Just come back, let's merge the new master

Merge branch 'master' into txh18/rnnlm

10 files changed, 71 insertions, 23 deletions

diff --git a/nerv/lib/matrix/cukernel.h b/nerv/lib/matrix/cukernel.h
index 7bb4c2c..2126c6f 100644
--- a/nerv/lib/matrix/cukernel.h
+++ b/nerv/lib/matrix/cukernel.h
@@ -18,4 +18,6 @@ void cudak_(cuda_rearrange_frm)(const Matrix *a, Matrix *b, int step);
 void cudak_(cuda_scale_rows_by_row)(const Matrix *a, Matrix *b);
 void cudak_(cuda_scale_rows_by_col)(const Matrix *a, Matrix *b);
 void cudak_(cuda_decompress)(const Matrix *a, Matrix *b);
+void cudak_(cuda_copy_rows_by_idx)(const Matrix *a, Matrix *b,
+                                    const Matrix *idx, int b_begin);
 #endif
diff --git a/nerv/lib/matrix/cumatrix.c b/nerv/lib/matrix/cumatrix.c
index aa81bfc..ff1168d 100644
--- a/nerv/lib/matrix/cumatrix.c
+++ b/nerv/lib/matrix/cumatrix.c
@@ -49,6 +49,7 @@ void nerv_cumatrix_init() {
 #define NERV_CUBLAS_(NAME) cublasS##NAME
 #define MATRIX_CUMATRIX_HOST_TNAME nerv_matrix_host_float_tname
 #include "generic/cumatrix.c"
+
 #undef NERV_CUBLAS_
 #undef cudak_
 #undef nerv_matrix_
diff --git a/nerv/lib/matrix/cumatrix.h b/nerv/lib/matrix/cumatrix.h
index 9f71507..e6def66 100644
--- a/nerv/lib/matrix/cumatrix.h
+++ b/nerv/lib/matrix/cumatrix.h
@@ -1,5 +1,6 @@
 #ifndef NERV_CUMATRIX_H
 #define NERV_CUMATRIX_H
+#include "matrix.h"
 void nerv_cumatrix_print_profile();
 void nerv_cumatrix_clear_profile();
 void nerv_cumatrix_init();
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 40a0030..770e503 100644
--- a/nerv/lib/matrix/generic/cumatrix.c
+++ b/nerv/lib/matrix/generic/cumatrix.c
@@ -315,16 +315,17 @@ 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);
     if (a->ncol != b->ncol)
         NERV_EXIT_STATUS(status, MAT_MISMATCH_DIM, 0);
+    PROFILE_START
     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);
@@ -339,6 +340,22 @@ void nerv_matrix_(copy_rows_fromh_by_idx)(Matrix *a, const Matrix *b,
         CUDA_SAFE_CALL(cudaStreamDestroy(streams[i]), status);
     }
     free(streams);
+    PROFILE_STOP
+    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);
 }
 
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);
diff --git a/nerv/lib/matrix/generic/matrix.c b/nerv/lib/matrix/generic/matrix.c
index e4afa37..4246751 100644
--- a/nerv/lib/matrix/generic/matrix.c
+++ b/nerv/lib/matrix/generic/matrix.c
@@ -3,7 +3,8 @@
 #include "matrix.h"
 /* FIXME: malloc failure detection */
 
-static void nerv_matrix_(data_free)(Matrix *self, Status *status) {
+void nerv_matrix_(data_free)(Matrix *self, Status *status) {
+    if(*self->data_ref == 0) return; /* FIXME: repeat free memory */
     assert(*self->data_ref > 0);
     if (--(*self->data_ref) == 0)
     {
@@ -18,7 +19,7 @@ static void nerv_matrix_(data_free)(Matrix *self, Status *status) {
     }
 }
 
-static void nerv_matrix_(data_retain)(Matrix *self) {
+void nerv_matrix_(data_retain)(Matrix *self) {
     (*self->data_ref)++;
 }
 
diff --git a/nerv/lib/matrix/generic/matrix.h b/nerv/lib/matrix/generic/matrix.h
index 9d44e6d..69b4e6d 100644
--- a/nerv/lib/matrix/generic/matrix.h
+++ b/nerv/lib/matrix/generic/matrix.h
@@ -2,3 +2,5 @@
 Matrix *nerv_matrix_(create)(long nrow, long ncol, Status *status);
 void nerv_matrix_(destroy)(Matrix *self, Status *status);
 Matrix *nerv_matrix_(getrow)(Matrix *self, int row);
+void nerv_matrix_(data_free)(Matrix *self, Status *status);
+void nerv_matrix_(data_retain)(Matrix *self);
diff --git a/nerv/lib/matrix/mmatrix.c b/nerv/lib/matrix/mmatrix.c
index 94f1ea8..b8157eb 100644
--- a/nerv/lib/matrix/mmatrix.c
+++ b/nerv/lib/matrix/mmatrix.c
@@ -7,6 +7,25 @@
 #define nerv_matrix_(NAME) nerv_matrix_host_float_##NAME
 #include "generic/matrix.h"
 #include "generic/mmatrix.c"
+
+Matrix *nerv_matrix_(perm_gen)(int ncol, Status *status) {
+    int i;
+    Matrix *self = nerv_matrix_(create)(1, ncol, status);
+    if (status->err_code != NERV_NORMAL)
+        return NULL;
+    float *prow = self->data.f;
+    for (i = 0; i < ncol; i++)
+        prow[i] = i;
+    for (i = ncol - 1; i >= 0; i--)
+    {
+        size_t j = rand() % (i + 1);
+        float tmp = prow[i];
+        prow[i] = prow[j];
+        prow[j] = tmp;
+    }
+    return self;
+}
+
 #undef nerv_matrix_
 #undef host_matrix_
 #undef MATRIX_USE_FLOAT
@@ -33,21 +52,3 @@
 #define host_matrix_(NAME) host_matrix_int_##NAME
 #define nerv_matrix_(NAME) nerv_matrix_host_int_##NAME
 #include "generic/mmatrix.c"
-
-Matrix *nerv_matrix_(perm_gen)(int ncol, Status *status) {
-    int i;
-    Matrix *self = nerv_matrix_(create)(1, ncol, status);
-    if (status->err_code != NERV_NORMAL)
-        return NULL;
-    long *prow = self->data.i;
-    for (i = 0; i < ncol; i++)
-        prow[i] = i;
-    for (i = ncol - 1; i >= 0; i--)
-    {
-        size_t j = rand() % (i + 1);
-        long tmp = prow[i];
-        prow[i] = prow[j];
-        prow[j] = tmp;
-    }
-    return self;
-}
diff --git a/nerv/lib/matrix/mmatrix.h b/nerv/lib/matrix/mmatrix.h
index df91e4c..31e7984 100644
--- a/nerv/lib/matrix/mmatrix.h
+++ b/nerv/lib/matrix/mmatrix.h
@@ -1,4 +1,5 @@
 #ifndef NERV_MMATRIX_H
 #define NERV_MMATRIX_H
-Matrix *nerv_matrix_(perm_gen)(int ncol, Status *status);
+#include "matrix.h"
+Matrix *nerv_matrix_host_float_perm_gen(int ncol, Status *status);
 #endif