ogl_beamforming

Ultrasound Beamforming Implemented with OpenGL
git clone anongit@rnpnr.xyz:ogl_beamforming.git
Log | Files | Refs | Feed | Submodules | README | LICENSE

beamformer_core.c (59713B)


      1 /* See LICENSE for license details. */
      2 /* TODO(rnp):
      3  * [ ]: refactor: DecodeMode_None should use a different mapping and optional conversion shader
      4  *      for rf only mode with no filter and demod/filter should gain the OutputFloats flag for iq
      5  *      case and rf mode with filter; this can also be used instead of first pass uniform
      6  * [ ]: refactor: replace UploadRF with just the scratch_rf_size variable,
      7  *      use below to spin wait in library
      8  * [ ]: utilize umonitor/umwait (intel), monitorx/mwaitx (amd), and wfe/sev (aarch64)
      9  *      for power efficient low latency waiting
     10  * [ ]: refactor: split decode into reshape and decode
     11  *      - the check for first pass reshaping is the last non constant check
     12  *        in the shader
     13  *      - this will also remove the need for the channel mapping in the decode shader
     14  * [X]: refactor: ui: reload only shader which is affected by the interaction
     15  * [ ]: BeamformWorkQueue -> BeamformerWorkQueue
     16  * [ ]: need to keep track of gpu memory in some way
     17  *      - want to be able to store more than 16 2D frames but limit 3D frames
     18  *      - maybe keep track of how much gpu memory is committed for beamformed images
     19  *        and use that to determine when to loop back over existing textures
     20  *      - to do this maybe use a circular linked list instead of a flat array
     21  *      - then have a way of querying how many frames are available for a specific point count
     22  * [ ]: bug: reinit cuda on hot-reload
     23  */
     24 
     25 #include "compiler.h"
     26 
     27 #if defined(BEAMFORMER_DEBUG) && !defined(BEAMFORMER_EXPORT) && OS_WINDOWS
     28   #define BEAMFORMER_EXPORT __declspec(dllexport)
     29 #endif
     30 
     31 #include "beamformer_internal.h"
     32 
     33 global f32 dt_for_frame;
     34 
     35 #define DECODE_FIRST_PASS_UNIFORM_LOC 1
     36 
     37 #define DAS_CYCLE_T_UNIFORM_LOC       2
     38 #define DAS_FAST_CHANNEL_UNIFORM_LOC  3
     39 
     40 #define MIN_MAX_MIPS_LEVEL_UNIFORM_LOC 1
     41 #define SUM_PRESCALE_UNIFORM_LOC       1
     42 
     43 #if !BEAMFORMER_RENDERDOC_HOOKS
     44 #define start_renderdoc_capture(...)
     45 #define end_renderdoc_capture(...)
     46 #define renderdoc_attached(...) (0)
     47 #else
     48 global renderdoc_start_frame_capture_fn *start_frame_capture;
     49 global renderdoc_end_frame_capture_fn   *end_frame_capture;
     50 #define start_renderdoc_capture(gl) if (start_frame_capture) start_frame_capture(gl, 0)
     51 #define end_renderdoc_capture(gl)   if (end_frame_capture)   end_frame_capture(gl, 0)
     52 #define renderdoc_attached(...)   (start_frame_capture != 0)
     53 #endif
     54 
     55 typedef struct {
     56 	BeamformerFrame *frames;
     57 	u32 capacity;
     58 	u32 offset;
     59 	u32 cursor;
     60 	u32 needed_frames;
     61 } ComputeFrameIterator;
     62 
     63 function void
     64 beamformer_compute_plan_release(BeamformerComputeContext *cc, u32 block)
     65 {
     66 	assert(block < countof(cc->compute_plans));
     67 	BeamformerComputePlan *cp = cc->compute_plans[block];
     68 	if (cp) {
     69 		glDeleteBuffers(countof(cp->ubos), cp->ubos);
     70 		glDeleteTextures(countof(cp->textures), cp->textures);
     71 		for (u32 i = 0; i < countof(cp->filters); i++)
     72 			glDeleteBuffers(1, &cp->filters[i].ssbo);
     73 		cc->compute_plans[block] = 0;
     74 		SLLPushFreelist(cp, cc->compute_plan_freelist);
     75 	}
     76 }
     77 
     78 function BeamformerComputePlan *
     79 beamformer_compute_plan_for_block(BeamformerComputeContext *cc, u32 block, Arena *arena)
     80 {
     81 	assert(block < countof(cc->compute_plans));
     82 	BeamformerComputePlan *result = cc->compute_plans[block];
     83 	if (!result) {
     84 		result = SLLPopFreelist(cc->compute_plan_freelist);
     85 		if (!result) result = push_struct_no_zero(arena, BeamformerComputePlan);
     86 		zero_struct(result);
     87 		cc->compute_plans[block] = result;
     88 
     89 		result->ui_voxel_transform = m4_identity();
     90 
     91 		glCreateBuffers(countof(result->ubos), result->ubos);
     92 
     93 		Stream label = arena_stream(*arena);
     94 		#define X(k, t, ...) \
     95 			glNamedBufferStorage(result->ubos[BeamformerComputeUBOKind_##k], sizeof(t), \
     96 			                     0, GL_DYNAMIC_STORAGE_BIT); \
     97 			stream_append_s8(&label, s8(#t "[")); \
     98 			stream_append_u64(&label, block);     \
     99 			stream_append_byte(&label, ']');      \
    100 			glObjectLabel(GL_BUFFER, result->ubos[BeamformerComputeUBOKind_##k], \
    101 			              label.widx, (c8 *)label.data); \
    102 			label.widx = 0;
    103 		BEAMFORMER_COMPUTE_UBO_LIST
    104 		#undef X
    105 
    106 		#define X(_k, t, ...) t,
    107 		GLenum gl_kind[] = {BEAMFORMER_COMPUTE_TEXTURE_LIST_FULL};
    108 		#undef X
    109 		read_only local_persist s8 tex_prefix[] = {
    110 			#define X(k, ...) s8_comp(#k "["),
    111 			BEAMFORMER_COMPUTE_TEXTURE_LIST_FULL
    112 			#undef X
    113 		};
    114 		glCreateTextures(GL_TEXTURE_1D, BeamformerComputeTextureKind_Count - 1, result->textures);
    115 		for (u32 i = 0; i < BeamformerComputeTextureKind_Count - 1; i++) {
    116 			/* TODO(rnp): this could be predicated on channel count for this compute plan */
    117 			glTextureStorage1D(result->textures[i], 1, gl_kind[i], BeamformerMaxChannelCount);
    118 			stream_append_s8(&label, tex_prefix[i]);
    119 			stream_append_u64(&label, block);
    120 			stream_append_byte(&label, ']');
    121 			glObjectLabel(GL_TEXTURE, result->textures[i], label.widx, (c8 *)label.data);
    122 			label.widx = 0;
    123 		}
    124 	}
    125 	return result;
    126 }
    127 
    128 function void
    129 beamformer_filter_update(BeamformerFilter *f, BeamformerFilterParameters fp, u32 block, u32 slot, Arena arena)
    130 {
    131 	Stream sb = arena_stream(arena);
    132 	stream_append_s8s(&sb,
    133 	                  beamformer_filter_kind_strings[fp.kind % countof(beamformer_filter_kind_strings)],
    134 	                  s8("Filter["));
    135 	stream_append_u64(&sb, block);
    136 	stream_append_s8(&sb, s8("]["));
    137 	stream_append_u64(&sb, slot);
    138 	stream_append_byte(&sb, ']');
    139 	s8 label = arena_stream_commit(&arena, &sb);
    140 
    141 	void *filter = 0;
    142 	switch (fp.kind) {
    143 	case BeamformerFilterKind_Kaiser:{
    144 		/* TODO(rnp): this should also support complex */
    145 		/* TODO(rnp): implement this as an IFIR filter instead to reduce computation */
    146 		filter = kaiser_low_pass_filter(&arena, fp.kaiser.cutoff_frequency, fp.sampling_frequency,
    147 		                                fp.kaiser.beta, (i32)fp.kaiser.length);
    148 		f->length     = (i32)fp.kaiser.length;
    149 		f->time_delay = (f32)f->length / 2.0f / fp.sampling_frequency;
    150 	}break;
    151 	case BeamformerFilterKind_MatchedChirp:{
    152 		typeof(fp.matched_chirp) *mc = &fp.matched_chirp;
    153 		f32 fs    = fp.sampling_frequency;
    154 		f->length = (i32)(mc->duration * fs);
    155 		if (fp.complex) {
    156 			filter = baseband_chirp(&arena, mc->min_frequency, mc->max_frequency, fs, f->length, 1, 0.5f);
    157 			f->time_delay = complex_filter_first_moment(filter, f->length, fs);
    158 		} else {
    159 			filter = rf_chirp(&arena, mc->min_frequency, mc->max_frequency, fs, f->length, 1);
    160 			f->time_delay = real_filter_first_moment(filter, f->length, fs);
    161 		}
    162 	}break;
    163 	InvalidDefaultCase;
    164 	}
    165 
    166 	f->parameters = fp;
    167 
    168 	glDeleteBuffers(1, &f->ssbo);
    169 	glCreateBuffers(1, &f->ssbo);
    170 	glNamedBufferStorage(f->ssbo, f->length * (i32)sizeof(f32) * (fp.complex? 2 : 1), filter, 0);
    171 	glObjectLabel(GL_BUFFER, f->ssbo, (i32)label.len, (c8 *)label.data);
    172 }
    173 
    174 function ComputeFrameIterator
    175 compute_frame_iterator(BeamformerCtx *ctx, u32 start_index, u32 needed_frames)
    176 {
    177 	start_index = start_index % ARRAY_COUNT(ctx->beamform_frames);
    178 
    179 	ComputeFrameIterator result;
    180 	result.frames        = ctx->beamform_frames;
    181 	result.offset        = start_index;
    182 	result.capacity      = ARRAY_COUNT(ctx->beamform_frames);
    183 	result.cursor        = 0;
    184 	result.needed_frames = needed_frames;
    185 	return result;
    186 }
    187 
    188 function BeamformerFrame *
    189 frame_next(ComputeFrameIterator *bfi)
    190 {
    191 	BeamformerFrame *result = 0;
    192 	if (bfi->cursor != bfi->needed_frames) {
    193 		u32 index = (bfi->offset + bfi->cursor++) % bfi->capacity;
    194 		result    = bfi->frames + index;
    195 	}
    196 	return result;
    197 }
    198 
    199 function b32
    200 beamformer_frame_compatible(BeamformerFrame *f, iv3 dim, GLenum gl_kind)
    201 {
    202 	b32 result = gl_kind == f->gl_kind && iv3_equal(dim, f->dim);
    203 	return result;
    204 }
    205 
    206 function iv3
    207 das_valid_points(iv3 points)
    208 {
    209 	iv3 result;
    210 	result.x = Max(points.x, 1);
    211 	result.y = Max(points.y, 1);
    212 	result.z = Max(points.z, 1);
    213 	return result;
    214 }
    215 
    216 function void
    217 alloc_beamform_frame(BeamformerFrame *out, iv3 out_dim, GLenum gl_kind, s8 name, Arena arena)
    218 {
    219 	out->dim = das_valid_points(out_dim);
    220 
    221 	/* NOTE: allocate storage for beamformed output data;
    222 	 * this is shared between compute and fragment shaders */
    223 	u32 max_dim = (u32)Max(out->dim.x, Max(out->dim.y, out->dim.z));
    224 	out->mips   = (i32)ctz_u32(round_up_power_of_2(max_dim)) + 1;
    225 
    226 	out->gl_kind = gl_kind;
    227 
    228 	Stream label = arena_stream(arena);
    229 	stream_append_s8(&label, name);
    230 	stream_append_byte(&label, '[');
    231 	stream_append_hex_u64(&label, out->id);
    232 	stream_append_byte(&label, ']');
    233 
    234 	glDeleteTextures(1, &out->texture);
    235 	glCreateTextures(GL_TEXTURE_3D, 1, &out->texture);
    236 	glTextureStorage3D(out->texture, out->mips, gl_kind, out->dim.x, out->dim.y, out->dim.z);
    237 
    238 	glTextureParameteri(out->texture, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
    239 	glTextureParameteri(out->texture, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
    240 
    241 	LABEL_GL_OBJECT(GL_TEXTURE, out->texture, stream_to_s8(&label));
    242 }
    243 
    244 function void
    245 update_hadamard_texture(BeamformerComputePlan *cp, i32 order, Arena arena)
    246 {
    247 	f32 *hadamard = make_hadamard_transpose(&arena, order);
    248 	if (hadamard) {
    249 		cp->hadamard_order = order;
    250 		u32 *texture = cp->textures + BeamformerComputeTextureKind_Hadamard;
    251 		glDeleteTextures(1, texture);
    252 		glCreateTextures(GL_TEXTURE_2D, 1, texture);
    253 		glTextureStorage2D(*texture, 1, GL_R32F, order, order);
    254 		glTextureSubImage2D(*texture, 0, 0, 0, order, order, GL_RED, GL_FLOAT, hadamard);
    255 
    256 		Stream label = arena_stream(arena);
    257 		stream_append_s8(&label, s8("Hadamard"));
    258 		stream_append_i64(&label, order);
    259 		LABEL_GL_OBJECT(GL_TEXTURE, *texture, stream_to_s8(&label));
    260 	}
    261 }
    262 
    263 function void
    264 alloc_shader_storage(BeamformerCtx *ctx, u32 decoded_data_size, Arena arena)
    265 {
    266 	BeamformerComputeContext *cc = &ctx->compute_context;
    267 	glDeleteBuffers(countof(cc->ping_pong_ssbos), cc->ping_pong_ssbos);
    268 	glCreateBuffers(countof(cc->ping_pong_ssbos), cc->ping_pong_ssbos);
    269 
    270 	cc->ping_pong_ssbo_size = decoded_data_size;
    271 
    272 	Stream label = arena_stream(arena);
    273 	stream_append_s8(&label, s8("PingPongSSBO["));
    274 	i32 s_widx = label.widx;
    275 	for (i32 i = 0; i < countof(cc->ping_pong_ssbos); i++) {
    276 		glNamedBufferStorage(cc->ping_pong_ssbos[i], (iz)decoded_data_size, 0, 0);
    277 		stream_append_i64(&label, i);
    278 		stream_append_byte(&label, ']');
    279 		LABEL_GL_OBJECT(GL_BUFFER, cc->ping_pong_ssbos[i], stream_to_s8(&label));
    280 		stream_reset(&label, s_widx);
    281 	}
    282 
    283 	/* TODO(rnp): (25.08.04) cuda lib is heavily broken atm. First there are multiple RF
    284 	 * buffers and cuda decode shouldn't assume that the data is coming from the rf_buffer
    285 	 * ssbo. Second each parameter block may need a different hadamard matrix so ideally
    286 	 * decode should just take the texture as a parameter. Third, none of these dimensions
    287 	 * need to be pre-known by the library unless its allocating GPU memory which it shouldn't
    288 	 * need to do. For now grab out of parameter block 0 but it is not correct */
    289 	BeamformerParameterBlock *pb = beamformer_parameter_block(ctx->shared_memory, 0);
    290 	/* NOTE(rnp): these are stubs when CUDA isn't supported */
    291 	cuda_register_buffers(cc->ping_pong_ssbos, countof(cc->ping_pong_ssbos), cc->rf_buffer.ssbo);
    292 	u32 decoded_data_dimension[3] = {pb->parameters.sample_count, pb->parameters.channel_count, pb->parameters.acquisition_count};
    293 	cuda_init(pb->parameters.raw_data_dimensions.E, decoded_data_dimension);
    294 }
    295 
    296 function void
    297 push_compute_timing_info(ComputeTimingTable *t, ComputeTimingInfo info)
    298 {
    299 	u32 index = atomic_add_u32(&t->write_index, 1) % countof(t->buffer);
    300 	t->buffer[index] = info;
    301 }
    302 
    303 function b32
    304 fill_frame_compute_work(BeamformerCtx *ctx, BeamformWork *work, BeamformerViewPlaneTag plane,
    305                         u32 parameter_block, b32 indirect)
    306 {
    307 	b32 result = work != 0;
    308 	if (result) {
    309 		u32 frame_id    = atomic_add_u32(&ctx->next_render_frame_index, 1);
    310 		u32 frame_index = frame_id % countof(ctx->beamform_frames);
    311 		work->kind      = indirect? BeamformerWorkKind_ComputeIndirect : BeamformerWorkKind_Compute;
    312 		work->lock      = BeamformerSharedMemoryLockKind_DispatchCompute;
    313 		work->compute_context.parameter_block = parameter_block;
    314 		work->compute_context.frame = ctx->beamform_frames + frame_index;
    315 		work->compute_context.frame->ready_to_present = 0;
    316 		work->compute_context.frame->view_plane_tag   = plane;
    317 		work->compute_context.frame->id               = frame_id;
    318 	}
    319 	return result;
    320 }
    321 
    322 function void
    323 do_sum_shader(BeamformerComputeContext *cc, u32 *in_textures, u32 in_texture_count,
    324               u32 out_texture, iv3 out_data_dim)
    325 {
    326 	/* NOTE: zero output before summing */
    327 	glClearTexImage(out_texture, 0, GL_RED, GL_FLOAT, 0);
    328 	glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT);
    329 
    330 	glBindImageTexture(0, out_texture, 0, GL_TRUE, 0, GL_READ_WRITE, GL_RG32F);
    331 	for (u32 i = 0; i < in_texture_count; i++) {
    332 		glBindImageTexture(1, in_textures[i], 0, GL_TRUE, 0, GL_READ_ONLY, GL_RG32F);
    333 		glDispatchCompute(ORONE((u32)out_data_dim.x / 32u),
    334 		                  ORONE((u32)out_data_dim.y),
    335 		                  ORONE((u32)out_data_dim.z / 32u));
    336 		glMemoryBarrier(GL_SHADER_IMAGE_ACCESS_BARRIER_BIT);
    337 	}
    338 }
    339 
    340 function void
    341 plan_compute_pipeline(BeamformerComputePlan *cp, BeamformerParameterBlock *pb)
    342 {
    343 	b32 run_cuda_hilbert = 0;
    344 	b32 demodulate       = 0;
    345 
    346 	for (u32 i = 0; i < pb->pipeline.shader_count; i++) {
    347 		switch (pb->pipeline.shaders[i]) {
    348 		case BeamformerShaderKind_CudaHilbert:{ run_cuda_hilbert = 1; }break;
    349 		case BeamformerShaderKind_Demodulate:{  demodulate = 1;       }break;
    350 		default:{}break;
    351 		}
    352 	}
    353 
    354 	if (demodulate) run_cuda_hilbert = 0;
    355 
    356 	BeamformerDataKind data_kind = pb->pipeline.data_kind;
    357 	cp->iq_pipeline = data_kind == BeamformerDataKind_Float32Complex ||
    358 	                  data_kind == BeamformerDataKind_Int16Complex   ||
    359 	                  demodulate || run_cuda_hilbert;
    360 
    361 	f32 sampling_frequency = pb->parameters.sampling_frequency;
    362 	u32 decimation_rate = Max(pb->parameters.decimation_rate, 1);
    363 	u32 sample_count    = pb->parameters.sample_count;
    364 	if (demodulate) {
    365 		sample_count       /= (2 * decimation_rate);
    366 		sampling_frequency /= 2 * (f32)decimation_rate;
    367 	}
    368 
    369 	cp->rf_size = sample_count * pb->parameters.channel_count * pb->parameters.acquisition_count;
    370 	if (cp->iq_pipeline) cp->rf_size *= 8;
    371 	else                 cp->rf_size *= 4;
    372 
    373 	u32 das_sample_stride   = 1;
    374 	u32 das_transmit_stride = sample_count;
    375 	u32 das_channel_stride  = sample_count * pb->parameters.acquisition_count;
    376 
    377 	f32 time_offset = pb->parameters.time_offset;
    378 
    379 	// TODO(rnp): subgroup size
    380 	u32 subgroup_size = gl_parameters.vendor_id == GLVendor_NVIDIA ? 32 : 64;
    381 
    382 	cp->pipeline.shader_count = 0;
    383 	for (u32 i = 0; i < pb->pipeline.shader_count; i++) {
    384 		BeamformerShaderParameters *sp = pb->pipeline.parameters + i;
    385 		u32 slot   = cp->pipeline.shader_count;
    386 		u32 shader = pb->pipeline.shaders[i];
    387 		b32 commit = 0;
    388 
    389 		BeamformerShaderDescriptor *ld = cp->shader_descriptors + slot - 1;
    390 		BeamformerShaderDescriptor *sd = cp->shader_descriptors + slot;
    391 		zero_struct(sd);
    392 
    393 		switch (shader) {
    394 		case BeamformerShaderKind_CudaHilbert:{ commit = run_cuda_hilbert; }break;
    395 		case BeamformerShaderKind_Decode:{
    396 			/* TODO(rnp): rework decode first and demodulate after */
    397 			b32 first = slot == 0;
    398 
    399 			sd->bake.data_kind = data_kind;
    400 			if (!first) {
    401 				if (data_kind == BeamformerDataKind_Int16) {
    402 					sd->bake.data_kind = BeamformerDataKind_Int16Complex;
    403 				} else {
    404 					sd->bake.data_kind = BeamformerDataKind_Float32Complex;
    405 				}
    406 			}
    407 
    408 			BeamformerShaderKind *last_shader = cp->pipeline.shaders + slot - 1;
    409 			assert(first || ((*last_shader == BeamformerShaderKind_Demodulate ||
    410 			                  *last_shader == BeamformerShaderKind_Filter)));
    411 
    412 			BeamformerShaderDecodeBakeParameters *db = &sd->bake.Decode;
    413 			db->decode_mode    = pb->parameters.decode_mode;
    414 			db->transmit_count = pb->parameters.acquisition_count;
    415 
    416 			u32 channel_stride         = pb->parameters.acquisition_count * pb->parameters.sample_count;
    417 			db->input_sample_stride    = first? 1                           : ld->bake.Filter.output_sample_stride;
    418 			db->input_channel_stride   = first? channel_stride              : ld->bake.Filter.output_channel_stride;
    419 			db->input_transmit_stride  = first? pb->parameters.sample_count : 1;
    420 
    421 			db->output_sample_stride   = das_sample_stride;
    422 			db->output_channel_stride  = das_channel_stride;
    423 			db->output_transmit_stride = das_transmit_stride;
    424 			if (first) {
    425 				db->output_channel_stride  *= decimation_rate;
    426 				db->output_transmit_stride *= decimation_rate;
    427 			}
    428 
    429 			if (run_cuda_hilbert) sd->bake.flags |= BeamformerShaderDecodeFlags_DilateOutput;
    430 
    431 			if (db->decode_mode == BeamformerDecodeMode_None) {
    432 				sd->layout = (uv3){{subgroup_size, 1, 1}};
    433 
    434 				sd->dispatch.x = (u32)ceil_f32((f32)sample_count                     / (f32)sd->layout.x);
    435 				sd->dispatch.y = (u32)ceil_f32((f32)pb->parameters.channel_count     / (f32)sd->layout.y);
    436 				sd->dispatch.z = (u32)ceil_f32((f32)pb->parameters.acquisition_count / (f32)sd->layout.z);
    437 			} else if (db->transmit_count > 40) {
    438 				sd->bake.flags |= BeamformerShaderDecodeFlags_UseSharedMemory;
    439 				db->to_process = 2;
    440 
    441 				if (db->transmit_count == 48)
    442 					db->to_process = db->transmit_count / 16;
    443 
    444 				b32 use_16z  = db->transmit_count == 48 || db->transmit_count == 80 ||
    445 				               db->transmit_count == 96 || db->transmit_count == 160;
    446 				sd->layout = (uv3){{4, 1, use_16z? 16 : 32}};
    447 
    448 				sd->dispatch.x = (u32)ceil_f32((f32)sample_count                     / (f32)sd->layout.x);
    449 				sd->dispatch.y = (u32)ceil_f32((f32)pb->parameters.channel_count     / (f32)sd->layout.y);
    450 				sd->dispatch.z = (u32)ceil_f32((f32)pb->parameters.acquisition_count / (f32)sd->layout.z / (f32)db->to_process);
    451 			} else {
    452 				db->to_process = 1;
    453 
    454 				/* NOTE(rnp): register caching. using more threads will cause the compiler to do
    455 				 * contortions to avoid spilling registers. using less gives higher performance */
    456 				sd->layout = (uv3){{subgroup_size / 2, 1, 1}};
    457 
    458 				sd->dispatch.x = (u32)ceil_f32((f32)sample_count                 / (f32)sd->layout.x);
    459 				sd->dispatch.y = (u32)ceil_f32((f32)pb->parameters.channel_count / (f32)sd->layout.y);
    460 				sd->dispatch.z = 1;
    461 			}
    462 
    463 			if (first) sd->dispatch.x *= decimation_rate;
    464 
    465 			/* NOTE(rnp): decode 2 samples per dispatch when data is i16 */
    466 			if (first && data_kind == BeamformerDataKind_Int16)
    467 				sd->dispatch.x = (u32)ceil_f32((f32)sd->dispatch.x / 2);
    468 
    469 			commit = first || db->decode_mode != BeamformerDecodeMode_None;
    470 		}break;
    471 		case BeamformerShaderKind_Demodulate:
    472 		case BeamformerShaderKind_Filter:
    473 		{
    474 			b32 first = slot == 0;
    475 			b32 demod = shader == BeamformerShaderKind_Demodulate;
    476 			BeamformerFilter *f = cp->filters + sp->filter_slot;
    477 
    478 			time_offset += f->time_delay;
    479 
    480 			BeamformerShaderFilterBakeParameters *fb = &sd->bake.Filter;
    481 			fb->filter_length = (u32)f->length;
    482 			if (demod)                 sd->bake.flags |= BeamformerShaderFilterFlags_Demodulate;
    483 			if (f->parameters.complex) sd->bake.flags |= BeamformerShaderFilterFlags_ComplexFilter;
    484 
    485 			sd->bake.data_kind = data_kind;
    486 			if (!first) sd->bake.data_kind = BeamformerDataKind_Float32;
    487 
    488 			/* NOTE(rnp): when we are demodulating we pretend that the sampler was alternating
    489 			 * between sampling the I portion and the Q portion of an IQ signal. Therefore there
    490 			 * is an implicit decimation factor of 2 which must always be included. All code here
    491 			 * assumes that the signal was sampled in such a way that supports this operation.
    492 			 * To recover IQ[n] from the sampled data (RF[n]) we do the following:
    493 			 *   I[n]  = RF[n]
    494 			 *   Q[n]  = RF[n + 1]
    495 			 *   IQ[n] = I[n] - j*Q[n]
    496 			 */
    497 			if (demod) {
    498 				fb->demodulation_frequency = pb->parameters.demodulation_frequency;
    499 				fb->sampling_frequency     = pb->parameters.sampling_frequency / 2;
    500 				fb->decimation_rate        = decimation_rate;
    501 				fb->sample_count           = pb->parameters.sample_count;
    502 
    503 				fb->output_channel_stride  = das_channel_stride;
    504 				fb->output_sample_stride   = das_sample_stride;
    505 				fb->output_transmit_stride = das_transmit_stride;
    506 
    507 				if (first) {
    508 					fb->input_channel_stride  = pb->parameters.sample_count * pb->parameters.acquisition_count / 2;
    509 					fb->input_sample_stride   = 1;
    510 					fb->input_transmit_stride = pb->parameters.sample_count / 2;
    511 
    512 					if (pb->parameters.decode_mode == BeamformerDecodeMode_None) {
    513 						sd->bake.flags |= BeamformerShaderFilterFlags_OutputFloats;
    514 					} else {
    515 						/* NOTE(rnp): output optimized layout for decoding */
    516 						fb->output_channel_stride  = das_channel_stride;
    517 						fb->output_sample_stride   = pb->parameters.acquisition_count;
    518 						fb->output_transmit_stride = 1;
    519 					}
    520 				} else {
    521 					assert(cp->pipeline.shaders[slot - 1] == BeamformerShaderKind_Decode);
    522 					fb->input_channel_stride  = ld->bake.Decode.output_channel_stride;
    523 					fb->input_sample_stride   = ld->bake.Decode.output_sample_stride;
    524 					fb->input_transmit_stride = ld->bake.Decode.output_transmit_stride;
    525 				}
    526 			} else {
    527 				fb->decimation_rate        = 1;
    528 				fb->output_channel_stride  = sample_count * pb->parameters.acquisition_count;
    529 				fb->output_sample_stride   = 1;
    530 				fb->output_transmit_stride = sample_count;
    531 				fb->input_channel_stride   = sample_count * pb->parameters.acquisition_count;
    532 				fb->input_sample_stride    = 1;
    533 				fb->input_transmit_stride  = sample_count;
    534 				fb->sample_count           = sample_count;
    535 			}
    536 
    537 			/* TODO(rnp): filter may need a different dispatch layout */
    538 			sd->layout     = (uv3){{128, 1, 1}};
    539 			sd->dispatch.x = (u32)ceil_f32((f32)sample_count                     / (f32)sd->layout.x);
    540 			sd->dispatch.y = (u32)ceil_f32((f32)pb->parameters.channel_count     / (f32)sd->layout.y);
    541 			sd->dispatch.z = (u32)ceil_f32((f32)pb->parameters.acquisition_count / (f32)sd->layout.z);
    542 
    543 			commit = 1;
    544 		}break;
    545 		case BeamformerShaderKind_DAS:{
    546 			sd->bake.data_kind = BeamformerDataKind_Float32;
    547 			if (cp->iq_pipeline)
    548 				sd->bake.data_kind = BeamformerDataKind_Float32Complex;
    549 
    550 			BeamformerShaderDASBakeParameters *db = &sd->bake.DAS;
    551 			BeamformerDASUBO *du = &cp->das_ubo_data;
    552 			du->xdc_element_pitch      = pb->parameters.xdc_element_pitch;
    553 			db->sampling_frequency     = sampling_frequency;
    554 			db->demodulation_frequency = pb->parameters.demodulation_frequency;
    555 			db->speed_of_sound         = pb->parameters.speed_of_sound;
    556 			db->time_offset            = time_offset;
    557 			db->f_number               = pb->parameters.f_number;
    558 			db->acquisition_kind       = pb->parameters.acquisition_kind;
    559 			db->sample_count           = sample_count;
    560 			db->channel_count          = pb->parameters.channel_count;
    561 			db->acquisition_count      = pb->parameters.acquisition_count;
    562 			db->interpolation_mode     = pb->parameters.interpolation_mode;
    563 			db->transmit_angle         = pb->parameters.focal_vector.E[0];
    564 			db->focus_depth            = pb->parameters.focal_vector.E[1];
    565 			db->transmit_receive_orientation = pb->parameters.transmit_receive_orientation;
    566 
    567 			// NOTE(rnp): old gcc will miscompile an assignment
    568 			mem_copy(du->voxel_transform.E, pb->parameters.das_voxel_transform.E, sizeof(du->voxel_transform));
    569 			mem_copy(du->xdc_transform.E,   pb->parameters.xdc_transform.E,       sizeof(du->xdc_transform));
    570 
    571 			du->voxel_transform = m4_mul(cp->ui_voxel_transform, du->voxel_transform);
    572 
    573 			u32 id = pb->parameters.acquisition_kind;
    574 
    575 			if (id == BeamformerAcquisitionKind_UFORCES || id == BeamformerAcquisitionKind_FORCES)
    576 				du->voxel_transform = m4_mul(du->xdc_transform, du->voxel_transform);
    577 
    578 			if (id == BeamformerAcquisitionKind_UFORCES || id == BeamformerAcquisitionKind_UHERCULES)
    579 				sd->bake.flags |= BeamformerShaderDASFlags_Sparse;
    580 
    581 			if (pb->parameters.single_focus)        sd->bake.flags |= BeamformerShaderDASFlags_SingleFocus;
    582 			if (pb->parameters.single_orientation)  sd->bake.flags |= BeamformerShaderDASFlags_SingleOrientation;
    583 			if (pb->parameters.coherency_weighting) sd->bake.flags |= BeamformerShaderDASFlags_CoherencyWeighting;
    584 			else                                    sd->bake.flags |= BeamformerShaderDASFlags_Fast;
    585 
    586 			sd->layout = (uv3){{1, 1, 1}};
    587 
    588 			b32 has_x = cp->output_points.x > 1;
    589 			b32 has_y = cp->output_points.y > 1;
    590 			b32 has_z = cp->output_points.z > 1;
    591 
    592 			u32 grid_3d_z_size = Max(1, subgroup_size / (4 * 4));
    593 			u32 grid_2d_y_size = Max(1, subgroup_size / 8);
    594 
    595 			switch (iv3_dimension(cp->output_points)) {
    596 
    597 			case 1:{
    598 				if (has_x) sd->layout.x = subgroup_size;
    599 				if (has_y) sd->layout.y = subgroup_size;
    600 				if (has_z) sd->layout.z = subgroup_size;
    601 			}break;
    602 
    603 			case 2:{
    604 				if (has_x && has_y) {sd->layout.x = 8; sd->layout.y = grid_2d_y_size;}
    605 				if (has_x && has_z) {sd->layout.x = 8; sd->layout.z = grid_2d_y_size;}
    606 				if (has_y && has_z) {sd->layout.y = 8; sd->layout.z = grid_2d_y_size;}
    607 			}break;
    608 
    609 			case 3:{sd->layout = (uv3){{4, 4, grid_3d_z_size}};}break;
    610 
    611 			InvalidDefaultCase;
    612 			}
    613 
    614 			sd->dispatch.x = (u32)ceil_f32((f32)cp->output_points.x / sd->layout.x);
    615 			sd->dispatch.y = (u32)ceil_f32((f32)cp->output_points.y / sd->layout.y);
    616 			sd->dispatch.z = (u32)ceil_f32((f32)cp->output_points.z / sd->layout.z);
    617 
    618 			commit = 1;
    619 		}break;
    620 		default:{ commit = 1; }break;
    621 		}
    622 
    623 		if (commit) {
    624 			u32 index = cp->pipeline.shader_count++;
    625 			cp->pipeline.shaders[index]    = shader;
    626 			cp->pipeline.parameters[index] = *sp;
    627 		}
    628 	}
    629 	cp->pipeline.data_kind = data_kind;
    630 }
    631 
    632 function void
    633 stream_push_shader_header(Stream *s, BeamformerShaderKind shader_kind, s8 header)
    634 {
    635 	stream_append_s8s(s, s8("#version 460 core\n\n"), header);
    636 
    637 	switch (shader_kind) {
    638 	case BeamformerShaderKind_DAS:{
    639 		stream_append_s8(s, s8(""
    640 		"layout(location = " str(DAS_CYCLE_T_UNIFORM_LOC)      ") uniform uint  u_cycle_t;\n"
    641 		"layout(location = " str(DAS_FAST_CHANNEL_UNIFORM_LOC) ") uniform int   u_channel;\n\n"
    642 		));
    643 	}break;
    644 	case BeamformerShaderKind_Decode:{
    645 		stream_append_s8s(s, s8(""
    646 		"layout(location = " str(DECODE_FIRST_PASS_UNIFORM_LOC) ") uniform bool u_first_pass;\n\n"
    647 		));
    648 	}break;
    649 	case BeamformerShaderKind_MinMax:{
    650 		stream_append_s8(s, s8("layout(location = " str(MIN_MAX_MIPS_LEVEL_UNIFORM_LOC)
    651 		                       ") uniform int u_mip_map;\n\n"));
    652 	}break;
    653 	case BeamformerShaderKind_Sum:{
    654 		stream_append_s8(s, s8("layout(location = " str(SUM_PRESCALE_UNIFORM_LOC)
    655 		                       ") uniform float u_sum_prescale = 1.0;\n\n"));
    656 	}break;
    657 	default:{}break;
    658 	}
    659 }
    660 
    661 function void
    662 load_compute_shader(BeamformerCtx *ctx, BeamformerComputePlan *cp, u32 shader_slot, Arena arena)
    663 {
    664 	read_only local_persist s8 compute_headers[BeamformerShaderKind_ComputeCount] = {
    665 		/* X(name, type, gltype) */
    666 		#define X(name, t, gltype) "\t" #gltype " " #name ";\n"
    667 		[BeamformerShaderKind_DAS] = s8_comp("layout(std140, binding = 0) uniform parameters {\n"
    668 			BEAMFORMER_DAS_UBO_PARAM_LIST
    669 			"};\n\n"
    670 		),
    671 		#undef X
    672 	};
    673 
    674 	BeamformerShaderKind shader = cp->pipeline.shaders[shader_slot];
    675 
    676 	u32 program          = 0;
    677 	i32 reloadable_index = beamformer_shader_reloadable_index_by_shader[shader];
    678 	if (reloadable_index != -1) {
    679 		BeamformerShaderKind base_shader = beamformer_reloadable_shader_kinds[reloadable_index];
    680 		s8 path;
    681 		if (!BakeShaders)
    682 			path = push_s8_from_parts(&arena, os_path_separator(), s8("shaders"),
    683 		                            beamformer_reloadable_shader_files[reloadable_index]);
    684 
    685 		Stream shader_stream = arena_stream(arena);
    686 		stream_push_shader_header(&shader_stream, base_shader, compute_headers[base_shader]);
    687 
    688 		i32  header_vector_length = beamformer_shader_header_vector_lengths[reloadable_index];
    689 		i32 *header_vector        = beamformer_shader_header_vectors[reloadable_index];
    690 		for (i32 index = 0; index < header_vector_length; index++)
    691 			stream_append_s8(&shader_stream, beamformer_shader_global_header_strings[header_vector[index]]);
    692 
    693 		BeamformerShaderDescriptor *sd = cp->shader_descriptors + shader_slot;
    694 
    695 		if (sd->layout.x != 0) {
    696 			stream_append_s8(&shader_stream,  s8("layout(local_size_x = "));
    697 			stream_append_u64(&shader_stream, sd->layout.x);
    698 			stream_append_s8(&shader_stream,  s8(", local_size_y = "));
    699 			stream_append_u64(&shader_stream, sd->layout.y);
    700 			stream_append_s8(&shader_stream,  s8(", local_size_z = "));
    701 			stream_append_u64(&shader_stream, sd->layout.z);
    702 			stream_append_s8(&shader_stream,  s8(") in;\n\n"));
    703 		}
    704 
    705 		u32 *parameters = (u32 *)&sd->bake;
    706 		s8  *names      = beamformer_shader_bake_parameter_names[reloadable_index];
    707 		u32  float_bits = beamformer_shader_bake_parameter_float_bits[reloadable_index];
    708 		i32  count      = beamformer_shader_bake_parameter_counts[reloadable_index];
    709 
    710 		for (i32 index = 0; index < count; index++) {
    711 			stream_append_s8s(&shader_stream, s8("#define "), names[index],
    712 			                  (float_bits & (1 << index))? s8(" uintBitsToFloat") : s8(" "), s8("(0x"));
    713 			stream_append_hex_u64(&shader_stream, parameters[index]);
    714 			stream_append_s8(&shader_stream, s8(")\n"));
    715 		}
    716 
    717 		stream_append_s8(&shader_stream, s8("#define DataKind (0x"));
    718 		stream_append_hex_u64(&shader_stream, sd->bake.data_kind);
    719 		stream_append_s8(&shader_stream, s8(")\n\n"));
    720 
    721 		s8  *flag_names = beamformer_shader_flag_strings[reloadable_index];
    722 		u32  flag_count = beamformer_shader_flag_strings_count[reloadable_index];
    723 		u32  flags      = sd->bake.flags;
    724 		for (u32 bit = 0; bit < flag_count; bit++) {
    725 			stream_append_s8s(&shader_stream, s8("#define "), flag_names[bit],
    726 			                  (flags & (1 << bit))? s8(" 1") : s8(" 0"), s8("\n"));
    727 		}
    728 
    729 		if (!renderdoc_attached())
    730 			stream_append_s8(&shader_stream, s8("\n#line 1\n"));
    731 
    732 		s8 shader_text;
    733 		if (BakeShaders) {
    734 			stream_append_s8(&shader_stream, beamformer_shader_data[reloadable_index]);
    735 			shader_text = arena_stream_commit(&arena, &shader_stream);
    736 		} else {
    737 			shader_text = arena_stream_commit(&arena, &shader_stream);
    738 			i64 length = os_read_entire_file((c8 *)path.data, arena.beg, arena_capacity(&arena, u8));
    739 			shader_text.len += length;
    740 			arena_commit(&arena, length);
    741 		}
    742 
    743 		/* TODO(rnp): instance name */
    744 		s8 shader_name = beamformer_shader_names[shader];
    745 		program = load_shader(arena, &shader_text, (u32 []){GL_COMPUTE_SHADER}, 1, shader_name);
    746 	}
    747 
    748 	glDeleteProgram(cp->programs[shader_slot]);
    749 	cp->programs[shader_slot] = program;
    750 }
    751 
    752 function void
    753 beamformer_commit_parameter_block(BeamformerCtx *ctx, BeamformerComputePlan *cp, u32 block, Arena arena)
    754 {
    755 	BeamformerParameterBlock *pb = beamformer_parameter_block_lock(ctx->shared_memory, block, -1);
    756 	for EachBit(pb->dirty_regions, region) {
    757 		switch (region) {
    758 		case BeamformerParameterBlockRegion_ComputePipeline:
    759 		case BeamformerParameterBlockRegion_Parameters:
    760 		{
    761 			cp->output_points  = das_valid_points(pb->parameters.output_points.xyz);
    762 			cp->average_frames = pb->parameters.output_points.E[3];
    763 
    764 			plan_compute_pipeline(cp, pb);
    765 
    766 			/* NOTE(rnp): these are both handled by plan_compute_pipeline() */
    767 			u32 mask = 1 << BeamformerParameterBlockRegion_ComputePipeline |
    768 			           1 << BeamformerParameterBlockRegion_Parameters;
    769 			pb->dirty_regions &= ~mask;
    770 
    771 			for (u32 shader_slot = 0; shader_slot < cp->pipeline.shader_count; shader_slot++) {
    772 				u128 hash = u128_hash_from_data(cp->shader_descriptors + shader_slot, sizeof(BeamformerShaderDescriptor));
    773 				if (!u128_equal(hash, cp->shader_hashes[shader_slot]))
    774 					cp->dirty_programs |= 1 << shader_slot;
    775 				cp->shader_hashes[shader_slot] = hash;
    776 			}
    777 
    778 			#define X(k, t, v) glNamedBufferSubData(cp->ubos[BeamformerComputeUBOKind_##k], \
    779 			                                        0, sizeof(t), &cp->v ## _ubo_data);
    780 			BEAMFORMER_COMPUTE_UBO_LIST
    781 			#undef X
    782 
    783 			cp->acquisition_count = pb->parameters.acquisition_count;
    784 			cp->acquisition_kind  = pb->parameters.acquisition_kind;
    785 
    786 			u32 decoded_data_size = cp->rf_size;
    787 			if (ctx->compute_context.ping_pong_ssbo_size < decoded_data_size)
    788 				alloc_shader_storage(ctx, decoded_data_size, arena);
    789 
    790 			if (cp->hadamard_order != (i32)cp->acquisition_count)
    791 				update_hadamard_texture(cp, (i32)cp->acquisition_count, arena);
    792 
    793 			mem_copy(cp->voxel_transform.E,  pb->parameters.das_voxel_transform.E, sizeof(cp->voxel_transform));
    794 
    795 			GLenum gl_kind = cp->iq_pipeline ? GL_RG32F : GL_R32F;
    796 			if (cp->average_frames > 1 && !beamformer_frame_compatible(ctx->averaged_frames + 0, cp->output_points, gl_kind)) {
    797 				alloc_beamform_frame(ctx->averaged_frames + 0, cp->output_points, gl_kind, s8("Averaged Frame"), arena);
    798 				alloc_beamform_frame(ctx->averaged_frames + 1, cp->output_points, gl_kind, s8("Averaged Frame"), arena);
    799 			}
    800 		}break;
    801 		case BeamformerParameterBlockRegion_ChannelMapping:{
    802 			cuda_set_channel_mapping(pb->channel_mapping);
    803 		}break;
    804 		case BeamformerParameterBlockRegion_FocalVectors:
    805 		case BeamformerParameterBlockRegion_SparseElements:
    806 		case BeamformerParameterBlockRegion_TransmitReceiveOrientations:
    807 		{
    808 			BeamformerComputeTextureKind texture_kind = 0;
    809 			u32 pixel_type = 0, texture_format = 0;
    810 			switch (region) {
    811 			#define X(kind, _gl, tf, pt, ...) \
    812 			case BeamformerParameterBlockRegion_## kind:{            \
    813 				texture_kind   = BeamformerComputeTextureKind_## kind; \
    814 				texture_format = tf;                                   \
    815 				pixel_type     = pt;                                   \
    816 			}break;
    817 			BEAMFORMER_COMPUTE_TEXTURE_LIST
    818 			#undef X
    819 			InvalidDefaultCase;
    820 			}
    821 			glTextureSubImage1D(cp->textures[texture_kind], 0, 0, BeamformerMaxChannelCount,
    822 			                    texture_format, pixel_type,
    823 			                    (u8 *)pb + BeamformerParameterBlockRegionOffsets[region]);
    824 		}break;
    825 		}
    826 	}
    827 	beamformer_parameter_block_unlock(ctx->shared_memory, block);
    828 }
    829 
    830 function void
    831 do_compute_shader(BeamformerCtx *ctx, BeamformerComputePlan *cp, BeamformerFrame *frame,
    832                   BeamformerShaderKind shader, u32 shader_slot, BeamformerShaderParameters *sp, Arena arena)
    833 {
    834 	BeamformerComputeContext *cc = &ctx->compute_context;
    835 
    836 	u32 program = cp->programs[shader_slot];
    837 	glUseProgram(program);
    838 
    839 	u32 output_ssbo_idx = !cc->last_output_ssbo_index;
    840 	u32 input_ssbo_idx  = cc->last_output_ssbo_index;
    841 
    842 	uv3 dispatch = cp->shader_descriptors[shader_slot].dispatch;
    843 	switch (shader) {
    844 	case BeamformerShaderKind_Decode:{
    845 		glBindImageTexture(0, cp->textures[BeamformerComputeTextureKind_Hadamard], 0, 0, 0, GL_READ_ONLY, GL_R32F);
    846 
    847 		BeamformerDecodeMode mode = cp->shader_descriptors[shader_slot].bake.Decode.decode_mode;
    848 		if (shader_slot == 0) {
    849 			if (mode != BeamformerDecodeMode_None) {
    850 				glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 2, cc->ping_pong_ssbos[input_ssbo_idx]);
    851 				glProgramUniform1ui(program, DECODE_FIRST_PASS_UNIFORM_LOC, 1);
    852 
    853 				glDispatchCompute(dispatch.x, dispatch.y, dispatch.z);
    854 				glMemoryBarrier(GL_SHADER_STORAGE_BARRIER_BIT);
    855 			}
    856 		}
    857 
    858 		if (mode != BeamformerDecodeMode_None)
    859 			glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 1, cc->ping_pong_ssbos[input_ssbo_idx]);
    860 
    861 		glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 3, cc->ping_pong_ssbos[output_ssbo_idx]);
    862 
    863 		glProgramUniform1ui(program, DECODE_FIRST_PASS_UNIFORM_LOC, 0);
    864 
    865 		glDispatchCompute(dispatch.x, dispatch.y, dispatch.z);
    866 		glMemoryBarrier(GL_SHADER_STORAGE_BARRIER_BIT);
    867 
    868 		cc->last_output_ssbo_index = !cc->last_output_ssbo_index;
    869 	}break;
    870 	case BeamformerShaderKind_CudaDecode:{
    871 		cuda_decode(0, output_ssbo_idx, 0);
    872 		cc->last_output_ssbo_index = !cc->last_output_ssbo_index;
    873 	}break;
    874 	case BeamformerShaderKind_CudaHilbert:{
    875 		cuda_hilbert(input_ssbo_idx, output_ssbo_idx);
    876 		cc->last_output_ssbo_index = !cc->last_output_ssbo_index;
    877 	}break;
    878 	case BeamformerShaderKind_Filter:
    879 	case BeamformerShaderKind_Demodulate:
    880 	{
    881 		if (shader_slot != 0)
    882 			glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 1, cc->ping_pong_ssbos[input_ssbo_idx]);
    883 		glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 2, cc->ping_pong_ssbos[output_ssbo_idx]);
    884 		glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 3, cp->filters[sp->filter_slot].ssbo);
    885 
    886 		glDispatchCompute(dispatch.x, dispatch.y, dispatch.z);
    887 		glMemoryBarrier(GL_SHADER_STORAGE_BARRIER_BIT);
    888 
    889 		cc->last_output_ssbo_index = !cc->last_output_ssbo_index;
    890 	}break;
    891 	case BeamformerShaderKind_MinMax:{
    892 		for (i32 i = 1; i < frame->mips; i++) {
    893 			glBindImageTexture(0, frame->texture, i - 1, GL_TRUE, 0, GL_READ_ONLY,  GL_RG32F);
    894 			glBindImageTexture(1, frame->texture, i - 0, GL_TRUE, 0, GL_WRITE_ONLY, GL_RG32F);
    895 			glProgramUniform1i(program, MIN_MAX_MIPS_LEVEL_UNIFORM_LOC, i);
    896 
    897 			u32 width  = (u32)frame->dim.x >> i;
    898 			u32 height = (u32)frame->dim.y >> i;
    899 			u32 depth  = (u32)frame->dim.z >> i;
    900 			glDispatchCompute(ORONE(width / 32), ORONE(height), ORONE(depth / 32));
    901 			glMemoryBarrier(GL_SHADER_IMAGE_ACCESS_BARRIER_BIT);
    902 		}
    903 	}break;
    904 	case BeamformerShaderKind_DAS:{
    905 		local_persist u32 das_cycle_t = 0;
    906 
    907 		BeamformerShaderBakeParameters *bp = &cp->shader_descriptors[shader_slot].bake;
    908 		b32 fast   = (bp->flags & BeamformerShaderDASFlags_Fast)   != 0;
    909 		b32 sparse = (bp->flags & BeamformerShaderDASFlags_Sparse) != 0;
    910 
    911 		if (fast) {
    912 			glClearTexImage(frame->texture, 0, GL_RED, GL_FLOAT, 0);
    913 			glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT);
    914 			glBindImageTexture(0, frame->texture, 0, GL_TRUE, 0, GL_READ_WRITE, cp->iq_pipeline ? GL_RG32F : GL_R32F);
    915 		} else {
    916 			glBindImageTexture(0, frame->texture, 0, GL_TRUE, 0, GL_WRITE_ONLY, cp->iq_pipeline ? GL_RG32F : GL_R32F);
    917 		}
    918 
    919 		u32 sparse_texture = cp->textures[BeamformerComputeTextureKind_SparseElements];
    920 		if (!sparse) sparse_texture = 0;
    921 
    922 		glBindBufferBase(GL_UNIFORM_BUFFER, 0, cp->ubos[BeamformerComputeUBOKind_DAS]);
    923 		glBindBufferRange(GL_SHADER_STORAGE_BUFFER, 1, cc->ping_pong_ssbos[input_ssbo_idx], 0, cp->rf_size);
    924 		glBindImageTexture(1, sparse_texture, 0, 0, 0, GL_READ_ONLY, GL_R16I);
    925 		glBindImageTexture(2, cp->textures[BeamformerComputeTextureKind_FocalVectors], 0, 0, 0, GL_READ_ONLY, GL_RG32F);
    926 		glBindImageTexture(3, cp->textures[BeamformerComputeTextureKind_TransmitReceiveOrientations], 0, 0, 0, GL_READ_ONLY, GL_R8I);
    927 
    928 		glProgramUniform1ui(program, DAS_CYCLE_T_UNIFORM_LOC, das_cycle_t++);
    929 
    930 		if (fast) {
    931 			i32 loop_end;
    932 			if (bp->DAS.acquisition_kind == BeamformerAcquisitionKind_RCA_VLS ||
    933 			    bp->DAS.acquisition_kind == BeamformerAcquisitionKind_RCA_TPW)
    934 			{
    935 				/* NOTE(rnp): to avoid repeatedly sampling the whole focal vectors
    936 				 * texture we loop over transmits for VLS/TPW */
    937 				loop_end = (i32)bp->DAS.acquisition_count;
    938 			} else {
    939 				loop_end = (i32)bp->DAS.channel_count;
    940 			}
    941 			f32 percent_per_step = 1.0f / (f32)loop_end;
    942 			cc->processing_progress = -percent_per_step;
    943 			for (i32 index = 0; index < loop_end; index++) {
    944 				cc->processing_progress += percent_per_step;
    945 				/* IMPORTANT(rnp): prevents OS from coalescing and killing our shader */
    946 				glFinish();
    947 				glProgramUniform1i(program, DAS_FAST_CHANNEL_UNIFORM_LOC, index);
    948 				glDispatchCompute(dispatch.x, dispatch.y, dispatch.z);
    949 				glMemoryBarrier(GL_SHADER_IMAGE_ACCESS_BARRIER_BIT);
    950 			}
    951 		} else {
    952 			glDispatchCompute(dispatch.x, dispatch.y, dispatch.z);
    953 		}
    954 		glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT|GL_SHADER_IMAGE_ACCESS_BARRIER_BIT);
    955 	}break;
    956 	case BeamformerShaderKind_Sum:{
    957 		u32 aframe_index = ctx->averaged_frame_index % ARRAY_COUNT(ctx->averaged_frames);
    958 		BeamformerFrame *aframe = ctx->averaged_frames + aframe_index;
    959 		aframe->id              = ctx->averaged_frame_index;
    960 		atomic_store_u32(&aframe->ready_to_present, 0);
    961 		/* TODO(rnp): hack we need a better way of specifying which frames to sum;
    962 		 * this is fine for rolling averaging but what if we want to do something else */
    963 		assert(frame >= ctx->beamform_frames);
    964 		assert(frame < ctx->beamform_frames + countof(ctx->beamform_frames));
    965 		u32 base_index   = (u32)(frame - ctx->beamform_frames);
    966 		u32 to_average   = (u32)cp->average_frames;
    967 		u32 frame_count  = 0;
    968 		u32 *in_textures = push_array(&arena, u32, BeamformerMaxSavedFrames);
    969 		ComputeFrameIterator cfi = compute_frame_iterator(ctx, 1 + base_index - to_average, to_average);
    970 		for (BeamformerFrame *it = frame_next(&cfi); it; it = frame_next(&cfi))
    971 			in_textures[frame_count++] = it->texture;
    972 
    973 		assert(to_average == frame_count);
    974 
    975 		glProgramUniform1f(program, SUM_PRESCALE_UNIFORM_LOC, 1 / (f32)frame_count);
    976 		do_sum_shader(cc, in_textures, frame_count, aframe->texture, aframe->dim);
    977 		mem_copy(aframe->voxel_transform.E,  frame->voxel_transform.E, sizeof(frame->voxel_transform));
    978 		aframe->compound_count   = frame->compound_count;
    979 		aframe->acquisition_kind = frame->acquisition_kind;
    980 	}break;
    981 	InvalidDefaultCase;
    982 	}
    983 }
    984 
    985 function s8
    986 shader_text_with_header(s8 header, s8 filepath, b32 has_file, BeamformerShaderKind shader_kind, Arena *arena)
    987 {
    988 	Stream sb = arena_stream(*arena);
    989 	stream_push_shader_header(&sb, shader_kind, header);
    990 	stream_append_s8(&sb, s8("\n#line 1\n"));
    991 
    992 	s8 result;
    993 	if (BakeShaders) {
    994 		/* TODO(rnp): better handling of shaders with no backing file */
    995 		if (has_file) {
    996 			i32 reloadable_index = beamformer_shader_reloadable_index_by_shader[shader_kind];
    997 			stream_append_s8(&sb, beamformer_shader_data[reloadable_index]);
    998 		}
    999 		result = arena_stream_commit(arena, &sb);
   1000 	} else {
   1001 		result = arena_stream_commit(arena, &sb);
   1002 		if (has_file) {
   1003 			i64 length = os_read_entire_file((c8 *)filepath.data, arena->beg, arena_capacity(arena, u8));
   1004 			result.len += length;
   1005 			arena_commit(arena, length);
   1006 		}
   1007 	}
   1008 
   1009 	return result;
   1010 }
   1011 
   1012 /* NOTE(rnp): currently this function is only handling rendering shaders.
   1013  * look at load_compute_shader for compute shaders */
   1014 function void
   1015 beamformer_reload_shader(BeamformerCtx *ctx, BeamformerShaderReloadContext *src, Arena arena, s8 shader_name)
   1016 {
   1017 	BeamformerShaderKind kind = beamformer_reloadable_shader_kinds[src->reloadable_info_index];
   1018 	assert(kind == BeamformerShaderKind_Render3D);
   1019 
   1020 	s8 path = push_s8_from_parts(&arena, os_path_separator(), s8("shaders"),
   1021 	                             beamformer_reloadable_shader_files[src->reloadable_info_index]);
   1022 
   1023 	i32 shader_count = 1;
   1024 	BeamformerShaderReloadContext *link = src->link;
   1025 	while (link != src) { shader_count++; link = link->link; }
   1026 
   1027 	s8  *shader_texts = push_array(&arena, s8,  shader_count);
   1028 	u32 *shader_types = push_array(&arena, u32, shader_count);
   1029 
   1030 	i32 index = 0;
   1031 	do {
   1032 		b32 has_file = link->reloadable_info_index >= 0;
   1033 		shader_texts[index] = shader_text_with_header(link->header, path, has_file, kind, &arena);
   1034 		shader_types[index] = link->gl_type;
   1035 		index++;
   1036 		link = link->link;
   1037 	} while (link != src);
   1038 
   1039 	u32 *shader = &ctx->frame_view_render_context.shader;
   1040 	glDeleteProgram(*shader);
   1041 	*shader = load_shader(arena, shader_texts, shader_types, shader_count, shader_name);
   1042 	ctx->frame_view_render_context.updated = 1;
   1043 }
   1044 
   1045 function void
   1046 complete_queue(BeamformerCtx *ctx, BeamformWorkQueue *q, Arena *arena, iptr gl_context)
   1047 {
   1048 	BeamformerComputeContext * cs = &ctx->compute_context;
   1049 	BeamformerSharedMemory *   sm = ctx->shared_memory;
   1050 
   1051 	BeamformWork *work = beamform_work_queue_pop(q);
   1052 	while (work) {
   1053 		b32 can_commit = 1;
   1054 		switch (work->kind) {
   1055 		case BeamformerWorkKind_ExportBuffer:{
   1056 			/* TODO(rnp): better way of handling DispatchCompute barrier */
   1057 			post_sync_barrier(ctx->shared_memory, BeamformerSharedMemoryLockKind_DispatchCompute);
   1058 			beamformer_shared_memory_take_lock(ctx->shared_memory, (i32)work->lock, (u32)-1);
   1059 			BeamformerExportContext *ec = &work->export_context;
   1060 			switch (ec->kind) {
   1061 			case BeamformerExportKind_BeamformedData:{
   1062 				BeamformerFrame *frame = ctx->latest_frame;
   1063 				if (frame) {
   1064 					assert(frame->ready_to_present);
   1065 					u32 texture  = frame->texture;
   1066 					iv3 dim      = frame->dim;
   1067 					u32 out_size = (u32)dim.x * (u32)dim.y * (u32)dim.z * 2 * sizeof(f32);
   1068 					if (out_size <= ec->size) {
   1069 						glGetTextureImage(texture, 0, GL_RG, GL_FLOAT, (i32)out_size,
   1070 						                  beamformer_shared_memory_scratch_arena(sm, ctx->shared_memory_size).beg);
   1071 					}
   1072 				}
   1073 			}break;
   1074 			case BeamformerExportKind_Stats:{
   1075 				ComputeTimingTable *table = ctx->compute_timing_table;
   1076 				/* NOTE(rnp): do a little spin to let this finish updating */
   1077 				spin_wait(table->write_index != atomic_load_u32(&table->read_index));
   1078 				ComputeShaderStats *stats = ctx->compute_shader_stats;
   1079 				if (sizeof(stats->table) <= ec->size)
   1080 					mem_copy(beamformer_shared_memory_scratch_arena(sm, ctx->shared_memory_size).beg,
   1081 					         &stats->table, sizeof(stats->table));
   1082 			}break;
   1083 			InvalidDefaultCase;
   1084 			}
   1085 			beamformer_shared_memory_release_lock(ctx->shared_memory, work->lock);
   1086 			post_sync_barrier(ctx->shared_memory, BeamformerSharedMemoryLockKind_ExportSync);
   1087 		}break;
   1088 		case BeamformerWorkKind_CreateFilter:{
   1089 			/* TODO(rnp): this should probably get deleted and moved to lazy loading */
   1090 			BeamformerCreateFilterContext *fctx = &work->create_filter_context;
   1091 			u32 block = fctx->parameter_block;
   1092 			u32 slot  = fctx->filter_slot;
   1093 			BeamformerComputePlan *cp = beamformer_compute_plan_for_block(cs, block, arena);
   1094 			beamformer_filter_update(cp->filters + slot, fctx->parameters, block, slot, *arena);
   1095 		}break;
   1096 		case BeamformerWorkKind_ComputeIndirect:{
   1097 			fill_frame_compute_work(ctx, work, work->compute_indirect_context.view_plane,
   1098 			                        work->compute_indirect_context.parameter_block, 1);
   1099 		} /* FALLTHROUGH */
   1100 		case BeamformerWorkKind_Compute:{
   1101 			DEBUG_DECL(glClearNamedBufferData(cs->ping_pong_ssbos[0], GL_RG32F, GL_RG, GL_FLOAT, 0);)
   1102 			DEBUG_DECL(glClearNamedBufferData(cs->ping_pong_ssbos[1], GL_RG32F, GL_RG, GL_FLOAT, 0);)
   1103 			DEBUG_DECL(glMemoryBarrier(GL_SHADER_STORAGE_BARRIER_BIT);)
   1104 
   1105 			push_compute_timing_info(ctx->compute_timing_table,
   1106 			                         (ComputeTimingInfo){.kind = ComputeTimingInfoKind_ComputeFrameBegin});
   1107 
   1108 			BeamformerComputePlan *cp = beamformer_compute_plan_for_block(cs, work->compute_context.parameter_block, arena);
   1109 			if (beamformer_parameter_block_dirty(sm, work->compute_context.parameter_block)) {
   1110 				u32 block = work->compute_context.parameter_block;
   1111 				beamformer_commit_parameter_block(ctx, cp, block, *arena);
   1112 				atomic_store_u32(&ctx->ui_dirty_parameter_blocks, (u32)(ctx->beamform_work_queue != q) << block);
   1113 			}
   1114 
   1115 			post_sync_barrier(ctx->shared_memory, work->lock);
   1116 
   1117 			u32 dirty_programs = atomic_swap_u32(&cp->dirty_programs, 0);
   1118 			static_assert(ISPOWEROF2(BeamformerMaxComputeShaderStages),
   1119 			              "max compute shader stages must be power of 2");
   1120 			assert((dirty_programs & ~((u32)BeamformerMaxComputeShaderStages - 1)) == 0);
   1121 			for EachBit(dirty_programs, slot)
   1122 				load_compute_shader(ctx, cp, (u32)slot, *arena);
   1123 
   1124 			atomic_store_u32(&cs->processing_compute, 1);
   1125 			start_renderdoc_capture(gl_context);
   1126 
   1127 			BeamformerFrame *frame = work->compute_context.frame;
   1128 
   1129 			GLenum gl_kind = cp->iq_pipeline ? GL_RG32F : GL_R32F;
   1130 			if (!beamformer_frame_compatible(frame, cp->output_points, gl_kind))
   1131 				alloc_beamform_frame(frame, cp->output_points, gl_kind, s8("Beamformed_Data"), *arena);
   1132 
   1133 			m4 voxel_transform = m4_mul(cp->ui_voxel_transform, cp->voxel_transform);
   1134 			mem_copy(frame->voxel_transform.E, voxel_transform.E, sizeof(voxel_transform));
   1135 			frame->acquisition_kind = cp->acquisition_kind;
   1136 			frame->compound_count   = cp->acquisition_count;
   1137 
   1138 			BeamformerComputeContext  *cc       = &ctx->compute_context;
   1139 			BeamformerComputePipeline *pipeline = &cp->pipeline;
   1140 			/* NOTE(rnp): first stage requires access to raw data buffer directly so we break
   1141 			 * it out into a separate step. This way data can get released as soon as possible */
   1142 			if (pipeline->shader_count > 0) {
   1143 				BeamformerRFBuffer *rf = &cs->rf_buffer;
   1144 				u32 slot = rf->compute_index % countof(rf->compute_syncs);
   1145 
   1146 				if (work->kind == BeamformerWorkKind_ComputeIndirect) {
   1147 					/* NOTE(rnp): compute indirect is used when uploading data. if compute thread
   1148 					 * preempts upload it must wait for the fence to exist. then it must tell the
   1149 					 * GPU to wait for upload to complete before it can start compute */
   1150 					spin_wait(!atomic_load_u64(rf->upload_syncs + slot));
   1151 
   1152 					glWaitSync(rf->upload_syncs[slot], 0, GL_TIMEOUT_IGNORED);
   1153 					glDeleteSync(rf->upload_syncs[slot]);
   1154 					rf->compute_index++;
   1155 				} else {
   1156 					slot = (rf->compute_index - 1) % countof(rf->compute_syncs);
   1157 				}
   1158 
   1159 				glBindBufferRange(GL_SHADER_STORAGE_BUFFER, 1, rf->ssbo, slot * rf->active_rf_size, rf->active_rf_size);
   1160 
   1161 				glBeginQuery(GL_TIME_ELAPSED, cc->shader_timer_ids[0]);
   1162 				do_compute_shader(ctx, cp, frame, pipeline->shaders[0], 0, pipeline->parameters + 0, *arena);
   1163 				glEndQuery(GL_TIME_ELAPSED);
   1164 
   1165 				if (work->kind == BeamformerWorkKind_ComputeIndirect) {
   1166 					atomic_store_u64(rf->compute_syncs + slot, glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, 0));
   1167 					atomic_store_u64(rf->upload_syncs + slot,  0);
   1168 				}
   1169 			}
   1170 
   1171 			b32 did_sum_shader = 0;
   1172 			for (u32 i = 1; i < pipeline->shader_count; i++) {
   1173 				did_sum_shader |= pipeline->shaders[i] == BeamformerShaderKind_Sum;
   1174 				glBeginQuery(GL_TIME_ELAPSED, cc->shader_timer_ids[i]);
   1175 				do_compute_shader(ctx, cp, frame, pipeline->shaders[i], i, pipeline->parameters + i, *arena);
   1176 				glEndQuery(GL_TIME_ELAPSED);
   1177 			}
   1178 
   1179 			/* NOTE(rnp): the first of these blocks until work completes */
   1180 			for (u32 i = 0; i < pipeline->shader_count; i++) {
   1181 				ComputeTimingInfo info = {0};
   1182 				info.kind   = ComputeTimingInfoKind_Shader;
   1183 				info.shader = pipeline->shaders[i];
   1184 				glGetQueryObjectui64v(cc->shader_timer_ids[i], GL_QUERY_RESULT, &info.timer_count);
   1185 				push_compute_timing_info(ctx->compute_timing_table, info);
   1186 			}
   1187 			cs->processing_progress = 1;
   1188 
   1189 			frame->ready_to_present = 1;
   1190 			if (did_sum_shader) {
   1191 				u32 aframe_index = ((ctx->averaged_frame_index++) % countof(ctx->averaged_frames));
   1192 				ctx->averaged_frames[aframe_index].view_plane_tag  = frame->view_plane_tag;
   1193 				ctx->averaged_frames[aframe_index].ready_to_present = 1;
   1194 				atomic_store_u64((u64 *)&ctx->latest_frame, (u64)(ctx->averaged_frames + aframe_index));
   1195 			} else {
   1196 				atomic_store_u64((u64 *)&ctx->latest_frame, (u64)frame);
   1197 			}
   1198 			cs->processing_compute  = 0;
   1199 
   1200 			push_compute_timing_info(ctx->compute_timing_table,
   1201 			                         (ComputeTimingInfo){.kind = ComputeTimingInfoKind_ComputeFrameEnd});
   1202 
   1203 			end_renderdoc_capture(gl_context);
   1204 		}break;
   1205 		InvalidDefaultCase;
   1206 		}
   1207 
   1208 		if (can_commit) {
   1209 			beamform_work_queue_pop_commit(q);
   1210 			work = beamform_work_queue_pop(q);
   1211 		}
   1212 	}
   1213 }
   1214 
   1215 function void
   1216 coalesce_timing_table(ComputeTimingTable *t, ComputeShaderStats *stats)
   1217 {
   1218 	/* TODO(rnp): we do not currently do anything to handle the potential for a half written
   1219 	 * info item. this could result in garbage entries but they shouldn't really matter */
   1220 
   1221 	u32 target = atomic_load_u32(&t->write_index);
   1222 	u32 stats_index = (stats->latest_frame_index + 1) % countof(stats->table.times);
   1223 
   1224 	static_assert(BeamformerShaderKind_Count + 1 <= 32, "timing coalescence bitfield test");
   1225 	u32 seen_info_test = 0;
   1226 
   1227 	while (t->read_index != target) {
   1228 		ComputeTimingInfo info = t->buffer[t->read_index % countof(t->buffer)];
   1229 		switch (info.kind) {
   1230 		case ComputeTimingInfoKind_ComputeFrameBegin:{
   1231 			assert(t->compute_frame_active == 0);
   1232 			t->compute_frame_active = 1;
   1233 			/* NOTE(rnp): allow multiple instances of same shader to accumulate */
   1234 			mem_clear(stats->table.times[stats_index], 0, sizeof(stats->table.times[stats_index]));
   1235 		}break;
   1236 		case ComputeTimingInfoKind_ComputeFrameEnd:{
   1237 			assert(t->compute_frame_active == 1);
   1238 			t->compute_frame_active = 0;
   1239 			stats->latest_frame_index = stats_index;
   1240 			stats_index = (stats_index + 1) % countof(stats->table.times);
   1241 		}break;
   1242 		case ComputeTimingInfoKind_Shader:{
   1243 			stats->table.times[stats_index][info.shader] += (f32)info.timer_count / 1.0e9f;
   1244 			seen_info_test |= (1u << info.shader);
   1245 		}break;
   1246 		case ComputeTimingInfoKind_RF_Data:{
   1247 			stats->latest_rf_index = (stats->latest_rf_index + 1) % countof(stats->table.rf_time_deltas);
   1248 			f32 delta = (f32)(info.timer_count - stats->last_rf_timer_count) / 1.0e9f;
   1249 			stats->table.rf_time_deltas[stats->latest_rf_index] = delta;
   1250 			stats->last_rf_timer_count = info.timer_count;
   1251 			seen_info_test |= (1 << BeamformerShaderKind_Count);
   1252 		}break;
   1253 		}
   1254 		/* NOTE(rnp): do this at the end so that stats table is always in a consistent state */
   1255 		atomic_add_u32(&t->read_index, 1);
   1256 	}
   1257 
   1258 	if (seen_info_test) {
   1259 		for EachEnumValue(BeamformerShaderKind, shader) {
   1260 			if (seen_info_test & (1 << shader)) {
   1261 				f32 sum = 0;
   1262 				for EachElement(stats->table.times, i)
   1263 					sum += stats->table.times[i][shader];
   1264 				stats->average_times[shader] = sum / countof(stats->table.times);
   1265 			}
   1266 		}
   1267 
   1268 		if (seen_info_test & (1 << BeamformerShaderKind_Count)) {
   1269 			f32 sum = 0;
   1270 			for EachElement(stats->table.rf_time_deltas, i)
   1271 				sum += stats->table.rf_time_deltas[i];
   1272 			stats->rf_time_delta_average = sum / countof(stats->table.rf_time_deltas);
   1273 		}
   1274 	}
   1275 }
   1276 
   1277 DEBUG_EXPORT BEAMFORMER_COMPLETE_COMPUTE_FN(beamformer_complete_compute)
   1278 {
   1279 	BeamformerCtx *ctx         = (BeamformerCtx *)user_context;
   1280 	BeamformerSharedMemory *sm = ctx->shared_memory;
   1281 	complete_queue(ctx, &sm->external_work_queue, arena, gl_context);
   1282 	complete_queue(ctx, ctx->beamform_work_queue, arena, gl_context);
   1283 }
   1284 
   1285 function void
   1286 beamformer_rf_buffer_allocate(BeamformerRFBuffer *rf, u32 rf_size, b32 nvidia)
   1287 {
   1288 	assert((rf_size % 64) == 0);
   1289 	if (!nvidia) glUnmapNamedBuffer(rf->ssbo);
   1290 	glDeleteBuffers(1, &rf->ssbo);
   1291 	glCreateBuffers(1, &rf->ssbo);
   1292 
   1293 	u32 buffer_flags = GL_DYNAMIC_STORAGE_BIT;
   1294 	if (!nvidia) buffer_flags |= GL_MAP_PERSISTENT_BIT|GL_MAP_WRITE_BIT;
   1295 
   1296 	glNamedBufferStorage(rf->ssbo, countof(rf->compute_syncs) * rf_size, 0, buffer_flags);
   1297 
   1298 	if (!nvidia) {
   1299 		u32 access = GL_MAP_PERSISTENT_BIT|GL_MAP_WRITE_BIT|GL_MAP_FLUSH_EXPLICIT_BIT|GL_MAP_UNSYNCHRONIZED_BIT;
   1300 		rf->buffer = glMapNamedBufferRange(rf->ssbo, 0, (GLsizei)(countof(rf->compute_syncs) * rf_size), access);
   1301 	}
   1302 
   1303 	LABEL_GL_OBJECT(GL_BUFFER, rf->ssbo, s8("Raw_RF_SSBO"));
   1304 	rf->size = rf_size;
   1305 }
   1306 
   1307 DEBUG_EXPORT BEAMFORMER_RF_UPLOAD_FN(beamformer_rf_upload)
   1308 {
   1309 	BeamformerSharedMemory *sm                  = ctx->shared_memory;
   1310 	BeamformerSharedMemoryLockKind scratch_lock = BeamformerSharedMemoryLockKind_ScratchSpace;
   1311 	BeamformerSharedMemoryLockKind upload_lock  = BeamformerSharedMemoryLockKind_UploadRF;
   1312 
   1313 	u64 rf_block_rf_size;
   1314 	if (atomic_load_u32(sm->locks + upload_lock) &&
   1315 	    (rf_block_rf_size = atomic_swap_u64(&sm->rf_block_rf_size, 0)))
   1316 	{
   1317 		beamformer_shared_memory_take_lock(ctx->shared_memory, (i32)scratch_lock, (u32)-1);
   1318 
   1319 		BeamformerRFBuffer       *rf = ctx->rf_buffer;
   1320 		BeamformerParameterBlock *b  = beamformer_parameter_block(sm, (u32)(rf_block_rf_size >> 32ULL));
   1321 		BeamformerParameters     *bp = &b->parameters;
   1322 		BeamformerDataKind data_kind = b->pipeline.data_kind;
   1323 
   1324 		b32 nvidia = gl_parameters.vendor_id == GLVendor_NVIDIA;
   1325 
   1326 		rf->active_rf_size = (u32)round_up_to(rf_block_rf_size & 0xFFFFFFFFULL, 64);
   1327 		if unlikely(rf->size < rf->active_rf_size)
   1328 			beamformer_rf_buffer_allocate(rf, rf->active_rf_size, nvidia);
   1329 
   1330 		u32 slot = rf->insertion_index++ % countof(rf->compute_syncs);
   1331 
   1332 		/* NOTE(rnp): if the rest of the code is functioning then the first
   1333 		 * time the compute thread processes an upload it must have gone
   1334 		 * through this path. therefore it is safe to spin until it gets processed */
   1335 		spin_wait(atomic_load_u64(rf->upload_syncs + slot));
   1336 
   1337 		if (atomic_load_u64(rf->compute_syncs + slot)) {
   1338 			GLenum sync_result = glClientWaitSync(rf->compute_syncs[slot], 0, 1000000000);
   1339 			if (sync_result == GL_TIMEOUT_EXPIRED || sync_result == GL_WAIT_FAILED) {
   1340 				// TODO(rnp): what do?
   1341 			}
   1342 			glDeleteSync(rf->compute_syncs[slot]);
   1343 		}
   1344 
   1345 		u32 size = bp->channel_count * bp->acquisition_count * bp->sample_count * beamformer_data_kind_byte_size[data_kind];
   1346 		u8 *data = beamformer_shared_memory_scratch_arena(sm, ctx->shared_memory_size).beg;
   1347 
   1348 		if (nvidia) glNamedBufferSubData(rf->ssbo, slot * rf->active_rf_size, (i32)size, data);
   1349 		else        memory_copy_non_temporal(rf->buffer + slot * rf->active_rf_size, data, size);
   1350 		store_fence();
   1351 
   1352 		beamformer_shared_memory_release_lock(ctx->shared_memory, (i32)scratch_lock);
   1353 		post_sync_barrier(ctx->shared_memory, upload_lock);
   1354 
   1355 		if (!nvidia)
   1356 			glFlushMappedNamedBufferRange(rf->ssbo, slot * rf->active_rf_size, (i32)rf->active_rf_size);
   1357 
   1358 		atomic_store_u64(rf->upload_syncs  + slot, glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, 0));
   1359 		atomic_store_u64(rf->compute_syncs + slot, 0);
   1360 
   1361 		os_wake_all_waiters(ctx->compute_worker_sync);
   1362 
   1363 		ComputeTimingInfo info = {.kind = ComputeTimingInfoKind_RF_Data};
   1364 		glGetQueryObjectui64v(rf->data_timestamp_query, GL_QUERY_RESULT, &info.timer_count);
   1365 		glQueryCounter(rf->data_timestamp_query, GL_TIMESTAMP);
   1366 		push_compute_timing_info(ctx->compute_timing_table, info);
   1367 	}
   1368 }
   1369 
   1370 function void
   1371 beamformer_queue_compute(BeamformerCtx *ctx, BeamformerFrame *frame, u32 parameter_block)
   1372 {
   1373 	BeamformerSharedMemory *sm = ctx->shared_memory;
   1374 	BeamformerSharedMemoryLockKind dispatch_lock = BeamformerSharedMemoryLockKind_DispatchCompute;
   1375 	if (!sm->live_imaging_parameters.active && beamformer_shared_memory_take_lock(sm, (i32)dispatch_lock, 0))
   1376 	{
   1377 		BeamformWork *work = beamform_work_queue_push(ctx->beamform_work_queue);
   1378 		BeamformerViewPlaneTag tag = frame ? frame->view_plane_tag : 0;
   1379 		if (fill_frame_compute_work(ctx, work, tag, parameter_block, 0))
   1380 			beamform_work_queue_push_commit(ctx->beamform_work_queue);
   1381 	}
   1382 	os_wake_all_waiters(&ctx->compute_worker.sync_variable);
   1383 }
   1384 
   1385 #include "ui.c"
   1386 
   1387 function void
   1388 beamformer_process_input_events(BeamformerCtx *ctx, BeamformerInput *input,
   1389                                 BeamformerInputEvent *events, u32 event_count)
   1390 {
   1391 	for (u32 index = 0; index < event_count; index++) {
   1392 		BeamformerInputEvent *event = events + index;
   1393 		switch (event->kind) {
   1394 
   1395 		case BeamformerInputEventKind_ExecutableReload:{
   1396 			ui_init(ctx, ctx->ui_backing_store);
   1397 
   1398 			#if BEAMFORMER_RENDERDOC_HOOKS
   1399 			start_frame_capture = input->renderdoc_start_frame_capture;
   1400 			end_frame_capture   = input->renderdoc_end_frame_capture;
   1401 			#endif
   1402 		}break;
   1403 
   1404 		case BeamformerInputEventKind_FileEvent:{
   1405 			BeamformerFileReloadContext *frc = event->file_watch_user_context;
   1406 			switch (frc->kind) {
   1407 			case BeamformerFileReloadKind_Shader:{
   1408 				BeamformerShaderReloadContext *src = frc->shader_reload_context;
   1409 				BeamformerShaderKind kind = beamformer_reloadable_shader_kinds[src->reloadable_info_index];
   1410 				beamformer_reload_shader(ctx, src, ctx->arena, beamformer_shader_names[kind]);
   1411 			}break;
   1412 			case BeamformerFileReloadKind_ComputeShader:{
   1413 				for EachElement(ctx->compute_context.compute_plans, block) {
   1414 					BeamformerComputePlan *cp = ctx->compute_context.compute_plans[block];
   1415 					for (u32 slot = 0; cp && slot < cp->pipeline.shader_count; slot++) {
   1416 						i32 shader_index = beamformer_shader_reloadable_index_by_shader[cp->pipeline.shaders[slot]];
   1417 						if (beamformer_reloadable_shader_kinds[shader_index] == frc->compute_shader_kind)
   1418 							atomic_or_u32(&cp->dirty_programs, 1 << slot);
   1419 					}
   1420 				}
   1421 
   1422 				if (ctx->latest_frame)
   1423 					beamformer_queue_compute(ctx, ctx->latest_frame, ctx->latest_frame->parameter_block);
   1424 			}break;
   1425 			InvalidDefaultCase;
   1426 			}
   1427 		}break;
   1428 
   1429 		InvalidDefaultCase;
   1430 		}
   1431 	}
   1432 }
   1433 
   1434 BEAMFORMER_EXPORT void
   1435 beamformer_frame_step(BeamformerInput *input)
   1436 {
   1437 	BeamformerCtx *ctx = BeamformerContextMemory(input->memory);
   1438 
   1439 	u64 current_time = os_timer_count();
   1440 	dt_for_frame = (f64)(current_time - ctx->frame_timestamp) / os_system_info()->timer_frequency;
   1441 	ctx->frame_timestamp = current_time;
   1442 
   1443 	if (IsWindowResized()) {
   1444 		ctx->window_size.h = GetScreenHeight();
   1445 		ctx->window_size.w = GetScreenWidth();
   1446 	}
   1447 
   1448 	coalesce_timing_table(ctx->compute_timing_table, ctx->compute_shader_stats);
   1449 
   1450 	beamformer_process_input_events(ctx, input, input->event_queue, input->event_count);
   1451 
   1452 	BeamformerSharedMemory *sm = ctx->shared_memory;
   1453 	if (atomic_load_u32(sm->locks + BeamformerSharedMemoryLockKind_UploadRF))
   1454 		os_wake_all_waiters(&ctx->upload_worker.sync_variable);
   1455 	if (atomic_load_u32(sm->locks + BeamformerSharedMemoryLockKind_DispatchCompute))
   1456 		os_wake_all_waiters(&ctx->compute_worker.sync_variable);
   1457 
   1458 	BeamformerFrame        *frame = ctx->latest_frame;
   1459 	BeamformerViewPlaneTag  tag   = frame? frame->view_plane_tag : 0;
   1460 	draw_ui(ctx, input, frame, tag);
   1461 
   1462 	ctx->frame_view_render_context.updated = 0;
   1463 }