[CPU] only allow per-oc or per-tensor FQ fusing into FC (openvinotool…

…kit#25530) ### Details: - Add a check to reject non-supported FakeQuantize from fusing into FC node, so they can run in standalone mode w/o causing exceptions when composing oneDNN postOps. - port from openvinotoolkit#23009 - add test case ### Tickets: - *CVS-131890* --------- Signed-off-by: HU Yuan2 <[email protected]> Co-authored-by: Li, Tingqian <[email protected]>
evkotov · Jul 25, 2024 · 0c598f4 · 0c598f4
1 parent b9d98cb
commit 0c598f4
Show file tree

Hide file tree

Showing 4 changed files with 86 additions and 1 deletion.
diff --git a/src/plugins/intel_cpu/src/nodes/fullyconnected.cpp b/src/plugins/intel_cpu/src/nodes/fullyconnected.cpp
@@ -25,6 +25,8 @@
 #include "utils/debug_capabilities.h"
 #include "utils/general_utils.h"
 
+#include "fake_quantize.h"
+
 using namespace dnnl;
 using namespace ov::element;
 
@@ -94,6 +96,25 @@ bool FullyConnected::canFuse(const NodePtr& node) const {
 #if defined(OV_CPU_WITH_SHL)
     return false;
 #endif
+    if (node->getType() == Type::FakeQuantize) {
+        auto* fq = dynamic_cast<FakeQuantize*>(node.get());
+        if (fq->getBroadcastingPolicy() != FakeQuantize::BroadcastingPolicy::PerTensor) {
+            const auto& dstShape = getOutputShapeAtPort(0);
+            auto dataRanks = dstShape.getRank();
+            // only per-OC or per-Tensor fakequantize can be postOps
+            if (fq->getAxis() != dataRanks - 1) {
+                DEBUG_LOG("reject FakeQuantize ",
+                          fq->getName(),
+                          "(axis=",
+                          fq->getAxis(),
+                          ") from fusing into ",
+                          getName(),
+                          " with dst shape ",
+                          dstShape);
+                return false;
+            }
+        }
+    }
     return canFuseSimpleOperation(node);
 }
 

diff --git a/src/plugins/intel_cpu/tests/functional/custom/single_layer_tests/instances/x64/matmul.cpp b/src/plugins/intel_cpu/tests/functional/custom/single_layer_tests/instances/x64/matmul.cpp
@@ -1108,6 +1108,45 @@ INSTANTIATE_TEST_SUITE_P(
     testParamsDynamicFusingFullUndefShapes,
     MatMulLayerCPUTest::getTestCaseName);
 
+class FCNotFuseFQCPUTest : public MatMulLayerCPUTest {
+    void SetUp() override {
+        MatMulLayerCPUTest::SetUp();
+        expectPostOpsToBeFused = false;
+    }
+};
+
+TEST_P(FCNotFuseFQCPUTest, CompareWithRefs) {
+    run();
+    CheckPluginRelatedResults(compiledModel, cpuNodeType);
+}
+
+const std::vector<ShapeRelatedParams>& notFuseSmoke() {
+    static const std::vector<ShapeRelatedParams> params = {
+        {static_shapes_to_test_representation({{59, 1}, {1, 120}}), {false, true}},
+        {static_shapes_to_test_representation({{59, 1}, {1, 120}}), {true, true}},
+
+        {static_shapes_to_test_representation({{59, 120}, {120, 1}}), {false, false}},
+        {static_shapes_to_test_representation({{59, 120}, {120, 1}}), {true, true}},
+
+        {static_shapes_to_test_representation({{71, 128}, {128, 20}}), {true, false}},
+        {static_shapes_to_test_representation({{71, 128}, {128, 20}}), {false, true}},
+    };
+    return params;
+}
+
+const auto notFuseTestParamsSmoke = ::testing::Combine(::testing::Combine(::testing::ValuesIn(notFuseSmoke()),
+                                                                          ::testing::Values(ElementType::f32),
+                                                                          ::testing::Values(ElementType::undefined),
+                                                                          ::testing::Values(ElementType::undefined),
+                                                                          ::testing::Values(utils::InputLayerType::CONSTANT),
+                                                                          ::testing::Values(ov::test::utils::DEVICE_CPU),
+                                                                          ::testing::Values(emptyAdditionalConfig())),
+                                                       ::testing::Values(MatMulNodeType::FullyConnected),
+                                                       ::testing::ValuesIn({fusingFakeQuantizePerBatch, fusingFakeQuantizeFullTensor}),
+                                                       ::testing::ValuesIn({CPUSpecificParams{{}, {}, {""}, "any_type"}}));
+
+INSTANTIATE_TEST_SUITE_P(smoke_FC, FCNotFuseFQCPUTest, notFuseTestParamsSmoke, FCNotFuseFQCPUTest::getTestCaseName);
+
 }  // namespace
 }  // namespace MatMul
 }  // namespace test

diff --git a/src/plugins/intel_cpu/tests/functional/utils/fusing_test_utils.cpp b/src/plugins/intel_cpu/tests/functional/utils/fusing_test_utils.cpp
@@ -58,7 +58,11 @@ void CpuTestWithFusing::CheckFusingResults(const std::shared_ptr<const ov::Model
             size_t pos = 0;
             for (const auto& fusedOp : fusedOps) {
                 pos = originalLayersNames.find(fusedOp, checkFusingPosition ? pos : 0);
-                ASSERT_TRUE(pos != std::string::npos) << "Fused op " << fusedOp << " has not been found!";
+                if (expectPostOpsToBeFused) {
+                    ASSERT_TRUE(pos != std::string::npos) << "Fused op " << fusedOp << " has not been found!";
+                } else {
+                    ASSERT_TRUE(pos == std::string::npos) << "op" << fusedOp << " should not be fused!";
+                }
             }
         }
     }

diff --git a/src/plugins/intel_cpu/tests/functional/utils/fusing_test_utils.hpp b/src/plugins/intel_cpu/tests/functional/utils/fusing_test_utils.hpp
@@ -93,6 +93,7 @@ class CpuTestWithFusing : public CPUTestsBase {
     std::shared_ptr<postOpMgr> postOpMgrPtr;
     std::vector<std::string> fusedOps;
     bool checkFusingPosition = true;
+    bool expectPostOpsToBeFused = true;
 };
 
 static int getChannelAxis(const ov::AxisSet &axes, bool keep_dims) {
@@ -304,6 +305,26 @@ const auto fusingFakeQuantizePerChannel = fusingSpecificParams{std::make_shared<
                 return ov::test::utils::make_fake_quantize(cfg.input, localPrc, 256, newShape);
             }, "FakeQuantize(PerChannel)"}}), {"FakeQuantize"}};
 
+const auto fusingFakeQuantizePerBatch = fusingSpecificParams{std::make_shared<postNodesMgr>(std::vector<postNodeBuilder>{
+            {[](postNodeConfig& cfg){
+                auto localPrc = cfg.input->get_element_type();
+                const auto shape = cfg.input->get_output_partial_shape(0);
+                ov::Shape perBatchSize(shape.size(), 1);
+                perBatchSize[0] = shape[0].get_length();
+                return ov::test::utils::make_fake_quantize(cfg.input, localPrc, 256, perBatchSize);
+            }, "FakeQuantize(PerBatch)"}}), {"FakeQuantize"}};
+
+const auto fusingFakeQuantizeFullTensor = fusingSpecificParams{std::make_shared<postNodesMgr>(std::vector<postNodeBuilder>{
+            {[](postNodeConfig& cfg){
+                auto localPrc = cfg.input->get_element_type();
+                const auto shape = cfg.input->get_output_partial_shape(0);
+                ov::Shape fullTensorShape(shape.size(), 1);
+                for (size_t axis = 0; axis < shape.size(); axis++) {
+                    fullTensorShape[axis] = shape[axis].get_length();
+                }
+                return ov::test::utils::make_fake_quantize(cfg.input, localPrc, 256, fullTensorShape);
+            }, "FakeQuantize(FullTensor)"}}), {"FakeQuantize"}};
+
 const auto fusingFakeQuantizePerChannelRelu = fusingSpecificParams{std::make_shared<postNodesMgr>(std::vector<postNodeBuilder>{
             {[](postNodeConfig& cfg){
                 auto localPrc = cfg.input->get_element_type();