give a update_select_rows a more proper name

1 files changed, 27 insertions, 3 deletions

diff --git a/nerv/lib/matrix/generic/cukernel.cu b/nerv/lib/matrix/generic/cukernel.cu
index 552f7a4..9244783 100644
--- a/nerv/lib/matrix/generic/cukernel.cu
+++ b/nerv/lib/matrix/generic/cukernel.cu
@@ -262,7 +262,7 @@ __global__ void cudak_(clip)(MATRIX_ELEM *a,
 }
 
 #ifdef __NERV_FUTURE_CUDA_7
-__global__ void cudak_(update_select_rows)(MATRIX_ELEM *c, const MATRIX_ELEM *a, const MATRIX_ELEM *idx,
+__global__ void cudak_(update_select_rows_by_rowidx)(MATRIX_ELEM *c, const MATRIX_ELEM *a, const MATRIX_ELEM *idx,
                             int nrow_a, int ncol_a, int nrow_c, 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;
@@ -275,6 +275,20 @@ __global__ void cudak_(update_select_rows)(MATRIX_ELEM *c, const MATRIX_ELEM *a,
     //c[j + i_c * stride_c] = c[j + i_c * stride_c] * (1 - beta * alpha) + a[j + i * stride_a] * alpha; 
     atomicAdd_nvidia(c + j + i_c * stride_c, c[j + i_c * stride_c] * (- beta * alpha) + a[j + i * stride_a] * alpha);
 }
+
+__global__ void cudak_(update_select_rows_by_colidx)(MATRIX_ELEM *c, const MATRIX_ELEM *a, const MATRIX_ELEM *idx,
+                            int nrow_a, int ncol_a, int nrow_c, int stride_c, int stride_a, int stride_idx, 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[stride_idx * i]);
+    if (i_c < 0 || i_c >= nrow_c) {
+        printf("ERROR inside kernel update_select_rows, i_c(%d) out of range!", i_c);
+    }
+    //critical: i_c could conflict among threads(same index in the idx array), so atomicAdd is used
+    //c[j + i_c * stride_c] = c[j + i_c * stride_c] * (1 - beta * alpha) + a[j + i * stride_a] * alpha; 
+    atomicAdd_nvidia(c + j + i_c * stride_c, c[j + i_c * stride_c] * (- beta * alpha) + a[j + i * stride_a] * alpha);
+}
 #endif
 
 __global__ void cudak_(expand_frm)(const MATRIX_ELEM *a, MATRIX_ELEM *b,
@@ -640,16 +654,26 @@ extern "C" {
     }
     
 #ifdef __NERV_FUTURE_CUDA_7
-    void cudak_(cuda_update_select_rows)(Matrix *c, const Matrix *a, const Matrix *idx, double alpha, double beta) {
+    void cudak_(cuda_update_select_rows_by_rowidx)(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>>> \
+        cudak_(update_select_rows_by_rowidx)<<<numBlocks, threadsPerBlock>>> \
             (MATRIX_ELEM_PTR(c), MATRIX_ELEM_PTR(a), MATRIX_ELEM_PTR(idx), 
              a->nrow, a->ncol, c->nrow, c->stride / sizeof(MATRIX_ELEM), 
              a->stride / sizeof(MATRIX_ELEM), alpha, beta);
         cudaStreamSynchronize(0);
     }
+    void cudak_(cuda_update_select_rows_by_colidx)(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_by_colidx)<<<numBlocks, threadsPerBlock>>> \
+            (MATRIX_ELEM_PTR(c), MATRIX_ELEM_PTR(a), MATRIX_ELEM_PTR(idx), 
+             a->nrow, a->ncol, c->nrow, c->stride / sizeof(MATRIX_ELEM), 
+             a->stride / sizeof(MATRIX_ELEM), idx->stride / sizeof(MATRIX_ELEM), alpha, beta);
+        cudaStreamSynchronize(0);
+    }
 #endif
 
     void cudak_(cuda_expand_frm)(const Matrix *a, Matrix *b, int context) {