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>

1 files changed, 36 insertions, 26 deletions

diff --git a/src/runtime/CL/functions/CLSoftmaxLayer.cpp b/src/runtime/CL/functions/CLSoftmaxLayer.cpp
index b0b2117cd9..71ccf9fa01 100644
--- a/src/runtime/CL/functions/CLSoftmaxLayer.cpp
+++ b/src/runtime/CL/functions/CLSoftmaxLayer.cpp
@@ -42,35 +42,38 @@ CLSoftmaxLayerGeneric<IS_LOG>::CLSoftmaxLayerGeneric(std::shared_ptr<IMemoryMana
 }
 
 template <bool IS_LOG>
-void CLSoftmaxLayerGeneric<IS_LOG>::configure_reshape_input_kernel(const ICLTensor *input, const ICLTensor *output, size_t axis)
+void CLSoftmaxLayerGeneric<IS_LOG>::configure_reshape_input_kernel(const ICLTensor *input, const ICLTensor *output, size_t first_n_reduce_axes)
 {
-    configure_reshape_input_kernel(CLKernelLibrary::get().get_compile_context(), input, output, axis);
+    configure_reshape_input_kernel(CLKernelLibrary::get().get_compile_context(), input, output, first_n_reduce_axes);
 }
 
 template <bool IS_LOG>
-void CLSoftmaxLayerGeneric<IS_LOG>::configure_reshape_input_kernel(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *output, size_t axis)
+void CLSoftmaxLayerGeneric<IS_LOG>::configure_reshape_input_kernel(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *output, size_t first_n_reduce_axes)
 {
     // Flatten the input
-    const TensorShape shape_flatten = misc::shape_calculator::compute_softmax_shape(input->info(), axis);
+    const TensorShape shape_flatten = misc::shape_calculator::compute_softmax_shape(input->info(), first_n_reduce_axes);
 
     // Initialize the flat input
     _input_flattened.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_flatten));
 
     // If we need to flatten the input, we can use CLFlattenKernel or CLReshapeKernel
-    // If flattening on the third axes, we use CLFlattenKernel.
+    // If the number of reduced axes is 3 (max dimension), which means collapsing all axes except the batch axis, we use CLFlattenKernel.
     // In all other cases we have to use CLReshapeKernel
-    if(axis != 3)
-    {
-        auto reshape_kernel_ptr = support::cpp14::make_unique<CLReshapeLayerKernel>();
-        reshape_kernel_ptr->configure(compile_context, input, &_input_flattened);
-        _flatten_kernel_ptr = std::move(reshape_kernel_ptr);
-    }
-    else
+    // Note that the "other cases" include both:
+    //   1. first_n_reduce_axes < 3: Reduce the first 1 (no need to reduce) or 2 dimensions (inclusive)
+    //   2. first_n_reduce_axes == 4: Reduce all 4 dimensions. This can only be handled by CLReshapeKernel instead of CLFlattenKernel.
+    if(first_n_reduce_axes == 3)
     {
         auto flatten_kernel_ptr = support::cpp14::make_unique<CLFlattenLayerKernel>();
         flatten_kernel_ptr->configure(compile_context, input, &_input_flattened);
         _flatten_kernel_ptr = std::move(flatten_kernel_ptr);
     }
+    else
+    {
+        auto reshape_kernel_ptr = support::cpp14::make_unique<CLReshapeLayerKernel>();
+        reshape_kernel_ptr->configure(compile_context, input, &_input_flattened);
+        _flatten_kernel_ptr = std::move(reshape_kernel_ptr);
+    }
 
     // We need to init the output tensor here. Indeed, the reshape kernel expects
     // both tensors to be already initialized
@@ -78,20 +81,23 @@ void CLSoftmaxLayerGeneric<IS_LOG>::configure_reshape_input_kernel(const CLCompi
 }
 
 template <bool IS_LOG>
-void CLSoftmaxLayerGeneric<IS_LOG>::configure(const ICLTensor *input, ICLTensor *output, float beta, size_t axis)
+void CLSoftmaxLayerGeneric<IS_LOG>::configure(const ICLTensor *input, ICLTensor *output, float beta, size_t reduce_end_axis)
 {
-    configure(CLKernelLibrary::get().get_compile_context(), input, output, beta, axis);
+    configure(CLKernelLibrary::get().get_compile_context(), input, output, beta, reduce_end_axis);
 }
 
 template <bool IS_LOG>
-void CLSoftmaxLayerGeneric<IS_LOG>::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, float beta, size_t axis)
+void CLSoftmaxLayerGeneric<IS_LOG>::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, float beta, size_t reduce_end_axis)
 {
     // Perform validation step
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
-    ARM_COMPUTE_ERROR_THROW_ON(CLSoftmaxLayerGeneric<IS_LOG>::validate(input->info(), output->info(), beta, axis));
+    ARM_COMPUTE_ERROR_THROW_ON(CLSoftmaxLayerGeneric<IS_LOG>::validate(input->info(), output->info(), beta, reduce_end_axis));
 
-    // We don't need flattening only in the case the input is 2D and axis is 1
-    _needs_flattening = axis != 1;
+    // Convert reduce-before axis (inclusive) to first n axes to reduce
+    size_t first_n_reduce_axes = dim_index_2_num_dims(reduce_end_axis, input->info()->num_dimensions());
+
+    // We only need flattening when the number of axes to reduce is greater than 1
+    _needs_flattening = first_n_reduce_axes > 1;
 
     // If we are dealing with a 4D tensor, we will:
     // - Flatten the input, so that we end up with a [width*height*depth] * batches 2D tensor
@@ -102,8 +108,8 @@ void CLSoftmaxLayerGeneric<IS_LOG>::configure(const CLCompileContext &compile_co
         // Add to the memory manager _input_flattened
         _memory_group.manage(&_input_flattened);
 
-        // Cofigure  _flatten_kernel and _input_flattened
-        configure_reshape_input_kernel(input, output, axis);
+        // Cofigure _flatten_kernel and _input_flattened
+        configure_reshape_input_kernel(input, output, first_n_reduce_axes);
     }
 
     // We want to deal with a 2D input. Either it is the flattened version of the original input (4D case)
@@ -165,11 +171,15 @@ void CLSoftmaxLayerGeneric<IS_LOG>::configure(const CLCompileContext &compile_co
 }
 
 template <bool IS_LOG>
-Status CLSoftmaxLayerGeneric<IS_LOG>::validate(const ITensorInfo *input, const ITensorInfo *output, float beta, size_t axis)
+Status CLSoftmaxLayerGeneric<IS_LOG>::validate(const ITensorInfo *input, const ITensorInfo *output, float beta, size_t reduce_end_axis)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->num_dimensions() > 4, "Only up to 4 dimensions are supported");
     ARM_COMPUTE_UNUSED(beta);
+    ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() <= reduce_end_axis);
+
+    // Convert reduce-before axis (inclusive) to first n axes to reduce
+    size_t first_n_reduce_axes = dim_index_2_num_dims(reduce_end_axis, input->num_dimensions());
 
     // Create intermediate tensor info
     DataType   tmp_data_type = is_data_type_quantized_asymmetric(input->data_type()) ? DataType::S32 : input->data_type();
@@ -180,20 +190,20 @@ Status CLSoftmaxLayerGeneric<IS_LOG>::validate(const ITensorInfo *input, const I
     TensorInfo tensor_info_max(input->clone()->set_tensor_shape(max_sum_shape).set_is_resizable(true));
     TensorInfo tensor_info_sum(input->clone()->set_tensor_shape(max_sum_shape).set_data_type(tmp_data_type).set_quantization_info(QuantizationInfo()).set_is_resizable(true));
 
-    const bool needs_flattening = (axis != 1);
+    const bool needs_flattening = (first_n_reduce_axes > 1);
 
     if(needs_flattening)
     {
-        const TensorShape shape_flatten = misc::shape_calculator::compute_softmax_shape(input, axis);
+        const TensorShape shape_flatten = misc::shape_calculator::compute_softmax_shape(input, first_n_reduce_axes);
         TensorInfo        tensor_info_flat(input->clone()->set_tensor_shape(shape_flatten).set_is_resizable(true));
 
-        if(axis != 3)
+        if(first_n_reduce_axes == 3)
         {
-            ARM_COMPUTE_RETURN_ON_ERROR(CLReshapeLayerKernel::validate(input, &tensor_info_flat));
+            ARM_COMPUTE_RETURN_ON_ERROR(CLFlattenLayerKernel::validate(input, &tensor_info_flat));
         }
         else
         {
-            ARM_COMPUTE_RETURN_ON_ERROR(CLFlattenLayerKernel::validate(input, &tensor_info_flat));
+            ARM_COMPUTE_RETURN_ON_ERROR(CLReshapeLayerKernel::validate(input, &tensor_info_flat));
         }
     }