Changed dot() so it doesn't call cublasSdot anymore since cublasSdot gives the

wrong outputs sometimes.

dlib/dnn/cublas_dlibapi.cpp

@@ -7,7 +7,6 @@
 
 #include "cublas_dlibapi.h"
 #include "cuda_utils.h"
-#include "tensor.h"
 
 #include <cublas_v2.h>
 
@@ -89,23 +88,6 @@ namespace dlib
             return c.get_handle();
         }
 
-    // -----------------------------------------------------------------------------------
-
-        float dot (
-            const tensor& a,
-            const tensor& b
-        )
-        {
-            DLIB_CASSERT(a.size() == b.size(), "");
-            float result = 0;
-            CHECK_CUBLAS(cublasSdot(context(), 
-                                    a.size(),
-                                    a.device(), 1,
-                                    b.device(), 1,
-                                    &result));
-            return result;
-        }
-
     // -----------------------------------------------------------------------------------
 
         void gemm (

dlib/dnn/cublas_dlibapi.h

@@ -5,30 +5,14 @@
 
 #ifdef DLIB_USE_CUDA
 
+#include "tensor.h"
 #include "cuda_errors.h"
 
 namespace dlib
 {
-    class tensor;
-
     namespace cuda 
     {
 
-    // -----------------------------------------------------------------------------------
-
-        float dot (
-            const tensor& a,
-            const tensor& b
-        );
-        /*!
-            requires
-                - a.size() == b.size()
-            ensures
-                - returns the dot product between a and b when they are both treated as
-                  a.size() dimensional vectors.  That is, this function pointwise
-                  multiplies the vectors together, then sums the result and returns it.
-        !*/
-
     // -----------------------------------------------------------------------------------
 
         void gemm (

dlib/dnn/tensor.h

@@ -7,7 +7,6 @@
 #include <cstring>
 #include "../matrix.h"
 #include "cudnn_dlibapi.h"
-#include "cublas_dlibapi.h"
 #include "gpu_data.h"
 #include <memory>
 
@@ -408,9 +407,7 @@ namespace dlib
         const tensor& b
     )
     {
-#ifdef DLIB_USE_CUDA
-        return cuda::dot(a,b);
-#else
+        // TODO, do on GPU?
         DLIB_CASSERT(a.size() == b.size(), "");
         const float* da = a.host();
         const float* db = b.host();
@@ -418,7 +415,6 @@ namespace dlib
         for (size_t i = 0; i < a.size(); ++i)
             sum += da[i]*db[i];
         return sum;
-#endif
     }
 
 // ----------------------------------------------------------------------------------------