COMPMID-1979: Fuse Activation Function in CLGEMM - part 3

Fused beta*bias in in the old cl gemm kernels
Fused activation function in the old cl gemm kernels

Change-Id: I695fb9189e6d4792010abd256784624982d17d79
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1587
Reviewed-by: Giuseppe Rossini <giuseppe.rossini@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

5 files changed, 1262 insertions, 1 deletions

diff --git a/tests/CL/Helper.h b/tests/CL/Helper.h
index ab2f8ccb22..0a4566be8d 100644
--- a/tests/CL/Helper.h
+++ b/tests/CL/Helper.h
@@ -53,6 +53,19 @@ public:
         k->configure(std::forward<Args>(args)...);
         _kernel = std::move(k);
     }
+    /** Configure the kernel setting the GPU target as well
+     *
+     * @param[in] gpu_target GPUTarget to set
+     * @param[in] args       Configuration arguments.
+     */
+    template <typename... Args>
+    void configure(GPUTarget gpu_target, Args &&... args)
+    {
+        auto k = arm_compute::support::cpp14::make_unique<K>();
+        k->set_target(gpu_target);
+        k->configure(std::forward<Args>(args)...);
+        _kernel = std::move(k);
+    }
     /** Validate input arguments
      *
      * @param[in] args Configuration arguments.
diff --git a/tests/validation/CL/GEMMMatrixMultiply.cpp b/tests/validation/CL/GEMMMatrixMultiply.cpp
new file mode 100644
index 0000000000..21fd7125ec
--- /dev/null
+++ b/tests/validation/CL/GEMMMatrixMultiply.cpp
@@ -0,0 +1,344 @@
+/*
+ * 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/CLGEMMMatrixMultiplyKernel.h"
+#include "arm_compute/core/KernelDescriptors.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "arm_compute/runtime/CL/CLTensor.h"
+#include "arm_compute/runtime/CL/CLTensorAllocator.h"
+#include "tests/CL/CLAccessor.h"
+#include "tests/CL/Helper.h"
+#include "tests/PaddingCalculator.h"
+#include "tests/datasets/ShapeDatasets.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/GEMMFixture.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+using namespace arm_compute::misc::shape_calculator;
+
+// Create function for CLGEMMMatrixMultiplyKernel
+using CLGEMMMatrixMultiplyNative = CLSynthetizeFunction<CLGEMMMatrixMultiplyKernel>;
+
+// Fixture for GEMMMatrixMultiplyValidationFixture
+template <typename T>
+using CLGEMMMatrixMultiplyNativeFixture = GEMMMatrixMultiplyValidationFixture<CLTensor, CLAccessor, T, CLGEMMMatrixMultiplyNative>;
+
+// Fixture for GEMMMatrixMultiply3DValidationFixture
+template <typename T>
+using CLGEMMMatrixMultiplyNative3DFixture = GEMMMatrixMultiply3DValidationFixture<CLTensor, CLAccessor, T, CLGEMMMatrixMultiplyNative>;
+
+namespace
+{
+// *INDENT-OFF*
+// clang-format off
+RelativeTolerance<float> rel_tolerance_f32(0.001f);
+constexpr float          abs_tolerance_f32(0.0001f);
+
+RelativeTolerance<half> rel_tolerance_f16(half(0.2));
+constexpr float         tolerance_num_f16 = 0.02f;
+
+/** Alpha values to test - Precommit */
+const auto alpha_values = framework::dataset::make("alpha", {0.0f, 1.0f, -0.75f} );
+
+/** Beta values to test - Precommit */
+const auto beta_values = framework::dataset::make("beta", {-0.75f, 0.0f} );
+
+/** M values to test - Precommit */
+const auto m_values_precommit = framework::dataset::make("M", {37, 1});
+
+/** N values to test - Precommit */
+const auto n_values_precommit = framework::dataset::make("N", 51);
+
+/** K values to test - Precommit */
+const auto k_values_precommit = framework::dataset::make("K", 23);
+
+/** M values to test - Nightly */
+const auto m_values_nightly = framework::dataset::make("M", {421, 1});
+
+/** N values to test - Nightly */
+const auto n_values_nightly = framework::dataset::make("N", {323, 1103});
+
+/** K values to test - Nightly */
+const auto k_values_nightly = framework::dataset::make("K", 207);
+
+/** M_W values to test - Precommit */
+const auto m_w_values_precommit = framework::dataset::make("M_W", 5);
+
+/** M_H values to test - Precommit */
+const auto m_h_values_precommit = framework::dataset::make("M_H", 7);
+
+/** M_W values to test - Nightly */
+const auto m_w_values_nightly = framework::dataset::make("M_W", 13);
+
+/** M_H values to test - Nightly */
+const auto m_h_values_nightly = framework::dataset::make("M_H", 27);
+
+/** Batch size values to test */
+const auto b_values = framework::dataset::make("batch_size", 1, 3);
+
+/** Activation values to test */
+const auto act_values = framework::dataset::make("Activation",
+{
+    ActivationLayerInfo(),
+    ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 8.f, 2.f),
+});
+
+/** Broadcast bias from vector to matrix */
+const auto broadcast_bias_values = framework::dataset::make("broadcast_bias", {false, true} );
+
+/** GPU architectures values to test */
+const auto gpu_arch_values = framework::dataset::make("GPUArch",
+{
+    GPUTarget::MIDGARD,
+    GPUTarget::BIFROST
+});
+
+/** Data types values to test in the configuration */
+const auto data_type_values = framework::dataset::make("DataType",
+{
+    DataType::F32,
+    DataType::F16
+});
+
+/** M values to test */
+const auto fp16_mixed_precision_values = framework::dataset::make("fp16_mixed_precision", {true, false});
+
+/** Configuration test */
+void validate_configuration(unsigned int m_value, unsigned int n_value, unsigned int k_value, unsigned int b_value, bool broadcast_bias, bool fp16_mixed_precision, const ActivationLayerInfo &act_info, DataType data_type, GPUTarget gpu_arch_value)
+{
+    GEMMReshapeInfo reshape_info(m_value, n_value, k_value, 1, 1, 0, false, broadcast_bias);
+
+    const TensorShape lhs_shape(k_value, m_value, b_value);
+    const TensorShape rhs_shape(n_value, k_value, b_value);
+
+    const TensorShape dst_shape = compute_mm_shape(TensorInfo(lhs_shape, 1, data_type),
+                                                   TensorInfo(rhs_shape, 1, data_type),
+                                                   reshape_info);
+
+    const TensorShape bias_shape(n_value,
+                                 broadcast_bias? 1 : m_value,
+                                 broadcast_bias? 1 : b_value);
+
+    // Create tensors
+    CLTensor lhs  = create_tensor<CLTensor>(lhs_shape, data_type);
+    CLTensor rhs  = create_tensor<CLTensor>(rhs_shape, data_type);
+    CLTensor bias = create_tensor<CLTensor>(bias_shape, data_type);
+    CLTensor dst  = create_tensor<CLTensor>(dst_shape, data_type);
+
+    ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+    ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+    ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS);
+    ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+    // Create and configure function
+    CLGEMMMatrixMultiplyNative gemm;
+    gemm.configure(gpu_arch_value, &lhs, &rhs, &bias, &dst, 1.0f, 2.0f, false, reshape_info, fp16_mixed_precision, act_info);
+}
+} // namespace
+
+TEST_SUITE(CL)
+TEST_SUITE(GEMMMatrixMultiply)
+TEST_SUITE(Float)
+TEST_SUITE(FP32)
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(combine(
+                                                                   m_values_precommit,
+                                                                   n_values_precommit),
+                                                                   k_values_precommit),
+                                                                   framework::dataset::make("batch_size", 1)),
+                                                                   broadcast_bias_values),
+                                                                   framework::dataset::make("fp16_mixed_precision", false)),
+                                                                   act_values),
+                                                                   data_type_values),
+                                                                   gpu_arch_values),
+m_value, n_value, k_value, b_value, broadcast_bias, fp16_mixed_precision_value, act_value, data_type_value, gpu_arch_value)
+{
+    validate_configuration(m_value, n_value, k_value, b_value, broadcast_bias, fp16_mixed_precision_value, act_value, data_type_value, gpu_arch_value);
+}
+
+FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyNativeFixture<float>, framework::DatasetMode::ALL,
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                                                                   m_values_precommit,
+                                                                   n_values_precommit),
+                                                                   k_values_precommit),
+                                                                   b_values),
+                                                                   alpha_values),
+                                                                   beta_values),
+                                                                   broadcast_bias_values),
+                                                                   framework::dataset::make("fp16_mixed_precision", false)),
+                                                                   act_values),
+                                                                   framework::dataset::make("DataType", DataType::F32)),
+                                                                   gpu_arch_values))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMMatrixMultiplyNativeFixture<float>, framework::DatasetMode::NIGHTLY,
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                                                                   m_values_nightly,
+                                                                   n_values_nightly),
+                                                                   k_values_nightly),
+                                                                   b_values),
+                                                                   alpha_values),
+                                                                   beta_values),
+                                                                   broadcast_bias_values),
+                                                                   framework::dataset::make("fp16_mixed_precision", false)),
+                                                                   act_values),
+                                                                   framework::dataset::make("DataType", DataType::F32)),
+                                                                   gpu_arch_values))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
+}
+
+FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMMatrixMultiplyNative3DFixture<float>, framework::DatasetMode::ALL,
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                                                                   m_w_values_precommit,
+                                                                   m_h_values_precommit),
+                                                                   n_values_precommit),
+                                                                   k_values_precommit),
+                                                                   b_values),
+                                                                   alpha_values),
+                                                                   beta_values),
+                                                                   broadcast_bias_values),
+                                                                   framework::dataset::make("fp16_mixed_precision", false)),
+                                                                   act_values),
+                                                                   framework::dataset::make("DataType", DataType::F32)),
+                                                                   gpu_arch_values))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyNative3DFixture<float>, framework::DatasetMode::NIGHTLY,
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                                                                   m_w_values_nightly,
+                                                                   m_h_values_nightly),
+                                                                   n_values_nightly),
+                                                                   k_values_nightly),
+                                                                   b_values),
+                                                                   alpha_values),
+                                                                   beta_values),
+                                                                   broadcast_bias_values),
+                                                                   framework::dataset::make("fp16_mixed_precision", false)),
+                                                                   act_values),
+                                                                   framework::dataset::make("DataType", DataType::F32)),
+                                                                   gpu_arch_values))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
+}
+
+TEST_SUITE_END() // FP32
+
+TEST_SUITE(FP16)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyNativeFixture<half>, framework::DatasetMode::ALL,
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                                                                   m_values_precommit,
+                                                                   n_values_precommit),
+                                                                   k_values_precommit),
+                                                                   b_values),
+                                                                   alpha_values),
+                                                                   beta_values),
+                                                                   broadcast_bias_values),
+                                                                   fp16_mixed_precision_values),
+                                                                   act_values),
+                                                                   framework::dataset::make("DataType", DataType::F16)),
+                                                                   gpu_arch_values))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMMatrixMultiplyNativeFixture<half>, framework::DatasetMode::NIGHTLY,
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                                                                   m_values_nightly,
+                                                                   n_values_nightly),
+                                                                   k_values_nightly),
+                                                                   b_values),
+                                                                   alpha_values),
+                                                                   beta_values),
+                                                                   broadcast_bias_values),
+                                                                   fp16_mixed_precision_values),
+                                                                   act_values),
+                                                                   framework::dataset::make("DataType", DataType::F16)),
+                                                                   gpu_arch_values))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16);
+}
+
+FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMMatrixMultiplyNative3DFixture<half>, framework::DatasetMode::ALL,
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                                                                   m_w_values_precommit,
+                                                                   m_h_values_precommit),
+                                                                   n_values_precommit),
+                                                                   k_values_precommit),
+                                                                   b_values),
+                                                                   alpha_values),
+                                                                   beta_values),
+                                                                   broadcast_bias_values),
+                                                                   fp16_mixed_precision_values),
+                                                                   act_values),
+                                                                   framework::dataset::make("DataType", DataType::F16)),
+                                                                   gpu_arch_values))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyNative3DFixture<half>, framework::DatasetMode::NIGHTLY,
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                                                                   m_w_values_nightly,
+                                                                   m_h_values_nightly),
+                                                                   n_values_nightly),
+                                                                   k_values_nightly),
+                                                                   b_values),
+                                                                   alpha_values),
+                                                                   beta_values),
+                                                                   broadcast_bias_values),
+                                                                   fp16_mixed_precision_values),
+                                                                   act_values),
+                                                                   framework::dataset::make("DataType", DataType::F16)),
+                                                                   gpu_arch_values))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16);
+}
+
+TEST_SUITE_END() // FP16
+TEST_SUITE_END() // Float
+TEST_SUITE_END() // GEMMMatrixMuliplty
+TEST_SUITE_END() // CL
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
\ No newline at end of file
diff --git a/tests/validation/CL/GEMMMatrixMultiplyInterleavedTransposed.cpp b/tests/validation/CL/GEMMMatrixMultiplyInterleavedTransposed.cpp
new file mode 100644
index 0000000000..cae94b2e15
--- /dev/null
+++ b/tests/validation/CL/GEMMMatrixMultiplyInterleavedTransposed.cpp
@@ -0,0 +1,404 @@
+/*
+ * 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/CLGEMMMatrixMultiplyKernel.h"
+#include "arm_compute/core/CL/kernels/CLGEMMReshapeLHSMatrixKernel.h"
+#include "arm_compute/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.h"
+#include "arm_compute/core/KernelDescriptors.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "arm_compute/runtime/CL/CLTensor.h"
+#include "arm_compute/runtime/CL/CLTensorAllocator.h"
+#include "tests/CL/CLAccessor.h"
+#include "tests/CL/Helper.h"
+#include "tests/PaddingCalculator.h"
+#include "tests/datasets/ShapeDatasets.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/GEMMFixture.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+using namespace arm_compute::misc::shape_calculator;
+
+// Create function for CLGEMMReshapeLHSMatrixKernel
+using CLGEMMReshapeLHSMatrix = CLSynthetizeFunction<CLGEMMReshapeLHSMatrixKernel>;
+
+// Create function for CLGEMMReshapeRHSMatrixKernel
+using CLGEMMReshapeRHSMatrix = CLSynthetizeFunction<CLGEMMReshapeRHSMatrixKernel>;
+
+// Create function for CLGEMMMatrixMultiplyKernel
+using CLGEMMMatrixMultiplyReshaped = CLSynthetizeFunction<CLGEMMMatrixMultiplyKernel>;
+
+// Fixture for GEMMMatrixMultiplyInterleavedTransposedValidationFixture
+template <typename T>
+using CLGEMMMatrixMultiplyReshapedFixture =
+    GEMMMatrixMultiplyInterleavedTransposedValidationFixture<CLTensor, CLAccessor, T, CLGEMMReshapeLHSMatrix, CLGEMMReshapeRHSMatrix, CLGEMMMatrixMultiplyReshaped>;
+
+// Fixture for GEMMMatrixMultiplyInterleavedTransposed3DValidationFixture
+template <typename T>
+using CLGEMMMatrixMultiplyReshaped3DFixture =
+    GEMMMatrixMultiplyInterleavedTransposed3DValidationFixture<CLTensor, CLAccessor, T, CLGEMMReshapeLHSMatrix, CLGEMMReshapeRHSMatrix, CLGEMMMatrixMultiplyReshaped>;
+
+namespace
+{
+// *INDENT-OFF*
+// clang-format off
+RelativeTolerance<float> rel_tolerance_f32(0.001f);
+constexpr float          abs_tolerance_f32(0.0001f);
+
+RelativeTolerance<half> rel_tolerance_f16(half(0.2));
+constexpr float         tolerance_num_f16 = 0.02f;
+
+/** Alpha values to test - Precommit */
+const auto alpha_values = framework::dataset::make("alpha", {0.0f, 1.0f, -0.75f} );
+
+/** Beta values to test - Precommit */
+const auto beta_values = framework::dataset::make("beta", {-0.75f, 0.0f} );
+
+/** M values to test - Precommit */
+const auto m_values_precommit = framework::dataset::make("M", 37);
+
+/** N values to test - Precommit */
+const auto n_values_precommit = framework::dataset::make("N", 51);
+
+/** K values to test - Precommit */
+const auto k_values_precommit = framework::dataset::make("K", 23);
+
+/** M values to test - Nightly */
+const auto m_values_nightly = framework::dataset::make("M", {421, 1});
+
+/** N values to test - Nightly */
+const auto n_values_nightly = framework::dataset::make("N", 323);
+
+/** K values to test - Nightly */
+const auto k_values_nightly = framework::dataset::make("K", 207);
+
+/** M_W values to test - Precommit */
+const auto m_w_values_precommit = framework::dataset::make("M_W", 5);
+
+/** M_H values to test - Precommit */
+const auto m_h_values_precommit = framework::dataset::make("M_H", 7);
+
+/** M_W values to test - Nightly */
+const auto m_w_values_nightly = framework::dataset::make("M_W", 13);
+
+/** M_H values to test - Nightly */
+const auto m_h_values_nightly = framework::dataset::make("M_H", 27);
+
+/** Batch size values to test */
+const auto b_values = framework::dataset::make("batch_size", 1, 3);
+
+/** Activation values to test */
+const auto act_values = framework::dataset::make("Activation",
+{
+    ActivationLayerInfo(),
+    ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 8.f, 2.f),
+});
+
+/** V0 values to test - Precommit */
+const auto v0_values_precommit = framework::dataset::make("V0", 2);
+
+/** H0 values to test - Precommit */
+const auto h0_values_precommit = framework::dataset::make("H0", 4);
+
+/** V0 values to test - Nightly */
+const auto v0_values_nightly = framework::dataset::make("V0", {2, 4});
+
+/** H0 values to test - Nightly */
+const auto h0_values_nightly = framework::dataset::make("H0", { 2, 4 });
+
+/** Broadcast bias from vector to matrix */
+const auto broadcast_bias_values = framework::dataset::make("broadcast_bias", {false, true} );
+
+/** GPU architectures values to test */
+const auto gpu_arch_values = framework::dataset::make("GPUArch",
+{
+    GPUTarget::MIDGARD,
+    GPUTarget::BIFROST
+});
+
+/** Data types values to test in the configuration */
+const auto data_type_values = framework::dataset::make("DataType",
+{
+    DataType::F32,
+    DataType::F16
+});
+
+/** M values to test */
+const auto fp16_mixed_precision_values = framework::dataset::make("fp16_mixed_precision", {true, false});
+
+/** Configuration test */
+void validate_configuration(unsigned int m_value, unsigned int n_value, unsigned int k_value, unsigned int b_value, unsigned int v0_value, unsigned int h0_value, bool broadcast_bias, bool fp16_mixed_precision, const ActivationLayerInfo &act_info, DataType data_type, GPUTarget gpu_arch_value)
+{
+    GEMMLHSMatrixInfo lhs_info;
+    lhs_info.m0         = 4;
+    lhs_info.k0         = 4;
+    lhs_info.v0         = v0_value;
+    lhs_info.interleave = true;
+    lhs_info.transpose  = true;
+
+    GEMMRHSMatrixInfo rhs_info;
+    rhs_info.n0         = data_type == DataType::F32? 4 : 8;
+    rhs_info.k0         = 1;
+    rhs_info.h0         = h0_value;
+    rhs_info.interleave = false;
+    rhs_info.transpose  = false;
+
+    GEMMReshapeInfo reshape_info(m_value, n_value, k_value, rhs_info.h0, lhs_info.v0, 0, false, broadcast_bias);
+
+    const TensorShape lhs_shape(k_value, m_value, b_value);
+    const TensorShape lhs_shape_reshaped = compute_lhs_reshaped_shape(TensorInfo(lhs_shape, 1, data_type),
+                                                                      lhs_info,
+                                                                      false);
+
+    const TensorShape rhs_shape(n_value, k_value, b_value);
+    const TensorShape rhs_shape_reshaped = compute_rhs_reshaped_shape(TensorInfo(rhs_shape, 1, data_type),
+                                                                      rhs_info);
+
+    const TensorShape dst_shape = compute_mm_shape(TensorInfo(lhs_shape_reshaped, 1, data_type),
+                                                   TensorInfo(rhs_shape_reshaped, 1, data_type),
+                                                   reshape_info);
+
+    const TensorShape bias_shape(n_value,
+                                 broadcast_bias? 1 : m_value,
+                                 broadcast_bias? 1 : b_value);
+
+    // Create tensors
+    CLTensor lhs_reshaped = create_tensor<CLTensor>(lhs_shape_reshaped, data_type);
+    CLTensor rhs_reshaped = create_tensor<CLTensor>(rhs_shape_reshaped, data_type);
+    CLTensor bias         = create_tensor<CLTensor>(bias_shape, data_type);
+    CLTensor dst          = create_tensor<CLTensor>(dst_shape, data_type);
+
+    ARM_COMPUTE_EXPECT(lhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS);
+    ARM_COMPUTE_EXPECT(rhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS);
+    ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS);
+    ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+    // Create and configure function
+    CLGEMMMatrixMultiplyReshaped gemm;
+    gemm.configure(gpu_arch_value, &lhs_reshaped, &rhs_reshaped, &bias, &dst, 1.0f, 2.0f, true, reshape_info, fp16_mixed_precision, act_info);
+}
+} // namespace
+
+TEST_SUITE(CL)
+TEST_SUITE(GEMMMatrixMultiplyInterleavedTransposed)
+TEST_SUITE(Float)
+TEST_SUITE(FP32)
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                                                                   m_values_precommit,
+                                                                   n_values_precommit),
+                                                                   k_values_precommit),
+                                                                   framework::dataset::make("batch_size", 1)),
+                                                                   v0_values_precommit),
+                                                                   h0_values_precommit),
+                                                                   broadcast_bias_values),
+                                                                   framework::dataset::make("fp16_mixed_precision", false)),
+                                                                   act_values),
+                                                                   data_type_values),
+                                                                   gpu_arch_values),
+m_value, n_value, k_value, b_value, v0_value, h0_value, broadcast_bias, fp16_mixed_precision_value, act_value, data_type_value, gpu_arch_value)
+{
+    validate_configuration(m_value, n_value, k_value, b_value, v0_value, h0_value, broadcast_bias, fp16_mixed_precision_value, act_value, data_type_value, gpu_arch_value);
+}
+
+FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyReshapedFixture<float>, framework::DatasetMode::ALL,
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                                                                   m_values_precommit,
+                                                                   n_values_precommit),
+                                                                   k_values_precommit),
+                                                                   b_values),
+                                                                   alpha_values),
+                                                                   beta_values),
+                                                                   v0_values_precommit),
+                                                                   h0_values_precommit),
+                                                                   broadcast_bias_values),
+                                                                   framework::dataset::make("fp16_mixed_precision", false)),
+                                                                   act_values),
+                                                                   framework::dataset::make("DataType", DataType::F32)),
+                                                                   gpu_arch_values))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMMatrixMultiplyReshapedFixture<float>, framework::DatasetMode::NIGHTLY,
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                                                                   m_values_nightly,
+                                                                   n_values_nightly),
+                                                                   k_values_nightly),
+                                                                   b_values),
+                                                                   alpha_values),
+                                                                   beta_values),
+                                                                   v0_values_nightly),
+                                                                   h0_values_nightly),
+                                                                   broadcast_bias_values),
+                                                                   framework::dataset::make("fp16_mixed_precision", false)),
+                                                                   act_values),
+                                                                   framework::dataset::make("DataType", DataType::F32)),
+                                                                   gpu_arch_values))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
+}
+
+FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMMatrixMultiplyReshaped3DFixture<float>, framework::DatasetMode::ALL,
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                                                                   m_w_values_precommit,
+                                                                   m_h_values_precommit),
+                                                                   n_values_precommit),
+                                                                   k_values_precommit),
+                                                                   b_values),
+                                                                   alpha_values),
+                                                                   beta_values),
+                                                                   v0_values_precommit),
+                                                                   h0_values_precommit),
+                                                                   broadcast_bias_values),
+                                                                   framework::dataset::make("fp16_mixed_precision", false)),
+                                                                   act_values),
+                                                                   framework::dataset::make("DataType", DataType::F32)),
+                                                                   gpu_arch_values))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyReshaped3DFixture<float>, framework::DatasetMode::NIGHTLY,
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                                                                   m_w_values_nightly,
+                                                                   m_h_values_nightly),
+                                                                   n_values_nightly),
+                                                                   k_values_nightly),
+                                                                   b_values),
+                                                                   alpha_values),
+                                                                   beta_values),
+                                                                   v0_values_nightly),
+                                                                   h0_values_nightly),
+                                                                   broadcast_bias_values),
+                                                                   framework::dataset::make("fp16_mixed_precision", false)),
+                                                                   act_values),
+                                                                   framework::dataset::make("DataType", DataType::F32)),
+                                                                   gpu_arch_values))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
+}
+
+TEST_SUITE_END() // FP32
+
+TEST_SUITE(FP16)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyReshapedFixture<half>, framework::DatasetMode::ALL,
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                                                                   m_values_precommit,
+                                                                   n_values_precommit),
+                                                                   k_values_precommit),
+                                                                   b_values),
+                                                                   alpha_values),
+                                                                   beta_values),
+                                                                   v0_values_precommit),
+                                                                   h0_values_precommit),
+                                                                   broadcast_bias_values),
+                                                                   fp16_mixed_precision_values),
+                                                                   act_values),
+                                                                   framework::dataset::make("DataType", DataType::F16)),
+                                                                   gpu_arch_values))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMMatrixMultiplyReshapedFixture<half>, framework::DatasetMode::NIGHTLY,
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                                                                   m_values_nightly,
+                                                                   n_values_nightly),
+                                                                   k_values_nightly),
+                                                                   b_values),
+                                                                   alpha_values),
+                                                                   beta_values),
+                                                                   v0_values_nightly),
+                                                                   h0_values_nightly),
+                                                                   broadcast_bias_values),
+                                                                   fp16_mixed_precision_values),
+                                                                   act_values),
+                                                                   framework::dataset::make("DataType", DataType::F16)),
+                                                                   gpu_arch_values))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16);
+}
+
+FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMMatrixMultiplyReshaped3DFixture<half>, framework::DatasetMode::ALL,
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                                                                   m_w_values_precommit,
+                                                                   m_h_values_precommit),
+                                                                   n_values_precommit),
+                                                                   k_values_precommit),
+                                                                   b_values),
+                                                                   alpha_values),
+                                                                   beta_values),
+                                                                   v0_values_precommit),
+                                                                   h0_values_precommit),
+                                                                   broadcast_bias_values),
+                                                                   fp16_mixed_precision_values),
+                                                                   act_values),
+                                                                   framework::dataset::make("DataType", DataType::F16)),
+                                                                   gpu_arch_values))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyReshaped3DFixture<half>, framework::DatasetMode::NIGHTLY,
+                combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+                                                                   m_w_values_nightly,
+                                                                   m_h_values_nightly),
+                                                                   n_values_nightly),
+                                                                   k_values_nightly),
+                                                                   b_values),
+                                                                   alpha_values),
+                                                                   beta_values),
+                                                                   v0_values_nightly),
+                                                                   h0_values_nightly),
+                                                                   broadcast_bias_values),
+                                                                   fp16_mixed_precision_values),
+                                                                   act_values),
+                                                                   framework::dataset::make("DataType", DataType::F16)),
+                                                                   gpu_arch_values))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16);
+}
+
+TEST_SUITE_END() // FP16
+TEST_SUITE_END() // Float
+TEST_SUITE_END() // GEMMMatrixMulipltyInterleavedTransposed
+TEST_SUITE_END() // CL
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
\ No newline at end of file
diff --git a/tests/validation/CL/GEMMMatrixMultiplyReshaped.cpp b/tests/validation/CL/GEMMMatrixMultiplyReshaped.cpp
index 99af2965d2..25221451ed 100644
--- a/tests/validation/CL/GEMMMatrixMultiplyReshaped.cpp
+++ b/tests/validation/CL/GEMMMatrixMultiplyReshaped.cpp
@@ -418,7 +418,7 @@ FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyReshaped3DFixture<half>,
 }
 TEST_SUITE_END() // FP16
 TEST_SUITE_END() // Float
-TEST_SUITE_END() // GEMMMatrixMulipltyReshaped
+TEST_SUITE_END() // GEMMMatrixMultiplyReshaped
 TEST_SUITE_END() // CL
 } // namespace validation
 } // namespace test
diff --git a/tests/validation/fixtures/GEMMFixture.h b/tests/validation/fixtures/GEMMFixture.h
index ac8ab2a949..b36bb99246 100644
--- a/tests/validation/fixtures/GEMMFixture.h
+++ b/tests/validation/fixtures/GEMMFixture.h
@@ -153,6 +153,506 @@ protected:
     SimpleTensor<T> _reference{};
 };
 
+template <typename TensorType, typename AccessorType, typename T, typename GEMMFunctionType>
+class GEMMMatrixMultiplyValidationFixture : public framework::Fixture
+{
+public:
+    template <typename...>
+    void setup(unsigned int m, unsigned int n, unsigned int k, unsigned int batch_size, float alpha, float beta, bool broadcast_bias, bool fp16_mixed_precision, const ActivationLayerInfo &act_info,
+               DataType data_type, GPUTarget gpu_arch)
+    {
+        // 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);
+        const TensorShape bias_shape(n,
+                                     broadcast_bias ? 1 : m,
+                                     broadcast_bias ? 1 : batch_size);
+
+        _target    = compute_target(lhs_shape, rhs_shape, bias_shape, data_type, alpha, beta, broadcast_bias, fp16_mixed_precision, act_info, gpu_arch);
+        _reference = compute_reference(lhs_shape, rhs_shape, bias_shape, data_type, alpha, beta, broadcast_bias, act_info);
+    }
+
+protected:
+    template <typename U>
+    void fill(U &&tensor, int i)
+    {
+        std::uniform_real_distribution<> distribution(-1.0f, 1.0f);
+        library->fill(tensor, distribution, i);
+
+        // Fill border with infinity in order to check the presence of NaN values (i.e. inf * 0)
+        std::uniform_real_distribution<> distribution_inf(std::numeric_limits<float>::infinity(), std::numeric_limits<float>::infinity());
+        library->fill_borders_with_garbage(tensor, distribution_inf, i);
+    }
+
+    TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, DataType data_type, float alpha, float beta, bool broadcast_bias,
+                              bool fp16_mixed_precision, const ActivationLayerInfo &act_info, GPUTarget gpu_arch)
+    {
+        // Create tensors
+        TensorType lhs  = create_tensor<TensorType>(lhs_shape, data_type, 1);
+        TensorType rhs  = create_tensor<TensorType>(rhs_shape, data_type, 1);
+        TensorType bias = create_tensor<TensorType>(bias_shape, data_type, 1);
+        TensorType dst;
+
+        const unsigned int m = lhs_shape[1];
+        const unsigned int n = rhs_shape[0];
+        const unsigned int k = lhs_shape[0];
+        GEMMReshapeInfo    reshape_info(m, n, k, 1, 1, 0, false, broadcast_bias);
+
+        // The output tensor will be auto-initialized within the function
+
+        // Create and configure function
+        GEMMFunctionType gemm;
+        gemm.configure(gpu_arch, &lhs, &rhs, &bias, &dst, alpha, beta, false, reshape_info, fp16_mixed_precision, act_info);
+
+        ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Allocate tensors
+        lhs.allocator()->allocate();
+        rhs.allocator()->allocate();
+        bias.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(!bias.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);
+        fill(AccessorType(bias), 2);
+
+        // Compute GEMM
+        gemm.run();
+
+        return dst;
+    }
+
+    SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, DataType data_type, float alpha, float beta, bool broadcast_bias,
+                                      const ActivationLayerInfo &act_info)
+    {
+        TensorShape dst_shape = lhs_shape;
+        dst_shape[0]          = rhs_shape[0];
+        dst_shape[1]          = lhs_shape[1];
+
+        // Create reference
+        SimpleTensor<T> lhs{ lhs_shape, data_type, 1 };
+        SimpleTensor<T> rhs{ rhs_shape, data_type, 1 };
+        SimpleTensor<T> bias{ dst_shape, data_type, 1 };
+
+        const int n          = rhs_shape[0];
+        const int m          = lhs_shape[1];
+        const int batch_size = lhs_shape[2];
+
+        // Fill reference
+        fill(lhs, 0);
+        fill(rhs, 1);
+        fill(bias, 2);
+
+        if(broadcast_bias)
+        {
+            // In case of broadcast, we need simply copy the first into the following "M" ones
+            for(int i = 1; i < m * batch_size; i++)
+            {
+                memcpy(bias.data() + i * n, bias.data(), n * sizeof(T));
+            }
+        }
+
+        return reference::activation_layer(reference::gemm<T>(lhs, rhs, bias, alpha, beta), act_info);
+    }
+
+    TensorType      _target{};
+    SimpleTensor<T> _reference{};
+};
+
+template <typename TensorType, typename AccessorType, typename T, typename GEMMFunctionType>
+class GEMMMatrixMultiply3DValidationFixture : 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, float alpha, float beta, bool broadcast_bias, bool fp16_mixed_precision,
+               const ActivationLayerInfo &act_info, DataType data_type, GPUTarget gpu_arch)
+    {
+        // 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);
+        const TensorShape bias_shape(n, 1, 1);
+
+        _target    = compute_target(lhs_shape, rhs_shape, bias_shape, data_type, alpha, beta, m_h, fp16_mixed_precision, act_info, gpu_arch);
+        _reference = compute_reference(lhs_shape, rhs_shape, bias_shape, data_type, alpha, beta, m_h, act_info);
+    }
+
+protected:
+    template <typename U>
+    void fill(U &&tensor, int i)
+    {
+        std::uniform_real_distribution<> distribution(-1.0f, 1.0f);
+        library->fill(tensor, distribution, i);
+    }
+
+    TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, DataType data_type, float alpha, float beta, unsigned int m_h,
+                              bool fp16_mixed_precision, const ActivationLayerInfo &act_info, GPUTarget gpu_arch)
+    {
+        // Create tensors
+        TensorType lhs  = create_tensor<TensorType>(lhs_shape, data_type, 1);
+        TensorType rhs  = create_tensor<TensorType>(rhs_shape, data_type, 1);
+        TensorType bias = create_tensor<TensorType>(bias_shape, data_type, 1);
+        TensorType dst;
+
+        const unsigned int m = lhs_shape[1];
+        const unsigned int n = rhs_shape[0];
+        const unsigned int k = lhs_shape[0];
+        GEMMReshapeInfo    reshape_info(m, n, k, 1, 1, m_h, false, true);
+
+        // The output tensor will be auto-initialized within the function
+
+        // Create and configure function
+        GEMMFunctionType gemm;
+        gemm.configure(gpu_arch, &lhs, &rhs, &bias, &dst, alpha, beta, false, reshape_info, fp16_mixed_precision, act_info);
+
+        ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Allocate tensors
+        lhs.allocator()->allocate();
+        rhs.allocator()->allocate();
+        bias.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(!bias.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);
+        fill(AccessorType(bias), 2);
+
+        // Compute GEMM
+        gemm.run();
+
+        return dst;
+    }
+
+    SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, DataType data_type, float alpha, float beta, unsigned int m_h,
+                                      const ActivationLayerInfo &act_info)
+    {
+        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<T> lhs{ lhs_shape, data_type, 1 };
+        SimpleTensor<T> rhs{ rhs_shape, data_type, 1 };
+        SimpleTensor<T> bias{ dst_shape, data_type, 1 };
+
+        const int n          = rhs_shape[0];
+        const int m          = lhs_shape[1];
+        const int batch_size = lhs_shape[2];
+
+        // Fill reference
+        fill(lhs, 0);
+        fill(rhs, 1);
+        fill(bias, 2);
+
+        // In case of broadcast, we need simply copy the first into the following "M" ones
+        for(int i = 1; i < m * batch_size; i++)
+        {
+            memcpy(bias.data() + i * n, bias.data(), n * sizeof(T));
+        }
+
+        return reference::activation_layer(reference::gemm<T>(lhs, rhs, bias, alpha, beta), act_info);
+    }
+
+    TensorType      _target{};
+    SimpleTensor<T> _reference{};
+};
+
+template <typename TensorType, typename AccessorType, typename T, typename ReshapeLHSFunctionType, typename ReshapeRHSFunctionType, typename GEMMFunctionType>
+class GEMMMatrixMultiplyInterleavedTransposedValidationFixture : public framework::Fixture
+{
+public:
+    template <typename...>
+    void setup(unsigned int m, unsigned int n, unsigned int k, unsigned int batch_size, float alpha, float beta, unsigned int v0, unsigned int h0, bool broadcast_bias, bool fp16_mixed_precision,
+               const ActivationLayerInfo &act_info, DataType data_type, GPUTarget gpu_arch)
+    {
+        GEMMLHSMatrixInfo lhs_info;
+        lhs_info.m0         = 4;
+        lhs_info.k0         = 4;
+        lhs_info.v0         = v0;
+        lhs_info.interleave = true;
+        lhs_info.transpose  = true;
+
+        GEMMRHSMatrixInfo rhs_info;
+        rhs_info.n0         = 16 / sizeof(T);
+        rhs_info.k0         = 1;
+        rhs_info.h0         = h0;
+        rhs_info.interleave = false;
+        rhs_info.transpose  = false;
+
+        // 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);
+        const TensorShape bias_shape(n,
+                                     broadcast_bias ? 1 : m,
+                                     broadcast_bias ? 1 : batch_size);
+
+        _target    = compute_target(lhs_shape, rhs_shape, bias_shape, lhs_info, rhs_info, data_type, alpha, beta, broadcast_bias, fp16_mixed_precision, act_info, gpu_arch);
+        _reference = compute_reference(lhs_shape, rhs_shape, bias_shape, data_type, alpha, beta, broadcast_bias, act_info);
+    }
+
+protected:
+    template <typename U>
+    void fill(U &&tensor, int i)
+    {
+        std::uniform_real_distribution<> distribution(-1.0f, 1.0f);
+        library->fill(tensor, distribution, i);
+
+        // Fill border with infinity in order to check the presence of NaN values (i.e. inf * 0)
+        std::uniform_real_distribution<> distribution_inf(std::numeric_limits<float>::infinity(), std::numeric_limits<float>::infinity());
+        library->fill_borders_with_garbage(tensor, distribution_inf, i);
+    }
+
+    TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info,
+                              DataType data_type, float alpha, float beta, bool broadcast_bias, bool fp16_mixed_precision, const ActivationLayerInfo &act_info, GPUTarget gpu_arch)
+    {
+        // Create tensors
+        TensorType lhs  = create_tensor<TensorType>(lhs_shape, data_type, 1);
+        TensorType rhs  = create_tensor<TensorType>(rhs_shape, data_type, 1);
+        TensorType bias = create_tensor<TensorType>(bias_shape, data_type, 1);
+        TensorType lhs_reshaped;
+        TensorType rhs_reshaped;
+        TensorType dst;
+
+        const unsigned int m = lhs_shape[1];
+        const unsigned int n = rhs_shape[0];
+        const unsigned int k = lhs_shape[0];
+        GEMMReshapeInfo    reshape_info(m, n, k, rhs_info.h0, lhs_info.v0, 0, false, broadcast_bias);
+
+        // The output tensor will be auto-initialized within the function
+
+        // Create and configure function
+        ReshapeLHSFunctionType reshape_lhs;
+        ReshapeRHSFunctionType reshape_rhs;
+        GEMMFunctionType       gemm;
+        reshape_lhs.configure(&lhs, &lhs_reshaped, lhs_info);
+        reshape_rhs.configure(&rhs, &rhs_reshaped, rhs_info);
+        gemm.configure(gpu_arch, &lhs_reshaped, &rhs_reshaped, &bias, &dst, alpha, beta, true, reshape_info, fp16_mixed_precision, act_info);
+
+        ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Allocate tensors
+        lhs.allocator()->allocate();
+        rhs.allocator()->allocate();
+        lhs_reshaped.allocator()->allocate();
+        rhs_reshaped.allocator()->allocate();
+        bias.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(!bias.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!lhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!rhs_reshaped.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);
+        fill(AccessorType(bias), 2);
+
+        // Compute GEMM
+        reshape_lhs.run();
+        reshape_rhs.run();
+        gemm.run();
+
+        return dst;
+    }
+
+    SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, DataType data_type, float alpha, float beta, bool broadcast_bias,
+                                      const ActivationLayerInfo &act_info)
+    {
+        TensorShape dst_shape = lhs_shape;
+        dst_shape[0]          = rhs_shape[0];
+        dst_shape[1]          = lhs_shape[1];
+
+        // Create reference
+        SimpleTensor<T> lhs{ lhs_shape, data_type, 1 };
+        SimpleTensor<T> rhs{ rhs_shape, data_type, 1 };
+        SimpleTensor<T> bias{ dst_shape, data_type, 1 };
+
+        const int n          = rhs_shape[0];
+        const int m          = lhs_shape[1];
+        const int batch_size = lhs_shape[2];
+
+        // Fill reference
+        fill(lhs, 0);
+        fill(rhs, 1);
+        fill(bias, 2);
+
+        if(broadcast_bias)
+        {
+            // In case of broadcast, we need simply copy the first into the following "M" ones
+            for(int i = 1; i < m * batch_size; i++)
+            {
+                memcpy(bias.data() + i * n, bias.data(), n * sizeof(T));
+            }
+        }
+
+        return reference::activation_layer(reference::gemm<T>(lhs, rhs, bias, alpha, beta), act_info);
+    }
+
+    TensorType      _target{};
+    SimpleTensor<T> _reference{};
+};
+
+template <typename TensorType, typename AccessorType, typename T, typename ReshapeLHSFunctionType, typename ReshapeRHSFunctionType, typename GEMMFunctionType>
+class GEMMMatrixMultiplyInterleavedTransposed3DValidationFixture : 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, float alpha, float beta, unsigned int v0, unsigned int h0, bool broadcast_bias,
+               bool fp16_mixed_precision, const ActivationLayerInfo &act_info, DataType data_type, GPUTarget gpu_arch)
+    {
+        GEMMLHSMatrixInfo lhs_info;
+        lhs_info.m0         = 4;
+        lhs_info.k0         = 4;
+        lhs_info.v0         = v0;
+        lhs_info.interleave = true;
+        lhs_info.transpose  = true;
+
+        GEMMRHSMatrixInfo rhs_info;
+        rhs_info.n0         = 16 / sizeof(T);
+        rhs_info.k0         = 1;
+        rhs_info.h0         = h0;
+        rhs_info.interleave = false;
+        rhs_info.transpose  = false;
+
+        // 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);
+        const TensorShape bias_shape(n, 1, 1);
+
+        _target    = compute_target(lhs_shape, rhs_shape, bias_shape, lhs_info, rhs_info, data_type, alpha, beta, m_h, fp16_mixed_precision, act_info, gpu_arch);
+        _reference = compute_reference(lhs_shape, rhs_shape, bias_shape, data_type, alpha, beta, m_h, act_info);
+    }
+
+protected:
+    template <typename U>
+    void fill(U &&tensor, int i)
+    {
+        std::uniform_real_distribution<> distribution(-1.0f, 1.0f);
+        library->fill(tensor, distribution, i);
+    }
+
+    TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info,
+                              DataType data_type, float alpha, float beta, unsigned int m_h, bool fp16_mixed_precision, const ActivationLayerInfo &act_info, GPUTarget gpu_arch)
+    {
+        // Create tensors
+        TensorType lhs  = create_tensor<TensorType>(lhs_shape, data_type, 1);
+        TensorType rhs  = create_tensor<TensorType>(rhs_shape, data_type, 1);
+        TensorType bias = create_tensor<TensorType>(bias_shape, data_type, 1);
+        TensorType lhs_reshaped;
+        TensorType rhs_reshaped;
+        TensorType dst;
+
+        const unsigned int m = lhs_shape[1];
+        const unsigned int n = rhs_shape[0];
+        const unsigned int k = lhs_shape[0];
+        GEMMReshapeInfo    reshape_info(m, n, k, rhs_info.h0, lhs_info.v0, m_h, false, true);
+
+        // The output tensor will be auto-initialized within the function
+
+        // Create and configure function
+        ReshapeLHSFunctionType reshape_lhs;
+        ReshapeRHSFunctionType reshape_rhs;
+        GEMMFunctionType       gemm;
+        reshape_lhs.configure(&lhs, &lhs_reshaped, lhs_info);
+        reshape_rhs.configure(&rhs, &rhs_reshaped, rhs_info);
+        gemm.configure(gpu_arch, &lhs_reshaped, &rhs_reshaped, &bias, &dst, alpha, beta, true, reshape_info, fp16_mixed_precision, act_info);
+
+        ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Allocate tensors
+        lhs.allocator()->allocate();
+        rhs.allocator()->allocate();
+        lhs_reshaped.allocator()->allocate();
+        rhs_reshaped.allocator()->allocate();
+        bias.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(!lhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!rhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!bias.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);
+        fill(AccessorType(bias), 2);
+
+        // Compute GEMM
+        reshape_lhs.run();
+        reshape_rhs.run();
+        gemm.run();
+
+        return dst;
+    }
+
+    SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, DataType data_type, float alpha, float beta, unsigned int m_h,
+                                      const ActivationLayerInfo &act_info)
+    {
+        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<T> lhs{ lhs_shape, data_type, 1 };
+        SimpleTensor<T> rhs{ rhs_shape, data_type, 1 };
+        SimpleTensor<T> bias{ dst_shape, data_type, 1 };
+
+        const int n          = rhs_shape[0];
+        const int m          = lhs_shape[1];
+        const int batch_size = lhs_shape[2];
+
+        // Fill reference
+        fill(lhs, 0);
+        fill(rhs, 1);
+        fill(bias, 2);
+
+        // In case of broadcast, we need simply copy the first into the following "M" ones
+        for(int i = 1; i < m * batch_size; i++)
+        {
+            memcpy(bias.data() + i * n, bias.data(), n * sizeof(T));
+        }
+
+        return reference::activation_layer(reference::gemm<T>(lhs, rhs, bias, alpha, beta), act_info);
+    }
+
+    TensorType      _target{};
+    SimpleTensor<T> _reference{};
+};
+
 template <typename TensorType, typename AccessorType, typename T, typename ReshapeLHSFunctionType, typename ReshapeRHSFunctionType, typename GEMMFunctionType>
 class GEMMMatrixMultiplyReshapedValidationFixture : public framework::Fixture
 {