[CPU] Enabled brconv implementation

Author: dmitrygo

src/plugins/intel_cpu/src/node.cpp

@@ -868,9 +868,8 @@ const std::vector<impl_desc_type>& MKLDNNNode::getPrimitivesPriority() {
             impl_desc_type::jit_avx512_amx_dw,
             impl_desc_type::jit_avx512_amx_1x1,
             impl_desc_type::jit_avx512_amx,
-            // Brgconv kernels disabled in order to prevent perf degradations on non AMX HW
-//            impl_desc_type::brgconv_avx512_1x1,
-//            impl_desc_type::brgconv_avx512,
+            impl_desc_type::brgconv_avx512_1x1,
+            impl_desc_type::brgconv_avx512,
             impl_desc_type::jit_uni_dw,
             impl_desc_type::jit_uni_1x1,
             impl_desc_type::jit_uni,

src/plugins/intel_cpu/src/nodes/conv.cpp

@@ -465,6 +465,13 @@ void MKLDNNConvolutionNode::getSupportedDescriptors() {
             auto inputShape = getInputShapeAtPort(0);
             auto outputShape = getOutputShapeAtPort(0);
 
+            if (one_of(inputDataType, memory::data_type::f32, memory::data_type::bf16) &&
+                    impl::cpu::x64::mayiuse(impl::cpu::x64::avx512_core)) {
+                in_candidate = std::make_shared<DnnlBlockedMemoryDesc>(inputShape, inputDataType, nspc);
+                out_candidate = std::make_shared<DnnlBlockedMemoryDesc>(outputShape, outputDataType, nspc);
+                createDescriptor({ in_candidate }, { out_candidate });
+            }
+
             if (IC == 1 && groupOC == 1) {
                 in_candidate = std::make_shared<DnnlBlockedMemoryDesc>(inputShape, inputDataType, ncsp);
                 out_candidate = std::make_shared<DnnlBlockedMemoryDesc>(outputShape, outputDataType, ncsp);
@@ -488,7 +495,9 @@ void MKLDNNConvolutionNode::getSupportedDescriptors() {
             out_candidate = std::make_shared<DnnlBlockedMemoryDesc>(outputShape, outputDataType, ncsp);
             createDescriptor({ in_candidate }, { out_candidate });
 
-            if (inputDataType != memory::data_type::bf16 && isNspcAvailable()) {
+            if ((inputDataType != memory::data_type::bf16 && isNspcAvailable()) ||
+                    (one_of(inputDataType, memory::data_type::f32, memory::data_type::bf16) &&
+                    impl::cpu::x64::mayiuse(impl::cpu::x64::avx512_core))) {
                 in_candidate = std::make_shared<DnnlBlockedMemoryDesc>(inputShape, inputDataType, nspc);
                 out_candidate = std::make_shared<DnnlBlockedMemoryDesc>(outputShape, outputDataType, nspc);
                 createDescriptor({ in_candidate }, { out_candidate });