SWDEV-456279 - Adding new hip flag to access contiguous memory and pass the flag to HSA API.

Change-Id: I1bafeaa3096395c729723af958d609bc41e7845c


[ROCm/clr commit: 1d48f2a1ab]

projects/clr/hipamd/src/hip_memory.cpp

@@ -589,6 +589,8 @@ hipError_t hipExtMallocWithFlags(void** ptr, size_t sizeBytes, unsigned int flag
     ihipFlags = CL_MEM_SVM_ATOMICS;
   } else if (flags == hipDeviceMallocUncached) {
     ihipFlags = CL_MEM_SVM_ATOMICS | ROCCLR_MEM_HSA_UNCACHED;
+  } else if (flags == hipDeviceMallocContiguous) {
+    ihipFlags = ROCCLR_MEM_HSA_CONTIGUOUS | ROCCLR_MEM_HSA_UNCACHED;
   } else if (flags == hipMallocSignalMemory) {
     ihipFlags = CL_MEM_SVM_ATOMICS | CL_MEM_SVM_FINE_GRAIN_BUFFER | ROCCLR_MEM_HSA_SIGNAL_MEMORY;
     if (sizeBytes != 8) {

projects/clr/rocclr/device/device.hpp

@@ -1772,7 +1772,7 @@ class Device : public RuntimeObject {
   }
 
   virtual void* deviceLocalAlloc(size_t size, bool atomics = false,
-                                 bool pseudo_fine_grain = false) const {
+                                 bool pseudo_fine_grain = false, bool contiguous = false) const {
     ShouldNotCallThis();
     return NULL;
   }

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

@@ -2331,7 +2331,8 @@ void Device::deviceVmemRelease(uint64_t mem_handle) const {
   }
 }
 
-void* Device::deviceLocalAlloc(size_t size, bool atomics, bool pseudo_fine_grain) const {
+void* Device::deviceLocalAlloc(size_t size, bool atomics, bool pseudo_fine_grain,
+                               bool contiguous) const {
   const hsa_amd_memory_pool_t& pool = (pseudo_fine_grain) ? gpu_ext_fine_grained_segment_
                                       : (atomics) ? gpu_fine_grained_segment_ : gpuvm_segment_;
 
@@ -2341,6 +2342,11 @@ void* Device::deviceLocalAlloc(size_t size, bool atomics, bool pseudo_fine_grain
     return nullptr;
   }
 
+  uint32_t hsa_mem_flags = 0;
+  if (contiguous) {
+    hsa_mem_flags = HSA_AMD_MEMORY_POOL_CONTIGUOUS_FLAG;
+  }
+
   void* ptr = nullptr;
   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);

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

@@ -453,7 +453,8 @@ class Device : public NullDevice {
   bool allowPeerAccess(device::Memory* memory) const;
   void deviceVmemRelease(uint64_t mem_handle) const;
   uint64_t deviceVmemAlloc(size_t size, uint64_t flags) const;
-  void* deviceLocalAlloc(size_t size, bool atomics = false, bool pseudo_fine_grain=false) const;
+  void* deviceLocalAlloc(size_t size, bool atomics = false, bool pseudo_fine_grain=false,
+                         bool contiguous = false) const;
 
   void memFree(void* ptr, size_t size) const;
 

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

@@ -856,7 +856,8 @@ 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,
-                                               (memFlags & ROCCLR_MEM_HSA_UNCACHED) != 0);
+                                               (memFlags & ROCCLR_MEM_HSA_UNCACHED) != 0,
+                                               (memFlags & ROCCLR_MEM_HSA_CONTIGUOUS) != 0);
       }
       owner()->setSvmPtr(deviceMemory_);
     } else {

projects/clr/rocclr/platform/memory.hpp

@@ -44,6 +44,7 @@
 #define ROCCLR_MEM_HSA_UNCACHED         (1u << 27)
 #define ROCCLR_MEM_INTERPROCESS         (1u << 26)
 #define ROCCLR_MEM_PHYMEM               (1u << 25)
+#define ROCCLR_MEM_HSA_CONTIGUOUS       (1u << 24)
 
 namespace device {
 class Memory;