[RUNTIME][OPENCL] OpenCL host pointer support to acheive zero copy

OpenCL supports device memory access to host by memory mapping.
OpenCL flag "CL_MEM_ALLOC_HOST_PTR" enable this while creating a memory object.

We enable this feature via compilation setting "USE_OPENCL_ENABLE_HOST_PTR"
followed by a new API "GetNativePtr" on DeviceAPI followed by NDArray class.

This allows application directly use hardware allocated memory while preparing the input.
From user side we allocate NDArray which same size as graph input, access native memory and
finally call set_input_zero_copy to set the input.

Psudo code looks like

auto narr = tvm::runtime::NDArray::Empty(shape, {kDLFloat, 32, 1}, {kDLOpenCL, 0});
void * nptr = narr.GetNativePtr();

... access memory pointed by nptr up to the tensor size ...

tvm::runtime::PackedFunc set_input = mod.GetFunction("set_input_zero_copy");
set_input(i, ninput);

Author: srkreddy1238

CMakeLists.txt

@@ -26,6 +26,7 @@ endif()
 # Alernatively, use cmake -DOPTION=VALUE through command-line.
 tvm_option(USE_CUDA "Build with CUDA" OFF)
 tvm_option(USE_OPENCL "Build with OpenCL" OFF)
+tvm_option(USE_OPENCL_ENABLE_HOST_PTR "Enable OpenCL memory object access to host" OFF)
 tvm_option(USE_OPENCL_GTEST "Path to OpenCL specific gtest version for runtime cpp tests." /path/to/opencl/gtest)
 tvm_option(USE_VULKAN "Build with Vulkan" OFF)
 

cmake/config.cmake

@@ -71,6 +71,11 @@ set(USE_AOCL OFF)
 # - /path/to/opencl-sdk: use specific path to opencl-sdk
 set(USE_OPENCL OFF)
 
+# Wheather to allow OPENCL cl_mem access to host
+# cl_mem will be allocated with CL_MEM_ALLOC_HOST_PTR
+# ndarray->GetHostPtr API returns the host accessible pointer
+set(USE_OPENCL_ENABLE_HOST_PTR OFF)
+
 # Whether enable Metal runtime
 set(USE_METAL OFF)
 

cmake/modules/OpenCL.cmake

@@ -64,6 +64,9 @@ if(USE_OPENCL)
     )
   endif()
   list(APPEND RUNTIME_SRCS ${RUNTIME_OPENCL_SRCS})
+  if(USE_OPENCL_ENABLE_HOST_PTR)
+    add_definitions(-DOPENCL_ENABLE_HOST_PTR)
+  endif(USE_OPENCL_ENABLE_HOST_PTR)
 else()
   list(APPEND COMPILER_SRCS src/target/opt/build_opencl_off.cc)
 endif(USE_OPENCL)

cmake/modules/contrib/CLML.cmake

@@ -54,5 +54,7 @@ if(USE_CLML_GRAPH_EXECUTOR)
     file(GLOB RUNTIME_OPENCL_SRCS src/runtime/opencl/*.cc)
     list(APPEND RUNTIME_SRCS ${RUNTIME_OPENCL_SRCS})
     set(USE_OPENCL ON)
-
+    if(USE_OPENCL_ENABLE_HOST_PTR)
+        add_definitions(-DOPENCL_ENABLE_HOST_PTR)
+    endif(USE_OPENCL_ENABLE_HOST_PTR)
 endif()

include/tvm/runtime/device_api.h

@@ -124,6 +124,11 @@ class TVM_DLL DeviceAPI {
    */
   virtual void* AllocDataSpace(Device dev, int ndim, const int64_t* shape, DLDataType dtype,
                                Optional<String> mem_scope = NullOpt);
+  /*!
+   * \brief Return host accessible native pointer if underlaying device supports.
+   * \return pointer to memory or nullptr if not supported
+   */
+  virtual void* GetNativePtr(Device dev, void* ptr);
   /*!
    * \brief Free a data space on device.
    * \param dev The device device to perform operation.

include/tvm/runtime/ndarray.h

@@ -155,6 +155,11 @@ class NDArray : public ObjectRef {
    */
   TVM_DLL static NDArray Empty(ShapeTuple shape, DLDataType dtype, Device dev,
                                Optional<String> mem_scope = NullOpt);
+  /*!
+   * \brief Return host accessible native pointer if underlaying device supports.
+   * \return pointer to memory or nullptr if not supported
+   */
+  TVM_DLL void* GetNativePtr();
   /*!
    * \brief Create a NDArray backed by an external DLTensor without memory copying.
    *

src/runtime/c_runtime_api.cc

@@ -171,6 +171,8 @@ void* DeviceAPI::AllocDataSpace(Device dev, int ndim, const int64_t* shape, DLDa
   return nullptr;
 }
 
+void * DeviceAPI::GetNativePtr(Device dev, void *ptr) { return nullptr; }
+
 void DeviceAPI::CopyDataFromTo(DLTensor* from, DLTensor* to, TVMStreamHandle stream) {
   // by default, we can always redirect to the flat memory copy operation.
   size_t nbytes = GetDataSize(*from);

src/runtime/ndarray.cc

@@ -205,6 +205,11 @@ NDArray NDArray::Empty(ShapeTuple shape, DLDataType dtype, Device dev, Optional<
   return ret;
 }
 
+void* NDArray::GetNativePtr() {
+  DLTensor dl_tensor = get_mutable()->dl_tensor;
+  return DeviceAPI::Get(dl_tensor.device)->GetNativePtr(dl_tensor.device, dl_tensor.data);
+}
+
 NDArray NDArray::FromExternalDLTensor(const DLTensor& dl_tensor) {
   ICHECK(::tvm::runtime::IsContiguous(dl_tensor)) << "External DLTensor must be contiguous.";
   ICHECK(IsAligned(dl_tensor)) << "Data in DLTensor is not aligned as required by NDArray";

src/runtime/opencl/opencl_common.h

@@ -213,6 +213,7 @@ inline cl_channel_type DTypeToOpenCLChannelType(DLDataType data_type) {
   }
 
 class OpenCLThreadEntry;
+struct BufferDescriptor;
 
 /*!
  * \brief Process global OpenCL workspace.
@@ -291,6 +292,7 @@ class OpenCLWorkspace : public DeviceAPI {
   void* AllocDataSpace(Device dev, size_t size, size_t alignment, DLDataType type_hint) final;
   void* AllocDataSpace(Device dev, int ndim, const int64_t* shape, DLDataType dtype,
                        Optional<String> mem_scope = NullOpt) final;
+  void* GetNativePtr(Device dev, void* ptr) final;
   void FreeDataSpace(Device dev, void* ptr) final;
   void StreamSync(Device dev, TVMStreamHandle stream) final;
   void* AllocWorkspace(Device dev, size_t size, DLDataType type_hint) final;
@@ -310,6 +312,8 @@ class OpenCLWorkspace : public DeviceAPI {
   static OpenCLWorkspace* Global();
 
   void CopyDataFromTo(DLTensor* from, DLTensor* to, TVMStreamHandle stream) final;
+
+  void* CreateHostPtrIfEnabled(BufferDescriptor *desc, Device dev, size_t size);
 };
 
 /*! \brief Thread local workspace */
@@ -371,6 +375,7 @@ struct BufferDescriptor {
   static String ScopeFromMemoryLayout(MemoryLayout mem_scope);
 
   cl_mem buffer{nullptr};
+  cl_uchar *host_ptr{nullptr};
   MemoryLayout layout{MemoryLayout::kBuffer1D};
 };
 }  // namespace cl

src/runtime/opencl/opencl_device_api.cc

@@ -27,6 +27,12 @@
 
 #include "opencl_common.h"
 
+#ifdef OPENCL_ENABLE_HOST_PTR
+#define CL_MEM_CREATE_FLAGS CL_MEM_READ_WRITE|CL_MEM_ALLOC_HOST_PTR
+#else
+#define CL_MEM_CREATE_FLAGS CL_MEM_READ_WRITE
+#endif
+
 namespace tvm {
 namespace runtime {
 namespace cl {
@@ -197,6 +203,17 @@ void OpenCLWorkspace::GetAttr(Device dev, DeviceAttrKind kind, TVMRetValue* rv)
   }
 }
 
+void * OpenCLWorkspace::CreateHostPtrIfEnabled(cl::BufferDescriptor* desc, Device dev, size_t size) {
+#if defined(OPENCL_ENABLE_HOST_PTR)
+    cl_int err_code;
+    desc->host_ptr = (cl_uchar *)clEnqueueMapBuffer(this->GetQueue(dev), desc->buffer, CL_TRUE,
+                                                    CL_MAP_WRITE, 0, sizeof(cl_uchar) * size, 0,
+                                                    NULL, NULL, &err_code);
+    OPENCL_CHECK_ERROR(err_code);
+#endif  // OPENCL_ENABLE_HOST_PTR
+    return desc;
+}
+
 void* OpenCLWorkspace::AllocDataSpace(Device dev, size_t size, size_t alignment,
                                       DLDataType type_hint) {
   this->Init();
@@ -207,10 +224,10 @@ void* OpenCLWorkspace::AllocDataSpace(Device dev, size_t size, size_t alignment,
   if (size == 0) {
     size = 1;
   }
-  desc->buffer = clCreateBuffer(this->context, CL_MEM_READ_WRITE, size, nullptr, &err_code);
+  desc->buffer = clCreateBuffer(this->context, CL_MEM_CREATE_FLAGS, size, nullptr, &err_code);
   desc->layout = cl::BufferDescriptor::MemoryLayout::kBuffer1D;
   OPENCL_CHECK_ERROR(err_code);
-  return desc;
+  return CreateHostPtrIfEnabled(desc, dev, size);
 }
 
 void* OpenCLWorkspace::AllocDataSpace(Device dev, int ndim, const int64_t* shape, DLDataType dtype,
@@ -232,12 +249,21 @@ void* OpenCLWorkspace::AllocDataSpace(Device dev, int ndim, const int64_t* shape
   return desc;
 }
 
+void* OpenCLWorkspace::GetNativePtr(Device dev, void* ptr) {
+  cl::BufferDescriptor* desc = static_cast<cl::BufferDescriptor*>(ptr);
+  return desc->host_ptr;
+}
+
 void OpenCLWorkspace::FreeDataSpace(Device dev, void* ptr) {
   // We have to make sure that the memory object is not in the command queue
   // for some OpenCL platforms.
   OPENCL_CALL(clFinish(this->GetQueue(dev)));
 
   cl::BufferDescriptor* desc = static_cast<cl::BufferDescriptor*>(ptr);
+  if(desc->host_ptr) {
+    clEnqueueUnmapMemObject(this->GetQueue(dev), desc->buffer, (void *) desc->host_ptr, 0, NULL,
+                            NULL);
+  }
   OPENCL_CALL(clReleaseMemObject(desc->buffer));
   delete desc;
 }
@@ -251,7 +277,7 @@ cl_mem OpenCLWorkspace::AllocTexture(Device dev, size_t width, size_t height,
   cl_image_format format = {CL_RGBA, cl_type};
   cl_image_desc descriptor = {CL_MEM_OBJECT_IMAGE2D, width, height, 0, 0, 0, 0, 0, 0};
   cl_mem mptr =
-      clCreateImage(this->context, CL_MEM_READ_WRITE, &format, &descriptor, nullptr, &err_code);
+      clCreateImage(this->context, CL_MEM_CREATE_FLAGS, &format, &descriptor, nullptr, &err_code);
   OPENCL_CHECK_ERROR(err_code);
   return mptr;
 }