...

Merge remote-tracking branch 'upstream/master'

8 files changed, 53 insertions, 35 deletions

diff --git a/matrix/cuda_helper.h b/matrix/cuda_helper.h
index d6effdb..fde6f18 100644
--- a/matrix/cuda_helper.h
+++ b/matrix/cuda_helper.h
@@ -62,17 +62,14 @@ static const char *cublasGetErrorString(cublasStatus_t err) {
 
 #define PROFILE_START \
     do { \
-        cudaEvent_t start, stop; \
-        cudaEventCreate(&start); \
-        cudaEventCreate(&stop); \
-        cudaEventRecord(start, 0);
+        cudaEventRecord(profile_start, 0);
 #define PROFILE_STOP \
-        cudaEventRecord(stop, 0); \
-        cudaEventSynchronize(stop); \
+        cudaEventRecord(profile_stop, 0); \
+        cudaEventSynchronize(profile_stop); \
         float milliseconds = 0; \
-        cudaEventElapsedTime(&milliseconds, start, stop); \
+        cudaEventElapsedTime(&milliseconds, profile_start, profile_stop); \
         accu_profile(__func__, milliseconds / 1000); \
     } while (0);
 
-#define PROFILE_END 
+#define PROFILE_END
 #endif
diff --git a/matrix/cukernel.cu b/matrix/cukernel.cu
index fbac369..a19030a 100644
--- a/matrix/cukernel.cu
+++ b/matrix/cukernel.cu
@@ -9,6 +9,7 @@
 #undef MATRIX_ELEM
 #undef MATRIX_ELEM_PTR
 #undef MATRIX_ELEM_FMT
+#undef MATRIX_ELEM_WRITE_FMT
 
 #define cudak_(NAME) cudak_double_ ## NAME
 #define MATRIX_USE_DOUBLE
diff --git a/matrix/cumatrix.c b/matrix/cumatrix.c
index 4ebc5ff..af34fb4 100644
--- a/matrix/cumatrix.c
+++ b/matrix/cumatrix.c
@@ -1,10 +1,14 @@
 #define NERV_GENERIC_CUMATRIX
 #include "../common.h"
 #include "cuda_helper.h"
+#include <string.h>
+#define PROFILE_HASHMAP_SIZE 123457
 static cublasHandle_t cublas_handle;
+static cudaEvent_t profile_start, profile_stop;
 static HashMap *profile;
 
-int print_profile(lua_State *L) {
+static int print_profile(lua_State *L) {
+    (void)L;
     size_t i;
     fprintf(stderr, "*** [nerv cumatrix profile] **\n");
     for (i = 0; i < profile->size; i++)
@@ -18,7 +22,8 @@ int print_profile(lua_State *L) {
     return 0;
 }
 
-int clear_profile(lua_State *L) {
+static int clear_profile(lua_State *L) {
+    (void)L;
     hashmap_clear(profile);
     return 0;
 }
@@ -34,6 +39,25 @@ void accu_profile(const char *name, float delta) {
     *val += delta;
 }
 
+static const luaL_Reg cumatrix_methods[] = {
+    {"print_profile", print_profile},
+    {"clear_profile", clear_profile},
+    {NULL, NULL}
+};
+
+extern void nerv_matrix_cuda_float_init(lua_State *L);
+extern void nerv_matrix_cuda_double_init(lua_State *L);
+
+void nerv_cumatrix_init(lua_State *L) {
+    luaL_register(L, NULL, cumatrix_methods);
+    cublasCreate(&cublas_handle);
+    cudaEventCreate(&profile_start);
+    cudaEventCreate(&profile_stop);
+    profile = hashmap_create(PROFILE_HASHMAP_SIZE, bkdr_hash, strcmp);
+    nerv_matrix_cuda_float_init(L);
+    nerv_matrix_cuda_double_init(L);
+}
+
 #define MATRIX_USE_FLOAT
 #define cuda_matrix_(NAME) cuda_matrix_float_##NAME
 #define nerv_matrix_(NAME) nerv_matrix_cuda_float_##NAME
@@ -50,6 +74,7 @@ const char *nerv_matrix_(tname) = "nerv.CuMatrixFloat";
 #undef MATRIX_ELEM
 #undef MATRIX_ELEM_PTR
 #undef MATRIX_ELEM_FMT
+#undef MATRIX_ELEM_WRITE_FMT
 #undef MATRIX_CUMATRIX_HOST_TNAME
 
 #define MATRIX_USE_DOUBLE
diff --git a/matrix/generic/cumatrix.c b/matrix/generic/cumatrix.c
index 956e1e6..a8e18e0 100644
--- a/matrix/generic/cumatrix.c
+++ b/matrix/generic/cumatrix.c
@@ -11,11 +11,9 @@
 #define MATRIX_BASE_TNAME nerv_matrix_cuda_tname
 #define NERV_GENERIC_MATRIX
 #define NERV_GENERIC_CUKERNEL
-#define PROFILE_HASHMAP_SIZE 123457
 #include "../../common.h"
 #include "../cukernel.h"
 #include "../cuda_helper.h"
-#include <string.h>
 
 Matrix *nerv_matrix_(new_)(lua_State *L, long nrow, long ncol);
 void nerv_matrix_(data_free)(lua_State *L, Matrix *self);
@@ -442,8 +440,6 @@ static const luaL_Reg nerv_matrix_(extra_methods)[] = {
 
 static void cuda_matrix_(init)(lua_State *L) {
     luaN_append_methods(L, nerv_matrix_(extra_methods));
-    cublasCreate(&cublas_handle);
-    profile = hashmap_create(PROFILE_HASHMAP_SIZE, bkdr_hash, strcmp);
 }
 
 static void cuda_matrix_(free)(lua_State *L, MATRIX_ELEM *ptr) {
diff --git a/matrix/generic/matrix.c b/matrix/generic/matrix.c
index d6b0aea..e17fb42 100644
--- a/matrix/generic/matrix.c
+++ b/matrix/generic/matrix.c
@@ -6,6 +6,7 @@ extern const char *nerv_matrix_(tname);
 extern const char *MATRIX_BASE_TNAME;
 
 void nerv_matrix_(data_free)(lua_State *L, Matrix *self) {
+    (void)L;
     assert(*self->data_ref > 0);
     if (--(*self->data_ref) == 0)
     {
diff --git a/matrix/generic/mmatrix.c b/matrix/generic/mmatrix.c
index 2045d65..b0f0791 100644
--- a/matrix/generic/mmatrix.c
+++ b/matrix/generic/mmatrix.c
@@ -87,15 +87,9 @@ int nerv_matrix_(save)(lua_State *L) {
         MATRIX_ELEM *row = MATRIX_ROW_PTR(self, i);
         for (j = 0; j < ncol; j++)
             if (fprintf(fp, MATRIX_ELEM_WRITE_FMT " ", row[j]) < 0)
-            {
-                free(self);
                 return 0;
-            }
         if (fprintf(fp, "\n") < 0)
-        {
-             free(self);
              return 0;
-        }
     }
     return 0;
 }
diff --git a/matrix/init.c b/matrix/init.c
index 7b7f478..c29d7e9 100644
--- a/matrix/init.c
+++ b/matrix/init.c
@@ -5,21 +5,14 @@ const char *nerv_matrix_tname = "nerv.Matrix";
 const char *nerv_matrix_cuda_tname = "nerv.CuMatrix";
 const char *nerv_matrix_host_tname = "nerv.MMatrix";
 
-void nerv_matrix_host_float_init(lua_State *L);
-void nerv_matrix_cuda_float_init(lua_State *L);
-void nerv_matrix_host_double_init(lua_State *L);
-void nerv_matrix_cuda_double_init(lua_State *L);
-void nerv_matrix_host_int_init(lua_State *L);
-int print_profile(lua_State *L);
-int clear_profile(lua_State *L);
+void nerv_cumatrix_init(lua_State *L);
+void nerv_mmatrix_init(lua_State *L);
 
 static const luaL_Reg matrix_methods[] = {
     {"__tostring__", nerv_error_method_not_implemented },
     {"__add__", nerv_error_method_not_implemented },
     {"__sub__", nerv_error_method_not_implemented },
     {"__mul__", nerv_error_method_not_implemented },
-    {"print_profile", print_profile},
-    {"clear_profile", clear_profile},
     {NULL, NULL}
 };
 
@@ -32,13 +25,11 @@ void nerv_matrix_init(lua_State *L) {
     /* CuMatrix inherits from Matrix */
     luaT_newmetatable(L, nerv_matrix_cuda_tname, nerv_matrix_tname,
                             NULL, NULL, NULL);
-    nerv_matrix_cuda_float_init(L);
-    nerv_matrix_cuda_double_init(L);
-
+    nerv_cumatrix_init(L);
+    lua_pop(L, 1);
     /* MMatrix inherits from Matrix */
     luaT_newmetatable(L, nerv_matrix_host_tname, nerv_matrix_tname,
                             NULL, NULL, NULL);
-    nerv_matrix_host_float_init(L);
-    nerv_matrix_host_double_init(L);
-    nerv_matrix_host_int_init(L);
+    nerv_mmatrix_init(L);
+    lua_pop(L, 1);
 }
diff --git a/matrix/mmatrix.c b/matrix/mmatrix.c
index 81f8dfc..ffc058d 100644
--- a/matrix/mmatrix.c
+++ b/matrix/mmatrix.c
@@ -1,4 +1,15 @@
 #define NERV_GENERIC_MMATRIX
+#include "../common.h"
+void nerv_matrix_host_float_init(lua_State *L);
+void nerv_matrix_host_double_init(lua_State *L);
+void nerv_matrix_host_int_init(lua_State *L);
+
+void nerv_mmatrix_init(lua_State *L) {
+    nerv_matrix_host_float_init(L);
+    nerv_matrix_host_double_init(L);
+    nerv_matrix_host_int_init(L);
+}
+
 #define MATRIX_USE_FLOAT
 #define host_matrix_(NAME) host_matrix_float_##NAME
 #define nerv_matrix_(NAME) nerv_matrix_host_float_##NAME
@@ -10,6 +21,7 @@ const char *nerv_matrix_(tname) = "nerv.MMatrixFloat";
 #undef MATRIX_ELEM
 #undef MATRIX_ELEM_PTR
 #undef MATRIX_ELEM_FMT
+#undef MATRIX_ELEM_WRITE_FMT
 
 #define NERV_GENERIC_MMATRIX
 #define MATRIX_USE_DOUBLE
@@ -23,6 +35,7 @@ const char *nerv_matrix_(tname) = "nerv.MMatrixDouble";
 #undef MATRIX_ELEM
 #undef MATRIX_ELEM_PTR
 #undef MATRIX_ELEM_FMT
+#undef MATRIX_ELEM_WRITE_FMT
 
 #define NERV_GENERIC_MMATRIX
 #define MATRIX_USE_INT