/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ #ifdef SOFTMAX_OP #include "operators/kernel/softmax_kernel.h" #include "operators/kernel/central-arm-func/softmax_arm_func.h" namespace paddle_mobile { namespace operators { template <> bool SoftmaxKernel::Init(SoftmaxParam *param) { auto input = const_cast(param->InputX()); auto dims = framework::vectorize(input->dims()); auto out = param->Out(); out->Resize(framework::make_ddim(dims)); out->mutable_data(framework::make_ddim(dims)); fpga::format_ofm(out); PADDLE_MOBILE_ENFORCE(input->dims().size() == 4, "Softmax should have 4-order input"); auto channel = dims[3]; if (channel == 1) { // This input is generated by FC op, dims = [N C 1 1] PADDLE_MOBILE_ENFORCE(dims[2] == 1, "Softmax input must come from FC op"); dims[3] = dims[1]; dims[1] = 1; } input->Resize(framework::make_ddim(dims)); if ((channel == 2) && (input->type() == type_id())) { auto input_ptr = input->data(); float Si = input->scale[0]; int16_t slope = fpga::fp32_2_fp16(Si / 127); fpga::format_ofm(out); fpga::BypassArgs args = {fpga::DATA_TYPE_FP16}; args.input_layout_type = fpga::LAYOUT_HWC; args.output_layout_type = fpga::LAYOUT_CHW; args.input_data_type = fpga::DATA_TYPE_FP16; args.output_data_type = fpga::DATA_TYPE_FP16; args.image.address = input_ptr; args.image.height = (uint32_t)input->dims()[1]; args.image.width = (uint32_t)input->dims()[2]; args.image.channels = (uint32_t)input->dims()[3]; args.output.address = out->data(); args.output.scale_address = out->scale; args.output.activation.activation_type = fpga::SOFTMAX; args.output.activation.leaky_relu_negative_slope = slope; param->SetFpgaArgs(args); } else if (input->type() == type_id()) { auto float_input_x = param->float_input_x_; float_input_x = std::make_shared(); float_input_x->Resize(input->dims()); float_input_x->init(type_id().hash_code()); fpga::format_ofm(float_input_x.get()); auto float_out = param->float_out; float_out = std::make_shared(); float_out->Resize(input->dims()); float_out->init(type_id().hash_code()); fpga::format_ofm(float_out.get()); } else { auto float_out = param->float_out; float_out = std::make_shared(); float_out->Resize(input->dims()); float_out->init(type_id().hash_code()); fpga::format_ofm(float_out.get()); } return true; } template <> void SoftmaxKernel::Compute(const SoftmaxParam ¶m) { auto *in_x = (param.InputX()); auto dims = in_x->dims(); auto n = dims[0]; auto h = dims[1]; auto w = dims[2]; auto c = dims[3]; if ((c == 2) && (in_x->type() == type_id())) { fpga::PerformBypass(param.FpgaArgs()); } else if (in_x->type() == type_id()) { auto in_data = in_x->data(); float Si = in_x->scale[0]; Tensor *out = param.Out(); out->Resize( {in_x->dims()[0], out->dims()[1], out->dims()[2], out->dims()[3]}); auto out_data = out->data(); auto float_input_x = param.float_input_x_; auto float_input_x_data = float_input_x->data(); int dataNum = n * h * fpga::align_to_x(w * c, IMAGE_ALIGNMENT); for (int i = 0; i < dataNum; i++) { float_input_x_data[i] = in_data[i] * Si / 127; } auto float_out = param.float_out; auto float_out_data = float_out->data(); math::SoftmaxFuntor()(float_input_x.get(), float_out.get()); for (int i = 0; i < dataNum; i++) { float tmp_out = float_out_data[i] * 127; out_data[i] = tmp_out < 0 ? (signed char)(tmp_out - 0.5) : (signed char)(tmp_out + 0.5); } fpga::fpga_flush(out_data, dataNum * sizeof(int8_t)); } else { Tensor *out = param.Out(); out->Resize( {in_x->dims()[0], out->dims()[1], out->dims()[2], out->dims()[3]}); auto out_data = out->data(); auto float_out = param.float_out; float_out = std::make_shared(); float_out->Resize(in_x->dims()); float_out->init(type_id().hash_code()); fpga::format_ofm(float_out.get()); math::SoftmaxFuntor()(in_x, float_out.get()); auto float_out_data = float_out->data(); int dataNum = n * h * fpga::align_to_x(w * c, IMAGE_ALIGNMENT); for (int i = 0; i < dataNum; i++) { float tmp_out = float_out_data[i] * 127; out_data[i] = tmp_out < 0 ? (signed char)(tmp_out - 0.5) : (signed char)(tmp_out + 0.5); } fpga::fpga_flush(out_data, dataNum * sizeof(int8_t)); } } } // namespace operators } // namespace paddle_mobile #endif