meta: Add tiler optimization extension documentation.

Signed-off-by: Hans-Kristian Arntzen <post@arntzen-software.no>

Author: HansKristian-Work

docs/tiler_optimization_extensions.md

@@ -0,0 +1,257 @@
+# Tiler Optimization Extensions
+
+To allow for optimal rendering performance on GPUs which support tiling, vkd3d-proton
+provides some special interfaces to expose this style of rendering by keeping render targets on-chip,
+allowing framebuffer fetch mechanisms which default D3D12 APIs do not expose by default.
+
+Applications which intend to run D3D12 over Proton on mobile hardware such as Adreno
+are able to leverage this new interface with minimal changes to the renderer.
+The expectation is that this interface will mostly be relevant for VR games
+which are more likely to cater to mobile concerns.
+
+## Alternative to D3D12 RenderPass API
+
+The default [D3D12 RenderPass API](https://microsoft.github.io/DirectX-Specs/d3d/RenderPasses.html)
+supports extensions for tiler optimizations with APIs like
+`D3D12_RENDER_PASS_ENDING_ACCESS_PRESERVE_LOCAL_SRV`.
+However, this API is fundamentally incompatible with Vulkan, and it is also unimplementable by
+virtually all mobile hardware, including mobile hardware that vkd3d-proton cares about.
+This vkd3d-proton interface should be supported on a wide range of hardware.
+
+It relies on [VK_KHR_dynamic_rendering_local_read](https://docs.vulkan.org/refpages/latest/refpages/source/VK_KHR_dynamic_rendering_local_read.html)
+as well as [VK_KHR_unified_image_layouts](https://docs.vulkan.org/refpages/latest/refpages/source/VK_KHR_unified_image_layouts.html).
+
+### Additional programmable blending support
+
+Unlike D3D12's RenderPass API, this extended API can express programmable blending
+where an attachment can be sampled from even when in `D3D12_RESOURCE_STATE_RENDER_TARGET`
+or `D3D12_RESOURCE_STATE_DEPTH_WRITE` resource states.
+
+### Shader reuse
+
+The intent is that existing shaders can be reused.
+For example, this shader could be used for programmable blending:
+
+```
+Texture2D<float4> RenderTarget : register(t5, space6);
+
+float4 blend(float4 dst, float4 src)
+{
+    // Or whatever you want.
+    return lerp(dst, src, src.a);
+}
+
+float4 main(float4 pos : SV_Position, float4 color : COLOR) : SV_Target
+{
+    float4 RT = RenderTarget.Load(int3(pos.xy, 0));
+    return blend(RT, color);
+}
+```
+
+Where we have new APIs for remapping `t5, space6` to e.g. RTV #0 in the root signature.
+Applications can also redirect normal Texture2D SRVs to the bound depth or stencil attachments.
+This allows for typical deferred rendering scenarios where the G-buffer is read from on-chip memory instead
+of textures.
+Multi-sampled attachments are also supported.
+This can be used for e.g. custom HDR resolves on-chip.
+
+Layered rendering and view instancing is also supported.
+However, in this case, a non-arrayed `Texture2D` or `Texture2DMS` is still used in the shader.
+The implementation samples from the corresponding layer implicitly.
+
+### `OMSetRenderTargets` support
+
+This new interface is compatible with both "immediate" `OMSetRenderTargets` style rendering
+as well as the more dedicated RenderPass APIs. However, to be as tiler friendly as possible,
+it is recommended to use the RenderPass API to get the most out of this interface.
+
+## New Device APIs
+
+```
+typedef struct D3D12_VK_INPUT_ATTACHMENT_MAPPING
+{
+    UINT RegisterSpace;
+    UINT ShaderRegister;
+} D3D12_VK_INPUT_ATTACHMENT_MAPPING;
+
+typedef struct D3D12_VK_INPUT_ATTACHMENT_MAPPINGS
+{
+    UINT NumRenderTargets;
+    D3D12_VK_INPUT_ATTACHMENT_MAPPING RenderTargets[D3D12_SIMULTANEOUS_RENDER_TARGET_COUNT];
+    BOOL EnableDepth;
+    BOOL EnableStencil;
+    D3D12_VK_INPUT_ATTACHMENT_MAPPING Depth;
+    D3D12_VK_INPUT_ATTACHMENT_MAPPING Stencil;
+} D3D12_VK_INPUT_ATTACHMENT_MAPPINGS;
+
+typedef enum D3D12_VK_TILER_OPTIMIZATION_TIER
+{
+    D3D12_VK_TILER_OPTIMIZATION_NOT_SUPPORTED = 0,
+    D3D12_VK_TILER_OPTIMIZATION_TIER_1 = 1,
+} D3D12_VK_TILER_OPTIMIZATION_TIER;
+
+[
+    uuid(b7798d22-9fce-434d-8eeb-c3cef1056125),
+    object,
+    local,
+    pointer_default(unique)
+]
+interface ID3D12DeviceExt2 : ID3D12DeviceExt1
+{
+    D3D12_VK_TILER_OPTIMIZATION_TIER GetTilerOptimizationTier();
+    HRESULT OptInToTilerOptimizations();
+    UINT GetInputAttachmentDescriptorsCount();
+    HRESULT CreateRootSignatureWithInputAttachments(
+            UINT node_mask,
+            const void *bytecode, SIZE_T bytecode_length,
+            const D3D12_VK_INPUT_ATTACHMENT_MAPPINGS *mappings,
+            REFIID riid, void **root_signature);
+    void CreateInputAttachmentDescriptors(D3D12_CPU_DESCRIPTOR_HANDLE base_descriptor,
+            UINT render_target_descriptor_count,
+            const D3D12_CPU_DESCRIPTOR_HANDLE *render_target_descriptors,
+            BOOL single_descriptor_handle,
+            const D3D12_CPU_DESCRIPTOR_HANDLE *depth_descriptor,
+            const D3D12_CPU_DESCRIPTOR_HANDLE *stencil_descriptor);
+}
+```
+
+### `GetTilerOptimizationTier()`
+
+This is a simple query to check if these APIs are supported by the device.
+There is currently only one feature tier.
+
+### `HRESULT OptInToTilerOptimizations()`
+
+This is intended to be called right after the ID3D12Device is created.
+Setting this modifes the implementation in certain ways to make it compatible with
+tiler optimizations without adding a lot of extra API churn.
+This call is not thread-safe and should not be called concurrently with any other API command.
+
+The differences are:
+
+- If a resource is created with `ALLOW_RENDER_TARGET` or `ALLOW_DEPTH_STENCIL`, `VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT` is
+  added automatically.
+- When creating RTV views, `VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT` is added automatically.
+- When creating DSV views, `VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT` is added automatically in some cases:
+  - For DSV views with a single plane, the usage is added automatically.
+  - For DSV views with both planes, the DSV is only input attachment enabled if
+    there is exactly one plane which is marked read-only with
+    `D3D12_DSV_FLAG_READ_ONLY_DEPTH` or `D3D12_DSV_FLAG_READ_ONLY_STENCIL`.
+    The read-only aspect is compatible with input attachments.
+    (This is somewhat awkward, but it removes a lot of extra API churn, and is very unlikely to come up in practice).
+
+### `UINT GetInputAttachmentDescriptorsCount()`
+
+To be able to read from on-chip memory, the application allocates special SRVs in the descriptor heap.
+Rather than normal texture SRVs, Vulkan requires the use of `VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT`.
+When writing descriptors to the heap, multiple descriptors are written together as a group in
+a layout which is opaque to the application. The expectation is that 10 CBV_SRV_UAV descriptors
+are consumed (8 RT + Depth + Stencil), but it may be different due to descriptor packing concerns.
+
+For simplicity and practicality of the implementation, the number of descriptors is fixed at the upper bound.
+New input attachment descriptors need only be allocated once per render pass.
+
+### `HRESULT CreateRootSignatureWithInputAttachments()`
+
+This is equivalent to CreateRootSignature, except that extra information can be added
+for input attachments. `mappings` can be `NULL` in which case the call is equivalent
+to `CreateRootSignature()`. (It is not reasonable to modify the encoded RootSignature payload to
+hack in support for this, so this was determined to be the most practical solution.)
+
+Input attachment mappings only work for non-arrayed descriptors. I.e., shaders which access
+the bound attachments through bindless means will not work with this interface
+since the compiler needs to statically map resource variables to a render target index to
+be able to take advantage of on-chip data.
+
+Input attachment mappings can conflict with normal descriptor table bindings,
+i.e. override existing descriptor table bindings in the root signature.
+In this case, the input attachment mapping takes precedence.
+This allows applications to keep using the normal SRV path on most implementations,
+but selectively "opt-in" to the fast path when supported without having to modify the shader code.
+
+When a `Texture2D` or `Texture2DMS` is mapped to an input attachment, that texture must only be used
+with simple `::Load()` functions or equivalent. It cannot be used with a sampler object.
+Misuse will lead to PSO creation failure.
+
+The coordinate except for sampler index is ignored, and replaced with the current pixel coordinate.
+To make this transformation transparent, the pixel shader can sample from `int2(SV_Position.xy)`.
+
+When mappings are used, the root signature must have at least one 1 DWORD available in the root signature
+for the implementation to pass down additional data.
+
+### `void CreateInputAttachmentDescriptors()`
+
+Takes an equivalent of `OMSetRenderTargets` and writes input attachment descriptors to them.
+`GetInputAttachmentDescriptorsCount()` number of consecutive CBV_SRV_UAV descriptors are consumed.
+
+The main difference is that depth and stencil descriptors are separate in this interface.
+
+The RTV or DSV descriptors need not be the exact same ones passed to `OMSetRenderTargets()`,
+but they must be equivalent except for any read-only DSV state.
+
+TODO: Add an interface for RenderPass API desc as well.
+
+NULL RTVs or DSVs are ignored, and the matching descriptor in the heap is not modified.
+Using input attachments to sample from a NULL RTV or DSV is undefined behavior.
+Just use normal SRVs instead.
+
+### PSO considerations
+
+An input attachment which intends to read from a render target must define that render target
+in the PSO by using a sufficiently large `NumRenderTargets`.
+If an SRV is mapped to render target `N`, and `N` is greater-or-equal to `NumRenderTargets`,
+the input attachment must not be read from.
+
+The RTV format can be `DXGI_FORMAT_UNKNOWN` if the render target is only used as an input attachment
+in the PSO.
+
+Depth-stencil input attachments can sample from input attachments even with `DSVFormat` equal to `DXGI_FORMAT_UNKNOWN`.
+
+## New CommandList APIs
+
+```
+[
+    uuid(9c228166-bf9e-464c-9078-ecf20a13271a),
+    object,
+    local,
+    pointer_default(unique)
+]
+interface ID3D12GraphicsCommandListExt2 : ID3D12GraphicsCommandListExt1
+{
+    void InputAttachmentPixelBarrier();
+    void SetRootSignatureInputAttachments(D3D12_GPU_DESCRIPTOR_HANDLE handle);
+    void SetInputAttachmentFeedback(UINT render_target_concurrent_mask, BOOL depth_concurrent, BOOL stencil_concurrent);
+}
+```
+
+### `void InputAttachmentPixelBarrier()`
+
+While an image as in `RENDER_TARGET` or `DEPTH_WRITE` resource states (or equivalent in enhanced barriers),
+it cannot be sampled from as an input attachment without performing a per-pixel barrier.
+This can be called at any time, even inside a render pass.
+Only render target writes before the pixel barrier are visible to input attachment reads after the barrier.
+
+NOTE: Unlike D3D12, Vulkan supports this use case in the `VK_IMAGE_LAYOUT_GENERAL` image layout,
+which is why this feature requires `VK_KHR_unified_image_layouts`.
+
+### `void SetRootSignatureInputAttachments()`
+
+Binds the descriptors for input attachments.
+Unlike normal root parameters, this argument is never invalidated by binding new root signatures.
+It can safely be called once per OMSetRenderTargets and forgotten about.
+The descriptor handle must point to the currently bound descriptor heap.
+
+### `void SetInputAttachmentFeedback()`
+
+Programmable blending use cases and G-buffer deferred rendering are similar, but have different data access patterns.
+
+In programmable blending, there is concurrent access of the render target while sampling from it.
+In typical G-buffer deferred there is no such issue since the data flow is clearly separated by a writing phase then a read-only phase.
+
+Even with appropriate barriers in place, there may be hazards when render targets are compressed leading to garbage pixels
+being read in the input attachment unless some care is taken.
+
+IMPORTANT: Calling this will end the render pass internally, so this should not be called last-minute while inside a render pass.
+For performance, set this state up front, alongside `OMSetRenderTargets()` or right before `BeginRenderPass()`.
+If you don't know up front, just enable full feedback for the render pass.
+It should be fine on most implementations anyway.