drivers/radeonsi/si_perfetto.cpp

/*
 * Copyright 2023 Advanced Micro Devices, Inc.
 *
 * SPDX-License-Identifier: MIT
 */
#include <stdio.h>
#include <stdarg.h>
#include <string.h>

#include "util/hash_table.h"
#include "util/u_process.h"
#include "util/hash_table.h"

#include "si_pipe.h"
#include "si_perfetto.h"
#include "si_tracepoints.h"

#ifdef HAVE_PERFETTO

#include "util/perf/u_perfetto.h"
#include "util/perf/u_perfetto_renderpass.h"

#include "si_tracepoints_perfetto.h"

/* Just naming stages */
static const struct {
   const char *name;

   /* The perfetto UI requires that there is a parent-child relationship
    * within a row of elements. Which means that all children elements must
    * end within the lifespan of their parent.
    *
    * Some elements like stalls and command buffers follow that relationship,
    * but not all. This tells us in which UI row the elements should live.
    */
   enum si_ds_queue_stage draw_stage;
} si_queue_stage_desc[SI_DS_QUEUE_STAGE_N_STAGES] = {
   /* Order must match the enum! */
   {
      "queue",
      SI_DS_QUEUE_STAGE_QUEUE,
   },
   {
      "compute",
      SI_DS_QUEUE_STAGE_COMPUTE,
   },
   {
      "draw",
      SI_DS_QUEUE_STAGE_DRAW,
   }
};

struct SIRenderpassIncrementalState {
   bool was_cleared = true;
};

struct SIRenderpassTraits : public perfetto::DefaultDataSourceTraits {
   using IncrementalStateType = SIRenderpassIncrementalState;
};

class SIRenderpassDataSource : public MesaRenderpassDataSource<SIRenderpassDataSource,
                                                               SIRenderpassTraits> {
};

PERFETTO_DECLARE_DATA_SOURCE_STATIC_MEMBERS(SIRenderpassDataSource);
PERFETTO_DEFINE_DATA_SOURCE_STATIC_MEMBERS(SIRenderpassDataSource);

using perfetto::protos::pbzero::InternedGpuRenderStageSpecification_RenderStageCategory;

static void sync_timestamp(SIRenderpassDataSource::TraceContext &ctx, struct si_ds_device *device)
{
   uint64_t cpu_ts = perfetto::base::GetBootTimeNs().count();
   uint64_t gpu_ts;

   struct si_context *sctx = container_of(device, struct si_context, ds);
   gpu_ts = sctx->screen->b.get_timestamp(&sctx->screen->b);


   cpu_ts = perfetto::base::GetBootTimeNs().count();

   if (cpu_ts < device->next_clock_sync_ns)
      return;

   PERFETTO_LOG("sending clocks gpu=0x%08x", device->gpu_clock_id);

   device->sync_gpu_ts = gpu_ts;
   device->next_clock_sync_ns = cpu_ts + 1000000000ull;
   MesaRenderpassDataSource<SIRenderpassDataSource, SIRenderpassTraits>::
      EmitClockSync(ctx, cpu_ts, gpu_ts, device->gpu_clock_id);
}

static void send_descriptors(SIRenderpassDataSource::TraceContext &ctx,
                             struct si_ds_device *device)
{
   PERFETTO_LOG("Sending renderstage descriptors");

   device->event_id = 0;
   list_for_each_entry_safe(struct si_ds_queue, queue, &device->queues, link) {
      for (uint32_t s = 0; s < ARRAY_SIZE(queue->stages); s++) {
         queue->stages[s].start_ns[0] = 0;
      }
   }

   {
      auto packet = ctx.NewTracePacket();

      packet->set_timestamp(perfetto::base::GetBootTimeNs().count());
      packet->set_timestamp_clock_id(perfetto::protos::pbzero::BUILTIN_CLOCK_BOOTTIME);
      packet->set_sequence_flags(perfetto::protos::pbzero::TracePacket::SEQ_INCREMENTAL_STATE_CLEARED);

      auto interned_data = packet->set_interned_data();

      {
         auto desc = interned_data->add_graphics_contexts();
         desc->set_iid(device->iid);
         desc->set_pid(getpid());
         switch (device->api) {
         case AMD_DS_API_OPENGL:
            desc->set_api(perfetto::protos::pbzero::InternedGraphicsContext_Api::OPEN_GL);
            break;
         case AMD_DS_API_VULKAN:
            desc->set_api(perfetto::protos::pbzero::InternedGraphicsContext_Api::VULKAN);
            break;
         default:
            break;
         }
      }

      /* Emit all the IID picked at device/queue creation. */
      list_for_each_entry_safe(struct si_ds_queue, queue, &device->queues, link) {
         for (unsigned s = 0; s < SI_DS_QUEUE_STAGE_N_STAGES; s++) {
            {
               /* We put the stage number in there so that all rows are order
                * by si_ds_queue_stage.
                */
               char name[100];
               snprintf(name, sizeof(name), "%.10s-%s-%u-%s", util_get_process_name(),
                        queue->name, s, si_queue_stage_desc[s].name);

               auto desc = interned_data->add_gpu_specifications();
               desc->set_iid(queue->stages[s].queue_iid);
               desc->set_name(name);
            }
            {
               auto desc = interned_data->add_gpu_specifications();
               desc->set_iid(queue->stages[s].stage_iid);
               desc->set_name(si_queue_stage_desc[s].name);
            }
         }
      }
   }

   device->next_clock_sync_ns = 0;
   sync_timestamp(ctx, device);
}

typedef void (*trace_payload_as_extra_func)(perfetto::protos::pbzero::GpuRenderStageEvent *,
                                            const void*);

static void begin_event(struct si_ds_queue *queue, uint64_t ts_ns, enum si_ds_queue_stage stage_id)
{
   PERFETTO_LOG("begin event called - ts_ns=%" PRIu64, ts_ns);
   uint32_t level = queue->stages[stage_id].level;
   /* If we haven't managed to calibrate the alignment between GPU and CPU
    * timestamps yet, then skip this trace, otherwise perfetto won't know
    * what to do with it.
    */
   if (!queue->device->sync_gpu_ts) {
      queue->stages[stage_id].start_ns[level] = 0;
      return;
   }

   if (level >= (ARRAY_SIZE(queue->stages[stage_id].start_ns) - 1))
      return;

   queue->stages[stage_id].start_ns[level] = ts_ns;
   queue->stages[stage_id].level++;
}

static void end_event(struct si_ds_queue *queue, uint64_t ts_ns, enum si_ds_queue_stage stage_id,
                      uint32_t submission_id, const char *app_event, const void* payload = nullptr,
                      trace_payload_as_extra_func payload_as_extra = nullptr)
{
   PERFETTO_LOG("end event called - ts_ns=%" PRIu64, ts_ns);
   struct si_ds_device *device = queue->device;

   /* If we haven't managed to calibrate the alignment between GPU and CPU
    * timestamps yet, then skip this trace, otherwise perfetto won't know
    * what to do with it.
    */
   if (!device->sync_gpu_ts)
      return;

   if (queue->stages[stage_id].level == 0)
      return;

   uint32_t level = --queue->stages[stage_id].level;
   struct si_ds_stage *stage = &queue->stages[stage_id];
   uint64_t start_ns = stage->start_ns[level];
   PERFETTO_LOG("end event called - start_ns=%" PRIu64 " ts_ns=%" PRIu64, start_ns, ts_ns);
   if (!start_ns || start_ns > ts_ns)
      return;

   SIRenderpassDataSource::Trace([=](SIRenderpassDataSource::TraceContext tctx) {
      if (auto state = tctx.GetIncrementalState(); state->was_cleared) {
         send_descriptors(tctx, queue->device);
         state->was_cleared = false;
      }

      sync_timestamp(tctx, queue->device);

      uint64_t evt_id = device->event_id++;

      /* If this is an application event, we might need to generate a new
       * stage_iid if not already seen. Otherwise, it's a driver event and we
       * have use the internal stage_iid.
       */
      uint64_t stage_iid = app_event ?
                           tctx.GetDataSourceLocked()->debug_marker_stage(tctx, app_event) :
                           stage->stage_iid;

      auto packet = tctx.NewTracePacket();

      packet->set_timestamp(start_ns);
      packet->set_timestamp_clock_id(queue->device->gpu_clock_id);

      assert(ts_ns >= start_ns);

      auto event = packet->set_gpu_render_stage_event();
      event->set_gpu_id(queue->device->gpu_id);

      event->set_hw_queue_iid(stage->queue_iid);
      event->set_stage_iid(stage_iid);
      event->set_context(queue->device->iid);
      event->set_event_id(evt_id);
      event->set_duration(ts_ns - start_ns);
      event->set_submission_id(submission_id);

      if (payload && payload_as_extra) {
         payload_as_extra(event, payload);
      }
   });

   stage->start_ns[level] = 0;
}

#endif /* HAVE_PERFETTO */

#ifdef __cplusplus
extern "C" {
#endif

#ifdef HAVE_PERFETTO

/*
 * Trace callbacks, called from u_trace once the timestamps from GPU have been
 * collected.
 */

#define CREATE_DUAL_EVENT_CALLBACK(event_name, stage)                                             \
void si_ds_begin_##event_name(struct si_ds_device *device, uint64_t ts_ns, uint16_t tp_idx,       \
                              const void *flush_data,                                             \
                              const struct trace_si_begin_##event_name *payload)                  \
{                                                                                                 \
   const struct si_ds_flush_data *flush = (const struct si_ds_flush_data *) flush_data;           \
   begin_event(flush->queue, ts_ns, stage);                                                       \
}                                                                                                 \
                                                                                                  \
void si_ds_end_##event_name(struct si_ds_device *device, uint64_t ts_ns, uint16_t tp_idx,         \
                            const void *flush_data,                                               \
                            const struct trace_si_end_##event_name *payload)                      \
{                                                                                                 \
   const struct si_ds_flush_data *flush =  (const struct si_ds_flush_data *) flush_data;          \
   end_event(flush->queue, ts_ns, stage, flush->submission_id, NULL, payload,                     \
             (trace_payload_as_extra_func)&trace_payload_as_extra_si_end_##event_name);           \
}                                                                                                 \

CREATE_DUAL_EVENT_CALLBACK(draw, SI_DS_QUEUE_STAGE_DRAW)
CREATE_DUAL_EVENT_CALLBACK(compute, SI_DS_QUEUE_STAGE_COMPUTE)

uint64_t si_ds_begin_submit(struct si_ds_queue *queue)
{
   return perfetto::base::GetBootTimeNs().count();
}

void si_ds_end_submit(struct si_ds_queue *queue, uint64_t start_ts)
{
   if (!u_trace_should_process(&queue->device->trace_context)) {
      queue->device->sync_gpu_ts = 0;
      queue->device->next_clock_sync_ns = 0;
      return;
   }

   uint64_t end_ts = perfetto::base::GetBootTimeNs().count();
   uint32_t submission_id = queue->submission_id++;

   SIRenderpassDataSource::Trace([=](SIRenderpassDataSource::TraceContext tctx) {
      if (auto state = tctx.GetIncrementalState(); state->was_cleared) {
         send_descriptors(tctx, queue->device);
         state->was_cleared = false;
      }

      sync_timestamp(tctx, queue->device);

      auto packet = tctx.NewTracePacket();

      packet->set_timestamp(start_ts);

      auto event = packet->set_vulkan_api_event();
      auto submit = event->set_vk_queue_submit();

      submit->set_duration_ns(end_ts - start_ts);
      submit->set_vk_queue((uintptr_t) queue);
      submit->set_submission_id(submission_id);
   });
}

#endif /* HAVE_PERFETTO */

static void si_driver_ds_init_once(void)
{
#ifdef HAVE_PERFETTO
   util_perfetto_init();
   perfetto::DataSourceDescriptor dsd;
   dsd.set_name("gpu.renderstages.amd");
   SIRenderpassDataSource::Register(dsd);
#endif
}

static once_flag si_driver_ds_once_flag = ONCE_FLAG_INIT;
static uint64_t iid = 1;

static uint64_t get_iid()
{
   return iid++;
}

static uint32_t si_pps_clock_id(uint32_t gpu_id)
{
   char buf[40];
   snprintf(buf, sizeof(buf), "org.freedesktop.mesa.amd.gpu%u", gpu_id);

   return _mesa_hash_string(buf) | 0x80000000;
}

void si_driver_ds_init(void)
{
   call_once(&si_driver_ds_once_flag, si_driver_ds_init_once);
   si_gpu_tracepoint_config_variable();
}

void si_ds_device_init(struct si_ds_device *device, const struct radeon_info *devinfo,
                       uint32_t gpu_id, enum amd_ds_api api)
{
   device->gpu_id = gpu_id;
   device->gpu_clock_id = si_pps_clock_id(gpu_id);
   device->info = devinfo;
   device->iid = get_iid();
   device->api = api;
   list_inithead(&device->queues);
}

void si_ds_device_fini(struct si_ds_device *device)
{
   u_trace_context_fini(&device->trace_context);
}

struct si_ds_queue * si_ds_device_init_queue(struct si_ds_device *device,
                                             struct si_ds_queue *queue,
                                             const char *fmt_name, ...)
{
   va_list ap;
   queue->device = device;

   va_start(ap, fmt_name);
   vsnprintf(queue->name, sizeof(queue->name), fmt_name, ap);
   va_end(ap);

   for (unsigned s = 0; s < SI_DS_QUEUE_STAGE_N_STAGES; s++) {
      queue->stages[s].queue_iid = get_iid();
      queue->stages[s].stage_iid = get_iid();
   }

   list_add(&queue->link, &device->queues);

   return queue;
}

void si_ds_flush_data_init(struct si_ds_flush_data *data, struct si_ds_queue *queue,
                           uint64_t submission_id)
{
   memset(data, 0, sizeof(*data));

   data->queue = queue;
   data->submission_id = submission_id;

   u_trace_init(&data->trace, &queue->device->trace_context);
}

void si_ds_flush_data_fini(struct si_ds_flush_data *data)
{
   u_trace_fini(&data->trace);
}

#ifdef __cplusplus
}
#endif