Add test for CLGEMM to test a batched matrix multiplication with variable input tensors

Resolves : [COMPMID-5502]

Signed-off-by: Mohammed Suhail Munshi <MohammedSuhail.Munshi@arm.com>
Change-Id: Ida001dc597973f9180468737a3e32e5022e6baee
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/c/VisualCompute/ComputeLibrary/+/450342
Tested-by: bsgcomp <bsgcomp@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Reviewed-by: Mohammed Suhail Munshi <mohammedsuhail.munshi@arm.com>
Comments-Addressed: bsgcomp <bsgcomp@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8224
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>

3 files changed, 68 insertions, 5 deletions

diff --git a/tests/datasets/SmallGEMMDataset.h b/tests/datasets/SmallGEMMDataset.h
index 23b46a6547..811064bb67 100644
--- a/tests/datasets/SmallGEMMDataset.h
+++ b/tests/datasets/SmallGEMMDataset.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2021 Arm Limited.
+ * Copyright (c) 2017-2022 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -87,6 +87,22 @@ public:
         add_config(TensorShape(16U, 16U, 5U, 3U), TensorShape(8U, 16U), TensorShape(8U), TensorShape(8U, 16U, 5U, 3U), 1.0f, 0.3f);
     }
 };
+
+class SmallBatchedMatMulDataset final : public GEMMDataset
+{
+public:
+    SmallBatchedMatMulDataset()
+    {
+        add_config(TensorShape(4U, 3U), TensorShape(2U, 4U), TensorShape(2U), TensorShape(2U, 3U), 1.0f, 0.0f);
+        add_config(TensorShape(12U, 15U), TensorShape(7U, 12U), TensorShape(7U), TensorShape(7U, 15U), 1.0f, 0.0f);
+        add_config(TensorShape(59U, 17U), TensorShape(36U, 59U), TensorShape(36U), TensorShape(36U, 17U), 1.0f, 0.0f);
+        add_config(TensorShape(2U, 4U, 3U), TensorShape(5U, 2U, 3U), TensorShape(5U), TensorShape(5U, 4U, 3U), 1.0f, 0.0f);
+        add_config(TensorShape(15U, 7U, 36U), TensorShape(29U, 15U, 36U), TensorShape(29U), TensorShape(29U, 7U, 36U), 1.0f, 0.0f);
+        add_config(TensorShape(56U, 17U, 32U), TensorShape(5U, 56U, 32U), TensorShape(5U), TensorShape(5U, 17U, 32U), 1.0f, 0.0f);
+        add_config(TensorShape(13U, 256U, 32U), TensorShape(19U, 13U, 32U), TensorShape(19U), TensorShape(19U, 256U, 32U), 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 838920c29d..3d1be77bd9 100644
--- a/tests/validation/CL/GEMM.cpp
+++ b/tests/validation/CL/GEMM.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2020 Arm Limited.
+ * Copyright (c) 2017-2022 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -71,6 +71,9 @@ using CLGEMMOutput3DFixture = GEMMValidationFixture<CLTensor, CLAccessor, CLGEMM
 template <typename T>
 using CLGEMMInputOutput3DFixture = GEMMValidationFixture<CLTensor, CLAccessor, CLGEMM, T, false, true, true>;
 
+template <typename T>
+using CLBatchedMatMulFixture = GEMMValidationFixture<CLTensor, CLAccessor, CLGEMM, T, true, false, false, false, false, true>;
+
 TEST_SUITE(Float)
 TEST_SUITE(FP16)
 FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallGEMMDataset(),
@@ -181,10 +184,32 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMOutput3DFixture<half>, framework::Dataset
     validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num);
 }
 TEST_SUITE_END() // FP16
-
 TEST_SUITE_END() // Float
 TEST_SUITE_END() // OUTPUT_3D
 
+TEST_SUITE(BATCHED_MATMUL)
+
+TEST_SUITE(FP32)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLBatchedMatMulFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallBatchedMatMulDataset(),
+                                                                                                                    framework::dataset::make("ReshapeWeights", { false })),
+                                                                                                            framework::dataset::make("DataType", DataType::F32)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_f32, tolerance_num);
+}
+TEST_SUITE_END()
+
+TEST_SUITE(FP16)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLBatchedMatMulFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallBatchedMatMulDataset(),
+                                                                                                                   framework::dataset::make("ReshapeWeights", { false })),
+                                                                                                           framework::dataset::make("DataType", DataType::F16)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num);
+}
+TEST_SUITE_END()
+TEST_SUITE_END() // BATCHED_MATMUL
+
 TEST_SUITE_END() // GEMM
 TEST_SUITE_END() // CL
 } // namespace validation
diff --git a/tests/validation/fixtures/GEMMFixture.h b/tests/validation/fixtures/GEMMFixture.h
index 55bbbdaf80..0682337c82 100644
--- a/tests/validation/fixtures/GEMMFixture.h
+++ b/tests/validation/fixtures/GEMMFixture.h
@@ -48,7 +48,7 @@ namespace test
 {
 namespace validation
 {
-template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool disable_c = false, bool reinterpret_input_as_3d = false, bool reinterpret_output_as_3d = false, bool pretranspose_a = false, bool pretranspose_b = false>
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool disable_c = false, bool reinterpret_input_as_3d = false, bool reinterpret_output_as_3d = false, bool pretranspose_a = false, bool pretranspose_b = false, bool run_twice = false>
 class GEMMValidationFixture : public framework::Fixture
 {
 public:
@@ -130,6 +130,18 @@ protected:
             fill(AccessorType(c), 2);
         }
 
+        // Run with variable inputs.
+        if(run_twice)
+        {
+            gemm.run();
+            fill(AccessorType(a), 3); // Fill tensors with new seed after run
+            fill(AccessorType(b), 4);
+            if(!disable_c)
+            {
+                fill(AccessorType(c), 5);
+            }
+        }
+
         // Compute GEMM function
         gemm.run();
 
@@ -195,10 +207,20 @@ protected:
             transpose_matrix<T>(b, b_transposed);
         }
 
+        // Run with variable inputs.
+        if(run_twice)
+        {
+            reference::gemm<T>((pretranspose_a) ? a_transposed : a, (pretranspose_b) ? b_transposed : b, c, alpha, disable_c ? 0.f : beta);
+            fill((pretranspose_a) ? a_transposed : a, 3);
+            fill((pretranspose_b) ? b_transposed : b, 4);
+            fill(c , 5);
+        }
+
         // Setting beta to 0 will effectively disable C for the
         // computation of the reference: alpha * A * B + 0 * C
         // Use transposed tensors if boolean enabled else use original tensors
-        return reference::gemm<T>((pretranspose_a) ? a_transposed : a, (pretranspose_b) ? b_transposed : b, c, alpha, disable_c ? 0.f : beta);
+        auto r =  reference::gemm<T>((pretranspose_a) ? a_transposed : a, (pretranspose_b) ? b_transposed : b, c, alpha, disable_c ? 0.f : beta);
+        return r;
     }
 
     TensorType      _target{};