Enable fat binary support

Changes our build system to allow building both Neon(TM) and SVE
kernels and package them in the same binary. This will allow
runtime selection of the underlying architecture.

Adds new build option, fat_binary, for enabling this feature.

Change-Id: I8e8386149773ce28e071a2fb7ddd8c8ae0f28a4a
Signed-off-by: Michalis Spyrou <michalis.spyrou@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5704
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 3 insertions, 94 deletions

diff --git a/src/core/cpu/kernels/add/sve/list.h b/src/core/cpu/kernels/add/sve/list.h
index 71dd875ad8..aebb43bb60 100644
--- a/src/core/cpu/kernels/add/sve/list.h
+++ b/src/core/cpu/kernels/add/sve/list.h
@@ -24,11 +24,12 @@
 #ifndef SRC_CORE_SVE_KERNELS_ADD_LIST_H
 #define SRC_CORE_SVE_KERNELS_ADD_LIST_H
 
-#if defined(__ARM_FEATURE_SVE)
+#if defined(ENABLE_SVE)
 #include "arm_compute/core/Types.h"
 #include "arm_compute/core/utils/misc/Traits.h"
 #include "src/core/NEON/SVEMath.h"
 #include "src/core/NEON/wrapper/intrinsics/intrinsics.h"
+#include "src/core/cpu/kernels/add/sve/impl.h"
 #include <arm_sve.h>
 
 namespace arm_compute
@@ -47,99 +48,7 @@ DECLARE_ADD_KERNEL(add_u8_u8_s16_sve);
 
 #undef DECLARE_ADD_KERNEL
 
-template <typename ScalarType>
-void add_same_sve(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
-{
-    const auto all_true_pg           = wrapper::svptrue<ScalarType>();
-    const auto window_start_x        = static_cast<int>(window.x().start());
-    const auto window_end_x          = static_cast<int>(window.x().end());
-    const bool is_broadcast_across_x = src0->info()->tensor_shape().x() != src1->info()->tensor_shape().x();
-    const bool is_sat                = (policy == ConvertPolicy::SATURATE);
-
-    // Clear X Dimension on execution window as we handle manually
-    Window win = window;
-    win.set(Window::DimX, Window::Dimension(0, 1, 1));
-
-    // Create input windows
-    Window input1_win = window.broadcast_if_dimension_le_one(src0->info()->tensor_shape());
-    Window input2_win = window.broadcast_if_dimension_le_one(src1->info()->tensor_shape());
-
-    Iterator input1(src0, window.broadcast_if_dimension_le_one(src0->info()->tensor_shape()));
-    Iterator input2(src1, window.broadcast_if_dimension_le_one(src1->info()->tensor_shape()));
-    Iterator output(dst, window);
-
-    if(is_broadcast_across_x)
-    {
-        const bool     is_broadcast_input_2 = input2_win.x().step() == 0;
-        Window         broadcast_win        = is_broadcast_input_2 ? input2_win : input1_win;
-        Window         non_broadcast_win    = !is_broadcast_input_2 ? input2_win : input1_win;
-        const ITensor *broadcast_tensor     = is_broadcast_input_2 ? src1 : src0;
-        const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? src1 : src0;
-
-        // Clear X Dimension on execution window as we handle manually
-        non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));
-
-        Iterator broadcast_input(broadcast_tensor, broadcast_win);
-        Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
-        Iterator output(dst, win);
-
-        execute_window_loop(win, [&](const Coordinates &)
-        {
-            const auto non_broadcast_input_ptr = reinterpret_cast<const ScalarType *>(non_broadcast_input.ptr());
-            const auto output_ptr              = reinterpret_cast<ScalarType *>(output.ptr());
-
-            const ScalarType broadcast_value     = *reinterpret_cast<const ScalarType *>(broadcast_input.ptr());
-            const auto       broadcast_value_vec = wrapper::svdup_n(broadcast_value);
-
-            int      x  = window_start_x;
-            svbool_t pg = wrapper::svwhilelt<ScalarType>(x, window_end_x);
-            do
-            {
-                const auto non_broadcast_v = svld1(pg, non_broadcast_input_ptr + x);
-                auto       res             = is_sat ? wrapper::svqadd(broadcast_value_vec, non_broadcast_v) : svadd_z(pg, broadcast_value_vec, non_broadcast_v);
-                svst1(pg, output_ptr + x, res);
-
-                x += wrapper::svcnt<ScalarType>();
-                pg = wrapper::svwhilelt<ScalarType>(x, window_end_x);
-            }
-            while(svptest_any(all_true_pg, pg));
-        },
-        broadcast_input, non_broadcast_input, output);
-    }
-    else
-    {
-        // Clear X Dimension on execution window as we handle manually
-        input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
-        input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
-
-        Iterator input1(src0, input1_win);
-        Iterator input2(src1, input2_win);
-        Iterator output(dst, win);
-
-        execute_window_loop(win, [&](const Coordinates &)
-        {
-            const auto input1_ptr = reinterpret_cast<const ScalarType *>(input1.ptr());
-            const auto input2_ptr = reinterpret_cast<const ScalarType *>(input2.ptr());
-            const auto output_ptr = reinterpret_cast<ScalarType *>(output.ptr());
-
-            int      x  = window_start_x;
-            svbool_t pg = wrapper::svwhilelt<ScalarType>(x, window_end_x);
-            do
-            {
-                const auto val1 = svld1(pg, input1_ptr + x);
-                const auto val2 = svld1(pg, input2_ptr + x);
-                const auto res  = is_sat ? wrapper::svqadd(val1, val2) : svadd_z(pg, val1, val2);
-                svst1(pg, output_ptr + x, res);
-
-                x += wrapper::svcnt<ScalarType>();
-                pg = wrapper::svwhilelt<ScalarType>(x, window_end_x);
-            }
-            while(svptest_any(all_true_pg, pg));
-        },
-        input1, input2, output);
-    }
-}
 } // namespace cpu
 } // namespace arm_compute
-#endif // defined(__ARM_FEATURE_SVE)
+#endif // defined(ENABLE_SVE)
 #endif // SRC_CORE_SVE_KERNELS_ADD_LIST_H
\ No newline at end of file