COMPMID-3870: Create ActivationLayer SVE/SVE2

Adds support for ActivationLayer for SVE and SVE2.
Datatypes supported:
*FP32
*FP16
*QASYMM8
*QASYMM8_SIGNED
*QSYMM16

Change-Id: Ia3583891795cda4ca2f9fa27c440731a5c27710d
Signed-off-by: Michalis Spyrou <michalis.spyrou@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4566
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

14 files changed, 792 insertions, 1 deletions

diff --git a/src/core/NEON/kernels/NEActivationLayerKernel.cpp b/src/core/NEON/kernels/NEActivationLayerKernel.cpp
index d969fd8e38..f215787bf6 100644
--- a/src/core/NEON/kernels/NEActivationLayerKernel.cpp
+++ b/src/core/NEON/kernels/NEActivationLayerKernel.cpp
@@ -56,6 +56,18 @@ struct ActivationKernel
 
 static const ActivationKernel available_kernels[] =
 {
+#if defined(__ARM_FEATURE_SVE)
+    {
+        "fp16_sve_activation",
+        [](const ActivationSelectorData & data) { return data.dt == DataType::F16; },
+        REGISTER_FP16_SVE(arm_compute::cpu::fp16_sve_activation)
+    },
+    {
+        "fp32_sve_activation",
+        [](const ActivationSelectorData & data) { return data.dt == DataType::F32; },
+        REGISTER_FP32_SVE(arm_compute::cpu::fp32_sve_activation)
+    },
+#else  /* !defined(__ARM_FEATURE_SVE) */
     {
         "fp16_neon_activation",
         [](const ActivationSelectorData & data) { return data.dt == DataType::F16; },
@@ -66,6 +78,25 @@ static const ActivationKernel available_kernels[] =
         [](const ActivationSelectorData & data) { return data.dt == DataType::F32; },
         REGISTER_FP32_NEON(arm_compute::cpu::fp32_neon_activation)
     },
+#endif /* defined(__ARM_FEATURE_SVE)  */
+
+#if defined(__ARM_FEATURE_SVE2) /* defined(__ARM_FEATURE_SVE2) */
+    {
+        "qasymm8_sve_activation",
+        [](const ActivationSelectorData & data) { return data.dt == DataType::QASYMM8; },
+        REGISTER_QASYMM8_SVE(arm_compute::cpu::qasymm8_sve_activation)
+    },
+    {
+        "qasymm8_signed_sve_activation",
+        [](const ActivationSelectorData & data) { return data.dt == DataType::QASYMM8_SIGNED; },
+        REGISTER_QASYMM8_SIGNED_SVE(arm_compute::cpu::qasymm8_signed_sve_activation)
+    },
+    {
+        "qsymm16_sve_activation",
+        [](const ActivationSelectorData & data) { return data.dt == DataType::QSYMM16; },
+        REGISTER_QSYMM16_SVE(arm_compute::cpu::qsymm16_sve_activation)
+    },
+#else  /* !defined(__ARM_FEATURE_SVE2) */
     {
         "qasymm8_neon_activation",
         [](const ActivationSelectorData & data) { return data.dt == DataType::QASYMM8; },
@@ -81,6 +112,7 @@ static const ActivationKernel available_kernels[] =
         [](const ActivationSelectorData & data) { return data.dt == DataType::QSYMM16; },
         REGISTER_QSYMM16_NEON(arm_compute::cpu::qsymm16_neon_activation)
     },
+#endif /* defined(__ARM_FEATURE_SVE2) */
 };
 
 const ActivationKernel *get_implementation(const ActivationSelectorData &data)
@@ -159,7 +191,6 @@ std::pair<Status, Window> validate_and_configure_window(const ITensorInfo *input
         // Output auto inizialitation if not yet initialized
         auto_init_if_empty(*output, *input->clone());
 
-        // NEActivationLayerKernel doesn't need padding so update_window_and_padding() can be skipped
         Coordinates coord;
         coord.set_num_dimensions(output->num_dimensions());
         output->set_valid_region(ValidRegion(coord, output->tensor_shape()));
diff --git a/src/core/NEON/kernels/activation/impl/fp16_neon_activation.cpp b/src/core/NEON/kernels/activation/impl/NEON/fp16.cpp
index 58e1cfcf23..58e1cfcf23 100644
--- a/src/core/NEON/kernels/activation/impl/fp16_neon_activation.cpp
+++ b/src/core/NEON/kernels/activation/impl/NEON/fp16.cpp
diff --git a/src/core/NEON/kernels/activation/impl/fp32_neon_activation.cpp b/src/core/NEON/kernels/activation/impl/NEON/fp32.cpp
index 610db05224..610db05224 100644
--- a/src/core/NEON/kernels/activation/impl/fp32_neon_activation.cpp
+++ b/src/core/NEON/kernels/activation/impl/NEON/fp32.cpp
diff --git a/src/core/NEON/kernels/activation/impl/qasymm8_neon_activation.cpp b/src/core/NEON/kernels/activation/impl/NEON/qasymm8.cpp
index 7b26441824..7b26441824 100644
--- a/src/core/NEON/kernels/activation/impl/qasymm8_neon_activation.cpp
+++ b/src/core/NEON/kernels/activation/impl/NEON/qasymm8.cpp
diff --git a/src/core/NEON/kernels/activation/impl/qasymm8_signed_neon_activation.cpp b/src/core/NEON/kernels/activation/impl/NEON/qasymm8_signed.cpp
index c616c5e27d..c616c5e27d 100644
--- a/src/core/NEON/kernels/activation/impl/qasymm8_signed_neon_activation.cpp
+++ b/src/core/NEON/kernels/activation/impl/NEON/qasymm8_signed.cpp
diff --git a/src/core/NEON/kernels/activation/impl/qsymm16_neon_activation.cpp b/src/core/NEON/kernels/activation/impl/NEON/qsymm16.cpp
index 0bef807db9..0bef807db9 100644
--- a/src/core/NEON/kernels/activation/impl/qsymm16_neon_activation.cpp
+++ b/src/core/NEON/kernels/activation/impl/NEON/qsymm16.cpp
diff --git a/src/core/NEON/kernels/activation/impl/SVE/fp16.cpp b/src/core/NEON/kernels/activation/impl/SVE/fp16.cpp
new file mode 100644
index 0000000000..8d6f4f2351
--- /dev/null
+++ b/src/core/NEON/kernels/activation/impl/SVE/fp16.cpp
@@ -0,0 +1,132 @@
+/*
+ * Copyright (c) 2020 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensorPack.h"
+#include "arm_compute/core/Window.h"
+#include "src/core/common/StdTypes.h"
+#include "src/core/common/Validate.h"
+
+#include <cmath>
+#include <cstddef>
+
+#if defined(__ARM_FEATURE_SVE)
+#include "src/core/NEON/SVEMath.h"
+#include <arm_sve.h>
+
+namespace arm_compute
+{
+namespace cpu
+{
+void fp16_sve_activation(const ITensor *src, ITensor *dst, const 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());
+    const ActivationLayerInfo::ActivationFunction act            = act_info.activation();
+
+    Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
+    win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    Iterator input(src, win_collapsed);
+    Iterator output(dst, win_collapsed);
+
+    const auto const_1     = svdup_n_f16(1.f);
+    const auto const_0     = svdup_n_f16(0.f);
+    const auto const_6     = svdup_n_f16(6.f);
+    const auto const_3     = svdup_n_f16(3.f);
+    const auto const_inv_6 = svdup_n_f16(0.166666667f);
+
+    const auto va = svdup_n_f16(act_info.a());
+    const auto vb = svdup_n_f16(act_info.b());
+    execute_window_loop(win_collapsed, [&](const Coordinates &)
+    {
+        const auto input_ptr  = reinterpret_cast<const float16_t *>(input.ptr());
+        const auto output_ptr = reinterpret_cast<float16_t *>(output.ptr());
+
+        svfloat16_t tmp;
+
+        int      x  = window_start_x;
+        svbool_t pg = svwhilelt_b16(x, window_end_x);
+        do
+        {
+            const auto vin = svld1_f16(pg, input_ptr + x);
+            switch(act)
+            {
+                case ActivationLayerInfo::ActivationFunction::ABS:
+                    tmp = svabs_f16_z(pg, vin);
+                    break;
+                case ActivationLayerInfo::ActivationFunction::LINEAR:
+                    tmp = svmla_f16_z(pg, vb, va, vin);
+                    break;
+                case ActivationLayerInfo::ActivationFunction::LOGISTIC:
+                    tmp = svinv_f16_z(pg, svadd_f16_z(pg, const_1, svexp_f16_z(pg, svneg_f16_z(pg, vin))));
+                    break;
+                case ActivationLayerInfo::ActivationFunction::RELU:
+                    tmp = svmax_f16_z(pg, const_0, vin);
+                    break;
+                case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU:
+                    tmp = svmin_f16_z(pg, va, svmax_f16_z(pg, const_0, vin));
+                    break;
+                case ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU:
+                    tmp = svmin_f16_z(pg, va, svmax_f16_z(pg, vb, vin));
+                    break;
+                case ActivationLayerInfo::ActivationFunction::LEAKY_RELU:
+                    tmp = svadd_f16_z(pg, svmul_f16_z(pg, svmin_f16_z(pg, vin, const_0), va), svmax_f16_z(pg, vin, const_0));
+                    break;
+                case ActivationLayerInfo::ActivationFunction::SOFT_RELU:
+                    tmp = svlog_f16_z(pg, svadd_f16_z(pg, const_1, svexp_f16_z(pg, vin)));
+                    break;
+                case ActivationLayerInfo::ActivationFunction::ELU:
+                    tmp = svsel_f16(svcmpgt_f16(pg, vin, const_0), vin, svmul_f16_z(pg, va, svsub_f16_z(pg, svexp_f16_z(pg, vin), const_1)));
+                    break;
+                case ActivationLayerInfo::ActivationFunction::SQRT:
+                    tmp = svsqrt_f16_z(pg, vin);
+                    break;
+                case ActivationLayerInfo::ActivationFunction::SQUARE:
+                    tmp = svmul_f16_z(pg, vin, vin);
+                    break;
+                case ActivationLayerInfo::ActivationFunction::TANH:
+                    tmp = svmul_f16_z(pg, va, svtanh_f16_z(pg, svmul_f16_z(pg, vb, vin)));
+                    break;
+                case ActivationLayerInfo::ActivationFunction::IDENTITY:
+                    tmp = vin;
+                    break;
+                case ActivationLayerInfo::ActivationFunction::HARD_SWISH:
+                    tmp = svmul_f16_z(pg, vin, svmul_f16_z(pg, const_inv_6, svmin_f16_z(pg, const_6, svmax_f16_z(pg, const_0, svadd_f16_z(pg, vin, const_3)))));
+                    break;
+                default:
+                    ARM_COMPUTE_ERROR("Unsupported activation function");
+            }
+            svst1_f16(pg, output_ptr + x, tmp);
+
+            x += svcnth();
+            pg = svwhilelt_b16(x, window_end_x);
+
+        }
+        while(svptest_any(svptrue_b16(), pg));
+    },
+    input, output);
+}
+} // namespace cpu
+} // namespace arm_compute
+#endif // __ARM_FEATURE_SVE
\ No newline at end of file
diff --git a/src/core/NEON/kernels/activation/impl/SVE/fp32.cpp b/src/core/NEON/kernels/activation/impl/SVE/fp32.cpp
new file mode 100644
index 0000000000..2c276028a0
--- /dev/null
+++ b/src/core/NEON/kernels/activation/impl/SVE/fp32.cpp
@@ -0,0 +1,133 @@
+/*
+ * Copyright (c) 2020 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensorPack.h"
+#include "arm_compute/core/Window.h"
+#include "src/core/NEON/SVEMath.h"
+#include "src/core/common/StdTypes.h"
+#include "src/core/common/Validate.h"
+
+#include <cmath>
+#include <cstddef>
+
+#if defined(__ARM_FEATURE_SVE)
+#include <arm_sve.h>
+
+namespace arm_compute
+{
+namespace cpu
+{
+void fp32_sve_activation(const ITensor *src, ITensor *dst, const 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());
+    const ActivationLayerInfo::ActivationFunction act            = act_info.activation();
+
+    Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
+    win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    Iterator input(src, win_collapsed);
+    Iterator output(dst, win_collapsed);
+
+    const auto const_1     = svdup_n_f32(1.f);
+    const auto const_0     = svdup_n_f32(0.f);
+    const auto const_6     = svdup_n_f32(6.f);
+    const auto const_3     = svdup_n_f32(3.f);
+    const auto const_inv_6 = svdup_n_f32(0.166666667f);
+
+    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());
+
+        svfloat32_t tmp;
+
+        // Compute S elements per iteration
+        int      x  = window_start_x;
+        svbool_t pg = svwhilelt_b32(x, window_end_x);
+        do
+        {
+            const auto vin = svld1_f32(pg, input_ptr + x);
+            switch(act)
+            {
+                case ActivationLayerInfo::ActivationFunction::ABS:
+                    tmp = svabs_f32_z(pg, vin);
+                    break;
+                case ActivationLayerInfo::ActivationFunction::LINEAR:
+                    tmp = svmla_f32_z(pg, vb, va, vin);
+                    break;
+                case ActivationLayerInfo::ActivationFunction::LOGISTIC:
+                    tmp = svinv_f32_z(pg, svadd_f32_z(pg, const_1, svexp_f32_z(pg, svneg_f32_z(pg, vin))));
+                    break;
+                case ActivationLayerInfo::ActivationFunction::RELU:
+                    tmp = svmax_f32_z(pg, const_0, vin);
+                    break;
+                case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU:
+                    tmp = svmin_f32_z(pg, va, svmax_f32_z(pg, const_0, vin));
+                    break;
+                case ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU:
+                    tmp = svmin_f32_z(pg, va, svmax_f32_z(pg, vb, vin));
+                    break;
+                case ActivationLayerInfo::ActivationFunction::LEAKY_RELU:
+                    tmp = svadd_f32_z(pg, svmul_f32_z(pg, svmin_f32_z(pg, vin, const_0), va), svmax_f32_z(pg, vin, const_0));
+                    break;
+                case ActivationLayerInfo::ActivationFunction::SOFT_RELU:
+                    tmp = svlog_f32_z(pg, svadd_f32_z(pg, const_1, svexp_f32_z(pg, vin)));
+                    break;
+                case ActivationLayerInfo::ActivationFunction::ELU:
+                    tmp = svsel_f32(svcmpgt_f32(pg, vin, const_0), vin, svmul_f32_z(pg, va, svsub_f32_z(pg, svexp_f32_z(pg, vin), const_1)));
+                    break;
+                case ActivationLayerInfo::ActivationFunction::SQRT:
+                    tmp = svsqrt_f32_z(pg, vin);
+                    break;
+                case ActivationLayerInfo::ActivationFunction::SQUARE:
+                    tmp = svmul_f32_z(pg, vin, vin);
+                    break;
+                case ActivationLayerInfo::ActivationFunction::TANH:
+                    tmp = svmul_f32_z(pg, va, svtanh_f32_z(pg, svmul_f32_z(pg, vb, vin)));
+                    break;
+                case ActivationLayerInfo::ActivationFunction::IDENTITY:
+                    tmp = vin;
+                    break;
+                case ActivationLayerInfo::ActivationFunction::HARD_SWISH:
+                    tmp = svmul_f32_z(pg, vin, svmul_f32_z(pg, const_inv_6, svmin_f32_z(pg, const_6, svmax_f32_z(pg, const_0, svadd_f32_z(pg, vin, const_3)))));
+                    break;
+                default:
+                    ARM_COMPUTE_ERROR("Unsupported activation function");
+            }
+            svst1_f32(pg, output_ptr + x, tmp);
+
+            x += svcntw();
+            pg = svwhilelt_b32(x, window_end_x);
+
+        }
+        while(svptest_any(svptrue_b32(), pg));
+    },
+    input, output);
+}
+} // namespace cpu
+} // namespace arm_compute
+#endif // __ARM_FEATURE_SVE
\ No newline at end of file
diff --git a/src/core/NEON/kernels/activation/impl/SVE/qasymm8.cpp b/src/core/NEON/kernels/activation/impl/SVE/qasymm8.cpp
new file mode 100644
index 0000000000..a49a562c84
--- /dev/null
+++ b/src/core/NEON/kernels/activation/impl/SVE/qasymm8.cpp
@@ -0,0 +1,185 @@
+/*
+ * Copyright (c) 2020 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/Window.h"
+#include "src/core/common/StdTypes.h"
+#include "src/core/common/Validate.h"
+
+#include <arm_neon.h>
+#include <cmath>
+#include <cstddef>
+
+#if defined(__ARM_FEATURE_SVE2)
+#include "src/core/NEON/SVEAsymm.h"
+#include "src/core/NEON/SVEMath.h"
+#include <arm_sve.h>
+
+namespace arm_compute
+{
+namespace cpu
+{
+void qasymm8_sve_activation(const ITensor *src, ITensor *dst, const 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());
+    const ActivationLayerInfo::ActivationFunction act            = act_info.activation();
+
+    Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
+    win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    Iterator input(src, win_collapsed);
+    Iterator output(dst, win_collapsed);
+
+    const UniformQuantizationInfo qi_in           = src->info()->quantization_info().uniform();
+    const UniformQuantizationInfo qi_out          = dst->info()->quantization_info().uniform();
+    const auto                    va              = svdup_n_u8(quantize_qasymm8(act_info.a(), qi_in));
+    const auto                    vb              = svdup_n_u8(quantize_qasymm8(act_info.b(), qi_in));
+    const auto                    const_0         = quantize_qasymm8(0.f, qi_in);
+    const auto                    vconst_0        = svdup_n_u8(const_0);
+    const auto                    vconst_1        = svdup_n_f32(1.f);
+    const auto                    va_f32          = svdup_n_f32(act_info.a());
+    const auto                    vb_f32          = svdup_n_f32(act_info.b());
+    const auto                    const_6_f32     = svdup_n_f32(6.f);
+    const auto                    const_0_f32     = svdup_n_f32(0.f);
+    const auto                    const_3_f32     = svdup_n_f32(3.f);
+    const auto                    const_inv_6_f32 = svdup_n_f32(0.166666667f);
+
+    // Initialise scale/offset for re-quantization
+    bool requant = true;
+    if(qi_in.scale == qi_out.scale && qi_in.offset == qi_out.offset)
+    {
+        requant = false;
+    }
+    float s  = qi_in.scale / qi_out.scale;
+    float o  = -qi_in.offset * s + qi_out.offset;
+    auto  vs = svdup_n_f32(s);
+    auto  vo = svdup_n_f32(o);
+
+    execute_window_loop(win_collapsed, [&](const Coordinates &)
+    {
+        const auto input_ptr  = reinterpret_cast<const uint8_t *>(input.ptr());
+        const auto output_ptr = reinterpret_cast<uint8_t *>(output.ptr());
+
+        svuint8_t tmp;
+
+        int      x  = window_start_x;
+        svbool_t pg = svwhilelt_b8(x, window_end_x);
+        do
+        {
+            const auto vin = svld1_u8(pg, input_ptr + x);
+            if(act == ActivationLayerInfo::ActivationFunction::RELU)
+            {
+                // Perform activation
+                tmp = svmax_u8_z(pg, vconst_0, vin);
+                // Re-quantize to new output space
+                tmp = requant ? svmla_qasymm8_z(pg, tmp, vs, vo) : tmp;
+            }
+            else if(act == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU)
+            {
+                // Perform activation
+                tmp = svmin_u8_z(pg, va, svmax_u8_z(pg, vconst_0, vin));
+                // Re-quantize to new output space
+                tmp = requant ? svmla_qasymm8_z(pg, tmp, vs, vo) : tmp;
+            }
+            else if(act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU)
+            {
+                // Perform activation
+                tmp = svmin_u8_z(pg, va, svmax_u8_z(pg, vb, vin));
+                // Re-quantize to new output space
+                tmp = svmla_qasymm8_z(pg, tmp, vs, vo);
+            }
+            else if(act == ActivationLayerInfo::ActivationFunction::LOGISTIC)
+            {
+                // De-quantize
+                const auto vin_deq = svdequantize_z(pg, vin, qi_in);
+                // Perform activation
+                const svfloat32x4_t tmp_dep =
+                {
+                    { {
+                            svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 0))))),
+                            svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 1))))),
+                            svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 2))))),
+                            svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 3))))),
+                        }
+                    }
+                };
+                // Re-quantize to new output space
+                tmp = svquantize_z(pg, tmp_dep, qi_out);
+            }
+            else if(act == ActivationLayerInfo::ActivationFunction::TANH)
+            {
+                // De-quantize
+                const auto vin_deq = svdequantize_z(pg, vin, qi_in);
+                // Perform activation
+                const svfloat32x4_t tmp_dep =
+                {
+                    { {
+                            svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 0), vb_f32))),
+                            svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 1), vb_f32))),
+                            svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 2), vb_f32))),
+                            svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 3), vb_f32))),
+                        }
+                    }
+                };
+                // Re-quantize to new output space
+                tmp = svquantize_z(pg, tmp_dep, qi_out);
+            }
+            else if(act == ActivationLayerInfo::ActivationFunction::HARD_SWISH)
+            {
+                // De-quantize
+                const auto vin_deq = svdequantize_z(pg, vin, qi_in);
+                // Perform activation
+                const svfloat32x4_t tmp_dep =
+                {
+                    { {
+                            svmul_f32_z(pg, svget4_f32(vin_deq, 0), svmul_f32_z(pg, const_inv_6_f32, svmin_f32_z(pg, const_6_f32, svmax_f32_z(pg, const_0_f32, svadd_f32_z(pg, svget4_f32(vin_deq, 0), const_3_f32))))),
+                            svmul_f32_z(pg, svget4_f32(vin_deq, 1), svmul_f32_z(pg, const_inv_6_f32, svmin_f32_z(pg, const_6_f32, svmax_f32_z(pg, const_0_f32, svadd_f32_z(pg, svget4_f32(vin_deq, 1), const_3_f32))))),
+                            svmul_f32_z(pg, svget4_f32(vin_deq, 2), svmul_f32_z(pg, const_inv_6_f32, svmin_f32_z(pg, const_6_f32, svmax_f32_z(pg, const_0_f32, svadd_f32_z(pg, svget4_f32(vin_deq, 2), const_3_f32))))),
+                            svmul_f32_z(pg, svget4_f32(vin_deq, 3), svmul_f32_z(pg, const_inv_6_f32, svmin_f32_z(pg, const_6_f32, svmax_f32_z(pg, const_0_f32, svadd_f32_z(pg, svget4_f32(vin_deq, 3), const_3_f32))))),
+                        }
+                    }
+                };
+                // Re-quantize to new output space
+                tmp = svquantize_z(pg, tmp_dep, qi_out);
+            }
+            else
+            {
+                ARM_COMPUTE_ERROR("Unsupported activation function");
+            }
+
+            svst1_u8(pg, output_ptr + x, tmp);
+
+            x += svcntb();
+            pg = svwhilelt_b8(x, window_end_x);
+
+        }
+        while(svptest_any(svptrue_b8(), pg));
+
+    },
+    input, output);
+}
+} // namespace cpu
+} // namespace arm_compute
+#endif /* defined(__ARM_FEATURE_SVE2) */
\ No newline at end of file
diff --git a/src/core/NEON/kernels/activation/impl/SVE/qasymm8_signed.cpp b/src/core/NEON/kernels/activation/impl/SVE/qasymm8_signed.cpp
new file mode 100644
index 0000000000..f34bee88fc
--- /dev/null
+++ b/src/core/NEON/kernels/activation/impl/SVE/qasymm8_signed.cpp
@@ -0,0 +1,183 @@
+/*
+ * Copyright (c) 2020 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/Window.h"
+#include "src/core/NEON/wrapper/wrapper.h"
+#include "src/core/common/StdTypes.h"
+#include "src/core/common/Validate.h"
+
+#include <cmath>
+#include <cstddef>
+
+#if defined(__ARM_FEATURE_SVE2)
+#include "src/core/NEON/SVEAsymm.h"
+#include "src/core/NEON/SVEMath.h"
+#include <arm_sve.h>
+
+namespace arm_compute
+{
+namespace cpu
+{
+void qasymm8_signed_sve_activation(const ITensor *src, ITensor *dst, const 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());
+    const ActivationLayerInfo::ActivationFunction act            = act_info.activation();
+
+    Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
+    win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    Iterator input(src, win_collapsed);
+    Iterator output(dst, win_collapsed);
+
+    const UniformQuantizationInfo qi_in           = src->info()->quantization_info().uniform();
+    const UniformQuantizationInfo qi_out          = dst->info()->quantization_info().uniform();
+    const auto                    va              = svdup_n_s8(quantize_qasymm8_signed(act_info.a(), qi_in));
+    const auto                    vb              = svdup_n_s8(quantize_qasymm8_signed(act_info.b(), qi_in));
+    const auto                    const_0         = quantize_qasymm8_signed(0.f, qi_in);
+    const auto                    vconst_0        = svdup_n_s8(const_0);
+    const auto                    vconst_1        = svdup_n_f32(1.f);
+    const auto                    va_f32          = svdup_n_f32(act_info.a());
+    const auto                    vb_f32          = svdup_n_f32(act_info.b());
+    const auto                    const_6_f32     = svdup_n_f32(6.f);
+    const auto                    const_0_f32     = svdup_n_f32(0.f);
+    const auto                    const_3_f32     = svdup_n_f32(3.f);
+    const auto                    const_inv_6_f32 = svdup_n_f32(0.166666667f);
+
+    // Initialise scale/offset for re-quantization
+    bool requant = true;
+    if(qi_in.scale == qi_out.scale && qi_in.offset == qi_out.offset)
+    {
+        requant = false;
+    }
+    float s  = qi_in.scale / qi_out.scale;
+    float o  = -qi_in.offset * s + qi_out.offset;
+    auto  vs = svdup_n_f32(s);
+    auto  vo = svdup_n_f32(o);
+
+    execute_window_loop(win_collapsed, [&](const Coordinates &)
+    {
+        const auto input_ptr  = reinterpret_cast<const int8_t *>(input.ptr());
+        const auto output_ptr = reinterpret_cast<int8_t *>(output.ptr());
+
+        svint8_t tmp;
+
+        int      x  = window_start_x;
+        svbool_t pg = svwhilelt_b8(x, window_end_x);
+        do
+        {
+            const auto vin = svld1_s8(pg, input_ptr + x);
+            if(act == ActivationLayerInfo::ActivationFunction::RELU)
+            {
+                // Perform activation
+                tmp = svmax_s8_z(pg, vconst_0, vin);
+                // Re-quantize to new output space
+                tmp = requant ? svmla_qasymm8_signed_z(pg, tmp, vs, vo) : tmp;
+            }
+            else if(act == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU)
+            {
+                // Perform activation
+                tmp = svmin_s8_z(pg, va, svmax_s8_z(pg, vconst_0, vin));
+                // Re-quantize to new output space
+                tmp = requant ? svmla_qasymm8_signed_z(pg, tmp, vs, vo) : tmp;
+            }
+            else if(act == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU)
+            {
+                // Perform activation
+                tmp = svmin_s8_z(pg, va, svmax_s8_z(pg, vb, vin));
+                // Re-quantize to new output space
+                tmp = requant ? svmla_qasymm8_signed_z(pg, tmp, vs, vo) : tmp;
+            }
+            else if(act == ActivationLayerInfo::ActivationFunction::LOGISTIC)
+            {
+                // De-quantize
+                const auto vin_deq = svdequantize_z(pg, vin, qi_in);
+                // Perform activation
+                const svfloat32x4_t tmp_dep =
+                {
+                    { {
+                            svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 0))))),
+                            svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 1))))),
+                            svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 2))))),
+                            svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget4_f32(vin_deq, 3))))),
+                        }
+                    }
+                };
+                // Re-quantize to new output space
+                tmp = svquantize_signed_z(pg, tmp_dep, qi_out);
+            }
+            else if(act == ActivationLayerInfo::ActivationFunction::TANH)
+            {
+                // De-quantize
+                const auto vin_deq = svdequantize_z(pg, vin, qi_in);
+                // Perform activation
+                const svfloat32x4_t tmp_dep =
+                {
+                    { {
+                            svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 0), vb_f32))),
+                            svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 1), vb_f32))),
+                            svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 2), vb_f32))),
+                            svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget4_f32(vin_deq, 3), vb_f32))),
+                        }
+                    }
+                };
+                // Re-quantize to new output space
+                tmp = svquantize_signed_z(pg, tmp_dep, qi_out);
+            }
+            else if(act == ActivationLayerInfo::ActivationFunction::HARD_SWISH)
+            {
+                // De-quantize
+                const auto vin_deq = svdequantize_z(pg, vin, qi_in);
+                // Perform activation
+                const svfloat32x4_t tmp_dep =
+                {
+                    { {
+                            svmul_f32_z(pg, svget4_f32(vin_deq, 0), svmul_f32_z(pg, const_inv_6_f32, svmin_f32_z(pg, const_6_f32, svmax_f32_z(pg, const_0_f32, svadd_f32_z(pg, svget4_f32(vin_deq, 0), const_3_f32))))),
+                            svmul_f32_z(pg, svget4_f32(vin_deq, 1), svmul_f32_z(pg, const_inv_6_f32, svmin_f32_z(pg, const_6_f32, svmax_f32_z(pg, const_0_f32, svadd_f32_z(pg, svget4_f32(vin_deq, 1), const_3_f32))))),
+                            svmul_f32_z(pg, svget4_f32(vin_deq, 2), svmul_f32_z(pg, const_inv_6_f32, svmin_f32_z(pg, const_6_f32, svmax_f32_z(pg, const_0_f32, svadd_f32_z(pg, svget4_f32(vin_deq, 2), const_3_f32))))),
+                            svmul_f32_z(pg, svget4_f32(vin_deq, 3), svmul_f32_z(pg, const_inv_6_f32, svmin_f32_z(pg, const_6_f32, svmax_f32_z(pg, const_0_f32, svadd_f32_z(pg, svget4_f32(vin_deq, 3), const_3_f32))))),
+                        }
+                    }
+                };
+                // Re-quantize to new output space
+                tmp = svquantize_signed_z(pg, tmp_dep, qi_out);
+            }
+            else
+            {
+                ARM_COMPUTE_ERROR("Unsupported activation function");
+            }
+
+            svst1_s8(pg, output_ptr + x, tmp);
+
+            x += svcntb();
+            pg = svwhilelt_b8(x, window_end_x);
+
+        }
+        while(svptest_any(svptrue_b8(), pg));
+    },
+    input, output);
+}
+} // namespace cpu
+} // namespace arm_compute
+#endif /* defined(__ARM_FEATURE_SVE2) */
diff --git a/src/core/NEON/kernels/activation/impl/SVE/qsymm16.cpp b/src/core/NEON/kernels/activation/impl/SVE/qsymm16.cpp
new file mode 100644
index 0000000000..1432e3bbdf
--- /dev/null
+++ b/src/core/NEON/kernels/activation/impl/SVE/qsymm16.cpp
@@ -0,0 +1,122 @@
+/*
+ * Copyright (c) 2020 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensorPack.h"
+#include "arm_compute/core/Window.h"
+#include "arm_compute/core/experimental/Types.h"
+#include "src/core/common/StdTypes.h"
+#include "src/core/common/Validate.h"
+
+#include <cmath>
+#include <cstddef>
+
+#if defined(__ARM_FEATURE_SVE2)
+#include "src/core/NEON/SVEMath.h"
+#include "src/core/NEON/SVESymm.h"
+#include <arm_sve.h>
+
+namespace arm_compute
+{
+namespace cpu
+{
+void qsymm16_sve_activation(const ITensor *src, ITensor *dst, const 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());
+    const ActivationLayerInfo::ActivationFunction act            = act_info.activation();
+
+    Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
+    win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    Iterator input(src, win_collapsed);
+    Iterator output(dst, win_collapsed);
+
+    const UniformQuantizationInfo qi_in    = src->info()->quantization_info().uniform();
+    const UniformQuantizationInfo qi_out   = dst->info()->quantization_info().uniform();
+    const auto                    vconst_1 = svdup_n_f32(1.f);
+    const auto                    va_f32   = svdup_n_f32(act_info.a());
+    const auto                    vb_f32   = svdup_n_f32(act_info.b());
+
+    execute_window_loop(win_collapsed, [&](const Coordinates &)
+    {
+        const auto input_ptr  = reinterpret_cast<const int16_t *>(input.ptr());
+        const auto output_ptr = reinterpret_cast<int16_t *>(output.ptr());
+
+        svint16_t tmp;
+
+        int      x  = window_start_x;
+        svbool_t pg = svwhilelt_b16(x, window_end_x);
+        do
+        {
+            const auto vin = svld1_s16(pg, input_ptr + x);
+            if(act == ActivationLayerInfo::ActivationFunction::LOGISTIC)
+            {
+                // De-quantize
+                auto vin_deq = svdequantize_qsymm16_z(pg, vin, qi_in.scale);
+                // Perform activation
+                const svfloat32x2_t tmp_dep =
+                {
+                    { {
+                            svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget2_f32(vin_deq, 0))))),
+                            svdiv_f32_z(pg, vconst_1, svadd_f32_z(pg, vconst_1, svexp_f32_z(pg, svneg_f32_z(pg, svget2_f32(vin_deq, 1))))),
+                        }
+                    }
+                };
+                // Re-quantize to new output space
+                tmp = svquantize_qsymm16_z(pg, tmp_dep, qi_out.scale);
+            }
+            else if(act == ActivationLayerInfo::ActivationFunction::TANH)
+            {
+                // De-quantize
+                auto vin_deq = svdequantize_qsymm16_z(pg, vin, qi_in.scale);
+                // Perform activation
+                const svfloat32x2_t tmp_dep =
+                {
+                    { {
+                            svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget2_f32(vin_deq, 0), vb_f32))),
+                            svmul_f32_z(pg, va_f32, svtanh_f32_z(pg, svmul_f32_z(pg, svget2_f32(vin_deq, 1), vb_f32))),
+                        }
+                    }
+                };
+                // Re-quantize to new output space
+                tmp = svquantize_qsymm16_z(pg, tmp_dep, qi_out.scale);
+            }
+            else
+            {
+                ARM_COMPUTE_ERROR("Unsupported activation function");
+            }
+
+            svst1_s16(pg, output_ptr + x, tmp);
+
+            x += svcnth();
+            pg = svwhilelt_b16(x, window_end_x);
+
+        }
+        while(svptest_any(svptrue_b16(), pg));
+    },
+    input, output);
+}
+} // namespace cpu
+} // namespace arm_compute
+#endif /* defined(__ARM_FEATURE_SVE2) */
diff --git a/src/core/NEON/kernels/activation/impl/list.h b/src/core/NEON/kernels/activation/impl/list.h
index 3b48ee3e22..db6c5b21b8 100644
--- a/src/core/NEON/kernels/activation/impl/list.h
+++ b/src/core/NEON/kernels/activation/impl/list.h
@@ -32,10 +32,15 @@ namespace cpu
     void func_name(const ITensor *src, ITensor *dst, const ActivationLayerInfo &act_info, const Window &window)
 
 DECLARE_ACTIVATION_KERNEL(qasymm8_neon_activation);
+DECLARE_ACTIVATION_KERNEL(qasymm8_sve_activation);
 DECLARE_ACTIVATION_KERNEL(qasymm8_signed_neon_activation);
+DECLARE_ACTIVATION_KERNEL(qasymm8_signed_sve_activation);
 DECLARE_ACTIVATION_KERNEL(qsymm16_neon_activation);
+DECLARE_ACTIVATION_KERNEL(qsymm16_sve_activation);
 DECLARE_ACTIVATION_KERNEL(fp16_neon_activation);
+DECLARE_ACTIVATION_KERNEL(fp16_sve_activation);
 DECLARE_ACTIVATION_KERNEL(fp32_neon_activation);
+DECLARE_ACTIVATION_KERNEL(fp32_sve_activation);
 
 #undef DECLARE_ACTIVATION_KERNEL
 } // namespace cpu
diff --git a/src/core/NEON/kernels/floor/impl/fp16_neon_floor.cpp b/src/core/NEON/kernels/floor/impl/NEON/fp16.cpp
index 4f56ca9daf..4f56ca9daf 100644
--- a/src/core/NEON/kernels/floor/impl/fp16_neon_floor.cpp
+++ b/src/core/NEON/kernels/floor/impl/NEON/fp16.cpp
diff --git a/src/core/NEON/kernels/floor/impl/fp32_neon_floor.cpp b/src/core/NEON/kernels/floor/impl/NEON/fp32.cpp
index 3f4b14b3e5..3f4b14b3e5 100644
--- a/src/core/NEON/kernels/floor/impl/fp32_neon_floor.cpp
+++ b/src/core/NEON/kernels/floor/impl/NEON/fp32.cpp