GPU and sparse kernels also calculate BMUs after last epoch

peterwittek · Aug 4, 2016 · 5aacea5 · 5aacea5
1 parent eb15da9
commit 5aacea5
Show file tree

Hide file tree

Showing 5 changed files with 34 additions and 12 deletions.
diff --git a/src/Python/setup.py b/src/Python/setup.py
@@ -34,10 +34,10 @@ def find_cuda():
     arch = int(platform.architecture()[0][0:2])
     if sys.platform.startswith('win'):
         os.path.join(libdir, "x"+str(arch))
-    if os.path.exists(os.path.join(home, libdir)):
-        cudaconfig['lib'] = libdir
-    elif os.path.exists(os.path.join(home, libdir + "64")):
+    if os.path.exists(os.path.join(home, libdir + "64")):
         cudaconfig['lib'] = libdir + "64"
+    elif os.path.exists(os.path.join(home, libdir)):
+        cudaconfig['lib'] = libdir
     else:
         raise EnvironmentError('The CUDA libraries could not be located')
     return cudaconfig

diff --git a/src/denseGpuKernels.cu b/src/denseGpuKernels.cu
@@ -322,9 +322,21 @@ void trainOneEpochDenseGPU(int itask, float *data, float *numerator,
                            unsigned int nDimensions, unsigned int nVectors,
                            unsigned int nVectorsPerRank, float radius,
                            float scale, string mapType, string gridType,
-                           bool compact_support, bool gaussian, int *globalBmus) {
+                           bool compact_support, bool gaussian,
+                           int *globalBmus, bool only_bmus) {
     unsigned int *bmus = new unsigned int[nVectorsPerRank * 2];
     getBmusOnGpu(bmus, codebook, nSomX, nSomY, nDimensions, nVectorsPerRank);
+    if (only_bmus) {
+#ifdef HAVE_MPI
+        MPI_Gather(bmus, nVectorsPerRank * 2, MPI_INT, globalBmus, nVectorsPerRank * 2, MPI_INT, 0, MPI_COMM_WORLD);
+#else
+        for (unsigned int i = 0; i < 2 * nVectorsPerRank; ++i) {
+            globalBmus[i] = bmus[i];
+        }
+#endif
+        delete [] bmus;
+        return;
+    }
     float *localNumerator = new float[nSomY * nSomX * nDimensions];
     float *localDenominator = new float[nSomY * nSomX];
 

diff --git a/src/somoclu.h b/src/somoclu.h
@@ -120,15 +120,15 @@ void trainOneEpochDenseCPU(int itask, float *data, float *numerator,
                            unsigned int nVectorsPerRank, float radius,
                            float scale, string mapType,
                            string gridType, bool compact_support, bool gaussian,
-                           int *globalBmus, bool only_bmus=false);
+                           int *globalBmus, bool only_bmus);
 void trainOneEpochSparseCPU(int itask, svm_node **sparseData, float *numerator,
                             float *denominator, float *codebook,
                             unsigned int nSomX, unsigned int nSomY,
                             unsigned int nDimensions, unsigned int nVectors,
                             unsigned int nVectorsPerRank, float radius,
                             float scale, string mapType,
                             string gridType, bool compact_support, bool gaussian,
-                            int *globalBmus);
+                            int *globalBmus, bool only_bmus);
 void initializeCodebook(unsigned int seed, float *codebook, unsigned int nSomX,
                         unsigned int nSomY, unsigned int nDimensions);
 
@@ -145,7 +145,7 @@ extern "C" {
                                unsigned int nVectorsPerRank, float radius,
                                float scale, string mapType,
                                string gridType, bool compact_support, bool gaussian,
-                               int *globalBmus);
+                               int *globalBmus, bool only_bmus);
 #endif
     void my_abort(string err);
 }

diff --git a/src/sparseCpuKernels.cpp b/src/sparseCpuKernels.cpp
@@ -84,7 +84,8 @@ void trainOneEpochSparseCPU(int itask, svm_node **sparseData, float *numerator,
                             unsigned int nDimensions, unsigned int nVectors,
                             unsigned int nVectorsPerRank, float radius,
                             float scale, string mapType, string gridType,
-                            bool compact_support, bool gaussian, int *globalBmus) {
+                            bool compact_support, bool gaussian,
+                            int *globalBmus, bool only_bmus) {
     int p1[2] = {0, 0};
     int *bmus = new int[nVectorsPerRank * 2];
 #ifdef _OPENMP
@@ -108,7 +109,17 @@ void trainOneEpochSparseCPU(int itask, svm_node **sparseData, float *numerator,
             }
         }
     }
-
+    if (only_bmus) {
+#ifdef HAVE_MPI
+        MPI_Gather(bmus, nVectorsPerRank * 2, MPI_INT, globalBmus, nVectorsPerRank * 2, MPI_INT, 0, MPI_COMM_WORLD);
+#else
+        for (unsigned int i = 0; i < 2 * nVectorsPerRank; ++i) {
+            globalBmus[i] = bmus[i];
+        }
+#endif
+        delete [] bmus;
+        return;
+    }
     float *localNumerator = new float[nSomY * nSomX * nDimensions];
     float *localDenominator = new float[nSomY * nSomX];
 #ifdef _OPENMP

diff --git a/src/training.cpp b/src/training.cpp
@@ -139,7 +139,6 @@ void train(int itask, float *data, svm_node **sparseData,
 #ifdef HAVE_MPI
         double epoch_time = MPI_Wtime();
 #endif
-
         trainOneEpoch(itask, data, sparseData, codebook, globalBmus,
                       nEpoch, currentEpoch,
                       nSomX, nSomY, nDimensions, nVectors, nVectorsPerRank,
@@ -322,7 +321,7 @@ void trainOneEpoch(int itask, float *data, svm_node **sparseData,
         trainOneEpochDenseGPU(itask, data, numerator, denominator,
                               codebook, nSomX, nSomY, nDimensions,
                               nVectors, nVectorsPerRank, radius, scale,
-                              mapType, gridType, compact_support, gaussian, globalBmus);
+                              mapType, gridType, compact_support, gaussian, globalBmus, only_bmus);
 #else
         my_abort("Compiled without CUDA!");
 #endif
@@ -332,7 +331,7 @@ void trainOneEpoch(int itask, float *data, svm_node **sparseData,
                                codebook, nSomX, nSomY, nDimensions,
                                nVectors, nVectorsPerRank, radius, scale,
                                mapType, gridType, compact_support, gaussian,
-                               globalBmus);
+                               globalBmus, only_bmus);
         break;
     }