Merge pull request openvinotoolkit#35 from riverlijunjie/river/solve_…

…4GB_issue [TEST] Try to solve 4GB memory issue
zhaixuejun1993 · Sep 26, 2024 · 2edf7d2 · 2edf7d2
2 parents ae500b5 + 113acb5
commit 2edf7d2
Show file tree

Hide file tree

Showing 3 changed files with 11 additions and 5 deletions.
diff --git a/src/plugins/intel_gpu/src/runtime/ocl/ocl_engine.cpp b/src/plugins/intel_gpu/src/runtime/ocl/ocl_engine.cpp
@@ -128,7 +128,7 @@ allocation_type ocl_engine::detect_usm_allocation_type(const void* memory) const
 bool ocl_engine::check_allocatable(const layout& layout, allocation_type type) {
     OPENVINO_ASSERT(supports_allocation(type) || type == allocation_type::cl_mem, "[GPU] Unsupported allocation type: ", type);
 
-    bool exceed_allocatable_mem_size = (layout.bytes_count() > get_device_info().max_alloc_mem_size);
+    bool exceed_allocatable_mem_size = false;// (layout.bytes_count() > get_device_info().max_alloc_mem_size);
 
     // When dynamic shape upper bound makes bigger buffer, then return false.
     if (exceed_allocatable_mem_size && layout.is_dynamic()) {

diff --git a/src/plugins/intel_gpu/src/runtime/ocl/ocl_ext.hpp b/src/plugins/intel_gpu/src/runtime/ocl/ocl_ext.hpp
@@ -8,6 +8,9 @@
 
 #pragma once
 
+#define CL_MEM_ALLOW_UNRESTRICTED_SIZE_INTEL (1 << 23)
+#define CL_MEM_FLAGS_INTEL 0x10001
+
 #include <array>
 
 #ifdef OV_GPU_USE_OPENCL_HPP
@@ -929,21 +932,24 @@ class UsmMemory {
 
     void allocateHost(size_t size) {
         cl_int error = CL_SUCCESS;
-        auto ptr = _usmHelper.allocate_host(nullptr, size, 0, &error);
+        cl_mem_properties_intel properties[] = {CL_MEM_FLAGS_INTEL, CL_MEM_ALLOW_UNRESTRICTED_SIZE_INTEL, 0};
+        auto ptr = _usmHelper.allocate_host(properties, size, 0, &error);
         _check_error(size, ptr, error, "Host");
         _allocate(ptr);
     }
 
     void allocateShared(size_t size) {
         cl_int error = CL_SUCCESS;
-        auto ptr = _usmHelper.allocate_shared(nullptr, size, 0, &error);
+        cl_mem_properties_intel properties[] = {CL_MEM_FLAGS_INTEL, CL_MEM_ALLOW_UNRESTRICTED_SIZE_INTEL, 0};
+        auto ptr = _usmHelper.allocate_shared(properties, size, 0, &error);
         _check_error(size, ptr, error, "Shared");
         _allocate(ptr);
     }
 
     void allocateDevice(size_t size) {
         cl_int error = CL_SUCCESS;
-        auto ptr = _usmHelper.allocate_device(nullptr, size, 0, &error);
+        cl_mem_properties_intel properties[] = {CL_MEM_FLAGS_INTEL, CL_MEM_ALLOW_UNRESTRICTED_SIZE_INTEL, 0};
+        auto ptr = _usmHelper.allocate_device(properties, size, 0, &error);
         _check_error(size, ptr, error, "Device");
         _allocate(ptr);
     }

diff --git a/src/plugins/intel_gpu/src/runtime/ocl/ocl_memory.cpp b/src/plugins/intel_gpu/src/runtime/ocl/ocl_memory.cpp
@@ -52,7 +52,7 @@ static int get_cl_map_type(mem_lock_type type) {
 gpu_buffer::gpu_buffer(ocl_engine* engine,
                        const layout& layout)
     : lockable_gpu_mem(), memory(engine, layout, allocation_type::cl_mem, nullptr)
-    , _buffer(engine->get_cl_context(), CL_MEM_READ_WRITE, size() < 8192 ? 8192 : size()) {
+    , _buffer(engine->get_cl_context(), CL_MEM_READ_WRITE | CL_MEM_ALLOW_UNRESTRICTED_SIZE_INTEL, size() < 8192 ? 8192 : size()) {
     m_mem_tracker = std::make_shared<MemoryTracker>(engine, _buffer.get(), layout.bytes_count(), allocation_type::cl_mem);
 }