vk/11/12: bind VBO only once per frame to save on draw calls.

Cargo.lock

@@ -8,6 +8,12 @@ version = "1.0.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "f26201604c87b1e01bd3d98f8d5d9a8fcbb815e8cedb41ffccbeb4bf593a35fe"
 
+[[package]]
+name = "array-concat"
+version = "0.5.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e9180feb72ccbc07cfe5ef7fa8bbf86ca71490d5dc9ef8ea02c7298ba94e7f7d"
+
 [[package]]
 name = "array-init"
 version = "2.1.0"
@@ -884,6 +890,7 @@ dependencies = [
 name = "librashader-runtime-d3d11"
 version = "0.1.0-beta.10"
 dependencies = [
+ "array-concat",
  "bytemuck",
  "gfx-maths",
  "librashader-common",
@@ -902,6 +909,7 @@ dependencies = [
 name = "librashader-runtime-d3d12"
 version = "0.1.0-beta.8"
 dependencies = [
+ "array-concat",
  "array-init",
  "bitvec",
  "bytemuck",

librashader-runtime-d3d11/Cargo.toml

@@ -23,6 +23,7 @@ spirv_cross = { package = "librashader-spirv-cross", version = "0.23" }
 rustc-hash = "1.1.0"
 bytemuck = "1.12.3"
 rayon = "1.6.1"
+array-concat = "0.5.2"
 
 [target.'cfg(windows)'.dependencies.windows]
 version = "0.44.0"

librashader-runtime-d3d11/src/draw_quad.rs

@@ -1,3 +1,4 @@
+use array_concat::concat_arrays;
 use crate::error;
 use crate::error::assume_d3d11_init;
 use bytemuck::offset_of;
@@ -21,7 +22,7 @@ struct D3D11Vertex {
 
 const CLEAR: [f32; 4] = [1.0, 1.0, 1.0, 1.0];
 
-static OFFSCREEN_VBO_DATA: &[D3D11Vertex; 4] = &[
+const OFFSCREEN_VBO_DATA: [D3D11Vertex; 4] = [
     D3D11Vertex {
         position: [-1.0, -1.0],
         texcoord: [0.0, 1.0],
@@ -44,7 +45,7 @@ static OFFSCREEN_VBO_DATA: &[D3D11Vertex; 4] = &[
     },
 ];
 
-static FINAL_VBO_DATA: &[D3D11Vertex; 4] = &[
+const FINAL_VBO_DATA: [D3D11Vertex; 4] = [
     D3D11Vertex {
         position: [0.0, 0.0],
         texcoord: [0.0, 1.0],
@@ -67,21 +68,21 @@ static FINAL_VBO_DATA: &[D3D11Vertex; 4] = &[
     },
 ];
 
+static VBO_DATA: &[D3D11Vertex; 8] = &concat_arrays!(OFFSCREEN_VBO_DATA, FINAL_VBO_DATA);
+
 pub(crate) struct DrawQuad {
-    final_vbo: ID3D11Buffer,
     context: ID3D11DeviceContext,
-    offset: u32,
     stride: u32,
-    offscreen_vbo: ID3D11Buffer,
+    vbo: ID3D11Buffer,
 }
 
 impl DrawQuad {
     pub fn new(device: &ID3D11Device, context: &ID3D11DeviceContext) -> error::Result<DrawQuad> {
         unsafe {
-            let mut final_vbo = None;
+            let mut vbo = None;
             device.CreateBuffer(
                 &D3D11_BUFFER_DESC {
-                    ByteWidth: std::mem::size_of::<[D3D11Vertex; 4]>() as u32,
+                    ByteWidth: 2 * std::mem::size_of::<[D3D11Vertex; 4]>() as u32,
                     Usage: D3D11_USAGE_IMMUTABLE,
                     BindFlags: D3D11_BIND_VERTEX_BUFFER,
                     CPUAccessFlags: Default::default(),
@@ -89,62 +90,48 @@ impl DrawQuad {
                     StructureByteStride: 0,
                 },
                 Some(&D3D11_SUBRESOURCE_DATA {
-                    pSysMem: FINAL_VBO_DATA.as_ptr().cast(),
+                    pSysMem: VBO_DATA.as_ptr().cast(),
                     SysMemPitch: 0,
                     SysMemSlicePitch: 0,
                 }),
-                Some(&mut final_vbo),
+                Some(&mut vbo),
             )?;
-            assume_d3d11_init!(final_vbo, "CreateBuffer");
+            assume_d3d11_init!(vbo, "CreateBuffer");
-
-            let mut offscreen_vbo = None;
-            device.CreateBuffer(
-                &D3D11_BUFFER_DESC {
-                    ByteWidth: std::mem::size_of::<[D3D11Vertex; 4]>() as u32,
-                    Usage: D3D11_USAGE_IMMUTABLE,
-                    BindFlags: D3D11_BIND_VERTEX_BUFFER,
-                    CPUAccessFlags: Default::default(),
-                    MiscFlags: Default::default(),
-                    StructureByteStride: 0,
-                },
-                Some(&D3D11_SUBRESOURCE_DATA {
-                    pSysMem: OFFSCREEN_VBO_DATA.as_ptr().cast(),
-                    SysMemPitch: 0,
-                    SysMemSlicePitch: 0,
-                }),
-                Some(&mut offscreen_vbo),
-            )?;
-            assume_d3d11_init!(offscreen_vbo, "CreateBuffer");
 
             Ok(DrawQuad {
-                final_vbo,
+                vbo,
-                offscreen_vbo,
                 context: context.clone(),
-                offset: 0,
                 stride: std::mem::size_of::<D3D11Vertex>() as u32,
             })
         }
     }
 
-    pub fn bind_vertices(&self, vbo_type: QuadType) {
+    pub fn bind_vbo_for_frame(&self) {
         unsafe {
             self.context
                 .IASetPrimitiveTopology(D3D11_PRIMITIVE_TOPOLOGY_TRIANGLESTRIP);
-            let buffer = match vbo_type {
-                QuadType::Offscreen => &self.offscreen_vbo,
-                QuadType::Final => &self.final_vbo,
-            };
 
             self.context.IASetVertexBuffers(
                 0,
                 1,
-                Some(&Some(buffer.clone())),
+                Some(&Some(self.vbo.clone())),
                 Some(&self.stride),
-                Some(&self.offset),
+                Some(&0),
             );
         }
     }
 
+    pub fn draw_quad(&self, context: &ID3D11DeviceContext, vbo_type: QuadType) {
+        let offset = match vbo_type {
+            QuadType::Offscreen => 0,
+            QuadType::Final => 4,
+        };
+
+        unsafe {
+            context.Draw(4, offset);
+        }
+    }
+
     pub fn get_spirv_cross_vbo_desc() -> [D3D11_INPUT_ELEMENT_DESC; 2] {
         [
             D3D11_INPUT_ELEMENT_DESC {

librashader-runtime-d3d11/src/filter_chain.rs

@@ -17,7 +17,7 @@ use crate::error::{assume_d3d11_init, FilterChainError};
 use crate::filter_pass::{ConstantBufferBinding, FilterPass};
 use crate::framebuffer::OwnedFramebuffer;
 use crate::options::{FilterChainOptionsD3D11, FrameOptionsD3D11};
-use crate::quad_render::DrawQuad;
+use crate::draw_quad::DrawQuad;
 use crate::render_target::RenderTarget;
 use crate::samplers::SamplerSet;
 use crate::util::d3d11_compile_bound_shader;
@@ -472,6 +472,8 @@ impl FilterChainD3D11 {
         let passes_len = passes.len();
         let (pass, last) = passes.split_at_mut(passes_len - 1);
 
+        self.common.draw_quad.bind_vbo_for_frame();
+
         for (index, pass) in pass.iter_mut().enumerate() {
             source.filter = pass.config.filter;
             source.wrap_mode = pass.config.wrap_mode;

librashader-runtime-d3d11/src/filter_pass.rs

@@ -159,7 +159,6 @@ impl FilterPass {
             }
         }
         unsafe {
-            parent.draw_quad.bind_vertices(vbo_type);
             context.IASetInputLayout(&self.vertex_layout);
             context.VSSetShader(&self.vertex_shader, None);
             context.PSSetShader(&self.pixel_shader, None);
@@ -243,10 +242,7 @@ impl FilterPass {
             context.RSSetViewports(Some(&[output.output.viewport]))
         }
 
-        unsafe {
+        parent.draw_quad.draw_quad(context, vbo_type);
-            // must be under primitive topology trianglestrip with quad
-            context.Draw(4, 0);
-        }
 
         unsafe {
             // unbind resources.

librashader-runtime-d3d11/src/lib.rs

@@ -15,7 +15,7 @@ mod filter_pass;
 mod framebuffer;
 pub mod options;
 mod parameters;
-mod quad_render;
+mod draw_quad;
 mod render_target;
 mod samplers;
 mod texture;
@@ -37,8 +37,9 @@ mod tests {
     // "../test/slang-shaders/presets/crt-geom-ntsc-upscale-sharp.slangp",
     // "../test/slang-shaders/bezel/Mega_Bezel/Presets/MBZ__0__SMOOTH-ADV.slangp",
     // "../test/null.slangp",
+    const FILTER_PATH: &str = "../test/slang-shaders/scalefx/scalefx-9x.slangp";
 
-    const FILTER_PATH: &str = "../test/slang-shaders/bezel/Mega_Bezel/Presets/MBZ__0__SMOOTH-ADV.slangp";
+    // const FILTER_PATH: &str = "../test/slang-shaders/bezel/Mega_Bezel/Presets/MBZ__0__SMOOTH-ADV.slangp";
     const IMAGE_PATH: &str = "../test/finalfightlong.png";
     #[test]
     fn triangle_d3d11_args() {

librashader-runtime-d3d12/Cargo.toml

@@ -25,6 +25,7 @@ bytemuck = { version = "1.12.3", features = ["derive"] }
 array-init = "2.1.0"
 bitvec = "1.0.1"
 widestring = "1.0.2"
+array-concat = "0.5.2"
 
 rayon = "1.6.1"
 

librashader-runtime-d3d12/src/filter_chain.rs

@@ -604,6 +604,9 @@ impl FilterChainD3D12 {
             cmd.SetGraphicsRootSignature(&self.common.root_signature.handle);
             self.common.mipmap_gen.pin_root_signature(cmd);
         }
+
+        self.common.draw_quad.bind_vertices_for_frame(cmd);
+
         for (index, pass) in pass.iter_mut().enumerate() {
             source.filter = pass.config.filter;
             source.wrap_mode = pass.config.wrap_mode;

librashader-runtime-d3d12/src/filter_pass.rs

@@ -142,7 +142,6 @@ impl FilterPass {
         output: &RenderTarget,
         vbo_type: QuadType,
     ) -> error::Result<()> {
-        parent.draw_quad.bind_vertices(cmd, vbo_type);
         unsafe {
             cmd.SetPipelineState(&self.pipeline.handle);
         }
@@ -212,8 +211,7 @@ impl FilterPass {
                 bottom: output.output.size.height as i32,
             }]);
 
-            // todo put this in drawquad
+            parent.draw_quad.draw_quad(&cmd, vbo_type)
-            cmd.DrawInstanced(4, 1, 0, 0)
         }
 
         unsafe { cmd.EndRenderPass() }

librashader-runtime-d3d12/src/quad_render.rs

@@ -1,13 +1,11 @@
+use array_concat::concat_arrays;
 use crate::buffer::D3D12Buffer;
 use crate::error;
 use bytemuck::{offset_of, Pod, Zeroable};
 use librashader_runtime::quad::QuadType;
 use windows::core::PCSTR;
 use windows::Win32::Graphics::Direct3D::D3D_PRIMITIVE_TOPOLOGY_TRIANGLESTRIP;
-use windows::Win32::Graphics::Direct3D12::{
+use windows::Win32::Graphics::Direct3D12::{ID3D12Device, ID3D12GraphicsCommandList, ID3D12Resource, D3D12_INPUT_CLASSIFICATION_PER_VERTEX_DATA, D3D12_INPUT_ELEMENT_DESC, D3D12_VERTEX_BUFFER_VIEW, ID3D12GraphicsCommandList4};
-    ID3D12Device, ID3D12GraphicsCommandList, ID3D12Resource,
-    D3D12_INPUT_CLASSIFICATION_PER_VERTEX_DATA, D3D12_INPUT_ELEMENT_DESC, D3D12_VERTEX_BUFFER_VIEW,
-};
 use windows::Win32::Graphics::Dxgi::Common::DXGI_FORMAT_R32G32_FLOAT;
 
 #[repr(C)]
@@ -20,7 +18,7 @@ struct D3D12Vertex {
 
 const CLEAR: [f32; 4] = [1.0, 1.0, 1.0, 1.0];
 
-static OFFSCREEN_VBO_DATA: &[D3D12Vertex; 4] = &[
+const OFFSCREEN_VBO_DATA: [D3D12Vertex; 4] = [
     D3D12Vertex {
         position: [-1.0, -1.0],
         texcoord: [0.0, 1.0],
@@ -43,7 +41,7 @@ static OFFSCREEN_VBO_DATA: &[D3D12Vertex; 4] = &[
     },
 ];
 
-static FINAL_VBO_DATA: &[D3D12Vertex; 4] = &[
+const FINAL_VBO_DATA: [D3D12Vertex; 4] = [
     D3D12Vertex {
         position: [0.0, 0.0],
         texcoord: [0.0, 1.0],
@@ -66,58 +64,49 @@ static FINAL_VBO_DATA: &[D3D12Vertex; 4] = &[
     },
 ];
 
+static VBO_DATA: &[D3D12Vertex; 8] = &concat_arrays!(OFFSCREEN_VBO_DATA, FINAL_VBO_DATA);
+
 pub(crate) struct DrawQuad {
-    offscreen_buffer: ID3D12Resource,
+    buffer: ID3D12Resource,
-    offscreen_view: D3D12_VERTEX_BUFFER_VIEW,
+    view: D3D12_VERTEX_BUFFER_VIEW,
-    final_buffer: ID3D12Resource,
-    final_view: D3D12_VERTEX_BUFFER_VIEW,
 }
 
 impl DrawQuad {
     pub fn new(device: &ID3D12Device) -> error::Result<DrawQuad> {
         let stride = std::mem::size_of::<D3D12Vertex>() as u32;
-        let size = std::mem::size_of::<[D3D12Vertex; 4]>() as u32;
+        let size = 2 * std::mem::size_of::<[D3D12Vertex; 4]>() as u32;
-        let mut offscreen_buffer = D3D12Buffer::new(device, size as usize)?;
+        let mut buffer = D3D12Buffer::new(device, size as usize)?;
-        offscreen_buffer
+        buffer
             .map(None)?
             .slice
-            .copy_from_slice(bytemuck::cast_slice(OFFSCREEN_VBO_DATA));
+            .copy_from_slice(bytemuck::cast_slice(VBO_DATA));
 
-        let offscreen_view = D3D12_VERTEX_BUFFER_VIEW {
+        let view = D3D12_VERTEX_BUFFER_VIEW {
-            BufferLocation: offscreen_buffer.gpu_address(),
+            BufferLocation: buffer.gpu_address(),
             SizeInBytes: size,
             StrideInBytes: stride,
         };
 
-        let offscreen_buffer = offscreen_buffer.into_raw();
+        let buffer = buffer.into_raw();
-
+        Ok(DrawQuad { buffer, view })
-        let mut final_buffer = D3D12Buffer::new(device, size as usize)?;
-        final_buffer
-            .map(None)?
-            .slice
-            .copy_from_slice(bytemuck::cast_slice(FINAL_VBO_DATA));
-
-        let final_view = D3D12_VERTEX_BUFFER_VIEW {
-            BufferLocation: final_buffer.gpu_address(),
-            SizeInBytes: size,
-            StrideInBytes: stride,
-        };
-
-        let final_buffer = final_buffer.into_raw();
-
-        Ok(DrawQuad { offscreen_buffer, offscreen_view, final_buffer, final_view })
     }
 
-    pub fn bind_vertices(&self, cmd: &ID3D12GraphicsCommandList, vbo_type: QuadType) {
+    pub fn bind_vertices_for_frame(&self, cmd: &ID3D12GraphicsCommandList) {
         unsafe {
             cmd.IASetPrimitiveTopology(D3D_PRIMITIVE_TOPOLOGY_TRIANGLESTRIP);
+            cmd.IASetVertexBuffers(0, Some(&[self.view]));
+        }
+    }
 
-            let view = match vbo_type {
+    // frame uses ID3D12GraphicsCommandList4 for renderpasses, don't need to bother with the cast.
-                QuadType::Offscreen => [self.offscreen_view],
+    pub fn draw_quad(&self, cmd: &ID3D12GraphicsCommandList4, vbo_type: QuadType) {
-                QuadType::Final => [self.final_view],
+        let offset = match vbo_type {
+            QuadType::Offscreen => 0,
+            QuadType::Final => 4,
         };
 
-            cmd.IASetVertexBuffers(0, Some(&view));
+        unsafe {
+            cmd.DrawInstanced(4, 1, offset, 0)
         }
     }
 

librashader-runtime-vk/src/draw_quad.rs

@@ -55,19 +55,25 @@ impl DrawQuad {
         })
     }
 
-    pub fn bind_vbo(&self, cmd: vk::CommandBuffer, vbo: QuadType) {
+    pub fn bind_vbo_for_frame(&self, cmd: vk::CommandBuffer) {
-        let offset = match vbo {
-            QuadType::Offscreen => 0,
-            QuadType::Final => std::mem::size_of::<[f32; 16]>(),
-        };
-
         unsafe {
             self.device.cmd_bind_vertex_buffers(
                 cmd,
                 0,
                 &[self.buffer.handle],
-                &[offset as vk::DeviceSize],
+                &[0 as vk::DeviceSize],
             )
         }
     }
+
+    pub fn draw_quad(&self, cmd: vk::CommandBuffer, vbo: QuadType) {
+        let offset = match vbo {
+            QuadType::Offscreen => 0,
+            QuadType::Final => 4,
+        };
+
+        unsafe {
+            self.device.cmd_draw(cmd, 4, 1, offset, 0);
+        }
+    }
 }

librashader-runtime-vk/src/filter_chain.rs

@@ -670,6 +670,7 @@ impl FilterChainVulkan {
 
         let frame_direction = options.map_or(1, |f| f.frame_direction);
 
+        self.common.draw_quad.bind_vbo_for_frame(cmd);
         for (index, pass) in pass.iter_mut().enumerate() {
             let target = &self.output_framebuffers[index];
             source.filter_mode = pass.config.filter;

librashader-runtime-vk/src/filter_pass.rs

@@ -172,7 +172,7 @@ impl FilterPass {
                 );
             }
 
-            parent.draw_quad.bind_vbo(cmd, vbo_type);
+            // parent.draw_quad.bind_vbo(cmd, vbo_type);
 
             parent.device.cmd_set_scissor(
                 cmd,
@@ -189,7 +189,7 @@ impl FilterPass {
             parent
                 .device
                 .cmd_set_viewport(cmd, 0, &[output.output.size.into()]);
-            parent.device.cmd_draw(cmd, 4, 1, 0, 0);
+            parent.draw_quad.draw_quad(cmd, vbo_type);
             self.graphics_pipeline.end_rendering(&parent.device, cmd);
         }
         Ok(residual)

librashader-runtime-vk/src/lib.rs

@@ -45,8 +45,8 @@ mod tests {
         dbg!("finished");
         let filter = FilterChainVulkan::load_from_path(
             &base,
-            // "../test/slang-shaders/border/gameboy-player/gameboy-player-crt-royale.slangp",
+            "../test/slang-shaders/border/gameboy-player/gameboy-player-crt-royale.slangp",
-            "../test/slang-shaders/bezel/Mega_Bezel/Presets/MBZ__2__ADV-NO-REFLECT.slangp",
+            // "../test/slang-shaders/bezel/Mega_Bezel/Presets/MBZ__2__ADV-NO-REFLECT.slangp",
             // "../test/basic.slangp",
             Some(&FilterChainOptionsVulkan {
                 frames_in_flight: 3,