COMPMID-3510 [Interface change] Fix definition of "axis" in NESoftmaxLayer and CLSoftmaxLayer

* [Interface change] "axis" argument is renamed to "reduce_end_axis"

* Unify the meaning of "axis"(now "reduce_end_axis") to be the last axis
  of the first n dimensions (inclusive)to reduce.
  This way the meaning of reduce_end_axis stays the same for both
  positive and negative values: it selects a dimension before which all
  dimensions (including the selected dimension) are reduced.

Change-Id: I4ab03bd8360b1cd8cac4998df0b1571064a9d4ed
Signed-off-by: SiCong Li <sicong.li@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3278
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

4 files changed, 92 insertions, 63 deletions

diff --git a/arm_compute/core/Helpers.h b/arm_compute/core/Helpers.h
index 09c672ecfa..8f1426a56e 100644
--- a/arm_compute/core/Helpers.h
+++ b/arm_compute/core/Helpers.h
@@ -801,6 +801,19 @@ inline T wrap_around(T x, T m)
     return x >= 0 ? x % m : (x % m + m) % m;
 }
 
+/** Convert a dimension axis to the number of dimensions in the range [0, @p dim_axis]
+ * Handle negative axis, negative axis is used to specify axis from the end (e.g. -1 for the last axis).
+ *
+ * @param[in] dim_axis The last axis (inclusive) in the range [0, @p dim_axis]
+ * @param[in] num_dims The total number of dimensions
+ *
+ * @return The number of dimensions in the range [0, @p dim_axis]
+ */
+inline size_t dim_index_2_num_dims(int32_t dim_axis, int32_t num_dims)
+{
+    return static_cast<size_t>(wrap_around(dim_axis, num_dims)) + 1;
+}
+
 /** Convert negative coordinates to positive in the range [0, num_dims_input]
  *
  * @param[out] coords    Array of coordinates to be converted.
diff --git a/arm_compute/runtime/CL/functions/CLSoftmaxLayer.h b/arm_compute/runtime/CL/functions/CLSoftmaxLayer.h
index fadbc430e6..231a56f712 100644
--- a/arm_compute/runtime/CL/functions/CLSoftmaxLayer.h
+++ b/arm_compute/runtime/CL/functions/CLSoftmaxLayer.h
@@ -50,6 +50,10 @@ class ICLTensor;
  * -# @ref CLLogits1DMaxKernel
  * -# @ref CLLogits1DShiftExpSumKernel
  * -# @ref CLLogits1DNormKernel
+ * And if the reduce_end_axis is not 0, the function will use one of the the following kernels to reshape the input and
+ * perform softmax on the reshaped input:
+ * -# @ref CLFlattenLayerKernel
+ * -# @ref CLReshapeLayerKernel
  */
 template <bool IS_LOG = false>
 class CLSoftmaxLayerGeneric : public IFunction
@@ -59,36 +63,39 @@ public:
     CLSoftmaxLayerGeneric(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
     /** Set the input and output tensors.
      *
-     * @param[in]  input  Source tensor. Data types supported: QASYMM8/F16/F32
-     * @param[out] output Destination tensor. Data types supported: same as @p input
-     * @param[in]  beta   (Optional) A scaling factor for the exponent. Defaults to 1.f
-     * @param[in]  axis   (Optional) Reduction axis. It has the purpose of squashing the first @p axis
-     *                    dimensions together. For instance, given a [4x4x4x4] image,
-     *                    when @p axis is 2, the Softmax reduction will be applied on each of the [4x4] planes of the input image.
+     * @param[in]  input           Source tensor. Data types supported: QASYMM8/F16/F32
+     * @param[out] output          Destination tensor. Data types supported: same as @p input
+     * @param[in]  beta            (Optional) A scaling factor for the exponent. Defaults to 1.f
+     * @param[in]  reduce_end_axis (Optional) The last axis of the first n dimensions (inclusive)to reduce. Defaults to 0.
+     *                   It has the purpose of squashing together the first n dimensions till (including) the @p reduce_end_axis. For instance, given a [2x3x4x5] image,
+     *                   when @p reduce_end_axis is 1, the reduction will be applied to axes 0 and 1, and the Softmax op will be applied on each of the [2x3] planes of the input image.
+     *                   Must be in range [0, input_num_dimensions).
      */
-    void configure(const ICLTensor *input, ICLTensor *output, float beta = 1.0f, size_t axis = 1);
+    void configure(const ICLTensor *input, ICLTensor *output, float beta = 1.0f, size_t reduce_end_axis = 0);
     /** Set the input and output tensors.
      *
      * @param[in]  compile_context The compile context to be used.
      * @param[in]  input           Source tensor. Data types supported: QASYMM8/F16/F32
      * @param[out] output          Destination tensor. Data types supported: same as @p input
      * @param[in]  beta            (Optional) A scaling factor for the exponent. Defaults to 1.f
-     * @param[in]  axis            (Optional) Reduction axis. It has the purpose of squashing the first @p axis
-     *                    dimensions together. For instance, given a [4x4x4x4] image,
-     *                    when @p axis is 2, the Softmax reduction will be applied on each of the [4x4] planes of the input image.
+     * @param[in]  reduce_end_axis (Optional) The last axis of the first n dimensions (inclusive)to reduce. Defaults to 0.
+     *                   It has the purpose of squashing together the first n dimensions till (including) the @p reduce_end_axis. For instance, given a [2x3x4x5] image,
+     *                   when @p reduce_end_axis is 1, the reduction will be applied to axes 0 and 1, and the Softmax op will be applied on each of the [2x3] planes of the input image.
+     *                   Must be in range [0, input_num_dimensions).
      */
-    void configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, float beta = 1.0f, size_t axis = 1);
+    void configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, float beta = 1.0f, size_t reduce_end_axis = 0);
     /** Static function to check if given info will lead to a valid configuration of @ref CLSoftmaxLayer
      *
-     * @param[in] input  Source tensor. Data types supported: QASYMM8/F16/F32
-     * @param[in] output Destination tensor. Data types supported: same as @p input
-     * @param[in] beta   (Optional) A scaling factor for the exponent. Defaults to 1.f
-     * @param[in] axis   (Optional) Reduction axis. It has the purpose of squashing the first @p axis
-     *                    dimensions together. For instance, given a [4x4x4x4] image,
-     *                    when @p axis is 2, the Softmax reduction will be applied on each of the [4x4] planes of the input image.
+     * @param[in] input           Source tensor. Data types supported: QASYMM8/F16/F32
+     * @param[in] output          Destination tensor. Data types supported: same as @p input
+     * @param[in] beta            (Optional) A scaling factor for the exponent. Defaults to 1.f
+     * @param[in] reduce_end_axis (Optional) The last axis of the first n dimensions (inclusive)to reduce. Defaults to 0.
+     *                   It has the purpose of squashing together the first n dimensions till (including) the @p reduce_end_axis. For instance, given a [2x3x4x5] image,
+     *                   when @p reduce_end_axis is 1, the reduction will be applied to axes 0 and 1, and the Softmax op will be applied on each of the [2x3] planes of the input image.
+     *                   Must be in range [0, input_num_dimensions).
      * @return a status
      */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *output, float beta = 1.0f, size_t axis = 1);
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, float beta = 1.0f, size_t reduce_end_axis = 0);
 
     // Inherited methods overridden:
     void run() override;
@@ -101,13 +108,14 @@ private:
      * it initializes the kernel @p _flatten_kernel and the tensors @p _input_flat and
      * @p _output_flat
      *
-     * @param[in] input  Original source tensor.
-     * @param[in] output Original destination tensor.
-     * @param[in] axis   (Optional) Reduction axis. It has the purpose of squashing the first @p axis
-     *                    dimensions together. For instance, given a [4x4x4x4] image,
-     *                    when @p axis is 2, the Softmax reduction will be applied on each of the [4x4] planes of the input image.
+     * @param[in] input           Original source tensor.
+     * @param[in] output          Original destination tensor.
+     * @param[in] reduce_end_axis (Optional) The last axis of the first n dimensions (inclusive)to reduce. Defaults to 0.
+     *                   It has the purpose of squashing together the first n dimensions till (including) the @p reduce_end_axis. For instance, given a [2x3x4x5] image,
+     *                   when @p reduce_end_axis is 1, the reduction will be applied to axes 0 and 1, and the Softmax op will be applied on each of the [2x3] planes of the input image.
+     *                   Must be in range [0, input_num_dimensions).
      */
-    void configure_reshape_input_kernel(const ICLTensor *input, const ICLTensor *output, size_t axis);
+    void configure_reshape_input_kernel(const ICLTensor *input, const ICLTensor *output, size_t reduce_end_axis);
     /** Utility method to configure the kernels needed to flatten the input
      * tensor.
      *
@@ -118,11 +126,12 @@ private:
      * @param[in] compile_context The compile context to be used.
      * @param[in] input           Original source tensor.
      * @param[in] output          Original destination tensor.
-     * @param[in] axis            (Optional) Reduction axis. It has the purpose of squashing the first @p axis
-     *                    dimensions together. For instance, given a [4x4x4x4] image,
-     *                    when @p axis is 2, the Softmax reduction will be applied on each of the [4x4] planes of the input image.
+     * @param[in] reduce_end_axis (Optional) The last axis of the first n dimensions (inclusive)to reduce. Defaults to 0.
+     *                   It has the purpose of squashing together the first n dimensions till (including) the @p reduce_end_axis. For instance, given a [2x3x4x5] image,
+     *                   when @p reduce_end_axis is 1, the reduction will be applied to axes 0 and 1, and the Softmax op will be applied on each of the [2x3] planes of the input image.
+     *                   Must be in range [0, input_num_dimensions).
      */
-    void configure_reshape_input_kernel(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *output, size_t axis);
+    void configure_reshape_input_kernel(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *output, size_t reduce_end_axis);
 
     MemoryGroup                    _memory_group;
     CLLogits1DMaxShiftExpSumKernel _max_shift_exp_sum_kernel;
diff --git a/arm_compute/runtime/GLES_COMPUTE/functions/GCSoftmaxLayer.h b/arm_compute/runtime/GLES_COMPUTE/functions/GCSoftmaxLayer.h
index 33faae5e06..e29322c052 100644
--- a/arm_compute/runtime/GLES_COMPUTE/functions/GCSoftmaxLayer.h
+++ b/arm_compute/runtime/GLES_COMPUTE/functions/GCSoftmaxLayer.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2019 ARM Limited.
+ * Copyright (c) 2017-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -50,16 +50,17 @@ public:
     GCSoftmaxLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
     /** Set the input and output tensors.
      *
-     * @param[in]  input  Source tensor. Data types supported: F16/F32
-     * @param[out] output Destination tensor. Data types supported: same as @p input
-     * @param[in]  beta   (Optional) A scaling factor for the exponent. Only beta = 1 is supported
-     * @param[in]  axis   (Optional) Reduction axis. It has the purpose of squashing the first @p axis
-     *                    dimensions together. For instance, given a [4x4x4x4] image,
-     *                    when @p axis is 2, the Softmax reduction will be applied on each of the [4x4] planes of the input image.
+     * @param[in]  input           Source tensor. Data types supported: F16/F32
+     * @param[out] output          Destination tensor. Data types supported: same as @p input
+     * @param[in]  beta            (Optional) A scaling factor for the exponent. Only beta = 1 is supported
+     * @param[in]  reduce_end_axis (Optional) The last axis of the first n dimensions (inclusive)to reduce. Defaults to 0.
+     *                   It has the purpose of squashing together the first n dimensions till (including) the @p reduce_end_axis. For instance, given a [2x3x4x5] image,
+     *                   when @p reduce_end_axis is 1, the reduction will be applied to axes 0 and 1, and the Softmax op will be applied on each of the [2x3] planes of the input image.
+     *                   Must be in range [0, input_num_dimensions).
      *
-     * @note The value of @p axis must be always 1 for GLES
+     * @note The value of @p reduce_end_axis must be always 0 for GLES
      */
-    void configure(const IGCTensor *input, IGCTensor *output, float beta = 1.0f, size_t axis = 1);
+    void configure(const IGCTensor *input, IGCTensor *output, float beta = 1.0f, size_t reduce_end_axis = 0);
 
     // Inherited methods overridden:
     void run() override;
diff --git a/arm_compute/runtime/NEON/functions/NESoftmaxLayer.h b/arm_compute/runtime/NEON/functions/NESoftmaxLayer.h
index b80ceaf25c..c5c83d8b5a 100644
--- a/arm_compute/runtime/NEON/functions/NESoftmaxLayer.h
+++ b/arm_compute/runtime/NEON/functions/NESoftmaxLayer.h
@@ -48,6 +48,10 @@ class ITensor;
  * -# @ref NEFillBorderKernel
  * -# @ref NELogits1DMaxKernel
  * -# @ref NELogits1DSoftmaxKernel
+ * And if the reduce_end_axis is not 0 or -input_num_dimensions, the function will use one of the the following kernels
+ * to reshape the input and perform softmax on the reshaped input:
+ * -# @ref NEFlattenLayerKernel
+ * -# @ref NEReshapeLayerKernel
  */
 template <bool IS_LOG = false>
 class NESoftmaxLayerGeneric : public IFunction
@@ -65,30 +69,31 @@ public:
     NESoftmaxLayerGeneric &operator=(NESoftmaxLayerGeneric &&) = default;
     /** Set the input and output tensors.
      *
-     * @param[in,out] input  Source tensor. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. If the width is not a
-     *                       multiple of the internal processing block size, @ref NEFillBorderKernel replicates the
-     *                       last value of each row to the nearest multiple.
-     * @param[out]    output Destination tensor. Data types supported: same as @p input.
-     * @param[in]     beta   (Optional) A scaling factor for the exponent.
-     * @param[in]     axis   (Optional) Reduction axis. Defaults to -1.
-     *                       Negative index is used to specify axis from the end (e.g. -1 for the last axis).Must be in range [-input_num_dimensions, input_num_dimensions).
-     *                       It has the purpose of squashing the first @p axis dimensions together. For instance, given a [4x4x4x4] image,
-     *                       when @p axis is 2, the Softmax reduction will be applied on each of the [4x4] planes of the input image.
+     * @param[in,out] input           Source tensor. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. If the width is not a
+     *                                multiple of the internal processing block size, @ref NEFillBorderKernel replicates the
+     *                                last value of each row to the nearest multiple.
+     * @param[out]    output          Destination tensor. Data types supported: same as @p input.
+     * @param[in]     beta            (Optional) A scaling factor for the exponent.
+     * @param[in]     reduce_end_axis (Optional) The last axis of the first n dimensions (inclusive)to reduce. Defaults to 0.
+     *                   It has the purpose of squashing together the first n dimensions till (including) the @p reduce_end_axis. For instance, given a [2x3x4x5] image,
+     *                   when @p reduce_end_axis is 1, the reduction will be applied to axes 0 and 1, and the Softmax op will be applied on each of the [2x3] planes of the input image.
+     *                   Negative index is used to specify axis from the end (e.g. -1 for the last axis).
+     *                   Must be in range [-input_num_dimensions, input_num_dimensions).
      */
-    void configure(ITensor *input, ITensor *output, float beta = 1.0f, int32_t axis = -1);
+    void configure(ITensor *input, ITensor *output, float beta = 1.0f, int32_t reduce_end_axis = 0);
     /** Static function to check if given info will lead to a valid configuration of @ref NESoftmaxLayer
      *
-     * @param[in] input  Source tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
-     * @param[in] output Destination tensor info. Data types supported: same as @p input
-     * @param[in] beta   (Optional) A scaling factor for the exponent.
-     * @param[in] axis   (Optional) Reduction axis. Defaults to -1.
-     *                   Negative index is used to specify axis from the end (e.g. -1 for the last axis).Must be in range [-input_num_dimensions, input_num_dimensions).
-     *                   It has the purpose of squashing the first @p axis dimensions together. For instance, given a [4x4x4x4] image,
-     *                   when @p axis is 2, the Softmax reduction will be applied on each of the [4x4] planes of the input image.
-     *
+     * @param[in] input           Source tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
+     * @param[in] output          Destination tensor info. Data types supported: same as @p input
+     * @param[in] beta            (Optional) A scaling factor for the exponent.
+     * @param[in] reduce_end_axis (Optional) The last axis of the first n dimensions (inclusive)to reduce. Defaults to 0.
+     *                   It has the purpose of squashing together the first n dimensions till (including) the @p reduce_end_axis. For instance, given a [2x3x4x5] image,
+     *                   when @p reduce_end_axis is 1, the reduction will be applied to axes 0 and 1, and the Softmax op will be applied on each of the [2x3] planes of the input image.
+     *                   Negative index is used to specify axis from the end (e.g. -1 for the last axis).
+     *                   Must be in range [-input_num_dimensions, input_num_dimensions).
      * @return a status
      */
-    static Status validate(const ITensorInfo *input, const ITensorInfo *output, float beta = 1.0f, int32_t axis = -1);
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, float beta = 1.0f, int32_t reduce_end_axis = 0);
 
     // Inherited methods overridden:
     void run() override;
@@ -101,14 +106,15 @@ private:
      * it initializes the kernel @p _flatten_kernel and the tensors @p _input_flat and
      * @p _output_flat
      *
-     * @param[in] input  Original source tensor.
-     * @param[in] output Original destination tensor.
-     * @param[in] axis   (Optional) Reduction axis. Defaults to -1.
-     *                   Negative index is used to specify axis from the end (e.g. -1 for the last axis).Must be in range [-input_num_dimensions, input_num_dimensions).
-     *                   It has the purpose of squashing the first @p axis dimensions together. For instance, given a [4x4x4x4] image,
-     *                   when @p axis is 2, the Softmax reduction will be applied on each of the [4x4] planes of the input image.
+     * @param[in] input           Original source tensor.
+     * @param[in] output          Original destination tensor.
+     * @param[in] reduce_end_axis (Optional) The last axis of the first n dimensions (inclusive)to reduce. Defaults to 0.
+     *                   It has the purpose of squashing together the first n dimensions till (including) the @p reduce_end_axis. For instance, given a [2x3x4x5] image,
+     *                   when @p reduce_end_axis is 1, the reduction will be applied to axes 0 and 1, and the Softmax op will be applied on each of the [2x3] planes of the input image.
+     *                   Negative index is used to specify axis from the end (e.g. -1 for the last axis).
+     *                   Must be in range [-input_num_dimensions, input_num_dimensions).
      */
-    void configure_reshape_input_kernel(const ITensor *input, const ITensor *output, int32_t axis);
+    void configure_reshape_input_kernel(const ITensor *input, const ITensor *output, int32_t reduce_end_axis);
 
     MemoryGroup                     _memory_group;
     NELogits1DMaxKernel             _max_kernel;