|
|
@ -148,7 +148,7 @@ class DensityPriorBoxOpCUDAKernel : public framework::OpKernel<T> {
|
|
|
|
// blockx is multiple of 32.
|
|
|
|
// blockx is multiple of 32.
|
|
|
|
int blockx = std::min(
|
|
|
|
int blockx = std::min(
|
|
|
|
static_cast<int64_t>(((feature_width * num_priors + 31) >> 5) << 5),
|
|
|
|
static_cast<int64_t>(((feature_width * num_priors + 31) >> 5) << 5),
|
|
|
|
512L);
|
|
|
|
static_cast<int64_t>(512L));
|
|
|
|
int gridx = (feature_width * num_priors + blockx - 1) / blockx;
|
|
|
|
int gridx = (feature_width * num_priors + blockx - 1) / blockx;
|
|
|
|
dim3 threads(blockx, 1);
|
|
|
|
dim3 threads(blockx, 1);
|
|
|
|
dim3 grids(gridx, feature_height);
|
|
|
|
dim3 grids(gridx, feature_height);
|
|
|
|