diff options
Diffstat (limited to 'src/routines/common.cpp')
-rw-r--r-- | src/routines/common.cpp | 40 |
1 files changed, 17 insertions, 23 deletions
diff --git a/src/routines/common.cpp b/src/routines/common.cpp index 2e82e04d..3969cf9f 100644 --- a/src/routines/common.cpp +++ b/src/routines/common.cpp @@ -22,23 +22,25 @@ namespace clblast { // Enqueues a kernel, waits for completion, and checks for errors StatusCode RunKernel(Kernel &kernel, Queue &queue, const Device &device, std::vector<size_t> global, const std::vector<size_t> &local, - EventPointer event, std::vector<Event>& waitForEvents) { + EventPointer event, const std::vector<Event> &waitForEvents) { - // Tests for validity of the local thread sizes - if (local.size() > device.MaxWorkItemDimensions()) { - return StatusCode::kInvalidLocalNumDimensions; - } - const auto max_work_item_sizes = device.MaxWorkItemSizes(); - for (auto i=size_t{0}; i<local.size(); ++i) { - if (local[i] > max_work_item_sizes[i]) { return StatusCode::kInvalidLocalThreadsDim; } - } - auto local_size = size_t{1}; - for (auto &item: local) { local_size *= item; } - if (local_size > device.MaxWorkGroupSize()) { return StatusCode::kInvalidLocalThreadsTotal; } + if (!local.empty()) { + // Tests for validity of the local thread sizes + if (local.size() > device.MaxWorkItemDimensions()) { + return StatusCode::kInvalidLocalNumDimensions; + } + const auto max_work_item_sizes = device.MaxWorkItemSizes(); + for (auto i=size_t{0}; i<local.size(); ++i) { + if (local[i] > max_work_item_sizes[i]) { return StatusCode::kInvalidLocalThreadsDim; } + } + auto local_size = size_t{1}; + for (auto &item: local) { local_size *= item; } + if (local_size > device.MaxWorkGroupSize()) { return StatusCode::kInvalidLocalThreadsTotal; } - // Make sure the global thread sizes are at least equal to the local sizes - for (auto i=size_t{0}; i<global.size(); ++i) { - if (global[i] < local[i]) { global[i] = local[i]; } + // Make sure the global thread sizes are at least equal to the local sizes + for (auto i=size_t{0}; i<global.size(); ++i) { + if (global[i] < local[i]) { global[i] = local[i]; } + } } // Tests for local memory usage @@ -69,13 +71,5 @@ StatusCode RunKernel(Kernel &kernel, Queue &queue, const Device &device, return StatusCode::kSuccess; } -// As above, but without an event waiting list -StatusCode RunKernel(Kernel &kernel, Queue &queue, const Device &device, - std::vector<size_t> global, const std::vector<size_t> &local, - EventPointer event) { - auto emptyWaitingList = std::vector<Event>(); - return RunKernel(kernel, queue, device, global, local, event, emptyWaitingList); -} - // ================================================================================================= } // namespace clblast |