working cuda -> lbfgs

3 files changed, 41 insertions, 16 deletions

diff --git a/generic/lbfgs.c b/generic/lbfgs.c
index ffe7bd0..7e39f7e 100644
--- a/generic/lbfgs.c
+++ b/generic/lbfgs.c
@@ -11,8 +11,7 @@ int THTensor_(copy_evaluate_start)(THTensor *parameters,
                                    const int nParameter)
 {
   THDoubleStorage *xs = THDoubleStorage_newWithData((double *)x,nParameter);
-  THStorage *ps = 
-    THStorage_(newWithData)(THTensor_(data)(parameters),nParameter);
+  THStorage *ps = THTensor_(storage)(parameters);
 
  
   /* copy given x (xs) -> parameters (ps) */
@@ -21,8 +20,6 @@ int THTensor_(copy_evaluate_start)(THTensor *parameters,
   /* only want to free the struct part of the storage not the data */
   xs->data = NULL;
   THDoubleStorage_free(xs);
-  ps->data = NULL;
-  THStorage_(free)(ps);
   return 0;
 }
 
@@ -31,8 +28,7 @@ int THTensor_(copy_evaluate_end)(lbfgsfloatval_t *g,
                                  const int nParameter)
 {
   THDoubleStorage *gs = THDoubleStorage_newWithData((double *)g,nParameter);
-  THStorage *gps = 
-    THStorage_(newWithData)(THTensor_(data)(gradParameters), nParameter);
+  THStorage *gps = THTensor_(storage)(gradParameters);
   
   /* copy gradParameters (gps) -> g (gs) */
 #ifdef TH_REAL_IS_FLOAT
@@ -47,8 +43,6 @@ int THTensor_(copy_evaluate_end)(lbfgsfloatval_t *g,
   /* only want to free the struct part of the storage not the data */
   gs->data = NULL;
   THDoubleStorage_free(gs);
-  gps->data = NULL;
-  THStorage_(free)(gps);
 
   return 0;
 }
@@ -59,8 +53,7 @@ int THTensor_(copy_init)(lbfgsfloatval_t *x,
                          const int nParameter) 
 {
   THDoubleStorage *xs = THDoubleStorage_newWithData((double *)x,nParameter);
-  THStorage *ps = 
-    THStorage_(newWithData)(THTensor_(data)(parameters),nParameter);
+  THStorage *ps = THTensor_(storage)(parameters);
 
   /* copy given parameters (ps) -> x (xs) */
 #ifdef TH_REAL_IS_FLOAT
@@ -75,9 +68,6 @@ int THTensor_(copy_init)(lbfgsfloatval_t *x,
   /* only want to free the struct part of the storage not the data */
   xs->data = NULL;
   THDoubleStorage_free(xs);
-  ps->data = NULL;
-  THStorage_(free)(ps);
-  printf("in copy_init : freed storage\n");
 
   /* done */
   return 0;
diff --git a/lbfgs.c b/lbfgs.c
index b9e1137..11f871e 100644
--- a/lbfgs.c
+++ b/lbfgs.c
@@ -66,7 +66,7 @@
 #endif
 
 #ifdef WITH_CUDA
-#include "THC.h"
+#include <THC/THC.h>
 #endif
 
 #include "TH.h"
@@ -1549,7 +1549,7 @@ int lbfgs_init(lua_State *L) {
     THFloatTensor_copy_init(x,(THFloatTensor *)parameters,nParameter);
 #ifdef WITH_CUDA
   else if ( current_torch_type = torch_CudaTensor_id ) 
-    THCudaTensor_copy_init(x,(THDoubleTensor *)parameters,nParameter);
+    THCudaTensor_copy_init(x,(THCudaTensor *)parameters,nParameter);
 #endif  
 
   /* initialize the parameters for the L-BFGS optimization */
diff --git a/nnx-1.0-1.rockspec b/nnx-1.0-1.rockspec
index 1b90f21..2650547 100644
--- a/nnx-1.0-1.rockspec
+++ b/nnx-1.0-1.rockspec
@@ -49,10 +49,45 @@ build = {
 
          set (CMAKE_INSTALL_RPATH_USE_LINK_PATH TRUE)
 
-         include_directories (${TORCH_INCLUDE_DIR} ${PROJECT_SOURCE_DIR})
+	include_directories (${TORCH_INCLUDE_DIR} ${PROJECT_SOURCE_DIR})
+
+	 SET(WITH_CUDA OFF CACHE BOOL "Compile qlua and associated packages")
+
+	IF(EXISTS "${TORCH_BIN_DIR}/../include/THC/THC.h")
+	FIND_PACKAGE(CUDA 4.0)
+	
+	IF (CUDA_FOUND)
+	   # bug on Apple
+	   IF(APPLE)
+	     LINK_DIRECTORIES("/usr/local/cuda/lib/")
+	   ENDIF(APPLE)
+
+	find_library (TORCH_THC THC ${TORCH_BIN_DIR}/../lib NO_DEFAULT_PATH)
+
+	
+	set (TORCH_LIBRARIES  ${TORCH_TH} ${TORCH_THC} ${TORCH_luaT} ${TORCH_lua})
+	set (CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DWITH_CUDA")
+
+         cuda_add_library (lbfgs SHARED lbfgs.c)
+         target_link_libraries (lbfgs ${TORCH_LIBRARIES})
+	 CUDA_ADD_CUBLAS_TO_TARGET(lbfgs)
+
+	ELSE (CUDA_FOUND)
+
+	    MESSAGE(STATUS "Disabling CUDA (CUDA 4.0 required, and not found)")
+	    MESSAGE(STATUS "If CUDA 4.0 is installed, then specify CUDA_TOOLKIT_ROOT_DIR")
+
+	    set (TORCH_LIBRARIES ${TORCH_TH} ${TORCH_luaT} ${TORCH_lua})
+
          add_library (lbfgs SHARED lbfgs.c)
          target_link_libraries (lbfgs ${TORCH_LIBRARIES})
+
+	ENDIF (CUDA_FOUND)
+	ENDIF(EXISTS "${TORCH_BIN_DIR}/../include/THC/THC.h")
+
          install_targets (/lib lbfgs)
+ 
+	
 
          include_directories (${TORCH_INCLUDE_DIR} ${PROJECT_SOURCE_DIR})
          add_library (nnx SHARED init.c)