IVGCVSW-5422 Add per-channel quantization tests to depthwise

Signed-off-by: Jan Eilers <jan.eilers@arm.com>
Change-Id: Ie51ce20540e5e7a704ce2b4be4e8cf64f91ea990

1 files changed, 360 insertions, 0 deletions

diff --git a/src/armnnTfLiteParser/test/DepthwiseConvolution2D.cpp b/src/armnnTfLiteParser/test/DepthwiseConvolution2D.cpp
index 5d220ebf23..7380d884fd 100644
--- a/src/armnnTfLiteParser/test/DepthwiseConvolution2D.cpp
+++ b/src/armnnTfLiteParser/test/DepthwiseConvolution2D.cpp
@@ -222,4 +222,364 @@ BOOST_FIXTURE_TEST_CASE(ParseDynamicDepthwiseConv2DSameBias, DynamicDepthwiseCon
                                                       true);
 }
 
+struct DepthwiseConvolution2dFixture2 : public ParserFlatbuffersFixture
+{
+    explicit DepthwiseConvolution2dFixture2(const std::string& inputShape,
+                                            const std::string& outputShape,
+                                            const std::string& filterShape,
+                                            const std::string& filterData,
+                                            const std::string& strides,
+                                            const std::string& paddingType,
+                                            const std::string  biasShape                = "",
+                                            const std::string  biasData                 = "",
+                                            const std::string  filter_quant_min         = "[ 0.0 ]",
+                                            const std::string  filter_quant_max         = "[ 255.0 ]",
+                                            const std::string  filter_quant_scale       = "[ 1.0 ]",
+                                            const std::string  filter_quant_zero_point  = "[ 0 ]",
+                                            const std::string  filter_quant_axis        = ""
+                                            )
+    {
+        std::string inputTensors = "[ 0, 2 ]";
+        std::string biasTensor   = "";
+        std::string biasBuffer   = "";
+        if (biasShape.size() > 0 && biasData.size() > 0)
+        {
+            inputTensors = "[ 0, 2, 3 ]";
+            biasTensor = R"(
+                        {
+                            "shape": )" + biasShape + R"( ,
+                            "type": "INT32",
+                            "buffer": 3,
+                            "name": "biasTensor",
+                            "quantization": {
+                                "min": [ 0.0 ],
+                                "max": [ 255.0 ],
+                                "scale": [ 1.0 ],
+                                "zero_point": [ 0 ],
+                            }
+                        } )";
+            biasBuffer = R"(
+                    { "data": )" + biasData + R"(, }, )";
+        }
+
+        std::string filter_qantization =
+                               R"(
+                                "min": )" + filter_quant_min + R"(,
+                                "max": )" + filter_quant_max + R"(,
+                                "scale": )" + filter_quant_scale + R"(,
+                                "zero_point": )" + filter_quant_zero_point;
+        // A given quantization axis indicates if per channel quantization is used for filters
+        if (filter_quant_axis.size() > 0)
+        {
+            filter_qantization +=
+                               R"(,
+                                "quantized_dimension": )" + filter_quant_axis;
+        }
+        m_JsonString = R"(
+            {
+                "version": 3,
+                "operator_codes": [ { "builtin_code": "DEPTHWISE_CONV_2D" } ],
+                "subgraphs": [ {
+                    "tensors": [
+                        {
+                            "shape": )" + inputShape + R"(,
+                            "type": "INT8",
+                            "buffer": 0,
+                            "name": "inputTensor",
+                            "quantization": {
+                                "min": [ 0.0 ],
+                                "max": [ 255.0 ],
+                                "scale": [ 1.0 ],
+                                "zero_point": [ 0 ],
+                            }
+                        },
+                        {
+                            "shape": )" + outputShape + R"(,
+                            "type": "INT8",
+                            "buffer": 1,
+                            "name": "outputTensor",
+                            "quantization": {
+                                "min": [ 0.0 ],
+                                "max": [ 511.0 ],
+                                "scale": [ 1.0 ],
+                                "zero_point": [ 0 ],
+                            }
+                        },
+                        {
+                            "shape": )" + filterShape + R"(,
+                            "type": "INT8",
+                            "buffer": 2,
+                            "name": "filterTensor",
+                            "quantization": {)" + filter_qantization + R"(
+                            }
+                        }, )" + biasTensor + R"(
+                    ],
+                    "inputs": [ 0 ],
+                    "outputs": [ 1 ],
+                    "operators": [
+                        {
+                            "opcode_index": 0,
+                            "inputs": )" + inputTensors + R"(,
+                            "outputs": [ 1 ],
+                            "builtin_options_type": "DepthwiseConv2DOptions",
+                            "builtin_options": {
+                                "padding": ")" + paddingType + R"(",
+                                "stride_w": )" + strides+ R"(,
+                                "stride_h": )" + strides+ R"(,
+                                "depth_multiplier": 1,
+                                "fused_activation_function": "NONE"
+                            },
+                            "custom_options_format": "FLEXBUFFERS"
+                        }
+                    ],
+                } ],
+                "buffers" : [
+                    { },
+                    { },
+                    { "data": )" + filterData + R"(, }, )"
+                    + biasBuffer + R"(
+                ]
+            }
+        )";
+        SetupSingleInputSingleOutput("inputTensor", "outputTensor");
+    }
+};
+
+
+// No quantization meaning scale=1.0 and offset=0.0 and tensor quantization
+struct DepthwiseConvolution2dNoQuantFixture : DepthwiseConvolution2dFixture2
+{
+    DepthwiseConvolution2dNoQuantFixture()
+    : DepthwiseConvolution2dFixture2("[ 1, 3, 3, 3 ]",           // inputShape
+                                     "[ 1, 3, 3, 3 ]",           // outputShape
+                                     "[ 1, 3, 3, 3 ]",           // filterShape
+                                     "[ 9,8,7, 6,5,4, 3,2,1, "
+                                       "9,8,7, 6,5,4, 3,2,1, "
+                                       "9,8,7, 6,5,4, 3,2,1 ]",  // filterData
+                                     "1",                        // stride w and h
+                                     "SAME",                     // padding type
+                                     "",                         // bias shape
+                                     ""                          // bias data
+                                    )
+    {}
+};
+
+// No quantization meaning scale=1.0 and offset=0.0 and tensor quantization
+BOOST_FIXTURE_TEST_CASE(ParseDepthwiseConv2DNoQuant, DepthwiseConvolution2dNoQuantFixture)
+{
+    RunTest<4, armnn::DataType::QAsymmS8>(
+        0,
+        { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1},
+        { 18, 14, 10, 36, 30, 24, 30, 26, 22, 27, 21, 15, 54, 45,
+          36, 45, 39, 33, 18, 14, 10, 36, 30, 24, 30, 26, 22});
+}
+
+// Uses per channel quantization on weights but with scales = 1.0 and offsets = 0.0
+struct DepthwiseConvolution2dNoChannelQuantFixture : DepthwiseConvolution2dFixture2
+{
+    DepthwiseConvolution2dNoChannelQuantFixture()
+    : DepthwiseConvolution2dFixture2("[ 1, 3, 3, 3 ]",           // inputShape
+                                     "[ 1, 3, 3, 3 ]",           // outputShape
+                                     "[ 1, 3, 3, 3 ]",           // filterShape
+                                     "[ 9,8,7, 6,5,4, 3,2,1, 9,8,7, 6,5,4, 3,2,1, 9,8,7, 6,5,4, 3,2,1 ]",  // filterData
+                                     "1",                        // stride w and h
+                                     "SAME",                     // padding type
+                                     "",                         // bias shape
+                                     "",                         // bias data
+                                     "[ 0.0 ]",                  // filter quantization min values
+                                     "[ 255.0 ]",                // filter quantization max values
+                                     "[ 1.0, 1.0, 1.0]",         // filter quantization scales
+                                     "[ 0, 0, 0]",               // filter quantization zero-points
+                                     "3"                         // filter quantized axis
+                                                                 // (in case of per channel quantization)
+                                    )
+    {}
+};
+
+// Uses per channel quantization on weights but with scales = 1.0 and offsets = 0.0
+BOOST_FIXTURE_TEST_CASE(ParseDepthwiseConv2DFilterNoChannelQuant, DepthwiseConvolution2dNoChannelQuantFixture)
+{
+    RunTest<4, armnn::DataType::QAsymmS8>(
+        0,
+        { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1},
+        { 18, 14, 10, 36, 30, 24, 30, 26, 22, 27, 21, 15, 54, 45,
+          36, 45, 39, 33, 18, 14, 10, 36, 30, 24, 30, 26, 22});
+}
+
+// Uses per channel quantization on weights but all scales are set to the same value
+struct DepthwiseConvolution2dWeightsPerChannelQuantFixture : DepthwiseConvolution2dFixture2
+{
+    DepthwiseConvolution2dWeightsPerChannelQuantFixture()
+    : DepthwiseConvolution2dFixture2("[ 1, 3, 3, 3 ]",           // inputShape
+                                     "[ 1, 3, 3, 3 ]",           // outputShape
+                                     "[ 1, 3, 3, 3 ]",           // filterShape
+                                     // filterData is [ 9,8,7, 6,5,4, 3,2,1, 9,8,7, 6,5,4, 3,2,1, 9,8,7, 6,5,4, 3,2,1 ]
+                                     // quantized per channel with q_dim=3
+                                     "[36, 32, 28, 24, 20, 16, 12,  8,  4, 36, 32, 28, 24, "
+                                      "20, 16, 12,  8,  4, 36, 32, 28, 24, 20, 16, 12,  8, 4]",
+                                     "1",                        // stride w and h
+                                     "SAME",                     // padding type
+                                     "",                         // bias shape
+                                     "",                         // bias data
+                                     "[ 0.0 ]",                  // filter quantization min values
+                                     "[ 255.0 ]",                // filter quantization max values
+                                     "[ 0.25, 0.25, 0.25]",      // filter quantization scales
+                                     "[ 0, 0, 0]",               // filter quantization zero-points
+                                     "3"                         // filter quantized axis
+                                                                 // (in case of per channel quantization)
+                                    )
+    {}
+};
+
+// Weights are per channel quantized but all scales are set to the same value
+BOOST_FIXTURE_TEST_CASE(ParseDepthwiseConv2DFilterWeightsPerChannelQuant,
+                        DepthwiseConvolution2dWeightsPerChannelQuantFixture)
+{
+    RunTest<4, armnn::DataType::QAsymmS8>(
+        0,
+        { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1},
+        { 18, 14, 10, 36, 30, 24, 30, 26, 22, 27, 21, 15, 54, 45,
+          36, 45, 39, 33, 18, 14, 10, 36, 30, 24, 30, 26, 22});
+}
+
+// Uses per channel quantization on weights all scales are different in this test
+struct DepthwiseConvolution2dWeightsPerChannelQuant1Fixture : DepthwiseConvolution2dFixture2
+{
+    DepthwiseConvolution2dWeightsPerChannelQuant1Fixture()
+    : DepthwiseConvolution2dFixture2("[ 1, 3, 3, 3 ]",           // inputShape
+                                     "[ 1, 3, 3, 3 ]",           // outputShape
+                                     "[ 1, 3, 3, 3 ]",           // filterShape
+                                     // filterData is [ 9,8,7, 6,5,4, 3,2,1, 9,8,7, 6,5,4, 3,2,1, 9,8,7, 6,5,4, 3,2,1 ]
+                                     // quantized per channel with q_dim=3
+                                     "[36, 40, 70, 24, 25, 40, 12, 10, 10, 36, 40, 70, 24, "
+                                      "25, 40, 12, 10, 10, 36, 40, 70, 24, 25, 40, 12, 10, 10]",
+                                     "1",                        // stride w and h
+                                     "SAME",                     // padding type
+                                     "",                         // bias shape
+                                     "",                         // bias data
+                                     "[ 0.0 ]",                  // filter quantization min values
+                                     "[ 255.0 ]",                // filter quantization max values
+                                     "[ 0.25, 0.2, 0.1]",        // filter quantization scales
+                                     "[ 0, 0, 0]",               // filter quantization zero-points
+                                     "3"                         // filter quantized axis
+                                                                 // (in case of per channel quantization)
+                                    )
+    {}
+};
+
+// Uses per channel quantization on weights all scales are different in this test
+BOOST_FIXTURE_TEST_CASE(ParseDepthwiseConv2DFilterWeightsPerChannelQuant1,
+                        DepthwiseConvolution2dWeightsPerChannelQuant1Fixture)
+{
+    RunTest<4, armnn::DataType::QAsymmS8>(
+        0,
+        { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1},
+        { 18, 14, 10, 36, 30, 24, 30, 26, 22, 27, 21, 15, 54, 45,
+          36, 45, 39, 33, 18, 14, 10, 36, 30, 24, 30, 26, 22});
+}
+
+
+// Uses per channel quantization on weights all scales are different in this test
+// Uses different shape for weights and input compared to the other tests above
+struct DepthwiseConvolution2dWeightsPerChannelQuant2Fixture : DepthwiseConvolution2dFixture2
+{
+    DepthwiseConvolution2dWeightsPerChannelQuant2Fixture()
+    : DepthwiseConvolution2dFixture2("[ 1, 4, 4, 4 ]",           // inputShape
+                                     "[ 1, 4, 4, 4 ]",           // outputShape
+                                     "[ 1, 2, 2, 4 ]",           // filterShape
+                                     // filterData is [ 9,8,7,6, 5,4,3,2, 1,9,8,7, 6,5,4,3 ]
+                                     // quantized per channel with q_dim=3
+                                     "[36, 40, 70, 20, 20, 20, 30, 6, 4, 45, 80, 23, 24, 25, 40, 10]",
+                                     "1",                        // stride w and h
+                                     "SAME",                     // padding type
+                                     "",                         // bias shape
+                                     "",                         // bias data
+                                     "[ 0.0 ]",                  // filter quantization min values
+                                     "[ 255.0 ]",                // filter quantization max values
+                                     "[ 0.25, 0.2, 0.1, 0.3]",   // filter quantization scales
+                                     "[ 0, 0, 0, 0]",            // filter quantization zero-points
+                                     "3"                         // filter quantized axis
+                                                                 // (in case of per channel quantization)
+                                    )
+    {}
+};
+
+// Uses per channel quantization on weights all scales are different in this test
+// Uses different shape for weights and input compared to the other tests above
+BOOST_FIXTURE_TEST_CASE(ParseDepthwiseConv2DFilterWeightsPerChannelQuant2,
+                        DepthwiseConvolution2dWeightsPerChannelQuant2Fixture)
+{
+    RunTest<4, armnn::DataType::QAsymmS8>(
+        0,
+        { 1,1,1,1, 1,1,1,1, 1,1,1,1, 1,1,1,1,
+          1,1,1,1, 1,1,1,1, 1,1,1,1, 1,1,1,1,
+          1,1,1,1, 1,1,1,1, 1,1,1,1, 1,1,1,1,
+          1,1,1,1, 1,1,1,1, 1,1,1,1, 1,1,1,1},
+        { 21, 26, 22, 18, 21, 26, 22, 18, 21, 26, 22, 18, 10, 17, 15, 13,
+          21, 26, 22, 18, 21, 26, 22, 18, 21, 26, 22, 18, 10, 17, 15, 13,
+          21, 26, 22, 18, 21, 26, 22, 18, 21, 26, 22, 18, 10, 17, 15, 13,
+          14, 12, 10,  8, 14, 12, 10,  8, 14, 12, 10,  8,  9,  8,  7,  6});
+}
+
+// Test for depthwise_multiplier different to one (M > 1)
+struct DepthwiseConvolution2dWeightsPerChannelQuant4Fixture : DepthwiseConvolution2dFixture2
+{
+    DepthwiseConvolution2dWeightsPerChannelQuant4Fixture()
+    : DepthwiseConvolution2dFixture2("[ 1, 4, 4, 4 ]",            // inputShape
+                                     "[ 1, 4, 4, 16 ]",           // outputShape
+                                     "[ 1, 2, 2, 16 ]",           // filterShape
+                                     // filter data is [ 9,8,7,6, 5,4,3,2, 1,9,8,7, 6,5,4,3,
+                                     //                  9,8,7,6, 5,4,3,2, 1,9,8,7, 6,5,4,3,
+                                     //                  9,8,7,6, 5,4,3,2, 1,9,8,7, 6,5,4,3,
+                                     //                  9,8,7,6, 5,4,3,2, 1,9,8,7, 6,5,4,3 ]
+                                     //                  quantized per channel with q_dim=3
+                                     "[36, 40, 70, 20, 20, 20, 30, 6, 4, 45, 80, 23, 24, 25, 40, 10, "
+                                      "36, 40, 70, 20, 20, 20, 30, 6, 4, 45, 80, 23, 24, 25, 40, 10, "
+                                      "36, 40, 70, 20, 20, 20, 30, 6, 4, 45, 80, 23, 24, 25, 40, 10, "
+                                      "36, 40, 70, 20, 20, 20, 30, 6, 4, 45, 80, 23, 24, 25, 40, 10]",
+                                     "1",                        // stride w and h
+                                     "SAME",                     // padding type
+                                     "",                         // bias shape
+                                     "",                         // bias data
+                                     "[ 0.0 ]",                  // filter quantization min values
+                                     "[ 255.0 ]",                // filter quantization max values
+                                     "[ 0.25, 0.2, 0.1, 0.3,"
+                                       "0.25, 0.2, 0.1, 0.3,"
+                                       "0.25, 0.2, 0.1, 0.3,"
+                                       "0.25, 0.2, 0.1, 0.3]",   // filter quantization scales
+                                     "[ 0, 0, 0, 0]",            // filter quantization zero-points
+                                     "3"                         // filter quantized axis
+                                                                 // (in case of per channel quantization)
+                                    )
+    {}
+};
+
+// Test for depthwise_multiplier different to one (M > 1)
+BOOST_FIXTURE_TEST_CASE(ParseDepthwiseConv2DFilterWeightsPerChannelQuant4,
+                        DepthwiseConvolution2dWeightsPerChannelQuant4Fixture)
+{
+    RunTest<4, armnn::DataType::QAsymmS8>(
+        0,
+        { 1,1,1,1, 1,1,1,1, 1,1,1,1, 1,1,1,1,
+          1,1,1,1, 1,1,1,1, 1,1,1,1, 1,1,1,1,
+          1,1,1,1, 1,1,1,1, 1,1,1,1, 1,1,1,1,
+          1,1,1,1, 1,1,1,1, 1,1,1,1, 1,1,1,1},
+        { 36, 32, 28, 24, 20, 16, 12,  8, 4, 36, 32, 28, 24, 20, 16, 12,
+          36, 32, 28, 24, 20, 16, 12,  8, 4, 36, 32, 28, 24, 20, 16, 12,
+          36, 32, 28, 24, 20, 16, 12,  8, 4, 36, 32, 28, 24, 20, 16, 12,
+          18, 16, 14, 12, 10,  8,  6,  4, 2, 18, 16, 14, 12, 10,  8,  6,
+          36, 32, 28, 24, 20, 16, 12,  8, 4, 36, 32, 28, 24, 20, 16, 12,
+          36, 32, 28, 24, 20, 16, 12,  8, 4, 36, 32, 28, 24, 20, 16, 12,
+          36, 32, 28, 24, 20, 16, 12,  8, 4, 36, 32, 28, 24, 20, 16, 12,
+          18, 16, 14, 12, 10,  8,  6,  4, 2, 18, 16, 14, 12, 10,  8,  6,
+          36, 32, 28, 24, 20, 16, 12,  8, 4, 36, 32, 28, 24, 20, 16, 12,
+          36, 32, 28, 24, 20, 16, 12,  8, 4, 36, 32, 28, 24, 20, 16, 12,
+          36, 32, 28, 24, 20, 16, 12,  8, 4, 36, 32, 28, 24, 20, 16, 12,
+          18, 16, 14, 12, 10,  8,  6,  4, 2, 18, 16, 14, 12, 10,  8,  6,
+          18, 16, 14, 12, 10,  8,  6,  4, 2, 18, 16, 14, 12, 10,  8,  6,
+          18, 16, 14, 12, 10,  8,  6,  4, 2, 18, 16, 14, 12, 10,  8,  6,
+          18, 16, 14, 12, 10,  8,  6,  4, 2, 18, 16, 14, 12, 10,  8,  6,
+           9,  8,  7,  6,  5,  4,  3,  2, 1,  9,  8,  7,  6,  5,  4,  3});
+}
+
 BOOST_AUTO_TEST_SUITE_END()