COMPMID-3871: Create BatchNormalization SVE/SVE2

1. Decouple data type for NHWC
2. Add NHWC SVE support for BachNormalization

Signed-off-by: Sheri Zhang <sheri.zhang@arm.com>
Change-Id: I0383b969b555b429d9acebb4efa17ecba9429ea7
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4755
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>

1 files changed, 146 insertions, 0 deletions

diff --git a/src/core/NEON/kernels/batchnormalization/impl/NEON/fp16.cpp b/src/core/NEON/kernels/batchnormalization/impl/NEON/fp16.cpp
new file mode 100644
index 0000000000..dfadef34f7
--- /dev/null
+++ b/src/core/NEON/kernels/batchnormalization/impl/NEON/fp16.cpp
@@ -0,0 +1,146 @@
+/*
+ * 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/NEMath.h"
+#include "src/core/NEON/kernels/detail/NEActivationFunctionDetail.h"
+#include "src/core/NEON/wrapper/wrapper.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_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS)
+namespace arm_compute
+{
+namespace
+{
+using BatchNomalizationPtr = void (*)(ITensor *src, ITensor *dst, const ITensor *mean, const ITensor *var, const ITensor *beta, const ITensor *gamma,
+                                      float epsilon, ActivationLayerInfo &act_info, const Window &window);
+
+template <typename T>
+void 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)
+{
+    /** NEON vector tag type. */
+    using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t<float16_t, wrapper::traits::BitWidth::W128>;
+
+    const int  window_step_x  = 8;
+    const auto window_start_x = static_cast<int>(window.x().start());
+    const auto window_end_x   = static_cast<int>(window.x().end());
+
+    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 input_mean  = reinterpret_cast<const float16_t *>(mean->ptr_to_element(Coordinates(0, 0)));
+    const auto input_var   = reinterpret_cast<const float16_t *>(var->ptr_to_element(Coordinates(0, 0)));
+    const auto input_gamma = (gamma != nullptr) ? reinterpret_cast<const float16_t *>(gamma->ptr_to_element(Coordinates(0, 0))) : nullptr;
+    const auto input_beta  = (beta != nullptr) ? reinterpret_cast<const float16_t *>(beta->ptr_to_element(Coordinates(0, 0))) : nullptr;
+
+    T activation_functor(act_info);
+
+    const auto epsilon_vec = wrapper::vdup_n(static_cast<float16_t>(epsilon), ExactTagType{});
+    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());
+
+        // Perform core calculations using vector operations
+        int x = window_start_x;
+        for(; x <= (window_end_x - window_step_x); x += window_step_x)
+        {
+            // Conctruct vectors
+            const auto mean_vec  = wrapper::vloadq(input_mean + x);
+            const auto var_vec   = wrapper::vloadq(input_var + x);
+            const auto gamma_vec = (input_gamma != nullptr) ? wrapper::vloadq(input_gamma + x) : wrapper::vdup_n(static_cast<float16_t>(1.f), ExactTagType{});
+            const auto beta_vec  = (input_beta != nullptr) ? wrapper::vloadq(input_beta + x) : wrapper::vdup_n(static_cast<float16_t>(0.f), ExactTagType{});
+
+            // Calculate denominator
+            const auto denominator = wrapper::vinvsqrt(wrapper::vadd(var_vec, epsilon_vec));
+
+            // Calculate x bar
+            const auto numerator = wrapper::vsub(wrapper::vloadq(input_ptr + x), mean_vec);
+            const auto x_bar     = wrapper::vmul(numerator, denominator);
+            auto       res       = wrapper::vmla(beta_vec, x_bar, gamma_vec);
+
+            // Perform fused activation
+            if(act_info.enabled())
+            {
+                activation_functor(res);
+            }
+
+            // Store results
+            wrapper::vstore(output_ptr + x, res);
+        }
+
+        // Compute left-over elements
+        for(; x < window_end_x; ++x)
+        {
+            // Conctruct vectors
+            const float16_t gamma = (input_gamma != nullptr) ? input_gamma[x] : 1.f;
+            const float16_t beta  = (input_beta != nullptr) ? input_beta[x] : 0.f;
+
+            const float16_t denominator = sqrt(input_var[x] + epsilon);
+            const float16_t numerator   = input_ptr[x] - input_mean[x];
+            const float16_t x_bar       = numerator / denominator;
+            float16_t       res         = beta + x_bar * gamma;
+
+            // Perform fused activation
+            if(act_info.enabled())
+            {
+                activation_functor(res);
+            }
+
+            // Store results
+            *reinterpret_cast<float16_t *>(output_ptr + x) = res;
+        }
+    },
+    input, output);
+}
+
+// Fused Batched Normalization with activation functions
+static std::map<ActivationLayerInfo::ActivationFunction, BatchNomalizationPtr> fused_map =
+{
+    { ActivationLayerInfo::ActivationFunction::RELU, &batch_normalization<detail::relu<float16_t, 8>> },
+    { ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, &batch_normalization<detail::brelu<float16_t, 8>> },
+    { ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, &batch_normalization<detail::lubrelu<float16_t, 8>> }
+};
+}
+namespace cpu
+{
+void fp16_neon_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)
+{
+    fused_map[act_info.activation()](src, dst, mean, var, beta, gamma, epsilon, act_info, window);
+}
+} // namespace cpu
+} // namespace arm_compute
+
+#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */
\ No newline at end of file