[CPU] Enable back INT32 for MatMul (#7898)

- Also enabled ngraph onnx matmul integer tests

inference-engine/src/mkldnn_plugin/nodes/mkldnn_matmul_node.cpp

@@ -4,6 +4,7 @@
 
 #include "mkldnn_matmul_node.h"
 
+#include "ie_precision.hpp"
 #include "memory_desc/cpu_blocked_memory_desc.h"
 #include "cpu_types.h"
 #include "mkldnn_eltwise_node.h"
@@ -120,6 +121,12 @@ void MKLDNNMatMulNode::getSupportedDescriptors() {
     if (firstInPortPrec.size() != secondInPortPrec.size())
         firstInPortPrec = secondInPortPrec = getMaxPrecision(getOriginalInputPrecisions());
 
+    // fallback to fp32 for any precision that cannot be handled natively
+    if ((!one_of(firstInPortPrec , Precision::U8, Precision::I8, Precision::BF16, Precision::FP32) ||
+         !one_of(secondInPortPrec , Precision::I8, Precision::BF16, Precision::FP32))) {
+        outPortPrec = firstInPortPrec = secondInPortPrec = Precision::FP32;
+    }
+
     if (!fusedWith.empty()) {
         outPortPrec = fusedWith[fusedWith.size() - 1]->getOriginalOutputPrecisionAtPort(0);
     }

ngraph/test/runtime/ie/unit_test.manifest

@@ -31,16 +31,6 @@ onnx_model_conv_integer_no_zero_point
 onnx_model_conv_integer_pads
 
 # Unsupported operator detected in the graph: QuantizedDot
-onnx_model_matmul_integer_2d_simple_zero_point
-onnx_model_matmul_integer_int8
-onnx_model_matmul_integer_vectorized_zero_point
-onnx_model_matmul_integer_no_zero_point
-onnx_model_matmul_integer_2d_x_3d
-onnx_model_matmul_integer_3d_x_2d
-onnx_model_matmul_integer_3d
-onnx_model_matmul_integer_4d
-onnx_model_matmul_integer_4d_zero_point
-onnx_model_matmul_integer_matrix_zero_point
 onnx_model_qlinear_matmul
 onnx_model_qlinear_matmul_3d
 

runtime/bindings/python/tests/test_onnx/test_backend.py

@@ -333,7 +333,6 @@ tests_expected_to_fail = [
     ),
     (
         xfail_issue_38722,
-        "OnnxBackendNodeModelTest.test_matmulinteger_cpu",
         "OnnxBackendNodeModelTest.test_qlinearmatmul_2D_cpu",
         "OnnxBackendNodeModelTest.test_qlinearmatmul_3D_cpu",
     ),