benchdnn: inputs: reorder: skip huge problems for slow impls

Author: dzarukin

tests/benchdnn/inputs/reorder/harness_reorder_large

@@ -0,0 +1,11 @@
+# test if jit kernels properly handle corner cases:
+# * large stride problems
+# * huge dimensions (UINT_MAX + 1)
+--reset
+--skip-impl=ref,simple # run only jit impl, won't iterate
+--sdt=f32
+--ddt=f32
+--stag=abx
+--dtag=aBx8b
+2x16x19200x19200
+1x4294967296x1

tests/benchdnn/inputs/reorder/test_reorder_all

@@ -60,16 +60,6 @@
 --stag=aBx4b,aBx8b --dtag=aBx16b 2x71x16x16 2x72x16x16 2x73x16x16
 --stag=aBx16b      --dtag=aBx8b  2x71x16x16 2x72x16x16 2x73x16x16
 
-# test if jit kernels properly handle corner cases:
-# * large stride problems
-# * huge dimensions (UINT_MAX + 1)
---reset
---skip-impl=ref,simple # ! test jit version only
---sdt=f32 --ddt=f32
---stag=abx --dtag=aBx8b 2x16x19200x19200
---skip-impl=
-1x4294967296x1
-
 # f16
 --batch=test_reorder_float16
 
@@ -102,3 +92,6 @@
 
 # Decompression quantization
 --batch=harness_reorder_decompression
+
+# large problems
+--batch=harness_reorder_large

tests/benchdnn/inputs/reorder/test_reorder_gpu

@@ -268,6 +268,7 @@
 
 # Catch overflows
 --reset
+--skip-impl=ref
 2147483648_n"int_overflow"
 4294967296_n"uint_overflow"
 2147483869_n"nd_range_overflow"