P4 to Git Change 1545895 by gandryey@gera-w8 on 2018/04/24 18:25:56

SWDEV-79445 - OCL generic changes and code clean-up - Replace dynamic memory object allocation for the managed buffer with a preallocated object Affected files ... ... //depot/stg/opencl/drivers/opencl/runtime/device/device.hpp#298 edit ... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palconstbuf.cpp#9 edit ... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palconstbuf.hpp#8 edit ... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palmemory.hpp#5 edit ... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palresource.cpp#64 edit ... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palresource.hpp#21 edit ... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palvirtual.cpp#94 edit ... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palvirtual.hpp#50 edit [ROCm/clr commit: 2f41f41977]
2018-04-24 18:35:32 -04:00
parent 564d236903
commit dacb39cafa
8 changed files with 61 additions and 39 deletions
@@ -838,15 +838,14 @@ class Memory : public amd::HeapObject {
  //! Decrement map count
  virtual void decIndMapCount() {}

+  size_t size_;   //!< Memory size
+
 private:
  //! Disable default copy constructor
-  Memory& operator=(const Memory&);
+  Memory& operator=(const Memory&) = delete;

  //! Disable operator=
-  Memory(const Memory&);
-
-  //! Our size
-  size_t size_;
+  Memory(const Memory&) = delete;
 };

 class Sampler : public amd::HeapObject {
@@ -135,39 +135,28 @@ uint64_t ConstantBuffer::UploadDataToHw(const void* sysmem, uint32_t size) const
 }

 // ================================================================================================
-XferBuffer::XferBuffer(ManagedBuffer& mbuf, uint32_t size)
-  : mbuf_(mbuf)
-  , size_(size)
-{}
+XferBuffer::XferBuffer(const Device& device, ManagedBuffer& mbuf, uint32_t size)
+  : buffer_view_(device, size)
+  , mbuf_(mbuf)
+  , size_(size) {
+  // Create a view for access
+  Resource::ViewParams params = {};
+  params.gpu_ = &mbuf_.gpu();
+  params.offset_ = 0;
+  params.size_ = size_;
+  params.resource_ = mbuf_.activeMemory();
+  bool result = buffer_view_.create(Resource::View, &params);
+  assert(result && "View creaiton should never return an error!");
+}

 // ================================================================================================
-Memory& XferBuffer::Acquire(uint32_t size) const
-{
+Memory& XferBuffer::Acquire(uint32_t size) {
  uint64_t  vm_address;
  // Reserve space in the managed buffer
  address   cpu_address = mbuf_.reserve(size, &vm_address);
-  // Create a view for access
-  Memory* mem = new Memory(mbuf_.gpu().dev(), static_cast<size_t>(size));
-  Resource::ViewParams params = {};
-  params.gpu_ = &mbuf_.gpu();
-  params.offset_ = vm_address - mbuf_.vmAddress();
-  params.size_ = size;
-  params.resource_ = mbuf_.activeMemory();
-  if (nullptr == mem || !mem->create(Resource::View, &params)) {
-    delete mem;
-    // If the suballocaiton failed for some reason, then return the top of the active buffer
-    return mbuf_.reserveAtTheTop(size);
-  }
-  return *mem;
-}
-
-// ================================================================================================
-void XferBuffer::Release(Memory& mem) const
-{
-  // Delete view
-  if (mem.desc().type_ == Resource::View) {
-    delete &mem;
-  }
+  // Update a view for access
+  buffer_view_.updateView(mbuf_.activeMemory(), vm_address - mbuf_.vmAddress(), size);
+  return buffer_view_;
 }

 }  // namespace pal
@@ -116,8 +116,9 @@ private:
 class XferBuffer : public amd::EmbeddedObject {
 public:
  //! Constructor for the ConstBuffer class
-  XferBuffer(ManagedBuffer& mbuf, //!< Managed buffer
-             uint32_t       size  //!< Maximum size of the transfer buffer
+  XferBuffer(const Device& device,  //!< Active GPU device 
+             ManagedBuffer& mbuf,   //!< Managed buffer
+             uint32_t       size    //!< Maximum size of the transfer buffer
  );

  //! Destructor for the ConstBuffer class
@@ -128,11 +129,11 @@ public:
  *  \return GPU memory object associated with free memory
  */
  Memory& Acquire(uint32_t size     //!< data size for transfers
-                  ) const;
+                  );

  //! Releases memory object used in the staging transfer
  void Release(Memory& mem  //!< Memory object for release
-               ) const;
+               ) { buffer_view_.updateView(nullptr, 0, 0); }

  size_t MaxSize() const { return static_cast<size_t>(size_); }

@@ -143,6 +144,7 @@ private:
  //! Disable operator=
  XferBuffer& operator=(const XferBuffer&) = delete;

+  Memory  buffer_view_;     //!< Buffer view returned in the acquire
  ManagedBuffer&  mbuf_;    //!< Managed buffer on GPU
  uint32_t  size_;          //!< Mx staging buffer size
 };
@@ -142,6 +142,12 @@ class Memory : public device::Memory, public Resource {
             (CL_MEM_ALLOC_HOST_PTR | CL_MEM_HOST_WRITE_ONLY | CL_MEM_READ_ONLY)));
  }

+  //! Quick view update for managed buffers. It should avoid expensive object allocations
+  void updateView(Resource* view, size_t offset, size_t size) {
+    size_ = size;
+    Resource::updateView(view, offset, size);
+  }
+
 protected:
  //! Decrement map count
  void decIndMapCount();
@@ -321,6 +321,8 @@ Resource::~Resource() {
    image_->Destroy();
    delete[] reinterpret_cast<char*>(image_);
  }
+
+  // Remove the current resource from the global resource list
  gpuDevice_.removeResource(this);
 }

@@ -1225,6 +1227,8 @@ void Resource::free()
  if (!desc().buffer_) {
    dev().srds().freeSrdSlot(hwSrd_);
  }
+
+  memRef_ = nullptr;
 }

 // ================================================================================================
@@ -359,6 +359,28 @@ class Resource : public amd::HeapObject {
  //! Erase an entry in the array for provided queue index
  void eraseGpuEvents(uint index) { events_.erase(events_.begin() + index); }

+  //! Quick view update for managed buffers. It should avoid expensive object allocations
+  //! If the base resource is null, then the view is released
+  void updateView(Resource* base, size_t offset, size_t size) {
+    if (base == nullptr) {
+      desc_.type_ = Empty;
+      memRef_->release();
+      memRef_ = nullptr;
+      viewOwner_ = nullptr;
+    } else {
+      desc_.type_ = View;
+      viewOwner_ = base;
+      offset_ = offset + viewOwner_->offset();
+      assert(viewOwner_->data() != nullptr && "CPU access must be provide for this call!");
+      address_ = viewOwner_->data() + offset;
+      desc_.cardMemory_ = viewOwner_->desc().cardMemory_;
+      memRef_ = viewOwner_->memRef_;
+      memRef_->retain();
+      desc_.width_ = amd::alignUp(size, Pal::Formats::BytesPerPixel(Pal::ChNumFormat::X32_Uint)) /
+        Pal::Formats::BytesPerPixel(Pal::ChNumFormat::X32_Uint);
+    }
+  }
+
 protected:
  /*! \brief Creates a PAL iamge object, associated with the resource
  *
@@ -688,7 +688,7 @@ VirtualGPU::VirtualGPU(Device& device)
      tsCache_(nullptr),
      dmaFlushMgmt_(device),
      managedBuffer_(*this, device.settings().stagedXferSize_ + 32 * Ki),
-      writeBuffer_(managedBuffer_, device.settings().stagedXferSize_),
+      writeBuffer_(device, managedBuffer_, device.settings().stagedXferSize_),
      hwRing_(0),
      readjustTimeGPU_(0),
      lastTS_(nullptr),
@@ -376,7 +376,7 @@ class VirtualGPU : public device::VirtualDevice {
                  );

  //! Return xfer buffer for staging operations
-  const XferBuffer& xferWrite() const { return writeBuffer_; }
+  XferBuffer& xferWrite() { return writeBuffer_; }

  //! Adds a pinned memory object into a map
  void addPinnedMem(amd::Memory* mem);