update LSTM layer

5 files changed, 45 insertions, 0 deletions

diff --git a/nerv/lib/matrix/generic/cukernel.cu b/nerv/lib/matrix/generic/cukernel.cu
index 51e3b6a..311a6ce 100644
--- a/nerv/lib/matrix/generic/cukernel.cu
+++ b/nerv/lib/matrix/generic/cukernel.cu
@@ -250,6 +250,14 @@ __global__ void cudak_(fill)(MATRIX_ELEM *a,
     a[j + i * stride] = val;
 }
 
+__global__ void cudak_(diagonalize)(MATRIX_ELEM *a,
+                            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 || i == j) return;
+    a[j + i * stride] = 0;
+}
+
 __global__ void cudak_(clip)(MATRIX_ELEM *a,
                             int nrow, int ncol, int stride, double val_1, double val_2) {
     int j = blockIdx.x * blockDim.x + threadIdx.x;
@@ -679,6 +687,16 @@ extern "C" {
         cudaStreamSynchronize(0);
     }
 
+    void cudak_(cuda_diagonalize)(Matrix *a) {
+        dim3 threadsPerBlock(CUDA_THREADS_N, CUDA_THREADS_N);
+        dim3 numBlocks(CEIL_DIV(a->ncol, threadsPerBlock.x),
+                CEIL_DIV(a->nrow, threadsPerBlock.y));
+        cudak_(diagonalize)<<<numBlocks, threadsPerBlock>>> \
+            (MATRIX_ELEM_PTR(a), a->nrow, a->ncol,
+            a->stride / sizeof(MATRIX_ELEM));
+        cudaStreamSynchronize(0);
+    }
+
     void cudak_(cuda_clip)(Matrix *a, double val_1, double val_2) {
         dim3 threadsPerBlock(CUDA_THREADS_N, CUDA_THREADS_N);
         dim3 numBlocks(CEIL_DIV(a->ncol, threadsPerBlock.x),
diff --git a/nerv/lib/matrix/generic/cumatrix.c b/nerv/lib/matrix/generic/cumatrix.c
index 7b70607..1c74866 100644
--- a/nerv/lib/matrix/generic/cumatrix.c
+++ b/nerv/lib/matrix/generic/cumatrix.c
@@ -494,6 +494,15 @@ void nerv_matrix_(prefixsum_row)(Matrix *a, const Matrix *b, Status *status) {
     NERV_SET_STATUS(status, NERV_NORMAL, 0);
 }
 
+void nerv_matrix_(diagonalize)(Matrix *a, Status *status) {
+    if (a->nrow != a->ncol)
+        NERV_EXIT_STATUS(status, MAT_MISMATCH_DIM, 0);
+    PROFILE_START
+    cudak_(cuda_diagonalize)(a);
+    PROFILE_STOP
+    NERV_SET_STATUS(status, NERV_NORMAL, 0);
+}
+
 static void cuda_matrix_(free)(MATRIX_ELEM *ptr, Status *status) {
     CUDA_SAFE_SYNC_CALL(cudaFree(ptr), status);
     NERV_SET_STATUS(status, NERV_NORMAL, 0);
diff --git a/nerv/lib/matrix/generic/cumatrix.h b/nerv/lib/matrix/generic/cumatrix.h
index f3c2df8..48d1f13 100644
--- a/nerv/lib/matrix/generic/cumatrix.h
+++ b/nerv/lib/matrix/generic/cumatrix.h
@@ -25,6 +25,7 @@ void nerv_matrix_(add_row)(Matrix *b, const Matrix *a, double beta,
                             Status *status);
 void nerv_matrix_(clip)(Matrix *self, double val_1, double val_2, Status *status);
 void nerv_matrix_(fill)(Matrix *self, double val, Status *status);
+void nerv_matrix_(diagonalize)(Matrix *self, Status *statut);
 void nerv_matrix_(copy_fromd)(Matrix *a, const Matrix *b,
                             int a_begin, int b_begin, int b_end,
                             Status *status);
diff --git a/nerv/lib/matrix/generic/mmatrix.c b/nerv/lib/matrix/generic/mmatrix.c
index fa1dc5f..3dabe0e 100644
--- a/nerv/lib/matrix/generic/mmatrix.c
+++ b/nerv/lib/matrix/generic/mmatrix.c
@@ -265,6 +265,22 @@ void nerv_matrix_(fill)(Matrix *self, double val, Status *status) {
     NERV_SET_STATUS(status, NERV_NORMAL, 0);
 }
 
+void nerv_matrix_(diagonalize)(Matrix *self, Status *status) {
+    if (self->nrow != self->ncol)
+        NERV_EXIT_STATUS(status, MAT_MISMATCH_DIM, 0);
+    int i, j;
+    size_t astride = self->stride;
+    MATRIX_ELEM *arow = MATRIX_ELEM_PTR(self);
+    for (i = 0; i < self->nrow; i++)
+    {
+        for (j = 0; j < self->ncol; j++)
+            if (i != j)
+                arow[j] = 0;
+        arow = MATRIX_NEXT_ROW_PTR(arow, astride);
+    }
+    NERV_SET_STATUS(status, NERV_NORMAL, 0);
+}
+
 void nerv_matrix_(sigmoid)(Matrix *a, const Matrix *b, Status *status) {
     CHECK_SAME_DIMENSION(a, b, status);
     int i, j;
diff --git a/nerv/lib/matrix/generic/mmatrix.h b/nerv/lib/matrix/generic/mmatrix.h
index c54c4e5..2cbca47 100644
--- a/nerv/lib/matrix/generic/mmatrix.h
+++ b/nerv/lib/matrix/generic/mmatrix.h
@@ -23,6 +23,7 @@ void nerv_matrix_(add_row)(Matrix *b, const Matrix *a, double beta,
                             Status *status);
 void nerv_matrix_(clip)(Matrix *self, double val_1, double val_2, Status *status);
 void nerv_matrix_(fill)(Matrix *self, double val, Status *status);
+void nerv_matrix_(diagonalize)(Matrix *self, Status *status);
 void nerv_matrix_(copy_fromh)(Matrix *a, const Matrix *b,
                             int a_begin, int b_begin, int b_end,
                             Status *status);