summaryrefslogtreecommitdiff
path: root/src/tuning/kernels/copy_fast.cpp
diff options
context:
space:
mode:
authorCedric Nugteren <web@cedricnugteren.nl>2018-03-22 21:01:02 +0100
committerCedric Nugteren <web@cedricnugteren.nl>2018-03-22 21:01:02 +0100
commit9fb6550dd02c54fafbb03e20516a394d9cd63f3f (patch)
tree143bd103186eb0d3fd9672ab8a41ebb076d1107d /src/tuning/kernels/copy_fast.cpp
parent7a2371213bb900c8c726c346f85920199f424d82 (diff)
Added the OpenCL local memory size constraint to the tuners
Diffstat (limited to 'src/tuning/kernels/copy_fast.cpp')
-rw-r--r--src/tuning/kernels/copy_fast.cpp10
1 files changed, 5 insertions, 5 deletions
diff --git a/src/tuning/kernels/copy_fast.cpp b/src/tuning/kernels/copy_fast.cpp
index 0314113c..13f7ef3c 100644
--- a/src/tuning/kernels/copy_fast.cpp
+++ b/src/tuning/kernels/copy_fast.cpp
@@ -22,11 +22,11 @@ using double2 = clblast::double2;
int main(int argc, char *argv[]) {
const auto command_line_args = clblast::RetrieveCommandLineArguments(argc, argv);
switch(clblast::GetPrecision(command_line_args)) {
- case clblast::Precision::kHalf: clblast::Tuner<half>(argc, argv, 0, clblast::CopyGetTunerDefaults, clblast::CopyGetTunerSettings<half>, clblast::CopyTestValidArguments<half>, clblast::CopySetConstraints, clblast::CopySetArguments<half>); break;
- case clblast::Precision::kSingle: clblast::Tuner<float>(argc, argv, 0, clblast::CopyGetTunerDefaults, clblast::CopyGetTunerSettings<float>, clblast::CopyTestValidArguments<float>, clblast::CopySetConstraints, clblast::CopySetArguments<float>); break;
- case clblast::Precision::kDouble: clblast::Tuner<double>(argc, argv, 0, clblast::CopyGetTunerDefaults, clblast::CopyGetTunerSettings<double>, clblast::CopyTestValidArguments<double>, clblast::CopySetConstraints, clblast::CopySetArguments<double>); break;
- case clblast::Precision::kComplexSingle: clblast::Tuner<float2>(argc, argv, 0, clblast::CopyGetTunerDefaults, clblast::CopyGetTunerSettings<float2>, clblast::CopyTestValidArguments<float2>, clblast::CopySetConstraints, clblast::CopySetArguments<float2>); break;
- case clblast::Precision::kComplexDouble: clblast::Tuner<double2>(argc, argv, 0, clblast::CopyGetTunerDefaults, clblast::CopyGetTunerSettings<double2>, clblast::CopyTestValidArguments<double2>, clblast::CopySetConstraints, clblast::CopySetArguments<double2>); break;
+ case clblast::Precision::kHalf: clblast::Tuner<half>(argc, argv, 0, clblast::CopyGetTunerDefaults, clblast::CopyGetTunerSettings<half>, clblast::CopyTestValidArguments<half>, clblast::CopySetConstraints, clblast::CopyComputeLocalMemSize<half>, clblast::CopySetArguments<half>); break;
+ case clblast::Precision::kSingle: clblast::Tuner<float>(argc, argv, 0, clblast::CopyGetTunerDefaults, clblast::CopyGetTunerSettings<float>, clblast::CopyTestValidArguments<float>, clblast::CopySetConstraints, clblast::CopyComputeLocalMemSize<float>, clblast::CopySetArguments<float>); break;
+ case clblast::Precision::kDouble: clblast::Tuner<double>(argc, argv, 0, clblast::CopyGetTunerDefaults, clblast::CopyGetTunerSettings<double>, clblast::CopyTestValidArguments<double>, clblast::CopySetConstraints, clblast::CopyComputeLocalMemSize<double>, clblast::CopySetArguments<double>); break;
+ case clblast::Precision::kComplexSingle: clblast::Tuner<float2>(argc, argv, 0, clblast::CopyGetTunerDefaults, clblast::CopyGetTunerSettings<float2>, clblast::CopyTestValidArguments<float2>, clblast::CopySetConstraints, clblast::CopyComputeLocalMemSize<float2>, clblast::CopySetArguments<float2>); break;
+ case clblast::Precision::kComplexDouble: clblast::Tuner<double2>(argc, argv, 0, clblast::CopyGetTunerDefaults, clblast::CopyGetTunerSettings<double2>, clblast::CopyTestValidArguments<double2>, clblast::CopySetConstraints, clblast::CopyComputeLocalMemSize<double2>, clblast::CopySetArguments<double2>); break;
}
return 0;
}