support gpu buffering

1 files changed, 20 insertions, 0 deletions

diff --git a/nerv/lib/matrix/generic/cukernel.cu b/nerv/lib/matrix/generic/cukernel.cu
index e337798..08feb59 100644
--- a/nerv/lib/matrix/generic/cukernel.cu
+++ b/nerv/lib/matrix/generic/cukernel.cu
@@ -284,6 +284,15 @@ __global__ void cudak_(gen_col_idx)(MATRIX_ELEM *b,
     b[j + i * stride] = j;
 }
 
+__global__ void cudak_(copy_rows_by_idx)(const MATRIX_ELEM *a, MATRIX_ELEM *b,
+                                    const MATRIX_ELEM *idx, int b_begin,
+                                    int nrow, int ncol, int stride) {
+    int j = blockIdx.x * blockDim.x + threadIdx.x;
+    int i = blockIdx.y * blockDim.y + threadIdx.y;
+    if (i >= nrow || j >= ncol) return;
+    b[j + i * stride] = a[j + lrintf(idx[i]) * stride];
+}
+
 extern "C" {
 #include "../cukernel.h"
     void cudak_(cuda_log_elem)(const Matrix *a, Matrix *b) {
@@ -589,5 +598,16 @@ extern "C" {
              b->stride / sizeof(MATRIX_ELEM));
         cudaStreamSynchronize(0);
     }
+
+    void cudak_(cuda_copy_rows_by_idx)(const Matrix *a, Matrix *b,
+                                        const Matrix *idx, int b_begin) {
+        dim3 threadsPerBlock(CUDA_THREADS_NN, 1);
+        dim3 numBlocks(CEIL_DIV(b->ncol, threadsPerBlock.x), b->nrow);
+        cudak_(copy_rows_by_idx)<<<numBlocks, threadsPerBlock>>> \
+            (MATRIX_ELEM_PTR(a), MATRIX_ELEM_PTR(b),
+            MATRIX_ELEM_PTR(idx) + b_begin,
+             b_begin, b->nrow, b->ncol, b->stride / sizeof(MATRIX_ELEM));
+        cudaStreamSynchronize(0);
+    }
 }
 #endif