aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/NEL2NormalizeLayerKernel.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/NEL2NormalizeLayerKernel.cpp')
-rw-r--r--src/core/NEON/kernels/NEL2NormalizeLayerKernel.cpp167
1 files changed, 71 insertions, 96 deletions
diff --git a/src/core/NEON/kernels/NEL2NormalizeLayerKernel.cpp b/src/core/NEON/kernels/NEL2NormalizeLayerKernel.cpp
index 9bda82d416..eea57a17d3 100644
--- a/src/core/NEON/kernels/NEL2NormalizeLayerKernel.cpp
+++ b/src/core/NEON/kernels/NEL2NormalizeLayerKernel.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2017-2021 Arm Limited.
+ * Copyright (c) 2017-2022 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -30,11 +30,14 @@
#include "arm_compute/core/Utils.h"
#include "arm_compute/core/Validate.h"
#include "arm_compute/core/Window.h"
-#include "src/core/NEON/NEMath.h"
+
+#include "src/common/cpuinfo/CpuIsaInfo.h"
+#include "src/core/common/Registrars.h"
#include "src/core/helpers/AutoConfiguration.h"
#include "src/core/helpers/WindowHelpers.h"
+#include "src/core/NEON/NEMath.h"
+#include "src/cpu/kernels/l2normlayer/list.h"
-#include "src/core/NEON/wrapper/wrapper.h"
#include <arm_neon.h>
#include <cmath>
@@ -44,93 +47,68 @@ namespace
{
constexpr int max_input_tensor_dim = 3;
-template <typename T, int S>
-void l2_normalize_X(const ITensor *in, const ITensor *sum, ITensor *out, float epsilon, const Window &window)
+struct L2NormalizeLayerSelectorData
{
- using ExactTagType = typename wrapper::traits::neon_vector<T, S>::tag_type;
+ DataType dt;
+ unsigned int actual_axis;
+ cpuinfo::CpuIsaInfo isa;
+};
- const int window_step_x = 16 / data_size_from_type(in->info()->data_type());
- const auto window_start_x = static_cast<int>(window.x().start());
- const auto window_end_x = static_cast<int>(window.x().end());
+using L2NormalizeLayerKernelSelctorPtr = std::add_pointer<bool(const L2NormalizeLayerSelectorData &data)>::type;
- Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
- win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
+using L2NormalizeLayerPtr = std::add_pointer<void(
+ const ITensor *in, const ITensor *sum, ITensor *out, float epsilon, const Window &window, size_t axis)>::type;
- Iterator input_it(in, win_collapsed);
- Iterator sum_it(sum, win_collapsed);
- Iterator output_it(out, win_collapsed);
-
- execute_window_loop(win_collapsed, [&](const Coordinates &)
+struct L2NormalizeLayerKernel
+{
+ const char *name;
+ const L2NormalizeLayerKernelSelctorPtr is_selected;
+ L2NormalizeLayerPtr ukernel;
+};
+
+static const L2NormalizeLayerKernel available_kernels[] = {
+ {"fp32_neon_l2normalize_x",
+ [](const L2NormalizeLayerSelectorData &data)
+ { return data.dt == DataType::F32 && data.actual_axis == Window::DimX; },
+ REGISTER_FP32_NEON(arm_compute::cpu::neon_fp32_l2_normalize_x)},
+ {"fp32_neon_l2normalize_yz",
+ [](const L2NormalizeLayerSelectorData &data)
+ { return data.dt == DataType::F32 && data.actual_axis != Window::DimX; },
+ REGISTER_FP32_NEON(arm_compute::cpu::neon_fp32_l2_normalize_yz)},
{
- const auto in_ptr = reinterpret_cast<const T *>(input_it.ptr());
- const auto out_ptr = reinterpret_cast<T *>(output_it.ptr());
-
- const T sum_value = *reinterpret_cast<const T *>(sum_it.ptr());
- const T norm_value = static_cast<T>(1.f) / std::sqrt(std::max(sum_value, static_cast<T>(epsilon)));
- const auto vec_norm_value = wrapper::vdup_n(norm_value, ExactTagType{});
-
- // Compute elements over vector steps
- int x = window_start_x;
- for(; x <= (window_end_x - window_step_x); x += window_step_x)
- {
- wrapper::vstore(out_ptr + x, wrapper::vmul(wrapper::vloadq(in_ptr + x), vec_norm_value));
- }
-
- // Compute left-over elements
- for(; x < window_end_x; ++x)
- {
- out_ptr[x] = in_ptr[x] * norm_value;
- }
+ "fp16_neon_l2normalize_x",
+ [](const L2NormalizeLayerSelectorData &data)
+ { return data.dt == DataType::F16 && data.isa.fp16 && data.actual_axis == Window::DimX; },
+ REGISTER_FP16_NEON(arm_compute::cpu::neon_fp16_l2_normalize_x),
},
- input_it, sum_it, output_it);
-}
+ {
+ "fp16_neon_l2normalize_yz",
+ [](const L2NormalizeLayerSelectorData &data)
+ { return data.dt == DataType::F16 && data.isa.fp16 && data.actual_axis != Window::DimX; },
+ REGISTER_FP16_NEON(arm_compute::cpu::neon_fp16_l2_normalize_yz),
+ },
+};
-template <typename T, int S>
-void l2_normalize_YZ(const ITensor *in, const ITensor *sum, ITensor *out, float epsilon, const Window &window, size_t axis)
+/** Micro-kernel selector
+ *
+ * @param[in] data Selection data passed to help pick the appropriate micro-kernel
+ *
+ * @return A matching micro-kernel else nullptr
+ */
+const L2NormalizeLayerKernel *get_implementation(const L2NormalizeLayerSelectorData &data)
{
- using ExactTagType = typename wrapper::traits::neon_vector<T, S>::tag_type;
-
- const int window_step_x = 16 / data_size_from_type(in->info()->data_type());
- const auto window_start_x = static_cast<int>(window.x().start());
- const auto window_end_x = static_cast<int>(window.x().end());
-
- Window win = window;
- win.set(Window::DimX, Window::Dimension(0, 1, 1));
-
- Window window_sum(win);
- window_sum.set(axis, Window::Dimension(0, 0, 0));
-
- Iterator input_it(in, win);
- Iterator sum_it(sum, window_sum);
- Iterator output_it(out, win);
-
- const auto vec_eps = wrapper::vdup_n(static_cast<T>(epsilon), ExactTagType{});
-
- execute_window_loop(win, [&](const Coordinates &)
+ for (const auto &uk : available_kernels)
{
- const auto in_ptr = reinterpret_cast<const T *>(input_it.ptr());
- const auto sum_ptr = reinterpret_cast<const T *>(sum_it.ptr());
- const auto out_ptr = reinterpret_cast<T *>(output_it.ptr());
-
- // Compute elements over vector steps
- int x = window_start_x;
- for(; x <= (window_end_x - window_step_x); x += window_step_x)
+ if (uk.is_selected(data))
{
- const auto vec_norm_value = wrapper::vinvsqrt(wrapper::vmax(wrapper::vloadq(sum_ptr + x), vec_eps));
- wrapper::vstore(out_ptr + x, wrapper::vmul(wrapper::vloadq(in_ptr + x), vec_norm_value));
+ return &uk;
}
-
- // Compute left-over elements
- for(; x < window_end_x; ++x)
- {
- const T norm_value = static_cast<T>(1.f) / std::sqrt(std::max(sum_ptr[x], static_cast<T>(epsilon)));
- out_ptr[x] = in_ptr[x] * norm_value;
- }
- },
- input_it, sum_it, output_it);
+ }
+ return nullptr;
}
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *sum, const ITensorInfo *output, int axis, float epsilon)
+Status
+validate_arguments(const ITensorInfo *input, const ITensorInfo *sum, const ITensorInfo *output, int axis, float epsilon)
{
ARM_COMPUTE_UNUSED(epsilon);
@@ -139,14 +117,15 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *sum, cons
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, sum);
ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
ARM_COMPUTE_RETURN_ERROR_ON_MSG(actual_axis > 2, "Actual axis greater than 2 is not supported");
- ARM_COMPUTE_RETURN_ERROR_ON_MSG(actual_axis >= TensorShape::num_max_dimensions, "Actual normalization axis greater than max number of dimensions");
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG(actual_axis >= TensorShape::num_max_dimensions,
+ "Actual normalization axis greater than max number of dimensions");
// Reduce shape on axis
TensorShape sum_shape = input->tensor_shape();
sum_shape.set(actual_axis, 1);
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(sum->tensor_shape(), sum_shape);
- if(output->total_size() != 0)
+ if (output->total_size() != 0)
{
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
@@ -175,7 +154,8 @@ NEL2NormalizeLayerKernel::NEL2NormalizeLayerKernel()
{
}
-void NEL2NormalizeLayerKernel::configure(const ITensor *input, const ITensor *sum, ITensor *output, int axis, float epsilon)
+void NEL2NormalizeLayerKernel::configure(
+ const ITensor *input, const ITensor *sum, ITensor *output, int axis, float epsilon)
{
ARM_COMPUTE_ERROR_ON_NULLPTR(input, sum, output);
ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), sum->info(), output->info(), axis, epsilon));
@@ -193,10 +173,12 @@ void NEL2NormalizeLayerKernel::configure(const ITensor *input, const ITensor *su
INEKernel::configure(std::get<1>(win_config));
}
-Status NEL2NormalizeLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *sum, const ITensorInfo *output, int axis, float epsilon)
+Status NEL2NormalizeLayerKernel::validate(
+ const ITensorInfo *input, const ITensorInfo *sum, const ITensorInfo *output, int axis, float epsilon)
{
ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, sum, output, axis, epsilon));
- ARM_COMPUTE_RETURN_ON_ERROR(std::get<0>(validate_and_configure_window(input->clone().get(), output->clone().get())));
+ ARM_COMPUTE_RETURN_ON_ERROR(
+ std::get<0>(validate_and_configure_window(input->clone().get(), output->clone().get())));
return Status{};
}
@@ -207,23 +189,16 @@ void NEL2NormalizeLayerKernel::run(const Window &window, const ThreadInfo &info)
ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
- if(_actual_axis > 2)
+ if (_actual_axis > 2)
{
ARM_COMPUTE_ERROR("Unsupported normalization axis");
}
- switch(_input->info()->data_type())
- {
- case DataType::F32:
- (_actual_axis == Window::DimX) ? l2_normalize_X<float, 4>(_input, _sum, _output, _epsilon, window) : l2_normalize_YZ<float, 4>(_input, _sum, _output, _epsilon, window, _actual_axis);
- break;
-#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
- case DataType::F16:
- (_actual_axis == Window::DimX) ? l2_normalize_X<float16_t, 8>(_input, _sum, _output, _epsilon, window) : l2_normalize_YZ<float16_t, 8>(_input, _sum, _output, _epsilon, window, _actual_axis);
- break;
-#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
- default:
- ARM_COMPUTE_ERROR("Not implemented");
- }
+ const auto *uk = get_implementation(
+ L2NormalizeLayerSelectorData{_output->info()->data_type(), _actual_axis, CPUInfo::get().get_isa()});
+ ARM_COMPUTE_ERROR_ON(uk == nullptr);
+ ARM_COMPUTE_ERROR_ON(uk->ukernel == nullptr);
+
+ uk->ukernel(_input, _sum, _output, _epsilon, window, _actual_axis);
}
} // namespace arm_compute
generated by cgit v1.2.1 (git 2.23.0) at 2024-05-04 04:09:40 +0000