gotch/wrapper/util.go

158 lines
4.7 KiB
Go
Raw Normal View History

2020-05-28 17:58:23 +01:00
package wrapper
2020-05-28 08:30:17 +01:00
// #include <stdlib.h>
import "C"
2020-05-28 08:30:17 +01:00
import (
"bytes"
"encoding/binary"
"fmt"
"reflect"
"unsafe"
// gotch "github.com/sugarme/gotch"
2020-05-28 08:30:17 +01:00
)
// nativeEndian is a ByteOrder for local platform.
// Ref. https://stackoverflow.com/a/53286786
// Ref. https://github.com/tensorflow/tensorflow/blob/master/tensorflow/go/tensor.go#L488-L505
2020-05-28 08:30:17 +01:00
var nativeEndian binary.ByteOrder
func init() {
buf := [2]byte{}
*(*uint16)(unsafe.Pointer(&buf[0])) = uint16(0xABCD)
switch buf {
case [2]byte{0xCD, 0xAB}:
nativeEndian = binary.LittleEndian
case [2]byte{0xAB, 0xCD}:
nativeEndian = binary.BigEndian
default:
panic("Could not determine native endianness.")
}
}
// CMalloc allocates a given number of bytes to C side memory.
// It returns
// - dataPtr: a C pointer type of `*void` (`unsafe.Pointer` in Go).
// - buf : a Go pointer points to a given bytes of buffer (empty) in C memory
// allocated by C waiting for writing data to.
//
// NOTE:
// 1. Go pointer is a pointer to Go memory. C pointer is a pointer to C memory.
// 2. General rule is Go code can use C pointers. Go code may pass Go pointer to C
// provided that the Go memory to which it points does NOT contain any Go
// pointers. BUT C code must not store any Go pointers in Go memory, even
// temporarily.
// 3. Some Go values contain Go pointers IMPLICITLY: strings, slices, maps,
// channels and function values. Thus, pointers to these values should not be
// passed to C side. Instead, data should be allocated to C memory and return a
// C pointer to it using `C.malloc`.
// Ref: https://github.com/golang/proposal/blob/master/design/12416-cgo-pointers.md
func CMalloc(nbytes int) (dataPtr unsafe.Pointer, buf *bytes.Buffer) {
dataPtr = C.malloc(C.size_t(nbytes))
// Recall: 1 << 30 = 1 * 2 * 30
// Ref. See more at https://stackoverflow.com/questions/48756732
dataSlice := (*[1 << 30]byte)(dataPtr)[:nbytes:nbytes]
buf = bytes.NewBuffer(dataSlice[:0:nbytes])
return dataPtr, buf
}
// EncodeTensor loads tensor data to C memory and returns a C pointer.
2020-05-28 17:58:23 +01:00
func EncodeTensor(w *bytes.Buffer, v reflect.Value, shape []int64) error {
2020-05-28 08:30:17 +01:00
switch v.Kind() {
case reflect.Bool:
b := byte(0)
if v.Bool() {
b = 1
}
if err := w.WriteByte(b); err != nil {
return err
}
case reflect.Uint8, reflect.Int8, reflect.Int16, reflect.Int32, reflect.Int64, reflect.Float32, reflect.Float64:
2020-05-28 08:30:17 +01:00
if err := binary.Write(w, nativeEndian, v.Interface()); err != nil {
return err
}
case reflect.Array, reflect.Slice:
// If current dimension is a slice, verify that it has the expected size
// Go's type system makes that guarantee for arrays.
if v.Kind() == reflect.Slice {
expected := int(shape[0])
if v.Len() != expected {
return fmt.Errorf("mismatched slice lengths: %d and %d", v.Len(), expected)
}
}
// Optimisation: if only one dimension is left we can use binary.Write() directly for this slice
if len(shape) == 1 && v.Len() > 0 {
switch v.Index(0).Kind() {
case reflect.Uint8, reflect.Int8, reflect.Int16, reflect.Int32, reflect.Int64, reflect.Float32, reflect.Float64:
2020-05-28 08:30:17 +01:00
return binary.Write(w, nativeEndian, v.Interface())
}
}
subShape := shape[1:]
for i := 0; i < v.Len(); i++ {
2020-05-28 17:58:23 +01:00
err := EncodeTensor(w, v.Index(i), subShape)
2020-05-28 08:30:17 +01:00
if err != nil {
return err
}
}
default:
return fmt.Errorf("unsupported type %v", v.Type())
}
return nil
}
// DecodeTensor decodes tensor value from a C memory buffer given
// C pointer, data type and shape and returns data value of type interface
2020-05-28 17:58:23 +01:00
func DecodeTensor(r *bytes.Reader, shape []int64, typ reflect.Type, ptr reflect.Value) error {
2020-05-28 08:30:17 +01:00
switch typ.Kind() {
case reflect.Bool:
b, err := r.ReadByte()
if err != nil {
return err
}
ptr.Elem().SetBool(b == 1)
case reflect.Uint8, reflect.Int8, reflect.Int16, reflect.Int32, reflect.Int64, reflect.Float32, reflect.Float64:
2020-05-28 08:30:17 +01:00
if err := binary.Read(r, nativeEndian, ptr.Interface()); err != nil {
return err
}
case reflect.Slice:
val := reflect.Indirect(ptr)
val.Set(reflect.MakeSlice(typ, int(shape[0]), int(shape[0])))
// Optimization: if only one dimension is left we can use binary.Read() directly for this slice
if len(shape) == 1 && val.Len() > 0 {
switch val.Index(0).Kind() {
case reflect.Uint8, reflect.Int8, reflect.Int16, reflect.Int32, reflect.Int64, reflect.Float32, reflect.Float64:
2020-05-28 08:30:17 +01:00
return binary.Read(r, nativeEndian, val.Interface())
}
}
for i := 0; i < val.Len(); i++ {
2020-05-28 17:58:23 +01:00
if err := DecodeTensor(r, shape[1:], typ.Elem(), val.Index(i).Addr()); err != nil {
2020-05-28 08:30:17 +01:00
return err
}
}
default:
return fmt.Errorf("unsupported type %v", typ)
}
return nil
}
// ElementCount counts number of element in the tensor given a shape
func ElementCount(shape []int64) int64 {
2020-05-28 17:58:23 +01:00
n := int64(1)
for _, d := range shape {
n *= d
}
return n
}