Add support for macOS (and other GPUs without hardware ray tracing su…

…pport) (#46) * It works.. I guess * Getting a device timeout * Adjust script * Switch to universal libdxcompiler.dynlib * Comment out rust post combine shader * Use a while loop in the rust post combine shader instead of a for loop * Switch back to the rust shader for post combine * Check updated rust shaders into this branch * Change commented out lines into cfg statements * Run cargo fmt * Remove the ray-tracing feature and replace it with a runtime extension check * Apply suggestions * Apply suggestions, should be good to merge as a MVP afaik * Make clippy happy Co-authored-by: Ashley <ashley@Ashleys-MacBook-Pro.local> Co-authored-by: Ashley Ruglys <ashley.ruglys@gmai.com>
EmbarkStudios · Feb 27, 2022 · 79b7a74 · 79b7a74
1 parent fdf3a4e
commit 79b7a74
Show file tree

Hide file tree

Showing 16 changed files with 333 additions and 208 deletions.
diff --git a/Cargo.lock b/Cargo.lock
diff --git a/README.md b/README.md
@@ -73,6 +73,10 @@ Operating systems:
 * `uuid-dev`
 * In case the bundled `libdxcompiler.so` doesn't work: https://github.com/microsoft/DirectXShaderCompiler#downloads
 
+### (Some) MacOS dependencies
+
+* `ossp-uuid` (`brew install ossp-uuid`)
+
 ## Building and running
 
 To build `kajiya` and its tools, [you need Rust](https://www.rust-lang.org/tools/install).

diff --git a/assets/fonts/LICENSE.txt b/assets/fonts/LICENSE.txt
diff --git a/crates/lib/kajiya-backend/Cargo.toml b/crates/lib/kajiya-backend/Cargo.toml
@@ -17,7 +17,7 @@ derive_builder = { version = "0.9", default-features = false }
 futures = "0.3"
 glam = "0.18"
 gpu-allocator = { git = "https://github.com/Traverse-Research/gpu-allocator.git", rev = "e66d062cbd73a6c98834fc3e3acef98318097156" }
-hassle-rs = "0.4"
+hassle-rs = "0.5"
 hotwatch = "0.4"
 lazy_static = "1.4"
 log = "0.4"
@@ -36,7 +36,5 @@ turbosloth = { git = "https://github.com/h3r2tic/turbosloth.git", rev = "92030af
 vk-sync = { git = "https://github.com/h3r2tic/vk-sync-rs", rev = "cb5bbf2" }
 
 [features]
-default = ["ray-tracing"]
 #default = []
-ray-tracing = []
 dlss = []
diff --git a/crates/lib/kajiya-backend/src/lib.rs b/crates/lib/kajiya-backend/src/lib.rs
@@ -16,9 +16,4 @@ pub use file::{canonical_path_from_vfs, normalized_path_from_vfs, set_vfs_mount_
 pub use gpu_allocator;
 pub use rspirv_reflect;
 pub use vk_sync;
-pub use vulkan::{
-    device::Device,
-    image::*,
-    shader::{MAX_BINDLESS_DESCRIPTOR_COUNT, MAX_DESCRIPTOR_SETS},
-    RenderBackend,
-};
+pub use vulkan::{device::Device, image::*, shader::MAX_DESCRIPTOR_SETS, RenderBackend};
diff --git a/crates/lib/kajiya-backend/src/vulkan/device.rs b/crates/lib/kajiya-backend/src/vulkan/device.rs
@@ -21,6 +21,11 @@ use std::{
     sync::Arc,
 };
 
+/// Descriptor count to subtract from the max bindless descriptor count,
+/// so that we don't overflow the max when using bindless _and_ non-bindless descriptors
+/// in the same shader stage.
+pub const RESERVED_DESCRIPTOR_COUNT: u32 = 32;
+
 pub struct Queue {
     pub raw: vk::Queue,
     pub family: QueueFamily,
@@ -145,6 +150,8 @@ pub struct Device {
     pub ray_tracing_pipeline_properties: vk::PhysicalDeviceRayTracingPipelinePropertiesKHR,
 
     frames: [Mutex<Arc<DeviceFrame>>; 2],
+
+    ray_tracing_enabled: bool,
 }
 
 // Allowing `Send` on `frames` is technically unsound. There are some checks
@@ -156,8 +163,26 @@ unsafe impl Send for Device {}
 unsafe impl Sync for Device {}
 
 impl Device {
-    fn extension_names(pdevice: &Arc<PhysicalDevice>) -> Vec<*const i8> {
-        let mut device_extension_names_raw = vec![
+    pub fn create(pdevice: &Arc<PhysicalDevice>) -> Result<Arc<Self>> {
+        let supported_extensions: HashSet<String> = unsafe {
+            let extension_properties = pdevice
+                .instance
+                .raw
+                .enumerate_device_extension_properties(pdevice.raw)?;
+            debug!("Extension properties:\n{:#?}", &extension_properties);
+
+            extension_properties
+                .iter()
+                .map(|ext| {
+                    std::ffi::CStr::from_ptr(ext.extension_name.as_ptr() as *const c_char)
+                        .to_string_lossy()
+                        .as_ref()
+                        .to_owned()
+                })
+                .collect()
+        };
+
+        let mut device_extension_names = vec![
             vk::ExtDescriptorIndexingFn::name().as_ptr(),
             vk::ExtScalarBlockLayoutFn::name().as_ptr(),
             vk::KhrMaintenance1Fn::name().as_ptr(),
@@ -168,10 +193,8 @@ impl Device {
             vk::KhrImagelessFramebufferFn::name().as_ptr(),
             vk::KhrImageFormatListFn::name().as_ptr(),
             vk::KhrDescriptorUpdateTemplateFn::name().as_ptr(),
-            vk::KhrDrawIndirectCountFn::name().as_ptr(),
             // Rust-GPU
             vk::KhrShaderFloat16Int8Fn::name().as_ptr(),
-            vk::KhrVulkanMemoryModelFn::name().as_ptr(),
             // DLSS
             #[cfg(feature = "dlss")]
             {
@@ -185,55 +208,47 @@ impl Device {
             vk::NvxImageViewHandleFn::name().as_ptr(),
         ];
 
-        #[cfg(feature = "ray-tracing")]
-        {
-            device_extension_names_raw.extend(
-                [
-                    vk::KhrPipelineLibraryFn::name().as_ptr(),        // rt dep
-                    vk::KhrDeferredHostOperationsFn::name().as_ptr(), // rt dep
-                    vk::KhrBufferDeviceAddressFn::name().as_ptr(),    // rt dep
-                    vk::KhrAccelerationStructureFn::name().as_ptr(),
-                    vk::KhrRayTracingPipelineFn::name().as_ptr(),
-                    //vk::KhrRayQueryFn::name().as_ptr(),
-                ]
-                .iter(),
-            );
-        }
+        let ray_tracing_extensions = [
+            vk::KhrVulkanMemoryModelFn::name().as_ptr(), // used in ray tracing shaders
+            vk::KhrPipelineLibraryFn::name().as_ptr(),   // rt dep
+            vk::KhrDeferredHostOperationsFn::name().as_ptr(), // rt dep
+            vk::KhrBufferDeviceAddressFn::name().as_ptr(), // rt dep
+            vk::KhrAccelerationStructureFn::name().as_ptr(),
+            vk::KhrRayTracingPipelineFn::name().as_ptr(),
+        ];
 
-        if pdevice.presentation_requested {
-            device_extension_names_raw.push(khr::Swapchain::name().as_ptr());
-        }
+        let ray_tracing_enabled = unsafe {
+            ray_tracing_extensions.iter().all(|ext| {
+                let ext = std::ffi::CStr::from_ptr(*ext).to_string_lossy();
 
-        device_extension_names_raw
-    }
+                let supported = supported_extensions.contains(ext.as_ref());
 
-    pub fn create(pdevice: &Arc<PhysicalDevice>) -> Result<Arc<Self>> {
-        let device_extension_names = Self::extension_names(pdevice);
+                if !supported {
+                    log::info!("Ray tracing extension not supported: {}", ext);
+                }
 
-        unsafe {
-            let extension_properties = pdevice
-                .instance
-                .raw
-                .enumerate_device_extension_properties(pdevice.raw)?;
-            debug!("Extension properties:\n{:#?}", &extension_properties);
+                supported
+            })
+        };
 
-            let supported_extensions: HashSet<String> = extension_properties
-                .iter()
-                .map(|ext| {
-                    std::ffi::CStr::from_ptr(ext.extension_name.as_ptr() as *const c_char)
-                        .to_string_lossy()
-                        .as_ref()
-                        .to_owned()
-                })
-                .collect();
+        if ray_tracing_enabled {
+            log::info!("All ray tracing extensions are supported");
+
+            device_extension_names.extend(ray_tracing_extensions.iter());
+        }
 
+        if pdevice.presentation_requested {
+            device_extension_names.push(khr::Swapchain::name().as_ptr());
+        }
+
+        unsafe {
             for &ext in &device_extension_names {
                 let ext = std::ffi::CStr::from_ptr(ext).to_string_lossy();
                 if !supported_extensions.contains(ext.as_ref()) {
                     panic!("Device extension not supported: {}", ext);
                 }
             }
-        };
+        }
 
         let priorities = [1.0];
 
@@ -264,11 +279,9 @@ impl Device {
         let mut get_buffer_device_address_features =
             ash::vk::PhysicalDeviceBufferDeviceAddressFeatures::default();
 
-        #[cfg(feature = "ray-tracing")]
         let mut acceleration_structure_features =
             ash::vk::PhysicalDeviceAccelerationStructureFeaturesKHR::default();
 
-        #[cfg(feature = "ray-tracing")]
         let mut ray_tracing_pipeline_features =
             ash::vk::PhysicalDeviceRayTracingPipelineFeaturesKHR::default();
 
@@ -283,8 +296,7 @@ impl Device {
                 .push_next(&mut vulkan_memory_model)
                 .push_next(&mut get_buffer_device_address_features);
 
-            #[cfg(feature = "ray-tracing")]
-            {
+            if ray_tracing_enabled {
                 features2 = features2
                     .push_next(&mut acceleration_structure_features)
                     .push_next(&mut ray_tracing_pipeline_features);
@@ -311,9 +323,7 @@ impl Device {
 
                 assert!(descriptor_indexing.shader_uniform_texel_buffer_array_dynamic_indexing != 0);
                 assert!(descriptor_indexing.shader_storage_texel_buffer_array_dynamic_indexing != 0);
-                assert!(descriptor_indexing.shader_uniform_buffer_array_non_uniform_indexing != 0);
                 assert!(descriptor_indexing.shader_sampled_image_array_non_uniform_indexing != 0);
-                assert!(descriptor_indexing.shader_storage_buffer_array_non_uniform_indexing != 0);
                 assert!(descriptor_indexing.shader_storage_image_array_non_uniform_indexing != 0);
                 assert!(descriptor_indexing.shader_uniform_texel_buffer_array_non_uniform_indexing != 0);
                 assert!(descriptor_indexing.shader_storage_texel_buffer_array_non_uniform_indexing != 0);
@@ -327,18 +337,20 @@ impl Device {
 
                 assert!(shader_float16_int8.shader_int8 != 0);
 
-                assert!(vulkan_memory_model.vulkan_memory_model != 0);
+                if ray_tracing_enabled {
+                    assert!(descriptor_indexing.shader_uniform_buffer_array_non_uniform_indexing != 0);
+                    assert!(descriptor_indexing.shader_storage_buffer_array_non_uniform_indexing != 0);
+
+                    assert!(vulkan_memory_model.vulkan_memory_model != 0);
 
-                #[cfg(feature = "ray-tracing")]
-                {
                     assert!(acceleration_structure_features.acceleration_structure != 0);
                     assert!(acceleration_structure_features.descriptor_binding_acceleration_structure_update_after_bind != 0);
 
                     assert!(ray_tracing_pipeline_features.ray_tracing_pipeline != 0);
                     assert!(ray_tracing_pipeline_features.ray_tracing_pipeline_trace_rays_indirect != 0);
-                }
 
-                assert!(get_buffer_device_address_features.buffer_device_address != 0);
+                    assert!(get_buffer_device_address_features.buffer_device_address != 0);
+                }
             }
 
             let device_create_info = vk::DeviceCreateInfo::builder()
@@ -412,6 +424,7 @@ impl Device {
                     Mutex::new(Arc::new(frame1)),
                     //Mutex::new(Arc::new(frame2)),
                 ],
+                ray_tracing_enabled,
             }))
         }
     }
@@ -601,6 +614,20 @@ impl Device {
     pub fn debug_utils(&self) -> Option<&DebugUtils> {
         self.instance.debug_utils.as_ref()
     }
+
+    pub fn max_bindless_descriptor_count(&self) -> u32 {
+        (512 * 1024).min(
+            self.pdevice
+                .properties
+                .limits
+                .max_per_stage_descriptor_sampled_images
+                - RESERVED_DESCRIPTOR_COUNT,
+        )
+    }
+
+    pub fn ray_tracing_enabled(&self) -> bool {
+        self.ray_tracing_enabled
+    }
 }
 
 impl Drop for Device {

diff --git a/crates/lib/kajiya-backend/src/vulkan/shader.rs b/crates/lib/kajiya-backend/src/vulkan/shader.rs
@@ -19,7 +19,6 @@ use std::{
 };
 
 pub const MAX_DESCRIPTOR_SETS: usize = 4;
-pub const MAX_BINDLESS_DESCRIPTOR_COUNT: usize = 512 * 1024;
 
 type DescriptorSetLayout = HashMap<u32, rspirv_reflect::DescriptorInfo>;
 type StageDescriptorSetLayouts = HashMap<u32, DescriptorSetLayout>;
@@ -196,7 +195,7 @@ pub fn create_descriptor_set_layouts(
                             rspirv_reflect::DescriptorDimensionality::Single => 1,
                             rspirv_reflect::DescriptorDimensionality::Array(size) => size,
                             rspirv_reflect::DescriptorDimensionality::RuntimeArray => {
-                                MAX_BINDLESS_DESCRIPTOR_COUNT as u32
+                                device.max_bindless_descriptor_count()
                             }
                         };
 

diff --git a/crates/lib/kajiya-rg/Cargo.toml b/crates/lib/kajiya-rg/Cargo.toml
@@ -15,8 +15,3 @@ log = "0.4"
 parking_lot = "0.11"
 puffin = "0.11.0"
 turbosloth = { git = "https://github.com/h3r2tic/turbosloth.git", rev = "92030af" }
-
-[features]
-default = ["ray-tracing"]
-#default = []
-ray-tracing = []
diff --git a/crates/lib/kajiya-rg/src/temporal.rs b/crates/lib/kajiya-rg/src/temporal.rs
@@ -131,6 +131,10 @@ impl TemporalRenderGraph {
             temporal_state: state,
         }
     }
+
+    pub fn device(&self) -> &Device {
+        self.device.as_ref()
+    }
 }
 
 pub trait GetOrCreateTemporal<Desc: ResourceDesc> {

diff --git a/crates/lib/kajiya/src/bindless_descriptor_set.rs b/crates/lib/kajiya/src/bindless_descriptor_set.rs
@@ -1,6 +1,6 @@
 use std::collections::HashMap;
 
-use kajiya_backend::{ash::vk, rspirv_reflect, vulkan::device, MAX_BINDLESS_DESCRIPTOR_COUNT};
+use kajiya_backend::{ash::vk, rspirv_reflect, vulkan::device};
 
 lazy_static::lazy_static! {
     pub static ref BINDLESS_DESCRIPTOR_SET_LAYOUT: HashMap<u32, rspirv_reflect::DescriptorInfo> = [
@@ -60,7 +60,7 @@ pub fn create_bindless_descriptor_set(device: &device::Device) -> vk::Descriptor
                             .build(),
                         vk::DescriptorSetLayoutBinding::builder()
                             .binding(2)
-                            .descriptor_count(MAX_BINDLESS_DESCRIPTOR_COUNT as _)
+                            .descriptor_count(device.max_bindless_descriptor_count() as _)
                             .descriptor_type(vk::DescriptorType::SAMPLED_IMAGE)
                             .stage_flags(vk::ShaderStageFlags::ALL)
                             .build(),
@@ -80,7 +80,7 @@ pub fn create_bindless_descriptor_set(device: &device::Device) -> vk::Descriptor
         },
         vk::DescriptorPoolSize {
             ty: vk::DescriptorType::SAMPLED_IMAGE,
-            descriptor_count: MAX_BINDLESS_DESCRIPTOR_COUNT as _,
+            descriptor_count: device.max_bindless_descriptor_count() as _,
         },
     ];
 
@@ -95,7 +95,7 @@ pub fn create_bindless_descriptor_set(device: &device::Device) -> vk::Descriptor
             .unwrap()
     };
 
-    let variable_descriptor_count = MAX_BINDLESS_DESCRIPTOR_COUNT as _;
+    let variable_descriptor_count = device.max_bindless_descriptor_count() as _;
     let mut variable_descriptor_count_allocate_info =
         vk::DescriptorSetVariableDescriptorCountAllocateInfo::builder()
             .descriptor_counts(std::slice::from_ref(&variable_descriptor_count))

diff --git a/crates/lib/kajiya/src/default_world_renderer.rs b/crates/lib/kajiya/src/default_world_renderer.rs
@@ -46,7 +46,9 @@ impl WorldRenderer {
         }
 
         // Build an empty TLAS to create the resources. We'll update it at runtime.
-        world_renderer.build_ray_tracing_top_level_acceleration();
+        if backend.device.ray_tracing_enabled() {
+            world_renderer.build_ray_tracing_top_level_acceleration();
+        }
 
         Ok(world_renderer)
     }