COMPMID-1405: Create our own gemm_native kernel / function.

Change-Id: Ie0a80bd6b4eb5632cac63ccf54bcb07d4309da19
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/140305
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

13 files changed, 598 insertions, 35 deletions

diff --git a/SConscript b/SConscript
index 983570b7dc..bfc3e78b70 100644
--- a/SConscript
+++ b/SConscript
@@ -190,6 +190,7 @@ if env['opencl']:
 if env['neon']:
     core_files += Glob('src/core/NEON/*.cpp')
     core_files += Glob('src/core/NEON/kernels/*.cpp')
+    core_files += Glob('src/core/NEON/kernels/assembly/*.cpp')
 
     core_files += Glob('src/core/NEON/kernels/arm_gemm/*.cpp')
 
@@ -209,6 +210,7 @@ if env['neon']:
 
     runtime_files += Glob('src/runtime/NEON/*.cpp')
     runtime_files += Glob('src/runtime/NEON/functions/*.cpp')
+    runtime_files += Glob('src/runtime/NEON/functions/assembly/*.cpp')
 
 if env['gles_compute']:
     if env['os'] != 'android':
diff --git a/arm_compute/core/Error.h b/arm_compute/core/Error.h
index 9f51fa234c..f137be6ecb 100644
--- a/arm_compute/core/Error.h
+++ b/arm_compute/core/Error.h
@@ -157,7 +157,7 @@ Status create_error(ErrorCode error_code, const char *function, const char *file
  *
  * @param[in] ... Variables which are unused.
  */
-#define ARM_COMPUTE_UNUSED(...) arm_compute::ignore_unused(__VA_ARGS__) // NOLINT
+#define ARM_COMPUTE_UNUSED(...) ::arm_compute::ignore_unused(__VA_ARGS__) // NOLINT
 
 /** Creates an error with a given message
  *
diff --git a/arm_compute/core/NEON/kernels/assembly/INEGEMMWrapperKernel.h b/arm_compute/core/NEON/kernels/assembly/INEGEMMWrapperKernel.h
new file mode 100644
index 0000000000..02e5b58c9d
--- /dev/null
+++ b/arm_compute/core/NEON/kernels/assembly/INEGEMMWrapperKernel.h
@@ -0,0 +1,106 @@
+/*
+ * Copyright (c) 2018 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_INEGEMMWRAPPERKERNEL_H__
+#define __ARM_COMPUTE_INEGEMMWRAPPERKERNEL_H__
+
+#include "arm_compute/core/NEON/INEKernel.h"
+
+namespace arm_compute
+{
+class ITensor;
+
+/** Common interface for all the arm_gemm Gemms
+ */
+class INEGEMMWrapperKernel : public INEKernel
+{
+public:
+    /** Parameters defining the dimensions of the matrices being multiplied */
+    struct Params
+    {
+        unsigned int M;       /**< Rows in output matrix C (and input matrix A). */
+        unsigned int N;       /**< Columns in output matrix C (and input matrix B). */
+        unsigned int K;       /**< Columns of input matrix A (= rows of input matrix B). */
+        unsigned int batches; /**< Number of "batched" GEMMs (unique A and C, shared B). */
+        unsigned int multis;  /**< Number of "multi" GEMMs (unique A, B and C). */
+    };
+
+    static Params extract_parameters(const ITensor *a, const ITensor *b, const ITensor *c);
+
+    /** Constructor */
+    INEGEMMWrapperKernel();
+    /** Prevent instances of this class from being copied */
+    INEGEMMWrapperKernel(const INEGEMMWrapperKernel &) = delete;
+    /** Prevent instances of this class from being copied */
+    INEGEMMWrapperKernel &operator=(const INEGEMMWrapperKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    INEGEMMWrapperKernel(INEGEMMWrapperKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    INEGEMMWrapperKernel &operator=(INEGEMMWrapperKernel &&) = default;
+    /** Initialise the kernel's input and output.
+     *
+     * @note The input and output tensor must have the same dimensions
+     *
+     * @param[in]  a     Input tensor (Matrix A)
+     * @param[in]  b     Input tensor (Matrix B)
+     * @param[out] c     Output tensor to store the result of matrix multiplication. Data type supported: same as @p input0.
+     * @param[in]  alpha Scalar multiplier to apply to AB matrix product.
+     * @param[in]  beta  Scalar multiplier to apply to input C matrix before adding product.
+     */
+    void configure(const ITensor *a, const ITensor *b, ITensor *c, float alpha, float beta);
+
+    // Inherited methods overridden:
+    void run(const Window &window, const ThreadInfo &info) override;
+
+protected:
+    /** Called as part of configure() after _a, _b, _c and _params have been set.
+     *
+     * @param[in] alpha Scalar multiplier to apply to AB matrix product.
+     * @param[in] beta  Scalar multiplier to apply to input C matrix before adding product.
+     *
+     * @return A 3D execution window.
+     */
+    virtual Window configure_internal(float alpha, float beta) = 0;
+
+    /** Run the kernel from the start to the end offset in window.
+     *
+     * @param[in] window       Window to use for the iteration
+     * @param[in] start_offset Where to start iterating from (In Window coordinates)
+     * @param[in] end_offset   Where to stop iterating (In Window coordinates).
+     * @param[in] info         Info about executing thread and CPU.
+     */
+    virtual void run_internal(const Window &window, const Coordinates &start_offset, const Coordinates &end_offset, const ThreadInfo &info) = 0;
+
+    const ITensor *_a;
+    const ITensor *_b;
+    ITensor       *_c;
+    Params         _params;
+
+private:
+    Window      _window3d;
+    TensorShape _window_shape;
+};
+
+} // namespace arm_compute
+
+#endif /* __ARM_COMPUTE_INEGEMMRAPPERKERNEL_H__ */
diff --git a/arm_compute/core/NEON/kernels/assembly/NEGEMMNativeWrapperKernel.h b/arm_compute/core/NEON/kernels/assembly/NEGEMMNativeWrapperKernel.h
new file mode 100644
index 0000000000..73a0d7f05f
--- /dev/null
+++ b/arm_compute/core/NEON/kernels/assembly/NEGEMMNativeWrapperKernel.h
@@ -0,0 +1,52 @@
+/*
+ * Copyright (c) 2018 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_NEGEMMNATIVEWRAPPERKERNEL_H__
+#define __ARM_COMPUTE_NEGEMMNATIVEWRAPPERKERNEL_H__
+
+#include "INEGEMMWrapperKernel.h"
+
+namespace arm_compute
+{
+/** Equivalent to arm_gemm::GemmNative but using Compute Library types.
+ */
+template <typename To, typename Tr>
+class NEGEMMNativeWrapperKernel : public INEGEMMWrapperKernel
+{
+public:
+    const char *name() const override
+    {
+        return "NEGEMMNativeWrapperKernel";
+    }
+
+protected:
+    // Inherited methods overridden:
+    Window configure_internal(float alpha, float beta) override;
+    void run_internal(const Window &window, const Coordinates &start_offset, const Coordinates &end_offset, const ThreadInfo &info) override;
+
+private:
+    Tr _beta{};
+};
+
+} // namespace arm_compute
+#endif /* __ARM_COMPUTE_NEGEMMNATIVEWRAPPERKERNEL_H__ */
diff --git a/arm_compute/core/WindowIterator.h b/arm_compute/core/WindowIterator.h
index 13e9973506..13cb9cf2f4 100644
--- a/arm_compute/core/WindowIterator.h
+++ b/arm_compute/core/WindowIterator.h
@@ -24,6 +24,7 @@
 #ifndef __ARM_COMPUTE_WINDOW_ITERATOR_H__
 #define __ARM_COMPUTE_WINDOW_ITERATOR_H__
 #include "arm_compute/core/Coordinates.h"
+#include "arm_compute/core/Error.h"
 #include "arm_compute/core/ITensor.h"
 #include "arm_compute/core/Window.h"
 
@@ -71,9 +72,20 @@ public:
      */
     inline size_t stride(size_t dim) const
     {
+        ARM_COMPUTE_ERROR_ON(_strides[dim] % sizeof(T) != 0);
         return _strides[dim] / sizeof(T);
     }
 
+    /** Manually set the stride of a dimension
+     *
+     * @param[in] dim  Dimension of the stride to set.
+     * @param[in] size Value to set the stride to (in bytes).
+     */
+    void set_stride(size_t dim, size_t size)
+    {
+        _strides[dim] = size;
+    }
+
     /** Returns a pointer to the element at coordinates (x,y,z,w)
      *
      * @param[in] x X coordinates
@@ -99,8 +111,8 @@ public:
     }
 
 private:
-    uint8_t       *_first;   /**< Pointer to the first element of the tensor.*/
-    const Strides &_strides; /**< Strides in bytes of the tensor */
+    uint8_t *_first;   /**< Pointer to the first element of the tensor.*/
+    Strides  _strides; /**< Strides in bytes of the tensor */
 };
 
 /** Iterate over a portion of a Window */
diff --git a/arm_compute/runtime/NEON/NEFunctions.h b/arm_compute/runtime/NEON/NEFunctions.h
index 6a3fabca67..fdb1450179 100644
--- a/arm_compute/runtime/NEON/NEFunctions.h
+++ b/arm_compute/runtime/NEON/NEFunctions.h
@@ -104,6 +104,7 @@
 #include "arm_compute/runtime/NEON/functions/NEReshapeLayer.h"
 #include "arm_compute/runtime/NEON/functions/NEScale.h"
 #include "arm_compute/runtime/NEON/functions/NEScharr3x3.h"
+#include "arm_compute/runtime/NEON/functions/NESimpleAssemblyFunction.h"
 #include "arm_compute/runtime/NEON/functions/NESobel3x3.h"
 #include "arm_compute/runtime/NEON/functions/NESobel5x5.h"
 #include "arm_compute/runtime/NEON/functions/NESobel7x7.h"
diff --git a/arm_compute/runtime/NEON/functions/NEGEMMAssemblyDispatch.h b/arm_compute/runtime/NEON/functions/NEGEMMAssemblyDispatch.h
index 4ac6a3cae2..1c9ecb088e 100644
--- a/arm_compute/runtime/NEON/functions/NEGEMMAssemblyDispatch.h
+++ b/arm_compute/runtime/NEON/functions/NEGEMMAssemblyDispatch.h
@@ -54,6 +54,20 @@ private:
     /** ACL Function */
     std::unique_ptr<IFunction> _function;
 
+    /** If supported create the ACL function corresponding to the GemmMethod provided to process the other passed parameters
+     *
+     * @param[in]  method             GemmMethod to use to perform the matrix multiplication.
+     * @param[in]  a                  Input tensor (Matrix A).
+     * @param[in]  b                  Input tensor (Matrix B).
+     * @param[out] d                  Output tensor to store the result of matrix multiplication. Data type supported: same as @p input0.
+     * @param[in]  alpha              Scalar multiplier to apply to AB matrix product.
+     * @param[in]  beta               Scalar multiplier to apply to input D matrix before adding product.
+     * @param[in]  pretransposed_hint Can the B tensor can be pretransposed (ie shared across invocations)?
+     *
+     * @return True if the method is supported and the function was successfully created, false otherwise.
+     */
+    bool create_function(arm_gemm::GemmMethod method, const ITensor *a, const ITensor *b, ITensor *d, float alpha, float beta, bool pretranspose_hint);
+
     //Fallback: use arm_gemm's AssemblyGemm:
     class Fallback
     {
@@ -63,7 +77,7 @@ private:
          *  The call to set_arrays is needed to deal with the input sizes containing batches (dims > 2)
          */
         void run();
-        void configure(const ITensor *a, const ITensor *b, ITensor *d, float alpha, float beta, bool pretranspose_hint, MemoryGroup &memory_group);
+        void configure(const ITensor *a, const ITensor *b, ITensor *d, arm_gemm::GemmArgs<TypeOutput> &args, MemoryGroup &memory_group);
         void prepare();
         bool is_configured() const;
 #endif /* DOXYGEN_SKIP_THIS */
@@ -102,7 +116,20 @@ private:
     } _arm_gemm;               /**< Fallback in case ACL doesn't have a function */
     MemoryGroup _memory_group; /**< Function memory group */
 public:
+    /** If supported create an ACL function else fallback to the arm_gemm function.
+     *
+     * @param[in]  a                 Input tensor (Matrix A)
+     * @param[in]  b                 Input tensor (Matrix B)
+     * @param[out] d                 Output tensor to store the result of matrix multiplication. Data type supported: same as @p input0.
+     * @param[in]  alpha             Scalar multiplier to apply to AB matrix product.
+     * @param[in]  beta              Scalar multiplier to apply to input D matrix before adding product.
+     * @param[in]  pretranspose_hint Can the B tensor can be pretransposed (ie shared across invocations)?
+     */
     void configure(const ITensor *a, const ITensor *b, ITensor *d, float alpha, float beta, bool pretranspose_hint);
+    /** Was the function successfully configured ?
+     *
+     * @return True if the function is configured and ready to run
+     */
     bool is_configured() const;
     // Inherited methods overridden:
     /** Runs a preparation step, usually for pre-transposing matrix b */
@@ -110,11 +137,11 @@ public:
     void run() override;
 };
 
-/** Float 32 assembly kernel glue */
+/** Float 32 assembly dispatch kernel */
 using NEGEMMAssemblyDispatchF32 = NEGEMMAssemblyDispatch<float, float>;
-/** Uint 8 to Uint 32 kernel glue */
+/** Uint 8 to Uint 32 assembly dispatch kernel */
 using NEGEMMAssemblyDispatchU8U32 = NEGEMMAssemblyDispatch<uint8_t, uint32_t>;
-/** Int 8 to Int 32 kernel glue */
+/** Int 8 to Int 32 assembly dispatch kernel */
 using NEGEMMAssemblyDispatchS8S32 = NEGEMMAssemblyDispatch<int8_t, int32_t>;
-}
+} // namespace arm_compute
 #endif /* __ARM_COMPUTE_NEGEMMASSEMBLYDISPATCH_H__ */
diff --git a/arm_compute/runtime/NEON/functions/NESimpleAssemblyFunction.h b/arm_compute/runtime/NEON/functions/NESimpleAssemblyFunction.h
new file mode 100644
index 0000000000..5a76c15627
--- /dev/null
+++ b/arm_compute/runtime/NEON/functions/NESimpleAssemblyFunction.h
@@ -0,0 +1,56 @@
+/*
+ * Copyright (c) 2018 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_NESIMPLEASSEMBLYFUNCTION_H__
+#define __ARM_COMPUTE_NESIMPLEASSEMBLYFUNCTION_H__
+
+#include "arm_compute/core/NEON/kernels/assembly/INEGEMMWrapperKernel.h"
+#include "arm_compute/runtime/IFunction.h"
+
+#include <memory>
+
+namespace arm_compute
+{
+/** Basic interface for functions which have a single NEON GEMM wrapper kernel to run */
+class NESimpleAssemblyFunction : public IFunction
+{
+public:
+    /** Constructor */
+    NESimpleAssemblyFunction();
+
+    /** Configure the function with the kernel to run
+     *
+     * @param[in] kernel GEMM Wrapper kernel configured and ready to run
+     *
+     * @note The kernel is expected to have a 1D window. The function will multi-thread this window across the X dimension.
+     */
+    void configure(std::unique_ptr<INEGEMMWrapperKernel> kernel);
+
+    // Inherited methods overridden:
+    void run() override final;
+
+protected:
+    std::unique_ptr<INEGEMMWrapperKernel> _kernel; /**< Kernel to run */
+};
+} //namespace arm_compute
+#endif /*__ARM_COMPUTE_NESIMPLEASSEMBLYFUNCTION_H__ */
diff --git a/src/core/NEON/kernels/assembly/INEGEMMWrapperKernel.cpp b/src/core/NEON/kernels/assembly/INEGEMMWrapperKernel.cpp
new file mode 100644
index 0000000000..0fc3610014
--- /dev/null
+++ b/src/core/NEON/kernels/assembly/INEGEMMWrapperKernel.cpp
@@ -0,0 +1,80 @@
+/*
+ * Copyright (c) 2018 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/NEON/kernels/assembly/INEGEMMWrapperKernel.h"
+
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/WindowIterator.h"
+
+using namespace arm_compute;
+
+INEGEMMWrapperKernel::INEGEMMWrapperKernel()
+    : _a(nullptr), _b(nullptr), _c(nullptr), _params(), _window3d(), _window_shape()
+{
+}
+
+INEGEMMWrapperKernel::Params INEGEMMWrapperKernel::extract_parameters(const ITensor *a, const ITensor *b, const ITensor *c)
+{
+    Params p;
+
+    ARM_COMPUTE_ERROR_ON_NULLPTR(a);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(b);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(c);
+
+    p.M       = c->info()->tensor_shape().y();
+    p.N       = c->info()->tensor_shape().x();
+    p.K       = a->info()->tensor_shape().x();
+    p.multis  = b->info()->tensor_shape().z();
+    p.batches = c->info()->tensor_shape().total_size_upper(2) / p.multis; //COMPMID-1423: Agree on and document the layout of gemm inputs/outputs
+
+    return p;
+}
+
+void INEGEMMWrapperKernel::configure(const ITensor *a, const ITensor *b, ITensor *c, float alpha, float beta)
+{
+    _params = extract_parameters(a, b, c);
+    _a      = a;
+    _b      = b;
+    _c      = c;
+
+    _window3d     = configure_internal(alpha, beta);
+    _window_shape = _window3d.shape();
+
+    // Convert the 3D window into a 1D window in order to allow the scheduler to arbitrary split it.
+    Window collapsed;
+    collapsed.set(0, Window::Dimension(0, _window3d.num_iterations_total()));
+
+    INEKernel::configure(collapsed);
+}
+
+void INEGEMMWrapperKernel::run(const Window &window, const ThreadInfo &info)
+{
+    const Coordinates start_offset = index2coords(_window_shape, window.x().start());
+    const Coordinates end_offset   = index2coords(_window_shape, window.x().end() - 1);
+
+    run_internal(_window3d, start_offset, end_offset, info);
+}
diff --git a/src/core/NEON/kernels/assembly/NEGEMMNativeWrapperKernel.cpp b/src/core/NEON/kernels/assembly/NEGEMMNativeWrapperKernel.cpp
new file mode 100644
index 0000000000..38eb40a06e
--- /dev/null
+++ b/src/core/NEON/kernels/assembly/NEGEMMNativeWrapperKernel.cpp
@@ -0,0 +1,119 @@
+/*
+ * Copyright (c) 2018 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/NEON/kernels/assembly/NEGEMMNativeWrapperKernel.h"
+
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/core/WindowIterator.h"
+
+#include "../arm_gemm/utils.hpp"
+#include "arm_gemm.hpp"
+
+#include "../arm_gemm/mergeresults.hpp"
+#include "../arm_gemm/transform.hpp"
+
+#include "../arm_gemm/kernels/a32_sgemm_8x6.hpp"
+#include "../arm_gemm/kernels/a64_sgemm_12x8.hpp"
+#include "../arm_gemm/kernels/a64_sgemm_native_16x4.hpp"
+#include "../arm_gemm/kernels/a64_sgemv_pretransposed.hpp"
+#include "../arm_gemm/kernels/a64_sgemv_trans.hpp"
+
+namespace arm_compute
+{
+template <typename To, typename Tr>
+struct Kernel
+{
+};
+
+#ifdef __aarch64__
+template <>
+struct Kernel<float, float>
+{
+    using strategy = arm_gemm::sgemm_native_16x4;
+};
+#endif /* __aarch64__ */
+
+template <typename To, typename Tr>
+Window NEGEMMNativeWrapperKernel<To, Tr>::configure_internal(float alpha, float beta)
+{
+    using strategy = typename Kernel<To, Tr>::strategy;
+
+    _beta = beta;
+
+    //Note: The window is shifted down by 1 dimension compare to the tensors
+    Window window;
+    window.set(Window::DimX, Window::Dimension(0, ceil_to_multiple(_params.M, strategy::out_height()), strategy::out_height()));
+    window.set(Window::DimY, Window::Dimension(0, _params.batches));
+    window.set(Window::DimZ, Window::Dimension(0, _params.multis));
+
+    return window;
+}
+
+template <typename To, typename Tr>
+void NEGEMMNativeWrapperKernel<To, Tr>::run_internal(const Window &window, const Coordinates &start_offset, const Coordinates &end_offset, const ThreadInfo &info)
+{
+    using strategy = typename Kernel<To, Tr>::strategy;
+
+    TensorAccessor<To> a(*_a);
+    TensorAccessor<To> b(*_b);
+    TensorAccessor<Tr> c(*_c);
+
+    if(_a->info()->data_layout() == DataLayout::NHWC)
+    {
+        // In the case of NHWC we want to interpret the output shape as 3D. Thus, the batch stride for A is
+        // the relevant multiple of the row stride.
+        const size_t nhwc_batch_stride = _a->info()->strides_in_bytes().y() * _c->info()->dimension(1);
+        a.set_stride(2, nhwc_batch_stride);
+    }
+
+    unsigned int m_end = 0;
+
+    strategy strat(info.cpu_info);
+    auto window_iterator = arm_compute::create_window_iterator(window, start_offset, end_offset, [&](const Coordinates & id)
+    {
+        const unsigned int y0    = id.x();
+        const unsigned int batch = id.y();
+        const unsigned int multi = id.z();
+        const unsigned int ymax  = std::min(y0 + strategy::out_height(), m_end);
+
+        strat.kernel(a(0, y0, batch, multi), a.stride(Window::DimY),
+                     b(0, 0, 0, multi), b.stride(Window::DimY),
+                     c(0, y0, batch, multi), c.stride(Window::DimY),
+                     _beta, (ymax - y0), _params.N, _params.K);
+    });
+
+    auto on_new_row_size = [&](unsigned int start, unsigned int end)
+    {
+        m_end = std::min(end, _params.M);
+    };
+
+    window_iterator.iterate_3D(on_new_row_size);
+}
+
+#ifdef __aarch64__
+template class NEGEMMNativeWrapperKernel<float, float>;
+#endif /* __aarch64__ */
+
+} // namespace arm_compute
diff --git a/src/runtime/NEON/functions/NEGEMMAssemblyDispatch.cpp b/src/runtime/NEON/functions/NEGEMMAssemblyDispatch.cpp
index e796a6a56e..f4710fab84 100644
--- a/src/runtime/NEON/functions/NEGEMMAssemblyDispatch.cpp
+++ b/src/runtime/NEON/functions/NEGEMMAssemblyDispatch.cpp
@@ -23,21 +23,74 @@
  */
 #include "arm_compute/runtime/NEON/functions/NEGEMMAssemblyDispatch.h"
 
+#include "arm_compute/core/NEON/kernels/assembly/NEGEMMNativeWrapperKernel.h"
 #include "arm_compute/runtime/NEON/NEScheduler.h"
+#include "arm_compute/runtime/NEON/functions/NESimpleAssemblyFunction.h"
 
-using namespace arm_compute;
-
+namespace arm_compute
+{
 template <typename TypeInput, typename TypeOutput>
 NEGEMMAssemblyDispatch<TypeInput, TypeOutput>::NEGEMMAssemblyDispatch(std::shared_ptr<IMemoryManager> memory_manager)
     : _function(nullptr), _arm_gemm(), _memory_group(std::move(memory_manager))
 {
 }
 
+template <>
+bool NEGEMMAssemblyDispatch<float, float>::create_function(arm_gemm::GemmMethod method, const ITensor *a, const ITensor *b, ITensor *d, float alpha, float beta, bool pretranspose_hint)
+{
+    ARM_COMPUTE_UNUSED(method);
+    ARM_COMPUTE_UNUSED(a);
+    ARM_COMPUTE_UNUSED(b);
+    ARM_COMPUTE_UNUSED(d);
+    ARM_COMPUTE_UNUSED(alpha);
+    ARM_COMPUTE_UNUSED(beta);
+    ARM_COMPUTE_UNUSED(pretranspose_hint);
+    switch(method)
+    {
+#ifdef __aarch64__
+        case arm_gemm::GemmMethod::GEMM_NATIVE:
+        {
+            auto kernel = support::cpp14::make_unique<NEGEMMNativeWrapperKernel<float, float>>();
+            kernel->configure(a, b, d, alpha, beta);
+            auto function = support::cpp14::make_unique<NESimpleAssemblyFunction>();
+            function->configure(std::move(kernel));
+            _function = std::move(function);
+            return true;
+        }
+#endif /* __aarch64__ */
+        default:
+            return false;
+    }
+}
+
+template <typename TypeInput, typename TypeOutput>
+bool NEGEMMAssemblyDispatch<TypeInput, TypeOutput>::create_function(arm_gemm::GemmMethod method, const ITensor *a, const ITensor *b, ITensor *d, float alpha, float beta, bool pretranspose_hint)
+{
+    ARM_COMPUTE_UNUSED(method);
+    ARM_COMPUTE_UNUSED(a);
+    ARM_COMPUTE_UNUSED(b);
+    ARM_COMPUTE_UNUSED(d);
+    ARM_COMPUTE_UNUSED(alpha);
+    ARM_COMPUTE_UNUSED(beta);
+    ARM_COMPUTE_UNUSED(pretranspose_hint);
+    return false;
+}
+
 template <typename TypeInput, typename TypeOutput>
 void NEGEMMAssemblyDispatch<TypeInput, TypeOutput>::configure(const ITensor *a, const ITensor *b, ITensor *d, float alpha, float beta, bool pretranspose_hint)
 {
-    //TODO(antbar01) Check heuristics here to figure out if we should use an ACL IFunction
-    _arm_gemm.configure(a, b, d, alpha, beta, pretranspose_hint, _memory_group);
+    INEGEMMWrapperKernel::Params p           = INEGEMMWrapperKernel::extract_parameters(a, b, d);
+    const CPUInfo               &ci          = NEScheduler::get().cpu_info();
+    unsigned int                 num_threads = NEScheduler::get().num_threads();
+
+    arm_gemm::GemmArgs<TypeOutput> args(&ci, p.M, p.N, p.K, p.batches, p.multis, false, false, alpha, beta, num_threads, pretranspose_hint);
+
+    //Try to create an ACL function:
+    if(!create_function(arm_gemm::get_gemm_method<TypeInput, TypeOutput>(args), a, b, d, alpha, beta, pretranspose_hint))
+    {
+        //Fallback onto arm_gemm function if ACL doesn't support this method.
+        _arm_gemm.configure(a, b, d, args, _memory_group);
+    }
 }
 
 template <typename TypeInput, typename TypeOutput>
@@ -75,10 +128,8 @@ void NEGEMMAssemblyDispatch<TypeInput, TypeOutput>::run()
 }
 
 #ifndef __aarch64__
-namespace arm_compute
-{
 template <>
-void NEGEMMAssemblyDispatch<uint8_t, uint32_t>::Fallback::configure(const ITensor *a, const ITensor *b, ITensor *d, float alpha, float beta, bool pretranspose_hint, MemoryGroup &memory_group)
+void NEGEMMAssemblyDispatch<uint8_t, uint32_t>::configure(const ITensor *a, const ITensor *b, ITensor *d, float alpha, float beta, bool pretranspose_hint)
 {
     // arm_gemm::gemm for 8bit only exists for aarch64
     ARM_COMPUTE_UNUSED(a);
@@ -87,11 +138,11 @@ void NEGEMMAssemblyDispatch<uint8_t, uint32_t>::Fallback::configure(const ITenso
     ARM_COMPUTE_UNUSED(alpha);
     ARM_COMPUTE_UNUSED(beta);
     ARM_COMPUTE_UNUSED(pretranspose_hint);
-    ARM_COMPUTE_UNUSED(memory_group);
+    ARM_COMPUTE_ERROR("Not supported for this architecture");
 }
 
 template <>
-void NEGEMMAssemblyDispatch<int8_t, int32_t>::Fallback::configure(const ITensor *a, const ITensor *b, ITensor *d, float alpha, float beta, bool pretranspose_hint, MemoryGroup &memory_group)
+void NEGEMMAssemblyDispatch<int8_t, int32_t>::configure(const ITensor *a, const ITensor *b, ITensor *d, float alpha, float beta, bool pretranspose_hint)
 {
     // arm_gemm::gemm for 8bit only exists for aarch64
     ARM_COMPUTE_UNUSED(a);
@@ -100,23 +151,37 @@ void NEGEMMAssemblyDispatch<int8_t, int32_t>::Fallback::configure(const ITensor
     ARM_COMPUTE_UNUSED(alpha);
     ARM_COMPUTE_UNUSED(beta);
     ARM_COMPUTE_UNUSED(pretranspose_hint);
+    ARM_COMPUTE_ERROR("Not supported for this architecture");
+}
+
+template <>
+void NEGEMMAssemblyDispatch<uint8_t, uint32_t>::Fallback::configure(const ITensor *a, const ITensor *b, ITensor *d, arm_gemm::GemmArgs<uint32_t> &args, MemoryGroup &memory_group)
+{
+    // arm_gemm::gemm for 8bit only exists for aarch64
+    ARM_COMPUTE_UNUSED(a);
+    ARM_COMPUTE_UNUSED(b);
+    ARM_COMPUTE_UNUSED(d);
+    ARM_COMPUTE_UNUSED(args);
     ARM_COMPUTE_UNUSED(memory_group);
+    ARM_COMPUTE_ERROR("Not supported for this architecture");
 }
 
-} //namespace arm_compute
+template <>
+void NEGEMMAssemblyDispatch<int8_t, int32_t>::Fallback::configure(const ITensor *a, const ITensor *b, ITensor *d, arm_gemm::GemmArgs<int32_t> &args, MemoryGroup &memory_group)
+{
+    // arm_gemm::gemm for 8bit only exists for aarch64
+    ARM_COMPUTE_UNUSED(a);
+    ARM_COMPUTE_UNUSED(b);
+    ARM_COMPUTE_UNUSED(d);
+    ARM_COMPUTE_UNUSED(args);
+    ARM_COMPUTE_UNUSED(memory_group);
+    ARM_COMPUTE_ERROR("Not supported for this architecture");
+}
 #endif // aarch64
 template <typename TypeInput, typename TypeOutput>
-void NEGEMMAssemblyDispatch<TypeInput, TypeOutput>::Fallback::configure(const ITensor *a, const ITensor *b, ITensor *d, float alpha, float beta, bool pretranspose_hint, MemoryGroup &memory_group)
+void NEGEMMAssemblyDispatch<TypeInput, TypeOutput>::Fallback::configure(const ITensor *a, const ITensor *b, ITensor *d, arm_gemm::GemmArgs<TypeOutput> &args, MemoryGroup &memory_group)
 {
-    const CPUInfo &ci          = NEScheduler::get().cpu_info();
-    const int      M           = d->info()->tensor_shape().y();
-    const int      N           = d->info()->tensor_shape().x();
-    const int      K           = a->info()->tensor_shape().x();
-    const int      batches     = d->info()->tensor_shape().total_size_upper(2);
-    const int      multis      = b->info()->tensor_shape().z();
-    unsigned int   num_threads = NEScheduler::get().num_threads();
-
-    _gemm_kernel_asm = arm_gemm::gemm<TypeInput, TypeOutput>(ci, M, N, K, batches, multis, false, false, alpha, beta, num_threads, pretranspose_hint);
+    _gemm_kernel_asm = arm_gemm::gemm<TypeInput, TypeOutput>(args, nullptr);
     if(_gemm_kernel_asm == nullptr)
     {
         //configuration not supported: Leave function unconfigured:
@@ -139,11 +204,10 @@ void NEGEMMAssemblyDispatch<TypeInput, TypeOutput>::Fallback::configure(const IT
     //if we disable this code below in brackets then ConvLayer deadlocks when threads > 1 and
     //the shapes are In=1x1x1024 Weights=1x1x1024x1001 Biases=1001 Out=1x1x1001
     {
-        const unsigned int window_size = _gemm_kernel_asm->get_window_size();
-        if(window_size < num_threads)
+        const int window_size = _gemm_kernel_asm->get_window_size();
+        if(window_size < args._maxthreads)
         {
-            num_threads = window_size;
-            _gemm_kernel_asm->set_nthreads(num_threads);
+            _gemm_kernel_asm->set_nthreads(window_size);
         }
     }
 
@@ -248,8 +312,6 @@ void NEGEMMAssemblyDispatch<TypeInput, TypeOutput>::Fallback::run()
     NEScheduler::get().schedule(_optimised_kernel.get(), Window::DimX);
 }
 
-namespace arm_compute
-{
 template class NEGEMMAssemblyDispatch<float, float>;
 template class NEGEMMAssemblyDispatch<uint8_t, uint32_t>;
 template class NEGEMMAssemblyDispatch<int8_t, int32_t>;
diff --git a/src/runtime/NEON/functions/NESimpleAssemblyFunction.cpp b/src/runtime/NEON/functions/NESimpleAssemblyFunction.cpp
new file mode 100644
index 0000000000..a4b0dfffaa
--- /dev/null
+++ b/src/runtime/NEON/functions/NESimpleAssemblyFunction.cpp
@@ -0,0 +1,46 @@
+/*
+ * Copyright (c) 2018 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/runtime/NEON/functions/NESimpleAssemblyFunction.h"
+
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/runtime/NEON/NEScheduler.h"
+
+using namespace arm_compute;
+
+NESimpleAssemblyFunction::NESimpleAssemblyFunction() // NOLINT
+    : _kernel()
+{
+}
+
+void NESimpleAssemblyFunction::run()
+{
+    NEScheduler::get().schedule(_kernel.get(), Window::DimX);
+}
+
+void NESimpleAssemblyFunction::configure(std::unique_ptr<INEGEMMWrapperKernel> kernel)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(kernel.get());
+    _kernel = std::move(kernel);
+    ARM_COMPUTE_ERROR_ON_WINDOW_DIMENSIONS_GTE(_kernel->window(), 1);
+}
diff --git a/tests/validation/NEON/ConvolutionLayer.cpp b/tests/validation/NEON/ConvolutionLayer.cpp
index 1d82ff0712..ed8c64ebb4 100644
--- a/tests/validation/NEON/ConvolutionLayer.cpp
+++ b/tests/validation/NEON/ConvolutionLayer.cpp
@@ -194,7 +194,7 @@ input_shape, weights_shape, bias_shape, output_shape, info, dilation, data_type,
 }
 
 template <typename T>
-using NEGEMMConvolutionLayerFixture = ConvolutionValidationFixture<Tensor, Accessor, NEConvolutionLayer, T>;
+using NEGEMMConvolutionLayerFixture = ConvolutionValidationFixture<Tensor, Accessor, NEGEMMConvolutionLayer, T>;
 
 TEST_SUITE(Float)
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC