Improve selection speed of CPU implementations

CPU micro-kernel to be used was picked during kernel execution.
Move selection during configuration to reduce runtime overhead.

Standardize kernel names as follows:
<simd_tech>_<data_type>_<data_layout>_<kernel_name>
e.g. sve_fp32_nhwc_scale

Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Change-Id: I544f1c08c8fef0f130a3bde61882ccb9a1f47f21
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5855
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 19 insertions, 13 deletions

diff --git a/src/core/cpu/kernels/CpuSoftmaxKernel.h b/src/core/cpu/kernels/CpuSoftmaxKernel.h
index aa10467965..2912098c30 100644
--- a/src/core/cpu/kernels/CpuSoftmaxKernel.h
+++ b/src/core/cpu/kernels/CpuSoftmaxKernel.h
@@ -38,7 +38,7 @@ class CpuLogits1DMaxKernel : public ICpuKernel
 {
 public:
     /** Constructor */
-    CpuLogits1DMaxKernel();
+    CpuLogits1DMaxKernel() = default;
     ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(CpuLogits1DMaxKernel);
     /** Set the input and output tensors.
      *
@@ -46,10 +46,9 @@ public:
      * @param[out] dst Destination tensor info. Data types supported: same as @p input
      */
     void configure(const ITensorInfo *src, ITensorInfo *dst);
-    /** Static function to check if given info will lead to a valid configuration of @ref CpuLogits1DMaxKernel
+    /** Static function to check if given info will lead to a valid configuration
      *
-     * @param[in] src Source tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
-     * @param[in] dst Destination tensor info. Data types supported: same as @p input
+     * Similar to CpuLogits1DMaxKernel::configure()
      *
      * @return a status
      */
@@ -58,6 +57,13 @@ public:
     // Inherited methods overridden:
     void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override;
     const char *name() const override;
+
+private:
+    using SoftmaxLogits1DMaxKernelPtr = std::add_pointer<void(const ITensor *, ITensor *, const Window &)>::type;
+
+private:
+    SoftmaxLogits1DMaxKernelPtr _run_method{ nullptr };
+    std::string                 _name{};
 };
 
 /** Interface for softmax computation for QASYMM8 with pre-computed max. */
@@ -66,7 +72,7 @@ class CpuLogits1DSoftmaxKernel : public ICpuKernel
 {
 public:
     /** Default constructor */
-    CpuLogits1DSoftmaxKernel();
+    CpuLogits1DSoftmaxKernel() = default;
     ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(CpuLogits1DSoftmaxKernel);
 
     /** Set the input and output tensors.
@@ -80,14 +86,9 @@ public:
      * @param      tmp    Auxiliary tensor info. Must be type F32 and same shape as the input.
      */
     void configure(const ITensorInfo *src, const ITensorInfo *max, ITensorInfo *dst, const float beta, ITensorInfo *tmp);
-    /** Static function to check if given info will lead to a valid configuration of @ref CpuLogits1DSoftmaxKernel
+    /** Static function to check if given info will lead to a valid configuration
      *
-     * @param[in] src  Source tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
-     * @param[in] max  Max values tensor info. Same shape as input with dimension 0 set to 1.
-     *                 Data types supported: same as @p input.
-     * @param[in] dst  Destination tensor info. Data types supported: same as @p input.
-     * @param[in] beta A scaling factor for the exponent.
-     * @param[in] tmp  Tensor info of auxiliary. Must be type F32 and same shape as the input.
+     * Similar to CpuLogits1DSoftmaxKernel::configure()
      *
      * @return a status
      */
@@ -99,7 +100,12 @@ public:
     const char *name() const override;
 
 private:
-    float _beta;
+    using SoftmaxLogits1DKernelPtr = std::add_pointer<void(const ITensor *, const ITensor *, void *const, ITensor *, float, bool, const Window &)>::type;
+
+private:
+    float                    _beta{ 1.0f };
+    SoftmaxLogits1DKernelPtr _run_method{ nullptr };
+    std::string              _name{};
 };
 } // namespace kernels
 } // namespace cpu