| 272 | } |
| 273 | |
| 274 | void vulkan_function::execute(const device_queue& cqueue, |
| 275 | vulkan_command_buffer* external_cmd_buffer, |
| 276 | const bool& is_cooperative, |
| 277 | const bool& wait_until_completion, |
| 278 | const uint32_t& dim floor_unused, |
| 279 | const uint3& global_work_size, |
| 280 | const uint3& local_work_size_, |
| 281 | const std::vector<device_function_arg>& args, |
| 282 | const std::vector<const device_fence*>& wait_fences_, |
| 283 | const std::vector<device_fence*>& signal_fences_, |
| 284 | const char* debug_label, |
| 285 | kernel_completion_handler_f&& completion_handler) const { |
| 286 | // no cooperative support yet |
| 287 | if (is_cooperative) { |
| 288 | log_error("cooperative kernel execution is not supported for Vulkan"); |
| 289 | return; |
| 290 | } |
| 291 | |
| 292 | // find entry for queue device |
| 293 | const auto function_iter = get_function(cqueue); |
| 294 | if (function_iter == functions.cend()) { |
| 295 | log_error("no function \"$\" for this compute queue/device exists!", function_name); |
| 296 | return; |
| 297 | } |
| 298 | |
| 299 | const auto& vk_dev = *function_iter->first; |
| 300 | const auto& vk_queue = (const vulkan_queue&)cqueue; |
| 301 | |
| 302 | // check work size |
| 303 | const uint3 block_dim = check_local_work_size(*function_iter->second, local_work_size_); |
| 304 | if (function_iter->second->info->has_valid_required_local_size() && |
| 305 | (function_iter->second->info->required_local_size != local_work_size_.maxed(1u)).any()) { |
| 306 | log_error("function $ has fixed compiled required local size of $, it may not be executed with a different local size of $", |
| 307 | function_iter->second->info->name, function_iter->second->info->required_local_size, local_work_size_); |
| 308 | return; |
| 309 | } |
| 310 | |
| 311 | const uint3 grid_dim_overflow { |
| 312 | global_work_size.x > 0 ? std::min(uint32_t(global_work_size.x % block_dim.x), 1u) : 0u, |
| 313 | global_work_size.y > 0 ? std::min(uint32_t(global_work_size.y % block_dim.y), 1u) : 0u, |
| 314 | global_work_size.z > 0 ? std::min(uint32_t(global_work_size.z % block_dim.z), 1u) : 0u |
| 315 | }; |
| 316 | uint3 grid_dim { (global_work_size / block_dim) + grid_dim_overflow }; |
| 317 | grid_dim.max(1u); |
| 318 | |
| 319 | // create command buffer ("encoder") for this function execution |
| 320 | const std::vector<const vulkan_function_entry*> shader_entries { |
| 321 | function_iter->second.get() |
| 322 | }; |
| 323 | bool encoder_success = false; |
| 324 | auto encoder = create_encoder(cqueue, external_cmd_buffer, |
| 325 | get_pipeline_spec(vk_dev, *function_iter->second, block_dim, {} /* use default or program defined */), |
| 326 | function_iter->second->pipeline_layout, |
| 327 | shader_entries, debug_label, encoder_success); |
| 328 | if (!encoder_success) { |
| 329 | log_error("failed to create Vulkan encoder / command buffer for function \"$\"", function_iter->second->info->name); |
| 330 | return; |
| 331 | } |
nothing calls this directly
no test coverage detected