fastest

1 files changed, 14 insertions, 14 deletions

diff --git a/src/amun/gpu/mblas/matrix_functions.h b/src/amun/gpu/mblas/matrix_functions.h
index c162910f..bd791797 100644
--- a/src/amun/gpu/mblas/matrix_functions.h
+++ b/src/amun/gpu/mblas/matrix_functions.h
@@ -151,16 +151,16 @@ __global__ void gBroadcast(Functor functor,
                            const MatrixWrapper<float> in2Wrap,
                            const VectorWrapper<uint> batchMappingWrap)
 {
-  //size_t srcSize = outWrap.dim(0);
-  //size_t inRows = in2Wrap.dim(0);
-  //size_t cols  = in1Wrap.dim(1);
+  uint srcSize = outWrap.dim(0);
+  uint inRows = in2Wrap.dim(0);
+  uint cols  = in1Wrap.dim(1);
 
   int id = threadIdx.x + blockIdx.x * blockDim.x;
   if (id < outWrap.size()) {
+    /*
     uint indices[SHAPE_SIZE];
     outWrap.id2Indices(id, indices);
 
-    /*
     int row = id / cols; // len * batch for in1
     int srcId = row % srcSize;  // source pos for in1
 
@@ -171,25 +171,25 @@ __global__ void gBroadcast(Functor functor,
                           in2Wrap(batchMappingIdx, indices[1], 0, 0) );
     */
 
-    //int row = id / cols;
-    //int stateIdx = id % cols;
-
-    //int beamIdx = row / srcSize;
-    //int srcId = row % srcSize;
+    uint row = id / cols;
+    uint stateIdx = id % cols;
+    uint beamIdx = row / srcSize;
+    uint srcId = row % srcSize;
+    uint batchIdx = batchMappingWrap[ beamIdx ];
 
-    uint batchIdx = batchMappingWrap[ indices[2] ];
+    //uint batchIdx = batchMappingWrap[ indices[2] ];
 
     //assert(srcId == indices[0]);
     //assert(stateIdx == indices[1]);
     //assert(beamIdx == indices[2]);
     //assert(0 == indices[3]);
 
-    //outWrap[id] = functor(in1Wrap[(batchIdx * srcSize + srcId) * cols + stateIdx],
-    //                      in2Wrap[beamIdx * cols + stateIdx]);
+    outWrap[id] = functor(in1Wrap[(batchIdx * srcSize + srcId) * cols + stateIdx],
+                          in2Wrap[beamIdx * cols + stateIdx]);
     //outWrap[id] = functor(in1Wrap(indices[0], indices[1], 0, batchIdx),
     //                      in2Wrap(indices[2], indices[1], 0, 0));
-    outWrap(indices[0], indices[1], indices[2], 0) = functor(in1Wrap(indices[0], indices[1], 0, batchIdx),
-                                                  in2Wrap(indices[2], indices[1], 0, 0));
+    //outWrap(srcId, stateIdx, beamIdx, 0) = functor(in1Wrap(srcId, stateIdx, 0, batchIdx),
+    //                                              in2Wrap(beamIdx, stateIdx, 0, 0));
   }
 }