Move NEGEMMAssemblyDispatch as an internal function

Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Change-Id: I89ee26c1595d510c5048904cae9422528b76cd45
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4662
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

6 files changed, 160 insertions, 28 deletions

diff --git a/src/runtime/NEON/functions/NEGEMM.cpp b/src/runtime/NEON/functions/NEGEMM.cpp
index 03f5aa37c1..6d83480cb9 100644
--- a/src/runtime/NEON/functions/NEGEMM.cpp
+++ b/src/runtime/NEON/functions/NEGEMM.cpp
@@ -31,7 +31,6 @@
 #include "arm_compute/core/Validate.h"
 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
 #include "arm_compute/runtime/NEON/NEScheduler.h"
-#include "arm_compute/runtime/NEON/functions/NEGEMMAssemblyDispatch.h"
 #include "arm_compute/runtime/TensorAllocator.h"
 #include "src/core/CPP/Validate.h"
 #include "src/core/NEON/kernels/NEGEMMInterleave4x4Kernel.h"
@@ -39,6 +38,7 @@
 #include "src/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.h"
 #include "src/core/NEON/kernels/NEGEMMTranspose1xWKernel.h"
 #include "src/core/helpers/AutoConfiguration.h"
+#include "src/runtime/NEON/functions/NEGEMMAssemblyDispatch.h"
 
 #include <cmath>
 
@@ -61,7 +61,7 @@ AsmGemmInfo init_assembly_metadata(const GEMMInfo &info)
 } // namespace
 
 NEGEMM::NEGEMM(std::shared_ptr<IMemoryManager> memory_manager, IWeightsManager *weights_manager)
-    : _memory_group(memory_manager), _weights_manager(weights_manager), _interleave_kernel(), _transpose_kernel(), _mm_kernel(), _asm_glue(memory_manager, weights_manager), _ma_kernel(),
+    : _memory_group(memory_manager), _weights_manager(weights_manager), _interleave_kernel(), _transpose_kernel(), _mm_kernel(), _asm_glue(std::make_unique<NEGEMMAssemblyDispatch>()), _ma_kernel(),
       _alpha_scale_func(nullptr), _add_bias(), _activation_func(), _tmp_a(), _tmp_b(), _tmp_d(), _original_b(nullptr), _run_vector_matrix_multiplication(false), _run_alpha_scale(false),
       _run_addition(false), _run_bias_addition(false), _run_activation(false), _reshape_b_only_on_first_run(false), _is_prepared(false)
 {
@@ -90,8 +90,8 @@ void NEGEMM::configure(const ITensor *a, const ITensor *b, const ITensor *c, ITe
     if(run_optimised)
     {
         const ITensor *c_to_use = is_c_bias ? c : nullptr;
-        _asm_glue.configure(a, b, c_to_use, d, asm_info);
-        ARM_COMPUTE_ERROR_ON(!_asm_glue.is_configured());
+        _asm_glue->configure(a, b, c_to_use, d, asm_info);
+        ARM_COMPUTE_ERROR_ON(!_asm_glue->is_configured());
 
         // Scale product by alpha
         if(_run_alpha_scale)
@@ -312,9 +312,9 @@ void NEGEMM::run()
 
     MemoryGroupResourceScope scope_mg(_memory_group);
 
-    if(_asm_glue.is_configured())
+    if(_asm_glue->is_configured())
     {
-        _asm_glue.run();
+        _asm_glue->run();
         if(_run_alpha_scale)
         {
             _alpha_scale_func.run();
@@ -361,20 +361,20 @@ void NEGEMM::prepare()
     if(!_is_prepared)
     {
         const bool original_b_managed_by_weights_manager = _weights_manager && _weights_manager->are_weights_managed(_original_b);
-        if(_asm_glue.is_configured())
+        if(_asm_glue->is_configured())
         {
             if(!original_b_managed_by_weights_manager)
             {
                 ARM_COMPUTE_ERROR_ON(!_original_b->is_used());
             }
 
-            _asm_glue.prepare();
+            _asm_glue->prepare();
             if(!original_b_managed_by_weights_manager)
             {
                 _original_b->mark_as_unused();
             }
         }
-        else if(_reshape_b_only_on_first_run && !_run_vector_matrix_multiplication && !_asm_glue.is_configured())
+        else if(_reshape_b_only_on_first_run && !_run_vector_matrix_multiplication && !_asm_glue->is_configured())
         {
             if(!original_b_managed_by_weights_manager)
             {
diff --git a/src/runtime/NEON/functions/NEGEMMAssemblyDispatch.cpp b/src/runtime/NEON/functions/NEGEMMAssemblyDispatch.cpp
index b54389cf5f..1c86393406 100644
--- a/src/runtime/NEON/functions/NEGEMMAssemblyDispatch.cpp
+++ b/src/runtime/NEON/functions/NEGEMMAssemblyDispatch.cpp
@@ -21,7 +21,7 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#include "arm_compute/runtime/NEON/functions/NEGEMMAssemblyDispatch.h"
+#include "src/runtime/NEON/functions/NEGEMMAssemblyDispatch.h"
 
 #include "arm_compute/runtime/NEON/NEScheduler.h"
 #include "src/core/CPP/Validate.h"
diff --git a/src/runtime/NEON/functions/NEGEMMAssemblyDispatch.h b/src/runtime/NEON/functions/NEGEMMAssemblyDispatch.h
new file mode 100644
index 0000000000..466e60183a
--- /dev/null
+++ b/src/runtime/NEON/functions/NEGEMMAssemblyDispatch.h
@@ -0,0 +1,125 @@
+/*
+ * Copyright (c) 2018-2020 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 SRC_NEGEMMASSEMBLYDISPATCH_H
+#define SRC_NEGEMMASSEMBLYDISPATCH_H
+
+#include "arm_compute/runtime/IFunction.h"
+#include "arm_compute/runtime/IMemoryManager.h"
+#include "arm_compute/runtime/IWeightsManager.h"
+#include "arm_compute/runtime/MemoryGroup.h"
+#include "arm_compute/runtime/Tensor.h"
+
+namespace arm_compute
+{
+/* Convolution method supported by the assembly gemm interface */
+enum class AsmConvMethod
+{
+    Im2Col,
+    Indirect,
+    Conv
+};
+
+struct AsmGemmInfo
+{
+    AsmConvMethod           method{ AsmConvMethod::Im2Col };
+    PadStrideInfo           ps_info{};
+    ActivationLayerInfo     activation_info{};
+    GEMMLowpOutputStageInfo output_stage{};
+    bool                    negated_offsets{ true };
+    bool                    reinterpret_input_as_3d{ false };
+    bool                    depth_output_gemm3d{ false };
+    int64_t                 padding_top{ 0 };
+    int64_t                 padding_left{ 0 };
+    float                   padding_value{ 0.f };
+};
+
+/** Assembly kernel glue */
+class NEGEMMAssemblyDispatch : public IFunction
+{
+public:
+    /** Constructor */
+    NEGEMMAssemblyDispatch(std::shared_ptr<IMemoryManager> memory_manager = nullptr, IWeightsManager *weights_manager = nullptr);
+    /** Prevent instances of this class from being copy constructed */
+    NEGEMMAssemblyDispatch(const NEGEMMAssemblyDispatch &) = delete;
+    /** Prevent instances of this class from being copied */
+    NEGEMMAssemblyDispatch &operator=(const NEGEMMAssemblyDispatch &) = delete;
+    NEGEMMAssemblyDispatch(NEGEMMAssemblyDispatch &&)                 = default;
+    NEGEMMAssemblyDispatch &operator=(NEGEMMAssemblyDispatch &&) = default;
+    ~NEGEMMAssemblyDispatch()                                    = default;
+
+    class IFallback
+    {
+    public:
+        virtual void run()                 = 0;
+        virtual void prepare()             = 0;
+        virtual bool is_configured() const = 0;
+        virtual ~IFallback()               = default;
+    };
+
+public:
+    /** If supported create a Compute Library function else fallback to the arm_gemm function.
+     *
+     * @param[in]  a    Input tensor (Matrix A)
+     * @param[in]  b    Input tensor (Matrix B)
+     * @param[in]  c    Input tensor (Matrix C) used to pass the bias for quantized calculations
+     * @param[out] d    Output tensor to store the result of matrix multiplication. Data type supported: same as @p input0.
+     * @param[in]  info GEMM meta-data
+     */
+    void configure(const ITensor *a, const ITensor *b, const ITensor *c, ITensor *d, const AsmGemmInfo &info);
+
+    /** Indicates whether or not this function can be used to process the given parameters.
+     *
+     * @param[in] a    Input tensor info (Matrix A)
+     * @param[in] b    Input tensor info (Matrix B)
+     * @param[in] c    Input tensor info (Matrix C) used to pass the bias for quantized calculations
+     * @param[in] d    Output tensor to store the result of matrix multiplication. Data type supported: same as @p input0.
+     * @param[in] info GEMM meta-data
+     *
+     * @return a status.
+     */
+    static Status validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *d, const AsmGemmInfo &info);
+    /** Checks if activation is supported by the gemm assembly dispatcher
+     *
+     * @param[in] activation Activation to check
+     *
+     * @return True if activation is supported else false
+     */
+    static bool is_activation_supported(const ActivationLayerInfo &activation);
+    /** Was the function successfully configured ?
+     *
+     * @return True if the function is configured and ready to run
+     */
+    bool is_configured() const;
+
+    // Inherited methods overridden:
+    void prepare() override;
+    void run() override;
+
+private:
+    std::unique_ptr<IFallback> _arm_gemm;        /** Interface for the arm_gemm fallback */
+    MemoryGroup                _memory_group;    /**< Function memory group */
+    IWeightsManager           *_weights_manager; /**< Pointer to the weights manager */
+};
+} // namespace arm_compute
+#endif /* SRC_NEGEMMASSEMBLYDISPATCH_H */
diff --git a/src/runtime/NEON/functions/NEGEMMConv2d.cpp b/src/runtime/NEON/functions/NEGEMMConv2d.cpp
index 860b6bb4e1..b8349d98db 100644
--- a/src/runtime/NEON/functions/NEGEMMConv2d.cpp
+++ b/src/runtime/NEON/functions/NEGEMMConv2d.cpp
@@ -22,9 +22,11 @@
  * SOFTWARE.
  */
 #include "arm_compute/runtime/NEON/functions/NEGEMMConv2d.h"
+
 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
 #include "arm_compute/core/utils/quantization/AsymmHelpers.h"
 #include "arm_compute/runtime/NEON/NEScheduler.h"
+#include "src/runtime/NEON/functions/NEGEMMAssemblyDispatch.h"
 
 #include <set>
 
@@ -81,9 +83,13 @@ AsmGemmInfo init_assembly_metadata(const Conv2dInfo &info, bool is_indirect)
 } // namespace
 
 NEGEMMConv2d::NEGEMMConv2d(const std::shared_ptr<IMemoryManager> &memory_manager)
-    : _gemm_asm_func(memory_manager), _activation_func(), _weights_permute_func(), _original_weights(nullptr), _permuted_weights(), _is_prepared(false), _run_activation(false)
+    : _gemm_asm_func(std::make_unique<NEGEMMAssemblyDispatch>(memory_manager)), _activation_func(), _weights_permute_func(), _original_weights(nullptr), _permuted_weights(), _is_prepared(false),
+      _run_activation(false)
 {
 }
+
+NEGEMMConv2d::~NEGEMMConv2d() = default;
+
 void NEGEMMConv2d::configure(ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const Conv2dInfo &info)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
@@ -101,10 +107,10 @@ void NEGEMMConv2d::configure(ITensor *input, const ITensor *weights, const ITens
     {
         asm_info.output_stage = calculate_output_stage_metadata(input->info(), weights->info(), output->info(), info.act_info);
     }
-    _gemm_asm_func.configure(input, &_permuted_weights, biases, output, asm_info);
+    _gemm_asm_func->configure(input, &_permuted_weights, biases, output, asm_info);
 
     // Configure activation
-    if(info.act_info.enabled() && !_gemm_asm_func.is_activation_supported(info.act_info))
+    if(info.act_info.enabled() && !_gemm_asm_func->is_activation_supported(info.act_info))
     {
         _activation_func.configure(output, nullptr, info.act_info);
         _run_activation = true;
@@ -150,7 +156,7 @@ void NEGEMMConv2d::run()
 {
     prepare();
 
-    _gemm_asm_func.run();
+    _gemm_asm_func->run();
     if(_run_activation)
     {
         _activation_func.run();
diff --git a/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp b/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp
index 50c7fe4c66..921626f0fe 100644
--- a/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp
+++ b/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp
@@ -42,6 +42,7 @@
 #include "src/core/NEON/kernels/NEGEMMLowpOffsetContributionOutputStageKernel.h"
 #include "src/core/NEON/kernels/NEGEMMLowpReductionKernel.h"
 #include "src/core/NEON/kernels/NEGEMMTranspose1xWKernel.h"
+#include "src/runtime/NEON/functions/NEGEMMAssemblyDispatch.h"
 
 namespace arm_compute
 {
@@ -65,10 +66,10 @@ using namespace arm_compute::misc::shape_calculator;
 NEGEMMLowpMatrixMultiplyCore::~NEGEMMLowpMatrixMultiplyCore() = default;
 
 NEGEMMLowpMatrixMultiplyCore::NEGEMMLowpMatrixMultiplyCore(std::shared_ptr<IMemoryManager> memory_manager, IWeightsManager *weights_manager)
-    : _memory_group(memory_manager), _weights_manager(weights_manager), _asm_glue(memory_manager, weights_manager), _mm_kernel(), _mtx_a_reshape_kernel(), _mtx_b_reshape_kernel(),
-      _mtx_a_reduction_kernel(), _mtx_b_reduction_kernel(), _offset_contribution_kernel(), _offset_contribution_output_stage_kernel(), _activation_func(), _convert_to_signed_asymm(),
-      _convert_from_signed_asymm(), _vector_sum_col(), _vector_sum_row(), _tmp_a(), _tmp_b(), _mm_result_s32(), _signed_a(), _signed_output(), _original_b(nullptr), _a_offset(0), _b_offset(0),
-      _run_vector_matrix_multiplication(false), _assembly_path(false), _fused_assembly_path(false), _reshape_b_only_on_first_run(false), _is_prepared(false), _fuse_output_stage(false),
+    : _memory_group(memory_manager), _weights_manager(weights_manager), _asm_glue(std::make_unique<NEGEMMAssemblyDispatch>(memory_manager, weights_manager)), _mm_kernel(), _mtx_a_reshape_kernel(),
+      _mtx_b_reshape_kernel(), _mtx_a_reduction_kernel(), _mtx_b_reduction_kernel(), _offset_contribution_kernel(), _offset_contribution_output_stage_kernel(), _activation_func(),
+      _convert_to_signed_asymm(), _convert_from_signed_asymm(), _vector_sum_col(), _vector_sum_row(), _tmp_a(), _tmp_b(), _mm_result_s32(), _signed_a(), _signed_output(), _original_b(nullptr), _a_offset(0),
+      _b_offset(0), _run_vector_matrix_multiplication(false), _assembly_path(false), _fused_assembly_path(false), _reshape_b_only_on_first_run(false), _is_prepared(false), _fuse_output_stage(false),
       _run_activation(false), _flip_signedness(false)
 {
 }
@@ -145,14 +146,14 @@ void NEGEMMLowpMatrixMultiplyCore::configure(const ITensor *a, const ITensor *b,
         {
             if(is_data_type_quantized_asymmetric(a_to_use->info()->data_type()) && info.gemmlowp_output_stage().type == GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT)
             {
-                _asm_glue.configure(a_to_use, b, c, output, asm_info);
-                _fused_assembly_path = _asm_glue.is_configured();
+                _asm_glue->configure(a_to_use, b, c, output, asm_info);
+                _fused_assembly_path = _asm_glue->is_configured();
             }
             else
             {
-                _asm_glue.configure(a_to_use, b, nullptr, _fuse_output_stage ? &_mm_result_s32 : output, asm_info);
+                _asm_glue->configure(a_to_use, b, nullptr, _fuse_output_stage ? &_mm_result_s32 : output, asm_info);
             }
-            _assembly_path = _asm_glue.is_configured();
+            _assembly_path = _asm_glue->is_configured();
             break;
         }
         default:
@@ -510,9 +511,9 @@ void NEGEMMLowpMatrixMultiplyCore::run()
     }
 
     // Run GEMM
-    if(_asm_glue.is_configured())
+    if(_asm_glue->is_configured())
     {
-        _asm_glue.run();
+        _asm_glue->run();
     }
     else
     {
@@ -575,21 +576,21 @@ void NEGEMMLowpMatrixMultiplyCore::prepare()
     {
         const bool original_b_managed_by_weights_manager = _weights_manager && _weights_manager->are_weights_managed(_original_b);
         // Run assembly reshape
-        if(_asm_glue.is_configured())
+        if(_asm_glue->is_configured())
         {
             if(!original_b_managed_by_weights_manager)
             {
                 ARM_COMPUTE_ERROR_ON(!_original_b->is_used());
             }
 
-            _asm_glue.prepare();
+            _asm_glue->prepare();
             if(!original_b_managed_by_weights_manager)
             {
                 _original_b->mark_as_unused();
             }
         }
         // Run non-assembly reshape
-        else if(_reshape_b_only_on_first_run && !_run_vector_matrix_multiplication && !_asm_glue.is_configured())
+        else if(_reshape_b_only_on_first_run && !_run_vector_matrix_multiplication && !_asm_glue->is_configured())
         {
             if(!original_b_managed_by_weights_manager)
             {
diff --git a/src/runtime/NEON/functions/NEWinogradConvolutionLayer.cpp b/src/runtime/NEON/functions/NEWinogradConvolutionLayer.cpp
index 265df9246f..bd3bdd6a26 100644
--- a/src/runtime/NEON/functions/NEWinogradConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEWinogradConvolutionLayer.cpp
@@ -28,13 +28,13 @@
 #include "arm_compute/core/Validate.h"
 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
 #include "arm_compute/runtime/NEON/NEScheduler.h"
-#include "arm_compute/runtime/NEON/functions/NEGEMMAssemblyDispatch.h"
 #include "src/core/CPP/Validate.h"
 #include "src/core/NEON/kernels/NEGEMMInterleave4x4Kernel.h"
 #include "src/core/NEON/kernels/NEGEMMMatrixAdditionKernel.h"
 #include "src/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.h"
 #include "src/core/NEON/kernels/NEGEMMTranspose1xWKernel.h"
 #include "src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.h"
+#include "src/runtime/NEON/functions/NEGEMMAssemblyDispatch.h"
 
 #include "src/core/NEON/kernels/convolution/common/utils.hpp"
 #include "src/core/NEON/kernels/convolution/winograd/winograd.hpp"