perf(sync): scope frame-loop barriers to swapchain image + real per-pass stages (#115)

The inter-pass and acquire/present barriers in the frame loop set both
stage masks to ALL_COMMANDS, and the inter-pass dependency used a
queue-wide VkMemoryBarrier — fully serialising against every pipeline
stage and flushing all caches every frame, when all the next pass needs
is the swapchain image the previous one wrote.

Replace the inter-pass global VkMemoryBarrier with an image memory
barrier scoped to the swapchain image's single colour subresource (as
the intra-pass UI barrier already does), and derive the barrier stage
masks per pass: RenderPass::SwapchainStage() is overridden by UIRenderer
(COMPUTE_SHADER) and RTPass (RAY_TRACING_SHADER), so a compute->compute
edge only serialises COMPUTE while an RT pass pulls in RAY_TRACING — the
acquire/present frame-edge masks use the real union of the frame's
passes (SwapchainStageUnion). The base default and the empty-passes
fallback are the conservative COMPUTE | RAY_TRACING | TRANSFER union, so
a polymorphic or un-overridden pass can only be over- not
under-synchronised.

Adds SwapchainBarrierScope (pure CPU) pinning the per-pass derivation,
the union narrowing, and the inter-pass barrier scope; FrameLoopSync
already drives the real GPU frame loop with validation enabled.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

tests/SwapchainBarrierScope/main.cpp

@@ -0,0 +1,174 @@
+/*
+Crafter®.Graphics
+Copyright (C) 2026 Catcrafts®
+catcrafts.net
+
+This library is free software; you can redistribute it and/or
+modify it under the terms of the GNU Lesser General Public
+License version 3.0 as published by the Free Software Foundation;
+
+This library is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+Lesser General Public License for more details.
+
+You should have received a copy of the GNU Lesser General Public
+License along with this library; if not, write to the Free Software
+Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301 USA
+*/
+
+// Issue #115: the frame loop's inter-pass and frame-edge (acquire/present)
+// barriers used to set BOTH stage masks to VK_PIPELINE_STAGE_ALL_COMMANDS_BIT
+// and the inter-pass dependency to a queue-wide VkMemoryBarrier — flushing
+// every cache and fully serialising against the whole pipeline every frame,
+// when all the next pass needs is the swapchain image the previous one wrote.
+//
+// The fix scopes the inter-pass dependency to the swapchain image's single
+// colour subresource (BuildSwapchainInterPassBarrier) and narrows the stage
+// masks to the stages that actually touch that image, derived per pass via the
+// polymorphic RenderPass::SwapchainStage() and unioned across the frame by
+// SwapchainStageUnion. The load-bearing correctness constraint is that this
+// derivation is per-pass: a compute pass writes via COMPUTE_SHADER, a ray-
+// tracing pass via RAY_TRACING_SHADER, so a hardcoded compute mask would
+// UNDER-synchronise an RT pass and corrupt the image. All three helpers are
+// pure CPU logic over the pass list, so this test drives them directly with no
+// GPU device at runtime — mirroring MipChainBarrierBatch / UploadStrategy.
+
+#include <cstdlib>
+#include "vulkan/vulkan.h"
+
+import Crafter.Graphics;
+import std;
+using namespace Crafter;
+
+namespace {
+
+int failures = 0;
+
+void Check(bool ok, std::string_view what) {
+    std::println("{}  {}", ok ? "PASS" : "FAIL", what);
+    if (!ok) ++failures;
+}
+
+// A pass that never overrides SwapchainStage(): stands in for any future pass
+// that forgets to narrow its stage, and must fall back to the conservative
+// writer union rather than silently picking one stage (which could under-sync).
+struct DefaultPass : RenderPass {
+    void Record(GraphicsCommandBuffer, std::uint32_t, Window&) override {}
+};
+
+VkImage SentinelImage() {
+    return reinterpret_cast<VkImage>(static_cast<std::uintptr_t>(0xC0FFEE));
+}
+
+} // namespace
+
+int main() {
+    // ─── per-pass SwapchainStage() polymorphism ─────────────────────────────
+    // The whole reason the fix is "moderate, not trivial": these two stages
+    // MUST differ, or scoping an RT pass's barrier to the compute stage would
+    // under-synchronise it. Pin that they are distinct and each correct.
+    Check(VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT != VK_PIPELINE_STAGE_RAY_TRACING_SHADER_BIT_KHR,
+          "compute and ray-tracing write stages are distinct (the under-sync trap)");
+
+    UIRenderer ui;                 // real compute UI pass
+    RTPass rt(nullptr);            // real RT pass (SwapchainStage doesn't touch the pipeline)
+    DefaultPass def;               // un-overridden fallback
+
+    // Access through the base pointer — the frame loop only ever sees RenderPass*.
+    RenderPass* uiBase  = &ui;
+    RenderPass* rtBase  = &rt;
+    RenderPass* defBase = &def;
+
+    Check(uiBase->SwapchainStage() == VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
+          "UIRenderer reports the COMPUTE_SHADER stage");
+    Check(rtBase->SwapchainStage() == VK_PIPELINE_STAGE_RAY_TRACING_SHADER_BIT_KHR,
+          "RTPass reports the RAY_TRACING_SHADER stage");
+    Check(defBase->SwapchainStage() == kSwapchainWriterStages,
+          "an un-overridden pass falls back to the conservative writer union");
+
+    // ─── the conservative writer union constant ─────────────────────────────
+    constexpr VkPipelineStageFlags expectedUnion =
+        VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT
+      | VK_PIPELINE_STAGE_RAY_TRACING_SHADER_BIT_KHR
+      | VK_PIPELINE_STAGE_TRANSFER_BIT;
+    Check(kSwapchainWriterStages == expectedUnion,
+          "writer union = COMPUTE_SHADER | RAY_TRACING_SHADER | TRANSFER");
+    // The point of the whole change: even the conservative fallback is far
+    // tighter than the ALL_COMMANDS mask it replaces.
+    Check(kSwapchainWriterStages != VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
+          "writer union is narrower than ALL_COMMANDS");
+
+    // ─── SwapchainStageUnion over the frame's pass list ─────────────────────
+    // Empty frame: no writers to derive from, so fall back to the conservative
+    // union (the image is still transitioned, just never written).
+    Check(SwapchainStageUnion(std::span<RenderPass* const>{}) == kSwapchainWriterStages,
+          "empty pass list falls back to the conservative writer union");
+
+    // All-compute frame: the union narrows to COMPUTE_SHADER ONLY — this is the
+    // real perf win, no RAY_TRACING / TRANSFER / ALL_COMMANDS dragged in.
+    {
+        std::array<RenderPass*, 2> computeOnly = {uiBase, uiBase};
+        VkPipelineStageFlags u = SwapchainStageUnion(computeOnly);
+        Check(u == VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
+              "all-compute frame narrows the union to COMPUTE_SHADER only");
+        Check((u & VK_PIPELINE_STAGE_RAY_TRACING_SHADER_BIT_KHR) == 0,
+              "all-compute frame does NOT pull in RAY_TRACING_SHADER");
+    }
+
+    // RT-only frame: narrows to RAY_TRACING_SHADER only.
+    {
+        std::array<RenderPass*, 1> rtOnly = {rtBase};
+        Check(SwapchainStageUnion(rtOnly) == VK_PIPELINE_STAGE_RAY_TRACING_SHADER_BIT_KHR,
+              "RT-only frame narrows the union to RAY_TRACING_SHADER only");
+    }
+
+    // Mixed frame: the union folds in BOTH stages — never under-syncs the RT
+    // pass, never over-syncs to ALL_COMMANDS.
+    {
+        std::array<RenderPass*, 2> mixed = {uiBase, rtBase};
+        VkPipelineStageFlags u = SwapchainStageUnion(mixed);
+        Check(u == (VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT | VK_PIPELINE_STAGE_RAY_TRACING_SHADER_BIT_KHR),
+              "mixed compute+RT frame unions COMPUTE_SHADER | RAY_TRACING_SHADER");
+        Check((u & VK_PIPELINE_STAGE_RAY_TRACING_SHADER_BIT_KHR) != 0,
+              "mixed frame keeps RAY_TRACING_SHADER (no RT under-sync)");
+    }
+
+    // A frame containing an un-overridden pass conservatively widens to the
+    // full writer union — safe by construction.
+    {
+        std::array<RenderPass*, 2> withDefault = {uiBase, defBase};
+        Check(SwapchainStageUnion(withDefault) == kSwapchainWriterStages,
+              "a frame with an un-overridden pass widens to the conservative union");
+    }
+
+    // ─── BuildSwapchainInterPassBarrier ─────────────────────────────────────
+    // Replaces the queue-wide VkMemoryBarrier: an image memory barrier scoped
+    // to the swapchain image's single colour subresource, GENERAL->GENERAL (no
+    // layout change — the storage image stays bound), matching the intra-pass
+    // UI barrier's access scope.
+    {
+        VkImageMemoryBarrier b = BuildSwapchainInterPassBarrier(SentinelImage());
+        Check(b.sType == VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER, "inter-pass barrier sType is IMAGE_MEMORY_BARRIER");
+        Check(b.image == SentinelImage(), "inter-pass barrier carries the swapchain image handle");
+        Check(b.oldLayout == VK_IMAGE_LAYOUT_GENERAL && b.newLayout == VK_IMAGE_LAYOUT_GENERAL,
+              "inter-pass barrier is GENERAL->GENERAL (pure memory dependency, no transition)");
+        Check(b.srcAccessMask == VK_ACCESS_SHADER_WRITE_BIT,
+              "inter-pass barrier src access is shader write");
+        Check(b.dstAccessMask == (VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT),
+              "inter-pass barrier dst access is shader read|write");
+        Check(b.srcQueueFamilyIndex == VK_QUEUE_FAMILY_IGNORED && b.dstQueueFamilyIndex == VK_QUEUE_FAMILY_IGNORED,
+              "inter-pass barrier is not a queue-family transfer");
+        Check(b.subresourceRange.aspectMask == VK_IMAGE_ASPECT_COLOR_BIT,
+              "inter-pass barrier covers the colour aspect");
+        Check(b.subresourceRange.levelCount == 1 && b.subresourceRange.layerCount == 1,
+              "inter-pass barrier is scoped to the single swapchain subresource");
+    }
+
+    if (failures != 0) {
+        std::println("{} check(s) failed", failures);
+        return EXIT_FAILURE;
+    }
+    std::println("all checks passed");
+    return EXIT_SUCCESS;
+}