Add suballocator for ordinary VRAM allocations smaller than 2MB.

Track pointer info for sub 2MB fragment allocations in allocation_map_.

Add fragment support to IPC.

Change-Id: I00cfc2e2fa289aac90a4718c392f9bb056a61a87

runtime/hsa-runtime/core/inc/amd_memory_region.h

@@ -49,6 +49,7 @@
 
 #include "core/inc/agent.h"
 #include "core/inc/memory_region.h"
+#include "core/util/simple_heap.h"
 
 #include "inc/hsa_ext_amd.h"
 
@@ -181,7 +182,21 @@ class MemoryRegion : public core::MemoryRegion {
   HSAuint64 virtual_size_;
 
   static const size_t kPageSize_ = 4096;
+
+  class BlockAllocator {
+   private:
+    MemoryRegion& region_;
+    static const size_t block_size_ = 2 * 1024 * 1024;  // 2MB blocks.
+   public:
+    explicit BlockAllocator(MemoryRegion& region) : region_(region) {}
+    void* alloc(size_t request_size, size_t& allocated_size) const;
+    void free(void* ptr, size_t length) const { region_.Free(ptr, length); }
+    size_t block_size() const { return block_size_; }
+  };
+
+  mutable SimpleHeap<BlockAllocator> fragment_allocator_;
 };
+
 }  // namespace
 
 #endif  // header guard

runtime/hsa-runtime/core/inc/memory_region.h

@@ -85,6 +85,7 @@ class MemoryRegion : public Checked<0x9C961F19EE175BB3> {
     AllocateRestrict = (1 << 0),    // Don't map system memory to GPU agents
     AllocateExecutable = (1 << 1),  // Set executable permission
     AllocateDoubleMap = (1 << 2),   // Map twice VA allocation to backing store
+    AllocateDirect = (1 << 3),      // Bypass fragment cache.
   };
 
   typedef uint32_t AllocateFlags;

runtime/hsa-runtime/core/inc/runtime.h

@@ -264,8 +264,14 @@ class Runtime {
 
   hsa_status_t InteropUnmap(void* ptr);
 
+  struct PtrInfoBlockData {
+    void* base;
+    size_t length;
+  };
+
   hsa_status_t PtrInfo(void* ptr, hsa_amd_pointer_info_t* info, void* (*alloc)(size_t),
-                       uint32_t* num_agents_accessible, hsa_agent_t** accessible);
+                       uint32_t* num_agents_accessible, hsa_agent_t** accessible,
+                       PtrInfoBlockData* block_info = nullptr);
 
   hsa_status_t SetPtrInfoData(void* ptr, void* userptr);
 
@@ -315,12 +321,13 @@ class Runtime {
   static void AsyncEventsLoop(void*);
 
   struct AllocationRegion {
-    AllocationRegion() : region(NULL), size(0) {}
+    AllocationRegion() : region(NULL), size(0), user_ptr(nullptr) {}
     AllocationRegion(const MemoryRegion* region_arg, size_t size_arg)
-        : region(region_arg), size(size_arg) {}
+        : region(region_arg), size(size_arg), user_ptr(nullptr) {}
 
     const MemoryRegion* region;
     size_t size;
+    void* user_ptr;
   };
 
   struct AsyncEventsControl {