Make Gpu Elementwise Operations kernels and functions state-less

Partially resolves COMPMID-4004

Change-Id: Ie99b6c676af5be4c2a67affc06fec378f3dff8ec
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4918
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

5 files changed, 317 insertions, 324 deletions

diff --git a/Android.bp b/Android.bp
index 7cfa817193..a1898b5432 100644
--- a/Android.bp
+++ b/Android.bp
@@ -797,6 +797,7 @@ cc_library_static {
         "src/runtime/gpu/cl/operators/ClActivation.cpp",
         "src/runtime/gpu/cl/operators/ClAdd.cpp",
         "src/runtime/gpu/cl/operators/ClConcatenate.cpp",
+        "src/runtime/gpu/cl/operators/ClElementwiseOperations.cpp",
         "src/runtime/gpu/cl/operators/ClFloor.cpp",
         "src/runtime/gpu/cl/operators/ClSub.cpp",
         "utils/CommonGraphOptions.cpp",
diff --git a/arm_compute/runtime/CL/functions/CLElementwiseOperations.h b/arm_compute/runtime/CL/functions/CLElementwiseOperations.h
index c8c7e0c587..2b291517f3 100644
--- a/arm_compute/runtime/CL/functions/CLElementwiseOperations.h
+++ b/arm_compute/runtime/CL/functions/CLElementwiseOperations.h
@@ -33,189 +33,6 @@ class ICLTensor;
 class CLCompileContext;
 class ITensorInfo;
 
-namespace experimental
-{
-/** Basic function to run @ref arm_compute::opencl::kernels::ClSaturatedArithmeticKernel for division
- *
- * @note The tensor data type for the inputs must be F16/F32.
- * @note The function performs an arithmetic division between two tensors.
- */
-class CLArithmeticDivision : public ICLOperator
-{
-public:
-    /** Default Constructor */
-    CLArithmeticDivision();
-    /** Initialise the kernel's inputs, output.
-     *
-     * @param[in]      compile_context The compile context to be used.
-     * @param[in, out] input1          First tensor input. Data types supported: F16/F32.
-     *                                 The input tensor is [in, out] because its TensorInfo might be modified inside the kernel in case of broadcasting of dimension 0.
-     * @param[in, out] input2          Second tensor input. Same as @p input1.
-     *                                 The input tensor is [in, out] because its TensorInfo might be modified inside the kernel in case of broadcasting of dimension 0.
-     * @param[out]     output          Output tensor. Data types supported: Same as @p input1.
-     * @param[in]      act_info        (Optional) Activation layer information in case of a fused activation.
-     */
-    void configure(const CLCompileContext &compile_context, ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, const ActivationLayerInfo &act_info = ActivationLayerInfo());
-    /** Static function to check if given info will lead to a valid configuration of @ref CLArithmeticDivision
-     *
-     * @param[in] input1   First tensor input info. Data types supported: F16/F32.
-     * @param[in] input2   Second tensor input info. Data types supported: Same as @p input1.
-     * @param[in] output   Output tensor info. Data types supported: Same as @p input1.
-     * @param[in] act_info (Optional) Activation layer information in case of a fused activation.
-     *
-     * @return a status
-     */
-    static Status validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, const ActivationLayerInfo &act_info = ActivationLayerInfo());
-
-    // Inherited methods overridden:
-    void run(ITensorPack &tensors) override;
-};
-
-/** Basic function to run @ref arm_compute::opencl::kernels::ClArithmeticKernel for max
- *
- * @note The tensor data type for the inputs must be U8/QASYMM8/S16/QSYMM16/S32/U32/F16/F32.
- * @note The function performs a max operation between two tensors.
- */
-class CLElementwiseMax : public ICLOperator
-{
-public:
-    /** Default Constructor */
-    CLElementwiseMax();
-    /** Initialise the kernel's inputs, output and conversion policy.
-     *
-     * @param[in]      compile_context The compile context to be used.
-     * @param[in, out] input1          First tensor input. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/S32/U32/F16/F32.
-     *                                 The input tensor is [in, out] because its TensorInfo might be modified inside the kernel in case of broadcasting of dimension 0.
-     * @param[in, out] input2          Second tensor input. Data types supported: same as @p input1.
-     *                                 The input tensor is [in, out] because its TensorInfo might be modified inside the kernel in case of broadcasting of dimension 0.
-     * @param[out]     output          Output tensor. Data types supported: same as @p input1.
-     * @param[in]      act_info        (Optional) Activation layer information in case of a fused activation.
-     */
-    void configure(const CLCompileContext &compile_context, ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, const ActivationLayerInfo &act_info = ActivationLayerInfo());
-    /** Static function to check if given info will lead to a valid configuration of @ref arm_compute::opencl::kernels::ClArithmeticKernel for max
-     *
-     * @param[in] input1   First tensor input info. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/S32/U32/F16/F32.
-     * @param[in] input2   Second tensor input info. Data types supported: same as @p input1.
-     * @param[in] output   Output tensor info. Data types supported: same as @p input1.
-     * @param[in] act_info (Optional) Activation layer information in case of a fused activation.
-     *
-     * @return a status
-     */
-    static Status validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, const ActivationLayerInfo &act_info = ActivationLayerInfo());
-
-    // Inherited methods overridden:
-    void run(ITensorPack &tensors) override;
-};
-
-/** Basic function to run @ref arm_compute::opencl::kernels::ClArithmeticKernel for min
- *
- * @note The tensor data type for the inputs must be U8/QASYMM8/S16/QSYMM16/S32/U32/F16/F32.
- * @note The function performs a max operation between two tensors.
- */
-class CLElementwiseMin : public ICLOperator
-{
-public:
-    /** Default Constructor */
-    CLElementwiseMin();
-    /** Initialise the kernel's inputs, output and conversion policy.
-     *
-     * @param[in]      compile_context The compile context to be used.
-     * @param[in, out] input1          First tensor input. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/S32/U32/F16/F32.
-     *                                 The input tensor is [in, out] because its TensorInfo might be modified inside the kernel in case of broadcasting of dimension 0.
-     * @param[in, out] input2          Second tensor input. Data types supported: same as @p input1.
-     *                                 The input tensor is [in, out] because its TensorInfo might be modified inside the kernel in case of broadcasting of dimension 0.
-     * @param[out]     output          Output tensor. Data types supported: same as @p input1.
-     * @param[in]      act_info        (Optional) Activation layer information in case of a fused activation.
-     */
-    void configure(const CLCompileContext &compile_context, ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, const ActivationLayerInfo &act_info = ActivationLayerInfo());
-    /** Static function to check if given info will lead to a valid configuration of @ref arm_compute::opencl::kernels::ClArithmeticKernel for min
-     *
-     * @param[in] input1   First tensor input info. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/S32/U32/F16/F32.
-     * @param[in] input2   Second tensor input info. Data types supported: same as @p input1.
-     * @param[in] output   Output tensor info. Data types supported: same as @p input1.
-     * @param[in] act_info (Optional) Activation layer information in case of a fused activation.
-     *
-     * @return a status
-     */
-    static Status validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, const ActivationLayerInfo &act_info = ActivationLayerInfo());
-
-    // Inherited methods overridden:
-    void run(ITensorPack &tensors) override;
-};
-
-/** Basic function to run @ref arm_compute::opencl::kernels::ClArithmeticKernel for squared difference
- *
- * @note The tensor data type for the inputs must be QASYMM8/U8/S16/QSYMM16/F16/F32.
- * @note The function performs a squared different operation between two tensors (i.e., out[i] = (in1[i] - in2[i])^2
- */
-class CLElementwiseSquaredDiff : public ICLOperator
-{
-public:
-    /** Default Constructor */
-    CLElementwiseSquaredDiff();
-    /** Initialise the kernel's inputs, output and conversion policy.
-     *
-     * @param[in]      compile_context The compile context to be used.
-     * @param[in, out] input1          First tensor input. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/F32.
-     *                                 The input tensor is [in, out] because its TensorInfo might be modified inside the kernel in case of broadcasting of dimension 0.
-     * @param[in, out] input2          Second tensor input. Data types supported: same as @p input1.
-     *                                 The input tensor is [in, out] because its TensorInfo might be modified inside the kernel in case of broadcasting of dimension 0.
-     * @param[out]     output          Output tensor. Data types supported: same as @p input1.
-     * @param[in]      act_info        (Optional) Activation layer information in case of a fused activation.
-     */
-    void configure(const CLCompileContext &compile_context, ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, const ActivationLayerInfo &act_info = ActivationLayerInfo());
-    /** Static function to check if given info will lead to a valid configuration of @ref arm_compute::opencl::kernels::ClArithmeticKernel for squared difference
-     *
-     * @param[in] input1   First tensor input info. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/F32.
-     * @param[in] input2   Second tensor input info. Data types supported: same as @p input1.
-     * @param[in] output   Output tensor info. Data types supported: same as @p input1.
-     * @param[in] act_info (Optional) Activation layer information in case of a fused activation.
-     *
-     * @return a status
-     */
-    static Status validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, const ActivationLayerInfo &act_info = ActivationLayerInfo());
-
-    // Inherited methods overridden:
-    void run(ITensorPack &tensors) override;
-};
-
-/** Basic function to run @ref arm_compute::opencl::kernels::ClArithmeticKernel for power
- *
- * @note The tensor data type for the inputs must be F16/F32.
- * @note The function performs an elementwise power of in1 to in2 (i.e., out[i] = in1[i] ^ in2[i])
- */
-class CLElementwisePower : public ICLOperator
-{
-public:
-    /** Default Constructor */
-    CLElementwisePower();
-    /** Initialise the kernel's inputs, output and conversion policy.
-     *
-     * @param[in]      compile_context The compile context to be used.
-     * @param[in, out] input1          First tensor input. Data types supported: F16/F32.
-     *                                 The input tensor is [in, out] because its TensorInfo might be modified inside the kernel in case of broadcasting of dimension 0.
-     * @param[in, out] input2          Second tensor input. Data types supported: F16/F32.
-     *                                 The input tensor is [in, out] because its TensorInfo might be modified inside the kernel in case of broadcasting of dimension 0.
-     * @param[out]     output          Output tensor. Data types supported:F16/F32.
-     * @param[in]      act_info        (Optional) Activation layer information in case of a fused activation.
-     */
-    void configure(const CLCompileContext &compile_context, ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, const ActivationLayerInfo &act_info = ActivationLayerInfo());
-    /** Static function to check if given info will lead to a valid configuration of @ref arm_compute::opencl::kernels::ClArithmeticKernel for power
-     *
-     * @param[in] input1   First tensor input info. Data types supported: F16/F32.
-     * @param[in] input2   Second tensor input info. Data types supported: F16/F32.
-     * @param[in] output   Output tensor info. Data types supported: F16/F32.
-     * @param[in] act_info (Optional) Activation layer information in case of a fused activation.
-     *
-     * @return a status
-     */
-    static Status validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, const ActivationLayerInfo &act_info = ActivationLayerInfo());
-
-    // Inherited methods overridden:
-    void run(ITensorPack &tensors) override;
-};
-} // namespace experimental
-
 /** Basic function to run @ref opencl::kernels::ClSaturatedArithmeticKernel for addition
  *
  * @note The tensor data type for the inputs must be U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/S32/F16/F32.
diff --git a/src/runtime/CL/functions/CLElementwiseOperations.cpp b/src/runtime/CL/functions/CLElementwiseOperations.cpp
index 9b809eebc7..60c699cbb8 100644
--- a/src/runtime/CL/functions/CLElementwiseOperations.cpp
+++ b/src/runtime/CL/functions/CLElementwiseOperations.cpp
@@ -23,124 +23,17 @@
  */
 #include "arm_compute/runtime/CL/functions/CLElementwiseOperations.h"
 
+#include "arm_compute/core/CL/CLKernelLibrary.h"
 #include "arm_compute/core/CL/ICLTensor.h"
-#include "arm_compute/runtime/CL/CLScheduler.h"
-#include "src/core/gpu/cl/kernels/ClElementwiseKernel.h"
+#include "arm_compute/core/Types.h"
+#include "src/core/CL/ICLKernel.h"
 
 #include "src/runtime/gpu/cl/operators/ClAdd.h"
+#include "src/runtime/gpu/cl/operators/ClElementwiseOperations.h"
 #include "src/runtime/gpu/cl/operators/ClSub.h"
 
-#include <utility>
-
 namespace arm_compute
 {
-namespace experimental
-{
-CLArithmeticDivision::CLArithmeticDivision()
-{
-}
-
-void CLArithmeticDivision::configure(const CLCompileContext &compile_context, ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, const ActivationLayerInfo &act_info)
-{
-    auto k = std::make_unique<arm_compute::opencl::kernels::ClArithmeticKernel>();
-    k->configure(compile_context, ArithmeticOperation::DIV, input1, input2, output, act_info);
-    _kernel = std::move(k);
-}
-
-Status CLArithmeticDivision::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, const ActivationLayerInfo &act_info)
-{
-    return arm_compute::opencl::kernels::ClArithmeticKernel::validate(ArithmeticOperation::DIV, input1, input2, output, act_info);
-}
-
-void CLArithmeticDivision::run(ITensorPack &tensors)
-{
-    ICLOperator::run(tensors);
-}
-
-CLElementwiseMax::CLElementwiseMax()
-{
-}
-
-void CLElementwiseMax::configure(const CLCompileContext &compile_context, ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, const ActivationLayerInfo &act_info)
-{
-    auto k = std::make_unique<arm_compute::opencl::kernels::ClArithmeticKernel>();
-    k->configure(compile_context, ArithmeticOperation::MAX, input1, input2, output, act_info);
-    _kernel = std::move(k);
-}
-
-Status CLElementwiseMax::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, const ActivationLayerInfo &act_info)
-{
-    return arm_compute::opencl::kernels::ClArithmeticKernel::validate(ArithmeticOperation::MAX, input1, input2, output, act_info);
-}
-
-void CLElementwiseMax::run(ITensorPack &tensors)
-{
-    ICLOperator::run(tensors);
-}
-
-CLElementwiseMin::CLElementwiseMin()
-{
-}
-
-void CLElementwiseMin::configure(const CLCompileContext &compile_context, ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, const ActivationLayerInfo &act_info)
-{
-    auto k = std::make_unique<arm_compute::opencl::kernels::ClArithmeticKernel>();
-    k->configure(compile_context, ArithmeticOperation::MIN, input1, input2, output, act_info);
-    _kernel = std::move(k);
-}
-
-Status CLElementwiseMin::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, const ActivationLayerInfo &act_info)
-{
-    return arm_compute::opencl::kernels::ClArithmeticKernel::validate(ArithmeticOperation::MIN, input1, input2, output, act_info);
-}
-
-void CLElementwiseMin::run(ITensorPack &tensors)
-{
-    ICLOperator::run(tensors);
-}
-
-CLElementwiseSquaredDiff::CLElementwiseSquaredDiff()
-{
-}
-
-void CLElementwiseSquaredDiff::configure(const CLCompileContext &compile_context, ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, const ActivationLayerInfo &act_info)
-{
-    auto k = std::make_unique<arm_compute::opencl::kernels::ClArithmeticKernel>();
-    k->configure(compile_context, ArithmeticOperation::SQUARED_DIFF, input1, input2, output, act_info);
-    _kernel = std::move(k);
-}
-
-Status CLElementwiseSquaredDiff::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, const ActivationLayerInfo &act_info)
-{
-    return arm_compute::opencl::kernels::ClArithmeticKernel::validate(ArithmeticOperation::SQUARED_DIFF, input1, input2, output, act_info);
-}
-
-void CLElementwiseSquaredDiff::run(ITensorPack &tensors)
-{
-    ICLOperator::run(tensors);
-}
-
-CLElementwisePower::CLElementwisePower()
-{
-}
-
-void CLElementwisePower::configure(const CLCompileContext &compile_context, ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, const ActivationLayerInfo &act_info)
-{
-    auto k = std::make_unique<arm_compute::opencl::kernels::ClArithmeticKernel>();
-    k->configure(compile_context, ArithmeticOperation::POWER, input1, input2, output, act_info);
-    _kernel = std::move(k);
-}
-
-Status CLElementwisePower::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, const ActivationLayerInfo &act_info)
-{
-    return arm_compute::opencl::kernels::ClArithmeticKernel::validate(ArithmeticOperation::POWER, input1, input2, output, act_info);
-}
-
-void CLElementwisePower::run(ITensorPack &tensors)
-{
-    ICLOperator::run(tensors);
-}
-} // namespace experimental
 struct CLArithmeticAddition::Impl
 {
     const ICLTensor               *src_0{ nullptr };
@@ -235,10 +128,10 @@ void CLArithmeticSubtraction::run()
 
 struct CLArithmeticDivision::Impl
 {
-    const ICLTensor                                    *src_0{ nullptr };
-    const ICLTensor                                    *src_1{ nullptr };
-    ICLTensor                                          *dst{ nullptr };
-    std::unique_ptr<experimental::CLArithmeticDivision> op{ nullptr };
+    const ICLTensor                               *src_0{ nullptr };
+    const ICLTensor                               *src_1{ nullptr };
+    ICLTensor                                     *dst{ nullptr };
+    std::unique_ptr<opencl::ClElementwiseDivision> op{ nullptr };
 };
 
 CLArithmeticDivision::CLArithmeticDivision()
@@ -259,13 +152,13 @@ void CLArithmeticDivision::configure(const CLCompileContext &compile_context, co
     _impl->src_0 = input1;
     _impl->src_1 = input2;
     _impl->dst   = output;
-    _impl->op    = std::make_unique<experimental::CLArithmeticDivision>();
+    _impl->op    = std::make_unique<opencl::ClElementwiseDivision>();
     _impl->op->configure(compile_context, input1->info(), input2->info(), output->info(), act_info);
 }
 
 Status CLArithmeticDivision::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, const ActivationLayerInfo &act_info)
 {
-    return experimental::CLArithmeticDivision::validate(input1, input2, output, act_info);
+    return opencl::ClElementwiseDivision::validate(input1, input2, output, act_info);
 }
 
 void CLArithmeticDivision::run()
@@ -280,10 +173,10 @@ void CLArithmeticDivision::run()
 
 struct CLElementwiseMax::Impl
 {
-    const ICLTensor                                *src_0{ nullptr };
-    const ICLTensor                                *src_1{ nullptr };
-    ICLTensor                                      *dst{ nullptr };
-    std::unique_ptr<experimental::CLElementwiseMax> op{ nullptr };
+    const ICLTensor                          *src_0{ nullptr };
+    const ICLTensor                          *src_1{ nullptr };
+    ICLTensor                                *dst{ nullptr };
+    std::unique_ptr<opencl::ClElementwiseMax> op{ nullptr };
 };
 
 CLElementwiseMax::CLElementwiseMax()
@@ -304,13 +197,13 @@ void CLElementwiseMax::configure(const CLCompileContext &compile_context, ICLTen
     _impl->src_0 = input1;
     _impl->src_1 = input2;
     _impl->dst   = output;
-    _impl->op    = std::make_unique<experimental::CLElementwiseMax>();
+    _impl->op    = std::make_unique<opencl::ClElementwiseMax>();
     _impl->op->configure(compile_context, input1->info(), input2->info(), output->info(), act_info);
 }
 
 Status CLElementwiseMax::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, const ActivationLayerInfo &act_info)
 {
-    return experimental::CLElementwiseMax::validate(input1, input2, output, act_info);
+    return opencl::ClElementwiseMax::validate(input1, input2, output, act_info);
 }
 
 void CLElementwiseMax::run()
@@ -325,10 +218,10 @@ void CLElementwiseMax::run()
 
 struct CLElementwiseMin::Impl
 {
-    const ICLTensor                                *src_0{ nullptr };
-    const ICLTensor                                *src_1{ nullptr };
-    ICLTensor                                      *dst{ nullptr };
-    std::unique_ptr<experimental::CLElementwiseMin> op{ nullptr };
+    const ICLTensor                          *src_0{ nullptr };
+    const ICLTensor                          *src_1{ nullptr };
+    ICLTensor                                *dst{ nullptr };
+    std::unique_ptr<opencl::ClElementwiseMin> op{ nullptr };
 };
 
 CLElementwiseMin::CLElementwiseMin()
@@ -349,13 +242,13 @@ void CLElementwiseMin::configure(const CLCompileContext &compile_context, ICLTen
     _impl->src_0 = input1;
     _impl->src_1 = input2;
     _impl->dst   = output;
-    _impl->op    = std::make_unique<experimental::CLElementwiseMin>();
+    _impl->op    = std::make_unique<opencl::ClElementwiseMin>();
     _impl->op->configure(compile_context, input1->info(), input2->info(), output->info(), act_info);
 }
 
 Status CLElementwiseMin::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, const ActivationLayerInfo &act_info)
 {
-    return experimental::CLElementwiseMin::validate(input1, input2, output, act_info);
+    return opencl::ClElementwiseMin::validate(input1, input2, output, act_info);
 }
 
 void CLElementwiseMin::run()
@@ -370,10 +263,10 @@ void CLElementwiseMin::run()
 
 struct CLElementwiseSquaredDiff::Impl
 {
-    const ICLTensor                                        *src_0{ nullptr };
-    const ICLTensor                                        *src_1{ nullptr };
-    ICLTensor                                              *dst{ nullptr };
-    std::unique_ptr<experimental::CLElementwiseSquaredDiff> op{ nullptr };
+    const ICLTensor                                  *src_0{ nullptr };
+    const ICLTensor                                  *src_1{ nullptr };
+    ICLTensor                                        *dst{ nullptr };
+    std::unique_ptr<opencl::ClElementwiseSquaredDiff> op{ nullptr };
 };
 
 CLElementwiseSquaredDiff::CLElementwiseSquaredDiff()
@@ -394,13 +287,13 @@ void CLElementwiseSquaredDiff::configure(const CLCompileContext &compile_context
     _impl->src_0 = input1;
     _impl->src_1 = input2;
     _impl->dst   = output;
-    _impl->op    = std::make_unique<experimental::CLElementwiseSquaredDiff>();
+    _impl->op    = std::make_unique<opencl::ClElementwiseSquaredDiff>();
     _impl->op->configure(compile_context, input1->info(), input2->info(), output->info(), act_info);
 }
 
 Status CLElementwiseSquaredDiff::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, const ActivationLayerInfo &act_info)
 {
-    return experimental::CLElementwiseSquaredDiff::validate(input1, input2, output, act_info);
+    return opencl::ClElementwiseSquaredDiff::validate(input1, input2, output, act_info);
 }
 
 void CLElementwiseSquaredDiff::run()
@@ -415,10 +308,10 @@ void CLElementwiseSquaredDiff::run()
 
 struct CLElementwisePower::Impl
 {
-    const ICLTensor                                  *src_0{ nullptr };
-    const ICLTensor                                  *src_1{ nullptr };
-    ICLTensor                                        *dst{ nullptr };
-    std::unique_ptr<experimental::CLElementwisePower> op{ nullptr };
+    const ICLTensor                            *src_0{ nullptr };
+    const ICLTensor                            *src_1{ nullptr };
+    ICLTensor                                  *dst{ nullptr };
+    std::unique_ptr<opencl::ClElementwisePower> op{ nullptr };
 };
 
 CLElementwisePower::CLElementwisePower()
@@ -439,13 +332,13 @@ void CLElementwisePower::configure(const CLCompileContext &compile_context, ICLT
     _impl->src_0 = input1;
     _impl->src_1 = input2;
     _impl->dst   = output;
-    _impl->op    = std::make_unique<experimental::CLElementwisePower>();
+    _impl->op    = std::make_unique<opencl::ClElementwisePower>();
     _impl->op->configure(compile_context, input1->info(), input2->info(), output->info(), act_info);
 }
 
 Status CLElementwisePower::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, const ActivationLayerInfo &act_info)
 {
-    return experimental::CLElementwisePower::validate(input1, input2, output, act_info);
+    return opencl::ClElementwisePower::validate(input1, input2, output, act_info);
 }
 
 void CLElementwisePower::run()
diff --git a/src/runtime/gpu/cl/operators/ClElementwiseOperations.cpp b/src/runtime/gpu/cl/operators/ClElementwiseOperations.cpp
new file mode 100644
index 0000000000..e5b836a0d8
--- /dev/null
+++ b/src/runtime/gpu/cl/operators/ClElementwiseOperations.cpp
@@ -0,0 +1,92 @@
+/*
+ * 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/gpu/cl/operators/ClElementwiseOperations.h"
+
+#include "src/core/gpu/cl/kernels/ClElementwiseKernel.h"
+
+namespace arm_compute
+{
+namespace opencl
+{
+void ClElementwiseDivision::configure(const ClCompileContext &compile_context, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst, const ActivationLayerInfo &act_info)
+{
+    auto k = std::make_unique<kernels::ClArithmeticKernel>();
+    k->configure(compile_context, ArithmeticOperation::DIV, src1, src2, dst, act_info);
+    _kernel = std::move(k);
+}
+
+Status ClElementwiseDivision::validate(const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, const ActivationLayerInfo &act_info)
+{
+    return kernels::ClArithmeticKernel::validate(ArithmeticOperation::DIV, src1, src2, dst, act_info);
+}
+
+void ClElementwiseMax::configure(const ClCompileContext &compile_context, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst, const ActivationLayerInfo &act_info)
+{
+    auto k = std::make_unique<kernels::ClArithmeticKernel>();
+    k->configure(compile_context, ArithmeticOperation::MAX, src1, src2, dst, act_info);
+    _kernel = std::move(k);
+}
+
+Status ClElementwiseMax::validate(const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, const ActivationLayerInfo &act_info)
+{
+    return kernels::ClArithmeticKernel::validate(ArithmeticOperation::MAX, src1, src2, dst, act_info);
+}
+
+void ClElementwiseMin::configure(const ClCompileContext &compile_context, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst, const ActivationLayerInfo &act_info)
+{
+    auto k = std::make_unique<kernels::ClArithmeticKernel>();
+    k->configure(compile_context, ArithmeticOperation::MIN, src1, src2, dst, act_info);
+    _kernel = std::move(k);
+}
+
+Status ClElementwiseMin::validate(const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, const ActivationLayerInfo &act_info)
+{
+    return kernels::ClArithmeticKernel::validate(ArithmeticOperation::MIN, src1, src2, dst, act_info);
+}
+
+void ClElementwiseSquaredDiff::configure(const ClCompileContext &compile_context, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst, const ActivationLayerInfo &act_info)
+{
+    auto k = std::make_unique<kernels::ClArithmeticKernel>();
+    k->configure(compile_context, ArithmeticOperation::SQUARED_DIFF, src1, src2, dst, act_info);
+    _kernel = std::move(k);
+}
+
+Status ClElementwiseSquaredDiff::validate(const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, const ActivationLayerInfo &act_info)
+{
+    return kernels::ClArithmeticKernel::validate(ArithmeticOperation::SQUARED_DIFF, src1, src2, dst, act_info);
+}
+
+void ClElementwisePower::configure(const ClCompileContext &compile_context, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst, const ActivationLayerInfo &act_info)
+{
+    auto k = std::make_unique<kernels::ClArithmeticKernel>();
+    k->configure(compile_context, ArithmeticOperation::POWER, src1, src2, dst, act_info);
+    _kernel = std::move(k);
+}
+
+Status ClElementwisePower::validate(const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, const ActivationLayerInfo &act_info)
+{
+    return kernels::ClArithmeticKernel::validate(ArithmeticOperation::POWER, src1, src2, dst, act_info);
+}
+} // namespace opencl
+} // namespace arm_compute
diff --git a/src/runtime/gpu/cl/operators/ClElementwiseOperations.h b/src/runtime/gpu/cl/operators/ClElementwiseOperations.h
new file mode 100644
index 0000000000..b9ab1405c8
--- /dev/null
+++ b/src/runtime/gpu/cl/operators/ClElementwiseOperations.h
@@ -0,0 +1,190 @@
+/*
+ * 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_CL_ELEMENTWISE_OPERATIONS_H
+#define ARM_COMPUTE_CL_ELEMENTWISE_OPERATIONS_H
+
+#include "src/core/gpu/cl/ClCompileContext.h"
+#include "src/runtime/gpu/cl/IClOperator.h"
+
+namespace arm_compute
+{
+namespace opencl
+{
+/** Basic function to run @ref opencl::kernels::ClArithmeticKernel for division
+ *
+ * @note The tensor data type for the inputs must be F16/F32.
+ * @note The function performs an arithmetic division between two tensors.
+ */
+class ClElementwiseDivision : public IClOperator
+{
+public:
+    /** Default Constructor */
+    ClElementwiseDivision() = default;
+    /** Configure function for a given list of arguments.
+     *
+     * @param[in]  compile_context The compile context to be used.
+     * @param[in]  src1            First source tensor info. Data types supported: F16/F32.
+     * @param[in]  src2            Second source tensor info. same as @p src1.
+     * @param[out] dst             Destination tensor info. Data types supported: same as @p src1.
+     * @param[in]  act_info        (Optional) Activation layer information in case of a fused activation.
+     */
+    void configure(const ClCompileContext &compile_context, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst, const ActivationLayerInfo &act_info = ActivationLayerInfo());
+    /** Static function to check if given info will lead to a valid configuration of @ref ClElementwiseDivision
+     *
+     * @param[in] src1     First source tensor info. Data types supported: F16/F32.
+     * @param[in] src2     Second source tensor info. Data types supported: same as @p src1.
+     * @param[in] dst      Destination tensor info. Data types supported: same as @p src1.
+     * @param[in] act_info (Optional) Activation layer information in case of a fused activation.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, const ActivationLayerInfo &act_info = ActivationLayerInfo());
+};
+
+/** Basic function to run @ref opencl::kernels::ClArithmeticKernel for max
+ *
+ * @note The tensor data type for the inputs must be U8/QASYMM8/S16/QSYMM16/S32/U32/F16/F32.
+ * @note The function performs a max operation between two tensors.
+ */
+class ClElementwiseMax : public IClOperator
+{
+public:
+    /** Default Constructor */
+    ClElementwiseMax() = default;
+    /** Configure function for a given list of arguments.
+     *
+     * @param[in]  compile_context The compile context to be used.
+     * @param[in]  src1            First source tensor info. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/S32/U32/F16/F32.
+     * @param[in]  src2            Second source tensor info. Data types supported: same as @p src1.
+     * @param[out] dst             Destination tensor info. Data types supported: same as @p src1.
+     * @param[in]  act_info        (Optional) Activation layer information in case of a fused activation.
+     */
+    void configure(const ClCompileContext &compile_context, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst, const ActivationLayerInfo &act_info = ActivationLayerInfo());
+    /** Static function to check if given info will lead to a valid configuration of @ref opencl::kernels::ClArithmeticKernel for max
+     *
+     * @param[in] src1     First source tensor info. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/S32/U32/F16/F32.
+     * @param[in] src2     Second source tensor info. Data types supported: same as @p src1.
+     * @param[in] dst      Destination tensor info. Data types supported: same as @p src1.
+     * @param[in] act_info (Optional) Activation layer information in case of a fused activation.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, const ActivationLayerInfo &act_info = ActivationLayerInfo());
+};
+
+/** Basic function to run @ref opencl::kernels::ClArithmeticKernel for min
+ *
+ * @note The tensor data type for the inputs must be U8/QASYMM8/S16/QSYMM16/S32/U32/F16/F32.
+ * @note The function performs a max operation between two tensors.
+ */
+class ClElementwiseMin : public IClOperator
+{
+public:
+    /** Default Constructor */
+    ClElementwiseMin() = default;
+    /** Configure function for a given list of arguments.
+     *
+     * @param[in]  compile_context The compile context to be used.
+     * @param[in]  src1            First source tensor info. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/S32/U32/F16/F32.
+     * @param[in]  src2            Second source tensor info. Data types supported: same as @p src1.
+     * @param[out] dst             Destination tensor info. Data types supported: same as @p src1.
+     * @param[in]  act_info        (Optional) Activation layer information in case of a fused activation.
+     */
+    void configure(const ClCompileContext &compile_context, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst, const ActivationLayerInfo &act_info = ActivationLayerInfo());
+    /** Static function to check if given info will lead to a valid configuration of @ref opencl::kernels::ClArithmeticKernel for min
+     *
+     * @param[in] src1     First source tensor info. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/S32/U32/F16/F32.
+     * @param[in] src2     Second source tensor info. Data types supported: same as @p src1.
+     * @param[in] dst      Destination tensor info. Data types supported: same as @p src1.
+     * @param[in] act_info (Optional) Activation layer information in case of a fused activation.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, const ActivationLayerInfo &act_info = ActivationLayerInfo());
+};
+
+/** Basic function to run @ref opencl::kernels::ClArithmeticKernel for squared difference
+ *
+ * @note The tensor data type for the inputs must be QASYMM8/U8/S16/QSYMM16/F16/F32.
+ * @note The function performs a squared different operation between two tensors (i.e., out[i] = (in1[i] - in2[i])^2
+ */
+class ClElementwiseSquaredDiff : public IClOperator
+{
+public:
+    /** Default Constructor */
+    ClElementwiseSquaredDiff() = default;
+    /** Configure function for a given list of arguments.
+     *
+     * @param[in]  compile_context The compile context to be used.
+     * @param[in]  src1            First source tensor info. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/F32.
+     * @param[in]  src2            Second source tensor info. Data types supported: same as @p src1.
+     * @param[out] dst             Destination tensor info. Data types supported: same as @p src1.
+     * @param[in]  act_info        (Optional) Activation layer information in case of a fused activation.
+     */
+    void configure(const ClCompileContext &compile_context, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst, const ActivationLayerInfo &act_info = ActivationLayerInfo());
+    /** Static function to check if given info will lead to a valid configuration of @ref opencl::kernels::ClArithmeticKernel for squared difference
+     *
+     * @param[in] src1     First source tensor info. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/F32.
+     * @param[in] src2     Second source tensor info. Data types supported: same as @p src1.
+     * @param[in] dst      Destination tensor info. Data types supported: same as @p src1.
+     * @param[in] act_info (Optional) Activation layer information in case of a fused activation.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, const ActivationLayerInfo &act_info = ActivationLayerInfo());
+};
+
+/** Basic function to run @ref opencl::kernels::ClArithmeticKernel for power
+ *
+ * @note The tensor data type for the inputs must be F16/F32.
+ * @note The function performs an elementwise power of in1 to in2 (i.e., out[i] = in1[i] ^ in2[i])
+ */
+class ClElementwisePower : public IClOperator
+{
+public:
+    /** Default Constructor */
+    ClElementwisePower() = default;
+    /** Configure function for a given list of arguments.
+     *
+     * @param[in]  compile_context The compile context to be used.
+     * @param[in]  src1            First source tensor info. Data types supported: F16/F32.
+     * @param[in]  src2            Second source tensor info. Data types supported: F16/F32.
+     * @param[out] dst             Destination tensor info. Data types supported:F16/F32.
+     * @param[in]  act_info        (Optional) Activation layer information in case of a fused activation.
+     */
+    void configure(const ClCompileContext &compile_context, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst, const ActivationLayerInfo &act_info = ActivationLayerInfo());
+    /** Static function to check if given info will lead to a valid configuration of @ref opencl::kernels::ClArithmeticKernel for power
+     *
+     * @param[in] src1     First source tensor info. Data types supported: F16/F32.
+     * @param[in] src2     Second source tensor info. Data types supported: F16/F32.
+     * @param[in] dst      Destination tensor info. Data types supported: F16/F32.
+     * @param[in] act_info (Optional) Activation layer information in case of a fused activation.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, const ActivationLayerInfo &act_info = ActivationLayerInfo());
+};
+} // namespace opencl
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_CL_ELEMENTWISE_OPERATIONS_H */