src: cpu: int8: dilated conv: fixup case when ih<=dh

Author: nivas-x86

src/cpu/jit_avx512_core_x8s8s32x_conv_kernel.cpp

@@ -505,7 +505,7 @@ void _jit_avx512_core_x8s8s32x_fwd_kernel<Vmm>::kh_loop(
         L(no_t_overflow_label);
     }
     mov(reg_kj, ptr[param1 + GET_OFF(kh_padding)]);
-    if ((jcp.signed_input)
+    if ((jcp.signed_input) || (jcp.dilate_h >= jcp.ih)
             || (!jcp.signed_input
                     && (jcp.kh - 1) * (jcp.dilate_h + 1)
                             < nstl::max(jcp.t_pad, jcp.b_pad))) {

tests/benchdnn/inputs/conv/test_conv_regression_general

@@ -130,3 +130,6 @@ mb1_g16ic16oc16_ih5oh2kh2sh3ph0
 
 # AVX JIT incorrectly calculates output when 'nb_oc_block' results in a prime number
 --reset --dir=FWD_I mb1ic3ih320oc51oh160kh7sh2ph3n"regression_oc-channel"
+
+# Dilated convolution with ih<=dh when there is _no_ compute work.
+--reset --conv --cfg=u8s8u8 mb1ic16ih1iw1oc16oh2ow1kh3kw1dh1ph1