Changes to InstanceNrom to enable fp16 in armv8a multi_isa builds

* Code guarded with __ARM_FEATURE_FP16_VECTOR_ARITHMETIC needs
  to be moved to an fp16.cpp file to allow compilation with
  -march=armv8.2-a+fp16

* Partially resolves MLCE-1102

Change-Id: If53ff1927948b3ad7c9e3c9347bc2af38764e342
Signed-off-by: Pablo Marquez Tello <pablo.tello@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/10243
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>

3 files changed, 137 insertions, 40 deletions

diff --git a/src/cpu/kernels/instancenorm/generic/neon/fp16.cpp b/src/cpu/kernels/instancenorm/generic/neon/fp16.cpp
index e9fcc84b35..2b7d91b144 100644
--- a/src/cpu/kernels/instancenorm/generic/neon/fp16.cpp
+++ b/src/cpu/kernels/instancenorm/generic/neon/fp16.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2022 Arm Limited.
+ * Copyright (c) 2022-2023 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -22,20 +22,153 @@
  * SOFTWARE.
  */
 #if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS)
+#include "arm_compute/core/Helpers.h"
+#include "src/core/NEON/wrapper/wrapper.h"
 #include "src/cpu/kernels/instancenorm/generic/neon/impl.h"
+
 namespace arm_compute
 {
 namespace cpu
 {
+namespace
+{
+template <typename InputType, typename AccType>
+void vector_float_sum_fp16(AccType &result, AccType &result_square, const InputType &inputs)
+{
+    result        = wrapper::vadd(result, inputs);
+    result_square = wrapper::vadd(result_square, wrapper::vmul(inputs, inputs));
+}
+
+template <typename InputType, typename AccType>
+InputType vector_float_norm_fp16(const InputType &inputs, const AccType &vec_mean, const AccType &vec_multip, const AccType &vec_beta)
+{
+    return wrapper::vadd(wrapper::vmul(wrapper::vsub(inputs, vec_mean), vec_multip), vec_beta);
+}
+
+template <>
+inline void vector_float_sum_fp16(float32x4_t &result, float32x4_t &result_square, const float16x8_t &inputs)
+{
+    vector_float_sum_fp16(result, result_square, wrapper::vcvt<float>(wrapper::vgetlow(inputs)));
+    vector_float_sum_fp16(result, result_square, wrapper::vcvt<float>(wrapper::vgethigh(inputs)));
+}
+template <>
+inline float16x8_t vector_float_norm_fp16(const float16x8_t &inputs, const float32x4_t &vec_mean, const float32x4_t &vec_multip, const float32x4_t &vec_beta)
+{
+    const auto  input_low   = wrapper::vcvt<float>(wrapper::vgetlow(inputs));
+    const auto  input_high  = wrapper::vcvt<float>(wrapper::vgethigh(inputs));
+    const auto  result_low  = wrapper::vcvt<float16_t>(vector_float_norm_fp16(input_low, vec_mean, vec_multip, vec_beta));
+    const auto  result_high = wrapper::vcvt<float16_t>(vector_float_norm_fp16(input_high, vec_mean, vec_multip, vec_beta));
+    float16x8_t result      = wrapper::vcombine(result_low, result_high);
+
+    return result;
+}
+
+template <typename AccType>
+void instance_normalization_nchw_fp16(const ITensor *input, ITensor *output, float gamma, float beta, float epsilon, const Window &window)
+{
+    /** SIMD vector tag type. */
+    using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t<float16_t, wrapper::traits::BitWidth::W128>;
+
+    // Clear X/Y dimensions on execution window as we handle the planes manually
+    Window win = window;
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+    win.set(Window::DimY, Window::Dimension(0, 1, 1));
+
+    constexpr int      window_step_x  = 16 / sizeof(float16_t);
+    const unsigned int elements_plane = input->info()->dimension(0) * output->info()->dimension(1);
+
+    Iterator input_it(input, win);
+    execute_window_loop(win, [&](const Coordinates & id)
+    {
+        Window win_plane = window;
+        win_plane.set(Window::DimX, Window::Dimension(0, 1, 1));
+        win_plane.set(Window::DimZ, Window::Dimension(id[2], id[2] + 1, 1));
+        win_plane.set(3, Window::Dimension(id[3], id[3] + 1, 1));
+
+        Iterator input_plane_it(input, win_plane);
+        Iterator output_plane_it(output, win_plane);
+
+        auto sum_h_w         = static_cast<AccType>(0.f);
+        auto sum_squares_h_w = static_cast<AccType>(0.f);
+
+        execute_window_loop(win_plane, [&](const Coordinates &)
+        {
+            const auto input_ptr = reinterpret_cast<const float16_t *>(input_plane_it.ptr());
+
+            auto vec_sum_h_w         = wrapper::vdup_n(static_cast<AccType>(0.f), ExactTagType{});
+            auto vec_sum_squares_h_w = wrapper::vdup_n(static_cast<AccType>(0.f), ExactTagType{});
+
+            // Compute S elements per iteration
+            int x = window.x().start();
+            for(; x <= (window.x().end() - window_step_x); x += window_step_x)
+            {
+                auto vec_input_val = wrapper::vloadq(input_ptr + x);
+                vector_float_sum_fp16(vec_sum_h_w, vec_sum_squares_h_w, vec_input_val);
+            }
+
+            auto vec2_sum_h_w         = wrapper::vpadd(wrapper::vgethigh(vec_sum_h_w), wrapper::vgetlow(vec_sum_h_w));
+            auto vec2_sum_squares_h_w = wrapper::vpadd(wrapper::vgethigh(vec_sum_squares_h_w), wrapper::vgetlow(vec_sum_squares_h_w));
+
+            vec2_sum_h_w         = wrapper::vpadd(vec2_sum_h_w, vec2_sum_h_w);
+            vec2_sum_squares_h_w = wrapper::vpadd(vec2_sum_squares_h_w, vec2_sum_squares_h_w);
+
+            sum_h_w += wrapper::vgetlane(vec2_sum_h_w, 0);
+            sum_squares_h_w += wrapper::vgetlane(vec2_sum_squares_h_w, 0);
+
+            // Compute left-over elements
+            for(; x < window.x().end(); ++x)
+            {
+                const auto value = static_cast<AccType>(*(input_ptr + x));
+                sum_h_w += value;
+                sum_squares_h_w += value * value;
+            }
+        },
+        input_plane_it, output_plane_it);
+
+        const auto mean_h_w = sum_h_w / elements_plane;
+        const auto var_h_w  = sum_squares_h_w / elements_plane - mean_h_w * mean_h_w;
+
+        const auto multip_h_w     = gamma / std::sqrt(var_h_w + epsilon);
+        const auto vec_mean_h_w   = wrapper::vdup_n(static_cast<AccType>(mean_h_w), ExactTagType{});
+        const auto vec_multip_h_w = wrapper::vdup_n(static_cast<AccType>(multip_h_w), ExactTagType{});
+        const auto vec_beta       = wrapper::vdup_n(static_cast<AccType>(beta), ExactTagType{});
+
+        execute_window_loop(win_plane, [&](const Coordinates &)
+        {
+            auto input_ptr  = reinterpret_cast<const float16_t *>(input_plane_it.ptr());
+            auto output_ptr = reinterpret_cast<float16_t *>(output_plane_it.ptr());
+
+            // Compute S elements per iteration
+            int x = window.x().start();
+            for(; x <= (window.x().end() - window_step_x); x += window_step_x)
+            {
+                const auto vec_val        = wrapper::vloadq(input_ptr + x);
+                const auto normalized_vec = vector_float_norm_fp16(vec_val, vec_mean_h_w, vec_multip_h_w, vec_beta);
+                wrapper::vstore(output_ptr + x, normalized_vec);
+            }
+
+            // Compute left-over elements
+            for(; x < window.x().end(); ++x)
+            {
+                const auto val    = static_cast<AccType>(*(input_ptr + x));
+                *(output_ptr + x) = static_cast<float16_t>((val - mean_h_w) * multip_h_w + beta);
+            }
+        },
+        input_plane_it, output_plane_it);
+    },
+    input_it);
+}
+}
+
 void neon_fp16_instancenorm(ITensor *input, ITensor *output, float gamma, float beta, float epsilon, bool use_mixed_precision, const Window &window)
 {
     if(use_mixed_precision)
     {
-        return instance_normalization_nchw<float16_t, float>(input, output, gamma, beta, epsilon, window);
+        return instance_normalization_nchw_fp16<float>(input, output, gamma, beta, epsilon, window);
     }
     else
     {
-        return instance_normalization_nchw<float16_t>(input, output, gamma, beta, epsilon, window);
+        return instance_normalization_nchw_fp16<float16_t>(input, output, gamma, beta, epsilon, window);
     }
 }
 } // namespace cpu
diff --git a/src/cpu/kernels/instancenorm/generic/neon/impl.cpp b/src/cpu/kernels/instancenorm/generic/neon/impl.cpp
index e35cf97608..483b6f568b 100644
--- a/src/cpu/kernels/instancenorm/generic/neon/impl.cpp
+++ b/src/cpu/kernels/instancenorm/generic/neon/impl.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2022 Arm Limited.
+ * Copyright (c) 2019-2023 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -37,35 +37,12 @@ void vector_float_sum(AccType &result, AccType &result_square, const InputType &
     result_square = wrapper::vadd(result_square, wrapper::vmul(inputs, inputs));
 }
 
-#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-template <>
-inline void vector_float_sum(float32x4_t &result, float32x4_t &result_square, const float16x8_t &inputs)
-{
-    vector_float_sum(result, result_square, wrapper::vcvt<float>(wrapper::vgetlow(inputs)));
-    vector_float_sum(result, result_square, wrapper::vcvt<float>(wrapper::vgethigh(inputs)));
-}
-template <>
-inline float16x8_t vector_float_norm(const float16x8_t &inputs, const float32x4_t &vec_mean, const float32x4_t &vec_multip, const float32x4_t &vec_beta)
-{
-    const auto  input_low   = wrapper::vcvt<float>(wrapper::vgetlow(inputs));
-    const auto  input_high  = wrapper::vcvt<float>(wrapper::vgethigh(inputs));
-    const auto  result_low  = wrapper::vcvt<float16_t>(vector_float_norm(input_low, vec_mean, vec_multip, vec_beta));
-    const auto  result_high = wrapper::vcvt<float16_t>(vector_float_norm(input_high, vec_mean, vec_multip, vec_beta));
-    float16x8_t result      = wrapper::vcombine(result_low, result_high);
-
-    return result;
-}
-#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-
 template <typename InputType, typename AccType>
 InputType vector_float_norm(const InputType &inputs, const AccType &vec_mean, const AccType &vec_multip, const AccType &vec_beta)
 {
     return wrapper::vadd(wrapper::vmul(wrapper::vsub(inputs, vec_mean), vec_multip), vec_beta);
 }
 
-#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-
-#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 template <typename T, typename AccType>
 void instance_normalization_nchw(ITensor *input, ITensor *output, float gamma, float beta, float epsilon, const Window &window)
 {
@@ -164,9 +141,5 @@ void instance_normalization_nchw(ITensor *input, ITensor *output, float gamma, f
 }
 
 template void instance_normalization_nchw<float>(ITensor *input, ITensor *output, float gamma, float beta, float epsilon, const Window &window);
-#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS)
-template void instance_normalization_nchw<float16_t, float>(ITensor *input, ITensor *output, float gamma, float beta, float epsilon, const Window &window);
-template void instance_normalization_nchw<float16_t>(ITensor *input, ITensor *output, float gamma, float beta, float epsilon, const Window &window);
-#endif //defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS)
 } // namespace cpu
 } // namespace arm_compute
diff --git a/src/cpu/kernels/instancenorm/generic/neon/impl.h b/src/cpu/kernels/instancenorm/generic/neon/impl.h
index 1d413a9bcd..0ddfcdd5ba 100644
--- a/src/cpu/kernels/instancenorm/generic/neon/impl.h
+++ b/src/cpu/kernels/instancenorm/generic/neon/impl.h
@@ -39,15 +39,6 @@ void vector_float_sum(AccType &result, AccType &result_square, const InputType &
 
 template <typename InputType, typename AccType = InputType>
 InputType vector_float_norm(const InputType &inputs, const AccType &vec_mean, const AccType &vec_multip, const AccType &vec_beta);
-
-#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS)
-template <>
-inline void vector_float_sum(float32x4_t &result, float32x4_t &result_square, const float16x8_t &inputs);
-
-template <>
-inline float16x8_t vector_float_norm(const float16x8_t &inputs, const float32x4_t &vec_mean, const float32x4_t &vec_multip, const float32x4_t &vec_beta);
-#endif //defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS)
-
 } // namespace cpu
 } // namespace arm_compute
 #endif //define SRC_CORE_SVE_KERNELS_INSTANCENORM_IMPL_H