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>

7 files changed, 262 insertions, 8 deletions

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__ */