diff options
Diffstat (limited to 'src/gpu/cl/kernels/ClScaleKernel.cpp')
| -rw-r--r-- | src/gpu/cl/kernels/ClScaleKernel.cpp | 68 |
1 files changed, 42 insertions, 26 deletions
diff --git a/src/gpu/cl/kernels/ClScaleKernel.cpp b/src/gpu/cl/kernels/ClScaleKernel.cpp index 4c4373a215..4305acad26 100644 --- a/src/gpu/cl/kernels/ClScaleKernel.cpp +++ b/src/gpu/cl/kernels/ClScaleKernel.cpp @@ -27,8 +27,9 @@ #include "arm_compute/core/TensorInfo.h" #include "arm_compute/core/Utils.h" #include "arm_compute/core/utils/helpers/AdjustVecSize.h" -#include "arm_compute/core/utils/StringUtils.h" #include "arm_compute/core/utils/InterpolationPolicyUtils.h" +#include "arm_compute/core/utils/StringUtils.h" + #include "src/core/AccessWindowStatic.h" #include "src/core/CL/CLValidate.h" #include "src/core/helpers/WindowHelpers.h" @@ -43,7 +44,8 @@ namespace kernels { namespace { -inline std::tuple<float, float> calculate_scale_factors(const ITensorInfo *src, const ITensorInfo *dst, DataLayout data_layout, bool align_corners) +inline std::tuple<float, float> +calculate_scale_factors(const ITensorInfo *src, const ITensorInfo *dst, DataLayout data_layout, bool align_corners) { const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); @@ -64,20 +66,25 @@ Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dst, const { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst); ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::U8, DataType::S16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, + DataType::U8, DataType::S16, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(src, dst); ARM_COMPUTE_RETURN_ERROR_ON(dst == src); - ARM_COMPUTE_RETURN_ERROR_ON(src->num_channels()!=1); - ARM_COMPUTE_RETURN_ERROR_ON(info.align_corners && !arm_compute::scale_utils::is_align_corners_allowed_sampling_policy(info.sampling_policy)); - ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized(src->data_type()) && !is_data_type_quantized_asymmetric(src->data_type())); + ARM_COMPUTE_RETURN_ERROR_ON(src->num_channels() != 1); + ARM_COMPUTE_RETURN_ERROR_ON( + info.align_corners && + !arm_compute::scale_utils::is_align_corners_allowed_sampling_policy(info.sampling_policy)); + ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized(src->data_type()) && + !is_data_type_quantized_asymmetric(src->data_type())); float scale_x = 0.f; float scale_y = 0.f; const DataLayout data_layout = info.data_layout == DataLayout::UNKNOWN ? src->data_layout() : info.data_layout; - std::tie(scale_x, scale_y) = calculate_scale_factors(src, dst, data_layout, info.align_corners); + std::tie(scale_x, scale_y) = calculate_scale_factors(src, dst, data_layout, info.align_corners); - ARM_COMPUTE_RETURN_ERROR_ON(info.interpolation_policy == InterpolationPolicy::AREA && (scale_x > 1.f || scale_y > 1.f)); + ARM_COMPUTE_RETURN_ERROR_ON(info.interpolation_policy == InterpolationPolicy::AREA && + (scale_x > 1.f || scale_y > 1.f)); return Status{}; } @@ -94,23 +101,26 @@ ClScaleKernel::ClScaleKernel() _type = CLKernelType::ELEMENTWISE; } -void ClScaleKernel::configure(const CLCompileContext &compile_context, ITensorInfo *src, ITensorInfo *dst, const ScaleKernelInfo &info) +void ClScaleKernel::configure(const CLCompileContext &compile_context, + ITensorInfo *src, + ITensorInfo *dst, + const ScaleKernelInfo &info) { ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst, info)); - auto padding_info = get_padding_info({ src, dst }); + auto padding_info = get_padding_info({src, dst}); // Info required for the static tuning _data_layout = info.data_layout == DataLayout::UNKNOWN ? src->data_layout() : info.data_layout; const bool is_nhwc = _data_layout == DataLayout::NHWC; - float scale_x = 0.f; - float scale_y = 0.f; + float scale_x = 0.f; + float scale_y = 0.f; std::tie(scale_x, scale_y) = calculate_scale_factors(src, dst, _data_layout, info.align_corners); // Area interpolation behaves as Nearest Neighbour in case of up-sampling auto interpolation_policy_to_use = info.interpolation_policy; - if(info.interpolation_policy == InterpolationPolicy::AREA && scale_x <= 1.f && scale_y <= 1.f) + if (info.interpolation_policy == InterpolationPolicy::AREA && scale_x <= 1.f && scale_y <= 1.f) { interpolation_policy_to_use = InterpolationPolicy::NEAREST_NEIGHBOR; } @@ -127,7 +137,7 @@ void ClScaleKernel::configure(const CLCompileContext &compile_context, ITensorIn unsigned int vec_size_leftover = 0; CLBuildOptions build_opts; - if(_data_layout == DataLayout::NHWC) + if (_data_layout == DataLayout::NHWC) { vec_size = adjust_vec_size(src->data_type() == DataType::F32 ? 4 : 8, dst_channels); vec_size_leftover = dst_channels % vec_size; @@ -135,7 +145,8 @@ void ClScaleKernel::configure(const CLCompileContext &compile_context, ITensorIn build_opts.add_option("-DSRC_DATA_TYPE=" + get_cl_type_from_data_type(src->data_type())); build_opts.add_option("-DDST_TENSOR_TYPE=BUFFER"); build_opts.add_option("-DDST_DATA_TYPE=" + get_cl_type_from_data_type(dst->data_type())); - build_opts.add_option("-DCONSTANT_VALUE=" + string_from_pixel_value(info.constant_border_value, src->data_type())); + build_opts.add_option("-DCONSTANT_VALUE=" + + string_from_pixel_value(info.constant_border_value, src->data_type())); build_opts.add_option("-DN0=" + support::cpp11::to_string(vec_size)); build_opts.add_option("-DPARTIAL_N0=" + support::cpp11::to_string(vec_size_leftover)); build_opts.add_option("-DSCALE_" + string_from_interpolation_policy(interpolation_policy_to_use)); @@ -144,27 +155,33 @@ void ClScaleKernel::configure(const CLCompileContext &compile_context, ITensorIn build_opts.add_option_if(info.border_mode == BorderMode::CONSTANT, "-DBORDER_MODE_CONSTANT"); build_opts.add_option_if(info.align_corners, "-DALIGN_CORNERS"); build_opts.add_option_if(is_data_type_float(src->data_type()), "-DIS_FLOATING_POINT"); - build_opts.add_option_if_else(info.sampling_policy == SamplingPolicy::CENTER, "-DSAMPLING_POLICY_CENTER", "-DSAMPLING_POLICY_TOP_LEFT"); + build_opts.add_option_if_else(info.sampling_policy == SamplingPolicy::CENTER, "-DSAMPLING_POLICY_CENTER", + "-DSAMPLING_POLICY_TOP_LEFT"); } - else if(_data_layout == DataLayout::NCHW) + else if (_data_layout == DataLayout::NCHW) { vec_size = adjust_vec_size(4, dst_width); vec_size_leftover = dst_width % vec_size; build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(src->data_type())); - build_opts.add_option("-DCONSTANT_VALUE=" + string_from_pixel_value(info.constant_border_value, src->data_type())); + build_opts.add_option("-DCONSTANT_VALUE=" + + string_from_pixel_value(info.constant_border_value, src->data_type())); build_opts.add_option("-DSRC_WIDTH=" + support::cpp11::to_string(src_width)); build_opts.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(src_height)); build_opts.add_option("-DSCALE_X=" + float_to_string_with_full_precision(scale_x)); build_opts.add_option("-DSCALE_Y=" + float_to_string_with_full_precision(scale_y)); build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(vec_size)); - build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + ((vec_size_leftover == 0) ? support::cpp11::to_string(vec_size) : support::cpp11::to_string(vec_size_leftover))); + build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + ((vec_size_leftover == 0) + ? support::cpp11::to_string(vec_size) + : support::cpp11::to_string(vec_size_leftover))); build_opts.add_option_if(info.border_mode == BorderMode::REPLICATE, "-DBORDER_MODE_REPLICATE"); build_opts.add_option_if(info.border_mode == BorderMode::CONSTANT, "-DBORDER_MODE_CONSTANT"); build_opts.add_option_if(info.align_corners, "-DALIGN_CORNERS"); - build_opts.add_option_if_else(info.sampling_policy == SamplingPolicy::CENTER, "-DSAMPLING_POLICY_CENTER", "-DSAMPLING_POLICY_TOP_LEFT"); + build_opts.add_option_if_else(info.sampling_policy == SamplingPolicy::CENTER, "-DSAMPLING_POLICY_CENTER", + "-DSAMPLING_POLICY_TOP_LEFT"); - const bool is_qasymm_bilinear = is_data_type_quantized_asymmetric(src->data_type()) && info.interpolation_policy == InterpolationPolicy::BILINEAR; - if(is_qasymm_bilinear) + const bool is_qasymm_bilinear = is_data_type_quantized_asymmetric(src->data_type()) && + info.interpolation_policy == InterpolationPolicy::BILINEAR; + if (is_qasymm_bilinear) { const UniformQuantizationInfo qinfo = src->quantization_info().uniform(); build_opts.add_option("-DSCALE=" + support::cpp11::to_string(qinfo.scale)); @@ -190,7 +207,7 @@ void ClScaleKernel::configure(const CLCompileContext &compile_context, ITensorIn ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info)); // Pass scale kernel arguments - if(is_nhwc) + if (is_nhwc) { unsigned int idx = 2 * num_arguments_per_4d_tensor_nhwc(); _kernel.setArg<cl_float>(idx++, scale_x); @@ -219,7 +236,7 @@ void ClScaleKernel::run_op(ITensorPack &tensors, const Window &window, cl::Comma auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC)); auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST)); - switch(_data_layout) + switch (_data_layout) { case DataLayout::NCHW: { @@ -231,8 +248,7 @@ void ClScaleKernel::run_op(ITensorPack &tensors, const Window &window, cl::Comma add_2D_tensor_argument(idx, src, slice); add_2D_tensor_argument(idx, dst, slice); enqueue(queue, *this, slice, lws_hint()); - } - while(window.slide_window_slice_2D(slice)); + } while (window.slide_window_slice_2D(slice)); break; } case DataLayout::NHWC: |