[CPU] Fix Reduce node bug regarding fusing with low precision (#18835)

src/plugins/intel_cpu/src/nodes/reduce.cpp

@@ -2154,7 +2154,7 @@ void Reduce::createPrimitive() {
 
     auto reduce_jcp = jcp;
     reduce_jcp.dst_dt = fuse_low_precision ? DnnlExtensionUtils::IEPrecisionToDataType(intermediate_prec) : jcp.dst_dt;
-    jcp.dst_data_size = DnnlExtensionUtils::sizeOfDataType(reduce_jcp.dst_dt);
+    reduce_jcp.dst_data_size = DnnlExtensionUtils::sizeOfDataType(reduce_jcp.dst_dt);
     create_reduce_kernel(reduce_kernel, reduce_jcp);
 
     // set_use_aux_kernel being false means this is a dynamic case, and prepareParams() hasn't been invoked yet.
@@ -2255,7 +2255,7 @@ void Reduce::reduce_type(const uint8_t *in_ptr, uint8_t *out_ptr) {
 }
 
 void Reduce::reduce_PLN(const uint8_t *in_ptr, uint8_t *out_ptr) {
-    output_info_reassign(out_ptr);
+    output_info_reassign(&out_ptr);
     init_dst_data(out_ptr, dst_size);
 
     if (ReduceN && !ReduceC && !ReduceD && !ReduceH && !ReduceW) {
@@ -2472,7 +2472,7 @@ void Reduce::reduce_PLN(const uint8_t *in_ptr, uint8_t *out_ptr) {
 void Reduce::reduce_BLK(const uint8_t *in_ptr, uint8_t *out_ptr) {
     size_t ICB = div_up(IC, blk_size);
     size_t OCB = div_up(OC, blk_size);
-    output_info_reassign(out_ptr);
+    output_info_reassign(&out_ptr);
     init_dst_data(out_ptr, dst_size);
 
     for (size_t ib = 0; ib < IB; ib++) {
@@ -2552,7 +2552,7 @@ void Reduce::reduce_BLK(const uint8_t *in_ptr, uint8_t *out_ptr) {
 void Reduce::reduce_BLK_concern_padding(const uint8_t *in_ptr, uint8_t *out_ptr) {
     size_t ICB = div_up(IC, blk_size);
     size_t OCB = div_up(OC, blk_size);
-    output_info_reassign(out_ptr);
+    output_info_reassign(&out_ptr);
     init_dst_data(out_ptr, dst_size);
 
     auto reduceSkipPadding = [&](const uint8_t *in_ptr_ncd, uint8_t *out_ptr_ncd, size_t ic) {
@@ -2720,10 +2720,10 @@ inline void Reduce::reduce_kernel_restore() {
     }
 }
 
-inline void Reduce::output_info_reassign(uint8_t *out_ptr) {
+inline void Reduce::output_info_reassign(uint8_t **out_ptr) {
     if (fuse_low_precision) {
-        tmp_ptr = out_ptr;
-        out_ptr = static_cast<uint8_t *>(&intermediate_buf[0]);
+        tmp_ptr = *out_ptr;
+        *out_ptr = static_cast<uint8_t *>(&intermediate_buf[0]);
         tmp_prec = output_prec;
         output_prec = intermediate_prec;
         tmp_data_size = dst_data_size;

src/plugins/intel_cpu/src/nodes/reduce.h

@@ -117,7 +117,7 @@ private:
     inline void reduce_kernel_post_process(uint8_t *out_ptr);
     inline void reduce_kernel_reassign();
     inline void reduce_kernel_restore();
-    inline void output_info_reassign(uint8_t *out_ptr);
+    inline void output_info_reassign(uint8_t **out_ptr);
     inline void output_info_restore(uint8_t **out_ptr);
     inline void init_dst_data(uint8_t *out_ptr, size_t dst_size);
     inline void create_hybrid_working_memory();

src/plugins/intel_cpu/tests/functional/single_layer_tests/instances/x64/reduce.cpp

@@ -6,6 +6,7 @@
 #include "shared_test_classes/single_layer/reduce_ops.hpp"
 #include "test_utils/cpu_test_utils.hpp"
 #include "test_utils/fusing_test_utils.hpp"
+#include "lpt_ngraph_functions/common/builders.hpp"
 
 using namespace InferenceEngine;
 using namespace CPUTestUtils;
@@ -110,6 +111,22 @@ const std::vector<ngraph::helpers::ReductionType> reductionTypesFusing = {
         ngraph::helpers::ReductionType::L2,
 };
 
+// This custom subgraph is used to test post-ops fusing case with U8/I8 precision on output,
+// since Transpose prevents dequantization part to be fused back into Reduce
+const auto fusingFakeQuantizeTranspose = fusingSpecificParams{std::make_shared<postNodesMgr>(std::vector<postNodeBuilder>{
+        {[](postNodeConfig& cfg){
+            auto localPrc = cfg.input->get_element_type();
+            ngraph::Shape newShape(cfg.input->get_output_partial_shape(0).size(), 1);
+            const auto fakeQuantize = ngraph::builder::makeFakeQuantize(cfg.input, localPrc, 256, newShape);
+            std::vector<size_t> order(newShape.size());
+            std::iota(order.begin(), order.end(), 0);
+            auto last = order[order.size() - 1];
+            order.pop_back();
+            order.insert(order.begin(), last);
+            const auto transpose = ngraph::builder::subgraph::Transpose(order);
+            return ngraph::builder::subgraph::makeTranspose(fakeQuantize, transpose);
+        }, "FakeQuantize(PerTensor)"}}), {"FakeQuantize"}};
+
 const std::vector<fusingSpecificParams> fusingParamsSet {
         /* activations */
         fusingSwish,
@@ -132,6 +149,10 @@ const std::vector<fusingSpecificParams> fusingParamsSet_KeepNoDims {
         fusingScaleShift
 };
 
+const std::vector<fusingSpecificParams> fusingParamsSet_LowPrecision {
+        fusingFakeQuantizeTranspose
+};
+
 /* ================================ 1.1 No fusion - Arithmetic ================================ */
 const auto params_OneAxis = testing::Combine(
         testing::Combine(
@@ -492,6 +513,20 @@ const auto params_MultiAxis_5D_fusing = testing::Combine(
         testing::ValuesIn(fusingParamsSet),
         testing::ValuesIn(additionalConfig()));
 
+const auto params_LowPrecision_fusing = testing::Combine(
+        testing::Combine(
+                testing::ValuesIn(axesNDFusing),
+                testing::Values(ov::test::utils::OpType::VECTOR),
+                testing::Values(true),
+                testing::ValuesIn(reductionTypesFusing),
+                testing::ValuesIn(inpOutPrc()),
+                testing::Values(ElementType::undefined),
+                testing::Values(ElementType::undefined),
+                testing::ValuesIn(inputShapes_dyn)),
+        testing::ValuesIn(filterCPUSpecificParams(cpuParams_4D)),
+        testing::ValuesIn(fusingParamsSet_LowPrecision),
+        testing::ValuesIn(additionalConfig()));
+
 INSTANTIATE_TEST_SUITE_P(
         smoke_Reduce_OneAxis_fusing_CPU,
         ReduceCPULayerTest,
@@ -513,6 +548,13 @@ INSTANTIATE_TEST_SUITE_P(
         ReduceCPULayerTest::getTestCaseName
 );
 
+INSTANTIATE_TEST_SUITE_P(
+        smoke_Reduce_LowPrecision_fusing_CPU,
+        ReduceCPULayerTest,
+        params_LowPrecision_fusing,
+        ReduceCPULayerTest::getTestCaseName
+);
+
 /* ================================ 2.2 Fusion - KeepNoDims ================================ */
 const auto params_OneAxis_fusing_KeepNoDims = testing::Combine(
         testing::Combine(