graphics: count DCC/CMASK metadata readbacks in the performance overlay

src/common/debugCounters.h

@@ -29,6 +29,8 @@ enum class Counter {
 	TextureUploadBytes,
 	TextureDownloadBytes,
 	TextureEvictions,
+	ColorMetadataReadbacks, // CPU blocked reading back GPU-written DCC/CMASK metadata.
+	ColorMetadataReadbackNs,
 	BufferUploadBytes,
 	BufferDownloadBytes,
 	BufferEvictions,

src/graphics/host_gpu/renderer/cache/textureCache.cpp

@@ -1070,6 +1070,9 @@ void TextureCache::MaterializeColorClear(ImageId id, const ImageDesc& desc,
 	// Finish native metadata writes before reading backing bytes. This can submit the scheduler,
 	// so discovery runs before final draw uploads and never holds the texture lock across it.
 	if (m_buffer_cache.IsRegionGpuModified(range.address, range.size)) {
+		Common::DebugCounters::ScopedTimer timer(
+		    Common::DebugCounters::Counter::ColorMetadataReadbacks,
+		    Common::DebugCounters::Counter::ColorMetadataReadbackNs);
 		m_buffer_cache.ReadMemory(range.address, range.size, false);
 	}
 	const auto slice_size = range.size / layers;

src/graphics/presentation/performanceOverlay.cpp

@@ -310,6 +310,9 @@ void GpuActivityRows(const Snapshot& shown) {
 	Row("CPU waiting on GPU",
 	    Fmt("%.0f ms/s  (%.0f waits/s)", Rate(shown, Counter::GpuWaitNs) / 1e6,
 	        Rate(shown, Counter::GpuWaits)));
+	Row("DCC/CMASK readbacks",
+	    Fmt("%.0f ms/s  (%.0f/s)", Rate(shown, Counter::ColorMetadataReadbackNs) / 1e6,
+	        Rate(shown, Counter::ColorMetadataReadbacks)));
 }
 
 void CompilationRows(const Snapshot& shown) {

tests/ShaderRecompilerComputeTests.cpp

@@ -1,6 +1,7 @@
 #define SPV_ENABLE_UTILITY_CODE
 
 #include "common/assert.h"
+#include "common/debugCounters.h"
 #include "common/emulatorConfig.h"
 #include "common/file.h"
 #include "common/hostException.h"
@@ -12266,7 +12267,15 @@ public:
         Require(name, "already GPU-owned metadata stays ordered",
                 resources.GetBufferCache().HasGpuDirtyBytes(dcc_address, metadata_size),
                 "a fill discarded pending native writes to its metadata pages");
+        Common::DebugCounters::SetEnabled(true);
+        const auto readbacks =
+            Common::DebugCounters::Get(Common::DebugCounters::Counter::ColorMetadataReadbacks);
         bind();
+        Common::DebugCounters::SetEnabled(false);
+        Require(name, "metadata readback counter",
+                Common::DebugCounters::Get(
+                    Common::DebugCounters::Counter::ColorMetadataReadbacks) == readbacks + 1,
+                "binding GPU-written metadata did not count exactly one readback");
         Require(name, "same-value raw native re-clear",
                 read_texel() == expected,
                 "a repeated metadata fill did not clear an already drawn target");