diff options
author | Felix Thomasmathibalan <felixjohnny.thomasmathibalan@arm.com> | 2023-09-27 17:46:17 +0100 |
---|---|---|
committer | felixjohnny.thomasmathibalan <felixjohnny.thomasmathibalan@arm.com> | 2023-09-28 12:08:05 +0000 |
commit | afd38f0c617d6f89b2b4532c6c44f116617e2b6f (patch) | |
tree | 03bc7d5a762099989b16a656fa8d397b490ed70e /src/cpu/kernels/select | |
parent | bdcb4c148ee2fdeaaddf4cf1e57bbb0de02bb894 (diff) | |
download | ComputeLibrary-afd38f0c617d6f89b2b4532c6c44f116617e2b6f.tar.gz |
Apply clang-format on repository
Code is formatted as per a revised clang format configuration
file(not part of this delivery). Version 14.0.6 is used.
Exclusion List:
- files with .cl extension
- files that are not strictly C/C++ (e.g. Android.bp, Sconscript ...)
And the following directories
- compute_kernel_writer/validation/
- tests/
- include/
- src/core/NEON/kernels/convolution/
- src/core/NEON/kernels/arm_gemm/
- src/core/NEON/kernels/arm_conv/
- data/
There will be a follow up for formatting of .cl files and the
files under tests/ and compute_kernel_writer/validation/.
Signed-off-by: Felix Thomasmathibalan <felixjohnny.thomasmathibalan@arm.com>
Change-Id: Ib7eb1fcf4e7537b9feaefcfc15098a804a3fde0a
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/10391
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Diffstat (limited to 'src/cpu/kernels/select')
-rw-r--r-- | src/cpu/kernels/select/generic/neon/fp16.cpp | 12 | ||||
-rw-r--r-- | src/cpu/kernels/select/generic/neon/fp32.cpp | 10 | ||||
-rw-r--r-- | src/cpu/kernels/select/generic/neon/impl.h | 111 | ||||
-rw-r--r-- | src/cpu/kernels/select/generic/neon/integer.cpp | 40 |
4 files changed, 105 insertions, 68 deletions
diff --git a/src/cpu/kernels/select/generic/neon/fp16.cpp b/src/cpu/kernels/select/generic/neon/fp16.cpp index b460213c72..38a58099bd 100644 --- a/src/cpu/kernels/select/generic/neon/fp16.cpp +++ b/src/cpu/kernels/select/generic/neon/fp16.cpp @@ -23,20 +23,22 @@ */ #if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) -#include "src/cpu/kernels/select/generic/neon/impl.h" - #include "arm_compute/core/Helpers.h" + #include "src/core/NEON/wrapper/wrapper.h" +#include "src/cpu/kernels/select/generic/neon/impl.h" namespace arm_compute { namespace cpu { -void neon_f16_select_same_rank(const ITensor *c, const ITensor *x, const ITensor *y, ITensor *output, const Window &window) +void neon_f16_select_same_rank( + const ITensor *c, const ITensor *x, const ITensor *y, ITensor *output, const Window &window) { return select_op_16<float16_t, uint16x8_t>(c, x, y, output, window); } -void neon_f16_select_not_same_rank(const ITensor *c, const ITensor *x, const ITensor *y, ITensor *output, const Window &window) +void neon_f16_select_not_same_rank( + const ITensor *c, const ITensor *x, const ITensor *y, ITensor *output, const Window &window) { return select_op_not_same_rank<float16_t>(c, x, y, output, window); } @@ -45,4 +47,4 @@ void neon_f16_select_not_same_rank(const ITensor *c, const ITensor *x, const ITe } // namespace arm_compute -#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */
\ No newline at end of file +#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */ diff --git a/src/cpu/kernels/select/generic/neon/fp32.cpp b/src/cpu/kernels/select/generic/neon/fp32.cpp index 63fd594901..50a80cb338 100644 --- a/src/cpu/kernels/select/generic/neon/fp32.cpp +++ b/src/cpu/kernels/select/generic/neon/fp32.cpp @@ -22,20 +22,22 @@ * SOFTWARE. */ -#include "src/cpu/kernels/select/generic/neon/impl.h" - #include "arm_compute/core/Helpers.h" + #include "src/core/NEON/wrapper/wrapper.h" +#include "src/cpu/kernels/select/generic/neon/impl.h" namespace arm_compute { namespace cpu { -void neon_f32_select_same_rank(const ITensor *c, const ITensor *x, const ITensor *y, ITensor *output, const Window &window) +void neon_f32_select_same_rank( + const ITensor *c, const ITensor *x, const ITensor *y, ITensor *output, const Window &window) { return select_op_32<float, uint32x4_t>(c, x, y, output, window); } -void neon_f32_select_not_same_rank(const ITensor *c, const ITensor *x, const ITensor *y, ITensor *output, const Window &window) +void neon_f32_select_not_same_rank( + const ITensor *c, const ITensor *x, const ITensor *y, ITensor *output, const Window &window) { return select_op_not_same_rank<float>(c, x, y, output, window); } diff --git a/src/cpu/kernels/select/generic/neon/impl.h b/src/cpu/kernels/select/generic/neon/impl.h index 6a6d9969f8..7ce640b6ff 100644 --- a/src/cpu/kernels/select/generic/neon/impl.h +++ b/src/cpu/kernels/select/generic/neon/impl.h @@ -25,6 +25,7 @@ #define ACL_SRC_CPU_KERNELS_SELECT_GENERIC_NEON_IMPL_H #include "arm_compute/core/TensorInfo.h" + #include "src/core/NEON/NEAsymm.h" #include "src/cpu/kernels/select/generic/neon/impl.h" @@ -37,8 +38,16 @@ namespace arm_compute namespace cpu { template <typename ScalarType, typename VectorType> -void select_op(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window, - const int window_step_x, const int window_start_x, const int window_end_x, const int limit, VectorType (*condition_conversion)(const uint8_t *)) +void select_op(const ITensor *cond, + const ITensor *in1, + const ITensor *in2, + ITensor *out, + const Window &window, + const int window_step_x, + const int window_start_x, + const int window_end_x, + const int limit, + VectorType (*condition_conversion)(const uint8_t *)) { Window win = window; win.set(Window::DimX, Window::Dimension(0, 1, 1)); @@ -48,30 +57,32 @@ void select_op(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITen Iterator input2(in2, win); Iterator output(out, win); - execute_window_loop(win, [&](const Coordinates &) - { - auto output_ptr = reinterpret_cast<ScalarType *>(output.ptr()); - const auto condition_ptr = reinterpret_cast<const uint8_t *>(condition.ptr()); - const auto input1_ptr = reinterpret_cast<const ScalarType *>(input1.ptr()); - const auto input2_ptr = reinterpret_cast<const ScalarType *>(input2.ptr()); - - int x = window_start_x; - for(; x <= limit; x += window_step_x) + execute_window_loop( + win, + [&](const Coordinates &) { - const auto c = (*condition_conversion)(condition_ptr + x); - const auto a = wrapper::vloadq(input1_ptr + x); - const auto b = wrapper::vloadq(input2_ptr + x); - wrapper::vstore(output_ptr + x, wrapper::vbsl(c, a, b)); - } - for(; x < window_end_x; ++x) - { - const auto c = *(condition_ptr + x); - const auto a = *(input1_ptr + x); - const auto b = *(input2_ptr + x); - *(output_ptr + x) = static_cast<bool>(c) ? a : b; - } - }, - condition, input1, input2, output); + auto output_ptr = reinterpret_cast<ScalarType *>(output.ptr()); + const auto condition_ptr = reinterpret_cast<const uint8_t *>(condition.ptr()); + const auto input1_ptr = reinterpret_cast<const ScalarType *>(input1.ptr()); + const auto input2_ptr = reinterpret_cast<const ScalarType *>(input2.ptr()); + + int x = window_start_x; + for (; x <= limit; x += window_step_x) + { + const auto c = (*condition_conversion)(condition_ptr + x); + const auto a = wrapper::vloadq(input1_ptr + x); + const auto b = wrapper::vloadq(input2_ptr + x); + wrapper::vstore(output_ptr + x, wrapper::vbsl(c, a, b)); + } + for (; x < window_end_x; ++x) + { + const auto c = *(condition_ptr + x); + const auto a = *(input1_ptr + x); + const auto b = *(input2_ptr + x); + *(output_ptr + x) = static_cast<bool>(c) ? a : b; + } + }, + condition, input1, input2, output); } template <typename ScalarType, typename VectorType> @@ -81,11 +92,14 @@ void select_op_8(const ITensor *cond, const ITensor *in1, const ITensor *in2, IT const auto window_start_x = static_cast<int>(window.x().start()); const auto window_end_x = static_cast<int>(window.x().end()); - select_op<ScalarType, VectorType>(cond, in1, in2, out, window, window_step_x, window_start_x, window_end_x, window_end_x - window_step_x, [](const uint8_t *condition_ptr) -> VectorType - { - static const auto zero = wrapper::vdup_n(static_cast<uint8_t>(0), arm_compute::wrapper::traits::vector_128_tag()); - return wrapper::vcgt(wrapper::vloadq(condition_ptr), zero); - }); + select_op<ScalarType, VectorType>( + cond, in1, in2, out, window, window_step_x, window_start_x, window_end_x, window_end_x - window_step_x, + [](const uint8_t *condition_ptr) -> VectorType + { + static const auto zero = + wrapper::vdup_n(static_cast<uint8_t>(0), arm_compute::wrapper::traits::vector_128_tag()); + return wrapper::vcgt(wrapper::vloadq(condition_ptr), zero); + }); } template <typename ScalarType, typename VectorType> @@ -95,11 +109,14 @@ void select_op_16(const ITensor *cond, const ITensor *in1, const ITensor *in2, I const auto window_start_x = static_cast<int>(window.x().start()); const auto window_end_x = static_cast<int>(window.x().end()); - select_op<ScalarType, VectorType>(cond, in1, in2, out, window, window_step_x, window_start_x, window_end_x, window_end_x - window_step_x, [](const uint8_t *condition_ptr) -> VectorType - { - static const auto zero = wrapper::vdup_n(static_cast<uint16_t>(0), arm_compute::wrapper::traits::vector_128_tag()); - return wrapper::vcgt(wrapper::vmovl(wrapper::vload(condition_ptr)), zero); - }); + select_op<ScalarType, VectorType>( + cond, in1, in2, out, window, window_step_x, window_start_x, window_end_x, window_end_x - window_step_x, + [](const uint8_t *condition_ptr) -> VectorType + { + static const auto zero = + wrapper::vdup_n(static_cast<uint16_t>(0), arm_compute::wrapper::traits::vector_128_tag()); + return wrapper::vcgt(wrapper::vmovl(wrapper::vload(condition_ptr)), zero); + }); } template <typename ScalarType, typename VectorType> @@ -109,15 +126,19 @@ void select_op_32(const ITensor *cond, const ITensor *in1, const ITensor *in2, I const auto window_start_x = static_cast<int>(window.x().start()); const auto window_end_x = static_cast<int>(window.x().end()); - select_op<ScalarType, VectorType>(cond, in1, in2, out, window, window_step_x, window_start_x, window_end_x, window_end_x - window_step_x, [](const uint8_t *condition_ptr) -> VectorType - { - static const auto zero = wrapper::vdup_n(static_cast<uint32_t>(0), arm_compute::wrapper::traits::vector_128_tag()); - return wrapper::vcgt(wrapper::vmovl(wrapper::vgetlow(wrapper::vmovl(wrapper::vload(condition_ptr)))), zero); - }); + select_op<ScalarType, VectorType>( + cond, in1, in2, out, window, window_step_x, window_start_x, window_end_x, window_end_x - window_step_x, + [](const uint8_t *condition_ptr) -> VectorType + { + static const auto zero = + wrapper::vdup_n(static_cast<uint32_t>(0), arm_compute::wrapper::traits::vector_128_tag()); + return wrapper::vcgt(wrapper::vmovl(wrapper::vgetlow(wrapper::vmovl(wrapper::vload(condition_ptr)))), zero); + }); } template <typename ScalarType> -void select_op_not_same_rank(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) +void select_op_not_same_rank( + const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) { ARM_COMPUTE_UNUSED(window); @@ -131,20 +152,20 @@ void select_op_not_same_rank(const ITensor *cond, const ITensor *in1, const ITen int offset = 0; const int step = 16 / in1->info()->element_size(); - for(int i = 0; i < outer_size; ++i) + for (int i = 0; i < outer_size; ++i) { int x = offset; const auto input_ptr = static_cast<bool>(*(condition_ptr + i)) ? input1_ptr : input2_ptr; - for(; x <= offset + inner_size - step; x += step) + for (; x <= offset + inner_size - step; x += step) { wrapper::vstore(output_ptr + x, wrapper::vloadq(input_ptr + x)); } - if(x <= offset + inner_size - (step / 2)) + if (x <= offset + inner_size - (step / 2)) { wrapper::vstore(output_ptr + x, wrapper::vload(input_ptr + x)); x += step / 2; } - for(; x < offset + inner_size; ++x) + for (; x < offset + inner_size; ++x) { *(output_ptr + x) = *(input_ptr + x); } diff --git a/src/cpu/kernels/select/generic/neon/integer.cpp b/src/cpu/kernels/select/generic/neon/integer.cpp index 71b2f0b933..135087c261 100644 --- a/src/cpu/kernels/select/generic/neon/integer.cpp +++ b/src/cpu/kernels/select/generic/neon/integer.cpp @@ -25,59 +25,71 @@ #include "arm_compute/core/TensorInfo.h" #include "arm_compute/core/Types.h" -#include <arm_neon.h> - #include "src/cpu/kernels/select/generic/neon/impl.h" +#include <arm_neon.h> + namespace arm_compute { namespace cpu { -void neon_s8_select_same_rank(const ITensor *c, const ITensor *x, const ITensor *y, ITensor *output, const Window &window) +void neon_s8_select_same_rank( + const ITensor *c, const ITensor *x, const ITensor *y, ITensor *output, const Window &window) { return select_op_8<int8_t, uint8x16_t>(c, x, y, output, window); } -void neon_s16_select_same_rank(const ITensor *c, const ITensor *x, const ITensor *y, ITensor *output, const Window &window) +void neon_s16_select_same_rank( + const ITensor *c, const ITensor *x, const ITensor *y, ITensor *output, const Window &window) { return select_op_16<int16_t, uint16x8_t>(c, x, y, output, window); } -void neon_s32_select_same_rank(const ITensor *c, const ITensor *x, const ITensor *y, ITensor *output, const Window &window) +void neon_s32_select_same_rank( + const ITensor *c, const ITensor *x, const ITensor *y, ITensor *output, const Window &window) { return select_op_32<int32_t, uint32x4_t>(c, x, y, output, window); } -void neon_s8_select_not_same_rank(const ITensor *c, const ITensor *x, const ITensor *y, ITensor *output, const Window &window) +void neon_s8_select_not_same_rank( + const ITensor *c, const ITensor *x, const ITensor *y, ITensor *output, const Window &window) { return select_op_not_same_rank<int8_t>(c, x, y, output, window); } -void neon_s16_select_not_same_rank(const ITensor *c, const ITensor *x, const ITensor *y, ITensor *output, const Window &window) +void neon_s16_select_not_same_rank( + const ITensor *c, const ITensor *x, const ITensor *y, ITensor *output, const Window &window) { return select_op_not_same_rank<int16_t>(c, x, y, output, window); } -void neon_s32_select_not_same_rank(const ITensor *c, const ITensor *x, const ITensor *y, ITensor *output, const Window &window) +void neon_s32_select_not_same_rank( + const ITensor *c, const ITensor *x, const ITensor *y, ITensor *output, const Window &window) { return select_op_not_same_rank<int32_t>(c, x, y, output, window); } -void neon_u8_select_same_rank(const ITensor *c, const ITensor *x, const ITensor *y, ITensor *output, const Window &window) +void neon_u8_select_same_rank( + const ITensor *c, const ITensor *x, const ITensor *y, ITensor *output, const Window &window) { return select_op_8<uint8_t, uint8x16_t>(c, x, y, output, window); } -void neon_u16_select_same_rank(const ITensor *c, const ITensor *x, const ITensor *y, ITensor *output, const Window &window) +void neon_u16_select_same_rank( + const ITensor *c, const ITensor *x, const ITensor *y, ITensor *output, const Window &window) { return select_op_16<uint16_t, uint16x8_t>(c, x, y, output, window); } -void neon_u32_select_same_rank(const ITensor *c, const ITensor *x, const ITensor *y, ITensor *output, const Window &window) +void neon_u32_select_same_rank( + const ITensor *c, const ITensor *x, const ITensor *y, ITensor *output, const Window &window) { return select_op_32<uint32_t, uint32x4_t>(c, x, y, output, window); } -void neon_u8_select_not_same_rank(const ITensor *c, const ITensor *x, const ITensor *y, ITensor *output, const Window &window) +void neon_u8_select_not_same_rank( + const ITensor *c, const ITensor *x, const ITensor *y, ITensor *output, const Window &window) { return select_op_not_same_rank<uint8_t>(c, x, y, output, window); } -void neon_u16_select_not_same_rank(const ITensor *c, const ITensor *x, const ITensor *y, ITensor *output, const Window &window) +void neon_u16_select_not_same_rank( + const ITensor *c, const ITensor *x, const ITensor *y, ITensor *output, const Window &window) { return select_op_not_same_rank<uint16_t>(c, x, y, output, window); } -void neon_u32_select_not_same_rank(const ITensor *c, const ITensor *x, const ITensor *y, ITensor *output, const Window &window) +void neon_u32_select_not_same_rank( + const ITensor *c, const ITensor *x, const ITensor *y, ITensor *output, const Window &window) { return select_op_not_same_rank<uint32_t>(c, x, y, output, window); } |