Multiview on DX12 (#8495)

Co-authored-by: Andreas Reich <r_andreas2@web.de>

Author: inner-daemons

CHANGELOG.md

@@ -82,7 +82,7 @@ can use this feature.
 
 Multiview is also called view instancing in DX12 land or vertex amplification in Metal land.
 
-Multiview has been reworked, adding support for Metal, and adding testing and validation to wgpu itself.
+Multiview has been reworked, adding support for Metal and DX12, and adding testing and validation to wgpu itself.
 This change also introduces a view bitmask, a new field in `RenderPassDescriptor` that allows a render pass to render to multiple non-adjacent layers
 when using the `SELECTIVE_MULTIVIEW` feature. Note that this also influences apps that don't use multiview, as they have to set this mask to `None`.
 ```diff

wgpu-hal/src/dx12/adapter.rs

@@ -601,13 +601,20 @@ impl super::Adapter {
             shader_barycentrics_supported,
         );
 
-        // Re-enable this when multiview is supported on DX12
-        // features.set(wgt::Features::MULTIVIEW, view_instancing);
-        // features.set(wgt::Features::SELECTIVE_MULTIVIEW, view_instancing);
+        features.set(
+            wgt::Features::MULTIVIEW,
+            view_instancing && shader_model >= naga::back::hlsl::ShaderModel::V6_1,
+        );
+        features.set(
+            wgt::Features::SELECTIVE_MULTIVIEW,
+            view_instancing && shader_model >= naga::back::hlsl::ShaderModel::V6_1,
+        );
 
         features.set(
             wgt::Features::EXPERIMENTAL_MESH_SHADER_MULTIVIEW,
-            mesh_shader_supported && view_instancing,
+            mesh_shader_supported
+                && view_instancing
+                && shader_model >= naga::back::hlsl::ShaderModel::V6_1,
         );
 
         // TODO: Determine if IPresentationManager is supported

wgpu-hal/src/dx12/device.rs

@@ -5,6 +5,7 @@ use alloc::{
     sync::Arc,
     vec::Vec,
 };
+use arrayvec::ArrayVec;
 use core::{ffi, num::NonZeroU32, ptr, time::Duration};
 use std::time::Instant;
 
@@ -983,7 +984,7 @@ impl crate::Device for super::Device {
         let mut ranges = Vec::with_capacity(total_non_dynamic_entries);
 
         let mut bind_group_infos =
-            arrayvec::ArrayVec::<super::BindGroupInfo, { crate::MAX_BIND_GROUPS }>::default();
+            ArrayVec::<super::BindGroupInfo, { crate::MAX_BIND_GROUPS }>::default();
         for (index, bgl) in desc.bind_group_layouts.iter().enumerate() {
             let mut info = super::BindGroupInfo {
                 tables: super::TableTypes::empty(),
@@ -1952,6 +1953,22 @@ impl crate::Device for super::Device {
         };
         let flags = Direct3D12::D3D12_PIPELINE_STATE_FLAG_NONE;
 
+        let mut view_instancing =
+            core::pin::pin!(ArrayVec::<Direct3D12::D3D12_VIEW_INSTANCE_LOCATION, 32>::new());
+        if let Some(mask) = desc.multiview_mask {
+            let mask = mask.get();
+            // This array is just what _could_ be rendered to. We actually apply the mask at
+            // renderpass creation time. The `view_index` passed to the shader depends on the
+            // view's index in this array, so if we include every view in this array, `view_index`
+            // actually the texture array layer, like in vulkan.
+            for i in 0..32 - mask.leading_zeros() {
+                view_instancing.push(Direct3D12::D3D12_VIEW_INSTANCE_LOCATION {
+                    ViewportArrayIndex: 0,
+                    RenderTargetArrayIndex: i,
+                });
+            }
+        }
+
         let mut stream_desc = RenderPipelineStateStreamDesc {
             // Shared by vertex and mesh pipelines
             root_signature: desc.layout.shared.signature.as_ref(),
@@ -1970,6 +1987,16 @@ impl crate::Device for super::Device {
             node_mask: 0,
             cached_pso,
             flags,
+            view_instancing: if !view_instancing.is_empty() {
+                Some(Direct3D12::D3D12_VIEW_INSTANCING_DESC {
+                    ViewInstanceCount: view_instancing.len() as u32,
+                    pViewInstanceLocations: view_instancing.as_ptr(),
+                    // This lets us hide/mask certain values later, at renderpass creation time.
+                    Flags: Direct3D12::D3D12_VIEW_INSTANCING_FLAG_ENABLE_VIEW_INSTANCE_MASKING,
+                })
+            } else {
+                None
+            },
 
             // Optional data that depends on the pipeline type (vertex vs mesh).
             vertex_shader: Default::default(),

wgpu-hal/src/dx12/pipeline_desc.rs

@@ -91,6 +91,7 @@ implement_stream_object! { unsafe D3D12_PIPELINE_STATE_FLAGS => D3D12_PIPELINE_S
 implement_stream_object! { unsafe D3D12_INPUT_LAYOUT_DESC => D3D12_PIPELINE_STATE_SUBOBJECT_TYPE_INPUT_LAYOUT }
 implement_stream_object! { unsafe D3D12_INDEX_BUFFER_STRIP_CUT_VALUE => D3D12_PIPELINE_STATE_SUBOBJECT_TYPE_IB_STRIP_CUT_VALUE }
 implement_stream_object! { unsafe D3D12_STREAM_OUTPUT_DESC => D3D12_PIPELINE_STATE_SUBOBJECT_TYPE_STREAM_OUTPUT }
+implement_stream_object! { unsafe D3D12_VIEW_INSTANCING_DESC => D3D12_PIPELINE_STATE_SUBOBJECT_TYPE_VIEW_INSTANCING }
 
 /// Implementaation of a pipeline state stream, which is a sequence of subobjects put into
 /// a byte array according to some basic alignment rules.
@@ -183,6 +184,7 @@ pub struct RenderPipelineStateStreamDesc<'a> {
     pub node_mask: u32,
     pub cached_pso: D3D12_CACHED_PIPELINE_STATE,
     pub flags: D3D12_PIPELINE_STATE_FLAGS,
+    pub view_instancing: Option<D3D12_VIEW_INSTANCING_DESC>,
 
     // Vertex pipeline specific
     pub vertex_shader: D3D12_SHADER_BYTECODE,
@@ -230,6 +232,9 @@ impl RenderPipelineStateStreamDesc<'_> {
             stream.add_object(self.cached_pso);
         }
         stream.add_object(self.flags);
+        if let Some(view_instancing) = self.view_instancing {
+            stream.add_object(view_instancing);
+        }
         if !self.pixel_shader.pShaderBytecode.is_null() {
             stream.add_object(PixelShader(self.pixel_shader));
         }

wgpu-types/src/features.rs

@@ -920,10 +920,9 @@ bitflags_array! {
         /// Supported platforms:
         /// - Vulkan
         /// - Metal
+        /// - DX12
         /// - OpenGL (web only)
         ///
-        /// DX12 support is a WIP.
-        ///
         /// This is a native only feature.
         const MULTIVIEW = 1 << 26;
         /// Enables using 64-bit types for vertex attributes.
@@ -1249,8 +1248,8 @@ bitflags_array! {
         ///
         /// Supported platforms
         /// - Vulkan
+        /// - DX12
         ///
-        /// DX12 will support this when it supports multiview in general.
         ///
         /// While metal supports this in theory, the behavior of `view_index` differs from vulkan and dx12 so the feature isn't exposed.
         const SELECTIVE_MULTIVIEW = 1 << 54;

wgpu/src/api/render_pass.rs

@@ -685,6 +685,9 @@ pub struct RenderPassDescriptor<'a> {
     /// 2nd layer, you would use 2=0b10. If you aren't using multiview this should be `None`.
     ///
     /// Note that setting bits higher than the number of texture layers is a validation error.
+    ///
+    /// This doesn't influence load/store/clear/etc operations, as those are defined for attachments,
+    /// therefore affecting all attachments. Meaning, this affects only any shaders executed on the `RenderPass`.
     pub multiview_mask: Option<NonZeroU32>,
 }
 #[cfg(send_sync)]