libvkd3d
- Try to use VK_EXT_buffer_device_address for GPU addresses (partially done - jkucia).
- Add tests for out-of-bounds access with vertex buffers.
- Improve multi-queue support.
- Queue family ownership transfer.
- Fix all Vulkan validation errors.
- Fix clears for A8_UNORM.
- Implement UAVs in vertex processing shader stages.
- Implement support for multiple viewports.
- Test and fix copying between BC and other compatible formats.
- State decay for resources and queue ownership transfer for resources.
- Reduce the number of queue submits. Queue submits are generally expensive, especially on recent kernels with amdgpu (mostly done - jkucia).
- Merge vkCmdPipelineBarrier() calls.
- Try to avoid fence worker thread when possible.
- Implement stencil aspect resource state transitions.
- Implement cache for SPIR-V shaders?
- Implement disk pipeline/shader cache?
- Implement sparse resources?
- Optimize barriers (render passes, ClearUnorderedAccessView, transform feedback).
- Memory residency.
- Raytracing?
- DirectML?
- Add support for sRGB swapchain views (VK_KHR_image_format_list).
- Add tests for validation of view format compatibility (RTV, SRV, UAV, DSV).
- D3D12 resource state tracking doesn't map very well to Vulkan
- D3D12 creates its resource directly in its initial state, while Vulkan requires a GPU operation for the first transition, which therefore has to be postponed to when a GPU queue is available (i.e., the first time the resource is used); see e.g. !838
- In D3D12 depth and stencil are independent subresources, so their state can be managed independently; in Vulkan there is just one layout that specifies what happens on both aspects; therefore when the state of only one subresource is changed in D3D12 we don't know what to do in Vulkan, unless we track the state of the other aspect (and we currently don't do that)
- D3D12 has implicit state promotion and decay so when a command is recorded we don't know for sure which (explicit) state that resource is supposed to be in; for example, the destination of a copy operation might be in the COPY_DEST state or in the GENERAL state implicitly promoted to COPY_DEST. But since promotion is implicit we don't know which actual layout the Vulkan image is in, and Vulkan copy commands want to know it.
- Correctly tracking D3D12 resource state cannot be done at D3D12 command list recording time, because it depends on the initial states of the resources, which depends on which order the command lists are submitted to a queue (and potentially even to many queues); in other words, we have to serialize command list submission and only at that point we have all the information to write Vulkan command buffers.
libvkd3d-shader
- Add support for capturing components with stream output.
- Finish the DXIL work (Conor).
- Add tests for OpenGL atomic counters.
- Import SM4 to GLSL shader translation from wined3d (Henri).
- Allow ignoring checking the checksum when parsing DXBC, both from the API and from the command line.
- Always check that memory allocation didn't fail; for example, calls to vkd3d_string_buffer_printf() usually are not checked.
- Finish API for scanning/gathering info from shaders (registers used, bindings used, parameters used, required limits, extensions).
- Implement ID3D11ShaderReflection et al.
- Should we put everything into the same (chained) vkd3d_shader_scan_dxbc_info struct, or split them up? (Split up the DXBC-specific ones from the rest?)
- I guess HLSL compile flags should be preserved verbatim...? I don't like the disparity we currently have.
- How to do things like constant table types?
- Add support for SSBO.
- Support HLSL -> sm6 compilation.
- Support direct HLSL -> spirv compilation. This will be easier as things are refactored to account for vsir.
- Consider adding a CUDA frontend.
TPF validation
Vkd3d-shader generally assumes that Direct3D shader bytecode is valid and doesn't do much validation. We should introduce some kind of validation pass. The validation potentially could be optional, e.g. used only in debug builds.
- Interpolation mode is the same for all variables in the input register.
- Number of indices for resources (SM5 vs SM5.1).
Tests
-
Make it possible to run all reasonable test combinations mixing these factors:
- the library used to compile the shaders: vkd3d-shader, d3dcompiler_xx.dll or dxcompiler.dll;
- the Shader Model to target;
- the API to use to run the shader (d3d8, d3d9, d3d10, d3d11, d3d12, Vulkan, GL);
- when running on Windows, the actual library to use to run the shader: vkd3d or a native implementation;
- possibly, one day, the library to use to recompile the shaders (it currently only makes sense on macOS once we support the Apple shader compiler for Metal; in all the other cases the choice is forced).
-
Run a decent subsets of those combinations in the CI.
-
Need more shader runners: OpenGL + GLSL (Henri has WIP), d3d10core, d3d8
-
Some HLSL functionality was accepted despite a lack of tests. We need tests for:
- all tex*() variants
- D3DCOLORtoUBYTE4() different results on sm1
- min()
- texture arrays
- centroid interpolation
- "linear centroid" keywords
- noperspective interpolation
- precise?
- texture types other than 2D
- nontrivial tests for 2D multisample textures
- GetDimensions()
- SampleCmp(), SampleCmpLevelZero()
- depth output
- GS SV_PrimitiveID
- For the sake of sm1 and its weird disjoint instruction sets, a lot of tests probably deserve VS variants as well.
- Uniform reflection, which depends on uniform reflection API design
- SM1 semantic matching (Zeb has this)
- Test that all of the SM1 input semantics work
- tbuffer
- We should properly gate features based on profile version, and test that we are correctly doing this.
None of this necessarily needs to be extensions to the shader runner; we can find a way to write C tests for these.
- We would also like a way to invoke shader runner logic from C. It's a bit unfortunate that the tests in hlsl_d3d12.c are d3d12-specific.
- Untested sm4 -> spirv paths:
- "lod" shader instruction
- bufinfo and resinfo with NULL SRV and UAV descriptors
- D3D discard semantics
DXGI
- Fix adapter LUIDs in Wine.
- Reimplement IDXGIAdapter3::QueryVideoMemoryInfo() on top of VK_EXT_memory_budget. Ideally, Vulkan implementations should return the memory budget based on total system load (memory used by all graphics APIs).
Vulkan
- Try to fix detached WSI surfaces after D3D9/D3D11 swapchain (OpenGL) is created for a window with D3D12 swapchain (Vulkan).
- Fix child window rendering for D3D12 RenderDoc.