Metal: track buffer allocations (#7556)

filament/backend/include/backend/Platform.h

@@ -193,9 +193,40 @@ public:
     size_t retrieveBlob(const void* UTILS_NONNULL key, size_t keySize,
             void* UTILS_NONNULL value, size_t valueSize);
 
+    using DebugUpdateStatFunc = utils::Invocable<void(const char* UTILS_NONNULL key, uint64_t value)>;
+
+    /**
+     * Sets the callback function that the backend can use to update backend-specific statistics
+     * to aid with debugging. This callback is guaranteed to be called on the Filament driver
+     * thread.
+     *
+     * @param debugUpdateStat   an Invocable that updates debug statistics
+     */
+    void setDebugUpdateStatFunc(DebugUpdateStatFunc&& debugUpdateStat) noexcept;
+
+    /**
+     * @return true if debugUpdateStat is valid.
+     */
+    bool hasDebugUpdateStatFunc() const noexcept;
+
+    /**
+     * To track backend-specific statistics, the backend implementation can call the
+     * application-provided callback function debugUpdateStatFunc to associate or update a value
+     * with a given key. It is possible for this function to be called multiple times with the
+     * same key, in which case newer values should overwrite older values.
+     *
+     * This function is guaranteed to be called only on a single thread, the Filament driver
+     * thread.
+     *
+     * @param key          a null-terminated C-string with the key of the debug statistic
+     * @param value        the updated value of key
+     */
+    void debugUpdateStat(const char* UTILS_NONNULL key, uint64_t value);
+
 private:
     InsertBlobFunc mInsertBlob;
     RetrieveBlobFunc mRetrieveBlob;
+    DebugUpdateStatFunc mDebugUpdateStat;
 };
 
 } // namespace filament

filament/backend/src/Platform.cpp

@@ -53,4 +53,18 @@ size_t Platform::retrieveBlob(void const* key, size_t keySize, void* value, size
     return 0;
 }
 
+void Platform::setDebugUpdateStatFunc(DebugUpdateStatFunc&& debugUpdateStat) noexcept {
+    mDebugUpdateStat = std::move(debugUpdateStat);
+}
+
+bool Platform::hasDebugUpdateStatFunc() const noexcept {
+    return bool(mDebugUpdateStat);
+}
+
+void Platform::debugUpdateStat(const char* key, uint64_t value) {
+    if (mDebugUpdateStat) {
+        mDebugUpdateStat(key, value);
+    }
+}
+
 } // namespace filament::backend

filament/backend/src/metal/MetalBuffer.h

@@ -18,7 +18,6 @@
 #define TNT_FILAMENT_DRIVER_METALBUFFER_H
 
 #include "MetalContext.h"
-#include "MetalBufferPool.h"
 
 #include <backend/DriverEnums.h>
 
@@ -28,9 +27,50 @@
 
 #include <utility>
 #include <memory>
+#include <atomic>
 
 namespace filament::backend {
 
+class TrackedMetalBuffer {
+public:
+    TrackedMetalBuffer() noexcept : mBuffer(nil) {}
+    TrackedMetalBuffer(id<MTLBuffer> buffer) noexcept : mBuffer(buffer) {
+        if (buffer) {
+            aliveBuffers++;
+        }
+    }
+    ~TrackedMetalBuffer() {
+        if (mBuffer) {
+            aliveBuffers--;
+        }
+    }
+
+    TrackedMetalBuffer(TrackedMetalBuffer&&) = delete;
+    TrackedMetalBuffer(TrackedMetalBuffer const&) = delete;
+    TrackedMetalBuffer& operator=(TrackedMetalBuffer const&) = delete;
+
+    TrackedMetalBuffer& operator=(TrackedMetalBuffer&& rhs) noexcept {
+        swap(rhs);
+        return *this;
+    }
+
+    id<MTLBuffer> get() const noexcept { return mBuffer; }
+    operator bool() const noexcept { return bool(mBuffer); }
+
+    static uint64_t getAliveBuffers() { return aliveBuffers; }
+
+private:
+    void swap(TrackedMetalBuffer& other) noexcept {
+        id<MTLBuffer> temp = mBuffer;
+        mBuffer = other.mBuffer;
+        other.mBuffer = temp;
+    }
+
+    id<MTLBuffer> mBuffer;
+
+    static std::atomic<uint64_t> aliveBuffers;
+};
+
 class MetalBuffer {
 public:
 
@@ -82,7 +122,7 @@ public:
 
 private:
 
-    id<MTLBuffer> mBuffer = nil;
+    TrackedMetalBuffer mBuffer;
     size_t mBufferSize = 0;
     void* mCpuBuffer = nullptr;
     MetalContext& mContext;
@@ -151,7 +191,7 @@ public:
             // finishes executing.
             mAuxBuffer = [mDevice newBufferWithLength:mSlotSizeBytes options:mBufferOptions];
             assert_invariant(mAuxBuffer);
-            return {mAuxBuffer, 0};
+            return {mAuxBuffer.get(), 0};
         }
         mCurrentSlot = (mCurrentSlot + 1) % mSlotCount;
         mOccupiedSlots->fetch_add(1, std::memory_order_relaxed);
@@ -180,9 +220,9 @@ public:
      */
     std::pair<id<MTLBuffer>, NSUInteger> getCurrentAllocation() const {
         if (UTILS_UNLIKELY(mAuxBuffer)) {
-            return { mAuxBuffer, 0 };
+            return { mAuxBuffer.get(), 0 };
         }
-        return { mBuffer, mCurrentSlot * mSlotSizeBytes };
+        return { mBuffer.get(), mCurrentSlot * mSlotSizeBytes };
     }
 
     bool canAccomodateLayout(MTLSizeAndAlign layout) const {
@@ -191,8 +231,8 @@ public:
 
 private:
     id<MTLDevice> mDevice;
-    id<MTLBuffer> mBuffer;
-    id<MTLBuffer> mAuxBuffer;
+    TrackedMetalBuffer mBuffer;
+    TrackedMetalBuffer mAuxBuffer;
 
     MTLResourceOptions mBufferOptions;
 

filament/backend/src/metal/MetalBuffer.mm

@@ -15,12 +15,15 @@
  */
 
 #include "MetalBuffer.h"
+#include "MetalBufferPool.h"
 
 #include "MetalContext.h"
 
 namespace filament {
 namespace backend {
 
+std::atomic<uint64_t> TrackedMetalBuffer::aliveBuffers = 0;
+
 MetalBuffer::MetalBuffer(MetalContext& context, BufferObjectBinding bindingType, BufferUsage usage,
         size_t size, bool forceGpuBuffer) : mBufferSize(size), mContext(context) {
     // If the buffer is less than 4K in size and is updated frequently, we don't use an explicit
@@ -61,7 +64,7 @@ void MetalBuffer::copyIntoBuffer(void* src, size_t size, size_t byteOffset) {
     // Acquire a staging buffer to hold the contents of this update.
     MetalBufferPool* bufferPool = mContext.bufferPool;
     const MetalBufferPoolEntry* const staging = bufferPool->acquireBuffer(size);
-    memcpy(staging->buffer.contents, src, size);
+    memcpy(staging->buffer.get().contents, src, size);
 
     // The blit below requires that byteOffset be a multiple of 4.
     ASSERT_PRECONDITION(!(byteOffset & 0x3u), "byteOffset must be a multiple of 4");
@@ -70,9 +73,9 @@ void MetalBuffer::copyIntoBuffer(void* src, size_t size, size_t byteOffset) {
     id<MTLCommandBuffer> cmdBuffer = getPendingCommandBuffer(&mContext);
     id<MTLBlitCommandEncoder> blitEncoder = [cmdBuffer blitCommandEncoder];
     blitEncoder.label = @"Buffer upload blit";
-    [blitEncoder copyFromBuffer:staging->buffer
+    [blitEncoder copyFromBuffer:staging->buffer.get()
                    sourceOffset:0
-                       toBuffer:mBuffer
+                       toBuffer:mBuffer.get()
               destinationOffset:byteOffset
                            size:size];
     [blitEncoder endEncoding];
@@ -93,7 +96,7 @@ id<MTLBuffer> MetalBuffer::getGpuBufferForDraw(id<MTLCommandBuffer> cmdBuffer) n
         return nil;
     }
     assert_invariant(mBuffer);
-    return mBuffer;
+    return mBuffer.get();
 }
 
 void MetalBuffer::bindBuffers(id<MTLCommandBuffer> cmdBuffer, id<MTLCommandEncoder> encoder,

filament/backend/src/metal/MetalBufferPool.h

@@ -19,6 +19,8 @@
 
 #include <Metal/Metal.h>
 
+#include "MetalBuffer.h"
+
 #include <map>
 #include <mutex>
 #include <unordered_set>
@@ -30,7 +32,7 @@ struct MetalContext;
 
 // Immutable POD representing a shared CPU-GPU buffer.
 struct MetalBufferPoolEntry {
-    id<MTLBuffer> buffer;
+    TrackedMetalBuffer buffer;
     size_t capacity;
     mutable uint64_t lastAccessed;
     mutable uint32_t referenceCount;

filament/backend/src/metal/MetalBufferPool.mm

@@ -45,12 +45,12 @@ MetalBufferPoolEntry const* MetalBufferPool::acquireBuffer(size_t numBytes) {
     id<MTLBuffer> buffer = [mContext.device newBufferWithLength:numBytes
                                                         options:MTLResourceStorageModeShared];
     ASSERT_POSTCONDITION(buffer, "Could not allocate Metal staging buffer of size %zu.", numBytes);
-    MetalBufferPoolEntry* stage = new MetalBufferPoolEntry({
+    MetalBufferPoolEntry* stage = new MetalBufferPoolEntry {
         .buffer = buffer,
         .capacity = numBytes,
         .lastAccessed = mCurrentFrame,
         .referenceCount = 1
-    });
+    };
     mUsedStages.insert(stage);
 
     return stage;

filament/backend/src/metal/MetalDriver.mm

@@ -20,6 +20,7 @@
 #include "metal/MetalDriver.h"
 
 #include "MetalBlitter.h"
+#include "MetalBufferPool.h"
 #include "MetalContext.h"
 #include "MetalDriverFactory.h"
 #include "MetalEnums.h"
@@ -36,6 +37,7 @@
 
 #include <utils/Log.h>
 #include <utils/Panic.h>
+#include <utils/sstream.h>
 
 #include <algorithm>
 
@@ -214,6 +216,9 @@ void MetalDriver::beginFrame(int64_t monotonic_clock_ns, uint32_t frameId) {
 #if defined(FILAMENT_METAL_PROFILING)
     os_signpost_interval_begin(mContext->log, mContext->signpostId, "Frame encoding", "%{public}d", frameId);
 #endif
+    if (mPlatform.hasDebugUpdateStatFunc()) {
+        mPlatform.debugUpdateStat("filament.metal.alive_buffers", TrackedMetalBuffer::getAliveBuffers());
+    }
 }
 
 void MetalDriver::setFrameScheduledCallback(Handle<HwSwapChain> sch,

filament/backend/src/metal/MetalHandles.mm

@@ -19,6 +19,7 @@
 #include "MetalBlitter.h"
 #include "MetalEnums.h"
 #include "MetalUtils.h"
+#include "MetalBufferPool.h"
 
 #include <filament/SwapChain.h>
 
@@ -770,13 +771,13 @@ void MetalTexture::loadWithCopyBuffer(uint32_t level, uint32_t slice, MTLRegion
         PixelBufferDescriptor const& data, const PixelBufferShape& shape) {
     const size_t stagingBufferSize = shape.totalBytes;
     auto entry = context.bufferPool->acquireBuffer(stagingBufferSize);
-    memcpy(entry->buffer.contents,
+    memcpy(entry->buffer.get().contents,
             static_cast<uint8_t*>(data.buffer) + shape.sourceOffset,
             stagingBufferSize);
     id<MTLCommandBuffer> blitCommandBuffer = getPendingCommandBuffer(&context);
     id<MTLBlitCommandEncoder> blitCommandEncoder = [blitCommandBuffer blitCommandEncoder];
     blitCommandEncoder.label = @"Texture upload buffer blit";
-    [blitCommandEncoder copyFromBuffer:entry->buffer
+    [blitCommandEncoder copyFromBuffer:entry->buffer.get()
                           sourceOffset:0
                      sourceBytesPerRow:shape.bytesPerRow
                    sourceBytesPerImage:shape.bytesPerSlice

libs/filamentapp/src/FilamentApp.cpp

@@ -674,6 +674,7 @@ FilamentApp::Window::Window(FilamentApp* filamentApp,
         mSwapChain = mFilamentApp->mEngine->createSwapChain(
                 nativeSwapChain, filament::SwapChain::CONFIG_HAS_STENCIL_BUFFER);
     }
+
     mRenderer = mFilamentApp->mEngine->createRenderer();
 
     // create cameras

libs/utils/include/utils/ostream.h

@@ -92,6 +92,7 @@ protected:
         std::pair<char*, size_t> grow(size_t s) noexcept;
         void advance(ssize_t n) noexcept;
         void reset() noexcept;
+        size_t length() const noexcept;
 
     private:
         void reserve(size_t newSize) noexcept;

libs/utils/include/utils/sstream.h

@@ -25,6 +25,7 @@ class sstream : public ostream {
 public:
     ostream& flush() noexcept override;
     const char* c_str() const noexcept;
+    size_t length() const noexcept;
 };
 
 } // namespace utils::io

libs/utils/src/ostream.cpp

@@ -267,6 +267,10 @@ void ostream::Buffer::reset() noexcept {
     size = capacity;
 }
 
+size_t ostream::Buffer::length() const noexcept {
+    return curr - buffer;
+}
+
 std::pair<char*, size_t> ostream::Buffer::grow(size_t s) noexcept {
     if (UTILS_UNLIKELY(size < s)) {
         size_t const used = curr - buffer;

libs/utils/src/sstream.cpp

@@ -28,4 +28,8 @@ const char* sstream::c_str() const noexcept {
     return buffer ? buffer : "";
 }
 
+size_t sstream::length() const noexcept {
+    return getBuffer().length();
+}
+
 } // namespace utils::io

libs/utils/test/test_sstream.cpp

@@ -144,6 +144,7 @@ TEST(sstream, LargeBuffer) {
     }
 
     EXPECT_EQ(1024 * 1024 * 16, strlen(ss.c_str()));
+    EXPECT_EQ(1024 * 1024 * 16, ss.length());
 }
 
 TEST(sstream, LargeString) {
@@ -158,6 +159,7 @@ TEST(sstream, LargeString) {
     ss << filler;
 
     EXPECT_EQ(size, strlen(ss.c_str()));
+    EXPECT_EQ(size, ss.length());
     EXPECT_STREQ(filler, ss.c_str());
 
     free(filler);
@@ -182,7 +184,18 @@ TEST(sstream, SeveralStrings) {
     ss << fillerB;
 
     EXPECT_EQ(sizeA + sizeB, strlen(ss.c_str()));
+    EXPECT_EQ(sizeA + sizeB, ss.length());
 
     free(fillerA);
     free(fillerB);
 }
+
+TEST(sstream, length) {
+    sstream ss;
+
+    EXPECT_EQ(0, ss.length());
+    ss << "Hello, world\n";
+    EXPECT_EQ(13, ss.length());
+    ss << "Foo bar\n";
+    EXPECT_EQ(13 + 8, ss.length());
+}