ogl_beamforming

beamformer_core.c (84914B)

---

 /* See LICENSE for license details. */
 /* TODO(rnp):
  * [ ]: backtrace dumping on SIGSEGV
  * [ ]: cooperative shared memory loading in decode shader
  * [ ]: refactor: save filter parameters with rest of parameters, whole slot thing is dumb
  * [ ]: upload previously exported data for display. maybe this is a UI thing but doing it
  *      programatically would be nice.
  * [ ]: Add interface for multi frame upload. RF upload already uses an offset into SM so
  *      that part works fine. We just need a way of specify a multi frame upload. (Data must
  *      be organized for simple offset access per frame).
  * [ ]: refactor: do_compute should build its own "command graph" which tracks
  *      dependencies better. It is very important that unnecessary barriers are
  *      not placed between compute stages which requires knowledge of the entire
  *      graph.
  * [ ]: refactor: replace UploadRF with just the scratch_rf_size variable,
  *      use below to spin wait in library
  * [ ]: utilize umonitor/umwait (intel), monitorx/mwaitx (amd), and wfe/sev (aarch64)
  *      for power efficient low latency waiting
  * [ ]: BeamformWorkQueue -> BeamformerWorkQueue
  * [ ]: refactor: work queue needs a cleanup, we should only have a single one
  *      - that queue isn't really considered hot so a lock is probably fine
  * [ ]: bug: reinit cuda on hot-reload
  *
  * [ ]: export for special frames/data
  *    - Color Map Data
  *    - Recursive Imaging Result
  *    - Incoherent sum
  *
  * [ ]: Tiled Array Handling
  *    [ ]: add tile count uv2
  *    [ ]: make xdc_transform an array
  *    [ ]: modify CPU DAS dispatch code to lookup xdc_transform based on tile index
  *    [ ]: modify CPU DAS dispatch code to set rf_element_offset based on tile index
  *       - need to check if this works for HERCULES or if the shader just needs to
  *         know about the extra tiles
  *    [ ]: write simple backpropagation code to optimize 2D tile position and 2D tilt
  *         (4 variables per tile).
  *       - two tests:
  *         1. Maximize value of wire target
  *         2. Maximize value of single cyst contrast
  *
  * [ ]: Recursive Imaging Handling
  *    [ ]: add array of image offsets to Compute Array Parameters
  *    [ ]: add array of weighting coeffiecents to Compute Array Parameters
  *    [ ]: Update 2D sum shader to use these
  *
  * [ ]: Power Doppler
  *    [ ]: also needs array of image offsets
  *    [ ]: make modified filter that grabs each sample from a different image offset
  *       - NOTE: this can't use existing striding mechanism because one image may
  *               be at the start of the ring buffer and another at the end of the
  *               ring buffer and the image size may not cleanly divide the ring buffer size.
  *       - Probably want this shader to just output the estimated power map directly.
  *    [ ]: make a modified render shader that takes two images: one structural and one that
  *         is used to index into a color map. (or second render pass that applies color overlay)
  */
 
 #include "base_platform.h"
 
 #if defined(BEAMFORMER_DEBUG) && !defined(BEAMFORMER_EXPORT) && OS_WINDOWS
   #define BEAMFORMER_EXPORT __declspec(dllexport)
 #endif
 
 #include "beamformer_internal.h"
 
 #define GPU_RESOURCE_HASH_TABLE_COUNT 256
 
 typedef struct GPUResource GPUResource;
 struct GPUResource {
 	str8 name;
 	u64  size;
 	u64  offset;
 	u64  alignment;
 
 	void *data;
 
 	u64  hash;
 
 	GPUResource *next;
 	GPUResource *hash_next, *hash_prev;
 };
 typedef struct {GPUResource *first, *last;} GPUResourceHashBucket;
 
 typedef struct {
 	Arena *arena;
 	u64    position;
 
 	GPUResource *resource_list;
 	GPUResourceHashBucket hash_table[GPU_RESOURCE_HASH_TABLE_COUNT];
 } GPUResourceBuilder;
 
 read_only BeamformerFrame       beamformer_nil_frame;
 read_only BeamformerComputePlan beamformer_nil_compute_plan;
 
 global BeamformerCtx   *beamformer_context;
 global BeamformerInput *beamformer_input;
 global f32 dt_for_frame;
 
 #define beamformer_frame_arena() (beamformer_context->frame_arenas[beamformer_context->frame_index % countof(beamformer_context->frame_arenas)])
 #define beamformer_registers() (&beamformer_context->registers->v)
 #define beamformer_push_registers(...) beamformer_push_registers_(&(BeamformerRegisters){beamformer_registers_init_literal __VA_ARGS__})
 #define BeamformerRegistersScope(...) DeferLoop(beamformer_push_registers(__VA_ARGS__), beamformer_pop_registers())
 #define beamformer_command(name, ...) beamformer_push_command(name, &(BeamformerRegisters){beamformer_registers_init_literal __VA_ARGS__})
 
 function BeamformerRegisters *
 beamformer_pop_registers(void)
 {
 	BeamformerRegisters *result = &beamformer_context->registers->v;
 	SLLStackPop(beamformer_context->registers, next);
 	if (beamformer_context->registers == 0)
 		beamformer_context->registers = &beamformer_context->base_registers;
 	return result;
 }
 
 function BeamformerRegisters *
 beamformer_push_registers_(BeamformerRegisters *registers)
 {
 	BeamformerRegistersNode *node   = push_struct(beamformer_frame_arena(), BeamformerRegistersNode);
 	BeamformerRegisters     *result = &node->v;
 	memory_copy(result, registers, sizeof(node->v));
 	SLLStackPush(beamformer_context->registers, node, next);
 	return result;
 }
 
 function void
 beamformer_command_list_push_new(Arena *arena, BeamformerCommandList *commands, str8 name, BeamformerRegisters *registers)
 {
 	BeamformerCommandNode *node = push_struct(arena, BeamformerCommandNode);
 	node->command.registers = push_struct_no_zero(arena, BeamformerRegisters);
 	node->command.name      = push_str8(arena, name);
 	memory_copy(node->command.registers, registers, sizeof(*registers));
 	DLLInsertLast(0, commands->first, commands->last, node, next, prev);
 	commands->count += 1;
 }
 
 function void
 beamformer_push_command(str8 name, BeamformerRegisters *registers)
 {
 	beamformer_command_list_push_new(beamformer_frame_arena(), beamformer_context->command_queues + 0,
 	                                 name, registers);
 }
 
 function BeamformerCommandKind
 beamformer_command_kind_from_string(str8 s)
 {
 	BeamformerCommandKind result = BeamformerCommandKind_Nil;
 	for EachElement(beamformer_command_infos, it) {
 		if (str8_equal(beamformer_command_infos[it].string, s)) {
 			result = (BeamformerCommandKind)it;
 			break;
 		}
 	}
 	return result;
 }
 
 function BeamformerPanelKind
 beamformer_panel_kind_from_string(str8 s)
 {
 	BeamformerPanelKind result = BeamformerPanelKind_Nil;
 	for EachElement(beamformer_panel_infos, it) {
 		if (str8_equal(beamformer_panel_infos[it].string, s)) {
 			result = (BeamformerPanelKind)it;
 			break;
 		}
 	}
 	return result;
 }
 
 function BeamformerFrame *
 beamformer_frame_from_index(u64 index)
 {
 	BeamformerFrame *result = (BeamformerFrame *)&beamformer_nil_frame;
 	if (index < countof(beamformer_context->compute_context.backlog.frames)) {
 		BeamformerFrame *frame = beamformer_context->compute_context.backlog.frames + index;
 		if (frame->timeline_valid_value != 0)
 			result = frame;
 	}
 	return result;
 }
 
 function b32
 beamformer_frame_valid(u64 index)
 {
 	b32 result = beamformer_frame_from_index(index) != &beamformer_nil_frame;
 	return result;
 }
 
 function void
 beamformer_compute_plan_release(BeamformerComputeContext *cc, u32 block)
 {
 	assert(block < countof(cc->compute_plans));
 	BeamformerComputePlan *cp = cc->compute_plans[block];
 	if (cp) {
 		gpu_buffer_release(&cp->gpu_temp_arena);
 		cc->compute_plans[block] = 0;
 		SLLPushFreelist(cp, cc->compute_plan_freelist);
 	}
 }
 
 function GPUResource *
 gpu_resource_from_hash(GPUResourceBuilder *rb, u64 hash)
 {
 	GPUResource *result = 0;
 
 	GPUResourceHashBucket *hb = rb->hash_table + (hash % GPU_RESOURCE_HASH_TABLE_COUNT);
 	for (GPUResource *r = hb->first; r; r = r->hash_next) {
 		if (hash == r->hash) {
 			result = r;
 			break;
 		}
 	}
 
 	return result;
 }
 
 typedef struct {
 	str8  name;
 	u64   align;
 	u64   size;
 	void *data;
 } GPUResourcePushInfo;
 #define gpu_resource_push(rb, t, count, ...) gpu_resource_push_(rb, (GPUResourcePushInfo){\
 	.align = Max(alignof(t), 16), \
 	.size  = sizeof(t) * count, \
 	__VA_ARGS__})
 
 function u32
 gpu_resource_push_(GPUResourceBuilder *rb, GPUResourcePushInfo info)
 {
 	assert(info.size > 0 && info.name.length > 0 && IsPowerOfTwo(info.align));
 
 	u64 hash = u64_hash_from_str8(info.name);
 	GPUResource *r = gpu_resource_from_hash(rb, hash);
 	if (!r) {
 		r = push_struct(rb->arena, GPUResource);
 		GPUResourceHashBucket *hb = rb->hash_table + (hash % GPU_RESOURCE_HASH_TABLE_COUNT);
 		DLLInsert(0, hb->first, hb->last, r, hash_next, hash_prev);
 		SLLStackPush(rb->resource_list, r, next);
 
 		r->hash      = hash;
 		r->name      = info.name;
 		r->alignment = Max(16, info.align);
 		r->offset    = AlignUpPowerOfTwo(rb->position, r->alignment);
 		r->size      = info.size;
 		rb->position = r->offset + r->size;
 	}
 
 	u32 result = r->offset;
 
 	// NOTE(rnp): if this is a new resource and no data is provided it is likely
 	// a temporary GPU side buffer. if this is not a new resource and no data
 	// is provided then maybe it is shared and someone else will provide it
 	if (info.data) r->data = info.data;
 
 	return result;
 }
 
 function GPUResourceBuilder *
 gpu_resource_build_begin(Arena *arena)
 {
 	GPUResourceBuilder *result = push_struct(arena, GPUResourceBuilder);
 	result->arena = arena;
 	return result;
 }
 
 function void
 gpu_resource_build_end(GPUResourceBuilder *rb, GPUBuffer *buffer)
 {
 	u64 size = gpu_round_up_to_sync_size(rb->position, 64);
 	if (size != (u64)buffer->size) {
 		gpu_buffer_allocate(buffer, (GPUBufferAllocateInfo){
 			.size  = size,
 			.flags = GPUUsageFlag_HostWrite,
 			.label = push_str8_f(rb->arena, "GPU Temp Arena [%p]", buffer),
 			.timelines_used = (GPUTimeline []){GPUTimeline_Compute},
 			.timeline_count = 1,
 		});
 	}
 
 	//////////////////////////////////////
 	// NOTE(rnp): upload data
 	u64 last_wait_value = 0;
 	for (GPUResource *r = rb->resource_list; r; r = r->next)
 		if (r->data)
 			last_wait_value = gpu_buffer_range_upload(buffer, r->data, r->offset, r->size, 0);
 
 	// TODO(rnp): cleanup this pointless stall
 	if (gpu_buffer_needs_sync(buffer))
 		gpu_host_wait_timeline(GPUTimeline_Transfer, last_wait_value, -1ULL);
 }
 
 function BeamformerComputePlan *
 beamformer_compute_plan_for_block(BeamformerComputeContext *cc, u32 block, Arena *arena)
 {
 	assert(block < countof(cc->compute_plans));
 	BeamformerComputePlan *result = cc->compute_plans[block];
 	if (!result) {
 		result = SLLPopFreelist(cc->compute_plan_freelist);
 		if (!result) result = push_struct_no_zero(arena, BeamformerComputePlan);
 		zero_struct(result);
 		cc->compute_plans[block] = result;
 
 		result->ui_voxel_transform = m4_identity();
 	}
 	return result;
 }
 
 function BeamformerFilter *
 beamformer_filter_create(Arena *arena, BeamformerFilterParameters fp)
 {
 	BeamformerFilter *result = push_struct(arena, BeamformerFilter);
 	switch (fp.kind) {
 	case BeamformerFilterKind_Kaiser:{
 		/* TODO(rnp): this should also support complex */
 		/* TODO(rnp): implement this as an IFIR filter instead to reduce computation */
 		result->data = kaiser_low_pass_filter(arena, fp.kaiser.cutoff_frequency, fp.sampling_frequency,
 		                                      fp.kaiser.beta, (i32)fp.kaiser.length);
 		result->length     = (i32)fp.kaiser.length;
 		result->time_delay = (f32)result->length / 2.0f / fp.sampling_frequency;
 	}break;
 
 	case BeamformerFilterKind_MatchedChirp:{
 		typeof(fp.matched_chirp) *mc = &fp.matched_chirp;
 		f32 fs = fp.sampling_frequency;
 		result->length = (i32)(mc->duration * fs);
 		if (fp.complex) {
 			result->data = baseband_chirp(arena, mc->min_frequency, mc->max_frequency, fs, result->length, 1, 0.5f);
 			result->time_delay = complex_filter_first_moment(result->data, result->length, fs);
 		} else {
 			result->data = rf_chirp(arena, mc->min_frequency, mc->max_frequency, fs, result->length, 1);
 			result->time_delay = real_filter_first_moment(result->data, result->length, fs);
 		}
 	}break;
 
 	InvalidDefaultCase;
 	}
 
 	result->parameters = fp;
 	return result;
 }
 
 function iv3
 das_valid_points(iv3 points)
 {
 	iv3 result;
 	result.x = Max(points.x, 1);
 	result.y = Max(points.y, 1);
 	result.z = Max(points.z, 1);
 	return result;
 }
 
 function GPUBuffer *
 beamformer_gpu_buffer_from_frame(BeamformerFrame *frame)
 {
 	GPUBuffer *result = beamformer_context->compute_context.backlog.buffer;
 	if (!Between(frame->gpu_pointer, result->gpu_pointer, result->gpu_pointer + result->size)) {
 		result = 0;
 		BeamformerComputePlan *cp = beamformer_context->compute_context.compute_plans[frame->parameter_block];
 		if (cp) {
 			result = &cp->gpu_temp_arena;
 			assert(Between(frame->gpu_pointer, result->gpu_pointer, result->gpu_pointer + result->size));
 		}
 	}
 	return result;
 }
 
 function u64
 beamformer_frame_byte_size(iv3 points, BeamformerDataKind kind)
 {
 	u64 result = points.x * points.y * points.z * beamformer_data_kind_byte_size[kind];
 	result = round_up_to(result, 64);
 	return result;
 }
 
 function u64
 beamformer_incoherent_frame_byte_size(iv3 points, BeamformerDataKind kind)
 {
 	u64 result = beamformer_frame_byte_size(points, kind) / beamformer_data_kind_element_count[kind];
 	return result;
 }
 
 function BeamformerFrame *
 beamformer_frame_next(BeamformerComputeContext *cc, iv3 output_points, b32 complex)
 {
 	BeamformerFrameBacklog *bl = &cc->backlog;
 
 	BeamformerDataKind kind = complex ? BeamformerDataKind_Float32Complex : BeamformerDataKind_Float32;
 	u64 frame_size = beamformer_frame_byte_size(output_points, kind);
 
 	// TODO(rnp): handle this somewhat gracefully (even it produces garbled output)
 	assert(frame_size <= (u64)bl->buffer->size);
 
 	if (bl->next_offset > (u64)bl->buffer->size - frame_size)
 		bl->next_offset = 0;
 
 	u64 id = bl->counter++;
 
 	BeamformerFrame *result = bl->frames + (id % countof(bl->frames));
 	atomic_store_u64(&result->timeline_valid_value, -1ULL);
 	result->id            = id & U32_MAX;
 	result->gpu_pointer   = bl->buffer->gpu_pointer + bl->next_offset;
 	result->points        = output_points;
 	result->data_kind     = kind;
 
 	bl->next_offset += frame_size;
 
 	return result;
 }
 
 function void
 push_compute_timing_info(ComputeTimingTable *t, ComputeTimingInfo info)
 {
 	u32 index = atomic_add_u32(&t->write_index, 1) % countof(t->buffer);
 	t->buffer[index] = info;
 }
 
 function uv3
 layout_for_output(iv3 points)
 {
 	uv3 result = {{1, 1, 1}};
 
 	b32 has_x = points.x > 1;
 	b32 has_y = points.y > 1;
 	b32 has_z = points.z > 1;
 
 	u32 subgroup_size  = gpu_info()->subgroup_size;
 	u32 grid_3d_z_size = Max(1, subgroup_size / (4 * 4));
 	u32 grid_2d_y_size = Max(1, subgroup_size / 8);
 
 	switch (iv3_dimension(points)) {
 	case 1:{
 		if (has_x) result.x = subgroup_size;
 		if (has_y) result.y = subgroup_size;
 		if (has_z) result.z = subgroup_size;
 	}break;
 
 	case 2:{
 		if (has_x && has_y) {result.x = 8; result.y = grid_2d_y_size;}
 		if (has_x && has_z) {result.x = 8; result.z = grid_2d_y_size;}
 		if (has_y && has_z) {result.y = 8; result.z = grid_2d_y_size;}
 	}break;
 
 	case 3:{result = (uv3){{4, 4, grid_3d_z_size}};}break;
 
 	InvalidDefaultCase;
 	}
 
 	return result;
 }
 
 function uv3
 dispatch_for_output(uv3 layout, iv3 points)
 {
 	uv3 result;
 	result.x = (u32)ceil_f32((f32)points.x / layout.x);
 	result.y = (u32)ceil_f32((f32)points.y / layout.y);
 	result.z = (u32)ceil_f32((f32)points.z / layout.z);
 	return result;
 }
 
 typedef struct BeamformerComputeGraphNode BeamformerComputeGraphNode;
 struct BeamformerComputeGraphNode {
 	// NOTE(rnp): will be BeamformerShaderKind_Count for root node
 	BeamformerShaderKind kind;
 
 	// NOTE(rnp): when any of input or output stride is assigned it is assumed that
 	// the shader requires a fixed layout for input, output, or both. When two adjacent
 	// nodes require incompatible layouts the second pass over the graph will insert
 	// Reshape shaders in between.
 	BeamformerDataKind   input_data_kind;
 	BeamformerDataLayout input_data_layout;
 
 	BeamformerDataKind   output_data_kind;
 	BeamformerDataLayout output_data_layout;
 
 	uv3 rf_dimensions;
 
 	b32 requires_fp_input;
 
 	i32 user_pipeline_index;
 
 	BeamformerComputeGraphNode *prev;
 	BeamformerComputeGraphNode *next;
 };
 
 typedef struct {
 	BeamformerComputeGraphNode *first;
 	BeamformerComputeGraphNode *last;
 	u64                         count;
 } BeamformerComputeGraph;
 
 function b32
 compute_plan_push_shader(BeamformerComputePlan *p, BeamformerComputeGraphNode *node, BeamformerShaderParameters *sp)
 {
 	b32 result = 0;
 	if (p->pipeline.shader_count < countof(p->pipeline.shaders)) {
 		u32 index = p->pipeline.shader_count++;
 		p->pipeline.shaders[index]    = node->kind;
 		zero_struct(p->shader_descriptors + index);
 		p->pipeline.parameters[index] = sp ? *sp : (BeamformerShaderParameters){0};
 
 		p->shader_descriptors[index].input_data_kind  = node->input_data_kind;
 		p->shader_descriptors[index].output_data_kind = node->output_data_kind;
 
 		result = 1;
 	}
 	return result;
 }
 
 function BeamformerComputeGraphNode *
 push_compute_graph_node(BeamformerComputeGraph *graph, BeamformerShaderKind kind, Arena *arena)
 {
 	BeamformerComputeGraphNode *result = push_struct(arena, BeamformerComputeGraphNode);
 	if (graph) {
 		if (graph->last) result->rf_dimensions = graph->last->rf_dimensions;
 		DLLInsertLast(0, graph->first, graph->last, result, next, prev);
 		graph->count++;
 	}
 	result->kind = kind;
 	result->user_pipeline_index = -1;
 	// NOTE(rnp): initially don't care data parameters
 	result->input_data_kind    = BeamformerDataKind_Count;
 	result->output_data_kind   = BeamformerDataKind_Count;
 	result->input_data_layout  = BeamformerDataLayout_Count;
 	result->output_data_layout = BeamformerDataLayout_Count;
 	return result;
 }
 
 function uv3
 beamformer_data_strides(BeamformerDataLayout layout, uv3 rf_dimensions)
 {
 	u32 channel_count = rf_dimensions.E[BeamformerRFDimension_Channels];
 	u32 event_count   = rf_dimensions.E[BeamformerRFDimension_Events];
 	u32 sample_count  = rf_dimensions.E[BeamformerRFDimension_Samples];
 
 	uv3 result = {0};
 	switch (layout) {
 	InvalidDefaultCase;
 
 	case BeamformerDataLayout_Image:{}break;
 
 	case BeamformerDataLayout_ChannelEventSample:{
 		result.E[BeamformerRFDimension_Channels] = event_count * sample_count;
 		result.E[BeamformerRFDimension_Events]   = sample_count;
 		result.E[BeamformerRFDimension_Samples]  = 1;
 	}break;
 
 	case BeamformerDataLayout_ChannelSampleEvent:{
 		result.E[BeamformerRFDimension_Channels] = event_count * sample_count;
 		result.E[BeamformerRFDimension_Events]   = 1;
 		result.E[BeamformerRFDimension_Samples]  = event_count;
 	}break;
 
 	case BeamformerDataLayout_EventChannelSample:{
 		result.E[BeamformerRFDimension_Events]   = channel_count * sample_count;
 		result.E[BeamformerRFDimension_Channels] = sample_count;
 		result.E[BeamformerRFDimension_Samples]  = 1;
 	}break;
 
 	case BeamformerDataLayout_SampleChannelEvent:{
 		result.E[BeamformerRFDimension_Samples]  = channel_count * event_count;
 		result.E[BeamformerRFDimension_Channels] = event_count;
 		result.E[BeamformerRFDimension_Events]   = 1;
 	}break;
 
 	}
 
 	return result;
 }
 
 function void
 plan_compute_pipeline(BeamformerComputePlan *cp, BeamformerParameterBlock *pb, Arena *scratch)
 {
 	b32 run_hilbert = 0;
 	b32 demodulate  = 0;
 
 	for (u32 i = 0; i < pb->pipeline.shader_count; i++) {
 		switch (pb->pipeline.shaders[i]) {
 		case BeamformerShaderKind_Hilbert:{run_hilbert = 1;}break;
 		case BeamformerShaderKind_Demodulate:{demodulate = 1;}break;
 		default:{}break;
 		}
 	}
 
 	if (demodulate) run_hilbert = 0;
 
 	f32 sampling_frequency = pb->parameters.sampling_frequency;
 	u32 input_sample_count = pb->parameters.sample_count;
 	u32 acquisition_count  = pb->parameters.acquisition_count;
 	u32 decimation_rate    = Max(pb->parameters.decimation_rate, 1);
 
 	cp->raw_channel_byte_stride = pb->parameters.sample_count * pb->parameters.acquisition_count
 	                              * beamformer_data_kind_byte_size[pb->pipeline.data_kind];
 
 	BeamformerDataKind input_data_kind = pb->pipeline.data_kind;
 	if (demodulate) {
 		switch (input_data_kind) {
 		case BeamformerDataKind_Int16:{  input_data_kind = BeamformerDataKind_Int16Complex;  }break;
 		case BeamformerDataKind_Float16:{input_data_kind = BeamformerDataKind_Float16Complex;}break;
 		case BeamformerDataKind_Float32:{input_data_kind = BeamformerDataKind_Float32Complex;}break;
 		default:{}break;
 		}
 		input_sample_count /= (2 * decimation_rate);
 		sampling_frequency /= (2 * decimation_rate);
 	}
 
 	cp->iq_pipeline = beamformer_data_kind_complex[input_data_kind] || run_hilbert;
 
 	BeamformerDataKind das_data_kind = cp->iq_pipeline ? BeamformerDataKind_Float32Complex
 	                                                   : BeamformerDataKind_Float32;
 
 	cp->channel_count = pb->parameters.channel_count;
 	u32 chunk_channel_count = Min(cp->channel_count, BeamformerChunkChannelCount);
 
 	cp->rf_size = input_sample_count * pb->parameters.acquisition_count * chunk_channel_count
 	              * beamformer_data_kind_byte_size[das_data_kind];
 
 	i64 buffer_size = PING_PONG_BUFFER_SLOTS * round_up_to(cp->rf_size, 64);
 	if (beamformer_context->compute_context.ping_pong_buffer.size < buffer_size) {
 		b32 cuda = cuda_supported();
 		GPUBufferAllocateInfo allocate_info = {
 			.size   = buffer_size,
 			.export = cuda ? &beamformer_context->compute_context.ping_pong_export_handle : 0,
 			.label  = str8("PingPongBuffer"),
 		};
 		gpu_buffer_allocate(&beamformer_context->compute_context.ping_pong_buffer, allocate_info);
 
 		// TODO(rnp): figure out how to share with CUDA
 		// IMPORTANT: on linux the handle is returned to os and should be cleared after import
 		// see usage of glImportMemoryFdEXT and surrounding code in ui.c for examples
 		if (cuda) {
 		}
 	}
 
 	read_only BeamformerDataKind data_kind_to_element_kind[] = {
 		[BeamformerDataKind_Int16]          = BeamformerDataKind_Float16,
 		[BeamformerDataKind_Float16]        = BeamformerDataKind_Float16,
 		[BeamformerDataKind_Float32]        = BeamformerDataKind_Float32,
 		[BeamformerDataKind_Int16Complex]   = BeamformerDataKind_Float16,
 		[BeamformerDataKind_Float16Complex] = BeamformerDataKind_Float16,
 		[BeamformerDataKind_Float32Complex] = BeamformerDataKind_Float32,
 	};
 
 	read_only BeamformerDataKind data_kind_to_fp_kind[] = {
 		[BeamformerDataKind_Int16]          = BeamformerDataKind_Float16,
 		[BeamformerDataKind_Float16]        = BeamformerDataKind_Float16,
 		[BeamformerDataKind_Float32]        = BeamformerDataKind_Float32,
 		[BeamformerDataKind_Int16Complex]   = BeamformerDataKind_Float16Complex,
 		[BeamformerDataKind_Float16Complex] = BeamformerDataKind_Float16Complex,
 		[BeamformerDataKind_Float32Complex] = BeamformerDataKind_Float32Complex,
 	};
 
 	//////////////////////////////////////
 	// NOTE(rnp): First Pass: build initial graph and insert hard layout constraints
 	BeamformerComputeGraph graph = {0};
 	BeamformerComputeGraphNode *root_node = push_compute_graph_node(&graph, BeamformerShaderKind_Count, scratch);
 	root_node->input_data_kind    = input_data_kind;
 	root_node->input_data_layout  = BeamformerDataLayout_ChannelEventSample;
 	root_node->output_data_kind   = input_data_kind;
 	root_node->output_data_layout = BeamformerDataLayout_ChannelEventSample;
 
 	root_node->rf_dimensions.E[BeamformerRFDimension_Channels] = chunk_channel_count;
 	root_node->rf_dimensions.E[BeamformerRFDimension_Events]   = pb->parameters.acquisition_count;
 	root_node->rf_dimensions.E[BeamformerRFDimension_Samples]  = pb->parameters.sample_count;
 
 	for EachIndex(pb->pipeline.shader_count, it) {
 		// NOTE(rnp): skip unnecessary shaders
 		switch (pb->pipeline.shaders[it]) {
 		case BeamformerShaderKind_Hilbert:{if (!run_hilbert) continue;}break;
 
 		case BeamformerShaderKind_Decode:{
 			if (pb->parameters.decode_mode == BeamformerDecodeMode_None)
 				continue;
 		}break;
 
 		case BeamformerShaderKind_Sum:
 		case BeamformerShaderKind_MinMax:
 		{
 			// NOTE(rnp): currently unsupported
 			continue;
 		}break;
 
 		default:{}break;
 		}
 
 		BeamformerComputeGraphNode *node = push_compute_graph_node(&graph, pb->pipeline.shaders[it], scratch);
 		node->user_pipeline_index = (i32)it;
 		switch (pb->pipeline.shaders[it]) {
 		case BeamformerShaderKind_Demodulate:{
 			node->rf_dimensions.E[BeamformerRFDimension_Samples] /= (2 * decimation_rate);
 		}break;
 
 		case BeamformerShaderKind_Decode:{
 			b32 low_precision   = beamformer_data_kind_element_size[input_data_kind] < 4;
 			b32 use_coop_matrix = gpu_info()->cooperative_matrix &&
 			                      low_precision &&
 			                      (acquisition_count   % 16 == 0) &&
 			                      (chunk_channel_count % 16 == 0);
 
 			node->requires_fp_input = 1;
 
 			// NOTE(rnp): fixed input layout required for reasonable performance
 			node->input_data_layout = BeamformerDataLayout_SampleChannelEvent;
 
 			if (use_coop_matrix) {
 				node->input_data_kind    = BeamformerDataKind_Float16;
 				node->output_data_kind   = data_kind_to_element_kind[das_data_kind];
 				node->output_data_layout = node->input_data_layout;
 			}
 		}break;
 
 		case BeamformerShaderKind_DAS:{
 			// NOTE(rnp): if we aren't decoding we can let the previous stage's
 			// data kind pass through as long as it is floating point
 			node->requires_fp_input = 1;
 			if (pb->parameters.decode_mode != BeamformerDecodeMode_None)
 				node->input_data_kind = das_data_kind;
 
 			node->input_data_layout = BeamformerDataLayout_ChannelEventSample;
 			if (pb->parameters.acquisition_kind == BeamformerAcquisitionKind_RCA_TPW ||
 			    pb->parameters.acquisition_kind == BeamformerAcquisitionKind_RCA_VLS)
 			{
 				node->input_data_layout = BeamformerDataLayout_EventChannelSample;
 			}
 
 			node->output_data_layout = BeamformerDataLayout_Image;
 			node->output_data_kind   = das_data_kind;
 
 			// NOTE(rnp): insert implicit CoherencyWeighting node
 			if (pb->parameters.coherency_weighting)
 				node = push_compute_graph_node(&graph, BeamformerShaderKind_CoherencyWeighting, scratch);
 		}break;
 
 		default:{}break;
 		}
 	}
 
 	//////////////////////////////////////
 	// NOTE(rnp): Second Pass: resolve layout constraints
 	for (BeamformerComputeGraphNode *node = root_node->next; node; node = node->next) {
 		b32 needs_reshape = 0;
 
 		// NOTE(rnp): data strides
 		{
 			b32 input_dont_care       = node->input_data_layout        == BeamformerDataLayout_Count;
 			b32 prev_output_dont_care = node->prev->output_data_layout == BeamformerDataLayout_Count;
 
 			if (prev_output_dont_care && !input_dont_care)
 				node->prev->output_data_layout = node->input_data_layout;
 
 			if (!prev_output_dont_care && input_dont_care)
 				node->input_data_layout = node->prev->output_data_layout;
 
 			if (prev_output_dont_care && input_dont_care)
 				node->input_data_layout = node->prev->output_data_layout = node->prev->input_data_layout;
 
 			needs_reshape |= node->input_data_layout != node->prev->output_data_layout;
 		}
 
 		// NOTE(rnp): data kinds
 		{
 			b32 input_dont_care       = node->input_data_kind        == BeamformerDataKind_Count;
 			b32 prev_output_dont_care = node->prev->output_data_kind == BeamformerDataKind_Count;
 
 			if (prev_output_dont_care && !input_dont_care)
 				node->prev->output_data_kind = node->input_data_kind;
 
 			if (!prev_output_dont_care && input_dont_care)
 				node->input_data_kind = node->prev->output_data_kind;
 
 			if (prev_output_dont_care && input_dont_care)
 				node->input_data_kind = node->prev->output_data_kind = node->prev->input_data_kind;
 
 			if (node->requires_fp_input) {
 				node->input_data_kind = data_kind_to_fp_kind[node->input_data_kind];
 				if (prev_output_dont_care)
 					node->prev->output_data_kind = node->input_data_kind;
 			}
 
 			needs_reshape |= node->input_data_kind != node->prev->output_data_kind;
 		}
 
 		// NOTE(rnp): insert reshape if needed
 		if (needs_reshape) {
 			BeamformerComputeGraphNode *new = push_compute_graph_node(0, BeamformerShaderKind_Reshape, scratch);
 			BeamformerComputeGraphNode *last  = node->prev;
 			DLLInsertLast(0, node, last, new, next, prev);
 			graph.count++;
 			new->rf_dimensions      = new->prev->rf_dimensions;
 			new->input_data_kind    = new->prev->output_data_kind;
 			new->input_data_layout  = new->prev->output_data_layout;
 			new->output_data_kind   = new->next->input_data_kind;
 			new->output_data_layout = new->next->input_data_layout;
 		}
 	}
 
 	// NOTE(rnp): ensure last node descriptor gets proper values for output data kind
 	if (graph.last->output_data_kind == BeamformerDataKind_Count)
 		graph.last->output_data_kind = graph.last->input_data_kind;
 
 	f32 time_offset   = pb->parameters.time_offset;
 	u32 subgroup_size = gpu_info()->subgroup_size;
 
 	cp->first_image_shader_index = 0;
 	cp->pipeline.shader_count = 0;
 
 	GPUResourceBuilder *resource_builder = gpu_resource_build_begin(scratch);
 	for (BeamformerComputeGraphNode *node = root_node->next; node; node = node->next) {
 		assert(node->prev->output_data_kind == node->input_data_kind);
 		assert(node->prev->output_data_layout == node->input_data_layout);
 
 		BeamformerShaderParameters *sp = 0;
 		if (node->user_pipeline_index >= 0)
 			sp = pb->pipeline.parameters + node->user_pipeline_index;
 
 		if (compute_plan_push_shader(cp, node, sp)) {
 			BeamformerShaderDescriptor *sd = cp->shader_descriptors + cp->pipeline.shader_count - 1;
 
 			uv3 output_stride = beamformer_data_strides(node->output_data_layout, node->rf_dimensions);
 			uv3 input_stride  = beamformer_data_strides(node->input_data_layout,  node->prev->rf_dimensions);
 
 			switch (node->kind) {
 			case BeamformerShaderKind_Decode:{
 				BeamformerDecodeBakeParameters *db = &sd->bake.Decode;
 
 				u32 decode_sample_count = node->rf_dimensions.E[BeamformerRFDimension_Samples];
 				db->DecodeMode        = pb->parameters.decode_mode;
 				db->TransmitCount     = node->rf_dimensions.E[BeamformerRFDimension_Events];
 				db->ChunkChannelCount = node->rf_dimensions.E[BeamformerRFDimension_Channels];
 
 				// NOTE(rnp): ignored when using coop matrices
 				db->OutputSampleStride   = output_stride.E[BeamformerRFDimension_Samples];
 				db->OutputChannelStride  = output_stride.E[BeamformerRFDimension_Channels];
 				db->OutputTransmitStride = output_stride.E[BeamformerRFDimension_Events];
 
 				db->ToProcess = 1;
 
 				b32 use_coop_matrix = gpu_info()->cooperative_matrix &&
 				                      node->input_data_kind == BeamformerDataKind_Float16 &&
 				                      (db->TransmitCount % 16 == 0) &&
 				                      (chunk_channel_count % 16 == 0);
 				if (use_coop_matrix) {
 					// TODO(rnp): shared memory for larger sizes
 					sd->layout = (uv3){{subgroup_size, 1, 1}};
 
 					if (demodulate)
 						decode_sample_count *= 2;
 
 					sd->compile_flags |= BeamformerDecodeCompileFlags_CooperativeMatrix;
 					db->CooperativeMatrixM = 16;
 					db->CooperativeMatrixN = 16;
 					db->CooperativeMatrixK = 16;
 
 					sd->dispatch.x = db->TransmitCount   / db->CooperativeMatrixN;
 					sd->dispatch.y = chunk_channel_count / db->CooperativeMatrixM;
 					sd->dispatch.z = decode_sample_count;
 				} else if (db->TransmitCount > 40) {
 					sd->compile_flags |= BeamformerDecodeCompileFlags_UseSharedMemory;
 
 					if (db->TransmitCount == 48)
 						db->ToProcess = db->TransmitCount / 16;
 
 					b32 use_16x  = db->TransmitCount == 48 || db->TransmitCount == 80 ||
 					               db->TransmitCount == 96 || db->TransmitCount == 160;
 					sd->layout.x = use_16x ? 16 : 32;
 					sd->layout.y = 4;
 					sd->layout.z = 1;
 
 					sd->dispatch.x = (u32)ceil_f32((f32)db->TransmitCount     / (f32)sd->layout.x / (f32)db->ToProcess);
 					sd->dispatch.y = (u32)ceil_f32((f32)db->ChunkChannelCount / (f32)sd->layout.y);
 					sd->dispatch.z = (u32)ceil_f32((f32)decode_sample_count   / (f32)sd->layout.z);
 				} else {
 					/* NOTE(rnp): register caching. using more threads will cause the compiler to do
 					 * contortions to avoid spilling registers. using less gives higher performance */
 					sd->layout = (uv3){{subgroup_size / 2, 1, 1}};
 
 					sd->dispatch.x = (u32)ceil_f32((f32)decode_sample_count   / (f32)sd->layout.x);
 					sd->dispatch.y = (u32)ceil_f32((f32)db->ChunkChannelCount / (f32)sd->layout.y);
 					sd->dispatch.z = 1;
 				}
 
 				sd->uses_heap = 1;
 				u32 order = db->TransmitCount;
 				db->Hadamard = gpu_resource_push(resource_builder, f16, order * order,
 				                                 .data = make_hadamard_transpose(scratch, order, use_coop_matrix),
 				                                 .name = str8("hadamard"));
 			}break;
 
 			case BeamformerShaderKind_Demodulate:
 			case BeamformerShaderKind_Filter:
 			{
 				b32 demod = node->kind == BeamformerShaderKind_Demodulate;
 				BeamformerFilter *f = beamformer_filter_create(scratch, cp->filter_parameters[sp->filter_slot]);
 
 				sd->compile_flags |= BeamformerFilterCompileFlags_Demodulate * demod;
 				sd->compile_flags |= BeamformerFilterCompileFlags_ComplexFilter * f->parameters.complex;
 
 				time_offset += f->time_delay;
 
 				BeamformerFilterBakeParameters *fb = &sd->bake.Filter;
 
 				sd->uses_heap = 1;
 				fb->FilterLength = (u32)f->length;
 				fb->FilterCoefficients = gpu_resource_push(resource_builder, f32, f->length * (f->parameters.complex ? 2 : 1),
 				                                           .data = f->data,
 				                                           .name = push_str8_f(scratch, "filter_%u", sp->filter_slot));
 
 				fb->SampleCount    = node->rf_dimensions.E[BeamformerRFDimension_Samples];
 				fb->DecimationRate = demod ? decimation_rate : 1;
 
 				b32 deinterleave =  beamformer_data_kind_complex[node->input_data_kind] &&
 				                   !beamformer_data_kind_complex[node->output_data_kind];
 				if (deinterleave)
 					fb->BatchSampleCount = node->rf_dimensions.x * node->rf_dimensions.y * node->rf_dimensions.z;
 
 				fb->OutputSampleStride   = output_stride.E[BeamformerRFDimension_Samples];
 				fb->OutputChannelStride  = output_stride.E[BeamformerRFDimension_Channels];
 				fb->OutputTransmitStride = output_stride.E[BeamformerRFDimension_Events];
 
 				fb->InputSampleStride    = input_stride.E[BeamformerRFDimension_Samples];
 				fb->InputChannelStride   = input_stride.E[BeamformerRFDimension_Channels];
 				fb->InputTransmitStride  = input_stride.E[BeamformerRFDimension_Events];
 
 				/* NOTE(rnp): when we are demodulating we pretend that the sampler was alternating
 				 * between sampling the I portion and the Q portion of an IQ signal. Therefore there
 				 * is an implicit decimation factor of 2 which must always be included. All code here
 				 * assumes that the signal was sampled in such a way that supports this operation.
 				 * To recover IQ[n] from the sampled data (RF[n]) we do the following:
 				 *   I[n]  = RF[n]
 				 *   Q[n]  = RF[n + 1]
 				 *   IQ[n] = I[n] - j*Q[n]
 				 */
 				if (demod) {
 					fb->DemodulationFrequency = pb->parameters.demodulation_frequency;
 					fb->SamplingFrequency     = pb->parameters.sampling_frequency / 2;
 				}
 
 				sd->layout     = (uv3){{subgroup_size, 1, 1}};
 				sd->dispatch.x = (u32)ceil_f32((f32)node->rf_dimensions.E[BeamformerRFDimension_Samples]  / (f32)sd->layout.x);
 				sd->dispatch.y = (u32)ceil_f32((f32)node->rf_dimensions.E[BeamformerRFDimension_Channels] / (f32)sd->layout.y);
 				sd->dispatch.z = (u32)ceil_f32((f32)node->rf_dimensions.E[BeamformerRFDimension_Events]   / (f32)sd->layout.z);
 			}break;
 
 			case BeamformerShaderKind_DAS:{
 				cp->first_image_shader_index = cp->pipeline.shader_count;
 
 				BeamformerDASBakeParameters *db = &sd->bake.DAS;
 				db->SamplingFrequency     = sampling_frequency;
 				db->DemodulationFrequency = pb->parameters.demodulation_frequency;
 				db->SpeedOfSound          = pb->parameters.speed_of_sound;
 				db->TimeOffset            = time_offset;
 				db->FNumber               = pb->parameters.f_number;
 				db->AcquisitionKind       = pb->parameters.acquisition_kind;
 				db->SampleCount           = node->rf_dimensions.E[BeamformerRFDimension_Samples];
 				db->ReceiveChannelCount   = pb->parameters.channel_count;
 				db->AcquisitionCount      = node->rf_dimensions.E[BeamformerRFDimension_Events];
 				db->ChunkChannelCount     = node->rf_dimensions.E[BeamformerRFDimension_Channels];
 				db->ChannelByteStride     = input_stride.E[BeamformerRFDimension_Channels]
 				                            * beamformer_data_kind_byte_size[node->input_data_kind];
 				db->AcquisitionByteStride = input_stride.E[BeamformerRFDimension_Events]
 				                            * beamformer_data_kind_byte_size[node->input_data_kind];
 				db->InterpolationMode     = pb->parameters.interpolation_mode;
 				db->TransmitAngle         = pb->parameters.focal_vector.E[0];
 				db->FocusDepth            = pb->parameters.focal_vector.E[1];
 				db->ReadiGroupCount       = pb->parameters.readi_group_count;
 				db->OutputSizeX           = cp->output_points.x;
 				db->OutputSizeY           = cp->output_points.y;
 				db->OutputSizeZ           = cp->output_points.z;
 				db->TransmitReceiveOrientation = pb->parameters.transmit_receive_orientation;
 
 				// NOTE(rnp): old gcc will miscompile an assignment
 				memory_copy(cp->xdc_transform.E, pb->parameters.xdc_transform.E, sizeof(cp->xdc_transform));
 
 				cp->voxel_transform   = m4_mul(cp->ui_voxel_transform, pb->parameters.das_voxel_transform);
 				cp->xdc_element_pitch = pb->parameters.xdc_element_pitch;
 
 				memory_copy(cp->das_voxel_transform.E, cp->voxel_transform.E, sizeof(cp->voxel_transform));
 
 				u32 id = pb->parameters.acquisition_kind;
 				b32 sparse = id == BeamformerAcquisitionKind_UFORCES || id == BeamformerAcquisitionKind_UHERCULES;
 				b32 single_focus       = pb->parameters.single_focus != 0;
 				b32 single_orientation = pb->parameters.single_orientation != 0;
 
 				sd->compile_flags |= BeamformerDASCompileFlags_CoherencyWeighting * (pb->parameters.coherency_weighting != 0);
 				sd->compile_flags |= BeamformerDASCompileFlags_SingleFocus        * single_focus;
 				sd->compile_flags |= BeamformerDASCompileFlags_SingleOrientation  * single_orientation;
 				sd->compile_flags |= BeamformerDASCompileFlags_Sparse             * sparse;
 
 				sd->layout   = layout_for_output(cp->output_points);
 				sd->dispatch = dispatch_for_output(sd->layout, cp->output_points);
 
 				if (id != BeamformerAcquisitionKind_UFORCES && id != BeamformerAcquisitionKind_FORCES && !single_focus) {
 					db->FocalVectors = gpu_resource_push(resource_builder, v2, db->AcquisitionCount,
 					                                     .data = pb->focal_vectors,
 					                                     .name = str8("focal_vectors"));
 					sd->uses_heap = 1;
 				}
 
 				if (id != BeamformerAcquisitionKind_UFORCES && id != BeamformerAcquisitionKind_FORCES && !single_orientation) {
 					db->TransmitReceiveOrientations = gpu_resource_push(resource_builder, u8, db->AcquisitionCount,
 					                                                    .data = pb->transmit_receive_orientations,
 					                                                    .name = str8("transmit_receive_orientations"));
 					sd->uses_heap = 1;
 				}
 
 				if (sparse) {
 					db->SparseElements = gpu_resource_push(resource_builder, i16, db->AcquisitionCount,
 					                                       .data = pb->sparse_elements,
 					                                       .name = str8("sparse_elements"));
 					sd->uses_heap = 1;
 				}
 
 				if (pb->parameters.coherency_weighting) {
 					db->IncoherentFrame = gpu_resource_push(resource_builder, u32, 0,
 					                                        .size = beamformer_incoherent_frame_byte_size(cp->output_points, das_data_kind),
 					                                        .name = str8("incoherent_buffer"));
 					sd->uses_heap = 1;
 				}
 
 				cp->readi_group = pb->parameters.readi_group;
 				if (db->ReadiGroupCount > 1) {
 					u32 order = db->ReadiGroupCount;
 					db->Hadamard = gpu_resource_push(resource_builder, f16, order * order,
 					                                 .data = make_hadamard_transpose(scratch, order, 0),
 					                                 .name = str8("readi_hadamard"));
 					sd->uses_heap = 1;
 				}
 			}break;
 
 			case BeamformerShaderKind_CoherencyWeighting:{
 				// NOTE(rnp): beamformed data is stored in linear order; making the layout 2D or 3D
 				// here is just slower
 				sd->layout   = (uv3){{subgroup_size, 1, 1}};
 				sd->dispatch = dispatch_for_output(sd->layout, cp->output_points);
 
 				BeamformerCoherencyWeightingBakeParameters *cw = &sd->bake.CoherencyWeighting;
 				cw->Scale         = 1.f;
 				cw->OutputVoxels  = cp->output_points.x * cp->output_points.y * cp->output_points.z;
 				cw->IncoherentSum = gpu_resource_push(resource_builder, u32, 0,
 				                                      .size = beamformer_incoherent_frame_byte_size(cp->output_points, das_data_kind),
 				                                      .name = str8("incoherent_buffer"));
 				sd->uses_heap = 1;
 			}break;
 
 			case BeamformerShaderKind_Reshape:{
 				BeamformerReshapeBakeParameters *rb = &sd->bake.Reshape;
 				b32 deinterleave =  beamformer_data_kind_complex[node->input_data_kind] &&
 				                   !beamformer_data_kind_complex[node->output_data_kind];
 				b32 interleave   = !beamformer_data_kind_complex[node->input_data_kind] &&
 				                    beamformer_data_kind_complex[node->output_data_kind];
 				assert(interleave == 0 || (interleave != deinterleave));
 				sd->compile_flags |= BeamformerReshapeCompileFlags_Deinterleave * deinterleave;
 				sd->compile_flags |= BeamformerReshapeCompileFlags_Interleave   * interleave;
 
 				rb->InputStrideX   = input_stride.x;
 				rb->InputStrideY   = input_stride.y;
 				rb->InputStrideZ   = input_stride.z;
 				rb->OutputStrideX  = output_stride.x;
 				rb->OutputStrideY  = output_stride.y;
 				rb->OutputStrideZ  = output_stride.z;
 
 				// NOTE(rnp): order doesn't really matter here but it must match the dispatch layout
 				rb->SizeX          = node->rf_dimensions.x;
 				rb->SizeY          = node->rf_dimensions.y;
 				rb->SizeZ          = node->rf_dimensions.z;
 
 				sd->layout.x = 1;
 				sd->layout.z = Min(subgroup_size, rb->SizeZ);
 				sd->layout.y = subgroup_size / sd->layout.z;
 
 				sd->dispatch.x = (u32)(ceil_f32((f32)rb->SizeX / sd->layout.x));
 				sd->dispatch.y = (u32)(ceil_f32((f32)rb->SizeY / sd->layout.y));
 				sd->dispatch.z = (u32)(ceil_f32((f32)rb->SizeZ / sd->layout.z));
 			}break;
 
 			default:{}break;
 
 			#if 0
 			case BeamformerShaderKind_Sum:{
 				sd->bake.data_kind = BeamformerDataKind_Float32;
 				if (cp->iq_pipeline)
 					sd->bake.data_kind = BeamformerDataKind_Float32Complex;
 
 				sd->layout   = layout_for_output(cp->output_points);
 				sd->dispatch = dispatch_for_output(sd->layout, cp->output_points);
 
 				commit = 1;
 			}break;
 			#endif
 
 			}
 		}
 	}
 
 	cp->pipeline.data_kind = input_data_kind;
 
 	if (cp->first_image_shader_index == 0)
 		cp->first_image_shader_index = cp->pipeline.shader_count;
 
 	gpu_resource_build_end(resource_builder, &cp->gpu_temp_arena);
 }
 
 function void
 stream_append_shader_header(Stream *s, i32 reloadable_index, u64 gpu_heap_pointer, BeamformerShaderDescriptor *sd, uv3 layout)
 {
 	stream_append_str8(s, str8("#version 460 core\n\n"
 	"#extension GL_EXT_buffer_reference : require\n"
 	"#extension GL_EXT_shader_16bit_storage : require\n"
 	"#extension GL_EXT_shader_explicit_arithmetic_types : require\n\n"
 	"#define f32     float32_t\n"
 	"#define f16     float16_t\n"
 	"#define s32     int32_t\n"
 	"#define u64     uint64_t\n"
 	"#define u32     uint32_t\n"
 	"#define s16     int16_t\n"
 	"#define u16     uint16_t\n"
 	"#define u8      uint8_t\n"
 	"#define s32vec2 i32vec2\n"
 	"#define s16vec2 i16vec2\n"
 	"\n"));
 
 	i32  header_vector_length = beamformer_shader_header_vector_lengths[reloadable_index];
 	i32 *header_vector        = (i32 *)beamformer_shader_header_vectors[reloadable_index];
 	for (i32 index = 0; index < header_vector_length; index++)
 		stream_append_str8(s, beamformer_shader_global_header_strings[header_vector[index]]);
 
 	if (layout.x != 0) {
 		stream_append_str8(s, str8("layout(local_size_x = "));
 		stream_append_u64(s,  layout.x);
 		stream_append_str8(s, str8(", local_size_y = "));
 		stream_append_u64(s,  layout.y);
 		stream_append_str8(s, str8(", local_size_z = "));
 		stream_append_u64(s,  layout.z);
 		stream_append_str8(s, str8(") in;\n\n"));
 	}
 
 	{
 		u32 max_length = 0;
 		for EachElement(beamformer_data_kind_str8, it)
 			max_length = Max(max_length, (u32)beamformer_data_kind_str8[it].length);
 
 		for EachElement(beamformer_data_kind_str8, it) {
 			stream_append_str8s(s, str8("#define DataKind_"), beamformer_data_kind_str8[it]);
 			stream_pad(s, ' ', max_length - beamformer_data_kind_str8[it].length + 1);
 			stream_append_u64(s, it);
 			stream_append_byte(s, '\n');
 		}
 		stream_append_byte(s, '\n');
 	}
 
 	if (gpu_heap_pointer) {
 		stream_append_str8(s, str8("#define HeapBase u64(0x"));
 		stream_append_hex_u64(s, gpu_heap_pointer);
 		stream_append_str8(s, str8("ul)\n"));
 	}
 
 	if (sd) {
 		BeamformerDataKind data_kinds[] = {sd->input_data_kind, sd->output_data_kind};
 		str8 line_prefixes[] = {str8_comp("Input"), str8_comp("Output")};
 		for EachElement(data_kinds, it) {
 			if (data_kinds[it] != BeamformerDataKind_Count) {
 				stream_append_str8s(s, str8("#define "), line_prefixes[it], str8("DataType "),
 				                    beamformer_data_kind_glsl_type[data_kinds[it]],
 				                    str8("\n#define "), line_prefixes[it], str8("DataKind DataKind_"),
 				                    beamformer_data_kind_str8[data_kinds[it]],
 				                    str8("\n#define "), line_prefixes[it], str8("DataKindByteSize "));
 				stream_append_u64(s, beamformer_data_kind_byte_size[data_kinds[it]]);
 				stream_append_byte(s, '\n');
 			}
 		}
 		stream_append_byte(s, '\n');
 
 		stream_append_str8(s, str8("#define CompileFlags (0x"));
 		stream_append_hex_u64_width(s, sd->compile_flags, 8);
 		stream_append_str8(s, str8(")\n"));
 
 		i32 struct_id = beamformer_base_shader_to_bake_struct_id[reloadable_index];
 		if (struct_id != -1) {
 			const str8             *names = meta_struct_member_names_by_id[struct_id];
 			const MetaStructInfo   *si    = meta_struct_info_by_id + struct_id;
 			const MetaStructMember *sm    = meta_struct_members_by_id[struct_id];
 			for (u32 index = 0; index < si->member_count; index++) {
 				str8 type = meta_kind_glsl_types[sm[index].type_id];
 				stream_append_str8(s, str8("layout(constant_id = "));
 				stream_append_u64(s, index);
 				stream_append_str8s(s, str8(") const "), type, str8(" "), names[index], str8(" = "), type, str8("(1);\n"));
 			}
 		}
 	}
 
 	if (!renderdoc_attached())
 		stream_append_str8(s, str8("\n\n#line 1\n"));
 }
 
 function void
 beamformer_reload_pipeline(VulkanHandle *pipeline, BeamformerShaderReloadInfo *sris, u32 count, Arena *scratch)
 {
 	assume(count <= 2);
 	str8 paths[2];
 	VulkanPipelineCreateInfo infos[2];
 
 	if (!BakeShaders) {
 		for (u32 i = 0; i < count; i++)
 			paths[i] = push_str8_from_parts(scratch, os_path_separator(), str8("shaders"), sris[i].filename_or_data);
 	}
 
 	u32 push_constants_size = 0;
 	for (u32 i = 0; i < count; i++) {
 		Stream shader_stream = arena_stream(scratch);
 		i32 reloadable_index = beamformer_shader_reloadable_index_by_shader[sris[i].shader];
 		if (i == 0) push_constants_size = beamformer_shader_push_constant_sizes[reloadable_index];
 		else        assert(push_constants_size == beamformer_shader_push_constant_sizes[reloadable_index]);
 
 		stream_append_shader_header(&shader_stream, reloadable_index, sris[i].gpu_heap_pointer,
 		                            sris[i].shader_descriptor, sris[i].layout);
 
 		str8 shader_text;
 		if (BakeShaders) {
 			stream_append_str8(&shader_stream, sris[i].filename_or_data);
 			shader_text = arena_stream_commit_zero(scratch, &shader_stream);
 		} else {
 			str8 stream_data = arena_stream_commit(scratch, &shader_stream);
 			str8 shader_data = os_read_entire_file(scratch, (c8 *)paths[i].data);
 			// NOTE(rnp): kinda sucky but need to make sure these are a contiguous string
 			shader_text = push_str8_from_parts(scratch, str8(""), stream_data, shader_data);
 		}
 
 		infos[i].kind = sris[i].shader_kind;
 		infos[i].text = shader_text;
 		infos[i].name = beamformer_shader_names[sris[i].shader];
 		infos[i].specialization_data      = sris[i].shader_descriptor ? &sris[i].shader_descriptor->bake : 0;
 		infos[i].specialization_struct_id = beamformer_base_shader_to_bake_struct_id[reloadable_index];
 
 		//str8 line = str8("---------------\n");
 		//str8 nl   = str8("\n");
 		//os_console_log(line.data, line.length);
 		//os_console_log(infos[i].name.data, infos[i].name.length);
 		//os_console_log(nl.data, nl.length);
 		//os_console_log(line.data, line.length);
 		//os_console_log(infos[i].text.data, infos[i].text.length);
 		//os_console_log(line.data, line.length);
 	}
 
 	vk_pipeline_release(*pipeline);
 	*pipeline = vk_pipeline(infos, count, push_constants_size);
 }
 
 function void
 beamformer_reload_render_pipeline(VulkanHandle *pipeline, BeamformerShaderKind shader, Arena *scratch)
 {
 	i32 index = beamformer_shader_reloadable_index_by_shader[shader];
 	BeamformerShaderReloadInfo infos[2] = {
 		{
 			.shader      = shader,
 			.shader_kind = beamformer_shader_primitive_is_vertex[index] ? VulkanShaderKind_Vertex : VulkanShaderKind_Mesh,
 			.filename_or_data = BakeShaders ? beamformer_shader_data[index][0]
 			                                : beamformer_reloadable_shader_files[index][0],
 		},
 		{
 			.shader           = shader,
 			.shader_kind      = VulkanShaderKind_Fragment,
 			.filename_or_data = BakeShaders ? beamformer_shader_data[index][1]
 			                                : beamformer_reloadable_shader_files[index][1],
 		},
 	};
 	beamformer_reload_pipeline(pipeline, infos, countof(infos), scratch);
 }
 
 function void
 beamformer_commit_parameter_block(BeamformerCtx *ctx, BeamformerComputePlan *cp, u32 block, Arena *scratch)
 {
 	BeamformerParameterBlock *pb;
 	DeferLoop(pb = beamformer_parameter_block_lock(ctx->shared_memory, block, -1),
 	          beamformer_parameter_block_unlock(ctx->shared_memory, block))
 	for EachBit(pb->region_update_flags, region)
 	{
 		pb->region_update_flags &= ~(1ul << region);
 		switch (region) {
 		case BeamformerParameterDirtyFlag_NotifyUI:{
 			atomic_store_u32(&ctx->ui_dirty_parameter_blocks, 1u << block);
 		}break;
 
 		case BeamformerParameterDirtyFlag_Parameters:{
 			cp->output_points  = das_valid_points(pb->parameters.output_points.xyz);
 			cp->average_frames = pb->parameters.output_points.E[3];
 
 			plan_compute_pipeline(cp, pb, scratch);
 
 			b32 heap_pointer_updated = cp->last_gpu_heap_pointer != cp->gpu_temp_arena.gpu_pointer;
 			cp->last_gpu_heap_pointer = cp->gpu_temp_arena.gpu_pointer;
 
 			for (u32 shader_slot = 0; shader_slot < cp->pipeline.shader_count; shader_slot++) {
 				u128 hash = u128_hash_from_data(cp->shader_descriptors + shader_slot, sizeof(BeamformerShaderDescriptor));
 				if (!u128_equal(hash, cp->shader_hashes[shader_slot]) ||
 				    (cp->shader_descriptors[shader_slot].uses_heap && heap_pointer_updated))
 				{
 					cp->dirty_programs |= 1 << shader_slot;
 				}
 				cp->shader_hashes[shader_slot] = hash;
 			}
 
 			cp->acquisition_count = pb->parameters.acquisition_count;
 			cp->acquisition_kind  = pb->parameters.acquisition_kind;
 			cp->contrast_mode     = pb->parameters.contrast_mode;
 		}break;
 		}
 	}
 }
 
 function void
 do_compute_shader(GPUCommandList cmd, BeamformerComputePlan *cp, BeamformerFrame *frame,
                   u64 input_pointer, u64 output_pointer, u32 shader_slot, u32 channel_offset)
 {
 	BeamformerComputeContext *cc = &beamformer_context->compute_context;
 
 	uv3 dispatch = cp->shader_descriptors[shader_slot].dispatch;
 
 	gpu_command_bind_pipeline(cmd, cp->vulkan_pipelines[shader_slot]);
 
 	switch (cp->pipeline.shaders[shader_slot]) {
 
 	case BeamformerShaderKind_Decode:{
 		BeamformerDecodePushConstants pc = {
 			.rf_buffer     = input_pointer,
 			.output_buffer = output_pointer,
 		};
 
 		gpu_command_push_constants(cmd, 0, sizeof(pc), &pc);
 		gpu_command_dispatch_compute(cmd, dispatch);
 
 		cc->ping_pong_input_index = !cc->ping_pong_input_index;
 	}break;
 
 	case BeamformerShaderKind_Hilbert:{
 		//cuda_hilbert(input_index, output_index);
 		//cc->ping_pong_input_index = !cc->ping_pong_input_index;
 	}break;
 
 	case BeamformerShaderKind_Filter:
 	case BeamformerShaderKind_Demodulate:
 	{
 		BeamformerFilterPushConstants pc = {
 			.input_buffer  = input_pointer,
 			.output_buffer = output_pointer,
 		};
 
 		gpu_command_push_constants(cmd, 0, sizeof(pc), &pc);
 		gpu_command_dispatch_compute(cmd, dispatch);
 
 		cc->ping_pong_input_index = !cc->ping_pong_input_index;
 	}break;
 
 	case BeamformerShaderKind_DAS:{
 		BeamformerDASPushConstants pc = {
 			.xdc_element_pitch = cp->xdc_element_pitch,
 			.rf_data           = input_pointer,
 			.output_frame      = frame->gpu_pointer,
 			.channel_offset    = channel_offset,
 			.readi_group       = cp->readi_group,
 		};
 		memory_copy(pc.voxel_transform.E, cp->das_voxel_transform.E, sizeof(pc.voxel_transform));
 		memory_copy(pc.xdc_transform.E,   cp->xdc_transform.E,       sizeof(pc.xdc_transform));
 
 		gpu_command_push_constants(cmd, 0, sizeof(pc), &pc);
 		gpu_command_dispatch_compute(cmd, dispatch);
 	}break;
 
 	case BeamformerShaderKind_CoherencyWeighting:{
 		BeamformerCoherencyWeightingPushConstants pc = {.coherent_sum = frame->gpu_pointer};
 		gpu_command_push_constants(cmd, 0, sizeof(pc), &pc);
 		gpu_command_dispatch_compute(cmd, dispatch);
 	}break;
 
 	case BeamformerShaderKind_Reshape:{
 		BeamformerDataKind input_data_kind = cp->shader_descriptors[shader_slot].input_data_kind;
 		BeamformerReshapeBakeParameters *rb = &cp->shader_descriptors[shader_slot].bake.Reshape;
 		BeamformerReshapePushConstants pc = {
 			.left_input_buffer  = input_pointer,
 			.right_input_buffer = input_pointer + rb->SizeX * rb->SizeY * rb->SizeZ
 			                                      * beamformer_data_kind_byte_size[input_data_kind],
 			.output_buffer      = output_pointer,
 		};
 
 		gpu_command_push_constants(cmd, 0, sizeof(pc), &pc);
 		gpu_command_dispatch_compute(cmd, dispatch);
 
 		cc->ping_pong_input_index = !cc->ping_pong_input_index;
 	}break;
 
 	// NOTE(rnp): invalid stages should be filtered in planning phase
 	InvalidDefaultCase;
 	}
 
 	#if 0
 	switch (shader) {
 	case BeamformerShaderKind_MinMax:{
 		for (u32 i = 1; i < frame->image.mip_map_levels; i++) {
 			glBindImageTexture(0, frame->texture, i - 1, GL_TRUE, 0, GL_READ_ONLY,  GL_RG32F);
 			glBindImageTexture(1, frame->texture, i - 0, GL_TRUE, 0, GL_WRITE_ONLY, GL_RG32F);
 			glProgramUniform1i(program, MIN_MAX_MIPS_LEVEL_UNIFORM_LOC, i);
 
 			u32 width  = (u32)frame->dim.x >> i;
 			u32 height = (u32)frame->dim.y >> i;
 			u32 depth  = (u32)frame->dim.z >> i;
 			glDispatchCompute(ORONE(width / 32), ORONE(height), ORONE(depth / 32));
 			glMemoryBarrier(GL_SHADER_IMAGE_ACCESS_BARRIER_BIT);
 		}
 	}break;
 	case BeamformerShaderKind_Sum:{
 		u32 aframe_index = ctx->averaged_frame_index % countof(ctx->averaged_frames);
 		BeamformerFrame *aframe = ctx->averaged_frames + aframe_index;
 		aframe->id              = ctx->averaged_frame_index;
 		atomic_store_u32(&aframe->ready_to_present, 0);
 		/* TODO(rnp): hack we need a better way of specifying which frames to sum;
 		 * this is fine for rolling averaging but what if we want to do something else */
 		assert(frame >= ctx->beamform_frames);
 		assert(frame < ctx->beamform_frames + countof(ctx->beamform_frames));
 		u32 base_index   = (u32)(frame - ctx->beamform_frames);
 		u32 to_average   = (u32)cp->average_frames;
 		u32 frame_count  = 0;
 		u32 *in_textures = push_array(&arena, u32, BeamformerMaxBacklogFrames);
 		ComputeFrameIterator cfi = compute_frame_iterator(ctx, 1 + base_index - to_average, to_average);
 		for (BeamformerFrame *it = frame_next(&cfi); it; it = frame_next(&cfi))
 			in_textures[frame_count++] = it->texture;
 
 		assert(to_average == frame_count);
 
 		glProgramUniform1f(program, SUM_PRESCALE_UNIFORM_LOC, 1 / (f32)frame_count);
 		/* NOTE: zero output before summing */
 		glClearTexImage(aframe->texture, 0, GL_RED, GL_FLOAT, 0);
 		glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT);
 
 		glBindImageTexture(0, out_texture, 0, GL_TRUE, 0, GL_READ_WRITE, GL_RG32F);
 		for (u32 i = 0; i < in_texture_count; i++) {
 			glBindImageTexture(1, in_textures[i], 0, GL_TRUE, 0, GL_READ_ONLY, GL_RG32F);
 			glDispatchCompute(dispatch.x, dispatch.y, dispatch.z);
 			glMemoryBarrier(GL_SHADER_IMAGE_ACCESS_BARRIER_BIT);
 		}
 
 		memory_copy(aframe->voxel_transform.E,  frame->voxel_transform.E, sizeof(frame->voxel_transform));
 		aframe->compound_count   = frame->compound_count;
 		aframe->acquisition_kind = frame->acquisition_kind;
 	}break;
 	}
 	#endif
 }
 
 function void
 complete_queue(BeamformerCtx *ctx, BeamformWorkQueue *q, Arena *arena)
 {
 	BeamformerComputeContext * cs = &ctx->compute_context;
 	BeamformerSharedMemory *   sm = ctx->shared_memory;
 
 	for (BeamformWork *work = beamform_work_queue_pop(q);
 	     work;
 	     beamform_work_queue_pop_commit(q), work = beamform_work_queue_pop(q))
 	{
 		switch (work->kind) {
 
 		case BeamformerWorkKind_ExportBuffer:{
 			/* TODO(rnp): better way of handling DispatchCompute barrier */
 			post_sync_barrier(ctx->shared_memory, BeamformerSharedMemoryLockKind_DispatchCompute);
 			beamformer_shared_memory_take_lock(ctx->shared_memory, (i32)work->lock, (u32)-1);
 			BeamformerExportContext *ec = &work->export_context;
 			switch (ec->kind) {
 			case BeamformerExportKind_BeamformedData:{
 				BeamformerFrameBacklog *bl = &ctx->compute_context.backlog;
 				u32 req_count = Clamp(ec->count, 1, bl->counter);
 				u32 frame_idx = bl->counter - req_count;
 				u8 *sm_output = beamformer_shared_memory_data_pointer(sm, ctx->shared_memory_size);
 				u64 exported_size = 0;
 				for (u32 export_count = 0; export_count < req_count; export_count++, frame_idx++) {
 					BeamformerFrame *f = bl->frames + frame_idx % countof(bl->frames);
 					u64 frame_size = beamformer_frame_byte_size(f->points, f->data_kind);
 					assert((frame_size & 63) == 0);
 					// NOTE(tkh) we don't want to assume that all req_count frames are the same size,
 					// so we either need to count the total size of all requested frames first or
 					// just fill up as much as possible.
 					if (exported_size + frame_size <= ec->size) {
 						u64 offset = f->gpu_pointer - bl->buffer->gpu_pointer;
 						gpu_host_wait_timeline(GPUTimeline_Compute, f->timeline_valid_value, -1ULL);
 						gpu_buffer_range_download(sm_output + exported_size, bl->buffer, offset, frame_size, 1);
 						exported_size += frame_size;
 					}
 				}
 			}break;
 
 			case BeamformerExportKind_Stats:{
 				ComputeTimingTable *table = ctx->compute_timing_table;
 				/* NOTE(rnp): do a little spin to let this finish updating */
 				spin_wait(table->write_index != atomic_load_u32(&table->read_index));
 				ComputeShaderStats *stats = ctx->compute_shader_stats;
 				if (sizeof(stats->table) <= ec->size)
 					memory_copy(beamformer_shared_memory_data_pointer(sm, ctx->shared_memory_size),
 					         &stats->table, sizeof(stats->table));
 			}break;
 			InvalidDefaultCase;
 			}
 			beamformer_shared_memory_release_lock(ctx->shared_memory, work->lock);
 			post_sync_barrier(ctx->shared_memory, BeamformerSharedMemoryLockKind_ExportSync);
 		}break;
 
 		case BeamformerWorkKind_CreateFilter:{
 			BeamformerCreateFilterContext *fctx = &work->create_filter_context;
 			u32 block = fctx->parameter_block;
 			u32 slot  = fctx->filter_slot;
 			BeamformerComputePlan *cp = beamformer_compute_plan_for_block(cs, block, arena);
 			cp->filter_parameters[slot] = fctx->parameters;
 		}break;
 
 		case BeamformerWorkKind_WaitThenCompute:
 		case BeamformerWorkKind_Compute:
 		{
 			push_compute_timing_info(ctx->compute_timing_table,
 			                         (ComputeTimingInfo){.kind = ComputeTimingInfoKind_ComputeFrameBegin});
 
 			BeamformerComputePlan *cp = beamformer_compute_plan_for_block(cs, work->compute_context.parameter_block, arena);
 			if unlikely(beamformer_parameter_block_dirty(sm, work->compute_context.parameter_block)) {
 				u32 block = work->compute_context.parameter_block;
 				Temp scratch = temp_begin(arena);
 				beamformer_commit_parameter_block(ctx, cp, block, arena);
 				temp_end(scratch);
 			}
 
 			post_sync_barrier(ctx->shared_memory, BeamformerSharedMemoryLockKind_DispatchCompute);
 
 			u32 dirty_programs = atomic_swap_u32(&cp->dirty_programs, 0);
 			static_assert(BeamformerMaxComputeShaderStages <= 32, "");
 			if unlikely(dirty_programs) {
 				for EachBit(dirty_programs, slot) {
 					assert(slot < BeamformerMaxComputeShaderStages);
 					Temp scratch;
 					DeferLoop(scratch = temp_begin(arena), temp_end(scratch))
 					{
 						u64 gpu_heap_pointer = cp->gpu_temp_arena.gpu_pointer;
 						i32 index = beamformer_shader_reloadable_index_by_shader[cp->pipeline.shaders[slot]];
 						BeamformerShaderReloadInfo info = {
 							.shader      = cp->pipeline.shaders[slot],
 							.shader_kind = VulkanShaderKind_Compute,
 							.shader_descriptor = cp->shader_descriptors + slot,
 							.gpu_heap_pointer  = cp->shader_descriptors[slot].uses_heap ? gpu_heap_pointer : 0,
 							.filename_or_data  = BakeShaders ? beamformer_shader_data[index][0]
 							                                 : beamformer_reloadable_shader_files[index][0],
 							.layout            = cp->shader_descriptors[slot].layout,
 						};
 						beamformer_reload_pipeline(cp->vulkan_pipelines + slot, &info, 1, arena);
 					}
 				}
 			}
 
 			atomic_store_u32(&cs->processing_compute, 1);
 
 			start_renderdoc_capture();
 
 			i32 das_index = -1;
 			i32 coherency_weighting = -1;
 			for (u32 i = 0; i < cp->pipeline.shader_count; i++) {
 				if (cp->pipeline.shaders[i] == BeamformerShaderKind_CoherencyWeighting)
 					coherency_weighting = (i32)i;
 
 				if (cp->pipeline.shaders[i] == BeamformerShaderKind_DAS)
 					das_index = (i32)i;
 			}
 
 			BeamformerFrame *frame  = beamformer_frame_next(cs, cp->output_points, cp->iq_pipeline);
 			frame->acquisition_kind = cp->acquisition_kind;
 			frame->contrast_mode    = cp->contrast_mode;
 			frame->compound_count   = cp->acquisition_count;
 			frame->parameter_block  = work->compute_context.parameter_block;
 			frame->view_plane_tag   = work->compute_context.view_plane;
 			memory_copy(frame->voxel_transform.E, cp->voxel_transform.E, sizeof(cp->voxel_transform));
 
 			GPUCommandList cmd = gpu_command_list_begin(GPUTimeline_Compute);
 			gpu_command_timestamp(cmd);
 
 			if (das_index >= 0) {
 				GPUBuffer *backlog = cs->backlog.buffer;
 				u64 frame_size = beamformer_frame_byte_size(frame->points, frame->data_kind);
 				u64 offset     = frame->gpu_pointer - backlog->gpu_pointer;
 				gpu_command_clear_buffer(cmd, backlog, offset, frame_size, 0);
 			}
 
 			if (coherency_weighting >= 0) {
 				BeamformerCoherencyWeightingBakeParameters *cw = &cp->shader_descriptors[coherency_weighting].bake.CoherencyWeighting;
 				GPUBuffer *gpu_arena = &cp->gpu_temp_arena;
 				u64 coherent_size = beamformer_incoherent_frame_byte_size(frame->points, frame->data_kind);
 				gpu_command_clear_buffer(cmd, gpu_arena, cw->IncoherentSum, coherent_size, 0);
 			}
 
 			BeamformerRFBuffer *rf = &cs->rf_buffer;
 			u64 compute_index = rf->compute_index;
 			u64 slot = compute_index % countof(rf->upload_complete_values);
 
 			// NOTE(rnp): the library/ui will ensure that any time a new dataset is uploaded
 			// the next time beamforming is issued it will be tagged as WaitThenCompute.
 			if (work->kind != BeamformerWorkKind_WaitThenCompute)
 				slot = (compute_index - 1) % countof(rf->upload_complete_values);
 
 			// NOTE(rnp): nvidia needs a memory barrier between pipeline stages
 			// for correct output. It doesn't seem to effect performance on nvidia cards.
 			// It does have a very minor performance decrease on AMD, but technically its
 			// against vulkan spec to skip them so for simplicity we will default to for now
 			b32 pipeline_memory_barrier = 1; //gpu_info()->vendor == GPUVendor_NVIDIA;
 
 			for (u32 channel_offset = 0;
 			     channel_offset < cp->channel_count;
 			     channel_offset += BeamformerChunkChannelCount)
 			{
 				u64 rf_pointer = rf->buffer.gpu_pointer + slot * rf->active_rf_size;
 				rf_pointer += cp->raw_channel_byte_stride * channel_offset;
 
 				// NOTE(rnp): handle the case where the channel count is not a multiple of
 				// BeamformerChunkChannelCount. We do not want to have to modify every shader
 				// that might run to care about this so here we run a copy/clear to pad the
 				// data for the last iteration.
 				b32 special_handling = (cp->channel_count - channel_offset) < BeamformerChunkChannelCount;
 				if unlikely(special_handling) {
 					u32 channel_pad_count = BeamformerChunkChannelCount - (cp->channel_count - channel_offset);
 					u32 input_index = cs->ping_pong_input_index;
 					u64 pp_size     = cs->ping_pong_buffer.size / PING_PONG_BUFFER_SLOTS;
 					u64 copy_size   = cp->raw_channel_byte_stride * (cp->channel_count - channel_offset);
 					u64 copy_offset = rf_pointer - rf->buffer.gpu_pointer;
 					u64 clear_size  = cp->raw_channel_byte_stride * channel_pad_count;
 
 					gpu_command_copy_buffer(cmd, &cs->ping_pong_buffer, input_index * pp_size, &rf->buffer, copy_offset, copy_size);
 					gpu_command_clear_buffer(cmd, &cs->ping_pong_buffer, input_index * pp_size + copy_size, clear_size, 0);
 					// TODO(rnp): technically this should be a transfer to compute barrier
 					// but a full memory barrier will also do the trick
 					gpu_command_pipeline_barrier(cmd, 1);
 				}
 
 				for (u32 i = 0; i < cp->first_image_shader_index; i++) {
 					u32 output_index      = !cs->ping_pong_input_index;
 					u32 input_index       =  cs->ping_pong_input_index;
 
 					u64 pp_size           = cs->ping_pong_buffer.size / PING_PONG_BUFFER_SLOTS;
 					u64 pp_input_pointer  = cs->ping_pong_buffer.gpu_pointer + input_index  * pp_size;
 					u64 pp_output_pointer = cs->ping_pong_buffer.gpu_pointer + output_index * pp_size;
 
 					u64 output_pointer = pp_output_pointer;
 					u64 input_pointer  = (i == 0 && !special_handling) ? rf_pointer : pp_input_pointer;
 
 					if (i != 0) gpu_command_pipeline_barrier(cmd, pipeline_memory_barrier);
 					do_compute_shader(cmd, cp, frame, input_pointer, output_pointer, i, channel_offset);
 					gpu_command_timestamp(cmd);
 				}
 			}
 
 			for (u32 i = cp->first_image_shader_index; i < cp->pipeline.shader_count; i++) {
 				gpu_command_pipeline_barrier(cmd, pipeline_memory_barrier);
 				do_compute_shader(cmd, cp, frame, 0, 0, i, 0);
 				gpu_command_timestamp(cmd);
 			}
 
 			u32 wait_info_count = 0;
 			GPUSemaphoreSignalInfo wait_info = {.semaphore = rf->upload_semaphore};
 			if (work->kind == BeamformerWorkKind_WaitThenCompute) {
 				spin_wait(atomic_load_u64(&rf->insertion_index) <= compute_index);
 				wait_info.value = atomic_load_u64(rf->upload_complete_values + slot);
 				wait_info_count = 1;
 			}
 
 			u64 end_timeline_value = gpu_command_list_end(cmd, &wait_info, wait_info_count, 0, 0);
 			atomic_store_u64(rf->compute_complete_values + slot, end_timeline_value);
 
 			if (work->kind == BeamformerWorkKind_WaitThenCompute)
 				atomic_add_u64(&rf->compute_index, 1);
 
 			atomic_store_u64(&frame->timeline_valid_value, end_timeline_value);
 
 			Temp scratch;
 			DeferLoop(scratch = temp_begin(arena), temp_end(scratch))
 			{
 				/* NOTE(rnp): this blocks until work completes */
 				u64  count       = 0;
 				u64 *timestamps  = gpu_read_timestamps(GPUTimeline_Compute, &count, arena);
 
 				i32 steps        = ((i32)cp->channel_count / BeamformerChunkChannelCount) - 1;
 				i32 step         = 0;
 				u32 shader_index = 0;
 				u64 last_time    = count > 0 ? timestamps[0] : 0;
 
 				for (u64 i = 1; i < count; i++) {
 					push_compute_timing_info(ctx->compute_timing_table, (ComputeTimingInfo){
 						.kind        = ComputeTimingInfoKind_Shader,
 						.shader      = cp->pipeline.shaders[shader_index],
 						.shader_slot = shader_index,
 						.timer_count = timestamps[i] - last_time,
 					});
 					last_time = timestamps[i];
 
 					shader_index++;
 					if (shader_index == cp->first_image_shader_index && step < steps) {
 						shader_index = 0;
 						step++;
 					}
 				}
 			}
 
 			cs->processing_progress = 1;
 
 			//if (has_sum) {
 			if (0) {
 				#if 0
 				u32 aframe_index = ((ctx->averaged_frame_index++) % countof(ctx->averaged_frames));
 				ctx->averaged_frames[aframe_index].view_plane_tag  = frame->view_plane_tag;
 				ctx->averaged_frames[aframe_index].ready_to_present = 1;
 				atomic_store_u64((u64 *)&ctx->latest_frame, (u64)(ctx->averaged_frames + aframe_index));
 				#endif
 			} else {
 				atomic_store_u64((u64 *)&ctx->latest_frame, (u64)frame);
 			}
 
 			atomic_store_u32(&cs->processing_compute, 0);
 
 			push_compute_timing_info(ctx->compute_timing_table,
 			                         (ComputeTimingInfo){.kind = ComputeTimingInfoKind_ComputeFrameEnd});
 
 			end_renderdoc_capture();
 		}break;
 		InvalidDefaultCase;
 		}
 	}
 }
 
 function void
 coalesce_timing_table(ComputeTimingTable *t, ComputeShaderStats *stats)
 {
 	/* TODO(rnp): we do not currently do anything to handle the potential for a half written
 	 * info item. this could result in garbage entries but they shouldn't really matter */
 
 	u32 target = atomic_load_u32(&t->write_index);
 	u32 stats_index = stats->latest_frame_index;
 
 	b32 has_rf = 0;
 	f32 gpu_clocks_to_nano = 1.0e-9f * gpu_info()->timestamp_period_ns;
 
 	// NOTE(rnp): not equal (the index may wrap)
 	while (t->read_index != target) {
 		ComputeTimingInfo info = t->buffer[t->read_index % countof(t->buffer)];
 		switch (info.kind) {
 
 		case ComputeTimingInfoKind_ComputeFrameBegin:{
 			assert(t->compute_frame_active == 0);
 			t->compute_frame_active = 1;
 			/* NOTE(rnp): allow multiple instances of same shader to accumulate */
 			t->in_flight_shader_count = 0;
 			memory_clear(t->in_flight_shader_ids, 0, sizeof(t->in_flight_shader_ids));
 			memory_clear(stats->table.times[stats_index], 0, sizeof(stats->table.times[stats_index]));
 		}break;
 
 		case ComputeTimingInfoKind_ComputeFrameEnd:{
 			assert(t->compute_frame_active == 1);
 			t->compute_frame_active = 0;
 			stats_index = stats->latest_frame_index = (stats_index + 1) % countof(stats->table.times);
 			stats->table.shader_count = t->in_flight_shader_count;
 			memory_copy(stats->table.shader_ids, t->in_flight_shader_ids, sizeof(t->in_flight_shader_ids));
 		}break;
 
 		case ComputeTimingInfoKind_Shader:{
 			t->in_flight_shader_count = Max(t->in_flight_shader_count, info.shader_slot + 1u);
 			t->in_flight_shader_ids[info.shader_slot] = info.shader;
 			stats->table.times[stats_index][info.shader_slot] += info.timer_count * gpu_clocks_to_nano;
 		}break;
 
 		case ComputeTimingInfoKind_RF_Data:{
 			stats->latest_rf_index = (stats->latest_rf_index + 1) % countof(stats->table.rf_time_deltas);
 			f32 delta = info.timer_count / (f32)os_system_info()->timer_frequency;
 			stats->table.rf_time_deltas[stats->latest_rf_index] = delta;
 			has_rf = 1;
 		}break;
 		}
 		/* NOTE(rnp): do this at the end so that stats table is always in a consistent state */
 		t->read_index++;
 	}
 
 	for (u32 i = 0; i < stats->table.shader_count; i++) {
 		f32 sum = 0;
 		for EachElement(stats->table.times, it)
 			sum += stats->table.times[it][i];
 		stats->average_times[i] = sum / countof(stats->table.times);
 	}
 
 	if (has_rf) {
 		f32 sum = 0;
 		for EachElement(stats->table.rf_time_deltas, i)
 			sum += stats->table.rf_time_deltas[i];
 		stats->rf_time_delta_average = sum / countof(stats->table.rf_time_deltas);
 	}
 }
 
 DEBUG_EXPORT BEAMFORMER_COMPLETE_COMPUTE_FN(beamformer_complete_compute)
 {
 	BeamformerSharedMemory *sm = ctx->shared_memory;
 	complete_queue(ctx, &sm->external_work_queue, arena);
 	complete_queue(ctx, ctx->beamform_work_queue, arena);
 }
 
 DEBUG_EXPORT BEAMFORMER_RF_UPLOAD_FN(beamformer_rf_upload)
 {
 	BeamformerSharedMemory *sm                  = ctx->shared_memory;
 	BeamformerSharedMemoryLockKind scratch_lock = BeamformerSharedMemoryLockKind_ScratchSpace;
 	BeamformerSharedMemoryLockKind upload_lock  = BeamformerSharedMemoryLockKind_UploadRF;
 
 	u64 rf_block_rf_size;
 	if (atomic_load_u32(sm->locks + upload_lock) &&
 	    (rf_block_rf_size = atomic_swap_u64(&sm->rf_block_rf_size, 0)))
 	{
 		beamformer_shared_memory_take_lock(ctx->shared_memory, (i32)scratch_lock, (u32)-1);
 
 		BeamformerRFBuffer *rf = ctx->rf_buffer;
 
 		rf->active_rf_size = gpu_round_up_to_sync_size(rf_block_rf_size & 0x00FFFFFFFFFFFFFFull, 64);
 		if unlikely((u64)rf->buffer.size < countof(rf->upload_complete_values) * rf->active_rf_size) {
 			gpu_buffer_allocate(&rf->buffer, (GPUBufferAllocateInfo){
 				.size  = countof(rf->upload_complete_values) * rf->active_rf_size,
 				.flags = GPUUsageFlag_HostWrite,
 				.label = str8("RawRFBuffer"),
 				.timelines_used = (GPUTimeline []){GPUTimeline_Compute},
 				.timeline_count = 1,
 			});
 		}
 
 		u64 slot = rf->insertion_index % countof(rf->upload_complete_values);
 
 		/* NOTE(rnp): don't overwrite slot if the compute thread hasn't processed it */
 		spin_wait(rf->insertion_index - atomic_load_u64(&rf->compute_index) >= countof(rf->upload_complete_values));
 		gpu_host_wait_timeline(GPUTimeline_Compute, rf->compute_complete_values[slot], -1ULL);
 
 		assert((ctx->shared_memory_size % os_system_info()->page_size) == 0 &&
 		       (os_system_info()->page_size % gpu_round_up_to_sync_size(1, 64)) == 0);
 
 		u64 wait_value = gpu_semaphore_value(rf->upload_semaphore) + 1;
 		atomic_store_u64(rf->upload_complete_values + slot, wait_value);
 		atomic_add_u64(&rf->insertion_index, 1);
 
 		void *host_pointer = gpu_buffer_host_pointer(&rf->buffer);
 		memory_copy_non_temporal((u8 *)host_pointer + slot * rf->active_rf_size,
 		                         beamformer_shared_memory_data_pointer(sm, ctx->shared_memory_size),
 		                         rf->active_rf_size);
 		store_fence();
 
 		GPUSemaphoreSignalInfo signal_info = {.semaphore = rf->upload_semaphore, .value = wait_value};
 		gpu_buffer_make_visible(&rf->buffer, slot * rf->active_rf_size, rf->active_rf_size, &signal_info, 1);
 
 		beamformer_shared_memory_release_lock(ctx->shared_memory, (i32)scratch_lock);
 		post_sync_barrier(ctx->shared_memory, upload_lock);
 		os_wake_all_waiters(ctx->compute_worker_sync);
 
 		u64 current_time = os_timer_count();
 		push_compute_timing_info(ctx->compute_timing_table, (ComputeTimingInfo){
 			.kind        = ComputeTimingInfoKind_RF_Data,
 			.timer_count = current_time - rf->timestamp,
 		});
 		rf->timestamp = current_time;
 	}
 }
 
 function void
 beamformer_queue_compute(BeamformerCtx *ctx, BeamformerFrame *frame, u32 parameter_block)
 {
 	BeamformerSharedMemory *sm = ctx->shared_memory;
 	BeamformerSharedMemoryLockKind dispatch_lock = BeamformerSharedMemoryLockKind_DispatchCompute;
 	if (!sm->live_imaging_parameters.active && beamformer_shared_memory_take_lock(sm, (i32)dispatch_lock, 0))
 	{
 		BeamformWork *work = beamform_work_queue_push(ctx->beamform_work_queue);
 		if (work) {
 			work->kind = BeamformerWorkKind_Compute;
 			work->compute_context.view_plane      = frame ? frame->view_plane_tag : 0;
 			work->compute_context.parameter_block = parameter_block;
 			beamform_work_queue_push_commit(ctx->beamform_work_queue);
 		}
 	}
 	os_wake_all_waiters(&ctx->compute_worker.sync_variable);
 }
 
 #include "ui.c"
 
 function void
 beamformer_process_input_events(BeamformerCtx *ctx, BeamformerInput *input,
                                 BeamformerInputEvent *events, u32 event_count)
 {
 	for (u32 index = 0; index < event_count; index++) {
 		BeamformerInputEvent *event = events + index;
 		switch (event->kind) {
 
 		// NOTE(rnp): ui will handle these
 		case BeamformerInputEventKind_ButtonPress:
 		case BeamformerInputEventKind_ButtonRelease:
 		case BeamformerInputEventKind_MouseScroll:
 		case BeamformerInputEventKind_WindowResize:
 		{}break;
 
 		case BeamformerInputEventKind_ExecutableReload:{
 			ui_init(ctx, ctx->ui_arena);
 		}break;
 
 		case BeamformerInputEventKind_FileEvent:{
 			BeamformerFileReloadContext *frc = event->file_watch_user_context;
 			switch (frc->kind) {
 			case BeamformerFileReloadKind_ComputeShader:{
 				for EachElement(ctx->compute_context.compute_plans, block) {
 					BeamformerComputePlan *cp = ctx->compute_context.compute_plans[block];
 					for (u32 slot = 0; cp && slot < cp->pipeline.shader_count; slot++) {
 						i32 shader_index = beamformer_shader_reloadable_index_by_shader[cp->pipeline.shaders[slot]];
 						if (beamformer_reloadable_shader_kinds[shader_index] == frc->shader_reload.shader)
 							atomic_or_u32(&cp->dirty_programs, 1 << slot);
 					}
 				}
 
 				// TODO(rnp): track latest parameter block
 				if (ctx->latest_frame)
 					beamformer_queue_compute(ctx, ctx->latest_frame, 0);
 			}break;
 
 			case BeamformerFileReloadKind_RenderShader:{
 				beamformer_reload_render_pipeline(frc->shader_reload.pipeline, frc->shader_reload.shader, ctx->arena);
 				ctx->render_shader_updated = 1;
 			}break;
 
 			InvalidDefaultCase;
 			}
 		}break;
 
 		InvalidDefaultCase;
 		}
 	}
 }
 
 function void
 beamformer_panel_group_insert_at(BeamformerUIPanel *group, BeamformerUIPanel *tab, u64 new_child_index)
 {
 	if (tab->parent) beamformer_ui_panel_unlink(tab);
 	new_child_index = Min(new_child_index, group->child_count);
 
 	tab->parent = group;
 	group->child_count++;
 	if (group->kind == BeamformerPanelKind_TabGroup) group->u.tab_focus = tab;
 
 	BeamformerUIPanel *previous_sibling = new_child_index == 0 ? 0 : group->first_child;
 	for (u64 child_index = 1; child_index < new_child_index; child_index++)
 		previous_sibling = previous_sibling->next_sibling;
 
 	if (previous_sibling) {
 		tab->previous_sibling = previous_sibling;
 		tab->next_sibling     = previous_sibling->next_sibling;
 		if (tab->next_sibling) tab->next_sibling->previous_sibling = tab;
 		previous_sibling->next_sibling = tab;
 		if (previous_sibling == group->last_child) group->last_child = tab;
 	} else {
 		DLLInsertFirst(0, group->first_child, group->last_child, tab, next_sibling, previous_sibling);
 	}
 }
 
 BEAMFORMER_EXPORT void
 beamformer_frame_step(void *memory, BeamformerInput *input)
 {
 	BeamformerCtx *ctx = beamformer_context = memory;
 	beamformer_input = input;
 
 	u64 current_time = os_timer_count();
 	dt_for_frame = (f64)(current_time - ctx->frame_timestamp) / os_system_info()->timer_frequency;
 	ctx->frame_timestamp = current_time;
 	ctx->frame_index++;
 
 	coalesce_timing_table(ctx->compute_timing_table, ctx->compute_shader_stats);
 
 	// NOTE(rnp): reset frame state
 	{
 		ctx->registers = &ctx->base_registers;
 		swap(ctx->command_queues[0], ctx->command_queues[1]);
 		zero_struct(ctx->command_queues + 0);
 		//zero_struct(ctx->registers);
 		arena_clear(beamformer_frame_arena());
 	}
 
 	beamformer_process_input_events(ctx, input, input->event_queue, input->event_count);
 
 	BeamformerSharedMemory *sm = ctx->shared_memory;
 	u32 live_imaging_active = atomic_load_u32(&sm->live_imaging_parameters.active);
 	if (live_imaging_active != ctx->live_imaging_active) {
 		if (ctx->live_imaging_active) {
 			if (ctx->auto_live_control_panel) {
 				BeamformerUIPanel *parent = ctx->auto_live_control_panel->parent;
 				beamformer_command(beamformer_command_infos[BeamformerCommandKind_CloseTab].string, .tree_node = (u64)ctx->auto_live_control_panel);
 				if (parent->child_count == 1)
 					beamformer_command(beamformer_command_infos[BeamformerCommandKind_CloseTab].string, .tree_node = (u64)parent);
 			}
 		} else {
 			if (beamformer_registers()->live_controls) {
 				beamformer_command(beamformer_command_infos[BeamformerCommandKind_FocusTab].string,
 				                   .tree_node = beamformer_registers()->live_controls);
 			} else {
 				ctx->auto_live_control_panel = beamformer_ui_push_panel(0, BeamformerPanelKind_LiveImagingControls);
 				beamformer_command(beamformer_command_infos[BeamformerCommandKind_SplitTree].string,
 				                   .tree_node        = (u64)ctx->auto_live_control_panel,
 				                   .split_axis       = Axis2_X,
 				                   .split_left_tree  = (u64)ui_context->tree,
 				                   .split_right_tree = 0,
 				                   .drop_target_tree = (u64)ui_context->tree);
 			}
 			ctx->live_imaging_active_frame = ctx->frame_index;
 		}
 		ctx->live_imaging_active = live_imaging_active;
 	}
 
 	if (atomic_load_u32(sm->locks + BeamformerSharedMemoryLockKind_UploadRF))
 		os_wake_all_waiters(&ctx->upload_worker.sync_variable);
 	if (atomic_load_u32(sm->locks + BeamformerSharedMemoryLockKind_DispatchCompute))
 		os_wake_all_waiters(&ctx->compute_worker.sync_variable);
 
 	beamformer_registers()->frame = (u64)(ctx->latest_frame - ctx->compute_context.backlog.frames);
 
 	beamformer_ui_frame();
 
 	// NOTE(rnp): execute commands
 	for (BeamformerCommandNode *node = ctx->command_queues[0].first;
 	     node;
 	     node = node == node->next ? 0 : node->next)
 	{
 		BeamformerRegistersScope()
 		{
 			memory_copy(beamformer_registers(), node->command.registers, sizeof(*node->command.registers));
 			BeamformerCommandKind kind = beamformer_command_kind_from_string(node->command.name);
 			switch (kind) {
 			InvalidDefaultCase;
 			case BeamformerCommandKind_CloseTab:{
 				BeamformerUIPanel *tab = (BeamformerUIPanel *)beamformer_registers()->tree_node;
 				ui_kill_panel(tab);
 			}break;
 
 			case BeamformerCommandKind_FocusTab:{
 				BeamformerUIPanel *tab = (BeamformerUIPanel *)beamformer_registers()->tree_node;
 				assert(tab->parent->kind == BeamformerPanelKind_TabGroup);
 				tab->parent->u.tab_focus = tab;
 			}break;
 
 			case BeamformerCommandKind_MoveTab:{
 				BeamformerUIPanel *move   = (BeamformerUIPanel *)beamformer_registers()->tree_node;
 				BeamformerUIPanel *group  = (BeamformerUIPanel *)beamformer_registers()->drop_target_tree;
 				BeamformerUIPanel *parent = move->parent;
 				u64 new_child_index = beamformer_registers()->drop_child_index;
 				beamformer_panel_group_insert_at(group, move, new_child_index);
 
 				if (move->kind == BeamformerPanelKind_LiveImagingControls) {
 					beamformer_context->base_registers.v.live_controls = (u64)move;
 					if (move == ctx->auto_live_control_panel)
 						ctx->auto_live_control_panel = 0;
 				}
 
 				if (parent->child_count == 0)
 					beamformer_command(beamformer_command_infos[BeamformerCommandKind_CloseTab].string, .tree_node = (u64)parent);
 			}break;
 
 			case BeamformerCommandKind_OpenTab:{
 				BeamformerUIPanel *panel = (BeamformerUIPanel *)beamformer_registers()->tree_node;
 				assert(panel->kind == BeamformerPanelKind_TabGroup);
 
 				BeamformerPanelKind new_panel_kind = beamformer_panel_kind_from_string(beamformer_registers()->string);
 				beamformer_ui_push_panel(panel, new_panel_kind);
 			}break;
 
 			case BeamformerCommandKind_SplitTree:{
 				BeamformerUIPanel *drag  = (BeamformerUIPanel *)beamformer_registers()->tree_node;
 				BeamformerUIPanel *left  = (BeamformerUIPanel *)beamformer_registers()->split_left_tree;
 				BeamformerUIPanel *right = (BeamformerUIPanel *)beamformer_registers()->split_right_tree;
 				Axis2 axis = beamformer_registers()->split_axis;
 
 				BeamformerUIPanel *new_split     = beamformer_ui_push_panel(0, BeamformerPanelKind_Split);
 				BeamformerUIPanel *new_tab_group = beamformer_ui_push_panel(0, BeamformerPanelKind_TabGroup);
 				beamformer_panel_group_insert_at(new_tab_group, drag, 0);
 
 				BeamformerUIPanel *target = 0;
 				u32 target_child_index = 0;
 				f32 new_split_pct = 0.5f;
 
 				if (left == 0 || right == 0) {
 					// NOTE(rnp): split on edge of window
 					target             = left ? left : right;
 					target_child_index = left ? 0 : 1;
 
 					if (target->kind == BeamformerPanelKind_TabGroup) {
 						new_split->kind        = BeamformerPanelKind_TabGroup;
 						new_split->u.tab_focus = target->u.tab_focus;
 					}
 
 					for (BeamformerUIPanel *child = target->last_child, *next; child; child = next) {
 						next = child->previous_sibling;
 						beamformer_panel_group_insert_at(new_split, child, 0);
 					}
 
 					beamformer_panel_group_insert_at(target, new_tab_group, 0);
 				} else if (((drag == left)  && right->kind == BeamformerPanelKind_Split) ||
 				           ((drag == right) && left->kind  == BeamformerPanelKind_Split))
 				{
 					// NOTE(rnp): split on internal split
 					target             = left == drag ? right : left;
 					target_child_index = 1;
 					new_split_pct      = 1.f / 3.f;
 					beamformer_panel_group_insert_at(new_split, new_tab_group, 0);
 					beamformer_panel_group_insert_at(new_split, target->last_child, 1);
 				} else {
 					// NOTE(rnp): TabGroup Split
 					target             = left == drag ? right : left;
 					target_child_index = left == drag ? 1 : 0;
 					assert(target->kind == BeamformerPanelKind_TabGroup);
 
 					BeamformerUIPanel *focus = target->u.tab_focus;
 					new_split->kind = BeamformerPanelKind_TabGroup;
 					for (BeamformerUIPanel *child = target->last_child, *next; child; child = next) {
 						next = child->previous_sibling;
 						beamformer_panel_group_insert_at(new_split, child, 0);
 					}
 					new_split->u.tab_focus = focus;
 
 					beamformer_panel_group_insert_at(target, new_tab_group, 0);
 				}
 
 				beamformer_panel_group_insert_at(target, new_split, target_child_index);
 				if (target->kind == BeamformerPanelKind_Split) {
 					new_split->u.split.axis     = target->u.split.axis;
 					new_split->u.split.fraction = target->u.split.fraction;
 				}
 				target->kind             = BeamformerPanelKind_Split;
 				target->u.split.axis     = axis;
 				target->u.split.fraction = new_split_pct;
 			}break;
 
 			}
 		}
 	}
 
 	ctx->render_shader_updated = 0;
 }