Replace the blit cascade in record_blur with shader-based dual-Kawase,

matching the OpenGL backend's algorithm. The blit cascade was a much weaker
   filter that degenerated badly on thin surfaces.

   The new path adds blur.vert / blur_down.frag (5-tap) / blur_up.frag (8-tap),
   a single-binding blur descriptor set layout, and per-format down/up pipelines.
   BLUR_SCRATCH_USAGE gains COLOR_ATTACHMENT so the scratch images can be both
   sampled and rendered into. Cache hit fast path and masked composite are
   unchanged.

build/vulkan/hash.rs

@@ -25,6 +25,9 @@ pub const TREES: &[Tree] = &[Tree {
         "rounded_tex.vert",
         "blur_composite.vert",
         "blur_composite.frag",
+        "blur.vert",
+        "blur_down.frag",
+        "blur_up.frag",
         "legacy/fill.frag",
         "legacy/fill.vert",
         "legacy/tex.vert",

src/gfx_apis/vulkan/blur.rs

@@ -2,8 +2,9 @@ use {
     crate::gfx_apis::vulkan::{
         VulkanError,
         image::{QueueFamily, QueueState, VulkanImage, VulkanImageMemory},
+        pipeline::VulkanPipeline,
         renderer::VulkanRenderer,
-        shaders::BlurCompositePushConstants,
+        shaders::{BlurCompositePushConstants, BlurPushConstants},
     },
     ash::vk::{
         AccessFlags2, AttachmentLoadOp, AttachmentStoreOp, BlitImageInfo2, CommandBuffer,
@@ -16,13 +17,14 @@ use {
     },
     gpu_alloc::UsageFlags,
     run_on_drop::on_drop,
-    std::{cell::Cell, collections::hash_map::Entry, rc::Rc, slice},
+    std::{cell::Cell, rc::Rc, slice},
 };
 
 const BLUR_SCRATCH_USAGE: ImageUsageFlags = ImageUsageFlags::from_raw(
     ImageUsageFlags::TRANSFER_SRC.as_raw()
         | ImageUsageFlags::TRANSFER_DST.as_raw()
-        | ImageUsageFlags::SAMPLED.as_raw(),
+        | ImageUsageFlags::SAMPLED.as_raw()
+        | ImageUsageFlags::COLOR_ATTACHMENT.as_raw(),
 );
 
 pub(super) struct BlurMaskRecord<'a> {
@@ -42,12 +44,17 @@ impl VulkanRenderer {
     ) -> Result<Rc<VulkanImage>, VulkanError> {
         let key = (width, height, format);
         let cached = &mut *self.blur_scratch.borrow_mut();
-        let entry = cached.entry(key);
+
-        if let Entry::Occupied(e) = &entry
+        if let Some(weak) = cached.get(&key) {
-            && let Some(img) = e.get().upgrade()
+            if let Some(img) = weak.upgrade() {
-        {
+                if Rc::strong_count(&img) == 1 {
+                    img.is_undefined.set(false);
+                    img.contents_are_undefined.set(false);
                     return Ok(img);
                 }
+            }
+        }
+
         let create_info = ImageCreateInfo::default()
             .image_type(ImageType::TYPE_2D)
             .format(format)
@@ -77,7 +84,6 @@ impl VulkanRenderer {
                 .bind_image_memory(image, allocation.memory, allocation.offset)
         };
         res.map_err(VulkanError::BindImageMemory)?;
-        // No view needed (we only blit), but VulkanImage requires one.
         let image_view_create_info = ImageViewCreateInfo::default()
             .image(image)
             .format(format)
@@ -96,8 +102,6 @@ impl VulkanRenderer {
         };
         let view = view.map_err(VulkanError::CreateImageView)?;
         destroy_image.forget();
-        // Reuse the BLEND_FORMAT placeholder; the format field is informational
-        // here, blit ops use the actual VkFormat above.
         let img = Rc::new(VulkanImage {
             renderer: self.clone(),
             format: crate::gfx_apis::vulkan::format::BLEND_FORMAT,
@@ -117,28 +121,30 @@ impl VulkanRenderer {
             sampled_image_descriptor: None,
             execution_version: Default::default(),
         });
-        match entry {
+        cached.insert(key, Rc::downgrade(&img));
-            Entry::Occupied(mut e) => {
-                e.insert(Rc::downgrade(&img));
-            }
-            Entry::Vacant(e) => {
-                e.insert(Rc::downgrade(&img));
-            }
-        }
         Ok(img)
     }
 
-    /// Records a backdrop blur of the given pixel rect on the target image.
+    /// Records a dual-Kawase backdrop blur of the given pixel rect on the target
-    /// Caller is responsible for ending the current dynamic render pass before
+    /// image. Caller is responsible for ending the current dynamic render pass
-    /// invoking, and for restarting it afterward (with LOAD).
+    /// before invoking, and for restarting it afterward (with LOAD).
+    ///
+    /// If `cached_blur` is Some, the cascade is skipped and that image is used
+    /// directly as the blurred input to the composite. The mask must also be Some
+    /// in that case, since cache+no-mask is just a no-op (blit-back of cached).
+    /// On a cache miss, the level-0 scratch image (holding the blurred result)
+    /// is returned via `out_blur_image` for the caller to store in the cache.
     pub(super) fn record_blur(
         self: &Rc<Self>,
         buf: CommandBuffer,
         target: &VulkanImage,
         rect: [i32; 4],
         passes: u8,
+        offset: f32,
         scratch_out: &mut Vec<Rc<VulkanImage>>,
         mask: Option<&BlurMaskRecord<'_>>,
+        cached_blur: Option<&Rc<VulkanImage>>,
+        out_blur_image: &mut Option<Rc<VulkanImage>>,
     ) -> Result<(), VulkanError> {
         let [x1, y1, x2, y2] = rect;
         let x1 = x1.max(0).min(target.width as i32);
@@ -150,7 +156,26 @@ impl VulkanRenderer {
         if w < 4 || h < 4 {
             return Ok(());
         }
-        let passes = passes.clamp(1, 6) as u32;
+        let passes = passes.clamp(1, 8) as u32;
+        let offset = offset.max(0.0);
+
+        let dev = &self.device.device;
+
+        // Cache hit fast path: skip cascade, just composite from cached image.
+        // The format check matters because BlurBarrier may run in either the
+        // BlendBuffer pass (linear format) or the FrameBuffer pass (gamma)
+        // depending on whether BB was elided. A cached image from one pass
+        // has the wrong format for the other.
+        if let (Some(cached), Some(mask)) = (cached_blur, mask)
+            && cached.width == w
+            && cached.height == h
+            && cached.format.vk_format == target.format.vk_format
+        {
+            self.record_blur_composite_only(buf, target, cached, mask, [x1, y1, x2, y2])?;
+            // Hold cached image alive for this frame's GPU execution.
+            scratch_out.push(cached.clone());
+            return Ok(());
+        }
 
         let format = target.format.vk_format;
         let mut levels: Vec<Rc<VulkanImage>> = Vec::with_capacity(passes as usize + 1);
@@ -163,7 +188,8 @@ impl VulkanRenderer {
             levels.push(self.acquire_blur_scratch(cw, ch, format)?);
         }
 
-        let dev = &self.device.device;
+        // After cascade, levels[0] holds the blurred result. Stash it for caching.
+        *out_blur_image = Some(levels[0].clone());
         let subres = ImageSubresourceLayers::default()
             .aspect_mask(ImageAspectFlags::COLOR)
             .layer_count(1)
@@ -203,7 +229,7 @@ impl VulkanRenderer {
         };
 
         // Step 1: target COLOR_ATTACHMENT -> TRANSFER_SRC.
-        // Step 1: levels[0] UNDEFINED -> TRANSFER_DST.
+        //         levels[0] -> TRANSFER_DST (discard prior contents).
         do_barriers(&[
             barrier(
                 target.image,
@@ -224,6 +250,7 @@ impl VulkanRenderer {
                 AccessFlags2::TRANSFER_WRITE,
             ),
         ]);
+        levels[0].is_undefined.set(false);
 
         // Step 2: blit target rect -> levels[0] full.
         let blit = ImageBlit2::default()
@@ -252,121 +279,9 @@ impl VulkanRenderer {
             dev.cmd_blit_image2(buf, &blit_info);
         }
 
-        // Down passes: levels[i-1] -> levels[i] with linear filter.
+        // Step 3: levels[0] TRANSFER_DST -> SHADER_READ_ONLY for sampling in
-        for i in 1..=passes as usize {
+        // the down pass.
-            let (src, dst) = (&levels[i - 1], &levels[i]);
+        do_barriers(&[barrier(
-            do_barriers(&[
-                barrier(
-                    src.image,
-                    ImageLayout::TRANSFER_DST_OPTIMAL,
-                    ImageLayout::TRANSFER_SRC_OPTIMAL,
-                    PipelineStageFlags2::TRANSFER,
-                    AccessFlags2::TRANSFER_WRITE,
-                    PipelineStageFlags2::TRANSFER,
-                    AccessFlags2::TRANSFER_READ,
-                ),
-                barrier(
-                    dst.image,
-                    ImageLayout::UNDEFINED,
-                    ImageLayout::TRANSFER_DST_OPTIMAL,
-                    PipelineStageFlags2::TOP_OF_PIPE,
-                    AccessFlags2::empty(),
-                    PipelineStageFlags2::TRANSFER,
-                    AccessFlags2::TRANSFER_WRITE,
-                ),
-            ]);
-            let blit = ImageBlit2::default()
-                .src_subresource(subres)
-                .dst_subresource(subres)
-                .src_offsets([
-                    Offset3D { x: 0, y: 0, z: 0 },
-                    Offset3D {
-                        x: src.width as i32,
-                        y: src.height as i32,
-                        z: 1,
-                    },
-                ])
-                .dst_offsets([
-                    Offset3D { x: 0, y: 0, z: 0 },
-                    Offset3D {
-                        x: dst.width as i32,
-                        y: dst.height as i32,
-                        z: 1,
-                    },
-                ]);
-            let blit_info = BlitImageInfo2::default()
-                .src_image(src.image)
-                .src_image_layout(ImageLayout::TRANSFER_SRC_OPTIMAL)
-                .dst_image(dst.image)
-                .dst_image_layout(ImageLayout::TRANSFER_DST_OPTIMAL)
-                .filter(Filter::LINEAR)
-                .regions(slice::from_ref(&blit));
-            unsafe {
-                dev.cmd_blit_image2(buf, &blit_info);
-            }
-        }
-
-        // Up passes: levels[i+1] -> levels[i] with linear filter.
-        for i in (0..passes as usize).rev() {
-            let (src, dst) = (&levels[i + 1], &levels[i]);
-            do_barriers(&[
-                barrier(
-                    src.image,
-                    ImageLayout::TRANSFER_DST_OPTIMAL,
-                    ImageLayout::TRANSFER_SRC_OPTIMAL,
-                    PipelineStageFlags2::TRANSFER,
-                    AccessFlags2::TRANSFER_WRITE,
-                    PipelineStageFlags2::TRANSFER,
-                    AccessFlags2::TRANSFER_READ,
-                ),
-                barrier(
-                    dst.image,
-                    ImageLayout::TRANSFER_SRC_OPTIMAL,
-                    ImageLayout::TRANSFER_DST_OPTIMAL,
-                    PipelineStageFlags2::TRANSFER,
-                    AccessFlags2::TRANSFER_READ,
-                    PipelineStageFlags2::TRANSFER,
-                    AccessFlags2::TRANSFER_WRITE,
-                ),
-            ]);
-            let blit = ImageBlit2::default()
-                .src_subresource(subres)
-                .dst_subresource(subres)
-                .src_offsets([
-                    Offset3D { x: 0, y: 0, z: 0 },
-                    Offset3D {
-                        x: src.width as i32,
-                        y: src.height as i32,
-                        z: 1,
-                    },
-                ])
-                .dst_offsets([
-                    Offset3D { x: 0, y: 0, z: 0 },
-                    Offset3D {
-                        x: dst.width as i32,
-                        y: dst.height as i32,
-                        z: 1,
-                    },
-                ]);
-            let blit_info = BlitImageInfo2::default()
-                .src_image(src.image)
-                .src_image_layout(ImageLayout::TRANSFER_SRC_OPTIMAL)
-                .dst_image(dst.image)
-                .dst_image_layout(ImageLayout::TRANSFER_DST_OPTIMAL)
-                .filter(Filter::LINEAR)
-                .regions(slice::from_ref(&blit));
-            unsafe {
-                dev.cmd_blit_image2(buf, &blit_info);
-            }
-        }
-
-        if let Some(mask) = mask {
-            // Masked composite path:
-            //   levels[0] (TRANSFER_DST) -> SHADER_READ_ONLY_OPTIMAL
-            //   target    (TRANSFER_SRC) -> COLOR_ATTACHMENT_OPTIMAL
-            //   draw composite shader sampling levels[0] + mask, blending onto fb
-            do_barriers(&[
-                barrier(
             levels[0].image,
             ImageLayout::TRANSFER_DST_OPTIMAL,
             ImageLayout::SHADER_READ_ONLY_OPTIMAL,
@@ -374,8 +289,135 @@ impl VulkanRenderer {
             AccessFlags2::TRANSFER_WRITE,
             PipelineStageFlags2::FRAGMENT_SHADER,
             AccessFlags2::SHADER_SAMPLED_READ,
-                ),
+        )]);
-                barrier(
+
+        let blur_down_pipeline = self.get_or_create_blur_down_pipeline(format)?;
+        let blur_up_pipeline = self.get_or_create_blur_up_pipeline(format)?;
+
+        // Helper to run one blur pass: sample `src`, draw into `dst`. Caller
+        // must have transitioned dst to COLOR_ATTACHMENT and src to
+        // SHADER_READ_ONLY before this. Layouts after: dst stays in
+        // COLOR_ATTACHMENT (caller transitions next).
+        let run_pass = |pipeline: &VulkanPipeline,
+                        src: &VulkanImage,
+                        dst: &VulkanImage|
+         -> Result<(), VulkanError> {
+            let color_attachment = RenderingAttachmentInfo::default()
+                .image_view(dst.texture_view)
+                .image_layout(ImageLayout::COLOR_ATTACHMENT_OPTIMAL)
+                .load_op(AttachmentLoadOp::DONT_CARE)
+                .store_op(AttachmentStoreOp::STORE);
+            let render_area = Rect2D {
+                offset: Offset2D { x: 0, y: 0 },
+                extent: Extent2D {
+                    width: dst.width,
+                    height: dst.height,
+                },
+            };
+            let rendering_info = RenderingInfo::default()
+                .render_area(render_area)
+                .layer_count(1)
+                .color_attachments(slice::from_ref(&color_attachment));
+            let viewport = Viewport {
+                x: 0.0,
+                y: 0.0,
+                width: dst.width as f32,
+                height: dst.height as f32,
+                min_depth: 0.0,
+                max_depth: 1.0,
+            };
+            let scissor = render_area;
+            let push = BlurPushConstants {
+                halfpixel: [0.5 / src.width as f32, 0.5 / src.height as f32],
+                offset,
+            };
+            let src_image_info = DescriptorImageInfo::default()
+                .image_view(src.texture_view)
+                .image_layout(ImageLayout::SHADER_READ_ONLY_OPTIMAL);
+            let writes = [WriteDescriptorSet::default()
+                .dst_binding(0)
+                .descriptor_type(DescriptorType::COMBINED_IMAGE_SAMPLER)
+                .image_info(slice::from_ref(&src_image_info))];
+            unsafe {
+                dev.cmd_begin_rendering(buf, &rendering_info);
+                dev.cmd_bind_pipeline(buf, PipelineBindPoint::GRAPHICS, pipeline.pipeline);
+                dev.cmd_set_viewport(buf, 0, slice::from_ref(&viewport));
+                dev.cmd_set_scissor(buf, 0, slice::from_ref(&scissor));
+                self.device.push_descriptor.cmd_push_descriptor_set(
+                    buf,
+                    PipelineBindPoint::GRAPHICS,
+                    pipeline.pipeline_layout,
+                    0,
+                    &writes,
+                );
+                dev.cmd_push_constants(
+                    buf,
+                    pipeline.pipeline_layout,
+                    ShaderStageFlags::FRAGMENT,
+                    0,
+                    uapi::as_bytes(&push),
+                );
+                dev.cmd_draw(buf, 4, 1, 0, 0);
+                dev.cmd_end_rendering(buf);
+            }
+            Ok(())
+        };
+
+        // Down passes: levels[i-1] (SHADER_READ_ONLY) -> levels[i] (COLOR_ATT).
+        // Each iteration transitions the destination to COLOR_ATTACHMENT,
+        // draws, then to SHADER_READ_ONLY for the next iteration's read.
+        for i in 1..=passes as usize {
+            do_barriers(&[barrier(
+                levels[i].image,
+                ImageLayout::UNDEFINED,
+                ImageLayout::COLOR_ATTACHMENT_OPTIMAL,
+                PipelineStageFlags2::TOP_OF_PIPE,
+                AccessFlags2::empty(),
+                PipelineStageFlags2::COLOR_ATTACHMENT_OUTPUT,
+                AccessFlags2::COLOR_ATTACHMENT_WRITE,
+            )]);
+            levels[i].is_undefined.set(false);
+            run_pass(&blur_down_pipeline, &levels[i - 1], &levels[i])?;
+            do_barriers(&[barrier(
+                levels[i].image,
+                ImageLayout::COLOR_ATTACHMENT_OPTIMAL,
+                ImageLayout::SHADER_READ_ONLY_OPTIMAL,
+                PipelineStageFlags2::COLOR_ATTACHMENT_OUTPUT,
+                AccessFlags2::COLOR_ATTACHMENT_WRITE,
+                PipelineStageFlags2::FRAGMENT_SHADER,
+                AccessFlags2::SHADER_SAMPLED_READ,
+            )]);
+        }
+
+        // Up passes: levels[i+1] (SHADER_READ_ONLY) -> levels[i] (COLOR_ATT).
+        for i in (0..passes as usize).rev() {
+            do_barriers(&[barrier(
+                levels[i].image,
+                ImageLayout::SHADER_READ_ONLY_OPTIMAL,
+                ImageLayout::COLOR_ATTACHMENT_OPTIMAL,
+                PipelineStageFlags2::FRAGMENT_SHADER,
+                AccessFlags2::SHADER_SAMPLED_READ,
+                PipelineStageFlags2::COLOR_ATTACHMENT_OUTPUT,
+                AccessFlags2::COLOR_ATTACHMENT_WRITE,
+            )]);
+            run_pass(&blur_up_pipeline, &levels[i + 1], &levels[i])?;
+            do_barriers(&[barrier(
+                levels[i].image,
+                ImageLayout::COLOR_ATTACHMENT_OPTIMAL,
+                ImageLayout::SHADER_READ_ONLY_OPTIMAL,
+                PipelineStageFlags2::COLOR_ATTACHMENT_OUTPUT,
+                AccessFlags2::COLOR_ATTACHMENT_WRITE,
+                PipelineStageFlags2::FRAGMENT_SHADER,
+                AccessFlags2::SHADER_SAMPLED_READ,
+            )]);
+        }
+
+        // After cascade: levels[0] in SHADER_READ_ONLY, target in TRANSFER_SRC.
+
+        if let Some(mask) = mask {
+            // Masked composite path: restore target to COLOR_ATTACHMENT and
+            // draw the composite shader sampling levels[0] + mask.
+            do_barriers(&[barrier(
                 target.image,
                 ImageLayout::TRANSFER_SRC_OPTIMAL,
                 ImageLayout::COLOR_ATTACHMENT_OPTIMAL,
@@ -383,8 +425,7 @@ impl VulkanRenderer {
                 AccessFlags2::TRANSFER_READ,
                 PipelineStageFlags2::COLOR_ATTACHMENT_OUTPUT,
                 AccessFlags2::COLOR_ATTACHMENT_WRITE | AccessFlags2::COLOR_ATTACHMENT_READ,
-                ),
+            )]);
-            ]);
 
             let pipeline = self.get_or_create_blur_composite_pipeline(target.format.vk_format)?;
 
@@ -470,14 +511,15 @@ impl VulkanRenderer {
                 dev.cmd_end_rendering(buf);
             }
         } else {
-            // Final blit: levels[0] -> target rect.
+            // Unmasked: transition levels[0] back to TRANSFER_SRC, target stays
+            // in TRANSFER_SRC, retarget target to TRANSFER_DST, blit-back.
             do_barriers(&[
                 barrier(
                     levels[0].image,
-                    ImageLayout::TRANSFER_DST_OPTIMAL,
+                    ImageLayout::SHADER_READ_ONLY_OPTIMAL,
                     ImageLayout::TRANSFER_SRC_OPTIMAL,
-                    PipelineStageFlags2::TRANSFER,
+                    PipelineStageFlags2::FRAGMENT_SHADER,
-                    AccessFlags2::TRANSFER_WRITE,
+                    AccessFlags2::SHADER_SAMPLED_READ,
                     PipelineStageFlags2::TRANSFER,
                     AccessFlags2::TRANSFER_READ,
                 ),
@@ -516,9 +558,11 @@ impl VulkanRenderer {
             unsafe {
                 dev.cmd_blit_image2(buf, &blit_info);
             }
-
+            // Restore target to COLOR_ATTACHMENT for the resumed render pass.
-            // Restore target to COLOR_ATTACHMENT_OPTIMAL.
+            // Also push levels[0] back to SHADER_READ_ONLY so its tracked layout
-            do_barriers(&[barrier(
+            // matches what the cache-hit fast path expects on next frame.
+            do_barriers(&[
+                barrier(
                     target.image,
                     ImageLayout::TRANSFER_DST_OPTIMAL,
                     ImageLayout::COLOR_ATTACHMENT_OPTIMAL,
@@ -526,11 +570,146 @@ impl VulkanRenderer {
                     AccessFlags2::TRANSFER_WRITE,
                     PipelineStageFlags2::COLOR_ATTACHMENT_OUTPUT,
                     AccessFlags2::COLOR_ATTACHMENT_WRITE | AccessFlags2::COLOR_ATTACHMENT_READ,
-            )]);
+                ),
+                barrier(
+                    levels[0].image,
+                    ImageLayout::TRANSFER_SRC_OPTIMAL,
+                    ImageLayout::SHADER_READ_ONLY_OPTIMAL,
+                    PipelineStageFlags2::TRANSFER,
+                    AccessFlags2::TRANSFER_READ,
+                    PipelineStageFlags2::FRAGMENT_SHADER,
+                    AccessFlags2::SHADER_SAMPLED_READ,
+                ),
+            ]);
         }
 
         // Hold the scratch images until the frame is submitted.
         scratch_out.extend(levels);
         Ok(())
     }
+
+    /// Cache-hit fast path. Cached image is already in SHADER_READ_ONLY_OPTIMAL
+    /// (the layout we leave it in after the previous frame's composite). We just
+    /// re-bind the composite pipeline with the cached image as input and draw.
+    fn record_blur_composite_only(
+        self: &Rc<Self>,
+        buf: CommandBuffer,
+        target: &VulkanImage,
+        cached: &Rc<VulkanImage>,
+        mask: &BlurMaskRecord<'_>,
+        rect: [i32; 4],
+    ) -> Result<(), VulkanError> {
+        let [x1, y1, x2, y2] = rect;
+        let w = (x2 - x1) as u32;
+        let h = (y2 - y1) as u32;
+        let dev = &self.device.device;
+        // The caller (BlurBarrier handler) has already cmd_end_rendering'd the
+        // pass that produced the underlying scene. Without an explicit barrier
+        // here, the new render pass's LOAD reads can race with those prior
+        // COLOR_ATTACHMENT_WRITEs and pull stale memory — which manifests as
+        // per-frame flicker in the blurred region.
+        let target_load_barrier = ImageMemoryBarrier2::default()
+            .image(target.image)
+            .old_layout(ImageLayout::COLOR_ATTACHMENT_OPTIMAL)
+            .new_layout(ImageLayout::COLOR_ATTACHMENT_OPTIMAL)
+            .subresource_range(ImageSubresourceRange {
+                aspect_mask: ImageAspectFlags::COLOR,
+                base_mip_level: 0,
+                level_count: 1,
+                base_array_layer: 0,
+                layer_count: 1,
+            })
+            .src_stage_mask(PipelineStageFlags2::COLOR_ATTACHMENT_OUTPUT)
+            .src_access_mask(AccessFlags2::COLOR_ATTACHMENT_WRITE)
+            .dst_stage_mask(PipelineStageFlags2::COLOR_ATTACHMENT_OUTPUT)
+            .dst_access_mask(
+                AccessFlags2::COLOR_ATTACHMENT_READ | AccessFlags2::COLOR_ATTACHMENT_WRITE,
+            )
+            .src_queue_family_index(ash::vk::QUEUE_FAMILY_IGNORED)
+            .dst_queue_family_index(ash::vk::QUEUE_FAMILY_IGNORED);
+        let dep = DependencyInfoKHR::default()
+            .image_memory_barriers(slice::from_ref(&target_load_barrier));
+        unsafe {
+            dev.cmd_pipeline_barrier2(buf, &dep);
+        }
+        let pipeline = self.get_or_create_blur_composite_pipeline(target.format.vk_format)?;
+        let target_render_view = target.render_view.unwrap_or(target.texture_view);
+        let color_attachment = RenderingAttachmentInfo::default()
+            .image_view(target_render_view)
+            .image_layout(ImageLayout::COLOR_ATTACHMENT_OPTIMAL)
+            .load_op(AttachmentLoadOp::LOAD)
+            .store_op(AttachmentStoreOp::STORE);
+        let render_area = Rect2D {
+            offset: Offset2D { x: 0, y: 0 },
+            extent: Extent2D {
+                width: target.width,
+                height: target.height,
+            },
+        };
+        let rendering_info = RenderingInfo::default()
+            .render_area(render_area)
+            .layer_count(1)
+            .color_attachments(slice::from_ref(&color_attachment));
+        let viewport = Viewport {
+            x: 0.0,
+            y: 0.0,
+            width: target.width as f32,
+            height: target.height as f32,
+            min_depth: 0.0,
+            max_depth: 1.0,
+        };
+        let scissor = Rect2D {
+            offset: Offset2D { x: x1, y: y1 },
+            extent: Extent2D {
+                width: w,
+                height: h,
+            },
+        };
+        let blurred_tc: [[f32; 2]; 4] = [[1.0, 0.0], [0.0, 0.0], [1.0, 1.0], [0.0, 1.0]];
+        let push = BlurCompositePushConstants {
+            pos: mask.target_points,
+            blurred_tex_pos: blurred_tc,
+            mask_tex_pos: mask.mask_source_points,
+            threshold: mask.threshold,
+        };
+        let blurred_image_info = DescriptorImageInfo::default()
+            .image_view(cached.texture_view)
+            .image_layout(ImageLayout::SHADER_READ_ONLY_OPTIMAL);
+        let mask_image_info = DescriptorImageInfo::default()
+            .image_view(mask.mask_view)
+            .image_layout(ImageLayout::SHADER_READ_ONLY_OPTIMAL);
+        let writes = [
+            WriteDescriptorSet::default()
+                .dst_binding(0)
+                .descriptor_type(DescriptorType::COMBINED_IMAGE_SAMPLER)
+                .image_info(slice::from_ref(&blurred_image_info)),
+            WriteDescriptorSet::default()
+                .dst_binding(1)
+                .descriptor_type(DescriptorType::COMBINED_IMAGE_SAMPLER)
+                .image_info(slice::from_ref(&mask_image_info)),
+        ];
+        unsafe {
+            dev.cmd_begin_rendering(buf, &rendering_info);
+            dev.cmd_bind_pipeline(buf, PipelineBindPoint::GRAPHICS, pipeline.pipeline);
+            dev.cmd_set_viewport(buf, 0, slice::from_ref(&viewport));
+            dev.cmd_set_scissor(buf, 0, slice::from_ref(&scissor));
+            self.device.push_descriptor.cmd_push_descriptor_set(
+                buf,
+                PipelineBindPoint::GRAPHICS,
+                pipeline.pipeline_layout,
+                0,
+                &writes,
+            );
+            dev.cmd_push_constants(
+                buf,
+                pipeline.pipeline_layout,
+                ShaderStageFlags::VERTEX | ShaderStageFlags::FRAGMENT,
+                0,
+                uapi::as_bytes(&push),
+            );
+            dev.cmd_draw(buf, 4, 1, 0, 0);
+            dev.cmd_end_rendering(buf);
+        }
+        Ok(())
+    }
 }

src/gfx_apis/vulkan/descriptor.rs

@@ -87,6 +87,31 @@ impl VulkanDevice {
         }))
     }
 
+    pub(super) fn create_blur_descriptor_set_layout(
+        self: &Rc<Self>,
+        sampler: &Rc<VulkanSampler>,
+    ) -> Result<Rc<VulkanDescriptorSetLayout>, VulkanError> {
+        let immutable_sampler = [sampler.sampler];
+        let binding = DescriptorSetLayoutBinding::default()
+            .binding(0)
+            .stage_flags(ShaderStageFlags::FRAGMENT)
+            .descriptor_count(1)
+            .descriptor_type(DescriptorType::COMBINED_IMAGE_SAMPLER)
+            .immutable_samplers(&immutable_sampler);
+        let create_info = DescriptorSetLayoutCreateInfo::default()
+            .bindings(slice::from_ref(&binding))
+            .flags(DescriptorSetLayoutCreateFlags::PUSH_DESCRIPTOR_KHR);
+        let layout = unsafe { self.device.create_descriptor_set_layout(&create_info, None) };
+        let layout = layout.map_err(VulkanError::CreateDescriptorSetLayout)?;
+        Ok(Rc::new(VulkanDescriptorSetLayout {
+            device: self.clone(),
+            layout,
+            size: 0,
+            offsets: Default::default(),
+            _sampler: Some(sampler.clone()),
+        }))
+    }
+
     pub(super) fn create_tex_sampler_descriptor_set_layout(
         self: &Rc<Self>,
         sampler: &Rc<VulkanSampler>,

src/gfx_apis/vulkan/renderer.rs

@@ -26,7 +26,8 @@ use {
             sampler::VulkanSampler,
             semaphore::VulkanSemaphore,
             shaders::{
-                BLUR_COMPOSITE_FRAG, BLUR_COMPOSITE_VERT, BlurCompositePushConstants,
+                BLUR_COMPOSITE_FRAG, BLUR_COMPOSITE_VERT, BLUR_DOWN_FRAG, BLUR_UP_FRAG, BLUR_VERT,
+                BlurCompositePushConstants, BlurPushConstants,
                 ColorManagementData, EotfArgs, FILL_FRAG, FILL_VERT, FillPushConstants,
                 InvEotfArgs, LEGACY_FILL_FRAG, LEGACY_FILL_VERT, LEGACY_ROUNDED_FILL_FRAG,
                 LEGACY_ROUNDED_FILL_VERT, LEGACY_ROUNDED_TEX_FRAG, LEGACY_ROUNDED_TEX_VERT,
@@ -123,6 +124,12 @@ pub struct VulkanRenderer {
     pub(super) blur_composite_frag_shader: Rc<VulkanShader>,
     pub(super) blur_composite_descriptor_set_layout: Rc<VulkanDescriptorSetLayout>,
     pub(super) blur_composite_pipelines: CopyHashMap<vk::Format, Rc<VulkanPipeline>>,
+    pub(super) blur_vert_shader: Rc<VulkanShader>,
+    pub(super) blur_down_frag_shader: Rc<VulkanShader>,
+    pub(super) blur_up_frag_shader: Rc<VulkanShader>,
+    pub(super) blur_descriptor_set_layout: Rc<VulkanDescriptorSetLayout>,
+    pub(super) blur_down_pipelines: CopyHashMap<vk::Format, Rc<VulkanPipeline>>,
+    pub(super) blur_up_pipelines: CopyHashMap<vk::Format, Rc<VulkanPipeline>>,
     pub(super) defunct: Cell<bool>,
     pub(super) pending_cpu_jobs: CopyHashMap<u64, PendingJob>,
     pub(super) shm_allocator: Rc<VulkanThreadedAllocator>,
@@ -229,13 +236,20 @@ enum VulkanOp {
 struct VulkanBlurOp {
     rect: crate::gfx_api::FramebufferRect,
     passes: u8,
+    offset: f32,
     mask: Option<VulkanBlurMask>,
+    cache: Option<Rc<RefCell<Option<crate::gfx_api::BlurCacheEntry>>>>,
+    cache_epoch: u64,
+    cache_pixel_rect: [i32; 4],
 }
 
 struct VulkanBlurMask {
     tex: Rc<VulkanImage>,
     source: crate::gfx_api::SampleRect,
     threshold: f32,
+    buffer_resv: Option<Rc<dyn BufferResv>>,
+    acquire_sync: Option<AcquireSync>,
+    release_sync: ReleaseSync,
 }
 
 struct VulkanTexOp {
@@ -246,6 +260,7 @@ struct VulkanTexOp {
     acquire_sync: Option<AcquireSync>,
     release_sync: ReleaseSync,
     alpha: f32,
+    discard_alpha: f32,
     source_type: TexSourceType,
     copy_type: TexCopyType,
     alpha_mode: AlphaMode,
@@ -286,6 +301,7 @@ struct VulkanRoundedTexOp {
     acquire_sync: Option<AcquireSync>,
     release_sync: ReleaseSync,
     alpha: f32,
+    discard_alpha: f32,
     source_type: TexSourceType,
     copy_type: TexCopyType,
     alpha_mode: AlphaMode,
@@ -406,6 +422,10 @@ impl VulkanDevice {
         let blur_composite_frag_shader = self.create_shader(BLUR_COMPOSITE_FRAG)?;
         let blur_composite_descriptor_set_layout =
             self.create_blur_composite_descriptor_set_layout(&sampler)?;
+        let blur_vert_shader = self.create_shader(BLUR_VERT)?;
+        let blur_down_frag_shader = self.create_shader(BLUR_DOWN_FRAG)?;
+        let blur_up_frag_shader = self.create_shader(BLUR_UP_FRAG)?;
+        let blur_descriptor_set_layout = self.create_blur_descriptor_set_layout(&sampler)?;
         let gfx_command_buffers = self.create_command_pool(self.graphics_queue_idx)?;
         let transfer_command_buffers = self
             .distinct_transfer_queue_family_idx
@@ -497,6 +517,12 @@ impl VulkanDevice {
             blur_composite_frag_shader,
             blur_composite_descriptor_set_layout,
             blur_composite_pipelines: Default::default(),
+            blur_vert_shader,
+            blur_down_frag_shader,
+            blur_up_frag_shader,
+            blur_descriptor_set_layout,
+            blur_down_pipelines: Default::default(),
+            blur_up_pipelines: Default::default(),
             defunct: Cell::new(false),
             pending_cpu_jobs: Default::default(),
             shm_allocator,
@@ -886,6 +912,128 @@ impl VulkanRenderer {
         }))
     }
 
+    pub(super) fn get_or_create_blur_down_pipeline(
+        &self,
+        format: vk::Format,
+    ) -> Result<Rc<VulkanPipeline>, VulkanError> {
+        if let Some(pl) = self.blur_down_pipelines.get(&format) {
+            return Ok(pl);
+        }
+        let pl = self.create_blur_pass_pipeline(format, &self.blur_down_frag_shader)?;
+        self.blur_down_pipelines.set(format, pl.clone());
+        Ok(pl)
+    }
+
+    pub(super) fn get_or_create_blur_up_pipeline(
+        &self,
+        format: vk::Format,
+    ) -> Result<Rc<VulkanPipeline>, VulkanError> {
+        if let Some(pl) = self.blur_up_pipelines.get(&format) {
+            return Ok(pl);
+        }
+        let pl = self.create_blur_pass_pipeline(format, &self.blur_up_frag_shader)?;
+        self.blur_up_pipelines.set(format, pl.clone());
+        Ok(pl)
+    }
+
+    fn create_blur_pass_pipeline(
+        &self,
+        format: vk::Format,
+        frag: &Rc<VulkanShader>,
+    ) -> Result<Rc<VulkanPipeline>, VulkanError> {
+        use ash::vk::{
+            ColorComponentFlags, CullModeFlags, DynamicState, FrontFace,
+            GraphicsPipelineCreateInfo, PipelineCache, PipelineColorBlendAttachmentState,
+            PipelineColorBlendStateCreateInfo, PipelineDynamicStateCreateInfo,
+            PipelineInputAssemblyStateCreateInfo, PipelineLayoutCreateInfo,
+            PipelineMultisampleStateCreateInfo, PipelineRasterizationStateCreateInfo,
+            PipelineRenderingCreateInfo, PipelineShaderStageCreateInfo,
+            PipelineVertexInputStateCreateInfo, PipelineViewportStateCreateInfo, PolygonMode,
+            PrimitiveTopology, PushConstantRange, SampleCountFlags,
+        };
+        let dev = &self.device.device;
+        let push_range = PushConstantRange::default()
+            .stage_flags(ShaderStageFlags::FRAGMENT)
+            .offset(0)
+            .size(size_of::<BlurPushConstants>() as u32);
+        let set_layouts = [self.blur_descriptor_set_layout.layout];
+        let layout_info = PipelineLayoutCreateInfo::default()
+            .push_constant_ranges(slice::from_ref(&push_range))
+            .set_layouts(&set_layouts);
+        let pipeline_layout = unsafe { dev.create_pipeline_layout(&layout_info, None) };
+        let pipeline_layout = pipeline_layout.map_err(VulkanError::CreatePipelineLayout)?;
+        let destroy_layout =
+            run_on_drop::on_drop(|| unsafe { dev.destroy_pipeline_layout(pipeline_layout, None) });
+        let stages = [
+            PipelineShaderStageCreateInfo::default()
+                .stage(ShaderStageFlags::VERTEX)
+                .module(self.blur_vert_shader.module)
+                .name(c"main"),
+            PipelineShaderStageCreateInfo::default()
+                .stage(ShaderStageFlags::FRAGMENT)
+                .module(frag.module)
+                .name(c"main"),
+        ];
+        let input_assembly_state = PipelineInputAssemblyStateCreateInfo::default()
+            .topology(PrimitiveTopology::TRIANGLE_STRIP);
+        let vertex_input_state = PipelineVertexInputStateCreateInfo::default();
+        let rasterization_state = PipelineRasterizationStateCreateInfo::default()
+            .polygon_mode(PolygonMode::FILL)
+            .cull_mode(CullModeFlags::NONE)
+            .line_width(1.0)
+            .front_face(FrontFace::COUNTER_CLOCKWISE);
+        let multisampling_state = PipelineMultisampleStateCreateInfo::default()
+            .sample_shading_enable(false)
+            .rasterization_samples(SampleCountFlags::TYPE_1);
+        let blending = PipelineColorBlendAttachmentState::default()
+            .color_write_mask(ColorComponentFlags::RGBA)
+            .blend_enable(false);
+        let color_blend_state =
+            PipelineColorBlendStateCreateInfo::default().attachments(slice::from_ref(&blending));
+        let dynamic_states = [DynamicState::VIEWPORT, DynamicState::SCISSOR];
+        let dynamic_state =
+            PipelineDynamicStateCreateInfo::default().dynamic_states(&dynamic_states);
+        let viewport_state = PipelineViewportStateCreateInfo::default()
+            .viewport_count(1)
+            .scissor_count(1);
+        let mut pipeline_rendering_create_info = PipelineRenderingCreateInfo::default()
+            .color_attachment_formats(slice::from_ref(&format));
+        let create_info = GraphicsPipelineCreateInfo::default()
+            .push_next(&mut pipeline_rendering_create_info)
+            .stages(&stages)
+            .input_assembly_state(&input_assembly_state)
+            .vertex_input_state(&vertex_input_state)
+            .rasterization_state(&rasterization_state)
+            .multisample_state(&multisampling_state)
+            .color_blend_state(&color_blend_state)
+            .dynamic_state(&dynamic_state)
+            .viewport_state(&viewport_state)
+            .layout(pipeline_layout);
+        let pipelines = unsafe {
+            dev.create_graphics_pipelines(
+                PipelineCache::null(),
+                slice::from_ref(&create_info),
+                None,
+            )
+        };
+        let mut pipelines = pipelines
+            .map_err(|e| e.1)
+            .map_err(VulkanError::CreatePipeline)?;
+        let pipeline = pipelines.pop().unwrap();
+        destroy_layout.forget();
+        Ok(Rc::new(VulkanPipeline {
+            vert: self.blur_vert_shader.clone(),
+            _frag: frag.clone(),
+            pipeline_layout,
+            pipeline,
+            _descriptor_set_layouts: {
+                let mut v = ArrayVec::new();
+                v.push(self.blur_descriptor_set_layout.clone());
+                v
+            },
+        }))
+    }
+
     pub(super) fn allocate_point(&self) -> u64 {
         self.last_point.fetch_add(1) + 1
     }
@@ -1263,6 +1411,7 @@ impl VulkanRenderer {
                             acquire_sync: Some(ct.acquire_sync.clone()),
                             release_sync: ct.release_sync,
                             alpha: ct.alpha.unwrap_or_default(),
+                            discard_alpha: ct.discard_alpha.unwrap_or(-1.0),
                             source_type,
                             copy_type,
                             alpha_mode: ct.alpha_mode,
@@ -1361,6 +1510,7 @@ impl VulkanRenderer {
                             acquire_sync: Some(ct.acquire_sync.clone()),
                             release_sync: ct.release_sync,
                             alpha: ct.alpha.unwrap_or_default(),
+                            discard_alpha: ct.discard_alpha.unwrap_or(-1.0),
                             source_type,
                             copy_type,
                             alpha_mode: ct.alpha_mode,
@@ -1376,9 +1526,6 @@ impl VulkanRenderer {
                     }
                 }
                 GfxApiOpt::BlurBackdrop(b) => {
-                    // Flush all pending ops in original order, then push a
-                    // barrier op to FrameBuffer pass that will end + restart
-                    // the render pass to do the blur work in between.
                     sync(memory);
                     let mask = if let Some(m) = &b.mask {
                         let tex = m.texture.clone().into_vk(&self.device.device)?;
@@ -1393,15 +1540,35 @@ impl VulkanRenderer {
                                 tex,
                                 source: m.source,
                                 threshold: m.threshold,
+                                buffer_resv: m.buffer_resv.clone(),
+                                acquire_sync: Some(m.acquire_sync.clone()),
+                                release_sync: m.release_sync,
                             })
                         }
                     } else {
                         None
                     };
-                    memory.ops[RenderPass::FrameBuffer].push(VulkanOp::BlurBarrier(VulkanBlurOp {
+                    // Route to whichever pass actually contains the scene.
+                    // The BlendBuffer holds the linearly-composed scene
+                    // (background + workspace + translucent layers) and is
+                    // copied into the FrameBuffer only at the end of the FB
+                    // pass. Reading FB before that copy would sample an
+                    // empty target and produce a black blur. If BB has been
+                    // elided (no blended content this frame), fall back to
+                    // FB which then carries the full scene itself.
+                    let target_pass = if !memory.paint_regions[RenderPass::BlendBuffer].is_empty() {
+                        RenderPass::BlendBuffer
+                    } else {
+                        RenderPass::FrameBuffer
+                    };
+                    memory.ops[target_pass].push(VulkanOp::BlurBarrier(VulkanBlurOp {
                         rect: b.rect,
                         passes: b.passes,
+                        offset: b.offset,
                         mask,
+                        cache: b.cache.clone(),
+                        cache_epoch: b.cache_epoch,
+                        cache_pixel_rect: b.cache_pixel_rect,
                     }));
                 }
             }
@@ -1586,7 +1753,7 @@ impl VulkanRenderer {
                         release_sync,
                     });
                 } else if let VulkanOp::BlurBarrier(b) = cmd
-                    && let Some(m) = &b.mask
+                    && let Some(m) = &mut b.mask
                 {
                     let tex = &m.tex;
                     if tex.execution_version.replace(execution) != execution {
@@ -1598,6 +1765,12 @@ impl VulkanRenderer {
                         if let VulkanImageMemory::DmaBuf(_) = &tex.ty {
                             memory.dmabuf_sample.push(tex.clone())
                         }
+                        memory.textures.push(UsedTexture {
+                            tex: tex.clone(),
+                            resv: m.buffer_resv.take(),
+                            acquire_sync: m.acquire_sync.take().unwrap(),
+                            release_sync: m.release_sync,
+                        });
                     }
                 }
             }
@@ -1929,6 +2102,7 @@ impl VulkanRenderer {
                         let push = TexPushConstants {
                             vertices: c.range_address,
                             alpha: c.alpha,
+                            discard_threshold: c.discard_alpha,
                         };
                         unsafe {
                             db.cmd_set_descriptor_buffer_offsets(
@@ -1966,6 +2140,7 @@ impl VulkanRenderer {
                                 pos,
                                 tex_pos,
                                 alpha: c.alpha,
+                                discard_threshold: c.discard_alpha,
                             };
                             unsafe {
                                 dev.cmd_push_constants(
@@ -2045,6 +2220,7 @@ impl VulkanRenderer {
                         let push = RoundedTexPushConstants {
                             vertices: c.range_address,
                             alpha: c.alpha,
+                            discard_threshold: c.discard_alpha,
                             size_x: c.size[0],
                             size_y: c.size[1],
                             corner_radius_tl: c.corner_radius[0],
@@ -2088,6 +2264,7 @@ impl VulkanRenderer {
                             pos: c.target,
                             tex_pos: c.source,
                             alpha: c.alpha,
+                            discard_threshold: c.discard_alpha,
                             size_x: c.size[0],
                             size_y: c.size[1],
                             corner_radius_tl: c.corner_radius[0],
@@ -2109,10 +2286,8 @@ impl VulkanRenderer {
                     }
                 }
                 VulkanOp::BlurBarrier(blur) => {
-                    // Blur is only meaningful in the FrameBuffer pass.
+                    // BlurBarrier is pushed to exactly one pass in convert_ops
-                    if pass != RenderPass::FrameBuffer {
+                    // (BB if present, else FB), so no per-pass gating is needed.
-                        continue;
-                    }
                     // End the current dynamic render pass, run the blur work
                     // (image-blit cascade between scratch images), and resume
                     // the render pass with LOAD so subsequent draws layer on
@@ -2132,14 +2307,55 @@ impl VulkanRenderer {
                                 threshold: m.threshold,
                                 _phantom: std::marker::PhantomData,
                             });
+
+                    // Cache lookup: a hit lets us skip the entire blur cascade.
+                    // Only masked blurs are cached. The masked path leaves the
+                    // blurred scratch image in SHADER_READ_ONLY_OPTIMAL, which
+                    // is the layout required by the cache-hit composite path.
+                    let cached_blur: Option<Rc<VulkanImage>> = mask_record
+                        .as_ref()
+                        .and_then(|_| blur.cache.as_ref())
+                        .and_then(|slot| {
+                            let slot_borrow = slot.borrow();
+                            slot_borrow.as_ref().and_then(|entry| {
+                                if entry.epoch == blur.cache_epoch
+                                    && entry.passes == blur.passes
+                                    && entry.offset == blur.offset
+                                    && entry.pixel_rect == blur.cache_pixel_rect
+                                {
+                                    entry.image.clone().into_vk(&self.device.device).ok()
+                                } else {
+                                    None
+                                }
+                            })
+                        });
+
+                    let mut produced_blur: Option<Rc<VulkanImage>> = None;
                     self.record_blur(
                         buf,
                         target,
                         rect_arr,
                         blur.passes,
+                        blur.offset,
                         &mut local_blur_scratch,
                         mask_record.as_ref(),
+                        cached_blur.as_ref(),
+                        &mut produced_blur,
                     )?;
+
+                    // On a masked cache miss, store the freshly-blurred image
+                    // for the next frame to reuse.
+                    if let (Some(_), Some(slot), Some(image)) =
+                        (mask_record.as_ref(), blur.cache.as_ref(), produced_blur)
+                    {
+                        *slot.borrow_mut() = Some(crate::gfx_api::BlurCacheEntry {
+                            pixel_rect: blur.cache_pixel_rect,
+                            passes: blur.passes,
+                            offset: blur.offset,
+                            epoch: blur.cache_epoch,
+                            image,
+                        });
+                    }
                     self.begin_rendering_load(buf, target);
                     // Pipeline state is invalidated across the render-pass
                     // break — force re-bind on next draw.
@@ -2679,6 +2895,7 @@ impl VulkanRenderer {
         let width = fb.width as f32;
         let height = fb.height as f32;
         let mut tag = 0;
+        let mut blur_rects: Vec<Rect> = Vec::new();
         for opt in opts.iter().rev() {
             let (opaque, fb_rect) = match opt {
                 GfxApiOpt::Sync => continue,
@@ -2706,7 +2923,10 @@ impl VulkanRenderer {
                     (false, rf.rect)
                 }
                 GfxApiOpt::RoundedCopyTexture(ct) => (false, ct.target),
-                GfxApiOpt::BlurBackdrop(_) => continue,
+                GfxApiOpt::BlurBackdrop(b) => {
+                    blur_rects.push(b.rect.to_rect(width, height));
+                    continue;
+                }
             };
             if opaque || bb.is_none() {
                 tag |= 1;
@@ -2719,6 +2939,20 @@ impl VulkanRenderer {
             }
             memory.regions_2.push(rect.with_tag(tag));
         }
+        // Force blur source rects into the effective damage region. The blur
+        // cascade reads its source from BB (or FB), and both buffers persist
+        // their contents across frames in undamaged regions. Without this,
+        // a cache-miss cascade can sample a stale composite — including the
+        // blur surface's own previously-drawn body — and re-blur it,
+        // producing visible double-shadow artifacts at the blur boundary.
+        let blur_region_owned = Region::from_rects2(&blur_rects);
+        let expanded_region;
+        let region: &Region = if blur_region_owned.is_empty() {
+            region
+        } else {
+            expanded_region = region.union_cow(&blur_region_owned).into_owned();
+            &expanded_region
+        };
         let clear_region = if clear.is_some() {
             let opaque_region = Region::from_rects2(&memory.regions_1);
             region.subtract_cow(&opaque_region)

src/gfx_apis/vulkan/shaders.rs

@@ -21,6 +21,9 @@ pub const ROUNDED_TEX_VERT: &[u8] = include_bytes!("shaders_bin/rounded_tex.vert
 pub const ROUNDED_TEX_FRAG: &[u8] = include_bytes!("shaders_bin/rounded_tex.frag.spv");
 pub const BLUR_COMPOSITE_VERT: &[u8] = include_bytes!("shaders_bin/blur_composite.vert.spv");
 pub const BLUR_COMPOSITE_FRAG: &[u8] = include_bytes!("shaders_bin/blur_composite.frag.spv");
+pub const BLUR_VERT: &[u8] = include_bytes!("shaders_bin/blur.vert.spv");
+pub const BLUR_DOWN_FRAG: &[u8] = include_bytes!("shaders_bin/blur_down.frag.spv");
+pub const BLUR_UP_FRAG: &[u8] = include_bytes!("shaders_bin/blur_up.frag.spv");
 pub const LEGACY_ROUNDED_FILL_VERT: &[u8] =
     include_bytes!("shaders_bin/legacy_rounded_fill.vert.spv");
 pub const LEGACY_ROUNDED_FILL_FRAG: &[u8] =
@@ -69,6 +72,7 @@ unsafe impl Packed for TexVertex {}
 pub struct TexPushConstants {
     pub vertices: DeviceAddress,
     pub alpha: f32,
+    pub discard_threshold: f32,
 }
 
 unsafe impl Packed for TexPushConstants {}
@@ -109,6 +113,7 @@ pub struct LegacyTexPushConstants {
     pub pos: [[f32; 2]; 4],
     pub tex_pos: [[f32; 2]; 4],
     pub alpha: f32,
+    pub discard_threshold: f32,
 }
 
 unsafe impl Packed for LegacyTexPushConstants {}
@@ -148,6 +153,7 @@ unsafe impl Packed for LegacyRoundedFillPushConstants {}
 pub struct RoundedTexPushConstants {
     pub vertices: DeviceAddress,
     pub alpha: f32,
+    pub discard_threshold: f32,
     pub size_x: f32,
     pub size_y: f32,
     pub corner_radius_tl: f32,
@@ -165,6 +171,7 @@ pub struct LegacyRoundedTexPushConstants {
     pub pos: [[f32; 2]; 4],
     pub tex_pos: [[f32; 2]; 4],
     pub alpha: f32,
+    pub discard_threshold: f32,
     pub size_x: f32,
     pub size_y: f32,
     pub corner_radius_tl: f32,
@@ -195,6 +202,15 @@ pub struct BlurCompositePushConstants {
 
 unsafe impl Packed for BlurCompositePushConstants {}
 
+#[derive(Copy, Clone, Debug)]
+#[repr(C)]
+pub struct BlurPushConstants {
+    pub halfpixel: [f32; 2],
+    pub offset: f32,
+}
+
+unsafe impl Packed for BlurPushConstants {}
+
 impl VulkanDevice {
     pub(super) fn create_shader(
         self: &Rc<Self>,

src/gfx_apis/vulkan/shaders/blur.vert

@@ -0,0 +1,26 @@
+#version 450
+
+layout(location = 0) out vec2 v_texcoord;
+
+void main() {
+	vec2 pos;
+	switch (gl_VertexIndex) {
+		case 0:
+			pos = vec2( 1.0, -1.0);
+			v_texcoord = vec2(1.0, 0.0);
+			break;
+		case 1:
+			pos = vec2(-1.0, -1.0);
+			v_texcoord = vec2(0.0, 0.0);
+			break;
+		case 2:
+			pos = vec2( 1.0,  1.0);
+			v_texcoord = vec2(1.0, 1.0);
+			break;
+		case 3:
+			pos = vec2(-1.0,  1.0);
+			v_texcoord = vec2(0.0, 1.0);
+			break;
+	}
+	gl_Position = vec4(pos, 0.0, 1.0);
+}

src/gfx_apis/vulkan/shaders/blur_down.frag

@@ -0,0 +1,21 @@
+#version 450
+
+layout(set = 0, binding = 0) uniform sampler2D tex;
+
+layout(push_constant, std430) uniform Data {
+	vec2 halfpixel;
+	float offset;
+} data;
+
+layout(location = 0) in vec2 v_texcoord;
+layout(location = 0) out vec4 out_color;
+
+void main() {
+	vec2 hp = data.halfpixel * data.offset;
+	vec4 sum = textureLod(tex, v_texcoord, 0.0) * 4.0;
+	sum += textureLod(tex, v_texcoord - hp, 0.0);
+	sum += textureLod(tex, v_texcoord + hp, 0.0);
+	sum += textureLod(tex, v_texcoord + vec2(hp.x, -hp.y), 0.0);
+	sum += textureLod(tex, v_texcoord - vec2(hp.x, -hp.y), 0.0);
+	out_color = sum / 8.0;
+}

src/gfx_apis/vulkan/shaders/blur_up.frag

@@ -0,0 +1,24 @@
+#version 450
+
+layout(set = 0, binding = 0) uniform sampler2D tex;
+
+layout(push_constant, std430) uniform Data {
+	vec2 halfpixel;
+	float offset;
+} data;
+
+layout(location = 0) in vec2 v_texcoord;
+layout(location = 0) out vec4 out_color;
+
+void main() {
+	vec2 hp = data.halfpixel * data.offset;
+	vec4 sum = textureLod(tex, v_texcoord + vec2(-hp.x * 2.0, 0.0), 0.0);
+	sum += textureLod(tex, v_texcoord + vec2(-hp.x, hp.y), 0.0) * 2.0;
+	sum += textureLod(tex, v_texcoord + vec2(0.0, hp.y * 2.0), 0.0);
+	sum += textureLod(tex, v_texcoord + vec2(hp.x, hp.y), 0.0) * 2.0;
+	sum += textureLod(tex, v_texcoord + vec2(hp.x * 2.0, 0.0), 0.0);
+	sum += textureLod(tex, v_texcoord + vec2(hp.x, -hp.y), 0.0) * 2.0;
+	sum += textureLod(tex, v_texcoord + vec2(0.0, -hp.y * 2.0), 0.0);
+	sum += textureLod(tex, v_texcoord + vec2(-hp.x, -hp.y), 0.0) * 2.0;
+	out_color = sum / 12.0;
+}

src/gfx_apis/vulkan/shaders_hash.txt

@@ -1,6 +1,9 @@
 302a9f250bdc4f8e0e71a9f77c9a8a7aa55fd003bc91c2422a700c4abd83f54e src/gfx_apis/vulkan/shaders/alpha_modes.glsl
+65acbe7a6496279fa22f520ad2036d3e14a7cb1707c6a509ce7858adc4a2dcba src/gfx_apis/vulkan/shaders/blur.vert
 16ad6f1eb029ccce5e0204a7d79709b05a8a708133feaf8bb20a24371de25ed7 src/gfx_apis/vulkan/shaders/blur_composite.frag
 6399e23afa2e07c98b9fd1a4e853ea974a9958547ce65734846483bd7cbc8461 src/gfx_apis/vulkan/shaders/blur_composite.vert
+a04b2453c39efb018754fc25d45a369b5813359c55fad1c99020804cbb3a18e0 src/gfx_apis/vulkan/shaders/blur_down.frag
+f6d51f3b5410387d1474529c44e71bfdc31ceb80174ea6e3e4c2df30d03f11c3 src/gfx_apis/vulkan/shaders/blur_up.frag
 b6a0df1e231fab533499329636b7a580384784418baee06c147af5fcc384cf5c src/gfx_apis/vulkan/shaders/eotfs.glsl
 8a38df18851cd13884499820f26939fb7319f45d913d867f254d8118d59fb117 src/gfx_apis/vulkan/shaders/fill.common.glsl
 21c488d12aa5ad2f109ec44cb856dfe837e02ea9025b5ed64439d742c17cbf30 src/gfx_apis/vulkan/shaders/fill.frag