added rand_uniform and thres_mask for cumatrix

1 files changed, 31 insertions, 0 deletions

diff --git a/nerv/lib/matrix/generic/cukernel.cu b/nerv/lib/matrix/generic/cukernel.cu
index 1a20b4f..b092e4a 100644
--- a/nerv/lib/matrix/generic/cukernel.cu
+++ b/nerv/lib/matrix/generic/cukernel.cu
@@ -20,6 +20,19 @@ __global__ void cudak_(log_elem)(const MATRIX_ELEM *a, MATRIX_ELEM *b,
     b[idx] = log(tmp);
 }
 
+__global__ void cudak_(thres_mask)(MATRIX_ELEM *a, double thres, double low, double high, 
+                                int nrow, int ncol, int stride) {
+    int j = blockIdx.x * blockDim.x + threadIdx.x;
+    int i = blockIdx.y * blockDim.y + threadIdx.y;
+    long idx;
+    if (i >= nrow || j >= ncol) return;
+    idx = j + i * stride;
+    if (a[idx] < thres) 
+        a[idx] = low;
+    else
+        a[idx] = high;
+}
+
 __global__ void cudak_(mul_elem)(const MATRIX_ELEM *a, const MATRIX_ELEM *b,
                                 MATRIX_ELEM *c, 
                                 int nrow, int ncol, int stride) {
@@ -376,6 +389,24 @@ extern "C" {
         cudaStreamSynchronize(0);
     }
 
+    void cudak_(cuda_rand_uniform)(Matrix *a) {
+        #ifdef MATRIX_USE_FLOAT
+        curandGenerateUniform(*(a->curand_gen), MATRIX_ELEM_PTR(a), a->nrow * a->stride / sizeof(MATRIX_ELEM));
+        #endif
+        #ifdef MATRIX_USE_DOUBLE
+        curandGenerateUniformDouble(*(a->curand_gen), MATRIX_ELEM_PTR(a), a->nrow * a->stride / sizeof(MATRIX_ELEM));
+        #endif
+    }
+
+    void cudak_(cuda_thres_mask)(const Matrix *a, double thres, double low, double high) {
+        dim3 threadsPerBlock(CUDA_THREADS_N, CUDA_THREADS_N);
+        dim3 numBlocks(CEIL_DIV(a->ncol, threadsPerBlock.x),
+                CEIL_DIV(a->nrow, threadsPerBlock.y));
+        cudak_(thres_mask)<<<numBlocks, threadsPerBlock>>> \
+            (MATRIX_ELEM_PTR(a), thres, low, high, a->nrow, a->ncol, a->stride / sizeof(MATRIX_ELEM));
+        cudaStreamSynchronize(0);
+    }
+
     void cudak_(cuda_tanh)(const Matrix *a, Matrix *b) {
         dim3 threadsPerBlock(CUDA_THREADS_N, CUDA_THREADS_N);
         dim3 numBlocks(CEIL_DIV(b->ncol, threadsPerBlock.x),