COMPMID-722 - Support for vector-matrix in GEMMLowp (NEON)

This patch includes COMPMID-716 as well

- Added vector-matrix case in NEGEMMLowpMatrixMultiplyKernel
- Added benchmarks for NEON and OpenCL

Change-Id: I715cd25e8668a4d6c8127e9a298a865e7713267f
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/111468
Tested-by: BSG Visual Compute Jenkins server to access repositories on http://mpd-gerrit.cambridge.arm.com <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

5 files changed, 78 insertions, 72 deletions

diff --git a/tests/benchmark/CL/GEMMLowp.cpp b/tests/benchmark/CL/GEMMLowp.cpp
new file mode 100644
index 0000000000..039695b18f
--- /dev/null
+++ b/tests/benchmark/CL/GEMMLowp.cpp
@@ -0,0 +1,52 @@
+/*
+ * Copyright (c) 2017 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/TensorShape.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/CL/CLTensor.h"
+#include "arm_compute/runtime/CL/CLTensorAllocator.h"
+#include "arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h"
+#include "tests/CL/CLAccessor.h"
+#include "tests/benchmark/fixtures/GEMMLowpFixture.h"
+#include "tests/datasets/GoogleNetGEMMDataset.h"
+#include "tests/datasets/MatrixMultiplyGEMMDataset.h"
+#include "tests/datasets/system_tests/googlenet/inceptionv1/GoogLeNetInceptionV1GEMMDataset.h"
+#include "tests/framework/Macros.h"
+#include "tests/framework/datasets/Datasets.h"
+#include "utils/TypePrinter.h"
+
+namespace arm_compute
+{
+namespace test
+{
+using CLGEMMLowpFixture = GEMMLowpMatrixMultiplyCoreFixture<CLTensor, CLGEMMLowpMatrixMultiplyCore, CLAccessor>;
+
+TEST_SUITE(CL)
+
+REGISTER_FIXTURE_DATA_TEST_CASE(GoogLeNetInceptionV1GEMMLowp, CLGEMMLowpFixture, framework::DatasetMode::ALL, datasets::GoogLeNetInceptionV1GEMMDataset());
+REGISTER_FIXTURE_DATA_TEST_CASE(MatrixMultiplyGEMMLowp, CLGEMMLowpFixture, framework::DatasetMode::ALL, datasets::MatrixMultiplyGEMMDataset());
+REGISTER_FIXTURE_DATA_TEST_CASE(GoogleNetGEMMLowp, CLGEMMLowpFixture, framework::DatasetMode::NIGHTLY, datasets::GoogleNetGEMMDataset());
+
+TEST_SUITE_END()
+} // namespace test
+} // namespace arm_compute
diff --git a/tests/benchmark/NEON/GEMMLowp.cpp b/tests/benchmark/NEON/GEMMLowp.cpp
index a0e5e694bd..e3ad1d43d2 100644
--- a/tests/benchmark/NEON/GEMMLowp.cpp
+++ b/tests/benchmark/NEON/GEMMLowp.cpp
@@ -23,42 +23,30 @@
  */
 #include "arm_compute/core/TensorShape.h"
 #include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h"
 #include "arm_compute/runtime/Tensor.h"
 #include "arm_compute/runtime/TensorAllocator.h"
 #include "tests/NEON/Accessor.h"
 #include "tests/benchmark/fixtures/GEMMLowpFixture.h"
+#include "tests/datasets/GoogleNetGEMMDataset.h"
+#include "tests/datasets/MatrixMultiplyGEMMDataset.h"
+#include "tests/datasets/system_tests/googlenet/inceptionv1/GoogLeNetInceptionV1GEMMDataset.h"
 #include "tests/framework/Macros.h"
 #include "tests/framework/datasets/Datasets.h"
 #include "utils/TypePrinter.h"
 
-#include "arm_compute/core/NEON/kernels/NEGEMMInterleaveBlockedKernel.h"
-#include "tests/NEON/Helper.h"
-
 namespace arm_compute
 {
 namespace test
 {
-const auto data_int_blk = framework::dataset::make("M", 800) * framework::dataset::make("N", 800) * framework::dataset::make("by", 8, 13) * framework::dataset::make("block", 4, 9);
+using NEGEMMLowpFixture = GEMMLowpMatrixMultiplyCoreFixture<Tensor, NEGEMMLowpMatrixMultiplyCore, Accessor>;
 
 TEST_SUITE(NEON)
 
-TEST_SUITE(INTERLEAVE_BLOCKED)
-using NEInterleaveBlocked            = NESynthetizeFunction<NEGEMMInterleaveBlockedKernel>;
-using NEGEMMInterleaveBlockedFixture = GEMMInterleaveBlockedFixture<Tensor, NEInterleaveBlocked, Accessor>;
-REGISTER_FIXTURE_DATA_TEST_CASE(InterleaveBlocked, NEGEMMInterleaveBlockedFixture, framework::DatasetMode::ALL, data_int_blk);
-TEST_SUITE_END()
-
-#if 0  //FIXME: enable when we update NEGEMMLowp interface to work without offsets
-TEST_SUITE(U32)
-using NEGEMMLowpFixture = GEMMLowpFixture<Tensor, NEGEMMLowp, Accessor>;
-REGISTER_FIXTURE_DATA_TEST_CASE(GEMMLowp, NEGEMMLowpFixture, framework::DatasetMode::ALL, framework::dataset::make("M", 100, 120) * framework::dataset::make("N", 100,
-                                110)
-                                * framework::dataset::make("K", 16, 20));
+REGISTER_FIXTURE_DATA_TEST_CASE(GoogLeNetInceptionV1GEMMLowp, NEGEMMLowpFixture, framework::DatasetMode::ALL, datasets::GoogLeNetInceptionV1GEMMDataset());
+REGISTER_FIXTURE_DATA_TEST_CASE(MatrixMultiplyGEMMLowp, NEGEMMLowpFixture, framework::DatasetMode::ALL, datasets::MatrixMultiplyGEMMDataset());
+REGISTER_FIXTURE_DATA_TEST_CASE(GoogleNetGEMMLowp, NEGEMMLowpFixture, framework::DatasetMode::NIGHTLY, datasets::GoogleNetGEMMDataset());
 
 TEST_SUITE_END()
-#endif // defined(__aarch64__)
-
-TEST_SUITE_END()
-
 } // namespace test
 } // namespace arm_compute
diff --git a/tests/benchmark/fixtures/GEMMLowpFixture.h b/tests/benchmark/fixtures/GEMMLowpFixture.h
index b640705990..4bd7dfd42f 100644
--- a/tests/benchmark/fixtures/GEMMLowpFixture.h
+++ b/tests/benchmark/fixtures/GEMMLowpFixture.h
@@ -21,8 +21,8 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#ifndef ARM_COMPUTE_TEST_GEMMFIXTURE
-#define ARM_COMPUTE_TEST_GEMMFIXTURE
+#ifndef ARM_COMPUTE_TEST_GEMMLOWPFIXTURE
+#define ARM_COMPUTE_TEST_GEMMLOWPFIXTURE
 
 #include "arm_compute/core/TensorShape.h"
 #include "arm_compute/core/Types.h"
@@ -34,59 +34,26 @@ namespace arm_compute
 {
 namespace test
 {
-template <typename TensorType, typename Function, typename Accessor, bool Transposed = false>
-class GEMMInterleaveBlockedFixture : public framework::Fixture
-{
-public:
-    template <typename...>
-    void setup(size_t x, size_t y, int int_by, int block)
-    {
-        const float       interleave_by_f32 = int_by;
-        const TensorShape shape_a(x, y);
-        const TensorShape shape_b(static_cast<size_t>(x * interleave_by_f32), static_cast<size_t>(std::ceil(y / interleave_by_f32)));
-        // Create tensors
-        a = create_tensor<TensorType>(shape_a, DataType::U8, 1);
-        b = create_tensor<TensorType>(shape_b, DataType::U8, 1);
-
-        // Create and configure function
-        f.configure(&a, &b, int_by, block, Transposed);
-
-        // Allocate tensors
-        a.allocator()->allocate();
-        b.allocator()->allocate();
-    }
-    void run()
-    {
-        f.run();
-    }
-
-    void teardown()
-    {
-        a.allocator()->free();
-        b.allocator()->free();
-    }
-
-private:
-    TensorType a{};
-    TensorType b{};
-    Function   f{};
-};
-
 /** Fixture that can be used for NEON and CL */
 template <typename TensorType, typename Function, typename Accessor>
-class GEMMLowpFixture : public framework::Fixture
+class GEMMLowpMatrixMultiplyCoreFixture : public framework::Fixture
 {
 public:
     template <typename...>
-    void setup(size_t m, size_t n, size_t k)
+    void setup(TensorShape shape_a, TensorShape shape_b, TensorShape shape_c, TensorShape shape_dst, float alpha, float beta)
     {
-        const TensorShape shape_a(k, m);
-        const TensorShape shape_b(n, k);
-        const TensorShape shape_c(n, m);
+        // TODO (COMPMID-717): The interface used for GEMMLowp is the same one used for GEMM in order to re-use the datasets
+        // However the interface for both GEMM and GEMMLowp should be reworked in order to accepts only the 3 dimensions M, N and K
+        ARM_COMPUTE_UNUSED(shape_c);
+        ARM_COMPUTE_UNUSED(alpha);
+        ARM_COMPUTE_UNUSED(beta);
+
+        // Note: The offsets for matrix A and matrix B are set to 0 in order to skip the computation for the offset contribution
+
         // Create tensors
-        a = create_tensor<TensorType>(shape_a, DataType::U8, 1);
-        b = create_tensor<TensorType>(shape_b, DataType::U8, 1);
-        c = create_tensor<TensorType>(shape_c, DataType::U32, 1);
+        a = create_tensor<TensorType>(shape_a, DataType::QASYMM8, 1, 0, QuantizationInfo(1.0f / 255.0f, 0));
+        b = create_tensor<TensorType>(shape_b, DataType::QASYMM8, 1, 0, QuantizationInfo(1.0f / 255.0f, 0));
+        c = create_tensor<TensorType>(shape_dst, DataType::S32, 1, 0, QuantizationInfo(1.0f / 255.0f, 0));
 
         // Create and configure function
         gemmlowp.configure(&a, &b, &c);
@@ -99,7 +66,6 @@ public:
         // Fill tensors
         library->fill_tensor_uniform(Accessor(a), 0);
         library->fill_tensor_uniform(Accessor(b), 1);
-        library->fill_tensor_uniform(Accessor(c), 2);
     }
     void run()
     {
@@ -122,4 +88,4 @@ private:
 
 } // namespace test
 } // namespace arm_compute
-#endif /* ARM_COMPUTE_TEST_GEMMFIXTURE */
+#endif /* ARM_COMPUTE_TEST_GEMMLOWPFIXTURE */
diff --git a/tests/datasets/LargeGEMMLowpDataset.h b/tests/datasets/LargeGEMMLowpDataset.h
index 87f879e70a..5c0230e262 100644
--- a/tests/datasets/LargeGEMMLowpDataset.h
+++ b/tests/datasets/LargeGEMMLowpDataset.h
@@ -42,7 +42,7 @@ class LargeGEMMLowpDataset final : public GEMMLowpDataset
 public:
     LargeGEMMLowpDataset()
     {
-        add_config(TensorShape(923U, 2U), TensorShape(871U, 923U), TensorShape(871U, 2U), 0, 0);
+        add_config(TensorShape(923U, 1U), TensorShape(871U, 923U), TensorShape(871U, 1U), 0, 0);
         add_config(TensorShape(923U, 429U), TensorShape(871U, 923U), TensorShape(871U, 429U), 0, 0);
         add_config(TensorShape(873U, 7U), TensorShape(784U, 873U), TensorShape(784U, 7U), -1, 3);
         add_config(TensorShape(873U, 513U), TensorShape(784U, 873U), TensorShape(784U, 513U), 0, 4);
diff --git a/tests/datasets/SmallGEMMLowpDataset.h b/tests/datasets/SmallGEMMLowpDataset.h
index 1d4ab53be5..b6651bdb42 100644
--- a/tests/datasets/SmallGEMMLowpDataset.h
+++ b/tests/datasets/SmallGEMMLowpDataset.h
@@ -42,7 +42,7 @@ class SmallGEMMLowpDataset final : public GEMMLowpDataset
 public:
     SmallGEMMLowpDataset()
     {
-        add_config(TensorShape(21U, 2U), TensorShape(43U, 21U), TensorShape(43U, 2U), 0, 0);
+        add_config(TensorShape(21U, 1U), TensorShape(43U, 21U), TensorShape(43U, 1U), 0, 0);
         add_config(TensorShape(21U, 13U), TensorShape(33U, 21U), TensorShape(33U, 13U), 0, 0);
         add_config(TensorShape(31U, 3U), TensorShape(72U, 31U), TensorShape(72U, 3U), -2, 13);
         add_config(TensorShape(52U, 13U), TensorShape(33U, 52U), TensorShape(33U, 13U), 0, 4);