COMPMID-3460: Refactor NEElementwiseUnaryKernel

Removed most of the templates and refactored the code.
Performance is the same but the libary size dropped by
52Kb.

Change-Id: I41ff0c0853c923d925cdaeb05f4a58c9086fff94
Signed-off-by: Michalis Spyrou <michalis.spyrou@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3190
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

7 files changed, 97 insertions, 126 deletions

diff --git a/arm_compute/core/NEON/kernels/NEElementwiseUnaryKernel.h b/arm_compute/core/NEON/kernels/NEElementwiseUnaryKernel.h
index 2a4a8f8e46..9a41cecf19 100644
--- a/arm_compute/core/NEON/kernels/NEElementwiseUnaryKernel.h
+++ b/arm_compute/core/NEON/kernels/NEElementwiseUnaryKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2019 ARM Limited.
+ * Copyright (c) 2018-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -78,23 +78,26 @@ public:
     // Inherited methods overridden:
     void run(const Window &window, const ThreadInfo &info) override;
 
+private:
     /** Common signature for all the specialised arithmetic functions
      *
-     * @param[in]  input  An input tensor. Data types supported: F16/F32.
-     * @param[out] output The output tensor. Data types supported: F16/F32.
-     * @param[in]  window Region on which to execute the kernel.
+     * @param[in] window Region on which to execute the kernel.
      */
-    using ElementwiseUnaryFunction = void(const ITensor *input, ITensor *output, const Window &window);
+    using ElementwiseUnaryPtr = void (NEElementwiseUnaryKernel::*)(const Window &window);
 
-protected:
-    // Inherited methods overridden:
-    static Status validate_arguments(ElementWiseUnary op, const ITensorInfo &input, const ITensorInfo &output);
-
-    /** Function to use for the particular tensor types passed to configure() */
-    std::function<void(const ITensor *input, ITensor *output, const Window &window)> _function;
+    /** Template function to run elementwise unary operation
+     *
+     * @tparam ScalarType Scalar datatype
+     *
+     * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()).
+     */
+    template <typename ScalarType>
+    void elementwise_op(const Window &window);
 
-    const ITensor *_input;
-    ITensor       *_output;
+    ElementwiseUnaryPtr _func;
+    const ITensor      *_input;
+    ITensor            *_output;
+    ElementWiseUnary    _op;
 };
 } // namespace arm_compute
 #endif /* ARM_COMPUTE_NEELEMENTWISEUNARYKERNEL_H */
diff --git a/arm_compute/core/NEON/wrapper/intrinsics/exp.h b/arm_compute/core/NEON/wrapper/intrinsics/exp.h
index f079af0ae2..4b17ebd93f 100644
--- a/arm_compute/core/NEON/wrapper/intrinsics/exp.h
+++ b/arm_compute/core/NEON/wrapper/intrinsics/exp.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2019 ARM Limited.
+ * Copyright (c) 2018-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -37,10 +37,18 @@ namespace wrapper
         return vexpq_##postfix(a);     \
     }
 
+#define VEXPQ_IMPL_INT(vtype, postfix)      \
+    inline vtype vexpq(const vtype &a)      \
+    {                                       \
+        ARM_COMPUTE_UNUSED(a);              \
+        ARM_COMPUTE_ERROR("Not supported"); \
+    }
+
 VEXPQ_IMPL(float32x4_t, f32)
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 VEXPQ_IMPL(float16x8_t, f16)
 #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+VEXPQ_IMPL_INT(int32x4_t, s32)
 #undef VEXPQ_IMPL
 
 } // namespace wrapper
diff --git a/arm_compute/core/NEON/wrapper/intrinsics/invsqrt.h b/arm_compute/core/NEON/wrapper/intrinsics/invsqrt.h
index 2bf9f52dbe..77adcf7b8c 100644
--- a/arm_compute/core/NEON/wrapper/intrinsics/invsqrt.h
+++ b/arm_compute/core/NEON/wrapper/intrinsics/invsqrt.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2019 ARM Limited.
+ * Copyright (c) 2018-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -37,10 +37,18 @@ namespace wrapper
         return prefix##_##postfix(a);                \
     }
 
+#define VINVSQRT_IMPL_INT(stype, vtype, prefix, postfix) \
+    inline vtype vinvsqrt(const vtype &a)                \
+    {                                                    \
+        ARM_COMPUTE_UNUSED(a);                           \
+        ARM_COMPUTE_ERROR("Not supported");              \
+    }
+
 VINVSQRT_IMPL(float, float32x2_t, vinvsqrt, f32)
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 VINVSQRT_IMPL(float16_t, float16x4_t, vinvsqrt, f16)
 #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+VINVSQRT_IMPL_INT(int, int32x4_t, vinvsqrt, s32)
 
 VINVSQRT_IMPL(float, float32x4_t, vinvsqrtq, f32)
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
diff --git a/arm_compute/core/NEON/wrapper/intrinsics/log.h b/arm_compute/core/NEON/wrapper/intrinsics/log.h
index bb4181ec93..682830c122 100644
--- a/arm_compute/core/NEON/wrapper/intrinsics/log.h
+++ b/arm_compute/core/NEON/wrapper/intrinsics/log.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2019 ARM Limited.
+ * Copyright (c) 2018-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -37,10 +37,19 @@ namespace wrapper
         return prefix##_##postfix(a);     \
     }
 
+#define VLOG_IMPL_INT(vtype, prefix, postfix) \
+    inline vtype vlog(const vtype &a)         \
+    {                                         \
+        ARM_COMPUTE_UNUSED(a);                \
+        ARM_COMPUTE_ERROR("Not supported");   \
+    }
+
 VLOG_IMPL(float32x4_t, vlogq, f32)
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 VLOG_IMPL(float16x8_t, vlogq, f16)
 #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+VLOG_IMPL_INT(int32x4_t, vlogq, s32)
+
 #undef VLOG_IMPL
 } // namespace wrapper
 } // namespace arm_compute
diff --git a/arm_compute/core/NEON/wrapper/intrinsics/round.h b/arm_compute/core/NEON/wrapper/intrinsics/round.h
index f3e0fe1ed8..d6f5a88689 100644
--- a/arm_compute/core/NEON/wrapper/intrinsics/round.h
+++ b/arm_compute/core/NEON/wrapper/intrinsics/round.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019 ARM Limited.
+ * Copyright (c) 2019-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -37,10 +37,18 @@ namespace wrapper
         return vroundq_rte_##postfix(a); \
     }
 
+#define VROUNDQ_IMPL_INT(vtype, postfix)    \
+    inline vtype vround(const vtype &a)     \
+    {                                       \
+        ARM_COMPUTE_UNUSED(a);              \
+        ARM_COMPUTE_ERROR("Not supported"); \
+    }
+
 VROUNDQ_IMPL(float32x4_t, f32)
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 VROUNDQ_IMPL(float16x8_t, f16)
 #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+VROUNDQ_IMPL_INT(int32x4_t, s32)
 #undef VROUNDQ_IMPL
 
 } // namespace wrapper
diff --git a/arm_compute/core/NEON/wrapper/intrinsics/sin.h b/arm_compute/core/NEON/wrapper/intrinsics/sin.h
index e0fe5fbff3..bca72db38a 100644
--- a/arm_compute/core/NEON/wrapper/intrinsics/sin.h
+++ b/arm_compute/core/NEON/wrapper/intrinsics/sin.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019 ARM Limited.
+ * Copyright (c) 2019-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -37,11 +37,20 @@ namespace wrapper
         return prefix##_##postfix(a);     \
     }
 
+#define VSIN_IMPL_INT(vtype, prefix, postfix) \
+    inline vtype vsin(const vtype &a)         \
+    {                                         \
+        ARM_COMPUTE_UNUSED(a);                \
+        ARM_COMPUTE_ERROR("Not supported");   \
+    }
+
 VSIN_IMPL(float32x4_t, vsinq, f32)
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 VSIN_IMPL(float16x8_t, vsinq, f16)
 #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 
+VSIN_IMPL_INT(int32x4_t, vsinq, s32)
+
 #undef vsub_IMPL
 } // namespace wrapper
 } // namespace arm_compute
diff --git a/src/core/NEON/kernels/NEElementwiseUnaryKernel.cpp b/src/core/NEON/kernels/NEElementwiseUnaryKernel.cpp
index 5d3af3b03d..ba93c37ba1 100644
--- a/src/core/NEON/kernels/NEElementwiseUnaryKernel.cpp
+++ b/src/core/NEON/kernels/NEElementwiseUnaryKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2019 ARM Limited.
+ * Copyright (c) 2018-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -26,30 +26,17 @@
 #include "arm_compute/core/CPP/Validate.h"
 #include "arm_compute/core/Error.h"
 #include "arm_compute/core/Helpers.h"
-#include "arm_compute/core/IAccessWindow.h"
 #include "arm_compute/core/ITensor.h"
-#include "arm_compute/core/NEON/NEAsymm.h"
-#include "arm_compute/core/NEON/NEFixedPoint.h"
-#include "arm_compute/core/NEON/NEMath.h"
 #include "arm_compute/core/NEON/wrapper/wrapper.h"
-#include "arm_compute/core/TensorInfo.h"
 #include "arm_compute/core/Validate.h"
 #include "support/ToolchainSupport.h"
 
-#include <algorithm>
-#include <arm_neon.h>
-#include <cstdint>
-#include <map>
-#include <string>
-
 namespace arm_compute
 {
-class Coordinates;
-
 namespace
 {
-template <ElementWiseUnary op, typename ScalarType>
-inline ScalarType elementwise_op_scalar(const ScalarType &a)
+template <typename ScalarType>
+inline ScalarType elementwise_op_scalar_imp(ElementWiseUnary op, const ScalarType &a)
 {
     switch(op)
     {
@@ -72,9 +59,8 @@ inline ScalarType elementwise_op_scalar(const ScalarType &a)
     }
 }
 
-/* Elementwise operations that are supported for float */
-template <ElementWiseUnary op, typename ScalarType, bool is_float, typename VectorType, typename std::enable_if<is_float, int>::type = 0>
-inline VectorType elementwise_op(const VectorType &a)
+template <typename ScalarType, typename VectorType>
+inline VectorType elementwise_op_imp(ElementWiseUnary op, const VectorType &a)
 {
     switch(op)
     {
@@ -96,24 +82,10 @@ inline VectorType elementwise_op(const VectorType &a)
             ARM_COMPUTE_ERROR("NOT_SUPPORTED!");
     }
 }
+} // namespace
 
-/* Elementwise operations that are supported for non floats */
-template < ElementWiseUnary op, typename ScalarType, bool is_float, typename VectorType, typename std::enable_if < !is_float, int >::type = 0 >
-inline VectorType elementwise_op(const VectorType &a)
-{
-    switch(op)
-    {
-        case ElementWiseUnary::NEG:
-            return wrapper::vneg(a);
-        case ElementWiseUnary::ABS:
-            return wrapper::vabs(a);
-        default:
-            ARM_COMPUTE_ERROR("NOT_SUPPORTED!");
-    }
-}
-
-template <ElementWiseUnary op, typename ScalarType, bool is_float>
-void elementwise_op(const ITensor *in, ITensor *out, const Window &window)
+template <typename ScalarType>
+void NEElementwiseUnaryKernel::elementwise_op(const Window &window)
 {
     const int  window_step_x  = 16 / sizeof(ScalarType);
     const auto window_start_x = static_cast<int>(window.x().start());
@@ -122,8 +94,8 @@ void elementwise_op(const ITensor *in, ITensor *out, const Window &window)
     Window win = window;
     win.set(Window::DimX, Window::Dimension(0, 1, 1));
 
-    Iterator input(in, win);
-    Iterator output(out, win);
+    Iterator input(_input, win);
+    Iterator output(_output, win);
 
     execute_window_loop(win, [&](const Coordinates &)
     {
@@ -133,55 +105,24 @@ void elementwise_op(const ITensor *in, ITensor *out, const Window &window)
         int x = window_start_x;
         for(; x <= window_end_x - window_step_x; x += window_step_x)
         {
-            wrapper::vstore(output_ptr + x, elementwise_op<op, ScalarType, is_float>(wrapper::vloadq(input_ptr + x)));
+            wrapper::vstore(output_ptr + x, elementwise_op_imp<ScalarType>(_op, wrapper::vloadq(input_ptr + x)));
         }
         for(; x < window_end_x; ++x)
         {
-            *(output_ptr + x) = elementwise_op_scalar<op>(*(input_ptr + x));
+            *(output_ptr + x) = elementwise_op_scalar_imp(_op, *(input_ptr + x));
         }
     },
     input, output);
 }
 
-template <ElementWiseUnary op>
-std::function<void(const ITensor *input, ITensor *output, const Window &window)>
-configure_func(const ITensor *input, ITensor *output)
-{
-    std::string function_to_call("op_");
-    function_to_call += string_from_data_type(input->info()->data_type()) + "_";
-    function_to_call += string_from_data_type(output->info()->data_type());
-
-    static std::map<std::string, NEElementwiseUnaryKernel::ElementwiseUnaryFunction *> map_function =
-    {
-        { "op_F32_F32", &elementwise_op<op, float, true> },
-        { "op_S32_S32", &elementwise_op<op, int32_t, false> },
-    };
-#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-    map_function["op_F16_F16"] = &elementwise_op<op, float16_t, true>;
-#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
-
-    auto it = map_function.find(function_to_call);
-
-    if(it != map_function.end())
-    {
-        auto func = it->second;
-        return [func](const ITensor * input, ITensor * output, const Window & window)
-        {
-            func(input, output, window);
-        };
-    }
-    return nullptr;
-}
-} // namespace
-
 NEElementwiseUnaryKernel::NEElementwiseUnaryKernel()
-    : _function(nullptr), _input(nullptr), _output(nullptr)
+    : _func(nullptr), _input(nullptr), _output(nullptr), _op()
 {
 }
 
 void NEElementwiseUnaryKernel::configure(ElementWiseUnary op, const ITensor *input, ITensor *output)
 {
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(op, *input->info(), *output->info()));
+    ARM_COMPUTE_ERROR_THROW_ON(validate(op, input->info(), output->info()));
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
 
     // Configure kernel window
@@ -196,40 +137,32 @@ void NEElementwiseUnaryKernel::configure(ElementWiseUnary op, const ITensor *inp
 
     _input  = input;
     _output = output;
+    _op     = op;
 
     INEKernel::configure(win);
 
-    switch(op)
+    switch(input->info()->data_type())
     {
-        case ElementWiseUnary::RSQRT:
-            _function = configure_func<ElementWiseUnary::RSQRT>(input, output);
+        case DataType::F32:
+            _func = &NEElementwiseUnaryKernel::elementwise_op<float>;
             break;
-        case ElementWiseUnary::EXP:
-            _function = configure_func<ElementWiseUnary::EXP>(input, output);
-            break;
-        case ElementWiseUnary::NEG:
-            _function = configure_func<ElementWiseUnary::NEG>(input, output);
-            break;
-        case ElementWiseUnary::LOG:
-            _function = configure_func<ElementWiseUnary::LOG>(input, output);
-            break;
-        case ElementWiseUnary::ABS:
-            _function = configure_func<ElementWiseUnary::ABS>(input, output);
-            break;
-        case ElementWiseUnary::ROUND:
-            _function = configure_func<ElementWiseUnary::ROUND>(input, output);
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+        case DataType::F16:
+            _func = &NEElementwiseUnaryKernel::elementwise_op<float16_t>;
+#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
             break;
-        case ElementWiseUnary::SIN:
-            _function = configure_func<ElementWiseUnary::SIN>(input, output);
+        case DataType::S32:
+            _func = &NEElementwiseUnaryKernel::elementwise_op<int32_t>;
             break;
         default:
-            ARM_COMPUTE_ERROR("NOT_SUPPORTED!");
+            ARM_COMPUTE_ERROR("DataType not supported");
     }
 }
 
-Status NEElementwiseUnaryKernel::validate_arguments(ElementWiseUnary op, const ITensorInfo &input, const ITensorInfo &output)
+Status NEElementwiseUnaryKernel::validate(ElementWiseUnary op, const ITensorInfo *input, const ITensorInfo *output)
 {
-    ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(&input);
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input);
     switch(op)
     {
         case ElementWiseUnary::EXP:
@@ -237,37 +170,30 @@ Status NEElementwiseUnaryKernel::validate_arguments(ElementWiseUnary op, const I
         case ElementWiseUnary::LOG:
         case ElementWiseUnary::ROUND:
         case ElementWiseUnary::SIN:
-            ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input, 1, DataType::F16, DataType::F32);
+            ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
             break;
         case ElementWiseUnary::NEG:
         case ElementWiseUnary::ABS:
-            ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input, 1, DataType::F16, DataType::F32, DataType::S32);
+            ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32, DataType::S32);
             break;
         default:
             ARM_COMPUTE_ERROR("ElementWiseUnary operation not supported");
     }
     // Validate in case of configured output
-    if(output.total_size() > 0)
+    if(output->total_size() > 0)
     {
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&input, &output);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
     }
 
     return Status{};
 }
 
-Status NEElementwiseUnaryKernel::validate(ElementWiseUnary op, const ITensorInfo *input, const ITensorInfo *output)
-{
-    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(op, *input, *output));
-    return Status{};
-}
-
 void NEElementwiseUnaryKernel::run(const Window &window, const ThreadInfo &info)
 {
     ARM_COMPUTE_UNUSED(info);
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
-    ARM_COMPUTE_ERROR_ON(_function == nullptr);
-    _function(_input, _output, window);
+    ARM_COMPUTE_ERROR_ON(_func == nullptr);
+    (this->*_func)(window);
 }
 } // namespace arm_compute