Fix RE's VulkanInterface destruction & unnecessary initialization
SurfaceFlinger's initialization of RE now:
- Only attempts to check for VK support if either the GaneshVk or
GraphiteVk flag is set.
- Caches the VulkanInterface created to check for VK support until a VK
instance of RE is created.
- Tears down a partially initialized VulkanInterface if some required
feature is unsupported.
Additionally, SkiaVkRenderEngine's destructor now tears down the static
VulkanInterfaces it uses, which necessitates:
- Caching whether VK is supported.
- Ensuring all Skia resources are destroyed *before* the VK resources
managed by VulkanInterface that they rely on are torn down. This
involves ensuring all textures are destroyed, and all Skia
context-like objects are destroyed.
This latter change means that tests in librenderengine_test that are
parameterized by Skia backend must now recreate RE's VulkanInterfaces
twice for each test case (once for GaneshVk, and again for GraphiteVk),
which results in a minor regression in test duration. However, this is
necessary because holding on to a VulkanInterface while attempting to
set up a test for GaneshGL will cause contention over the real-time
GPU context priority, which can only be held exclusively by either a GL
context OR a VK context on some hardware.
Many thanks to joseph.cheng@imgtec.com for raising the issue of SF's
init of RE not tearing down the VulkanInterface used to check for
support, and for proposing an initial patch (b/333477752) which this
change builds upon. And thanks to scroggo@google.com for proposing the
reordering of SF's flag vs. VK support checks.
Bug: b/293371537
Bug: b/333477752
Test: librenderengine_test && manual boot validation across 3 backends
Flag: com.android.graphics.surfaceflinger.flags.graphite_renderengine-READ-ONLY
Change-Id: I289c3f7699d16707d1462179f4d5e8c54e4bb049
diff --git a/libs/renderengine/skia/compat/GraphiteGpuContext.h b/libs/renderengine/skia/compat/GraphiteGpuContext.h
index 685f899..413817f 100644
--- a/libs/renderengine/skia/compat/GraphiteGpuContext.h
+++ b/libs/renderengine/skia/compat/GraphiteGpuContext.h
@@ -26,7 +26,7 @@
class GraphiteGpuContext : public SkiaGpuContext {
public:
GraphiteGpuContext(std::unique_ptr<skgpu::graphite::Context> context);
- ~GraphiteGpuContext() override = default;
+ ~GraphiteGpuContext() override;
std::shared_ptr<skgpu::graphite::Context> graphiteContext() override;
std::shared_ptr<skgpu::graphite::Recorder> graphiteRecorder() override;
@@ -45,7 +45,6 @@
// functions yet, as its design may evolve.)
void setResourceCacheLimit(size_t maxResourceBytes) override{};
- void finishRenderingAndAbandonContext() override;
// TODO: b/293371537 - Triple-check and validate that no cleanup is necessary when switching
// contexts.
// No-op (unnecessary during context switch for Graphite's client-budgeted memory model).
@@ -58,7 +57,7 @@
private:
DISALLOW_COPY_AND_ASSIGN(GraphiteGpuContext);
- const std::shared_ptr<skgpu::graphite::Context> mContext;
+ std::shared_ptr<skgpu::graphite::Context> mContext;
std::shared_ptr<skgpu::graphite::Recorder> mRecorder;
};