P4 to Git Change 1537540 by gandryey@gera-w8 on 2018/04/06 12:30:49

SWDEV-79445 - OCL generic changes and code clean-up
	- Reduce the managed buffer alignment to 128 bytes
	- Add direct upload function to the constant buffer
	- Add heap preference print into the active allocations

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palconstbuf.cpp#7 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palconstbuf.hpp#6 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palkernel.cpp#49 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palvirtual.cpp#87 edit


[ROCm/clr commit: c006b7b262]

projects/clr/rocclr/runtime/device/pal/palconstbuf.cpp

@@ -53,7 +53,8 @@ bool ManagedBuffer::create(Resource::MemoryType type) {
 
 // ================================================================================================
 address ManagedBuffer::reserve(uint32_t size, uint64_t* gpu_address) {
-  static constexpr uint32_t MemAlignment = 256;
+  // Align to the maximum data size available in OpenCL
+  static constexpr uint32_t MemAlignment = sizeof(cl_double16);
 
   // Align reserve size on the vector's boundary
   uint32_t count = amd::alignUp(size, MemAlignment);
@@ -103,8 +104,7 @@ bool ConstantBuffer::Create() {
 }
 
 // ================================================================================================
-uint64_t ConstantBuffer::UploadDataToHw(uint32_t size) const
-{
+uint64_t ConstantBuffer::UploadDataToHw(uint32_t size) const {
   uint64_t  vm_address;
   address   cpu_address = mbuf_.reserve(size, &vm_address);
   // Update memory with new CB data
@@ -112,4 +112,13 @@ uint64_t ConstantBuffer::UploadDataToHw(uint32_t size) const
   return vm_address;
 }
 
+// ================================================================================================
+uint64_t ConstantBuffer::UploadDataToHw(const void* sysmem, uint32_t size) const {
+  uint64_t  vm_address;
+  address   cpu_address = mbuf_.reserve(size, &vm_address);
+  // Update memory with new CB data
+  memcpy(cpu_address, sysmem, size);
+  return vm_address;
+}
+
 }  // namespace pal