From 01217d0383ee4b75e5a017eb25ebc6a5d357f6fe Mon Sep 17 00:00:00 2001
From: German <German.Andryeyev@amd.com>
Date: Wed, 29 Nov 2023 16:53:09 -0500
Subject: [PATCH] SWDEV-435015 - Disable sysmem allocation in abstraction layer

Change-Id: Idefdabd2997d618b962c42545ae67768c628c5d6
---
 rocclr/device/rocm/rocmemory.cpp | 12 ++++++++++--
 rocclr/platform/memory.cpp       |  9 ++++++++-
 2 files changed, 18 insertions(+), 3 deletions(-)

diff --git a/rocclr/device/rocm/rocmemory.cpp b/rocclr/device/rocm/rocmemory.cpp
index 757aafcc68..bb926e38e7 100644
--- a/rocclr/device/rocm/rocmemory.cpp
+++ b/rocclr/device/rocm/rocmemory.cpp
@@ -951,7 +951,7 @@ bool Buffer::create(bool alloc_local) {
 
     return deviceMemory_ != nullptr;
   }
-  assert(owner()->getHostMem() != nullptr);
+  assert(owner()->getHostMem() != nullptr || (owner()->getContext().devices().size() == 1));
 
   flags_ |= HostMemoryDirectAccess;
 
@@ -965,7 +965,15 @@ bool Buffer::create(bool alloc_local) {
     return deviceMemory_ != nullptr;
   }
 
-  if (owner()->getSvmPtr() != owner()->getHostMem()) {
+  // Just one device and allocation must be done in the backend
+  if ((memFlags & CL_MEM_ALLOC_HOST_PTR) && (owner()->getContext().devices().size() == 1)) {
+    deviceMemory_ = dev().hostAlloc(size(), 1, Device::MemorySegment::kNoAtomics);
+    // Copy original data to the allocated host memory
+    if (memFlags & CL_MEM_COPY_HOST_PTR) {
+      memcpy(deviceMemory_, owner()->getHostMem(), owner()->getSize());
+    }
+    owner()->setHostMem(deviceMemory_);
+  } else if (owner()->getSvmPtr() != owner()->getHostMem()) {
     if (memFlags & (CL_MEM_USE_HOST_PTR | CL_MEM_ALLOC_HOST_PTR)) {
       hsa_amd_memory_pool_t pool = (memFlags & CL_MEM_SVM_ATOMICS) ?
                                     dev().SystemSegment() :
diff --git a/rocclr/platform/memory.cpp b/rocclr/platform/memory.cpp
index 5bd1c11dcd..01955cfb28 100644
--- a/rocclr/platform/memory.cpp
+++ b/rocclr/platform/memory.cpp
@@ -214,7 +214,14 @@ bool Memory::allocHostMemory(void* initFrom, bool allocHostMem, bool forceCopy)
   // This allocation is necessary to use coherency mechanism
   // for the initialization
   if (getMemFlags() & (CL_MEM_COPY_HOST_PTR | CL_MEM_ALLOC_HOST_PTR)) {
-    allocHostMem = true;
+    // Extra system memory allocation and copy can be very expensive.
+    // Thus, avoid it if runtime doesn't perform deferred allocations
+    if ((devices.size() == 1) || DISABLE_DEFERRED_ALLOC) {
+      allocHostMem = false;
+      setHostMem(initFrom);
+    } else {
+      allocHostMem = true;
+    }
   }
 
   // Did application request to use host memory?