Merge "Implement functions for managing Graphite memory budget." into main

Author: egdaniel

libs/renderengine/skia/compat/GraphiteGpuContext.cpp

@@ -110,8 +110,40 @@ bool GraphiteGpuContext::isAbandonedOrDeviceLost() {
     return mContext->isDeviceLost();
 }
 
+void GraphiteGpuContext::setResourceCacheLimit(size_t maxResourceBytes) {
+    // Graphite has a separate budget for its Context and its Recorder. For now the majority of
+    // memory that Graphite will allocate will be on the Recorder and minimal amount on the Context.
+    // The main allocations on the Context are MSAA buffers (not often, if ever used in
+    // RenderEngine) and stencil buffers. However, both of these should be "memoryless" in Vulkan on
+    // tiled GPUs, so they don't actually use GPU memory. However, in Vulkan there are scenarios
+    // where Vulkan could end up using real memory for them. Skia will regularly query the device to
+    // get the real memory usage and update the budgeted appropriately. Though for all real usage
+    // patterns we don't expect to ever trigger the device to allocate real memory.
+    //
+    // Therefore, we set the full maxResourceBytes budget on the Recorder. However, in the rare
+    // chance that the devcies does allocate real memory we don't want to immediately kill device
+    // performance by constantly trashing allocations on the Context. Thus we set the Context's
+    // budget to be 50% of the total budget to make sure we allow the MSAA or Stencil buffers to be
+    // allocated in Skia and not immediately discarded. But even with this extra 50% budget, as
+    // described above, this shouldn't result in actual GPU memory usage.
+    //
+    // TODO: We will need to revise this strategy for GLES which does not have the same memoryless
+    // textures.
+    // TODO: Work in Graphite has started to move a lot more of its scratch resources to be owned
+    // by the Context and not on Recorders. This will mean most memory is actually owned by the
+    // Context and thus the budgeting here will need to be updated.
+    mContext->setMaxBudgetedBytes(maxResourceBytes / 2);
+    mRecorder->setMaxBudgetedBytes(maxResourceBytes);
+}
+
+void GraphiteGpuContext::purgeUnlockedScratchResources() {
+    mContext->freeGpuResources();
+    mRecorder->freeGpuResources();
+}
+
 void GraphiteGpuContext::dumpMemoryStatistics(SkTraceMemoryDump* traceMemoryDump) const {
     mContext->dumpMemoryStatistics(traceMemoryDump);
+    mRecorder->dumpMemoryStatistics(traceMemoryDump);
 }
 
 } // namespace android::renderengine::skia

libs/renderengine/skia/compat/GraphiteGpuContext.h

@@ -39,16 +39,10 @@ class GraphiteGpuContext : public SkiaGpuContext {
     size_t getMaxRenderTargetSize() const override;
     size_t getMaxTextureSize() const override;
     bool isAbandonedOrDeviceLost() override;
-    // No-op (large resources like textures, surfaces, images, etc. created by clients don't count
-    // towards Graphite's internal caching budgets, so adjusting its limits based on display change
-    // events should be unnecessary. Additionally, Graphite doesn't expose many cache tweaking
-    // functions yet, as its design may evolve.)
-    void setResourceCacheLimit(size_t maxResourceBytes) 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).
-    void purgeUnlockedScratchResources() override{};
+    void setResourceCacheLimit(size_t maxResourceBytes) override;
+    void purgeUnlockedScratchResources() override;
+
     // No-op (only applicable to GL).
     void resetContextIfApplicable() override{};