CpuSubKernel changes to enable fp16 in armv8a multi_isa builds

    * FP16 kernels must be instantiated in fp16.cpp.

    * Partially resolves MLCE-1102

Change-Id: I497fe0ba6e84493a5072c3e80bbba7ecd5de8095
Signed-off-by: Pablo Marquez Tello <pablo.tello@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/10448
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>

8 files changed, 227 insertions, 132 deletions

diff --git a/Android.bp b/Android.bp
index 1bd70c4f36..5760b7d08e 100644
--- a/Android.bp
+++ b/Android.bp
@@ -569,6 +569,7 @@ cc_library_static {
         "src/cpu/kernels/softmax/generic/neon/impl.cpp",
         "src/cpu/kernels/softmax/generic/neon/qasymm8.cpp",
         "src/cpu/kernels/softmax/generic/neon/qasymm8_signed.cpp",
+        "src/cpu/kernels/sub/neon/fp16.cpp",
         "src/cpu/kernels/sub/neon/qasymm8.cpp",
         "src/cpu/kernels/sub/neon/qasymm8_signed.cpp",
         "src/cpu/kernels/sub/neon/qsymm16.cpp",
diff --git a/filelist.json b/filelist.json
index bb90fcf4bb..aeb54c6d9d 100644
--- a/filelist.json
+++ b/filelist.json
@@ -2249,13 +2249,17 @@
           "common": [
             "src/cpu/operators/CpuSub.cpp",
             "src/cpu/kernels/CpuSubKernel.cpp",
-            "src/runtime/NEON/functions/NEArithmeticSubtraction.cpp",
-            "src/cpu/kernels/sub/neon/qasymm8.cpp",
-            "src/cpu/kernels/sub/neon/qasymm8_signed.cpp",
-            "src/cpu/kernels/sub/neon/qsymm16.cpp"
-          ]
+            "src/runtime/NEON/functions/NEArithmeticSubtraction.cpp"
+          ],
+          "neon": {
+            "fp16":["src/cpu/kernels/sub/neon/fp16.cpp"],
+            "qasymm8": ["src/cpu/kernels/sub/neon/qasymm8.cpp"],
+            "qasymm8_signed": ["src/cpu/kernels/sub/neon/qasymm8_signed.cpp"],
+            "qsymm16": ["src/cpu/kernels/sub/neon/qsymm16.cpp"]
+          }
         }
       },
+
       "Tile": {
         "files": {
           "common": [
diff --git a/src/BUILD.bazel b/src/BUILD.bazel
index 2c56f26c11..7c644fc25f 100644
--- a/src/BUILD.bazel
+++ b/src/BUILD.bazel
@@ -818,6 +818,7 @@ filegroup(
 	"cpu/kernels/softmax/generic/neon/impl.cpp",
 	"cpu/kernels/softmax/generic/neon/qasymm8.cpp",
 	"cpu/kernels/softmax/generic/neon/qasymm8_signed.cpp",
+	"cpu/kernels/sub/neon/fp16.cpp",
 	"cpu/kernels/sub/neon/qasymm8.cpp",
 	"cpu/kernels/sub/neon/qasymm8_signed.cpp",
 	"cpu/kernels/sub/neon/qsymm16.cpp",
diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
index 4db76b1328..103f2d7c62 100644
--- a/src/CMakeLists.txt
+++ b/src/CMakeLists.txt
@@ -809,6 +809,7 @@ target_sources(
 	cpu/kernels/softmax/generic/neon/impl.cpp
 	cpu/kernels/softmax/generic/neon/qasymm8.cpp
 	cpu/kernels/softmax/generic/neon/qasymm8_signed.cpp
+	cpu/kernels/sub/neon/fp16.cpp
 	cpu/kernels/sub/neon/qasymm8.cpp
 	cpu/kernels/sub/neon/qasymm8_signed.cpp
 	cpu/kernels/sub/neon/qsymm16.cpp
diff --git a/src/cpu/kernels/CpuSubKernel.cpp b/src/cpu/kernels/CpuSubKernel.cpp
index 2b2c6f2e92..c8706ff651 100644
--- a/src/cpu/kernels/CpuSubKernel.cpp
+++ b/src/cpu/kernels/CpuSubKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2021-2022 Arm Limited.
+ * Copyright (c) 2021-2023 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -31,6 +31,7 @@
 #include "src/core/helpers/AutoConfiguration.h"
 #include "src/core/helpers/WindowHelpers.h"
 #include "src/cpu/kernels/add/generic/neon/impl.h"
+#include "src/cpu/kernels/sub/neon/impl.h"
 #include "src/cpu/kernels/sub/neon/list.h"
 
 #if defined(ENABLE_FP32_KERNELS)
@@ -57,7 +58,7 @@ static const std::vector<CpuSubKernel::SubKernel> available_kernels = {
      REGISTER_FP32_NEON(arm_compute::cpu::sub_same_neon<float>)},
     {"neon_fp16_sub",
      [](const CpuSubKernelDataTypeISASelectorData &data) { return (data.dt == DataType::F16) && data.isa.fp16; },
-     REGISTER_FP16_NEON(arm_compute::cpu::sub_same_neon<float16_t>)},
+     REGISTER_FP16_NEON(arm_compute::cpu::sub_same_neon_fp16)},
     {"neon_u8_sub", [](const CpuSubKernelDataTypeISASelectorData &data) { return (data.dt == DataType::U8); },
      REGISTER_INTEGER_NEON(arm_compute::cpu::sub_same_neon<uint8_t>)},
     {"neon_s16_sub", [](const CpuSubKernelDataTypeISASelectorData &data) { return (data.dt == DataType::S16); },
diff --git a/src/cpu/kernels/sub/neon/fp16.cpp b/src/cpu/kernels/sub/neon/fp16.cpp
new file mode 100644
index 0000000000..023068817b
--- /dev/null
+++ b/src/cpu/kernels/sub/neon/fp16.cpp
@@ -0,0 +1,43 @@
+/*
+ * Copyright (c) 2023 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/ITensor.h"
+#include "arm_compute/core/Types.h"
+
+#include "src/cpu/kernels/sub/neon/impl.h"
+
+#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS)
+
+namespace arm_compute
+{
+namespace cpu
+{
+void sub_same_neon_fp16(
+    const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
+{
+    sub_same_neon<float16_t>(src0, src1, dst, policy, window);
+}
+} // namespace cpu
+} // namespace arm_compute
+
+#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */
diff --git a/src/cpu/kernels/sub/neon/impl.h b/src/cpu/kernels/sub/neon/impl.h
new file mode 100644
index 0000000000..6123f7e25a
--- /dev/null
+++ b/src/cpu/kernels/sub/neon/impl.h
@@ -0,0 +1,164 @@
+/*
+ * Copyright (c) 2021-2023 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.
+ */
+
+#ifndef ACL_SRC_CPU_KERNELS_SUB_NEON_IMPL_H
+#define ACL_SRC_CPU_KERNELS_SUB_NEON_IMPL_H
+
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/utils/misc/Traits.h"
+
+#include "src/core/helpers/WindowHelpers.h"
+#include "src/core/NEON/wrapper/intrinsics/intrinsics.h"
+#include "src/core/NEON/wrapper/scalar/sub.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+template <typename T>
+void sub_same_neon(
+    const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
+{
+    /** SIMD vector tag type. */
+    using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t<T, wrapper::traits::BitWidth::W128>;
+
+    bool is_sat = policy == ConvertPolicy::SATURATE;
+
+    // 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());
+
+    // Clear X Dimension on execution window as we handle manually
+    Window win = window;
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    constexpr int window_step_x         = 16 / sizeof(T);
+    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();
+
+    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 T *>(non_broadcast_input.ptr());
+                const auto output_ptr              = reinterpret_cast<T *>(output.ptr());
+
+                const T    broadcast_value     = *reinterpret_cast<const T *>(broadcast_input.ptr());
+                const auto broadcast_value_vec = wrapper::vdup_n(broadcast_value, ExactTagType{});
+
+                // Compute S elements per iteration
+                int x = window_start_x;
+                for (; x <= (window_end_x - window_step_x); x += window_step_x)
+                {
+                    const auto non_broadcast_v = wrapper::vloadq(non_broadcast_input_ptr + x);
+                    auto       res             = is_sat ? wrapper::vqsub(broadcast_value_vec, non_broadcast_v)
+                                                        : wrapper::vsub(broadcast_value_vec, non_broadcast_v);
+                    if (is_broadcast_input_2)
+                    {
+                        res = wrapper::vmul(res, wrapper::vdup_n(static_cast<T>(-1), ExactTagType{}));
+                    }
+                    wrapper::vstore(output_ptr + x, res);
+                }
+
+                // Compute left-over elements
+                for (; x < window_end_x; ++x)
+                {
+                    const auto non_broadcast_v = *(non_broadcast_input_ptr + x);
+                    auto       res =
+                        is_sat ? wrapper::sub_sat(broadcast_value, non_broadcast_v) : broadcast_value - non_broadcast_v;
+                    if (is_broadcast_input_2)
+                    {
+                        res = static_cast<T>(-1) * res;
+                    }
+
+                    *(output_ptr + x) = res;
+                }
+            },
+            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 T *>(input1.ptr());
+                const auto input2_ptr = reinterpret_cast<const T *>(input2.ptr());
+                const auto output_ptr = reinterpret_cast<T *>(output.ptr());
+
+                // Compute S elements per iteration
+                int x = window_start_x;
+                for (; x <= (window_end_x - window_step_x); x += window_step_x)
+                {
+                    const auto val1 = wrapper::vloadq(input1_ptr + x);
+                    const auto val2 = wrapper::vloadq(input2_ptr + x);
+                    const auto res  = is_sat ? wrapper::vqsub(val1, val2) : wrapper::vsub(val1, val2);
+                    wrapper::vstore(output_ptr + x, res);
+                }
+
+                // Compute left-over elements
+                for (; x < window_end_x; ++x)
+                {
+                    const auto val1   = *(input1_ptr + x);
+                    const auto val2   = *(input2_ptr + x);
+                    *(output_ptr + x) = is_sat ? wrapper::sub_sat(val1, val2) : val1 - val2;
+                }
+            },
+            input1, input2, output);
+    }
+}
+} // namespace cpu
+} // namespace arm_compute
+
+#endif // ACL_SRC_CPU_KERNELS_SUB_NEON_IMPL_H
diff --git a/src/cpu/kernels/sub/neon/list.h b/src/cpu/kernels/sub/neon/list.h
index 9f6c92271f..f29571f122 100644
--- a/src/cpu/kernels/sub/neon/list.h
+++ b/src/cpu/kernels/sub/neon/list.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2021-2022 Arm Limited.
+ * Copyright (c) 2021-2023 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -21,8 +21,8 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#ifndef SRC_CORE_NEON_KERNELS_SUB_LIST_H
-#define SRC_CORE_NEON_KERNELS_SUB_LIST_H
+#ifndef ACL_SRC_CPU_KERNELS_SUB_NEON_LIST_H
+#define ACL_SRC_CPU_KERNELS_SUB_NEON_LIST_H
 
 #include "arm_compute/core/Types.h"
 #include "arm_compute/core/utils/misc/Traits.h"
@@ -42,129 +42,9 @@ DECLARE_SUB_KERNEL(sub_qasymm8_signed_neon_fixedpoint);
 DECLARE_SUB_KERNEL(sub_qasymm8_neon);
 DECLARE_SUB_KERNEL(sub_qasymm8_signed_neon);
 DECLARE_SUB_KERNEL(sub_qsymm16_neon);
+DECLARE_SUB_KERNEL(sub_same_neon_fp16);
 
 #undef DECLARE_SUB_KERNEL
-
-template <typename T>
-void sub_same_neon(
-    const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
-{
-    /** SIMD vector tag type. */
-    using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t<T, wrapper::traits::BitWidth::W128>;
-
-    bool is_sat = policy == ConvertPolicy::SATURATE;
-
-    // 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());
-
-    // Clear X Dimension on execution window as we handle manually
-    Window win = window;
-    win.set(Window::DimX, Window::Dimension(0, 1, 1));
-
-    constexpr int window_step_x         = 16 / sizeof(T);
-    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();
-
-    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 T *>(non_broadcast_input.ptr());
-                const auto output_ptr              = reinterpret_cast<T *>(output.ptr());
-
-                const T    broadcast_value     = *reinterpret_cast<const T *>(broadcast_input.ptr());
-                const auto broadcast_value_vec = wrapper::vdup_n(broadcast_value, ExactTagType{});
-
-                // Compute S elements per iteration
-                int x = window_start_x;
-                for (; x <= (window_end_x - window_step_x); x += window_step_x)
-                {
-                    const auto non_broadcast_v = wrapper::vloadq(non_broadcast_input_ptr + x);
-                    auto       res             = is_sat ? wrapper::vqsub(broadcast_value_vec, non_broadcast_v)
-                                                        : wrapper::vsub(broadcast_value_vec, non_broadcast_v);
-                    if (is_broadcast_input_2)
-                    {
-                        res = wrapper::vmul(res, wrapper::vdup_n(static_cast<T>(-1), ExactTagType{}));
-                    }
-                    wrapper::vstore(output_ptr + x, res);
-                }
-
-                // Compute left-over elements
-                for (; x < window_end_x; ++x)
-                {
-                    const auto non_broadcast_v = *(non_broadcast_input_ptr + x);
-                    auto       res =
-                        is_sat ? wrapper::sub_sat(broadcast_value, non_broadcast_v) : broadcast_value - non_broadcast_v;
-                    if (is_broadcast_input_2)
-                    {
-                        res = static_cast<T>(-1) * res;
-                    }
-
-                    *(output_ptr + x) = res;
-                }
-            },
-            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 T *>(input1.ptr());
-                const auto input2_ptr = reinterpret_cast<const T *>(input2.ptr());
-                const auto output_ptr = reinterpret_cast<T *>(output.ptr());
-
-                // Compute S elements per iteration
-                int x = window_start_x;
-                for (; x <= (window_end_x - window_step_x); x += window_step_x)
-                {
-                    const auto val1 = wrapper::vloadq(input1_ptr + x);
-                    const auto val2 = wrapper::vloadq(input2_ptr + x);
-                    const auto res  = is_sat ? wrapper::vqsub(val1, val2) : wrapper::vsub(val1, val2);
-                    wrapper::vstore(output_ptr + x, res);
-                }
-
-                // Compute left-over elements
-                for (; x < window_end_x; ++x)
-                {
-                    const auto val1   = *(input1_ptr + x);
-                    const auto val2   = *(input2_ptr + x);
-                    *(output_ptr + x) = is_sat ? wrapper::sub_sat(val1, val2) : val1 - val2;
-                }
-            },
-            input1, input2, output);
-    }
-}
 } // namespace cpu
 } // namespace arm_compute
-#endif // SRC_CORE_NEON_KERNELS_SUB_LIST_H
+#endif // ACL_SRC_CPU_KERNELS_SUB_NEON_LIST_H