kscience/kmath
9
0
Fork 1
Code Issues Pull Requests 6 Packages Projects Releases 16 Wiki Activity

Dropping support for buffered NDStructures

Browse Source
This commit is contained in:
rgrit91 2021-01-09 17:13:38 +00:00
parent 9b1a958491
commit 0cb2c3f0da
22 changed files with 508 additions and 676 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

87
kmath-core/src/commonMain/kotlin/kscience/kmath/structures/NDStructure.kt
View File

@ -148,28 +148,23 @@ public interface Strides {
/** /**
* Array strides * Array strides
*/ */
public val strides: IntArray public val strides: List<Int>
/**
* The size of linear buffer to accommodate all elements of ND-structure corresponding to strides
*/
public val linearSize: Int
/** /**
* Get linear index from multidimensional index * Get linear index from multidimensional index
*/ */
public fun offset(index: IntArray): Int = index.mapIndexed { i, value -> public fun offset(index: IntArray): Int
if (value < 0 || value >= this.shape[i])
throw IndexOutOfBoundsException("Index $value out of shape bounds: (0,${this.shape[i]})")
value * strides[i]
}.sum()
/** /**
* Get multidimensional from linear * Get multidimensional from linear
*/ */
public fun index(offset: Int): IntArray public fun index(offset: Int): IntArray
/**
* The size of linear buffer to accommodate all elements of ND-structure corresponding to strides
*/
public val linearSize: Int
// TODO introduce a fast way to calculate index of the next element? // TODO introduce a fast way to calculate index of the next element?
/** /**
@ -188,7 +183,7 @@ public class DefaultStrides private constructor(override val shape: IntArray) :
/** /**
* Strides for memory access * Strides for memory access
*/ */
override val strides: IntArray by lazy { override val strides: List<Int> by lazy {
sequence { sequence {
var current = 1 var current = 1
yield(1) yield(1)
@ -197,9 +192,16 @@ public class DefaultStrides private constructor(override val shape: IntArray) :
current *= it current *= it
yield(current) yield(current)
} }
}.toList().toIntArray() }.toList()
} }
override fun offset(index: IntArray): Int = index.mapIndexed { i, value ->
if (value < 0 || value >= this.shape[i])
throw IndexOutOfBoundsException("Index $value out of shape bounds: (0,${this.shape[i]})")
value * strides[i]
}.sum()
override fun index(offset: Int): IntArray { override fun index(offset: Int): IntArray {
val res = IntArray(shape.size) val res = IntArray(shape.size)
var current = offset var current = offset
@ -239,19 +241,17 @@ public class DefaultStrides private constructor(override val shape: IntArray) :
* Trait for [NDStructure] over [Buffer]. * Trait for [NDStructure] over [Buffer].
* *
* @param T the type of items * @param T the type of items
* @param BufferImpl implementation of [Buffer].
*/ */
public abstract class NDBufferTrait<T, out BufferImpl : Buffer<T>, out StridesImpl: Strides> : public abstract class NDBuffer<T> : NDStructure<T> {
NDStructure<T> {
/** /**
* The underlying buffer. * The underlying buffer.
*/ */
public abstract val buffer: BufferImpl public abstract val buffer: Buffer<T>
/** /**
* The strides to access elements of [Buffer] by linear indices. * The strides to access elements of [Buffer] by linear indices.
*/ */
public abstract val strides: StridesImpl public abstract val strides: Strides
override operator fun get(index: IntArray): T = buffer[strides.offset(index)] override operator fun get(index: IntArray): T = buffer[strides.offset(index)]
@ -259,10 +259,6 @@ public abstract class NDBufferTrait<T, out BufferImpl : Buffer<T>, out StridesIm
override fun elements(): Sequence<Pair<IntArray, T>> = strides.indices().map { it to this[it] } override fun elements(): Sequence<Pair<IntArray, T>> = strides.indices().map { it to this[it] }
public fun checkStridesBufferCompatibility(): Unit = require(strides.linearSize == buffer.size) {
"Expected buffer side of ${strides.linearSize}, but found ${buffer.size}"
}
override fun hashCode(): Int { override fun hashCode(): Int {
var result = strides.hashCode() var result = strides.hashCode()
result = 31 * result + buffer.hashCode() result = 31 * result + buffer.hashCode()
@ -286,36 +282,10 @@ public abstract class NDBufferTrait<T, out BufferImpl : Buffer<T>, out StridesIm
} }
return "NDBuffer(shape=${shape.contentToString()}, buffer=$bufferRepr)" return "NDBuffer(shape=${shape.contentToString()}, buffer=$bufferRepr)"
} }
} }
/**
* Trait for [MutableNDStructure] over [MutableBuffer].
*
* @param T the type of items
* @param MutableBufferImpl implementation of [MutableBuffer].
*/
public abstract class MutableNDBufferTrait<T, out MutableBufferImpl : MutableBuffer<T>, out StridesImpl: Strides> :
NDBufferTrait<T, MutableBufferImpl, StridesImpl>(), MutableNDStructure<T> {
override fun hashCode(): Int = 0
override fun equals(other: Any?): Boolean = false
override operator fun set(index: IntArray, value: T): Unit =
buffer.set(strides.offset(index), value)
}
/**
* Default representation of [NDStructure] over [Buffer].
*
* @param T the type of items.
*/
public abstract class NDBuffer<T> : NDBufferTrait<T, Buffer<T>, Strides>()
/**
* Default representation of [MutableNDStructure] over [MutableBuffer].
*
* @param T the type of items.
*/
public abstract class MutableNDBuffer<T> : MutableNDBufferTrait<T, MutableBuffer<T>, Strides>()
/** /**
* Boxing generic [NDStructure] * Boxing generic [NDStructure]
*/ */
@ -324,7 +294,9 @@ public class BufferNDStructure<T>(
override val buffer: Buffer<T>, override val buffer: Buffer<T>,
) : NDBuffer<T>() { ) : NDBuffer<T>() {
init { init {
checkStridesBufferCompatibility() if (strides.linearSize != buffer.size) {
error("Expected buffer side of ${strides.linearSize}, but found ${buffer.size}")
}
} }
} }
@ -344,17 +316,22 @@ public inline fun <T, reified R : Any> NDStructure<T>.mapToBuffer(
} }
/** /**
* Boxing generic [MutableNDStructure]. * Mutable ND buffer based on linear [MutableBuffer].
*/ */
public class MutableBufferNDStructure<T>( public class MutableBufferNDStructure<T>(
override val strides: Strides, override val strides: Strides,
override val buffer: MutableBuffer<T>, override val buffer: MutableBuffer<T>,
) : MutableNDBuffer<T>() { ) : NDBuffer<T>(), MutableNDStructure<T> {
init { init {
checkStridesBufferCompatibility() require(strides.linearSize == buffer.size) {
"Expected buffer side of ${strides.linearSize}, but found ${buffer.size}"
} }
} }
override operator fun set(index: IntArray, value: T): Unit = buffer.set(strides.offset(index), value)
}
public inline fun <reified T : Any> NDStructure<T>.combine( public inline fun <reified T : Any> NDStructure<T>.combine(
struct: NDStructure<T>, struct: NDStructure<T>,
crossinline block: (T, T) -> T, crossinline block: (T, T) -> T,

26
kmath-torch/README.md
View File

@ -78,28 +78,20 @@ pluginManagement {
Tensors implement the buffer protocol over `MutableNDStructure`. They can only be instantiated through provided factory methods and require scoping: Tensors implement the buffer protocol over `MutableNDStructure`. They can only be instantiated through provided factory methods and require scoping:
```kotlin ```kotlin
memScoped { TorchTensorRealAlgebra {
val intTensor: TorchTensorInt = TorchTensor.copyFromIntArray(
scope = this,
array = (1..24).toList().toIntArray(),
shape = intArrayOf(3, 2, 4)
)
println(intTensor)
val floatTensor: TorchTensorFloat = TorchTensor.copyFromFloatArray( val realTensor: TorchTensorReal = copyFromArray(
scope = this, array = (1..10).map { it + 50.0 }.toList().toDoubleArray(),
array = (1..10).map { it + 50f }.toList().toFloatArray(), shape = intArrayOf(2,5)
shape = intArrayOf(10)
) )
println(floatTensor) println(realTensor)
val gpuFloatTensor: TorchTensorFloatGPU = TorchTensor.copyFromFloatArrayToGPU( val gpuRealTensor: TorchTensorReal = copyFromArray(
scope = this, array = (1..8).map { it * 2.5 }.toList().toDoubleArray(),
array = (1..8).map { it * 2f }.toList().toFloatArray(),
shape = intArrayOf(2, 2, 2), shape = intArrayOf(2, 2, 2),
device = 0 device = TorchDevice.TorchCUDA(0)
) )
println(gpuFloatTensor) println(gpuRealTensor)
} }
``` ```

1
kmath-torch/build.gradle.kts
View File

@ -145,6 +145,7 @@ kotlin {
} }
val nativeGPUTest by creating { val nativeGPUTest by creating {
dependsOn(nativeMain) dependsOn(nativeMain)
dependsOn(nativeTest)
} }

57
kmath-torch/ctorch/include/ctorch.h
View File

@ -18,49 +18,46 @@ extern "C"
void set_seed(int seed); void set_seed(int seed);
TorchTensorHandle copy_from_blob_double(double *data, int *shape, int dim); TorchTensorHandle copy_from_blob_double(double *data, int *shape, int dim, int device);
TorchTensorHandle copy_from_blob_float(float *data, int *shape, int dim); TorchTensorHandle copy_from_blob_float(float *data, int *shape, int dim, int device);
TorchTensorHandle copy_from_blob_long(long *data, int *shape, int dim); TorchTensorHandle copy_from_blob_long(long *data, int *shape, int dim, int device);
TorchTensorHandle copy_from_blob_int(int *data, int *shape, int dim); TorchTensorHandle copy_from_blob_int(int *data, int *shape, int dim, int device);
TorchTensorHandle copy_from_blob_to_gpu_double(double *data, int *shape, int dim, int device);
TorchTensorHandle copy_from_blob_to_gpu_float(float *data, int *shape, int dim, int device);
TorchTensorHandle copy_from_blob_to_gpu_long(long *data, int *shape, int dim, int device);
TorchTensorHandle copy_from_blob_to_gpu_int(int *data, int *shape, int dim, int device);
TorchTensorHandle copy_tensor(TorchTensorHandle tensor_handle); TorchTensorHandle copy_tensor(TorchTensorHandle tensor_handle);
TorchTensorHandle copy_to_device(TorchTensorHandle tensor_handle, int device);
double *get_data_double(TorchTensorHandle tensor_handle); double get_item_double(TorchTensorHandle tensor_handle);
float *get_data_float(TorchTensorHandle tensor_handle); float get_item_float(TorchTensorHandle tensor_handle);
long *get_data_long(TorchTensorHandle tensor_handle); long get_item_long(TorchTensorHandle tensor_handle);
int *get_data_int(TorchTensorHandle tensor_handle); int get_item_int(TorchTensorHandle tensor_handle);
int get_numel(TorchTensorHandle tensor_handle);
int get_dim(TorchTensorHandle tensor_handle); int get_dim(TorchTensorHandle tensor_handle);
int *get_shape(TorchTensorHandle tensor_handle); int get_numel(TorchTensorHandle tensor_handle);
int *get_strides(TorchTensorHandle tensor_handle); int get_shape_at(TorchTensorHandle tensor_handle, int d);
int get_stride_at(TorchTensorHandle tensor_handle, int d);
int get_device(TorchTensorHandle tensor_handle);
char *tensor_to_string(TorchTensorHandle tensor_handle); char *tensor_to_string(TorchTensorHandle tensor_handle);
void dispose_int_array(int *ptr);
void dispose_char(char *ptr); void dispose_char(char *ptr);
void dispose_tensor(TorchTensorHandle tensor_handle); void dispose_tensor(TorchTensorHandle tensor_handle);
// Workaround for GPU tensors
double get_at_offset_double(TorchTensorHandle tensor_handle, int offset);
float get_at_offset_float(TorchTensorHandle tensor_handle, int offset);
long get_at_offset_long(TorchTensorHandle tensor_handle, int offset);
int get_at_offset_int(TorchTensorHandle tensor_handle, int offset);
void set_at_offset_double(TorchTensorHandle tensor_handle, int offset, double value);
void set_at_offset_float(TorchTensorHandle tensor_handle, int offset, float value);
void set_at_offset_long(TorchTensorHandle tensor_handle, int offset, long value);
void set_at_offset_int(TorchTensorHandle tensor_handle, int offset, int value);
TorchTensorHandle copy_to_cpu(TorchTensorHandle tensor_handle); double get_double(TorchTensorHandle tensor_handle, int* index);
TorchTensorHandle copy_to_gpu(TorchTensorHandle tensor_handle, int device); float get_float(TorchTensorHandle tensor_handle, int* index);
long get_long(TorchTensorHandle tensor_handle, int* index);
int get_int(TorchTensorHandle tensor_handle, int* index);
void set_double(TorchTensorHandle tensor_handle, int* index, double value);
void set_float(TorchTensorHandle tensor_handle, int* index, float value);
void set_long(TorchTensorHandle tensor_handle, int* index, long value);
void set_int(TorchTensorHandle tensor_handle, int* index, int value);
TorchTensorHandle randn_float(int* shape, int shape_size);
TorchTensorHandle randn_double(int* shape, int shape_size, int device);
TorchTensorHandle rand_double(int* shape, int shape_size, int device);
TorchTensorHandle randn_float(int* shape, int shape_size, int device);
TorchTensorHandle rand_float(int* shape, int shape_size, int device);
TorchTensorHandle matmul(TorchTensorHandle lhs, TorchTensorHandle rhs); TorchTensorHandle matmul(TorchTensorHandle lhs, TorchTensorHandle rhs);
void matmul_assign(TorchTensorHandle lhs, TorchTensorHandle rhs);
#ifdef __cplusplus #ifdef __cplusplus
} }

60
kmath-torch/ctorch/include/utils.hh
View File

@ -33,6 +33,16 @@ namespace ctorch
return *static_cast<torch::Tensor *>(tensor_handle); return *static_cast<torch::Tensor *>(tensor_handle);
} }
inline int device_to_int(const torch::Tensor &tensor)
{
return (tensor.device().type() == torch::kCPU) ? 0 : 1 + tensor.device().index();
}
inline torch::Device int_to_device(int device_int)
{
return (device_int == 0) ? torch::kCPU : torch::Device(torch::kCUDA, device_int - 1);
}
inline std::vector<int64_t> to_vec_int(int *arr, int arr_size) inline std::vector<int64_t> to_vec_int(int *arr, int arr_size)
{ {
auto vec = std::vector<int64_t>(arr_size); auto vec = std::vector<int64_t>(arr_size);
@ -40,46 +50,34 @@ namespace ctorch
return vec; return vec;
} }
inline std::vector<at::indexing::TensorIndex> to_index(int *arr, int arr_size)
{
std::vector<at::indexing::TensorIndex> index;
for (int i = 0; i < arr_size; i++)
{
index.emplace_back(arr[i]);
}
return index;
}
template <typename Dtype> template <typename Dtype>
inline torch::Tensor copy_from_blob(Dtype *data, std::vector<int64_t> shape, torch::Device device) inline torch::Tensor copy_from_blob(Dtype *data, std::vector<int64_t> shape, torch::Device device)
{ {
return torch::from_blob(data, shape, dtype<Dtype>()).to(torch::TensorOptions().layout(torch::kStrided).device(device), false, true); return torch::from_blob(data, shape, dtype<Dtype>()).to(torch::TensorOptions().layout(torch::kStrided).device(device), false, true);
} }
inline int *to_dynamic_ints(const c10::IntArrayRef &arr) template <typename NumType>
inline NumType get(const TorchTensorHandle &tensor_handle, int *index)
{ {
size_t n = arr.size(); auto ten = ctorch::cast(tensor_handle);
int *res = (int *)malloc(sizeof(int) * n); return ten.index(to_index(index, ten.dim())).item<NumType>();
for (size_t i = 0; i < n; i++)
{
res[i] = arr[i];
}
return res;
}
inline std::vector<at::indexing::TensorIndex> offset_to_index(int offset, const c10::IntArrayRef &strides)
{
std::vector<at::indexing::TensorIndex> index;
for (const auto &stride : strides)
{
index.emplace_back(offset / stride);
offset %= stride;
}
return index;
} }
template <typename NumType> template <typename NumType>
inline NumType get_at_offset(const TorchTensorHandle &tensor_handle, int offset) inline void set(TorchTensorHandle &tensor_handle, int *index, NumType value)
{ {
auto ten = ctorch::cast(tensor_handle); auto ten = ctorch::cast(tensor_handle);
return ten.index(ctorch::offset_to_index(offset, ten.strides())).item<NumType>(); ten.index(to_index(index, ten.dim())) = value;
}
template <typename NumType>
inline void set_at_offset(TorchTensorHandle &tensor_handle, int offset, NumType value)
{
auto ten = ctorch::cast(tensor_handle);
ten.index(offset_to_index(offset, ten.strides())) = value;
} }
template <typename Dtype> template <typename Dtype>
@ -88,4 +86,10 @@ namespace ctorch
return torch::randn(shape, torch::TensorOptions().dtype(dtype<Dtype>()).layout(torch::kStrided).device(device)); return torch::randn(shape, torch::TensorOptions().dtype(dtype<Dtype>()).layout(torch::kStrided).device(device));
} }
template <typename Dtype>
inline torch::Tensor rand(std::vector<int64_t> shape, torch::Device device)
{
return torch::rand(shape, torch::TensorOptions().dtype(dtype<Dtype>()).layout(torch::kStrided).device(device));
}
} // namespace ctorch } // namespace ctorch

183
kmath-torch/ctorch/src/ctorch.cc
View File

@ -25,80 +25,50 @@ void set_seed(int seed)
torch::manual_seed(seed); torch::manual_seed(seed);
} }
TorchTensorHandle copy_from_blob_double(double *data, int *shape, int dim)
{
return new torch::Tensor(ctorch::copy_from_blob<double>(data, ctorch::to_vec_int(shape, dim), torch::kCPU));
}
TorchTensorHandle copy_from_blob_float(float *data, int *shape, int dim)
{
return new torch::Tensor(ctorch::copy_from_blob<float>(data, ctorch::to_vec_int(shape, dim), torch::kCPU));
}
TorchTensorHandle copy_from_blob_long(long *data, int *shape, int dim)
{
return new torch::Tensor(ctorch::copy_from_blob<long>(data, ctorch::to_vec_int(shape, dim), torch::kCPU));
}
TorchTensorHandle copy_from_blob_int(int *data, int *shape, int dim)
{
return new torch::Tensor(ctorch::copy_from_blob<int>(data, ctorch::to_vec_int(shape, dim), torch::kCPU));
}
TorchTensorHandle copy_from_blob_to_gpu_double(double *data, int *shape, int dim, int device)
{
return new torch::Tensor(ctorch::copy_from_blob<double>(data, ctorch::to_vec_int(shape, dim), torch::Device(torch::kCUDA, device)));
}
TorchTensorHandle copy_from_blob_to_gpu_float(float *data, int *shape, int dim, int device)
{
return new torch::Tensor(ctorch::copy_from_blob<float>(data, ctorch::to_vec_int(shape, dim), torch::Device(torch::kCUDA, device)));
}
TorchTensorHandle copy_from_blob_to_gpu_long(long *data, int *shape, int dim, int device)
{
return new torch::Tensor(ctorch::copy_from_blob<long>(data, ctorch::to_vec_int(shape, dim), torch::Device(torch::kCUDA, device)));
}
TorchTensorHandle copy_from_blob_to_gpu_int(int *data, int *shape, int dim, int device)
{
return new torch::Tensor(ctorch::copy_from_blob<int>(data, ctorch::to_vec_int(shape, dim), torch::Device(torch::kCUDA, device)));
}
TorchTensorHandle copy_tensor(TorchTensorHandle tensor_handle)
{
return new torch::Tensor(ctorch::cast(tensor_handle).clone());
}
double *get_data_double(TorchTensorHandle tensor_handle)
{
return ctorch::cast(tensor_handle).data_ptr<double>();
}
float *get_data_float(TorchTensorHandle tensor_handle)
{
return ctorch::cast(tensor_handle).data_ptr<float>();
}
long *get_data_long(TorchTensorHandle tensor_handle)
{
return ctorch::cast(tensor_handle).data_ptr<long>();
}
int *get_data_int(TorchTensorHandle tensor_handle)
{
return ctorch::cast(tensor_handle).data_ptr<int>();
}
int get_numel(TorchTensorHandle tensor_handle)
{
return ctorch::cast(tensor_handle).numel();
}
int get_dim(TorchTensorHandle tensor_handle) int get_dim(TorchTensorHandle tensor_handle)
{ {
return ctorch::cast(tensor_handle).dim(); return ctorch::cast(tensor_handle).dim();
} }
int get_numel(TorchTensorHandle tensor_handle)
int *get_shape(TorchTensorHandle tensor_handle)
{ {
return ctorch::to_dynamic_ints(ctorch::cast(tensor_handle).sizes()); return ctorch::cast(tensor_handle).numel();
}
int get_shape_at(TorchTensorHandle tensor_handle, int d)
{
return ctorch::cast(tensor_handle).size(d);
}
int get_stride_at(TorchTensorHandle tensor_handle, int d)
{
return ctorch::cast(tensor_handle).stride(d);
}
int get_device(TorchTensorHandle tensor_handle)
{
return ctorch::device_to_int(ctorch::cast(tensor_handle));
} }
int *get_strides(TorchTensorHandle tensor_handle) TorchTensorHandle copy_from_blob_double(double *data, int *shape, int dim, int device)
{ {
return ctorch::to_dynamic_ints(ctorch::cast(tensor_handle).strides()); return new torch::Tensor(ctorch::copy_from_blob<double>(data, ctorch::to_vec_int(shape, dim), ctorch::int_to_device(device)));
}
TorchTensorHandle copy_from_blob_float(float *data, int *shape, int dim, int device)
{
return new torch::Tensor(ctorch::copy_from_blob<float>(data, ctorch::to_vec_int(shape, dim), ctorch::int_to_device(device)));
}
TorchTensorHandle copy_from_blob_long(long *data, int *shape, int dim, int device)
{
return new torch::Tensor(ctorch::copy_from_blob<long>(data, ctorch::to_vec_int(shape, dim), ctorch::int_to_device(device)));
}
TorchTensorHandle copy_from_blob_int(int *data, int *shape, int dim, int device)
{
return new torch::Tensor(ctorch::copy_from_blob<int>(data, ctorch::to_vec_int(shape, dim), ctorch::int_to_device(device)));
}
TorchTensorHandle copy_tensor(TorchTensorHandle tensor_handle)
{
return new torch::Tensor(ctorch::cast(tensor_handle).clone());
}
TorchTensorHandle copy_to_device(TorchTensorHandle tensor_handle, int device)
{
return new torch::Tensor(ctorch::cast(tensor_handle).to(ctorch::int_to_device(device), false, true));
} }
char *tensor_to_string(TorchTensorHandle tensor_handle) char *tensor_to_string(TorchTensorHandle tensor_handle)
@ -110,68 +80,89 @@ char *tensor_to_string(TorchTensorHandle tensor_handle)
std::strcpy(crep, rep.c_str()); std::strcpy(crep, rep.c_str());
return crep; return crep;
} }
void dispose_int_array(int *ptr)
{
free(ptr);
}
void dispose_char(char *ptr) void dispose_char(char *ptr)
{ {
free(ptr); free(ptr);
} }
void dispose_tensor(TorchTensorHandle tensor_handle) void dispose_tensor(TorchTensorHandle tensor_handle)
{ {
delete static_cast<torch::Tensor *>(tensor_handle); delete static_cast<torch::Tensor *>(tensor_handle);
} }
double get_at_offset_double(TorchTensorHandle tensor_handle, int offset) double get_double(TorchTensorHandle tensor_handle, int *index)
{ {
return ctorch::get_at_offset<double>(tensor_handle, offset); return ctorch::get<double>(tensor_handle, index);
} }
float get_at_offset_float(TorchTensorHandle tensor_handle, int offset) float get_float(TorchTensorHandle tensor_handle, int *index)
{ {
return ctorch::get_at_offset<float>(tensor_handle, offset); return ctorch::get<float>(tensor_handle, index);
} }
long get_at_offset_long(TorchTensorHandle tensor_handle, int offset) long get_long(TorchTensorHandle tensor_handle, int *index)
{ {
return ctorch::get_at_offset<long>(tensor_handle, offset); return ctorch::get<long>(tensor_handle, index);
} }
int get_at_offset_int(TorchTensorHandle tensor_handle, int offset) int get_int(TorchTensorHandle tensor_handle, int *index)
{ {
return ctorch::get_at_offset<int>(tensor_handle, offset); return ctorch::get<int>(tensor_handle, index);
} }
void set_at_offset_double(TorchTensorHandle tensor_handle, int offset, double value) void set_double(TorchTensorHandle tensor_handle, int *index, double value)
{ {
ctorch::set_at_offset<double>(tensor_handle, offset, value); ctorch::set<double>(tensor_handle, index, value);
} }
void set_at_offset_float(TorchTensorHandle tensor_handle, int offset, float value) void set_float(TorchTensorHandle tensor_handle, int *index, float value)
{ {
ctorch::set_at_offset<float>(tensor_handle, offset, value); ctorch::set<float>(tensor_handle, index, value);
} }
void set_at_offset_long(TorchTensorHandle tensor_handle, int offset, long value) void set_long(TorchTensorHandle tensor_handle, int *index, long value)
{ {
ctorch::set_at_offset<long>(tensor_handle, offset, value); ctorch::set<long>(tensor_handle, index, value);
} }
void set_at_offset_int(TorchTensorHandle tensor_handle, int offset, int value) void set_int(TorchTensorHandle tensor_handle, int *index, int value)
{ {
ctorch::set_at_offset<int>(tensor_handle, offset, value); ctorch::set<int>(tensor_handle, index, value);
} }
TorchTensorHandle copy_to_cpu(TorchTensorHandle tensor_handle) double get_item_double(TorchTensorHandle tensor_handle)
{ {
return new torch::Tensor(ctorch::cast(tensor_handle).to(torch::kCPU,false, true)); return ctorch::cast(tensor_handle).item<double>();
} }
TorchTensorHandle copy_to_gpu(TorchTensorHandle tensor_handle, int device) float get_item_float(TorchTensorHandle tensor_handle)
{ {
return new torch::Tensor(ctorch::cast(tensor_handle).to(torch::Device(torch::kCUDA, device),false, true)); return ctorch::cast(tensor_handle).item<float>();
}
long get_item_long(TorchTensorHandle tensor_handle)
{
return ctorch::cast(tensor_handle).item<long>();
}
int get_item_int(TorchTensorHandle tensor_handle)
{
return ctorch::cast(tensor_handle).item<int>();
} }
TorchTensorHandle randn_float(int* shape, int shape_size){ TorchTensorHandle randn_double(int *shape, int shape_size, int device)
return new torch::Tensor(ctorch::randn<float>(ctorch::to_vec_int(shape, shape_size), torch::kCPU)); {
return new torch::Tensor(ctorch::randn<double>(ctorch::to_vec_int(shape, shape_size), ctorch::int_to_device(device)));
}
TorchTensorHandle rand_double(int *shape, int shape_size, int device)
{
return new torch::Tensor(ctorch::rand<double>(ctorch::to_vec_int(shape, shape_size), ctorch::int_to_device(device)));
}
TorchTensorHandle randn_float(int *shape, int shape_size, int device)
{
return new torch::Tensor(ctorch::randn<float>(ctorch::to_vec_int(shape, shape_size), ctorch::int_to_device(device)));
}
TorchTensorHandle rand_float(int *shape, int shape_size, int device)
{
return new torch::Tensor(ctorch::rand<float>(ctorch::to_vec_int(shape, shape_size), ctorch::int_to_device(device)));
} }
TorchTensorHandle matmul(TorchTensorHandle lhs, TorchTensorHandle rhs){ TorchTensorHandle matmul(TorchTensorHandle lhs, TorchTensorHandle rhs)
{
return new torch::Tensor(torch::matmul(ctorch::cast(lhs), ctorch::cast(rhs))); return new torch::Tensor(torch::matmul(ctorch::cast(lhs), ctorch::cast(rhs)));
} }
void matmul_assign(TorchTensorHandle lhs, TorchTensorHandle rhs)
{
auto lhs_tensor = ctorch::cast(lhs);
lhs_tensor = lhs_tensor.matmul(ctorch::cast(rhs));
}

13
kmath-torch/src/nativeGPUTest/kotlin/kscience/kmath/torch/TestTorchTensorAlgebraGPU.kt Normal file
View File

@ -0,0 +1,13 @@
package kscience.kmath.torch
import kotlin.test.*
class TestTorchTensorAlgebraGPU {
@Test
fun testScalarProduct() = testingScalarProduct(device = TorchDevice.TorchCUDA(0))
@Test
fun testMatrixMultiplication() = testingMatrixMultiplication(device = TorchDevice.TorchCUDA(0))
}

25
kmath-torch/src/nativeGPUTest/kotlin/kscience/kmath/torch/TestTorchTensorGPU.kt
View File

@ -1,25 +1,22 @@
package kscience.kmath.torch package kscience.kmath.torch
import kscience.kmath.structures.asBuffer
import kotlinx.cinterop.memScoped
import kotlin.test.* import kotlin.test.*
class TestTorchTensorGPU { class TestTorchTensorGPU {
@Test @Test
fun cudaAvailability() { fun testCopyFromArray() = testingCopyFromArray(TorchDevice.TorchCUDA(0))
assertTrue(cudaAvailable())
}
@Test @Test
fun floatGPUTensorLayout() = memScoped { fun testCopyToDevice() = TorchTensorRealAlgebra {
val array = (1..8).map { it * 2f }.toList().toFloatArray() setSeed(SEED)
val shape = intArrayOf(2, 2, 2) val normalCpu = randNormal(intArrayOf(2, 3))
val tensor = TorchTensor.copyFromFloatArrayToGPU(this, array, shape, 0) val normalGpu = normalCpu.copyToDevice(TorchDevice.TorchCUDA(0))
tensor.elements().forEach { assertTrue(normalCpu.copyToArray() contentEquals normalGpu.copyToArray())
assertEquals(tensor[it.first], it.second)
} val uniformGpu = randUniform(intArrayOf(3,2),TorchDevice.TorchCUDA(0))
assertTrue(tensor.asBuffer().contentEquals(array.asBuffer())) val uniformCpu = uniformGpu.copyToDevice(TorchDevice.TorchCPU)
assertTrue(uniformGpu.copyToArray() contentEquals uniformCpu.copyToArray())
} }
} }

33
kmath-torch/src/nativeGPUTest/kotlin/kscience/kmath/torch/TestUtilsGPU.kt Normal file
View File

@ -0,0 +1,33 @@
package kscience.kmath.torch
import kotlin.test.*
internal class TestUtilsGPU {
@Test
fun testCudaAvailable() {
assertTrue(cudaAvailable())
}
@Test
fun testSetSeed() = testingSetSeed(TorchDevice.TorchCUDA(0))
@Test
fun testReadmeFactory() = TorchTensorRealAlgebra {
val realTensor: TorchTensorReal = copyFromArray(
array = (1..10).map { it + 50.0 }.toList().toDoubleArray(),
shape = intArrayOf(2,5)
)
println(realTensor)
val gpuRealTensor: TorchTensorReal = copyFromArray(
array = (1..8).map { it * 2.5 }.toList().toDoubleArray(),
shape = intArrayOf(2, 2, 2),
device = TorchDevice.TorchCUDA(0)
)
println(gpuRealTensor)
}
}

2
kmath-torch/src/nativeInterop/cinterop/libctorch.def
View File

@ -1,2 +1,2 @@
package=ctorch package=kscience.kmath.ctorch
headers=ctorch.h headers=ctorch.h

18
kmath-torch/src/nativeMain/kotlin/kscience.kmath.torch/TorchDevice.kt Normal file
View File

@ -0,0 +1,18 @@
package kscience.kmath.torch
public sealed class TorchDevice {
public object TorchCPU: TorchDevice()
public data class TorchCUDA(val index: Int): TorchDevice()
public fun toInt(): Int {
when(this) {
is TorchCPU -> return 0
is TorchCUDA -> return this.index + 1
}
}
public companion object {
public fun fromInt(deviceInt: Int): TorchDevice {
return if (deviceInt == 0) TorchCPU else TorchCUDA(deviceInt-1)
}
}
}

93
kmath-torch/src/nativeMain/kotlin/kscience.kmath.torch/TorchTensor.kt Normal file
View File

@ -0,0 +1,93 @@
package kscience.kmath.torch
import kscience.kmath.structures.MutableNDStructure
import kotlinx.cinterop.*
import kscience.kmath.ctorch.*
public sealed class TorchTensor<T> constructor(
internal val scope: DeferScope,
internal val tensorHandle: COpaquePointer
) : MutableNDStructure<T> {
init {
scope.defer(::close)
}
private fun close(): Unit = dispose_tensor(tensorHandle)
protected abstract fun item(): T
internal abstract fun wrap(outScope: DeferScope, outTensorHandle: COpaquePointer): TorchTensor<T>
override val dimension: Int get() = get_dim(tensorHandle)
override val shape: IntArray
get() = (1..dimension).map{get_shape_at(tensorHandle, it-1)}.toIntArray()
public val strides: IntArray
get() = (1..dimension).map{get_stride_at(tensorHandle, it-1)}.toIntArray()
public val size: Int get() = get_numel(tensorHandle)
public val device: TorchDevice get() = TorchDevice.fromInt(get_device(tensorHandle))
override fun equals(other: Any?): Boolean = false
override fun hashCode(): Int = 0
override fun toString(): String {
val nativeStringRepresentation: CPointer<ByteVar> = tensor_to_string(tensorHandle)!!
val stringRepresentation = nativeStringRepresentation.toKString()
dispose_char(nativeStringRepresentation)
return stringRepresentation
}
override fun elements(): Sequence<Pair<IntArray, T>> {
if (dimension == 0) {
return emptySequence()
}
val indices = (1..size).asSequence().map { indexFromOffset(it - 1, strides, dimension) }
return indices.map { it to get(it) }
}
public fun value(): T {
check(dimension == 0) {
"This tensor has shape ${shape.toList()}"
}
return item()
}
public fun copy(): TorchTensor<T> =
wrap(
outScope = scope,
outTensorHandle = copy_tensor(tensorHandle)!!
)
public fun copyToDevice(device: TorchDevice): TorchTensor<T> =
wrap(
outScope = scope,
outTensorHandle = copy_to_device(tensorHandle, device.toInt())!!
)
}
public class TorchTensorReal internal constructor(
scope: DeferScope,
tensorHandle: COpaquePointer
) : TorchTensor<Double>(scope, tensorHandle) {
override fun item(): Double = get_item_double(tensorHandle)
override fun wrap(outScope: DeferScope, outTensorHandle: COpaquePointer
): TorchTensorReal = TorchTensorReal(scope = outScope, tensorHandle = outTensorHandle)
override fun get(index: IntArray): Double = get_double(tensorHandle, index.toCValues())
override fun set(index: IntArray, value: Double) {
set_double(tensorHandle, index.toCValues(), value)
}
}
private inline fun indexFromOffset(offset: Int, strides: IntArray, nDim: Int): IntArray {
val res = IntArray(nDim)
var current = offset
var strideIndex = 0
while (strideIndex < nDim) {
res[strideIndex] = (current / strides[strideIndex])
current %= strides[strideIndex]
strideIndex++
}
return res
}

66
kmath-torch/src/nativeMain/kotlin/kscience.kmath.torch/TorchTensorAlgebra.kt Normal file
View File

@ -0,0 +1,66 @@
package kscience.kmath.torch
import kotlinx.cinterop.*
import kscience.kmath.ctorch.*
public sealed class TorchTensorAlgebra<T, PrimitiveArrayType> constructor(
internal val scope: DeferScope
) {
internal abstract fun wrap(tensorHandle: COpaquePointer): TorchTensor<T>
public abstract fun copyFromArray(
array: PrimitiveArrayType,
shape: IntArray,
device: TorchDevice = TorchDevice.TorchCPU
): TorchTensor<T>
public abstract fun TorchTensor<T>.copyToArray(): PrimitiveArrayType
public infix fun TorchTensor<T>.dot(other: TorchTensor<T>): TorchTensor<T> =
wrap(matmul(this.tensorHandle, other.tensorHandle)!!)
public infix fun TorchTensor<T>.dotAssign(other: TorchTensor<T>): Unit {
matmul_assign(this.tensorHandle, other.tensorHandle)
}
}
public sealed class TorchTensorFieldAlgebra<T, PrimitiveArrayType>(scope: DeferScope) :
TorchTensorAlgebra<T, PrimitiveArrayType>(scope) {
public abstract fun randNormal(shape: IntArray, device: TorchDevice = TorchDevice.TorchCPU): TorchTensor<T>
public abstract fun randUniform(shape: IntArray, device: TorchDevice = TorchDevice.TorchCPU): TorchTensor<T>
}
public class TorchTensorRealAlgebra(scope: DeferScope) : TorchTensorFieldAlgebra<Double, DoubleArray>(scope) {
override fun wrap(tensorHandle: COpaquePointer): TorchTensorReal =
TorchTensorReal(scope = scope, tensorHandle = tensorHandle)
override fun TorchTensor<Double>.copyToArray(): DoubleArray =
this.elements().map { it.second }.toList().toDoubleArray()
override fun copyFromArray(
array: DoubleArray,
shape: IntArray,
device: TorchDevice
): TorchTensorReal =
TorchTensorReal(
scope = scope,
tensorHandle = copy_from_blob_double(
array.toCValues(),
shape.toCValues(),
shape.size,
device.toInt()
)!!
)
override fun randNormal(shape: IntArray, device: TorchDevice): TorchTensorReal = TorchTensorReal(
scope = scope,
tensorHandle = randn_double(shape.toCValues(), shape.size, device.toInt())!!
)
override fun randUniform(shape: IntArray, device: TorchDevice): TorchTensorReal = TorchTensorReal(
scope = scope,
tensorHandle = rand_double(shape.toCValues(), shape.size, device.toInt())!!
)
}
public fun <R> TorchTensorRealAlgebra(block: TorchTensorRealAlgebra.() -> R): R =
memScoped { TorchTensorRealAlgebra(this).block() }

2
kmath-torch/src/nativeMain/kotlin/kscience/kmath/torch/Utils.kt → kmath-torch/src/nativeMain/kotlin/kscience.kmath.torch/Utils.kt
View File

@ -1,7 +1,7 @@
package kscience.kmath.torch package kscience.kmath.torch
import kotlinx.cinterop.* import kotlinx.cinterop.*
import ctorch.* import kscience.kmath.ctorch.*
public fun getNumThreads(): Int { public fun getNumThreads(): Int {
return get_num_threads() return get_num_threads()

18
kmath-torch/src/nativeMain/kotlin/kscience/kmath/torch/TorchMemoryHolder.kt
View File

@ -1,18 +0,0 @@
package kscience.kmath.torch
import kotlinx.cinterop.*
import ctorch.*
public abstract class TorchMemoryHolder internal constructor(
internal val scope: DeferScope,
internal var tensorHandle: COpaquePointer?
){
init {
scope.defer(::close)
}
protected fun close() {
dispose_tensor(tensorHandle)
tensorHandle = null
}
}

119
kmath-torch/src/nativeMain/kotlin/kscience/kmath/torch/TorchTensor.kt
View File

@ -1,119 +0,0 @@
package kscience.kmath.torch
import kscience.kmath.structures.*
import kotlinx.cinterop.*
import ctorch.*
public sealed class TorchTensor<T, out TorchTensorBufferImpl : TorchTensorBuffer<T>> :
MutableNDBufferTrait<T, TorchTensorBufferImpl, TorchTensorStrides>() {
public fun asBuffer(): MutableBuffer<T> = buffer
public companion object {
public fun copyFromFloatArray(scope: DeferScope, array: FloatArray, shape: IntArray): TorchTensorFloat {
val tensorHandle: COpaquePointer = copy_from_blob_float(
array.toCValues(), shape.toCValues(), shape.size
)!!
return TorchTensorFloat(
scope = scope,
tensorHandle = tensorHandle,
strides = populateStridesFromNative(tensorHandle, rawShape = shape)
)
}
public fun copyFromIntArray(scope: DeferScope, array: IntArray, shape: IntArray): TorchTensorInt {
val tensorHandle: COpaquePointer = copy_from_blob_int(
array.toCValues(), shape.toCValues(), shape.size
)!!
return TorchTensorInt(
scope = scope,
tensorHandle = tensorHandle,
strides = populateStridesFromNative(tensorHandle, rawShape = shape)
)
}
public fun copyFromFloatArrayToGPU(
scope: DeferScope,
array: FloatArray,
shape: IntArray,
device: Int
): TorchTensorFloatGPU {
val tensorHandle: COpaquePointer = copy_from_blob_to_gpu_float(
array.toCValues(), shape.toCValues(), shape.size, device
)!!
return TorchTensorFloatGPU(
scope = scope,
tensorHandle = tensorHandle,
strides = populateStridesFromNative(tensorHandle, rawShape = shape)
)
}
}
override fun toString(): String {
val nativeStringRepresentation: CPointer<ByteVar> = tensor_to_string(buffer.tensorHandle!!)!!
val stringRepresentation = nativeStringRepresentation.toKString()
dispose_char(nativeStringRepresentation)
return stringRepresentation
}
protected abstract fun wrap(
outScope: DeferScope,
outTensorHandle: COpaquePointer,
outStrides: TorchTensorStrides
): TorchTensor<T, TorchTensorBufferImpl>
public fun copy(): TorchTensor<T, TorchTensorBufferImpl> = wrap(
outScope = buffer.scope,
outTensorHandle = copy_tensor(buffer.tensorHandle!!)!!,
outStrides = strides
)
}
public class TorchTensorFloat internal constructor(
scope: DeferScope,
tensorHandle: COpaquePointer,
override val strides: TorchTensorStrides
) : TorchTensor<Float, TorchTensorBufferFloat>() {
override val buffer: TorchTensorBufferFloat = TorchTensorBufferFloat(scope, tensorHandle)
override fun wrap(
outScope: DeferScope,
outTensorHandle: COpaquePointer,
outStrides: TorchTensorStrides
): TorchTensorFloat = TorchTensorFloat(
scope = outScope, tensorHandle = outTensorHandle, strides = outStrides
)
}
public class TorchTensorInt internal constructor(
scope: DeferScope,
tensorHandle: COpaquePointer,
override val strides: TorchTensorStrides
) : TorchTensor<Int, TorchTensorBufferInt>() {
override val buffer: TorchTensorBufferInt = TorchTensorBufferInt(scope, tensorHandle)
override fun wrap(
outScope: DeferScope,
outTensorHandle: COpaquePointer,
outStrides: TorchTensorStrides
): TorchTensorInt = TorchTensorInt(
scope = outScope, tensorHandle = outTensorHandle, strides = outStrides
)
}
public class TorchTensorFloatGPU internal constructor(
scope: DeferScope,
tensorHandle: COpaquePointer,
override val strides: TorchTensorStrides
) : TorchTensor<Float, TorchTensorBufferFloatGPU>() {
override val buffer: TorchTensorBufferFloatGPU = TorchTensorBufferFloatGPU(scope, tensorHandle)
override fun wrap(
outScope: DeferScope,
outTensorHandle: COpaquePointer,
outStrides: TorchTensorStrides
): TorchTensorFloatGPU =
TorchTensorFloatGPU(
scope = outScope, tensorHandle = outTensorHandle, strides = outStrides
)
}

65
kmath-torch/src/nativeMain/kotlin/kscience/kmath/torch/TorchTensorAlgebra.kt
View File

@ -1,65 +0,0 @@
package kscience.kmath.torch
import kotlinx.cinterop.*
import ctorch.*
public sealed class TorchTensorAlgebra<
T,
TorchTensorBufferImpl : TorchTensorBuffer<T>,
PrimitiveArrayType>
constructor(
internal val scope: DeferScope
) {
protected abstract fun wrap(
outTensorHandle: COpaquePointer,
outStrides: TorchTensorStrides
): TorchTensor<T, TorchTensorBufferImpl>
public infix fun TorchTensor<T, TorchTensorBufferImpl>.swap(other: TorchTensor<T, TorchTensorBufferImpl>): Unit {
check(this.shape contentEquals other.shape) {
"Attempt to swap tensors with different shapes"
}
this.buffer.tensorHandle = other.buffer.tensorHandle.also {
other.buffer.tensorHandle = this.buffer.tensorHandle
}
}
public abstract fun copyFromArray(array: PrimitiveArrayType, shape: IntArray): TorchTensor<T, TorchTensorBufferImpl>
public infix fun TorchTensor<T, TorchTensorBufferImpl>.dot(other: TorchTensor<T, TorchTensorBufferImpl>):
TorchTensor<T, TorchTensorBufferImpl> {
val resultHandle = matmul(this.buffer.tensorHandle, other.buffer.tensorHandle)!!
val strides = populateStridesFromNative(tensorHandle = resultHandle)
return wrap(resultHandle, strides)
}
}
public sealed class TorchTensorField<T, TorchTensorBufferImpl : TorchTensorBuffer<T>, PrimitiveArrayType>
constructor(scope: DeferScope) : TorchTensorAlgebra<T, TorchTensorBufferImpl, PrimitiveArrayType>(scope) {
public abstract fun randn(shape: IntArray): TorchTensor<T, TorchTensorBufferImpl>
}
public class TorchTensorFloatAlgebra(scope: DeferScope) :
TorchTensorField<Float, TorchTensorBufferFloat, FloatArray>(scope) {
override fun wrap(
outTensorHandle: COpaquePointer,
outStrides: TorchTensorStrides
): TorchTensorFloat = TorchTensorFloat(scope = scope, tensorHandle = outTensorHandle, strides = outStrides)
override fun randn(shape: IntArray): TorchTensor<Float, TorchTensorBufferFloat> {
val tensorHandle = randn_float(shape.toCValues(), shape.size)!!
val strides = populateStridesFromNative(tensorHandle = tensorHandle, rawShape = shape)
return wrap(tensorHandle, strides)
}
override fun copyFromArray(array: FloatArray, shape: IntArray): TorchTensorFloat =
TorchTensor.copyFromFloatArray(scope, array, shape)
}
public fun <R> TorchTensorFloatAlgebra(block: TorchTensorFloatAlgebra.() -> R): R =
memScoped { TorchTensorFloatAlgebra(this).block() }

98
kmath-torch/src/nativeMain/kotlin/kscience/kmath/torch/TorchTensorBuffer.kt
View File

@ -1,98 +0,0 @@
package kscience.kmath.torch
import kscience.kmath.structures.MutableBuffer
import kotlinx.cinterop.*
import ctorch.*
public sealed class TorchTensorBuffer<T> constructor(
scope: DeferScope,
tensorHandle: COpaquePointer?
) : MutableBuffer<T>, TorchMemoryHolder(scope, tensorHandle) {
override val size: Int
get(){
return get_numel(tensorHandle!!)
}
internal abstract fun wrap(outScope: DeferScope, outTensorHandle: COpaquePointer): TorchTensorBuffer<T>
override fun copy(): TorchTensorBuffer<T> = wrap(
outScope = scope,
outTensorHandle = copy_tensor(tensorHandle!!)!!
)
}
public class TorchTensorBufferFloat internal constructor(
scope: DeferScope,
tensorHandle: COpaquePointer
) : TorchTensorBuffer<Float>(scope, tensorHandle) {
private val tensorData: CPointer<FloatVar>
get(){
return get_data_float(tensorHandle!!)!!
}
override operator fun get(index: Int): Float = tensorData[index]
override operator fun set(index: Int, value: Float) {
tensorData[index] = value
}
override operator fun iterator(): Iterator<Float> = (1..size).map { tensorData[it - 1] }.iterator()
override fun wrap(outScope: DeferScope, outTensorHandle: COpaquePointer) = TorchTensorBufferFloat(
scope = outScope,
tensorHandle = outTensorHandle
)
}
public class TorchTensorBufferInt internal constructor(
scope: DeferScope,
tensorHandle: COpaquePointer
) : TorchTensorBuffer<Int>(scope, tensorHandle) {
private val tensorData: CPointer<IntVar>
get(){
return get_data_int(tensorHandle!!)!!
}
override operator fun get(index: Int): Int = tensorData[index]
override operator fun set(index: Int, value: Int) {
tensorData[index] = value
}
override operator fun iterator(): Iterator<Int> = (1..size).map { tensorData[it - 1] }.iterator()
override fun wrap(outScope: DeferScope, outTensorHandle: COpaquePointer) = TorchTensorBufferInt(
scope = outScope,
tensorHandle = outTensorHandle
)
}
public class TorchTensorBufferFloatGPU internal constructor(
scope: DeferScope,
tensorHandle: COpaquePointer
) : TorchTensorBuffer<Float>(scope, tensorHandle) {
override operator fun get(index: Int): Float = get_at_offset_float(tensorHandle!!, index)
override operator fun set(index: Int, value: Float) {
set_at_offset_float(tensorHandle!!, index, value)
}
override operator fun iterator(): Iterator<Float> {
val cpuCopy = copy_to_cpu(tensorHandle!!)!!
val tensorCpuData = get_data_float(cpuCopy)!!
val iteratorResult = (1..size).map { tensorCpuData[it - 1] }.iterator()
dispose_tensor(cpuCopy)
return iteratorResult
}
override fun wrap(outScope: DeferScope, outTensorHandle: COpaquePointer) = TorchTensorBufferFloatGPU(
scope = outScope,
tensorHandle = outTensorHandle
)
}

55
kmath-torch/src/nativeMain/kotlin/kscience/kmath/torch/TorchTensorStrides.kt
View File

@ -1,55 +0,0 @@
package kscience.kmath.torch
import kscience.kmath.structures.Strides
import kotlinx.cinterop.*
import ctorch.*
public class TorchTensorStrides internal constructor(
override val shape: IntArray,
override val strides: IntArray,
override val linearSize: Int
) : Strides {
override fun index(offset: Int): IntArray {
val nDim = shape.size
val res = IntArray(nDim)
var current = offset
var strideIndex = 0
while (strideIndex < nDim) {
res[strideIndex] = (current / strides[strideIndex])
current %= strides[strideIndex]
strideIndex++
}
return res
}
}
private inline fun intPointerToArrayAndClean(ptr: CPointer<IntVar>, nDim: Int): IntArray {
val res: IntArray = (1 .. nDim).map{ptr[it-1]}.toIntArray()
dispose_int_array(ptr)
return res
}
private inline fun getShapeFromNative(tensorHandle: COpaquePointer, nDim: Int): IntArray{
return intPointerToArrayAndClean(get_shape(tensorHandle)!!, nDim)
}
private inline fun getStridesFromNative(tensorHandle: COpaquePointer, nDim: Int): IntArray{
return intPointerToArrayAndClean(get_strides(tensorHandle)!!, nDim)
}
internal inline fun populateStridesFromNative(
tensorHandle: COpaquePointer,
rawShape: IntArray? = null,
rawStrides: IntArray? = null,
rawLinearSize: Int? = null
): TorchTensorStrides {
val nDim = rawShape?.size?: rawStrides?.size?: get_dim(tensorHandle)
return TorchTensorStrides(
shape = rawShape?: getShapeFromNative(tensorHandle, nDim),
strides = rawStrides?: getStridesFromNative(tensorHandle, nDim),
linearSize = rawLinearSize?: get_numel(tensorHandle)
)
}

51
kmath-torch/src/nativeTest/kotlin/kscience/kmath/torch/TestTorchTensor.kt
View File

@ -1,45 +1,22 @@
package kscience.kmath.torch package kscience.kmath.torch
import kscience.kmath.structures.asBuffer
import kotlinx.cinterop.memScoped
import kotlin.test.* import kotlin.test.*
internal fun testingCopyFromArray(device: TorchDevice = TorchDevice.TorchCPU): Unit {
TorchTensorRealAlgebra {
val array = (1..24).map { 10.0 * it * it }.toDoubleArray()
val shape = intArrayOf(2, 3, 4)
val tensor = copyFromArray(array, shape = shape, device = device)
val copyOfTensor = tensor.copy()
tensor[intArrayOf(0, 0)] = 0.1
assertTrue(copyOfTensor.copyToArray() contentEquals array)
assertEquals(0.1, tensor[intArrayOf(0, 0)])
}
}
internal class TestTorchTensor {
class TestTorchTensor {
@Test @Test
fun intTensorLayout() = memScoped { fun testCopyFromArray() = testingCopyFromArray()
val array = (1..24).toList().toIntArray()
val shape = intArrayOf(4, 6)
val tensor = TorchTensor.copyFromIntArray(scope = this, array = array, shape = shape)
tensor.elements().forEach {
assertEquals(tensor[it.first], it.second)
}
assertTrue(tensor.asBuffer().contentEquals(array.asBuffer()))
}
@Test
fun floatTensorLayout() = memScoped {
val array = (1..10).map { it + 50f }.toList().toFloatArray()
val shape = intArrayOf(10)
val tensor = TorchTensor.copyFromFloatArray(this, array, shape)
tensor.elements().forEach {
assertEquals(tensor[it.first], it.second)
}
assertTrue(tensor.asBuffer().contentEquals(array.asBuffer()))
}
@Test
fun mutableStructure() = memScoped {
val array = (1..10).map { 1f * it }.toList().toFloatArray()
val shape = intArrayOf(10)
val tensor = TorchTensor.copyFromFloatArray(this, array, shape)
val tensorCopy = tensor.copy()
tensor[intArrayOf(0)] = 99f
assertEquals(99f, tensor[intArrayOf(0)])
assertEquals(1f, tensorCopy[intArrayOf(0)])
}
} }

66
kmath-torch/src/nativeTest/kotlin/kscience/kmath/torch/TestTorchTensorAlgebra.kt
View File

@ -1,36 +1,52 @@
package kscience.kmath.torch package kscience.kmath.torch
import kscience.kmath.linear.RealMatrixContext
import kscience.kmath.operations.invoke
import kscience.kmath.structures.Matrix
import kotlin.math.abs
import kotlin.test.* import kotlin.test.*
import kotlin.time.measureTime
internal fun testingScalarProduct(device: TorchDevice = TorchDevice.TorchCPU): Unit {
TorchTensorRealAlgebra {
val lhs = randUniform(shape = intArrayOf(10), device = device)
val rhs = randUniform(shape = intArrayOf(10), device = device)
val product = lhs dot rhs
var expected = 0.0
lhs.elements().forEach {
expected += it.second * rhs[it.first]
}
assertTrue(abs(expected - product.value()) < TOLERANCE)
}
}
class TestTorchTensorAlgebra { internal fun testingMatrixMultiplication(device: TorchDevice = TorchDevice.TorchCPU): Unit {
TorchTensorRealAlgebra {
setSeed(SEED)
val lhsTensor = randNormal(shape = intArrayOf(20, 20), device = device)
val rhsTensor = randNormal(shape = intArrayOf(20, 20), device = device)
val product = lhsTensor dot rhsTensor
val expected: Matrix<Double> = RealMatrixContext {
val lhs = produce(20, 20) { i, j -> lhsTensor[intArrayOf(i, j)] }
val rhs = produce(20, 20) { i, j -> rhsTensor[intArrayOf(i, j)] }
lhs dot rhs
}
var error: Double = 0.0
product.elements().forEach {
error += abs(expected[it.first] - it.second)
}
assertTrue(error < TOLERANCE)
}
}
internal class TestTorchTensorAlgebra {
@Test @Test
fun swappingTensors() = TorchTensorFloatAlgebra { fun testScalarProduct() = testingScalarProduct()
val tensorA = copyFromArray(floatArrayOf(1f, 2f, 3f), intArrayOf(3))
val tensorB = tensorA.copy()
val tensorC = copyFromArray(floatArrayOf(4f, 5f, 6f), intArrayOf(3))
tensorA swap tensorC
assertTrue(tensorB.asBuffer().contentEquals(tensorC.asBuffer()))
}
@Test @Test
fun dotOperation() = TorchTensorFloatAlgebra { fun testMatrixMultiplication() = testingMatrixMultiplication()
setSeed(987654)
var tensorA = randn(intArrayOf(1000, 1000))
val tensorB = randn(intArrayOf(1000, 1000))
measureTime {
repeat(100) {
TorchTensorFloatAlgebra {
tensorA swap (tensorA dot tensorB)
}
}
}.also(::println)
assertTrue(tensorA.shape contentEquals tensorB.shape)
}
} }

28
kmath-torch/src/nativeTest/kotlin/kscience/kmath/torch/TestUtils.kt
View File

@ -3,19 +3,31 @@ package kscience.kmath.torch
import kotlin.test.* import kotlin.test.*
internal val SEED = 987654
internal val TOLERANCE = 1e-6
internal fun testingSetSeed(device: TorchDevice = TorchDevice.TorchCPU): Unit {
TorchTensorRealAlgebra {
setSeed(SEED)
val normal = randNormal(IntArray(0), device = device).value()
val uniform = randUniform(IntArray(0), device = device).value()
setSeed(SEED)
val nextNormal = randNormal(IntArray(0), device = device).value()
val nextUniform = randUniform(IntArray(0), device = device).value()
assertEquals(normal, nextNormal)
assertEquals(uniform, nextUniform)
}
}
internal class TestUtils { internal class TestUtils {
@Test @Test
fun settingTorchThreadsCount() { fun testSetNumThreads() {
val numThreads = 2 val numThreads = 2
setNumThreads(numThreads) setNumThreads(numThreads)
assertEquals(numThreads, getNumThreads()) assertEquals(numThreads, getNumThreads())
} }
@Test @Test
fun seedSetting() = TorchTensorFloatAlgebra { fun testSetSeed() = testingSetSeed()
setSeed(987654)
val tensorA = randn(intArrayOf(2,3))
setSeed(987654)
val tensorB = randn(intArrayOf(2,3))
assertTrue(tensorA.asBuffer().contentEquals(tensorB.asBuffer()))
}
} }