gotch/ts/tensor.go

package ts

//#include "stdlib.h"
//#include "stdbool.h"
//#include<stdio.h>
import "C"

import (
	"bytes"
	"encoding/binary"
	"fmt"
	"log"
	"reflect"
	"runtime"
	"sync"
	"sync/atomic"
	"time"
	"unsafe"

	gotch "github.com/sugarme/gotch"
	lib "github.com/sugarme/gotch/libtch"
)

var (
	TensorCount  int64 // incremental counting created tensors
	ScalarCount  int64 // incremental counting created scalars
	AllocatedMem int64 // bytes - keeping track of memory created and still occupied by gotch/tensor (excluding mem allocated by libtorch at C side)

	ExistingTensors map[string]struct{} = make(map[string]struct{}) // keep track of existing tensors by name
	ExistingScalars map[string]struct{} = make(map[string]struct{}) // keep track of existing scalar by name
	lock            sync.Mutex
)

// NOTE. None is an undefined tensor.
// It can be used in optional tensor parameter where 'None' value used.
// `ts.MustDefined()` function is used for checking 'null'
var None = NewTensor()

type bigStruct struct {
	lots [1e5]byte // 100k - always on host memory.
}

// Tensor is a Go wrapper of a "C tensor pointer" - 8 Bytes (64-bits OS)
// or 4 Bytes (32-bits OS).
// `ctensor` is just a "C pointer" to `torch::Tensor` (torch::Tensor *lib.Ctensor)
//
// NOTE.Tensor should be big enough to be in heap memory.
// (yes, we choose to place tensor consistently in heap memory so that
// it can be targeted by Go garbage collector).
//
// For heap allocation see. https://stackoverflow.com/questions/10866195
type Tensor struct {
	d          *bigStruct
	name       string
	ctensor    lib.Ctensor
	calledFrom string
}

func newTensor(ctensor lib.Ctensor, nameOpt ...string) *Tensor {
	if len(nameOpt) == 0 {
		nameOpt = []string{}
	}

	x := new(Tensor)
	x.ctensor = ctensor
	x.d = new(bigStruct)

	atomic.AddInt64(&TensorCount, 1)
	nbytes := x.nbytes()
	atomic.AddInt64(&AllocatedMem, nbytes)
	lock.Lock()
	name := newName(nameOpt...)
	if _, ok := ExistingTensors[name]; ok {
		name = fmt.Sprintf("%s_%09d", name, TensorCount)
	}
	ExistingTensors[name] = struct{}{}
	lock.Unlock()

	x.name = name

	if gotch.Debug {
		log.Printf("INFO: Added tensor %q - Allocated memory: %d bytes.\n", x.name, nbytes)
	}

	x.calledFrom = "newTensor()"

	runtime.SetFinalizer(x, freeCTensor)

	return x
}

// New creates new tensor from C tensor.
func New(ctensor lib.Ctensor, nameOpt ...string) *Tensor {
	return newTensor(ctensor, nameOpt...)
}

func CheckCMemLeak() string {
	tensors := []string{}
	lock.Lock()
	for n := range ExistingTensors {
		tensors = append(tensors, n)
	}
	memUsed := AllocatedMem
	lock.Unlock()

	var msg string
	msg += fmt.Sprintf("============================= C MEMORY CHECK RESULT ==================================\n")
	msg += fmt.Sprintf("C memory allocated not been released: %v bytes\n", memUsed)
	msg += fmt.Sprintf("Tensors not been released: %q\n", tensors)
	msg += fmt.Sprintf("======================================================================================\n")

	return msg
}

// CleanUp calls double runtime.GC() with sleep time in between.
func CleanUp(sleepTimeOpt ...int) {
	sleepTime := time.Duration(1000) // 1 second
	if len(sleepTimeOpt) > 0 {
		sleepTime = time.Duration(sleepTimeOpt[0])
	}

	runtime.GC()
	time.Sleep(time.Millisecond * sleepTime)
	runtime.GC()
}

// Ctensor return C pointer value.
func (ts *Tensor) Ctensor() unsafe.Pointer {
	return unsafe.Pointer(ts.ctensor)
}

// free releases C allocated memory.
func freeCTensor(ts *Tensor) error {
	if ts == nil || ts.ctensor == nil {
		return nil
	}

	lock.Lock()
	defer lock.Unlock()

	if _, ok := ExistingTensors[ts.name]; !ok {
		log.Printf("WARNING: Probably double free tensor %q. Called from %q. Just skipping...\n", ts.name, ts.calledFrom)

		return nil
	}

	if gotch.Debug {
		shape, err := ts.Size()
		if err != nil {
			err = fmt.Errorf("ERROR: failed to release tensor %q: %w\n", ts.name, err)
		}
		log.Printf(err.Error())

		numel := uint(FlattenDim(shape))
		dtype := ts.DType()
		nbytes := int64(numel * dtype.Size())
		atomic.AddInt64(&AllocatedMem, -nbytes)

		log.Printf("INFO: Released tensor %q - C memory(%d bytes).\n", ts.name, nbytes)
	}

	lib.AtFree(ts.ctensor)
	if err := TorchErr(); err != nil {
		err := fmt.Errorf("ERROR: failed to release tensor %q - %w", ts.name, err)
		return err
	}

	delete(ExistingTensors, ts.name)

	// IMPORTANT. make it nil so won't double free.
	ts.ctensor = nil

	return nil
}

func newName(nameOpt ...string) string {
	var name string
	if len(nameOpt) > 0 {
		name = nameOpt[0]
	} else {
		name = fmt.Sprintf("tensor_%09d", TensorCount)
	}

	return name
}

// NewTensor creates a new tensor
func NewTensor(nameOpt ...string) *Tensor {
	ctensor := lib.AtNewTensor()

	return newTensor(ctensor, nameOpt...)
}

func FromCtensor(ctensor unsafe.Pointer) *Tensor {
	cts := (lib.Ctensor)(ctensor)

	return newTensor(cts)
}

func (ts *Tensor) Name() string {
	return ts.name
}

func (ts *Tensor) Dim() uint64 {
	dim := lib.AtDim(ts.ctensor)
	if err := TorchErr(); err != nil {
		log.Fatal(err)
	}
	return dim
}

// Size return shape of the tensor
//
// NOTE: C++ libtorch calls at_shape() -> t.sizes()
// And returns a slice of sizes or shape using given pointer
// to that slice.
func (ts *Tensor) Size() ([]int64, error) {
	dim := lib.AtDim(ts.ctensor)
	if dim < 0 || dim > 100 {
		err := fmt.Errorf("Invalid dim: %v\n", dim)
		return nil, err
	}

	sz := make([]int64, dim)
	szPtr, err := DataAsPtr(sz)
	if err != nil {
		return nil, err
	}
	defer C.free(unsafe.Pointer(szPtr))

	lib.AtShape(ts.ctensor, szPtr)
	if err = TorchErr(); err != nil {
		return nil, err
	}

	shape := decodeSize(szPtr, dim)

	return shape, nil
}

func (ts *Tensor) MustSize() []int64 {
	shape, err := ts.Size()
	if err != nil {
		log.Fatal(err)
	}

	return shape
}

func (ts *Tensor) Stride() ([]int64, error) {
	dim := lib.AtDim(ts.ctensor)
	sz := make([]int64, dim)
	szPtr, err := DataAsPtr(sz)
	if err != nil {
		return nil, err
	}
	defer C.free(unsafe.Pointer(szPtr))

	lib.AtStride(ts.ctensor, szPtr)
	if err = TorchErr(); err != nil {
		return nil, err
	}

	strides := decodeSize(szPtr, dim)

	return strides, nil
}

func (ts *Tensor) MustStride() []int64 {
	strides, err := ts.Stride()
	if err != nil {
		log.Fatal(err)
	}

	return strides
}

// Size1 returns the tensor size for 1D tensors.
func (ts *Tensor) Size1() (int64, error) {
	shape, err := ts.Size()
	if err != nil {
		return 0, err
	}

	if len(shape) != 1 {
		err = fmt.Errorf("Expected one dim, got %v\n", len(shape))
		return 0, err
	}

	return shape[0], nil
}

// Size2 returns the tensor size for 2D tensors.
func (ts *Tensor) Size2() ([]int64, error) {
	shape, err := ts.Size()
	if err != nil {
		return nil, err
	}

	if len(shape) != 2 {
		err = fmt.Errorf("Expected two dims, got %v\n", len(shape))
		return nil, err
	}

	return shape, nil
}

// Size3 returns the tensor size for 3D tensors.
func (ts *Tensor) Size3() ([]int64, error) {
	shape, err := ts.Size()
	if err != nil {
		return nil, err
	}

	if len(shape) != 3 {
		err = fmt.Errorf("Expected three dims, got %v\n", len(shape))
		return nil, err
	}

	return shape, nil
}

// Size4 returns the tensor size for 4D tensors.
func (ts *Tensor) Size4() ([]int64, error) {
	shape, err := ts.Size()
	if err != nil {
		return nil, err
	}

	if len(shape) != 4 {
		err = fmt.Errorf("Expected four dims, got %v\n", len(shape))
		return nil, err
	}

	return shape, nil
}

// nbytes calculates tensor data size in bytes.
func (ts *Tensor) nbytes() int64 {
	numel := ts.Numel()
	if numel == 0 {
		return 0 // ts.None
	}

	return int64(numel * ts.DType().Size())
}

func decodeSize(ptr unsafe.Pointer, nsize uint64) []int64 {
	dtype := gotch.Int64 // tensor size dtype = int64
	nbytes := int(nsize) * int(dtype.Size())
	dataSlice := (*[1 << 30]byte)(ptr)[:nbytes:nbytes]
	r := bytes.NewReader(dataSlice)
	dataIn := make([]int64, nsize)
	if err := binary.Read(r, nativeEndian, dataIn); err != nil {
		log.Fatal(err)
	}

	return dataIn
}

// TensorOptions constructs options to build/rebuild tensor.
type TensorOptions struct {
	Name      string
	DType     gotch.DType
	Quantized bool
	// TODO. can expand as needed
}

type TensorOpt func(*TensorOptions)

func DefaultTensorOptions() *TensorOptions {
	return &TensorOptions{
		Name:      "",
		DType:     gotch.Float,
		Quantized: false,
	}
}

func WithName(v string) TensorOpt {
	return func(o *TensorOptions) {
		o.Name = v
	}
}

func WithDType(v gotch.DType) TensorOpt {
	return func(o *TensorOptions) {
		o.DType = v
	}
}

func WithQuantized(v bool) TensorOpt {
	return func(o *TensorOptions) {
		o.Quantized = v
	}
}

// OfSlice creates tensor from a slice data
func OfSlice(data interface{}, opts ...TensorOpt) (*Tensor, error) {
	o := DefaultTensorOptions()
	for _, opt := range opts {
		opt(o)
	}

	// convert []int -> int32. `binary.Write()` can't write `[]int` because it's not fixed-size!
	if reflect.TypeOf(data).String() == "[]int" {
		data = sliceIntToInt32(data.([]int))
	}

	v := reflect.ValueOf(data)
	kind := v.Kind().String()
	if kind != "slice" && kind != "array" {
		err := fmt.Errorf("Expected slice data. Got %q", kind)
		return nil, err
	}

	elementKind := reflect.TypeOf(data).Elem().Kind()
	dataLen := v.Len()

	dtype, err := gotch.GoKind2DType(elementKind, gotch.HalfDTypePref(o.DType), gotch.WithQuantized(o.Quantized))
	if err != nil {
		return nil, err
	}

	shape := []int64{int64(dataLen)}
	elementNum := ElementCount(shape)

	nbytes := int(dtype.Size()) * elementNum

	dataPtr, buff := CMalloc(nbytes)
	defer C.free(unsafe.Pointer(dataPtr))

	if err = EncodeTensor(buff, reflect.ValueOf(data), shape); err != nil {
		return nil, err
	}

	ctensor := lib.AtTensorOfData(dataPtr, shape, uint(len(shape)), uint(dtype.Size()), int(dtype.CKind()))
	if err = TorchErr(); err != nil {
		return nil, err
	}

	return newTensor(ctensor, o.Name), nil
	// return newTensor(ctensor), nil
}

// OfDataSize creates Tensor from input byte data, shape and dtype.
func OfDataSize(data []byte, shape []int64, dtype gotch.DType, opts ...TensorOpt) (*Tensor, error) {
	o := DefaultTensorOptions()
	for _, opt := range opts {
		opt(o)
	}

	elementNum := ElementCount(shape)

	nbytes := elementNum * int(dtype.Size())

	if nbytes != len(data) {
		err := fmt.Errorf("data and shape mismatched for dtype (%v): byte data (%v) - shape (%v).\n", dtype, len(data), shape)
		return nil, err
	}

	dataPtr, buff := CMalloc(nbytes)
	defer C.free(unsafe.Pointer(dataPtr))

	if err := binary.Write(buff, nativeEndian, data); err != nil {
		return nil, err
	}

	ctensor := lib.AtTensorOfData(dataPtr, shape, uint(len(shape)), dtype.Size(), int(dtype.CKind()))
	if err := TorchErr(); err != nil {
		return nil, err
	}

	return newTensor(ctensor, o.Name), nil
	// return newTensor(ctensor), nil
}

// MustOfDataSize create Tensor from input byte data and specified shape and dtype
// or panic if error
func MustOfDataSize(data []byte, size []int64, dtype gotch.DType, opts ...TensorOpt) *Tensor {
	ts, err := OfDataSize(data, size, dtype, opts...)
	if err != nil {
		log.Fatal(err)
	}

	return ts
}

// MustOfSlice create a tensor from slice of data. It will be panic if error.
func MustOfSlice(data interface{}, opts ...TensorOpt) *Tensor {
	ts, err := OfSlice(data, opts...)
	if err != nil {
		log.Fatal(err)
	}

	return ts
}

// TensorFrom create a tensor from slice of data. It will be panic if error.
func TensorFrom(data interface{}, opts ...TensorOpt) *Tensor {
	ts, err := OfSlice(data, opts...)
	if err != nil {
		log.Fatal(err)
	}
	return ts
}

// Print prints tensor values to console.
//
// NOTE: it is printed from C and will print ALL elements of tensor
// with no truncation at all.
func (ts *Tensor) Print() {
	lib.AtPrint(ts.ctensor)
	if err := TorchErr(); err != nil {
		log.Fatal(err)
	}
}

// NewTensorFromData creates tensor from given data and shape
func NewTensorFromData(data interface{}, shape []int64, opts ...TensorOpt) (*Tensor, error) {
	o := DefaultTensorOptions()
	for _, opt := range opts {
		opt(o)
	}

	// 1. Check whether data and shape match
	elementNum, err := DataDim(data)
	if err != nil {
		return nil, err
	}

	nflattend := FlattenDim(shape)

	if elementNum != nflattend {
		err = fmt.Errorf("Number of data elements (%v) and flatten shape (%v) dimension mismatched.\n", elementNum, nflattend)
		return nil, err
	}

	// 2. Write raw data to C memory and get C pointer
	dataPtr, err := DataAsPtr(data)
	defer C.free(unsafe.Pointer(dataPtr))
	if err != nil {
		return nil, err
	}

	// 3. Create tensor with pointer and shape
	dtype, err := gotch.DTypeFromData(data)
	if err != nil {
		return nil, err
	}

	ctensor := lib.AtTensorOfData(dataPtr, shape, uint(len(shape)), dtype.Size(), int(dtype.CKind()))
	if err = TorchErr(); err != nil {
		return nil, err
	}

	return newTensor(ctensor, o.Name), nil
}

func (ts *Tensor) DType() gotch.DType {
	cint := lib.AtScalarType(ts.ctensor)

	return gotch.CKind2DType(cint)
}

func (ts *Tensor) Device() (gotch.Device, error) {
	var (
		retVal gotch.Device
		err    error
	)
	cInt := lib.AtDevice(ts.ctensor)

	if err = TorchErr(); err != nil {
		return retVal, err
	}

	var device gotch.Device

	return device.OfCInt(int32(cInt)), nil
}

func (ts *Tensor) MustDevice() gotch.Device {
	device, err := ts.Device()
	if err != nil {
		log.Fatal(err)
	}

	return device
}

/*
 * func (ts Tensor) Eq1(other Tensor, del bool) (retVal Tensor, err error) {
 *
 *   // Get a C null pointer
 *   // https://stackoverflow.com/a/2022369
 *   ptr := (*lib.Ctensor)(unsafe.Pointer(C.malloc(0)))
 *   if del {
 *     defer ts.MustDrop()
 *   }
 *
 *   lib.AtgEq1(ptr, ts.ctensor, other.ctensor)
 *   if err = TorchErr(); err != nil {
 *     return retVal, err
 *   }
 *
 *   return Tensor{ctensor: *ptr}, nil
 *
 * }
 *
 * func (ts Tensor) MustEq1(other Tensor, del bool) (retVal Tensor) {
 *   retVal, err := ts.Eq1(other, del)
 *   if err != nil {
 *     log.Fatal(err)
 *   }
 *
 *   return retVal
 * }
 *  */

// Float64Value returns a float value on tensors holding a single element.
// An error is returned otherwise.
// double at_double_value_at_indexes(tensor, int64_t *indexes, int indexes_len);
func (ts *Tensor) Float64Value(idx []int64) (float64, error) {

	idxPtr, err := DataAsPtr(idx)
	if err != nil {
		return 0, err
	}
	defer C.free(unsafe.Pointer(idxPtr))

	f64Val := lib.AtDoubleValueAtIndexes(ts.ctensor, idxPtr, len(idx))
	if err = TorchErr(); err != nil {
		return 0, err
	}

	return f64Val, err
}

func (ts *Tensor) MustFloat64Value(idx []int64) float64 {
	f64Val, err := ts.Float64Value(idx)
	if err != nil {
		log.Fatal(err)
	}
	return f64Val
}

// Int64Value returns an int value on tensors holding a single element. An error is
// returned otherwise.
func (ts *Tensor) Int64Value(idx []int64) (int64, error) {

	var (
		retVal int64
		err    error
	)
	idxPtr, err := DataAsPtr(idx)
	if err != nil {
		return retVal, err
	}
	defer C.free(unsafe.Pointer(idxPtr))

	int64Val := lib.AtInt64ValueAtIndexes(ts.ctensor, idxPtr, len(idx))
	if err = TorchErr(); err != nil {
		return 0, err
	}

	return int64Val, err
}

func (ts *Tensor) MustInt64Value(idx []int64) int64 {
	int64Val, err := ts.Int64Value(idx)
	if err != nil {
		log.Fatal(err)
	}
	return int64Val
}

// RequiresGrad returns true if gradient are currently tracked for this tensor.
func (ts *Tensor) RequiresGrad() (bool, error) {
	state := lib.AtRequiresGrad(ts.ctensor)

	if err := TorchErr(); err != nil {
		return false, err
	}

	return state, nil
}

func (ts *Tensor) MustRequiresGrad() bool {
	state, err := ts.RequiresGrad()
	if err != nil {
		log.Fatal(err)
	}

	return state
}

// DataPtr returns the address of the first element of this tensor.
func (ts *Tensor) DataPtr() (unsafe.Pointer, error) {

	datPtr := lib.AtDataPtr(ts.ctensor)

	if err := TorchErr(); err != nil {
		return nil, err
	}

	return datPtr, nil
}

func (ts *Tensor) MustDataPtr() unsafe.Pointer {
	p, err := ts.DataPtr()
	if err != nil {
		panic(err)
	}

	return p
}

// Defined returns true is the tensor is defined.
func (ts *Tensor) Defined() (bool, error) {
	state := lib.AtDefined(ts.ctensor)

	if err := TorchErr(); err != nil {
		return false, err
	}

	return state, nil
}

func (ts *Tensor) MustDefined() bool {
	state, err := ts.Defined()
	if err != nil {
		log.Fatal(err)
	}

	return state
}

// IsSparse returns true is the tensor is spare.
func (ts *Tensor) IsSparse() (bool, error) {
	state := lib.AtIsSparse(ts.ctensor)

	if err := TorchErr(); err != nil {
		return false, err
	}

	return state, nil
}
func (ts *Tensor) MustIsSparse() bool {
	state, err := ts.IsSparse()
	if err != nil {
		log.Fatal(err)
	}

	return state
}

// IsContiguous returns true is the tensor is contiguous.
func (ts *Tensor) IsContiguous() (bool, error) {
	state := lib.AtIsContiguous(ts.ctensor)

	if err := TorchErr(); err != nil {
		return false, err
	}

	return state, nil
}
func (ts *Tensor) MustIsContiguous() bool {
	state, err := ts.IsContiguous()
	if err != nil {
		log.Fatal(err)
	}

	return state
}

// IsMkldnn returns true is the tensor is mkldnn.
func (ts *Tensor) IsMkldnn() (bool, error) {
	state := lib.AtIsMkldnn(ts.ctensor)

	if err := TorchErr(); err != nil {
		return false, err
	}

	return state, nil
}
func (ts *Tensor) MustIsMkldnn() bool {
	state, err := ts.IsMkldnn()
	if err != nil {
		log.Fatal(err)
	}

	return state
}

// ZeroGrad zeroes the gradient tensor attached to this tensor if defined.
func (ts *Tensor) ZeroGrad() {
	grad := ts.MustGrad(false)
	if grad.MustDefined() {
		grad.Detach_()
		grad.Zero_()
	}
}

// Backward runs the backward pass, populating the gradient tensors for tensors
// which gradients are tracked.
//
// Gradients tracking can be turned on via `SetRequiresGrad`.
func (ts *Tensor) Backward() error {
	lib.AtBackward(ts.ctensor, 0, 0)
	if err := TorchErr(); err != nil {
		return err
	}

	return nil
}

func (ts *Tensor) MustBackward() {
	if err := ts.Backward(); err != nil {
		log.Fatal(err)
	}
}

// RunBackward runs the backward ...
func RunBackward(tensors []*Tensor, inputs []*Tensor, keepGraphB bool, createGraphB bool) ([]*Tensor, error) {
	// NOTE: outputs is a slice of tensors with length = len(inputs)
	var outputsPtr []*lib.Ctensor
	// Are they allocated contigously??? Definitely not.
	// TODO. calculate C memory size = C pointer size x n pointers
	// Then C.malloc such calculated amount
	// NOTE. replace with the following code and test.
	/*
	 *   ntensors := len(inputs)
	 *   nbytes := C.size_t(ntensors) * C.size_t(unsafe.Sizeof(uintptr(0)))
	 *   ctensorsPtr := (*[1 << 30]lib.Ctensor)(C.malloc(nbytes))
	 *   for i :=0; i < ntensors; i++ {
	 *     outputPtr := (*lib.Ctensor)(unsafe.Pointer(C.malloc(0)))
	 *     outputsPtr[i] = outputPtr
	 *   }
	 *  */
	for i := 0; i < len(inputs); i++ {
		outputPtr := (*lib.Ctensor)(unsafe.Pointer(C.malloc(0)))
		defer C.free(unsafe.Pointer(outputPtr))
		outputsPtr = append(outputsPtr, outputPtr)
	}

	// Get first element pointer
	ctensor := tensors[0].ctensor
	cinput := inputs[0].ctensor
	tensorsPtr := (*lib.Ctensor)(unsafe.Pointer(&ctensor))
	inputsPtr := (*lib.Ctensor)(unsafe.Pointer(&cinput))
	var keepGraph int = 0
	if keepGraphB {
		keepGraph = 1
	}
	var createGraph int = 0
	if createGraphB {
		createGraph = 1
	}

	lib.AtRunBackward(tensorsPtr, len(tensors), inputsPtr, len(inputs), outputsPtr[0], keepGraph, createGraph)
	if err := TorchErr(); err != nil {
		return nil, err
	}

	var oTensors []*Tensor
	for i := 0; i < len(inputs); i++ {
		outputPtr := outputsPtr[i]
		oTensors = append(oTensors, newTensor(*outputPtr))
	}

	return oTensors, nil
}

// CopyDataUint8 copies `numel` elements from `self` to `dst`.
//
// NOTE: `dst` located in Go memory. Should it be?
func (ts *Tensor) CopyDataUint8(dst []uint8, numel uint) error {
	// NOTE: we must make sure that `dst` has same len as `numel`. Otherwise,
	// there will be memory leak and or out of range error.
	if len(dst) < int(numel) {
		err := fmt.Errorf("CopyDataUint8 Error: length of destination slice data (%v) is smaller than \nnumber of elements to be copied (%v)", len(dst), numel)
		return err
	}

	vs := unsafe.Pointer(&dst[0])
	dtype := gotch.Uint8
	lib.AtCopyData(ts.ctensor, vs, numel, dtype.Size())
	if err := TorchErr(); err != nil {
		return err
	}

	return nil
}

func (ts *Tensor) MustCopyDataUint8(dst []uint8, numel uint) {
	err := ts.CopyDataUint8(dst, numel)
	if err != nil {
		log.Fatal(err)
	}
}

// CopyData copies `numel` elements from `self` to `dst`.
// `dst` should be a slice of Go type equivalent to tensor type.
//
// NOTE: `dst` located in Go memory. Should it be?
// We will render Go pointer of first element of `dst` slice
// and number of elements to C land. This may break in the future
// if Go policy changes.
func (ts *Tensor) CopyData(dst interface{}, numel uint) error {
	dtype, dlen, err := DataCheck(dst)
	if dlen < int(numel) {
		err = fmt.Errorf("ts.CopyData() failed: length of destination slice data (%v) is smaller than \nnumber of elements to be copied (%v)", dlen, numel)
		return err
	}
	if ts.DType() != dtype {
		err = fmt.Errorf("ts.CopyData() failed: Type mismatched: `dst` type: %v, tensor DType: %v", dtype, ts.DType())
		return err
	}

	// Get data pointer
	dataPtr := reflect.ValueOf(dst).UnsafePointer()

	lib.AtCopyData(ts.ctensor, dataPtr, numel, dtype.Size())
	if err = TorchErr(); err != nil {
		return err
	}

	return nil
}

// MustCopyData copies number of elements from tensor to a slice of data
//
// NOTE: `dst` is a slice with length = numel and Go type equavalent to tensor
// DType
func (ts *Tensor) MustCopyData(dst interface{}, numel uint) {
	err := ts.CopyData(dst, numel)
	if err != nil {
		log.Fatal(err)
	}
}

// Numel returns the total number of elements stored in a tensor.
func (ts *Tensor) Numel() uint {
	if !ts.MustDefined() {
		return 0 // ts.None case
	}

	shape := ts.MustSize()
	return uint(FlattenDim(shape))
}

// ShallowClone returns a new tensor that share storage with the input tensor.
func (ts *Tensor) ShallowClone() (*Tensor, error) {
	ctensor := lib.AtShallowClone(ts.ctensor)

	if err := TorchErr(); err != nil {
		return nil, err
	}

	name := fmt.Sprintf("%s_cloned", ts.name)
	return newTensor(ctensor, name), nil
}

// MustShallowClone returns a new tensor that share storage with the input
// tensor. It will panic if error occurred
func (ts *Tensor) MustShallowClone() *Tensor {
	newTs, err := ts.ShallowClone()
	if err != nil {
		log.Fatal(err)
	}

	return newTs
}

// Get gets the sub-tensor at the given index.
func (ts *Tensor) Get(index int) (*Tensor, error) {

	ctensor := lib.AtGet(ts.ctensor, index)
	if err := TorchErr(); err != nil {
		return nil, err
	}

	return newTensor(ctensor), nil
}

// MustGet gets the sub-tensor at the given index. It will panic if error
// occurred.
func (ts *Tensor) MustGet(index int) *Tensor {

	subTs, err := ts.Get(index)
	if err != nil {
		log.Fatal(err)
	}

	return subTs
}

// Copy_ copies in-place values from the argument tensor to the input tensor.
func Copy_(self, src *Tensor) {

	lib.AtCopy_(self.ctensor, src.ctensor)
	if err := TorchErr(); err != nil {
		log.Fatal(err)
	}
}

// Copy_ copies in-place values from the argument tensor to existing tensor
func (ts *Tensor) Copy_(src *Tensor) {

	lib.AtCopy_(ts.ctensor, src.ctensor)
	if err := TorchErr(); err != nil {
		log.Fatal(err)
	}
}

// Save saves a tensor to a file.
func (ts *Tensor) Save(path string) error {

	lib.AtSave(ts.ctensor, path)
	if err := TorchErr(); err != nil {
		return err
	}

	return nil
}

// MustSave saves a tensor to a file. It will panic if error
func (ts *Tensor) MustSave(path string) {
	if err := ts.Save(path); err != nil {
		log.Fatal(err)
	}
}

// Load loads a tensor from a file.
func Load(path string, nameOpt ...string) (*Tensor, error) {

	ctensor := lib.AtLoad(path)
	if err := TorchErr(); err != nil {
		return nil, err
	}

	return newTensor(ctensor, nameOpt...), nil
}

// MustLoad loads a tensor to a file. It will panic if error
func MustLoad(path string, nameOpt ...string) *Tensor {
	ts, err := Load(path, nameOpt...)
	if err != nil {
		log.Fatal(err)
	}

	return ts
}

// NamedTensor wraps C tensor and its name
type NamedTensor struct {
	Name   string
	Tensor *Tensor
}

// SaveMulti saves some named tensors to a file
//
// The file format is the same as the one used by the PyTorch C++ API.
// NOTE. This method is depreciated and will be replaced with `SaveMultiNew`
func SaveMulti(namedTensors []NamedTensor, path string) error {
	var ctensors []lib.Ctensor
	var names []string

	for _, ts := range namedTensors {
		ctensors = append(ctensors, ts.Tensor.ctensor)
		names = append(names, ts.Name)
	}

	lib.AtSaveMulti(ctensors, names, len(namedTensors), path)
	if err := TorchErr(); err != nil {
		return err
	}

	return nil
}

// MustSaveMulti saves some named tensors to a file. It will panic if error
//
// NOTE. This method is depreciated and will be replaced with `MustSaveMultiNew`
func MustSaveMulti(namedTensors []NamedTensor, path string) {
	err := SaveMulti(namedTensors, path)
	if err != nil {
		log.Fatal(err)
	}
}

// LoadMulti loads some named tensors from a file
//
// The file format is the same as the one used by the PyTorch C++ API.
func LoadMulti(path string) ([]NamedTensor, error) {

	var data lib.LoadData
	dataPtr := lib.PStore.Set(&data)
	lib.AtLoadCallback(path, dataPtr)
	if err := TorchErr(); err != nil {
		return nil, err
	}

	var namedTensors []NamedTensor
	for _, v := range data.NamedCtensors {
		namedTensor := NamedTensor{
			Name:   v.Name,
			Tensor: newTensor(v.Ctensor, v.Name),
		}

		namedTensors = append(namedTensors, namedTensor)
	}

	return namedTensors, nil
}

// MustLoadMulti loads some named tensors from a file. It will panic if error
func MustLoadMulti(path string) []NamedTensor {
	namedTensors, err := LoadMulti(path)
	if err != nil {
		log.Fatal(err)
	}

	return namedTensors
}

// LoadMultiWithDevice loads some named tensors from a file to a given device
//
// The file format is the same as the one used by the PyTorch C++ API.
func LoadMultiWithDevice(path string, device gotch.Device) ([]NamedTensor, error) {
	var data lib.LoadData
	dataPtr := lib.PStore.Set(&data)

	lib.AtLoadCallbackWithDevice(path, dataPtr, device.CInt())
	if err := TorchErr(); err != nil {
		return nil, err
	}

	var namedTensors []NamedTensor
	for _, v := range data.NamedCtensors {
		namedTensor := NamedTensor{
			Name:   v.Name,
			Tensor: newTensor(v.Ctensor, v.Name),
		}

		namedTensors = append(namedTensors, namedTensor)
	}

	return namedTensors, nil
}

// MustLoadMulti loads some named tensors from a file. It will panic if error
func MustLoadMultiWithDevice(path string, device gotch.Device) []NamedTensor {
	namedTensors, err := LoadMultiWithDevice(path, device)
	if err != nil {
		log.Fatal(err)
	}

	return namedTensors
}

// ToString returns a string representation for the tensor.
//
// lw : line width (size)
// NOTE: The representation will contain all the tensor element hence may be huge for
// large tensors.
func (ts *Tensor) ToString(lw int64) (string, error) {
	tensorStr := lib.AtToString(ts.ctensor, lw)
	if err := TorchErr(); err != nil {
		return "", err
	}

	return tensorStr, nil
}

// MustToString returns a string representation for the tensor. It will be panic
// if error.
// lw : line width (size)
func (ts *Tensor) MustToString(lw int64) string {
	tensorStr, err := ts.ToString(lw)
	if err != nil {
		log.Fatal(err)
	}

	return tensorStr
}

// Drop drops (frees) the tensor
func (ts *Tensor) Drop() error {
	if ts.ctensor == nil {
		return nil
	}

	// Clear SetFinalizer on ts so no double free tensor.
	// Ref. https://pkg.go.dev/runtime#SetFinalizer
	runtime.SetFinalizer(ts, nil)

	ts.calledFrom = "ts.Drop()"
	return freeCTensor(ts)
}

// MustDrop drops the tensor. It will be panic if error
func (ts *Tensor) MustDrop() {
	if err := ts.Drop(); err != nil {
		panic(err)
	}
}

// GradSetEnabled sets globally whether GradMode gradient accumulation is enable or not.
// It returns PREVIOUS state of Grad before setting.
func GradSetEnabled(b bool) (bool, error) {

	var cbool, cretVal int
	switch b {
	case true:
		cbool = 1
	case false:
		cbool = 0
	}

	var (
		err   error
		state bool
	)
	cretVal = lib.AtGradSetEnabled(cbool)
	if err = TorchErr(); err != nil {
		return false, err
	}

	switch cretVal {
	case 0:
		state = false
		break
	case 1:
		state = true
		break
		// case -1: // should be unreachable as error is captured above with TorchrErr()
		// err = fmt.Errorf("Cannot set grad enable. \n")
		// return retVal, err
		// default: // should be unreachable as error is captured above with TorchrErr()
		// err = fmt.Errorf("Cannot set grad enable. \n")
		// return retVal, err
	}

	return state, nil
}

// MustGradSetEnabled sets globally whether GradMode gradient accumuation is enable or not.
// It returns PREVIOUS state of Grad before setting. It will be panic if error
func MustGradSetEnabled(b bool) bool {
	state, err := GradSetEnabled(b)
	if err != nil {
		log.Fatal(err)
	}

	return state
}

// NoGrad runs a closure without keeping track of gradients.
func NoGrad(fn func()) {
	// Switch off Grad
	MustGradSetEnabled(false)

	fn()

	// Switch on Grad
	MustGradSetEnabled(true)
}

func NoGrad1(fn func() interface{}) interface{} {
	newTs := NewTensor()
	newTs.Drop()

	// Switch off Grad
	prev := MustGradSetEnabled(false)

	retVal := fn()

	// Switch on Grad
	_ = MustGradSetEnabled(prev)

	return retVal
}

// NoGradGuard is a RAII guard that prevents gradient tracking until deallocated.
// It actually sets a global flag that is checked by the backend whenever an op is done on a variable.
// The guard itself saved the current status and set it to false in the constructor.
// And restore the saved status in it’s destructor.
// That way it is similar to a with torch.no_grad(): block in python.
// Ref. https://discuss.pytorch.org/t/how-does-nogradguard-works-in-cpp/34960/2
//
// TODO: should we implement Go `mutex` here???
type NoGradGuard struct {
	enabled bool
}

// Init NoGradGuard and disables gradient tracking
func NewNoGradGuard() *NoGradGuard {
	return noGradGuardInit()
}

// Disables gradient tracking, this will be enabled back when the
// returned value gets deallocated.
func noGradGuardInit() *NoGradGuard {
	return &NoGradGuard{enabled: MustGradSetEnabled(false)}
}

// Drop drops the NoGradGuard state.
func (ngg *NoGradGuard) Drop() {
	ngg.enabled = true
	_ = MustGradSetEnabled(ngg.enabled)
}

func (ngg *NoGradGuard) Enable() {
	ngg.enabled = false
	_ = MustGradSetEnabled(ngg.enabled)
}

const (
	// Do not reduce
	ReductionNone int64 = 0
	// Mean of losses
	ReductionMean int64 = 1
	// Sum of losses
	ReductionSum int64 = 2
	// Escape hatch in case new options become available
	ReductionOther int64 = 3
)

// func (r Reduction) ToInt() int {
// switch r {
// case ReductionNone:
// return 0
// case ReductionMean:
// return 1
// case ReductionSum:
// return 2
// case ReductionOther:
// return 3
// }
//
// // NOTE. should it be panic here instead of returning -1?
// return -1
// }

// Float64Values returns values of tensor in a slice of float64.
func (ts *Tensor) Float64Values(delOpt ...bool) []float64 {
	del := false
	if len(delOpt) > 0 {
		del = delOpt[0]
	}
	numel := ts.Numel()
	vec := make([]float64, numel)

	float64Ts := ts.MustTotype(gotch.Double, false)

	float64Ts.MustCopyData(vec, numel)
	// float64Ts.MustDrop()

	if del {
		ts.MustDrop()
	}

	return vec
}

// Int64Values returns values of tensor in a slice of int64.
func (ts *Tensor) Int64Values(delOpt ...bool) []int64 {
	del := false
	if len(delOpt) > 0 {
		del = delOpt[0]
	}
	numel := ts.Numel()
	vec := make([]int64, numel)

	int64Ts := ts.MustTotype(gotch.Int64, false)

	int64Ts.MustCopyData(vec, numel)
	int64Ts.MustDrop()

	if del {
		ts.MustDrop()
	}

	return vec
}

// Vals returns tensor values in a slice
// NOTE: need a type insersion to get runtime type
// E.g. res := xs.Vals().([]int64)
func (ts *Tensor) Vals() interface{} {
	dtype := ts.DType()
	numel := int(ts.Numel())

	typ, err := dtype.GoType()
	if err != nil {
		log.Fatal(err)
	}

	dataSlice := reflect.MakeSlice(reflect.SliceOf(typ), numel, numel).Interface()

	ts.CopyData(dataSlice, uint(numel))

	return dataSlice
}

// FlatView flattens a tensor.
//
// This returns a flattened version of the given tensor. The first dimension
// is preserved as it is assumed to be the mini-batch dimension.
func (ts *Tensor) FlatView() *Tensor {
	batchSize := ts.MustSize()[0]
	return ts.MustView([]int64{batchSize, -1}, false)
}

func (ts *Tensor) ZeroPad2d(left, right, top, bottom int64, del bool) (*Tensor, error) {
	if ts.Dim() != 4 {
		err := fmt.Errorf("Expected a 4 dimension tensor, got %v\n", ts.MustSize())
		return nil, err
	}

	return ts.ConstantPadNd([]int64{left, right, top, bottom}, del)
}

func (ts *Tensor) MustZeroPad2d(left, right, top, bottom int64, del bool) *Tensor {
	retVal, err := ts.ZeroPad2d(left, right, top, bottom, del)
	if err != nil {
		log.Fatal(err)
	}

	return retVal
}

// Onehot converts a tensor to a one-hot encoded version.
//
// If the input has a size [N1, N2, ..., Nk], the returned tensor has a size
// [N1, ..., Nk, labels]. The returned tensor uses float values.
// Elements of the input vector are expected to be between 0 and labels-1.
//
// NOTE: There's other `ts.OneHot` and `ts.MustOneHot` generated from Atg C++ API
func (ts *Tensor) Onehot(labels int64) *Tensor {
	dims := ts.MustSize()
	dims = append(dims, labels)
	unsqueezeTs := ts.MustUnsqueeze(-1, false)
	inputTs := unsqueezeTs.MustTotype(gotch.Int64, true)

	zerosTs := MustZeros(dims, gotch.Float, gotch.CPU)
	retVal := zerosTs.MustScatterValue(-1, inputTs, FloatScalar(1.0), true)
	inputTs.MustDrop()

	return retVal
}

func (ts *Tensor) Swish() *Tensor {
	sig := ts.MustSigmoid(false)
	mulTs := ts.MustMul(sig, false)
	sig.MustDrop()
	return mulTs
}

func (ts *Tensor) AvgPool2DDefault(ksize int64, del bool) *Tensor {
	return ts.MustAvgPool2d([]int64{ksize, ksize}, []int64{ksize, ksize}, []int64{0, 0}, false, true, []int64{1}, del)
}

// SaveMultiNew saves a slice of named tensors to the given file path.
func SaveMultiNew(namedTensors []NamedTensor, path string) error {
	var (
		tensors []lib.Ctensor
		names   []string
	)

	for _, nts := range namedTensors {
		tensors = append(tensors, nts.Tensor.ctensor)
		names = append(names, nts.Name)
	}

	lib.AtSaveMultiNew(tensors, names, path)
	if err := TorchErr(); err != nil {
		return err
	}

	return nil
}

func (ts *Tensor) ConstantPadNdWithVal(pad []int64, value *Scalar, del bool) (retVal *Tensor, err error) {
	if del {
		defer ts.MustDrop()
	}
	ptr := (*lib.Ctensor)(unsafe.Pointer(C.malloc(0)))

	lib.AtoConstantPadNd(ptr, ts.ctensor, pad, len(pad), value.cscalar)
	if err = TorchErr(); err != nil {
		return retVal, err
	}

	retVal = newTensor(*ptr)

	return retVal, err
}

func (ts *Tensor) MustConstantPadNdWithVal(pad []int64, value *Scalar, del bool) (retVal *Tensor) {

	retVal, err := ts.ConstantPadNdWithVal(pad, value, del)
	if err != nil {
		log.Fatal(err)
	}

	return retVal
}

// TT. Added some torch.cuda APIs for handling CUDA qt

// CudaCurrentDevice get device index of current CUDA device.
func CudaCurrentDevice() (int, error) {
	currentDeviceIndex := lib.AtcGetDevice()
	if err := TorchErr(); err != nil {
		err = fmt.Errorf("ts.CudaCurrentDevice() failed: %w\n", err)
		return -99, err
	}

	return currentDeviceIndex, nil
}

// CudaSetDevice set new cuda device index and returns previous cuda index.
func CudaSetDevice(cudaDeviceIndex int) (int, error) {
	currentDeviceIndex, err := CudaCurrentDevice()
	if err != nil {
		err = fmt.Errorf("ts.CudaSetDevice() failed: %w\n", err)
		return -99, err
	}

	lib.AtcSetDevice(cudaDeviceIndex)
	if err := TorchErr(); err != nil {
		err = fmt.Errorf("ts.CudaSetDevice() failed: %w\n", err)
		return -99, err
	}
	return currentDeviceIndex, nil
}

// CudaSynchronize waits for all kernels in all streams on a CUDA device to complete.
func CudaSynchronize(cudaDeviceIndexOpt ...int) error {
	var cudaDeviceIndex int
	var err error
	if len(cudaDeviceIndexOpt) > 0 {
		cudaDeviceIndex = cudaDeviceIndexOpt[0]
	} else {
		cudaDeviceIndex, err = CudaCurrentDevice()
		if err != nil {
			err := fmt.Errorf("ts.CudaSynchronize() failed: %w\n", err)
			return err
		}
	}

	lib.AtcSynchronize(int64(cudaDeviceIndex))

	return TorchErr()
}