From 6d878878f06b39bcdafff3e9569c807b1bf9c34f Mon Sep 17 00:00:00 2001 From: Pablo Marquez Tello Date: Fri, 22 Sep 2023 10:29:46 +0100 Subject: Select changes to enable fp16 in armv8a multi_isa builds * Code guarded with __ARM_FEATURE_FP16_VECTOR_ARITHMETIC needs to be moved to an fp16.cpp file to allow compilation with -march=armv8.2-a+fp16 * fp16.cpp needs to use the template select_op() which had to be moved from impl.cpp to fp16.cpp * Partially resolves MLCE-1102 Change-Id: Ic9e73e121482fcc5e4fcbe8ae1ecd23649cbd3d1 Signed-off-by: Pablo Marquez Tello Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/10359 Benchmark: Arm Jenkins Tested-by: Arm Jenkins Reviewed-by: Jakub Sujak --- src/cpu/kernels/select/generic/neon/impl.h | 134 +++++++++++++++++++++++++---- 1 file changed, 118 insertions(+), 16 deletions(-) (limited to 'src/cpu/kernels/select/generic/neon/impl.h') diff --git a/src/cpu/kernels/select/generic/neon/impl.h b/src/cpu/kernels/select/generic/neon/impl.h index 2bbc38b744..6a6d9969f8 100644 --- a/src/cpu/kernels/select/generic/neon/impl.h +++ b/src/cpu/kernels/select/generic/neon/impl.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2022 Arm Limited. + * Copyright (c) 2022-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -21,34 +21,136 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#ifndef SRC_CORE_NEON_KERNELS_SELECT_IMPL_H -#define SRC_CORE_NEON_KERNELS_SELECT_IMPL_H +#ifndef ACL_SRC_CPU_KERNELS_SELECT_GENERIC_NEON_IMPL_H +#define ACL_SRC_CPU_KERNELS_SELECT_GENERIC_NEON_IMPL_H -#include +#include "arm_compute/core/TensorInfo.h" +#include "src/core/NEON/NEAsymm.h" +#include "src/cpu/kernels/select/generic/neon/impl.h" + +#include +#include +#include namespace arm_compute { -class ITensor; -class Window; - namespace cpu { -template -void select_op_not_same_rank(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window); - template -void select_op_32(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window); +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)); + + Iterator condition(cond, win); + Iterator input1(in1, win); + Iterator input2(in2, win); + Iterator output(out, win); + + execute_window_loop(win, [&](const Coordinates &) + { + auto output_ptr = reinterpret_cast(output.ptr()); + const auto condition_ptr = reinterpret_cast(condition.ptr()); + const auto input1_ptr = reinterpret_cast(input1.ptr()); + const auto input2_ptr = reinterpret_cast(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(c) ? a : b; + } + }, + condition, input1, input2, output); +} template -void select_op_16(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window); +void select_op_8(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) +{ + const auto window_step_x = 16 / sizeof(ScalarType); + const auto window_start_x = static_cast(window.x().start()); + const auto window_end_x = static_cast(window.x().end()); + + select_op(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(0), arm_compute::wrapper::traits::vector_128_tag()); + return wrapper::vcgt(wrapper::vloadq(condition_ptr), zero); + }); +} template -void select_op_8(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window); +void select_op_16(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) +{ + const auto window_step_x = 16 / sizeof(ScalarType); + const auto window_start_x = static_cast(window.x().start()); + const auto window_end_x = static_cast(window.x().end()); + + select_op(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(0), arm_compute::wrapper::traits::vector_128_tag()); + return wrapper::vcgt(wrapper::vmovl(wrapper::vload(condition_ptr)), zero); + }); +} template -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_32(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) +{ + const auto window_step_x = 16 / sizeof(ScalarType); + const auto window_start_x = static_cast(window.x().start()); + const auto window_end_x = static_cast(window.x().end()); + + select_op(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(0), arm_compute::wrapper::traits::vector_128_tag()); + return wrapper::vcgt(wrapper::vmovl(wrapper::vgetlow(wrapper::vmovl(wrapper::vload(condition_ptr)))), zero); + }); +} + +template +void select_op_not_same_rank(const ITensor *cond, const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) +{ + ARM_COMPUTE_UNUSED(window); + + auto output_ptr = reinterpret_cast(out->buffer()); + const auto condition_ptr = reinterpret_cast(cond->buffer()); + const auto input1_ptr = reinterpret_cast(in1->buffer()); + const auto input2_ptr = reinterpret_cast(in2->buffer()); + + const int outer_size = cond->info()->total_size() / cond->info()->element_size(); + const int inner_size = (in1->info()->total_size() / in1->info()->element_size()) / outer_size; + int offset = 0; + const int step = 16 / in1->info()->element_size(); + for(int i = 0; i < outer_size; ++i) + { + int x = offset; + const auto input_ptr = static_cast(*(condition_ptr + i)) ? input1_ptr : input2_ptr; + 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)) + { + wrapper::vstore(output_ptr + x, wrapper::vload(input_ptr + x)); + x += step / 2; + } + for(; x < offset + inner_size; ++x) + { + *(output_ptr + x) = *(input_ptr + x); + } + offset += inner_size; + } +} } // namespace cpu } // namespace arm_compute -#endif //SRC_CORE_NEON_KERNELS_SELECT_IMPL_H \ No newline at end of file +#endif // ACL_SRC_CPU_KERNELS_SELECT_GENERIC_NEON_IMPL_H -- cgit v1.2.1