forked from clj-python/libpython-clj
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathnp_array.clj
199 lines (174 loc) · 6.67 KB
/
np_array.clj
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
(ns libpython-clj2.python.np-array
"Bindings for deeper intergration of numpy into the tech.v3.datatype system. This
allows seamless usage of numpy arrays in datatype and tensor functionality such as
enabling the tech.v3.tensor/ensure-tensor call to work with numpy arrays -- using
zero copying when possible.
All users need to do is call require this namespace; then as-jvm will convert a numpy
array into a tech tensor in-place."
(:require [libpython-clj2.python.ffi :as py-ffi]
[libpython-clj2.python.fn :as py-fn]
[libpython-clj2.python.protocols :as py-proto]
[libpython-clj2.python.copy :as py-copy]
[libpython-clj2.python.bridge-as-jvm :as py-bridge]
[libpython-clj2.python.base :as py-base]
[libpython-clj2.python.gc :as pygc]
[tech.v3.tensor :as dtt]
[tech.v3.datatype.protocols :as dtype-proto]
[tech.v3.datatype.casting :as casting]
[tech.v3.datatype.argops :as argops]
[tech.v3.datatype :as dtype]
[clojure.set :as set])
(:import [tech.v3.datatype NDBuffer]))
(def py-dtype->dtype-map
(->> (concat (for [bit-width [8 16 32 64]
unsigned? [true false]]
(str (if unsigned?
"uint"
"int")
bit-width))
["float32" "float64"])
(map (juxt identity keyword))
(into {})))
(def dtype->py-dtype-map
(set/map-invert py-dtype->dtype-map))
(defn obj-dtype->dtype
[py-dtype]
(when-let [fields (py-proto/get-attr py-dtype "fields")]
(throw (ex-info (format "Cannot convert numpy object with fields: %s"
(py-fn/call-attr fields "__str__" nil))
{})))
(if-let [retval (->> (py-proto/get-attr py-dtype "name")
(get py-dtype->dtype-map))]
retval
(throw (ex-info (format "Unable to find datatype: %s"
(py-proto/get-attr py-dtype "name"))
{}))))
(defn numpy->desc
[np-obj]
(py-ffi/with-gil
(let [ctypes (py-proto/as-jvm (py-proto/get-attr np-obj "ctypes") {})
np-dtype (-> (py-proto/as-jvm (py-proto/get-attr np-obj "dtype") {})
(obj-dtype->dtype))
shape (-> (delay (py-proto/get-attr ctypes "shape"))
(py-bridge/generic-python-as-list)
vec)
strides (-> (delay (py-proto/get-attr ctypes "strides"))
(py-bridge/generic-python-as-list)
vec)
long-addr (py-proto/get-attr ctypes "data")]
{:ptr long-addr
:elemwise-datatype np-dtype
:shape shape
:strides strides
:type :numpy
:ctypes ctypes})))
(defmethod py-proto/pyobject->jvm :ndarray
[pyobj opts]
(pygc/with-stack-context
(-> (numpy->desc pyobj)
(dtt/nd-buffer-descriptor->tensor)
(dtt/clone))))
(defmethod py-proto/pyobject-as-jvm :ndarray
[pyobj opts]
(let [pyobj* (delay pyobj)]
(py-bridge/bridge-pyobject
pyobj*
Iterable
(iterator [this] (py-bridge/python->jvm-iterator @pyobj* py-base/as-jvm))
dtype-proto/PToTensor
(as-tensor [item]
(-> (numpy->desc item)
(dtt/nd-buffer-descriptor->tensor)))
dtype-proto/PElemwiseDatatype
(elemwise-datatype
[this]
(py-ffi/with-gil
(-> (py-proto/get-attr pyobj "dtype")
(py-proto/as-jvm {})
(obj-dtype->dtype))))
dtype-proto/PECount
(ecount [this] (apply * (dtype-proto/shape this)))
dtype-proto/PShape
(shape
[this]
(py-ffi/with-gil
(-> (py-proto/get-attr @pyobj* "shape")
(py-proto/->jvm {}))))
dtype-proto/PToNativeBuffer
(convertible-to-native-buffer? [item] true)
(->native-buffer
[item]
(py-ffi/with-gil
(dtype-proto/->native-buffer
(dtype-proto/as-tensor item))))
dtype-proto/PSubBuffer
(sub-buffer
[buffer offset length]
(py-ffi/with-gil
(-> (dtype-proto/as-tensor buffer)
(dtype-proto/sub-buffer offset length))))
dtype-proto/PToNDBufferDesc
(convertible-to-nd-buffer-desc? [item] true)
(->nd-buffer-descriptor
[item]
(py-ffi/with-gil
(numpy->desc item))))))
(defn datatype->ptr-type-name
[dtype]
(case dtype
:int8 "c_byte"
:uint8 "c_ubyte"
:int16 "c_short"
:uint16 "c_ushort"
:int32 "c_int"
:uint32 "c_uint"
:int64 "c_longlong"
:uint64 "c_ulonglong"
:float32 "c_float"
:float64 "c_double"))
(defn descriptor->numpy
[{:keys [ptr shape strides elemwise-datatype] :as buffer-desc}]
(py-ffi/with-gil
(let [stride-tricks (-> (py-ffi/import-module "numpy.lib.stride_tricks")
(py-base/as-jvm))
ctypes (-> (py-ffi/import-module "ctypes")
(py-base/as-jvm))
np-ctypes (-> (py-ffi/import-module "numpy.ctypeslib")
(py-base/as-jvm))
dtype-size (casting/numeric-byte-width elemwise-datatype)
max-stride-idx (argops/argmax strides)
buffer-len (* (long (dtype/get-value shape max-stride-idx))
(long (dtype/get-value strides max-stride-idx)))
n-elems (quot buffer-len dtype-size)
lvalue (long ptr)
void-p (py-fn/call-attr ctypes "c_void_p" [lvalue])
actual-ptr (py-fn/call-attr
ctypes "cast"
[void-p
(py-fn/call-attr
ctypes "POINTER"
[(py-proto/get-attr
ctypes
(datatype->ptr-type-name elemwise-datatype))])])
initial-buffer (py-fn/call-attr
np-ctypes "as_array"
[actual-ptr (py-copy/->py-tuple [n-elems])])
retval (py-fn/call-attr stride-tricks "as_strided"
[initial-buffer
(py-copy/->py-tuple shape)
(py-copy/->py-tuple strides)])]
;;Ensure we have metadata that allows the GC to track both buffer
;;desc and retval
(vary-meta retval assoc
:nd-buffer-descriptor buffer-desc))))
;;Efficient conversion from jvm to python
(extend-type NDBuffer
py-proto/PCopyToPython
(->python [item opts]
(-> (dtt/ensure-nd-buffer-descriptor item)
(descriptor->numpy)))
py-proto/PBridgeToJVM
(as-python [item opts]
(when (dtype-proto/convertible-to-nd-buffer-desc? item)
(-> (dtype-proto/->nd-buffer-descriptor item)
(descriptor->numpy)))))