summaryrefslogtreecommitdiff
path: root/libs/ode-0.16.1/ode/src/fastlsolve_impl.h
diff options
context:
space:
mode:
Diffstat (limited to 'libs/ode-0.16.1/ode/src/fastlsolve_impl.h')
-rw-r--r--libs/ode-0.16.1/ode/src/fastlsolve_impl.h1610
1 files changed, 1610 insertions, 0 deletions
diff --git a/libs/ode-0.16.1/ode/src/fastlsolve_impl.h b/libs/ode-0.16.1/ode/src/fastlsolve_impl.h
new file mode 100644
index 0000000..f14ada7
--- /dev/null
+++ b/libs/ode-0.16.1/ode/src/fastlsolve_impl.h
@@ -0,0 +1,1610 @@
+
+
+/*************************************************************************
+ * *
+ * Open Dynamics Engine, Copyright (C) 2001,2002 Russell L. Smith. *
+ * All rights reserved. Email: russ@q12.org Web: www.q12.org *
+ * *
+ * This library is free software; you can redistribute it and/or *
+ * modify it under the terms of EITHER: *
+ * (1) The GNU Lesser General Public License as published by the Free *
+ * Software Foundation; either version 2.1 of the License, or (at *
+ * your option) any later version. The text of the GNU Lesser *
+ * General Public License is included with this library in the *
+ * file LICENSE.TXT. *
+ * (2) The BSD-style license that is included with this library in *
+ * the file LICENSE-BSD.TXT. *
+ * *
+ * This library is distributed in the hope that it will be useful, *
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of *
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the files *
+ * LICENSE.TXT and LICENSE-BSD.TXT for more details. *
+ * *
+ *************************************************************************/
+
+/*
+ * Code style improvements and optimizations by Oleh Derevenko ????-2019
+ * L1Straight cooperative solving code of ThreadedEquationSolverLDLT copyright (c) 2017-2019 Oleh Derevenko, odar@eleks.com (change all "a" to "e")
+ */
+
+#ifndef _ODE_FASTLSOLVE_IMPL_H_
+#define _ODE_FASTLSOLVE_IMPL_H_
+
+
+/* solve L*X=B, with B containing 1 right hand sides.
+ * L is an n*n lower triangular matrix with ones on the diagonal.
+ * L is stored by rows and its leading dimension is lskip.
+ * B is an n*1 matrix that contains the right hand sides.
+ * B is stored by columns and its leading dimension is also lskip.
+ * B is overwritten with X.
+ * this processes blocks of 4*4.
+ * if this is in the factorizer source file, n must be a multiple of 4.
+ */
+
+template<unsigned int b_stride>
+void solveL1Straight (const dReal *L, dReal *B, unsigned rowCount, unsigned rowSkip)
+{
+ dIASSERT(rowCount != 0);
+
+ /* compute all 4 x 1 blocks of X */
+ unsigned blockStartRow = 0;
+ bool subsequentPass = false;
+ bool goForLoopX4 = rowCount >= 4;
+ const unsigned loopX4LastRow = goForLoopX4 ? rowCount - 4 : 0;
+ for (; goForLoopX4; subsequentPass = true, goForLoopX4 = (blockStartRow += 4) <= loopX4LastRow)
+ {
+ /* declare variables - Z matrix, p and q vectors, etc */
+ const dReal *ptrLElement;
+ dReal *ptrBElement;
+
+ dReal Z11, Z21, Z31, Z41;
+
+ /* compute all 4 x 1 block of X, from rows i..i+4-1 */
+ if (subsequentPass)
+ {
+ ptrLElement = L + (1 + blockStartRow) * rowSkip;
+ ptrBElement = B;
+ /* set the Z matrix to 0 */
+ Z11 = 0; Z21 = 0; Z31 = 0; Z41 = 0;
+
+ /* the inner loop that computes outer products and adds them to Z */
+ for (unsigned columnCounter = blockStartRow; ; )
+ {
+ dReal q1, p1, p2, p3, p4;
+
+ /* load p and q values */
+ q1 = ptrBElement[0 * b_stride];
+ p1 = (ptrLElement - rowSkip)[0];
+ p2 = ptrLElement[0];
+ ptrLElement += rowSkip;
+ p3 = ptrLElement[0];
+ p4 = ptrLElement[0 + rowSkip];
+
+ /* compute outer product and add it to the Z matrix */
+ Z11 += p1 * q1;
+ Z21 += p2 * q1;
+ Z31 += p3 * q1;
+ Z41 += p4 * q1;
+
+ /* load p and q values */
+ q1 = ptrBElement[1 * b_stride];
+ p3 = ptrLElement[1];
+ p4 = ptrLElement[1 + rowSkip];
+ ptrLElement -= rowSkip;
+ p1 = (ptrLElement - rowSkip)[1];
+ p2 = ptrLElement[1];
+
+ /* compute outer product and add it to the Z matrix */
+ Z11 += p1 * q1;
+ Z21 += p2 * q1;
+ Z31 += p3 * q1;
+ Z41 += p4 * q1;
+
+ /* load p and q values */
+ q1 = ptrBElement[2 * b_stride];
+ p1 = (ptrLElement - rowSkip)[2];
+ p2 = ptrLElement[2];
+ ptrLElement += rowSkip;
+ p3 = ptrLElement[2];
+ p4 = ptrLElement[2 + rowSkip];
+
+ /* compute outer product and add it to the Z matrix */
+ Z11 += p1 * q1;
+ Z21 += p2 * q1;
+ Z31 += p3 * q1;
+ Z41 += p4 * q1;
+
+ /* load p and q values */
+ q1 = ptrBElement[3 * b_stride];
+ p3 = ptrLElement[3];
+ p4 = ptrLElement[3 + rowSkip];
+ ptrLElement -= rowSkip;
+ p1 = (ptrLElement - rowSkip)[3];
+ p2 = ptrLElement[3];
+
+ /* compute outer product and add it to the Z matrix */
+ Z11 += p1 * q1;
+ Z21 += p2 * q1;
+ Z31 += p3 * q1;
+ Z41 += p4 * q1;
+
+ if (columnCounter > 12)
+ {
+ columnCounter -= 12;
+
+ /* advance pointers */
+ ptrLElement += 12;
+ ptrBElement += 12 * b_stride;
+
+ /* load p and q values */
+ q1 = ptrBElement[-8 * (int)b_stride];
+ p1 = (ptrLElement - rowSkip)[-8];
+ p2 = ptrLElement[-8];
+ ptrLElement += rowSkip;
+ p3 = ptrLElement[-8];
+ p4 = ptrLElement[-8 + rowSkip];
+
+ /* compute outer product and add it to the Z matrix */
+ Z11 += p1 * q1;
+ Z21 += p2 * q1;
+ Z31 += p3 * q1;
+ Z41 += p4 * q1;
+
+ /* load p and q values */
+ q1 = ptrBElement[-7 * (int)b_stride];
+ p3 = ptrLElement[-7];
+ p4 = ptrLElement[-7 + rowSkip];
+ ptrLElement -= rowSkip;
+ p1 = (ptrLElement - rowSkip)[-7];
+ p2 = ptrLElement[-7];
+
+ /* compute outer product and add it to the Z matrix */
+ Z11 += p1 * q1;
+ Z21 += p2 * q1;
+ Z31 += p3 * q1;
+ Z41 += p4 * q1;
+
+ /* load p and q values */
+ q1 = ptrBElement[-6 * (int)b_stride];
+ p1 = (ptrLElement - rowSkip)[-6];
+ p2 = ptrLElement[-6];
+ ptrLElement += rowSkip;
+ p3 = ptrLElement[-6];
+ p4 = ptrLElement[-6 + rowSkip];
+
+ /* compute outer product and add it to the Z matrix */
+ Z11 += p1 * q1;
+ Z21 += p2 * q1;
+ Z31 += p3 * q1;
+ Z41 += p4 * q1;
+
+ /* load p and q values */
+ q1 = ptrBElement[-5 * (int)b_stride];
+ p3 = ptrLElement[-5];
+ p4 = ptrLElement[-5 + rowSkip];
+ ptrLElement -= rowSkip;
+ p1 = (ptrLElement - rowSkip)[-5];
+ p2 = ptrLElement[-5];
+
+ /* compute outer product and add it to the Z matrix */
+ Z11 += p1 * q1;
+ Z21 += p2 * q1;
+ Z31 += p3 * q1;
+ Z41 += p4 * q1;
+
+ /* load p and q values */
+ q1 = ptrBElement[-4 * (int)b_stride];
+ p1 = (ptrLElement - rowSkip)[-4];
+ p2 = ptrLElement[-4];
+ ptrLElement += rowSkip;
+ p3 = ptrLElement[-4];
+ p4 = ptrLElement[-4 + rowSkip];
+
+ /* compute outer product and add it to the Z matrix */
+ Z11 += p1 * q1;
+ Z21 += p2 * q1;
+ Z31 += p3 * q1;
+ Z41 += p4 * q1;
+
+ /* load p and q values */
+ q1 = ptrBElement[-3 * (int)b_stride];
+ p3 = ptrLElement[-3];
+ p4 = ptrLElement[-3 + rowSkip];
+ ptrLElement -= rowSkip;
+ p1 = (ptrLElement - rowSkip)[-3];
+ p2 = ptrLElement[-3];
+
+ /* compute outer product and add it to the Z matrix */
+ Z11 += p1 * q1;
+ Z21 += p2 * q1;
+ Z31 += p3 * q1;
+ Z41 += p4 * q1;
+
+ /* load p and q values */
+ q1 = ptrBElement[-2 * (int)b_stride];
+ p1 = (ptrLElement - rowSkip)[-2];
+ p2 = ptrLElement[-2];
+ ptrLElement += rowSkip;
+ p3 = ptrLElement[-2];
+ p4 = ptrLElement[-2 + rowSkip];
+
+ /* compute outer product and add it to the Z matrix */
+ Z11 += p1 * q1;
+ Z21 += p2 * q1;
+ Z31 += p3 * q1;
+ Z41 += p4 * q1;
+
+ /* load p and q values */
+ q1 = ptrBElement[-1 * (int)b_stride];
+ p3 = ptrLElement[-1];
+ p4 = ptrLElement[-1 + rowSkip];
+ ptrLElement -= rowSkip;
+ p1 = (ptrLElement - rowSkip)[-1];
+ p2 = ptrLElement[-1];
+
+ /* compute outer product and add it to the Z matrix */
+ Z11 += p1 * q1;
+ Z21 += p2 * q1;
+ Z31 += p3 * q1;
+ Z41 += p4 * q1;
+ }
+ else
+ {
+ /* advance pointers */
+ ptrLElement += 4;
+ ptrBElement += 4 * b_stride;
+
+ if ((columnCounter -= 4) == 0)
+ {
+ break;
+ }
+ }
+ /* end of inner loop */
+ }
+ }
+ else
+ {
+ ptrLElement = L + rowSkip/* + blockStartRow * rowSkip*/; dIASSERT(blockStartRow == 0);
+ ptrBElement = B;
+ /* set the Z matrix to 0 */
+ Z11 = 0; Z21 = 0; Z31 = 0; Z41 = 0;
+ }
+
+ /* finish computing the X(i) block */
+ dReal Y11, Y21, Y31, Y41;
+ {
+ Y11 = ptrBElement[0 * b_stride] - Z11;
+ ptrBElement[0 * b_stride] = Y11;
+ }
+ {
+ dReal p2 = ptrLElement[0];
+ Y21 = ptrBElement[1 * b_stride] - Z21 - p2 * Y11;
+ ptrBElement[1 * b_stride] = Y21;
+ }
+ ptrLElement += rowSkip;
+ {
+ dReal p3 = ptrLElement[0];
+ dReal p3_1 = ptrLElement[1];
+ Y31 = ptrBElement[2 * b_stride] - Z31 - p3 * Y11 - p3_1 * Y21;
+ ptrBElement[2 * b_stride] = Y31;
+ }
+ {
+ dReal p4 = ptrLElement[rowSkip];
+ dReal p4_1 = ptrLElement[1 + rowSkip];
+ dReal p4_2 = ptrLElement[2 + rowSkip];
+ Y41 = ptrBElement[3 * b_stride] - Z41 - p4 * Y11 - p4_1 * Y21 - p4_2 * Y31;
+ ptrBElement[3 * b_stride] = Y41;
+ }
+ /* end of outer loop */
+ }
+
+ /* compute rows at end that are not a multiple of block size */
+ for (; !subsequentPass || blockStartRow < rowCount; subsequentPass = true, ++blockStartRow)
+ {
+ /* compute all 1 x 1 block of X, from rows i..i+1-1 */
+ dReal *ptrBElement;
+
+ dReal Z11, Z12;
+
+ if (subsequentPass)
+ {
+ ptrBElement = B;
+ /* set the Z matrix to 0 */
+ Z11 = 0; Z12 = 0;
+
+ const dReal *ptrLElement = L + blockStartRow * rowSkip;
+
+ /* the inner loop that computes outer products and adds them to Z */
+ unsigned columnCounter = blockStartRow;
+ for (bool exitLoop = columnCounter < 4; !exitLoop; exitLoop = false)
+ {
+ dReal p1, p2, q1, q2;
+
+ /* load p and q values */
+ p1 = ptrLElement[0];
+ p2 = ptrLElement[1];
+ q1 = ptrBElement[0 * b_stride];
+ q2 = ptrBElement[1 * b_stride];
+
+ /* compute outer product and add it to the Z matrix */
+ Z11 += p1 * q1;
+ Z12 += p2 * q2;
+
+ /* load p and q values */
+ p1 = ptrLElement[2];
+ p2 = ptrLElement[3];
+ q1 = ptrBElement[2 * b_stride];
+ q2 = ptrBElement[3 * b_stride];
+
+ /* compute outer product and add it to the Z matrix */
+ Z11 += p1 * q1;
+ Z12 += p2 * q2;
+
+ if (columnCounter >= (12 + 4))
+ {
+ columnCounter -= 12;
+
+ /* advance pointers */
+ ptrLElement += 12;
+ ptrBElement += 12 * b_stride;
+
+ /* load p and q values */
+ p1 = ptrLElement[-8];
+ p2 = ptrLElement[-7];
+ q1 = ptrBElement[-8 * (int)b_stride];
+ q2 = ptrBElement[-7 * (int)b_stride];
+
+ /* compute outer product and add it to the Z matrix */
+ Z11 += p1 * q1;
+ Z12 += p2 * q2;
+
+ /* load p and q values */
+ p1 = ptrLElement[-6];
+ p2 = ptrLElement[-5];
+ q1 = ptrBElement[-6 * (int)b_stride];
+ q2 = ptrBElement[-5 * (int)b_stride];
+
+ /* compute outer product and add it to the Z matrix */
+ Z11 += p1 * q1;
+ Z12 += p2 * q2;
+
+ /* load p and q values */
+ p1 = ptrLElement[-4];
+ p2 = ptrLElement[-3];
+ q1 = ptrBElement[-4 * (int)b_stride];
+ q2 = ptrBElement[-3 * (int)b_stride];
+
+ /* compute outer product and add it to the Z matrix */
+ Z11 += p1 * q1;
+ Z12 += p2 * q2;
+
+ /* load p and q values */
+ p1 = ptrLElement[-2];
+ p2 = ptrLElement[-1];
+ q1 = ptrBElement[-2 * (int)b_stride];
+ q2 = ptrBElement[-1 * (int)b_stride];
+
+ /* compute outer product and add it to the Z matrix */
+ Z11 += p1 * q1;
+ Z12 += p2 * q2;
+ }
+ else
+ {
+ /* advance pointers */
+ ptrLElement += 4;
+ ptrBElement += 4 * b_stride;
+
+ if ((columnCounter -= 4) < 4)
+ {
+ break;
+ }
+ }
+ /* end of inner loop */
+ }
+
+ /* compute left-over iterations */
+ if ((columnCounter & 2) != 0)
+ {
+ dReal p1, p2, q1, q2;
+
+ /* load p and q values */
+ p1 = ptrLElement[0];
+ p2 = ptrLElement[1];
+ q1 = ptrBElement[0 * b_stride];
+ q2 = ptrBElement[1 * b_stride];
+
+ /* compute outer product and add it to the Z matrix */
+ Z11 += p1 * q1;
+ Z12 += p2 * q2;
+
+ /* advance pointers */
+ ptrLElement += 2;
+ ptrBElement += 2 * b_stride;
+ }
+
+ if ((columnCounter & 1) != 0)
+ {
+ dReal p1, q1;
+
+ /* load p and q values */
+ p1 = ptrLElement[0];
+ q1 = ptrBElement[0 * b_stride];
+
+ /* compute outer product and add it to the Z matrix */
+ Z11 += p1 * q1;
+
+ /* advance pointers */
+ // ptrLElement += 1; -- not needed any more
+ ptrBElement += 1 * b_stride;
+ }
+
+ /* finish computing the X(i) block */
+ dReal Y11 = ptrBElement[0 * b_stride] - (Z11 + Z12);
+ ptrBElement[0 * b_stride] = Y11;
+ }
+ }
+}
+
+
+template<unsigned int block_step>
+/*static */
+sizeint ThreadedEquationSolverLDLT::estimateCooperativelySolvingL1StraightMemoryRequirement(unsigned rowCount, SolvingL1StraightMemoryEstimates &ref_solvingMemoryEstimates)
+{
+ unsigned blockCount = deriveSolvingL1StraightBlockCount(rowCount, block_step);
+ sizeint descriptorSizeRequired = dEFFICIENT_SIZE(sizeof(cellindexint) * blockCount);
+ sizeint contextSizeRequired = dEFFICIENT_SIZE(sizeof(SolveL1StraightCellContext) * (CCI__MAX + 1) * blockCount);
+ ref_solvingMemoryEstimates.assignData(descriptorSizeRequired, contextSizeRequired);
+
+ sizeint totalSizeRequired = descriptorSizeRequired + contextSizeRequired;
+ return totalSizeRequired;
+}
+
+template<unsigned int block_step>
+/*static */
+void ThreadedEquationSolverLDLT::initializeCooperativelySolveL1StraightMemoryStructures(unsigned rowCount,
+ atomicord32 &out_blockCompletionProgress, cellindexint *blockProgressDescriptors, SolveL1StraightCellContext *dUNUSED(cellContexts))
+{
+ unsigned blockCount = deriveSolvingL1StraightBlockCount(rowCount, block_step);
+
+ out_blockCompletionProgress = 0;
+ memset(blockProgressDescriptors, 0, blockCount * sizeof(*blockProgressDescriptors));
+}
+
+template<unsigned int block_step, unsigned int b_stride>
+void ThreadedEquationSolverLDLT::participateSolvingL1Straight(const dReal *L, dReal *B, unsigned rowCount, unsigned rowSkip,
+ volatile atomicord32 &refBlockCompletionProgress/*=0*/, volatile cellindexint *blockProgressDescriptors/*=[blockCount]*/,
+ SolveL1StraightCellContext *cellContexts/*=[CCI__MAX x blockCount] + [blockCount]*/, unsigned ownThreadIndex)
+{
+ const unsigned lookaheadRange = 32;
+ const unsigned blockCount = deriveSolvingL1StraightBlockCount(rowCount, block_step), lastBlock = blockCount - 1;
+ /* compute rows at end that are not a multiple of block size */
+ const unsigned loopX1RowCount = rowCount % block_step;
+
+ BlockProcessingState blockProcessingState = BPS_NO_BLOCKS_PROCESSED;
+
+ unsigned completedBlocks = refBlockCompletionProgress;
+ unsigned currentBlock = completedBlocks;
+ dIASSERT(completedBlocks <= blockCount);
+
+ for (bool exitLoop = completedBlocks == blockCount; !exitLoop; exitLoop = false)
+ {
+ bool goForLockedBlockPrimaryCalculation = false, goForLockedBlockDuplicateCalculation = false;
+ bool goAssigningTheResult = false, stayWithinTheBlock = false;
+
+ dReal Z[block_step];
+ dReal Y[block_step];
+
+ dReal *ptrBElement;
+
+ CellContextInstance previousContextInstance;
+ unsigned completedColumnBlock;
+ bool partialBlock;
+
+ for (cellindexint testDescriptor = blockProgressDescriptors[currentBlock]; ; )
+ {
+ if (testDescriptor == INVALID_CELLDESCRIPTOR)
+ {
+ // Invalid descriptor is the indication that the row has been fully calculated
+ // Test if this was the last row and break out if so.
+ if (currentBlock + 1 == blockCount)
+ {
+ exitLoop = true;
+ break;
+ }
+
+ // Treat detected row advancement as a row processed
+ // blockProcessingState = BPS_SOME_BLOCKS_PROCESSED; <-- performs better without it
+ break;
+ }
+
+ CooperativeAtomics::AtomicReadReorderBarrier();
+ // It is necessary to read up to date completedBblocks value after the descriptor retrieval
+ // as otherwise the logic below breaks
+ completedBlocks = refBlockCompletionProgress;
+
+ if (!GET_CELLDESCRIPTOR_ISLOCKED(testDescriptor))
+ {
+ completedColumnBlock = GET_CELLDESCRIPTOR_COLUMNINDEX(testDescriptor);
+ dIASSERT(completedColumnBlock < currentBlock || (completedColumnBlock == currentBlock && currentBlock == 0)); // Otherwise, why would the calculation have had stopped if the final column is reachable???
+ dIASSERT(completedColumnBlock <= completedBlocks); // Since the descriptor is not locked
+
+ if (completedColumnBlock == completedBlocks && currentBlock != completedBlocks)
+ {
+ dIASSERT(completedBlocks < currentBlock);
+ break;
+ }
+
+ if (CooperativeAtomics::AtomicCompareExchangeCellindexint(&blockProgressDescriptors[currentBlock], testDescriptor, MARK_CELLDESCRIPTOR_LOCKED(testDescriptor)))
+ {
+ if (completedColumnBlock != 0)
+ {
+ CellContextInstance contextInstance = GET_CELLDESCRIPTOR_CONTEXTINSTANCE(testDescriptor);
+ previousContextInstance = contextInstance;
+
+ const SolveL1StraightCellContext &sourceContext = buildBlockContextRef(cellContexts, currentBlock, contextInstance);
+ sourceContext.loadPrecalculatedZs(Z);
+ }
+ else
+ {
+ previousContextInstance = CCI__MIN;
+ SolveL1StraightCellContext::initializePrecalculatedZs(Z);
+ }
+
+ goForLockedBlockPrimaryCalculation = true;
+ break;
+ }
+
+ if (blockProcessingState != BPS_COMPETING_FOR_A_BLOCK)
+ {
+ break;
+ }
+
+ testDescriptor = blockProgressDescriptors[currentBlock];
+ }
+ else
+ {
+ if (blockProcessingState != BPS_COMPETING_FOR_A_BLOCK)
+ {
+ break;
+ }
+
+ cellindexint verificativeDescriptor;
+ bool verificationFailure = false;
+
+ completedColumnBlock = GET_CELLDESCRIPTOR_COLUMNINDEX(testDescriptor);
+ dIASSERT(completedColumnBlock != currentBlock || currentBlock == 0); // There is no reason for computations to stop at the very end other than being the initial value at the very first block
+
+ if (completedColumnBlock != 0)
+ {
+ CellContextInstance contextInstance = GET_CELLDESCRIPTOR_CONTEXTINSTANCE(testDescriptor);
+ const SolveL1StraightCellContext &sourceContext = buildBlockContextRef(cellContexts, currentBlock, contextInstance);
+ sourceContext.loadPrecalculatedZs(Z);
+ }
+ else
+ {
+ SolveL1StraightCellContext::initializePrecalculatedZs(Z);
+ }
+
+ if (completedColumnBlock != 0 && completedColumnBlock <= currentBlock)
+ {
+ // Make sure the descriptor is re-read after the precalculates
+ CooperativeAtomics::AtomicReadReorderBarrier();
+ }
+
+ if (completedColumnBlock <= currentBlock)
+ {
+ verificativeDescriptor = blockProgressDescriptors[currentBlock];
+ verificationFailure = verificativeDescriptor != testDescriptor;
+ }
+
+ if (!verificationFailure)
+ {
+ dIASSERT(completedColumnBlock <= currentBlock + 1);
+
+ goForLockedBlockDuplicateCalculation = true;
+ break;
+ }
+
+ testDescriptor = verificativeDescriptor;
+ }
+ }
+
+ if (exitLoop)
+ {
+ break;
+ }
+
+ if (goForLockedBlockPrimaryCalculation)
+ {
+ blockProcessingState = BPS_SOME_BLOCKS_PROCESSED;
+
+ // Declare and assign the variables at the top to not interfere with any branching -- the compiler is going to eliminate them anyway.
+ bool handleComputationTakenOver = false, rowEndReached = false;
+
+ const dReal *ptrLElement;
+ unsigned finalColumnBlock;
+
+ /* check if this is not the partial block of fewer rows */
+ if (currentBlock != lastBlock || loopX1RowCount == 0)
+ {
+ partialBlock = false;
+
+ if (currentBlock != 0)
+ {
+ ptrLElement = L + (sizeint)(1 + currentBlock * block_step) * rowSkip + completedColumnBlock * block_step;
+ ptrBElement = B + (sizeint)(completedColumnBlock * block_step) * b_stride;
+
+ /* the inner loop that computes outer products and adds them to Z */
+ finalColumnBlock = dMACRO_MIN(currentBlock, completedBlocks);
+ dIASSERT(completedColumnBlock != finalColumnBlock/* || currentBlock == 0*/);
+
+ for (unsigned columnCounter = finalColumnBlock - completedColumnBlock; ; )
+ {
+ dReal q1, p1, p2, p3, p4;
+
+ /* load p and q values */
+ q1 = ptrBElement[0 * b_stride];
+ p1 = (ptrLElement - rowSkip)[0];
+ p2 = ptrLElement[0];
+ ptrLElement += rowSkip;
+ p3 = ptrLElement[0];
+ p4 = ptrLElement[0 + rowSkip];
+
+ /* compute outer product and add it to the Z matrix */
+ Z[0] += p1 * q1;
+ Z[1] += p2 * q1;
+ Z[2] += p3 * q1;
+ Z[3] += p4 * q1;
+
+ /* load p and q values */
+ q1 = ptrBElement[1 * b_stride];
+ p3 = ptrLElement[1];
+ p4 = ptrLElement[1 + rowSkip];
+ ptrLElement -= rowSkip;
+ p1 = (ptrLElement - rowSkip)[1];
+ p2 = ptrLElement[1];
+
+ /* compute outer product and add it to the Z matrix */
+ Z[0] += p1 * q1;
+ Z[1] += p2 * q1;
+ Z[2] += p3 * q1;
+ Z[3] += p4 * q1;
+
+ /* load p and q values */
+ q1 = ptrBElement[2 * b_stride];
+ p1 = (ptrLElement - rowSkip)[2];
+ p2 = ptrLElement[2];
+ ptrLElement += rowSkip;
+ p3 = ptrLElement[2];
+ p4 = ptrLElement[2 + rowSkip];
+
+ /* compute outer product and add it to the Z matrix */
+ Z[0] += p1 * q1;
+ Z[1] += p2 * q1;
+ Z[2] += p3 * q1;
+ Z[3] += p4 * q1;
+
+ /* load p and q values */
+ q1 = ptrBElement[3 * b_stride];
+ p3 = ptrLElement[3];
+ p4 = ptrLElement[3 + rowSkip];
+ ptrLElement -= rowSkip;
+ p1 = (ptrLElement - rowSkip)[3];
+ p2 = ptrLElement[3];
+
+ /* compute outer product and add it to the Z matrix */
+ Z[0] += p1 * q1;
+ Z[1] += p2 * q1;
+ Z[2] += p3 * q1;
+ Z[3] += p4 * q1;
+ dSASSERT(block_step == 4);
+
+ if (columnCounter > 3)
+ {
+ columnCounter -= 3;
+
+ ptrLElement += 3 * block_step;
+ ptrBElement += 3 * block_step * b_stride;
+
+ /* load p and q values */
+ q1 = ptrBElement[-8 * (int)b_stride];
+ p1 = (ptrLElement - rowSkip)[-8];
+ p2 = ptrLElement[-8];
+ ptrLElement += rowSkip;
+ p3 = ptrLElement[-8];
+ p4 = ptrLElement[-8 + rowSkip];
+
+ /* compute outer product and add it to the Z matrix */
+ Z[0] += p1 * q1;
+ Z[1] += p2 * q1;
+ Z[2] += p3 * q1;
+ Z[3] += p4 * q1;
+
+ /* load p and q values */
+ q1 = ptrBElement[-7 * (int)b_stride];
+ p3 = ptrLElement[-7];
+ p4 = ptrLElement[-7 + rowSkip];
+ ptrLElement -= rowSkip;
+ p1 = (ptrLElement - rowSkip)[-7];
+ p2 = ptrLElement[-7];
+
+ /* compute outer product and add it to the Z matrix */
+ Z[0] += p1 * q1;
+ Z[1] += p2 * q1;
+ Z[2] += p3 * q1;
+ Z[3] += p4 * q1;
+
+ /* load p and q values */
+ q1 = ptrBElement[-6 * (int)b_stride];
+ p1 = (ptrLElement - rowSkip)[-6];
+ p2 = ptrLElement[-6];
+ ptrLElement += rowSkip;
+ p3 = ptrLElement[-6];
+ p4 = ptrLElement[-6 + rowSkip];
+
+ /* compute outer product and add it to the Z matrix */
+ Z[0] += p1 * q1;
+ Z[1] += p2 * q1;
+ Z[2] += p3 * q1;
+ Z[3] += p4 * q1;
+
+ /* load p and q values */
+ q1 = ptrBElement[-5 * (int)b_stride];
+ p3 = ptrLElement[-5];
+ p4 = ptrLElement[-5 + rowSkip];
+ ptrLElement -= rowSkip;
+ p1 = (ptrLElement - rowSkip)[-5];
+ p2 = ptrLElement[-5];
+
+ /* compute outer product and add it to the Z matrix */
+ Z[0] += p1 * q1;
+ Z[1] += p2 * q1;
+ Z[2] += p3 * q1;
+ Z[3] += p4 * q1;
+
+ /* load p and q values */
+ q1 = ptrBElement[-4 * (int)b_stride];
+ p1 = (ptrLElement - rowSkip)[-4];
+ p2 = ptrLElement[-4];
+ ptrLElement += rowSkip;
+ p3 = ptrLElement[-4];
+ p4 = ptrLElement[-4 + rowSkip];
+
+ /* compute outer product and add it to the Z matrix */
+ Z[0] += p1 * q1;
+ Z[1] += p2 * q1;
+ Z[2] += p3 * q1;
+ Z[3] += p4 * q1;
+
+ /* load p and q values */
+ q1 = ptrBElement[-3 * (int)b_stride];
+ p3 = ptrLElement[-3];
+ p4 = ptrLElement[-3 + rowSkip];
+ ptrLElement -= rowSkip;
+ p1 = (ptrLElement - rowSkip)[-3];
+ p2 = ptrLElement[-3];
+
+ /* compute outer product and add it to the Z matrix */
+ Z[0] += p1 * q1;
+ Z[1] += p2 * q1;
+ Z[2] += p3 * q1;
+ Z[3] += p4 * q1;
+
+ /* load p and q values */
+ q1 = ptrBElement[-2 * (int)b_stride];
+ p1 = (ptrLElement - rowSkip)[-2];
+ p2 = ptrLElement[-2];
+ ptrLElement += rowSkip;
+ p3 = ptrLElement[-2];
+ p4 = ptrLElement[-2 + rowSkip];
+
+ /* compute outer product and add it to the Z matrix */
+ Z[0] += p1 * q1;
+ Z[1] += p2 * q1;
+ Z[2] += p3 * q1;
+ Z[3] += p4 * q1;
+
+ /* load p and q values */
+ q1 = ptrBElement[-1 * (int)b_stride];
+ p3 = ptrLElement[-1];
+ p4 = ptrLElement[-1 + rowSkip];
+ ptrLElement -= rowSkip;
+ p1 = (ptrLElement - rowSkip)[-1];
+ p2 = ptrLElement[-1];
+
+ /* compute outer product and add it to the Z matrix */
+ Z[0] += p1 * q1;
+ Z[1] += p2 * q1;
+ Z[2] += p3 * q1;
+ Z[3] += p4 * q1;
+ dSASSERT(block_step == 4);
+ }
+ else
+ {
+ ptrLElement += block_step;
+ ptrBElement += block_step * b_stride;
+
+ if (--columnCounter == 0)
+ {
+ if (finalColumnBlock == currentBlock)
+ {
+ rowEndReached = true;
+ break;
+ }
+
+ // Take a look if any more columns have been completed...
+ completedBlocks = refBlockCompletionProgress;
+ dIASSERT(completedBlocks >= finalColumnBlock);
+
+ if (completedBlocks == finalColumnBlock)
+ {
+ break;
+ }
+
+ // ...continue if so.
+ unsigned columnCompletedSoFar = finalColumnBlock;
+ finalColumnBlock = dMACRO_MIN(currentBlock, completedBlocks);
+ columnCounter = finalColumnBlock - columnCompletedSoFar;
+ }
+ }
+ /* end of inner loop */
+ }
+ }
+ else
+ {
+ ptrLElement = L + (sizeint)(1/* + currentBlock * block_step*/) * rowSkip/* + completedColumnBlock * block_step*/;
+ ptrBElement = B/* + (sizeint)(completedColumnBlock * block_step) * b_stride*/;
+ dIASSERT(completedColumnBlock == 0);
+
+ rowEndReached = true;
+ }
+ }
+ else
+ {
+ partialBlock = true;
+
+ if (currentBlock != 0)
+ {
+ dReal tempZ[dMACRO_MAX(block_step - 1U, 1U)] = { REAL(0.0), };
+
+ ptrLElement = L + (sizeint)(/*1 + */currentBlock * block_step) * rowSkip + completedColumnBlock * block_step;
+ ptrBElement = B + (sizeint)(completedColumnBlock * block_step) * b_stride;
+
+ /* the inner loop that computes outer products and adds them to Z */
+ finalColumnBlock = dMACRO_MIN(currentBlock, completedBlocks);
+ dIASSERT(completedColumnBlock != finalColumnBlock/* || currentBlock == 0*/);
+
+ for (unsigned partialRow = 0, columnCompletedSoFar = completedColumnBlock; ; )
+ {
+ dReal Z1 = 0, Z2 = 0, Z3 = 0, Z4 = 0;
+
+ for (unsigned columnCounter = finalColumnBlock - columnCompletedSoFar; ; )
+ {
+ dReal q1, q2, q3, q4, p1, p2, p3, p4;
+
+ /* load p and q values */
+ q1 = ptrBElement[0 * b_stride];
+ q2 = ptrBElement[1 * b_stride];
+ q3 = ptrBElement[2 * b_stride];
+ q4 = ptrBElement[3 * b_stride];
+ p1 = ptrLElement[0];
+ p2 = ptrLElement[1];
+ p3 = ptrLElement[2];
+ p4 = ptrLElement[3];
+
+ /* compute outer product and add it to the Z matrix */
+ Z1 += p1 * q1;
+ Z2 += p2 * q2;
+ Z3 += p3 * q3;
+ Z4 += p4 * q4;
+ dSASSERT(block_step == 4);
+
+ if (columnCounter > 3)
+ {
+ columnCounter -= 3;
+
+ ptrLElement += 3 * block_step;
+ ptrBElement += 3 * block_step * b_stride;
+
+ /* load p and q values */
+ q1 = ptrBElement[-8 * (int)b_stride];
+ q2 = ptrBElement[-7 * (int)b_stride];
+ q3 = ptrBElement[-6 * (int)b_stride];
+ q4 = ptrBElement[-5 * (int)b_stride];
+ p1 = ptrLElement[-8];
+ p2 = ptrLElement[-7];
+ p3 = ptrLElement[-6];
+ p4 = ptrLElement[-5];
+
+ /* compute outer product and add it to the Z matrix */
+ Z1 += p1 * q1;
+ Z2 += p2 * q2;
+ Z3 += p3 * q3;
+ Z4 += p4 * q4;
+
+ /* load p and q values */
+ q1 = ptrBElement[-4 * (int)b_stride];
+ q2 = ptrBElement[-3 * (int)b_stride];
+ q3 = ptrBElement[-2 * (int)b_stride];
+ q4 = ptrBElement[-1 * (int)b_stride];
+ p1 = ptrLElement[-4];
+ p2 = ptrLElement[-3];
+ p3 = ptrLElement[-2];
+ p4 = ptrLElement[-1];
+
+ /* compute outer product and add it to the Z matrix */
+ Z1 += p1 * q1;
+ Z2 += p2 * q2;
+ Z3 += p3 * q3;
+ Z4 += p4 * q4;
+ dSASSERT(block_step == 4);
+ }
+ else
+ {
+ ptrLElement += block_step;
+ ptrBElement += block_step * b_stride;
+
+ if (--columnCounter == 0)
+ {
+ break;
+ }
+ }
+ /* end of inner loop */
+ }
+
+ tempZ[partialRow] += Z1 + Z2 + Z3 + Z4;
+
+ if (++partialRow == loopX1RowCount)
+ {
+ // Here switch is used to avoid accessing Z by parametrized index.
+ // So far all the accesses were performed by explicit constants
+ // what lets the compiler treat Z elements as individual variables
+ // rather than array elements.
+ Z[0] += tempZ[0];
+
+ if (loopX1RowCount >= 2)
+ {
+ Z[1] += tempZ[1];
+
+ if (loopX1RowCount > 2)
+ {
+ Z[2] += tempZ[2];
+ }
+ }
+ dSASSERT(block_step == 4);
+
+ if (finalColumnBlock == currentBlock)
+ {
+ if (loopX1RowCount > 2)
+ {
+ // Correct the LElement so that it points to the second row
+ //
+ // Note, that ff there is just one partial row, it does not matter that
+ // the LElement will remain pointing at the first row,
+ // since the former is not going to be used in that case.
+ ptrLElement -= /*(sizeint)*/rowSkip/* * (loopX1RowCount - 2)*/; dIASSERT(loopX1RowCount == 3);
+ }
+ dSASSERT(block_step == 4);
+
+ rowEndReached = true;
+ break;
+ }
+
+ // Take a look if any more columns have been completed...
+ completedBlocks = refBlockCompletionProgress;
+ dIASSERT(completedBlocks >= finalColumnBlock);
+
+ if (completedBlocks == finalColumnBlock)
+ {
+ break;
+ }
+
+ std::fill(tempZ, tempZ + loopX1RowCount, REAL(0.0));
+ partialRow = 0;
+
+ // Correct the LElement pointer to continue at the first partial row
+ ptrLElement -= (sizeint)rowSkip * (loopX1RowCount - 1);
+
+ // ...continue if so.
+ columnCompletedSoFar = finalColumnBlock;
+ finalColumnBlock = dMACRO_MIN(currentBlock, completedBlocks);
+ }
+ else
+ {
+ ptrLElement += rowSkip - (finalColumnBlock - columnCompletedSoFar) * block_step;
+ ptrBElement -= (sizeint)((finalColumnBlock - columnCompletedSoFar) * block_step) * b_stride;
+ }
+ /* end of loop by individual rows */
+ }
+ }
+ else
+ {
+ ptrLElement = L + (sizeint)(1/* + currentBlock * block_step*/) * rowSkip/* + completedColumnBlock * block_step*/;
+ ptrBElement = B/* + (sizeint)(completedColumnBlock * block_step) * b_stride*/;
+ dIASSERT(completedColumnBlock == 0);
+
+ rowEndReached = true;
+ }
+ }
+
+ if (rowEndReached)
+ {
+ // Check whether there is still a need to proceed or if the computation has been taken over by another thread
+ cellindexint oldDescriptor = MAKE_CELLDESCRIPTOR(completedColumnBlock, previousContextInstance, true);
+
+ if (blockProgressDescriptors[currentBlock] == oldDescriptor)
+ {
+ /* finish computing the X(i) block */
+ if (!partialBlock)
+ {
+ Y[0] = ptrBElement[0 * b_stride] - Z[0];
+
+ dReal p2 = ptrLElement[0];
+ Y[1] = ptrBElement[1 * b_stride] - Z[1] - p2 * Y[0];
+
+ ptrLElement += rowSkip;
+
+ dReal p3 = ptrLElement[0];
+ dReal p3_1 = ptrLElement[1];
+ Y[2] = ptrBElement[2 * b_stride] - Z[2] - p3 * Y[0] - p3_1 * Y[1];
+
+ dReal p4 = ptrLElement[rowSkip];
+ dReal p4_1 = ptrLElement[1 + rowSkip];
+ dReal p4_2 = ptrLElement[2 + rowSkip];
+ Y[3] = ptrBElement[3 * b_stride] - Z[3] - p4 * Y[0] - p4_1 * Y[1] - p4_2 * Y[2];
+ dSASSERT(block_step == 4);
+ }
+ else
+ {
+ Y[0] = ptrBElement[0 * b_stride] - Z[0];
+
+ if (loopX1RowCount >= 2)
+ {
+ dReal p2 = ptrLElement[0];
+ Y[1] = ptrBElement[1 * b_stride] - Z[1] - p2 * Y[0];
+
+ if (loopX1RowCount > 2)
+ {
+ dReal p3 = ptrLElement[0 + rowSkip];
+ dReal p3_1 = ptrLElement[1 + rowSkip];
+ Y[2] = ptrBElement[2 * b_stride] - Z[2] - p3 * Y[0] - p3_1 * Y[1];
+ }
+ }
+ dSASSERT(block_step == 4);
+ }
+
+ // Use atomic memory barrier to make sure memory reads of ptrBElement[] and blockProgressDescriptors[] are not swapped
+ CooperativeAtomics::AtomicReadReorderBarrier();
+
+ // The descriptor has not been altered yet - this means the ptrBElement[] values used above were not modified yet
+ // and the computation result is valid.
+ if (blockProgressDescriptors[currentBlock] == oldDescriptor)
+ {
+ // Assign the results to the result context (possibly in parallel with other threads
+ // that could and ought to be assigning exactly the same values)
+ SolveL1StraightCellContext &resultContext = buildResultContextRef(cellContexts, currentBlock, blockCount);
+ resultContext.storePrecalculatedZs(Y);
+
+ // Assign the result assignment progress descriptor
+ cellindexint newDescriptor = MAKE_CELLDESCRIPTOR(currentBlock + 1, CCI__MIN, true);
+ CooperativeAtomics::AtomicCompareExchangeCellindexint(&blockProgressDescriptors[currentBlock], oldDescriptor, newDescriptor); // the result is to be ignored
+
+ // Whether succeeded or not, the result is valid, so go on trying to assign it to the matrix
+ goAssigningTheResult = true;
+ }
+ else
+ {
+ // Otherwise, go on competing for copying the results
+ handleComputationTakenOver = true;
+ }
+ }
+ else
+ {
+ handleComputationTakenOver = true;
+ }
+ }
+ else
+ {
+ // If the final column has not been reached yet, store current values to the context.
+ // Select the other context instance as the previous one might be read by other threads.
+ CellContextInstance nextContextInstance = buildNextContextInstance(previousContextInstance);
+ SolveL1StraightCellContext &destinationContext = buildBlockContextRef(cellContexts, currentBlock, nextContextInstance);
+ destinationContext.storePrecalculatedZs(Z);
+
+ // Unlock the row until more columns can be used
+ cellindexint oldDescriptor = MAKE_CELLDESCRIPTOR(completedColumnBlock, previousContextInstance, true);
+ cellindexint newDescriptor = MAKE_CELLDESCRIPTOR(finalColumnBlock, nextContextInstance, false);
+ // The descriptor might have been updated by a competing thread
+ if (!CooperativeAtomics::AtomicCompareExchangeCellindexint(&blockProgressDescriptors[currentBlock], oldDescriptor, newDescriptor))
+ {
+ // Adjust the ptrBElement to point to the result area...
+ ptrBElement = B + (sizeint)(currentBlock * block_step) * b_stride;
+ // ...and go on handling the case
+ handleComputationTakenOver = true;
+ }
+ }
+
+ if (handleComputationTakenOver)
+ {
+ cellindexint existingDescriptor = blockProgressDescriptors[currentBlock];
+ // This can only happen if the row was (has become) the uppermost not fully completed one
+ // and the competing thread is at final stage of calculation (i.e., it has reached the currentBlock column).
+ if (existingDescriptor != INVALID_CELLDESCRIPTOR)
+ {
+ // If not fully completed this must be the final stage of the result assignment into the matrix
+ dIASSERT(existingDescriptor == MAKE_CELLDESCRIPTOR(currentBlock + 1, CCI__MIN, true));
+
+ // Go on competing copying the result as anyway the block is the topmost not completed one
+ // and since there was competition for it, there is no other work that can be done right now.
+ const SolveL1StraightCellContext &resultContext = buildResultContextRef(cellContexts, currentBlock, blockCount);
+ resultContext.loadPrecalculatedZs(Y);
+
+ goAssigningTheResult = true;
+ }
+ else
+ {
+ // everything is over -- just go handling next blocks
+ }
+ }
+ }
+ else if (goForLockedBlockDuplicateCalculation)
+ {
+ blockProcessingState = BPS_SOME_BLOCKS_PROCESSED;
+
+ bool skipToHandlingSubsequentRows = false, skiptoCopyingResult = false;
+
+ /* declare variables */
+ const dReal *ptrLElement;
+
+ if (completedColumnBlock < currentBlock)
+ {
+ /* check if this is not the partial block of fewer rows */
+ if (currentBlock != lastBlock || loopX1RowCount == 0)
+ {
+ partialBlock = false;
+
+ ptrLElement = L + (sizeint)(1 + currentBlock * block_step) * rowSkip + completedColumnBlock * block_step;
+ ptrBElement = B + (sizeint)(completedColumnBlock * block_step) * b_stride;
+
+ /* the inner loop that computes outer products and adds them to Z */
+ unsigned finalColumnBlock = currentBlock;
+ dIASSERT(currentBlock == completedBlocks); // Why would we be competing for a row otherwise?
+
+ unsigned lastCompletedColumn = completedColumnBlock;
+ unsigned columnCounter = finalColumnBlock - completedColumnBlock;
+ for (bool exitInnerLoop = false; !exitInnerLoop; exitInnerLoop = --columnCounter == 0)
+ {
+ dReal q1, p1, p2, p3, p4;
+
+ /* load p and q values */
+ q1 = ptrBElement[0 * b_stride];
+ p1 = (ptrLElement - rowSkip)[0];
+ p2 = ptrLElement[0];
+ ptrLElement += rowSkip;
+ p3 = ptrLElement[0];
+ p4 = ptrLElement[0 + rowSkip];
+
+ /* compute outer product and add it to the Z matrix */
+ Z[0] += p1 * q1;
+ Z[1] += p2 * q1;
+ Z[2] += p3 * q1;
+ Z[3] += p4 * q1;
+
+ /* load p and q values */
+ q1 = ptrBElement[1 * b_stride];
+ p3 = ptrLElement[1];
+ p4 = ptrLElement[1 + rowSkip];
+ ptrLElement -= rowSkip;
+ p1 = (ptrLElement - rowSkip)[1];
+ p2 = ptrLElement[1];
+
+ /* compute outer product and add it to the Z matrix */
+ Z[0] += p1 * q1;
+ Z[1] += p2 * q1;
+ Z[2] += p3 * q1;
+ Z[3] += p4 * q1;
+
+ /* load p and q values */
+ q1 = ptrBElement[2 * b_stride];
+ p1 = (ptrLElement - rowSkip)[2];
+ p2 = ptrLElement[2];
+ ptrLElement += rowSkip;
+ p3 = ptrLElement[2];
+ p4 = ptrLElement[2 + rowSkip];
+
+ /* compute outer product and add it to the Z matrix */
+ Z[0] += p1 * q1;
+ Z[1] += p2 * q1;
+ Z[2] += p3 * q1;
+ Z[3] += p4 * q1;
+
+ /* load p and q values */
+ q1 = ptrBElement[3 * b_stride];
+ p3 = ptrLElement[3];
+ p4 = ptrLElement[3 + rowSkip];
+ ptrLElement -= rowSkip;
+ p1 = (ptrLElement - rowSkip)[3];
+ p2 = ptrLElement[3];
+
+ /* compute outer product and add it to the Z matrix */
+ Z[0] += p1 * q1;
+ Z[1] += p2 * q1;
+ Z[2] += p3 * q1;
+ Z[3] += p4 * q1;
+ dSASSERT(block_step == 4);
+
+ // Check if the primary solver thread has not made any progress
+ cellindexint descriptorVerification = blockProgressDescriptors[currentBlock];
+ unsigned newCompletedColumn = GET_CELLDESCRIPTOR_COLUMNINDEX(descriptorVerification);
+
+ if (newCompletedColumn != lastCompletedColumn)
+ {
+ // Check, this is the first change the current thread detects.
+ // There is absolutely no reason in code for the computation to stop/resume twice
+ // while the current thread is competing.
+ dIASSERT(lastCompletedColumn == completedColumnBlock);
+
+ if (descriptorVerification == INVALID_CELLDESCRIPTOR)
+ {
+ skipToHandlingSubsequentRows = true;
+ break;
+ }
+
+ if (newCompletedColumn == currentBlock + 1)
+ {
+ skiptoCopyingResult = true;
+ break;
+ }
+
+ // Check if the current thread is behind
+ if (newCompletedColumn > finalColumnBlock - columnCounter)
+ {
+ // If so, go starting over one more time
+ blockProcessingState = BPS_COMPETING_FOR_A_BLOCK;
+ stayWithinTheBlock = true;
+ skipToHandlingSubsequentRows = true;
+ break;
+ }
+
+ // If current thread is ahead, just save new completed column for further comparisons and go on calculating
+ lastCompletedColumn = newCompletedColumn;
+ }
+
+ /* advance pointers */
+ ptrLElement += block_step;
+ ptrBElement += block_step * b_stride;
+ /* end of inner loop */
+ }
+ }
+ else
+ {
+ partialBlock = true;
+
+ dReal tempZ[dMACRO_MAX(block_step - 1U, 1U)] = { REAL(0.0), };
+
+ ptrLElement = L + (sizeint)(/*1 + */currentBlock * block_step) * rowSkip + completedColumnBlock * block_step;
+ ptrBElement = B + (sizeint)(completedColumnBlock * block_step) * b_stride;
+
+ /* the inner loop that computes outer products and adds them to Z */
+ unsigned finalColumnBlock = currentBlock;
+ dIASSERT(currentBlock == completedBlocks); // Why would we be competing for a row otherwise?
+
+ unsigned lastCompletedColumn = completedColumnBlock;
+ for (unsigned columnCounter = finalColumnBlock - completedColumnBlock; ; )
+ {
+ dReal q1, q2, q3, q4;
+
+ /* load q values */
+ q1 = ptrBElement[0 * b_stride];
+ q2 = ptrBElement[1 * b_stride];
+ q3 = ptrBElement[2 * b_stride];
+ q4 = ptrBElement[3 * b_stride];
+
+ for (unsigned partialRow = 0; ; )
+ {
+ dReal p1, p2, p3, p4;
+
+ /* load p values */
+ p1 = ptrLElement[0];
+ p2 = ptrLElement[1];
+ p3 = ptrLElement[2];
+ p4 = ptrLElement[3];
+
+ /* compute outer product and add it to the Z matrix */
+ tempZ[partialRow] += p1 * q1 + p2 * q2 + p3 * q3 + p4 * q4;
+ dSASSERT(block_step == 4);
+
+ if (++partialRow == loopX1RowCount)
+ {
+ break;
+ }
+
+ ptrLElement += rowSkip;
+ }
+
+ // Check if the primary solver thread has not made any progress
+ cellindexint descriptorVerification = blockProgressDescriptors[currentBlock];
+ unsigned newCompletedColumn = GET_CELLDESCRIPTOR_COLUMNINDEX(descriptorVerification);
+
+ if (newCompletedColumn != lastCompletedColumn)
+ {
+ // Check, this is the first change the current thread detects.
+ // There is absolutely no reason in code for the computation to stop/resume twice
+ // while the current thread is competing.
+ dIASSERT(lastCompletedColumn == completedColumnBlock);
+
+ if (descriptorVerification == INVALID_CELLDESCRIPTOR)
+ {
+ skipToHandlingSubsequentRows = true;
+ break;
+ }
+
+ if (newCompletedColumn == currentBlock + 1)
+ {
+ skiptoCopyingResult = true;
+ break;
+ }
+
+ // Check if the current thread is behind
+ if (newCompletedColumn > finalColumnBlock - columnCounter)
+ {
+ // If so, go starting over one more time
+ blockProcessingState = BPS_COMPETING_FOR_A_BLOCK;
+ stayWithinTheBlock = true;
+ skipToHandlingSubsequentRows = true;
+ break;
+ }
+
+ // If current thread is ahead, just save new completed column for further comparisons and go on calculating
+ lastCompletedColumn = newCompletedColumn;
+ }
+
+ ptrLElement += block_step;
+ ptrBElement += block_step * b_stride;
+
+ if (--columnCounter == 0)
+ {
+ // Here switch is used to avoid accessing Z by parametrized index.
+ // So far all the accesses were performed by explicit constants
+ // what lets the compiler treat Z elements as individual variables
+ // rather than array elements.
+ Z[0] += tempZ[0];
+
+ if (loopX1RowCount >= 2)
+ {
+ Z[1] += tempZ[1];
+
+ if (loopX1RowCount > 2)
+ {
+ Z[2] += tempZ[2];
+
+ // Correct the LElement so that it points to the second row
+ //
+ // Note, that if there is just one partial row, it does not matter that
+ // the LElement will remain pointing at the first row,
+ // since the former is not going to be used in that case.
+ ptrLElement -= /*(sizeint)*/rowSkip/* * (loopX1RowCount - 2)*/; dIASSERT(loopX1RowCount == 3);
+ }
+ }
+ dSASSERT(block_step == 4);
+
+ break;
+ }
+
+ /* advance pointers */
+ ptrLElement -= (sizeint)rowSkip * (loopX1RowCount - 1);
+ /* end of inner loop */
+ }
+ }
+ }
+ else if (completedColumnBlock > currentBlock)
+ {
+ dIASSERT(completedColumnBlock == currentBlock + 1);
+
+ partialBlock = currentBlock == lastBlock && loopX1RowCount != 0;
+
+ skiptoCopyingResult = true;
+ }
+ else
+ {
+ dIASSERT(currentBlock == 0); // Execution can get here within the very first block only
+
+ partialBlock = rowCount < block_step;
+
+ /* assign the pointers appropriately and go on computing the results */
+ ptrLElement = L + (sizeint)(1/* + currentBlock * block_step*/) * rowSkip/* + completedColumnBlock * block_step*/;
+ ptrBElement = B/* + (sizeint)(completedColumnBlock * block_step) * b_stride*/;
+ }
+
+ if (!skipToHandlingSubsequentRows)
+ {
+ if (!skiptoCopyingResult)
+ {
+ if (!partialBlock)
+ {
+ Y[0] = ptrBElement[0 * b_stride] - Z[0];
+
+ dReal p2 = ptrLElement[0];
+ Y[1] = ptrBElement[1 * b_stride] - Z[1] - p2 * Y[0];
+
+ ptrLElement += rowSkip;
+
+ dReal p3 = ptrLElement[0];
+ dReal p3_1 = ptrLElement[1];
+ Y[2] = ptrBElement[2 * b_stride] - Z[2] - p3 * Y[0] - p3_1 * Y[1];
+
+ dReal p4 = ptrLElement[rowSkip];
+ dReal p4_1 = ptrLElement[1 + rowSkip];
+ dReal p4_2 = ptrLElement[2 + rowSkip];
+ Y[3] = ptrBElement[3 * b_stride] - Z[3] - p4 * Y[0] - p4_1 * Y[1] - p4_2 * Y[2];
+ dSASSERT(block_step == 4);
+ }
+ else
+ {
+ Y[0] = ptrBElement[0 * b_stride] - Z[0];
+
+ if (loopX1RowCount >= 2)
+ {
+ dReal p2 = ptrLElement[0];
+ Y[1] = ptrBElement[1 * b_stride] - Z[1] - p2 * Y[0];
+
+ if (loopX1RowCount > 2)
+ {
+ dReal p3 = ptrLElement[0 + rowSkip];
+ dReal p3_1 = ptrLElement[1 + rowSkip];
+ Y[2] = ptrBElement[2 * b_stride] - Z[2] - p3 * Y[0] - p3_1 * Y[1];
+ }
+ }
+ dSASSERT(block_step == 4);
+ }
+
+ CooperativeAtomics::AtomicReadReorderBarrier();
+
+ // Use atomic load to make sure memory reads of ptrBElement[] and blockProgressDescriptors[] are not swapped
+ cellindexint existingDescriptor = blockProgressDescriptors[currentBlock];
+
+ if (existingDescriptor == INVALID_CELLDESCRIPTOR)
+ {
+ // Everything is over -- proceed to subsequent rows
+ skipToHandlingSubsequentRows = true;
+ }
+ else if (existingDescriptor == MAKE_CELLDESCRIPTOR(currentBlock + 1, CCI__MIN, true))
+ {
+ // The values computed above may not be valid. Copy the values already in the result context.
+ skiptoCopyingResult = true;
+ }
+ else
+ {
+ // The descriptor has not been altered yet - this means the ptrBElement[] values used above were not modified yet
+ // and the computation result is valid.
+ cellindexint newDescriptor = MAKE_CELLDESCRIPTOR(currentBlock + 1, CCI__MIN, true); // put the computation at the top so that the evaluation result from the expression above is reused
+
+ // Assign the results to the result context (possibly in parallel with other threads
+ // that could and ought to be assigning exactly the same values)
+ SolveL1StraightCellContext &resultContext = buildResultContextRef(cellContexts, currentBlock, blockCount);
+ resultContext.storePrecalculatedZs(Y);
+
+ // Assign the result assignment progress descriptor
+ CooperativeAtomics::AtomicCompareExchangeCellindexint(&blockProgressDescriptors[currentBlock], existingDescriptor, newDescriptor); // the result is to be ignored
+
+ // Whether succeeded or not, the result is valid, so go on trying to assign it to the matrix
+ }
+ }
+
+ if (!skipToHandlingSubsequentRows)
+ {
+ if (skiptoCopyingResult)
+ {
+ // Extract the result values stored in the result context
+ const SolveL1StraightCellContext &resultContext = buildResultContextRef(cellContexts, currentBlock, blockCount);
+ resultContext.loadPrecalculatedZs(Y);
+
+ ptrBElement = B + (sizeint)(currentBlock * block_step) * b_stride;
+ }
+
+ goAssigningTheResult = true;
+ }
+ }
+ }
+
+ if (goAssigningTheResult)
+ {
+ cellindexint existingDescriptor = blockProgressDescriptors[currentBlock];
+ // Check if the assignment has not been completed yet
+ if (existingDescriptor != INVALID_CELLDESCRIPTOR)
+ {
+ // Assign the computation results to the B vector
+ if (!partialBlock)
+ {
+ ptrBElement[0 * b_stride] = Y[0];
+ ptrBElement[1 * b_stride] = Y[1];
+ ptrBElement[2 * b_stride] = Y[2];
+ ptrBElement[3 * b_stride] = Y[3];
+ dSASSERT(block_step == 4);
+ }
+ else
+ {
+ ptrBElement[0 * b_stride] = Y[0];
+
+ if (loopX1RowCount >= 2)
+ {
+ ptrBElement[1 * b_stride] = Y[1];
+
+ if (loopX1RowCount > 2)
+ {
+ ptrBElement[2 * b_stride] = Y[2];
+ }
+ }
+ dSASSERT(block_step == 4);
+ }
+
+ ThrsafeIncrementIntUpToLimit(&refBlockCompletionProgress, currentBlock + 1);
+ dIASSERT(refBlockCompletionProgress >= currentBlock + 1);
+
+ // And assign the completed status no matter what
+ CooperativeAtomics::AtomicStoreCellindexint(&blockProgressDescriptors[currentBlock], INVALID_CELLDESCRIPTOR);
+ }
+ else
+ {
+ // everything is over -- just go handling next blocks
+ }
+ }
+
+ if (!stayWithinTheBlock)
+ {
+ completedBlocks = refBlockCompletionProgress;
+
+ if (completedBlocks == blockCount)
+ {
+ break;
+ }
+
+ currentBlock += 1;
+
+ bool lookaheadBoundaryReached = false;
+
+ if (currentBlock == blockCount || completedBlocks == 0)
+ {
+ lookaheadBoundaryReached = true;
+ }
+ else if (currentBlock >= completedBlocks + lookaheadRange)
+ {
+ lookaheadBoundaryReached = blockProcessingState > BPS_NO_BLOCKS_PROCESSED;
+ }
+ else if (currentBlock < completedBlocks)
+ {
+ // Treat detected row advancement as a row processed
+ // blockProcessingState = BPS_SOME_BLOCKS_PROCESSED; <-- performs better without it
+
+ currentBlock = completedBlocks;
+ }
+
+ if (lookaheadBoundaryReached)
+ {
+ dIASSERT(blockProcessingState != BPS_COMPETING_FOR_A_BLOCK); // Why did not we compete???
+
+ // If no row has been processed in the previous pass, compete for the next row to avoid cycling uselessly
+ if (blockProcessingState <= BPS_NO_BLOCKS_PROCESSED)
+ {
+ // Abandon job if too few blocks remain
+ if (blockCount - completedBlocks <= ownThreadIndex)
+ {
+ break;
+ }
+
+ blockProcessingState = BPS_COMPETING_FOR_A_BLOCK;
+ }
+ else
+ {
+ // If there was some progress, just continue to the next pass
+ blockProcessingState = BPS_NO_BLOCKS_PROCESSED;
+ }
+
+ currentBlock = completedBlocks;
+ }
+ }
+ }
+}
+
+
+#endif
generated by cgit v1.2.1 (git 2.18.0) at 2024-11-01 02:21:14 +0000