Merge branch 'master' of github.com:liuq901/nerv into liuq901-master

7 files changed, 72 insertions, 3 deletions

diff --git a/nerv/lib/matrix/cumatrix.c b/nerv/lib/matrix/cumatrix.c
index aec4d60..43448bf 100644
--- a/nerv/lib/matrix/cumatrix.c
+++ b/nerv/lib/matrix/cumatrix.c
@@ -9,15 +9,19 @@
 void nerv_cuda_context_print_profile(CuContext *context) {
     HashMap *profile = context->profile;
     size_t i;
+    float tmp, tot = 0;
     fprintf(stderr, "*** [nerv cumatrix profile] **\n");
     for (i = 0; i < profile->size; i++)
     {
         HashNode *ptr;
         for (ptr = profile->bucket[i]; ptr; ptr = ptr->next)
         {
-            fprintf(stderr, "%s:\t%.6f\n", ptr->key, *(float *)ptr->val);
+            tmp = *(float *)ptr->val;
+            fprintf(stderr, "%s:\t%.6f\n", ptr->key, tmp);
+            tot += tmp;
         }
     }
+    fprintf(stderr, "Total time:\t%.6f\n", tot);
 }
 
 void nerv_cuda_context_clear_profile(CuContext *context) {
diff --git a/nerv/lib/matrix/generic/cukernel.cu b/nerv/lib/matrix/generic/cukernel.cu
index fc630ad..cf9d213 100644
--- a/nerv/lib/matrix/generic/cukernel.cu
+++ b/nerv/lib/matrix/generic/cukernel.cu
@@ -328,6 +328,15 @@ __global__ void cudak_(rearrange_frm)(const MATRIX_ELEM *a, MATRIX_ELEM *b,
     b[j + i * stride] = a[j / step + (j % step) * orig_dim + i * stride];
 }
 
+__global__ void cudak_(set_values_by_mask)(const MATRIX_ELEM *a, MATRIX_ELEM *b,
+                                        int nrow, int ncol,
+                                        int astride, int bstride, double val) {
+    int j = blockIdx.x * blockDim.x + threadIdx.x;
+    int i = blockIdx.y * blockDim.y + threadIdx.y;
+    if (i >= nrow || j >= ncol || a[i * astride] != 0.0) return;
+    b[j + i * bstride] = val;
+}
+
 __global__ void cudak_(scale_rows_by_col)(const MATRIX_ELEM *a, MATRIX_ELEM *b,
                                         int nrow, int ncol,
                                         int astride, int bstride) {
@@ -772,6 +781,19 @@ extern "C" {
         cudaStreamSynchronize(0);
     }
 
+    void cudak_(cuda_set_values_by_mask)(const Matrix *a, Matrix *b, double val) {
+        dim3 threadsPerBlock(CUDA_THREADS_N, CUDA_THREADS_N);
+        dim3 numBlocks(CEIL_DIV(b->ncol, threadsPerBlock.x),
+                CEIL_DIV(b->nrow, threadsPerBlock.y));
+        cudak_(set_values_by_mask)<<<numBlocks, threadsPerBlock>>> \
+            (MATRIX_ELEM_PTR(a), MATRIX_ELEM_PTR(b),
+            b->nrow, b->ncol,
+            a->stride / sizeof(MATRIX_ELEM),
+            b->stride / sizeof(MATRIX_ELEM),
+            val);
+        cudaStreamSynchronize(0);
+    }
+
     void cudak_(cuda_scale_rows_by_row)(const Matrix *a, Matrix *b) {
         dim3 threadsPerBlock(CUDA_THREADS_N, CUDA_THREADS_N);
         dim3 numBlocks(CEIL_DIV(b->ncol, threadsPerBlock.x),
diff --git a/nerv/lib/matrix/generic/cumatrix.c b/nerv/lib/matrix/generic/cumatrix.c
index 6d84663..bc5f285 100644
--- a/nerv/lib/matrix/generic/cumatrix.c
+++ b/nerv/lib/matrix/generic/cumatrix.c
@@ -515,7 +515,7 @@ void nerv_matrix_(prefixsum_row)(Matrix *a, const Matrix *b,
     NERV_SET_STATUS(status, NERV_NORMAL, 0);
 }
 
-void nerv_matrix_(diagonalize)(Matrix *a, CuContext * context, Status *status) {
+void nerv_matrix_(diagonalize)(Matrix *a, CuContext *context, Status *status) {
     if (a->nrow != a->ncol)
         NERV_EXIT_STATUS(status, MAT_MISMATCH_DIM, 0);
     PROFILE_START
@@ -524,6 +524,18 @@ void nerv_matrix_(diagonalize)(Matrix *a, CuContext * context, Status *status) {
     NERV_SET_STATUS(status, NERV_NORMAL, 0);
 }
 
+void nerv_matrix_(set_values_by_mask)(Matrix *a, const Matrix *b, double val,
+                                    CuContext *context, Status *status) {
+    if (a->nrow != b->nrow)
+        NERV_EXIT_STATUS(status, MAT_MISMATCH_DIM, 0);
+    if (b->ncol != 1)
+        NERV_EXIT_STATUS(status, MAT_COL_VECTOR_EXP, 0);
+    PROFILE_START
+    cudak_(cuda_set_values_by_mask)(b, a, val);
+    PROFILE_STOP
+    NERV_SET_STATUS(status, NERV_NORMAL, 0);
+}
+
 static void cuda_matrix_(free)(MATRIX_ELEM *ptr, CuContext *context, Status *status) {
     CUDA_SAFE_SYNC_CALL(cudaFree(ptr), status);
     NERV_SET_STATUS(status, NERV_NORMAL, 0);
diff --git a/nerv/lib/matrix/generic/cumatrix.h b/nerv/lib/matrix/generic/cumatrix.h
index de3a09e..79bfc76 100644
--- a/nerv/lib/matrix/generic/cumatrix.h
+++ b/nerv/lib/matrix/generic/cumatrix.h
@@ -35,6 +35,8 @@ void nerv_matrix_(fill)(Matrix *self, double val,
                         CuContext *context, Status *status);
 void nerv_matrix_(diagonalize)(Matrix *self,
                         CuContext *context, Status *status);
+void nerv_matrix_(set_values_by_mask)(Matrix *self, Matrix *mask, double val,
+                        CuContext *context, Status *status);
 void nerv_matrix_(copy_fromd)(Matrix *a, const Matrix *b,
                             int a_begin, int b_begin, int b_end,
                             CuContext *context, Status *status);
diff --git a/nerv/lib/matrix/generic/mmatrix.c b/nerv/lib/matrix/generic/mmatrix.c
index badddbd..e356de7 100644
--- a/nerv/lib/matrix/generic/mmatrix.c
+++ b/nerv/lib/matrix/generic/mmatrix.c
@@ -507,6 +507,29 @@ void nerv_matrix_(scale_rows_by_col)(Matrix *a, const Matrix *b,
     NERV_SET_STATUS(status, NERV_NORMAL, 0);
 }
 
+void nerv_matrix_(set_values_by_mask)(Matrix *a, const Matrix *b, double val, 
+                                    MContext *context, Status *status) {
+    if (a->nrow != b->nrow)
+        NERV_EXIT_STATUS(status, MAT_MISMATCH_DIM, 0);
+    if (b->ncol != 1)
+        NERV_EXIT_STATUS(status, MAT_COL_VECTOR_EXP, 0);
+    int i, j;
+    size_t astride = a->stride, bstride = b->stride;
+    MATRIX_ELEM *arow = MATRIX_ELEM_PTR(a),
+                *brow = MATRIX_ELEM_PTR(b);
+    for (i = 0; i < a->nrow; i++)
+    {
+        if (brow[0] == 0.0)
+        {
+            for (j = 0; j < a->ncol; j++)
+                arow[j] = val;
+        }
+        arow = MATRIX_NEXT_ROW_PTR(arow, astride);
+        brow = MATRIX_NEXT_ROW_PTR(brow, bstride);
+    }
+    NERV_SET_STATUS(status, NERV_NORMAL, 0);
+}
+
 static void host_matrix_(free)(MATRIX_ELEM *ptr, MContext *context, Status *status) {
     free(ptr);
     NERV_SET_STATUS(status, NERV_NORMAL, 0);
diff --git a/nerv/lib/matrix/generic/mmatrix.h b/nerv/lib/matrix/generic/mmatrix.h
index 6d17c99..41c39f6 100644
--- a/nerv/lib/matrix/generic/mmatrix.h
+++ b/nerv/lib/matrix/generic/mmatrix.h
@@ -48,6 +48,8 @@ void nerv_matrix_(expand_frm)(Matrix *a, const Matrix *b,
                             int cont, MContext *context, Status *status);
 void nerv_matrix_(rearrange_frm)(Matrix *a, const Matrix *b,
                                 int step, MContext *context, Status *status);
+void nerv_matrix_(set_values_by_mask)(Matrix *a, const Matrix *b, double val,
+                                    MContext *context, Status *status);
 void nerv_matrix_(scale_rows_by_col)(Matrix *a, const Matrix *b,
                                     MContext *context, Status *status);
 void nerv_matrix_(scale_rows_by_row)(Matrix *a, const Matrix *b,
diff --git a/nerv/lib/matrix/mmatrix.c b/nerv/lib/matrix/mmatrix.c
index e40b160..006735d 100644
--- a/nerv/lib/matrix/mmatrix.c
+++ b/nerv/lib/matrix/mmatrix.c
@@ -8,15 +8,19 @@
 void nerv_host_context_print_profile(MContext *context) {
     HashMap *profile = context->profile;
     size_t i;
+    float tmp, tot = 0;
     fprintf(stderr, "*** [nerv mmatrix profile] **\n");
     for (i = 0; i < profile->size; i++)
     {
         HashNode *ptr;
         for (ptr = profile->bucket[i]; ptr; ptr = ptr->next)
         {
-            fprintf(stderr, "%s:\t%.6f\n", ptr->key, *(float *)ptr->val);
+            tmp = *(float *)ptr->val;
+            fprintf(stderr, "%s:\t%.6f\n", ptr->key, tmp);
+            tot += tmp;
         }
     }
+    fprintf(stderr, "Total time:\t%.6f\n", tot);
 }
 
 void nerv_host_context_clear_profile(MContext *context) {