Basic function to compute a SoftmaxLayer and a Log SoftmaxLayer. More...

#include <CpuSoftmax.h>

Collaboration diagram for CpuSoftmaxGeneric< IS_LOG >:

Public Member Functions
	CpuSoftmaxGeneric ()

void	configure (const ITensorInfo src, ITensorInfo dst, float beta=1.0f, int32_t axis=0)
	Set the input and output tensors. More...

void	run (ITensorPack &tensors) override
	Run the kernels contained in the function. More...

experimental::MemoryRequirements	workspace () const override
	Return the memory requirements required by the workspace. More...

Public Member Functions inherited from INEOperator
	INEOperator (IRuntimeContext *ctx=nullptr)
	Constructor. More...

	INEOperator (const INEOperator &)=delete
	Prevent instances of this class from being copied (As this class contains pointers) More...

	INEOperator (INEOperator &&)=default
	Default move constructor. More...

INEOperator &	operator= (const INEOperator &)=delete
	Prevent instances of this class from being copied (As this class contains pointers) More...

INEOperator &	operator= (INEOperator &&)=default
	Default move assignment operator. More...

	~INEOperator ()
	Default destructor. More...

void	prepare (ITensorPack &constants) override
	Prepare the function for executing. More...

Public Member Functions inherited from IOperator
virtual	~IOperator ()=default
	Destructor. More...

Static Public Member Functions
static Status	validate (const ITensorInfo src, const ITensorInfo dst, float beta=1.0f, int32_t axis=0)
	Static function to check if given info will lead to a valid configuration. More...

Detailed Description

template<bool IS_LOG = false>
class arm_compute::cpu::CpuSoftmaxGeneric< IS_LOG >

Basic function to compute a SoftmaxLayer and a Log SoftmaxLayer.

Softmax is calculated by :

\[ out = exp((x - max(x)) * beta) / sum(exp((x - max(x)) * beta)) \]

Log Softmax is calculated by :

\[ out = (x - max(x) * beta) - log(\sum{e^{x - max(x) * beta}}) \]

This function runs the following function/kernels:

Definition at line 57 of file CpuSoftmax.h.

Constructor & Destructor Documentation

◆ CpuSoftmaxGeneric()

CpuSoftmaxGeneric

Definition at line 44 of file CpuSoftmax.cpp.

     : _permute_input(),
       _permute_output(),
       _max_kernel(),
       _softmax_kernel(),
       _max(),
       _tmp(),
       _input_permuted(),
       _output_permuted(),
       _needs_permute(false),
       _aux_mem(InternalTensorIdx::COUNT)
 {
 }

Member Function Documentation

◆ configure()

void configure	(	const ITensorInfo *	src,
		ITensorInfo *	dst,
		float	beta = `1.0f`,
		int32_t	axis = `0`
	)

Set the input and output tensors.

Parameters

[in,out]	src	Source tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. last value of each row to the nearest multiple.
[out]	dst	Destination tensor ifo. Data types supported: same as `input`.
[in]	beta	(Optional) A scaling factor for the exponent.
[in]	axis	(Optional) The dimension in which to apply the function. E.g. for input of shape 4x5x6 and axis=1, softmax will be applied to 4x6=24 vectors of size 5. Defaults to 0

Definition at line 59 of file CpuSoftmax.cpp.

 {
     // Perform validation step
     ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
     ARM_COMPUTE_ERROR_THROW_ON(CpuSoftmaxGeneric::validate(src, dst, beta, axis));
     ARM_COMPUTE_LOG_PARAMS(src, dst, beta, axis);
  
     const unsigned int actual_axis = static_cast<unsigned int>(wrap_around(axis, static_cast<int32_t>(src->num_dimensions())));
  
     _needs_permute = actual_axis > 0;
  
     if(_needs_permute)
     {
         _permute_input.configure(src, &_input_permuted, softmax_helpers::get_permutation_vector_from_softmax_axis(actual_axis));
     }
  
     // We want to deal with a 2D input. Either it is the permuted version of the original input (4D case)
     // or it is the original input case (2D case)
     const ITensorInfo *tmp_input = (_needs_permute ? &_input_permuted : src);
  
     // Create intermediate tensors shapes
     TensorShape max_sum_shape = tmp_input->tensor_shape();
     max_sum_shape.set(0, 1);
     const TensorInfo input_info    = tmp_input->clone()->reset_padding().set_is_resizable(true);
     DataType         tmp_data_type = is_data_type_quantized_asymmetric(tmp_input->data_type()) ? DataType::F32 : tmp_input->data_type();
     TensorInfo       tensor_info_tmp(input_info.clone()->set_data_type(tmp_data_type));
     TensorInfo       max_info(tmp_input->clone()->set_tensor_shape(max_sum_shape));
  
     // Init intermediate tensors
     _max = TensorInfo(max_info);
     _tmp = TensorInfo(tensor_info_tmp);
  
     // Configure kernels
     auto mk = std::make_unique<kernels::CpuLogits1DMaxKernel>();
     mk->configure(tmp_input, &_max);
     _max_kernel = std::move(mk);
  
     auto sm = std::make_unique<kernels::CpuLogits1DSoftmaxKernel<IS_LOG>>();
     if(_needs_permute)
     {
         // The normalization kernel stores the result in a permuted output tensor
         sm->configure(tmp_input, &_max, &_output_permuted, beta, &_tmp);
  
         // Re-permute the permuted output into the requested (4D) output
         _permute_output.configure(&_output_permuted, dst, softmax_helpers::get_permutation_vector_from_softmax_axis(actual_axis));
     }
     else
     {
         // Softmax 2D case
         sm->configure(tmp_input, &_max, dst, beta, &_tmp);
     }
     _softmax_kernel = std::move(sm);
  
     _aux_mem[InternalTensorIdx::MAX] = MemoryInfo(offset_int_vec(InternalTensorIdx::MAX), MemoryLifetime::Temporary, _max.total_size());
     _aux_mem[InternalTensorIdx::TMP] = MemoryInfo(offset_int_vec(InternalTensorIdx::TMP), MemoryLifetime::Temporary, _tmp.total_size());
  
     _aux_mem[InternalTensorIdx::PERMUTED_SRC] = MemoryInfo(offset_int_vec(InternalTensorIdx::PERMUTED_SRC), MemoryLifetime::Temporary, _input_permuted.total_size());
     _aux_mem[InternalTensorIdx::PERMUTED_DST] = MemoryInfo(offset_int_vec(InternalTensorIdx::PERMUTED_DST), MemoryLifetime::Temporary, _output_permuted.total_size());
 }

References ARM_COMPUTE_ERROR_ON_NULLPTR, ARM_COMPUTE_ERROR_THROW_ON, ARM_COMPUTE_LOG_PARAMS, ICloneable< T >::clone(), ITensorInfo::data_type(), arm_compute::test::validation::dst, arm_compute::F32, arm_compute::softmax_helpers::get_permutation_vector_from_softmax_axis(), arm_compute::test::validation::input_info, arm_compute::is_data_type_quantized_asymmetric(), TensorShape::set(), arm_compute::test::validation::src, ITensorInfo::tensor_shape(), CpuSoftmaxGeneric< IS_LOG >::validate(), and arm_compute::wrap_around().

◆ run()

void run ( ITensorPack & tensors )

overridevirtual

Run the kernels contained in the function.

Parameters

[in] tensors Vector that contains the tensors to operate on.

Reimplemented from INEOperator.

Definition at line 158 of file CpuSoftmax.cpp.

 {
     ARM_COMPUTE_ERROR_ON_MSG(tensors.empty(), "No inputs provided");
  
     auto src = tensors.get_const_tensor(TensorType::ACL_SRC);
     auto dst = tensors.get_tensor(TensorType::ACL_DST);
  
     CpuAuxTensorHandler tmp(offset_int_vec(InternalTensorIdx::TMP), _tmp, tensors, true);
     CpuAuxTensorHandler max(offset_int_vec(InternalTensorIdx::MAX), _max, tensors, true);
  
     CpuAuxTensorHandler input_permuted(offset_int_vec(InternalTensorIdx::PERMUTED_SRC), _input_permuted, tensors, true);
     CpuAuxTensorHandler output_permuted(offset_int_vec(InternalTensorIdx::PERMUTED_DST), _output_permuted, tensors, true);
  
     ITensorPack max_pack;
     ITensorPack softmax_pack;
  
     if(_needs_permute)
     {
         ITensorPack permute_in_pack = { { TensorType::ACL_SRC, src }, { TensorType::ACL_DST, input_permuted.get() } };
         _permute_input.run(permute_in_pack);
  
         max_pack = { { TensorType::ACL_SRC, input_permuted.get() }, { TensorType::ACL_DST, max.get() } };
  
         softmax_pack =
         {
             { TensorType::ACL_SRC_0, input_permuted.get() },
             { TensorType::ACL_SRC_1, max.get() },
             { TensorType::ACL_DST_0, output_permuted.get() },
             { TensorType::ACL_DST_1, tmp.get() }
         };
     }
     else
     {
         max_pack = { { TensorType::ACL_SRC, src }, { TensorType::ACL_DST, max.get() } };
  
         softmax_pack =
         {
             { TensorType::ACL_SRC_0, src },
             { TensorType::ACL_SRC_1, max.get() },
             { TensorType::ACL_DST_0, dst },
             { TensorType::ACL_DST_1, tmp.get() }
         };
     }
  
     NEScheduler::get().schedule_op(_max_kernel.get(), Window::DimY, _max_kernel->window(), max_pack);
     NEScheduler::get().schedule_op(_softmax_kernel.get(), Window::DimY, _softmax_kernel->window(), softmax_pack);
  
     if(_needs_permute)
     {
         ITensorPack permute_out_pack;
         permute_out_pack.add_tensor(TensorType::ACL_SRC, output_permuted.get());
         permute_out_pack.add_tensor(TensorType::ACL_DST, dst);
         _permute_output.run(permute_out_pack);
     }
 }

References arm_compute::ACL_DST, arm_compute::ACL_DST_0, arm_compute::ACL_DST_1, arm_compute::ACL_SRC, arm_compute::ACL_SRC_0, arm_compute::ACL_SRC_1, ITensorPack::add_tensor(), ARM_COMPUTE_ERROR_ON_MSG, Window::DimY, arm_compute::test::validation::dst, ITensorPack::empty(), Scheduler::get(), CpuAuxTensorHandler::get(), ITensorPack::get_const_tensor(), ITensorPack::get_tensor(), MAX, arm_compute::offset_int_vec(), IScheduler::schedule_op(), and arm_compute::test::validation::src.

◆ validate()

Status validate	(	const ITensorInfo *	src,
		const ITensorInfo *	dst,
		float	beta = `1.0f`,
		int32_t	axis = `0`
	)

static

Static function to check if given info will lead to a valid configuration.

Returns: a status

Definition at line 120 of file CpuSoftmax.cpp.

 {
     // Perform validation step
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst);
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(src->num_dimensions() > 4, "Only up to 4 dimensions are supported");
     ARM_COMPUTE_UNUSED(beta);
     ARM_COMPUTE_RETURN_ERROR_ON(axis < static_cast<int32_t>(-src->num_dimensions()) || static_cast<int32_t>(src->num_dimensions()) <= axis);
  
     // Create intermediate tensor info
     DataType         tmp_data_type = src->data_type();
     const TensorInfo tensor_info_tmp(src->clone()->set_data_type(tmp_data_type).set_is_resizable(true));
  
     TensorShape max_sum_shape = src->tensor_shape();
     max_sum_shape.set(0, 1);
     const TensorInfo tensor_info_max_sum(src->clone()->set_tensor_shape(max_sum_shape).set_data_type(tmp_data_type).set_quantization_info(src->quantization_info()).set_is_resizable(true));
     const TensorInfo dont_care;
  
     const unsigned int actual_axis = static_cast<unsigned int>(wrap_around(axis, static_cast<int32_t>(src->num_dimensions())));
  
     const bool needs_permute = actual_axis > 0;
  
     if(needs_permute)
     {
         const PermutationVector permutation_vector = softmax_helpers::get_permutation_vector_from_softmax_axis(actual_axis);
         const TensorShape       permuted_shape     = misc::shape_calculator::compute_permutation_output_shape(*src, permutation_vector);
         TensorInfo              input_permuted(src->clone()->set_tensor_shape(permuted_shape));
         ARM_COMPUTE_RETURN_ON_ERROR(CpuPermute::validate(src, &input_permuted, permutation_vector));
         TensorInfo output_permuted(dst->clone()->set_tensor_shape(permuted_shape));
         ARM_COMPUTE_RETURN_ON_ERROR(CpuPermute::validate(&output_permuted, dst, permutation_vector));
     }
  
     ARM_COMPUTE_RETURN_ON_ERROR(kernels::CpuLogits1DMaxKernel::validate(src, &tensor_info_max_sum));
     ARM_COMPUTE_RETURN_ON_ERROR(kernels::CpuLogits1DSoftmaxKernel<IS_LOG>::validate(&tensor_info_tmp, &tensor_info_max_sum, dst, beta, &dont_care));
  
     return Status{};
 }

References ARM_COMPUTE_RETURN_ERROR_ON, ARM_COMPUTE_RETURN_ERROR_ON_MSG, ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR, ARM_COMPUTE_RETURN_ON_ERROR, ARM_COMPUTE_UNUSED, arm_compute::misc::shape_calculator::compute_permutation_output_shape(), arm_compute::test::validation::dst, arm_compute::softmax_helpers::get_permutation_vector_from_softmax_axis(), TensorShape::set(), arm_compute::test::validation::src, CpuPermute::validate(), CpuLogits1DMaxKernel::validate(), and arm_compute::wrap_around().

Referenced by CpuSoftmaxGeneric< IS_LOG >::configure().

◆ workspace()

experimental::MemoryRequirements workspace ( ) const

overridevirtual

Return the memory requirements required by the workspace.

Reimplemented from INEOperator.

Definition at line 215 of file CpuSoftmax.cpp.

 {
     return _aux_mem;
 }

The documentation for this class was generated from the following files:

src/cpu/operators/CpuSoftmax.h
src/cpu/operators/CpuSoftmax.cpp

Public Member Functions

Static Public Member Functions