diff options
Diffstat (limited to 'src/core/CL/CLUtils.cpp')
-rw-r--r-- | src/core/CL/CLUtils.cpp | 114 |
1 files changed, 2 insertions, 112 deletions
diff --git a/src/core/CL/CLUtils.cpp b/src/core/CL/CLUtils.cpp index 03f78697bc..7e56a3ba18 100644 --- a/src/core/CL/CLUtils.cpp +++ b/src/core/CL/CLUtils.cpp @@ -23,16 +23,14 @@ */ #include "src/core/CL/CLUtils.h" -#include "arm_compute/core/utils/ActivationFunctionUtils.h" #include "arm_compute/core/CL/CLCompileContext.h" #include "arm_compute/core/CL/CLKernelLibrary.h" #include "arm_compute/core/CL/ICLTensor.h" #include "arm_compute/core/Validate.h" +#include "arm_compute/core/utils/ActivationFunctionUtils.h" #include "arm_compute/core/utils/StringUtils.h" #include "support/StringSupport.h" -#include "src/core/experimental/PostOpUtils.h" - namespace arm_compute { cl::Image2D create_image2d_from_tensor(const ICLTensor *tensor, CLImage2DType image_type) @@ -40,7 +38,7 @@ cl::Image2D create_image2d_from_tensor(const ICLTensor *tensor, CLImage2DType im ARM_COMPUTE_ERROR_ON_NULLPTR(tensor); const cl::Context &ctx = CLKernelLibrary::get().context(); - const cl::Buffer &buffer = tensor->cl_buffer(); + const cl::Buffer &buffer = tensor->cl_buffer(); const ITensorInfo *info = tensor->info(); ARM_COMPUTE_ERROR_ON_MSG(info->lock_paddings(), "Tensor paddings must not be locked to allow extending paddings to satisfy cl_image pitch alignment requirement"); @@ -113,112 +111,4 @@ cl::Image2D create_image2d_from_buffer(const cl::Context &ctx, const cl::Buffer return cl::Image2D(cl_image); } - -namespace experimental -{ -PostOpCLKernelUtils::PostOpCLKernelUtils(const Config &supported_config) - : _supported_config(supported_config) -{ - ARM_COMPUTE_ERROR_ON_MSG(supported_config.empty(), "Empty PostOp CL kernel support configuration is not allowed"); - for(auto it = _supported_config.begin(); it != _supported_config.end(); ++it) - { - auto post_op_sequence = it->first; - auto post_op_slots = std::get<1>(it->second); - ARM_COMPUTE_ERROR_ON_MSG(post_op_sequence.size() != post_op_slots.size(), "The number of PostOps must be the same as that of the assigned slots"); - } -} - -bool PostOpCLKernelUtils::are_post_op_shapes_compliant(const ITensorInfo *dst, const experimental::PostOpList<ITensorInfo *> &post_ops) -{ - for(const auto &op : post_ops.get_list()) - { - for(const auto &tensor : op->arguments()) - { - const TensorShape &out_shape = TensorShape::broadcast_shape(dst->tensor_shape(), (*tensor)->tensor_shape()); - // All post ops must be elementwise and must not alter the shape of the original dst tensor after broadcasting - if(detail::have_different_dimensions(out_shape, dst->tensor_shape(), 0)) - { - return false; - } - // NOTE: Kernel limitation: currently only the following broadcasting types are supported: - // 1. Post op arg is scalar, broadcast in both first and second dims - // 2. Post op arg is of shape: second dim=1, first dim=N, broadcast only in second dim - // This means this case: Post op arg is of shape: second dim=M, first dim=1, broadcast only in first dim, is NOT supported - if(dst->dimension(0) > 1 && dst->dimension(1) > 1 && (*tensor)->dimension(0) == 1 && (*tensor)->dimension(1) > 1) - { - return false; - } - } - } - return true; -} - -bool PostOpCLKernelUtils::is_post_op_sequence_supported(const PostOpList<ITensorInfo *> &post_ops) const -{ - if(post_ops.size() == 0) - { - return true; // Always support cases where no post op is specified - } - const auto post_op_sequence = get_post_op_sequence(post_ops); - - return _supported_config.find(post_op_sequence) != _supported_config.end(); -} - -void PostOpCLKernelUtils::set_post_ops_cl_build_options(CLBuildOptions &build_opts, const PostOpList<ITensorInfo *> &post_ops) const -{ - const auto post_op_sequence = get_post_op_sequence(post_ops); - const auto slots = std::get<1>(_supported_config.at(post_op_sequence)); - for(size_t post_op_id = 0; post_op_id < post_ops.size(); ++post_op_id) - { - const auto &post_op = post_ops.get_list().at(post_op_id); - const auto slot_prefix = "-DP" + support::cpp11::to_string(slots[post_op_id]); - if(post_op->type() == experimental::PostOpType::Activation) - { - const auto _post_op = utils::cast::polymorphic_downcast<const experimental::PostOpAct<ITensorInfo *> *>(post_op.get()); - const auto act_type = slot_prefix + "_ACTIVATION_TYPE=" + lower_string(string_from_activation_func(_post_op->_act_info.activation())); - const auto act_a_val = slot_prefix + "_ACTIVATION_A_VAL=" + float_to_string_with_full_precision(_post_op->_act_info.a()); - const auto act_b_val = slot_prefix + "_ACTIVATION_B_VAL=" + float_to_string_with_full_precision(_post_op->_act_info.b()); - build_opts.add_option(act_type); - build_opts.add_option(act_a_val); - build_opts.add_option(act_b_val); - } - else if(post_op->type() == experimental::PostOpType::Eltwise_Add) - { - size_t arg_id = 1; - const auto eltwise_op = slot_prefix + "_ELTWISE_OP=ADD" + "_X_POS_" + support::cpp11::to_string(post_op->prev_dst_pos()); - build_opts.add_option(eltwise_op); - for(const auto &tensor : post_op->arguments()) - { - const auto height = slot_prefix + "_ELTWISE_ARG" + support::cpp11::to_string(arg_id) + "_HEIGHT=" + support::cpp11::to_string((*tensor)->dimension(1)); - const auto width = slot_prefix + "_ELTWISE_ARG" + support::cpp11::to_string(arg_id) + "_WIDTH=" + support::cpp11::to_string((*tensor)->dimension(0)); - build_opts.add_option(height); - build_opts.add_option(width); - ++arg_id; - } - } - else if(post_op->type() == experimental::PostOpType::Eltwise_PRelu) - { - size_t arg_id = 1; - const auto eltwise_op = slot_prefix + "_ELTWISE_OP=PRELU" + "_X_POS_" + support::cpp11::to_string(post_op->prev_dst_pos()); - build_opts.add_option(eltwise_op); - for(const auto &tensor : post_op->arguments()) - { - const auto height = slot_prefix + "_ELTWISE_ARG" + support::cpp11::to_string(arg_id) + "_HEIGHT=" + support::cpp11::to_string((*tensor)->dimension(1)); - const auto width = slot_prefix + "_ELTWISE_ARG" + support::cpp11::to_string(arg_id) + "_WIDTH=" + support::cpp11::to_string((*tensor)->dimension(0)); - build_opts.add_option(height); - build_opts.add_option(width); - ++arg_id; - } - } - } -} - -void PostOpCLKernelUtils::set_post_ops_cl_kernel_name(std::string &kernel_name, const PostOpList<ITensorInfo *> &post_ops) const -{ - const auto post_op_sequence = get_post_op_sequence(post_ops); - const auto postfix = std::get<0>(_supported_config.at(post_op_sequence)); - kernel_name += postfix; -} -} // namespace experimental - } // namespace arm_compute |