tests: benchdnn: inputs: eltwise: add large buffer gpu test

tests/benchdnn/inputs/eltwise/harness_eltwise_large_buffer

@@ -0,0 +1,17 @@
+--reset
+
+# Use inplace to reduce memory consumption, most eltwise implementations do not
+# use different implementations for inplace. Use smallest supported data type to
+# minimize memory usage as well.
+
+--inplace=true
+
+--dt=u8
+--dir=FWD_D
+--attr-post-ops=,add:f32:per_tensor
+--batch=shapes_large_buffer
+
+--reset
+--dt=bf16
+--dir=BWD_D
+--batch=shapes_large_buffer

tests/benchdnn/inputs/eltwise/shapes_large_buffer

@@ -0,0 +1,2 @@
+# Test dimensions with 2 ^ 32 + 1 size to catch integer overflow
+4294967297n"large_dim1"

tests/benchdnn/inputs/eltwise/test_eltwise_gpu

@@ -34,3 +34,6 @@
 # Test layers of some key GPU DL Frameworks
 --reset
 --batch=option_set_fwks_key_gpu
+
+--reset
+--batch=harness_eltwise_large_buffer