Home
last modified time | relevance | path

Searched refs:point_loop_bound (Results 1 – 2 of 2) sorted by relevance

/libCEED/include/ceed/jit-source/cuda/
H A Dcuda-shared-basis-tensor-at-points.h60 …const CeedInt point_loop_bound = (blockDim.x * blockDim.y) * ceil(1.0 * BASIS_NUM_PTS / (blockDim.… in InterpAtPoints() local
62 …for (CeedInt i = threadIdx.x + threadIdx.y * blockDim.x; i < point_loop_bound; i += blockDim.x * b… in InterpAtPoints()
117 …const CeedInt point_loop_bound = (blockDim.x * blockDim.y) * ceil(1.0 * BASIS_NUM_PTS / (blockDim.… in InterpTransposeAtPoints() local
119 …for (CeedInt i = threadIdx.x + threadIdx.y * blockDim.x; i < point_loop_bound; i += blockDim.x * b… in InterpTransposeAtPoints()
176 …const CeedInt point_loop_bound = (blockDim.x * blockDim.y) * ceil(1.0 * BASIS_NUM_PTS / (blockDim.… in InterpTransposeAddAtPoints() local
178 …for (CeedInt i = threadIdx.x + threadIdx.y * blockDim.x; i < point_loop_bound; i += blockDim.x * b… in InterpTransposeAddAtPoints()
247 …const CeedInt point_loop_bound = (blockDim.x * blockDim.y) * ceil(1.0 * BASIS_NUM_PTS / (blockDim.… in GradAtPoints() local
249 …for (CeedInt i = threadIdx.x + threadIdx.y * blockDim.x; i < point_loop_bound; i += blockDim.x * b… in GradAtPoints()
304 …const CeedInt point_loop_bound = (blockDim.x * blockDim.y) * ceil(1.0 * BASIS_NUM_PTS / (blockDim.… in GradTransposeAtPoints() local
306 …for (CeedInt i = threadIdx.x + threadIdx.y * blockDim.x; i < point_loop_bound; i += blockDim.x * b… in GradTransposeAtPoints()
[all …]
/libCEED/include/ceed/jit-source/hip/
H A Dhip-shared-basis-tensor-at-points.h61 …const CeedInt point_loop_bound = (blockDim.x * blockDim.y) * ceil(1.0 * BASIS_NUM_PTS / (blockDim.… in __launch_bounds__() local
63 …for (CeedInt i = threadIdx.x + threadIdx.y * blockDim.x; i < point_loop_bound; i += blockDim.x * b… in __launch_bounds__()
118 …const CeedInt point_loop_bound = (blockDim.x * blockDim.y) * ceil(1.0 * BASIS_NUM_PTS / (blockDim.… in __launch_bounds__() local
120 …for (CeedInt i = threadIdx.x + threadIdx.y * blockDim.x; i < point_loop_bound; i += blockDim.x * b… in __launch_bounds__()
177 …const CeedInt point_loop_bound = (blockDim.x * blockDim.y) * ceil(1.0 * BASIS_NUM_PTS / (blockDim.… in __launch_bounds__() local
179 …for (CeedInt i = threadIdx.x + threadIdx.y * blockDim.x; i < point_loop_bound; i += blockDim.x * b… in __launch_bounds__()
249 …const CeedInt point_loop_bound = (blockDim.x * blockDim.y) * ceil(1.0 * BASIS_NUM_PTS / (blockDim.… in __launch_bounds__() local
251 …for (CeedInt i = threadIdx.x + threadIdx.y * blockDim.x; i < point_loop_bound; i += blockDim.x * b… in __launch_bounds__()
306 …const CeedInt point_loop_bound = (blockDim.x * blockDim.y) * ceil(1.0 * BASIS_NUM_PTS / (blockDim.… in __launch_bounds__() local
308 …for (CeedInt i = threadIdx.x + threadIdx.y * blockDim.x; i < point_loop_bound; i += blockDim.x * b… in __launch_bounds__()
[all …]