Blend translucent Wayland surfaces

2026-10-14 19:27:09 -04:00 · 2026-09-27 10:02:23 -04:00
12 changed files with 578 additions and 45 deletions
@@ -18,6 +18,7 @@ add_library(bannerwayland SHARED
        src/wl_toplevel_icon.c
        src/wl_color_mgmt.c
        src/hdr_compose.c
+        src/blend_pass.c
        generated/color-management-v1-protocol.c
        generated/xdg-shell-protocol.c
        generated/linux-dmabuf-v1-protocol.c
@@ -0,0 +1,9 @@
+#version 450
+// Wayland buffers carry premultiplied alpha; the pipeline blends ONE, ONE_MINUS_SRC_ALPHA.
+layout(binding = 0) uniform sampler2D Surface;
+layout(location = 0) in vec2 uv;
+layout(location = 0) out vec4 outColor;
+
+void main() {
+    outColor = texture(Surface, uv);
+}
@@ -0,0 +1,10 @@
+#version 450
+// One surface as a quad: dst = the target rectangle in NDC, src = the buffer rectangle in UV.
+layout(push_constant) uniform PC { vec4 dst; vec4 src; } pc;
+layout(location = 0) out vec2 uv;
+
+void main() {
+    vec2 t = vec2(gl_VertexIndex & 1, gl_VertexIndex >> 1);
+    gl_Position = vec4(mix(pc.dst.xy, pc.dst.zw, t), 0.0, 1.0);
+    uv = mix(pc.src.xy, pc.src.zw, t);
+}
@@ -0,0 +1,22 @@
+	// 1115.1.0
+	 #pragma once
+const uint32_t blend_frag_code[] = {
+	0x07230203,0x00010000,0x0008000b,0x00000014,0x00000000,0x00020011,0x00000001,0x0006000b,
+	0x00000001,0x4c534c47,0x6474732e,0x3035342e,0x00000000,0x0003000e,0x00000000,0x00000001,
+	0x0007000f,0x00000004,0x00000004,0x6e69616d,0x00000000,0x00000009,0x00000011,0x00030010,
+	0x00000004,0x00000007,0x00030003,0x00000002,0x000001c2,0x00040005,0x00000004,0x6e69616d,
+	0x00000000,0x00050005,0x00000009,0x4374756f,0x726f6c6f,0x00000000,0x00040005,0x0000000d,
+	0x66727553,0x00656361,0x00030005,0x00000011,0x00007675,0x00040047,0x00000009,0x0000001e,
+	0x00000000,0x00040047,0x0000000d,0x00000021,0x00000000,0x00040047,0x0000000d,0x00000022,
+	0x00000000,0x00040047,0x00000011,0x0000001e,0x00000000,0x00020013,0x00000002,0x00030021,
+	0x00000003,0x00000002,0x00030016,0x00000006,0x00000020,0x00040017,0x00000007,0x00000006,
+	0x00000004,0x00040020,0x00000008,0x00000003,0x00000007,0x0004003b,0x00000008,0x00000009,
+	0x00000003,0x00090019,0x0000000a,0x00000006,0x00000001,0x00000000,0x00000000,0x00000000,
+	0x00000001,0x00000000,0x0003001b,0x0000000b,0x0000000a,0x00040020,0x0000000c,0x00000000,
+	0x0000000b,0x0004003b,0x0000000c,0x0000000d,0x00000000,0x00040017,0x0000000f,0x00000006,
+	0x00000002,0x00040020,0x00000010,0x00000001,0x0000000f,0x0004003b,0x00000010,0x00000011,
+	0x00000001,0x00050036,0x00000002,0x00000004,0x00000000,0x00000003,0x000200f8,0x00000005,
+	0x0004003d,0x0000000b,0x0000000e,0x0000000d,0x0004003d,0x0000000f,0x00000012,0x00000011,
+	0x00050057,0x00000007,0x00000013,0x0000000e,0x00000012,0x0003003e,0x00000009,0x00000013,
+	0x000100fd,0x00010038
+};
@@ -0,0 +1,315 @@
+/* Alpha composition: one surface drawn source-over onto the scene - see blend_pass.h. Compositor thread. */
+#define _POSIX_C_SOURCE 200809L
+#include "blend_pass.h"
+#include "vk_loader.h"
+#include "vk_present.h" /* banner_log, vkp_gpu_name */
+#include <string.h>
+
+/* SPIR-V, pre-compiled (glslangValidator -V blend.vert --vn blend_vert_code, the same for blend.frag)
+ * and committed next to the GLSL: the NDK build compiles no shaders. */
+#include "blend_vert.h"
+#include "blend_frag.h"
+
+struct pc_blend { float dst[4]; float src[4]; };
+
+static VkDevice g_dev;
+static int g_ready;                 /* 0 = not built, 1 = ok, -1 = unavailable (said once) */
+static VkSampler g_sampler;
+static VkDescriptorSetLayout g_dsl;
+static VkPipelineLayout g_pl;
+static VkDescriptorPool g_dpool;
+static VkShaderModule g_vert, g_frag;
+static unsigned g_frame = 1;
+
+/* One render pass + pipeline per target format: the screen swapchain's, the scene image's, the HDR
+ * composition's. */
+#define NFMT 4
+static struct { VkFormat fmt; VkRenderPass rp; VkPipeline pipe; } g_out[NFMT];
+
+#define NSRC 32
+static struct { VkImage img; VkImageView view; VkDescriptorSet ds; unsigned frame; } g_src[NSRC];
+#define NDST 12
+static struct { VkImage img; VkFormat fmt; int w, h; VkImageView view; VkFramebuffer fb; unsigned frame; } g_dst[NDST];
+
+void blendp_bind_device(VkDevice dev) { g_dev = dev; }
+
+void blendp_begin_frame(void) { g_frame++; }
+
+static void src_release(int i) {
+    if (g_src[i].ds) g_vk.FreeDescriptorSets(g_dev, g_dpool, 1, &g_src[i].ds);
+    if (g_src[i].view) g_vk.DestroyImageView(g_dev, g_src[i].view, NULL);
+    memset(&g_src[i], 0, sizeof(g_src[i]));
+}
+
+static void dst_release(int i) {
+    if (g_dst[i].fb) g_vk.DestroyFramebuffer(g_dev, g_dst[i].fb, NULL);
+    if (g_dst[i].view) g_vk.DestroyImageView(g_dev, g_dst[i].view, NULL);
+    memset(&g_dst[i], 0, sizeof(g_dst[i]));
+}
+
+void blendp_forget_image(VkImage img) {
+    if (!g_dev || !img) return;
+    for (int i = 0; i < NSRC; i++) if (g_src[i].img == img) src_release(i);
+    for (int i = 0; i < NDST; i++) if (g_dst[i].img == img) dst_release(i);
+}
+
+static VkShaderModule shader(const uint32_t *code, size_t size) {
+    VkShaderModule m = VK_NULL_HANDLE;
+    VkShaderModuleCreateInfo ci = {.sType = VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO, .codeSize = size, .pCode = code};
+    if (g_vk.CreateShaderModule(g_dev, &ci, NULL, &m) != VK_SUCCESS) return VK_NULL_HANDLE;
+    return m;
+}
+
+/* The target keeps what the blits put there (LOAD) and stays in TRANSFER_DST_OPTIMAL for the blits
+ * that follow. The dependencies order the pass between them. */
+static VkRenderPass make_rp(VkFormat fmt) {
+    VkAttachmentDescription att = {.format = fmt, .samples = VK_SAMPLE_COUNT_1_BIT,
+                                   .loadOp = VK_ATTACHMENT_LOAD_OP_LOAD, .storeOp = VK_ATTACHMENT_STORE_OP_STORE,
+                                   .stencilLoadOp = VK_ATTACHMENT_LOAD_OP_DONT_CARE,
+                                   .stencilStoreOp = VK_ATTACHMENT_STORE_OP_DONT_CARE,
+                                   .initialLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
+                                   .finalLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL};
+    VkAttachmentReference ref = {0, VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL};
+    VkSubpassDescription sub = {.pipelineBindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS, .colorAttachmentCount = 1,
+                                .pColorAttachments = &ref};
+    VkSubpassDependency deps[2] = {
+        {.srcSubpass = VK_SUBPASS_EXTERNAL, .dstSubpass = 0,
+         .srcStageMask = VK_PIPELINE_STAGE_TRANSFER_BIT | VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT,
+         .dstStageMask = VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT,
+         .srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT | VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT,
+         .dstAccessMask = VK_ACCESS_COLOR_ATTACHMENT_READ_BIT | VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT},
+        {.srcSubpass = 0, .dstSubpass = VK_SUBPASS_EXTERNAL,
+         .srcStageMask = VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT,
+         .dstStageMask = VK_PIPELINE_STAGE_TRANSFER_BIT | VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT,
+         .srcAccessMask = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT,
+         .dstAccessMask = VK_ACCESS_TRANSFER_READ_BIT | VK_ACCESS_TRANSFER_WRITE_BIT |
+                          VK_ACCESS_COLOR_ATTACHMENT_READ_BIT | VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT}};
+    VkRenderPassCreateInfo rci = {.sType = VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO, .attachmentCount = 1,
+                                  .pAttachments = &att, .subpassCount = 1, .pSubpasses = &sub,
+                                  .dependencyCount = 2, .pDependencies = deps};
+    VkRenderPass rp = VK_NULL_HANDLE;
+    if (g_vk.CreateRenderPass(g_dev, &rci, NULL, &rp) != VK_SUCCESS) return VK_NULL_HANDLE;
+    return rp;
+}
+
+static VkPipeline make_pipe(VkRenderPass rp) {
+    VkPipelineShaderStageCreateInfo st[2] = {
+        {.sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO, .stage = VK_SHADER_STAGE_VERTEX_BIT,
+         .module = g_vert, .pName = "main"},
+        {.sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO, .stage = VK_SHADER_STAGE_FRAGMENT_BIT,
+         .module = g_frag, .pName = "main"}};
+    VkPipelineVertexInputStateCreateInfo vi = {.sType = VK_STRUCTURE_TYPE_PIPELINE_VERTEX_INPUT_STATE_CREATE_INFO};
+    VkPipelineInputAssemblyStateCreateInfo ia = {.sType = VK_STRUCTURE_TYPE_PIPELINE_INPUT_ASSEMBLY_STATE_CREATE_INFO,
+                                                 .topology = VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP};
+    VkDynamicState dyn[2] = {VK_DYNAMIC_STATE_VIEWPORT, VK_DYNAMIC_STATE_SCISSOR};
+    VkPipelineDynamicStateCreateInfo ds = {.sType = VK_STRUCTURE_TYPE_PIPELINE_DYNAMIC_STATE_CREATE_INFO,
+                                           .dynamicStateCount = 2, .pDynamicStates = dyn};
+    VkPipelineViewportStateCreateInfo vp = {.sType = VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO,
+                                            .viewportCount = 1, .scissorCount = 1};
+    VkPipelineRasterizationStateCreateInfo rs = {.sType = VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_STATE_CREATE_INFO,
+                                                 .polygonMode = VK_POLYGON_MODE_FILL, .cullMode = VK_CULL_MODE_NONE,
+                                                 .frontFace = VK_FRONT_FACE_COUNTER_CLOCKWISE, .lineWidth = 1.0f};
+    VkPipelineMultisampleStateCreateInfo ms = {.sType = VK_STRUCTURE_TYPE_PIPELINE_MULTISAMPLE_STATE_CREATE_INFO,
+                                               .rasterizationSamples = VK_SAMPLE_COUNT_1_BIT};
+    /* The target's alpha is left alone: the screen is opaque whatever is drawn onto it. */
+    VkPipelineColorBlendAttachmentState ba = {
+        .blendEnable = VK_TRUE, .srcColorBlendFactor = VK_BLEND_FACTOR_ONE,
+        .dstColorBlendFactor = VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA, .colorBlendOp = VK_BLEND_OP_ADD,
+        .srcAlphaBlendFactor = VK_BLEND_FACTOR_ONE, .dstAlphaBlendFactor = VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA,
+        .alphaBlendOp = VK_BLEND_OP_ADD,
+        .colorWriteMask = VK_COLOR_COMPONENT_R_BIT | VK_COLOR_COMPONENT_G_BIT | VK_COLOR_COMPONENT_B_BIT};
+    VkPipelineColorBlendStateCreateInfo cb = {.sType = VK_STRUCTURE_TYPE_PIPELINE_COLOR_BLEND_STATE_CREATE_INFO,
+                                              .attachmentCount = 1, .pAttachments = &ba};
+    VkGraphicsPipelineCreateInfo pi = {.sType = VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO, .stageCount = 2,
+                                       .pStages = st, .pVertexInputState = &vi, .pInputAssemblyState = &ia,
+                                       .pViewportState = &vp, .pRasterizationState = &rs, .pMultisampleState = &ms,
+                                       .pColorBlendState = &cb, .pDynamicState = &ds, .layout = g_pl,
+                                       .renderPass = rp, .subpass = 0};
+    VkPipeline p = VK_NULL_HANDLE;
+    if (g_vk.CreateGraphicsPipelines(g_dev, VK_NULL_HANDLE, 1, &pi, NULL, &p) != VK_SUCCESS) return VK_NULL_HANDLE;
+    return p;
+}
+
+static void objects_destroy(void) {
+    if (g_vert) g_vk.DestroyShaderModule(g_dev, g_vert, NULL);
+    if (g_frag) g_vk.DestroyShaderModule(g_dev, g_frag, NULL);
+    if (g_dpool) g_vk.DestroyDescriptorPool(g_dev, g_dpool, NULL);
+    if (g_pl) g_vk.DestroyPipelineLayout(g_dev, g_pl, NULL);
+    if (g_dsl) g_vk.DestroyDescriptorSetLayout(g_dev, g_dsl, NULL);
+    if (g_sampler) g_vk.DestroySampler(g_dev, g_sampler, NULL);
+    g_vert = g_frag = VK_NULL_HANDLE; g_dpool = VK_NULL_HANDLE; g_pl = VK_NULL_HANDLE;
+    g_dsl = VK_NULL_HANDLE; g_sampler = VK_NULL_HANDLE;
+}
+
+static int objects_ensure(void) {
+    if (g_ready) return g_ready == 1 ? 0 : -1;
+    g_ready = -1;
+    if (!g_dev || !g_vk.CreateGraphicsPipelines) goto fail;
+    VkSamplerCreateInfo sci = {.sType = VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO, .magFilter = VK_FILTER_LINEAR,
+                               .minFilter = VK_FILTER_LINEAR, .mipmapMode = VK_SAMPLER_MIPMAP_MODE_NEAREST,
+                               .addressModeU = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE,
+                               .addressModeV = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE,
+                               .addressModeW = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE};
+    if (g_vk.CreateSampler(g_dev, &sci, NULL, &g_sampler) != VK_SUCCESS) goto fail;
+    VkDescriptorSetLayoutBinding b = {.binding = 0, .descriptorType = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
+                                      .descriptorCount = 1, .stageFlags = VK_SHADER_STAGE_FRAGMENT_BIT};
+    VkDescriptorSetLayoutCreateInfo dli = {.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO,
+                                           .bindingCount = 1, .pBindings = &b};
+    if (g_vk.CreateDescriptorSetLayout(g_dev, &dli, NULL, &g_dsl) != VK_SUCCESS) goto fail;
+    VkPushConstantRange pcr = {.stageFlags = VK_SHADER_STAGE_VERTEX_BIT, .offset = 0, .size = sizeof(struct pc_blend)};
+    VkPipelineLayoutCreateInfo pli = {.sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO, .setLayoutCount = 1,
+                                      .pSetLayouts = &g_dsl, .pushConstantRangeCount = 1, .pPushConstantRanges = &pcr};
+    if (g_vk.CreatePipelineLayout(g_dev, &pli, NULL, &g_pl) != VK_SUCCESS) goto fail;
+    VkDescriptorPoolSize ps = {VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, NSRC};
+    VkDescriptorPoolCreateInfo dpi = {.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_CREATE_INFO,
+                                      .flags = VK_DESCRIPTOR_POOL_CREATE_FREE_DESCRIPTOR_SET_BIT, .maxSets = NSRC,
+                                      .poolSizeCount = 1, .pPoolSizes = &ps};
+    if (g_vk.CreateDescriptorPool(g_dev, &dpi, NULL, &g_dpool) != VK_SUCCESS) goto fail;
+    if (!(g_vert = shader(blend_vert_code, sizeof(blend_vert_code)))) goto fail;
+    if (!(g_frag = shader(blend_frag_code, sizeof(blend_frag_code)))) goto fail;
+    g_ready = 1;
+    banner_log("gpu", "alpha composition ready on %s: translucent surfaces are blended over what is under them",
+            vkp_gpu_name());
+    return 0;
+fail:
+    objects_destroy();
+    banner_log("error", "gpu: the alpha composition pass could not be built on this driver; translucent surfaces "
+               "are drawn opaque this session");
+    return -1;
+}
+
+static int out_slot(VkFormat fmt) {
+    int i;
+    for (i = 0; i < NFMT && g_out[i].pipe; i++) if (g_out[i].fmt == fmt) return i;
+    if (i == NFMT) return -1;
+    VkRenderPass rp = make_rp(fmt);
+    VkPipeline pipe = rp ? make_pipe(rp) : VK_NULL_HANDLE;
+    if (!pipe) {
+        if (rp) g_vk.DestroyRenderPass(g_dev, rp, NULL);
+        return -1;
+    }
+    g_out[i].fmt = fmt; g_out[i].rp = rp; g_out[i].pipe = pipe;
+    return i;
+}
+
+static VkImageView make_view(VkImage img, VkFormat fmt) {
+    VkImageView v = VK_NULL_HANDLE;
+    VkImageViewCreateInfo ci = {.sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO, .image = img,
+                                .viewType = VK_IMAGE_VIEW_TYPE_2D, .format = fmt,
+                                .subresourceRange = {VK_IMAGE_ASPECT_COLOR_BIT, 0, 1, 0, 1}};
+    if (g_vk.CreateImageView(g_dev, &ci, NULL, &v) != VK_SUCCESS) return VK_NULL_HANDLE;
+    return v;
+}
+
+/* The slot to recycle: a free one, else the one unused for longest - never one this command buffer
+ * already refers to. -1 = every slot is in this frame. */
+#define OLDEST(arr, n, out) do { \
+        unsigned age_ = 0; \
+        (out) = -1; \
+        for (int k_ = 0; k_ < (n); k_++) { \
+            if (!(arr)[k_].img) { (out) = k_; break; } \
+            if ((arr)[k_].frame != g_frame && g_frame - (arr)[k_].frame > age_) { age_ = g_frame - (arr)[k_].frame; (out) = k_; } \
+        } \
+    } while (0)
+
+static VkDescriptorSet src_ensure(VkImage img, VkFormat fmt, VkImageLayout layout) {
+    int i;
+    for (i = 0; i < NSRC; i++)
+        if (g_src[i].img == img) { g_src[i].frame = g_frame; return g_src[i].ds; }
+    OLDEST(g_src, NSRC, i);
+    if (i < 0) return VK_NULL_HANDLE;
+    src_release(i);
+    if (!(g_src[i].view = make_view(img, fmt))) return VK_NULL_HANDLE;
+    VkDescriptorSetAllocateInfo ai = {.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_ALLOCATE_INFO,
+                                      .descriptorPool = g_dpool, .descriptorSetCount = 1, .pSetLayouts = &g_dsl};
+    if (g_vk.AllocateDescriptorSets(g_dev, &ai, &g_src[i].ds) != VK_SUCCESS) {
+        g_src[i].ds = VK_NULL_HANDLE;
+        src_release(i);
+        return VK_NULL_HANDLE;
+    }
+    VkDescriptorImageInfo ii = {.sampler = g_sampler, .imageView = g_src[i].view, .imageLayout = layout};
+    VkWriteDescriptorSet wr = {.sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET, .dstSet = g_src[i].ds, .dstBinding = 0,
+                               .descriptorCount = 1, .descriptorType = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
+                               .pImageInfo = &ii};
+    g_vk.UpdateDescriptorSets(g_dev, 1, &wr, 0, NULL);
+    g_src[i].img = img; g_src[i].frame = g_frame;
+    return g_src[i].ds;
+}
+
+static VkFramebuffer dst_ensure(VkImage img, VkFormat fmt, int w, int h, VkRenderPass rp) {
+    int i;
+    for (i = 0; i < NDST; i++) {
+        if (g_dst[i].img != img) continue;
+        if (g_dst[i].fmt == fmt && g_dst[i].w == w && g_dst[i].h == h) { g_dst[i].frame = g_frame; return g_dst[i].fb; }
+        if (g_dst[i].frame == g_frame) return VK_NULL_HANDLE;
+        break;
+    }
+    if (i == NDST) OLDEST(g_dst, NDST, i);
+    if (i < 0) return VK_NULL_HANDLE;
+    dst_release(i);
+    if (!(g_dst[i].view = make_view(img, fmt))) return VK_NULL_HANDLE;
+    VkFramebufferCreateInfo fi = {.sType = VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO, .renderPass = rp,
+                                  .attachmentCount = 1, .pAttachments = &g_dst[i].view, .width = (uint32_t)w,
+                                  .height = (uint32_t)h, .layers = 1};
+    if (g_vk.CreateFramebuffer(g_dev, &fi, NULL, &g_dst[i].fb) != VK_SUCCESS) {
+        g_dst[i].fb = VK_NULL_HANDLE;
+        dst_release(i);
+        return VK_NULL_HANDLE;
+    }
+    g_dst[i].img = img; g_dst[i].fmt = fmt; g_dst[i].w = w; g_dst[i].h = h; g_dst[i].frame = g_frame;
+    return g_dst[i].fb;
+}
+
+static void src_layout(VkCommandBuffer cmd, VkImage img, VkImageLayout from, VkImageLayout to) {
+    int sampling = to == VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
+    VkImageMemoryBarrier b = {.sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER, .oldLayout = from, .newLayout = to,
+                              .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED, .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+                              .image = img, .subresourceRange = {VK_IMAGE_ASPECT_COLOR_BIT, 0, 1, 0, 1},
+                              .srcAccessMask = sampling ? VK_ACCESS_TRANSFER_READ_BIT : VK_ACCESS_SHADER_READ_BIT,
+                              .dstAccessMask = sampling ? VK_ACCESS_SHADER_READ_BIT : VK_ACCESS_TRANSFER_READ_BIT};
+    g_vk.CmdPipelineBarrier(cmd, sampling ? VK_PIPELINE_STAGE_TRANSFER_BIT : VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT,
+                            sampling ? VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT : VK_PIPELINE_STAGE_TRANSFER_BIT,
+                            0, 0, NULL, 0, NULL, 1, &b);
+}
+
+int blendp_draw(VkCommandBuffer cmd, VkImage src, VkFormat src_fmt, int src_w, int src_h, int src_general,
+                VkImage dst, VkFormat dst_fmt, int dst_w, int dst_h, const VkImageBlit *blit) {
+    if (!src || !dst || src_w <= 0 || src_h <= 0 || dst_w <= 0 || dst_h <= 0 || objects_ensure() != 0) return -1;
+    int slot = out_slot(dst_fmt);
+    if (slot < 0) return -1;
+    VkImageLayout sampled = src_general ? VK_IMAGE_LAYOUT_GENERAL : VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
+    VkDescriptorSet ds = src_ensure(src, src_fmt, sampled);
+    if (!ds) return -1;
+    VkFramebuffer fb = dst_ensure(dst, dst_fmt, dst_w, dst_h, g_out[slot].rp);
+    if (!fb) return -1;
+
+    const VkOffset3D *d = blit->dstOffsets, *s = blit->srcOffsets;
+    int x0 = d[0].x < d[1].x ? d[0].x : d[1].x, x1 = d[0].x < d[1].x ? d[1].x : d[0].x;
+    int y0 = d[0].y < d[1].y ? d[0].y : d[1].y, y1 = d[0].y < d[1].y ? d[1].y : d[0].y;
+    if (x0 < 0) x0 = 0;
+    if (y0 < 0) y0 = 0;
+    if (x1 > dst_w) x1 = dst_w;
+    if (y1 > dst_h) y1 = dst_h;
+    if (x1 <= x0 || y1 <= y0) return 0;
+
+    struct pc_blend pc = {
+        .dst = {2.0f * d[0].x / dst_w - 1.0f, 2.0f * d[0].y / dst_h - 1.0f,
+                2.0f * d[1].x / dst_w - 1.0f, 2.0f * d[1].y / dst_h - 1.0f},
+        .src = {(float)s[0].x / src_w, (float)s[0].y / src_h, (float)s[1].x / src_w, (float)s[1].y / src_h}};
+
+    if (!src_general) src_layout(cmd, src, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, sampled);
+    VkRenderPassBeginInfo rpi = {.sType = VK_STRUCTURE_TYPE_RENDER_PASS_BEGIN_INFO, .renderPass = g_out[slot].rp,
+                                 .framebuffer = fb,
+                                 .renderArea = {{x0, y0}, {(uint32_t)(x1 - x0), (uint32_t)(y1 - y0)}}};
+    VkViewport vp = {0, 0, (float)dst_w, (float)dst_h, 0, 1};
+    g_vk.CmdBeginRenderPass(cmd, &rpi, VK_SUBPASS_CONTENTS_INLINE);
+    g_vk.CmdSetViewport(cmd, 0, 1, &vp);
+    g_vk.CmdSetScissor(cmd, 0, 1, &rpi.renderArea);
+    g_vk.CmdBindPipeline(cmd, VK_PIPELINE_BIND_POINT_GRAPHICS, g_out[slot].pipe);
+    g_vk.CmdBindDescriptorSets(cmd, VK_PIPELINE_BIND_POINT_GRAPHICS, g_pl, 0, 1, &ds, 0, NULL);
+    g_vk.CmdPushConstants(cmd, g_pl, VK_SHADER_STAGE_VERTEX_BIT, 0, sizeof(pc), &pc);
+    g_vk.CmdDraw(cmd, 4, 1, 0, 0);
+    g_vk.CmdEndRenderPass(cmd);
+    if (!src_general) src_layout(cmd, src, sampled, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL);
+    return 0;
+}
@@ -0,0 +1,28 @@
+#ifndef BLEND_PASS_H
+#define BLEND_PASS_H
+/*
+ * Alpha composition. The scene is composed with blits, which replace what is under them; a surface
+ * with translucent pixels (gamescope's overlay and notification planes over a game, a toolkit's
+ * shadowed popup) is drawn through this pass instead, source-over with premultiplied alpha.
+ *
+ * Compositor thread. Objects are built on the first use; if the driver refuses them the pass is
+ * unavailable for the session (logged once) and the caller keeps blitting.
+ */
+#define VK_USE_PLATFORM_ANDROID_KHR
+#include <vulkan/vulkan.h>
+
+/* Called from dev_init (device up). */
+void blendp_bind_device(VkDevice dev);
+/* A source or target image is about to be destroyed: drop what is cached for it. Frames are fenced,
+ * so nothing submitted still uses it. */
+void blendp_forget_image(VkImage img);
+/* A new command buffer is being recorded: what the previous one used may be recycled again. */
+void blendp_begin_frame(void);
+
+/* Record one surface over `dst`. `dst` is in TRANSFER_DST_OPTIMAL before and after. `src` was created
+ * with SAMPLED usage and is in GENERAL (src_general) or TRANSFER_SRC_OPTIMAL, and is left there.
+ * `blit` is the draw as the blit path maps it. 0 = recorded, -1 = unavailable (nothing recorded). */
+int blendp_draw(VkCommandBuffer cmd, VkImage src, VkFormat src_fmt, int src_w, int src_h, int src_general,
+                VkImage dst, VkFormat dst_fmt, int dst_w, int dst_h, const VkImageBlit *blit);
+
+#endif
@@ -0,0 +1,55 @@
+	// 1115.1.0
+	 #pragma once
+const uint32_t blend_vert_code[] = {
+	0x07230203,0x00010000,0x0008000b,0x0000003a,0x00000000,0x00020011,0x00000001,0x0006000b,
+	0x00000001,0x4c534c47,0x6474732e,0x3035342e,0x00000000,0x0003000e,0x00000000,0x00000001,
+	0x0008000f,0x00000000,0x00000004,0x6e69616d,0x00000000,0x0000000c,0x0000001b,0x00000031,
+	0x00030003,0x00000002,0x000001c2,0x00040005,0x00000004,0x6e69616d,0x00000000,0x00030005,
+	0x00000009,0x00000074,0x00060005,0x0000000c,0x565f6c67,0x65747265,0x646e4978,0x00007865,
+	0x00060005,0x00000019,0x505f6c67,0x65567265,0x78657472,0x00000000,0x00060006,0x00000019,
+	0x00000000,0x505f6c67,0x7469736f,0x006e6f69,0x00070006,0x00000019,0x00000001,0x505f6c67,
+	0x746e696f,0x657a6953,0x00000000,0x00070006,0x00000019,0x00000002,0x435f6c67,0x4470696c,
+	0x61747369,0x0065636e,0x00070006,0x00000019,0x00000003,0x435f6c67,0x446c6c75,0x61747369,
+	0x0065636e,0x00030005,0x0000001b,0x00000000,0x00030005,0x0000001d,0x00004350,0x00040006,
+	0x0000001d,0x00000000,0x00747364,0x00040006,0x0000001d,0x00000001,0x00637273,0x00030005,
+	0x0000001f,0x00006370,0x00030005,0x00000031,0x00007675,0x00040047,0x0000000c,0x0000000b,
+	0x0000002a,0x00030047,0x00000019,0x00000002,0x00050048,0x00000019,0x00000000,0x0000000b,
+	0x00000000,0x00050048,0x00000019,0x00000001,0x0000000b,0x00000001,0x00050048,0x00000019,
+	0x00000002,0x0000000b,0x00000003,0x00050048,0x00000019,0x00000003,0x0000000b,0x00000004,
+	0x00030047,0x0000001d,0x00000002,0x00050048,0x0000001d,0x00000000,0x00000023,0x00000000,
+	0x00050048,0x0000001d,0x00000001,0x00000023,0x00000010,0x00040047,0x00000031,0x0000001e,
+	0x00000000,0x00020013,0x00000002,0x00030021,0x00000003,0x00000002,0x00030016,0x00000006,
+	0x00000020,0x00040017,0x00000007,0x00000006,0x00000002,0x00040020,0x00000008,0x00000007,
+	0x00000007,0x00040015,0x0000000a,0x00000020,0x00000001,0x00040020,0x0000000b,0x00000001,
+	0x0000000a,0x0004003b,0x0000000b,0x0000000c,0x00000001,0x0004002b,0x0000000a,0x0000000e,
+	0x00000001,0x00040017,0x00000015,0x00000006,0x00000004,0x00040015,0x00000016,0x00000020,
+	0x00000000,0x0004002b,0x00000016,0x00000017,0x00000001,0x0004001c,0x00000018,0x00000006,
+	0x00000017,0x0006001e,0x00000019,0x00000015,0x00000006,0x00000018,0x00000018,0x00040020,
+	0x0000001a,0x00000003,0x00000019,0x0004003b,0x0000001a,0x0000001b,0x00000003,0x0004002b,
+	0x0000000a,0x0000001c,0x00000000,0x0004001e,0x0000001d,0x00000015,0x00000015,0x00040020,
+	0x0000001e,0x00000009,0x0000001d,0x0004003b,0x0000001e,0x0000001f,0x00000009,0x00040020,
+	0x00000020,0x00000009,0x00000015,0x0004002b,0x00000006,0x00000029,0x00000000,0x0004002b,
+	0x00000006,0x0000002a,0x3f800000,0x00040020,0x0000002e,0x00000003,0x00000015,0x00040020,
+	0x00000030,0x00000003,0x00000007,0x0004003b,0x00000030,0x00000031,0x00000003,0x00050036,
+	0x00000002,0x00000004,0x00000000,0x00000003,0x000200f8,0x00000005,0x0004003b,0x00000008,
+	0x00000009,0x00000007,0x0004003d,0x0000000a,0x0000000d,0x0000000c,0x000500c7,0x0000000a,
+	0x0000000f,0x0000000d,0x0000000e,0x0004006f,0x00000006,0x00000010,0x0000000f,0x0004003d,
+	0x0000000a,0x00000011,0x0000000c,0x000500c3,0x0000000a,0x00000012,0x00000011,0x0000000e,
+	0x0004006f,0x00000006,0x00000013,0x00000012,0x00050050,0x00000007,0x00000014,0x00000010,
+	0x00000013,0x0003003e,0x00000009,0x00000014,0x00050041,0x00000020,0x00000021,0x0000001f,
+	0x0000001c,0x0004003d,0x00000015,0x00000022,0x00000021,0x0007004f,0x00000007,0x00000023,
+	0x00000022,0x00000022,0x00000000,0x00000001,0x00050041,0x00000020,0x00000024,0x0000001f,
+	0x0000001c,0x0004003d,0x00000015,0x00000025,0x00000024,0x0007004f,0x00000007,0x00000026,
+	0x00000025,0x00000025,0x00000002,0x00000003,0x0004003d,0x00000007,0x00000027,0x00000009,
+	0x0008000c,0x00000007,0x00000028,0x00000001,0x0000002e,0x00000023,0x00000026,0x00000027,
+	0x00050051,0x00000006,0x0000002b,0x00000028,0x00000000,0x00050051,0x00000006,0x0000002c,
+	0x00000028,0x00000001,0x00070050,0x00000015,0x0000002d,0x0000002b,0x0000002c,0x00000029,
+	0x0000002a,0x00050041,0x0000002e,0x0000002f,0x0000001b,0x0000001c,0x0003003e,0x0000002f,
+	0x0000002d,0x00050041,0x00000020,0x00000032,0x0000001f,0x0000000e,0x0004003d,0x00000015,
+	0x00000033,0x00000032,0x0007004f,0x00000007,0x00000034,0x00000033,0x00000033,0x00000000,
+	0x00000001,0x00050041,0x00000020,0x00000035,0x0000001f,0x0000000e,0x0004003d,0x00000015,
+	0x00000036,0x00000035,0x0007004f,0x00000007,0x00000037,0x00000036,0x00000036,0x00000002,
+	0x00000003,0x0004003d,0x00000007,0x00000038,0x00000009,0x0008000c,0x00000007,0x00000039,
+	0x00000001,0x0000002e,0x00000034,0x00000037,0x00000038,0x0003003e,0x00000031,0x00000039,
+	0x000100fd,0x00010038
+};
@@ -170,6 +170,9 @@ struct client_info {
    /* An OpenGL program whose EGL gave up on the GPU: it asked for dma-buf feedback (EGL's
     * Wayland GPU path always does), never made a dma-buf buffer, and draws wl_shm frames. */
    unsigned asked_feedback : 1, shm_gl_said : 1;
+    /* Clients that describe opaque regions can use alpha outside those regions. Wine windows
+     * never describe them, and their unused alpha channel must not make them translucent. */
+    unsigned declares_opaque : 1;
    unsigned dmabuf_buffers, shm_frames;
    struct client_info *next;
 };
@@ -248,6 +251,8 @@ struct surface {
    struct dmabuf_buffer *dmabuf_buf;       /* its imported image, kept until replaced or the surface goes */
    struct wl_listener dmabuf_destroy;
    int buf_w, buf_h, has_content;
+    int buf_alpha;
+    int opaque[4], pending_opaque[4], pending_opaque_set; /* x, y, w, h; w = 0: none */
    int src_set, dst_set;
    float src[4];
    int dst[2];
@@ -743,6 +748,7 @@ static void take_shm(struct surface *s, struct wl_shm_buffer *shm, struct wl_res
    wl_buffer_send_release(buffer);
    s->buf_w = w;
    s->buf_h = h;
+    s->buf_alpha = wl_shm_buffer_get_format(shm) == WL_SHM_FORMAT_ARGB8888;
    s->has_content = s->shm_img != NULL;
    g_stat_shm++;
    struct client_info *ci = client_info_of(wl_resource_get_client(s->resource));
@@ -794,6 +800,7 @@ static void take_dmabuf(struct surface *s, struct dmabuf_buffer *b, struct wl_re
    }
    s->buf_w = b->width;
    s->buf_h = b->height;
+    s->buf_alpha = (b->format & 0xff) == 'A';
    s->has_content = b->img != NULL;
    g_stat_dmabuf++;
    const int announced_now = !s->announced_vulkan;
@@ -965,8 +972,27 @@ static void surface_frame(struct wl_client *c, struct wl_resource *r, uint32_t c
    wl_resource_set_implementation(callback, NULL, NULL, frame_callback_destroy);
    wl_list_insert(s->pending_frames.prev, wl_resource_get_link(callback));
 }
+/* The bounding box serves pointer confinement. The exact flag also tells opaque-region handling
+ * whether the bounding box itself is the whole region. */
+struct region { int set, exact; int x, y, w, h; };
+
+/* A wl_surface's opaque region is pending until its next commit. An exact region that covers
+ * the surface lets the blit path keep handling an otherwise alpha-capable buffer. */
 static void surface_set_opaque(struct wl_client *c, struct wl_resource *r,
-                               struct wl_resource *region) {}
+                               struct wl_resource *region) {
+    struct surface *s = wl_resource_get_user_data(r);
+    struct region *rg = region ? wl_resource_get_user_data(region) : NULL;
+    memset(s->pending_opaque, 0, sizeof(s->pending_opaque));
+    s->pending_opaque_set = 1;
+    if (rg && rg->set && rg->exact) {
+        s->pending_opaque[0] = rg->x;
+        s->pending_opaque[1] = rg->y;
+        s->pending_opaque[2] = rg->w;
+        s->pending_opaque[3] = rg->h;
+    }
+    struct client_info *ci = client_info_of(c);
+    if (ci) ci->declares_opaque = 1;
+}
 static void surface_set_input(struct wl_client *c, struct wl_resource *r,
                              struct wl_resource *region) {}

@@ -995,6 +1021,10 @@ static void surface_commit(struct wl_client *c, struct wl_resource *r) {
        memcpy(s->dst, s->pending_dst, sizeof(s->dst));
        s->pending_dst_set = 0;
    }
+    if (s->pending_opaque_set) {
+        memcpy(s->opaque, s->pending_opaque, sizeof(s->opaque));
+        s->pending_opaque_set = 0;
+    }

    if (s->pending_attach) {
        struct wl_resource *buffer = s->pending_buffer;
@@ -1139,24 +1169,29 @@ static void surface_resource_destroy(struct wl_resource *r) {

 /* ------------------------------------------------------------------ wl_region */

-/* A region is kept as the bounding box of its rectangles: enough for pointer confinement,
- * where winewayland sends one rectangle (the ClipCursor area). Subtractions are ignored. */
-struct region { int set; int x, y, w, h; };
-
 static void region_destroy(struct wl_client *c, struct wl_resource *r) { wl_resource_destroy(r); }
 static void region_add(struct wl_client *c, struct wl_resource *r,
                       int32_t x, int32_t y, int32_t w, int32_t h) {
    struct region *rg = wl_resource_get_user_data(r);
    if (!rg || w <= 0 || h <= 0) return;
-    if (!rg->set) { rg->x = x; rg->y = y; rg->w = w; rg->h = h; rg->set = 1; return; }
-    int x2 = rg->x + rg->w > x + w ? rg->x + rg->w : x + w;
-    int y2 = rg->y + rg->h > y + h ? rg->y + rg->h : y + h;
+    if (!rg->set) { rg->x = x; rg->y = y; rg->w = w; rg->h = h; rg->set = rg->exact = 1; return; }
+    long long x2 = (long long)rg->x + rg->w, y2 = (long long)rg->y + rg->h;
+    long long nx2 = (long long)x + w, ny2 = (long long)y + h;
+    int inside = x >= rg->x && y >= rg->y && nx2 <= x2 && ny2 <= y2;
+    int around = x <= rg->x && y <= rg->y && nx2 >= x2 && ny2 >= y2;
+    if (!inside && !around) rg->exact = 0;
+    if (nx2 > x2) x2 = nx2;
+    if (ny2 > y2) y2 = ny2;
    if (x < rg->x) rg->x = x;
    if (y < rg->y) rg->y = y;
-    rg->w = x2 - rg->x; rg->h = y2 - rg->y;
+    rg->w = (int)(x2 - rg->x > INT32_MAX ? INT32_MAX : x2 - rg->x);
+    rg->h = (int)(y2 - rg->y > INT32_MAX ? INT32_MAX : y2 - rg->y);
 }
 static void region_subtract(struct wl_client *c, struct wl_resource *r,
-                            int32_t x, int32_t y, int32_t w, int32_t h) {}
+                            int32_t x, int32_t y, int32_t w, int32_t h) {
+    struct region *rg = wl_resource_get_user_data(r);
+    if (rg && w > 0 && h > 0) rg->exact = 0;
+}
 static void region_resource_destroy(struct wl_resource *r) { free(wl_resource_get_user_data(r)); }
 static const struct wl_region_interface region_impl = {
    .destroy = region_destroy,
@@ -1942,6 +1977,17 @@ static void note_hdr_unimported(const struct draw_list *dl, struct surface *s, i
    g_hdr_unimported_below = dl->n;
 }

+/* A client that describes opaque regions can intentionally leave part of an alpha buffer see-through.
+ * Gamescope uses this for its full-size Steam notification and overlay planes. */
+static int surface_translucent(const struct surface *s, int w, int h) {
+    if (!s->buf_alpha) return 0;
+    const struct client_info *ci = client_info_of(wl_resource_get_client(s->resource));
+    if (!ci || !ci->declares_opaque) return 0;
+    const int *o = s->opaque;
+    return !(o[2] > 0 && o[0] <= 0 && o[1] <= 0 &&
+             (long long)o[0] + o[2] >= w && (long long)o[1] + o[3] >= h);
+}
+
 static void add_surface(struct draw_list *dl, struct surface *s, int ox, int oy) {
    struct vkp_image *img = surface_image(s);
    float sx = 0, sy = 0, sw = (float)s->buf_w, sh = (float)s->buf_h;
@@ -1961,7 +2007,8 @@ static void add_surface(struct draw_list *dl, struct surface *s, int ox, int oy)
        dl->d = d;
        dl->cap = cap;
    }
-    dl->d[dl->n++] = (struct vkp_draw){img, sx, sy, sw, sh, ox, oy, dw, dh};
+    dl->d[dl->n++] = (struct vkp_draw){img, sx, sy, sw, sh, ox, oy, dw, dh,
+                                      surface_translucent(s, dw, dh)};
    s->drawn = 1;
 }

@@ -2032,6 +2079,7 @@ static int on_frame_timer(void *data) {
 static int layer_candidate(const struct draw_list *dl, int scene_w, int scene_h) {
    for (int i = dl->n - 1; i >= 0 && i >= dl->n - 2; i--) {
        const struct vkp_draw *d = &dl->d[i];
+        if (d->blend) continue; /* the translucent plane must not replace the opaque game layer */
        if (!vkp_image_is_dmabuf(d->img)) continue;
        if (d->dx != 0 || d->dy != 0 || d->dw != scene_w || d->dh != scene_h) continue;
        if (d->sx != 0 || d->sy != 0 || (int)d->sw != vkp_image_width(d->img) ||
@@ -2370,7 +2418,7 @@ static void render_scene(void) {
             * its own layer, cropped and placed by the display. */
            int go[8], ov = 0;
            if (over == 1 && li >= 0 && vkp_map_draw(&dl.d[li + 1], go))
-                ov = sc_layer_present_overlay(dl.d[li + 1].img, go) == 0 ? 1 : -1;
+                ov = sc_layer_present_overlay(dl.d[li + 1].img, go, dl.d[li + 1].blend) == 0 ? 1 : -1;
            if (ov <= 0) sc_layer_hide_overlay();
            if (ov < 0) { /* the overlay layer refused it: draw the whole scene the old way */
                sc_layer_hide();
@@ -85,6 +85,7 @@ static struct {

 #define ASC_VISIBILITY_HIDE 0
 #define ASC_VISIBILITY_SHOW 1
+#define ASC_TRANSPARENCY_TRANSLUCENT 1
 #define ASC_TRANSPARENCY_OPAQUE 2

 static int load_api(void) {
@@ -766,7 +767,8 @@ static ASurfaceControl *swap_sc_begin(struct layer *l) {

 /* The transaction that puts pool slot `idx` of layer `l` on screen at `r`, as `color` (NULL = no
 * description). 0 = applied. */
-static int present_slot(struct layer *l, int idx, const int r[8], const struct banner_color *color) {
+static int present_slot(struct layer *l, int idx, const int r[8], const struct banner_color *color,
+                        int translucent) {
    struct slot *s = &l->slots[idx];
    ASurfaceTransaction *tx = api.txCreate();
    if (!tx) return -1;
@@ -774,9 +776,8 @@ static int present_slot(struct layer *l, int idx, const int r[8], const struct b
    if (old) api.setZOrder(tx, l->sc, l->z);
    /* The blit was waited for on the CPU, so no acquire fence is needed (-1). */
    api.setBuffer(tx, l->sc, s->ahb, -1);
-    /* The compositor composes with blits, which overwrite: nothing is ever alpha-blended on the
-     * copy path either, so every layer is opaque and the picture matches it pixel for pixel. */
-    api.setBufferTransparency(tx, l->sc, ASC_TRANSPARENCY_OPAQUE);
+    api.setBufferTransparency(tx, l->sc,
+                              translucent ? ASC_TRANSPARENCY_TRANSLUCENT : ASC_TRANSPARENCY_OPAQUE);
    apply_geometry(tx, l, r);
    apply_frame_rate(tx, l);
    apply_colour(tx, l, color);
@@ -883,7 +884,7 @@ int sc_layer_present(struct vkp_image *src, int scene_w, int scene_h, const stru
    int idx = take_free_slot(l, sw, sh, AHB_RGBA8, &wait_fd);
    if (idx < 0) { log_drop(l); return 0; }
    if (vkp_blit_image(src, l->slots[idx].img, wait_fd) != 0) return -1;
-    if (present_slot(l, idx, r, color) != 0) return -1;
+    if (present_slot(l, idx, r, color, 0) != 0) return -1;
    if (color && color->dataspace) banner_color_frame_shown(color, BANNER_HDR_LAYER_COPY, AHB_RGBA8);
    if (!l->first_logged) {
        l->first_logged = 1;
@@ -908,7 +909,7 @@ int sc_layer_present_pass(const struct vkp_draw *draws, int n, int scene_w, int
    int idx = take_free_slot(l, rw, rh, AHB_RGBA8, &wait_fd);
    if (idx < 0) { vkp_pass_abort(); log_drop(l); return 0; }
    if (vkp_pass_copy_to(l->slots[idx].img, wait_fd) != 0) return -1;
-    if (present_slot(l, idx, r, NULL) != 0) return -1; /* the effects chain's result is 8-bit sRGB */
+    if (present_slot(l, idx, r, NULL, 0) != 0) return -1; /* the effects chain's result is 8-bit sRGB */
    if (!l->first_logged) {
        l->first_logged = 1;
        banner_log("layer", "presenting %dx%d frames on their own SurfaceControl layer (%s pool, %d buffers); "
@@ -950,7 +951,7 @@ int sc_layer_present_hdr_scene(const struct vkp_draw *draws, int n, const struct
    }
    if (idx < 0) { vkp_pass_abort(); log_drop(l); return 0; }
    if (vkp_pass_copy_to(l->slots[idx].img, wait_fd) != 0) return -1;
-    if (present_slot(l, idx, r, tm ? NULL : color) != 0) return -1; /* tone-mapped = plain sRGB: no tag */
+    if (present_slot(l, idx, r, tm ? NULL : color, 0) != 0) return -1; /* tone-mapped = plain sRGB: no tag */
    if (color && color->dataspace)
        banner_color_frame_shown(color, tm ? BANNER_HDR_TONEMAPPED : BANNER_HDR_COMPOSED, tm ? AHB_RGBA8 : g_hdr_pool_fmt);
    static int said = -1;
@@ -1031,7 +1032,7 @@ int sc_layer_overlay_affordable(void) {
    return ok;
 }

-int sc_layer_present_overlay(struct vkp_image *src, const int geo[8]) {
+int sc_layer_present_overlay(struct vkp_image *src, const int geo[8], int translucent) {
    struct layer *l = layer_of(SC_LAYER_OVERLAY);
    if (!src || !geo) return -1;
    /* The caller decides with sc_layer_overlay_affordable() BEFORE it commits the game to a layer;
@@ -1046,7 +1047,7 @@ int sc_layer_present_overlay(struct vkp_image *src, const int geo[8]) {
    /* The window is copied into the layer buffer 1:1; `geo` crops it and places it, so the layer
     * is exactly the window's rectangle on screen and nothing else is blended anywhere. */
    if (vkp_blit_image(src, l->slots[idx].img, wait_fd) != 0) return -1;
-    if (present_slot(l, idx, geo, NULL) != 0) return -1;
+    if (present_slot(l, idx, geo, NULL, translucent) != 0) return -1;
    if (!l->first_logged) {
        l->first_logged = 1;
        log_layer_count();
@@ -150,8 +150,9 @@ int sc_layer_overlay_affordable(void);

 /* OVERLAY layer: show `src` (one window's imported frame) above the game layer, at the placement
 * `geo` = {src x0,y0,x1,y1 in image pixels, dst x0,y0,x1,y1 in output pixels} from vkp_map_draw.
+ * `translucent` preserves a client's alpha over the game for Steam notifications and overlays.
 * 0 = shown or dropped, -1 = unavailable (the caller must fall back to the copy path). */
-int sc_layer_present_overlay(struct vkp_image *src, const int geo[8]);
+int sc_layer_present_overlay(struct vkp_image *src, const int geo[8], int translucent);

 /* The scene is not a single fullscreen window this frame: hide every layer that is up. */
 void sc_layer_hide(void);
@@ -7,6 +7,7 @@
 #include "effects_chain.h"
 #include "framegen_bridge.h"
 #include "hdr_compose.h"
+#include "blend_pass.h"
 #include "banner_color.h"
 #include <stdio.h>
 #include <stdlib.h>
@@ -32,6 +33,8 @@ struct vkp_image {
    void *map;                /* shm images: persistently mapped linear memory */
    VkDeviceSize offset, row_pitch;
    int in_general;           /* shm images: moved from PREINITIALIZED to GENERAL */
+    VkFormat fmt;
+    int sampled;              /* created with SAMPLED usage: the alpha pass can read it */
 };

 static ANativeWindow *g_window;  /* the window frames go to; compositor thread only */
@@ -230,6 +233,7 @@ static void destroy_swapchain(void) {
    if (g_dev_state != 1) return;
    g_vk.DeviceWaitIdle(g_dev);
    reset_sync();
+    for (uint32_t i = 0; i < g_nimg; i++) blendp_forget_image(g_images[i]);
    if (g_swapchain) g_vk.DestroySwapchainKHR(g_dev, g_swapchain, NULL);
    g_swapchain = VK_NULL_HANDLE;
    g_swap_md_identity = 0; /* a new swapchain gets the metadata again */
@@ -525,6 +529,7 @@ static int dev_init(void) {

    vkp_effects_bind_device(g_dev, g_pd, &g_memprops);
    hdrc_bind_device(g_dev, &g_memprops);
+    blendp_bind_device(g_dev);
    g_dev_state = 1;
    reset_sync();
    init_black_image();
@@ -873,9 +878,8 @@ const char *vkp_modifier_name(uint64_t modifier) {
    return other;
 }

-/* Can the driver create the image vkp_image_import_dmabuf() creates for this format+modifier
- * (2D, DRM_FORMAT_MODIFIER tiling, TRANSFER_SRC, dma-buf memory) - and import it? */
-static int modifier_importable(VkFormat fmt, uint64_t modifier) {
+/* Can the driver create and import this format/modifier with the requested image usage? */
+static int modifier_importable(VkFormat fmt, uint64_t modifier, VkImageUsageFlags usage) {
    if (!g_vk.GetPhysicalDeviceImageFormatProperties2) return 1; /* can't ask; the create will tell */
    VkPhysicalDeviceExternalImageFormatInfo ext = {
        .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_IMAGE_FORMAT_INFO,
@@ -886,7 +890,7 @@ static int modifier_importable(VkFormat fmt, uint64_t modifier) {
    VkPhysicalDeviceImageFormatInfo2 info = {
        .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGE_FORMAT_INFO_2, .pNext = &mod,
        .format = fmt, .type = VK_IMAGE_TYPE_2D, .tiling = VK_IMAGE_TILING_DRM_FORMAT_MODIFIER_EXT,
-        .usage = VK_IMAGE_USAGE_TRANSFER_SRC_BIT};
+        .usage = usage};
    VkExternalImageFormatProperties extp = {.sType = VK_STRUCTURE_TYPE_EXTERNAL_IMAGE_FORMAT_PROPERTIES};
    VkImageFormatProperties2 props = {.sType = VK_STRUCTURE_TYPE_IMAGE_FORMAT_PROPERTIES_2, .pNext = &extp};
    if (g_vk.GetPhysicalDeviceImageFormatProperties2(g_pd, &info, &props) != VK_SUCCESS) return 0;
@@ -914,7 +918,7 @@ int vkp_dmabuf_modifiers(uint32_t drm_format, uint64_t *out, int max) {
        if (m != VKP_MOD_LINEAR && m != VKP_MOD_QCOM_COMPRESSED) why = "unknown layout";
        else if (props[i].drmFormatModifierPlaneCount != 1) why = "not single-plane";
        else if (!(props[i].drmFormatModifierTilingFeatures & VK_FORMAT_FEATURE_BLIT_SRC_BIT)) why = "no blit source";
-        else if (!modifier_importable(fmt, m)) why = "not importable as a dma-buf";
+        else if (!modifier_importable(fmt, m, VK_IMAGE_USAGE_TRANSFER_SRC_BIT)) why = "not importable as a dma-buf";
        if (why) {
            LOGI("dmabuf: %c%c%c%c %s reported by the driver but not advertised: %s",
                 drm_format & 0xff, (drm_format >> 8) & 0xff, (drm_format >> 16) & 0xff,
@@ -942,6 +946,10 @@ struct vkp_image *vkp_image_import_dmabuf(int fd, uint32_t drm_format, uint64_t
    struct vkp_image *img = calloc(1, sizeof(*img));
    if (!img) return NULL;
    img->w = w; img->h = h; img->dmabuf = 1; img->blit_dst = as_blit_dst ? 1 : 0;
+    img->fmt = drm_to_vk(drm_format);
+    const VkImageUsageFlags sampled_usage = VK_IMAGE_USAGE_TRANSFER_SRC_BIT | VK_IMAGE_USAGE_SAMPLED_BIT;
+    img->sampled = !as_blit_dst && g_vk.GetPhysicalDeviceImageFormatProperties2 &&
+                   modifier_importable(img->fmt, modifier, sampled_usage);

    VkSubresourceLayout plane = {.offset = offset, .rowPitch = stride};
    VkImageDrmFormatModifierExplicitCreateInfoEXT modInfo = {
@@ -952,10 +960,11 @@ struct vkp_image *vkp_image_import_dmabuf(int fd, uint32_t drm_format, uint64_t
        .handleTypes = VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT};
    VkImageCreateInfo ici = {
        .sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO, .pNext = &extImg,
-        .imageType = VK_IMAGE_TYPE_2D, .format = drm_to_vk(drm_format), .extent = {w, h, 1},
+        .imageType = VK_IMAGE_TYPE_2D, .format = img->fmt, .extent = {w, h, 1},
        .mipLevels = 1, .arrayLayers = 1, .samples = VK_SAMPLE_COUNT_1_BIT,
        .tiling = VK_IMAGE_TILING_DRM_FORMAT_MODIFIER_EXT,
-        .usage = as_blit_dst ? VK_IMAGE_USAGE_TRANSFER_DST_BIT : VK_IMAGE_USAGE_TRANSFER_SRC_BIT,
+        .usage = as_blit_dst ? VK_IMAGE_USAGE_TRANSFER_DST_BIT
+                 : img->sampled ? sampled_usage : VK_IMAGE_USAGE_TRANSFER_SRC_BIT,
        .sharingMode = VK_SHARING_MODE_EXCLUSIVE, .initialLayout = VK_IMAGE_LAYOUT_UNDEFINED};
    VkResult cr = g_vk.CreateImage(g_dev, &ici, NULL, &img->image);
    if (cr != VK_SUCCESS) {
@@ -1008,18 +1017,35 @@ struct vkp_image *vkp_image_import_dmabuf(int fd, uint32_t drm_format, uint64_t
    return img;
 }

+/* Linear wl_shm images can use the blend pass only if the driver samples them with a linear filter. */
+static int shm_sampled(void) {
+    static int known;
+    if (!known) {
+        const VkFormatFeatureFlags need = VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT |
+                                          VK_FORMAT_FEATURE_SAMPLED_IMAGE_FILTER_LINEAR_BIT;
+        VkFormatProperties2 fp = {.sType = VK_STRUCTURE_TYPE_FORMAT_PROPERTIES_2};
+        if (g_vk.GetPhysicalDeviceFormatProperties2)
+            g_vk.GetPhysicalDeviceFormatProperties2(g_pd, VK_FORMAT_B8G8R8A8_UNORM, &fp);
+        known = (fp.formatProperties.linearTilingFeatures & need) == need ? 1 : -1;
+    }
+    return known == 1;
+}
+
 struct vkp_image *vkp_image_create_shm(int w, int h) {
    if (w <= 0 || h <= 0 || dev_init() != 0) return NULL;

    struct vkp_image *img = calloc(1, sizeof(*img));
    if (!img) return NULL;
    img->w = w; img->h = h;
+    img->fmt = VK_FORMAT_B8G8R8A8_UNORM;
+    img->sampled = shm_sampled();

    VkImageCreateInfo ici = {
        .sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO, .imageType = VK_IMAGE_TYPE_2D,
-        .format = VK_FORMAT_B8G8R8A8_UNORM, .extent = {w, h, 1}, .mipLevels = 1, .arrayLayers = 1,
+        .format = img->fmt, .extent = {w, h, 1}, .mipLevels = 1, .arrayLayers = 1,
        .samples = VK_SAMPLE_COUNT_1_BIT, .tiling = VK_IMAGE_TILING_LINEAR,
-        .usage = VK_IMAGE_USAGE_TRANSFER_SRC_BIT, .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
+        .usage = VK_IMAGE_USAGE_TRANSFER_SRC_BIT | (img->sampled ? VK_IMAGE_USAGE_SAMPLED_BIT : 0),
+        .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
        .initialLayout = VK_IMAGE_LAYOUT_PREINITIALIZED};
    if (g_vk.CreateImage(g_dev, &ici, NULL, &img->image) != VK_SUCCESS) { free(img); return NULL; }

@@ -1062,6 +1088,7 @@ int vkp_image_height(const struct vkp_image *img) { return img ? img->h : 0; }
 void vkp_image_destroy(struct vkp_image *img) {
    if (!img) return;
    if (g_dev) {
+        blendp_forget_image(img->image);
        if (img->map) g_vk.UnmapMemory(g_dev, img->mem);
        if (img->image) g_vk.DestroyImage(g_dev, img->image, NULL);
        if (img->mem) g_vk.FreeMemory(g_dev, img->mem, NULL);
@@ -1111,8 +1138,21 @@ static int draw_to_scene_blit(const struct vkp_draw *d, int scene_w, int scene_h
    return map_draw(d, 1.0f, 1.0f, 0.0f, 0.0f, 0.0f, 0.0f, (float)scene_w, (float)scene_h, blit);
 }

+/* A translucent draw needs source-over blending; every other draw keeps the existing blit path. */
+static void record_draw(VkCommandBuffer cmd, const struct vkp_draw *d, const VkImageBlit *blit, VkImage target,
+                        VkFormat target_fmt, int target_w, int target_h, VkFilter filter) {
+    const struct vkp_image *im = d->img;
+    if (d->blend && im->sampled &&
+        blendp_draw(cmd, im->image, im->fmt, im->w, im->h, !im->dmabuf, target, target_fmt, target_w, target_h,
+                    blit) == 0)
+        return;
+    g_vk.CmdBlitImage(cmd, im->image, im->dmabuf ? VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL : VK_IMAGE_LAYOUT_GENERAL,
+                      target, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, 1, blit, filter);
+}
+
 static void destroy_scene_image(void) {
    if (g_scene.img) hdrc_forget_image(g_scene.img); /* the HDR pass may have drawn into it */
+    if (g_scene.img) blendp_forget_image(g_scene.img);
    if (g_scene.img) g_vk.DestroyImage(g_dev, g_scene.img, NULL);
    if (g_scene.mem) g_vk.FreeMemory(g_dev, g_scene.mem, NULL);
    memset(&g_scene, 0, sizeof(g_scene));
@@ -1150,6 +1190,7 @@ static int ensure_scene_image(int w, int h) {
 static int ensure_img(struct vkp_img_slot *s, int w, int h, VkFormat fmt, VkImageUsageFlags usage, const char *what) {
    if (s->img && s->w == w && s->h == h && s->fmt == fmt) return 0;
    if (s->img) hdrc_forget_image(s->img);
+    if (s->img) blendp_forget_image(s->img);
    if (s->img) g_vk.DestroyImage(g_dev, s->img, NULL);
    if (s->mem) g_vk.FreeMemory(g_dev, s->mem, NULL);
    memset(s, 0, sizeof(*s));
@@ -1190,7 +1231,8 @@ static int compose_hdr(VkCommandBuffer cmd, const struct vkp_draw *draws, int n,
                       int scene_w, int scene_h, VkImage out, VkFormat out_fmt, enum hdrc_mode mode) {
    if (!hf || !hf->is_hdr || n <= 0) return -1;
    if (ensure_img(&g_mixed, scene_w, scene_h, HDR_FMT,
-                   VK_IMAGE_USAGE_TRANSFER_DST_BIT | VK_IMAGE_USAGE_SAMPLED_BIT | VK_IMAGE_USAGE_TRANSFER_SRC_BIT,
+                   VK_IMAGE_USAGE_TRANSFER_DST_BIT | VK_IMAGE_USAGE_SAMPLED_BIT | VK_IMAGE_USAGE_TRANSFER_SRC_BIT |
+                       VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT,
                   "HDR composition (mixed)") != 0)
        return -1;
    VkImageSubresourceRange range = {VK_IMAGE_ASPECT_COLOR_BIT, 0, 1, 0, 1};
@@ -1214,9 +1256,7 @@ static int compose_hdr(VkCommandBuffer cmd, const struct vkp_draw *draws, int n,
        VkImageBlit blit;
        if (!draws[i].img || !draw_to_scene_blit(&draws[i], scene_w, scene_h, &blit)) continue;
        /* A blit converts UNORM to UNORM by value: 8-bit sRGB and 10-bit PQ both land unchanged. */
-        g_vk.CmdBlitImage(cmd, draws[i].img->image,
-                          draws[i].img->dmabuf ? VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL : VK_IMAGE_LAYOUT_GENERAL,
-                          g_mixed.img, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, 1, &blit, VK_FILTER_LINEAR);
+        record_draw(cmd, &draws[i], &blit, g_mixed.img, HDR_FMT, scene_w, scene_h, VK_FILTER_LINEAR);
        if (hf->is_hdr[i] && lowest_hdr < 0) lowest_hdr = i;
    }
    static int truncated_said;
@@ -1249,6 +1289,7 @@ static void device_lost(const char *where) {
    vkp_framegen_device_lost();
    banner_log("error", "GPU device lost (VK_ERROR_DEVICE_LOST in %s): the compositor has stopped presenting; "
               "restart the session", where);
+    for (uint32_t i = 0; i < g_nimg; i++) blendp_forget_image(g_images[i]);
    if (g_swapchain) g_vk.DestroySwapchainKHR(g_dev, g_swapchain, NULL);
    g_swapchain = VK_NULL_HANDLE;
    if (g_surface) g_vk.DestroySurfaceKHR(g_inst, g_surface, NULL);
@@ -1547,6 +1588,7 @@ static int render_impl(int scene_w, int scene_h, const struct vkp_draw *draws, i
    VkCommandBufferBeginInfo bi = {.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO,
                                   .flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT};
    g_vk.BeginCommandBuffer(cmd, &bi);
+    blendp_begin_frame();
    VkImageSubresourceRange range = {VK_IMAGE_ASPECT_COLOR_BIT, 0, 1, 0, 1};

    /* Source images: take dmabufs from the client's queue family, move shm images to
@@ -1589,8 +1631,8 @@ static int render_impl(int scene_w, int scene_h, const struct vkp_draw *draws, i
    free(bars);

    /* Clear the composition target to black (the letterbox on the direct path, the desktop's
-     * background on the pass) before the draws land on it. A draw that covers the whole scene
-     * overwrites every pixel of its rectangle (blits never blend), so then only what lies outside it
+     * background on the pass) before the draws land on it. An opaque draw that covers the whole scene
+     * overwrites every pixel of its rectangle, so then only what lies outside it
     * needs black: the letterbox bars on the direct path, nothing at all on the pass (the scene image
     * is exactly the scene). Anything else - no such draw, the HDR composition - clears the whole
     * target as before. */
@@ -1600,7 +1642,8 @@ static int render_impl(int scene_w, int scene_h, const struct vkp_draw *draws, i
    if (!(hf && pass)) {
        for (int i = n - 1; i >= 0 && !covered; i--) {
            const struct vkp_draw *d = &draws[i];
-            if (!d->img || d->dx > 0 || d->dy > 0 || d->dx + d->dw < scene_w || d->dy + d->dh < scene_h) continue;
+            if (!d->img || d->blend || d->dx > 0 || d->dy > 0 || d->dx + d->dw < scene_w ||
+                d->dy + d->dh < scene_h) continue;
            covered = pass ? draw_to_scene_blit(d, scene_w, scene_h, &cover_blit) : draw_to_blit(d, &cover_blit);
        }
    }
@@ -1633,10 +1676,9 @@ static int render_impl(int scene_w, int scene_h, const struct vkp_draw *draws, i
        VkImageBlit blit;
        if (!draws[i].img) continue;
        if (pass ? !draw_to_scene_blit(&draws[i], scene_w, scene_h, &blit) : !draw_to_blit(&draws[i], &blit)) continue;
-        g_vk.CmdBlitImage(cmd, draws[i].img->image,
-                          draws[i].img->dmabuf ? VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL : VK_IMAGE_LAYOUT_GENERAL,
-                          target, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, 1, &blit,
-                          pass ? VK_FILTER_LINEAR : blit_filter);
+        record_draw(cmd, &draws[i], &blit, target, pass ? scene_fmt : g_swap_fmt,
+                    pass ? scene_w : (int)g_extent.width, pass ? scene_h : (int)g_extent.height,
+                    pass ? VK_FILTER_LINEAR : blit_filter);
        drawn++;
    }

@@ -2073,6 +2115,7 @@ static int pass_begin_impl(int scene_w, int scene_h, const struct vkp_draw *draw
    VkCommandBufferBeginInfo bi = {.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO,
                                   .flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT};
    g_vk.BeginCommandBuffer(cmd, &bi);
+    blendp_begin_frame();
    VkImageSubresourceRange range = {VK_IMAGE_ASPECT_COLOR_BIT, 0, 1, 0, 1};

    /* Same acquisition as vkp_render's compositor pass: the sources come from the client's queue
@@ -2132,9 +2175,8 @@ static int pass_begin_impl(int scene_w, int scene_h, const struct vkp_draw *draw
        for (int i = 0; i < n; i++) {
            VkImageBlit blit;
            if (!draws[i].img || !draw_to_scene_blit(&draws[i], scene_w, scene_h, &blit)) continue;
-            g_vk.CmdBlitImage(cmd, draws[i].img->image,
-                              draws[i].img->dmabuf ? VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL : VK_IMAGE_LAYOUT_GENERAL,
-                              g_scene.img, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, 1, &blit, VK_FILTER_LINEAR);
+            record_draw(cmd, &draws[i], &blit, g_scene.img, SCENE_FMT, scene_w, scene_h,
+                        VK_FILTER_LINEAR);
        }
        if (vkp_effects_active()) vkp_effects_set_formats(VK_FORMAT_R8G8B8A8_UNORM, VK_FORMAT_R8G8B8A8_UNORM);
    }
@@ -79,6 +79,7 @@ struct vkp_draw {
    struct vkp_image *img;
    float sx, sy, sw, sh;
    int dx, dy, dw, dh;
+    int blend;                /* translucent: composed over what is under it (blend_pass.h) */
 };

 // The GPU the renderer runs on ("Adreno (TM) 750"), empty before the device is up.