COMPMID-1276 - Allow GEMM to work with 3D input tensor

Skipped im2col in CLGEMMConvolutionLayer for 1x1 convolutions with NHWC data layout

Change-Id: I894e6b952ed8605e8f3ffc0ffc25c24730d4664c
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/141909
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

4 files changed, 101 insertions, 35 deletions

diff --git a/tests/datasets/LargeGEMMDataset.h b/tests/datasets/LargeGEMMDataset.h
index bc9a056e7c..9c45cb0657 100644
--- a/tests/datasets/LargeGEMMDataset.h
+++ b/tests/datasets/LargeGEMMDataset.h
@@ -48,18 +48,34 @@ public:
         add_config(TensorShape(941U, 1U), TensorShape(623U, 941U), TensorShape(623U, 1U), TensorShape(623U, 1U), 0.4f, 0.7f);
     }
 };
-class LargeGEMM3DDataset final : public GEMMDataset
+
+class LargeGEMMOutput3DDataset final : public GEMMDataset
 {
 public:
-    LargeGEMM3DDataset()
+    LargeGEMMOutput3DDataset()
     {
         add_config(TensorShape(923U, 429U), TensorShape(871U, 923U), TensorShape(871U, 143U, 3U), TensorShape(871U, 143U, 3U), 1.0f, 0.0f);
         add_config(TensorShape(681U, 1025U), TensorShape(213U, 681U), TensorShape(213U, 205U, 5U), TensorShape(213U, 205U, 5U), 1.0f, 0.0f);
-        add_config(TensorShape(364, 3025), TensorShape(96, 364), TensorShape(96, 605, 5), TensorShape(96, 605, 5), 1.0f, 0.0f);
-        add_config(TensorShape(1201, 729), TensorShape(128, 1201), TensorShape(128, 243, 3), TensorShape(128, 243, 3), 1.0f, 0.0f);
-        add_config(TensorShape(2305, 169), TensorShape(384, 2305), TensorShape(384, 13, 13), TensorShape(384, 13, 13), 1.0f, 0.0f);
-        add_config(TensorShape(1729, 170), TensorShape(192, 1729), TensorShape(192, 85, 2), TensorShape(192, 85, 2), 1.0f, 0.0f);
-        add_config(TensorShape(1729, 170), TensorShape(128, 1729), TensorShape(128, 17, 10), TensorShape(128, 17, 10), 1.0f, 0.0f);
+        add_config(TensorShape(364U, 3025U), TensorShape(96U, 364U), TensorShape(96U, 605U, 5U), TensorShape(96U, 605U, 5U), 1.0f, 0.0f);
+        add_config(TensorShape(1201U, 729U), TensorShape(128U, 1201U), TensorShape(128U, 243U, 3U), TensorShape(128U, 243U, 3U), 1.0f, 0.0f);
+        add_config(TensorShape(2305U, 169U), TensorShape(384U, 2305U), TensorShape(384U, 13U, 13U), TensorShape(384U, 13U, 13U), 1.0f, 0.0f);
+        add_config(TensorShape(1729U, 170U), TensorShape(192U, 1729U), TensorShape(192U, 85U, 2U), TensorShape(192U, 85U, 2U), 1.0f, 0.0f);
+        add_config(TensorShape(1729U, 170U), TensorShape(128U, 1729U), TensorShape(128U, 17U, 10U), TensorShape(128U, 17U, 10U), 1.0f, 0.0f);
+    }
+};
+
+class LargeGEMMInputOutput3DDataset final : public GEMMDataset
+{
+public:
+    LargeGEMMInputOutput3DDataset()
+    {
+        add_config(TensorShape(923U, 143U, 3U), TensorShape(871U, 923U), TensorShape(871U, 143U, 3U), TensorShape(871U, 143U, 3U), 1.0f, 0.0f);
+        add_config(TensorShape(681U, 205U, 5U), TensorShape(213U, 681U), TensorShape(213U, 205U, 5U), TensorShape(213U, 205U, 5U), 1.0f, 0.0f);
+        add_config(TensorShape(364U, 605U, 5U), TensorShape(96U, 364U), TensorShape(96U, 605U, 5U), TensorShape(96U, 605U, 5U), 1.0f, 0.0f);
+        add_config(TensorShape(1201U, 243U, 3U), TensorShape(128U, 1201U), TensorShape(128U, 243U, 3U), TensorShape(128U, 243U, 3U), 1.0f, 0.0f);
+        add_config(TensorShape(2305U, 13U, 13U), TensorShape(384U, 2305U), TensorShape(384U, 13U, 13U), TensorShape(384U, 13U, 13U), 1.0f, 0.0f);
+        add_config(TensorShape(1729U, 85U, 2U, 2U), TensorShape(192U, 1729U, 2U), TensorShape(192U, 85U, 2U, 2U), TensorShape(192U, 85U, 2U, 2U), 1.0f, 0.0f);
+        add_config(TensorShape(1729U, 17U, 10U, 3U), TensorShape(128U, 1729U, 3U), TensorShape(128U, 17U, 10U, 3U), TensorShape(128U, 17U, 10U, 3U), 1.0f, 0.0f);
     }
 };
 } // namespace datasets
diff --git a/tests/datasets/SmallGEMMDataset.h b/tests/datasets/SmallGEMMDataset.h
index c9bf674ad3..e108fcc1ca 100644
--- a/tests/datasets/SmallGEMMDataset.h
+++ b/tests/datasets/SmallGEMMDataset.h
@@ -49,10 +49,10 @@ public:
         add_config(TensorShape(32U, 1U), TensorShape(17U, 32U), TensorShape(17U, 1U), TensorShape(17U, 1U), 0.4f, 0.7f);
     }
 };
-class SmallGEMM3DDataset final : public GEMMDataset
+class SmallGEMMOutput3DDataset final : public GEMMDataset
 {
 public:
-    SmallGEMM3DDataset()
+    SmallGEMMOutput3DDataset()
     {
         add_config(TensorShape(21U, 14U), TensorShape(34U, 21U), TensorShape(34U, 7U, 2U), TensorShape(34U, 7U, 2U), 1.0f, 0.0f);
         add_config(TensorShape(31U, 1U), TensorShape(23U, 31U), TensorShape(23U, 1U, 1U), TensorShape(23U, 1U, 1U), 1.0f, 0.0f);
@@ -62,6 +62,20 @@ public:
         add_config(TensorShape(16U, 16U, 5U), TensorShape(8U, 16U, 5U), TensorShape(8U, 8U, 2U, 5U), TensorShape(8U, 8U, 2U, 5U), 1.0f, 0.0f);
     }
 };
+
+class SmallGEMMInputOutput3DDataset final : public GEMMDataset
+{
+public:
+    SmallGEMMInputOutput3DDataset()
+    {
+        add_config(TensorShape(21U, 14U, 13U), TensorShape(34U, 21U), TensorShape(34U, 14U, 13U), TensorShape(34U, 14U, 13U), 1.0f, 0.0f);
+        add_config(TensorShape(31U, 1U, 3U), TensorShape(23U, 31U), TensorShape(23U, 1U, 3U), TensorShape(23U, 1U, 3U), 1.0f, 0.0f);
+        add_config(TensorShape(38U, 12U, 2U), TensorShape(21U, 38U), TensorShape(21U, 12U, 2U), TensorShape(21U, 12U, 2U), 0.2f, 1.2f);
+        add_config(TensorShape(32U, 1U, 4U, 3U), TensorShape(17U, 32U), TensorShape(17U, 1U, 4U, 3U), TensorShape(17U, 1U, 4U, 3U), 0.4f, 0.7f);
+        add_config(TensorShape(16U, 16U, 3U, 2U), TensorShape(8U, 16U), TensorShape(8U, 16U, 3U, 2U), TensorShape(8U, 16U, 3U, 2U), 1.0f, 0.0f);
+        add_config(TensorShape(16U, 16U, 5U, 3U), TensorShape(8U, 16U, 3U), TensorShape(8U, 16U, 5U, 3U), TensorShape(8U, 16U, 5U, 3U), 1.0f, 0.0f);
+    }
+};
 } // namespace datasets
 } // namespace test
 } // namespace arm_compute
diff --git a/tests/validation/CL/GEMM.cpp b/tests/validation/CL/GEMM.cpp
index 639182030e..ff2071a756 100644
--- a/tests/validation/CL/GEMM.cpp
+++ b/tests/validation/CL/GEMM.cpp
@@ -107,6 +107,12 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(framework::da
 template <typename T>
 using CLGEMMFixture = GEMMValidationFixture<CLTensor, CLAccessor, CLGEMM, T>;
 
+template <typename T>
+using CLGEMMOutput3DFixture = GEMMValidationFixture<CLTensor, CLAccessor, CLGEMM, T, false, true>;
+
+template <typename T>
+using CLGEMMInputOutput3DFixture = GEMMValidationFixture<CLTensor, CLAccessor, CLGEMM, T, true, true>;
+
 TEST_SUITE(TRANSPOSE_1XW)
 using CLGEMMTranspose1xW        = CLSynthetizeFunctionWithZeroConstantBorder<CLGEMMTranspose1xWKernel, 4>;
 using CLGEMMTranspose1xWFixture = GEMMTranspose1xWValidationFixture<CLTensor, CLAccessor, CLGEMMTranspose1xW, float>;
@@ -149,17 +155,53 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMFixture<float>, framework::DatasetMode::N
 TEST_SUITE_END()
 TEST_SUITE_END()
 
+TEST_SUITE(INPUT_OUTPUT_3D)
+TEST_SUITE(Float)
+TEST_SUITE(FP32)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMInputOutput3DFixture<float>, framework::DatasetMode::PRECOMMIT, combine(datasets::SmallGEMMInputOutput3DDataset(),
+                                                                                                               framework::dataset::make("DataType", DataType::F32)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_f32);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMInputOutput3DFixture<float>, framework::DatasetMode::NIGHTLY, combine(datasets::LargeGEMMInputOutput3DDataset(),
+                                                                                                             framework::dataset::make("DataType", DataType::F32)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32);
+}
+TEST_SUITE_END() // FP32
+
+TEST_SUITE(FP16)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMInputOutput3DFixture<half>, framework::DatasetMode::PRECOMMIT, combine(datasets::SmallGEMMInputOutput3DDataset(),
+                                                                                                              framework::dataset::make("DataType", DataType::F16)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMInputOutput3DFixture<half>, framework::DatasetMode::NIGHTLY, combine(datasets::LargeGEMMInputOutput3DDataset(),
+                                                                                                            framework::dataset::make("DataType",
+                                                                                                                    DataType::F16)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num);
+}
+TEST_SUITE_END() // FP16
+
+TEST_SUITE_END() // Float
+TEST_SUITE_END() // INPUT_OUTPUT_3D
+
 TEST_SUITE(OUTPUT_3D)
 TEST_SUITE(Float)
 TEST_SUITE(FP32)
-FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMFixture<float>, framework::DatasetMode::PRECOMMIT, combine(datasets::SmallGEMM3DDataset(),
-                                                                                                  framework::dataset::make("DataType", DataType::F32)))
+FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMOutput3DFixture<float>, framework::DatasetMode::PRECOMMIT, combine(datasets::SmallGEMMOutput3DDataset(),
+                                                                                                          framework::dataset::make("DataType", DataType::F32)))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, tolerance_f32);
 }
-FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMFixture<float>, framework::DatasetMode::NIGHTLY, combine(datasets::LargeGEMM3DDataset(),
-                                                                                                framework::dataset::make("DataType", DataType::F32)))
+FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMOutput3DFixture<float>, framework::DatasetMode::NIGHTLY, combine(datasets::LargeGEMMOutput3DDataset(),
+                                                                                                        framework::dataset::make("DataType", DataType::F32)))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32);
@@ -167,15 +209,15 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMFixture<float>, framework::DatasetMode::N
 TEST_SUITE_END() // FP32
 
 TEST_SUITE(FP16)
-FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMFixture<half>, framework::DatasetMode::PRECOMMIT, combine(datasets::SmallGEMM3DDataset(),
-                                                                                                 framework::dataset::make("DataType", DataType::F16)))
+FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMOutput3DFixture<half>, framework::DatasetMode::PRECOMMIT, combine(datasets::SmallGEMMOutput3DDataset(),
+                                                                                                         framework::dataset::make("DataType", DataType::F16)))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num);
 }
-FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMFixture<half>, framework::DatasetMode::NIGHTLY, combine(datasets::LargeGEMM3DDataset(),
-                                                                                               framework::dataset::make("DataType",
-                                                                                                                        DataType::F16)))
+FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMOutput3DFixture<half>, framework::DatasetMode::NIGHTLY, combine(datasets::LargeGEMMOutput3DDataset(),
+                                                                                                       framework::dataset::make("DataType",
+                                                                                                               DataType::F16)))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num);
diff --git a/tests/validation/fixtures/GEMMFixture.h b/tests/validation/fixtures/GEMMFixture.h
index e4762cc5be..255b12c0ed 100644
--- a/tests/validation/fixtures/GEMMFixture.h
+++ b/tests/validation/fixtures/GEMMFixture.h
@@ -42,8 +42,8 @@ namespace test
 {
 namespace validation
 {
-template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
-class GEMMValidationFixedPointFixture : public framework::Fixture
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool reinterpret_input_as_3d = false, bool reinterpret_ouput_as_3d = false>
+class GEMMValidationFixture : public framework::Fixture
 {
 public:
     template <typename...>
@@ -87,10 +87,7 @@ protected:
         // The GEMMinfo includes the values of the depth in case of reinterpreted 3d output.
         // If the output shape has the same number of dimensions of the input the method called is a 2D matrix multiplication (depth_output_reinterpreted_as_3D = 1),
         // in the other case we have to use the reinterpreted version of GEMM (depth_output_reinterpreted_as_3D = depth of the 3D output).
-        bool is_output_reinterpreted_as_3D = output_shape.num_dimensions() > shape_a.num_dimensions();
-        gemm.configure(&a, &b, &c, &dst, alpha, beta,
-                       GEMMInfo(false, false, false, is_output_reinterpreted_as_3D ? output_shape[2] : 1));
-
+        gemm.configure(&a, &b, &c, &dst, alpha, beta, GEMMInfo(false, false, false, (reinterpret_ouput_as_3d ? output_shape[2] : 1), reinterpret_input_as_3d));
         ARM_COMPUTE_EXPECT(a.info()->is_resizable(), framework::LogLevel::ERRORS);
         ARM_COMPUTE_EXPECT(b.info()->is_resizable(), framework::LogLevel::ERRORS);
         ARM_COMPUTE_EXPECT(c.info()->is_resizable(), framework::LogLevel::ERRORS);
@@ -121,8 +118,15 @@ protected:
     SimpleTensor<T> compute_reference(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_c, const TensorShape &output_shape, float alpha, float beta,
                                       DataType data_type)
     {
+        TensorShape shape_a_to_use = shape_a;
+        if(reinterpret_input_as_3d)
+        {
+            // Collapse the second and third dimension if the input is 3D
+            shape_a_to_use.collapse(2U, 1U);
+        }
+
         // Create reference
-        SimpleTensor<T> a{ shape_a, data_type, 1 };
+        SimpleTensor<T> a{ shape_a_to_use, data_type, 1 };
         SimpleTensor<T> b{ shape_b, data_type, 1 };
         SimpleTensor<T> c{ shape_c, data_type, 1 };
 
@@ -139,16 +143,6 @@ protected:
     DataType        _data_type{};
 };
 
-template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
-class GEMMValidationFixture : public GEMMValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T>
-{
-public:
-    template <typename...>
-    void setup(TensorShape shape_a, TensorShape shape_b, TensorShape shape_c, TensorShape output_shape, float alpha, float beta, DataType data_type)
-    {
-        GEMMValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T>::setup(shape_a, shape_b, shape_c, output_shape, alpha, beta, data_type);
-    }
-};
 } // namespace validation
 } // namespace test
 } // namespace arm_compute