added update_select_rows for select_linear:update speed-up

4 files changed, 34 insertions, 0 deletions

diff --git a/nerv/lib/matrix/cukernel.h b/nerv/lib/matrix/cukernel.h
index 2126c6f..fffe0bc 100644
--- a/nerv/lib/matrix/cukernel.h
+++ b/nerv/lib/matrix/cukernel.h
@@ -13,6 +13,7 @@ void cudak_(cuda_softmax_final)(const Matrix *a, const Matrix *max, const Matrix
 void cudak_(cuda_add_row)(const Matrix *a, Matrix *b, double beta);
 void cudak_(cuda_fill)(Matrix *a, double val);
 void cudak_(cuda_clip)(Matrix *a, double val_1, double val_2);
+void cudak_(cuda_update_select_rows)(Matrix *c, const Matrix *a, const Matrix *idx, double alpha, double beta);
 void cudak_(cuda_expand_frm)(const Matrix *a, Matrix *b, int context);
 void cudak_(cuda_rearrange_frm)(const Matrix *a, Matrix *b, int step);
 void cudak_(cuda_scale_rows_by_row)(const Matrix *a, Matrix *b);
diff --git a/nerv/lib/matrix/generic/cukernel.cu b/nerv/lib/matrix/generic/cukernel.cu
index 08feb59..6c8e64a 100644
--- a/nerv/lib/matrix/generic/cukernel.cu
+++ b/nerv/lib/matrix/generic/cukernel.cu
@@ -225,6 +225,15 @@ __global__ void cudak_(clip)(MATRIX_ELEM *a,
             a[j + i * stride] = val_1;
 }
 
+__global__ void cudak_(update_select_rows)(MATRIX_ELEM *c, const MATRIX_ELEM *a, const MATRIX_ELEM *idx,
+                            int nrow_a, int ncol_a, int stride_c, int stride_a, double alpha, double beta) {
+    int j = blockIdx.x * blockDim.x + threadIdx.x;
+    int i = blockIdx.y * blockDim.y + threadIdx.y;
+    if (i >= nrow_a || j >= ncol_a) return;
+    int i_c = lrintf(idx[i]);
+    c[j + i_c * stride_c] = c[j + i_c * stride_c] * (1 - beta * alpha) + a[j + i * stride_a] * alpha;
+}
+
 __global__ void cudak_(expand_frm)(const MATRIX_ELEM *a, MATRIX_ELEM *b,
                                     int nrow, int ncol,
                                     int enrow, int encol,
@@ -540,6 +549,17 @@ extern "C" {
             a->stride / sizeof(MATRIX_ELEM), val_1, val_2);
         cudaStreamSynchronize(0);
     }
+    
+    void cudak_(cuda_update_select_rows)(Matrix *c, const Matrix *a, const Matrix *idx, double alpha, double beta) {
+        dim3 threadsPerBlock(CUDA_THREADS_N, CUDA_THREADS_N);
+        dim3 numBlocks(CEIL_DIV(a->ncol, threadsPerBlock.x),
+                CEIL_DIV(a->nrow, threadsPerBlock.y));
+        cudak_(update_select_rows)<<<numBlocks, threadsPerBlock>>> \
+            (MATRIX_ELEM_PTR(c), MATRIX_ELEM_PTR(a), MATRIX_ELEM_PTR(idx), 
+             a->nrow, a->ncol, c->stride / sizeof(MATRIX_ELEM), 
+             a->stride / sizeof(MATRIX_ELEM), alpha, beta);
+        cudaStreamSynchronize(0);
+    }
 
     void cudak_(cuda_expand_frm)(const Matrix *a, Matrix *b, int context) {
         dim3 threadsPerBlock(CUDA_THREADS_N, CUDA_THREADS_N);
diff --git a/nerv/lib/matrix/generic/cumatrix.c b/nerv/lib/matrix/generic/cumatrix.c
index 770e503..2dc5899 100644
--- a/nerv/lib/matrix/generic/cumatrix.c
+++ b/nerv/lib/matrix/generic/cumatrix.c
@@ -359,6 +359,18 @@ void nerv_matrix_(copy_rows_fromd_by_idx)(Matrix *a, const Matrix *b,
     NERV_SET_STATUS(status, NERV_NORMAL, 0);
 }
 
+void nerv_matrix_(update_select_rows)(Matrix *c, const Matrix *a, const Matrix *idx, double alpha, double beta, Status *status) {
+    long nrow = a->nrow;
+    if (idx->nrow != 1)
+        NERV_EXIT_STATUS(status, MAT_IDX_VECTOR_EXP, 0);
+    if (a->ncol != c->ncol)
+        NERV_EXIT_STATUS(status, MAT_MISMATCH_DIM, 0);
+    PROFILE_START
+    cudak_(cuda_update_select_rows)(c, a, idx, alpha, beta);
+    PROFILE_STOP
+    NERV_SET_STATUS(status, NERV_NORMAL, 0);
+}
+
 void nerv_matrix_(expand_frm)(Matrix *a, const Matrix *b,
                             int context, Status *status) {
     if (a->nrow != b->nrow)
diff --git a/nerv/lib/matrix/generic/cumatrix.h b/nerv/lib/matrix/generic/cumatrix.h
index 04e8c5a..21c29b7 100644
--- a/nerv/lib/matrix/generic/cumatrix.h
+++ b/nerv/lib/matrix/generic/cumatrix.h
@@ -42,6 +42,7 @@ void nerv_matrix_(copy_rows_fromh_by_idx)(Matrix *a, const Matrix *b,
                             const Matrix *idx, int b_begin, Status *status);
 void nerv_matrix_(copy_rows_fromd_by_idx)(Matrix *a, const Matrix *b,
                             const Matrix *idx, int b_begin, Status *status);
+void nerv_matrix_(update_select_rows)(Matrix *c, const Matrix *a, const Matrix *idx, double alpha, double beta, Status *status);
 
 void nerv_matrix_(expand_frm)(Matrix *a, const Matrix *b,
                             int context, Status *status);