| 31 | #include <cstddef> |
| 32 | |
| 33 | void arm_compute::enqueue(cl::CommandQueue &queue, |
| 34 | ICLKernel &kernel, |
| 35 | const Window &window, |
| 36 | const cl::NDRange &lws_hint, |
| 37 | bool use_dummy_work_items) |
| 38 | { |
| 39 | if (kernel.kernel()() == nullptr) |
| 40 | { |
| 41 | return; |
| 42 | } |
| 43 | |
| 44 | for (unsigned int i = 0; i < Coordinates::num_max_dimensions; ++i) |
| 45 | { |
| 46 | ARM_COMPUTE_ERROR_ON(window[i].step() == 0); |
| 47 | // Make sure that dimensions > Z are 1 |
| 48 | ARM_COMPUTE_ERROR_ON((i >= 3) && ((window[i].end() - window[i].start()) != 1)); |
| 49 | } |
| 50 | |
| 51 | cl::NDRange gws = ICLKernel::gws_from_window(window, use_dummy_work_items); |
| 52 | |
| 53 | // Check for empty NDRange |
| 54 | if (gws.dimensions() == 0) |
| 55 | { |
| 56 | return; |
| 57 | } |
| 58 | |
| 59 | kernel.cache_gws(gws); |
| 60 | |
| 61 | cl::NDRange valid_lws; |
| 62 | if (lws_hint[0] * lws_hint[1] * lws_hint[2] > kernel.get_max_workgroup_size()) |
| 63 | { |
| 64 | valid_lws = cl::NullRange; |
| 65 | } |
| 66 | else |
| 67 | { |
| 68 | valid_lws = lws_hint; |
| 69 | } |
| 70 | |
| 71 | cl::NDRange lws = cl::NullRange; |
| 72 | |
| 73 | if ((valid_lws[0] <= gws[0]) && (valid_lws[1] <= gws[1]) && (valid_lws[2] <= gws[2])) |
| 74 | { |
| 75 | lws = valid_lws; |
| 76 | } |
| 77 | |
| 78 | if (CLKernelLibrary::get().is_wbsm_supported()) |
| 79 | { |
| 80 | set_wbsm(kernel.kernel(), kernel.wbsm_hint()); |
| 81 | } |
| 82 | queue.enqueueNDRangeKernel(kernel.kernel(), cl::NullRange, gws, lws); |
| 83 | } |
| 84 | |
| 85 | namespace arm_compute |
| 86 | { |
no test coverage detected