SWDEV-371904 - Adding pseudo fine grain flag to hsa memory allocation for device fine grained memory.

Change-Id: I8cada90f0e3880dfbc5bf5a3fac4554e7a0cb08e


[ROCm/clr commit: e56a611b92]

projects/clr/rocclr/device/rocm/rocdevice.cpp

@@ -2081,8 +2081,10 @@ bool Device::allowPeerAccess(device::Memory* memory) const {
   return true;
 }
 
-void* Device::deviceLocalAlloc(size_t size, bool atomics) const {
-  const hsa_amd_memory_pool_t& pool = (atomics)? gpu_fine_grained_segment_ : gpuvm_segment_;
+void* Device::deviceLocalAlloc(size_t size, bool atomics, bool pseudo_fine_grain) const {
+  const hsa_amd_memory_pool_t& pool = (atomics) ? gpu_fine_grained_segment_ : gpuvm_segment_;
+  uint32_t hsa_mem_flags = (atomics && pseudo_fine_grain) ? HSA_AMD_MEMORY_POOL_PCIE_FLAG
+                                                          : HSA_AMD_MEMORY_POOL_STANDARD_FLAG;
 
   if (pool.handle == 0 || gpuvm_segment_max_alloc_ == 0) {
     DevLogPrintfError("Invalid argument, pool_handle: 0x%x , max_alloc: %u \n",
@@ -2091,7 +2093,7 @@ void* Device::deviceLocalAlloc(size_t size, bool atomics) const {
   }
 
   void* ptr = nullptr;
-  hsa_status_t stat = hsa_amd_memory_pool_allocate(pool, size, 0, &ptr);
+  hsa_status_t stat = hsa_amd_memory_pool_allocate(pool, size, hsa_mem_flags, &ptr);
   ClPrint(amd::LOG_DEBUG, amd::LOG_MEM, "Allocate hsa device memory %p, size 0x%zx", ptr, size);
   if (stat != HSA_STATUS_SUCCESS) {
     LogError("Fail allocation local memory");

projects/clr/rocclr/device/rocm/rocdevice.hpp

@@ -422,7 +422,7 @@ class Device : public NullDevice {
 
   bool allowPeerAccess(device::Memory* memory) const;
 
-  void* deviceLocalAlloc(size_t size, bool atomics = false) const;
+  void* deviceLocalAlloc(size_t size, bool atomics = false, bool pseudo_fine_grain=false) const;
 
   void memFree(void* ptr, size_t size) const;
 

projects/clr/rocclr/device/rocm/rocmemory.cpp

@@ -796,7 +796,8 @@ bool Buffer::create(bool alloc_local) {
         }
       } else {
         assert(!isHostMemDirectAccess() && "Runtime doesn't support direct access to GPU memory!");
-        deviceMemory_ = dev().deviceLocalAlloc(size(), (memFlags & CL_MEM_SVM_ATOMICS) != 0);
+        deviceMemory_ = dev().deviceLocalAlloc(size(), (memFlags & CL_MEM_SVM_ATOMICS) != 0,
+                                               (memFlags & ROCCLR_MEM_HSA_PSEUDO_FINE_GRAIN) != 0);
       }
       owner()->setSvmPtr(deviceMemory_);
     } else {

projects/clr/rocclr/platform/memory.hpp

@@ -41,6 +41,7 @@
 #define ROCCLR_MEM_HSA_SIGNAL_MEMORY                (1u << 30)
 #define ROCCLR_MEM_INTERNAL_MEMORY                  (1u << 29)
 #define CL_MEM_VA_RANGE_AMD                         (1u << 28)
+#define ROCCLR_MEM_HSA_PSEUDO_FINE_GRAIN            (1u << 27)
 
 namespace device {
 class Memory;