other sites

the linking will only help you if the code can be written and compiled once. If you're letting the user write new code directly then this won't work. Otherwise the functionality could be split across a large number of class which can be bound in any arbitrary order from the runtime to meet the currently requested functionality. The shader will have to have all the code available when it compiles but you will have the ability to change how the shader operates from the runtime side. Perhaps it wold help to know more about your tool/application.

The tool is an "implicit function modeller". The idea is to compositionally build complex models out of code rather than traditional modelling. The user writes code (at the moment in HLSL, but could be a more friendly special purpose language in the future that gets translated to HLSL on the fly), and the app shows what that looks in 3D like at the press of a key. To do so, it evaluates the function for however many locations in 3D space such that a good looking marching cubes mesh results (the loops in the above code for example sample the function many times to get accurate location of the isosurface on an edge, and anti-aliased color sampling).

One thing I could for a preview is a version that has less loops and thus looks uglier.

So the HLSL code changes constantly, and moreover, can become quite big is someone tries to model something complex (say a building, using many "union" operators). So compile time is critical, and would only get worse from this small example.

One way I could imagine hiding compile time latency is to continuously compile the code the user is editing, and keep the last one that compiled without errors, but that is still pointless if the compile takes several minutes.

Share this post


Link to post
Share on other sites

There is one way I've spotted that you could effectively split some of that shader into multiple pieces that you can 'link together - convert some of the functions to texture lookups. For example you could try creating a 2D texture which you sample to calculate the results of randint(x, y).


Yeah, there are specific code size optimizations I could do, but that doesn't help me in the general case (see previous post). Besides, replacing randint by the number 42 does not change compile time much.

Share this post


Link to post
Share on other sites
Sign in to follow this  

  • Advertisement
  • Advertisement
  • Popular Tags

  • Advertisement
  • Popular Now

  • Similar Content

    • By evelyn4you
      hi,
      after implementing skinning with compute shader i want to implement skinning with VertexShader Streamout method to compare performance.
      The following Thread is a discussion about it.
      Here's the recommended setup:
      Use a pass-through geometry shader (point->point), setup the streamout and set topology to point list. Draw the whole buffer with context->Draw(). This gives a 1:1 mapping of the vertices. Later bind the stream out buffer as vertex buffer. Bind the index buffer of the original mesh. draw with DrawIndexed like you would with the original mesh (or whatever draw call you had). I know the reason why a point list as input is used, because when using the normal vertex topology as input the output would be a stream of "each of his own" primitives that would blow up the vertexbuffer. I assume a indexbuffer then would be needless ?
      But how can you transform position and normal in one step when feeding the pseudo Vertex/Geometry Shader with a point list ?
      In my VertexShader i first calculate the resulting transform matrix from bone indexes(4) und weights (4) and transform position and normal with the same resulting transform Matrix.
      Do i have to run 2 passes ? One for transforming position and one for transforming normal ?
      I think it could be done better ?
      thanks for any help
       
    • By derui
      i am new to directx. i just followed some tutorials online and started to program. It had been well till i faced this problem of loading my own 3d models from 3ds max exported as .x which is supported by directx. I am using c++ on visual studio 2010 and directX9. i really tried to find help on the net but i couldn't find which can solve my problem. i don't know where exactly the problem is. i run most of samples and examples all worked well. can anyone give me the hint or solution for my problem ?
      thanks in advance!
    • By DiligentDev
      This article uses material originally posted on Diligent Graphics web site.
      Introduction
      Graphics APIs have come a long way from small set of basic commands allowing limited control of configurable stages of early 3D accelerators to very low-level programming interfaces exposing almost every aspect of the underlying graphics hardware. Next-generation APIs, Direct3D12 by Microsoft and Vulkan by Khronos are relatively new and have only started getting widespread adoption and support from hardware vendors, while Direct3D11 and OpenGL are still considered industry standard. New APIs can provide substantial performance and functional improvements, but may not be supported by older hardware. An application targeting wide range of platforms needs to support Direct3D11 and OpenGL. New APIs will not give any advantage when used with old paradigms. It is totally possible to add Direct3D12 support to an existing renderer by implementing Direct3D11 interface through Direct3D12, but this will give zero benefits. Instead, new approaches and rendering architectures that leverage flexibility provided by the next-generation APIs are expected to be developed.
      There are at least four APIs (Direct3D11, Direct3D12, OpenGL/GLES, Vulkan, plus Apple's Metal for iOS and osX platforms) that a cross-platform 3D application may need to support. Writing separate code paths for all APIs is clearly not an option for any real-world application and the need for a cross-platform graphics abstraction layer is evident. The following is the list of requirements that I believe such layer needs to satisfy:
      Lightweight abstractions: the API should be as close to the underlying native APIs as possible to allow an application leverage all available low-level functionality. In many cases this requirement is difficult to achieve because specific features exposed by different APIs may vary considerably. Low performance overhead: the abstraction layer needs to be efficient from performance point of view. If it introduces considerable amount of overhead, there is no point in using it. Convenience: the API needs to be convenient to use. It needs to assist developers in achieving their goals not limiting their control of the graphics hardware. Multithreading: ability to efficiently parallelize work is in the core of Direct3D12 and Vulkan and one of the main selling points of the new APIs. Support for multithreading in a cross-platform layer is a must. Extensibility: no matter how well the API is designed, it still introduces some level of abstraction. In some cases the most efficient way to implement certain functionality is to directly use native API. The abstraction layer needs to provide seamless interoperability with the underlying native APIs to provide a way for the app to add features that may be missing. Diligent Engine is designed to solve these problems. Its main goal is to take advantages of the next-generation APIs such as Direct3D12 and Vulkan, but at the same time provide support for older platforms via Direct3D11, OpenGL and OpenGLES. Diligent Engine exposes common C++ front-end for all supported platforms and provides interoperability with underlying native APIs. It also supports integration with Unity and is designed to be used as graphics subsystem in a standalone game engine, Unity native plugin or any other 3D application. Full source code is available for download at GitHub and is free to use.
      Overview
      Diligent Engine API takes some features from Direct3D11 and Direct3D12 as well as introduces new concepts to hide certain platform-specific details and make the system easy to use. It contains the following main components:
      Render device (IRenderDevice  interface) is responsible for creating all other objects (textures, buffers, shaders, pipeline states, etc.).
      Device context (IDeviceContext interface) is the main interface for recording rendering commands. Similar to Direct3D11, there are immediate context and deferred contexts (which in Direct3D11 implementation map directly to the corresponding context types). Immediate context combines command queue and command list recording functionality. It records commands and submits the command list for execution when it contains sufficient number of commands. Deferred contexts are designed to only record command lists that can be submitted for execution through the immediate context.
      An alternative way to design the API would be to expose command queue and command lists directly. This approach however does not map well to Direct3D11 and OpenGL. Besides, some functionality (such as dynamic descriptor allocation) can be much more efficiently implemented when it is known that a command list is recorded by a certain deferred context from some thread.
      The approach taken in the engine does not limit scalability as the application is expected to create one deferred context per thread, and internally every deferred context records a command list in lock-free fashion. At the same time this approach maps well to older APIs.
      In current implementation, only one immediate context that uses default graphics command queue is created. To support multiple GPUs or multiple command queue types (compute, copy, etc.), it is natural to have one immediate contexts per queue. Cross-context synchronization utilities will be necessary.
      Swap Chain (ISwapChain interface). Swap chain interface represents a chain of back buffers and is responsible for showing the final rendered image on the screen.
      Render device, device contexts and swap chain are created during the engine initialization.
      Resources (ITexture and IBuffer interfaces). There are two types of resources - textures and buffers. There are many different texture types (2D textures, 3D textures, texture array, cubmepas, etc.) that can all be represented by ITexture interface.
      Resources Views (ITextureView and IBufferView interfaces). While textures and buffers are mere data containers, texture views and buffer views describe how the data should be interpreted. For instance, a 2D texture can be used as a render target for rendering commands or as a shader resource.
      Pipeline State (IPipelineState interface). GPU pipeline contains many configurable stages (depth-stencil, rasterizer and blend states, different shader stage, etc.). Direct3D11 uses coarse-grain objects to set all stage parameters at once (for instance, a rasterizer object encompasses all rasterizer attributes), while OpenGL contains myriad functions to fine-grain control every individual attribute of every stage. Both methods do not map very well to modern graphics hardware that combines all states into one monolithic state under the hood. Direct3D12 directly exposes pipeline state object in the API, and Diligent Engine uses the same approach.
      Shader Resource Binding (IShaderResourceBinding interface). Shaders are programs that run on the GPU. Shaders may access various resources (textures and buffers), and setting correspondence between shader variables and actual resources is called resource binding. Resource binding implementation varies considerably between different API. Diligent Engine introduces a new object called shader resource binding that encompasses all resources needed by all shaders in a certain pipeline state.
      API Basics
      Creating Resources
      Device resources are created by the render device. The two main resource types are buffers, which represent linear memory, and textures, which use memory layouts optimized for fast filtering. Graphics APIs usually have a native object that represents linear buffer. Diligent Engine uses IBuffer interface as an abstraction for a native buffer. To create a buffer, one needs to populate BufferDesc structure and call IRenderDevice::CreateBuffer() method as in the following example:
      BufferDesc BuffDesc; BufferDesc.Name = "Uniform buffer"; BuffDesc.BindFlags = BIND_UNIFORM_BUFFER; BuffDesc.Usage = USAGE_DYNAMIC; BuffDesc.uiSizeInBytes = sizeof(ShaderConstants); BuffDesc.CPUAccessFlags = CPU_ACCESS_WRITE; m_pDevice->CreateBuffer( BuffDesc, BufferData(), &m_pConstantBuffer ); While there is usually just one buffer object, different APIs use very different approaches to represent textures. For instance, in Direct3D11, there are ID3D11Texture1D, ID3D11Texture2D, and ID3D11Texture3D objects. In OpenGL, there is individual object for every texture dimension (1D, 2D, 3D, Cube), which may be a texture array, which may also be multisampled (i.e. GL_TEXTURE_2D_MULTISAMPLE_ARRAY). As a result there are nine different GL texture types that Diligent Engine may create under the hood. In Direct3D12, there is only one resource interface. Diligent Engine hides all these details in ITexture interface. There is only one  IRenderDevice::CreateTexture() method that is capable of creating all texture types. Dimension, format, array size and all other parameters are specified by the members of the TextureDesc structure:
      TextureDesc TexDesc; TexDesc.Name = "My texture 2D"; TexDesc.Type = TEXTURE_TYPE_2D; TexDesc.Width = 1024; TexDesc.Height = 1024; TexDesc.Format = TEX_FORMAT_RGBA8_UNORM; TexDesc.Usage = USAGE_DEFAULT; TexDesc.BindFlags = BIND_SHADER_RESOURCE | BIND_RENDER_TARGET | BIND_UNORDERED_ACCESS; TexDesc.Name = "Sample 2D Texture"; m_pRenderDevice->CreateTexture( TexDesc, TextureData(), &m_pTestTex ); If native API supports multithreaded resource creation, textures and buffers can be created by multiple threads simultaneously.
      Interoperability with native API provides access to the native buffer/texture objects and also allows creating Diligent Engine objects from native handles. It allows applications seamlessly integrate native API-specific code with Diligent Engine.
      Next-generation APIs allow fine level-control over how resources are allocated. Diligent Engine does not currently expose this functionality, but it can be added by implementing IResourceAllocator interface that encapsulates specifics of resource allocation and providing this interface to CreateBuffer() or CreateTexture() methods. If null is provided, default allocator should be used.
      Initializing the Pipeline State
      As it was mentioned earlier, Diligent Engine follows next-gen APIs to configure the graphics/compute pipeline. One big Pipelines State Object (PSO) encompasses all required states (all shader stages, input layout description, depth stencil, rasterizer and blend state descriptions etc.). This approach maps directly to Direct3D12/Vulkan, but is also beneficial for older APIs as it eliminates pipeline misconfiguration errors. With many individual calls tweaking various GPU pipeline settings it is very easy to forget to set one of the states or assume the stage is already properly configured when in fact it is not. Using pipeline state object helps avoid these problems as all stages are configured at once.
      Creating Shaders
      While in earlier APIs shaders were bound separately, in the next-generation APIs as well as in Diligent Engine shaders are part of the pipeline state object. The biggest challenge when authoring shaders is that Direct3D and OpenGL/Vulkan use different shader languages (while Apple uses yet another language in their Metal API). Maintaining two versions of every shader is not an option for real applications and Diligent Engine implements shader source code converter that allows shaders authored in HLSL to be translated to GLSL. To create a shader, one needs to populate ShaderCreationAttribs structure. SourceLanguage member of this structure tells the system which language the shader is authored in:
      SHADER_SOURCE_LANGUAGE_DEFAULT - The shader source language matches the underlying graphics API: HLSL for Direct3D11/Direct3D12 mode, and GLSL for OpenGL and OpenGLES modes. SHADER_SOURCE_LANGUAGE_HLSL - The shader source is in HLSL. For OpenGL and OpenGLES modes, the source code will be converted to GLSL. SHADER_SOURCE_LANGUAGE_GLSL - The shader source is in GLSL. There is currently no GLSL to HLSL converter, so this value should only be used for OpenGL and OpenGLES modes. There are two ways to provide the shader source code. The first way is to use Source member. The second way is to provide a file path in FilePath member. Since the engine is entirely decoupled from the platform and the host file system is platform-dependent, the structure exposes pShaderSourceStreamFactory member that is intended to provide the engine access to the file system. If FilePath is provided, shader source factory must also be provided. If the shader source contains any #include directives, the source stream factory will also be used to load these files. The engine provides default implementation for every supported platform that should be sufficient in most cases. Custom implementation can be provided when needed.
      When sampling a texture in a shader, the texture sampler was traditionally specified as separate object that was bound to the pipeline at run time or set as part of the texture object itself. However, in most cases it is known beforehand what kind of sampler will be used in the shader. Next-generation APIs expose new type of sampler called static sampler that can be initialized directly in the pipeline state. Diligent Engine exposes this functionality: when creating a shader, textures can be assigned static samplers. If static sampler is assigned, it will always be used instead of the one initialized in the texture shader resource view. To initialize static samplers, prepare an array of StaticSamplerDesc structures and initialize StaticSamplers and NumStaticSamplers members. Static samplers are more efficient and it is highly recommended to use them whenever possible. On older APIs, static samplers are emulated via generic sampler objects.
      The following is an example of shader initialization:
      ShaderCreationAttribs Attrs; Attrs.Desc.Name = "MyPixelShader"; Attrs.FilePath = "MyShaderFile.fx"; Attrs.SearchDirectories = "shaders;shaders\\inc;"; Attrs.EntryPoint = "MyPixelShader"; Attrs.Desc.ShaderType = SHADER_TYPE_PIXEL; Attrs.SourceLanguage = SHADER_SOURCE_LANGUAGE_HLSL; BasicShaderSourceStreamFactory BasicSSSFactory(Attrs.SearchDirectories); Attrs.pShaderSourceStreamFactory = &BasicSSSFactory; ShaderVariableDesc ShaderVars[] = {     {"g_StaticTexture", SHADER_VARIABLE_TYPE_STATIC},     {"g_MutableTexture", SHADER_VARIABLE_TYPE_MUTABLE},     {"g_DynamicTexture", SHADER_VARIABLE_TYPE_DYNAMIC} }; Attrs.Desc.VariableDesc = ShaderVars; Attrs.Desc.NumVariables = _countof(ShaderVars); Attrs.Desc.DefaultVariableType = SHADER_VARIABLE_TYPE_STATIC; StaticSamplerDesc StaticSampler; StaticSampler.Desc.MinFilter = FILTER_TYPE_LINEAR; StaticSampler.Desc.MagFilter = FILTER_TYPE_LINEAR; StaticSampler.Desc.MipFilter = FILTER_TYPE_LINEAR; StaticSampler.TextureName = "g_MutableTexture"; Attrs.Desc.NumStaticSamplers = 1; Attrs.Desc.StaticSamplers = &StaticSampler; ShaderMacroHelper Macros; Macros.AddShaderMacro("USE_SHADOWS", 1); Macros.AddShaderMacro("NUM_SHADOW_SAMPLES", 4); Macros.Finalize(); Attrs.Macros = Macros; RefCntAutoPtr<IShader> pShader; m_pDevice->CreateShader( Attrs, &pShader );
      Creating the Pipeline State Object
      After all required shaders are created, the rest of the fields of the PipelineStateDesc structure provide depth-stencil, rasterizer, and blend state descriptions, the number and format of render targets, input layout format, etc. For instance, rasterizer state can be described as follows:
      PipelineStateDesc PSODesc; RasterizerStateDesc &RasterizerDesc = PSODesc.GraphicsPipeline.RasterizerDesc; RasterizerDesc.FillMode = FILL_MODE_SOLID; RasterizerDesc.CullMode = CULL_MODE_NONE; RasterizerDesc.FrontCounterClockwise = True; RasterizerDesc.ScissorEnable = True; RasterizerDesc.AntialiasedLineEnable = False; Depth-stencil and blend states are defined in a similar fashion.
      Another important thing that pipeline state object encompasses is the input layout description that defines how inputs to the vertex shader, which is the very first shader stage, should be read from the memory. Input layout may define several vertex streams that contain values of different formats and sizes:
      // Define input layout InputLayoutDesc &Layout = PSODesc.GraphicsPipeline.InputLayout; LayoutElement TextLayoutElems[] = {     LayoutElement( 0, 0, 3, VT_FLOAT32, False ),     LayoutElement( 1, 0, 4, VT_UINT8, True ),     LayoutElement( 2, 0, 2, VT_FLOAT32, False ), }; Layout.LayoutElements = TextLayoutElems; Layout.NumElements = _countof( TextLayoutElems ); Finally, pipeline state defines primitive topology type. When all required members are initialized, a pipeline state object can be created by IRenderDevice::CreatePipelineState() method:
      // Define shader and primitive topology PSODesc.GraphicsPipeline.PrimitiveTopologyType = PRIMITIVE_TOPOLOGY_TYPE_TRIANGLE; PSODesc.GraphicsPipeline.pVS = pVertexShader; PSODesc.GraphicsPipeline.pPS = pPixelShader; PSODesc.Name = "My pipeline state"; m_pDev->CreatePipelineState(PSODesc, &m_pPSO); When PSO object is bound to the pipeline, the engine invokes all API-specific commands to set all states specified by the object. In case of Direct3D12 this maps directly to setting the D3D12 PSO object. In case of Direct3D11, this involves setting individual state objects (such as rasterizer and blend states), shaders, input layout etc. In case of OpenGL, this requires a number of fine-grain state tweaking calls. Diligent Engine keeps track of currently bound states and only calls functions to update these states that have actually changed.
      Binding Shader Resources
      Direct3D11 and OpenGL utilize fine-grain resource binding models, where an application binds individual buffers and textures to certain shader or program resource binding slots. Direct3D12 uses a very different approach, where resource descriptors are grouped into tables, and an application can bind all resources in the table at once by setting the table in the command list. Resource binding model in Diligent Engine is designed to leverage this new method. It introduces a new object called shader resource binding that encapsulates all resource bindings required for all shaders in a certain pipeline state. It also introduces the classification of shader variables based on the frequency of expected change that helps the engine group them into tables under the hood:
      Static variables (SHADER_VARIABLE_TYPE_STATIC) are variables that are expected to be set only once. They may not be changed once a resource is bound to the variable. Such variables are intended to hold global constants such as camera attributes or global light attributes constant buffers. Mutable variables (SHADER_VARIABLE_TYPE_MUTABLE) define resources that are expected to change on a per-material frequency. Examples may include diffuse textures, normal maps etc. Dynamic variables (SHADER_VARIABLE_TYPE_DYNAMIC) are expected to change frequently and randomly. Shader variable type must be specified during shader creation by populating an array of ShaderVariableDesc structures and initializing ShaderCreationAttribs::Desc::VariableDesc and ShaderCreationAttribs::Desc::NumVariables members (see example of shader creation above).
      Static variables cannot be changed once a resource is bound to the variable. They are bound directly to the shader object. For instance, a shadow map texture is not expected to change after it is created, so it can be bound directly to the shader:
      PixelShader->GetShaderVariable( "g_tex2DShadowMap" )->Set( pShadowMapSRV ); Mutable and dynamic variables are bound via a new Shader Resource Binding object (SRB) that is created by the pipeline state (IPipelineState::CreateShaderResourceBinding()):
      m_pPSO->CreateShaderResourceBinding(&m_pSRB); Note that an SRB is only compatible with the pipeline state it was created from. SRB object inherits all static bindings from shaders in the pipeline, but is not allowed to change them.
      Mutable resources can only be set once for every instance of a shader resource binding. Such resources are intended to define specific material properties. For instance, a diffuse texture for a specific material is not expected to change once the material is defined and can be set right after the SRB object has been created:
      m_pSRB->GetVariable(SHADER_TYPE_PIXEL, "tex2DDiffuse")->Set(pDiffuseTexSRV); In some cases it is necessary to bind a new resource to a variable every time a draw command is invoked. Such variables should be labeled as dynamic, which will allow setting them multiple times through the same SRB object:
      m_pSRB->GetVariable(SHADER_TYPE_VERTEX, "cbRandomAttribs")->Set(pRandomAttrsCB); Under the hood, the engine pre-allocates descriptor tables for static and mutable resources when an SRB objcet is created. Space for dynamic resources is dynamically allocated at run time. Static and mutable resources are thus more efficient and should be used whenever possible.
      As you can see, Diligent Engine does not expose low-level details of how resources are bound to shader variables. One reason for this is that these details are very different for various APIs. The other reason is that using low-level binding methods is extremely error-prone: it is very easy to forget to bind some resource, or bind incorrect resource such as bind a buffer to the variable that is in fact a texture, especially during shader development when everything changes fast. Diligent Engine instead relies on shader reflection system to automatically query the list of all shader variables. Grouping variables based on three types mentioned above allows the engine to create optimized layout and take heavy lifting of matching resources to API-specific resource location, register or descriptor in the table.
      This post gives more details about the resource binding model in Diligent Engine.
      Setting the Pipeline State and Committing Shader Resources
      Before any draw or compute command can be invoked, the pipeline state needs to be bound to the context:
      m_pContext->SetPipelineState(m_pPSO); Under the hood, the engine sets the internal PSO object in the command list or calls all the required native API functions to properly configure all pipeline stages.
      The next step is to bind all required shader resources to the GPU pipeline, which is accomplished by IDeviceContext::CommitShaderResources() method:
      m_pContext->CommitShaderResources(m_pSRB, COMMIT_SHADER_RESOURCES_FLAG_TRANSITION_RESOURCES); The method takes a pointer to the shader resource binding object and makes all resources the object holds available for the shaders. In the case of D3D12, this only requires setting appropriate descriptor tables in the command list. For older APIs, this typically requires setting all resources individually.
      Next-generation APIs require the application to track the state of every resource and explicitly inform the system about all state transitions. For instance, if a texture was used as render target before, while the next draw command is going to use it as shader resource, a transition barrier needs to be executed. Diligent Engine does the heavy lifting of state tracking.  When CommitShaderResources() method is called with COMMIT_SHADER_RESOURCES_FLAG_TRANSITION_RESOURCES flag, the engine commits and transitions resources to correct states at the same time. Note that transitioning resources does introduce some overhead. The engine tracks state of every resource and it will not issue the barrier if the state is already correct. But checking resource state is an overhead that can sometimes be avoided. The engine provides IDeviceContext::TransitionShaderResources() method that only transitions resources:
      m_pContext->TransitionShaderResources(m_pPSO, m_pSRB); In some scenarios it is more efficient to transition resources once and then only commit them.
      Invoking Draw Command
      The final step is to set states that are not part of the PSO, such as render targets, vertex and index buffers. Diligent Engine uses Direct3D11-syle API that is translated to other native API calls under the hood:
      ITextureView *pRTVs[] = {m_pRTV}; m_pContext->SetRenderTargets(_countof( pRTVs ), pRTVs, m_pDSV); // Clear render target and depth buffer const float zero[4] = {0, 0, 0, 0}; m_pContext->ClearRenderTarget(nullptr, zero); m_pContext->ClearDepthStencil(nullptr, CLEAR_DEPTH_FLAG, 1.f); // Set vertex and index buffers IBuffer *buffer[] = {m_pVertexBuffer}; Uint32 offsets[] = {0}; Uint32 strides[] = {sizeof(MyVertex)}; m_pContext->SetVertexBuffers(0, 1, buffer, strides, offsets, SET_VERTEX_BUFFERS_FLAG_RESET); m_pContext->SetIndexBuffer(m_pIndexBuffer, 0); Different native APIs use various set of function to execute draw commands depending on command details (if the command is indexed, instanced or both, what offsets in the source buffers are used etc.). For instance, there are 5 draw commands in Direct3D11 and more than 9 commands in OpenGL with something like glDrawElementsInstancedBaseVertexBaseInstance not uncommon. Diligent Engine hides all details with single IDeviceContext::Draw() method that takes takes DrawAttribs structure as an argument. The structure members define all attributes required to perform the command (primitive topology, number of vertices or indices, if draw call is indexed or not, if draw call is instanced or not, if draw call is indirect or not, etc.). For example:
      DrawAttribs attrs; attrs.IsIndexed = true; attrs.IndexType = VT_UINT16; attrs.NumIndices = 36; attrs.Topology = PRIMITIVE_TOPOLOGY_TRIANGLE_LIST; pContext->Draw(attrs); For compute commands, there is IDeviceContext::DispatchCompute() method that takes DispatchComputeAttribs structure that defines compute grid dimension.
      Source Code
      Full engine source code is available on GitHub and is free to use. The repository contains two samples, asteroids performance benchmark and example Unity project that uses Diligent Engine in native plugin.
      AntTweakBar sample is Diligent Engine’s “Hello World” example.

       
      Atmospheric scattering sample is a more advanced example. It demonstrates how Diligent Engine can be used to implement various rendering tasks: loading textures from files, using complex shaders, rendering to multiple render targets, using compute shaders and unordered access views, etc.

      Asteroids performance benchmark is based on this demo developed by Intel. It renders 50,000 unique textured asteroids and allows comparing performance of Direct3D11 and Direct3D12 implementations. Every asteroid is a combination of one of 1000 unique meshes and one of 10 unique textures.

      Finally, there is an example project that shows how Diligent Engine can be integrated with Unity.

      Future Work
      The engine is under active development. It currently supports Windows desktop, Universal Windows and Android platforms. Direct3D11, Direct3D12, OpenGL/GLES backends are now feature complete. Vulkan backend is coming next, and support for more platforms is planned.
    • By kan123
      Hello,
      DX9Ex. I have the problem with driver stability in time of serial renderings, which i try to use for image processing in memory with fragment shaders. For big bitmaps the video driver sometimes becomes unstable ("Display driver stopped responding and has recovered") and, for instance, if the media player runs video in background, it sometimes freezes and distorts. I tried to use next methods of IDirect3DDevice9Ex:
      SetGPUThreadPriority(-7);
      WaitForVBlank(0);
      EvictManagedResources();
      with purpose to give some time for GPU between scenes, but it seems to be has not notable effect in this case. I don't want to reinitilialize subsystem for every step to avoid performance loss.
      So, my question is next: does some common practice exists to avoid overloading of GPU by running tasks? Many thanks in advance.
       
    • By AxeGuywithanAxe
      I wanted to see how others are currently handling descriptor heap updates and management.
      I've read a few articles and there tends to be three major strategies :
      1 ) You split up descriptor heaps per shader stage ( i.e one for vertex shader , pixel , hull, etc)
      2) You have one descriptor heap for an entire pipeline
      3) You split up descriptor heaps for update each update frequency (i.e EResourceSet_PerInstance , EResourceSet_PerPass , EResourceSet_PerMaterial, etc)
      The benefits of the first two approaches is that it makes it easier to port current code, and descriptor / resource descriptor management and updating tends to be easier to manage, but it seems to be not as efficient.
      The benefits of the third approach seems to be that it's the most efficient because you only manage and update objects when they change.
  • Advertisement