summaryrefslogtreecommitdiff
path: root/src/kernels/level3/transpose_pad.opencl
diff options
context:
space:
mode:
authorCedric Nugteren <web@cedricnugteren.nl>2017-10-14 17:13:54 +0200
committerCedric Nugteren <web@cedricnugteren.nl>2017-10-14 17:13:54 +0200
commitb06bc01da90983ce484fded4e1a87f5fcd5c4eca (patch)
treebd835ea8612fc15747c84736c355b4ceb3fb70ff /src/kernels/level3/transpose_pad.opencl
parentd9456306e0ee16f18f788106cd8ba74a81bf2e31 (diff)
Make local memory pointers a define in OpenCL; some fixes to the recently changed transpose kernel code
Diffstat (limited to 'src/kernels/level3/transpose_pad.opencl')
-rw-r--r--src/kernels/level3/transpose_pad.opencl4
1 files changed, 2 insertions, 2 deletions
diff --git a/src/kernels/level3/transpose_pad.opencl b/src/kernels/level3/transpose_pad.opencl
index 49c5b9a3..ba9a6a56 100644
--- a/src/kernels/level3/transpose_pad.opencl
+++ b/src/kernels/level3/transpose_pad.opencl
@@ -24,7 +24,7 @@ R"(
// Transposes a matrix from source to destination. The output is padded with zero values in case the
// destination matrix dimensions are larger than the transposed source matrix dimensions.
-INLINE_FUNC void _TransposePadMatrix(__local real* tile,
+INLINE_FUNC void _TransposePadMatrix(LOCAL_PTR real* tile,
const int src_one, const int src_two,
const int src_ld, const int src_offset,
__global const real* restrict src,
@@ -105,7 +105,7 @@ void TransposePadMatrix(const int src_one, const int src_two,
// Transposes a matrix, while considering possible padding in the source matrix. Data is read from a
// padded source matrix, but only the actual data is written back to the transposed destination
// matrix. This kernel optionally checks for upper/lower triangular matrices.
-INLINE_FUNC void _TransposeMatrix(__local real* tile,
+INLINE_FUNC void _TransposeMatrix(LOCAL_PTR real* tile,
const int src_one, const int src_two,
const int src_ld, const int src_offset,
__global const real* restrict src,