ogl_beamforming

beamformer.c (13669B)

---

 /* See LICENSE for license details. */
 
 #include "beamformer_internal.h"
 
 /* NOTE(rnp): magic variables to force discrete GPU usage on laptops with multiple devices */
 EXPORT i32 NvOptimusEnablement = 1;
 EXPORT i32 AmdPowerXpressRequestHighPerformance = 1;
 
 #if !BEAMFORMER_DEBUG
 #include "beamformer_core.c"
 #else
 
 typedef void beamformer_frame_step_fn(BeamformerInput *);
 
 #define BEAMFORMER_DEBUG_ENTRY_POINTS \
 	X(beamformer_debug_ui_deinit)  \
 	X(beamformer_complete_compute) \
 	X(beamformer_frame_step)       \
 	X(beamformer_rf_upload)        \
 
 #define X(name) global name ##_fn *name;
 BEAMFORMER_DEBUG_ENTRY_POINTS
 #undef X
 
 BEAMFORMER_EXPORT void
 beamformer_debug_hot_release(BeamformerInput *input)
 {
 	BeamformerCtx *ctx = BeamformerContextMemory(input->memory);
 
 	// TODO(rnp): this will deadlock if live imaging is active
 	/* NOTE(rnp): spin until compute thread finishes its work (we will probably
 	 * never reload while compute is in progress but just incase). */
 	spin_wait(atomic_load_u32(&ctx->upload_worker.awake));
 	spin_wait(atomic_load_u32(&ctx->compute_worker.awake));
 }
 
 BEAMFORMER_EXPORT void
 beamformer_debug_hot_reload(OSLibrary library, BeamformerInput *input)
 {
 	#define X(name) name = os_lookup_symbol(library, #name);
 	BEAMFORMER_DEBUG_ENTRY_POINTS
 	#undef X
 
 	str8 info = beamformer_info("reloaded main executable");
 	os_console_log(info.data, info.length);
 }
 
 #endif /* BEAMFORMER_DEBUG */
 
 function no_return void
 fatal(str8 message)
 {
 	os_fatal(message.data, message.length);
 	unreachable();
 }
 
 #include "vulkan.c"
 
 // TODO(rnp): this doesn't belong here, but will be removed
 // once vulkan migration is complete
 void * glfwGetProcAddress(char *);
 
 function void
 gl_debug_logger(u32 src, u32 type, u32 id, u32 lvl, i32 len, const char *msg, const void *userctx)
 {
 	Stream *e = (Stream *)userctx;
 	stream_append_str8s(e, str8("[OpenGL] "), (str8){.length = len, .data = (u8 *)msg}, str8("\n"));
 	os_console_log(e->data, e->widx);
 	stream_reset(e, 0);
 }
 
 function void
 load_gl(Stream *err)
 {
 	#define X(name, ret, params) name = (name##_fn *)glfwGetProcAddress(#name);
 	OGLProcedureList
 	OGLRequiredExtensionProcedureList
 	#undef X
 
 	stream_reset(err, 0);
 	#define X(name, ret, params) if (!name) stream_append_str8(err, str8("missing required GL function: " #name "\n"));
 	OGLProcedureList
 	OGLRequiredExtensionProcedureListBase
 	#if OS_WINDOWS
 	  OGLRequiredExtensionProcedureListW32
 	#else
 	  OGLRequiredExtensionProcedureListLinux
 	#endif
 	#undef X
 
 	if (err->widx) fatal(stream_to_str8(err));
 }
 
 function void
 beamformer_load_cuda_library(BeamformerCtx *ctx, OSLibrary cuda, Arena arena)
 {
 	/* TODO(rnp): (25.10.30) registering the rf buffer with CUDA is currently
 	 * causing a major performance regression. for now we are disabling its use
 	 * altogether. it will be reenabled once the issue can be fixed */
 	b32 result = 0 && vk_gpu_info()->vendor == GPUVendor_NVIDIA && ValidHandle(cuda);
 	if (result) {
 		Stream err = arena_stream(arena);
 
 		stream_append_str8(&err, beamformer_info("loading CUDA library functions"));
 		#define X(name, symname) cuda_## name = os_lookup_symbol(cuda, symname);
 		CUDALibraryProcedureList
 		#undef X
 
 		os_console_log(err.data, err.widx);
 	}
 
 	#define X(name, symname) if (!cuda_## name) cuda_## name = cuda_ ## name ## _stub;
 	CUDALibraryProcedureList
 	#undef X
 }
 
 function void
 worker_thread_sleep(GLWorkerThreadContext *ctx, BeamformerSharedMemory *sm)
 {
 	for (;;) {
 		i32 expected = 0;
 		if (atomic_cas_u32(&ctx->sync_variable, &expected, 1) ||
 		    atomic_load_u32(&sm->live_imaging_parameters.active))
 		{
 			break;
 		}
 
 		/* TODO(rnp): clean this crap up; we shouldn't need two values to communicate this */
 		atomic_store_u32(&ctx->awake, 0);
 		os_wait_on_address(&ctx->sync_variable, 1, (u32)-1);
 		atomic_store_u32(&ctx->awake, 1);
 	}
 }
 
 function OS_THREAD_ENTRY_POINT_FN(compute_worker_thread_entry_point)
 {
 	GLWorkerThreadContext *ctx = user_context;
 
 	BeamformerCtx *beamformer = (BeamformerCtx *)ctx->user_context;
 
 	for (;;) {
 		worker_thread_sleep(ctx, beamformer->shared_memory);
 		asan_poison_region(ctx->arena.beg, ctx->arena.end - ctx->arena.beg);
 		beamformer_complete_compute(beamformer, &ctx->arena);
 	}
 
 	unreachable();
 
 	return 0;
 }
 
 function OS_THREAD_ENTRY_POINT_FN(beamformer_upload_entry_point)
 {
 	GLWorkerThreadContext         *ctx = user_context;
 	BeamformerUploadThreadContext *up  = (typeof(up))ctx->user_context;
 
 	for (;;) {
 		worker_thread_sleep(ctx, up->shared_memory);
 		beamformer_rf_upload(up);
 	}
 
 	unreachable();
 
 	return 0;
 }
 
 BEAMFORMER_EXPORT void
 beamformer_init(BeamformerInput *input)
 {
 	Arena  memory        = arena_from_memory(input->memory, input->memory_size);
 	Arena  compute_arena = sub_arena_end(&memory, MB(2), KB(4));
 	Arena  upload_arena  = sub_arena_end(&memory, KB(4), KB(4));
 	Arena  ui_arena      = sub_arena_end(&memory, MB(2), KB(4));
 	Stream error         = arena_stream(sub_arena_end(&memory, MB(1), 1));
 
 	BeamformerCtx *ctx   = push_struct(&memory, BeamformerCtx);
 
 	for EachElement(ctx->frame_arenas, it) {
 		ctx->frame_arenas[it]           = sub_arena(&memory, KB(64), KB(4));
 		ctx->frame_arena_savepoints[it] = begin_temp_arena(ctx->frame_arenas + it);
 	}
 
 	str8 window_title = str8("VK Beamformer");
 	ctx->main_window  = os_window_create(window_title.data, window_title.length, 1280, 840);
 	ctx->window_size  = (iv2){{1280, 840}};
 
 	Arena scratch = {.beg = memory.end - 4096L, .end = memory.end};
 	memory.end = scratch.beg;
 
 	ctx->error_stream          = error;
 	ctx->ui_backing_store      = ui_arena;
 	ctx->compute_worker.arena  = compute_arena;
 	ctx->upload_worker.arena   = upload_arena;
 
 	#if BEAMFORMER_RENDERDOC_HOOKS
 	start_frame_capture       = input->renderdoc_start_frame_capture;
 	end_frame_capture         = input->renderdoc_end_frame_capture;
 	set_capture_path_template = input->renderdoc_set_capture_file_path_template;
 	#endif
 
 	vk_load(input->vulkan_library_handle, &memory, &ctx->error_stream);
 
 	BeamformerComputeContext *cs = &ctx->compute_context;
 
 	// NOTE(rnp): allocate beamformed image ring buffer
 	{
 		u64 gpu_heap_size = vk_gpu_info()->gpu_heap_size;
 		u64 trial_sizes[] = {
 			GB(4),
 			GB(2),
 			GB(1) + MB(512),
 			GB(1),
 		};
 
 		u32 base_index = 0;
 		for EachElement(trial_sizes, it) {
 			if (gpu_heap_size >= 2 * trial_sizes[it])
 				break;
 			base_index++;
 		}
 
 		for (u32 i = base_index; i < countof(trial_sizes); i++) {
 			// TODO(rnp): it may be better to download data from this using the transfer queue
 			VulkanTimeline timelines[] = {VulkanTimeline_Compute, VulkanTimeline_Graphics};
 			GPUBufferAllocateInfo allocate_info = {
 				.size            = trial_sizes[i],
 				.flags           = VulkanUsageFlag_TransferDestination|VulkanUsageFlag_TransferSource|VulkanUsageFlag_HostReadWrite,
 				.timeline_count  = countof(timelines),
 				.timelines_used  = timelines,
 				.label           = str8("BeamformedData"),
 			};
 			vk_buffer_allocate(cs->backlog.buffer, &allocate_info);
 			if (cs->backlog.buffer->size > 0)
 				break;
 		}
 		if (cs->backlog.buffer->size == 0) {
 			// NOTE(rnp): if this becomes an issue we may be able to get by in some other way
 			fatal(str8("Failed to allocate space for beamformed data\n"));
 		}
 
 		BeamformerShaderResourceInfo shader_resource_infos[] = {
 			{
 				.kind   = BeamformerShaderResourceKind_Buffer,
 				.handle = cs->backlog.buffer->handle,
 				.slot   = BeamformerShaderBufferSlot_BeamformedData,
 			},
 		};
 		vk_bind_shader_resources(shader_resource_infos, countof(shader_resource_infos));
 	}
 
 	beamformer_load_cuda_library(ctx, input->cuda_library_handle, memory);
 
 	load_gl(&ctx->error_stream);
 
 	ctx->shared_memory      = input->shared_memory;
 	ctx->shared_memory_size = input->shared_memory_size;
 	if (ctx->shared_memory_size < (i64)sizeof(*ctx->shared_memory))
 		fatal(str8("Get more ram lol\n"));
 	zero_struct(ctx->shared_memory);
 
 	ctx->shared_memory->version = BEAMFORMER_SHARED_MEMORY_VERSION;
 	ctx->shared_memory->reserved_parameter_blocks = 1;
 
 	ctx->shared_memory->beamformed_frame_buffer_size = cs->backlog.buffer->size;
 
 	// TODO(rnp): dynamic rf data buffer slot usage
 	// NOTE(rnp): will be same as the max size we were able to get for the frame buffer
 	ctx->shared_memory->capabilities.max_rf_data_size = cs->backlog.buffer->size
 	                                                    / BeamformerMaxRawDataFramesInFlight;
 
 	ctx->shared_memory->capabilities.cuda    = cuda_supported();
 	// TODO(rnp): re-enable hilbert support, with and without cuda
 	ctx->shared_memory->capabilities.hilbert = 0;
 
 	/* TODO(rnp): I'm not sure if its a good idea to pre-reserve a bunch of semaphores
 	 * on w32 but thats what we are doing for now */
 	#if OS_WINDOWS
 	{
 		Stream sb = arena_stream(memory);
 		stream_append(&sb, input->shared_memory_name, input->shared_memory_name_length);
 		stream_append_str8(&sb, str8("_lock_"));
 		i32 start_index = sb.widx;
 		for EachElement(os_w32_shared_memory_semaphores, it) {
 			stream_reset(&sb, start_index);
 			stream_append_u64(&sb, it);
 			stream_append_byte(&sb, 0);
 			os_w32_shared_memory_semaphores[it] = os_w32_create_semaphore((c8 *)sb.data, 1, 1);
 			if InvalidHandle(os_w32_shared_memory_semaphores[it])
 				fatal(beamformer_info("init: failed to create w32 shared memory semaphore\n"));
 
 			/* NOTE(rnp): hacky garbage because CreateSemaphore will just open an existing
 			 * semaphore without any indication. Sometimes the other side of the shared memory
 			 * will provide incorrect parameters or will otherwise fail and its faster to
 			 * restart this program than to get that application to release the semaphores */
 			/* TODO(rnp): figure out something more robust */
 			os_w32_semaphore_release(os_w32_shared_memory_semaphores[it], 1);
 		}
 	}
 	#endif
 
 	GLWorkerThreadContext *worker = &ctx->compute_worker;
 	/* TODO(rnp): we should lock this down after we have something working */
 	worker->user_context = (iptr)ctx;
 	worker->handle       = os_create_thread("[compute]", worker, compute_worker_thread_entry_point);
 
 	GLWorkerThreadContext         *upload = &ctx->upload_worker;
 	BeamformerUploadThreadContext *upctx  = push_struct(&memory, typeof(*upctx));
 	upload->user_context        = (iptr)upctx;
 	upctx->rf_buffer            = &cs->rf_buffer;
 	upctx->shared_memory        = ctx->shared_memory;
 	upctx->shared_memory_size   = ctx->shared_memory_size;
 	upctx->compute_timing_table = ctx->compute_timing_table;
 	upctx->compute_worker_sync  = &ctx->compute_worker.sync_variable;
 	upload->handle = os_create_thread("[upload]", upload, beamformer_upload_entry_point);
 
 	/* NOTE: set up OpenGL debug logging */
 	Stream *gl_error_stream = push_struct(&memory, Stream);
 	*gl_error_stream        = stream_alloc(&memory, 1024);
 	glDebugMessageCallback(gl_debug_logger, gl_error_stream);
 #ifdef _DEBUG
 	glEnable(GL_DEBUG_OUTPUT);
 #endif
 
 	if (!BakeShaders)
 	{
 		for EachElement(beamformer_reloadable_compute_shader_info_indices, it) {
 			i32   index = beamformer_reloadable_compute_shader_info_indices[it];
 			Arena temp  = scratch;
 			str8 file = push_str8_from_parts(&temp, os_path_separator(), str8("shaders"),
 			                                 beamformer_reloadable_shader_files[index][0]);
 			BeamformerFileReloadContext *frc = push_struct(&memory, typeof(*frc));
 			frc->kind                 = BeamformerFileReloadKind_ComputeShader;
 			frc->shader_reload.shader = beamformer_reloadable_shader_kinds[index];
 			os_add_file_watch((char *)file.data, file.length, frc);
 		}
 
 		for EachElement(beamformer_reloadable_compute_helpers_shader_info_indices, it) {
 			i32   index = beamformer_reloadable_compute_helpers_shader_info_indices[it];
 			Arena temp  = scratch;
 			str8 file = push_str8_from_parts(&temp, os_path_separator(), str8("shaders"),
 			                                 beamformer_reloadable_shader_files[index][0]);
 			BeamformerFileReloadContext *frc = push_struct(&memory, typeof(*frc));
 			frc->kind                 = BeamformerFileReloadKind_ComputeShader;
 			frc->shader_reload.shader = beamformer_reloadable_shader_kinds[index];
 			os_add_file_watch((char *)file.data, file.length, frc);
 		}
 	}
 
 	memory.end = scratch.end;
 	ctx->arena = memory;
 	ctx->state = BeamformerState_Running;
 }
 
 BEAMFORMER_EXPORT void
 beamformer_terminate(BeamformerInput *input)
 {
 	/* NOTE(rnp): work around pebkac when the beamformer is closed while we are doing live
 	 * imaging. if the verasonics is blocked in an external function (calling the library
 	 * to start compute) it is impossible for us to get it to properly shut down which
 	 * will sometimes result in us needing to power cycle the system. set the shared memory
 	 * into an error state and release dispatch lock so that future calls will error instead
 	 * of blocking.
 	 */
 	BeamformerCtx *          ctx = BeamformerContextMemory(input->memory);
 	BeamformerSharedMemory * sm  = input->shared_memory;
 	if (ctx->state != BeamformerState_Terminated) {
 		if (sm) {
 			BeamformerSharedMemoryLockKind lock = BeamformerSharedMemoryLockKind_DispatchCompute;
 			atomic_store_u32(&sm->invalid, 1);
 			atomic_store_u32(&sm->external_work_queue.ridx, sm->external_work_queue.widx);
 			DEBUG_DECL(if (sm->locks[lock])) {
 				beamformer_shared_memory_release_lock(sm, (i32)lock);
 			}
 
 			atomic_or_u32(&sm->live_imaging_dirty_flags, BeamformerLiveImagingDirtyFlags_StopImaging);
 		}
 
 		beamformer_debug_ui_deinit(ctx);
 
 		ctx->state = BeamformerState_Terminated;
 	}
 }
 
 BEAMFORMER_EXPORT u32
 beamformer_should_close(BeamformerInput *input)
 {
 	BeamformerCtx * ctx = BeamformerContextMemory(input->memory);
 	if (ctx->state == BeamformerState_ShouldClose)
 		beamformer_terminate(input);
 	return ctx->state == BeamformerState_Terminated;
 }