Implementation of sptrsv + spmv

.gitignore

@@ -0,0 +1 @@
+tools/HyperDAG_generator/generator

tools/HyperDAG_generator/README.txt → tools/HyperDAG_generator/README.md

@@ -57,10 +57,14 @@ fixedIn: generates a random DAG where each node has an indegree that is equal to
 
 expectedIn: generates a random DAG where each node has an indegree that is equal to the "indegree" parameter in expectation. That is, for node i, there is an edge from each node 0,1,...,(i-1) to node i independently with a fixed probability p, and p is always chosen such that the expected number of incoming edges of node i is "indegree". If i<"indegree", the node i has an edge from all previous nodes, and its degree is only i. Additionally, if the sourceProb parameter is used, then each node is chosen as a source with this given probability, meaning that it does not have any incoming edges at all (corresponding to e.g. new input scalars in a real-world computation). Possible parameters: -output, -N, -indegree, -sourceProb
 
-SpMV: generates a DAG that corresponds to the fine-grained representation of a sparse matrix - dense vector multiplication. By default, it generates an input matrix A of size NxN, where each cell is independently nonzero with a given probability, specified by the "nonzeroProb" parameter. The matrix A is then multiplied by a dense vector of length N. In the end, nodes of the DAG which are no connected to the final vector (e.g. cells of the vector that correspond to a matrix column that is entirely zero) are removed. It is also possible to read the input matrix from a file if the "input" parameter is used. Possible parameters: -output, -N, -input, -nonzeroProb
+SpMV: generates a DAG that corresponds to the fine-grained representation of a sparse matrix - dense vector multiplication (out-of-place). By default, it generates an input matrix A of size NxN, where each cell is independently nonzero with a given probability, specified by the "nonzeroProb" parameter. The matrix A is then multiplied by a dense vector of length N. In the end, nodes of the DAG which are no connected to the final vector (e.g. cells of the vector that correspond to a matrix column that is entirely zero) are removed. It is also possible to read the input matrix from a file if the "input" parameter is used. Possible parameters: -output, -N, -input, -nonzeroProb
 
 SpMVExp: generates a DAG that corresponds to the fine-grained representation of the operation where we multiply a sparse matrix A by a dense vector v, K times consecutively (A^k * v). By default, it generates an input matrix A of size NxN, where each cell is independently nonzero with a given probability, specified by the "nonzeroProb" parameter. The matrix A is then multiplied by a dense vector (of length N) K times consecutively. In the end, nodes of the DAG which are not predecessors of a nonzero value in the final vector are removed. Note that this step can remove larger parts of the DAG when some intermediate values are computed, but then always only multiplied by zeros (in some cases, the number of nodes in the DAG can even decrease when K is incremented!). It is also possible to read the input matrix from a file if the "input" parameter is used. Possible parameters: -output, -N, -K, -input, -nonzeroProb
 
+LLtSolver: generates a DAG that corresponds to the fine-grained representation of: z =  inv(trans(L)) . (inv(L) . x). By default, it generates an input matrix A of size NxN, where each cell is independently nonzero with a given probability, specified by the "nonzeroProb" parameter. The matrix L (lower triangular matrix) is then multiplied by a dense vector of length N, then the result is re-multiplied by this same matrix L but transposed. In the end, nodes of the DAG which are no connected to the final vector (e.g. cells of the vector that correspond to a matrix column that is entirely zero) are removed. It is also possible to read the input matrix from a file if the "input" parameter is used. Possible parameters: -output, -N, -input, -nonzeroProb
+
+LUSolver: generates a DAG that corresponds to the fine-grained representation of: z =  inv(U) . (inv(L) . x). By default, it generates an input matrix A of size NxN, where each cell is independently nonzero with a given probability, specified by the "nonzeroProb" parameter. The matrix L (lower triangular matrix) is then multiplied by a dense vector of length N, then the result is re-multiplied by this same matrix L but transposed. In the end, nodes of the DAG which are no connected to the final vector (e.g. cells of the vector that correspond to a matrix column that is entirely zero) are removed. It is also possible to read the input matrix from a file if the "input" parameter is used. Possible parameters: -output, -N, -input, -nonzeroProb
+
 kNN: generates a DAG that corresponds to the fine-grained representation of the operation where we multiply a sparse matrix A by a sparse vector v, K times consecutively (A^k * v). By default, it generates an input matrix A of size NxN, where each cell is independently nonzero with a given probability, specified by the "nonzeroProb" parameter. The vector "v" only has one nonzero cell (at a randomly chosen index) in the beginning. The matrix A is then multiplied by the vector v, and this is done K times consecutively. In the end, nodes of the DAG which are no connected to the final vector are removed. It is also possible to read the input matrix from a file if the "input" parameter is used; in this case, we can also specify the index of v where the nonzero cell is located, through the "sourceNode" parameter. Possible parameters: -output, -N, -K, -input, -nonzeroProb, -sourceNode
 
 CG: generates a DAG that corresponds to the fine-grained representation of K iterations of the conjugate gradient method for a sparse matrix A (corresponding to a simple version of the iterative method, as described in the pseudocode in the appropriate Wikipedia article). By default, it generates an input matrix A of size NxN, where each cell is independently nonzero with a given probability, specified by the "nonzeroProb" parameter. It is also possible to read the input matrix from a file if the "input" parameter is used. Possible parameters: -output, -N, -K, -input, -nonzeroProb
@@ -69,7 +73,7 @@ CG: generates a DAG that corresponds to the fine-grained representation of K ite
 POSSIBLE PARAMETERS (case sensitive):
 -------------------------------------
 
-"-mode" : compulsory parameter, describes the method used for generating a hyperDAG. Possible values are: ER, fixedIn, expectedIn, SpMV, SpMVExp, kNN, CG
+"-mode" : compulsory parameter, describes the method used for generating a hyperDAG. Possible values are: ER, fixedIn, expectedIn, SpMV, SpMVExp, LLtSolver, kNN, CG
 
 "-output": describes the name of the output file where the hyperDAG is written. If not specified, the default name is "output.txt"
 

tools/HyperDAG_generator/compile.sh

@@ -0,0 +1,4 @@
+command="g++ generator.cpp -o generator -g -Ofast -ltbb -std=c++17"
+
+echo $command
+eval $command