[GPU] Fix not to request 0 bytes layout allocation to OpenCL

Author: kelvinchoi-intel

src/plugins/intel_gpu/src/runtime/ocl/ocl_engine.cpp

@@ -128,6 +128,10 @@ 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);
 
+    // if (layout.bytes_count() == 0) {
+    //     return false;
+    // }
+
     bool exceed_allocatable_mem_size = (layout.bytes_count() > get_device_info().max_alloc_mem_size);
 
     // When dynamic shape upper bound makes bigger buffer, then return false.
@@ -172,25 +176,46 @@ bool ocl_engine::check_allocatable(const layout& layout, allocation_type type) {
 memory::ptr ocl_engine::allocate_memory(const layout& layout, allocation_type type, bool reset) {
     OPENVINO_ASSERT(!layout.is_dynamic() || layout.has_upper_bound(), "[GPU] Can't allocate memory for dynamic layout");
 
-    check_allocatable(layout, type);
+    if (!check_allocatable(layout, type))
+        return nullptr;
+
+    auto zero_bytes_layout = false;
+    auto non_zero_layout = layout;
+    if (layout.bytes_count() == 0) {
+        cldnn::layout zero_dim_layout = layout;
+        auto mem_ps = zero_dim_layout.get_partial_shape();
+        for (size_t k = 0; k < mem_ps.size(); k++) {
+            if (mem_ps[k] == 0)
+                mem_ps[k] = 1;
+        }
+
+        non_zero_layout = cldnn::layout(mem_ps, zero_dim_layout.data_type, zero_dim_layout.format);
+        zero_bytes_layout = true;
+    }
+
+
 
     try {
         memory::ptr res = nullptr;
         if (layout.format.is_image_2d()) {
-            res = std::make_shared<ocl::gpu_image2d>(this, layout);
+            res = std::make_shared<ocl::gpu_image2d>(this, non_zero_layout);
         } else if (type == allocation_type::cl_mem) {
-            res = std::make_shared<ocl::gpu_buffer>(this, layout);
+            res = std::make_shared<ocl::gpu_buffer>(this, non_zero_layout);
         } else {
-            res = std::make_shared<ocl::gpu_usm>(this, layout, type);
+            res = std::make_shared<ocl::gpu_usm>(this, non_zero_layout, type);
         }
 
-        if (reset || res->is_memory_reset_needed(layout)) {
+        if (reset || res->is_memory_reset_needed(non_zero_layout)) {
             auto ev = res->fill(get_service_stream());
             if (ev) {
                 get_service_stream().wait_for_events({ev});
             }
         }
 
+        if (zero_bytes_layout) {
+            res = reinterpret_buffer(*res, layout);
+        }
+
         return res;
     } catch (const cl::Error& clErr) {
         switch (clErr.err()) {