Port NEGEMM to memory injecting interface (Part 3)

- Complete porting of NEGEMM to the new API

Resolves: COMPMID-4402

Change-Id: I14904102b25332dbb4fc048d45dca068a15b6eca
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5890
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

3 files changed, 515 insertions, 0 deletions

diff --git a/src/runtime/cpu/operators/CpuGemm.cpp b/src/runtime/cpu/operators/CpuGemm.cpp
new file mode 100644
index 0000000000..9a4d171ce6
--- /dev/null
+++ b/src/runtime/cpu/operators/CpuGemm.cpp
@@ -0,0 +1,366 @@
+/*
+ * Copyright (c) 2021 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 "src/runtime/cpu/operators/CpuGemm.h"
+
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "arm_compute/runtime/NEON/NEScheduler.h"
+#include "src/core/CPP/Validate.h"
+#include "src/core/helpers/AutoConfiguration.h"
+#include "src/core/helpers/MemoryHelpers.h"
+#include "src/runtime/cpu/utils/CpuAuxTensorHandler.h"
+
+using namespace arm_compute::experimental;
+using namespace arm_compute::misc::shape_calculator;
+
+namespace arm_compute
+{
+namespace cpu
+{
+namespace
+{
+cpu::AsmGemmInfo init_assembly_metadata(const GEMMInfo &info)
+{
+    cpu::AsmGemmInfo asm_info;
+    asm_info.method                  = cpu::AsmConvMethod::Im2Col;
+    asm_info.reinterpret_input_as_3d = info.reinterpret_input_as_3d();
+    asm_info.depth_output_gemm3d     = info.depth_output_gemm3d();
+    asm_info.activation_info         = info.activation_info();
+
+    return asm_info;
+}
+} // namespace
+
+void CpuGemm::configure(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, ITensorInfo *d, float alpha, float beta, const GEMMInfo &gemm_info)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(a, b, d);
+    ARM_COMPUTE_ERROR_THROW_ON(CpuGemm::validate(a, b, c, d, alpha, beta, gemm_info));
+
+    const cpu::AsmGemmInfo asm_info      = init_assembly_metadata(gemm_info);
+    const bool             is_c_bias     = gemm_info.reshape_b_only_on_first_run();
+    bool                   run_optimised = bool(cpu::CpuGemmAssemblyDispatch::validate(a, b, (is_c_bias) ? c : nullptr, d, asm_info));
+
+    // Check if we need to reshape the matrix B only on the first run
+    _is_prepared                      = false;
+    _reshape_b_only_on_first_run      = gemm_info.reshape_b_only_on_first_run();
+    _run_vector_matrix_multiplication = a->dimension(1) < 2;
+    _run_alpha_scale                  = alpha != 1.f;
+    _run_bias_addition                = c != nullptr && gemm_info.reshape_b_only_on_first_run();
+    _run_addition                     = beta != 0 && c != nullptr && !gemm_info.reshape_b_only_on_first_run();
+    _run_activation                   = gemm_info.activation_info().enabled() && (!run_optimised || (run_optimised
+                                                                                                     && !cpu::CpuGemmAssemblyDispatch::is_activation_supported(gemm_info.activation_info())));
+
+    if(run_optimised)
+    {
+        const ITensorInfo *c_to_use = is_c_bias ? c : nullptr;
+        _asm_glue                   = std::make_unique<cpu::CpuGemmAssemblyDispatch>();
+        _asm_glue->configure(a, b, c_to_use, d, asm_info);
+        ARM_COMPUTE_ERROR_ON(!_asm_glue->is_configured());
+
+        auto asm_mem_req           = _asm_glue->workspace();
+        _aux_mem[AsmGemmWorkspace] = asm_mem_req[AsmGemmWorkspace];
+        _aux_mem[Pretraspose]      = asm_mem_req[Pretraspose];
+
+        // Scale product by alpha
+        if(_run_alpha_scale)
+        {
+            _alpha_scale_func = std::make_unique<cpu::CpuActivation>();
+            _alpha_scale_func->configure(d, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, alpha, 0.f));
+        }
+    }
+    else
+    {
+        // Pick output tensor in case bias addition should be performed
+        ITensorInfo *gemm_output_to_use = (_run_bias_addition) ? &_tmp_d : d;
+
+        _mm_kernel = std::make_unique<cpu::kernels::CpuGemmMatrixMultiplyKernel>();
+
+        // Select between GEMV and GEMM
+        if(_run_vector_matrix_multiplication)
+        {
+            // Configure the matrix multiply kernel
+            _mm_kernel->configure(a, b, gemm_output_to_use, alpha, false);
+        }
+        else
+        {
+            const int m = a->dimension(1);
+            const int n = b->dimension(0);
+            const int k = a->dimension(0);
+
+            // Configure interleave kernel
+            _interleave_kernel = std::make_unique<cpu::kernels::CpuGemmInterleave4x4Kernel>();
+            _interleave_kernel->configure(a, &_tmp_a);
+            _aux_mem[InterleavedLHS] = MemoryInfo(offset_int_vec(InterleavedLHS), MemoryLifetime::Temporary, _tmp_a.total_size());
+
+            // Configure transpose kernel
+            _transpose_kernel = std::make_unique<cpu::kernels::CpuGemmTranspose1xWKernel>();
+            _transpose_kernel->configure(b, &_tmp_b);
+            _aux_mem[TransposedRHS] = MemoryInfo(offset_int_vec(TransposedRHS), MemoryLifetime::Persistent, _tmp_b.total_size());
+
+            // Configure matrix multiplication kernel
+            _mm_kernel->configure(&_tmp_a, &_tmp_b, gemm_output_to_use, alpha, true, GEMMReshapeInfo(m, n, k));
+        }
+
+        if(_run_bias_addition)
+        {
+            _add_bias = std::make_unique<cpu::CpuAdd>();
+            _add_bias->configure(gemm_output_to_use, c, d, ConvertPolicy::SATURATE);
+            _aux_mem[TempResult] = MemoryInfo(offset_int_vec(TempResult), MemoryLifetime::Persistent, _tmp_d.total_size());
+        }
+    }
+
+    // Configure matrix addition kernel
+    if(_run_addition)
+    {
+        _ma_kernel = std::make_unique<cpu::kernels::CpuGemmMatrixAdditionKernel>();
+        _ma_kernel->configure(c, d, beta);
+    }
+
+    // Configure activation
+    if(_run_activation)
+    {
+        _activation_func = std::make_unique<cpu::CpuActivation>();
+        _activation_func->configure(d, nullptr, gemm_info.activation_info());
+    }
+}
+
+Status CpuGemm::validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *d, float alpha, float beta, const GEMMInfo &gemm_info)
+{
+    ARM_COMPUTE_UNUSED(alpha);
+    const bool is_c_bias = gemm_info.reshape_b_only_on_first_run();
+
+    ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(a);
+    ARM_COMPUTE_RETURN_ERROR_ON_CPU_BF16_UNSUPPORTED(a);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::BFLOAT16, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(a, b);
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(a->dimension(0) != b->dimension(1), "The product AB is defined only if the number of columns in A is equal to the number of rows in B");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.is_a_reshaped(), "Matrix A already reshaped is not supported");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.is_b_reshaped(), "Matrix B already reshaped is not supported");
+    if(a->data_type() != DataType::BFLOAT16)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(a, d);
+    }
+
+    if(c != nullptr && !is_c_bias)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON(gemm_info.depth_output_gemm3d() != 0);
+        ARM_COMPUTE_RETURN_ERROR_ON(gemm_info.reinterpret_input_as_3d());
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(c, d);
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(a->dimension(1) != c->dimension(1), "The C matrix must have the same number of rows as the matrix A");
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(b->dimension(0) != c->dimension(0), "The C matrix must have the same number of columns as the matrix B");
+    }
+
+    if(d->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON(b->dimension(0) != d->dimension(0));
+        if(gemm_info.depth_output_gemm3d() != 0)
+        {
+            if(gemm_info.reinterpret_input_as_3d())
+            {
+                ARM_COMPUTE_RETURN_ERROR_ON(a->dimension(1) != d->dimension(1));
+                ARM_COMPUTE_RETURN_ERROR_ON(a->dimension(2) != d->dimension(2));
+            }
+            else
+            {
+                ARM_COMPUTE_RETURN_ERROR_ON(a->dimension(1) != d->dimension(1) * d->dimension(2));
+            }
+        }
+        else
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON(a->dimension(1) != d->dimension(1));
+        }
+    }
+
+    // Check if we need to run the optimized assembly kernel
+    cpu::AsmGemmInfo asm_info      = init_assembly_metadata(gemm_info);
+    const bool       run_optimised = bool(cpu::CpuGemmAssemblyDispatch::validate(a, b, is_c_bias ? c : nullptr, d, asm_info));
+
+    if(!run_optimised)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.reinterpret_input_as_3d(), "CpuGemm cannot reinterpret the input tensor as 3D");
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.depth_output_gemm3d() != 0, "CpuGemm cannot reinterpret the output tensor as 3D");
+
+        // Check if the first input tensor is a vector.
+        const bool run_vector_matrix_multiplication = a->dimension(1) < 2;
+        // Check if we need to reshape the matrix A and matrix B
+        const bool run_interleave_transpose = !run_vector_matrix_multiplication && !(gemm_info.reshape_b_only_on_first_run());
+
+        // Arguments used by GEMMReshapeInfo
+        // If we pass the matrix A and matrix B reshaped to CpuGemmMatrixMultiplyKernel, we need to pass m, n, k, mult_transpose1xW_width and mult_interleave4x4_height to GEMMReshapeInfo
+        // in order to know how the matrices have been reshaped
+        const int m                         = a->dimension(1);
+        const int n                         = b->dimension(0);
+        const int k                         = a->dimension(0);
+        int       mult_transpose1xW_width   = 1;
+        int       mult_interleave4x4_height = 1;
+
+        const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(m, n, k, mult_transpose1xW_width, mult_interleave4x4_height, gemm_info.depth_output_gemm3d());
+
+        const ITensorInfo *matrix_a_info = a;
+        const ITensorInfo *matrix_b_info = b;
+
+        TensorInfo tmp_a_info{};
+        TensorInfo tmp_b_info{};
+        TensorInfo tmp_output_info = *d->clone();
+
+        if(run_interleave_transpose)
+        {
+            matrix_a_info = &tmp_a_info;
+            matrix_b_info = &tmp_b_info;
+
+            // Validate interleave kernel
+            auto_init_if_empty(tmp_a_info, a->clone()->set_tensor_shape(compute_interleaved_shape(*a, mult_interleave4x4_height, gemm_info.reinterpret_input_as_3d())));
+            ARM_COMPUTE_RETURN_ON_ERROR(cpu::kernels::CpuGemmInterleave4x4Kernel::validate(a, &tmp_a_info));
+
+            // Validate transpose kernel
+            auto_init_if_empty(tmp_b_info, b->clone()->set_tensor_shape(compute_transpose1xW_with_element_size_shape(*b, mult_transpose1xW_width)));
+            ARM_COMPUTE_RETURN_ON_ERROR(cpu::kernels::CpuGemmTranspose1xWKernel::validate(b, &tmp_b_info));
+        }
+
+        // Validate matrix multiply
+        auto_init_if_empty(tmp_output_info, matrix_a_info->clone()->set_tensor_shape(compute_mm_shape(*matrix_a_info, *matrix_b_info, run_interleave_transpose, reshape_info)));
+        ARM_COMPUTE_RETURN_ON_ERROR(cpu::kernels::CpuGemmMatrixMultiplyKernel::validate(matrix_a_info, matrix_b_info, &tmp_output_info, alpha, run_interleave_transpose, reshape_info));
+
+        if(c != nullptr && gemm_info.reshape_b_only_on_first_run())
+        {
+            ARM_COMPUTE_RETURN_ON_ERROR(cpu::CpuAdd::validate(&tmp_output_info, c, d, ConvertPolicy::SATURATE));
+        }
+    }
+
+    // Validate matrix addition kernel
+    if(beta != 0 && c != nullptr && !is_c_bias)
+    {
+        ARM_COMPUTE_RETURN_ON_ERROR(cpu::kernels::CpuGemmMatrixAdditionKernel::validate(c, d, beta));
+    }
+
+    // Validate activation
+    const ActivationLayerInfo &activation = gemm_info.activation_info();
+    if(activation.enabled())
+    {
+        ARM_COMPUTE_RETURN_ON_ERROR(cpu::CpuActivation::validate(d, nullptr, activation));
+    }
+
+    return Status{};
+}
+
+void CpuGemm::run(ITensorPack &tensors)
+{
+    prepare(tensors);
+
+    auto a = tensors.get_const_tensor(ACL_SRC_0);
+    auto b = tensors.get_const_tensor(ACL_SRC_1);
+    auto c = tensors.get_const_tensor(ACL_SRC_2);
+    auto d = tensors.get_tensor(ACL_DST);
+
+    if(_asm_glue->is_configured())
+    {
+        // Pass c to asm dispatch only if it's the bias tensor
+        ITensorPack asm_pack = tensors;
+        asm_pack.add_const_tensor(ACL_SRC_2, (_reshape_b_only_on_first_run) ? c : nullptr);
+        _asm_glue->run(asm_pack);
+        if(_run_alpha_scale)
+        {
+            ITensorPack pack{ { ACL_SRC, d }, { ACL_DST, d } };
+            _alpha_scale_func->run(pack);
+        }
+    }
+    else
+    {
+        CpuAuxTensorHandler interleaved_a(offset_int_vec(InterleavedLHS), _tmp_a, tensors, true);
+        CpuAuxTensorHandler transposed_b(offset_int_vec(TransposedRHS), _tmp_b, tensors, true);
+        CpuAuxTensorHandler temp_d(offset_int_vec(TempResult), _tmp_d, tensors, true);
+
+        ITensorPack mm_pack{ { ACL_SRC_0, a }, { ACL_SRC_1, b }, { ACL_DST, (_run_bias_addition) ? temp_d.get() : d } };
+        if(!_run_vector_matrix_multiplication)
+        {
+            // Run interleave kernel
+            ITensorPack interleave_pack{ { ACL_SRC, a }, { ACL_DST, interleaved_a.get() } };
+            NEScheduler::get().schedule_op(_interleave_kernel.get(), Window::DimY, _interleave_kernel->window(), interleave_pack);
+
+            if(!_reshape_b_only_on_first_run)
+            {
+                // Run transpose kernel
+                ITensorPack transpose_pack{ { ACL_SRC, b }, { ACL_DST, transposed_b.get() } };
+                NEScheduler::get().schedule_op(_transpose_kernel.get(), Window::DimY, _transpose_kernel->window(), transpose_pack);
+            }
+
+            // Use reshaped matrices
+            mm_pack.add_const_tensor(ACL_SRC_0, interleaved_a.get());
+            mm_pack.add_const_tensor(ACL_SRC_1, transposed_b.get());
+        }
+
+        NEScheduler::get().schedule_op(_mm_kernel.get(), _run_vector_matrix_multiplication ? Window::DimX : Window::DimY, _mm_kernel->window(), mm_pack);
+
+        // Run bias addition kernel
+        if(_run_bias_addition)
+        {
+            ITensorPack pack{ { ACL_SRC_0, temp_d.get() }, { ACL_SRC_1, c }, { ACL_DST, d } };
+            _add_bias->run(pack);
+        }
+    }
+
+    // Run matrix addition kernel
+    if(_run_addition)
+    {
+        ITensorPack c_add_pack{ { ACL_SRC, c }, { ACL_DST, d } };
+        NEScheduler::get().schedule_op(_ma_kernel.get(), Window::DimY, _ma_kernel->window(), c_add_pack);
+    }
+
+    // Run activation function
+    if(_run_activation)
+    {
+        ITensorPack pack{ { ACL_SRC, d }, { ACL_DST, d } };
+        _activation_func->run(pack);
+    }
+}
+
+void CpuGemm::prepare(ITensorPack &tensors)
+{
+    if(!_is_prepared)
+    {
+        if(_asm_glue->is_configured())
+        {
+            _asm_glue->prepare(tensors);
+        }
+        else if(_reshape_b_only_on_first_run && !_run_vector_matrix_multiplication)
+        {
+            const ITensor *b     = tensors.get_const_tensor(ACL_SRC_1);
+            ITensor       *b_aux = utils::cast::polymorphic_cast<ITensor *>(tensors.get_tensor(offset_int_vec(TransposedRHS)));
+            ARM_COMPUTE_ERROR_ON_NULLPTR(b, b_aux);
+
+            CpuAuxTensorHandler transposed_b(_tmp_b, *b_aux);
+            ITensorPack         transpose_pack{ { ACL_SRC, b }, { ACL_DST, transposed_b.get() } };
+            NEScheduler::get().schedule_op(_transpose_kernel.get(), Window::DimY, _transpose_kernel->window(), transpose_pack);
+        }
+        _is_prepared = true;
+    }
+}
+
+experimental::MemoryRequirements CpuGemm::workspace() const
+{
+    return _aux_mem;
+}
+} // namespace cpu
+} // namespace arm_compute
diff --git a/src/runtime/cpu/operators/CpuGemm.h b/src/runtime/cpu/operators/CpuGemm.h
new file mode 100644
index 0000000000..8d859791f5
--- /dev/null
+++ b/src/runtime/cpu/operators/CpuGemm.h
@@ -0,0 +1,145 @@
+/*
+ * Copyright (c) 2021 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.
+ */
+#ifndef ARM_COMPUTE_CPU_GEMM_H
+#define ARM_COMPUTE_CPU_GEMM_H
+
+#include "src/runtime/cpu/ICpuOperator.h"
+
+#include "arm_compute/core/ITensorPack.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Types.h"
+#include "src/core/cpu/kernels/CpuGemmInterleave4x4Kernel.h"
+#include "src/core/cpu/kernels/CpuGemmMatrixAdditionKernel.h"
+#include "src/core/cpu/kernels/CpuGemmMatrixMultiplyKernel.h"
+#include "src/core/cpu/kernels/CpuGemmTranspose1xWKernel.h"
+#include "src/runtime/cpu/operators/CpuActivation.h"
+#include "src/runtime/cpu/operators/CpuAdd.h"
+#include "src/runtime/cpu/operators/internal/CpuGemmAssemblyDispatch.h"
+
+#include <memory>
+
+namespace arm_compute
+{
+namespace cpu
+{
+/** Basic function to execute GEMM. This function calls the following kernels:
+ *
+ * If optimized assembly is available:
+ *  -# @ref cpu::CpuGemmAssemblyDispatch
+ *  -# @ref cpu::CpuActivation (if alpha != 1.0)
+ * Else:
+ *  -# @ref cpu::kernels::CpuGemmInterleave4x4Kernel (if the output tensor is a matrix)
+ *  -# @ref cpu::kernels::CpuGemmTranspose1xWKernel (if the output tensor is a matrix)
+ *  -# @ref cpu::kernels::CpuGemmMatrixMultiplyKernel
+ * In both cases:
+ *  -# @ref cpu::kernels::CpuGemmMatrixAdditionKernel (if c != nullptr and beta != 0.0 and is not reshaped once)
+ * Else:
+ *  -# @ref cpu::CpuAdd (if c != nullptr and is reshaped once and not optimized assembly in place)
+ *
+ *  -# @ref cpu::CpuActivation (if activation is specified in GEMMInfo)
+ */
+class CpuGemm : public ICpuOperator
+{
+public:
+    /** Default constructor */
+    CpuGemm() = default;
+    /** Default destructor */
+    ~CpuGemm() = default;
+    /** Configure operator for a given list of arguments
+     *
+     * Valid data layouts:
+     * - All
+     *
+     * Valid data type configurations:
+     * |src0         |src1        |src2      |dst            |
+     * |:------------|:-----------|:---------|:--------------|
+     * |F32          |F32         |F32       |F32            |
+     * |F16          |F16         |F16       |F16            |
+     * |BFLOAT16     |BFLOAT16    |BFLOAT16  |BFLOAT16       |
+     *
+     * @note GEMM: General Matrix Multiply - [alpha * A * B + beta * C].
+     * @note GEMM: The tensors a, b, c, d must have the same data type. You should not mix data types when calling this function.
+     *
+     * @param[in]  a         First input tensor info (Matrix A or Vector A). Data type supported: BFLOAT16/F16/F32
+     * @param[in]  b         Second input tensor info (Matrix B). Data type supported: same as @p a
+     * @param[in]  c         Third input tensor info (Matrix C). It can be a nullptr if just the multiplication between @p a and @p b is needed. Data type supported: same as @p a
+     * @param[out] d         Output tensor info. Data type supported: same as @p a
+     * @param[in]  alpha     Weight of the matrix product
+     * @param[in]  beta      Weight of matrix C
+     * @param[in]  gemm_info (Optional) Specifies if the matrix A and/or matrix B have been reshaped and
+     *                       if the reshape of matrix B should happen only for the first run
+     */
+    void configure(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, ITensorInfo *d,
+                   float alpha, float beta, const GEMMInfo &gemm_info = GEMMInfo());
+    /** Static function to check if given info will lead to a valid configuration of @ref CpuGemm.
+     *
+     * Similar to @ref CpuGemm::configure()
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *d,
+                           float alpha, float beta, const GEMMInfo &gemm_info = GEMMInfo());
+
+    // Inherited methods overridden:
+    void run(ITensorPack &tensors) override;
+    void prepare(ITensorPack &constants) override;
+    experimental::MemoryRequirements workspace() const override;
+
+private:
+    enum AuxTensorIdx
+    {
+        AsmGemmWorkspace = 0,
+        Pretraspose,
+        InterleavedLHS,
+        TransposedRHS,
+        TempResult,
+        Count
+    };
+
+    std::unique_ptr<kernels::CpuGemmInterleave4x4Kernel>  _interleave_kernel{ nullptr };
+    std::unique_ptr<kernels::CpuGemmTranspose1xWKernel>   _transpose_kernel{ nullptr };
+    std::unique_ptr<kernels::CpuGemmMatrixMultiplyKernel> _mm_kernel{ nullptr };
+    std::unique_ptr<CpuGemmAssemblyDispatch>              _asm_glue{ nullptr };
+    std::unique_ptr<kernels::CpuGemmMatrixAdditionKernel> _ma_kernel{ nullptr };
+    std::unique_ptr<CpuActivation>                        _alpha_scale_func{ nullptr };
+    std::unique_ptr<CpuAdd>                               _add_bias{ nullptr };
+    std::unique_ptr<CpuActivation>                        _activation_func{ nullptr };
+
+    TensorInfo _tmp_a{};
+    TensorInfo _tmp_b{};
+    TensorInfo _tmp_d{};
+
+    bool _run_vector_matrix_multiplication{ false };
+    bool _run_alpha_scale{ false };
+    bool _run_addition{ false };
+    bool _run_bias_addition{ false };
+    bool _run_activation{ false };
+    bool _reshape_b_only_on_first_run{ false };
+    bool _is_prepared{ false };
+
+    experimental::MemoryRequirements _aux_mem{ Count };
+};
+} // namespace cpu
+} // namespace arm_compute
+#endif /*ARM_COMPUTE_CPU_GEMM_H */
diff --git a/src/runtime/cpu/utils/CpuAuxTensorHandler.h b/src/runtime/cpu/utils/CpuAuxTensorHandler.h
index 644018a718..0d1c927b5a 100644
--- a/src/runtime/cpu/utils/CpuAuxTensorHandler.h
+++ b/src/runtime/cpu/utils/CpuAuxTensorHandler.h
@@ -41,6 +41,10 @@ public:
     CpuAuxTensorHandler(int slot_id, TensorInfo &info, ITensorPack &pack, bool pack_inject = false)
         : _tensor()
     {
+        if(info.total_size() == 0)
+        {
+            return;
+        }
         _tensor.allocator()->soft_init(info);
 
         ITensor *packed_tensor = utils::cast::polymorphic_downcast<ITensor *>(pack.get_tensor(slot_id));