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, 75 deletions

diff --git a/src/core/cpu/kernels/elementwise/sve/elementwise_unary_list.h b/src/core/cpu/kernels/elementwise/sve/elementwise_unary_list.h
index 23502c71e5..63490421e9 100644
--- a/src/core/cpu/kernels/elementwise/sve/elementwise_unary_list.h
+++ b/src/core/cpu/kernels/elementwise/sve/elementwise_unary_list.h
@@ -25,87 +25,15 @@
 #define SRC_CORE_SVE_KERNELS_ELEMENTWISE_UNARY_LIST_H
 
 #include "arm_compute/core/Types.h"
-#include "arm_compute/core/utils/misc/Traits.h"
-#include "src/core/NEON/wrapper/intrinsics/intrinsics.h"
-#if defined(__ARM_FEATURE_SVE)
-#include "src/core/NEON/SVEMath.h"
-#include <arm_sve.h>
+#if defined(ENABLE_SVE)
 
 namespace arm_compute
 {
 namespace cpu
 {
-template <typename ScalarType, typename VectorType>
-inline typename std::enable_if<utils::traits::is_floating_point<ScalarType>::value, VectorType>::type elementwise_op_sve_imp(svbool_t pg, ElementWiseUnary op, const VectorType &a)
-{
-    switch(op)
-    {
-        case ElementWiseUnary::RSQRT:
-            return svinvsqrt(pg, a);
-        case ElementWiseUnary::EXP:
-            return wrapper::svexp_z(pg, a);
-        case ElementWiseUnary::NEG:
-            return svneg_z(pg, a);
-        case ElementWiseUnary::LOG:
-            return wrapper::svlog_z(pg, a);
-        case ElementWiseUnary::ABS:
-            return svabs_z(pg, a);
-        case ElementWiseUnary::ROUND:
-            return svrintn_z(pg, a);
-        case ElementWiseUnary::SIN:
-            return wrapper::svsin_z(pg, a);
-        default:
-            ARM_COMPUTE_ERROR("NOT_SUPPORTED");
-    }
-}
-
-template <typename ScalarType, typename VectorType>
-inline typename std::enable_if<std::is_integral<ScalarType>::value, VectorType>::type elementwise_op_sve_imp(svbool_t pg, ElementWiseUnary op, const VectorType &a)
-{
-    switch(op)
-    {
-        case ElementWiseUnary::NEG:
-            return svneg_z(pg, a);
-        case ElementWiseUnary::ABS:
-            return svabs_z(pg, a);
-        default:
-            ARM_COMPUTE_ERROR("NOT_SUPPORTED");
-    }
-}
-
 template <typename ScalarType>
-void elementwise_sve_op(const ITensor *in, ITensor *out, const Window &window, ElementWiseUnary op)
-{
-    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());
-
-    Window win = window;
-    win.set(Window::DimX, Window::Dimension(0, 1, 1));
-
-    Iterator input(in, win);
-    Iterator output(out, win);
-
-    execute_window_loop(win, [&](const Coordinates &)
-    {
-        auto       output_ptr = reinterpret_cast<ScalarType *>(output.ptr());
-        const auto input_ptr  = reinterpret_cast<const ScalarType *>(input.ptr());
-        int        x          = window_start_x;
-
-        svbool_t pg = wrapper::svwhilelt<ScalarType>(x, window_end_x);
-        do
-        {
-            const auto vin = svld1(pg, input_ptr + x);
-            svst1(pg, output_ptr + x, elementwise_op_sve_imp<ScalarType, decltype(vin)>(pg, op, vin));
-            x += wrapper::svcnt<ScalarType>();
-            pg = wrapper::svwhilelt<ScalarType>(x, window_end_x);
-        }
-        while(svptest_any(all_true_pg, pg));
-    },
-    input, output);
-}
-
+void elementwise_sve_op(const ITensor *in, ITensor *out, const Window &window, ElementWiseUnary op);
 } // namespace cpu
 } // namespace arm_compute
-#endif // defined(__ARM_FEATURE_SVE)
+#endif // defined(ENABLE_SVE)
 #endif // SRC_CORE_NEON_KERNELS_ELEMENTWISE_UNARY_LIST_H
\ No newline at end of file