support gpu buffering

3 files changed, 39 insertions, 2 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
diff --git a/nerv/lib/matrix/generic/cumatrix.c b/nerv/lib/matrix/generic/cumatrix.c
index 2cb3563..770e503 100644
--- a/nerv/lib/matrix/generic/cumatrix.c
+++ b/nerv/lib/matrix/generic/cumatrix.c
@@ -315,7 +315,7 @@ void nerv_matrix_(copy_rows_fromh_by_idx)(Matrix *a, const Matrix *b,
     long nrow = a->nrow;
     if (!(0 <= b_begin && b_begin + nrow <= idx->ncol))
         NERV_EXIT_STATUS(status, MAT_INVALID_COPY_INTERVAL, 0);
-    long *idx_ptr = idx->data.i;
+    float *idx_ptr = idx->data.f;
     int i;
     if (idx->nrow != 1)
         NERV_EXIT_STATUS(status, MAT_IDX_VECTOR_EXP, 0);
@@ -325,7 +325,7 @@ void nerv_matrix_(copy_rows_fromh_by_idx)(Matrix *a, const Matrix *b,
     cudaStream_t *streams = (cudaStream_t*)malloc(sizeof(cudaStream_t) * nrow);
     for (i = 0; i < nrow; i++)
     {
-        int src_row = idx_ptr[b_begin + i];
+        int src_row = (int)idx_ptr[b_begin + i];
         if (!(0 <= src_row && src_row < b->nrow))
             NERV_EXIT_STATUS(status, MAT_INVALID_IDX, 0);
         CUDA_SAFE_CALL(cudaStreamCreate(streams + i), status);
@@ -344,6 +344,21 @@ void nerv_matrix_(copy_rows_fromh_by_idx)(Matrix *a, const Matrix *b,
     NERV_SET_STATUS(status, NERV_NORMAL, 0);
 }
 
+void nerv_matrix_(copy_rows_fromd_by_idx)(Matrix *a, const Matrix *b,
+                            const Matrix *idx, int b_begin, Status *status) {
+    long nrow = a->nrow;
+    if (!(0 <= b_begin && b_begin + nrow <= idx->ncol))
+        NERV_EXIT_STATUS(status, MAT_INVALID_COPY_INTERVAL, 0);
+    if (idx->nrow != 1)
+        NERV_EXIT_STATUS(status, MAT_IDX_VECTOR_EXP, 0);
+    if (a->ncol != b->ncol)
+        NERV_EXIT_STATUS(status, MAT_MISMATCH_DIM, 0);
+    PROFILE_START
+    cudak_(cuda_copy_rows_by_idx)(b, a, idx, b_begin);
+    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 3f1f8a3..04e8c5a 100644
--- a/nerv/lib/matrix/generic/cumatrix.h
+++ b/nerv/lib/matrix/generic/cumatrix.h
@@ -40,6 +40,8 @@ void nerv_matrix_(log_elem)(Matrix *b, const Matrix *a, Status *status);
 Matrix *nerv_matrix_(decompress)(const Matrix *a, int orig_col, Status *status);
 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_(expand_frm)(Matrix *a, const Matrix *b,
                             int context, Status *status);