FEAT-47686 - Changes to support new uncached memory segment in ROCr.

Change-Id: I0ba8769d6737cdf1cc8a8644b2e82109f584a430


[ROCm/clr commit: aca7d9e14a]

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

@@ -901,6 +901,12 @@ hsa_status_t Device::iterateGpuMemoryPoolCallback(hsa_amd_memory_pool_t pool, vo
           return stat;
         }
 
+        // If the flag set is ext scoped fine grain, break the loop
+        if ((global_flag & HSA_REGION_GLOBAL_FLAG_EXTENDED_SCOPE_FINE_GRAINED) != 0) {
+          dev->gpu_ext_fine_grained_segment_ = pool;
+          break;
+        }
+
         if ((global_flag & HSA_REGION_GLOBAL_FLAG_FINE_GRAINED) != 0) {
           dev->gpu_fine_grained_segment_ = pool;
         } else if ((global_flag & HSA_REGION_GLOBAL_FLAG_COARSE_GRAINED) != 0) {
@@ -959,6 +965,12 @@ hsa_status_t Device::iterateCpuMemoryPoolCallback(hsa_amd_memory_pool_t pool, vo
         break;
       }
 
+      // If the flag set is ext scoped fine grain, break the loop
+      if ((global_flag & HSA_REGION_GLOBAL_FLAG_EXTENDED_SCOPE_FINE_GRAINED) != 0) {
+        agentInfo->ext_fine_grain_pool_ = pool;
+        break;
+      }
+
       if ((global_flag & HSA_REGION_GLOBAL_FLAG_FINE_GRAINED) != 0) {
         if (agentInfo->fine_grain_pool.handle == 0) {
           agentInfo->fine_grain_pool = pool;
@@ -2158,9 +2170,8 @@ bool Device::allowPeerAccess(device::Memory* memory) const {
 }
 
 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;
+  const hsa_amd_memory_pool_t& pool = (pseudo_fine_grain) ? gpu_ext_fine_grained_segment_
+                                      : (atomics) ? gpu_fine_grained_segment_ : gpuvm_segment_;
 
   if (pool.handle == 0 || gpuvm_segment_max_alloc_ == 0) {
     DevLogPrintfError("Invalid argument, pool_handle: 0x%x , max_alloc: %u \n",
@@ -2169,7 +2180,7 @@ void* Device::deviceLocalAlloc(size_t size, bool atomics, bool pseudo_fine_grain
   }
 
   void* ptr = nullptr;
-  hsa_status_t stat = hsa_amd_memory_pool_allocate(pool, size, hsa_mem_flags, &ptr);
+  hsa_status_t stat = hsa_amd_memory_pool_allocate(pool, size, 0, &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

@@ -288,6 +288,7 @@ struct AgentInfo {
   hsa_amd_memory_pool_t fine_grain_pool;
   hsa_amd_memory_pool_t coarse_grain_pool;
   hsa_amd_memory_pool_t kern_arg_pool;
+  hsa_amd_memory_pool_t ext_fine_grain_pool_;
 };
 
 //! A HSA device ordinal (physical HSA device)
@@ -590,6 +591,7 @@ class Device : public NullDevice {
   hsa_amd_memory_pool_t system_kernarg_segment_;
   hsa_amd_memory_pool_t gpuvm_segment_;
   hsa_amd_memory_pool_t gpu_fine_grained_segment_;
+  hsa_amd_memory_pool_t gpu_ext_fine_grained_segment_;
   hsa_signal_t prefetch_signal_;    //!< Prefetch signal, used to explicitly prefetch SVM on device
   std::atomic<int> cache_state_;    //!< State of cache, kUnknown/kFlushedToDevice/kFlushedToSystem