cpu: avx512_core: fix access to padded areas in dw bwd/wu convs

Author: nshustrov

src/cpu/jit_avx512_core_bf16_dw_conv_kernel.cpp

@@ -557,6 +557,8 @@ inline void jit_avx512_dw_conv_bwd_weights_kernel_bf16::compute_ow_step_unroll(
         if (i_ur == 0) {
             for (int c = 0; c < input_overlap; ++c) {
                 int off_input = (c - pad_offset) * jcp.ch_block;
+                if (off_input < 0 && unroll_w == jcp.ow)
+                        continue;
                 Zmm zmm_input = get_input_reg(c);
                 vpmovzxwd(zmm_input,
                         ptr[reg_tmp_input + off_input * jcp.typesize_in]);
@@ -565,6 +567,8 @@ inline void jit_avx512_dw_conv_bwd_weights_kernel_bf16::compute_ow_step_unroll(
             for (int c = 0; c < cascade_input; ++c) {
                 int overlap = (i_ur - 1) * jcp.stride_w + input_overlap;
                 int off_input = (overlap + c - pad_offset) * jcp.ch_block;
+                if (off_input < 0 || overlap + c + l_pad > right_border)
+                    continue;
                 Zmm zmm_input = get_input_reg(overlap + c);
                 vpmovzxwd(zmm_input,
                         ptr[reg_tmp_input + off_input * jcp.typesize_in]);

src/cpu/jit_uni_dw_conv_kernel_f32.cpp

@@ -557,6 +557,8 @@ inline void jit_uni_dw_conv_bwd_weights_kernel_f32<isa>::compute_ow_step_unroll(
                 for (int c = 0; c < input_overlap; ++c) {
                     int off_input
                             = ((c - pad_offset) * reg_repeats + r) * simd_w;
+                    if (off_input < 0 && unroll_w == jcp.ow)
+                        continue;
                     Vmm vmm_input
                             = get_input_reg((c % jcp.kw) * reg_repeats + r);
                     uni_vmovups(vmm_input,
@@ -568,6 +570,8 @@ inline void jit_uni_dw_conv_bwd_weights_kernel_f32<isa>::compute_ow_step_unroll(
                     int off_input
                             = ((overlap + c - pad_offset) * reg_repeats + r)
                             * simd_w;
+                    if (off_input < 0 || overlap + c + l_pad > right_border)
+                        continue;
                     Vmm vmm_input = get_input_reg(
                             ((overlap + c) % jcp.kw) * reg_repeats + r);
                     uni_vmovups(vmm_input,