Merge pull request 'perf(sync): scope frame-loop barriers to swapchain image + real per-pass stages (#115)' (#137) from claude/issue-115 into master

This commit is contained in:
catbot 2026-06-17 21:45:13 +02:00
commit 82fd6916d9
6 changed files with 316 additions and 8 deletions

View file

@ -35,6 +35,13 @@ export namespace Crafter {
RTPass(PipelineRTVulkan* p) : pipeline(p) {}
// An RT pass writes the swapchain image from the ray-tracing pipeline,
// so the frame loop's barriers must wait on RAY_TRACING_SHADER — using
// the compute stage here would under-synchronise and corrupt the image.
VkPipelineStageFlags SwapchainStage() const override {
return VK_PIPELINE_STAGE_RAY_TRACING_SHADER_BIT_KHR;
}
void Record(VkCommandBuffer cmd, std::uint32_t frameIdx, Window& window) override {
vkCmdBindPipeline(cmd, VK_PIPELINE_BIND_POINT_RAY_TRACING_KHR, pipeline->pipeline);
// NVIDIA descriptor-heap AS-read workaround (issue #15 / #7): feed

View file

@ -16,6 +16,10 @@ 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
*/
module;
#ifndef CRAFTER_GRAPHICS_WINDOW_DOM
#include "vulkan/vulkan.h"
#endif // !CRAFTER_GRAPHICS_WINDOW_DOM
export module Crafter.Graphics:RenderPass;
import std;
import :GraphicsTypes;
@ -23,8 +27,74 @@ import :GraphicsTypes;
export namespace Crafter {
struct Window;
#ifndef CRAFTER_GRAPHICS_WINDOW_DOM
// Conservative union of every pipeline stage that can write the swapchain
// storage image: a compute pass via COMPUTE_SHADER, a ray-tracing pass via
// RAY_TRACING_SHADER, plus TRANSFER for any blit/copy writer. Used as the
// RenderPass default (a pass that doesn't narrow its own stage) and as the
// fallback when there are no passes to derive a real union from. Far tighter
// than ALL_COMMANDS, but never under-synchronises a polymorphic pass.
inline constexpr VkPipelineStageFlags kSwapchainWriterStages =
VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT
| VK_PIPELINE_STAGE_RAY_TRACING_SHADER_BIT_KHR
| VK_PIPELINE_STAGE_TRANSFER_BIT;
#endif // !CRAFTER_GRAPHICS_WINDOW_DOM
struct RenderPass {
virtual void Record(GraphicsCommandBuffer cmd, std::uint32_t frameIdx, Window& window) = 0;
virtual ~RenderPass() = default;
#ifndef CRAFTER_GRAPHICS_WINDOW_DOM
// Pipeline stage at which this pass reads and writes the swapchain
// storage image. The frame loop uses this to scope the inter-pass and
// frame-edge barriers to the stages that actually touch the image
// instead of ALL_COMMANDS. A subclass MUST narrow this to its real
// stage (RTPass -> RAY_TRACING_SHADER, a compute pass -> COMPUTE_SHADER):
// the default is the conservative writer union so an un-overridden pass
// can never be under-synchronised, only over-synchronised.
virtual VkPipelineStageFlags SwapchainStage() const { return kSwapchainWriterStages; }
#endif // !CRAFTER_GRAPHICS_WINDOW_DOM
};
#ifndef CRAFTER_GRAPHICS_WINDOW_DOM
// Union of the swapchain-access stages declared by `passes` — the "real"
// per-pass stage union the frame-edge barriers (acquire dst / present src)
// narrow to: an all-compute frame yields COMPUTE_SHADER only, a frame with
// any RT pass folds in RAY_TRACING_SHADER, and so on. Falls back to the
// conservative writer union when there are no passes (the image is still
// transitioned, just never written), so the barrier is always well-formed.
inline VkPipelineStageFlags SwapchainStageUnion(std::span<RenderPass* const> passes) {
if (passes.empty()) return kSwapchainWriterStages;
VkPipelineStageFlags stages = 0;
for (RenderPass* pass : passes) stages |= pass->SwapchainStage();
return stages;
}
// The inter-pass swapchain barrier (replacing the old queue-wide
// VkMemoryBarrier at the #115 location). Scoped to the swapchain image's
// single colour subresource so only that image's shader caches round-trip —
// unrelated buffers/images stay resident — exactly as the intra-pass UI
// barrier already does. The image is bound as a storage image and stays in
// VK_IMAGE_LAYOUT_GENERAL, so this is a pure memory dependency (no layout
// transition). The execution scope (stage masks) is supplied per pass pair
// by the caller via vkCmdPipelineBarrier.
inline VkImageMemoryBarrier BuildSwapchainInterPassBarrier(VkImage image) {
return {
.sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
.srcAccessMask = VK_ACCESS_SHADER_WRITE_BIT,
.dstAccessMask = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT,
.oldLayout = VK_IMAGE_LAYOUT_GENERAL,
.newLayout = VK_IMAGE_LAYOUT_GENERAL,
.srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
.dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
.image = image,
.subresourceRange = {
.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
.baseMipLevel = 0,
.levelCount = 1,
.baseArrayLayer = 0,
.layerCount = 1,
},
};
}
#endif // !CRAFTER_GRAPHICS_WINDOW_DOM
}

View file

@ -201,6 +201,15 @@ export namespace Crafter {
void Record(GraphicsCommandBuffer cmd, std::uint32_t frameIdx, Window& window) override;
#ifndef CRAFTER_GRAPHICS_WINDOW_DOM
// A UI pass draws into the swapchain image entirely from compute
// shaders (see Dispatch), so the frame loop's inter-pass / frame-edge
// barriers only need to synchronise the COMPUTE_SHADER stage for it.
VkPipelineStageFlags SwapchainStage() const override {
return VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT;
}
#endif // !CRAFTER_GRAPHICS_WINDOW_DOM
UIDispatchHeader FillHeader(std::uint32_t itemBufferSlot,
std::uint32_t itemCount,
std::array<float,4> clipRectPx = {0.0f, 0.0f, 1e9f, 1e9f},