Buffer protocol for torch tensors

.gitignore

@@ -2,6 +2,7 @@
 build/
 out/
 .idea/
+.vscode/
 
 # Avoid ignoring Gradle wrapper jar file (.jar files are usually ignored)
 !gradle-wrapper.jar

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

@@ -148,23 +148,28 @@ public interface Strides {
     /**
      * Array strides
      */
-    public val strides: List<Int>
-
-    /**
-     * Get linear index from multidimensional index
-     */
-    public fun offset(index: IntArray): Int
-
-    /**
-     * Get multidimensional from linear
-     */
-    public fun index(offset: Int): IntArray
+    public val strides: IntArray
 
     /**
      * 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
+     */
+    public 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()
+
+    /**
+     * Get multidimensional from linear
+     */
+    public fun index(offset: Int): IntArray
+
     // TODO introduce a fast way to calculate index of the next element?
 
     /**
@@ -183,7 +188,7 @@ public class DefaultStrides private constructor(override val shape: IntArray) :
     /**
      * Strides for memory access
      */
-    override val strides: List<Int> by lazy {
+    override val strides: IntArray by lazy {
         sequence {
             var current = 1
             yield(1)
@@ -192,16 +197,9 @@ public class DefaultStrides private constructor(override val shape: IntArray) :
                 current *= it
                 yield(current)
             }
-        }.toList()
+        }.toList().toIntArray()
     }
 
-    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 {
         val res = IntArray(shape.size)
         var current = offset
@@ -238,20 +236,22 @@ public class DefaultStrides private constructor(override val shape: IntArray) :
 }
 
 /**
- * Represents [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 NDBuffer<T> : NDStructure<T> {
+public abstract class NDBufferTrait<T, out BufferImpl : Buffer<T>, out StridesImpl: Strides> :
+    NDStructure<T> {
     /**
      * The underlying buffer.
      */
-    public abstract val buffer: Buffer<T>
+    public abstract val buffer: BufferImpl
 
     /**
      * The strides to access elements of [Buffer] by linear indices.
      */
-    public abstract val strides: Strides
+    public abstract val strides: StridesImpl
 
     override operator fun get(index: IntArray): T = buffer[strides.offset(index)]
 
@@ -259,8 +259,8 @@ public abstract class NDBuffer<T> : NDStructure<T> {
 
     override fun elements(): Sequence<Pair<IntArray, T>> = strides.indices().map { it to this[it] }
 
-    override fun equals(other: Any?): Boolean {
-        return NDStructure.equals(this, other as? NDStructure<*> ?: return false)
+    public fun checkStridesBufferCompatibility(): Unit = require(strides.linearSize == buffer.size) {
+        "Expected buffer side of ${strides.linearSize}, but found ${buffer.size}"
     }
 
     override fun hashCode(): Int {
@@ -269,6 +269,10 @@ public abstract class NDBuffer<T> : NDStructure<T> {
         return result
     }
 
+    override fun equals(other: Any?): Boolean {
+        return NDStructure.equals(this, other as? NDStructure<*> ?: return false)
+    }
+
     override fun toString(): String {
         val bufferRepr: String = when (shape.size) {
             1 -> buffer.asSequence().joinToString(prefix = "[", postfix = "]", separator = ", ")
@@ -282,10 +286,36 @@ public abstract class NDBuffer<T> : NDStructure<T> {
         }
         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]
  */
@@ -294,9 +324,7 @@ public class BufferNDStructure<T>(
     override val buffer: Buffer<T>,
 ) : NDBuffer<T>() {
     init {
-        if (strides.linearSize != buffer.size) {
-            error("Expected buffer side of ${strides.linearSize}, but found ${buffer.size}")
-        }
+        checkStridesBufferCompatibility()
     }
 }
 
@@ -316,20 +344,15 @@ public inline fun <T, reified R : Any> NDStructure<T>.mapToBuffer(
 }
 
 /**
- * Mutable ND buffer based on linear [MutableBuffer].
+ * Boxing generic [MutableNDStructure].
  */
 public class MutableBufferNDStructure<T>(
     override val strides: Strides,
     override val buffer: MutableBuffer<T>,
-) : NDBuffer<T>(), MutableNDStructure<T> {
-
+) : MutableNDBuffer<T>() {
     init {
-        require(strides.linearSize == buffer.size) {
-            "Expected buffer side of ${strides.linearSize}, but found ${buffer.size}"
+        checkStridesBufferCompatibility()
     }
-    }
-
-    override operator fun set(index: IntArray, value: T): Unit = buffer.set(strides.offset(index), value)
 }
 
 public inline fun <reified T : Any> NDStructure<T>.combine(

kmath-torch/README.md

@@ -0,0 +1,95 @@
+# LibTorch extension (`kmath-torch`)
+
+This is a `Kotlin/Native` module, with only `linuxX64` supported so far. This library wraps some of the [PyTorch C++ API](https://pytorch.org/cppdocs), focusing on integrating `Aten` & `Autograd` with `KMath`. 
+
+## Installation
+To install the library, you have to build & publish locally `kmath-core`, `kmath-memory` with `kmath-torch`:
+```
+./gradlew -q :kmath-core:publishToMavenLocal :kmath-memory:publishToMavenLocal :kmath-torch:publishToMavenLocal
+```
+
+This builds `ctorch`, a C wrapper for `LibTorch` placed inside:
+
+`~/.konan/third-party/kmath-torch-0.2.0-dev-4/cpp-build` 
+
+You will have to link against it in your own project. Here is an example of build script for a standalone application:
+```kotlin
+//build.gradle.kts
+plugins {
+    id("ru.mipt.npm.mpp")
+}
+
+repositories {
+    jcenter()
+    mavenLocal()
+}
+
+val home = System.getProperty("user.home")
+val kver = "0.2.0-dev-4"
+val cppBuildDir = "$home/.konan/third-party/kmath-torch-$kver/cpp-build"
+
+kotlin {
+    explicitApiWarning()
+
+    val nativeTarget = linuxX64("your.app")
+    nativeTarget.apply {
+        binaries {
+            executable {
+                entryPoint = "your.app.main"
+            }
+            all {
+                linkerOpts(
+                    "-L$cppBuildDir",
+                    "-Wl,-rpath=$cppBuildDir",
+                    "-lctorch"
+                )
+            }
+        }
+    }
+
+    val main by nativeTarget.compilations.getting
+
+    sourceSets {
+        val nativeMain by creating {
+            dependencies {
+                implementation("kscience.kmath:kmath-torch:$kver")
+            }
+        }
+        main.defaultSourceSet.dependsOn(nativeMain)
+    }
+}
+```
+```kotlin
+//settings.gradle.kts
+pluginManagement {
+    repositories {
+        gradlePluginPortal()
+        jcenter()
+        maven("https://dl.bintray.com/mipt-npm/dev")
+    }
+    plugins {
+        id("ru.mipt.npm.mpp") version "0.7.1"
+        kotlin("jvm") version "1.4.21"
+    }
+}
+```
+
+## Usage
+
+Tensors implement the buffer protocol over `MutableNDStructure`. They can only be instantiated through provided factory methods and require scoping:
+```kotlin
+memScoped {
+    val intTensor: TorchTensorInt = TorchTensor.copyFromIntArray(
+        scope = this,
+        array = intArrayOf(7,8,9,2,6,5), 
+        shape = intArrayOf(3,2))
+    println(intTensor)
+    
+    val floatTensor: TorchTensorFloat = TorchTensor.copyFromFloatArray(
+        scope = this,
+        array = floatArrayOf(7f,8.9f,2.6f,5.6f),
+        shape = intArrayOf(4))
+    println(intTensor)
+}
+```
+

kmath-torch/build.gradle.kts

@@ -0,0 +1,146 @@
+import de.undercouch.gradle.tasks.download.Download
+import org.jetbrains.kotlin.gradle.plugin.mpp.KotlinNativeTarget
+
+
+plugins {
+    id("ru.mipt.npm.mpp")
+    id("de.undercouch.download")
+}
+
+
+val home = System.getProperty("user.home")
+val thirdPartyDir = "$home/.konan/third-party/kmath-torch-${project.property("version")}"
+val cppBuildDir = "$thirdPartyDir/cpp-build"
+
+val cmakeArchive = "cmake-3.19.2-Linux-x86_64"
+val torchArchive = "libtorch"
+
+val cmakeCmd = "$thirdPartyDir/$cmakeArchive/bin/cmake"
+val ninjaCmd = "$thirdPartyDir/ninja"
+
+val downloadCMake by tasks.registering(Download::class) {
+    val tarFile = "$cmakeArchive.tar.gz"
+    src("https://github.com/Kitware/CMake/releases/download/v3.19.2/$tarFile")
+    dest(File(thirdPartyDir, tarFile))
+    overwrite(false)
+}
+
+val downloadNinja by tasks.registering(Download::class) {
+    src("https://github.com/ninja-build/ninja/releases/download/v1.10.2/ninja-linux.zip")
+    dest(File(thirdPartyDir, "ninja-linux.zip"))
+    overwrite(false)
+}
+
+val downloadTorch by tasks.registering(Download::class) {
+    val zipFile = "$torchArchive-cxx11-abi-shared-with-deps-1.7.1%2Bcu110.zip"
+    src("https://download.pytorch.org/libtorch/cu110/$zipFile")
+    dest(File(thirdPartyDir, "$torchArchive.zip"))
+    overwrite(false)
+}
+
+val extractCMake by tasks.registering(Copy::class) {
+    dependsOn(downloadCMake)
+    from(tarTree(resources.gzip(downloadCMake.get().dest)))
+    into(thirdPartyDir)
+}
+
+val extractTorch by tasks.registering(Copy::class) {
+    dependsOn(downloadTorch)
+    from(zipTree(downloadTorch.get().dest))
+    into(thirdPartyDir)
+}
+
+val extractNinja by tasks.registering(Copy::class) {
+    dependsOn(downloadNinja)
+    from(zipTree(downloadNinja.get().dest))
+    into(thirdPartyDir)
+}
+
+val configureCpp by tasks.registering {
+    dependsOn(extractCMake)
+    dependsOn(extractNinja)
+    dependsOn(extractTorch)
+    onlyIf { !file(cppBuildDir).exists() }
+    doLast {
+        exec {
+            workingDir(thirdPartyDir)
+            commandLine("mkdir", "-p", cppBuildDir)
+        }
+        exec {
+            workingDir(cppBuildDir)
+            commandLine(
+                cmakeCmd,
+                projectDir.resolve("ctorch"),
+                "-GNinja",
+                "-DCMAKE_MAKE_PROGRAM=$ninjaCmd",
+                "-DCMAKE_PREFIX_PATH=$thirdPartyDir/$torchArchive",
+                "-DCMAKE_BUILD_TYPE=Release"
+            )
+        }
+    }
+}
+
+val cleanCppBuild by tasks.registering {
+    onlyIf { file(cppBuildDir).exists() }
+    doLast {
+        exec {
+            workingDir(thirdPartyDir)
+            commandLine("rm", "-rf", cppBuildDir)
+        }
+    }
+}
+
+val buildCpp by tasks.registering {
+    dependsOn(configureCpp)
+    doLast {
+        exec {
+            workingDir(cppBuildDir)
+            commandLine(cmakeCmd, "--build", ".", "--config", "Release")
+        }
+    }
+}
+
+kotlin {
+    explicitApiWarning()
+
+    val nativeTarget = linuxX64("torch")
+    nativeTarget.apply {
+        binaries {
+            all {
+                linkerOpts(
+                    "-L$cppBuildDir",
+                    "-Wl,-rpath=$cppBuildDir",
+                    "-lctorch"
+                )
+            }
+        }
+    }
+
+    val main by nativeTarget.compilations.getting {
+        cinterops {
+            val libctorch by creating {
+                includeDirs(projectDir.resolve("ctorch/include"))
+            }
+        }
+    }
+
+    val test by nativeTarget.compilations.getting
+
+    sourceSets {
+        val nativeMain by creating {
+            dependencies {
+                api(project(":kmath-core"))
+            }
+        }
+        val nativeTest by creating {
+            dependsOn(nativeMain)
+        }
+
+        main.defaultSourceSet.dependsOn(nativeMain)
+        test.defaultSourceSet.dependsOn(nativeTest)
+    }
+}
+
+val torch: KotlinNativeTarget by kotlin.targets
+tasks[torch.compilations["main"].cinterops["libctorch"].interopProcessingTaskName]
+    .dependsOn(buildCpp)

kmath-torch/ctorch/CMakeLists.txt

@@ -0,0 +1,25 @@
+cmake_minimum_required(VERSION 3.12)
+
+project(CTorch LANGUAGES C CXX)
+
+# Require C++17
+set(CMAKE_CXX_STANDARD 17)
+
+# Build configuration
+if(NOT CMAKE_BUILD_TYPE)
+  set(CMAKE_BUILD_TYPE Release CACHE STRING "Build type" FORCE)
+endif()
+message(STATUS "Build type: ${CMAKE_BUILD_TYPE}")
+
+find_package(Torch REQUIRED)
+
+add_library(ctorch SHARED src/ctorch.cc)
+target_include_directories(ctorch PRIVATE include)
+target_link_libraries(ctorch PRIVATE torch)
+target_compile_options(ctorch PRIVATE -Wall -Wextra -Wpedantic -O3 -fPIC)
+
+include(GNUInstallDirs)
+set_target_properties(ctorch PROPERTIES PUBLIC_HEADER include/ctorch.h)
+install(TARGETS ctorch
+    LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR}
+    PUBLIC_HEADER DESTINATION ${CMAKE_INSTALL_INCLUDEDIR})

kmath-torch/ctorch/include/ctorch.h

@@ -0,0 +1,48 @@
+#ifndef CTORCH
+#define CTORCH
+
+#include <stdbool.h>
+
+#ifdef __cplusplus
+extern "C"
+{
+#endif
+
+    typedef void *TorchTensorHandle;
+
+    int get_num_threads();
+
+    void set_num_threads(int num_threads);
+
+    bool cuda_is_available();
+
+    void set_seed(int seed);
+
+    TorchTensorHandle copy_from_blob_double(double *data, int *shape, int dim);
+    TorchTensorHandle copy_from_blob_float(float *data, int *shape, int dim);
+    TorchTensorHandle copy_from_blob_long(long *data, int *shape, int dim);
+    TorchTensorHandle copy_from_blob_int(int *data, int *shape, int dim);
+
+    TorchTensorHandle copy_tensor(TorchTensorHandle tensor_handle);
+
+    double *get_data_double(TorchTensorHandle tensor_handle);
+    float *get_data_float(TorchTensorHandle tensor_handle);
+    long *get_data_long(TorchTensorHandle tensor_handle);
+    int *get_data_int(TorchTensorHandle tensor_handle);
+
+    int get_numel(TorchTensorHandle tensor_handle);
+    int get_dim(TorchTensorHandle tensor_handle);
+    int *get_shape(TorchTensorHandle tensor_handle);
+    int *get_strides(TorchTensorHandle tensor_handle);
+
+    char *tensor_to_string(TorchTensorHandle tensor_handle);
+
+    void dispose_int_array(int *ptr);
+    void dispose_char(char *ptr);
+    void dispose_tensor(TorchTensorHandle tensor_handle);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif //CTORCH

kmath-torch/ctorch/include/utils.hh

@@ -0,0 +1,56 @@
+#include <torch/torch.h>
+
+#include "ctorch.h"
+
+namespace ctorch
+{
+    template <typename Dtype>
+    inline c10::ScalarType dtype()
+    {
+        return torch::kFloat64;
+    }
+
+    template <>
+    inline c10::ScalarType dtype<float>()
+    {
+        return torch::kFloat32;
+    }
+
+    template <>
+    inline c10::ScalarType dtype<long>()
+    {
+        return torch::kInt64;
+    }
+
+    template <>
+    inline c10::ScalarType dtype<int>()
+    {
+        return torch::kInt32;
+    }
+
+    inline torch::Tensor &cast(TorchTensorHandle tensor_handle)
+    {
+        return *static_cast<torch::Tensor *>(tensor_handle);
+    }
+
+    template <typename Dtype>
+    inline torch::Tensor copy_from_blob(Dtype *data, int *shape, int dim)
+    {
+        auto shape_vec = std::vector<int64_t>(dim);
+        shape_vec.assign(shape, shape + dim);
+        return torch::from_blob(data, shape_vec, dtype<Dtype>()).clone();
+    }
+
+    template <typename IntArray>
+    inline int *to_dynamic_ints(IntArray arr)
+    {
+        size_t n = arr.size();
+        int *res = (int *)malloc(sizeof(int) * n);
+        for (size_t i = 0; i < n; i++)
+        {
+            res[i] = arr[i];
+        }
+        return res;
+    }
+
+} // namespace ctorch

kmath-torch/ctorch/src/ctorch.cc

@@ -0,0 +1,110 @@
+#include <torch/torch.h>
+#include <iostream>
+#include <stdlib.h>
+
+#include "ctorch.h"
+#include "utils.hh"
+
+int get_num_threads()
+{
+  return torch::get_num_threads();
+}
+
+void set_num_threads(int num_threads)
+{
+  torch::set_num_threads(num_threads);
+}
+
+bool cuda_is_available()
+{
+  return torch::cuda::is_available();
+}
+
+void set_seed(int 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, shape, dim));
+}
+TorchTensorHandle copy_from_blob_float(float *data, int *shape, int dim)
+{
+  return new torch::Tensor(ctorch::copy_from_blob<float>(data, shape, dim));
+}
+TorchTensorHandle copy_from_blob_long(long *data, int *shape, int dim)
+{
+  return new torch::Tensor(ctorch::copy_from_blob<long>(data, shape, dim));
+}
+TorchTensorHandle copy_from_blob_int(int *data, int *shape, int dim)
+{
+  return new torch::Tensor(ctorch::copy_from_blob<int>(data, shape, dim));
+}
+
+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)
+{
+  return ctorch::cast(tensor_handle).dim();
+}
+
+int *get_shape(TorchTensorHandle tensor_handle)
+{
+  return ctorch::to_dynamic_ints(ctorch::cast(tensor_handle).sizes());
+}
+
+int *get_strides(TorchTensorHandle tensor_handle)
+{
+  return ctorch::to_dynamic_ints(ctorch::cast(tensor_handle).strides());
+}
+
+char *tensor_to_string(TorchTensorHandle tensor_handle)
+{
+  std::stringstream bufrep;
+  bufrep << ctorch::cast(tensor_handle);
+  auto rep = bufrep.str();
+  char *crep = (char *)malloc(rep.length() + 1);
+  std::strcpy(crep, rep.c_str());
+  return crep;
+}
+
+void dispose_int_array(int *ptr)
+{
+  free(ptr);
+}
+
+void dispose_char(char *ptr)
+{
+  free(ptr);
+}
+
+void dispose_tensor(TorchTensorHandle tensor_handle)
+{
+  delete static_cast<torch::Tensor *>(tensor_handle);
+}

kmath-torch/src/nativeInterop/cinterop/libctorch.def

@@ -0,0 +1,2 @@
+package=ctorch
+headers=ctorch.h

kmath-torch/src/nativeMain/kotlin/kscience/kmath/torch/TorchTensor.kt

@@ -0,0 +1,52 @@
+package kscience.kmath.torch
+
+import kscience.kmath.structures.*
+
+import kotlinx.cinterop.*
+import ctorch.*
+
+public abstract class TorchTensor<T,
+        TVar : CPrimitiveVar,
+        TorchTensorBufferImpl : TorchTensorBuffer<T, TVar>> :
+    MutableNDBufferTrait<T, TorchTensorBufferImpl, TorchTensorStrides>() {
+
+    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(populateStridesFromNative(tensorHandle, rawShape = shape), scope, tensorHandle)
+        }
+        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(populateStridesFromNative(tensorHandle, rawShape = shape), scope, tensorHandle)
+        }
+    }
+
+    override fun toString(): String {
+        val nativeStringRepresentation: CPointer<ByteVar> = tensor_to_string(buffer.tensorHandle)!!
+        val stringRepresentation = nativeStringRepresentation.toKString()
+        dispose_char(nativeStringRepresentation)
+        return stringRepresentation
+    }
+
+}
+
+public class TorchTensorFloat internal constructor(
+    override val strides: TorchTensorStrides,
+    scope: DeferScope,
+    tensorHandle: COpaquePointer
+): TorchTensor<Float, FloatVar, TorchTensorBufferFloat>() {
+    override val buffer: TorchTensorBufferFloat = TorchTensorBufferFloat(scope, tensorHandle)
+}
+
+public class TorchTensorInt internal constructor(
+    override val strides: TorchTensorStrides,
+    scope: DeferScope,
+    tensorHandle: COpaquePointer
+): TorchTensor<Int, IntVar, TorchTensorBufferInt>() {
+    override val buffer: TorchTensorBufferInt = TorchTensorBufferInt(scope, tensorHandle)
+}
+

kmath-torch/src/nativeMain/kotlin/kscience/kmath/torch/TorchTensorBuffer.kt

@@ -0,0 +1,67 @@
+package kscience.kmath.torch
+
+import kscience.kmath.structures.MutableBuffer
+
+import kotlinx.cinterop.*
+import ctorch.*
+
+public abstract class TorchTensorBuffer<T, TVar : CPrimitiveVar> internal constructor(
+    internal val scope: DeferScope,
+    internal val tensorHandle: COpaquePointer
+) : MutableBuffer<T> {
+    init {
+        scope.defer(::close)
+    }
+
+    internal fun close() {
+        dispose_tensor(tensorHandle)
+    }
+
+    protected abstract val tensorData: CPointer<TVar>
+
+    override val size: Int
+        get() = get_numel(tensorHandle)
+
+}
+
+
+public class TorchTensorBufferFloat internal constructor(
+    scope: DeferScope,
+    tensorHandle: COpaquePointer
+) : TorchTensorBuffer<Float, FloatVar>(scope, tensorHandle) {
+    override val tensorData: CPointer<FloatVar> = 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 copy(): TorchTensorBufferFloat = TorchTensorBufferFloat(
+        scope = scope,
+        tensorHandle = copy_tensor(tensorHandle)!!
+    )
+}
+
+public class TorchTensorBufferInt internal constructor(
+    scope: DeferScope,
+    tensorHandle: COpaquePointer
+) : TorchTensorBuffer<Int, IntVar>(scope, tensorHandle) {
+    override val tensorData: CPointer<IntVar> = 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 copy(): TorchTensorBufferInt = TorchTensorBufferInt(
+        scope = scope,
+        tensorHandle = copy_tensor(tensorHandle)!!
+    )
+}
+

kmath-torch/src/nativeMain/kotlin/kscience/kmath/torch/TorchTensorStrides.kt

@@ -0,0 +1,55 @@
+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)
+    )
+}

kmath-torch/src/nativeMain/kotlin/kscience/kmath/torch/Utils.kt

@@ -0,0 +1,20 @@
+package kscience.kmath.torch
+
+import kotlinx.cinterop.*
+import ctorch.*
+
+public fun getNumThreads(): Int {
+    return get_num_threads()
+}
+
+public fun setNumThreads(numThreads: Int): Unit {
+    set_num_threads(numThreads)
+}
+
+public fun cudaAvailable(): Boolean {
+    return cuda_is_available()
+}
+
+public fun setSeed(seed: Int): Unit {
+    set_seed(seed)
+}

kmath-torch/src/nativeTest/kotlin/kscience/kmath/torch/TestTorchTensor.kt

@@ -0,0 +1,33 @@
+package kscience.kmath.torch
+
+import kscience.kmath.structures.asBuffer
+
+import kotlinx.cinterop.memScoped
+import kotlin.test.*
+
+
+internal class TestTorchTensor {
+
+    @Test
+    fun intTensorLayout() = memScoped {
+        val array = intArrayOf(7,8,9,2,6,5)
+        val shape = intArrayOf(3,2)
+        val tensor = TorchTensor.copyFromIntArray(scope=this, array=array, shape=shape)
+        tensor.elements().forEach {
+            assertEquals(tensor[it.first], it.second)
+        }
+        assertTrue(tensor.buffer.contentEquals(array.asBuffer()))
+    }
+
+    @Test
+    fun floatTensorLayout() = memScoped {
+        val array = floatArrayOf(7.5f,8.2f,9f,2.58f,6.5f,5f)
+        val shape = intArrayOf(2,3)
+        val tensor = TorchTensor.copyFromFloatArray(this, array, shape)
+        tensor.elements().forEach {
+            assertEquals(tensor[it.first], it.second)
+        }
+        assertTrue(tensor.buffer.contentEquals(array.asBuffer()))
+    }
+
+}

kmath-torch/src/nativeTest/kotlin/kscience/kmath/torch/TestUtils.kt

@@ -0,0 +1,19 @@
+package kscience.kmath.torch
+
+import kotlin.test.Test
+import kotlin.test.assertEquals
+import kotlin.test.assertTrue
+
+
+internal class TestUtils {
+    @Test
+    fun settingTorchThreadsCount(){
+        val numThreads = 2
+        setNumThreads(numThreads)
+        assertEquals(numThreads, getNumThreads())
+    }
+    @Test
+    fun cudaAvailability(){
+        assertTrue(cudaAvailable())
+    }
+}

settings.gradle.kts

@@ -42,3 +42,7 @@ include(
     ":kmath-kotlingrad",
     ":examples"
 )
+
+if(System.getProperty("os.name") == "Linux"){
+    include(":kmath-torch")
+}