[FORK][FEATURE] Enable jit sse41 NxN convolution for grayscale input

Author: dmitry-gorokhov

src/cpu/x64/jit_sse41_conv_kernel_f32.cpp

@@ -468,7 +468,7 @@ status_t jit_sse41_conv_fwd_kernel_f32::init_conf(jit_conv_conf_t &jcp,
             sum_requires_scale_one, sum_requires_zp_zero));
     VDISPATCH_CONV_IC(post_ops_ok_, VERBOSE_UNSUPPORTED_POSTOP);
 
-    const bool flat = jcp.ic == 3;
+    const bool flat = jcp.ic == 3 || jcp.ic == 1;
     const bool mimo = !flat;
 
     const bool tag_ok = true

src/cpu/x64/jit_sse41_convolution.hpp

@@ -88,7 +88,7 @@ struct jit_sse41_convolution_fwd_t : public primitive_t {
                     && IMPLICATION(curr_dst_tag != dat_tag_nxc,
                             dst_d.format_kind() == format_kind::any)
                     && utils::one_of(dat_tag_nxc, curr_src_tag, curr_dst_tag);
-            const bool flat = IC() == 3;
+            const bool flat = IC() == 3 || IC() == 1;;
             auto src_tag = is_data_layout_nxc ? dat_tag_nxc
                     : flat                    ? dat_tag_ncx
                                               : dat_tag_nCx8c;