SWDEV-326397 - P2P copies to take SDMA path if there is no pending dispatch

Change-Id: I50cfb8d77f7882151a20a1de7aaf5219b1695b7d

rocclr/device/rocm/rocblit.cpp

@@ -2051,7 +2051,7 @@ bool KernelBlitManager::copyBuffer(device::Memory& srcMemory, device::Memory& ds
   amd::ScopedLock k(lockXferOps_);
   bool result = false;
   bool p2p = (&gpuMem(srcMemory).dev() != &gpuMem(dstMemory).dev()) &&
-             (sizeIn[0] > ROC_P2P_SDMA_SIZE * Ki);
+             ((sizeIn[0] > ROC_P2P_SDMA_SIZE * Ki) || !gpu().IsPendingDispatch());
   bool asan = false;
 #if defined(__clang__)
 #if __has_feature(address_sanitizer)

rocclr/device/rocm/rocvirtual.hpp

@@ -382,6 +382,7 @@ class VirtualGPU : public device::VirtualDevice {
   void enableSyncBlit() const;
 
   void hasPendingDispatch() { hasPendingDispatch_ = true; }
+  bool IsPendingDispatch() const { return (hasPendingDispatch_) ? true : false; }
   void addSystemScope() { addSystemScope_ = true; }
   void SetCopyCommandType(cl_command_type type) { copy_command_type_ = type; }
 

rocclr/utils/flags.hpp

@@ -267,7 +267,7 @@ release(bool, AMD_CPU_AFFINITY, false,                                        \
 release(bool, ROC_USE_FGS_KERNARG, true,                                      \
         "Use fine grain kernel args segment for supported asics")             \
 release(uint, ROC_P2P_SDMA_SIZE, 1024,                                        \
-        "The minimum size in MB for P2P transfer with SDMA")                  \
+        "The minimum size in KB for P2P transfer with SDMA")                  \
 release(uint, ROC_AQL_QUEUE_SIZE, 4096,                                       \
         "AQL queue size in AQL packets")                                      \
 release(bool, ROC_SKIP_KERNEL_ARG_COPY, false,                                \