kmath-core/src/commonMain/kotlin/space/kscience/kmath/tensors/BufferedTensor.kt

@@ -1,5 +1,7 @@
 package space.kscience.kmath.tensors
 
+import space.kscience.kmath.linear.Matrix
+import space.kscience.kmath.nd.*
 import space.kscience.kmath.structures.*
 
 
@@ -29,59 +31,50 @@ public open class BufferedTensor<T>(
 
     override fun hashCode(): Int = 0
 
-}
-
-/*
-//todo make generator mb nextMatrixIndex?
-public class InnerMatrix<T>(private val tensor: BufferedTensor<T>){
-    private var offset: Int = 0
-    private val n : Int = tensor.shape.size
-    //stride?
-    private val step = tensor.shape[n - 1] * tensor.shape[n - 2]
-
-    public operator fun get(i: Int, j: Int): T {
-        val index = tensor.strides.index(offset)
-        index[n - 2] = i
-        index[n - 1] = j
-        return tensor[index]
-    }
-
-    public operator fun set(i: Int, j: Int, value: T): Unit {
-        val index = tensor.strides.index(offset)
-        index[n - 2] = i
-        index[n - 1] = j
-        tensor[index] = value
-    }
-
-    public fun makeStep(){
-        offset += step
+    // todo rename to vector
+    public inline fun forEachVector(vectorAction : (MutableStructure1D<T>) -> Unit): Unit {
+        check(shape.size >= 1) {"todo"}
+        val vectorOffset = strides.strides[0]
+        val vectorShape = intArrayOf(shape.last())
+        for (offset in 0 until numel step vectorOffset) {
+            val vector = BufferedTensor<T>(vectorShape, buffer, offset).as1D()
+            vectorAction(vector)
         }
     }
 
-public class InnerVector<T>(private val tensor: BufferedTensor<T>){
-    private var offset: Int = 0
-    private val n : Int = tensor.shape.size
-    //stride?
-    private val step = tensor.shape[n - 1]
+    public inline fun forEachMatrix(matrixAction : (MutableStructure2D<T>) -> Unit): Unit {
+        check(shape.size >= 2) {"todo"}
+        val matrixOffset = strides.strides[1]
+        val matrixShape = intArrayOf(shape[shape.size - 2], shape.last()) //todo better way?
+        for (offset in 0 until numel step matrixOffset) {
+            val matrix = BufferedTensor<T>(matrixShape, buffer, offset).as2D()
+            matrixAction(matrix)
+        }
+    }
+    // todo remove code copy-pasting
 
-    public operator fun get(i: Int): T {
-        val index = tensor.strides.index(offset)
-        index[n - 1] = i
-        return tensor[index]
+    public fun vectorSequence(): Sequence<MutableStructure1D<T>> = sequence {
+        check(shape.size >= 1) {"todo"}
+        val vectorOffset = strides.strides[0]
+        val vectorShape = intArrayOf(shape.last())
+        for (offset in 0 until numel step vectorOffset) {
+            val vector = BufferedTensor<T>(vectorShape, buffer, offset).as1D()
+            yield(vector)
+        }
     }
 
-    public operator fun set(i: Int, value: T): Unit {
-        val index = tensor.strides.index(offset)
-        index[n - 1] = i
-        tensor[index] = value
+    public fun matrixSequence(): Sequence<MutableStructure2D<T>> = sequence {
+        check(shape.size >= 2) {"todo"}
+        val matrixOffset = strides.strides[1]
+        val matrixShape = intArrayOf(shape[shape.size - 2], shape.last()) //todo better way?
+        for (offset in 0 until numel step matrixOffset) {
+            val matrix = BufferedTensor<T>(matrixShape, buffer, offset).as2D()
+            yield(matrix)
+        }
     }
 
-    public fun makeStep(){
-        offset += step
 }
-}
-//todo default buffer = arrayOf(0)???
- */
+
 
 public class IntTensor(
     shape: IntArray,

kmath-core/src/commonMain/kotlin/space/kscience/kmath/tensors/DoubleLinearOpsTensorAlgebra.kt

@@ -9,27 +9,21 @@ public class DoubleLinearOpsTensorAlgebra :
     }
 
     override fun DoubleTensor.lu(): Pair<DoubleTensor, IntTensor> {
-        /*
+
         // todo checks
+
         val luTensor = this.copy()
-        val lu = InnerMatrix(luTensor)
-        //stride TODO!!! move to generator?
-        var matCnt = 1
-        for (i in 0 until this.shape.size - 2) {
-            matCnt *= this.shape[i]
-        }
+
         val n = this.shape.size
-        val m = this.shape[n - 1]
-        val pivotsShape = IntArray(n - 1) { i ->
-            this.shape[i]
-        }
+        val m = this.shape.last()
+        val pivotsShape = IntArray(n - 1) { i -> this.shape[i] }
         val pivotsTensor = IntTensor(
             pivotsShape,
-            IntArray(matCnt * m) { 0 }
+            IntArray(pivotsShape.reduce(Int::times)) { 0 } //todo default???
         )
-        val pivot = InnerVector(pivotsTensor)
-        for (i in 0 until matCnt) {
-            for (row in 0 until m) pivot[row] = row
+
+        for ((lu, pivots) in luTensor.matrixSequence().zip(pivotsTensor.vectorSequence())){
+            for (row in 0 until m) pivots[row] = row
 
             for (i in 0 until m) {
                 var maxA = -1.0
@@ -47,9 +41,9 @@ public class DoubleLinearOpsTensorAlgebra :
 
                 if (iMax != i) {
 
-                    val j = pivot[i]
-                    pivot[i] = pivot[iMax]
-                    pivot[iMax] = j
+                    val j = pivots[i]
+                    pivots[i] = pivots[iMax]
+                    pivots[iMax] = j
 
                     for (k in 0 until m) {
                         val tmp = lu[i, k]
@@ -66,26 +60,28 @@ public class DoubleLinearOpsTensorAlgebra :
                     }
                 }
             }
-            lu.makeStep()
-            pivot.makeStep()
         }
 
-        return Pair(luTensor, pivotsTensor)*/
 
-        TODO("Andrei, use view, get, as2D, as1D")
+        return Pair(luTensor, pivotsTensor)
+
     }
 
-    override fun luPivot(lu: DoubleTensor, pivots: IntTensor): Triple<DoubleTensor, DoubleTensor, DoubleTensor> {
-        /*
+    override fun luPivot(luTensor: DoubleTensor, pivotsTensor: IntTensor): Triple<DoubleTensor, DoubleTensor, DoubleTensor> {
         //todo checks
-        val n = lu.shape[0]
-        val p = lu.zeroesLike()
-        pivots.buffer.unsafeToIntArray().forEachIndexed { i, pivot ->
-            p[i, pivot] = 1.0
+        val n = luTensor.shape.last()
+        val pTensor = luTensor.zeroesLike()
+        for ((p, pivot) in pTensor.matrixSequence().zip(pivotsTensor.vectorSequence())){
+            for (i in 0 until n){
+                p[i, pivot[i]] = 1.0
+            }
         }
-        val l = lu.zeroesLike()
-        val u = lu.zeroesLike()
 
+        val lTensor = luTensor.zeroesLike()
+        val uTensor = luTensor.zeroesLike()
+
+        for ((pairLU, lu) in lTensor.matrixSequence().zip(uTensor.matrixSequence()).zip(luTensor.matrixSequence())){
+            val (l, u) = pairLU
             for (i in 0 until n) {
                 for (j in 0 until n) {
                     if (i == j) {
@@ -99,9 +95,10 @@ public class DoubleLinearOpsTensorAlgebra :
                     }
                 }
             }
+        }
+
+        return Triple(pTensor, lTensor, uTensor)
 
-        return Triple(p, l, u)*/
-        TODO("Andrei, first we need implement get(Int)")
     }
 
     override fun DoubleTensor.cholesky(): DoubleTensor {

kmath-core/src/commonMain/kotlin/space/kscience/kmath/tensors/TensorAlgebra.kt

@@ -5,7 +5,7 @@ public interface TensorAlgebra<T, TensorType : TensorStructure<T>> {
 
 
     public fun zeros(shape: IntArray): TensorType
-    public fun TensorType.zeroesLike(): TensorType
+    public fun TensorType.zeroesLike(): TensorType // mb it shouldn't be tensor but algebra method (like in numpy/torch) ?
     public fun ones(shape: IntArray): TensorType
     public fun TensorType.onesLike(): TensorType