Merge branch 'master' into 84-mip-levels

Cargo.toml

@@ -25,7 +25,7 @@ ash-window = "0.13"
 derive_builder = "0.20"
 gpu-allocator = "0.27"
 log = "0.4"
-ordered-float = "4.1"
+ordered-float = "4.2"
 parking_lot = { version = "0.12", optional = true }
 paste = "1.0"
 profiling = "1.0"
@@ -39,10 +39,10 @@ ash-molten = "0.19"
 [dev-dependencies]
 anyhow = "1.0"
 bmfont = { version = "0.3", default-features = false }
-bytemuck = "1.14"
+bytemuck = "1.16"
 clap = { version = "4.5", features = ["derive"] }
-glam = { version = "0.27", features = ["bytemuck"] }
-half = { version = "2.3", features = ["bytemuck"] }
+glam = { version = "0.28", features = ["bytemuck"] }
+half = { version = "2.4", features = ["bytemuck"] }
 hassle-rs = "0.11"
 image = "0.25"
 inline-spirv = "0.2"

contrib/screen-13-fx/src/image_loader.rs

@@ -6,10 +6,6 @@ use {
 #[cfg(debug_assertions)]
 use log::warn;
 
-fn align_up_u32(val: u32, atom: u32) -> u32 {
-    (val + atom - 1) & !(atom - 1)
-}
-
 /// Describes the channels and pixel stride of an image format
 #[derive(Clone, Copy, Debug)]
 pub enum ImageFormat {
@@ -133,7 +129,7 @@ impl ImageLoader {
         );
 
         #[cfg(debug_assertions)]
-        if pixels.len() > align_up_u32(format.stride() as u32 * width * height, 4) as usize {
+        if pixels.len() > (format.stride() as u32 * width * height).next_multiple_of(4) as usize {
             warn!("unused data");
         }
 
@@ -156,7 +152,7 @@ impl ImageLoader {
 
                 //trace!("{bitmap_width}x{bitmap_height} Stride={bitmap_stride}");
 
-                let pixel_buf_stride = align_up_u32(stride, 12);
+                let pixel_buf_stride = stride.next_multiple_of(12);
                 let pixel_buf_len = (pixel_buf_stride * height) as vk::DeviceSize;
 
                 //trace!("pixel_buf_len={pixel_buf_len} pixel_buf_stride={pixel_buf_stride}");

examples/cpu_readback.rs

@@ -52,5 +52,7 @@ fn main() -> Result<(), DriverError> {
     println!("Waited {}μs", (Instant::now() - started).as_micros());
 
     // It is now safe to read back what we did!
-    Ok(println!("{:?}", Buffer::mapped_slice(&dst_buf)))
+    println!("{:?}", Buffer::mapped_slice(&dst_buf));
+
+    Ok(())
 }

examples/font_bmp.rs

@@ -142,11 +142,7 @@ fn main() -> anyhow::Result<()> {
             .record_compute(move |compute, _| {
                 compute
                     .push_constants(&elapsed_time.as_secs_f32().to_ne_bytes())
-                    .dispatch(
-                        (frame.width + subgroup_size - 1) / subgroup_size,
-                        frame.height,
-                        1,
-                    );
+                    .dispatch(frame.width.div_ceil(subgroup_size), frame.height, 1);
             });
 
         // Print some text onto the image

examples/ray_omni.rs

@@ -148,30 +148,31 @@ fn create_blas(
     device: &Arc<Device>,
     models: &[&Model],
 ) -> Result<Arc<AccelerationStructure>, DriverError> {
-    let info = AccelerationStructureGeometryInfo {
-        ty: vk::AccelerationStructureTypeKHR::BOTTOM_LEVEL,
-        flags: vk::BuildAccelerationStructureFlagsKHR::PREFER_FAST_TRACE,
-        geometries: models
+    let info = AccelerationStructureGeometryInfo::blas(
+        models
             .iter()
-            .map(|model| AccelerationStructureGeometry {
-                max_primitive_count: model.index_count / 3,
-                flags: vk::GeometryFlagsKHR::OPAQUE,
-                geometry: AccelerationStructureGeometryData::Triangles {
-                    index_data: DeviceOrHostAddress::DeviceAddress(Buffer::device_address(
-                        &model.index_buf,
-                    )),
-                    index_type: vk::IndexType::UINT32,
-                    max_vertex: model.vertex_count,
-                    transform_data: None,
-                    vertex_data: DeviceOrHostAddress::DeviceAddress(Buffer::device_address(
-                        &model.vertex_buf,
-                    )),
-                    vertex_format: vk::Format::R32G32B32_SFLOAT,
-                    vertex_stride: 24,
-                },
+            .map(|model| {
+                (
+                    AccelerationStructureGeometry {
+                        max_primitive_count: model.index_count / 3,
+                        flags: vk::GeometryFlagsKHR::OPAQUE,
+                        geometry: AccelerationStructureGeometryData::triangles(
+                            Buffer::device_address(&model.index_buf),
+                            vk::IndexType::UINT32,
+                            model.vertex_count,
+                            None,
+                            Buffer::device_address(&model.vertex_buf),
+                            vk::Format::R32G32B32_SFLOAT,
+                            24,
+                        ),
+                    },
+                    vk::AccelerationStructureBuildRangeInfoKHR::default()
+                        .primitive_count(model.index_count / 3),
+                )
             })
-            .collect(),
-    };
+            .collect::<Box<_>>(),
+    )
+    .flags(vk::BuildAccelerationStructureFlagsKHR::PREFER_FAST_TRACE);
     let size = AccelerationStructure::size_of(device, &info);
 
     let mut render_graph = RenderGraph::new();
@@ -196,15 +197,7 @@ fn create_blas(
         .to_builder()
         .alignment(accel_struct_scratch_offset_alignment),
     )?);
-    let build_ranges = models
-        .iter()
-        .map(|model| vk::AccelerationStructureBuildRangeInfoKHR {
-            primitive_count: model.index_count / 3,
-            primitive_offset: 0,
-            first_vertex: 0,
-            transform_offset: 0,
-        })
-        .collect::<Box<_>>();
+    let scratch_data = render_graph.node_device_address(scratch_buf);
 
     let mut pass = render_graph.begin_pass("Build BLAS");
 
@@ -219,7 +212,7 @@ fn create_blas(
     pass.access_node(blas, AccessType::AccelerationStructureBuildWrite)
         .access_node(scratch_buf, AccessType::AccelerationStructureBufferWrite)
         .record_acceleration(move |accel, _| {
-            accel.build_structure(blas, scratch_buf, &info, &build_ranges);
+            accel.build_structure(&info, blas, scratch_data);
         });
 
     let blas = render_graph.unbind_node(blas);
@@ -358,18 +351,14 @@ fn create_tlas(
         buffer
     });
 
-    let info = AccelerationStructureGeometryInfo {
-        ty: vk::AccelerationStructureTypeKHR::TOP_LEVEL,
-        flags: vk::BuildAccelerationStructureFlagsKHR::PREFER_FAST_TRACE,
-        geometries: vec![AccelerationStructureGeometry {
-            max_primitive_count: 2,
-            flags: vk::GeometryFlagsKHR::OPAQUE,
-            geometry: AccelerationStructureGeometryData::Instances {
-                array_of_pointers: false,
-                data: DeviceOrHostAddress::DeviceAddress(Buffer::device_address(&instance_buf)),
-            },
-        }],
-    };
+    let info = AccelerationStructureGeometryInfo::tlas([(
+        AccelerationStructureGeometry::opaque(
+            2,
+            AccelerationStructureGeometryData::instances(Buffer::device_address(&instance_buf)),
+        ),
+        vk::AccelerationStructureBuildRangeInfoKHR::default().primitive_count(1),
+    )])
+    .flags(vk::BuildAccelerationStructureFlagsKHR::PREFER_FAST_TRACE);
     let size = AccelerationStructure::size_of(device, &info);
     let tlas =
         render_graph.bind_node(pool.lease(AccelerationStructureInfo::tlas(size.create_size))?);
@@ -391,6 +380,7 @@ fn create_tlas(
             .alignment(accel_struct_scratch_offset_alignment),
         )?,
     );
+    let scratch_data = render_graph.node_device_address(scratch_buf);
     let blas = render_graph.bind_node(blas);
     let instance_buf = render_graph.bind_node(instance_buf);
 
@@ -401,17 +391,7 @@ fn create_tlas(
         .access_node(scratch_buf, AccessType::AccelerationStructureBufferWrite)
         .access_node(tlas, AccessType::AccelerationStructureBuildWrite)
         .record_acceleration(move |accel, _| {
-            accel.build_structure(
-                tlas,
-                scratch_buf,
-                &info,
-                &[vk::AccelerationStructureBuildRangeInfoKHR {
-                    first_vertex: 0,
-                    primitive_count: 1,
-                    primitive_offset: 0,
-                    transform_offset: 0,
-                }],
-            )
+            accel.build_structure(&info, tlas, scratch_data);
         });
 
     Ok(tlas)