aboutsummaryrefslogtreecommitdiff
path: root/src/cpu/kernels/add/generic/sve/impl.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/cpu/kernels/add/generic/sve/impl.cpp')
-rw-r--r--src/cpu/kernels/add/generic/sve/impl.cpp106
1 files changed, 59 insertions, 47 deletions
diff --git a/src/cpu/kernels/add/generic/sve/impl.cpp b/src/cpu/kernels/add/generic/sve/impl.cpp
index e8606436fd..ca850fcef4 100644
--- a/src/cpu/kernels/add/generic/sve/impl.cpp
+++ b/src/cpu/kernels/add/generic/sve/impl.cpp
@@ -23,17 +23,21 @@
*/
#include "src/cpu/kernels/add/generic/sve/impl.h"
+
#include "arm_compute/core/Helpers.h"
#include "arm_compute/core/utils/misc/Traits.h"
+
#include "src/core/NEON/SVEMath.h"
#include "src/core/NEON/wrapper/intrinsics/intrinsics.h"
+
#include <arm_sve.h>
namespace arm_compute
{
namespace cpu
{
template <typename ScalarType>
-void add_same_sve(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
+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());
@@ -53,7 +57,7 @@ void add_same_sve(const ITensor *src0, const ITensor *src1, ITensor *dst, const
Iterator input2(src1, window.broadcast_if_dimension_le_one(src1->info()->tensor_shape()));
Iterator output(dst, window);
- if(is_broadcast_across_x)
+ 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;
@@ -68,28 +72,30 @@ void add_same_sve(const ITensor *src0, const ITensor *src1, ITensor *dst, const
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());
+ 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);
+ 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);
+ 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
{
@@ -101,35 +107,41 @@ void add_same_sve(const ITensor *src0, const ITensor *src1, ITensor *dst, const
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
+ execute_window_loop(
+ win,
+ [&](const Coordinates &)
{
- 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);
+ 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);
}
}
-template void add_same_sve<float>(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window);
-template void add_same_sve<uint8_t>(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window);
-template void add_same_sve<int16_t>(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window);
-template void add_same_sve<int32_t>(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window);
+template void add_same_sve<float>(
+ const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window);
+template void add_same_sve<uint8_t>(
+ const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window);
+template void add_same_sve<int16_t>(
+ const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window);
+template void add_same_sve<int32_t>(
+ const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window);
#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS)
-template void add_same_sve<float16_t>(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window);
+template void add_same_sve<float16_t>(
+ const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window);
#endif /* (__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */
} // namespace cpu
} // namespace arm_compute
generated by cgit v1.2.1 (git 2.23.0) at 2024-05-20 00:51:51 +0000