Fixed review comments

dmitry-gorokhov · Aug 3, 2023 · ca54646 · ca54646
1 parent ec9cbd9
commit ca54646
Show file tree

Hide file tree

Showing 4 changed files with 11 additions and 9 deletions.
diff --git a/src/plugins/intel_cpu/src/graph_optimizer.cpp b/src/plugins/intel_cpu/src/graph_optimizer.cpp
@@ -287,15 +287,15 @@ void GraphOptimizer::FuseFCAndWeightsDecompression(Graph &graph) {
         return node->getType() == expectedType && node->getChildEdges().size() == 1;
     };
 
+    if (!impl::cpu::x64::mayiuse(impl::cpu::x64::avx2))
+        return;
+
     auto& graphNodes = graph.GetNodes();
     for (size_t i = 0; i < graphNodes.size(); i++) {
         const auto fcNode = dynamic_cast<node::FullyConnected*>(graphNodes[i].get());
         if (fcNode == nullptr)
             continue;
 
-        if (!impl::cpu::x64::mayiuse(impl::cpu::x64::avx2))
-            continue;
-
         const auto parent = fcNode->getParentEdgesAtPort(1)[0]->getParent();
         const bool withTranspose = parent->getType() == Type::Transpose;
         const NodePtr transposeNode = withTranspose ? parent : nullptr;

diff --git a/src/plugins/intel_cpu/src/nodes/fullyconnected.cpp b/src/plugins/intel_cpu/src/nodes/fullyconnected.cpp
@@ -590,7 +590,8 @@ void FullyConnected::setPostOps(dnnl::primitive_attr& attr, const VectorDims& di
     DnnlPostOpsComposer dnnlpoc(getEngine(), attr, ops, postOpsArgs, dims, dims.size() - 1, canBeExecutedInInt8(),
                                     1 << 0,  getDQScales(), withBiases);
 
-    dnnlpoc.appendDecompressionScales(decompressionMultiply);
+    if (!decompressionMultiply.empty())
+        dnnlpoc.appendDecompressionScales(decompressionMultiply);
     if (!decompressionSubtract.empty())
         dnnlpoc.appendDecompressionZeroPoints(decompressionSubtract);
 
@@ -680,10 +681,9 @@ void FullyConnected::createDescriptorInternal(const dnnl::memory::desc &inputDes
     dnnl::memory::data_type wdt = indt;
     dnnl::memory::data_type bdt = outdt;
 
-    dnnl::memory::data_type original_wdt = DnnlExtensionUtils::IEPrecisionToDataType(getOriginalInputPrecisionAtPort(WEIGHTS_ID));
-    if (indt == dnnl::memory::data_type::f32 && original_wdt == dnnl::memory::data_type::u8) {
+    if (useWeightsDecompressionImpl) {
         // Weights decompression case
-        wdt = original_wdt;
+        wdt = DnnlExtensionUtils::IEPrecisionToDataType(getOriginalInputPrecisionAtPort(WEIGHTS_ID));
     } else if (one_of(indt, dnnl::memory::data_type::bf16, dnnl::memory::data_type::f16)) {
 #if defined(OPENVINO_ARCH_X86_64)
         bdt = dnnl::memory::data_type::f32;

diff --git a/src/plugins/intel_cpu/tests/functional/subgraph_tests/src/matmul_weights_decompression.cpp b/src/plugins/intel_cpu/tests/functional/subgraph_tests/src/matmul_weights_decompression.cpp
@@ -189,7 +189,9 @@ class MatmulWeightsDecompression : public testing::WithParamInterface<MatmulWeig
         }
 
         std::map<std::string, std::string> additional_config = std::get<5>(test_param);
-        const size_t expected_count = additional_config[PluginConfigParams::KEY_ENFORCE_BF16] == PluginConfigParams::YES ? 1 : 0;
+        const size_t expected_count =
+            InferenceEngine::with_cpu_x86_avx2() &&
+            additional_config[PluginConfigParams::KEY_ENFORCE_BF16] != PluginConfigParams::YES ? 0 : 1;
         CheckNumberOfNodesWithType(compiledModel, "Convert", expected_count);
         CheckNumberOfNodesWithType(compiledModel, "Eltwise", expected_count);
         CheckNumberOfNodesWithType(compiledModel, "Subgraph", 0);

diff --git a/src/plugins/intel_cpu/thirdparty/onednn b/src/plugins/intel_cpu/thirdparty/onednn