aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/batchnormalization/impl/SVE/fp32.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/batchnormalization/impl/SVE/fp32.cpp')
-rw-r--r--src/core/NEON/kernels/batchnormalization/impl/SVE/fp32.cpp120
1 files changed, 67 insertions, 53 deletions
diff --git a/src/core/NEON/kernels/batchnormalization/impl/SVE/fp32.cpp b/src/core/NEON/kernels/batchnormalization/impl/SVE/fp32.cpp
index 7cc570d8aa..df4fbfe607 100644
--- a/src/core/NEON/kernels/batchnormalization/impl/SVE/fp32.cpp
+++ b/src/core/NEON/kernels/batchnormalization/impl/SVE/fp32.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2020-2021 Arm Limited.
+ * Copyright (c) 2020-2021,2023 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -24,20 +24,29 @@
#include "arm_compute/core/Helpers.h"
#include "arm_compute/core/ITensorPack.h"
#include "arm_compute/core/Window.h"
+#include "arm_compute/function_info/ActivationLayerInfo.h"
+
#include "src/core/NEON/SVEMath.h"
#include <cmath>
#include <cstddef>
-#if defined(ENABLE_SVE)
+#if defined(ARM_COMPUTE_ENABLE_SVE)
#include <arm_sve.h>
namespace arm_compute
{
namespace cpu
{
-void fp32_sve_batch_normalization(ITensor *src, ITensor *dst, const ITensor *mean, const ITensor *var, const ITensor *beta, const ITensor *gamma,
- float epsilon, ActivationLayerInfo &act_info, const Window &window)
+void fp32_sve_batch_normalization(ITensor *src,
+ ITensor *dst,
+ const ITensor *mean,
+ const ITensor *var,
+ const ITensor *beta,
+ const ITensor *gamma,
+ float epsilon,
+ ActivationLayerInfo &act_info,
+ const Window &window)
{
const auto window_start_x = static_cast<int>(window.x().start());
const auto window_end_x = static_cast<int>(window.x().end());
@@ -48,69 +57,74 @@ void fp32_sve_batch_normalization(ITensor *src, ITensor *dst, const ITensor *mea
Iterator input(src, win_collapsed);
Iterator output(dst, win_collapsed);
- const auto input_mean = reinterpret_cast<const float *>(mean->ptr_to_element(Coordinates(0, 0)));
- const auto input_var = reinterpret_cast<const float *>(var->ptr_to_element(Coordinates(0, 0)));
- const auto input_gamma = (gamma != nullptr) ? reinterpret_cast<const float *>(gamma->ptr_to_element(Coordinates(0, 0))) : nullptr;
- const auto input_beta = (beta != nullptr) ? reinterpret_cast<const float *>(beta->ptr_to_element(Coordinates(0, 0))) : nullptr;
+ const auto input_mean = reinterpret_cast<const float *>(mean->ptr_to_element(Coordinates(0, 0)));
+ const auto input_var = reinterpret_cast<const float *>(var->ptr_to_element(Coordinates(0, 0)));
+ const auto input_gamma =
+ (gamma != nullptr) ? reinterpret_cast<const float *>(gamma->ptr_to_element(Coordinates(0, 0))) : nullptr;
+ const auto input_beta =
+ (beta != nullptr) ? reinterpret_cast<const float *>(beta->ptr_to_element(Coordinates(0, 0))) : nullptr;
const auto epsilon_vec = svdup_n_f32(epsilon);
const auto const_1 = svdup_n_f32(1.f);
const auto const_0 = svdup_n_f32(0.f);
const auto va = svdup_n_f32(act_info.a());
const auto vb = svdup_n_f32(act_info.b());
- execute_window_loop(win_collapsed, [&](const Coordinates &)
- {
- const auto input_ptr = reinterpret_cast<const float *>(input.ptr());
- const auto output_ptr = reinterpret_cast<float *>(output.ptr());
-
- // Compute S elements per iteration
- int x = window_start_x;
- svbool_t pg = svwhilelt_b32(x, window_end_x);
- do
+ execute_window_loop(
+ win_collapsed,
+ [&](const Coordinates &)
{
- // Conctruct vectors
- const auto mean_vec = svld1_f32(pg, input_mean + x);
- const auto var_vec = svld1_f32(pg, input_var + x);
- const auto gamma_vec = (input_gamma != nullptr) ? svld1_f32(pg, input_gamma + x) : const_1;
- const auto beta_vec = (input_beta != nullptr) ? svld1_f32(pg, input_beta + x) : const_0;
+ const auto input_ptr = reinterpret_cast<const float *>(input.ptr());
+ const auto output_ptr = reinterpret_cast<float *>(output.ptr());
- // Calculate denominator
- const auto tmp = svadd_f32_z(pg, var_vec, epsilon_vec);
- auto denominator = svrsqrte_f32(tmp);
- denominator = svmul_f32_z(pg, svrsqrts_f32(svmul_f32_z(pg, tmp, denominator), denominator), denominator);
- denominator = svmul_f32_z(pg, svrsqrts_f32(svmul_f32_z(pg, tmp, denominator), denominator), denominator);
+ // Compute S elements per iteration
+ int x = window_start_x;
+ svbool_t pg = svwhilelt_b32(x, window_end_x);
+ do
+ {
+ // Conctruct vectors
+ const auto mean_vec = svld1_f32(pg, input_mean + x);
+ const auto var_vec = svld1_f32(pg, input_var + x);
+ const auto gamma_vec = (input_gamma != nullptr) ? svld1_f32(pg, input_gamma + x) : const_1;
+ const auto beta_vec = (input_beta != nullptr) ? svld1_f32(pg, input_beta + x) : const_0;
- // Calculate x bar
- const auto numerator = svsub_f32_z(pg, svld1_f32(pg, input_ptr + x), mean_vec);
- const auto x_bar = svmul_f32_z(pg, numerator, denominator);
- auto res = svmla_f32_z(pg, beta_vec, x_bar, gamma_vec);
+ // Calculate denominator
+ const auto tmp = svadd_f32_z(pg, var_vec, epsilon_vec);
+ auto denominator = svrsqrte_f32(tmp);
+ denominator =
+ svmul_f32_z(pg, svrsqrts_f32(svmul_f32_z(pg, tmp, denominator), denominator), denominator);
+ denominator =
+ svmul_f32_z(pg, svrsqrts_f32(svmul_f32_z(pg, tmp, denominator), denominator), denominator);
- // Perform fused activation
- if(act_info.enabled())
- {
- if(act_info.activation() == ActivationLayerInfo::ActivationFunction::RELU)
- {
- res = svmax_f32_z(pg, const_0, res);
- }
- else if(act_info.activation() == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU)
- {
- res = svmin_f32_z(pg, va, svmax_f32_z(pg, const_0, res));
- }
- else if(act_info.activation() == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU)
+ // Calculate x bar
+ const auto numerator = svsub_f32_z(pg, svld1_f32(pg, input_ptr + x), mean_vec);
+ const auto x_bar = svmul_f32_z(pg, numerator, denominator);
+ auto res = svmla_f32_z(pg, beta_vec, x_bar, gamma_vec);
+
+ // Perform fused activation
+ if (act_info.enabled())
{
- res = svmin_f32_z(pg, va, svmax_f32_z(pg, vb, res));
+ if (act_info.activation() == ActivationLayerInfo::ActivationFunction::RELU)
+ {
+ res = svmax_f32_z(pg, const_0, res);
+ }
+ else if (act_info.activation() == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU)
+ {
+ res = svmin_f32_z(pg, va, svmax_f32_z(pg, const_0, res));
+ }
+ else if (act_info.activation() == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU)
+ {
+ res = svmin_f32_z(pg, va, svmax_f32_z(pg, vb, res));
+ }
}
- }
- // Store results
- svst1_f32(pg, output_ptr + x, res);
+ // Store results
+ svst1_f32(pg, output_ptr + x, res);
- x += svcntw();
- pg = svwhilelt_b32(x, window_end_x);
- }
- while(svptest_any(svptrue_b32(), pg));
- },
- input, output);
+ x += svcntw();
+ pg = svwhilelt_b32(x, window_end_x);
+ } while (svptest_any(svptrue_b32(), pg));
+ },
+ input, output);
}
} // namespace cpu
} // namespace arm_compute
generated by cgit v1.2.1 (git 2.23.0) at 2024-05-03 16:09:33 +0000