vk: write uniforms directly to host-visible mapped memory

librashader-runtime-vk/src/draw_quad.rs

@@ -1,6 +1,7 @@
 use crate::error;
 use crate::vulkan_primitives::VulkanBuffer;
 use ash::vk;
+use std::sync::Arc;
 
 #[rustfmt::skip]
 pub(crate) static VBO_DEFAULT_FINAL: &[f32; 16] = &[
@@ -13,12 +14,12 @@ pub(crate) static VBO_DEFAULT_FINAL: &[f32; 16] = &[
 
 pub struct DrawQuad {
     buffer: VulkanBuffer,
-    device: ash::Device,
+    device: Arc<ash::Device>,
 }
 
 impl DrawQuad {
     pub fn new(
-        device: &ash::Device,
+        device: &Arc<ash::Device>,
         mem_props: &vk::PhysicalDeviceMemoryProperties,
     ) -> error::Result<DrawQuad> {
         let mut buffer = VulkanBuffer::new(

librashader-runtime-vk/src/filter_chain.rs

@@ -8,7 +8,8 @@ use crate::queue_selection::get_graphics_queue;
 use crate::render_target::{RenderTarget, DEFAULT_MVP};
 use crate::samplers::SamplerSet;
 use crate::texture::{InputImage, OwnedImage, OwnedImageLayout, VulkanImage};
-use crate::ubo_ring::VkUboRing;
+// use crate::ubo_ring::VkUboRing;
+use crate::vulkan_primitives::RawVulkanBuffer;
 use crate::vulkan_state::VulkanGraphicsPipeline;
 use crate::{error, util};
 use ash::vk;
@@ -352,8 +353,13 @@ impl FilterChainVulkan {
                 .as_ref()
                 .map(|ubo| ubo.size as usize)
                 .unwrap_or(0);
-            let uniform_storage = UniformStorage::new(
+            let uniform_storage = UniformStorage::new_with_storage(
-                ubo_size,
+                RawVulkanBuffer::new(
+                    &vulkan.device,
+                    &vulkan.memory_properties,
+                    vk::BufferUsageFlags::UNIFORM_BUFFER,
+                    ubo_size,
+                )?,
                 reflection
                     .push_constant
                     .as_ref()
@@ -383,13 +389,12 @@ impl FilterChainVulkan {
                 frames_in_flight,
             )?;
 
-            let ubo_ring = VkUboRing::new(
+            // let ubo_ring = VkUboRing::new(
-                &vulkan.device,
+            //     &vulkan.device,
-                &vulkan.memory_properties,
+            //     &vulkan.memory_properties,
-                frames_in_flight as usize,
+            //     frames_in_flight as usize,
-                ubo_size,
+            //     ubo_size,
-            )?;
+            // )?;
-            // shader_vulkan: 2026
             filters.push(FilterPass {
                 device: vulkan.device.clone(),
                 reflection,
@@ -399,7 +404,7 @@ impl FilterChainVulkan {
                 source,
                 config,
                 graphics_pipeline,
-                ubo_ring,
+                // ubo_ring,
                 frames_in_flight,
             });
         }

librashader-runtime-vk/src/filter_pass.rs

@@ -2,7 +2,7 @@ use crate::filter_chain::FilterCommon;
 use crate::render_target::RenderTarget;
 use crate::samplers::SamplerSet;
 use crate::texture::InputImage;
-use crate::ubo_ring::VkUboRing;
+use crate::vulkan_primitives::RawVulkanBuffer;
 use crate::vulkan_state::VulkanGraphicsPipeline;
 use crate::{error, VulkanImage};
 use ash::vk;
@@ -14,7 +14,7 @@ use librashader_reflect::reflect::semantics::{
 };
 use librashader_reflect::reflect::ShaderReflection;
 use librashader_runtime::binding::{BindSemantics, TextureInput};
-use librashader_runtime::uniforms::{UniformStorage, UniformStorageAccess};
+use librashader_runtime::uniforms::{NoUniformBinder, UniformStorage, UniformStorageAccess};
 use rustc_hash::FxHashMap;
 use std::sync::Arc;
 
@@ -22,12 +22,12 @@ pub struct FilterPass {
     pub device: Arc<ash::Device>,
     pub reflection: ShaderReflection,
     // pub(crate) compiled: ShaderCompilerOutput<Vec<u32>>,
-    pub(crate) uniform_storage: UniformStorage,
+    pub(crate) uniform_storage: UniformStorage<NoUniformBinder, Option<()>, RawVulkanBuffer>,
     pub uniform_bindings: FxHashMap<UniformBinding, MemberOffset>,
     pub source: ShaderSource,
     pub config: ShaderPassConfig,
     pub graphics_pipeline: VulkanGraphicsPipeline,
-    pub ubo_ring: VkUboRing,
+    // pub ubo_ring: VkUboRing,
     pub frames_in_flight: u32,
 }
 
@@ -37,7 +37,7 @@ impl TextureInput for InputImage {
     }
 }
 
-impl BindSemantics for FilterPass {
+impl BindSemantics<NoUniformBinder, Option<()>, RawVulkanBuffer> for FilterPass {
     type InputTexture = InputImage;
     type SamplerSet = SamplerSet;
     type DescriptorSet<'a> = vk::DescriptorSet;
@@ -113,10 +113,21 @@ impl FilterPass {
         );
 
         if let Some(ubo) = &self.reflection.ubo {
-            // shader_vulkan: 2554 (ra uses uses one big buffer)
+            // RetroArch uses uses one big buffer for this and handles alignment themselves.
-            // itll be simpler for us if we just use a RingBuffer<vk::Buffer> tbh.
+            // It's simpler here if we just use a buffer for each pass but not ideal.
-            self.ubo_ring
+            //
-                .bind_to_descriptor_set(descriptor, ubo.binding, &self.uniform_storage)?;
+            // May want to switch to something like VMA in the future.
+
+            // VkUboRing is no longer used, we just write directly to the storage now.
+
+            // self.ubo_ring
+            //     .bind_to_descriptor_set(descriptor, ubo.binding, &self.uniform_storage)?;
+
+            self.uniform_storage.inner_ubo().bind_to_descriptor_set(
+                descriptor,
+                ubo.binding,
+                &self.uniform_storage,
+            )?;
         }
 
         output.output.begin_pass(cmd);

librashader-runtime-vk/src/hello_triangle/pipeline.rs

@@ -12,6 +12,8 @@ use std::mem::align_of;
 const ENTRY_POINT: &CStr = unsafe { CStr::from_bytes_with_nul_unchecked(b"main\0") };
 
 #[derive(Default, Clone, Debug, Copy)]
+#[repr(C)]
+#[allow(dead_code)]
 struct Vertex {
     pos: [f32; 4],
     color: [f32; 4],

librashader-runtime-vk/src/lib.rs

@@ -14,7 +14,6 @@ mod queue_selection;
 mod render_target;
 mod samplers;
 mod texture;
-mod ubo_ring;
 mod util;
 mod vulkan_primitives;
 mod vulkan_state;

librashader-runtime-vk/src/vulkan_primitives.rs

@@ -1,6 +1,10 @@
 use crate::{error, util};
 use ash::vk;
+use librashader_runtime::uniforms::UniformStorageAccess;
 use std::ffi::c_void;
+use std::mem::ManuallyDrop;
+use std::ops::{Deref, DerefMut};
+use std::ptr::NonNull;
 use std::sync::Arc;
 
 pub struct VulkanImageMemory {
@@ -36,9 +40,9 @@ impl Drop for VulkanImageMemory {
 
 pub struct VulkanBuffer {
     pub handle: vk::Buffer,
-    device: ash::Device,
+    device: Arc<ash::Device>,
-    pub memory: vk::DeviceMemory,
+    memory: vk::DeviceMemory,
-    pub size: vk::DeviceSize,
+    size: vk::DeviceSize,
 }
 
 pub struct VulkanBufferMapHandle<'a> {
@@ -48,7 +52,7 @@ pub struct VulkanBufferMapHandle<'a> {
 
 impl VulkanBuffer {
     pub fn new(
-        device: &ash::Device,
+        device: &Arc<ash::Device>,
         mem_props: &vk::PhysicalDeviceMemoryProperties,
         usage: vk::BufferUsageFlags,
         size: usize,
@@ -112,13 +116,15 @@ impl Drop for VulkanBuffer {
 
 impl<'a> VulkanBufferMapHandle<'a> {
     pub unsafe fn copy_from(&mut self, offset: usize, src: &[u8]) {
-        std::ptr::copy_nonoverlapping(
+        let mut align = ash::util::Align::new(
-            src.as_ptr(),
             self.ptr
                 .map_addr(|original| original.wrapping_add(offset))
                 .cast(),
-            src.len(),
+            std::mem::align_of::<u8>() as u64,
+            self.buffer.size,
         );
+
+        align.copy_from_slice(src);
     }
 }
 
@@ -127,3 +133,87 @@ impl<'a> Drop for VulkanBufferMapHandle<'a> {
         unsafe { self.buffer.device.unmap_memory(self.buffer.memory) }
     }
 }
+
+pub struct RawVulkanBuffer {
+    buffer: ManuallyDrop<VulkanBuffer>,
+    ptr: NonNull<c_void>,
+}
+
+impl RawVulkanBuffer {
+    pub fn new(
+        device: &Arc<ash::Device>,
+        mem_props: &vk::PhysicalDeviceMemoryProperties,
+        usage: vk::BufferUsageFlags,
+        size: usize,
+    ) -> error::Result<Self> {
+        let buffer = ManuallyDrop::new(VulkanBuffer::new(device, mem_props, usage, size)?);
+        let ptr = unsafe {
+            NonNull::new_unchecked(device.map_memory(
+                buffer.memory,
+                0,
+                buffer.size,
+                vk::MemoryMapFlags::empty(),
+            )?)
+        };
+
+        Ok(RawVulkanBuffer { buffer, ptr })
+    }
+
+    pub fn bind_to_descriptor_set(
+        &self,
+        descriptor_set: vk::DescriptorSet,
+        binding: u32,
+        storage: &impl UniformStorageAccess,
+    ) -> error::Result<()> {
+        unsafe {
+            let buffer_info = [vk::DescriptorBufferInfo::builder()
+                .buffer(self.buffer.handle)
+                .offset(0)
+                .range(storage.ubo_slice().len() as vk::DeviceSize)
+                .build()];
+
+            let write_info = [vk::WriteDescriptorSet::builder()
+                .descriptor_type(vk::DescriptorType::UNIFORM_BUFFER)
+                .dst_set(descriptor_set)
+                .dst_binding(binding)
+                .dst_array_element(0)
+                .buffer_info(&buffer_info)
+                .build()];
+
+            self.buffer.device.update_descriptor_sets(&write_info, &[])
+        }
+        Ok(())
+    }
+}
+
+impl Drop for RawVulkanBuffer {
+    fn drop(&mut self) {
+        unsafe {
+            self.buffer.device.unmap_memory(self.buffer.memory);
+            self.ptr = NonNull::dangling();
+            if self.buffer.memory != vk::DeviceMemory::null() {
+                self.buffer.device.free_memory(self.buffer.memory, None);
+            }
+
+            if self.buffer.handle != vk::Buffer::null() {
+                self.buffer.device.destroy_buffer(self.buffer.handle, None);
+            }
+        }
+    }
+}
+
+impl Deref for RawVulkanBuffer {
+    type Target = [u8];
+
+    fn deref(&self) -> &Self::Target {
+        unsafe { std::slice::from_raw_parts(self.ptr.as_ptr().cast(), self.buffer.size as usize) }
+    }
+}
+
+impl DerefMut for RawVulkanBuffer {
+    fn deref_mut(&mut self) -> &mut Self::Target {
+        unsafe {
+            std::slice::from_raw_parts_mut(self.ptr.as_ptr().cast(), self.buffer.size as usize)
+        }
+    }
+}

librashader-runtime-vk/src/vulkan_state.rs

@@ -183,7 +183,6 @@ impl VulkanGraphicsPipeline {
         reflection: &ShaderReflection,
         replicas: u32,
     ) -> error::Result<VulkanGraphicsPipeline> {
-        // shader_vulkan 1927 (init_pipeline_layout)
         let pipeline_layout = PipelineLayoutObjects::new(reflection, replicas, device)?;
 
         let input_assembly = vk::PipelineInputAssemblyStateCreateInfo::builder()

librashader-runtime/src/binding.rs

@@ -6,6 +6,7 @@ use librashader_reflect::reflect::semantics::{
 };
 use std::collections::HashMap;
 use std::hash::BuildHasher;
+use std::ops::{Deref, DerefMut};
 
 /// Trait for input textures used during uniform binding,
 pub trait TextureInput {
@@ -47,8 +48,9 @@ where
 }
 
 /// Trait that abstracts binding of semantics to shader uniforms.
-pub trait BindSemantics<H = NoUniformBinder, C = Option<()>>
+pub trait BindSemantics<H = NoUniformBinder, C = Option<()>, S = Box<[u8]>>
 where
+    S: Deref<Target = [u8]> + DerefMut,
     H: BindUniform<C, f32>,
     H: BindUniform<C, u32>,
     H: BindUniform<C, i32>,
@@ -84,7 +86,7 @@ where
     fn bind_semantics<'a>(
         device: &Self::DeviceContext,
         sampler_set: &Self::SamplerSet,
-        uniform_storage: &mut UniformStorage<H, C>,
+        uniform_storage: &mut UniformStorage<H, C, S>,
         descriptor_set: &mut Self::DescriptorSet<'a>,
         mvp: &[f32; 16],
         frame_count: u32,

librashader-runtime/src/uniforms.rs

@@ -1,5 +1,6 @@
 use librashader_reflect::reflect::semantics::MemberOffset;
 use std::marker::PhantomData;
+use std::ops::{Deref, DerefMut};
 
 /// A scalar value that is valid as a uniform member
 pub trait UniformScalar: Copy + bytemuck::Pod {}
@@ -36,7 +37,10 @@ pub trait UniformStorageAccess {
     fn push_slice(&self) -> &[u8];
 }
 
-impl<T, H> UniformStorageAccess for UniformStorage<T, H> {
+impl<T, H, S> UniformStorageAccess for UniformStorage<T, H, S>
+where
+    S: Deref<Target = [u8]> + DerefMut,
+{
     fn ubo_pointer(&self) -> *const u8 {
         self.ubo.as_ptr()
     }
@@ -64,24 +68,29 @@ impl<T> BindUniform<Option<()>, T> for NoUniformBinder {
 }
 
 /// A helper to bind uniform variables to UBO or Push Constant Buffers.
-pub struct UniformStorage<H = NoUniformBinder, C = Option<()>> {
+pub struct UniformStorage<H = NoUniformBinder, C = Option<()>, S = Box<[u8]>>
-    ubo: Box<[u8]>,
+where
+    S: Deref<Target = [u8]> + DerefMut,
+{
+    ubo: S,
     push: Box<[u8]>,
     _h: PhantomData<H>,
     _c: PhantomData<C>,
 }
 
-impl<H, C> UniformStorage<H, C> {
+impl<H, C, S> UniformStorage<H, C, S>
-    /// Create a new `UniformStorage` with the given size for UBO and Push Constant Buffer sizes.
+where
-    pub fn new(ubo_size: usize, push_size: usize) -> Self {
+    S: Deref<Target = [u8]> + DerefMut,
-        UniformStorage {
+{
-            ubo: vec![0u8; ubo_size].into_boxed_slice(),
+    pub fn inner_ubo(&self) -> &S {
-            push: vec![0u8; push_size].into_boxed_slice(),
+        &self.ubo
-            _h: Default::default(),
-            _c: Default::default(),
-        }
     }
+}
 
+impl<H, C, S> UniformStorage<H, C, S>
+where
+    S: Deref<Target = [u8]> + DerefMut,
+{
     #[inline(always)]
     fn write_scalar_inner<T: UniformScalar>(buffer: &mut [u8], value: T, ctx: C)
     where
@@ -100,8 +109,8 @@ impl<H, C> UniformStorage<H, C> {
         H: BindUniform<C, T>,
     {
         let (buffer, offset) = match offset {
-            MemberOffset::Ubo(offset) => (&mut self.ubo, offset),
+            MemberOffset::Ubo(offset) => (self.ubo.deref_mut(), offset),
-            MemberOffset::PushConstant(offset) => (&mut self.push, offset),
+            MemberOffset::PushConstant(offset) => (self.push.deref_mut(), offset),
         };
 
         Self::write_scalar_inner(
@@ -110,10 +119,33 @@ impl<H, C> UniformStorage<H, C> {
             ctx,
         )
     }
+
+    /// Create a new `UniformStorage` with the given backing storage
+    pub fn new_with_storage(storage: S, push_size: usize) -> UniformStorage<H, C, S> {
+        UniformStorage {
+            ubo: storage,
+            push: vec![0u8; push_size].into_boxed_slice(),
+            _h: Default::default(),
+            _c: Default::default(),
+        }
+    }
 }
 
-impl<H, C> UniformStorage<H, C>
+impl<H, C> UniformStorage<H, C, Box<[u8]>> {
+    /// Create a new `UniformStorage` with the given size for UBO and Push Constant Buffer sizes.
+    pub fn new(ubo_size: usize, push_size: usize) -> UniformStorage<H, C, Box<[u8]>> {
+        UniformStorage {
+            ubo: vec![0u8; ubo_size].into_boxed_slice(),
+            push: vec![0u8; push_size].into_boxed_slice(),
+            _h: Default::default(),
+            _c: Default::default(),
+        }
+    }
+}
+
+impl<H, C, S> UniformStorage<H, C, S>
 where
+    S: Deref<Target = [u8]> + DerefMut,
     H: for<'a> BindUniform<C, &'a [f32; 4]>,
 {
     #[inline(always)]
@@ -128,8 +160,8 @@ where
     #[inline(always)]
     pub fn bind_vec4(&mut self, offset: MemberOffset, value: impl Into<[f32; 4]>, ctx: C) {
         let (buffer, offset) = match offset {
-            MemberOffset::Ubo(offset) => (&mut self.ubo, offset),
+            MemberOffset::Ubo(offset) => (self.ubo.deref_mut(), offset),
-            MemberOffset::PushConstant(offset) => (&mut self.push, offset),
+            MemberOffset::PushConstant(offset) => (self.push.deref_mut(), offset),
         };
 
         Self::write_vec4_inner(
@@ -140,8 +172,9 @@ where
     }
 }
 
-impl<H, C> UniformStorage<H, C>
+impl<H, C, S> UniformStorage<H, C, S>
 where
+    S: Deref<Target = [u8]> + DerefMut,
     H: for<'a> BindUniform<C, &'a [f32; 16]>,
 {
     #[inline(always)]
@@ -156,8 +189,8 @@ where
     #[inline(always)]
     pub fn bind_mat4(&mut self, offset: MemberOffset, value: &[f32; 16], ctx: C) {
         let (buffer, offset) = match offset {
-            MemberOffset::Ubo(offset) => (&mut self.ubo, offset),
+            MemberOffset::Ubo(offset) => (self.ubo.deref_mut(), offset),
-            MemberOffset::PushConstant(offset) => (&mut self.push, offset),
+            MemberOffset::PushConstant(offset) => (self.push.deref_mut(), offset),
         };
         Self::write_mat4_inner(
             &mut buffer[offset..][..16 * std::mem::size_of::<f32>()],