WIP

openvinotoolkit · Jul 19, 2021 · a3cffa1 · a3cffa1
1 parent ad59d7e
commit a3cffa1
Show file tree

Hide file tree

Showing 3 changed files with 27 additions and 25 deletions.
diff --git a/inference-engine/tests/functional/plugin/cpu/single_layer_tests/convolution.cpp b/inference-engine/tests/functional/plugin/cpu/single_layer_tests/convolution.cpp
@@ -77,6 +77,15 @@ class ConvolutionLayerCPUTest : public testing::WithParamInterface<convLayerCPUT
         ASSERT_TRUE(foundConv) << "Can't find Convolution node";
     }
 
+    int calculateQuantizeInHigh(const InferenceEngine::SizeVector& kernel, const int ic, const int maxIn0 = 10, const int maxIn1 = 10) const {
+        auto quantizeInHigh = maxIn0 * maxIn1;
+        quantizeInHigh *= ic;
+        for (int i = 0; i < kernel.size(); i++) {
+            quantizeInHigh *= kernel[i];
+        }
+        return quantizeInHigh;
+    }
+
     void SetUp() override {
         using namespace ngraph;
         convLayerTestParamsSet basicParamsSet;
@@ -123,23 +132,19 @@ class ConvolutionLayerCPUTest : public testing::WithParamInterface<convLayerCPUT
         auto convolutionNode = builder::makeConvolutionRelaxed(paramOuts.front(), weiPrc, kernel, stride, padBegin,
             padEnd, dilation, padType, convOutChannels);
 
-        // todo: scale calculation
-        scale = inputShape[1];
-        for (int i = 0; i < kernel.size(); i++) {
-            scale *= kernel[i];
-        }
-        scale *= 10 * 10;
-        scale += 1;
-
-
-        // todo: FuncTestUtils::PrecisionUtils::convertIE2nGraphPrc(outPrc)?
-        function = makeNgraphFunction(element::f32, inputParams, convolutionNode, "Convolution", FuncTestUtils::PrecisionUtils::convertIE2nGraphPrc(outPrc));
-
         if (inPrc == Precision::U8 || inPrc == Precision::I8) {
+            threshold = 1.001f;
+            outElemType = FuncTestUtils::PrecisionUtils::convertIE2nGraphPrc(outPrc);
+            quantizeInHigh = calculateQuantizeInHigh(kernel, inputShape[1]);
             additionalPasses.push_back(std::make_shared<pass::ConvertPrecision<element::i8, element::f32>>());
             additionalPasses.push_back(std::make_shared<pass::ConvertPrecision<element::u8, element::f32>>());
         }
 
+        // todo: FuncTestUtils::PrecisionUtils::convertIE2nGraphPrc(outPrc)?
+        function = makeNgraphFunction(element::f32, inputParams, convolutionNode, "Convolution");
+
+
+
         // todo: is it needed?
 //        if (outPrc != Precision::FP32 && outPrc != Precision::BF16) {
 //            additionalPasses.push_back(std::make_shared<ConvertPrecision<opset1::Convolution>>());
@@ -277,8 +282,6 @@ INSTANTIATE_TEST_SUITE_P(smoke_Conv_2D_GEMM_BF16, ConvolutionLayerCPUTest,
         ::testing::Values(cpuBF16PluginConfig)),
     ConvolutionLayerCPUTest::getTestCaseName);
 
-// todo:
-
 
 /* ============= Convolution (GEMM 3D) ============= */
 const auto convParams_ExplicitPadding_GEMM_3D = ::testing::Combine(
@@ -528,8 +531,6 @@ const std::vector<CPUSpecificParams> CPUParams_1x1_2D = {
     conv_avx512_2D_1x1_nspc
 };
 
-
-
 INSTANTIATE_TEST_SUITE_P(smoke_Conv_2D_1x1_FP32, ConvolutionLayerCPUTest,
     ::testing::Combine(
         ::testing::Combine(

diff --git a/inference-engine/tests/functional/plugin/cpu/test_utils/cpu_test_utils.cpp b/inference-engine/tests/functional/plugin/cpu/test_utils/cpu_test_utils.cpp
@@ -241,20 +241,19 @@ CPUTestsBase::makeCPUInfo(std::vector<cpu_memory_format_t> inFmts, std::vector<c
 
 std::shared_ptr<ngraph::Function>
 CPUTestsBase::makeNgraphFunction(const ngraph::element::Type &ngPrc, ngraph::ParameterVector &params,
-                                 const std::shared_ptr<ngraph::Node> &lastNode, std::string name,
-                                 const ngraph::element::Type outPrc) const {
+                                 const std::shared_ptr<ngraph::Node> &lastNode, std::string name) const {
    auto newLastNode = modifyGraph(ngPrc, params, lastNode);
    ngraph::ResultVector results;
 
-   if (outPrc == ngraph::element::u8) {
+   if (outElemType == ngraph::element::u8) {
        const std::vector<float> inLow = {0};
-       const std::vector<float> inHigh = {static_cast<float>(scale)};
+       const std::vector<float> inHigh = {static_cast<float>(quantizeInHigh)};
        const std::vector<float> outLow = {0};
        const std::vector<float> outHigh = {255};
        newLastNode = ngraph::builder::makeFakeQuantize(newLastNode, ngPrc, 256, {1, 1, 1, 1}, inLow, inHigh, outLow, outHigh);
-   } else if (outPrc == ngraph::element::i8) {
+   } else if (outElemType == ngraph::element::i8) {
        const std::vector<float> inLow = {0};
-       const std::vector<float> inHigh = {static_cast<float>(scale)};
+       const std::vector<float> inHigh = {static_cast<float>(quantizeInHigh)};
        const std::vector<float> outLow = {-127};
        const std::vector<float> outHigh = {127};
        newLastNode = ngraph::builder::makeFakeQuantize(newLastNode, ngPrc, 255, {1, 1, 1, 1}, inLow, inHigh, outLow, outHigh);

diff --git a/inference-engine/tests/functional/plugin/cpu/test_utils/cpu_test_utils.hpp b/inference-engine/tests/functional/plugin/cpu/test_utils/cpu_test_utils.hpp
@@ -125,8 +125,7 @@ class CPUTestsBase {
     std::shared_ptr<ngraph::Function> makeNgraphFunction(const ngraph::element::Type &ngPrc,
                                                          ngraph::ParameterVector &params,
                                                          const std::shared_ptr<ngraph::Node> &lastNode,
-                                                         std::string name,
-                                                         const ngraph::element::Type outPrc = ngraph::element::f32) const;
+                                                         std::string name) const;
 
 protected:
     virtual void CheckPluginRelatedResults(InferenceEngine::ExecutableNetwork &execNet, std::string nodeType) const;
@@ -146,7 +145,10 @@ class CPUTestsBase {
     std::vector<cpu_memory_format_t> inFmts, outFmts;
     std::vector<std::string> priority;
     std::string selectedType;
-    int scale = 1;
+
+    ngraph::element::Type outElemType = ngraph::element::f32;
+    // only for int8 testing
+    int quantizeInHigh = 1;
 };
 
 template <typename NodeType>