1 files changed, 59 insertions, 29 deletions

diff --git a/nerv/lib/matrix/cumatrix.c b/nerv/lib/matrix/cumatrix.c
index d998871..2fbe7d8 100644
--- a/nerv/lib/matrix/cumatrix.c
+++ b/nerv/lib/matrix/cumatrix.c
@@ -1,23 +1,12 @@
 #define NERV_GENERIC_CUMATRIX
+#define MATRIX_CONTEXT CuContext
 #include "cumatrix.h"
 #include "cuda_helper.h"
 #include <string.h>
 #include <time.h>
-#define PROFILE_HASHMAP_SIZE 123457
-static cublasHandle_t cublas_handle;
-static cudaEvent_t profile_start, profile_stop;
-curandGenerator_t curand_gen;
-static HashMap *profile;
 
-void nerv_cumatrix_select_gpu(int dev, Status *status) {
-    fprintf(stderr, "** selecting GPU %d\n", dev);
-    NERV_SET_STATUS(status, NERV_NORMAL, 0);
-    CUDA_SAFE_SYNC_CALL(cudaSetDevice(dev), status);
-    CUDA_SAFE_SYNC_CALL(cublasDestroy(cublas_handle), status);
-    CUDA_SAFE_SYNC_CALL(cublasCreate(&cublas_handle), status);
-}
-
-void nerv_cumatrix_print_profile() {
+void nerv_cuda_context_print_profile(CuContext *context) {
+    HashMap *profile = context->profile;
     size_t i;
     fprintf(stderr, "*** [nerv cumatrix profile] **\n");
     for (i = 0; i < profile->size; i++)
@@ -30,28 +19,72 @@ void nerv_cumatrix_print_profile() {
     }
 }
 
-void nerv_cumatrix_clear_profile() {
-    hashmap_clear(profile);
+void nerv_cuda_context_clear_profile(CuContext *context) {
+    nerv_hashmap_clear(context->profile);
 }
 
-void accu_profile(const char *name, float delta) {
-    float *val = hashmap_getval(profile, name);
+void nerv_cuda_context_accu_profile(CuContext *context,
+                                    const char *name, float delta) {
+    HashMap *profile = context->profile;
+    float *val = nerv_hashmap_getval(profile, name);
     if (!val)
     {
         val = malloc(sizeof(float));
         *val = 0;
-        hashmap_setval(profile, name, val);
+        nerv_hashmap_setval(profile, name, val);
     }
     *val += delta;
 }
 
-void nerv_cumatrix_init() {
-    cublasCreate(&cublas_handle);
-    curandCreateGenerator(&curand_gen, CURAND_RNG_PSEUDO_DEFAULT);
-    curandSetPseudoRandomGeneratorSeed(curand_gen, time(NULL));
-    cudaEventCreate(&profile_start);
-    cudaEventCreate(&profile_stop);
-    profile = hashmap_create(PROFILE_HASHMAP_SIZE, bkdr_hash, strcmp);
+static void new_cuda_handles(CuContext *context, Status *status) {
+    CUBLAS_SAFE_SYNC_CALL(cublasCreate(&(context->cublas_handle)), status);
+    CURAND_SAFE_SYNC_CALL(curandCreateGenerator(&(context->curand_gen),
+                            CURAND_RNG_PSEUDO_DEFAULT), status);
+    CURAND_SAFE_SYNC_CALL(
+            curandSetPseudoRandomGeneratorSeed(context->curand_gen, time(NULL)),
+            status);
+    CUDA_SAFE_SYNC_CALL(cudaEventCreate(&(context->profile_start)), status);
+    CUDA_SAFE_SYNC_CALL(cudaEventCreate(&(context->profile_stop)), status);
+    NERV_SET_STATUS(status, NERV_NORMAL, 0);
+}
+
+static void free_cuda_handles(CuContext *context, Status *status) {
+    CUBLAS_SAFE_SYNC_CALL(cublasDestroy(context->cublas_handle), status);
+    CURAND_SAFE_SYNC_CALL(curandDestroyGenerator(context->curand_gen), status);
+    CUDA_SAFE_SYNC_CALL(cudaEventDestroy(context->profile_start), status);
+    CUDA_SAFE_SYNC_CALL(cudaEventDestroy(context->profile_stop), status);
+    NERV_SET_STATUS(status, NERV_NORMAL, 0);
+}
+
+CuContext *nerv_cuda_context_create(Status *status) {
+    CuContext *context = (CuContext *)malloc(sizeof(CuContext));
+    new_cuda_handles(context, status);
+    if (status->err_code != NERV_NORMAL)
+        return NULL;
+    context->profile = nerv_hashmap_create(PROFILE_HASHMAP_SIZE, bkdr_hash, strcmp);
+    NERV_SET_STATUS(status, NERV_NORMAL, 0);
+    return context;
+}
+
+void nerv_cuda_context_destroy(CuContext *context, Status *status) {
+    free_cuda_handles(context, status);
+    if (status->err_code != NERV_NORMAL)
+        return;
+    nerv_hashmap_destroy(context->profile);
+    free(context);
+    NERV_SET_STATUS(status, NERV_NORMAL, 0);
+}
+
+void nerv_cuda_context_select_gpu(CuContext *context,
+                                    int dev, Status *status) {
+    free_cuda_handles(context, status);
+    if (status->err_code != NERV_NORMAL)
+        return;
+    CUDA_SAFE_SYNC_CALL(cudaSetDevice(dev), status);
+    new_cuda_handles(context, status);
+    if (status->err_code != NERV_NORMAL)
+        return;
+    NERV_SET_STATUS(status, NERV_NORMAL, 0);
 }
 
 #define MATRIX_USE_FLOAT
@@ -59,7 +92,6 @@ void nerv_cumatrix_init() {
 #define nerv_matrix_(NAME) nerv_matrix_cuda_float_##NAME
 #define cudak_(NAME) cudak_float_ ## NAME
 #define NERV_CUBLAS_(NAME) cublasS##NAME
-#define MATRIX_CUMATRIX_HOST_TNAME nerv_matrix_host_float_tname
 #include "generic/cumatrix.c"
 
 #undef NERV_CUBLAS_
@@ -72,12 +104,10 @@ void nerv_cumatrix_init() {
 #undef MATRIX_ELEM_PTR_BASE
 #undef MATRIX_ELEM_FMT
 #undef MATRIX_ELEM_WRITE_FMT
-#undef MATRIX_CUMATRIX_HOST_TNAME
 
 #define MATRIX_USE_DOUBLE
 #define cuda_matrix_(NAME) cuda_matrix_double_##NAME
 #define nerv_matrix_(NAME) nerv_matrix_cuda_double_##NAME
 #define cudak_(NAME) cudak_double_ ## NAME
 #define NERV_CUBLAS_(NAME) cublasD##NAME
-#define MATRIX_CUMATRIX_HOST_TNAME nerv_matrix_host_double_tname
 #include "generic/cumatrix.c"