!4392 use builtin float16 for arm

Merge pull request !4392 from xychow/use-float16-in-arm-neon

mindspore/ccsrc/backend/optimizer/ascend/ir_fission/topk_split.cc

@@ -46,13 +46,13 @@ tensor::TensorPtr CreateTensor(const AnfNodePtr &node) {
   // 2 set value of tensor
   auto data_ptr = indices_tensor->data_c();
   MS_EXCEPTION_IF_NULL(data_ptr);
-  std::vector<Eigen::half> half_data;
+  std::vector<float16> half_data;
   for (size_t i = 0; i < last_dim; ++i) {
-    half_data.emplace_back(Eigen::half(static_cast<float>(i)));
+    half_data.emplace_back(float16(static_cast<float>(i)));
   }
   for (size_t i = 0; i < last_dim; ++i) {
-    auto gap = static_cast<int>(i) - static_cast<int>(Eigen::half(static_cast<float>(i)));
-    half_data.emplace_back(Eigen::half(static_cast<float>(gap)));
+    auto gap = static_cast<int>(i) - static_cast<int>(float16(static_cast<float>(i)));
+    half_data.emplace_back(float16(static_cast<float>(gap)));
   }
   auto elem_num = last_dim * kFloat16Len * 2;
   auto ret_code = memcpy_s(data_ptr, static_cast<size_t>(indices_tensor->data().nbytes()), half_data.data(), elem_num);

mindspore/ccsrc/backend/optimizer/ascend/ir_fusion/fused_batch_norm_fusion.cc

@@ -91,9 +91,9 @@ ValuePtr FusedBatchNormFusion::GetFactor(const EquivPtr &equiv) const {
   auto tensor_ptr = value->cast<tensor::TensorPtr>();
   MS_EXCEPTION_IF_NULL(tensor_ptr);
   if (tensor_ptr->data_type() == kNumberTypeFloat16) {
-    auto *half_data = static_cast<const Eigen::half *>(tensor_ptr->data_c());
+    auto *half_data = static_cast<const float16 *>(tensor_ptr->data_c());
     MS_EXCEPTION_IF_NULL(half_data);
-    float float_data = Eigen::half_impl::half_to_float(half_data[0]);
+    float float_data = half_to_float(half_data[0]);
     return MakeValue(float_data);
   } else if (tensor_ptr->data_type() == kNumberTypeFloat32) {
     auto *tensor_data = static_cast<const float *>(tensor_ptr->data_c());

mindspore/ccsrc/common/trans.cc

@@ -138,9 +138,9 @@ template <typename SrcT>
 void TransDataSrc2Fp16(const TypeIdArgs &args, void *dst, const size_t data_size) {
   CheckMemSize(args);
   auto src_data = static_cast<const SrcT *>(args.data);
-  auto half_data = static_cast<Eigen::half *>(dst);
+  auto half_data = static_cast<float16 *>(dst);
   for (size_t i = 0; i < data_size; i++) {
-    half_data[i] = Eigen::half(src_data[i]);
+    half_data[i] = float16(src_data[i]);
   }
 }
 

mindspore/ccsrc/minddata/dataset/core/data_type.h

@@ -25,8 +25,7 @@
 #include "minddata/dataset/core/pybind_support.h"
 namespace py = pybind11;
 #else
-#include "Eigen/Core"
-using float16 = Eigen::half;
+#include "base/float16.h"
 #endif
 #include "minddata/dataset/core/constants.h"
 namespace mindspore {

mindspore/ccsrc/minddata/dataset/core/pybind_support.h

@@ -21,10 +21,9 @@
 
 #include "pybind11/numpy.h"
 #include "pybind11/pybind11.h"
-#include "Eigen/Core"
+#include "base/float16.h"
 
 namespace py = pybind11;
-using float16 = Eigen::half;
 
 namespace pybind11 {
 namespace detail {

mindspore/ccsrc/minddata/dataset/engine/datasetops/batch_op.cc

@@ -27,8 +27,6 @@
 #include "minddata/dataset/engine/opt/pass.h"
 #include "minddata/dataset/kernels/data/data_utils.h"
 
-using float16 = Eigen::half;
-
 namespace mindspore {
 namespace dataset {
 BatchOp::Builder::Builder(int32_t batch_size) : builder_drop_(false), builder_pad_(false), builder_pad_map_({}) {

mindspore/ccsrc/minddata/dataset/kernels/data/data_utils.cc

@@ -345,14 +345,14 @@ Status ToFloat16(const std::shared_ptr<Tensor> &input, std::shared_ptr<Tensor> *
 
   for (; out_itr != out_end; in_itr++, out_itr++) {
     float element = *in_itr;
-    float float16_max = static_cast<float>(std::numeric_limits<Eigen::half>::max());
-    float float16_min = static_cast<float>(std::numeric_limits<Eigen::half>::lowest());
+    float float16_max = static_cast<float>(std::numeric_limits<float16>::max());
+    float float16_min = static_cast<float>(std::numeric_limits<float16>::lowest());
     if (element > float16_max || element < float16_min) {
       RETURN_STATUS_UNEXPECTED("Value " + std::to_string(element) + " is outside of valid float16 range [" +
                                std::to_string(float16_max) + ", " + std::to_string(float16_min) + "].");
     }
 
-    *out_itr = Eigen::half(*in_itr);
+    *out_itr = float16(*in_itr);
   }
 
   return Status::OK();

mindspore/ccsrc/runtime/device/ascend/ascend_device_address.cc

@@ -111,7 +111,7 @@ bool FloatToHalfAndSyncHostToDevice(void *dst, size_t dst_size, const void *src,
     MS_EXCEPTION(ArgumentError) << "FloatToHalf failed. size not match src_size[" << src_size << "], dst_size["
                                 << dst_size << "]";
   }
-  std::vector<Eigen::half> half_data(elem_num);
+  std::vector<float16> half_data(elem_num);
   FloatToHalf(half_data.data(), src, elem_num);
   SyncMemory(dst, half_data.data(), dst_size, RT_MEMCPY_HOST_TO_DEVICE);
   return true;
@@ -134,7 +134,7 @@ bool SyncDeviceToHostAndHalfToFloat(void *dst, size_t dst_size, const void *src,
     MS_EXCEPTION(ArgumentError) << "HalfToFloat failed. size not match src_size[" << src_size << "], dst_size["
                                 << dst_size << "]";
   }
-  std::vector<Eigen::half> half_data(elem_num);
+  std::vector<float16> half_data(elem_num);
   SyncMemory(half_data.data(), src, src_size, RT_MEMCPY_DEVICE_TO_HOST);
   HalfToFloat(dst, half_data.data(), elem_num);
   return true;

mindspore/ccsrc/runtime/device/convert_tensor_utils.cc

@@ -18,19 +18,19 @@
 namespace mindspore {
 namespace device {
 void HalfToFloat(void *dst, const void *src, size_t elem_num) {
-  auto half_data = static_cast<const Eigen::half *>(src);
+  auto half_data = static_cast<const float16 *>(src);
   auto float_data = static_cast<float *>(dst);
   for (size_t i = 0; i < elem_num; ++i) {
-    float tmp = Eigen::half_impl::half_to_float(half_data[i]);
+    float tmp = half_to_float(half_data[i]);
     float_data[i] = tmp;
   }
 }
 
 void FloatToHalf(void *dst, const void *src, size_t elem_num) {
   auto float_data = static_cast<const float *>(src);
-  auto half_data = static_cast<Eigen::half *>(dst);
+  auto half_data = static_cast<float16 *>(dst);
   for (size_t i = 0; i < elem_num; ++i) {
-    half_data[i] = Eigen::half(float_data[i]);
+    half_data[i] = float16(float_data[i]);
   }
 }
 

mindspore/ccsrc/utils/load_onnx/anf_model_parser.h

@@ -29,7 +29,6 @@ namespace lite {
 using int32 = int32_t;
 using int64 = int64_t;
 using uint64 = uint64_t;
-using float16 = Eigen::half;
 class MSANFModelParser {
  public:
   MSANFModelParser() : producer_name_(""), model_version_(0), ir_version_(0) {}

mindspore/core/base/float16.h

+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * 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.
+ */
+#ifndef MINDSPORE_CORE_BASE_FLOAT16_H_
+#define MINDSPORE_CORE_BASE_FLOAT16_H_
+
+#if defined(ENABLE_ARM32) || defined(ENABLE_ARM64)
+// Built for lite and ARM
+#include <arm_neon.h>
+
+using float16 = float16_t;
+inline float half_to_float(float16 h) { return static_cast<float>(h); }
+#else
+#include <functional>
+#include "Eigen/Core"
+
+using float16 = Eigen::half;
+using HalfToFloat = std::function<float(float16)>;
+const inline HalfToFloat half_to_float = Eigen::half_impl::half_to_float;
+#endif
+#endif  // MINDSPORE_CORE_BASE_FLOAT16_H_

mindspore/core/ir/tensor.h

@@ -22,12 +22,10 @@
 #include <vector>
 #include <numeric>
 
-#include "Eigen/Core"
 #include "ir/device_sync.h"
 #include "ir/meta_tensor.h"
 #include "utils/log_adapter.h"
-
-using float16 = Eigen::half;
+#include "base/float16.h"
 
 // brief mindspore namespace.
 //