Repository: mahout Updated Branches: refs/heads/master 84e90ed23 -> 034790cce
http://git-wip-us.apache.org/repos/asf/mahout/blob/034790cc/viennacl/src/main/scala/org/apache/mahout/viennacl/opencl/javacpp/SrMatDnMatProdExpression.scala ---------------------------------------------------------------------- diff --git a/viennacl/src/main/scala/org/apache/mahout/viennacl/opencl/javacpp/SrMatDnMatProdExpression.scala b/viennacl/src/main/scala/org/apache/mahout/viennacl/opencl/javacpp/SrMatDnMatProdExpression.scala new file mode 100644 index 0000000..24d2c7b --- /dev/null +++ b/viennacl/src/main/scala/org/apache/mahout/viennacl/opencl/javacpp/SrMatDnMatProdExpression.scala @@ -0,0 +1,33 @@ +/** + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +package org.apache.mahout.viennacl.opencl.javacpp + +import org.bytedeco.javacpp.Pointer +import org.bytedeco.javacpp.annotation.{Name, Namespace, Platform, Properties} + + +@Properties(inherit = Array(classOf[Context]), + value = Array(new Platform( + library = "jniViennaCL") + )) +@Namespace("viennacl") +@Name(Array("matrix_expression<const viennacl::compressed_matrix<double>, " + + "const viennacl::matrix_base<double>, " + + "viennacl::op_prod>")) +class SrMatDnMatProdExpression extends Pointer { + +} http://git-wip-us.apache.org/repos/asf/mahout/blob/034790cc/viennacl/src/main/scala/org/apache/mahout/viennacl/opencl/javacpp/VCLVector.scala ---------------------------------------------------------------------- diff --git a/viennacl/src/main/scala/org/apache/mahout/viennacl/opencl/javacpp/VCLVector.scala b/viennacl/src/main/scala/org/apache/mahout/viennacl/opencl/javacpp/VCLVector.scala new file mode 100644 index 0000000..f0e3010 --- /dev/null +++ b/viennacl/src/main/scala/org/apache/mahout/viennacl/opencl/javacpp/VCLVector.scala @@ -0,0 +1,133 @@ +/** + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +package org.apache.mahout.viennacl.opencl.javacpp + +import org.bytedeco.javacpp._ +import org.bytedeco.javacpp.annotation._ + +import scala.collection.mutable.ArrayBuffer + + +@Properties(inherit = Array(classOf[Context]), + value = Array(new Platform( + library="jniViennaCL" + ))) +@Name(Array("viennacl::vector<double>")) +final class VCLVector(defaultCtr: Boolean = true) extends VectorBase { + + if (defaultCtr) allocate() + + def this(){ + this(false) + allocate() + } + + def this(i: Int){ + this(false) + allocate(i) + } + + def this(size: Int, ctx: Context = new Context(Context.MAIN_MEMORY)) { + this(false) + allocate(size, ctx) + } + + def this(@Const @ByRef ve: VecMultExpression) { + this(false) + allocate(ve) + } + + def this(@Const @ByRef vmp: MatVecProdExpression) { + this(false) + allocate(vmp) + } + +// conflicting with the next signature as MemHandle is a pointer and so is a DoublePointer.. +// leave out for now. +// +// def this(h: MemHandle , vec_size: Int, vec_start: Int = 0, vec_stride: Int = 1) { +// this(false) +// allocate(h, vec_size, vec_start, vec_stride) +// } + + def this(ptr_to_mem: DoublePointer, + @Cast(Array("viennacl::memory_types"))mem_type : Int, + vec_size: Int, + start: Int = 0, + stride: Int = 1) { + + this(false) + allocate(ptr_to_mem, mem_type, vec_size, start, stride) + ptrs += ptr_to_mem + } + + def this(@Const @ByRef vc: VCLVector) { + this(false) + allocate(vc) + } + def this(@Const @ByRef vb: VectorBase) { + this(false) + allocate(vb) + } + + @native protected def allocate() + + @native protected def allocate(size: Int) + + @native protected def allocate(size: Int, @ByVal ctx: Context) + + @native protected def allocate(@Const @ByRef ve: VecMultExpression) + + @native protected def allocate(@Const @ByRef ve: MatVecProdExpression) + + @native protected def allocate(@Const @ByRef vb: VCLVector) + + @native protected def allocate(@Const @ByRef vb: VectorBase) + + +// @native protected def allocate(h: MemHandle , vec_size: Int, +// vec_start: Int, +// vec_stride: Int) + + @native protected def allocate(ptr_to_mem: DoublePointer, + @Cast(Array("viennacl::memory_types"))mem_type : Int, + vec_size: Int, + start: Int, + stride: Int) + + @Name(Array("viennacl::vector<double>::self_type")) + def selfType:VectorBase = this.asInstanceOf[VectorBase] + + + @native def switch_memory_context(@ByVal context: Context): Unit + +// Swaps the handles of two vectors by swapping the OpenCL handles only, no data copy. +// @native def fast_swap(@ByVal other: VCLVector): VectorBase + +// add this operator in for tests many more can be added +// @Name(Array("operator*")) +// @native @ByPtr def *(i: Int): VectorMultExpression + + + +} + +object VCLVector { + Context.loadLib() +} + + http://git-wip-us.apache.org/repos/asf/mahout/blob/034790cc/viennacl/src/main/scala/org/apache/mahout/viennacl/opencl/javacpp/VecMultExpression.scala ---------------------------------------------------------------------- diff --git a/viennacl/src/main/scala/org/apache/mahout/viennacl/opencl/javacpp/VecMultExpression.scala b/viennacl/src/main/scala/org/apache/mahout/viennacl/opencl/javacpp/VecMultExpression.scala new file mode 100644 index 0000000..1904151 --- /dev/null +++ b/viennacl/src/main/scala/org/apache/mahout/viennacl/opencl/javacpp/VecMultExpression.scala @@ -0,0 +1,32 @@ +/** + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +package org.apache.mahout.viennacl.opencl.javacpp + +import org.bytedeco.javacpp.Pointer +import org.bytedeco.javacpp.annotation.{Name, Namespace, Platform, Properties} + + +@Properties(inherit = Array(classOf[Context]), + value = Array(new Platform( + library = "jniViennaCL") + )) +@Namespace("viennacl") +@Name(Array("vector_expression<const viennacl::vector_base<double>," + + "const double, viennacl::op_mult >")) +class VecMultExpression extends Pointer { + +} http://git-wip-us.apache.org/repos/asf/mahout/blob/034790cc/viennacl/src/main/scala/org/apache/mahout/viennacl/opencl/javacpp/VectorBase.scala ---------------------------------------------------------------------- diff --git a/viennacl/src/main/scala/org/apache/mahout/viennacl/opencl/javacpp/VectorBase.scala b/viennacl/src/main/scala/org/apache/mahout/viennacl/opencl/javacpp/VectorBase.scala new file mode 100644 index 0000000..43ae39d --- /dev/null +++ b/viennacl/src/main/scala/org/apache/mahout/viennacl/opencl/javacpp/VectorBase.scala @@ -0,0 +1,57 @@ +/** + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +package org.apache.mahout.viennacl.opencl.javacpp + +import java.nio._ + +import org.bytedeco.javacpp._ +import org.bytedeco.javacpp.annotation._ + +import scala.collection.mutable.ArrayBuffer + + +@Properties(inherit = Array(classOf[Context]), + value = Array(new Platform( + library="jniViennaCL" + ))) +@Name(Array("viennacl::vector_base<double>")) +class VectorBase extends Pointer { + + protected val ptrs = new ArrayBuffer[Pointer]() + + override def deallocate(deallocate: Boolean): Unit = { + super.deallocate(deallocate) + ptrs.foreach(_.close()) + } + + // size of the vec elements + @native @Const def size(): Int + + // size of the vec elements + padding + @native @Const def internal_size(): Int + + // handle to the vec element buffer + @native @Const @ByRef def handle: MemHandle + +// // add this operator in for tests many more can be added +// @Name(Array("operator* ")) +// @native def *(i: Int): VectorMultExpression + + +} + + http://git-wip-us.apache.org/repos/asf/mahout/blob/034790cc/viennacl/src/main/scala/org/apache/mahout/viennacl/opencl/package.scala ---------------------------------------------------------------------- diff --git a/viennacl/src/main/scala/org/apache/mahout/viennacl/opencl/package.scala b/viennacl/src/main/scala/org/apache/mahout/viennacl/opencl/package.scala new file mode 100644 index 0000000..8c3743a --- /dev/null +++ b/viennacl/src/main/scala/org/apache/mahout/viennacl/opencl/package.scala @@ -0,0 +1,434 @@ +package org.apache.mahout.viennacl + +import java.nio._ + +import org.apache.mahout.math._ +import scalabindings._ +import RLikeOps._ +import org.apache.mahout.math.backend.incore._ +import scala.collection.JavaConversions._ +import org.apache.mahout.viennacl.opencl.javacpp.{CompressedMatrix, Context, DenseRowMatrix, Functions, VCLVector} +import org.apache.mahout.viennacl.opencl.javacpp.Context +import org.bytedeco.javacpp.{DoublePointer, IntPointer} + + + +package object opencl { + + type IntConvertor = Int => Int + + def toVclDenseRM(src: Matrix, vclCtx: Context = new Context(Context.MAIN_MEMORY)): DenseRowMatrix = { + vclCtx.memoryType match { + case Context.MAIN_MEMORY â + val vclMx = new DenseRowMatrix( + data = repackRowMajor(src, src.nrow, src.ncol), + nrow = src.nrow, + ncol = src.ncol, + ctx = vclCtx + ) + vclMx + case _ â + val vclMx = new DenseRowMatrix(src.nrow, src.ncol, vclCtx) + fastCopy(src, vclMx) + vclMx + } + } + + + /** + * Convert a dense row VCL matrix to mahout matrix. + * + * @param src + * @return + */ + def fromVclDenseRM(src: DenseRowMatrix): Matrix = { + val nrowIntern = src.internalnrow + val ncolIntern = src.internalncol + + // A technical debt here: + + // We do double copying here, this is obviously suboptimal, but hopefully we'll compensate + // this with gains from running superlinear algorithms in VCL. + val dbuff = new DoublePointer(nrowIntern * ncolIntern) + Functions.fastCopy(src, dbuff) + var srcOffset = 0 + val ncol = src.ncol + val rows = for (irow â 0 until src.nrow) yield { + + val rowvec = new Array[Double](ncol) + dbuff.position(srcOffset).get(rowvec) + + srcOffset += ncolIntern + rowvec + } + + // Always! use shallow = true to avoid yet another copying. + new DenseMatrix(rows.toArray, true) + } + + def fastCopy(mxSrc: Matrix, dst: DenseRowMatrix) = { + val nrowIntern = dst.internalnrow + val ncolIntern = dst.internalncol + + assert(nrowIntern >= mxSrc.nrow && ncolIntern >= mxSrc.ncol) + + val rmajorData = repackRowMajor(mxSrc, nrowIntern, ncolIntern) + Functions.fastCopy(rmajorData, new DoublePointer(rmajorData).position(rmajorData.limit()), dst) + + rmajorData.close() + } + + private def repackRowMajor(mx: Matrix, nrowIntern: Int, ncolIntern: Int): DoublePointer = { + + assert(mx.nrow <= nrowIntern && mx.ncol <= ncolIntern) + + val dbuff = new DoublePointer(nrowIntern * ncolIntern) + + mx match { + case dm: DenseMatrix â + val valuesF = classOf[DenseMatrix].getDeclaredField("values") + valuesF.setAccessible(true) + val values = valuesF.get(dm).asInstanceOf[Array[Array[Double]]] + var dstOffset = 0 + for (irow â 0 until mx.nrow) { + val rowarr = values(irow) + dbuff.position(dstOffset).put(rowarr, 0, rowarr.size min ncolIntern) + dstOffset += ncolIntern + } + dbuff.position(0) + case _ â + // Naive copying. Could be sped up for a DenseMatrix. TODO. + for (row â mx) { + val dstOffset = row.index * ncolIntern + for (el â row.nonZeroes) dbuff.put(dstOffset + el.index, el) + } + } + + dbuff + } + + /** + * + * @param mxSrc + * @param ctx + * @return + */ + def toVclCmpMatrixAlt(mxSrc: Matrix, ctx: Context): CompressedMatrix = { + + // use repackCSR(matrix, ctx) to convert all ints to unsigned ints if Context is Ocl + // val (jumpers, colIdcs, els) = repackCSRAlt(mxSrc) + val (jumpers, colIdcs, els) = repackCSR(mxSrc, ctx) + + val compMx = new CompressedMatrix(mxSrc.nrow, mxSrc.ncol, els.capacity().toInt, ctx) + compMx.set(jumpers, colIdcs, els, mxSrc.nrow, mxSrc.ncol, els.capacity().toInt) + compMx + } + + private def repackCSRAlt(mx: Matrix): (IntPointer, IntPointer, DoublePointer) = { + val nzCnt = mx.map(_.getNumNonZeroElements).sum + val jumpers = new IntPointer(mx.nrow + 1L) + val colIdcs = new IntPointer(nzCnt + 0L) + val els = new DoublePointer(nzCnt) + var posIdx = 0 + + var sortCols = false + + // Row-wise loop. Rows may not necessarily come in order. But we have to have them in-order. + for (irow â 0 until mx.nrow) { + + val row = mx(irow, ::) + jumpers.put(irow.toLong, posIdx) + + // Remember row start index in case we need to restart conversion of this row if out-of-order + // column index is detected + val posIdxStart = posIdx + + // Retry loop: normally we are done in one pass thru it unless we need to re-run it because + // out-of-order column was detected. + var done = false + while (!done) { + + // Is the sorting mode on? + if (sortCols) { + + // Sorting of column indices is on. So do it. + row.nonZeroes() + // Need to convert to a strict collection out of iterator + .map(el â el.index â el.get) + // Sorting requires Sequence api + .toSeq + // Sort by column index + .sortBy(_._1) + // Flush to the CSR buffers. + .foreach { case (index, v) â + colIdcs.put(posIdx.toLong, index) + els.put(posIdx.toLong, v) + posIdx += 1 + } + + // Never need to retry if we are already in the sorting mode. + done = true + + } else { + + // Try to run unsorted conversion here, switch lazily to sorted if out-of-order column is + // detected. + var lastCol = 0 + val nzIter = row.nonZeroes().iterator() + var abortNonSorted = false + + while (nzIter.hasNext && !abortNonSorted) { + + val el = nzIter.next() + val index = el.index + + if (index < lastCol) { + + // Out of order detected: abort inner loop, reset posIdx and retry with sorting on. + abortNonSorted = true + sortCols = true + posIdx = posIdxStart + + } else { + + // Still in-order: save element and column, continue. + els.put(posIdx, el) + colIdcs.put(posIdx.toLong, index) + posIdx += 1 + + // Remember last column seen. + lastCol = index + } + } // inner non-sorted + + // Do we need to re-run this row with sorting? + done = !abortNonSorted + + } // if (sortCols) + + } // while (!done) retry loop + + } // row-wise loop + + // Make sure Mahout matrix did not cheat on non-zero estimate. + assert(posIdx == nzCnt) + + jumpers.put(mx.nrow.toLong, nzCnt) + + (jumpers, colIdcs, els) + } + + // same as repackCSRAlt except converts to jumpers, colIdcs to unsigned ints before setting + private def repackCSR(mx: Matrix, context: Context): (IntPointer, IntPointer, DoublePointer) = { + val nzCnt = mx.map(_.getNumNonZeroElements).sum + val jumpers = new IntPointer(mx.nrow + 1L) + val colIdcs = new IntPointer(nzCnt + 0L) + val els = new DoublePointer(nzCnt) + var posIdx = 0 + + var sortCols = false + + def convertInt: IntConvertor = if(context.memoryType == Context.OPENCL_MEMORY) { + int2cl_uint + } else { + i: Int => i: Int + } + + // Row-wise loop. Rows may not necessarily come in order. But we have to have them in-order. + for (irow â 0 until mx.nrow) { + + val row = mx(irow, ::) + jumpers.put(irow.toLong, posIdx) + + // Remember row start index in case we need to restart conversion of this row if out-of-order + // column index is detected + val posIdxStart = posIdx + + // Retry loop: normally we are done in one pass thru it unless we need to re-run it because + // out-of-order column was detected. + var done = false + while (!done) { + + // Is the sorting mode on? + if (sortCols) { + + // Sorting of column indices is on. So do it. + row.nonZeroes() + // Need to convert to a strict collection out of iterator + .map(el â el.index â el.get) + // Sorting requires Sequence api + .toIndexedSeq + // Sort by column index + .sortBy(_._1) + // Flush to the CSR buffers. + .foreach { case (index, v) â + // convert to cl_uint if context is OCL + colIdcs.put(posIdx.toLong, convertInt(index)) + els.put(posIdx.toLong, v) + posIdx += 1 + } + + // Never need to retry if we are already in the sorting mode. + done = true + + } else { + + // Try to run unsorted conversion here, switch lazily to sorted if out-of-order column is + // detected. + var lastCol = 0 + val nzIter = row.nonZeroes().iterator() + var abortNonSorted = false + + while (nzIter.hasNext && !abortNonSorted) { + + val el = nzIter.next() + val index = el.index + + if (index < lastCol) { + + // Out of order detected: abort inner loop, reset posIdx and retry with sorting on. + abortNonSorted = true + sortCols = true + posIdx = posIdxStart + + } else { + + // Still in-order: save element and column, continue. + els.put(posIdx, el) + // convert to cl_uint if context is OCL + colIdcs.put(posIdx.toLong, convertInt(index)) + posIdx += 1 + + // Remember last column seen. + lastCol = index + } + } // inner non-sorted + + // Do we need to re-run this row with sorting? + done = !abortNonSorted + + } // if (sortCols) + + } // while (!done) retry loop + + } // row-wise loop + + // Make sure Mahout matrix did not cheat on non-zero estimate. + assert(posIdx == nzCnt) + + // convert to cl_uint if context is OCL + jumpers.put(mx.nrow.toLong, convertInt(nzCnt)) + + (jumpers, colIdcs, els) + } + + + + def fromVclCompressedMatrix(src: CompressedMatrix): Matrix = { + val m = src.size1 + val n = src.size2 + val NNz = src.nnz + + val row_ptr_handle = src.handle1 + val col_idx_handle = src.handle2 + val element_handle = src.handle + + val row_ptr = new IntPointer((m + 1).toLong) + val col_idx = new IntPointer(NNz.toLong) + val values = new DoublePointer(NNz.toLong) + + Functions.memoryReadInt(row_ptr_handle, 0, (m + 1) * 4, row_ptr, false) + Functions.memoryReadInt(col_idx_handle, 0, NNz * 4, col_idx, false) + Functions.memoryReadDouble(element_handle, 0, NNz * 8, values, false) + + val rowPtr = row_ptr.asBuffer() + val colIdx = col_idx.asBuffer() + val vals = values.asBuffer() + + rowPtr.rewind() + colIdx.rewind() + vals.rewind() + + + val srMx = new SparseRowMatrix(m, n) + + // read the values back into the matrix + var j = 0 + // row wise, copy any non-zero elements from row(i-1,::) + for (i <- 1 to m) { + // for each nonzero element, set column col(idx(j) value to vals(j) + while (j < rowPtr.get(i)) { + srMx(i - 1, colIdx.get(j)) = vals.get(j) + j += 1 + } + } + srMx + } + + def toVclVec(vec: Vector, ctx: Context): VCLVector = { + + vec match { + case vec: DenseVector => { + val valuesF = classOf[DenseVector].getDeclaredField("values") + valuesF.setAccessible(true) + val values = valuesF.get(vec).asInstanceOf[Array[Double]] + val el_ptr = new DoublePointer(values.length.toLong) + el_ptr.put(values, 0, values.length) + + new VCLVector(el_ptr, ctx.memoryType, values.length) + } + + case vec: SequentialAccessSparseVector => { + val it = vec.iterateNonZero + val size = vec.size() + val el_ptr = new DoublePointer(size.toLong) + while (it.hasNext) { + val el: Vector.Element = it.next + el_ptr.put(el.index, el.get()) + } + new VCLVector(el_ptr, ctx.memoryType, size) + } + + case vec: RandomAccessSparseVector => { + val it = vec.iterateNonZero + val size = vec.size() + val el_ptr = new DoublePointer(size.toLong) + while (it.hasNext) { + val el: Vector.Element = it.next + el_ptr.put(el.index, el.get()) + } + new VCLVector(el_ptr, ctx.memoryType, size) + } + case _ => throw new IllegalArgumentException("Vector sub-type not supported.") + } + + } + + def fromVClVec(vclVec: VCLVector): Vector = { + val size = vclVec.size + val element_handle = vclVec.handle + val ele_ptr = new DoublePointer(size) + Functions.memoryReadDouble(element_handle, 0, size * 8, ele_ptr, false) + + // for now just assume its dense since we only have one flavor of + // VCLVector + val mVec = new DenseVector(size) + for (i <- 0 until size) { + mVec.setQuick(i, ele_ptr.get(i + 0L)) + } + + mVec + } + + + // TODO: Fix this? cl_uint must be an unsigned int per each machine's representation of such. + // this is currently not working anyways. + // cl_uint is needed for OpenCl sparse Buffers + // per https://www.khronos.org/registry/cl/sdk/1.1/docs/man/xhtml/scalarDataTypes.html + // it is simply an unsigned int, so strip the sign. + def int2cl_uint(i: Int): Int = { + ((i >>> 1) << 1) + (i & 1) + } + + +} http://git-wip-us.apache.org/repos/asf/mahout/blob/034790cc/viennacl/src/test/scala/org/apache/mahout/viennacl/opencl/ViennaCLSuiteVCL.scala ---------------------------------------------------------------------- diff --git a/viennacl/src/test/scala/org/apache/mahout/viennacl/opencl/ViennaCLSuiteVCL.scala b/viennacl/src/test/scala/org/apache/mahout/viennacl/opencl/ViennaCLSuiteVCL.scala new file mode 100644 index 0000000..c433534 --- /dev/null +++ b/viennacl/src/test/scala/org/apache/mahout/viennacl/opencl/ViennaCLSuiteVCL.scala @@ -0,0 +1,427 @@ +/** + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +package org.apache.mahout.opencl.viennacl + +import org.apache.mahout.math._ +import org.apache.mahout.math.scalabindings.RLikeOps._ +import org.apache.mahout.viennacl.opencl.javacpp.CompressedMatrix +import org.apache.mahout.viennacl.opencl._ +import org.apache.mahout.viennacl.opencl.javacpp.Functions._ +import org.apache.mahout.viennacl.opencl.javacpp.LinalgFunctions._ +import org.apache.mahout.viennacl.opencl.javacpp.{Context, LinalgFunctions, VCLVector, _} +import org.bytedeco.javacpp.DoublePointer +import org.scalatest.{FunSuite, Matchers} + +import scala.util.Random + +class ViennaCLSuiteVCL extends FunSuite with Matchers { + + test("row-major viennacl::matrix") { + + // Just to make sure the javacpp library is loaded: + Context.loadLib() + + val m = 20 + val n = 30 + val data = new DoublePointer(m * n) + val buff = data.asBuffer() + // Fill with some noise + while (buff.remaining() > 0) buff.put(Random.nextDouble()) + + // Create row-major matrix with OpenCL + val openClCtx = new Context(Context.OPENCL_MEMORY) + val hostClCtx = new Context(Context.MAIN_MEMORY) + val oclMx = new DenseRowMatrix(m, n, openClCtx) + val cpuMx = new DenseRowMatrix(data = data, nrow = m, ncol = n, hostClCtx) + + oclMx.memoryDomain shouldBe Context.OPENCL_MEMORY + + // Apparently, this doesn't really switch any contexts? any how, uncommenting this causes + // subsequent out-of-resources OCL error for me in other tests. Perhaps we shouldnt' really + // do cross-memory-domain assigns? + + // oclMx := cpuMx + + // Did it change memory domain? that may explain the OCL resource leak. + info(s"OCL matrix memory domain after assgn=${oclMx.memoryDomain}") + oclMx.memoryDomain shouldBe Context.OPENCL_MEMORY + + + // And free. + cpuMx.close() + oclMx.close() + + } + + test("dense vcl mmul with fast_copy") { + + import LinalgFunctions._ + + val vclCtx = new Context(Context.OPENCL_MEMORY) + + val m = 20 + val n = 30 + val s = 40 + + val r = new Random(1234) + + // Dense row-wise + val mxA = new DenseMatrix(m, s) + val mxB = new DenseMatrix(s, n) + + // add some data + mxA := { (_, _, _) => r.nextDouble() } + mxB := { (_, _, _) => r.nextDouble() } + + // time Mahout MMul + // mxC = mxA %*% mxB via Mahout MMul + val mxCControl = mxA %*% mxB + + val vclA = toVclDenseRM(mxA, vclCtx) + val vclB = toVclDenseRM(mxB, vclCtx) + + val vclC = new DenseRowMatrix(prod(vclA, vclB)) + + val mxC = fromVclDenseRM(vclC) + + vclA.close() + vclB.close() + vclC.close() + + // So did we compute it correctly? + (mxC - mxA %*% mxB).norm / m / n should be < 1e-16 + + vclCtx.deallocate() + vclCtx.close() + + } + + test("mmul microbenchmark") { + val oclCtx = new Context(Context.OPENCL_MEMORY) + val memCtx = new Context(Context.MAIN_MEMORY) + + val m = 3000 + val n = 3000 + val s = 1000 + + val r = new Random(1234) + + // Dense row-wise + val mxA = new DenseMatrix(m, s) + val mxB = new DenseMatrix(s, n) + + // add some data + mxA := { (_, _, _) => r.nextDouble() } + mxB := { (_, _, _) => r.nextDouble() } + + var ms = System.currentTimeMillis() + mxA %*% mxB + ms = System.currentTimeMillis() - ms + info(s"Mahout multiplication time: $ms ms.") + + import LinalgFunctions._ + + // openCL time, including copying: + { + ms = System.currentTimeMillis() + val oclA = toVclDenseRM(mxA, oclCtx) + val oclB = toVclDenseRM(mxB, oclCtx) + val oclC = new DenseRowMatrix(prod(oclA, oclB)) + val mxC = fromVclDenseRM(oclC) + ms = System.currentTimeMillis() - ms + info(s"ViennaCL/OpenCL multiplication time: $ms ms.") + + oclA.close() + oclB.close() + oclC.close() + } + + // openMP/cpu time, including copying: + { + ms = System.currentTimeMillis() + val ompA = toVclDenseRM(mxA, memCtx) + val ompB = toVclDenseRM(mxB, memCtx) + val ompC = new DenseRowMatrix(prod(ompA, ompB)) + val mxC = fromVclDenseRM(ompC) + ms = System.currentTimeMillis() - ms + info(s"ViennaCL/cpu/OpenMP multiplication time: $ms ms.") + + ompA.close() + ompB.close() + ompC.close() + } + oclCtx.deallocate() + oclCtx.close() + + + } + + test("trans") { + + val oclCtx = new Context(Context.OPENCL_MEMORY) + val ompCtx = new Context(Context.MAIN_MEMORY) + + + val m = 20 + val n = 30 + + val r = new Random(1234) + + // Dense row-wise + val mxA = new DenseMatrix(m, n) + + // add some data + mxA := { (_, _, _) => r.nextDouble() } + + // Test transposition in OpenCL + { + val oclA = toVclDenseRM(src = mxA, oclCtx) + val oclAt = new DenseRowMatrix(trans(oclA)) + + val mxAt = fromVclDenseRM(oclAt) + oclA.close() + oclAt.close() + + (mxAt - mxA.t).norm / m / n should be < 1e-16 + } + + // Test transposition in OpenMP + { + val ompA = toVclDenseRM(src = mxA, ompCtx) + val ompAt = new DenseRowMatrix(trans(ompA)) + + val mxAt = fromVclDenseRM(ompAt) + ompA.close() + ompAt.close() + + (mxAt - mxA.t).norm / m / n should be < 1e-16 + } + oclCtx.deallocate() + oclCtx.close() + + + } + + test("sparse mmul microbenchmark") { + + val oclCtx = new Context(Context.OPENCL_MEMORY) + val ompCtx = new Context(Context.MAIN_MEMORY) + + + val m = 3000 + val n = 3000 + val s = 1000 + + val r = new Random(1234) + + // sparse row-wise + val mxA = new SparseRowMatrix(m, s, false) + val mxB = new SparseRowMatrix(s, n, true) + + // add some sparse data with a 20% threshold + mxA := { (_, _, v) => if (r.nextDouble() < .20) r.nextDouble() else v } + mxB := { (_, _, v) => if (r.nextDouble() < .20) r.nextDouble() else v } + + var ms = System.currentTimeMillis() + val mxC = mxA %*% mxB + ms = System.currentTimeMillis() - ms + info(s"Mahout Sparse multiplication time: $ms ms.") + +// Test multiplication in OpenCL + { + + ms = System.currentTimeMillis() + val oclA = toVclCmpMatrixAlt(mxA, oclCtx) + val oclB = toVclCmpMatrixAlt(mxB, oclCtx) + + val oclC = new CompressedMatrix(prod(oclA, oclB)) + ms = System.currentTimeMillis() - ms + info(s"ViennaCL/OpenCL Sparse multiplication time: $ms ms.") + + val oclMxC = fromVclCompressedMatrix(oclC) + val ompMxC = fromVclCompressedMatrix(oclC) + (mxC - oclMxC).norm / mxC.nrow / mxC.ncol should be < 1e-16 + + oclA.close() + oclB.close() + oclC.close() + } + + // Test multiplication in OpenMP + { + ms = System.currentTimeMillis() + // val ompA = toVclCompressedMatrix(src = mxA, ompCtx) + // val ompB = toVclCompressedMatrix(src = mxB, ompCtx) + + val ompA = toVclCmpMatrixAlt(mxA, ompCtx) + val ompB = toVclCmpMatrixAlt(mxB, ompCtx) + + val ompC = new CompressedMatrix(prod(ompA, ompB)) + + ms = System.currentTimeMillis() - ms + info(s"ViennaCL/cpu/OpenMP Sparse multiplication time: $ms ms.") + + val ompMxC = fromVclCompressedMatrix(ompC) + (mxC - ompMxC).norm / mxC.nrow / mxC.ncol should be < 1e-16 + + ompA.close() + ompB.close() + ompC.close() + + } + oclCtx.deallocate() + oclCtx.close() + + } + + test("VCL Dense Matrix %*% Dense vector") { + + val oclCtx = new Context(Context.OPENCL_MEMORY) + val ompCtx = new Context(Context.MAIN_MEMORY) + + + val m = 30 + val s = 10 + + val r = new Random(1234) + + // Dense row-wise + val mxA = new DenseMatrix(m, s) + val dvecB = new DenseVector(s) + + // add some random data + mxA := { (_,_,_) => r.nextDouble() } + dvecB := { (_,_) => r.nextDouble() } + + //test in matrix %*% vec + var ms = System.currentTimeMillis() + val mDvecC = mxA %*% dvecB + ms = System.currentTimeMillis() - ms + info(s"Mahout dense matrix %*% dense vector multiplication time: $ms ms.") + + + /* TODO: CL_OUT_OF_RESOURCES error thrown when trying to read data out of OpenCl GPU Vectors */ + //Test multiplication in OpenCL +// { +// +// ms = System.currentTimeMillis() +// val oclA = toVclDenseRM(mxA, oclCtx) +// val oclVecB = toVclVec(dvecB, oclCtx) +// +// val oclVecC = new VCLVector(prod(oclA, oclVecB)) +// val oclDvecC = fromVClVec(oclVecC) +//// +//// ms = System.currentTimeMillis() - ms +//// info(s"ViennaCL/OpenCL dense matrix %*% dense vector multiplication time: $ms ms.") +//// (oclDvecC.toColMatrix - mDvecC.toColMatrix).norm / s should be < 1e-16 +// +// oclA.close() +// oclVecB.close() +// oclVecC.close() +// } + + //Test multiplication in OpenMP + { + + ms = System.currentTimeMillis() + val ompMxA = toVclDenseRM(mxA, ompCtx) + val ompVecB = toVclVec(dvecB, ompCtx) + + val ompVecC = new VCLVector(prod(ompMxA, ompVecB)) + val ompDvecC = fromVClVec(ompVecC) + + ms = System.currentTimeMillis() - ms + info(s"ViennaCL/cpu/OpenMP dense matrix %*% dense vector multiplication time: $ms ms.") + (ompDvecC.toColMatrix - mDvecC.toColMatrix).norm / s should be < 1e-16 + + ompMxA.close() + ompVecB.close() + ompVecC.close() + } + + oclCtx.deallocate() + oclCtx.close() + + + } + + + test("Sparse %*% Dense mmul microbenchmark") { + val oclCtx = new Context(Context.OPENCL_MEMORY) + val memCtx = new Context(Context.MAIN_MEMORY) + + val m = 3000 + val n = 3000 + val s = 1000 + + val r = new Random(1234) + + // Dense row-wise + val mxSr = new SparseMatrix(m, s) + val mxDn = new DenseMatrix(s, n) + + // add some data + mxSr := { (_, _, v) => if (r.nextDouble() < .20) r.nextDouble() else v } + mxDn := { (_, _, _) => r.nextDouble() } + + var ms = System.currentTimeMillis() + mxSr %*% mxDn + ms = System.currentTimeMillis() - ms + info(s"Mahout multiplication time: $ms ms.") + + import LinalgFunctions._ + + // For now, since our dense matrix is fully dense lets just assume that our result is dense. + // openCL time, including copying: + { + ms = System.currentTimeMillis() + val oclA = toVclCmpMatrixAlt(mxSr, oclCtx) + val oclB = toVclDenseRM(mxDn, oclCtx) + val oclC = new DenseRowMatrix(prod(oclA, oclB)) + val mxC = fromVclDenseRM(oclC) + ms = System.currentTimeMillis() - ms + info(s"ViennaCL/OpenCL multiplication time: $ms ms.") + + oclA.close() + oclB.close() + oclC.close() + } + + // openMP/cpu time, including copying: + { + ms = System.currentTimeMillis() + val ompA = toVclCmpMatrixAlt(mxSr, memCtx) + val ompB = toVclDenseRM(mxDn, memCtx) + val ompC = new DenseRowMatrix(prod(ompA, ompB)) + val mxC = fromVclDenseRM(ompC) + ms = System.currentTimeMillis() - ms + info(s"ViennaCL/cpu/OpenMP multiplication time: $ms ms.") + + ompA.close() + ompB.close() + ompC.close() + } + + oclCtx.deallocate() + oclCtx.close() + + + } + + + +}
