Add extra acceleration structure vertex formats.

Vecvec · Vecvec · commit 91dd80140e59 · 2025-04-20T07:37:24.000+12:00
diff --git a/tests/tests/wgpu-gpu/ray_tracing/as_build.rs b/tests/tests/wgpu-gpu/ray_tracing/as_build.rs
@@ -1,4 +1,4 @@
-use std::iter;
+use std::{iter, mem};
 
 use crate::ray_tracing::AsBuildContext;
 use wgpu::util::{BufferInitDescriptor, DeviceExt};
@@ -604,3 +604,204 @@ fn only_tlas_vertex_return(ctx: TestingContext) {
         None,
     );
 }
+
+#[gpu_test]
+static EXTRA_FORMAT_BUILD: GpuTestConfiguration = GpuTestConfiguration::new()
+    .parameters(
+        TestParameters::default()
+            .test_features_limits()
+            .features(
+                wgpu::Features::EXPERIMENTAL_RAY_TRACING_ACCELERATION_STRUCTURE
+                    | wgpu::Features::EXTENDED_ACCELERATION_STRUCTURE_VERTEX_FORMATS,
+            )
+            // https://github.com/gfx-rs/wgpu/issues/6727
+            .skip(FailureCase::backend_adapter(wgpu::Backends::VULKAN, "AMD")),
+    )
+    .run_sync(extra_format_build);
+
+fn extra_format_build(ctx: TestingContext) {
+    let vertices = ctx.device.create_buffer_init(&BufferInitDescriptor {
+        label: None,
+        contents: &[0; mem::size_of::<[[i16; 3]; 3]>()],
+        usage: BufferUsages::BLAS_INPUT,
+    });
+
+    let blas_size = BlasTriangleGeometrySizeDescriptor {
+        // The fourth component is ignored, and it allows us to have a smaller stride.
+        vertex_format: VertexFormat::Snorm16x4,
+        vertex_count: 3,
+        index_format: None,
+        index_count: None,
+        flags: wgpu::AccelerationStructureGeometryFlags::empty(),
+    };
+
+    let blas = ctx.device.create_blas(
+        &CreateBlasDescriptor {
+            label: Some("BLAS"),
+            flags: wgpu::AccelerationStructureFlags::PREFER_FAST_TRACE,
+            update_mode: AccelerationStructureUpdateMode::Build,
+        },
+        BlasGeometrySizeDescriptors::Triangles {
+            descriptors: vec![blas_size.clone()],
+        },
+    );
+
+    let mut command_encoder = ctx
+        .device
+        .create_command_encoder(&CommandEncoderDescriptor {
+            label: Some("BLAS_1"),
+        });
+    command_encoder.build_acceleration_structures(
+        &[BlasBuildEntry {
+            blas: &blas,
+            geometry: BlasGeometries::TriangleGeometries(vec![BlasTriangleGeometry {
+                size: &blas_size,
+                vertex_buffer: &vertices,
+                first_vertex: 0,
+                vertex_stride: mem::size_of::<[i16; 3]>() as BufferAddress,
+                index_buffer: None,
+                first_index: None,
+                transform_buffer: None,
+                transform_buffer_offset: None,
+            }]),
+        }],
+        &[],
+    );
+    ctx.queue.submit([command_encoder.finish()]);
+}
+
+#[gpu_test]
+static MISALIGNED_BUILD: GpuTestConfiguration = GpuTestConfiguration::new()
+    .parameters(
+        TestParameters::default()
+            .test_features_limits()
+            .features(wgpu::Features::EXPERIMENTAL_RAY_TRACING_ACCELERATION_STRUCTURE)
+            // https://github.com/gfx-rs/wgpu/issues/6727
+            .skip(FailureCase::backend_adapter(wgpu::Backends::VULKAN, "AMD")),
+    )
+    .run_sync(misaligned_build);
+
+fn misaligned_build(ctx: TestingContext) {
+    let vertices = ctx.device.create_buffer_init(&BufferInitDescriptor {
+        label: None,
+        contents: &[0; mem::size_of::<[[i16; 3]; 3]>()],
+        usage: BufferUsages::BLAS_INPUT,
+    });
+
+    let blas_size = BlasTriangleGeometrySizeDescriptor {
+        // The fourth component is ignored, and it allows us to have a smaller stride.
+        vertex_format: VertexFormat::Float32x3,
+        vertex_count: 3,
+        index_format: None,
+        index_count: None,
+        flags: wgpu::AccelerationStructureGeometryFlags::empty(),
+    };
+
+    let blas = ctx.device.create_blas(
+        &CreateBlasDescriptor {
+            label: Some("BLAS"),
+            flags: wgpu::AccelerationStructureFlags::PREFER_FAST_TRACE,
+            update_mode: AccelerationStructureUpdateMode::Build,
+        },
+        BlasGeometrySizeDescriptors::Triangles {
+            descriptors: vec![blas_size.clone()],
+        },
+    );
+
+    let mut command_encoder = ctx
+        .device
+        .create_command_encoder(&CommandEncoderDescriptor {
+            label: Some("BLAS_1"),
+        });
+    fail(
+        &ctx.device,
+        || {
+            command_encoder.build_acceleration_structures(
+                &[BlasBuildEntry {
+                    blas: &blas,
+                    geometry: BlasGeometries::TriangleGeometries(vec![BlasTriangleGeometry {
+                        size: &blas_size,
+                        vertex_buffer: &vertices,
+                        first_vertex: 0,
+                        // Not aligned to four bytes like it should be
+                        vertex_stride: 13,
+                        index_buffer: None,
+                        first_index: None,
+                        transform_buffer: None,
+                        transform_buffer_offset: None,
+                    }]),
+                }],
+                &[],
+            )
+        },
+        None,
+    );
+}
+
+#[gpu_test]
+static TOO_SMALL_STRIDE_BUILD: GpuTestConfiguration = GpuTestConfiguration::new()
+    .parameters(
+        TestParameters::default()
+            .test_features_limits()
+            .features(wgpu::Features::EXPERIMENTAL_RAY_TRACING_ACCELERATION_STRUCTURE)
+            // https://github.com/gfx-rs/wgpu/issues/6727
+            .skip(FailureCase::backend_adapter(wgpu::Backends::VULKAN, "AMD")),
+    )
+    .run_sync(too_small_stride_build);
+
+fn too_small_stride_build(ctx: TestingContext) {
+    let vertices = ctx.device.create_buffer_init(&BufferInitDescriptor {
+        label: None,
+        contents: &[0; mem::size_of::<[[i16; 3]; 3]>()],
+        usage: BufferUsages::BLAS_INPUT,
+    });
+
+    let blas_size = BlasTriangleGeometrySizeDescriptor {
+        // The fourth component is ignored, and it allows us to have a smaller stride.
+        vertex_format: VertexFormat::Float32x3,
+        vertex_count: 3,
+        index_format: None,
+        index_count: None,
+        flags: wgpu::AccelerationStructureGeometryFlags::empty(),
+    };
+
+    let blas = ctx.device.create_blas(
+        &CreateBlasDescriptor {
+            label: Some("BLAS"),
+            flags: wgpu::AccelerationStructureFlags::PREFER_FAST_TRACE,
+            update_mode: AccelerationStructureUpdateMode::Build,
+        },
+        BlasGeometrySizeDescriptors::Triangles {
+            descriptors: vec![blas_size.clone()],
+        },
+    );
+
+    let mut command_encoder = ctx
+        .device
+        .create_command_encoder(&CommandEncoderDescriptor {
+            label: Some("BLAS_1"),
+        });
+    fail(
+        &ctx.device,
+        || {
+            command_encoder.build_acceleration_structures(
+                &[BlasBuildEntry {
+                    blas: &blas,
+                    geometry: BlasGeometries::TriangleGeometries(vec![BlasTriangleGeometry {
+                        size: &blas_size,
+                        vertex_buffer: &vertices,
+                        first_vertex: 0,
+                        // Aligned to four bytes but too small
+                        vertex_stride: 8,
+                        index_buffer: None,
+                        first_index: None,
+                        transform_buffer: None,
+                        transform_buffer_offset: None,
+                    }]),
+                }],
+                &[],
+            )
+        },
+        None,
+    );
+}
diff --git a/wgpu-core/src/command/ray_tracing.rs b/wgpu-core/src/command/ray_tracing.rs
@@ -898,6 +898,35 @@ fn iter_blas<'a>(
                         ));
                     }
 
+                    if size_desc
+                        .vertex_format
+                        .acceleration_structure_vertex_readable_size()
+                        > mesh.vertex_stride
+                    {
+                        return Err(BuildAccelerationStructureError::VertexStrideTooSmall(
+                            blas.error_ident(),
+                            size_desc
+                                .vertex_format
+                                .acceleration_structure_vertex_readable_size(),
+                            mesh.vertex_stride,
+                        ));
+                    }
+
+                    if mesh.vertex_stride
+                        % size_desc
+                            .vertex_format
+                            .acceleration_structure_stride_alignment()
+                        != 0
+                    {
+                        return Err(BuildAccelerationStructureError::VertexStrideUnaligned(
+                            blas.error_ident(),
+                            size_desc
+                                .vertex_format
+                                .acceleration_structure_stride_alignment(),
+                            mesh.vertex_stride,
+                        ));
+                    }
+
                     match (size_desc.index_count, mesh.size.index_count) {
                         (Some(_), None) | (None, Some(_)) => {
                             return Err(
diff --git a/wgpu-core/src/ray_tracing.rs b/wgpu-core/src/ray_tracing.rs
@@ -102,6 +102,12 @@ pub enum BuildAccelerationStructureError {
     #[error("Blas {0:?} vertex formats are different, creation format: {1:?}, provided: {2:?}")]
     DifferentBlasVertexFormats(ResourceErrorIdent, VertexFormat, VertexFormat),
 
+    #[error("Blas {0:?} stride was required to be at least {1} but stride given was {2}")]
+    VertexStrideTooSmall(ResourceErrorIdent, u64, u64),
+
+    #[error("Blas {0:?} stride was required to be a multiple of {1} but stride given was {2}")]
+    VertexStrideUnaligned(ResourceErrorIdent, u64, u64),
+
     #[error("Blas {0:?} index count was provided at creation or building, but not the other")]
     BlasIndexCountProvidedMismatch(ResourceErrorIdent),
 
diff --git a/wgpu-hal/src/dx12/adapter.rs b/wgpu-hal/src/dx12/adapter.rs
@@ -471,7 +471,8 @@ impl super::Adapter {
         // Once ray tracing pipelines are supported they also will go here
         features.set(
             wgt::Features::EXPERIMENTAL_RAY_QUERY
-                | wgt::Features::EXPERIMENTAL_RAY_TRACING_ACCELERATION_STRUCTURE,
+                | wgt::Features::EXPERIMENTAL_RAY_TRACING_ACCELERATION_STRUCTURE
+                | wgt::Features::EXTENDED_ACCELERATION_STRUCTURE_VERTEX_FORMATS,
             features5.RaytracingTier == Direct3D12::D3D12_RAYTRACING_TIER_1_1
                 && shader_model >= naga::back::hlsl::ShaderModel::V6_5
                 && has_features5,
diff --git a/wgpu-hal/src/vulkan/adapter.rs b/wgpu-hal/src/vulkan/adapter.rs
@@ -779,7 +779,8 @@ impl PhysicalDeviceFeatures {
         features.set(F::DEPTH32FLOAT_STENCIL8, texture_d32_s8);
 
         features.set(
-            F::EXPERIMENTAL_RAY_TRACING_ACCELERATION_STRUCTURE,
+            F::EXPERIMENTAL_RAY_TRACING_ACCELERATION_STRUCTURE
+                | F::EXTENDED_ACCELERATION_STRUCTURE_VERTEX_FORMATS,
             caps.supports_extension(khr::deferred_host_operations::NAME)
                 && caps.supports_extension(khr::acceleration_structure::NAME)
                 && caps.supports_extension(khr::buffer_device_address::NAME),
diff --git a/wgpu-types/src/features.rs b/wgpu-types/src/features.rs
@@ -1193,6 +1193,15 @@ bitflags_array! {
         ///
         /// This is a native only feature.
         const EXPERIMENTAL_MESH_SHADER_MULTIVIEW = 1 << 49;
+
+        /// Allows usage of additional vertex formats in [BlasTriangleGeometrySizeDescriptor::vertex_format]
+        ///
+        /// Supported platforms
+        /// - Vulkan
+        /// - DX12
+        ///
+        /// [BlasTriangleGeometrySizeDescriptor::vertex_format]: super::BlasTriangleGeometrySizeDescriptor
+        const EXTENDED_ACCELERATION_STRUCTURE_VERTEX_FORMATS = 1 << 50;
     }
 
     /// Features that are not guaranteed to be supported.
@@ -1440,6 +1449,13 @@ impl Features {
         if self.contains(Self::EXPERIMENTAL_RAY_TRACING_ACCELERATION_STRUCTURE) {
             formats.push(VertexFormat::Float32x3);
         }
+        if self.contains(Self::EXTENDED_ACCELERATION_STRUCTURE_VERTEX_FORMATS) {
+            formats.push(VertexFormat::Float32x2);
+            formats.push(VertexFormat::Float16x2);
+            formats.push(VertexFormat::Float16x4);
+            formats.push(VertexFormat::Snorm16x2);
+            formats.push(VertexFormat::Snorm16x4);
+        }
         formats
     }
 }
diff --git a/wgpu-types/src/lib.rs b/wgpu-types/src/lib.rs
@@ -4918,6 +4918,29 @@ impl VertexFormat {
             Self::Float64x4 => 32,
         }
     }
+
+    /// Returns the readable size of the vertex format in and acceleration structure build (this is
+    /// slightly different from [`Self::size`])
+    #[must_use]
+    pub const fn acceleration_structure_vertex_readable_size(&self) -> u64 {
+        match self {
+            Self::Float16x2 | Self::Snorm16x2 => 4,
+            Self::Float32x3 => 12,
+            Self::Float32x2 => 8,
+            Self::Float16x4 | Self::Snorm16x4 => 6,
+            _ => unreachable!(),
+        }
+    }
+
+    /// Returns the alignment required for `wgpu::BlasTriangleGeometry::vertex_stride`
+    #[must_use]
+    pub const fn acceleration_structure_stride_alignment(&self) -> u64 {
+        match self {
+            Self::Float16x4 | Self::Float16x2 | Self::Snorm16x4 | Self::Snorm16x2 => 2,
+            Self::Float32x2 | Self::Float32x3 => 4,
+            _ => unreachable!(),
+        }
+    }
 }
 
 bitflags::bitflags! {
@@ -7378,7 +7401,7 @@ impl Default for ShaderRuntimeChecks {
 pub struct BlasTriangleGeometrySizeDescriptor {
     /// Format of a vertex position, must be [`VertexFormat::Float32x3`]
     /// with just [`Features::EXPERIMENTAL_RAY_TRACING_ACCELERATION_STRUCTURE`]
-    /// but later features may add more formats.
+    /// but [`Features::EXTENDED_ACCELERATION_STRUCTURE_VERTEX_FORMATS`] adds more.
     pub vertex_format: VertexFormat,
     /// Number of vertices.
     pub vertex_count: u32,
diff --git a/wgpu/src/api/blas.rs b/wgpu/src/api/blas.rs
@@ -100,7 +100,8 @@ pub struct BlasTriangleGeometry<'a> {
     pub vertex_buffer: &'a Buffer,
     /// Offset into the vertex buffer as a factor of the vertex stride.
     pub first_vertex: u32,
-    /// Vertex stride.
+    /// Vertex stride, must be greater than [`wgpu_types::VertexFormat::acceleration_structure_vertex_readable_size`]
+    /// of the format and must be a multiple of [`wgpu_types::VertexFormat::acceleration_structure_stride_alignment`].
     pub vertex_stride: wgt::BufferAddress,
     /// Index buffer (optional).
     pub index_buffer: Option<&'a Buffer>,
