TheRock/rocm-systems/projects/clr/rocclr/platform/activity.cpp

/*
 * Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
 *
 * SPDX-License-Identifier: MIT
 */

#include "platform/activity.hpp"
#include "platform/command.hpp"
#include "platform/commandqueue.hpp"
#include "platform/command_utils.hpp"

#include <atomic>

namespace amd::activity_prof {

decltype(report_activity) report_activity{nullptr};

// Reserved sentinel pointer value (0x1) used to signal roctracer that CLR commits
// to delivering an activity record for this operation. roctracer's TracerCallback
// distinguishes this from an IsEnabled query (nullptr) and a real record (valid ptr).
// See TracerCallback in roctracer.cpp, ACTIVITY_DOMAIN_HIP_OPS case.
static void* const kCommitRecordSentinel = reinterpret_cast<void*>(uintptr_t{1});

void CommitRecord(OpId operation_id) {
  auto function = report_activity.load(std::memory_order_acquire);
  if (function) function(ACTIVITY_DOMAIN_HIP_OPS, operation_id, kCommitRecordSentinel);
}

#if defined(__linux__)
__thread activity_correlation_id_t correlation_id __attribute__((tls_model("initial-exec"))) = 0;
#elif defined(_WIN32)
__declspec(thread) activity_correlation_id_t correlation_id = 0;
#endif  // defined(_WIN32)

static inline size_t linearSize(const amd::Coord3D& size3d) {
  size_t size = size3d[0];
  if (size3d[1] != 0) size *= size3d[1];
  if (size3d[2] != 0) size *= size3d[2];
  return size;
}

bool IsEnabled(OpId operation_id) {
  if (operation_id < OP_ID_NUMBER)
    if (auto report = report_activity.load(std::memory_order_acquire))
      return report(ACTIVITY_DOMAIN_HIP_OPS, operation_id, nullptr) == 0;
  return false;
}

void ReportActivity(const amd::Command& command) {
  assert(command.profilingInfo().enabled_ && "Profiling must be enabled for this command");
  activity_op_t operation_id = OperationId(command.type());
  if (operation_id >= OP_ID_NUMBER) {
    // hipEventRecord enqueues an EventMarker with command type 0 (internal marker) rather than
    // CL_COMMAND_MARKER, so OperationId() can't classify it. It still dispatches a real barrier
    // packet carrying a GPU timestamp (marker_ts_), so report it as a barrier to surface
    // event-record sync points on the GPU timeline. Generic internal markers (batch-flush) have
    // marker_ts_ == false and stay unreported.
    if (command.type() == 0 && command.profilingInfo().marker_ts_) {
      operation_id = OP_ID_BARRIER;
    } else {
      // This command does not translate into a profiler activity (dispatch, memcopy, etc...),
      // there is nothing to report to the profiler.
      return;
    }
  }

  auto function = report_activity.load(std::memory_order_acquire);
  if (!function) return;

  const auto* queue = command.queue();
  assert(queue != nullptr);
  activity_record_t record{
      ACTIVITY_DOMAIN_HIP_OPS,                  // activity domain
      command.type(),                           // activity kind
      operation_id,                             // operation id
      command.profilingInfo().correlation_id_,  // activity correlation id
      command.profilingInfo().start_,           // begin timestamp, ns
      command.profilingInfo().end_,             // end timestamp, ns
      {{
          static_cast<int>(queue->device().info().driverNodeId_),  // device id
          queue->vdev()->index()                                   // queue id
      }},
      {}  // copied data size for memcpy, or kernel name for dispatch
  };

  switch (command.type()) {
    case CL_COMMAND_NDRANGE_KERNEL:
      record.kernel_name =
          static_cast<const amd::NDRangeKernelCommand&>(command).kernel().name().c_str();
      break;
    case CL_COMMAND_READ_BUFFER:
    case CL_COMMAND_READ_BUFFER_RECT:
      record.bytes = linearSize(static_cast<const amd::ReadMemoryCommand&>(command).size());
      break;
    case CL_COMMAND_WRITE_BUFFER:
    case CL_COMMAND_WRITE_BUFFER_RECT:
      record.bytes = linearSize(static_cast<const amd::WriteMemoryCommand&>(command).size());
      break;
    case CL_COMMAND_COPY_BUFFER:
    case CL_COMMAND_COPY_BUFFER_RECT:
      record.bytes = linearSize(static_cast<const amd::CopyMemoryCommand&>(command).size());
      break;
    case CL_COMMAND_FILL_BUFFER:
      record.bytes = linearSize(static_cast<const amd::FillMemoryCommand&>(command).size());
      break;
    case ROCCLR_COMMAND_BATCH_COPY_BUFFER: {
      const auto& ops = static_cast<const amd::BatchCopyMemoryCommand&>(command).copyOps();
      size_t total = 0;
      for (const auto& op : ops) total += op.size;
      record.bytes = total;
      break;
    }
    case ROCCLR_COMMAND_BATCH_WRITE_BUFFER: {
      const std::vector<amd::BatchWriteMemoryOp>& ops =
          static_cast<const amd::BatchWriteMemoryCommand&>(command).WriteOps();
      size_t total = 0;
      for (const amd::BatchWriteMemoryOp& op : ops) {
        total += op.size;
      }
      record.bytes = total;
      break;
    }
    case ROCCLR_COMMAND_BATCH_READ_BUFFER: {
      const std::vector<amd::BatchReadMemoryOp>& ops =
          static_cast<const amd::BatchReadMemoryCommand&>(command).ReadOps();
      size_t total = 0;
      for (const amd::BatchReadMemoryOp& op : ops) {
        total += op.size;
      }
      record.bytes = total;
      break;
    }
    default:
      break;
  }

  if (command.type() == CL_COMMAND_TASK) {
    auto timestamps = static_cast<const amd::AccumulateCommand&>(command).getTimestamps();
    const auto& kernel_names =
        static_cast<const amd::AccumulateCommand&>(command).getKernelNames();
    // kernel_names and timestamps are both populated only when profiling is active
    // at dispatch time. Walk the shorter of the two as a safety bound.
    for (uint32_t i = 0; i < kernel_names.size() && i < timestamps.size(); i++) {
      auto& ts = timestamps[i];
      record.begin_ns = ts.start;
      record.end_ns = ts.end;
      record.kernel_name = kernel_names[i];
      // Use per-packet queue_index when available so each kernel is assigned
      // to the internal parallel stream it ran on, not the launch stream.
      if (ts.queue_index != UINT32_MAX) {
        record.queue_id = static_cast<uint64_t>(ts.queue_index);
      }
      function(ACTIVITY_DOMAIN_HIP_OPS, operation_id, &record);
    }
  } else {
    record.begin_ns = command.profilingInfo().start_;
    record.end_ns = command.profilingInfo().end_;
    function(ACTIVITY_DOMAIN_HIP_OPS, operation_id, &record);
  }
}


#define CASE_STRING(X, C)                                                                          \
  case X:                                                                                          \
    return #C

const char* getOclCommandKindString(cl_command_type commandType) {
  switch (commandType) {
    CASE_STRING(0, InternalMarker);
    CASE_STRING(CL_COMMAND_MARKER, Marker);
    CASE_STRING(CL_COMMAND_NDRANGE_KERNEL, KernelExecution);
    CASE_STRING(CL_COMMAND_READ_BUFFER, CopyDeviceToHost);
    CASE_STRING(CL_COMMAND_WRITE_BUFFER, CopyHostToDevice);
    CASE_STRING(CL_COMMAND_COPY_BUFFER, CopyDeviceToDevice);
    CASE_STRING(CL_COMMAND_READ_BUFFER_RECT, CopyDeviceToHost2D);
    CASE_STRING(CL_COMMAND_WRITE_BUFFER_RECT, CopyHostToDevice2D);
    CASE_STRING(CL_COMMAND_COPY_BUFFER_RECT, CopyDeviceToDevice2D);
    CASE_STRING(CL_COMMAND_FILL_BUFFER, FillBuffer);
    CASE_STRING(CL_COMMAND_TASK, Task);
    CASE_STRING(CL_COMMAND_NATIVE_KERNEL, NativeKernel);
    CASE_STRING(CL_COMMAND_READ_IMAGE, ReadImage);
    CASE_STRING(CL_COMMAND_WRITE_IMAGE, WriteImage);
    CASE_STRING(CL_COMMAND_COPY_IMAGE, CopyImage);
    CASE_STRING(CL_COMMAND_COPY_IMAGE_TO_BUFFER, CopyImageToBuffer);
    CASE_STRING(CL_COMMAND_COPY_BUFFER_TO_IMAGE, CopyBufferToImage);
    CASE_STRING(CL_COMMAND_MAP_BUFFER, MapBuffer);
    CASE_STRING(CL_COMMAND_MAP_IMAGE, MapImage);
    CASE_STRING(CL_COMMAND_UNMAP_MEM_OBJECT, UnmapMemObject);
    CASE_STRING(CL_COMMAND_ACQUIRE_GL_OBJECTS, AcquireGLObjects);
    CASE_STRING(CL_COMMAND_RELEASE_GL_OBJECTS, ReleaseGLObjects);
    CASE_STRING(CL_COMMAND_USER, User);
    CASE_STRING(CL_COMMAND_BARRIER, Barrier);
    CASE_STRING(CL_COMMAND_MIGRATE_MEM_OBJECTS, MigrateMemObjects);
    CASE_STRING(CL_COMMAND_FILL_IMAGE, FillImage);
    CASE_STRING(CL_COMMAND_SVM_FREE, SvmFree);
    CASE_STRING(CL_COMMAND_SVM_MEMCPY, SvmMemcpy);
    CASE_STRING(CL_COMMAND_SVM_MEMFILL, SvmMemFill);
    CASE_STRING(CL_COMMAND_SVM_MAP, SvmMap);
    CASE_STRING(CL_COMMAND_SVM_UNMAP, SvmUnmap);
    CASE_STRING(ROCCLR_COMMAND_STREAM_WAIT_VALUE, StreamWait);
    CASE_STRING(ROCCLR_COMMAND_STREAM_WRITE_VALUE, StreamWrite);
    CASE_STRING(ROCCLR_COMMAND_BATCH_STREAM, BatchStreamOp);
    CASE_STRING(ROCCLR_COMMAND_BATCH_COPY_BUFFER, BatchCopyBuffer);
    CASE_STRING(ROCCLR_COMMAND_BATCH_WRITE_BUFFER, BatchWriteBuffer);
    CASE_STRING(ROCCLR_COMMAND_BATCH_READ_BUFFER, BatchReadBuffer);
    default:
      break;
  };
  return "Unknown command kind";
};
}  // namespace amd::activity_prof