[NumPy]Set numpy default dtype (#17283)

* draft 1

Preliminary completion

fix rebase mistake

* add new change

* use global flag & fix part of CI error

* fix CI error

* modify docs

* rebase

* consistent with FFI

* ones/full ffi

* arange ffi

* identity ffi

* depart set_np_default_dtype from set_np to test CI

fix sanity error

fix sanity error

* test CI

* fix CI error : use InitNumpyType instead of InitType / remove full ffi

fix sanity error

fix rebase error

* try to pass CI

* set the second output's dtype of normal to be float32

* modify ffi op

* update & rebase

* test windows CI

* comment test when platform is windows

* resolve comment

* add to set_np

benchmark/python/einsum/benchmark_einsum.py

@@ -83,5 +83,5 @@ def test_np_einsum():
 
 
 if __name__ == "__main__":
-    npx.set_np()
+    npx.set_np(dtype=False)
     test_np_einsum()

benchmark/python/ffi/benchmark_ffi.py

@@ -51,6 +51,9 @@ def generate_workloads():
 def prepare_workloads():
     pool = generate_workloads()
     OpArgMngr.add_workload("zeros", (2, 2))
+    OpArgMngr.add_workload("full", (2, 2), 10)
+    OpArgMngr.add_workload("identity", 3)
+    OpArgMngr.add_workload("ones", (2, 2))
     OpArgMngr.add_workload("einsum", "ii", pool['2x2'], optimize=False)
     OpArgMngr.add_workload("unique", pool['1'], return_index=True, return_inverse=True, return_counts=True, axis=-1)
     OpArgMngr.add_workload("dstack", (pool['2x1'], pool['2x1'], pool['2x1'], pool['2x1']))
@@ -244,7 +247,7 @@ def show_results(results):
     import numpy as onp
     from mxnet import np as dnp
 
-    mx.npx.set_np()
+    mx.npx.set_np(dtype=False)
     packages = {
         "onp": {
             "module": onp,

include/mxnet/c_api.h

@@ -1305,6 +1305,20 @@ MXNET_DLL int MXIsNumpyShape(int* curr);
  * \return 0 when success, -1 when failure happens
  */
 MXNET_DLL int MXSetIsNumpyShape(int is_np_shape, int* prev);
+/*!
+ * \brief get numpy default data type
+ * \param curr returns the current status
+ * \return 0 when success, -1 when failure happens
+ */
+MXNET_DLL int MXIsNumpyDefaultDtype(bool* curr);
+/*!
+ * \brief set numpy default data type
+ * \param dtype_flag false when default dtype is flaot32,
+ *                   true when default dtype is flaot64.
+ * \param prev returns the previous status before this set
+ * \return 0 when success, -1 when failure happens
+ */
+MXNET_DLL int MXSetIsNumpyDefaultDtype(bool dtype_flag, bool* prev);
 /*!
  * \brief mark NDArrays as variables to compute gradient for autograd
  * \param num_var number of variable NDArrays

include/mxnet/imperative.h

@@ -46,6 +46,7 @@ namespace mxnet {
    *   turn off numpy shape flag globally.
    * */
   enum NumpyShape{Off, ThreadLocalOn, GlobalOn};
+  typedef NumpyShape NumpyDefaultDtype;
 /*! \brief runtime functions for NDArray */
 class Imperative {
  public:
@@ -189,9 +190,11 @@ class Imperative {
    * */
   int is_np_shape() const {
     if (is_np_shape_global_) {
-      return 2;
+      return NumpyShape::GlobalOn;
     }
-    return is_np_shape_thread_local_ ? 1 : 0;
+    return is_np_shape_thread_local_ ?
+           NumpyShape::ThreadLocalOn :
+           NumpyShape::Off;
   }
   /*! \brief specify numpy compatibility off, thread local on or global on. */
   bool set_is_np_shape(int is_np_shape) {
@@ -212,6 +215,24 @@ class Imperative {
     }
     return old;
   }
+  /*! \brief return current numpy default dtype compatibility status.
+   * */
+  bool is_np_default_dtype() const {
+    if (is_np_default_dtype_global_) {
+      return true;
+    }
+    return false;
+  }
+  /*! \brief specify numpy default dtype off or global on. */
+  bool set_is_np_default_dtype(bool is_np_default_dtype) {
+    bool old = this->is_np_default_dtype();
+    if (is_np_default_dtype) {
+      is_np_default_dtype_global_ = true;
+    } else {
+      is_np_default_dtype_global_ = false;
+    }
+    return old;
+  }
   /*! \brief to record operator, return corresponding node. */
   void RecordOp(nnvm::NodeAttrs&& attrs,
                 const std::vector<NDArray*>& inputs,
@@ -301,6 +322,7 @@ class Imperative {
   static MX_THREAD_LOCAL bool is_np_shape_thread_local_;
 #endif
   bool is_np_shape_global_{false};
+  bool is_np_default_dtype_global_{false};
   /*! \brief node count used for naming */
   std::atomic<uint64_t> node_count_{0};
   /*! \brief variable count used for naming */

python/mxnet/__init__.py

@@ -25,6 +25,7 @@
 from .base import MXNetError
 from .util import is_np_shape, set_np_shape, np_shape, use_np_shape
 from .util import is_np_array, np_array, use_np_array, use_np
+from .util import is_np_default_dtype, np_default_dtype, use_np_default_dtype
 from . import base
 
 # version info

python/mxnet/gluon/data/dataloader.py

@@ -407,7 +407,7 @@ def __len__(self):
 
 def _thread_worker_initializer(active_shape, active_array):
     """Initializer for ThreadPool."""
-    set_np(shape=active_shape, array=active_array)
+    set_np(shape=active_shape, array=active_array, dtype=False)
 
 
 _worker_dataset = None
@@ -418,7 +418,7 @@ def _worker_initializer(dataset, active_shape, active_array):
     # can be passed as argument
     global _worker_dataset
     _worker_dataset = dataset
-    set_np(shape=active_shape, array=active_array)
+    set_np(shape=active_shape, array=active_array, dtype=False)
 
 def _worker_fn(samples, batchify_fn, dataset=None):
     """Function for processing data in worker process."""