COMPMID-1381: Cleaned up the AssemblyHelper interface

Introduced a new IFunction for when we'll fork the arm_gemm functions
Increased encapsulation and abstraction of which method is used

Change-Id: I5fd8b14b5c77e7f8ecb09029b5e2eccd10dbdcf4
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/139108
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-by: Pablo Tello <pablo.tello@arm.com>

5 files changed, 136 insertions, 25 deletions

diff --git a/arm_compute/runtime/NEON/functions/NEGEMM.h b/arm_compute/runtime/NEON/functions/NEGEMM.h
index cf059e5c4d..523f1d33a1 100644
--- a/arm_compute/runtime/NEON/functions/NEGEMM.h
+++ b/arm_compute/runtime/NEON/functions/NEGEMM.h
@@ -32,10 +32,9 @@
 #include "arm_compute/runtime/IFunction.h"
 #include "arm_compute/runtime/IMemoryManager.h"
 #include "arm_compute/runtime/MemoryGroup.h"
+#include "arm_compute/runtime/NEON/functions/NEGEMMAssemblyDispatch.h"
 #include "arm_compute/runtime/Tensor.h"
 
-#include "arm_compute/runtime/NEON/AssemblyHelper.h"
-
 #include <memory>
 
 namespace arm_compute
@@ -86,12 +85,10 @@ private:
     NEGEMMInterleave4x4Kernel  _interleave_kernel;
     NEGEMMTranspose1xWKernel   _transpose_kernel;
     NEGEMMMatrixMultiplyKernel _mm_kernel;
-    AssemblyKernelGlueF32      _asm_glue;
+    NEGEMMAssemblyDispatchF32  _asm_glue;
     NEGEMMMatrixAdditionKernel _ma_kernel;
     Tensor                     _tmp_a;
     Tensor                     _tmp_b;
-    Tensor                     _workspace;
-    Tensor                     _B_pretransposed;
     const ITensor             *_original_b;
     bool                       _run_vector_matrix_multiplication;
     bool                       _run_addition;
diff --git a/arm_compute/runtime/NEON/functions/NEGEMMAssemblyDispatch.h b/arm_compute/runtime/NEON/functions/NEGEMMAssemblyDispatch.h
new file mode 100644
index 0000000000..4ac6a3cae2
--- /dev/null
+++ b/arm_compute/runtime/NEON/functions/NEGEMMAssemblyDispatch.h
@@ -0,0 +1,120 @@
+/*
+ * 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_NEGEMMASSEMBLYDISPATCH_H__
+#define __ARM_COMPUTE_NEGEMMASSEMBLYDISPATCH_H__
+
+#include "arm_compute/core/NEON/kernels/assembly/NEGEMMAssemblyWrapperKernel.h"
+#include "arm_compute/runtime/IFunction.h"
+#include "arm_compute/runtime/IMemoryManager.h"
+#include "arm_compute/runtime/MemoryGroup.h"
+#include "arm_compute/runtime/Tensor.h"
+
+#include "arm_compute/core/NEON/kernels/assembly/arm_gemm.hpp"
+
+namespace arm_compute
+{
+/** Assembly kernel glue */
+template <typename TypeInput, typename TypeOutput>
+class NEGEMMAssemblyDispatch : public IFunction
+{
+public:
+    /** Default constructor */
+    NEGEMMAssemblyDispatch(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
+
+    /** Prevent instances of this class from being copy constructed */
+    NEGEMMAssemblyDispatch(const NEGEMMAssemblyDispatch<TypeInput, TypeOutput> &) = delete;
+    /** Prevent instances of this class from being copied */
+    NEGEMMAssemblyDispatch<TypeInput, TypeOutput> &operator=(const NEGEMMAssemblyDispatch<TypeInput, TypeOutput> &) = delete;
+    NEGEMMAssemblyDispatch(NEGEMMAssemblyDispatch<TypeInput, TypeOutput> &&) = default;
+    NEGEMMAssemblyDispatch<TypeInput, TypeOutput> &operator=(NEGEMMAssemblyDispatch<TypeInput, TypeOutput> &&) = default;
+    ~NEGEMMAssemblyDispatch() = default;
+
+private:
+    /** ACL Function */
+    std::unique_ptr<IFunction> _function;
+
+    //Fallback: use arm_gemm's AssemblyGemm:
+    class Fallback
+    {
+#ifndef DOXYGEN_SKIP_THIS
+    public:
+        /** Configures the arrays pointers and strides in the assembly kernel and executes the assembly kernel.
+         *  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 prepare();
+        bool is_configured() const;
+#endif /* DOXYGEN_SKIP_THIS */
+
+    private:
+        /** Allocate a workspace tensor.
+         *
+         * @param[in] workspace_size Size to allocate.
+         * @param[in] memory_group   Tensor memory group.
+         * @param[in] alignment      Workspace memory alignment.
+         */
+        void allocate_workspace(size_t workspace_size, MemoryGroup *memory_group, size_t alignment);
+
+        /** Assembly Gemm kernel */
+        std::unique_ptr<arm_gemm::GemmCommon<TypeInput, TypeOutput>> _gemm_kernel_asm{ nullptr };
+        /** Optimised NEON kernel */
+        std::unique_ptr<INEKernel> _optimised_kernel{ nullptr };
+        /** Input A */
+        const ITensor *_a
+        {
+            nullptr
+        };
+        /** Input B */
+        const ITensor *_b
+        {
+            nullptr
+        };
+        /** Output */
+        ITensor *_d{ nullptr };
+        /** GEMM workspace */
+        Tensor _workspace{};
+        /** Pre-transpose tensor */
+        Tensor _pretranspose{};
+        /** Prepared flag */
+        bool _is_prepared{ false };
+    } _arm_gemm;               /**< Fallback in case ACL doesn't have a function */
+    MemoryGroup _memory_group; /**< Function memory group */
+public:
+    void configure(const ITensor *a, const ITensor *b, ITensor *d, float alpha, float beta, bool pretranspose_hint);
+    bool is_configured() const;
+    // Inherited methods overridden:
+    /** Runs a preparation step, usually for pre-transposing matrix b */
+    void prepare() override;
+    void run() override;
+};
+
+/** Float 32 assembly kernel glue */
+using NEGEMMAssemblyDispatchF32 = NEGEMMAssemblyDispatch<float, float>;
+/** Uint 8 to Uint 32 kernel glue */
+using NEGEMMAssemblyDispatchU8U32 = NEGEMMAssemblyDispatch<uint8_t, uint32_t>;
+/** Int 8 to Int 32 kernel glue */
+using NEGEMMAssemblyDispatchS8S32 = NEGEMMAssemblyDispatch<int8_t, int32_t>;
+}
+#endif /* __ARM_COMPUTE_NEGEMMASSEMBLYDISPATCH_H__ */
diff --git a/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h b/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h
index 68e1145e35..1564b6c983 100644
--- a/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h
@@ -37,8 +37,8 @@
 #include "arm_compute/core/NEON/kernels/NEWeightsReshapeKernel.h"
 #include "arm_compute/core/Types.h"
 #include "arm_compute/runtime/MemoryGroup.h"
-#include "arm_compute/runtime/NEON/AssemblyHelper.h"
 #include "arm_compute/runtime/NEON/functions/NEActivationLayer.h"
+#include "arm_compute/runtime/NEON/functions/NEGEMMAssemblyDispatch.h"
 #include "arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h"
 #include "arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h"
 #include "arm_compute/runtime/Tensor.h"
@@ -168,8 +168,8 @@ private:
     void configure_mm(const ITensor *input, const ITensor *weights, ITensor *output, bool is_interleaved, const GEMMReshapeInfo &reshape_info = GEMMReshapeInfo());
 
 private:
-    AssemblyKernelGlueF32                               _asm_glue;
     MemoryGroup                                         _memory_group;
+    NEGEMMAssemblyDispatchF32                           _asm_glue;
     NEIm2ColKernel                                      _input_im2col_kernel;
     NEGEMMInterleave4x4Kernel                           _input_interleave_kernel;
     NEConvolutionLayerReshapeWeights                    _reshape_weights;
@@ -187,8 +187,6 @@ private:
     Tensor _weights_reshaped;
     Tensor _gemm_output;
     Tensor _tmp_output;
-    Tensor _workspace;
-    Tensor _B_pretransposed;
 
     DataLayout _data_layout;
     bool       _append_bias;
diff --git a/arm_compute/runtime/NEON/functions/NEGEMMLowpAssemblyMatrixMultiplyCore.h b/arm_compute/runtime/NEON/functions/NEGEMMLowpAssemblyMatrixMultiplyCore.h
index 11ca1bc313..b6672d7584 100644
--- a/arm_compute/runtime/NEON/functions/NEGEMMLowpAssemblyMatrixMultiplyCore.h
+++ b/arm_compute/runtime/NEON/functions/NEGEMMLowpAssemblyMatrixMultiplyCore.h
@@ -29,7 +29,7 @@
 #include "arm_compute/runtime/IFunction.h"
 #include "arm_compute/runtime/IMemoryManager.h"
 #include "arm_compute/runtime/MemoryGroup.h"
-#include "arm_compute/runtime/NEON/AssemblyHelper.h"
+#include "arm_compute/runtime/NEON/functions/NEGEMMAssemblyDispatch.h"
 #include "arm_compute/runtime/Tensor.h"
 
 #include <memory>
@@ -58,16 +58,14 @@ public:
     void run() override;
 
 private:
-    MemoryGroup                _memory_group;
-    AssemblyKernelGlueU8U32    _asm_glue_unsigned;
-    AssemblyKernelGlueS8S32    _asm_glue_signed;
-    std::unique_ptr<INEKernel> _mm_kernel;
-    std::unique_ptr<INEKernel> _mtx_a_reshape_kernel;
-    std::unique_ptr<INEKernel> _mtx_b_reshape_kernel;
-    Tensor                     _tmp_a;
-    Tensor                     _tmp_b;
-    Tensor                     _workspace;
-    Tensor                     _B_pretransposed;
+    MemoryGroup                 _memory_group;
+    NEGEMMAssemblyDispatchU8U32 _asm_glue_unsigned;
+    NEGEMMAssemblyDispatchS8S32 _asm_glue_signed;
+    std::unique_ptr<INEKernel>  _mm_kernel;
+    std::unique_ptr<INEKernel>  _mtx_a_reshape_kernel;
+    std::unique_ptr<INEKernel>  _mtx_b_reshape_kernel;
+    Tensor                      _tmp_a;
+    Tensor                      _tmp_b;
 };
 }
 #endif /*__ARM_COMPUTE_NEGEMMLOWPASSEMBLYMATRIXMULTIPLYCORE_H__ */
diff --git a/arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h b/arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h
index f32eb3c757..96ac7bb7e0 100644
--- a/arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h
+++ b/arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h
@@ -30,7 +30,7 @@
 #include "arm_compute/runtime/IFunction.h"
 #include "arm_compute/runtime/IMemoryManager.h"
 #include "arm_compute/runtime/MemoryGroup.h"
-#include "arm_compute/runtime/NEON/AssemblyHelper.h"
+#include "arm_compute/runtime/NEON/functions/NEGEMMAssemblyDispatch.h"
 #include "arm_compute/runtime/Tensor.h"
 
 #include <memory>
@@ -98,8 +98,8 @@ public:
 
 private:
     MemoryGroup                        _memory_group;
-    AssemblyKernelGlueU8U32            _asm_glue_unsigned;
-    AssemblyKernelGlueS8S32            _asm_glue_signed;
+    NEGEMMAssemblyDispatchU8U32        _asm_glue_unsigned;
+    NEGEMMAssemblyDispatchS8S32        _asm_glue_signed;
     std::unique_ptr<INEKernel>         _mm_kernel;
     std::unique_ptr<INEKernel>         _mtx_a_reshape_kernel;
     std::unique_ptr<INEKernel>         _mtx_b_reshape_kernel;
@@ -110,8 +110,6 @@ private:
     Tensor                             _vector_sum_row;
     Tensor                             _tmp_a;
     Tensor                             _tmp_b;
-    Tensor                             _workspace;
-    Tensor                             _B_pretranspose;
     const ITensor                     *_original_b;
     int32_t                            _a_offset;
     int32_t                            _b_offset;