COMPMID-2094: Implement CLGEMMLowpNative

Change-Id: I2a80eec28baf9e83bfc67a930e2a140642e0b09e
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1285
Reviewed-by: Giuseppe Rossini <giuseppe.rossini@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

2 files changed, 341 insertions, 0 deletions

diff --git a/tests/validation/CL/GEMMLowpMatrixMultiplyNative.cpp b/tests/validation/CL/GEMMLowpMatrixMultiplyNative.cpp
new file mode 100644
index 0000000000..1fc8cc47c4
--- /dev/null
+++ b/tests/validation/CL/GEMMLowpMatrixMultiplyNative.cpp
@@ -0,0 +1,152 @@
+/*
+ * Copyright (c) 2019 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/CL/kernels/CLGEMMLowpMatrixMultiplyNativeKernel.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "tests/CL/CLAccessor.h"
+#include "tests/CL/Helper.h"
+#include "tests/framework/Asserts.h"
+#include "tests/framework/Macros.h"
+#include "tests/framework/datasets/Datasets.h"
+#include "tests/validation/Validation.h"
+#include "tests/validation/fixtures/GEMMLowpFixture.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+using namespace arm_compute::misc::shape_calculator;
+
+// Create function for CLGEMMMatrixMultiplyNativeKernel
+using CLGEMMLowpMatrixMultiplyNative = CLSynthetizeFunction<CLGEMMLowpMatrixMultiplyNativeKernel>;
+
+// Fixture for CLGEMMLowpMatrixMultiplyNative
+using CLGEMMLowpMatrixMultiplyNativeFixture = GEMMLowpMatrixMultiplyNativeValidationFixture<CLTensor, CLAccessor, CLGEMMLowpMatrixMultiplyNative>;
+
+// Fixture for CLGEMMMatrixMultiplyNative3D
+using CLGEMMLowpMatrixMultiplyNative3DFixture = GEMMLowpMatrixMultiplyNative3DValidationFixture<CLTensor, CLAccessor, CLGEMMLowpMatrixMultiplyNative>;
+
+namespace
+{
+// *INDENT-OFF*
+// clang-format off
+/** M values to test */
+const auto m_values = framework::dataset::make("M", 37);
+
+/** M_W values to test */
+const auto m_w_values = framework::dataset::make("M_W", 5);
+
+/** M_H values to test */
+const auto m_h_values = framework::dataset::make("M_H", 7);
+
+/** N values to test */
+const auto n_values = framework::dataset::make("N", 51);
+
+/** K values to test */
+const auto k_values = framework::dataset::make("K", 23);
+
+/** Batch size values to test */
+const auto b_values = framework::dataset::make("batch_size", 1, 3);
+
+/** M0 values to test - Precommit */
+const auto m0_values_precommit = framework::dataset::make("M0", {4, 6});
+
+/** N0 values to test - Precommit */
+const auto n0_values_precommit = framework::dataset::make("N0", { 2, 4 });
+
+/** K0 values to test - Precommit */
+const auto k0_values_precommit = framework::dataset::make("K0", { 4 });
+
+/** M0 values to test - Nightly */
+const auto m0_values_nightly = framework::dataset::make("M0", 2, 7);
+
+/** N0 values to test - Nightly */
+const auto n0_values_nightly = framework::dataset::make("N0", { 2, 3, 4, 8 });
+
+/** K0 values to test - Nightly */
+const auto k0_values_nightly = framework::dataset::make("K0", { 2, 3, 4, 8 });
+} // namespace
+
+TEST_SUITE(CL)
+TEST_SUITE(GEMMLowpMatrixMultiplyNative)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpMatrixMultiplyNativeFixture, framework::DatasetMode::ALL,
+                combine(combine(combine(combine(combine(combine(m_values,
+                                                                n_values),
+                                                                k_values),
+                                                                b_values),
+                                                                m0_values_precommit),
+                                                                n0_values_precommit),
+                                                                k0_values_precommit))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMLowpMatrixMultiplyNativeFixture, framework::DatasetMode::NIGHTLY,
+                combine(combine(combine(combine(combine(combine(m_values,
+                                                                n_values),
+                                                                k_values),
+                                                                b_values),
+                                                                m0_values_nightly),
+                                                                n0_values_nightly),
+                                                                k0_values_nightly))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+
+FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMLowpMatrixMultiplyNative3DFixture, framework::DatasetMode::ALL,
+                combine(combine(combine(combine(combine(combine(combine(m_w_values,
+                                                                        m_h_values),
+                                                                        n_values),
+                                                                        k_values),
+                                                                        b_values),
+                                                                        m0_values_precommit),
+                                                                        n0_values_precommit),
+                                                                        k0_values_precommit))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMLowpMatrixMultiplyNative3DFixture, framework::DatasetMode::NIGHTLY,
+                combine(combine(combine(combine(combine(combine(combine(m_w_values,
+                                                                        m_h_values),
+                                                                        n_values),
+                                                                        k_values),
+                                                                        b_values),
+                                                                        m0_values_nightly),
+                                                                        n0_values_nightly),
+                                                                        k0_values_nightly))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+TEST_SUITE_END() // GEMMLowpMatrixMultiplyNative
+TEST_SUITE_END() // CL
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
diff --git a/tests/validation/fixtures/GEMMLowpFixture.h b/tests/validation/fixtures/GEMMLowpFixture.h
index 5793ebdd2d..ad5acfc418 100644
--- a/tests/validation/fixtures/GEMMLowpFixture.h
+++ b/tests/validation/fixtures/GEMMLowpFixture.h
@@ -819,6 +819,195 @@ protected:
     TensorType            _target{};
     SimpleTensor<int32_t> _reference{};
 };
+
+template <typename TensorType, typename AccessorType, typename GEMMFunctionType>
+class GEMMLowpMatrixMultiplyNativeValidationFixture : public framework::Fixture
+{
+public:
+    template <typename...>
+    void setup(unsigned int m, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0)
+    {
+        GEMMLHSMatrixInfo lhs_info;
+        lhs_info.m0 = m0;
+        lhs_info.k0 = k0;
+
+        GEMMRHSMatrixInfo rhs_info;
+        rhs_info.n0 = n0;
+        rhs_info.k0 = k0;
+
+        // Set the tensor shapes for LHS and RHS matrices
+        const TensorShape lhs_shape(k, m, batch_size);
+        const TensorShape rhs_shape(n, k, batch_size);
+
+        _target    = compute_target(lhs_shape, rhs_shape, lhs_info, rhs_info);
+        _reference = compute_reference(lhs_shape, rhs_shape);
+    }
+
+protected:
+    template <typename U>
+    void fill(U &&tensor, int i)
+    {
+        // Between 1 and 254 in order to avoid having -128 and 128 for the DOT product path
+        std::uniform_int_distribution<> distribution(1, 254);
+        library->fill(tensor, distribution, i);
+    }
+
+    TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info)
+    {
+        // Create tensors
+        TensorType lhs = create_tensor<TensorType>(lhs_shape, DataType::QASYMM8, 1);
+        TensorType rhs = create_tensor<TensorType>(rhs_shape, DataType::QASYMM8, 1);
+        TensorType dst;
+
+        const unsigned int M = lhs_shape[1];
+        const unsigned int N = rhs_shape[0];
+        const unsigned int K = lhs_shape[0];
+
+        // The output tensor will be auto-initialized within the function
+
+        // Create and configure function
+        GEMMFunctionType gemm;
+        gemm.configure(&lhs, &rhs, &dst, lhs_info, rhs_info, GEMMReshapeInfo(M, N, K));
+
+        ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Allocate tensors
+        lhs.allocator()->allocate();
+        rhs.allocator()->allocate();
+        dst.allocator()->allocate();
+
+        ARM_COMPUTE_EXPECT(!lhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!rhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Fill tensors
+        fill(AccessorType(lhs), 0);
+        fill(AccessorType(rhs), 1);
+
+        // Compute GEMM
+        gemm.run();
+
+        return dst;
+    }
+
+    SimpleTensor<int32_t> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape)
+    {
+        TensorShape dst_shape = lhs_shape;
+        dst_shape[0]          = rhs_shape[0];
+        dst_shape[1]          = lhs_shape[1];
+
+        // Create reference
+        SimpleTensor<uint8_t> lhs{ lhs_shape, DataType::QASYMM8, 1 };
+        SimpleTensor<uint8_t> rhs{ rhs_shape, DataType::QASYMM8, 1 };
+
+        // Fill reference
+        fill(lhs, 0);
+        fill(rhs, 1);
+
+        return reference::gemmlowp_matrix_multiply_core<int32_t, uint8_t>(lhs, rhs, dst_shape, 0, 0);
+    }
+
+    TensorType            _target{};
+    SimpleTensor<int32_t> _reference{};
+};
+
+template <typename TensorType, typename AccessorType, typename GEMMFunctionType>
+class GEMMLowpMatrixMultiplyNative3DValidationFixture : public framework::Fixture
+{
+public:
+    template <typename...>
+    void setup(unsigned int m_w, unsigned int m_h, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0)
+    {
+        GEMMLHSMatrixInfo lhs_info;
+        lhs_info.m0 = m0;
+        lhs_info.k0 = k0;
+
+        GEMMRHSMatrixInfo rhs_info;
+        rhs_info.n0 = n0;
+        rhs_info.k0 = k0;
+
+        // In case of GEMM3D, m is the product between m_w and m_h
+        const unsigned int m = m_w * m_h;
+
+        // Set the tensor shapes for LHS and RHS matrices
+        const TensorShape lhs_shape(k, m, batch_size);
+        const TensorShape rhs_shape(n, k, batch_size);
+
+        _target    = compute_target(lhs_shape, rhs_shape, lhs_info, rhs_info, m_h);
+        _reference = compute_reference(lhs_shape, rhs_shape, m_h);
+    }
+
+protected:
+    template <typename U>
+    void fill(U &&tensor, int i)
+    {
+        // Between 1 and 254 in order to avoid having -128 and 128 for the DOT product path
+        std::uniform_int_distribution<> distribution(1, 254);
+        library->fill(tensor, distribution, i);
+    }
+
+    TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, unsigned int m_h)
+    {
+        // Create tensors
+        TensorType lhs = create_tensor<TensorType>(lhs_shape, DataType::QASYMM8, 1);
+        TensorType rhs = create_tensor<TensorType>(rhs_shape, DataType::QASYMM8, 1);
+        TensorType dst;
+
+        const unsigned int M = lhs_shape[1];
+        const unsigned int N = rhs_shape[0];
+        const unsigned int K = lhs_shape[0];
+
+        // The output tensor will be auto-initialized within the function
+
+        // Create and configure function
+        GEMMFunctionType gemm;
+        gemm.configure(&lhs, &rhs, &dst, lhs_info, rhs_info, GEMMReshapeInfo(M, N, K, 1, 1, m_h));
+
+        ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Allocate tensors
+        lhs.allocator()->allocate();
+        rhs.allocator()->allocate();
+        dst.allocator()->allocate();
+
+        ARM_COMPUTE_EXPECT(!lhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!rhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Fill tensors
+        fill(AccessorType(lhs), 0);
+        fill(AccessorType(rhs), 1);
+
+        // Compute GEMM
+        gemm.run();
+
+        return dst;
+    }
+
+    SimpleTensor<int32_t> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, unsigned int m_h)
+    {
+        TensorShape dst_shape = lhs_shape;
+        dst_shape.set(0, rhs_shape[0]);
+        dst_shape.set(1, lhs_shape[1] / m_h);
+        dst_shape.set(2, m_h);
+        dst_shape.set(3, lhs_shape[2]);
+
+        // Create reference
+        SimpleTensor<uint8_t> lhs{ lhs_shape, DataType::QASYMM8, 1 };
+        SimpleTensor<uint8_t> rhs{ rhs_shape, DataType::QASYMM8, 1 };
+
+        // Fill reference
+        fill(lhs, 0);
+        fill(rhs, 1);
+
+        return reference::gemmlowp_matrix_multiply_core<int32_t, uint8_t>(lhs, rhs, dst_shape, 0, 0);
+    }
+
+    TensorType            _target{};
+    SimpleTensor<int32_t> _reference{};
+};
 } // namespace validation
 } // namespace test
 } // namespace arm_compute