aboutsummaryrefslogtreecommitdiff
path: root/src/cpu/operators/CpuSoftmax.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/cpu/operators/CpuSoftmax.cpp')
-rw-r--r--src/cpu/operators/CpuSoftmax.cpp101
1 files changed, 57 insertions, 44 deletions
diff --git a/src/cpu/operators/CpuSoftmax.cpp b/src/cpu/operators/CpuSoftmax.cpp
index bf4c2fa3a2..e55d7f903e 100644
--- a/src/cpu/operators/CpuSoftmax.cpp
+++ b/src/cpu/operators/CpuSoftmax.cpp
@@ -25,9 +25,10 @@
#include "arm_compute/core/Helpers.h"
#include "arm_compute/core/TensorInfo.h"
-#include "arm_compute/core/Validate.h"
#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "arm_compute/core/Validate.h"
#include "arm_compute/runtime/NEON/NEScheduler.h"
+
#include "src/common/utils/Log.h"
#include "src/core/helpers/MemoryHelpers.h"
#include "src/core/helpers/SoftmaxHelpers.h"
@@ -63,13 +64,15 @@ void CpuSoftmaxGeneric<IS_LOG>::configure(const ITensorInfo *src, ITensorInfo *d
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())));
+ 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)
+ if (_needs_permute)
{
- _permute_input.configure(src, &_input_permuted, softmax_helpers::get_permutation_vector_from_softmax_axis(actual_axis));
+ _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)
@@ -79,10 +82,11 @@ void CpuSoftmaxGeneric<IS_LOG>::configure(const ITensorInfo *src, ITensorInfo *d
// 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));
+ 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);
@@ -94,13 +98,14 @@ void CpuSoftmaxGeneric<IS_LOG>::configure(const ITensorInfo *src, ITensorInfo *d
_max_kernel = std::move(mk);
auto sm = std::make_unique<kernels::CpuLogits1DSoftmaxKernel<IS_LOG>>();
- if(_needs_permute)
+ 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));
+ _permute_output.configure(&_output_permuted, dst,
+ softmax_helpers::get_permutation_vector_from_softmax_axis(actual_axis));
}
else
{
@@ -109,11 +114,15 @@ void CpuSoftmaxGeneric<IS_LOG>::configure(const ITensorInfo *src, ITensorInfo *d
}
_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::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());
+ _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());
}
template <bool IS_LOG>
@@ -123,7 +132,8 @@ Status CpuSoftmaxGeneric<IS_LOG>::validate(const ITensorInfo *src, const ITensor
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);
+ 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();
@@ -131,25 +141,33 @@ Status CpuSoftmaxGeneric<IS_LOG>::validate(const ITensorInfo *src, const ITensor
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 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 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)
+ 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));
+ 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));
+ ARM_COMPUTE_RETURN_ON_ERROR(kernels::CpuLogits1DSoftmaxKernel<IS_LOG>::validate(
+ &tensor_info_tmp, &tensor_info_max_sum, dst, beta, &dont_care));
return Status{};
}
@@ -166,43 +184,38 @@ void CpuSoftmaxGeneric<IS_LOG>::run(ITensorPack &tensors)
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);
+ CpuAuxTensorHandler output_permuted(offset_int_vec(InternalTensorIdx::PERMUTED_DST), _output_permuted, tensors,
+ true);
ITensorPack max_pack;
ITensorPack softmax_pack;
- if(_needs_permute)
+ if (_needs_permute)
{
- ITensorPack permute_in_pack = { { TensorType::ACL_SRC, src }, { TensorType::ACL_DST, input_permuted.get() } };
+ 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() } };
+ 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() }
- };
+ 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() }
- };
+ 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)
+ if (_needs_permute)
{
ITensorPack permute_out_pack;
permute_out_pack.add_tensor(TensorType::ACL_SRC, output_permuted.get());
@@ -211,7 +224,7 @@ void CpuSoftmaxGeneric<IS_LOG>::run(ITensorPack &tensors)
}
}
-template <bool IS_LOG>
+template <bool IS_LOG>
experimental::MemoryRequirements CpuSoftmaxGeneric<IS_LOG>::workspace() const
{
return _aux_mem;
generated by cgit v1.2.1 (git 2.23.0) at 2024-05-19 22:35:36 +0000