More flexible buffer pool's allocation policy

Buffer could be allocated in shared memory, or GPU memory
or automatically chosen based on size

Change-Id: Iab8ce9313c9ff2a7e0f062cd874d1b24a54d7de4

Author: kakashidinho

src/libANGLE/renderer/metal/BufferMtl.mm

@@ -367,17 +367,20 @@
         case gl::BufferUsage::DynamicRead:
         case gl::BufferUsage::StreamRead:
             maxBuffers = 1;  // static/read buffer doesn't need high speed data update
+            mBufferPool.setAlwaysUseGPUMem();
             break;
         default:
             // dynamic buffer, allow up to 10 update per frame/encoding without
             // waiting for GPU.
             if (adjustedSize <= kDynamicBufferPoolMaxBufSize)
             {
                 maxBuffers = 10;
+                mBufferPool.setAlwaysUseSharedMem();
             }
             else
             {
                 maxBuffers = 1;
+                mBufferPool.setAlwaysUseGPUMem();
             }
             break;
     }

src/libANGLE/renderer/metal/mtl_buffer_pool.h

@@ -20,6 +20,13 @@ class ContextMtl;
 namespace mtl
 {
 
+enum class BufferPoolMemPolicy
+{
+    AlwaysSharedMem,  // Always allocate buffer in shared memory, useful for dynamic small buffer
+    AlwaysGPUMem,     // Always allocate buffer in GPU dedicated memory.
+    Auto,             // Auto allocate buffer in shared memory if it is small. GPU otherwise.
+};
+
 // A buffer pool is conceptually an infinitely long buffer. Each time you write to the buffer,
 // you will always write to a previously unused portion. After a series of writes, you must flush
 // the buffer data to the device. Buffer lifetime currently assumes that each new allocation will
@@ -39,7 +46,7 @@ class BufferPool
     // - alwaysUseSharedMem: indicate the allocated buffers should be in shared memory or not.
     // If this flag is false. Buffer pool will automatically use shared mem if buffer size is small.
     BufferPool(bool alwaysAllocNewBuffer = false);
-    BufferPool(bool alwaysAllocNewBuffer, bool alwaysUseSharedMem);
+    BufferPool(bool alwaysAllocNewBuffer, BufferPoolMemPolicy memPolicy);
     ~BufferPool();
 
     // Init is called after the buffer creation so that the alignment can be specified later.
@@ -79,10 +86,16 @@ class BufferPool
     // buffer or not. Default is false.
     void setAlwaysAllocateNewBuffer(bool e) { mAlwaysAllocateNewBuffer = e; }
 
-    // Set whether all subsequent allocated buffers should always use shared memory or not
-    void setAlwaysUseSharedMem(bool e) { mAlwaysUseSharedMem = e; }
+    void setMemoryPolicy(BufferPoolMemPolicy policy) { mMemPolicy = policy; }
+
+    // Set all subsequent allocated buffers should always use shared memory
+    void setAlwaysUseSharedMem() { setMemoryPolicy(BufferPoolMemPolicy::AlwaysSharedMem); }
+
+    // Set all subsequent allocated buffers should always use GPU memory
+    void setAlwaysUseGPUMem() { setMemoryPolicy(BufferPoolMemPolicy::AlwaysGPUMem); }
 
   private:
+    bool shouldAllocateInSharedMem() const;
     void reset();
     angle::Result allocateNewBuffer(ContextMtl *contextMtl);
     void destroyBufferList(ContextMtl *contextMtl, std::vector<BufferRef> *buffers);
@@ -98,8 +111,8 @@ class BufferPool
 
     size_t mBuffersAllocated;
     size_t mMaxBuffers;
+    BufferPoolMemPolicy mMemPolicy = BufferPoolMemPolicy::Auto;
     bool mAlwaysAllocateNewBuffer;
-    bool mAlwaysUseSharedMem;
 };
 
 }  // namespace mtl

src/libANGLE/renderer/metal/mtl_buffer_pool.mm

@@ -20,21 +20,23 @@
 namespace
 {
 // The max size of a buffer that will be allocated in shared memory
-constexpr size_t kSharedMemBufferMaxBufSize = 128 * 1024;
+constexpr size_t kSharedMemBufferMaxBufSizeHint = 128 * 1024;
 }
 
 // BufferPool implementation.
-BufferPool::BufferPool(bool alwaysAllocNewBuffer) : BufferPool(alwaysAllocNewBuffer, false) {}
-BufferPool::BufferPool(bool alwaysAllocNewBuffer, bool alwaysUseSharedMem)
+BufferPool::BufferPool(bool alwaysAllocNewBuffer)
+    : BufferPool(alwaysAllocNewBuffer, BufferPoolMemPolicy::Auto)
+{}
+BufferPool::BufferPool(bool alwaysAllocNewBuffer, BufferPoolMemPolicy policy)
     : mInitialSize(0),
       mBuffer(nullptr),
       mNextAllocationOffset(0),
       mSize(0),
       mAlignment(1),
       mBuffersAllocated(0),
       mMaxBuffers(0),
-      mAlwaysAllocateNewBuffer(alwaysAllocNewBuffer),
-      mAlwaysUseSharedMem(alwaysUseSharedMem)
+      mMemPolicy(policy),
+      mAlwaysAllocateNewBuffer(alwaysAllocNewBuffer)
 
 {}
 
@@ -65,7 +67,7 @@
             {
                 continue;
             }
-            bool useSharedMem = mAlwaysUseSharedMem || mSize <= kSharedMemBufferMaxBufSize;
+            bool useSharedMem = shouldAllocateInSharedMem();
             if (IsError(buffer->reset(contextMtl, useSharedMem, mSize, nullptr)))
             {
                 mBufferFreeList.clear();
@@ -90,6 +92,19 @@
 
 BufferPool::~BufferPool() {}
 
+bool BufferPool::shouldAllocateInSharedMem() const
+{
+    switch (mMemPolicy)
+    {
+        case BufferPoolMemPolicy::AlwaysSharedMem:
+            return true;
+        case BufferPoolMemPolicy::AlwaysGPUMem:
+            return false;
+        default:
+            return mSize <= kSharedMemBufferMaxBufSizeHint;
+    }
+}
+
 angle::Result BufferPool::allocateNewBuffer(ContextMtl *contextMtl)
 {
     if (mMaxBuffers > 0 && mBuffersAllocated >= mMaxBuffers)
@@ -120,7 +135,7 @@
         return angle::Result::Continue;
     }
 
-    bool useSharedMem = mAlwaysUseSharedMem || mSize <= kSharedMemBufferMaxBufSize;
+    bool useSharedMem = shouldAllocateInSharedMem();
     ANGLE_TRY(Buffer::MakeBuffer(contextMtl, useSharedMem, mSize, nullptr, &mBuffer));
 
     ASSERT(mBuffer);
@@ -225,7 +240,12 @@
     for (auto &toRelease : mInFlightBuffers)
     {
         // If the dynamic buffer was resized we cannot reuse the retained buffer.
-        if (toRelease->size() < mSize)
+        if (toRelease->size() < mSize
+#if TARGET_OS_OSX || TARGET_OS_MACCATALYST
+            // Also release buffer if it was allocated in different policy
+            || toRelease->useSharedMem() != shouldAllocateInSharedMem()
+#endif
+        )
         {
             toRelease = nullptr;
             mBuffersAllocated--;

src/libANGLE/renderer/metal/mtl_resources.h

@@ -238,6 +238,7 @@ class Buffer final : public Resource,
     void unmap(ContextMtl *context, size_t offsetWritten, size_t sizeWritten);
 
     size_t size() const;
+    bool useSharedMem() const;
 
     // Explicitly sync content between CPU and GPU
     void syncContent(ContextMtl *context, mtl::BlitCommandEncoder *encoder);

src/libANGLE/renderer/metal/mtl_resources.mm

@@ -661,5 +661,11 @@ void EnsureContentSynced(ContextMtl *context, const std::shared_ptr<T> &resource
 {
     return get().length;
 }
+
+bool Buffer::useSharedMem() const
+{
+    return get().storageMode == MTLStorageModeShared;
+}
+
 }
 }