#include #include #include #include #include #include #include #include #include #include namespace migraphx { inline namespace MIGRAPHX_INLINE_NS { static bool try_compute_shape(instruction_ref ins, const std::vector& inputs, const std::vector& mods) { try { shape new_shape = ins->get_operator().compute_shape(inputs, mods); // If the output shape is a standard shape, no need to try its output if(new_shape.standard()) { return true; } // if no changes for the shape, the contiguous can also be removed if(new_shape == ins->get_shape()) { return true; } auto outputs = ins->outputs(); // If the current instruction has no output, it means it is the last // instruction and generates a non-standard output shape, and the last // output shape is different from the case with the contiguous operator if(outputs.empty()) { return false; } for(auto output : outputs) { auto args = output->inputs(); std::vector input_shapes(args.size()); std::transform(args.begin(), args.end(), input_shapes.begin(), [&](auto& arg) { return (arg == ins) ? new_shape : arg->get_shape(); }); if(!try_compute_shape(output, input_shapes, mods)) { return false; } } } catch(...) { return false; } return true; } static bool try_compute_shape(instruction_ref ins, const std::vector& args, const std::vector& mods) { auto inputs = to_shapes(args); return try_compute_shape(ins, inputs, mods); } void eliminate_contiguous::apply(module& m) const { std::vector const_instruction; for(auto ins : iterator_for(m)) { // return instruction should have inputs with standard shape if(ins->name() == "@return") continue; // Make a copy so we can modify it while we iterate auto args = ins->inputs(); auto new_args = args; auto mod_args = ins->module_inputs(); for(auto arg : ins->inputs()) { if(arg->name() == op_name) { auto prev = arg->inputs().front(); replace(new_args, arg, prev); if(try_compute_shape(ins, new_args, mod_args)) { instruction::replace_argument(ins, arg, prev); } else if(prev->can_eval()) { const_instruction.push_back(arg); } } } } // Perform evaluations in parallel std::vector literals(const_instruction.size()); par_for(const_instruction.size(), 1, [&](const auto i) { auto c = op::contiguous{}; auto prev = const_instruction[i]->inputs().front(); literals[i] = c.compute(c.compute_shape({prev->get_shape()}), {prev->eval()}); }); for(size_t i = 0; i < const_instruction.size(); i++) { auto l = m.add_literal(literals[i].get_shape(), literals[i].data()); m.replace_instruction(const_instruction[i], l); } } } // namespace MIGRAPHX_INLINE_NS } // namespace migraphx