MPI implementation of matrix multiplication - tested with OpenMPI

gemm/mpi/CMakeLists.txt

+
+# ==================================================================================================
+# This file is part of the CodeVault project. The project is licensed under Apache Version 2.0.
+# CodeVault is part of the EU-project PRACE-4IP (WP7.3.C).
+#
+# Author(s):
+#   Valeriu Codreanu <valeriu.codreanu@surfsara.nl>
+#
+# ==================================================================================================
+
+# Packages are optional: if they are not present, certain code samples are not compiled
+find_package(MPI)     # Built-in in CMake
+include(${CMAKE_CURRENT_SOURCE_DIR}/../../../cmake/common.cmake)
+# ==================================================================================================
+if ("${DWARF_PREFIX}" STREQUAL "")
+  set(DWARF_PREFIX 1_dense)
+endif()
+
+# C++ compiler settings
+
+find_package(Common)
+select_compiler_flags(cxx_flags
+  GNU "-march=native"   # I suggest remove "-O3" as this is controlled by the CMAKE_BUILD_TYPE
+  CLANG "-march=native" # same here
+  Intel "-axavx2,avx")
+set(CXX_FLAGS ${cxx_flags})
+if("${CMAKE_CXX_COMPILER_ID}" STREQUAL "GNU")
+  set(CXX_FLAGS "${CXX_FLAGS} -Wall -Wno-comment")
+  if(APPLE)
+    set(CXX_FLAGS "${CXX_FLAGS} -Wa,-q")
+  endif()
+endif()
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${CXX_FLAGS}")
+set(CMAKE_CXX_COMPILE_FLAGS ${CMAKE_CXX_COMPILE_FLAGS} ${MPI_COMPILE_FLAGS})
+set(CMAKE_CXX_LINK_FLAGS ${CMAKE_CXX_LINK_FLAGS} ${MPI_LINK_FLAGS})
+# ==================================================================================================
+
+# GEMM with the CUDA library
+set(NAME ${DWARF_PREFIX}_gemm_mpi)
+if (MPI_FOUND)
+  add_executable(${NAME} src/gemm_mpi.cpp)
+  target_link_libraries(${NAME} ${MPI_LIBRARIES})
+  install(TARGETS ${NAME} DESTINATION bin)
+else()
+  message("** Skipping '${NAME}': no MPI")
+  dummy_install(${NAME} "MPI")
+endif()
+
+unset(NAME)
+
+# ==================================================================================================

gemm/mpi/README.md

+=======
+README
+=======
+
+# 1. Code sample name
+gemm
+
+# 2. Description of the code sample package
+This example demonstrates an MPI implementation
+
+Additional pre-requisites:
+* Open MPI there is an issue with Intel or mvapich. Will be fixed in future releases. For now it works also with those if  something like "mpirun_rsh -np 2 node001 node001 ./runnable" is used
+
+# 3. Release date
+6 Dec 2016
+
+# 4. Version history 
+1.0
+
+Damian Podareanu <damian.podareanu@surfsara.nl>
+
+# 6. Copyright / License of the code sample
+Apache 2.0
+
+# 7. Language(s) 
+C++
+
+# 8. Parallelisation Implementation(s)
+MPI
+
+# 9. Level of the code sample complexity 
+Basic level
+
+# 10. Instructions on how to compile the code
+Uses the CodeVault CMake infrastructure, see main README.md
+
+# 11. Instructions on how to run the code
+Run the executable. Matrix size is defined in src/gemm_mpi.cpp
+
+# 12. Sample input(s)
+Input-data is generated automatically when running the program.
\ No newline at end of file

gemm/mpi/src/gemm_mpi.cpp

+#include "stdio.h"
+#include "mpi.h"
+
+const int size = 1000;
+
+float a[size][size];
+float b[size][size];
+float c[size][size];
+
+void multiply(int istart, int iend)
+{
+    for (int i = istart; i <= iend; ++i) {
+        for (int j = 0; j < size; ++j) {
+            for (int k = 0; k < size; ++k) {
+                c[i][j] += a[i][k] * b[k][j];
+            }
+        }
+    }
+}
+
+int main(int argc, char* argv[])
+{
+    int rank, nproc;
+    int istart, iend;
+    double start, end;
+
+    MPI_Init(&argc, &argv);
+    MPI_Comm_size(MPI_COMM_WORLD, &nproc);
+    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
+
+    // MPI_Barrier(MPI_COMM_WORLD);
+    // start = MPI_Wtime();
+
+    if (rank == 0) {
+        // Initialize buffers.
+        for (int i = 0; i < size; ++i) {
+            for (int j = 0; j < size; ++j) {
+                a[i][j] = (float)i + j;
+                b[i][j] = (float)i - j;
+                c[i][j] = 0.0f;
+            }
+        }
+    }
+
+    // Broadcast matrices to all workers.
+    MPI_Bcast(a, size*size, MPI_FLOAT, 0,MPI_COMM_WORLD);
+    MPI_Bcast(b, size*size, MPI_FLOAT, 0,MPI_COMM_WORLD);
+    MPI_Bcast(c, size*size, MPI_FLOAT, 0,MPI_COMM_WORLD);
+
+    // Partition work by i-for-loop.
+    istart = (size / nproc) * rank;
+    iend = (size / nproc) * (rank + 1) - 1;
+
+    // Compute matrix multiplication in [istart,iend]
+    // of i-for-loop.
+    // C <- C + A x B
+    multiply(istart, iend);
+
+    // Gather computed results.
+    MPI_Gather(c + (size/nproc*rank),
+               size*size/nproc,
+               MPI_FLOAT,
+               c + (size/nproc*rank),
+               size*size/nproc,
+               MPI_FLOAT,
+               0,
+               MPI_COMM_WORLD);
+
+    if (rank == 0) {
+        // Compute remaining multiplications
+        // when size % nproc > 0.
+        if (size % nproc > 0) {
+            multiply((size/nproc)*nproc, size-1);
+        }
+    }
+    
+    // MPI_Barrier(MPI_COMM_WORLD);
+    // end = MPI_Wtime();
+
+    MPI_Finalize();
+
+    // if (rank == 0) { /* use time on master node */
+    //     float msec_total = 0.0f;
+
+    //     // Compute and print the performance
+    //     float msec_per_matrix_mul = end-start;
+    //     double flops_per_matrix_mul = 2.0 * (double)size * (double)size * (double)size;
+    //     double giga_flops = (flops_per_matrix_mul * 1.0e-9f) / (msec_per_matrix_mul / 1000.0f);
+    //     printf(
+    //         "Performance= %.2f GFlop/s, Time= %.3f msec, Size= %.0f Ops\n",
+    //         giga_flops,
+    //         msec_per_matrix_mul,
+    //         flops_per_matrix_mul);
+    // }
+   
+
+    return 0;
+}
\ No newline at end of file