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

#include "device/devhostcall.hpp"
#include "device/rocm/rocdevice.hpp"
#include "device/rocm/rocvirtual.hpp"
#include "device/rocm/rockernel.hpp"
#include "utils/nontemporal.hpp"
#include "device/rocm/rocmemory.hpp"
#include "device/rocm/rocblit.hpp"
#include "device/rocm/roccounters.hpp"
#include "platform/activity.hpp"
#include "platform/kernel.hpp"
#include "platform/context.hpp"
#include "platform/command.hpp"
#include "platform/command_utils.hpp"
#include "platform/memory.hpp"
#include "platform/sampler.hpp"
#include "utils/debug.hpp"
#include "os/os.hpp"

#include <fstream>
#include <limits>

#ifdef _WIN32
#include "device/rocm/rocd3d10interop.hpp"
#include "device/rocm/rocd3d11interop.hpp"
#include "platform/interop_d3d10.hpp"
#include "platform/interop_d3d11.hpp"
#endif
#include <memory>
#include <string>
#include <thread>
#include <vector>
#include <atomic>
#include <cinttypes>
#include <cstdarg>
#include <mutex>


/**
 * HSA image object size in bytes (see HSA spec)
 */
#define HSA_IMAGE_OBJECT_SIZE 48

/**
 * HSA image object alignment in bytes (see HSA spec)
 */
#define HSA_IMAGE_OBJECT_ALIGNMENT 16

/**
 * HSA sampler object size in bytes (see HSA spec)
 */
#define HSA_SAMPLER_OBJECT_SIZE 32

/**
 * HSA sampler object alignment in bytes (see HSA spec)
 */
#define HSA_SAMPLER_OBJECT_ALIGNMENT 16

namespace amd::roc {
// (HSA_FENCE_SCOPE_AGENT << HSA_PACKET_HEADER_ACQUIRE_FENCE_SCOPE) invalidates I, K and L1
// (HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_RELEASE_FENCE_SCOPE) invalidates L1, L2 and flushes
// L2

static constexpr uint16_t kInvalidAql = (HSA_PACKET_TYPE_INVALID << HSA_PACKET_HEADER_TYPE);

static constexpr uint16_t kBarrierPacketHeader =
    (HSA_PACKET_TYPE_BARRIER_AND << HSA_PACKET_HEADER_TYPE) | (1 << HSA_PACKET_HEADER_BARRIER) |
    (HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_ACQUIRE_FENCE_SCOPE) |
    (HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_RELEASE_FENCE_SCOPE);

static constexpr uint16_t kNopPacketHeader =
    (HSA_PACKET_TYPE_BARRIER_AND << HSA_PACKET_HEADER_TYPE) | (1 << HSA_PACKET_HEADER_BARRIER) |
    (HSA_FENCE_SCOPE_NONE << HSA_PACKET_HEADER_ACQUIRE_FENCE_SCOPE) |
    (HSA_FENCE_SCOPE_NONE << HSA_PACKET_HEADER_RELEASE_FENCE_SCOPE);

static constexpr uint16_t kBarrierPacketAcquireHeader =
    (HSA_PACKET_TYPE_BARRIER_AND << HSA_PACKET_HEADER_TYPE) | (1 << HSA_PACKET_HEADER_BARRIER) |
    (HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_ACQUIRE_FENCE_SCOPE) |
    (HSA_FENCE_SCOPE_NONE << HSA_PACKET_HEADER_RELEASE_FENCE_SCOPE);

static constexpr uint16_t kBarrierPacketReleaseHeader =
    (HSA_PACKET_TYPE_BARRIER_AND << HSA_PACKET_HEADER_TYPE) | (1 << HSA_PACKET_HEADER_BARRIER) |
    (HSA_FENCE_SCOPE_NONE << HSA_PACKET_HEADER_ACQUIRE_FENCE_SCOPE) |
    (HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_RELEASE_FENCE_SCOPE);

static constexpr uint16_t kBarrierVendorPacketHeader =
    (HSA_PACKET_TYPE_VENDOR_SPECIFIC << HSA_PACKET_HEADER_TYPE) | (1 << HSA_PACKET_HEADER_BARRIER) |
    (HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_ACQUIRE_FENCE_SCOPE) |
    (HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_RELEASE_FENCE_SCOPE);

static constexpr uint16_t kBarrierVendorPacketNopScopeHeader =
    (HSA_PACKET_TYPE_VENDOR_SPECIFIC << HSA_PACKET_HEADER_TYPE) | (1 << HSA_PACKET_HEADER_BARRIER) |
    (HSA_FENCE_SCOPE_NONE << HSA_PACKET_HEADER_ACQUIRE_FENCE_SCOPE) |
    (HSA_FENCE_SCOPE_NONE << HSA_PACKET_HEADER_RELEASE_FENCE_SCOPE);

static constexpr hsa_barrier_and_packet_t kBarrierAcquirePacket = {
    kBarrierPacketAcquireHeader, 0, 0, {{0}}, 0, {0}};

static constexpr hsa_barrier_and_packet_t kBarrierReleasePacket = {
    kBarrierPacketReleaseHeader, 0, 0, {{0}}, 0, {0}};

namespace {

void CL_CALLBACK ReleasePinnedMemoryCallback(cl_event event, int32_t command_exec_status,
                                             void* user_data) {
  (void)event;
  (void)command_exec_status;
  std::vector<amd::Memory*>* pinned_memories =
      reinterpret_cast<std::vector<amd::Memory*>*>(user_data);
  for (amd::Memory* pinned_memory : *pinned_memories) {
    pinned_memory->release();
  }
  delete pinned_memories;
}

}  // namespace

double Timestamp::ticksToTime_ = 0;

static unsigned extractAqlBits(unsigned v, unsigned pos, unsigned width) {
  return (v >> pos) & ((1 << width) - 1);
};

// Build the " virtual_pipe_id=N," log prefix when queue-pipe distribution is on,
// otherwise an empty string. Writes into the caller-provided buffer so the
// returned pointer stays valid for the enclosing ClPrint call.
static inline const char* formatVirtualPipePrefix(char* buf, size_t bufLen, const Device& dev,
                                                  const hsa_queue_t* queue) {
  if (dev.settings().queue_pipe_dist_) {
    snprintf(buf, bufLen, " virtual_pipe_id=%zu,", queue->id % dev.NumHwPipes());
    return buf;
  }
  return "";
}

// AQL diagnostic lines are emitted either through the debug log (gated by log level)
// or unconditionally to stderr for hang analysis (VirtualGPU::AnalyzeAqlQueue). The
// only difference between the two is the output routine, so the formatting lives in
// one place and the sink selects fprintf vs ClPrint.
enum class AqlLogSink { kLog, kStderr };

// Format a printf-style string into a std::string.
static inline std::string aqlFormat(const char* format, ...) {
  va_list args;
  va_start(args, format);
  va_list args_copy;
  va_copy(args_copy, args);
  const int len = vsnprintf(nullptr, 0, format, args);
  va_end(args);
  std::string result;
  if (len > 0) {
    result.resize(static_cast<size_t>(len));
    vsnprintf(&result[0], static_cast<size_t>(len) + 1, format, args_copy);
  }
  va_end(args_copy);
  return result;
}

static inline void emitAqlLine(AqlLogSink sink, amd::LogLevel level, const std::string& msg) {
  if (sink == AqlLogSink::kStderr) {
    fprintf(stderr, "%s\n", msg.c_str());
  } else {
    ClPrint(level, amd::LOG_AQL, "%s", msg.c_str());
  }
}

static inline void logAqlDispatchPacket(const Device& dev, const hsa_queue_t* queue,
                                        uint16_t header, const hsa_kernel_dispatch_packet_t* pkt,
                                        uint64_t wptr, amd::CommandQueue::Priority priority,
                                        AqlLogSink sink = AqlLogSink::kLog) {
  char buf[32];
  const char* prefix = formatVirtualPipePrefix(buf, sizeof(buf), dev, queue);
  const uint64_t rptr = Hsa::queue_load_read_index_relaxed(queue);

  emitAqlLine(sink, amd::LOG_DETAIL_DEBUG, aqlFormat(
          "SWq=0x%zx, HWq=0x%zx, id=%d, priority=%d,%s Dispatch Header = "
          "0x%x (type=%d, barrier=%d, acquire=%d, release=%d), "
          "setup=%d, grid=[%u, %u, %u], workgroup=[%u, %u, %u], "
          "private_seg_size=%u, group_seg_size=%u, kernel_obj=0x%zx, "
          "kernarg_address=0x%zx, completion_signal=0x%zx, "
          "rptr=%" PRIu64 ", wptr=%" PRIu64,
          queue, queue->base_address, queue->id, static_cast<int>(priority), prefix, header,
          extractAqlBits(header, HSA_PACKET_HEADER_TYPE, HSA_PACKET_HEADER_WIDTH_TYPE),
          extractAqlBits(header, HSA_PACKET_HEADER_BARRIER, HSA_PACKET_HEADER_WIDTH_BARRIER),
          extractAqlBits(header, HSA_PACKET_HEADER_SCACQUIRE_FENCE_SCOPE,
                         HSA_PACKET_HEADER_WIDTH_SCACQUIRE_FENCE_SCOPE),
          extractAqlBits(header, HSA_PACKET_HEADER_SCRELEASE_FENCE_SCOPE,
                         HSA_PACKET_HEADER_WIDTH_SCRELEASE_FENCE_SCOPE),
          pkt->setup, pkt->grid_size_x, pkt->grid_size_y, pkt->grid_size_z,
          pkt->workgroup_size_x, pkt->workgroup_size_y, pkt->workgroup_size_z,
          pkt->private_segment_size, pkt->group_segment_size, pkt->kernel_object,
          pkt->kernarg_address, pkt->completion_signal.handle,
          rptr, wptr));
}

static inline void logAqlMetadataPacket(const Device& dev, const hsa_queue_t* queue,
                                        const hsa_amd_metadata_kernel_dispatch_packet_t* meta,
                                        uint64_t wptr, AqlLogSink sink = AqlLogSink::kLog) {
  if (meta == nullptr) {
    return;
  }
  const auto& desc = meta->kernel_descriptor;

  // Dump the launch descriptor as a raw dword blob (13 dwords / 52 bytes).
  constexpr size_t kNumLaunchDescriptorDwords =
      sizeof(meta->launch_descriptor) / sizeof(uint32_t);
  const uint32_t* launchDwords = reinterpret_cast<const uint32_t*>(&meta->launch_descriptor);
  char launchDescriptorBlob[kNumLaunchDescriptorDwords * 11 + 1];
  size_t offset = 0;
  for (size_t dword_idx = 0; dword_idx < kNumLaunchDescriptorDwords; ++dword_idx) {
    offset += snprintf(launchDescriptorBlob + offset, sizeof(launchDescriptorBlob) - offset,
                       "%s0x%08x", (dword_idx == 0) ? "" : " ", launchDwords[dword_idx]);
  }

  emitAqlLine(sink, amd::LOG_EXTRA_DEBUG, aqlFormat(
          "SWq=0x%zx, HWq=0x%zx, id=%d, Metadata Packet [wptr=%" PRIu64 "]: "
          "header=[0x%x, 0x%x, 0x%x, 0x%x], event_id=0x%x, "
          "kernel_code_entry_byte_offset=0x%llx, compute_pgm_rsrc1=0x%x, compute_pgm_rsrc2=0x%x, "
          "compute_pgm_rsrc3=0x%x, kernel_code_properties=0x%x, "
          "kernarg_preload=[length=%u, offset=%u], launch_descriptor=[%s]",
          queue, queue->base_address, queue->id, wptr,
          meta->header0, meta->header1, meta->header2, meta->header3, meta->event_id,
          static_cast<unsigned long long>(desc.kernel_code_entry_byte_offset),
          desc.compute_pgm_rsrc1, desc.compute_pgm_rsrc2, desc.compute_pgm_rsrc3,
          desc.kernel_code_properties, desc.kernarg_preload.length, desc.kernarg_preload.offset,
          launchDescriptorBlob));
}

static inline void logAqlDispatchPacketExtended(
    const Device& dev, const hsa_queue_t* queue, uint16_t header,
    const hsa_amd_ext_kernel_dispatch_packet_t* pkt, uint64_t wptr,
    amd::CommandQueue::Priority priority,
    const hsa_amd_metadata_kernel_dispatch_packet_t* meta = nullptr,
    AqlLogSink sink = AqlLogSink::kLog) {
  char buf[32];
  const char* prefix = formatVirtualPipePrefix(buf, sizeof(buf), dev, queue);
  const uint64_t rptr = Hsa::queue_load_read_index_relaxed(queue);

  emitAqlLine(sink, amd::LOG_DETAIL_DEBUG, aqlFormat(
          "SWq=0x%zx, HWq=0x%zx, id=%d, priority=%d,%s Dispatch Header = "
          "0x%x (type=%d, barrier=%d, acquire=%d, release=%d), "
          "setup=%d, cluster_count=[%u, %u, %u], cluster_size=[%u, %u, %u], "
          "workgroup=[%u, %u, %u], private_seg_size=%u, group_seg_size=%u, kernel_obj=0x%zx, "
          "kernarg_address=0x%zx, dep_signal=0x%zx, completion_signal=0x%zx, "
          "rptr=%" PRIu64 ", wptr=%" PRIu64,
          queue, queue->base_address, queue->id, static_cast<int>(priority), prefix, header,
          extractAqlBits(header, HSA_PACKET_HEADER_TYPE, HSA_PACKET_HEADER_WIDTH_TYPE),
          extractAqlBits(header, HSA_PACKET_HEADER_BARRIER, HSA_PACKET_HEADER_WIDTH_BARRIER),
          extractAqlBits(header, HSA_PACKET_HEADER_SCACQUIRE_FENCE_SCOPE,
                         HSA_PACKET_HEADER_WIDTH_SCACQUIRE_FENCE_SCOPE),
          extractAqlBits(header, HSA_PACKET_HEADER_SCRELEASE_FENCE_SCOPE,
                         HSA_PACKET_HEADER_WIDTH_SCRELEASE_FENCE_SCOPE),
          pkt->setup, pkt->cluster_count_x, pkt->cluster_count_y, pkt->cluster_count_z,
          pkt->cluster_size_x, pkt->cluster_size_y, pkt->cluster_size_z, pkt->workgroup_size_x,
          pkt->workgroup_size_y, pkt->workgroup_size_z, pkt->private_segment_size,
          pkt->group_segment_size, pkt->kernel_object, pkt->kernarg_address,
          pkt->dep_signal.handle, pkt->completion_signal.handle, rptr, wptr));

  logAqlMetadataPacket(dev, queue, meta, wptr, sink);
}

static inline void logAqlBarrierPacket(const Device& dev, const hsa_queue_t* queue,
                                       uint16_t header, const hsa_barrier_and_packet_t* pkt,
                                       uint64_t wptr, amd::CommandQueue::Priority priority,
                                       AqlLogSink sink = AqlLogSink::kLog) {
  char buf[32];
  const char* prefix = formatVirtualPipePrefix(buf, sizeof(buf), dev, queue);
  const uint64_t rptr = Hsa::queue_load_read_index_relaxed(queue);

  uint16_t pktType = extractAqlBits(header, HSA_PACKET_HEADER_TYPE, HSA_PACKET_HEADER_WIDTH_TYPE);
  const char* typeStr = (pktType == HSA_PACKET_TYPE_BARRIER_OR) ? "Barrier-OR" : "Barrier-AND";
  emitAqlLine(sink, amd::LOG_DETAIL_DEBUG, aqlFormat(
          "SWq=0x%zx, HWq=0x%zx, id=%d, priority=%d,%s %s Header = "
          "0x%x (type=%d, barrier=%d, acquire=%d, release=%d), "
          "dep_signal=[0x%zx, 0x%zx, 0x%zx, 0x%zx, 0x%zx], "
          "completion_signal=0x%zx, rptr=%" PRIu64 ", wptr=%" PRIu64,
          queue, queue->base_address, queue->id, static_cast<int>(priority), prefix, typeStr, header,
          pktType,
          extractAqlBits(header, HSA_PACKET_HEADER_BARRIER, HSA_PACKET_HEADER_WIDTH_BARRIER),
          extractAqlBits(header, HSA_PACKET_HEADER_SCACQUIRE_FENCE_SCOPE,
                         HSA_PACKET_HEADER_WIDTH_SCACQUIRE_FENCE_SCOPE),
          extractAqlBits(header, HSA_PACKET_HEADER_SCRELEASE_FENCE_SCOPE,
                         HSA_PACKET_HEADER_WIDTH_SCRELEASE_FENCE_SCOPE),
          pkt->dep_signal[0].handle, pkt->dep_signal[1].handle,
          pkt->dep_signal[2].handle, pkt->dep_signal[3].handle,
          pkt->dep_signal[4].handle,
          pkt->completion_signal.handle, rptr, wptr));
}

static inline void logAqlBarrierValuePacket(const Device& dev, const hsa_queue_t* queue,
                                            uint16_t header,
                                            const hsa_amd_barrier_value_packet_t* pkt,
                                            uint64_t wptr, amd::CommandQueue::Priority priority,
                                            AqlLogSink sink = AqlLogSink::kLog) {
  char buf[32];
  const char* prefix = formatVirtualPipePrefix(buf, sizeof(buf), dev, queue);
  const uint64_t rptr = Hsa::queue_load_read_index_relaxed(queue);

  emitAqlLine(sink, amd::LOG_DETAIL_DEBUG, aqlFormat(
          "SWq=0x%zx, HWq=0x%zx, id=%d, priority=%d,%s BarrierValue Header = 0x%x AmdFormat = 0x%x "
          "(type=%d, barrier=%d, acquire=%d, release=%d), "
          "signal=0x%zx, value=0x%llx, mask=0x%llx, cond=%s, "
          "completion_signal=0x%zx, rptr=%" PRIu64 ", wptr=%" PRIu64,
          queue, queue->base_address, queue->id, static_cast<int>(priority), prefix,
          header, HSA_AMD_PACKET_TYPE_BARRIER_VALUE,
          extractAqlBits(header, HSA_PACKET_HEADER_TYPE, HSA_PACKET_HEADER_WIDTH_TYPE),
          extractAqlBits(header, HSA_PACKET_HEADER_BARRIER, HSA_PACKET_HEADER_WIDTH_BARRIER),
          extractAqlBits(header, HSA_PACKET_HEADER_SCACQUIRE_FENCE_SCOPE,
                         HSA_PACKET_HEADER_WIDTH_SCACQUIRE_FENCE_SCOPE),
          extractAqlBits(header, HSA_PACKET_HEADER_SCRELEASE_FENCE_SCOPE,
                         HSA_PACKET_HEADER_WIDTH_SCRELEASE_FENCE_SCOPE),
          pkt->signal.handle, pkt->value, pkt->mask,
          pkt->cond == 0 ? "EQ" : pkt->cond == 1 ? "NE" : pkt->cond == 2 ? "LT" : "GTE",
          pkt->completion_signal.handle, rptr, wptr));
}

// ================================================================================================
void ProfilingSignal::CacheTimingData(hsa_agent_t gpu_device) {
  // Lock needed as async handler thread can also touch this structure
  std::scoped_lock lock(lock_);

  // Return if timing is already cached
  if (cached_timing_.valid_) {
    return;
  }

  // Wait for this signal to complete if not already done
  if (Hsa::signal_load_relaxed(signal_) > 0) {
    WaitForSignal(signal_);
  }

  // Extract timing and cache it
  if (IsSdmaEngine(engine_)) {
    hsa_amd_profiling_async_copy_time_t time = {};
    Hsa::profiling_get_async_copy_time(signal_, &time);
    cached_timing_.start_ = time.start;
    cached_timing_.end_ = time.end;
  } else {
    hsa_amd_profiling_dispatch_time_t time = {};
    Hsa::profiling_get_dispatch_time(gpu_device, signal_, &time);
    cached_timing_.start_ = time.start;
    cached_timing_.end_ = time.end;
  }
  cached_timing_.valid_ = true;
}

// ================================================================================================
// Process GPU timing for signals
// If single_signal is nullptr, processes all signals and clears the list
// If single_signal is provided, processes only that signal with merge enabled
void Timestamp::checkGpuTime(ProfilingSignal* single_signal) {
  std::scoped_lock s(lock_);

  // For single signal mode, validate it exists in the list
  if (single_signal != nullptr) {
    auto it = std::find(signals_.begin(), signals_.end(), single_signal);
    if (it == signals_.end()) {
      return;
    }
  }

  if (HwProfiling()) {
    uint64_t start = std::numeric_limits<uint64_t>::max();
    uint64_t end = 0;
    uint64_t sdmaStart = std::numeric_limits<uint64_t>::max();
    uint64_t sdmaEnd = 0;

    // Process either single signal or all signals
    auto process_signal = [&](ProfilingSignal* sig) {
      // Skip signals already processed
      if (sig->flags_.done_) {
        return;
      }

      // Ignore the wait if runtime processes API callback, because the signal value is bigger
      // than expected and the value reset will occur after API callback is done
      if (GetCallbackSignal().handle == 0 || GetBlocking() == false) {
        ExtractSignalTiming(sig, start, end, sdmaStart, sdmaEnd);
      }

      if (IsLogEnabled(amd::LOG_INFO, amd::LOG_TS) &&
          (command().GetBatchHead() == nullptr || command().profilingInfo().marker_ts_ ||
           command().type() == CL_COMMAND_TASK)) {
        uint64_t sig_start, sig_end;
        sig->GetCachedTiming(sig_start, sig_end);
        amd_signal_t* amdSignal = reinterpret_cast<amd_signal_t*>(sig->signal_.handle);

        ClPrint(amd::LOG_INFO, amd::LOG_TS,
                "Signal = (0x%lx), Translated start/end = %ld / %ld, Elapsed = %ld ns, "
                "ticks start/end = %ld / %ld, Ticks elapsed = %ld, Engine = %u",
                sig->signal_.handle, sig_start, sig_end, sig_end - sig_start,
                amdSignal->start_ts, amdSignal->end_ts, amdSignal->end_ts - amdSignal->start_ts,
                sig->engine_);
      }
    };

    if (single_signal != nullptr) {
      process_signal(single_signal);
      // Remove the signal from the list after extracting its timing.
      // This prevents a stale entry when ActiveSignal() recycles and re-adds
      // the same ProfilingSignal — without this, the recycled signal appears
      // twice in signals_ and the first (stale) entry re-reads wrong HW timestamps.
      auto it = std::find(signals_.begin(), signals_.end(), single_signal);
      if (it != signals_.end()) {
        signals_.erase(it);
      }
    } else {
      for (auto it : signals_) {
        process_signal(it);
      }
      signals_.clear();
    }

    // Aggregate timing based on command type
    if (end != 0 || sdmaEnd != 0) {
      uint64_t final_start, final_end;
      const auto cmd_type = command().type();

      if (cmd_type == CL_COMMAND_COPY_BUFFER || cmd_type == CL_COMMAND_READ_BUFFER ||
          cmd_type == CL_COMMAND_WRITE_BUFFER || cmd_type == CL_COMMAND_COPY_BUFFER_RECT ||
          cmd_type == CL_COMMAND_READ_BUFFER_RECT || cmd_type == CL_COMMAND_WRITE_BUFFER_RECT) {
        // Copy/Read/Write — prefer SDMA timing, fall back to compute
        final_start = ((sdmaEnd != 0) ? sdmaStart : start) * ticksToTime_;
        final_end = ((sdmaEnd != 0) ? sdmaEnd : end) * ticksToTime_;
      } else {
        // Batch copy and all other commands — min/max across all engines
        final_start = std::min(sdmaEnd != 0 ? sdmaStart : start,
                               end != 0 ? start : sdmaStart) * ticksToTime_;
        final_end = std::max(sdmaEnd, end) * ticksToTime_;
      }

      if (!accum_ena_) {
        start_ = final_start;
        accum_ena_ = true;
      } else {
        start_ = std::min(start_, final_start);
      }
      end_ = std::max(end_, final_end);
    }
  }
}

// ================================================================================================
// Extract timing from a single signal and update accumulators
void Timestamp::ExtractSignalTiming(ProfilingSignal* signal,
                                    uint64_t& start, uint64_t& end,
                                    uint64_t& sdmaStart, uint64_t& sdmaEnd) {
  // Ensure timing data is cached
  if (!signal->IsTimingCached()) {
    signal->CacheTimingData(gpu()->gpu_device());
  }

  // Get cached timing
  uint64_t sig_start, sig_end;
  signal->GetCachedTiming(sig_start, sig_end);

  // Lock signal for accessing engine_ and flags_
  std::scoped_lock sig_lock(signal->LockSignalOps());

  // Guard against invalid timestamps from GPU (e.g. HW didn't write start_ts/end_ts).
  // TranslateTime returns 0 for these cases. Skip accumulation to avoid corrupting timing.
  if (sig_start == 0 || sig_end == 0 || sig_end < sig_start) {
    ClPrint(amd::LOG_WARNING, amd::LOG_TS,
            "Invalid signal timing: start=%lu, end=%lu, signal=0x%lx",
            sig_start, sig_end, signal->signal_.handle);
    signal->flags_.done_ = true;
    return;
  }

  // Update appropriate accumulators based on engine type
  if (IsSdmaEngine(signal->engine_)) {
    sdmaStart = std::min(sig_start, sdmaStart);
    sdmaEnd = std::max(sig_end, sdmaEnd);
  } else {
    start = std::min(sig_start, start);
    end = std::max(sig_end, end);
  }

  // Handle AccumulateCommand timestamps (convert ticks to system time).
  // Pass signal->queue_index_ so ReportActivity can assign each kernel to
  // the internal parallel stream it actually ran on, not the launch stream.
  if ((command().type() == CL_COMMAND_TASK) && (signal->flags_.isPacketDispatch_ == true)) {
    static_cast<amd::AccumulateCommand&>(command()).addTimestamps(
        static_cast<uint64_t>(sig_start * ticksToTime_),
        static_cast<uint64_t>(sig_end * ticksToTime_),
        signal->queue_index_);
  }

  signal->flags_.done_ = true;
}

// True while this thread is executing inside HsaAmdSignalHandler (the ROCr
// async-events thread). Device::releaseQueue uses it to defer the blocking
// queue_destroy off this thread and avoid a self-deadlock.
namespace {
thread_local bool tls_in_async_handler = false;
struct AsyncHandlerScope {
  AsyncHandlerScope() { tls_in_async_handler = true; }
  ~AsyncHandlerScope() { tls_in_async_handler = false; }
};
}  // namespace

bool InAsyncSignalHandler() { return tls_in_async_handler; }

// ================================================================================================
bool HsaAmdSignalHandler(hsa_signal_value_t value, void* arg) {
  AsyncHandlerScope async_scope;
  Timestamp* ts = reinterpret_cast<Timestamp*>(arg);

  VirtualGPU* const gpu = ts->gpu();

  // Save callback signal
  hsa_signal_t callback_signal = ts->GetCallbackSignal();

  ClPrint(amd::LOG_INFO, amd::LOG_SIG, "Handler: value(%d), timestamp(%p), callback_signal(0x%lx)",
          static_cast<uint32_t>(value), arg, callback_signal.handle);

  bool isBlocking = ts->GetBlocking();

  // Update the batch, since signal is complete
  gpu->updateCommandsState(ts->command().GetBatchHead());

  // Opportunistically try to release the HW queue if it's now idle
  // This helps reclaim queues in async workloads without explicit sync
  gpu->ReleaseHwQueue();

  // Reset API callback signal. It will release AQL queue and start commands processing
  if (callback_signal.handle != 0 && isBlocking) {
    Hsa::signal_subtract_relaxed(callback_signal, 1);
  }

  // Return false, so the callback will not be called again for this signal
  gpu->QueuedAsyncHandlers()--;
  gpu->release();
  return false;
}

// ================================================================================================
bool VirtualGPU::MemoryDependency::create(size_t numMemObj) {
  if (numMemObj > 0) {
    // Allocate the array of memory objects for dependency tracking
    memObjectsInQueue_ = new MemoryState[numMemObj];
    if (nullptr == memObjectsInQueue_) {
      return false;
    }
    memset(memObjectsInQueue_, 0, sizeof(MemoryState) * numMemObj);
    maxMemObjectsInQueue_ = numMemObj;
  }

  return true;
}

// ================================================================================================
void VirtualGPU::MemoryDependency::validate(VirtualGPU& gpu, const Memory* memory, bool readOnly) {
  bool flushL1Cache = false;

  if (maxMemObjectsInQueue_ == 0) {
    // Sync AQL packets
    gpu.setAqlHeader(gpu.dispatchPacketHeader_);
    return;
  }

  uint64_t curStart = reinterpret_cast<uint64_t>(memory->getDeviceMemory());
  uint64_t curEnd = curStart + memory->size();

  // Loop through all memory objects in the queue and find dependency
  // @note don't include objects from the current kernel
  for (size_t j = 0; j < endMemObjectsInQueue_; ++j) {
    // Check if the queue already contains this mem object and
    // GPU operations aren't readonly
    uint64_t busyStart = memObjectsInQueue_[j].start_;
    uint64_t busyEnd = memObjectsInQueue_[j].end_;

    // Check if the start inside the busy region
    if ((((curStart >= busyStart) && (curStart < busyEnd)) ||
         // Check if the end inside the busy region
         ((curEnd > busyStart) && (curEnd <= busyEnd)) ||
         // Check if the start/end cover the busy region
         ((curStart <= busyStart) && (curEnd >= busyEnd))) &&
        // If the buys region was written or the current one is for write
        (!memObjectsInQueue_[j].readOnly_ || !readOnly)) {
      flushL1Cache = true;
      break;
    }
  }

  // Did we reach the limit?
  if (maxMemObjectsInQueue_ <= numMemObjectsInQueue_) {
    flushL1Cache = true;
  }

  if (flushL1Cache) {
    // Sync AQL packets
    gpu.setAqlHeader(gpu.dispatchPacketHeader_);

    // Clear memory dependency state
    const static bool All = true;
    clear(!All);
  }

  // Insert current memory object into the queue always,
  // since runtime calls flush before kernel execution and it has to keep
  // current kernel in tracking
  memObjectsInQueue_[numMemObjectsInQueue_].start_ = curStart;
  memObjectsInQueue_[numMemObjectsInQueue_].end_ = curEnd;
  memObjectsInQueue_[numMemObjectsInQueue_].readOnly_ = readOnly;
  numMemObjectsInQueue_++;
}

// ================================================================================================
void VirtualGPU::MemoryDependency::clear(bool all) {
  if (numMemObjectsInQueue_ > 0) {
    if (all) {
      endMemObjectsInQueue_ = numMemObjectsInQueue_;
    }

    // If the current launch didn't start from the beginning, then move the data
    if (0 != endMemObjectsInQueue_) {
      // Preserve all objects from the current kernel
      size_t i, j;
      for (i = 0, j = endMemObjectsInQueue_; j < numMemObjectsInQueue_; i++, j++) {
        memObjectsInQueue_[i].start_ = memObjectsInQueue_[j].start_;
        memObjectsInQueue_[i].end_ = memObjectsInQueue_[j].end_;
        memObjectsInQueue_[i].readOnly_ = memObjectsInQueue_[j].readOnly_;
      }
    } else if (numMemObjectsInQueue_ >= maxMemObjectsInQueue_) {
      // note: The array growth shouldn't occur under the normal conditions,
      // but in a case when SVM path sends the amount of SVM ptrs over
      // the max size of kernel arguments
      MemoryState* ptr = new MemoryState[maxMemObjectsInQueue_ << 1];
      if (nullptr == ptr) {
        numMemObjectsInQueue_ = 0;
        return;
      }
      maxMemObjectsInQueue_ <<= 1;
      memcpy(ptr, memObjectsInQueue_, sizeof(MemoryState) * numMemObjectsInQueue_);
      delete[] memObjectsInQueue_;
      memObjectsInQueue_ = ptr;
    }

    numMemObjectsInQueue_ -= endMemObjectsInQueue_;
    endMemObjectsInQueue_ = 0;
  }
}

// ================================================================================================
VirtualGPU::HwQueueTracker::~HwQueueTracker() {
  for (auto& signal : signal_list_) {
    CpuWaitForSignal(signal);
    signal->release();
  }
  // Destroy all extra signals. Note: these signals must be idle already
  while (signal_pool_.size() != 0) {
    signal_pool_.top()->release();
    signal_pool_.pop();
  }
  while (signal_pool_irq_.size() != 0) {
    signal_pool_irq_.top()->release();
    signal_pool_irq_.pop();
  }
}

// ================================================================================================
bool VirtualGPU::HwQueueTracker::CreateSignal(ProfilingSignal* signal, bool interrupt) const {
  const Settings& settings = gpu_.dev().settings();
  // Use interrupts when active wait is disabled to avoid extra polling on the CPU.
  interrupt |= !gpu_.dev().ActiveWait();
  // Check if the interrupt was requested for the signal
  if (interrupt && settings.system_scope_signal_) {
    if (HSA_STATUS_SUCCESS != Hsa::signal_create(0, 0, nullptr, &signal->signal_)) {
      return false;
    }
  } else {
    if (HSA_STATUS_SUCCESS !=
        Hsa::signal_create(0, 0, nullptr, HSA_AMD_SIGNAL_AMD_GPU_ONLY, &signal->signal_)) {
      return false;
    }
  }
  signal->flags_.interrupt_ = interrupt;
  return true;
}

// ================================================================================================
bool VirtualGPU::HwQueueTracker::Create() {
  const uint kSignalListSize = ROC_SIGNAL_POOL_SIZE;
  signal_list_.resize(kSignalListSize);
  for (uint i = 0; i < kSignalListSize; ++i) {
    std::unique_ptr<ProfilingSignal> signal(new ProfilingSignal());
    if ((signal == nullptr) || !CreateSignal(signal.get())) {
      return false;
    }
    signal_list_[i] = signal.release();
  }
  // Add extra signals with the interrupts for the callbacks
  if (gpu_.dev().ActiveWait()) {
    for (uint32_t i = 0; i < 5; ++i) {
      std::unique_ptr<ProfilingSignal> signal(new ProfilingSignal());
      constexpr bool kEnableInterrupt = true;
      if ((signal == nullptr) || !CreateSignal(signal.get(), kEnableInterrupt)) {
        return false;
      }
      signal_pool_irq_.push(signal.release());
    }
  }
  return true;
}

// ================================================================================================
hsa_signal_t VirtualGPU::HwQueueTracker::ActiveSignal(hsa_signal_value_t init_val, Timestamp* ts,
                                                      bool attach_signal) {
  amd::Command* cmd = gpu_.command();
  // If no signal is needed, decrement the refcount and clear the hw_event of current command
  if (!attach_signal) {
    if (nullptr != cmd) {
      if (cmd->HwEvent() != nullptr) {
        reinterpret_cast<ProfilingSignal*>(cmd->HwEvent())->release();
      }
      cmd->SetHwEvent(nullptr);
    }
    return hsa_signal_t{0};
  }

  bool new_signal = false;

  // Peep signal +2 ahead to see if its done
  auto temp_id = (current_id_ + 2) % signal_list_.size();

  // If GPU is still busy with processing, then add more signals to avoid more frequent stalls
  if (Hsa::signal_load_relaxed(signal_list_[temp_id]->signal_) > 0) {
    std::unique_ptr<ProfilingSignal> signal(new ProfilingSignal());
    if ((signal != nullptr) && CreateSignal(signal.get())) {
      // Find valid new index
      ++current_id_ %= signal_list_.size();
      // Insert the new signal into the current slot and ignore any wait
      signal_list_.insert(signal_list_.begin() + current_id_, signal.release());
      new_signal = true;
    }
  }

  // If it's the new signal, then the wait can be avoided.
  // That will allow to grow the list of signals without stalls
  if (!new_signal) {
    // Find valid index
    ++current_id_ %= signal_list_.size();
    // Make sure the previous operation on the current signal is done
    WaitCurrent();

    // Have to wait the next signal in the queue to avoid a race condition between
    // a GPU waiter(which may be not triggered yet) and CPU signal reset below
    WaitNext();
  }

  if (signal_list_[current_id_]->referenceCount() > 1) {
    // The signal was assigned to the global marker's event, hence runtime can't reuse it
    // and needs a new signal
    std::unique_ptr<ProfilingSignal> signal(new ProfilingSignal());

    // Ensure that signals of the same type are created with the same interrupt flag,
    // as the tracking list depends on this for reuse.
    if ((signal != nullptr) && CreateSignal(signal.get(), signal_list_[current_id_]->flags_.interrupt_)) {
      signal_list_[current_id_]->release();
      signal_list_[current_id_] = signal.release();
    } else {
      assert(!"ProfilingSignal reallocation failed! Marker has a conflict with signal reuse!");
    }
  }

  bool enqueHandler = false;
  if (ts != nullptr) {
    enqueHandler =
        (ts->command().Callback() != nullptr || ts->command().GetBatchHead() != nullptr) &&
        !ts->command().CpuWaitRequested();
  }
  bool use_irq = enqueHandler || (IS_WINDOWS && gpu_.ForceIrq());
  // Check if the signal doesn't match the requested one.
  // Note: runtime needs the interrupts for the callbacks and the marker events, but it can reuse
  // the non-interrupt signals for the regular dispatches
  if ((signal_list_[current_id_]->flags_.interrupt_ != use_irq) && gpu_.dev().ActiveWait()) {
    // Use different stacks if an interrupt is required or not.
    // @note: if runtime needs an interrupt, then the tracking list replaces the original signal
    // with the interrupt signal and saves the signal without interrupt, or vise versa
    auto& pool_get = (use_irq) ? signal_pool_irq_ : signal_pool_;
    auto& pool_save = (use_irq) ? signal_pool_ : signal_pool_irq_;

    // Check if a free signal in the pop stack isn't available
    if (pool_get.empty()) {
      std::unique_ptr<ProfilingSignal> signal(new ProfilingSignal());
      if ((signal != nullptr) && CreateSignal(signal.get(), use_irq)) {
        pool_get.push(signal.release());
      }
    }
    // Make sure a free signal exists and replace it in the current slot
    if (!pool_get.empty()) {
      pool_save.push(signal_list_[current_id_]);
      signal_list_[current_id_] = pool_get.top();
      pool_get.pop();
    }
  }
  ProfilingSignal* prof_signal = signal_list_[current_id_];
  // Reset the signal and return
  Hsa::signal_silent_store_relaxed(prof_signal->signal_, init_val);
  prof_signal->flags_.done_ = false;
  prof_signal->engine_ = engine_;
  prof_signal->flags_.isPacketDispatch_ = false;
  prof_signal->ResetCachedTiming();
  prof_signal->queue_index_ = gpu_.index();

  // Release any existing HwEvent before setting new one for the same command
  VirtualGPU::AttachHwEvent(cmd, prof_signal);

  if (ts != nullptr) {
    // Save HSA signal earlier to make sure the possible callback will have a valid
    // value for processing
    ts->retain();
    prof_signal->ts_ = ts;
    ts->AddProfilingSignal(prof_signal);
    // If an enqueue handler is requested, set up marker/callback handling that updates the batch
    // upon HSA signal completion
    if (enqueHandler) {
      uint32_t init_value = kInitSignalValueOne;
      // If API callback is enabled, then use a blocking signal for AQL queue.
      // HSA signal will be acquired in SW and released after HSA signal callback
      if (ts->command().Callback() != nullptr) {
        bool blocking = ts->command().Callback()->blocking_;
        ts->SetCallbackSignal(prof_signal->signal_, blocking);
        // Blocks AQL queue from further processing
        if (blocking) {
          Hsa::signal_add_relaxed(prof_signal->signal_, 1);
          init_value += 1;
        }
      }
      gpu_.QueuedAsyncHandlers()++;
      ts->gpu()->retain();
      hsa_status_t result = Hsa::signal_async_handler(
          prof_signal->signal_, HSA_SIGNAL_CONDITION_LT, init_value, &HsaAmdSignalHandler, ts);
      if (HSA_STATUS_SUCCESS != result) {
        gpu_.QueuedAsyncHandlers()--;
        ts->gpu()->release();
        LogError("hsa_amd_signal_async_handler() failed to set the handler!");
      } else {
        ClPrint(amd::LOG_INFO, amd::LOG_SIG,
                "Set Handler: handle(0x%lx), timestamp(%p), blocking CB=%d",
                prof_signal->signal_.handle, prof_signal,
                ts->command().Callback() != nullptr && ts->GetBlocking());
      }
    }
  }
  return prof_signal->signal_;
}

// ================================================================================================
std::vector<hsa_signal_t>& VirtualGPU::HwQueueTracker::WaitingSignal(HwQueueEngine engine) {
  bool explicit_wait = false;
  // Reset all current waiting signals
  waiting_signals_.clear();

  // Does runtime switch the active engine?
  if (engine != engine_) {
    // Yes, return the signal from the previous operation for a wait
    engine_ = engine;
    explicit_wait = true;
  } else {
    // Unknown engine in use, hence return a wait signal always
    if (engine == HwQueueEngine::Unknown) {
      explicit_wait = true;
    } else {
      // Check if skip wait optimization is enabled. It will try to predict the same engine in ROCr
      // and ignore the signal wait, relying on in-order engine execution
      const Settings& settings = gpu_.dev().settings();
      if (engine != HwQueueEngine::Compute) {
        explicit_wait = true;
      }
    }
  }

  // Check if a wait is required
  if (explicit_wait) {
    bool skip_internal_signal = false;

    for (uint32_t i = 0; i < external_signals_.size(); ++i) {
      // If external signal matches internal one, then skip it
      if (external_signals_[i]->signal_.handle == signal_list_[current_id_]->signal_.handle) {
        skip_internal_signal = true;
      }
    }
    // Add the oldest signal into the tracking for a wait
    if (!skip_internal_signal) {
      external_signals_.push_back(signal_list_[current_id_]);
    }
  }

  // Validate all signals for the wait and skip already completed
  for (uint32_t i = 0; i < external_signals_.size(); ++i) {
    // Early signal status check
    if (Hsa::signal_load_relaxed(external_signals_[i]->signal_) > 0) {
      const Settings& settings = gpu_.dev().settings();
      if (settings.cpu_wait_for_signal_) {
        // Wait on CPU for completion if requested
        CpuWaitForSignal(external_signals_[i]);
      } else {
        // Add HSA signal for tracking on GPU
        waiting_signals_.push_back(external_signals_[i]->signal_);
      }
    }
  }
  external_signals_.clear();

  // Append raw signals added via AddDynamicQueueWait (e.g. IPC dep signals)
  for (auto& s : dynamic_queue_waits_) {
    waiting_signals_.push_back(s);
  }
  dynamic_queue_waits_.clear();

  // Return the array of waiting HSA signals
  return waiting_signals_;
}

// ================================================================================================
bool VirtualGPU::HwQueueTracker::CpuWaitForSignal(ProfilingSignal* signal) {
  // Wait for the current signal to complete
  if (Hsa::signal_load_relaxed(signal->signal_) > 0) {
    ClPrint(amd::LOG_DEBUG, amd::LOG_COPY, "Host wait on completion_signal=0x%zx",
            signal->signal_.handle);
    if (!WaitForSignal(signal->signal_, gpu_.ActiveWait())) {
      LogPrintfError("Failed signal [0x%lx] wait", signal->signal_);
      return false;
    }
  }

  // Process this signal's timing before signal reuse
  // This copies timing to the Timestamp
  if (signal->ts_ != nullptr) {
    signal->ts_->checkGpuTime(signal);
    signal->ts_->release();
    signal->ts_ = nullptr;
  } else {
    // No timestamp - just mark signal as done
    std::scoped_lock lock(signal->LockSignalOps());
    signal->flags_.done_ = true;
  }

  return true;
}

// ================================================================================================
bool VirtualGPU::HwQueueTracker::WaitCurrent() {
  ProfilingSignal* signal = signal_list_[current_id_];
  return CpuWaitForSignal(signal);
}

// ================================================================================================
void VirtualGPU::HwQueueTracker::WaitNext() {
  size_t next = (current_id_ + 1) % signal_list_.size();
  ProfilingSignal* signal = signal_list_[next];
  // Only wait, there is no need to save timestamp for the next signal
  // It will be saved when the signal is actually used
  WaitForSignal(signal->signal_, gpu_.ActiveWait());
}

// ================================================================================================
void VirtualGPU::HwQueueTracker::ResetCurrentSignal() {
  // Reset the signal and return
  Hsa::signal_silent_store_relaxed(signal_list_[current_id_]->signal_, 0);
  // Fallback to the previous signal
  current_id_ = (current_id_ == 0) ? (signal_list_.size() - 1) : (current_id_ - 1);
}

// ================================================================================================
bool VirtualGPU::processMemObjects(const amd::Kernel& kernel, const_address params,
                                   size_t& ldsAddress, bool cooperativeGroups,
                                   bool& imageBufferWrtBack,
                                   std::vector<device::Memory*>& wrtBackImageBuffer) {
  Kernel& hsaKernel =
      const_cast<Kernel&>(static_cast<const Kernel&>(*(kernel.getDeviceKernel(dev()))));
  const amd::KernelSignature& signature = kernel.signature();
  const amd::KernelParameters& kernelParams = kernel.parameters();

  if (!cooperativeGroups && memoryDependency().maxMemObjectsInQueue() != 0) {
    // AQL packets
    setAqlHeader(dispatchPacketHeaderNoSync_);
  }

  amd::Memory* const* memories =
      reinterpret_cast<amd::Memory* const*>(params + kernelParams.memoryObjOffset());

  // HIP shouldn't use cache coherency layer at any time
  if (!amd::IS_HIP) {
    // Process cache coherency first, since the extra transfers may affect
    // other mem dependency tracking logic: TS and signalWrite()
    for (uint i = 0; i < signature.numMemories(); ++i) {
      amd::Memory* mem = memories[i];
      if (mem != nullptr) {
        roc::Memory* gpuMem = dev().getGpuMemory(mem);
        // Don't sync for internal objects, since they are not shared between devices
        if (gpuMem->owner()->getVirtualDevice() == nullptr) {
          // Synchronize data with other memory instances if necessary
          gpuMem->syncCacheFromHost(*this);
        }
      }
    }
  }

  // Mark the tracker with a new kernel, so it can avoid checks of the aliased objects
  memoryDependency().newKernel();

  bool deviceSupportFGS = 0 != dev().isFineGrainedSystem(true);
  bool supportFineGrainedSystem = deviceSupportFGS;
  FGSStatus status = kernelParams.getSvmSystemPointersSupport();
  switch (status) {
    case FGS_YES:
      if (!deviceSupportFGS) {
        return false;
      }
      supportFineGrainedSystem = true;
      break;
    case FGS_NO:
      supportFineGrainedSystem = false;
      break;
    case FGS_DEFAULT:
    default:
      break;
  }

  size_t count = kernelParams.getNumberOfSvmPtr();
  size_t execInfoOffset = kernelParams.getTotalSize();
  bool sync = true;

  amd::Memory* memory = nullptr;
  // get svm non arugment information
  void* const* svmPtrArray = reinterpret_cast<void* const*>(params + execInfoOffset);
  for (size_t i = 0; i < count; i++) {
    memory = amd::MemObjMap::FindMemObj(svmPtrArray[i]);
    if (nullptr == memory) {
      if (!supportFineGrainedSystem) {
        return false;
      } else if (sync) {
        // Sync AQL packets
        setAqlHeader(dispatchPacketHeader_);
        // Clear memory dependency state
        const static bool All = true;
        memoryDependency().clear(!All);
        continue;
      }
    } else {
      Memory* rocMemory = static_cast<Memory*>(memory->getDeviceMemory(dev()));
      if (nullptr != rocMemory) {
        // Synchronize data with other memory instances if necessary
        rocMemory->syncCacheFromHost(*this);

        const static bool IsReadOnly = false;
        // Validate SVM passed in the non argument list
        memoryDependency().validate(*this, rocMemory, IsReadOnly);
      } else {
        return false;
      }
    }
  }

  // Check all parameters for the current kernel
  for (size_t i = 0; i < signature.numParameters(); ++i) {
    const amd::KernelParameterDescriptor& desc = signature.at(i);
    Memory* gpuMem = nullptr;
    amd::Memory* mem = nullptr;

    // Find if current argument is a buffer
    if (desc.type_ == T_POINTER) {
      if (desc.addressQualifier_ == CL_KERNEL_ARG_ADDRESS_LOCAL) {
        // Align the LDS on the alignment requirement of type pointed to
        ldsAddress = amd::alignUp(ldsAddress, desc.info_.arrayIndex_);
        if (desc.size_ == 8) {
          // Save the original LDS size
          uint64_t ldsSize = *reinterpret_cast<const uint64_t*>(params + desc.offset_);
          // Patch the LDS address in the original arguments with an LDS address(offset)
          WriteAqlArgAt(const_cast<address>(params), ldsAddress, desc.size_, desc.offset_);
          // Add the original size
          ldsAddress += ldsSize;
        } else {
          // Save the original LDS size
          uint32_t ldsSize = *reinterpret_cast<const uint32_t*>(params + desc.offset_);
          // Patch the LDS address in the original arguments with an LDS address(offset)
          uint32_t ldsAddr = ldsAddress;
          WriteAqlArgAt(const_cast<address>(params), ldsAddr, desc.size_, desc.offset_);
          // Add the original size
          ldsAddress += ldsSize;
        }
      } else {
        uint32_t index = desc.info_.arrayIndex_;
        mem = memories[index];
        const void* globalAddress = *reinterpret_cast<const void* const*>(params + desc.offset_);
        if (mem == nullptr) {
          ClPrint(amd::LOG_DEBUG, amd::LOG_KERN, "Arg%d: %s %s = ptr:%p ", i, desc.typeName_.c_str(),
                  desc.name_.c_str(), globalAddress);
          //! This condition is for SVM fine-grain
          if (dev().isFineGrainedSystem(true)) {
            // Sync AQL packets
            setAqlHeader(dispatchPacketHeader_);
            // Clear memory dependency state
            const static bool All = true;
            memoryDependency().clear(!All);
          }
        } else {
          gpuMem = static_cast<Memory*>(mem->getDeviceMemory(dev()));

          const void* globalAddress = *reinterpret_cast<const void* const*>(params + desc.offset_);
          ClPrint(amd::LOG_DEBUG, amd::LOG_KERN, "Arg%d: %s %s = ptr:%p obj:[%p-%p]", i,
                  desc.typeName_.c_str(), desc.name_.c_str(), globalAddress,
                  gpuMem->getDeviceMemory(),
                  reinterpret_cast<address>(gpuMem->getDeviceMemory()) + mem->getSize());

          // Validate memory for a dependency in the queue
          memoryDependency().validate(*this, gpuMem, (desc.info_.readOnly_ == 1));

          assert((desc.addressQualifier_ == CL_KERNEL_ARG_ADDRESS_GLOBAL ||
                  desc.addressQualifier_ == CL_KERNEL_ARG_ADDRESS_CONSTANT) &&
                 "Unsupported address qualifier");

          const bool readOnly = (desc.typeQualifier_ == CL_KERNEL_ARG_TYPE_CONST) ||
                                ((mem->getMemFlags() & CL_MEM_READ_ONLY) != 0);

          if (!readOnly) {
            mem->signalWrite(&dev());
          }

          if (desc.info_.oclObject_ == amd::KernelParameterDescriptor::ImageObject) {
            Image* image = static_cast<Image*>(mem->getDeviceMemory(dev()));

            const uint64_t image_srd = image->getHsaImageObject().handle;
            assert(amd::isMultipleOf(image_srd, sizeof(image_srd)));
            WriteAqlArgAt(const_cast<address>(params), image_srd, sizeof(image_srd), desc.offset_);

            // Check if synchronization has to be performed
            if (image->CopyImageBuffer() != nullptr) {
              Memory* devBuf = dev().getGpuMemory(mem->parent());
              amd::Coord3D offs(0);
              Image* devCpImg = static_cast<Image*>(dev().getGpuMemory(image->CopyImageBuffer()));
              amd::Image* img = mem->asImage();

              // Copy memory from the original image buffer into the backing store image
              bool result =
                  blitMgr().copyBufferToImage(*devBuf, *devCpImg, offs, offs, img->getRegion(),
                                              true, img->getRowPitch(), img->getSlicePitch());
              // Make sure the copy operation is done
              setAqlHeader(dispatchPacketHeader_);
              // Use backing store SRD as the replacment
              const uint64_t srd = devCpImg->getHsaImageObject().handle;
              WriteAqlArgAt(const_cast<address>(params), srd, sizeof(srd), desc.offset_);

              // If it's not a read only resource, then runtime has to write back
              if (!desc.info_.readOnly_) {
                wrtBackImageBuffer.push_back(mem->getDeviceMemory(dev()));
                imageBufferWrtBack = true;
              }
            }
          }
        }
      }
    } else if (desc.type_ == T_QUEUE) {
      uint32_t index = desc.info_.arrayIndex_;
      const amd::DeviceQueue* queue =
          reinterpret_cast<amd::DeviceQueue* const*>(params + kernelParams.queueObjOffset())[index];

      if (!createVirtualQueue(queue->size()) || !createSchedulerParam()) {
        return false;
      }
      uint64_t vqVA = getVQVirtualAddress();
      WriteAqlArgAt(const_cast<address>(params), vqVA, sizeof(vqVA), desc.offset_);
    } else if (desc.type_ == T_VOID) {
      const_address srcArgPtr = params + desc.offset_;
      if (desc.info_.oclObject_ == amd::KernelParameterDescriptor::ReferenceObject) {
        void* mem = allocKernArg(desc.size_, 128);
        memcpy(mem, srcArgPtr, desc.size_);
        const auto it = hsaKernel.patch().find(desc.offset_);
        WriteAqlArgAt(const_cast<address>(params), mem, sizeof(void*), it->second);
      }

      if (IsLogEnabled(amd::LOG_INFO, amd::LOG_KERN)) {
        if (desc.size_ > 8) {
          std::string bytes = "0x";
          constexpr size_t kMaxBytes = 64;
          for (size_t j = 0; j < std::min(desc.size_, kMaxBytes); j++) {
            char byteStr[4];
            snprintf(byteStr, sizeof(byteStr), "%02x ",
                     reinterpret_cast<const uint8_t*>(srcArgPtr)[j]);
            bytes += byteStr;
          }
          if (desc.size_ > kMaxBytes) {
            bytes += "...";
          }
          ClPrint(amd::LOG_DEBUG, amd::LOG_KERN, "Arg%d: %s %s = %s (size:0x%x)", i,
                  desc.typeName_.c_str(), desc.name_.c_str(), bytes.c_str(), desc.size_);
        } else {
          ClPrint(amd::LOG_DEBUG, amd::LOG_KERN, "Arg%d: %s %s = val:0x%lx (size:0x%x)", i,
                  desc.typeName_.c_str(), desc.name_.c_str(),
                  (desc.size_ == 1)   ? *reinterpret_cast<const uint8_t*>(srcArgPtr)
                  : (desc.size_ == 2) ? *reinterpret_cast<const uint16_t*>(srcArgPtr)
                  : (desc.size_ == 4) ? *reinterpret_cast<const uint32_t*>(srcArgPtr)
                  : (desc.size_ == 8) ? *reinterpret_cast<const uint64_t*>(srcArgPtr)
                                      : 0LL,
                  desc.size_);
        }
      }
    } else if (desc.type_ == T_SAMPLER) {
      uint32_t index = desc.info_.arrayIndex_;
      const amd::Sampler* sampler =
          reinterpret_cast<amd::Sampler* const*>(params + kernelParams.samplerObjOffset())[index];

      device::Sampler* devSampler = sampler->getDeviceSampler(dev());

      uint64_t sampler_srd = devSampler->hwSrd();
      WriteAqlArgAt(const_cast<address>(params), sampler_srd, sizeof(sampler_srd), desc.offset_);
    }
  }

  if (hsaKernel.program()->hasGlobalStores()) {
    // Sync AQL packets
    setAqlHeader(dispatchPacketHeader_);
    // Clear memory dependency state
    const static bool All = true;
    memoryDependency().clear(!All);
  }

  return true;
}

// ================================================================================================
void VirtualGPU::SetGpuQueue(hsa_queue_t* queue) {
  gpu_queue_ = queue;

  cached_read_dispatch_id_ = 0;
  // The cached queue-progress state belongs to the previously assigned HW queue. When the HW
  // queue changes (released back to / reacquired from the shared dynamic-queue pool), that state
  // is stale: the write indices refer to a different queue. Reset to the "empty queue" sentinel
  // so IsQueueIdle() reports idle for the freshly assigned queue.
  last_write_index_ = kInvalidQueueIndex;
  last_packet_with_signal_index_ = kInvalidQueueIndex;

  if (queue) {
    auto extras = roc_device_.GetQueueExtras(queue);
    device_mem_ring_buf_ = extras.deviceMemRingBuf;
    use_movdir64b_ = roc_device_.info().movdir64b_ && device_mem_ring_buf_;
    doorbell_ptr_ = extras.doorbellPtr;
    metadata_preloader_.SetQueueBase(extras.metadataRingBuffer,
                                     roc_device_.MetadataVersionHeader());
  } else {
    device_mem_ring_buf_ = false;
    use_movdir64b_ = false;
    doorbell_ptr_ = nullptr;
    metadata_preloader_.SetQueueBase(nullptr, roc_device_.MetadataVersionHeader());
  }
}

// ================================================================================================
void VirtualGPU::AcquireQueueWithPreference() {
  std::scoped_lock lock(execution());
  if (!dedicated_queue_ && gpu_queue_ == nullptr && last_hwq_ != nullptr) {
    SetGpuQueue(roc_device_.AcquireActiveQueue(priority_, last_hwq_));
    last_hwq_ = nullptr;
  }
}

// ================================================================================================
bool VirtualGPU::ReacquireQueueExcluding(const std::unordered_set<uint64_t>& excluded_ids) {
  std::scoped_lock lock(execution());
  if (gpu_queue_ != nullptr) {
    // Detach from the current queue: decrements refCount in the pool but never
    // destroys the queue.
    // Unlike ReleaseActiveQueue, this is unconditional — we are switching queues,
    // not conditionally reclaiming under pressure.
    roc_device_.releaseQueue(gpu_queue_, std::vector<uint32_t>{}, false, true);
    gpu_queue_ = nullptr;
  }
  SetGpuQueue(roc_device_.AcquireActiveQueue(priority_, nullptr, &excluded_ids));
  return gpu_queue_ != nullptr;
}

// ================================================================================================
uint64_t VirtualGPU::getQueueID() {
  std::scoped_lock lock(execution());
  // Dedicated queues keep their HW queue, never acquire from pool
  if (!dedicated_queue_ && gpu_queue_ == nullptr) {
    SetGpuQueue(roc_device_.AcquireActiveQueue(priority_));
  }
  return gpu_queue_->id;
}

// ================================================================================================
static inline void packet_store_release(uint32_t* packet, uint16_t header, uint16_t rest) {
#if IS_WINDOWS
  std::atomic_ref<uint32_t> atomic_header(*packet);
  atomic_header.store(header | (rest << 16), std::memory_order_release);
#else
  __atomic_store_n(packet, header | (rest << 16), __ATOMIC_RELEASE);
#endif
}

// ================================================================================================
std::string VirtualGPU::AnalyzeAqlQueue() const {
  std::string kernelName = "<not identified>";
  const uint32_t queueMask = gpu_queue_->size - 1;
  const uint64_t index = Hsa::queue_load_write_index_relaxed(gpu_queue_);
  const uint64_t read = Hsa::queue_load_read_index_relaxed(gpu_queue_);

  if (index <= read) {
    fprintf(stderr, "VGPU(%p) Queue(%p) is idle\n", this, gpu_queue_);
    return kernelName;
  }

  // Skip the drained (INVALID) packets to find the stuck one.
  int valid_packet_idx = 0;
  constexpr int kAqlSearchWindow = 32;
  while (valid_packet_idx < kAqlSearchWindow) {
    auto* aql_loc = &(reinterpret_cast<hsa_kernel_dispatch_packet_t*>(
        gpu_queue_->base_address))[(read + valid_packet_idx) & queueMask];
    if (extractAqlBits(aql_loc->header, HSA_PACKET_HEADER_TYPE, HSA_PACKET_HEADER_WIDTH_TYPE) ==
        HSA_PACKET_TYPE_INVALID) {
      valid_packet_idx++;
    } else {
      break;
    }
  }
  if (valid_packet_idx == kAqlSearchWindow) {
    fprintf(stderr, "VGPU(%p) Queue(%p). Couldn't find the hang AQL packet!\n", this, gpu_queue_);
    return kernelName;
  }

  const uint64_t slot = (read + valid_packet_idx) & queueMask;
  auto* aql_loc =
      &(reinterpret_cast<hsa_kernel_dispatch_packet_t*>(gpu_queue_->base_address))[slot];
  const hsa_kernel_dispatch_packet_t packet = *aql_loc;
  const uint16_t header = packet.header;
  const auto pkt_type = extractAqlBits(header, HSA_PACKET_HEADER_TYPE, HSA_PACKET_HEADER_WIDTH_TYPE);
  const auto* meta = metadata_preloader_.GetMetadataPacket(slot);

  auto lookupKernelName = [&](uint64_t kernel_object) {
    auto it = dev().KernelMap().find(kernel_object);
    if (it != dev().KernelMap().end()) {
      kernelName = it->second.getDemangledName();
    } else {
      fprintf(stderr, "VGPU(%p) Queue(%p). Couldn't find kernel\n", this, gpu_queue_);
    }
  };

  // Reuse the shared AQL packet formatters with the stderr sink so hang analysis
  // and the debug log print identical packet detail, including the metadata blob.
  fprintf(stderr, "VGPU(%p) hang analysis:\n", this);
  if (pkt_type == HSA_PACKET_TYPE_VENDOR_SPECIFIC) {
    auto* vendor_hdr = reinterpret_cast<const hsa_amd_vendor_packet_header_t*>(aql_loc);
    if (vendor_hdr->AmdFormat == HSA_AMD_PACKET_TYPE_EXT_KERNEL_DISPATCH) {
      auto* ext = reinterpret_cast<const hsa_amd_ext_kernel_dispatch_packet_t*>(aql_loc);
      lookupKernelName(ext->kernel_object);
      logAqlDispatchPacketExtended(dev(), gpu_queue_, header, ext, index, priority_, meta,
                                   AqlLogSink::kStderr);
    } else {
      fprintf(stderr, "VGPU(%p) Queue(%p) rptr=%" PRIu64 ", wptr=%" PRIu64
             ". Vendor packet (amd_format=%d)\n",
             this, gpu_queue_, read, index, static_cast<int>(vendor_hdr->AmdFormat));
    }
  } else if (pkt_type == HSA_PACKET_TYPE_KERNEL_DISPATCH ||
             (index == read && packet.kernel_object != 0)) {
    lookupKernelName(packet.kernel_object);
    logAqlDispatchPacket(dev(), gpu_queue_, header, &packet, index, priority_, AqlLogSink::kStderr);
    logAqlMetadataPacket(dev(), gpu_queue_, meta, index, AqlLogSink::kStderr);
  } else {
    logAqlBarrierPacket(dev(), gpu_queue_, header,
                        reinterpret_cast<const hsa_barrier_and_packet_t*>(aql_loc), index,
                        priority_, AqlLogSink::kStderr);
  }
  return kernelName;
}

// ================================================================================================
// Write an AQL packet to the ring buffer with metadata prefetch.
//
// MOVDIR64B + device memory path (2 sfences):
//   metadata MOVDIR64B → sfence → AQL MOVDIR64B → ROCr doorbell sfence → doorbell
//
// Legacy + device memory (2 sfences):
//   AQL body NT → metadata 4x64B NT (staged) → sfence → AQL header → sfence → doorbell
//   (body+header within each 64B NT copy share one WC buffer fill — no internal sfence)
//
// Legacy + system memory (3 sfences):
//   AQL body NT → metadata direct writes → sfence(body→hdr) → metadata headers →
//   sfence → AQL header → sfence → doorbell
template <typename AqlPacket>
void VirtualGPU::writePacketToRingBuffer(AqlPacket* aql_loc, AqlPacket* packet,
                                         uint16_t header, uint16_t rest, uint64_t slot_index) {
  if (use_movdir64b_) {
    metadata_preloader_.SetMetadata(packet, header, slot_index, true, device_mem_ring_buf_);
    amd::nontemporalStoreFence();
    amd::nontemporalWriteAQL(aql_loc, packet, header, rest);
  } else {
    amd::nontemporalCopyAQL(aql_loc, packet);
    metadata_preloader_.SetMetadata(packet, header, slot_index, false, device_mem_ring_buf_);
    amd::nontemporalStoreFence();
    packet_store_release(reinterpret_cast<uint32_t*>(aql_loc), header, rest);
  }
}

// ================================================================================================
void VirtualGPU::ringQueueDoorbell(uint64_t index) {
  if (doorbell_ptr_) {
    amd::ringDoorbell(doorbell_ptr_, index, use_movdir64b_);
  } else {
    Hsa::signal_store_screlease(gpu_queue_->doorbell_signal, index);
  }
}

// ================================================================================================
template <typename AqlPacket>
bool VirtualGPU::dispatchGenericAqlPacket(AqlPacket* packet, uint16_t header, uint16_t rest,
                                          bool blocking, bool attach_signal, bool cluster_launch) {
  const uint32_t queueSize = gpu_queue_->size;
  const uint32_t queueMask = queueSize - 1;
  const uint32_t sw_queue_size = queueMask;

  // Check for queue full and wait if needed.
  uint64_t index = Hsa::queue_add_write_index_screlease(gpu_queue_, 1);
  adjustHeader(header);

  bool attachSignal = timestamp_ != nullptr || attach_signal;
  // Get active signal for current dispatch if profiling is necessary
  packet->completion_signal =
      Barriers().ActiveSignal(kInitSignalValueOne, timestamp_, attachSignal);

  if (timestamp_ != nullptr) {
    // If profiling is enabled, store the correlation ID in the dispatch packet. The profiler
    // can retrieve this correlation ID to attribute waves to specific dispatch locations.
    // reserved2 exists only on the standard (non-ext) dispatch packet.
    if constexpr (std::is_same_v<AqlPacket, hsa_kernel_dispatch_packet_t>) {
      if (amd::activity_prof::IsEnabled(OP_ID_DISPATCH) &&
          !dev().settings().ext_dispatch_packet_) {
        packet->reserved2 = timestamp_->command().profilingInfo().correlation_id_;
      }
    }

    ProfilingSignal* current_signal = Barriers().GetLastSignal();
    current_signal->flags_.isPacketDispatch_ = true;
  }

  // Make sure the slot is free for usage
  WaitForQueueSlot(index, sw_queue_size);

  // Add blocking command if the original value of read index was behind of the queue size.
  // Note: direct dispatch relies on the slot stall above to keep the forward progress
  // of the app if a dispatched kernel requires some CPU input for completion
  if (blocking) {
    if (packet->completion_signal.handle == 0) {
      packet->completion_signal = Barriers().ActiveSignal();
    }
    blocking = true;
  }

  TrackQueueProgress(*packet, index);

  AqlPacket* aql_loc = &((AqlPacket*)(gpu_queue_->base_address))[index & queueMask];
  writePacketToRingBuffer(aql_loc, packet, header, rest, index & queueMask);

  if (IsLogEnabled(amd::LOG_DETAIL_DEBUG, amd::LOG_AQL)) {
    if (dev().settings().ext_dispatch_packet_) {
      logAqlDispatchPacketExtended(
          roc_device_, gpu_queue_, header,
          reinterpret_cast<hsa_amd_ext_kernel_dispatch_packet_t*>(packet), index, priority_,
          metadata_preloader_.GetMetadataPacket(index & queueMask));
    } else {
      logAqlDispatchPacket(roc_device_, gpu_queue_, header,
                           reinterpret_cast<hsa_kernel_dispatch_packet_t*>(packet), index,
                           priority_);
    }
  }
  // Optimization for native AQL path in Windows has problems with PM4 emulation,
  // skipping the doorbell will not wake up the AQL worker thread
  uint32_t skip_limit = DEBUG_CLR_DOORBELL_SKIP;
  bool ring_for_non_profiler_signal = attach_signal && (packet->completion_signal.handle != 0);
  bool ring_doorbell = IS_LINUX || dev().IsPm4Emulation() || blocking ||
                       (skippedDispatches_ >= skip_limit) || ring_for_non_profiler_signal;
  if (ring_doorbell) {
    ringQueueDoorbell(index);
    skippedDispatches_ = 0;
  } else {
    ++skippedDispatches_;
  }

  // Mark the flag indicating if a dispatch is outstanding.
  // We are not waiting after every dispatch.
  hasPendingDispatch_ = true;

  // Wait on signal ?
  if (blocking) {
    LogInfo("Runtime reached the AQL queue limit. SW is much ahead of HW. Blocking AQL queue!");
    if (!Barriers().WaitCurrent()) {
      LogPrintfError("Failed blocking queue wait with signal [0x%lx]",
                     packet->completion_signal.handle);
      return false;
    }
  }

  return true;
}

// ================================================================================================
void VirtualGPU::adjustHeader(uint16_t& header) {
  setFenceDirty(true);

  if (addSystemScope_) {
    header &= ~(HSA_FENCE_SCOPE_AGENT << HSA_PACKET_HEADER_SCACQUIRE_FENCE_SCOPE |
                HSA_FENCE_SCOPE_AGENT << HSA_PACKET_HEADER_SCRELEASE_FENCE_SCOPE);
    header |= (HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_SCACQUIRE_FENCE_SCOPE |
               HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_SCRELEASE_FENCE_SCOPE);
    addSystemScope_ = false;
  }

  auto expected_fence_state = extractAqlBits(header, HSA_PACKET_HEADER_SCRELEASE_FENCE_SCOPE,
                                             HSA_PACKET_HEADER_WIDTH_SCRELEASE_FENCE_SCOPE);

  if (expected_fence_state == amd::Device::kCacheStateSystem) {
    setFenceDirty(false);
  }

  if (fence_state_ == amd::Device::kCacheStateSystem &&
      expected_fence_state == amd::Device::kCacheStateSystem) {
    header = dispatchPacketHeader_;
    setFenceDirty(true);
  }

  fence_state_ = static_cast<Device::CacheState>(expected_fence_state);
}

// ================================================================================================
void VirtualGPU::dispatchBlockingWait(hsa_kernel_dispatch_packet_t* packet) {
  auto wait_signals = Barriers().WaitingSignal();
  if (dev().settings().ext_dispatch_packet_ && wait_signals.size() == 1 && packet != nullptr) {
      // The Ext Dispatch Packet supports only one dependent signal
      auto ext_packet = reinterpret_cast<hsa_amd_ext_kernel_dispatch_packet_t*>(packet);
      ext_packet->dep_signal = wait_signals[0];
  } else {
    // AQL dispatch doesn't support dependent signals and extra barrier packet must be generated
    for (uint32_t i = 0; i < wait_signals.size(); ++i) {
      uint32_t j = i % 5;
      barrier_packet_.dep_signal[j] = wait_signals[i];
      constexpr bool kSkipSignal = true;
      // If runtime reached the packet limit or the count limit, then flush the barrier
      if ((j == 4) || ((i + 1) == wait_signals.size())) {
        dispatchBarrierPacket(kNopPacketHeader, kSkipSignal);
      }
    }
  }
}

// ================================================================================================
template <typename AqlPacket>
bool VirtualGPU::dispatchAqlPacket(AqlPacket* packet, uint16_t header,
                                   uint16_t rest, bool blocking, bool attach_signal) {
  if (command_ != nullptr && command_->getPktCapturingState()) {
    adjustHeader(header);
    packet->header = header;
    if constexpr (std::is_same_v<AqlPacket, hsa_amd_ext_kernel_dispatch_packet_t>) {
      // For ext dispatch packets the first 32 bits are {header(16), amd_format(8), setup(8)}.
      // rest already encodes (amd_format | (setup << 8)).
      packet->amd_format = static_cast<hsa_amd_packet_type8_t>(rest & 0xFF);
      packet->setup = static_cast<uint8_t>((rest >> 8) & 0xFF);
    } else {
      packet->setup = rest;
    }
    std::memcpy(const_cast<uint8_t*>(command_->getAqlPacket()), packet, sizeof(AqlPacket));

    if (metadata_preloader_.HasMetadataQueue()) {
      uint8_t* metaBuf = command_->getMetadataPacket();
      if (metaBuf != nullptr) {
        metadata_preloader_.CaptureMetadata(packet, header, metaBuf);
      }
    }
    return true;
  } else {
    dispatchBlockingWait(reinterpret_cast<hsa_kernel_dispatch_packet_t*>(packet));
    return dispatchGenericAqlPacket(packet, header, rest, blocking, attach_signal);
  }
}
// ================================================================================================
bool VirtualGPU::dispatchAqlPacket(hsa_barrier_and_packet_t* packet, uint16_t header, uint16_t rest,
                                   bool blocking, bool attach_signal) {
  return dispatchGenericAqlPacket(packet, header, rest, blocking, attach_signal);
}

// ================================================================================================
// Publish one metadata-prefetch packet (256B = 4 x 64B segments) to the metadata ring at |dst|.
// |src| is the captured metadata packet with valid headers already baked into each segment
// (INVALID for barrier/empty slots).  Each 64B segment carries its own header dword together
// with its body, so the header/body ordering within a segment is guaranteed by the single 64B
// store and no intra-packet fence is needed.  This mirrors the single-dispatch staged path
// (MetaDataPreloader::SetPacket): MOVDIR64B when available, otherwise non-temporal stores.
static inline void writeMetadataPacketToRing(uint8_t* dst, const uint8_t* src,
                                             bool use_movdir64b) {
  static constexpr size_t kMetaSize = 256;
  for (size_t off = 0; off < kMetaSize; off += 64) {
    if (use_movdir64b) {
      amd::movdir64b_copy64(dst + off, src + off);
    } else {
      amd::nontemporalMemcpy(dst + off, src + off, 64);
    }
  }
}

// ================================================================================================
// Unified flat-buffer graph dispatch.  Reserves all N queue slots with a single wptr bump,
// then processes them in kPeriod-sized chunks: yield-until-free → memcpy → per-packet fixups
// (profiling signals, correlation IDs, kernel-name printing) → valid-header writes → doorbell.
// For graphs that fit in the queue the yield never fires.  For oversized graphs the GPU drains
// earlier chunks while the CPU copies later ones, avoiding a deadlock.  Packet-0's header is
// committed last in the first chunk per the AQL protocol.
//
// Two ring-write paths, selected by |use_movdir64b| (device-memory ring + HW support +
// DEBUG_CLR_USE_MOVDIR64B):
//   * Non-temporal (default/baseline): bulk NT copy of packet bodies (headers held INVALID in
//     the flat data), one sfence, then scattered 4-byte valid-header release stores.
//   * MOVDIR64B: one sfence, then publish each packet as a single atomic 64B write (body +
//     signal + valid header).  Reserved slots already carry INVALID headers from the CP's
//     previous consumption, so publish order is irrelevant and no packet-0-last dance is
//     needed.  The metadata ring uses the same NT-vs-MOVDIR64B split.
bool VirtualGPU::dispatchAqlPacketBatchFlat(const amd::AlignedVector64<uint8_t>& flatPacketData,
                                            const std::vector<uint32_t>& validFullHeaders,
                                            amd::AccumulateCommand* vcmd, bool attach_signal,
                                            bool pre_patched, bool blocking,
                                            const std::vector<uint8_t>* flatMetadataData) {
  // Both base and ext kernel dispatch packets place kernel_object at the same offset.
  static_assert(offsetof(hsa_kernel_dispatch_packet_t, kernel_object) ==
                    offsetof(hsa_amd_ext_kernel_dispatch_packet_t, kernel_object),
                "kernel_object offset mismatch between base and ext dispatch packets");
  if (vcmd == nullptr || flatPacketData.empty() || validFullHeaders.empty()) {
    return false;
  }

  const size_t numPackets = validFullHeaders.size();
  if (flatPacketData.size() != numPackets * 64) {
    return false;
  }

  static constexpr size_t kMetaPktSize =
      sizeof(hsa_amd_metadata_kernel_dispatch_packet_t);
  if (flatMetadataData != nullptr && !flatMetadataData->empty()) {
    if (flatMetadataData->size() != numPackets * kMetaPktSize) {
      return false;
    }
  }

  std::scoped_lock lock(execution());
  profilingBegin(*vcmd);
  dispatchBlockingWait(nullptr);

  const uint32_t queueSize = gpu_queue_->size;
  const uint32_t queueMask = queueSize - 1;
  const uint32_t sw_queue_size = queueMask;
  static constexpr size_t kPacketSize = sizeof(hsa_kernel_dispatch_packet_t);

  // Unpack first/last headers; apply system-scope once for the whole batch.
  // validFullHeaders stores the AQL full_header dword: low 16 = header, high 16 = setup.
  uint16_t firstHeader = static_cast<uint16_t>(validFullHeaders[0]);
  uint16_t firstSetup  = static_cast<uint16_t>(validFullHeaders[0] >> 16);
  uint16_t lastHeader  = static_cast<uint16_t>(validFullHeaders[numPackets - 1]);
  if (addSystemScope_) {
    firstHeader &= ~(HSA_FENCE_SCOPE_AGENT << HSA_PACKET_HEADER_SCACQUIRE_FENCE_SCOPE);
    firstHeader |= (HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_SCACQUIRE_FENCE_SCOPE);
    lastHeader &= ~(HSA_FENCE_SCOPE_AGENT << HSA_PACKET_HEADER_SCRELEASE_FENCE_SCOPE);
    lastHeader |= (HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_SCRELEASE_FENCE_SCOPE);
    // For a single-packet batch, firstHeader and lastHeader describe the same
    // packet, but only firstHeader is committed by packet_store_release().  Put
    // the system release fence on firstHeader too.
    if (numPackets == 1) {
      firstHeader &= ~(HSA_FENCE_SCOPE_AGENT << HSA_PACKET_HEADER_SCRELEASE_FENCE_SCOPE);
      firstHeader |= (HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_SCRELEASE_FENCE_SCOPE);
    }
    addSystemScope_ = false;
  }

  uint8_t* queueBase = static_cast<uint8_t*>(gpu_queue_->base_address);

  const bool use_movdir64b = use_movdir64b_;
  // Reserve ALL slots with a single wptr bump, then submit in kPeriod-sized chunks.
  // Per-chunk: yield if the queue is full (handles graphs larger than the queue), then
  // memcpy + per-packet fixups + headers + doorbell.  For graphs that fit in the queue
  // the yield never fires.
  uint64_t startIndex = Hsa::queue_add_write_index_screlease(gpu_queue_, numPackets);
  setFenceDirty(true);

  // Update cached fence state from the last packet's release scope.
  // Clear fence dirty if the last packet has system-scope release, matching
  // the single-dispatch path in dispatchGenericAqlPacket (set dirty on reserve,
  // then conditionally clear if system scope).
  auto expected_fence_state =
      extractAqlBits(lastHeader, HSA_PACKET_HEADER_SCRELEASE_FENCE_SCOPE,
                     HSA_PACKET_HEADER_WIDTH_SCRELEASE_FENCE_SCOPE);
  if (expected_fence_state == amd::Device::kCacheStateSystem) {
    setFenceDirty(false);
  }
  fence_state_ = static_cast<Device::CacheState>(expected_fence_state);

  const size_t kPeriod = DEBUG_HIP_GRAPH_BATCH_SIZE;
  // Ramp-up: submit a small first (lead) chunk so the doorbell is rung after
  // copying only kLead packets instead of a full kPeriod. The GPU starts
  // executing the lead while the CPU keeps copying the remaining
  // packet+metadata payload, overlapping the bulk copy with GPU execution and
  // cutting the one-time startup idle that grew with the 256B/packet metadata.
  constexpr size_t kLead = 8;
  auto* first_loc = reinterpret_cast<uint32_t*>(
      queueBase + (startIndex & queueMask) * kPacketSize);

  // Attach profiling / completion signals to one packet.  Used by the MOVDIR64B path,
  // which assembles the full packet (body + signal + valid header) in a host staging
  // buffer before the atomic 64B store, so signals must be written into |pkt| (staging)
  // rather than patched into the ring slot after the body copy (the NT path does the
  // latter inline below).  |isLast| marks the final packet of the whole batch.
  auto attachPacketSignal = [&](hsa_kernel_dispatch_packet_t* pkt, size_t i, bool isLast) {
    const uint16_t hdr = static_cast<uint16_t>(validFullHeaders[i]);
    const uint8_t pktType =
        extractAqlBits(hdr, HSA_PACKET_HEADER_TYPE, HSA_PACKET_HEADER_WIDTH_TYPE);
    const uint8_t amdFormat = static_cast<uint8_t>((validFullHeaders[i] >> 16) & 0xFF);
    const bool isBaseKernelDispatch = (pktType == HSA_PACKET_TYPE_KERNEL_DISPATCH);
    const bool isKernelDispatch =
        isBaseKernelDispatch ||
        (pktType == HSA_PACKET_TYPE_VENDOR_SPECIFIC &&
         amdFormat == HSA_AMD_PACKET_TYPE_EXT_KERNEL_DISPATCH);
    if (timestamp_ != nullptr) {
      // When pre_patched, keep any completion_signal already written by
      // ApplyHwEventPatches (carried into staging via the flat-buffer copy).
      bool has_prepatched_signal = pre_patched && (pkt->completion_signal.handle != 0);
      if (!has_prepatched_signal) {
        pkt->completion_signal =
            Barriers().ActiveSignal(kInitSignalValueOne, timestamp_, true);
        if (isKernelDispatch) {
          if (isBaseKernelDispatch && amd::activity_prof::IsEnabled(OP_ID_DISPATCH)) {
            pkt->reserved2 = timestamp_->command().profilingInfo().correlation_id_;
          }
          Barriers().GetLastSignal()->flags_.isPacketDispatch_ = true;
        }
      } else if (has_prepatched_signal && isBaseKernelDispatch &&
                 amd::activity_prof::IsEnabled(OP_ID_DISPATCH)) {
        pkt->reserved2 = timestamp_->command().profilingInfo().correlation_id_;
      }
    } else if (isLast && (attach_signal || blocking)) {
      pkt->completion_signal = Barriers().ActiveSignal();
    }
  };

  // Kernel-name collection is required when dispatch activity tracing is on; detailed
  // packet logging when LOG_KERN2 / LOG_AQL is on.  Both are handled by logBatchPacket.
  const bool needKernelNamesReported = amd::activity_prof::IsEnabled(OP_ID_DISPATCH);
  const bool kLogBatch = IsLogEnabled(amd::LOG_DETAIL_DEBUG, amd::LOG_KERN2) ||
                         IsLogEnabled(amd::LOG_DETAIL_DEBUG, amd::LOG_AQL);

  // All packet fields are read from the host-side flat buffer (|flatPacketData|), never
  // from the ring slot: the device-resident ring is write-combining, so reading
  // kernel_object back from it is unreliable.  Kernel names are resolved from the device
  // KernelMap by kernel_object and, when activity tracing is on, recorded into the command
  // via addKernelName().  getDemangledName() returns a reference to a name cached for the
  // device's lifetime, so the borrowed pointer stays valid.
  auto logBatchPacket = [&](size_t i, uint64_t slotIdx) {
    const auto* hostPkt = reinterpret_cast<const hsa_kernel_dispatch_packet_t*>(
        flatPacketData.data() + i * kPacketSize);
    const uint16_t hdr = static_cast<uint16_t>(validFullHeaders[i]);
    const uint8_t pktType =
        extractAqlBits(hdr, HSA_PACKET_HEADER_TYPE, HSA_PACKET_HEADER_WIDTH_TYPE);
    const uint8_t amdFormat = static_cast<uint8_t>((validFullHeaders[i] >> 16) & 0xFF);
    const bool isBaseKernelDispatch = (pktType == HSA_PACKET_TYPE_KERNEL_DISPATCH);
    const bool isKernelDispatch =
        isBaseKernelDispatch ||
        (pktType == HSA_PACKET_TYPE_VENDOR_SPECIFIC &&
         amdFormat == HSA_AMD_PACKET_TYPE_EXT_KERNEL_DISPATCH);
    if (isKernelDispatch) {
      auto kit = dev().KernelMap().find(hostPkt->kernel_object);
      const char* kname = kit != dev().KernelMap().end()
                              ? kit->second.getDemangledName().c_str()
                              : "<unknown>";
      if (needKernelNamesReported) {
        vcmd->addKernelName(kname);
      }
      if (kLogBatch) {
        ClPrint(amd::LOG_DETAIL_DEBUG, amd::LOG_KERN2,
                "Graph ShaderName : %s, device id : %u", kname, dev().index());
        if (isBaseKernelDispatch) {
          logAqlDispatchPacket(roc_device_, gpu_queue_, hdr, hostPkt, slotIdx, priority_);
        } else {
          logAqlDispatchPacketExtended(
              roc_device_, gpu_queue_, hdr,
              reinterpret_cast<const hsa_amd_ext_kernel_dispatch_packet_t*>(hostPkt), slotIdx,
              priority_);
        }
      }
    } else if (kLogBatch && pktType == HSA_PACKET_TYPE_VENDOR_SPECIFIC &&
               amdFormat == HSA_AMD_PACKET_TYPE_BARRIER_VALUE) {
      ClPrint(amd::LOG_DETAIL_DEBUG, amd::LOG_KERN2,
              "Graph ShaderName : [Graph vendor barrier value], device id : %u", dev().index());
      logAqlBarrierValuePacket(
          roc_device_, gpu_queue_, hdr,
          reinterpret_cast<const hsa_amd_barrier_value_packet_t*>(hostPkt), slotIdx, priority_);
    } else if (kLogBatch &&
               (pktType == HSA_PACKET_TYPE_BARRIER_AND || pktType == HSA_PACKET_TYPE_BARRIER_OR)) {
      // Inline barriers placed in the batch by BuildSyncPlan never go through
      // dispatchBarrierPacket, so log them here. Classify by patched fields:
      // dep_signal set -> cross-dep barrier, else completion_signal -> per-segment barrier.
      const auto* bpkt = reinterpret_cast<const hsa_barrier_and_packet_t*>(hostPkt);
      bool has_dep = false;
      for (int k = 0; k < 5 && !has_dep; ++k) {
        if (bpkt->dep_signal[k].handle != 0) has_dep = true;
      }
      const char* tag = has_dep
          ? " [Graph cross dep barrier]"
          : (bpkt->completion_signal.handle != 0 ? " [Graph completion barrier]"
                                                 : " [Graph batch barrier]");
      ClPrint(amd::LOG_DETAIL_DEBUG, amd::LOG_KERN2, "Graph ShaderName :%s, device id : %u", tag,
              dev().index());
      logAqlBarrierPacket(roc_device_, gpu_queue_, hdr, bpkt, slotIdx, priority_);
    }
  };

  for (size_t chunkStart = 0; chunkStart < numPackets; ) {
    const size_t period = (chunkStart == 0) ? kLead : kPeriod;
    const size_t chunkEnd  = std::min(chunkStart + period, numPackets);
    const size_t thisChunk = chunkEnd - chunkStart;
    const bool isFirstChunk = (chunkStart == 0);
    const bool isLastChunk  = (chunkEnd == numPackets);

    // Yield until this chunk's physical slots are free.
    WaitForQueueSlot(startIndex + chunkEnd - 1, sw_queue_size);

    const size_t chunkSlot = (startIndex + chunkStart) & queueMask;

    // Publish Bodies with invalid headers for NT store only. MOVDIR64B may not need this
    if (!use_movdir64b) {
      // Non-temporal path: copy this chunk's packet bodies to the ring using NT stores.
      // Headers within the flat data are invalid; real headers are armed via
      // packet_store_release below, after the sfence.
      const uint8_t* srcData = flatPacketData.data() + chunkStart * kPacketSize;
      if (chunkSlot + thisChunk <= queueSize) {
        amd::nontemporalMemcpy(queueBase + chunkSlot * kPacketSize, srcData,
                               thisChunk * kPacketSize);
      } else {
        const size_t firstCount = queueSize - chunkSlot;
        amd::nontemporalMemcpy(queueBase + chunkSlot * kPacketSize, srcData,
                               firstCount * kPacketSize);
        amd::nontemporalMemcpy(queueBase, srcData + firstCount * kPacketSize,
                               (thisChunk - firstCount) * kPacketSize);
      }
    }

    // Publish this chunk's metadata-prefetch packets. The captured flat buffer already
    // carries the armed header value for real slots and INVALID for barrier/empty slots;
    // each 64B segment is written with its header and body together (NT or MOVDIR64B).
    if (flatMetadataData != nullptr && metadata_preloader_.HasMetadataQueue()) {
      auto* metaBase = static_cast<uint8_t*>(metadata_preloader_.GetQueueBase());
      for (size_t j = 0; j < thisChunk; ++j) {
        const uint64_t slot = (startIndex + chunkStart + j) & queueMask;
        writeMetadataPacketToRing(metaBase + slot * kMetaPktSize,
                                  flatMetadataData->data() + (chunkStart + j) * kMetaPktSize,
                                  use_movdir64b);
      }
    }

    // Core path to publish headers
    if (!use_movdir64b) {
      // Non-temporal publish: signals patched in place, then armed headers
      // Attach signal to the last packet when requested (before per-packet logging).
      auto* lastSlotPtr = reinterpret_cast<hsa_kernel_dispatch_packet_t*>(
          queueBase + ((startIndex + chunkEnd - 1) & queueMask) * kPacketSize);
      if (isLastChunk && (attach_signal || blocking) && timestamp_ == nullptr) {
        lastSlotPtr->completion_signal = Barriers().ActiveSignal();
      }

      // Per-packet fixups: profiling signals, kernel-name printing, inline barrier logging.
      if (timestamp_ != nullptr || needKernelNamesReported || kLogBatch) {
        for (size_t i = chunkStart; i < chunkEnd; ++i) {
          const uint64_t slotIdx = (startIndex + i) & queueMask;
          auto* slot = reinterpret_cast<hsa_kernel_dispatch_packet_t*>(
              queueBase + slotIdx * kPacketSize);
          if (timestamp_ != nullptr) {
            attachPacketSignal(slot, i, i == numPackets - 1);
          }
          if (needKernelNamesReported || kLogBatch) {
            logBatchPacket(i, slotIdx);
          }
        }
      }

      // Ordering point: drain all preceding NT body writes from the WC buffer
      // before writing any valid headers (diagram steps #6a, #7, #9).
      amd::nontemporalStoreFence();

      // Write valid headers. Hold global packet-0's header back in the first chunk so the
      // GPU sees a fully-committed batch before starting (AQL protocol). Subsequent chunks
      // write in forward order — packet 0 is already committed.
      for (size_t i = (isFirstChunk ? 1 : chunkStart); i < chunkEnd; ++i) {
        const uint64_t idx = startIndex + i;
        auto* aql_loc =
            reinterpret_cast<uint32_t*>(queueBase + (idx & queueMask) * kPacketSize);
        const uint32_t dword = validFullHeaders[i];
        const uint16_t hdr = (i == numPackets - 1) ? lastHeader : static_cast<uint16_t>(dword);
        packet_store_release(aql_loc, hdr, static_cast<uint16_t>(dword >> 16));
      }
      if (isFirstChunk) {
        packet_store_release(first_loc, firstHeader, firstSetup);
      }
    } else {
      // MOVDIR64B publish: full 64B atomic writes (body + signal + valid header).
      // sfence orders the preceding metadata MOVDIR64B writes (WC-ordered) so they
      // are globally visible before any dispatch packet becomes visible to the CP.
      // Every reserved slot already carries an INVALID header from the CP's previous
      // consumption (guaranteed by WaitForQueueSlot), so publish order is irrelevant —
      // the CP stalls on any not-yet-written slot — and no packet-0-last dance is needed.
      amd::nontemporalStoreFence();
      for (size_t i = chunkStart; i < chunkEnd; ++i) {
        const uint64_t slotIdx = (startIndex + i) & queueMask;
        alignas(64) hsa_kernel_dispatch_packet_t stg;
        std::memcpy(&stg, flatPacketData.data() + i * kPacketSize, kPacketSize);
        attachPacketSignal(&stg, i, i == numPackets - 1);
        const uint32_t dword = validFullHeaders[i];
        const uint16_t hdr = (i == 0)                 ? firstHeader
                             : (i == numPackets - 1)  ? lastHeader
                                                      : static_cast<uint16_t>(dword);
        const uint16_t setup = (i == 0) ? firstSetup : static_cast<uint16_t>(dword >> 16);
        *reinterpret_cast<uint32_t*>(&stg) = hdr | (static_cast<uint32_t>(setup) << 16);
        auto* dst = queueBase + slotIdx * kPacketSize;
        amd::movdir64b_copy64(dst, &stg);
        if (needKernelNamesReported || kLogBatch) {
          logBatchPacket(i, slotIdx);
        }
      }
    }

    if (doorbell_ptr_) {
      amd::ringDoorbell(doorbell_ptr_, startIndex + chunkEnd - 1);
    } else {
      Hsa::signal_store_screlease(gpu_queue_->doorbell_signal, startIndex + chunkEnd - 1);
    }

    chunkStart = chunkEnd;
  }

  hasPendingDispatch_ = true;

  auto* finalLastSlot = reinterpret_cast<hsa_kernel_dispatch_packet_t*>(
      queueBase + ((startIndex + numPackets - 1) & queueMask) * kPacketSize);

  // Skip the pending dispatch only when both conditions are met: a completion
  // signal tracks the last packet and the fence is already clean (system scope).
  if (finalLastSlot->completion_signal.handle != 0 && !isFenceDirty()) {
    hasPendingDispatch_ = false;
  }

  TrackQueueProgress(*finalLastSlot, startIndex + numPackets - 1, pre_patched);

  if (blocking) {
    LogInfo("Running serialized as blocking is requested");
    if (!Barriers().WaitCurrent()) {
      LogPrintfError("Failed blocking queue wait with signal [0x%lx]",
                     finalLastSlot->completion_signal.handle);
      profilingEnd();
      return false;
    }
  }

  profilingEnd();
  return true;
}

// ================================================================================================
bool VirtualGPU::dispatchCounterAqlPacket(hsa_ext_amd_aql_pm4_packet_t* packet,
                                          const uint32_t gfxVersion, bool blocking,
                                          const hsa_ven_amd_aqlprofile_1_00_pfn_t* extApi) {
  // PM4 IB packet submission is different between GFX8 and GFX9:
  //  In GFX8 the PM4 IB packet blob is writing directly to AQL queue
  //  In GFX9 the PM4 IB is submitting by AQL Vendor Specific packet and
  switch (gfxVersion) {
    case PerfCounter::ROC_GFX9:
    case PerfCounter::ROC_GFX10: {
      uint16_t hdr = HSA_PACKET_TYPE_VENDOR_SPECIFIC << HSA_PACKET_HEADER_TYPE;
      return dispatchGenericAqlPacket(packet, hdr, 0, blocking);
    } break;
  }

  return false;
}

// ================================================================================================
void VirtualGPU::WaitCompleteSignal(hsa_signal_t signal) {
  // Add the signal as a dynamic dependency so WaitingSignal() includes it
  // alongside any pending external signals in the barrier's dep_signal list.
  Barriers().AddDynamicQueueWait(signal);
  dispatchBarrierPacket(kBarrierPacketHeader, false);
}

// ================================================================================================
void VirtualGPU::dispatchBarrierPacket(uint16_t packetHeader, bool skipSignal,
                                       hsa_signal_t signal) {
  const uint32_t queueSize = gpu_queue_->size;
  const uint32_t queueMask = queueSize - 1;

  if (!skipSignal) {
    // Make sure the wait is issued before queue index reservation
    auto wait_signals = Barriers().WaitingSignal();
    for (uint32_t i = 0; i < wait_signals.size(); ++i) {
      uint32_t j = i % 5;
      barrier_packet_.dep_signal[j] = wait_signals[i];
      constexpr bool kSkipSignal = true;
      // If runtime reached the packet limit and signals left, then flush the barrier
      if ((j == 4) && ((i + 1) < wait_signals.size())) {
        dispatchBarrierPacket(kNopPacketHeader, kSkipSignal);
      }
    }
  }

  uint64_t index = Hsa::queue_add_write_index_screlease(gpu_queue_, 1);

  setFenceDirty(true);
  auto cache_state = extractAqlBits(packetHeader, HSA_PACKET_HEADER_SCRELEASE_FENCE_SCOPE,
                                    HSA_PACKET_HEADER_WIDTH_SCRELEASE_FENCE_SCOPE);
  // Tracker-owned only when the signal comes from Barriers().ActiveSignal() below; an externally
  // provided signal (managed-buffer pool, IPC) is a different object, so it is not usable for idle
  // detection and must be treated like skip_signal.
  const bool external_signal = skipSignal || (signal.handle != 0);
  if (!external_signal) {
    // Get active signal for current dispatch if profiling is necessary
    barrier_packet_.completion_signal = Barriers().ActiveSignal(kInitSignalValueOne, timestamp_);
  } else {
    // Attach external signal to the packet
    barrier_packet_.completion_signal = signal;
  }

  TrackQueueProgress(barrier_packet_, index, external_signal);

  // Reset fence_dirty_ flag if we submit a barrier with system scopes
  if (cache_state == amd::Device::kCacheStateSystem) {
    setFenceDirty(false);
  }

  WaitForQueueSlot(index, queueMask);
  hsa_barrier_and_packet_t* aql_loc =
      &(reinterpret_cast<hsa_barrier_and_packet_t*>(gpu_queue_->base_address))[index & queueMask];

  writePacketToRingBuffer(aql_loc, &barrier_packet_, packetHeader, uint16_t{0}, index & queueMask);
  ringQueueDoorbell(index);
  if (IsLogEnabled(amd::LOG_DETAIL_DEBUG, amd::LOG_AQL)) {
    logAqlBarrierPacket(roc_device_, gpu_queue_, packetHeader, &barrier_packet_, index, priority_);
  }

  // Clear dependent signals for the next packet
  barrier_packet_.dep_signal[0] = hsa_signal_t{};
  barrier_packet_.dep_signal[1] = hsa_signal_t{};
  barrier_packet_.dep_signal[2] = hsa_signal_t{};
  barrier_packet_.dep_signal[3] = hsa_signal_t{};
  barrier_packet_.dep_signal[4] = hsa_signal_t{};
}

// ================================================================================================
void VirtualGPU::dispatchBarrierValuePacket(uint16_t packetHeader, bool resolveDepSignal,
                                            hsa_signal_t signal, hsa_signal_value_t value,
                                            hsa_signal_value_t mask, hsa_signal_condition32_t cond,
                                            bool skipTs, hsa_signal_t completionSignal) {
  uint16_t rest = HSA_AMD_PACKET_TYPE_BARRIER_VALUE;
  const uint32_t queueSize = gpu_queue_->size;
  const uint32_t queueMask = queueSize - 1;

  barrier_value_packet_.signal = signal;
  barrier_value_packet_.value = value;
  barrier_value_packet_.mask = mask;
  barrier_value_packet_.cond = cond;

  // Dependent signal and external signal cant be true at the same time
  assert((resolveDepSignal && (signal.handle != 0)) == false);
  if (resolveDepSignal) {
    auto wait_signals = Barriers().WaitingSignal();
    if (wait_signals.size() > 0) {
      barrier_value_packet_.signal = wait_signals[0];
      barrier_value_packet_.value = kInitSignalValueOne;
      barrier_value_packet_.mask = std::numeric_limits<int64_t>::max();
      barrier_value_packet_.cond = HSA_SIGNAL_CONDITION_LT;
      for (uint32_t i = 1; i < wait_signals.size(); ++i) {
        uint32_t j = (i - 1) % 5;
        barrier_packet_.dep_signal[j] = wait_signals[i];
        constexpr bool kSkipSignal = true;
        // If runtime reached the packet limit or the count limit, then flush the barrier
        if ((j == 4) || ((i + 1) == wait_signals.size())) {
          dispatchBarrierPacket(kNopPacketHeader, kSkipSignal);
        }
      }
    }
  }

  setFenceDirty(true);
  auto cache_state = extractAqlBits(packetHeader, HSA_PACKET_HEADER_SCRELEASE_FENCE_SCOPE,
                                    HSA_PACKET_HEADER_WIDTH_SCRELEASE_FENCE_SCOPE);

  // Tracker-owned only when the signal comes from Barriers().ActiveSignal() below; an externally
  // provided completion signal is a different object, so it is not usable for idle detection and
  // must be treated like skip_signal.
  const bool external_signal = (completionSignal.handle != 0);
  if (!external_signal) {
    // Get active signal for current dispatch if profiling is necessary
    barrier_value_packet_.completion_signal =
        Barriers().ActiveSignal(kInitSignalValueOne, skipTs ? nullptr : timestamp_);
  } else {
    // Attach external signal to the packet
    barrier_value_packet_.completion_signal = completionSignal;
  }

  // Reset fence_dirty_ flag if we submit a barrier
  if (cache_state == amd::Device::kCacheStateSystem) {
    setFenceDirty(false);
  }

  uint64_t index = Hsa::queue_add_write_index_screlease(gpu_queue_, 1);

  TrackQueueProgress(barrier_value_packet_, index, external_signal);

  WaitForQueueSlot(index, queueMask);
  hsa_amd_barrier_value_packet_t* aql_loc = &(reinterpret_cast<hsa_amd_barrier_value_packet_t*>(
      gpu_queue_->base_address))[index & queueMask];
  writePacketToRingBuffer(aql_loc, &barrier_value_packet_, packetHeader, rest, index & queueMask);
  ringQueueDoorbell(index);

  if (IsLogEnabled(amd::LOG_DETAIL_DEBUG, amd::LOG_AQL)) {
    logAqlBarrierValuePacket(roc_device_, gpu_queue_, packetHeader, &barrier_value_packet_, index,
                             priority_);
  }
  // Clear dependent signals for the next packet
  barrier_value_packet_.signal = hsa_signal_t{};
}

// ================================================================================================
bool VirtualGPU::IsQueueIdle() const {
  if (gpu_queue_ == nullptr) {
    return true;
  }

  // Nothing has been submitted to this HW queue yet.
  if (last_write_index_ == kInvalidQueueIndex) {
    return true;
  }

  // The last packet must have carried the queue-owned completion signal; otherwise idleness
  // cannot be proven.
  if (last_packet_with_signal_index_ != last_write_index_) {
    return false;
  }

  // Read the tracker-owned completion signal instead of a cached handle. The tracker owns this
  // signal for the lifetime of the HW queue, so this never dereferences a recycled/destroyed
  // signal (unlike a copied hsa_signal_t handle, which was the source of the teardown segfault).
  const ProfilingSignal* signal = barriers_.GetLastSignal();
  return (signal == nullptr) || (Hsa::signal_load_relaxed(signal->signal_) == 0);
}

// ================================================================================================
void VirtualGPU::ResetQueueStates() {
  // Release all memory dependencies
  memoryDependency().clear();

  // Release the pool, since runtime just completed a barrier
  // @note: Runtime can reset kernel arg pool only if the barrier with L2 invalidation was issued
  resetKernArgPool();
}

// ================================================================================================
bool VirtualGPU::releaseGpuMemoryFence(bool skip_cpu_wait) {
  if (hasPendingDispatch_ || isFenceDirty() || !Barriers().IsExternalSignalListEmpty()) {
    // Dispatch barrier packet into the queue
    dispatchBarrierPacket(kBarrierPacketHeader);
    hasPendingDispatch_ = false;
    skippedDispatches_ = 0;
    retainExternalSignals_ = false;
  }
  // Any fence/flush ends the coalescing window for event records.
  SetCoalesceWindow(0, nullptr);

  // Check if runtime could skip CPU wait
  if (!skip_cpu_wait) {
    Barriers().WaitCurrent();

    ResetQueueStates();
  }
  return true;
}

// ================================================================================================
VirtualGPU::VirtualGPU(Device& device, bool profiling, bool cooperative,
                       const std::vector<uint32_t>& cuMask, amd::CommandQueue::Priority priority,
                       bool dedicated_queue)
    : device::VirtualDevice(device),
      state_(0),
      gpu_queue_(nullptr),
      roc_device_(device),
      virtualQueue_(nullptr),
      deviceQueueSize_(0),
      maskGroups_(0),
      schedulerThreads_(0),
      schedulerQueue_(nullptr),
      barriers_(*this),
      managed_buffer_(*this, kStagingPoolNumSignals * device.settings().stagedXferSize_, kStagingPoolNumSignals),
      managed_kernarg_buffer_(*this, device.settings().kernargPoolSize_, kKernArgPoolNumSignals),
      cuMask_(cuMask),
      priority_(priority),
      copy_command_type_(0),
      fence_state_(Device::CacheState::kCacheStateInvalid),
      fence_dirty_(false),
      dedicated_queue_(dedicated_queue),
      schedulerQueueThreadRunning_(false),
      hostcallBuffer_(nullptr) {
  index_ = device.numOfVgpus_++;
  gpu_device_ = device.getBackendDevice();
  printfdbg_ = nullptr;

  // Initialize the last signal and dispatch flags
  timestamp_ = nullptr;
  command_ = nullptr;
  hasPendingDispatch_ = false;
  skippedDispatches_ = 0;
  profiling_ = profiling;
  cooperative_ = cooperative;

  // Initialize barrier and barrier value packets
  barrier_packet_.header = kInvalidAql;
  barrier_value_packet_.header.header = kInvalidAql;

  constexpr uint16_t kernelDispatchHBits =
      (HSA_PACKET_TYPE_KERNEL_DISPATCH << HSA_PACKET_HEADER_TYPE);
  constexpr uint16_t barrierHBits = (1 << HSA_PACKET_HEADER_BARRIER);
  constexpr uint16_t agentScopeHBits =
      (HSA_FENCE_SCOPE_AGENT << HSA_PACKET_HEADER_SCACQUIRE_FENCE_SCOPE) |
      (HSA_FENCE_SCOPE_AGENT << HSA_PACKET_HEADER_SCRELEASE_FENCE_SCOPE);
  constexpr uint16_t systemScopeHBits =
      (HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_SCACQUIRE_FENCE_SCOPE) |
      (HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_SCRELEASE_FENCE_SCOPE);
  // acquire=SYSTEM, release=AGENT (needed for GFX12)
  constexpr uint16_t sysAcquireAgentReleaseHBits =
      (HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_SCACQUIRE_FENCE_SCOPE) |
      (HSA_FENCE_SCOPE_AGENT << HSA_PACKET_HEADER_SCRELEASE_FENCE_SCOPE);
  constexpr uint16_t vendorSpecificHBits = (HSA_PACKET_TYPE_VENDOR_SPECIFIC
                                            << HSA_PACKET_HEADER_TYPE);

  if (device.settings().fenceScopeAgent_) {
    const auto& isa = device.isa();
    const bool isGfx12 = (isa.versionMajor() == 12) && (isa.versionMinor() == 0) &&
                         (isa.versionStepping() == 0 || isa.versionStepping() == 1);

    dispatchPacketHeaderNoSync_ =
        ((device.settings().ext_dispatch_packet_ ? vendorSpecificHBits : kernelDispatchHBits) |
         (isGfx12 ? sysAcquireAgentReleaseHBits : agentScopeHBits));
    dispatchPacketHeader_ =
        ((device.settings().ext_dispatch_packet_ ? vendorSpecificHBits : kernelDispatchHBits) |
         barrierHBits | (isGfx12 ? sysAcquireAgentReleaseHBits : agentScopeHBits));
  } else {
    dispatchPacketHeaderNoSync_ =
        ((device.settings().ext_dispatch_packet_ ? vendorSpecificHBits : kernelDispatchHBits) |
         systemScopeHBits);
    dispatchPacketHeader_ =
        ((device.settings().ext_dispatch_packet_ ? vendorSpecificHBits : kernelDispatchHBits) |
         barrierHBits | systemScopeHBits);
  }

  aqlHeader_ = dispatchPacketHeader_;

  // Note: Virtual GPU device creation must be a thread safe operation
  roc_device_.vgpus_.resize(roc_device_.numOfVgpus_);
  roc_device_.vgpus_[index()] = this;
}

// ================================================================================================
VirtualGPU::~VirtualGPU() {
  // Release SDMA engine assignment for this VirtualGPU
  ReleaseSdmaEngines();

  delete blitMgr_;

  if (tracking_created_) {
    std::scoped_lock l(execution());
    // Dedicated queues keep their HW queue, never acquire from pool
    if (!dedicated_queue_ && gpu_queue_ == nullptr) {
      SetGpuQueue(roc_device_.AcquireActiveQueue(priority_));
    }
    // Windows requires an interrupt in more cases than Linux for OS fence updates
    force_irq_ = IS_WINDOWS;
    // Force extra barrier to make sure OS gets an interrupt,
    // but avoid if the PM4 emulation, since PM4 path can deadlock during device destruction
    hasPendingDispatch_ |= IS_WINDOWS && !dev().IsPm4Emulation();
    // Release the resources of signal
    releaseGpuMemoryFence();
  }

  // Release any retained coalescing-window signal not freed via releaseGpuMemoryFence.
  {
    std::scoped_lock l(execution());
    SetCoalesceWindow(0, nullptr);
  }

  if (timestamp_ != nullptr) {
    timestamp_->release();
    timestamp_ = nullptr;
    LogError("There was a timestamp that was not used; deleting.");
  }

  delete printfdbg_;

  if (nullptr != schedulerQueue_) {
#if defined(_WIN32)
    // Stop the monitor thread before destroying the queue
    if (schedulerQueueThread_.joinable()) {
      schedulerQueueThreadRunning_.store(false, std::memory_order_release);
      scheduler_cv_.notify_one();
      schedulerQueueThread_.join();
    }
#endif  // _WIN32
    Hsa::queue_destroy(schedulerQueue_);
    schedulerQueue_ = nullptr;
    schedulerDoorbell_ = nullptr;
  }

  if (nullptr != virtualQueue_) {
    virtualQueue_->release();
  }

  {
    // Lock the device to make the following thread safe
    std::scoped_lock lock(roc_device_.vgpusAccess());

    --roc_device_.numOfVgpus_;  // Virtual gpu unique index decrementing
    roc_device_.vgpus_.erase(roc_device_.vgpus_.begin() + index());
    for (uint idx = index(); idx < roc_device_.vgpus().size(); ++idx) {
      roc_device_.vgpus()[idx]->index_--;
    }
  }

  if (gpu_queue_ != nullptr) {
    roc_device_.releaseQueue(gpu_queue_, cuMask_, cooperative_);
  }

  if (hostcallBuffer_) {
    ClPrint(amd::LOG_DETAIL_DEBUG, amd::LOG_QUEUE, "Deleting hostcall buffer %p", hostcallBuffer_);
    amd::disableHostcalls(hostcallBuffer_);
    roc_device_.hostFree(hostcallBuffer_, hostcallBufferSize_);
  }
}

// ================================================================================================
bool VirtualGPU::create() {
  // Pick a reasonable queue size
  uint32_t queue_size = ROC_AQL_QUEUE_SIZE;
  SetGpuQueue(roc_device_.acquireQueue(queue_size, cooperative_, cuMask_, priority_, false,
                                       dedicated_queue_));
  if (!gpu_queue_) return false;

  if (dev().isa().versionMajor() == 12 && dev().isa().versionMinor() >= 5) {
    metadata_preloader_.SetLaunchDescriptorVersion(AMD_LAUNCH_DESCRIPTOR_VERSION_GFX1250);
  }

  if (!managed_kernarg_buffer_.Create(Device::MemorySegment::kKernArg)) {
    LogError("Couldn't allocate arguments/signals for the queue");
    return false;
  }

  device::BlitManager::Setup blitSetup;
  blitMgr_ = new KernelBlitManager(*this, blitSetup);
  if ((nullptr == blitMgr_) || !blitMgr_->create(roc_device_)) {
    LogError("Could not create BlitManager!");
    return false;
  }

  // Create a object of PrintfDbg
  printfdbg_ = new PrintfDbg(roc_device_);
  if (nullptr == printfdbg_) {
    LogError("Could not create printfDbg Object!");
    return false;
  }

  // Initialize timestamp conversion factor
  if (Timestamp::getGpuTicksToTime() == 0) {
    uint64_t frequency;
    Hsa::system_get_info(HSA_SYSTEM_INFO_TIMESTAMP_FREQUENCY, &frequency);
    Timestamp::setGpuTicksToTime(1e9 / double(frequency));
  }

  if (!memoryDependency().create(GPU_NUM_MEM_DEPENDENCY)) {
    LogError("Could not create the array of memory objects!");
    return false;
  }

  // Allocate signal tracker for ROCr copy queue
  tracking_created_ = Barriers().Create();
  if (!tracking_created_) {
    LogError("Could not create signal for copy queue!");
    return false;
  }
  // Create managed buffer for staging copies
  if (!managed_buffer_.Create(Device::MemorySegment::kNoAtomics)) {
    LogError("Could not create managed buffer for this queue!");
    return false;
  }
  // Release HW queue until the first usage
  ReleaseHwQueue();
  return true;
}

// ================================================================================================
VirtualGPU::ManagedBuffer::~ManagedBuffer() {
  for (auto& it : pool_signal_) {
    if (it.handle != 0) {
      Hsa::signal_destroy(it);
    }
  }
  if (pool_base_ != nullptr) {
    gpu_.dev().hostFree(pool_base_, pool_size_);
  }
}

// ================================================================================================
bool VirtualGPU::ManagedBuffer::Create(Device::MemorySegment mem_segment) {
  pool_chunk_end_ = pool_size_ / num_chunk_signals_;
  active_chunk_ = 0;
  // Allocate memory for managed buffer
  if (mem_segment == Device::MemorySegment::kKernArg &&
      (gpu_.dev().settings().kernel_arg_impl_ != KernelArgImpl::HostKernelArgs) &&
      gpu_.dev().info().largeBar_) {
    amd::Device::AllocationFlags flags = {};
    flags.executable_ = true;
    pool_base_ = reinterpret_cast<address>(gpu_.dev().deviceLocalAlloc(pool_size_, flags, false));
    if (pool_base_ != nullptr) {
      // @note Workaround first access penalty.
      // KFD/TTM maps CPU PTEs lazily, and TTM prefaults 16 4KiB pages
      // from the faulting address.
      constexpr size_t kCpuPtePrefaultSpan = 64 * Ki;
      volatile unsigned char* ptr = pool_base_;
      for (size_t offset = 0; offset < pool_size_; offset += kCpuPtePrefaultSpan) {
        ptr[offset] = 0;
      }
    }
  } else {
    pool_base_ = reinterpret_cast<address>(gpu_.dev().hostAlloc(pool_size_, 0, mem_segment, nullptr, false));
  }
  if (pool_base_ == nullptr) {
    return false;
  }
  hsa_agent_t agent = gpu_.dev().getBackendDevice();
  for (auto& it : pool_signal_) {
    if (HSA_STATUS_SUCCESS != Hsa::signal_create(0, 1, &agent, HSA_AMD_SIGNAL_AMD_GPU_ONLY, &it)) {
      return false;
    }
  }
  return true;
}

// ================================================================================================
address VirtualGPU::ManagedBuffer::Acquire(uint32_t size) {
  return Acquire(size, gpu_.dev().info().globalMemCacheLineSize_);
}

// ================================================================================================
address VirtualGPU::ManagedBuffer::Acquire(uint32_t size, uint32_t alignment) {
  assert(alignment != 0);
  address result = nullptr;
  result = amd::alignUp(pool_base_ + pool_cur_offset_, alignment);
  const size_t pool_new_usage = (result + size) - pool_base_;
  if (pool_new_usage <= pool_chunk_end_) {
    pool_cur_offset_ = pool_new_usage;
    return result;
  } else {
    // Reset the signal for the barrier packet
    Hsa::signal_silent_store_relaxed(pool_signal_[active_chunk_], kInitSignalValueOne);
    ClPrint(amd::LOG_DETAIL_DEBUG, amd::LOG_KERN, "Issue barrier to flush chunk %d",
            active_chunk_);
    // Currently don't skip wait signal check, because SDMA engine cna be used in staging copy
    constexpr bool kSkipSignal = false;
    // Dispatch a barrier packet into the queue
    gpu_.dispatchBarrierPacket(kBarrierPacketHeader, kSkipSignal, pool_signal_[active_chunk_]);
    // Get the next chunk
    active_chunk_ = ++active_chunk_ % num_chunk_signals_;
    // Make sure the new active chunk is free
    bool test = WaitForSignal(pool_signal_[active_chunk_], gpu_.ActiveWait());
    assert(test && "Runtime can't fail a wait for chunk!");
    // Make sure the current offset matches the new chunk to avoid possible overlaps
    // between chunks and issues during recycle
    pool_cur_offset_ = (active_chunk_ == 0) ? 0 : pool_chunk_end_;
    pool_chunk_end_ = pool_cur_offset_ + pool_size_ / num_chunk_signals_;
    result = amd::alignUp(pool_base_ + pool_cur_offset_, alignment);
    pool_cur_offset_ = (result + size) - pool_base_;
  }

  return result;
}

// ================================================================================================
void VirtualGPU::ManagedBuffer::ResetPool() {
  pool_cur_offset_ = 0;
  pool_chunk_end_ = pool_size_ / num_chunk_signals_;
  active_chunk_ = 0;
}

// ================================================================================================
void* VirtualGPU::allocKernArg(size_t size, size_t alignment) {
  return managed_kernarg_buffer_.Acquire(size, alignment);
}

// ================================================================================================
address VirtualGPU::allocKernelArguments(size_t size, size_t alignment) {
  if (ROC_SKIP_KERNEL_ARG_COPY) {
    // Make sure VirtualGPU has an exclusive access to the resources
    std::scoped_lock lock(execution());
    return reinterpret_cast<address>(allocKernArg(size, alignment));
  } else {
    return nullptr;
  }
}

// ================================================================================================
void VirtualGPU::ReleaseSdmaEngines() {
  if (!hasAssignedSdmaEngine()) {
    return;
  }
  // Release SDMA engine assignment when queue is idle
  // This allows the engine to be reassigned to other active streams
  dev().ReleaseSdmaEngine(this);
  ClearAssignedSdmaEngine();
}

// ================================================================================================
void VirtualGPU::ReleaseAllHwQueues() {
  if (roc_device_.settings().dynamic_queues_) {
    // Check if any priority level exceeds max_hw_queues_
    bool should_release = false;
    for (uint qIdx = 0; qIdx < Device::QueuePriority::Total; ++qIdx) {
      if (roc_device_.NumQueues(qIdx) > roc_device_.settings().max_hw_queues_) {
        should_release = true;
        break;
      }
    }
    if (should_release) {
      // Lock the device to make the following thread safe
      std::scoped_lock lock(roc_device_.vgpusAccess());
      for (uint idx = 0; idx < roc_device_.vgpus().size(); ++idx) {
        roc_device_.vgpus()[idx]->ReleaseHwQueue();
      }
    }
  }
}

// ================================================================================================
void VirtualGPU::ReleaseHwQueue() {
  // Dedicated queues and pinned graph queues keep their HW queue
  if (dedicated_queue_ || queue_pinned_) {
    return;
  }

  // Try to release queue to the pool of active queues.
  // Use tryLock() since this may be called from the HsaAmdSignalHandler
  // and blocking here could cause deadlock
  if (roc_device_.settings().dynamic_queues_ > 0 && !cooperative_ &&
      (cuMask_.size() == 0)) {
    // If tryLock fails, skip the release - the queue will be released
    // on next opportunity
    if (execution().try_lock()) {
      if (gpu_queue_ != nullptr) {
        if (IsQueueIdle()) {
          last_hwq_ = gpu_queue_;
          if (roc_device_.ReleaseActiveQueue(gpu_queue_, priority_)) {
            SetGpuQueue(nullptr);
          }
        }
      }
      execution().unlock();
    }
  }
}

// ================================================================================================
// hsa_amd_profiling_async_copy_enable() is a process-global toggle shared by
// every agent/queue, so it cannot be scoped per-stream. Multiple VirtualGPUs
// (one per stream, potentially on different host threads) may want SDMA
// profiling enabled for their in-flight command at the same time. Guard the
// global with a process-wide refcount under a lock: enable on the 0->1
// transition, disable on the 1->0 transition, so it stays on exactly while at
// least one profiled command is outstanding and is never toggled off from
// under a concurrent profiled command on another stream.
namespace {
std::mutex g_sdmaProfilingLock;
uint32_t g_sdmaProfilingRefCount = 0;

void acquireSdmaProfiling() {
  std::lock_guard<std::mutex> lock(g_sdmaProfilingLock);
  if (g_sdmaProfilingRefCount++ == 0) {
    Hsa::profiling_async_copy_enable(true);
  }
}

void releaseSdmaProfiling() {
  std::lock_guard<std::mutex> lock(g_sdmaProfilingLock);
  assert(g_sdmaProfilingRefCount > 0 && "unbalanced SDMA profiling release");
  if (--g_sdmaProfilingRefCount == 0) {
    Hsa::profiling_async_copy_enable(false);
  }
}
}  // namespace

// ================================================================================================
/* profilingBegin, when profiling is enabled, creates a timestamp to save in
 * virtualgpu's timestamp_, saves the pointer timestamp_ to the command's data
 * and then calls start() to get the current host timestamp.
 */
void VirtualGPU::profilingBegin(amd::Command& command, bool sdmaProfiling) {
  // Dedicated queues keep their HW queue, never acquire from pool
  if (!dedicated_queue_ && gpu_queue_ == nullptr) {
    SetGpuQueue(roc_device_.AcquireActiveQueue(priority_));
  }
  // Track the current command
  command_ = &command;

  // Any non-marker command is intervening work that ends the event coalescing window.
  if (!command.isMarkerCommand()) {
    SetCoalesceWindow(0, nullptr);
  }

  // Disable profiling when command is being captured to prevent memory leak from created timestamp_
  // which won't get freed, since the command is not being executed until graph launch
  if (!command.getPktCapturingState() && command.profilingInfo().enabled_) {
    if (timestamp_ != nullptr) {
      LogWarning(
          "Trying to create a second timestamp in VirtualGPU. \
                  This could have unintended consequences.");
      return;
    }

    // Without barrier profiling will wait for each individual signal
    timestamp_ = new Timestamp(this, command);
    command.data().emplace_back(timestamp_);
    timestamp_->start();
    // Enable SDMA profiling for this command only; disabled in profilingEnd().
    // The underlying HSA async-copy profiling flag is process-global, so use a
    // refcounted, locked acquire instead of a per-stream bool (which races
    // across concurrent streams and could disable profiling out from under
    // another stream's in-flight profiled copy).
    if (sdmaProfiling) {
      acquireSdmaProfiling();
      sdma_profiling_for_cmd_ = true;
    }
  }

  if (!retainExternalSignals_) {
    Barriers().ClearExternalSignals();
  }
  for (auto it = command.eventWaitList().begin(); it < command.eventWaitList().end(); ++it) {
    void* hw_event =
        ((*it)->NotifyEvent() != nullptr) ? (*it)->NotifyEvent()->HwEvent() : (*it)->HwEvent();
    if (hw_event != nullptr) {
      Barriers().AddExternalSignal(reinterpret_cast<ProfilingSignal*>(hw_event));
    } else if (static_cast<amd::Command*>(*it)->queue() != command.queue() &&
                ((*it)->status() != CL_COMPLETE)) {
      LogPrintfError("Waiting event(%p) doesn't have a HSA signal!\n", *it);
    } else {
      // Assume serialization on the same queue...
    }

    // Check if the waiting event's queue has a dirty fence and propagate it
    if (!isFenceDirty()) {
      amd::Command* wait_cmd = static_cast<amd::Command*>(*it);
      if (wait_cmd->queue() != nullptr && wait_cmd->queue() != command.queue()) {
        device::VirtualDevice* wait_vdev = wait_cmd->queue()->vdev();
        if (wait_vdev != nullptr && wait_vdev->isFenceDirty()) {
          setFenceDirty(true);
        }
      }
    }
  }
  for (auto it = command.getDepHwEvents().begin(); it < command.getDepHwEvents().end(); ++it) {
    ClPrint(amd::LOG_DEBUG, amd::LOG_SIG, "Adding dep hw event signal: 0x%lx",
            reinterpret_cast<ProfilingSignal*>(*it)->signal_.handle);
    Barriers().AddExternalSignal(reinterpret_cast<ProfilingSignal*>(*it));
  }
  command.clearDepHwEvents();
}

// ================================================================================================
/* profilingEnd, when profiling is enabled, checks to see if a signal was
 * created for whatever command we are running and calls end() to get the
 * current host timestamp if no signal is available.
 */
void VirtualGPU::profilingEnd(bool clearHwEvent) {
  if (!command_->getPktCapturingState() && command_->profilingInfo().enabled_) {
    if (timestamp_ != nullptr) {
      if (timestamp_->HwProfiling() == false) {
        timestamp_->end();
      }
      timestamp_ = nullptr;
    }
  }

  if (sdma_profiling_for_cmd_) {
    releaseSdmaProfiling();
    sdma_profiling_for_cmd_ = false;
  }

  // Certain commands like map/unmap memory may not need hw_events as its not a
  // queue operation. In such cases clear already set events which may have been for sync
  // before some memory map/unmap operation
  if (clearHwEvent) {
    if (command_->HwEvent() != nullptr) {
      reinterpret_cast<ProfilingSignal*>(command_->HwEvent())->release();
      command_->SetHwEvent(nullptr);
    }
  }

  // Clear the command tracking
  command_ = nullptr;
}

// ================================================================================================
void VirtualGPU::SetCoalesceWindow(uint64_t event_id, void* hw_event) {
  last_barrier_coalesce_event_ = event_id;
  if (hw_event != last_barrier_hw_event_) {
    if (hw_event != nullptr) {
      reinterpret_cast<ProfilingSignal*>(hw_event)->retain();
    }
    if (last_barrier_hw_event_ != nullptr) {
      reinterpret_cast<ProfilingSignal*>(last_barrier_hw_event_)->release();
    }
    last_barrier_hw_event_ = hw_event;
  }
}

// ================================================================================================
void VirtualGPU::AttachHwEvent(amd::Command* cmd, void* hw_event) {
  if (cmd == nullptr || hw_event == nullptr) {
    return;
  }
  if (cmd->HwEvent() != nullptr) {
    reinterpret_cast<ProfilingSignal*>(cmd->HwEvent())->release();
  }
  cmd->SetHwEvent(hw_event);
  reinterpret_cast<ProfilingSignal*>(hw_event)->retain();
}

// ================================================================================================
void VirtualGPU::updateCommandsState(amd::Command* list) const {
  Timestamp* ts = nullptr;

  amd::Command* current = list;
  amd::Command* next = nullptr;

  if (current == nullptr) {
    return;
  }

  uint64_t endTimeStamp = 0;
  uint64_t startTimeStamp = endTimeStamp;

  if (current->profilingInfo().enabled_) {
    // TODO: use GPU timestamp when available.
    endTimeStamp = amd::Os::timeNanos();
    startTimeStamp = endTimeStamp;

    // This block gets the first valid timestamp from the first command
    // that has one. This timestamp is used below to mark any command that
    // came before it to start and end with this first valid start time.
    current = list;
    while (current != nullptr) {
      if (!current->data().empty()) {
        ts = reinterpret_cast<Timestamp*>(current->data().back());
        ts->getTime(&startTimeStamp, &endTimeStamp);
        break;
      }
      current = current->getNext();
    }
  }

  // Iterate through the list of commands, and set timestamps as appropriate
  // Note, if a command does not have a timestamp, it does one of two things:
  // - if the command (without a timestamp), A, precedes another command, C,
  // that _does_ contain a valid timestamp, command A will set RUNNING and
  // COMPLETE with the RUNNING (start) timestamp from command C. This would
  // also be true for command B, which is between A and C. These timestamps
  // are actually retrieved in the block above (startTimeStamp, endTimeStamp).
  // - if the command (without a timestamp), C, follows another command, A,
  // that has a valid timestamp, command C will be set RUNNING and COMPLETE
  // with the COMPLETE (end) timestamp of the previous command, A. This is
  // also true for any command B, which falls between A and C.
  current = list;
  while (current != nullptr) {
    if (current->profilingInfo().enabled_) {
      if (!current->data().empty()) {
        for (auto i = 0; i < current->data().size(); i++) {
          // Since this is a valid command to get a timestamp, we use the
          // timestamp provided by the runtime (saved in the data())
          ts = reinterpret_cast<Timestamp*>(current->data()[i]);
          ts->getTime(&startTimeStamp, &endTimeStamp);
          ts->release();
        }
        current->data().clear();
      } else {
        // If we don't have a command that contains a valid timestamp,
        // we simply use the end timestamp of the previous command.
        // Note, if this is a command before the first valid timestamp,
        // this will be equal to the start timestamp of the first valid
        // timestamp at this point.
        startTimeStamp = endTimeStamp;
      }
    }

    if (current->status() == CL_SUBMITTED) {
      current->setStatus(CL_RUNNING, startTimeStamp);
      current->setStatus(CL_COMPLETE, endTimeStamp);
    } else if (current->status() != CL_COMPLETE) {
      LogPrintfError("Unexpected command status - %d.", current->status());
    }

    next = current->getNext();
    current->release();
    current = next;
  }
}

// ================================================================================================

void VirtualGPU::submitReadMemory(amd::ReadMemoryCommand& cmd) {
  // Make sure VirtualGPU has an exclusive access to the resources
  std::scoped_lock lock(execution());

  profilingBegin(cmd, true);

  size_t offset = 0;
  // Find if virtual address is a CL allocation
  device::Memory* hostMemory = dev().findMemoryFromVA(cmd.destination(), &offset);

  Memory* devMem = dev().getRocMemory(&cmd.source());
  // Synchronize data with other memory instances if necessary
  devMem->syncCacheFromHost(*this);

  void* dst = cmd.destination();
  amd::Coord3D size = cmd.size();

  //! @todo: add multi-devices synchronization when supported.

  cl_command_type type = cmd.type();
  bool result = false;
  bool imageBuffer = false;

  // Force buffer read for IMAGE1D_BUFFER
  if ((type == CL_COMMAND_READ_IMAGE) && (cmd.source().getType() == CL_MEM_OBJECT_IMAGE1D_BUFFER)) {
    type = CL_COMMAND_READ_BUFFER;
    imageBuffer = true;
  }

  switch (type) {
    case CL_COMMAND_READ_BUFFER: {
      amd::Coord3D origin(cmd.origin()[0]);
      if (imageBuffer) {
        size_t elemSize = cmd.source().asImage()->getImageFormat().getElementSize();
        origin.c[0] *= elemSize;
        size.c[0] *= elemSize;
      }
      if (hostMemory != nullptr) {
        // Accelerated transfer without pinning
        amd::Coord3D dstOrigin(offset);
        result = blitMgr().copyBuffer(*devMem, *hostMemory, origin, dstOrigin, size,
                                      cmd.isEntireMemory(), cmd.copyMetadata());
      } else {
        result = blitMgr().readBuffer(*devMem, dst, origin, size, cmd.isEntireMemory(),
                                      cmd.copyMetadata());
      }
      break;
    }
    case CL_COMMAND_READ_BUFFER_RECT: {
      amd::BufferRect hostbufferRect;
      amd::Coord3D region(0);
      amd::Coord3D hostOrigin(cmd.hostRect().start_ + offset);
      hostbufferRect.create(hostOrigin.c, size.c, cmd.hostRect().rowPitch_,
                            cmd.hostRect().slicePitch_);
      if (hostMemory != nullptr) {
        result = blitMgr().copyBufferRect(*devMem, *hostMemory, cmd.bufRect(), hostbufferRect, size,
                                          cmd.isEntireMemory(), cmd.copyMetadata());
      } else {
        result = blitMgr().readBufferRect(*devMem, dst, cmd.bufRect(), cmd.hostRect(), size,
                                          cmd.isEntireMemory(), cmd.copyMetadata());
      }
      break;
    }
    case CL_COMMAND_READ_IMAGE: {
      if ((cmd.source().parent() != nullptr) &&
          (cmd.source().parent()->getType() == CL_MEM_OBJECT_BUFFER)) {
        Image* imageBuffer = static_cast<Image*>(devMem);
        // Check if synchronization has to be performed
        if (nullptr != imageBuffer->CopyImageBuffer()) {
          amd::Memory* memory = imageBuffer->CopyImageBuffer();
          devMem = dev().getGpuMemory(memory);
          Memory* buffer = dev().getGpuMemory(imageBuffer->owner()->parent());
          amd::Image* image = imageBuffer->owner()->asImage();
          amd::Coord3D offs(0);
          // Copy memory from the original image buffer into the backing store image
          result = blitMgr().copyBufferToImage(*buffer, *devMem, offs, offs, image->getRegion(),
                                               true, image->getRowPitch(), image->getSlicePitch());
        }
      }
      if (hostMemory != nullptr) {
        // Accelerated image to buffer transfer without pinning
        amd::Coord3D dstOrigin(offset);
        result = blitMgr().copyImageToBuffer(*devMem, *hostMemory, cmd.origin(), dstOrigin, size,
                                             cmd.isEntireMemory(), cmd.rowPitch(), cmd.slicePitch(),
                                             cmd.copyMetadata());
      } else {
        result = blitMgr().readImage(*devMem, dst, cmd.origin(), size, cmd.rowPitch(),
                                     cmd.slicePitch(), cmd.isEntireMemory(), cmd.copyMetadata());
      }
      break;
    }
    default:
      ShouldNotReachHere();
      break;
  }

  if (!result) {
    LogError("submitReadMemory failed!");
    cmd.setStatus(CL_OUT_OF_RESOURCES);
  }

  profilingEnd();
}

void VirtualGPU::submitWriteMemory(amd::WriteMemoryCommand& cmd) {
  // Make sure VirtualGPU has an exclusive access to the resources
  std::scoped_lock lock(execution());

  profilingBegin(cmd, true);

  size_t offset = 0;
  // Find if virtual address is a CL allocation
  device::Memory* hostMemory = dev().findMemoryFromVA(cmd.source(), &offset);

  Memory* devMem = dev().getRocMemory(&cmd.destination());

  // Synchronize memory from host if necessary
  device::Memory::SyncFlags syncFlags;
  syncFlags.skipEntire_ = cmd.isEntireMemory();
  devMem->syncCacheFromHost(*this, syncFlags);

  const char* src = static_cast<const char*>(cmd.source());
  amd::Coord3D size = cmd.size();

  //! @todo add multi-devices synchronization when supported.

  cl_command_type type = cmd.type();
  bool result = false;
  bool imageBuffer = false;

  // Force buffer write for IMAGE1D_BUFFER
  if ((type == CL_COMMAND_WRITE_IMAGE) &&
      (cmd.destination().getType() == CL_MEM_OBJECT_IMAGE1D_BUFFER)) {
    type = CL_COMMAND_WRITE_BUFFER;
    imageBuffer = true;
  }

  switch (type) {
    case CL_COMMAND_WRITE_BUFFER: {
      amd::Coord3D origin(cmd.origin()[0]);
      if (imageBuffer) {
        size_t elemSize = cmd.destination().asImage()->getImageFormat().getElementSize();
        origin.c[0] *= elemSize;
        size.c[0] *= elemSize;
      }
      if (hostMemory != nullptr) {
        // Accelerated transfer without pinning
        amd::Coord3D srcOrigin(offset);
        result = blitMgr().copyBuffer(*hostMemory, *devMem, srcOrigin, origin, size,
                                      cmd.isEntireMemory(), cmd.copyMetadata());
      } else {
        result = blitMgr().writeBuffer(src, *devMem, origin, size, cmd.isEntireMemory(),
                                       cmd.copyMetadata());
      }
      break;
    }
    case CL_COMMAND_WRITE_BUFFER_RECT: {
      amd::BufferRect hostbufferRect;
      amd::Coord3D region(0);
      amd::Coord3D hostOrigin(cmd.hostRect().start_ + offset);
      hostbufferRect.create(hostOrigin.c, size.c, cmd.hostRect().rowPitch_,
                            cmd.hostRect().slicePitch_);
      if (hostMemory != nullptr) {
        result = blitMgr().copyBufferRect(*hostMemory, *devMem, hostbufferRect, cmd.bufRect(), size,
                                          cmd.isEntireMemory(), cmd.copyMetadata());
      } else {
        result = blitMgr().writeBufferRect(src, *devMem, cmd.hostRect(), cmd.bufRect(), size,
                                           cmd.isEntireMemory(), cmd.copyMetadata());
      }
      break;
    }
    case CL_COMMAND_WRITE_IMAGE: {
      if (hostMemory != nullptr) {
        // Accelerated buffer to image transfer without pinning
        amd::Coord3D srcOrigin(offset);
        result = blitMgr().copyBufferToImage(*hostMemory, *devMem, srcOrigin, cmd.origin(), size,
                                             cmd.isEntireMemory(), cmd.rowPitch(), cmd.slicePitch(),
                                             cmd.copyMetadata());
      } else {
        result = blitMgr().writeImage(src, *devMem, cmd.origin(), size, cmd.rowPitch(),
                                      cmd.slicePitch(), cmd.isEntireMemory(), cmd.copyMetadata());
      }
      break;
    }
    default:
      ShouldNotReachHere();
      break;
  }

  if (!result) {
    LogError("submitWriteMemory failed!");
    cmd.setStatus(CL_OUT_OF_RESOURCES);
  } else {
    cmd.destination().signalWrite(&dev());
  }

  profilingEnd();
}

// ================================================================================================
void VirtualGPU::submitSvmFreeMemory(amd::SvmFreeMemoryCommand& cmd) {
  // Make sure VirtualGPU has an exclusive access to the resources
  std::scoped_lock lock(execution());

  // in-order semantics: previous commands need to be done before we start
  releaseGpuMemoryFence();

  profilingBegin(cmd);
  const std::vector<void*>& svmPointers = cmd.svmPointers();
  if (cmd.pfnFreeFunc() == nullptr) {
    // pointers allocated using clSVMAlloc
    for (uint32_t i = 0; i < svmPointers.size(); i++) {
      amd::SvmBuffer::free(cmd.context(), svmPointers[i]);
    }
  } else {
    cmd.pfnFreeFunc()(as_cl(cmd.queue()->asCommandQueue()), svmPointers.size(),
                      (void**)(&(svmPointers[0])), cmd.userData());
  }
  profilingEnd();
}

// ================================================================================================
void VirtualGPU::submitSvmPrefetchAsync(amd::SvmPrefetchAsyncCommand& cmd) {
  // Make sure VirtualGPU has an exclusive access to the resources
  std::scoped_lock lock(execution());
  profilingBegin(cmd);

  if (dev().info().hmmSupported_) {
    // Initialize signal for the barrier
    auto wait_events = Barriers().WaitingSignal(HwQueueEngine::Unknown);
    hsa_signal_t active = Barriers().ActiveSignal(kInitSignalValueOne, timestamp_);

    // Find the requested agent for the transfer
    hsa_agent_t agent =
        (cmd.cpu_access() || (dev().settings().hmmFlags_ & Settings::Hmm::EnableSystemMemory))
            ? dev().getCpuAgent(cmd.numa_id())
            : (static_cast<const roc::Device*>(cmd.device()))->getBackendDevice();

    // Initiate a prefetch command
    hsa_status_t status =
        Hsa::svm_prefetch_async(const_cast<void*>(cmd.dev_ptr()), cmd.count(), agent,
                                wait_events.size(), wait_events.data(), active);
    ClPrint(amd::LOG_DEBUG, amd::LOG_COPY,
            "HSA prefetch async dev_ptr=0x%zx, count=%d, wait_event=0x%zx, "
            "completion_signal=0x%zx",
            const_cast<void*>(cmd.dev_ptr()), cmd.count(),
            (wait_events.size() != 0) ? wait_events[0].handle : 0, active.handle);

    if ((status != HSA_STATUS_SUCCESS)) {
      Barriers().ResetCurrentSignal();
      LogError("hsa_amd_svm_prefetch_async failed");
      cmd.setStatus(CL_INVALID_OPERATION);
    }

    // Add system scope, since the prefetch scope is unclear
    addSystemScope();
  } else {
    LogWarning("hsa_amd_svm_prefetch_async is ignored, because no HMM support");
  }
  profilingEnd();
}

// ================================================================================================
void VirtualGPU::SubmitSvmPrefetchBatchAsync(amd::SvmPrefetchBatchAsyncCommand& command) {
  std::scoped_lock lock(execution());
  profilingBegin(command);

  auto wait_events = Barriers().WaitingSignal(HwQueueEngine::Unknown);
  hsa_signal_t active = Barriers().ActiveSignal(command.Count(), timestamp_);

  const bool enable_system_memory =
      (dev().settings().hmmFlags_ & Settings::Hmm::EnableSystemMemory) != 0;

  for (size_t idx = 0; idx < command.Count(); idx++) {
    void* dev_ptr = command.DevicePointers()[idx];
    size_t size = command.Sizes()[idx];
    const roc::Device* target_dev = static_cast<const roc::Device*>(command.TargetDevices()[idx]);
    bool cpu_access = target_dev == nullptr;

    hsa_agent_t agent = (cpu_access || enable_system_memory) ? dev().getCpuAgent(CpuDeviceId)
                                                             : target_dev->getBackendDevice();

    hsa_status_t status = Hsa::svm_prefetch_async(dev_ptr, size, agent, wait_events.size(),
                                                  wait_events.data(), active);
    ClPrint(amd::LOG_DEBUG, amd::LOG_COPY,
            "HSA prefetch batch async[%zu] dev_ptr=0x%zx, size=%zu, wait_event=0x%zx, "
            "completion_signal=0x%zx",
            idx, dev_ptr, size, wait_events.empty() ? 0 : wait_events[0].handle, active.handle);

    if (status != HSA_STATUS_SUCCESS) {
      Barriers().ResetCurrentSignal();
      LogError("HSA prefetch batch async failed in batch operation");
      command.setStatus(CL_INVALID_OPERATION);
      profilingEnd();
      return;
    }
  }

  addSystemScope();
  profilingEnd();
}

// ================================================================================================
void VirtualGPU::SubmitSvmDiscardBatchAsync(amd::SvmDiscardBatchAsyncCommand& command) {
  std::scoped_lock lock(execution());
  profilingBegin(command);

  auto wait_events = Barriers().WaitingSignal(HwQueueEngine::Unknown);
  hsa_signal_t active = Barriers().ActiveSignal(kInitSignalValueOne, timestamp_);

  // Collect all pointers and sizes for batch operation
  std::vector<void*> dev_ptrs_vec;
  std::vector<size_t> sizes_vec;
  for (size_t idx = 0; idx < command.Count(); idx++) {
    dev_ptrs_vec.push_back(const_cast<void*>(command.DevicePointers()[idx]));
    sizes_vec.push_back(command.Sizes()[idx]);
  }

  hsa_status_t status = Hsa::svm_discard_batch_async(dev_ptrs_vec.data(), sizes_vec.data(),
                                                      command.Count(), wait_events.size(),
                                                      wait_events.data(), active);
  ClPrint(amd::LOG_DEBUG, amd::LOG_COPY,
          "HSA discard batch async count=%zu, wait_event=0x%zx, completion_signal=0x%zx",
          command.Count(), wait_events.empty() ? 0 : wait_events[0].handle, active.handle);

  if (status != HSA_STATUS_SUCCESS) {
    Barriers().ResetCurrentSignal();
    LogError("HSA discard batch async failed");
    command.setStatus(CL_INVALID_OPERATION);
    profilingEnd();
    return;
  }

  addSystemScope();
  profilingEnd();
}

// ================================================================================================
bool VirtualGPU::copyMemory(cl_command_type type, amd::Memory& srcMem, amd::Memory& dstMem,
                            bool entire, const amd::Coord3D& srcOrigin,
                            const amd::Coord3D& dstOrigin, const amd::Coord3D& size,
                            const amd::BufferRect& srcRect, const amd::BufferRect& dstRect,
                            amd::CopyMetadata copyMetadata) {
  Memory* srcDevMem = dev().getRocMemory(&srcMem);
  Memory* dstDevMem = dev().getRocMemory(&dstMem);
  if (srcDevMem == nullptr || dstDevMem == nullptr) {
    LogError("submitCopyMemory failed!");
    return false;
  }
  // Synchronize source and destination memory
  device::Memory::SyncFlags syncFlags;
  syncFlags.skipEntire_ = entire;
  dstDevMem->syncCacheFromHost(*this, syncFlags);
  srcDevMem->syncCacheFromHost(*this);

  bool result = false;
  amd::Memory* bufferFromImageSrc = nullptr;
  amd::Memory* bufferFromImageDst = nullptr;

  // Force buffer copy for IMAGE1D_BUFFER
  if (srcMem.getType() == CL_MEM_OBJECT_IMAGE1D_BUFFER) {
    bufferFromImageSrc = createBufferFromImage(srcMem);
    if (nullptr == bufferFromImageSrc) {
      LogError("We should not fail buffer creation from image_buffer!");
    } else {
      srcDevMem = dev().getRocMemory(bufferFromImageSrc);
    }
  }
  // Force buffer write for IMAGE1D_BUFFER
  if (dstMem.getType() == CL_MEM_OBJECT_IMAGE1D_BUFFER) {
    bufferFromImageDst = createBufferFromImage(dstMem);
    if (nullptr == bufferFromImageDst) {
      LogError("We should not fail buffer creation from image_buffer!");
    } else {
      dstDevMem = dev().getRocMemory(bufferFromImageDst);
    }
  }
  type = getCopyCommandType(type, srcMem.getType(), dstMem.getType());
  switch (type) {
    case CL_COMMAND_SVM_MEMCPY:
    case CL_COMMAND_COPY_BUFFER: {
      amd::Coord3D realSrcOrigin(srcOrigin[0]);
      amd::Coord3D realDstOrigin(dstOrigin[0]);
      amd::Coord3D realSize(size.c[0], size.c[1], size.c[2]);

      if (nullptr != bufferFromImageSrc) {
        const size_t elemSize = srcMem.asImage()->getImageFormat().getElementSize();
        realSrcOrigin.c[0] *= elemSize;
        if (nullptr != bufferFromImageDst) {
          realDstOrigin.c[0] *= elemSize;
        }
        realSize.c[0] *= elemSize;
      } else if (nullptr != bufferFromImageDst) {
        const size_t elemSize = dstMem.asImage()->getImageFormat().getElementSize();
        realDstOrigin.c[0] *= elemSize;
        realSize.c[0] *= elemSize;
      }

      result = blitMgr().copyBuffer(*srcDevMem, *dstDevMem, realSrcOrigin, realDstOrigin, realSize,
                                    entire, copyMetadata);
      break;
    }
    case CL_COMMAND_COPY_BUFFER_RECT: {
      result = blitMgr().copyBufferRect(*srcDevMem, *dstDevMem, srcRect, dstRect, size, entire,
                                        copyMetadata);
      break;
    }
    case CL_COMMAND_COPY_IMAGE: {
      result = blitMgr().copyImage(*srcDevMem, *dstDevMem, srcOrigin, dstOrigin, size, entire,
                                   copyMetadata);
      break;
    }
    case CL_COMMAND_COPY_IMAGE_TO_BUFFER: {
      amd::Coord3D realDstOrigin(dstOrigin);
      if (nullptr != bufferFromImageDst) {
        const size_t elemSize = dstMem.asImage()->getImageFormat().getElementSize();
        realDstOrigin.c[0] *= elemSize;
      }
      result = blitMgr().copyImageToBuffer(*srcDevMem, *dstDevMem, srcOrigin, realDstOrigin, size,
                                   entire, dstRect.rowPitch_, dstRect.slicePitch_, copyMetadata);
      break;
    }
    case CL_COMMAND_COPY_BUFFER_TO_IMAGE: {
      amd::Coord3D realSrcOrigin(srcOrigin);
      if (nullptr != bufferFromImageSrc) {
        const size_t elemSize = srcMem.asImage()->getImageFormat().getElementSize();
        realSrcOrigin.c[0] *= elemSize;
      }
      result = blitMgr().copyBufferToImage(*srcDevMem, *dstDevMem, realSrcOrigin, dstOrigin, size,
                                   entire, srcRect.rowPitch_, srcRect.slicePitch_, copyMetadata);
      break;
    }
    default:
      ShouldNotReachHere();
      break;
  }
  if (nullptr != bufferFromImageSrc) {
    bufferFromImageSrc->release();
  }
  if (nullptr != bufferFromImageDst) {
    bufferFromImageDst->release();
  }
  if (!result) {
    LogError("submitCopyMemory failed!");
    return false;
  }

  // Mark this as the most-recently written cache of the destination
  dstMem.signalWrite(&dev());
  return true;
}

// ================================================================================================
void VirtualGPU::submitCopyMemory(amd::CopyMemoryCommand& cmd) {
  // Make sure VirtualGPU has an exclusive access to the resources
  std::scoped_lock lock(execution());

  profilingBegin(cmd, true);

  cl_command_type type = cmd.type();
  bool entire = cmd.isEntireMemory();

  if (!copyMemory(type, cmd.source(), cmd.destination(), entire, cmd.srcOrigin(), cmd.dstOrigin(),
                  cmd.size(), cmd.srcRect(), cmd.dstRect(), cmd.copyMetadata())) {
    cmd.setStatus(CL_INVALID_OPERATION);
  }

  // Runtime may change the command type to report a more accurate info in ROC profiler
  if (copy_command_type_ != 0) {
    cmd.OverrrideCommandType(copy_command_type_);
    copy_command_type_ = 0;
  }
  profilingEnd();
}

// ================================================================================================
void VirtualGPU::submitSvmCopyMemory(amd::SvmCopyMemoryCommand& cmd) {
  // Make sure VirtualGPU has an exclusive access to the resources
  std::scoped_lock lock(execution());

  profilingBegin(cmd, true);
  // no op for FGS supported device
  if (!dev().isFineGrainedSystem(true)) {
    amd::Coord3D srcOrigin(0, 0, 0);
    amd::Coord3D dstOrigin(0, 0, 0);
    amd::Coord3D size(cmd.srcSize(), 1, 1);
    amd::BufferRect srcRect;
    amd::BufferRect dstRect;

    bool result = false;
    amd::Memory* srcMem = amd::MemObjMap::FindMemObj(cmd.src());
    amd::Memory* dstMem = amd::MemObjMap::FindMemObj(cmd.dst());

    device::Memory::SyncFlags syncFlags;
    if (nullptr != srcMem) {
      srcOrigin.c[0] =
          static_cast<const_address>(cmd.src()) - static_cast<address>(srcMem->getSvmPtr());
      if (!(srcMem->validateRegion(srcOrigin, size))) {
        cmd.setStatus(CL_INVALID_OPERATION);
        return;
      }
    }
    if (nullptr != dstMem) {
      dstOrigin.c[0] =
          static_cast<const_address>(cmd.dst()) - static_cast<address>(dstMem->getSvmPtr());
      if (!(dstMem->validateRegion(dstOrigin, size))) {
        cmd.setStatus(CL_INVALID_OPERATION);
        return;
      }
    }

    if ((nullptr == srcMem && nullptr == dstMem) ||  // both not in svm space
        (nullptr != srcMem && dev().forceFineGrain(srcMem)) ||
        (nullptr != dstMem && dev().forceFineGrain(dstMem))) {
      // Wait on a kernel if one is outstanding
      releaseGpuMemoryFence();

      // If these are from different contexts, then one of them could be in the device memory
      // This is fine, since spec doesn't allow for copies with pointers from different contexts
      std::memcpy(cmd.dst(), cmd.src(), cmd.srcSize());
      result = true;
    } else if (nullptr == srcMem && nullptr != dstMem) {  // src not in svm space
      Memory* memory = dev().getRocMemory(dstMem);
      // Synchronize source and destination memory
      syncFlags.skipEntire_ = dstMem->isEntirelyCovered(dstOrigin, size);
      memory->syncCacheFromHost(*this, syncFlags);

      result = blitMgr().writeBuffer(cmd.src(), *memory, dstOrigin, size,
                                     dstMem->isEntirelyCovered(dstOrigin, size));
      // Mark this as the most-recently written cache of the destination
      dstMem->signalWrite(&dev());
    } else if (nullptr != srcMem && nullptr == dstMem) {  // dst not in svm space
      Memory* memory = dev().getRocMemory(srcMem);
      // Synchronize source and destination memory
      memory->syncCacheFromHost(*this);

      result = blitMgr().readBuffer(*memory, cmd.dst(), srcOrigin, size,
                                    srcMem->isEntirelyCovered(srcOrigin, size));
    } else if (nullptr != srcMem && nullptr != dstMem) {  // both in svm space
      bool entire =
          srcMem->isEntirelyCovered(srcOrigin, size) && dstMem->isEntirelyCovered(dstOrigin, size);
      result = copyMemory(cmd.type(), *srcMem, *dstMem, entire, srcOrigin, dstOrigin, size, srcRect,
                          dstRect);
    }

    if (!result) {
      cmd.setStatus(CL_INVALID_OPERATION);
    }
  } else {
    // Stall GPU for CPU access to memory
    releaseGpuMemoryFence();
    // direct memcpy for FGS enabled system
    amd::SvmBuffer::memFill(cmd.dst(), cmd.src(), cmd.srcSize(), 1);
  }
  profilingEnd();
}

// ================================================================================================
void VirtualGPU::submitCopyMemoryP2P(amd::CopyMemoryP2PCommand& cmd) {
  // Make sure VirtualGPU has an exclusive access to the resources
  std::scoped_lock lock(execution());

  profilingBegin(cmd, true);

  Memory* srcDevMem = static_cast<roc::Memory*>(
      cmd.source().getDeviceMemory(*cmd.source().getContext().devices()[0]));
  Memory* dstDevMem = static_cast<roc::Memory*>(
      cmd.destination().getDeviceMemory(*cmd.destination().getContext().devices()[0]));

  bool p2pAllowed = false;
  // Loop through all available P2P devices for the destination buffer
  for (auto agent : dstDevMem->dev().p2pAgents()) {
    // Find the device, which is matching the current
    if (agent.handle == dev().getBackendDevice().handle) {
      p2pAllowed = true;
      break;
    }

    for (auto agent : srcDevMem->dev().p2pAgents()) {
      if (agent.handle == dev().getBackendDevice().handle) {
        p2pAllowed = true;
        break;
      }
    }
  }

  // Synchronize source and destination memory
  device::Memory::SyncFlags syncFlags;
  syncFlags.skipEntire_ = cmd.isEntireMemory();
  amd::Coord3D size = cmd.size();

  bool result = false;
  switch (cmd.type()) {
    case CL_COMMAND_COPY_BUFFER: {
      amd::Coord3D srcOrigin(cmd.srcOrigin()[0]);
      amd::Coord3D dstOrigin(cmd.dstOrigin()[0]);

      if (p2pAllowed) {
        result = blitMgr().copyBuffer(*srcDevMem, *dstDevMem, srcOrigin, dstOrigin, size,
                                      cmd.isEntireMemory(), cmd.copyMetadata());
      } else {
        // Sync the current queue, since P2P staging uses the device queues for transfer
        releaseGpuMemoryFence();

        std::scoped_lock lock(dev().P2PStageOps());
        Memory* dstStgMem = static_cast<Memory*>(
            dev().P2PStage()->getDeviceMemory(*cmd.source().getContext().devices()[0]));
        Memory* srcStgMem = static_cast<Memory*>(
            dev().P2PStage()->getDeviceMemory(*cmd.destination().getContext().devices()[0]));

        size_t copy_size = Device::kP2PStagingSize;
        size_t left_size = size[0];
        result = true;
        do {
          if (left_size <= copy_size) {
            copy_size = left_size;
          }
          left_size -= copy_size;
          amd::Coord3D stageOffset(0);
          amd::Coord3D cpSize(copy_size);

          // Perform 2 step transfer with staging buffer
          result &= srcDevMem->dev().xferMgr().copyBuffer(*srcDevMem, *dstStgMem, srcOrigin,
                                                          stageOffset, cpSize, false,
                                                          cmd.copyMetadata());
          srcOrigin.c[0] += copy_size;
          result &= dstDevMem->dev().xferMgr().copyBuffer(*srcStgMem, *dstDevMem, stageOffset,
                                                          dstOrigin, cpSize, false,
                                                          cmd.copyMetadata());
          dstOrigin.c[0] += copy_size;
        } while (left_size > 0);
      }
      break;
    }
    case CL_COMMAND_COPY_BUFFER_RECT: {
      if (p2pAllowed) {
        result = blitMgr().copyBufferRect(*srcDevMem, *dstDevMem, cmd.srcRect(), cmd.dstRect(),
                                          size, cmd.isEntireMemory(), cmd.copyMetadata());
      } else {
        // Sync the current queue, since P2P staging uses the device queues for transfer
        releaseGpuMemoryFence();

        std::scoped_lock lock(dev().P2PStageOps());
        Memory* dstStgMem = static_cast<Memory*>(
            dev().P2PStage()->getDeviceMemory(*cmd.source().getContext().devices()[0]));
        Memory* srcStgMem = static_cast<Memory*>(
            dev().P2PStage()->getDeviceMemory(*cmd.destination().getContext().devices()[0]));

        if ((cmd.srcRect().slicePitch_ * size[2]) <= Device::kP2PStagingSize) {
          result = true;
          // Perform 2 step transfer with staging buffer
          result &= srcDevMem->dev().xferMgr().copyBufferRect(*srcDevMem, *dstStgMem, cmd.srcRect(),
                                                              cmd.srcRect(), size, false,
                                                              cmd.copyMetadata());

          result &= dstDevMem->dev().xferMgr().copyBufferRect(*srcStgMem, *dstDevMem, cmd.srcRect(),
                                                              cmd.dstRect(), size, false,
                                                              cmd.copyMetadata());
        } else {
          size_t srcOffset;
          size_t dstOffset;
          result = true;

          for (size_t z = 0; z < size[2]; ++z) {
            for (size_t y = 0; y < size[1]; ++y) {
              srcOffset = cmd.srcRect().offset(0, y, z);
              dstOffset = cmd.dstRect().offset(0, y, z);

              amd::Coord3D srcOrigin(srcOffset);
              amd::Coord3D dstOrigin(dstOffset);
              size_t copy_size = Device::kP2PStagingSize;
              size_t left_size = size[0];
              amd::Coord3D stageOffset(0);
              do {
                if (left_size <= copy_size) {
                  copy_size = left_size;
                }
                left_size -= copy_size;

                // Perform 2 step transfer with staging buffer
                result &= srcDevMem->dev().xferMgr().copyBuffer(*srcDevMem, *dstStgMem, srcOrigin,
                                                                stageOffset, copy_size);

                result &= dstDevMem->dev().xferMgr().copyBuffer(*srcStgMem, *dstDevMem, stageOffset,
                                                                dstOrigin, copy_size);

                srcOrigin.c[0] += copy_size;
                dstOrigin.c[0] += copy_size;
              } while (left_size > 0);
            }
          }
        }
      }
      break;
    }
    case CL_COMMAND_COPY_IMAGE:
    case CL_COMMAND_COPY_IMAGE_TO_BUFFER:
    case CL_COMMAND_COPY_BUFFER_TO_IMAGE:
      LogError("Unsupported P2P type!");
      break;
    default:
      ShouldNotReachHere();
      break;
  }

  if (!result) {
    LogError("submitCopyMemoryP2P failed!");
    cmd.setStatus(CL_OUT_OF_RESOURCES);
  }

  cmd.destination().signalWrite(&dstDevMem->dev());

  profilingEnd();
}

// ================================================================================================
void VirtualGPU::submitBatchCopyMemory(amd::BatchCopyMemoryCommand& cmd) {
  // Make sure VirtualGPU has an exclusive access to the resources
  std::scoped_lock lock(execution());

  profilingBegin(cmd, true);

  auto& copyOps = cmd.copyOps();
  if (copyOps.empty()) {
    profilingEnd();
    return;
  }

  bool result = true;

  // Sync caches for all ops
  device::Memory::SyncFlags syncFlags;
  syncFlags.skipEntire_ = false;

  for (const auto& op : copyOps) {
    Memory* srcDevMem = dev().getRocMemory(op.srcMemory);
    Memory* dstDevMem = dev().getRocMemory(op.dstMemory);

    if (srcDevMem == nullptr || dstDevMem == nullptr) {
      LogError("submitBatchCopyMemory: Invalid memory objects!");
      cmd.setStatus(CL_INVALID_MEM_OBJECT);
      profilingEnd();
      return;
    }

    dstDevMem->syncCacheFromHost(*this, syncFlags);
    srcDevMem->syncCacheFromHost(*this);
  }

  // KernelBlitManager::copyBufferBatch handles the D2D/D2H/H2D/P2P split:
  // D2D copies use copyBuffer (kernel blit), D2H/H2D/P2P use DMA batch
  std::vector<amd::BatchCopyOp> batchOps(copyOps.begin(), copyOps.end());
  if (!blitMgr().copyBufferBatch(batchOps)) {
    LogError("submitBatchCopyMemory: Batch copy failed!");
    result = false;
  }

  // copyBufferBatch records every completion signal it produced. The last one sits at
  // current_id_ and is attached to this command; any signals that are not implicitly
  // ordered behind it (mixed shader + SDMA paths, or multiple SDMA engine groups running
  // in parallel) are queued as external signals. Only then is a barrier needed, to join
  // them into a single completion signal for this command -- otherwise the next command's
  // profilingBegin() clears those external signals and the dependency is lost.
  //
  // When there are no extra signals (e.g. a pure P2P/D2D batch that collapses to one
  // engine-group op), the sole completion signal already represents the whole batch and
  // the next op waits on it through the engine switch, so the barrier is pure overhead.
  if (result && !Barriers().IsExternalSignalListEmpty()) {
    dispatchBarrierPacket(kNopPacketHeader);
  }

  if (!result) {
    LogError("submitBatchCopyMemory failed!");
    cmd.setStatus(CL_OUT_OF_RESOURCES);
  } else {
    // Mark all destinations as written
    for (const auto& op : copyOps) {
      op.dstMemory->signalWrite(&dev());
    }
  }

  profilingEnd();
}

// ================================================================================================
void VirtualGPU::SchedulePinnedMemoryRelease(amd::HostQueue& queue,
                                             std::vector<amd::Memory*> pinned_memory) {
  if (pinned_memory.empty()) {
    return;
  }

  std::unique_ptr<std::vector<amd::Memory*>> pinned_memory_ptr =
      std::make_unique<std::vector<amd::Memory*>>(std::move(pinned_memory));

  amd::Command* marker = new amd::Marker(queue, false);
  if (marker == nullptr) {
    releaseGpuMemoryFence();
    for (amd::Memory* pinned : *pinned_memory_ptr) {
      pinned->release();
    }
    return;
  }

  marker->setCommandEntryScope(amd::Device::kCacheStateIgnore);
  constexpr bool kNonBlockingCallback = false;
  if (!marker->setCallback(CL_COMPLETE, ReleasePinnedMemoryCallback, pinned_memory_ptr.get(),
                           kNonBlockingCallback)) {
    marker->release();
    releaseGpuMemoryFence();
    for (amd::Memory* pinned : *pinned_memory_ptr) {
      pinned->release();
    }
    return;
  }

  marker->enqueue();
  marker->release();
  pinned_memory_ptr.release();
}

// ================================================================================================
void VirtualGPU::SubmitBatchWriteMemory(amd::BatchWriteMemoryCommand& cmd) {
  // Make sure VirtualGPU has an exclusive access to the resources
  std::scoped_lock lock(execution());

  profilingBegin(cmd, true);

  const std::vector<amd::BatchWriteMemoryOp>& write_ops = cmd.WriteOps();

  if (!amd::IS_HIP) {
    device::Memory::SyncFlags sync_flags;
    sync_flags.skipEntire_ = false;

    for (const amd::BatchWriteMemoryOp& op : write_ops) {
      dev().getRocMemory(op.dst_memory)->syncCacheFromHost(*this, sync_flags);
    }
  }

  if (!blitMgr().WriteBufferBatch(write_ops)) {
    LogError("SubmitBatchWriteMemory: Batch write failed!");
    cmd.setStatus(CL_OUT_OF_RESOURCES);
    profilingEnd();
    return;
  }

  if (!amd::IS_HIP) {
    for (const amd::BatchWriteMemoryOp& op : write_ops) {
      op.dst_memory->signalWrite(&dev());
    }
  }

  profilingEnd();
  SchedulePinnedMemoryRelease(*cmd.queue(), cmd.TakePinnedMemory());
}

// ================================================================================================
void VirtualGPU::SubmitBatchReadMemory(amd::BatchReadMemoryCommand& cmd) {
  // Make sure VirtualGPU has an exclusive access to the resources
  std::scoped_lock lock(execution());

  profilingBegin(cmd, true);

  const std::vector<amd::BatchReadMemoryOp>& read_ops = cmd.ReadOps();

  if (!amd::IS_HIP) {
    for (const amd::BatchReadMemoryOp& op : read_ops) {
      dev().getRocMemory(op.src_memory)->syncCacheFromHost(*this);
    }
  }

  if (!blitMgr().ReadBufferBatch(read_ops)) {
    LogError("SubmitBatchReadMemory: Batch read failed!");
    cmd.setStatus(CL_OUT_OF_RESOURCES);
    profilingEnd();
    return;
  }

  profilingEnd();
  SchedulePinnedMemoryRelease(*cmd.queue(), cmd.TakePinnedMemory());
}

// ================================================================================================
void VirtualGPU::submitSvmMapMemory(amd::SvmMapMemoryCommand& cmd) {
  // Make sure VirtualGPU has an exclusive access to the resources
  std::scoped_lock lock(execution());

  profilingBegin(cmd, true);

  // no op for FGS supported device
  if (!dev().isFineGrainedSystem(true) && !dev().forceFineGrain(cmd.getSvmMem())) {
    // Make sure we have memory for the command execution
    Memory* memory = dev().getRocMemory(cmd.getSvmMem());

    memory->saveMapInfo(cmd.svmPtr(), cmd.origin(), cmd.size(), cmd.mapFlags(),
                        cmd.isEntireMemory());

    if (memory->mapMemory() != nullptr) {
      if (cmd.mapFlags() & (CL_MAP_READ | CL_MAP_WRITE)) {
        Memory* hsaMapMemory = dev().getRocMemory(memory->mapMemory());

        if (!blitMgr().copyBuffer(*memory, *hsaMapMemory, cmd.origin(), cmd.origin(), cmd.size(),
                                  cmd.isEntireMemory())) {
          LogError("submitSVMMapMemory() - copy failed");
          cmd.setStatus(CL_MAP_FAILURE);
        }
        // Wait on a kernel if one is outstanding
        releaseGpuMemoryFence();
        const void* mappedPtr = hsaMapMemory->owner()->getHostMem();
        std::memcpy(cmd.svmPtr(), mappedPtr, cmd.size()[0]);
      }
    } else {
      LogError("Unhandled svm map!");
    }
  }

  profilingEnd();
}

// ================================================================================================
void VirtualGPU::submitSvmUnmapMemory(amd::SvmUnmapMemoryCommand& cmd) {
  // Make sure VirtualGPU has an exclusive access to the resources
  std::scoped_lock lock(execution());

  profilingBegin(cmd, true);

  // no op for FGS supported device
  if (!dev().isFineGrainedSystem(true) && !dev().forceFineGrain(cmd.getSvmMem())) {
    Memory* memory = dev().getRocMemory(cmd.getSvmMem());
    const device::Memory::WriteMapInfo* writeMapInfo = memory->writeMapInfo(cmd.svmPtr());

    if (memory->mapMemory() != nullptr) {
      if (writeMapInfo->isUnmapWrite()) {
        // Wait on a kernel if one is outstanding
        releaseGpuMemoryFence();
        amd::Coord3D srcOrigin(0, 0, 0);
        Memory* hsaMapMemory = dev().getRocMemory(memory->mapMemory());

        void* mappedPtr = hsaMapMemory->owner()->getHostMem();
        std::memcpy(mappedPtr, cmd.svmPtr(), writeMapInfo->region_[0]);
        // Target is a remote resource, so copy
        if (!blitMgr().copyBuffer(*hsaMapMemory, *memory, writeMapInfo->origin_,
                                  writeMapInfo->origin_, writeMapInfo->region_,
                                  writeMapInfo->isEntire())) {
          LogError("submitSvmUnmapMemory() - copy failed");
          cmd.setStatus(CL_OUT_OF_RESOURCES);
        }
      }
    } else {
      LogError("Unhandled svm map!");
    }

    memory->clearUnmapInfo(cmd.svmPtr());
  }

  profilingEnd();
}

// ================================================================================================
void VirtualGPU::submitMapMemory(amd::MapMemoryCommand& cmd) {
  // Make sure VirtualGPU has an exclusive access to the resources
  std::scoped_lock lock(execution());

  profilingBegin(cmd, true);

  //! @todo add multi-devices synchronization when supported.

  roc::Memory* devMemory =
      reinterpret_cast<roc::Memory*>(cmd.memory().getDeviceMemory(dev(), false));

  cl_command_type type = cmd.type();
  bool imageBuffer = (cmd.memory().getType() == CL_MEM_OBJECT_IMAGE1D_BUFFER);
  if (imageBuffer) {
    type = CL_COMMAND_MAP_BUFFER;
  }

  // Save map requirement.
  cl_map_flags mapFlag = cmd.mapFlags();

  // Treat no map flag as read-write.
  if (mapFlag == 0) {
    mapFlag = CL_MAP_READ | CL_MAP_WRITE;
  }

  devMemory->saveMapInfo(cmd.mapPtr(), cmd.origin(), cmd.size(), mapFlag, cmd.isEntireMemory());

  // Sync to the map target.
  // If we have host memory, use it
  if ((devMemory->owner()->getHostMem() != nullptr) &&
      (devMemory->owner()->getSvmPtr() == nullptr)) {
    // Target is the backing store, so just ensure that owner is up-to-date
    devMemory->owner()->cacheWriteBack(this);

    if (devMemory->isHostMemDirectAccess()) {
      // Add memory to VA cache, so rutnime can detect direct access to VA
      dev().addVACache(devMemory);
    }
  } else if (devMemory->IsPersistentDirectMap()) {
    // Persistent memory - NOP map
  } else if (mapFlag & (CL_MAP_READ | CL_MAP_WRITE)) {
    bool result = false;
    roc::Memory* hsaMemory = static_cast<roc::Memory*>(devMemory);

    amd::Memory* mapMemory = hsaMemory->mapMemory();
    void* hostPtr =
        mapMemory == nullptr ? hsaMemory->owner()->getHostMem() : mapMemory->getHostMem();

    if (type == CL_COMMAND_MAP_BUFFER) {
      amd::Coord3D origin(cmd.origin()[0]);
      amd::Coord3D size(cmd.size()[0]);
      amd::Coord3D dstOrigin(cmd.origin()[0], 0, 0);
      if (imageBuffer) {
        size_t elemSize = cmd.memory().asImage()->getImageFormat().getElementSize();
        origin.c[0] *= elemSize;
        size.c[0] *= elemSize;
      }

      if (mapMemory != nullptr) {
        roc::Memory* hsaMapMemory =
            static_cast<roc::Memory*>(mapMemory->getDeviceMemory(dev(), false));
        result = blitMgr().copyBuffer(*hsaMemory, *hsaMapMemory, origin, dstOrigin, size,
                                      cmd.isEntireMemory());
        void* svmPtr = devMemory->owner()->getSvmPtr();
        if ((svmPtr != nullptr) && (hostPtr != svmPtr)) {
          // Wait on a kernel if one is outstanding
          releaseGpuMemoryFence();
          std::memcpy(svmPtr, hostPtr, size[0]);
        }
      } else {
        result = blitMgr().readBuffer(*hsaMemory, static_cast<char*>(hostPtr) + origin[0], origin,
                                      size, cmd.isEntireMemory());
      }
    } else if (type == CL_COMMAND_MAP_IMAGE) {
      amd::Image* image = cmd.memory().asImage();
      if (mapMemory != nullptr) {
        roc::Memory* hsaMapMemory =
            static_cast<roc::Memory*>(mapMemory->getDeviceMemory(dev(), false));
        result =
            blitMgr().copyImageToBuffer(*hsaMemory, *hsaMapMemory, cmd.origin(),
                                        amd::Coord3D(0, 0, 0), cmd.size(), cmd.isEntireMemory());
      } else {
        result = blitMgr().readImage(*hsaMemory, hostPtr, amd::Coord3D(0), image->getRegion(),
                                     image->getRowPitch(), image->getSlicePitch(), true);
      }
    } else {
      ShouldNotReachHere();
    }

    if (!result) {
      LogError("submitMapMemory failed!");
      cmd.setStatus(CL_OUT_OF_RESOURCES);
    }
  }

  profilingEnd();
}

// ================================================================================================
void VirtualGPU::submitUnmapMemory(amd::UnmapMemoryCommand& cmd) {
  // Make sure VirtualGPU has an exclusive access to the resources
  std::scoped_lock lock(execution());

  roc::Memory* devMemory = static_cast<roc::Memory*>(cmd.memory().getDeviceMemory(dev(), false));

  const device::Memory::WriteMapInfo* mapInfo = devMemory->writeMapInfo(cmd.mapPtr());
  if (nullptr == mapInfo) {
    LogError("Unmap without map call");
    return;
  }

  profilingBegin(cmd, true);

  // Force buffer write for IMAGE1D_BUFFER
  bool imageBuffer = (cmd.memory().getType() == CL_MEM_OBJECT_IMAGE1D_BUFFER);

  // We used host memory
  if ((devMemory->owner()->getHostMem() != nullptr) &&
      (devMemory->owner()->getSvmPtr() == nullptr)) {
    if (mapInfo->isUnmapWrite()) {
      // Target is the backing store, so sync
      devMemory->owner()->signalWrite(nullptr);
      devMemory->syncCacheFromHost(*this);
    }
    if (devMemory->isHostMemDirectAccess()) {
      // Remove memory from VA cache
      dev().removeVACache(devMemory);
    }
  } else if (devMemory->IsPersistentDirectMap()) {
    // Persistent memory - NOP unmap
  } else if (mapInfo->isUnmapWrite()) {
    // Commit the changes made by the user.
    if (!devMemory->isHostMemDirectAccess()) {
      bool result = false;

      amd::Memory* mapMemory = devMemory->mapMemory();
      if (cmd.memory().asImage() && !imageBuffer) {
        amd::Image* image = cmd.memory().asImage();
        if (mapMemory != nullptr) {
          roc::Memory* hsaMapMemory =
              static_cast<roc::Memory*>(mapMemory->getDeviceMemory(dev(), false));
          result =
              blitMgr().copyBufferToImage(*hsaMapMemory, *devMemory, amd::Coord3D(0, 0, 0),
                                          mapInfo->origin_, mapInfo->region_, mapInfo->isEntire());
        } else {
          void* hostPtr = devMemory->owner()->getHostMem();

          result = blitMgr().writeImage(hostPtr, *devMemory, amd::Coord3D(0), image->getRegion(),
                                        image->getRowPitch(), image->getSlicePitch(), true);
        }
      } else {
        amd::Coord3D origin(mapInfo->origin_[0]);
        amd::Coord3D size(mapInfo->region_[0]);
        amd::Coord3D dstOrigin(mapInfo->origin_[0]);
        if (imageBuffer) {
          size_t elemSize = cmd.memory().asImage()->getImageFormat().getElementSize();
          origin.c[0] *= elemSize;
          size.c[0] *= elemSize;
        }
        if (mapMemory != nullptr) {
          roc::Memory* hsaMapMemory =
              static_cast<roc::Memory*>(mapMemory->getDeviceMemory(dev(), false));

          const void* svmPtr = devMemory->owner()->getSvmPtr();
          void* hostPtr = mapMemory->getHostMem();
          if ((svmPtr != nullptr) && (hostPtr != svmPtr)) {
            // Wait on a kernel if one is outstanding
            releaseGpuMemoryFence();
            std::memcpy(hostPtr, svmPtr, size[0]);
          }
          result = blitMgr().copyBuffer(*hsaMapMemory, *devMemory, origin, dstOrigin, size,
                                        mapInfo->isEntire());
        } else {
          result = blitMgr().writeBuffer(cmd.mapPtr(), *devMemory, origin, size);
        }
      }
      if (!result) {
        LogError("submitMapMemory failed!");
        cmd.setStatus(CL_OUT_OF_RESOURCES);
      }
    }

    cmd.memory().signalWrite(&dev());
  }

  devMemory->clearUnmapInfo(cmd.mapPtr());

  profilingEnd();
}

// ================================================================================================
bool VirtualGPU::fillMemory(cl_command_type type, amd::Memory* amdMemory, const void* pattern,
                            size_t patternSize, const amd::Coord3D& surface,
                            const amd::Coord3D& origin, const amd::Coord3D& size, bool forceBlit) {
  // Make sure VirtualGPU has an exclusive access to the resources
  std::scoped_lock lock(execution());

  Memory* memory = dev().getRocMemory(amdMemory);

  bool entire = amdMemory->isEntirelyCovered(origin, size);
  // Synchronize memory from host if necessary
  device::Memory::SyncFlags syncFlags;
  syncFlags.skipEntire_ = entire;
  memory->syncCacheFromHost(*this, syncFlags);

  bool result = false;
  bool imageBuffer = false;
  float fillValue[4];

  // Force fill buffer for IMAGE1D_BUFFER
  if ((type == CL_COMMAND_FILL_IMAGE) && (amdMemory->getType() == CL_MEM_OBJECT_IMAGE1D_BUFFER)) {
    type = CL_COMMAND_FILL_BUFFER;
    imageBuffer = true;
  }

  // Find the the right fill operation
  switch (type) {
    case CL_COMMAND_SVM_MEMFILL:
    case CL_COMMAND_FILL_BUFFER: {
      amd::Coord3D realSurf(surface[0], surface[1], surface[2]);
      amd::Coord3D realOrigin(origin[0], origin[1], origin[2]);
      amd::Coord3D realSize(size[0], size[1], size[2]);
      // Reprogram fill parameters if it's an IMAGE1D_BUFFER object
      if (imageBuffer) {
        size_t elemSize = amdMemory->asImage()->getImageFormat().getElementSize();
        realOrigin.c[0] *= elemSize;
        realSize.c[0] *= elemSize;
        memset(fillValue, 0, sizeof(fillValue));
        amdMemory->asImage()->getImageFormat().formatColor(pattern, fillValue);
        pattern = fillValue;
        patternSize = elemSize;
      }
      result = blitMgr().fillBuffer(*memory, pattern, patternSize, realSurf, realOrigin, realSize,
                                    entire, forceBlit);
      break;
    }
    case CL_COMMAND_FILL_IMAGE: {
      result = blitMgr().fillImage(*memory, pattern, origin, size, entire);
      break;
    }
    default:
      ShouldNotReachHere();
      break;
  }

  if (!result) {
    LogError("submitFillMemory failed!");
  }

  amdMemory->signalWrite(&dev());
  return true;
}

// ================================================================================================
void VirtualGPU::submitFillMemory(amd::FillMemoryCommand& cmd) {
  // Make sure VirtualGPU has an exclusive access to the resources
  std::scoped_lock lock(execution());

  profilingBegin(cmd);

  bool force_blit = false;
  if (amd::IS_HIP) {
    // Always use blit for memset for HIP.
    force_blit = true;
  }

  if (!fillMemory(cmd.type(), &cmd.memory(), cmd.pattern(), cmd.patternSize(), cmd.surface(),
                  cmd.origin(), cmd.size(), force_blit)) {
    cmd.setStatus(CL_INVALID_OPERATION);
  }
  profilingEnd();
}

// ================================================================================================
void VirtualGPU::submitStreamOperation(amd::StreamOperationCommand& cmd) {
  // Make sure VirtualGPU has an exclusive access to the resources
  std::scoped_lock lock(execution());
  profilingBegin(cmd);

  const cl_command_type type = cmd.type();
  const uint64_t value = cmd.value();
  const uint64_t mask = cmd.mask();
  const unsigned int flags = cmd.flags();
  const size_t sizeBytes = cmd.sizeBytes();
  const size_t offset = cmd.offset();

  amd::Memory* amdMemory = &cmd.memory();
  Memory* memory = dev().getRocMemory(amdMemory);

  if (type == ROCCLR_COMMAND_STREAM_WAIT_VALUE) {
    if (GPU_STREAMOPS_CP_WAIT) {
      uint16_t header = kBarrierVendorPacketHeader;
      Buffer* buff = static_cast<Buffer*>(memory);
      hsa_signal_t signal = buff->getSignal();

      // mask is always applied on value at signal before performing
      // the comparision defiend by 'condition'
      switch (flags) {
        case ROCCLR_STREAM_WAIT_VALUE_GTE: {
          dispatchBarrierValuePacket(header, false, signal, value, mask, HSA_SIGNAL_CONDITION_GTE,
                                     true);
          break;
        }
        case ROCCLR_STREAM_WAIT_VALUE_EQ: {
          dispatchBarrierValuePacket(header, false, signal, value, mask, HSA_SIGNAL_CONDITION_EQ,
                                     true);
          break;
        }
        case ROCCLR_STREAM_WAIT_VALUE_AND: {
          dispatchBarrierValuePacket(header, false, signal, 0, (value & mask),
                                     HSA_SIGNAL_CONDITION_NE, true);
          break;
        }
        case ROCCLR_STREAM_WAIT_VALUE_NOR: {
          uint64_t norValue = ~value & mask;
          dispatchBarrierValuePacket(header, false, signal, norValue, norValue,
                                     HSA_SIGNAL_CONDITION_NE, true);
          break;
        }
        default:
          ShouldNotReachHere();
          break;
      }
    }
    // Use a blit kernel to perform the wait operation
    else {
      // Even though the blit kernel uses an atomic with system scope, we still need to add system scope on
      // the AQLPacket because atomics on kernels can bypass L2 cache on some hardware.
      addSystemScope();

      // mask is applied on value before performing
      // the comparision defined by 'condition'
      bool result = blitMgr().streamOpsWait(*memory, value, offset, sizeBytes, flags, mask);
      ClPrint(amd::LOG_DEBUG, amd::LOG_COPY,
              "Waiting for value: 0x%lx."
              " Flags: 0x%lx mask: 0x%lx",
              value, flags, mask);
      if (!result) {
        LogError("submitStreamOperation: Wait failed!");
      }
    }
  } else if (type == ROCCLR_COMMAND_STREAM_WRITE_VALUE) {
    // Even though the blit kernel uses system scope atomic, we still need to add system scope on
    // the AQLPacket because atomics on kernels can bypass L2 cache on some hardware.
    addSystemScope();

    bool result;
    switch (flags) {
      case ROCCLR_STREAM_WRITE_VALUE_DEFAULT: {
        result = blitMgr().streamOpsWrite(*memory, value, offset, sizeBytes);
        break;
      }
      case ROCCLR_STREAM_WRITE_VALUE_INCREMENT: {
        result = blitMgr().streamOpsIncrement(*memory, value, offset, sizeBytes);
        break;
      }
      case ROCCLR_STREAM_WRITE_VALUE_DECREMENT: {
        result = blitMgr().streamOpsDecrement(*memory, value, offset, sizeBytes);
        break;
      }
      default: {
        ShouldNotReachHere();
        break;
      }
    }

    if (!result) {
      LogError("submitStreamOperation: Write failed!");
    }
  } else {
    ShouldNotReachHere();
  }
  profilingEnd();
}

// ================================================================================================
void VirtualGPU::submitBatchMemoryOperation(amd::BatchMemoryOperationCommand& cmd) {
  // Make sure VirtualGPU has an exclusive access to the resources
  std::scoped_lock lock(execution());
  profilingBegin(cmd);

  bool result = blitMgr().batchMemOps(cmd.getParamPtr(), cmd.paramSize(), cmd.count());
  if (!result) {
    LogError("submitBatchMemoryOperation failed!");
  }
  profilingEnd();
}

// ================================================================================================
void VirtualGPU::submitVirtualMap(amd::VirtualMapCommand& vcmd) {
  // Make sure VirtualGPU has an exclusive access to the resources
  std::scoped_lock lock(execution());

  profilingBegin(vcmd);

  // Find the amd::Memory object for virtual ptr. vcmd.ptr() is vaddr.
  amd::Memory* vaddr_base_obj = amd::MemObjMap::FindVirtualMemObj(vcmd.ptr());
  if (vaddr_base_obj == nullptr || !(vaddr_base_obj->getMemFlags() & CL_MEM_VA_RANGE_AMD)) {
    profilingEnd();
    return;
  }

  // Get the amd::Memory object for the physical address
  amd::Memory* phys_mem_obj = vcmd.memory();
  hsa_status_t hsa_status = HSA_STATUS_SUCCESS;

  // If Physical address is not set, then it is map command. If set, it is unmap command.
  if (phys_mem_obj != nullptr) {
    amd::Memory* vaddr_sub_obj =
        dev().MapMemObjBookkeeping(phys_mem_obj, const_cast<void*>(vcmd.ptr()), vcmd.size());
    if (vaddr_sub_obj == nullptr) {
      LogError("HSA Command: MapMemObjBookkeeping failed!");
      profilingEnd();
      return;
    }
    // Map the physical to virtual address the hsa api
    hsa_amd_vmem_alloc_handle_t opaque_hsa_handle;
    opaque_hsa_handle.handle = phys_mem_obj->getUserData().hsa_handle;
    if ((hsa_status = Hsa::vmem_map(vaddr_sub_obj->getSvmPtr(), vcmd.size(),
                                       vaddr_sub_obj->getOffset(), opaque_hsa_handle, 0)) ==
        HSA_STATUS_SUCCESS) {
      constexpr bool kImportVmmForInterprocess = true;
      dev().FinalizeMapMemObjBookkeeping(vaddr_sub_obj, phys_mem_obj, const_cast<void*>(vcmd.ptr()),
                                         kImportVmmForInterprocess);
      // The VA is now backed, so the true owning agent (a peer device for imported memory)
      // can finally be resolved. create() ran before the mapping existed, so cache it now.
      if (auto* devMem = static_cast<Memory*>(vaddr_sub_obj->getDeviceMemory(dev()))) {
        devMem->refreshOwningAgentFromPointerInfo();
      }
    } else {
      LogError("HSA Command: hsa_amd_vmem_map failed!");
    }
  } else {
    dispatchBarrierPacket(kBarrierPacketHeader, false);
    Barriers().WaitCurrent();

    amd::Memory* vaddr_sub_obj = amd::MemObjMap::FindMemObj(vcmd.ptr());
    assert(vaddr_sub_obj != nullptr);

    // Unmap the object, since the physical addr is set.
    if ((hsa_status = Hsa::vmem_unmap(vaddr_sub_obj->getSvmPtr(), vcmd.size())) ==
        HSA_STATUS_SUCCESS) {
      constexpr bool kDestroyVirtualBuffer = true;
      constexpr bool kReleaseSubObj = true;
      dev().UnmapMemObjBookkeeping(vaddr_sub_obj, const_cast<void*>(vcmd.ptr()),
                                   kDestroyVirtualBuffer, kReleaseSubObj);
    } else {
      LogError("HSA Command: hsa_amd_vmem_unmap failed");
    }
  }

  // Since this is a memory operation, the HW event set for barrier packet
  // may not encapsulate what the command wants to do. Hence clear the hw_event
  constexpr bool kClearHwEvent = true;
  profilingEnd(kClearHwEvent);
}

// ================================================================================================
void VirtualGPU::submitSvmFillMemory(amd::SvmFillMemoryCommand& cmd) {
  // Make sure VirtualGPU has an exclusive access to the resources
  std::scoped_lock lock(execution());

  profilingBegin(cmd);

  amd::Memory* dstMemory = amd::MemObjMap::FindMemObj(cmd.dst());

  if (!dev().isFineGrainedSystem(true) ||
      ((dstMemory != nullptr) && !dev().forceFineGrain(dstMemory))) {
    size_t patternSize = cmd.patternSize();
    size_t fillSize = patternSize * cmd.times();

    size_t offset = reinterpret_cast<uintptr_t>(cmd.dst()) -
                    reinterpret_cast<uintptr_t>(dstMemory->getSvmPtr());

    Memory* memory = dev().getRocMemory(dstMemory);

    amd::Coord3D origin(offset, 0, 0);
    amd::Coord3D size(fillSize, 1, 1);

    assert((dstMemory->validateRegion(origin, size)) && "The incorrect fill size!");

    if (!fillMemory(cmd.type(), dstMemory, cmd.pattern(), cmd.patternSize(), size, origin, size,
                    true)) {
      cmd.setStatus(CL_INVALID_OPERATION);
    }
  } else {
    // Stall GPU for CPU access to memory
    releaseGpuMemoryFence();
    // for FGS capable device, fill CPU memory directly
    amd::SvmBuffer::memFill(cmd.dst(), cmd.pattern(), cmd.patternSize(), cmd.times());
  }

  profilingEnd();
}

// ================================================================================================
void VirtualGPU::submitMigrateMemObjects(amd::MigrateMemObjectsCommand& vcmd) {
  // Make sure VirtualGPU has an exclusive access to the resources
  std::scoped_lock lock(execution());

  profilingBegin(vcmd);

  for (auto itr : vcmd.memObjects()) {
    // Find device memory
    Memory* memory = dev().getRocMemory(&(*itr));

    if (vcmd.migrationFlags() & CL_MIGRATE_MEM_OBJECT_HOST) {
      if (!memory->isHostMemDirectAccess()) {
        // Make sure GPU finished operation before synchronization with the backing store
        releaseGpuMemoryFence();
      }
      memory->mgpuCacheWriteBack(*this);
    } else if (vcmd.migrationFlags() & CL_MIGRATE_MEM_OBJECT_CONTENT_UNDEFINED) {
      // Synchronize memory from host if necessary.
      // The sync function will perform memory migration from
      // another device if necessary
      device::Memory::SyncFlags syncFlags;
      memory->syncCacheFromHost(*this, syncFlags);
    } else {
      LogWarning("Unknown operation for memory migration!");
    }
  }

  profilingEnd();
}

// ================================================================================================
bool VirtualGPU::createSchedulerParam() {
  if (nullptr != schedulerQueue_) {
    return true;
  }

  while (true) {
    // The queue is written by multiple threads of the scheduler kernel
    if (HSA_STATUS_SUCCESS !=
        Hsa::queue_create(gpu_device(), 2048, HSA_QUEUE_TYPE_MULTI, callbackQueue, &roc_device_,
                          std::numeric_limits<uint>::max(), std::numeric_limits<uint>::max(),
                          &schedulerQueue_)) {
      break;
    }

    schedulerDoorbell_ = nullptr;
    if (DEBUG_CLR_DIRECT_DOORBELL) {
      uint64_t db_id = 0;
      if (hsa_amd_queue_get_info(schedulerQueue_, HSA_AMD_QUEUE_INFO_DOORBELL_ID, &db_id) ==
              HSA_STATUS_SUCCESS &&
          db_id != 0) {
        schedulerDoorbell_ = reinterpret_cast<volatile uint64_t*>(db_id);
      }
    }

#if defined(_WIN32)
  if (!isSchedulerQueueThreadRunning()) {
    std::call_once(scheduler_thread_init_, [this]() { startSchedulerQueueThread(); });
  }
#endif  // _WIN32
    return true;
  }

  if (nullptr != schedulerQueue_) {
    Hsa::queue_destroy(schedulerQueue_);
    schedulerQueue_ = nullptr;
  }

  return false;
}

void VirtualGPU::startSchedulerQueueThread() {
  schedulerQueueThreadRunning_.store(true, std::memory_order_release);

  schedulerQueueThread_ = std::thread([this]() {
    uint64_t updated_write_index = 0;
    while (isSchedulerQueueThreadRunning()) {

      // Wait until scheduler events are added or thread termination
      {
        std::unique_lock<std::mutex> lock(scheduler_mutex_);
        scheduler_cv_.wait(lock, [this]() {
          return !pendingSchedulerEvents_.empty() ||
                 !isSchedulerQueueThreadRunning();
        });
      }

      // Actively monitor the scheduler queue while any sync event is pending.
      bool has_active_events = true;
      while (has_active_events && isSchedulerQueueThreadRunning()) {
        uint64_t write_index = Hsa::queue_load_write_index_scacquire(schedulerQueue_);

        if (write_index > updated_write_index) {
          // New packets in the scheduler queue, ringing the doorbell
          if (schedulerDoorbell_) {
            amd::ringDoorbell(schedulerDoorbell_, write_index - 1);
          } else {
            Hsa::signal_store_screlease(schedulerQueue_->doorbell_signal, write_index - 1);
          }
          updated_write_index = write_index;
        } else {
          // Yield briefly before re-checking.
          amd::Os::yield();
        }
        // Check all scheduler completion signals, remove the completed ones
        {
          std::lock_guard<std::mutex> lock(scheduler_mutex_);
          pendingSchedulerEvents_.erase(
            std::remove_if(pendingSchedulerEvents_.begin(),
                           pendingSchedulerEvents_.end(),
                           [](hsa_signal_t signal) {
                             return (Hsa::signal_load_relaxed(signal) == 0);
                           }),
                    pendingSchedulerEvents_.end());
          has_active_events = !pendingSchedulerEvents_.empty();
        }
      }
      // All scheduler completion signals completed, go back to wait
    }
  });
}

// ================================================================================================
uint64_t VirtualGPU::getVQVirtualAddress() {
  Memory* vqMem = dev().getRocMemory(virtualQueue_);
  return reinterpret_cast<uint64_t>(vqMem->getDeviceMemory());
}

// ================================================================================================
bool VirtualGPU::createVirtualQueue(uint deviceQueueSize) {
  uint MinDeviceQueueSize = 16 * 1024;
  deviceQueueSize = std::max(deviceQueueSize, MinDeviceQueueSize);

  maskGroups_ = deviceQueueSize / (512 * Ki);
  maskGroups_ = (maskGroups_ == 0) ? 1 : maskGroups_;

  // Align the queue size for the multiple dispatch scheduler.
  // Each thread works with 32 entries * maskGroups
  uint extra = deviceQueueSize % (sizeof(AmdAqlWrap) * DeviceQueueMaskSize * maskGroups_);
  if (extra != 0) {
    deviceQueueSize += (sizeof(AmdAqlWrap) * DeviceQueueMaskSize * maskGroups_) - extra;
  }

  if (deviceQueueSize_ == deviceQueueSize) {
    return true;
  } else {
    if (0 != deviceQueueSize_) {
      virtualQueue_->release();
      virtualQueue_ = nullptr;
      deviceQueueSize_ = 0;
      schedulerThreads_ = 0;
    }
  }

  uint numSlots = deviceQueueSize / sizeof(AmdAqlWrap);
  uint allocSize = deviceQueueSize;

  // Add the virtual queue header
  allocSize += sizeof(AmdVQueueHeader);
  allocSize = amd::alignUp(allocSize, sizeof(AmdAqlWrap));

  uint argOffs = allocSize;

  // Add the kernel arguments and wait events
  uint singleArgSize = amd::alignUp(
      dev().info().maxParameterSize_ + 64 + dev().settings().numWaitEvents_ * sizeof(uint64_t),
      sizeof(AmdAqlWrap));
  allocSize += singleArgSize * numSlots;

  uint eventsOffs = allocSize;
  // Add the device events
  allocSize += dev().settings().numDeviceEvents_ * sizeof(AmdEvent);

  uint eventMaskOffs = allocSize;
  // Add mask array for events
  allocSize += amd::alignUp(dev().settings().numDeviceEvents_, DeviceQueueMaskSize) / 8;

  uint slotMaskOffs = allocSize;
  // Add mask array for AmdAqlWrap slots
  allocSize += amd::alignUp(numSlots, DeviceQueueMaskSize) / 8;

  // Align size to 64 bytes for more efficient fill operation
  allocSize = amd::alignUp(allocSize, 8 * sizeof(uint64_t));

  // CL_MEM_ALLOC_HOST_PTR/CL_MEM_READ_WRITE
  virtualQueue_ = new (dev().context()) amd::Buffer(dev().context(), CL_MEM_READ_WRITE, allocSize);

  if ((nullptr != virtualQueue_) && !virtualQueue_->create(nullptr)) {
    virtualQueue_->release();
    return false;
  }

  Memory* vqMem = dev().getRocMemory(virtualQueue_);

  if (nullptr == vqMem) {
    return false;
  }

  uint64_t vqVA = reinterpret_cast<uint64_t>(vqMem->getDeviceMemory());

  // Use shadow to prepare the data structure in host.
  auto shadow = std::make_unique<uint8_t[]>(allocSize);

  std::memset(&shadow[0], 0, allocSize);

  AmdVQueueHeader* header = reinterpret_cast<AmdVQueueHeader*>(&shadow[0]);
  // Initialize the virtual queue header
  header->aql_slot_num = numSlots;
  header->event_slot_num = dev().settings().numDeviceEvents_;
  header->event_slot_mask = vqVA + eventMaskOffs;
  header->event_slots = vqVA + eventsOffs;
  header->aql_slot_mask = vqVA + slotMaskOffs;
  header->wait_size = dev().settings().numWaitEvents_;
  header->arg_size = dev().info().maxParameterSize_ + 64;
  header->mask_groups = maskGroups_;

  // Go over all slots and perform initialization
  size_t offset = sizeof(AmdVQueueHeader);
  for (uint i = 0; i < numSlots; ++i) {
    AmdAqlWrap* slot = reinterpret_cast<AmdAqlWrap*>(&shadow[0] + offset);
    uint64_t argStart = vqVA + argOffs + i * singleArgSize;

    slot->aql.kernarg_address = reinterpret_cast<void*>(argStart);
    slot->wait_list = argStart + dev().info().maxParameterSize_ + 64;

    offset += sizeof(AmdAqlWrap);
  }

  amd::Coord3D origin(0, 0, 0);
  amd::Coord3D region(allocSize, 1, 1);

  // copy the data structure from host to GPU
  if (!dev().xferMgr().writeBuffer(&shadow[0], *vqMem, origin, region)) {
    return false;
  }

  deviceQueueSize_ = deviceQueueSize;
  schedulerThreads_ = numSlots / (DeviceQueueMaskSize * maskGroups_);

  return true;
}


void VirtualGPU::HiddenHeapInit() {
  // We don't really need its id, just want to ensure the queue is created.
  (void)getQueueID();
  const_cast<Device&>(dev()).HiddenHeapInit(*this);
}

// ================================================================================================
bool VirtualGPU::submitKernelInternal(const amd::NDRangeContainer& sizes, const amd::Kernel& kernel,
                                      const_address parameters, void* event_handle,
                                      uint32_t sharedMemBytes, amd::NDRangeKernelCommand* vcmd,
                                      hsa_kernel_dispatch_packet_t* aql_packet,
                                      bool attach_signal) {
  device::Kernel* devKernel = const_cast<device::Kernel*>(kernel.getDeviceKernel(dev()));
  Kernel& gpuKernel = static_cast<Kernel&>(*devKernel);
  size_t ldsUsage = gpuKernel.WorkgroupGroupSegmentByteSize();
  bool imageBufferWrtBack = false;                  // Image buffer write back is required
  std::vector<device::Memory*> wrtBackImageBuffer;  // Array of images for write back

  // Check memory dependency and SVM objects
  bool coopGroups = (vcmd != nullptr) ? vcmd->cooperativeGroups() : false;
  if (!processMemObjects(kernel, parameters, ldsUsage, coopGroups, imageBufferWrtBack,
                         wrtBackImageBuffer)) {
    LogError("Wrong memory objects!");
    return false;
  }

  // Init PrintfDbg object if printf is enabled.
  bool printfEnabled = (gpuKernel.printfInfo().size() > 0) ? true : false;
  if (!printfDbg()->init(printfEnabled)) {
    LogError("\nPrintfDbg object initialization failed!");
    return false;
  }

  const amd::KernelSignature& signature = kernel.signature();
  const amd::KernelParameters& kernelParams = kernel.parameters();

  bool isGraphCapture = command_ != nullptr && command_->getPktCapturingState();

  ClPrint(amd::LOG_INFO, amd::LOG_KERN2, "ShaderName : %s", gpuKernel.getDemangledName().c_str());
  ClPrint(amd::LOG_DETAIL_DEBUG, amd::LOG_KERN,
          "argSize = %zu, KernargSegmentByteSize = %zu, KernargSegmentAlignment = %zu",
          std::min(gpuKernel.KernargSegmentByteSize(), signature.paramsSize()),
          gpuKernel.KernargSegmentByteSize(), gpuKernel.KernargSegmentAlignment());

  amd::NDRange local_size(sizes.local());
  address hidden_arguments = const_cast<address>(parameters);
  // Calculate local size if it wasn't provided
  devKernel->FindLocalWorkSize(sizes.dimensions(), sizes.global(), local_size);

  uint16_t local[3] = {1, 1, 1};
  uint32_t global[3] = {1, 1, 1};
  for (uint i = 0; i < sizes.dimensions(); i++) {
    global[i] = static_cast<uint32_t>(sizes.global()[i]);
    local[i] = static_cast<uint16_t>(local_size[i]);
  }
  uint64_t spVA = 0;
  // Check if runtime has to setup hidden arguments
  for (uint32_t i = signature.numParameters(); i < signature.numParametersAll(); ++i) {
    const auto& it = signature.at(i);
    switch (it.info_.oclObject_) {
      case amd::KernelParameterDescriptor::HiddenNone:
        break;
      case amd::KernelParameterDescriptor::HiddenGlobalOffsetX: {
        WriteAqlArgAt(hidden_arguments, sizes.offset()[0], it.size_, it.offset_);
        break;
      }
      case amd::KernelParameterDescriptor::HiddenGlobalOffsetY: {
        if (sizes.dimensions() >= 2) {
          WriteAqlArgAt(hidden_arguments, sizes.offset()[1], it.size_, it.offset_);
        }
        break;
      }
      case amd::KernelParameterDescriptor::HiddenGlobalOffsetZ: {
        if (sizes.dimensions() >= 3) {
          WriteAqlArgAt(hidden_arguments, sizes.offset()[2], it.size_, it.offset_);
        }
        break;
      }
      case amd::KernelParameterDescriptor::HiddenPrintfBuffer: {
        uintptr_t bufferPtr = reinterpret_cast<uintptr_t>(printfDbg()->dbgBuffer());
        if (printfEnabled && bufferPtr) {
          WriteAqlArgAt(hidden_arguments, bufferPtr, it.size_, it.offset_);
        }
        break;
      }
      case amd::KernelParameterDescriptor::HiddenHostcallBuffer: {
        if (amd::IS_HIP) {
          if (dev().info().pcie_atomics_) {
            uintptr_t buffer = reinterpret_cast<uintptr_t>(getOrCreateHostcallBuffer());
            if (!buffer) {
              LogError("Kernel expects a hostcall buffer, but none found");
              return false;
            }
            WriteAqlArgAt(hidden_arguments, buffer, it.size_, it.offset_);
          } else {
            LogError("Pcie atomics not enabled, hostcall not supported");
            return false;
          }
        }
        break;
      }
      case amd::KernelParameterDescriptor::HiddenDefaultQueue: {
        uint64_t vqVA = 0;
        amd::DeviceQueue* defQueue = kernel.program().context().defDeviceQueue(dev());
        if (nullptr != defQueue && devKernel->dynamicParallelism()) {
          if (!createVirtualQueue(defQueue->size()) || !createSchedulerParam()) {
            return false;
          }
          vqVA = getVQVirtualAddress();
        }
        WriteAqlArgAt(hidden_arguments, vqVA, it.size_, it.offset_);
        break;
      }
      case amd::KernelParameterDescriptor::HiddenCompletionAction: {
        if (devKernel->dynamicParallelism()) {
          auto params = allocKernArg(sizeof(AmdAqlWrap), 64);
          AmdAqlWrap* wrap = reinterpret_cast<AmdAqlWrap*>(params);
          memset(wrap, 0, sizeof(AmdAqlWrap));
          wrap->state = AQL_WRAP_DONE;
          spVA = reinterpret_cast<uint64_t>(wrap);
        }
        WriteAqlArgAt(hidden_arguments, spVA, it.size_, it.offset_);
        break;
      }
      case amd::KernelParameterDescriptor::HiddenMultiGridSync: {
        bool multiGridSync = (vcmd != nullptr) ? vcmd->cooperativeMultiDeviceGroups() : false;
        bool singleGridSync = (vcmd != nullptr) ? vcmd->cooperativeGroups() : false;
        Device::MGSyncInfo* syncInfo = nullptr;
        if (multiGridSync) {
          // Find CPU pointer to the right sync info structure. It should be after MGSyncData
          syncInfo = reinterpret_cast<Device::MGSyncInfo*>(
              dev().MGSync() + Device::kMGInfoSizePerDevice * dev().index() +
              Device::kMGSyncDataSize);
          // Update sync data address. Use the offset adjustment to the right location
          syncInfo->mgs = reinterpret_cast<Device::MGSyncData*>(
              dev().MGSync() + Device::kMGInfoSizePerDevice * vcmd->firstDevice());
        } else if (singleGridSync) {
          syncInfo = reinterpret_cast<Device::MGSyncInfo*>(allocKernArg(Device::kSGInfoSize, 64));
          syncInfo->mgs = nullptr;
        }
        if (multiGridSync || singleGridSync) {
          // Update sync data address.
          syncInfo->sgs = {0};
          // Fill rest of sync info fields
          syncInfo->grid_id = vcmd->gridId();
          syncInfo->num_grids = vcmd->numGrids();
          syncInfo->prev_sum = vcmd->prevGridSum();
          syncInfo->all_sum = vcmd->allGridSum();
          syncInfo->num_wg = vcmd->numWorkgroups();
        }
        // Update GPU address for grid sync info. Use the offset adjustment for the right
        // location
        WriteAqlArgAt(hidden_arguments, reinterpret_cast<uint64_t>(syncInfo), it.size_, it.offset_);
        break;
      }
      case amd::KernelParameterDescriptor::HiddenHeap:
        // Allocate hidden heap for HIP applications only
        if ((amd::IS_HIP) && (dev().HeapBuffer() == nullptr)) {
          const_cast<Device&>(dev()).HiddenHeapAlloc(*this);
        }
        if (dev().HeapBuffer() != nullptr) {
          // Initialize hidden heap buffer
          if (!isGraphCapture) {
            const_cast<Device&>(dev()).HiddenHeapInit(*this);
            if (!heap_init_fence_emitted_) {
              addSystemScope();
              heap_init_fence_emitted_ = true;
            }
          }
          // Add heap pointer to the code
          size_t heap_ptr = static_cast<size_t>(dev().HeapBuffer()->virtualAddress());
          WriteAqlArgAt(hidden_arguments, heap_ptr, it.size_, it.offset_);
        }
        break;
      case amd::KernelParameterDescriptor::HiddenBlockCountX:
        WriteAqlArgAt(hidden_arguments, global[0] / local[0], it.size_, it.offset_);
        break;
      case amd::KernelParameterDescriptor::HiddenBlockCountY:
        WriteAqlArgAt(hidden_arguments, global[1] / local[1], it.size_, it.offset_);
        break;
      case amd::KernelParameterDescriptor::HiddenBlockCountZ:
        WriteAqlArgAt(hidden_arguments, global[2] / local[2], it.size_, it.offset_);
        break;
      case amd::KernelParameterDescriptor::HiddenGroupSizeX:
        WriteAqlArgAt(hidden_arguments, local[0], it.size_, it.offset_);
        break;
      case amd::KernelParameterDescriptor::HiddenGroupSizeY:
        WriteAqlArgAt(hidden_arguments, local[1], it.size_, it.offset_);
        break;
      case amd::KernelParameterDescriptor::HiddenGroupSizeZ:
        WriteAqlArgAt(hidden_arguments, local[2], it.size_, it.offset_);
        break;
      case amd::KernelParameterDescriptor::HiddenRemainderX:
        WriteAqlArgAt(hidden_arguments, static_cast<uint16_t>(global[0] % local[0]), it.size_,
                      it.offset_);
        break;
      case amd::KernelParameterDescriptor::HiddenRemainderY:
        if (sizes.dimensions() >= 2) {
          WriteAqlArgAt(hidden_arguments, static_cast<uint16_t>(global[1] % local[1]), it.size_,
                        it.offset_);
        }
        break;
      case amd::KernelParameterDescriptor::HiddenRemainderZ:
        if (sizes.dimensions() >= 3) {
          WriteAqlArgAt(hidden_arguments, static_cast<uint16_t>(global[2] % local[2]), it.size_,
                        it.offset_);
        }
        break;
      case amd::KernelParameterDescriptor::HiddenGridDims:
        WriteAqlArgAt(hidden_arguments, static_cast<uint16_t>(sizes.dimensions()), it.size_,
                      it.offset_);
        break;
      case amd::KernelParameterDescriptor::HiddenPrivateBase:
        WriteAqlArgAt(hidden_arguments,
                      reinterpret_cast<amd_queue_t*>(gpu_queue_)->private_segment_aperture_base_hi,
                      it.size_, it.offset_);
        break;
      case amd::KernelParameterDescriptor::HiddenSharedBase:
        WriteAqlArgAt(hidden_arguments,
                      reinterpret_cast<amd_queue_t*>(gpu_queue_)->group_segment_aperture_base_hi,
                      it.size_, it.offset_);
        break;
      case amd::KernelParameterDescriptor::HiddenQueuePtr:
        WriteAqlArgAt(hidden_arguments, gpu_queue_, it.size_, it.offset_);
        break;
      case amd::KernelParameterDescriptor::HiddenDynamicLdsSize:
        WriteAqlArgAt(hidden_arguments, sharedMemBytes, it.size_, it.offset_);
        break;
    }
  }
  address argBuffer = hidden_arguments;
  size_t argSize = std::min(gpuKernel.KernargSegmentByteSize(), signature.paramsSize());

  // Find all parameters for the current kernel
  if (!kernel.parameters().deviceKernelArgs() || gpuKernel.isInternalKernel()) {
    // Allocate buffer to hold kernel arguments
    if (isGraphCapture) {
      argBuffer = command_->getGraphKernArg(gpuKernel.KernargSegmentByteSize(),
                                            gpuKernel.KernargSegmentAlignment(), dev().index());
      command_->SetKernelName(gpuKernel.getDemangledName());
    } else {
      argBuffer = reinterpret_cast<address>(
          allocKernArg(gpuKernel.KernargSegmentByteSize(), gpuKernel.KernargSegmentAlignment()));
    }

    amd::nontemporalMemcpy(argBuffer, parameters, argSize);
    // When the queue ring buffer is in device memory, kernarg NT stores and
    // the dispatch packet all travel to the same PCIe endpoint.  PCIe posted
    // write ordering guarantees the kernargs arrive before the dispatch
    // header, so no HDP flush or readback workaround is needed — the sfence
    // before packet_store_release is sufficient.
    if (roc_device_.info().largeBar_ && !isGraphCapture && !device_mem_ring_buf_) {
      const auto kernArgImpl = dev().settings().kernel_arg_impl_;
      if (kernArgImpl == KernelArgImpl::DeviceKernelArgsReadback && argSize != 0) {
#if defined(ATI_ARCH_X86)
        _mm_sfence();
#else
        __atomic_thread_fence(__ATOMIC_SEQ_CST);
#endif
        *(argBuffer + argSize - 1) = *(parameters + argSize - 1);
#if defined(ATI_ARCH_X86)
        _mm_mfence();
#else
        __atomic_thread_fence(__ATOMIC_SEQ_CST);
#endif
        auto kSentinel = *reinterpret_cast<volatile unsigned char*>(argBuffer + argSize - 1);
      }
    }
  }
  // Check for group memory overflow
  //! @todo Check should be in HSA - here we should have at most an assert
  assert(dev().info().localMemSizePerCU_ > 0);
  if (ldsUsage > dev().info().localMemSizePerCU_) {
    LogError("No local memory available\n");
    return false;
  }

      // Initialize the dispatch Packet
    static_assert(sizeof(hsa_kernel_dispatch_packet_t)
                  == sizeof(hsa_amd_ext_kernel_dispatch_packet_t));

    alignas(64) union {
      hsa_kernel_dispatch_packet_t kernelDispatch;
      hsa_amd_ext_kernel_dispatch_packet_t extKernelDispatch;
    } dispatchPacketUnion;

    auto& dispatchPacket = dispatchPacketUnion.kernelDispatch;
    memset(&dispatchPacket, 0, sizeof(dispatchPacket));

    uint32_t newGlobalSize[3] = {global[0], global[1], global[2]};

    dispatchPacket.header = kInvalidAql;
    dispatchPacket.kernel_object = gpuKernel.KernelCodeHandle();

    // dispatchPacket.header = aqlHeader_;
    // dispatchPacket.setup |= sizes.dimensions() << HSA_KERNEL_DISPATCH_PACKET_SETUP_DIMENSIONS;

    bool extDispatchPacket =
        (sizes.dimensions() > 0 && sizes.cluster()[0] > 1) ||
        (sizes.dimensions() > 1 && sizes.cluster()[1] > 1) ||
        (sizes.dimensions() > 2 && sizes.cluster()[2] > 1) ||
        dev().settings().ext_dispatch_packet_;

    if (extDispatchPacket) {
      auto& dispatchPacketExt = dispatchPacketUnion.extKernelDispatch;

      dispatchPacketExt.cluster_size_x = sizes.dimensions() > 0 ? sizes.cluster()[0] : 1;
      dispatchPacketExt.cluster_size_y = sizes.dimensions() > 1 ? sizes.cluster()[1] : 1;
      dispatchPacketExt.cluster_size_z = sizes.dimensions() > 2 ? sizes.cluster()[2] : 1;

      // Already validated in HIP Launch Params that newGlobalSize is perfectly divisible by local
      // and it is divisible by cluster size.
      dispatchPacketExt.cluster_count_x = sizes.dimensions() > 0
                                          ? (newGlobalSize[0] / local[0] / sizes.cluster()[0]) : 1;
      dispatchPacketExt.cluster_count_y = sizes.dimensions() > 1
                                          ? (newGlobalSize[1] / local[1] / sizes.cluster()[1]) : 1;
      dispatchPacketExt.cluster_count_z = sizes.dimensions() > 2
                                          ? (newGlobalSize[2] / local[2] / sizes.cluster()[2]) : 1;

    } else {
      dispatchPacket.grid_size_x = sizes.dimensions() > 0 ? newGlobalSize[0] : 1;
      dispatchPacket.grid_size_y = sizes.dimensions() > 1 ? newGlobalSize[1] : 1;
      dispatchPacket.grid_size_z = sizes.dimensions() > 2 ? newGlobalSize[2] : 1;
    }

    if (dev().settings().groupMemCarveout_) {
      uint8_t percent = devKernel->workGroupInfo()->groupMemCarveout_
          ? devKernel->workGroupInfo()->groupMemCarveout_
          : dev().GetGroupMemCarveout();
      auto& dispatchPacketExt = dispatchPacketUnion.extKernelDispatch;
      // Encodings [1, 127] represent a range from 0% (no group memory) to 100% (maximum
      // group memory)
      if (dev().isa().versionMajor() == 12 && dev().isa().versionMinor() == 5) {
        dispatchPacketExt.perf_hint.group_mem_carveout = 127;
      } else {
        dispatchPacketExt.perf_hint.group_mem_carveout = (percent + 1) * 1.26F;
      }
    }

    dispatchPacket.workgroup_size_x = sizes.dimensions() > 0 ? local[0] : 1;
    dispatchPacket.workgroup_size_y = sizes.dimensions() > 1 ? local[1] : 1;
    dispatchPacket.workgroup_size_z = sizes.dimensions() > 2 ? local[2] : 1;

    dispatchPacket.kernarg_address = argBuffer;
    dispatchPacket.group_segment_size = ldsUsage + sharedMemBytes;
    dispatchPacket.private_segment_size = devKernel->workGroupInfo()->privateMemSize_;

    if ((devKernel->workGroupInfo()->usedStackSize_ & 0x1) == 0x1) {
      dispatchPacket.private_segment_size =
          std::max<uint64_t>(dev().StackSize(), dispatchPacket.private_segment_size);
      const size_t maxStackSize = dev().MaxStackSize();
      // we return an explicit error when we exceed the max stack size limit
      if (dispatchPacket.private_segment_size > maxStackSize) {
        LogPrintfError("Scratch size (%u) exceeds max allowed (%zu) for kernel : %s",
                       dispatchPacket.private_segment_size, maxStackSize,
                       gpuKernel.getDemangledName().c_str());
        return false;
      }
    }

    // Pass the header accordingly
    auto aqlHeaderWithOrder = aqlHeader_;

  if (vcmd != nullptr) {
    if (vcmd->getAnyOrderLaunchFlag()) {
      constexpr uint32_t kAqlHeaderMask = ~(1 << HSA_PACKET_HEADER_BARRIER);
      aqlHeaderWithOrder &= kAqlHeaderMask;
    }
    if (vcmd->getCommandEntryScope() == amd::Device::kCacheStateSystem) {
      addSystemScope_ = true;
    }
  }

  // Copy scheduler's AQL packet for possible relaunch from the scheduler itself
    if (aql_packet != nullptr) {
      *aql_packet = dispatchPacket;
      if (extDispatchPacket) {
        aql_packet->header = (HSA_PACKET_TYPE_VENDOR_SPECIFIC << HSA_PACKET_HEADER_TYPE) |
                              (1 << HSA_PACKET_HEADER_BARRIER) |
                              (HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_ACQUIRE_FENCE_SCOPE) |
                              (HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_RELEASE_FENCE_SCOPE);
        // For an ext-dispatch vendor packet byte+2 is amd_format, not setup; set
        // amd_format=EXT_KERNEL_DISPATCH and shift setup (dimensions) into byte+3.
        aql_packet->setup = static_cast<uint16_t>(HSA_AMD_PACKET_TYPE_EXT_KERNEL_DISPATCH
                              | ((sizes.dimensions()
                                  << HSA_KERNEL_DISPATCH_PACKET_SETUP_DIMENSIONS) << 8));
      } else {
        aql_packet->header = (HSA_PACKET_TYPE_KERNEL_DISPATCH << HSA_PACKET_HEADER_TYPE) |
                              (1 << HSA_PACKET_HEADER_BARRIER) |
                              (HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_ACQUIRE_FENCE_SCOPE) |
                              (HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_RELEASE_FENCE_SCOPE);
        aql_packet->setup = static_cast<uint16_t>(sizes.dimensions()
                              << HSA_KERNEL_DISPATCH_PACKET_SETUP_DIMENSIONS);
      }
    }


    uint16_t rest = 0;
    if (extDispatchPacket) {
      // When launching an AQL packet, the 32 bits has to be written atomically for CP to track,
      // on normal dispatch packet, first 32 bits are header & setup. In ext dispatch packet,
      // the first 32 bits are header, amd_format, setup. Update the "rest" of the 32 bits, so we
      // can commit it atomically in packet_store_release.
      rest = (HSA_AMD_PACKET_TYPE_EXT_KERNEL_DISPATCH
              | ((sizes.dimensions() << HSA_KERNEL_DISPATCH_PACKET_SETUP_DIMENSIONS) << 8));
    } else {
      rest = (sizes.dimensions() << HSA_KERNEL_DISPATCH_PACKET_SETUP_DIMENSIONS);
    }

    metadata_preloader_.PrepareDispatch(gpuKernel.MetadataKernelDescriptor(),
                                        gpuKernel.MetadataPreloadLength(),
                                        gpuKernel.MetadataPreloadOffset());

    bool dispatched = extDispatchPacket
        ? dispatchAqlPacket(&dispatchPacketUnion.extKernelDispatch, aqlHeaderWithOrder, rest,
                            GPU_FLUSH_ON_EXECUTION, attach_signal)
        : dispatchAqlPacket(&dispatchPacket, aqlHeaderWithOrder, rest,
                            GPU_FLUSH_ON_EXECUTION, attach_signal);
    if (!dispatched) {
      return false;
    }

  // Output printf buffer
  if (!printfDbg()->output(*this, printfEnabled, gpuKernel.printfInfo())) {
    LogError("\nCould not print data from the printf buffer!");
    return false;
  }

  if (gpuKernel.dynamicParallelism()) {
    dispatchBarrierPacket(kBarrierPacketHeader, true);
    if (virtualQueue_ != nullptr) {
      static_cast<KernelBlitManager&>(blitMgr()).runScheduler(
          getVQVirtualAddress(), schedulerQueue_, schedulerThreads_, spVA);
    }
  }

  // Check if image buffer write back is required
  if (imageBufferWrtBack) {
    // Make sure the original kernel execution is done
    releaseGpuMemoryFence();
    for (const auto imageBuffer : wrtBackImageBuffer) {
      Memory* buffer = dev().getGpuMemory(imageBuffer->owner()->parent());
      amd::Image* image = imageBuffer->owner()->asImage();
      Image* devImage = static_cast<Image*>(dev().getGpuMemory(imageBuffer->owner()));
      Memory* cpyImage = dev().getGpuMemory(devImage->CopyImageBuffer());
      amd::Coord3D offs(0);
      // Copy memory from the the backing store image into original buffer
      bool result = blitMgr().copyImageToBuffer(*cpyImage, *buffer, offs, offs, image->getRegion(),
                                                true, image->getRowPitch(), image->getSlicePitch());
    }
  }
  return true;
}

/**
 * @brief Api to dispatch a kernel for execution. The implementation
 * parses the input object, an instance of virtual command to obtain
 * the parameters of global size, work group size, offsets of work
 * items, enable/disable profiling, etc.
 *
 * It also parses the kernel arguments buffer to inject into Hsa Runtime
 * the list of kernel parameters.
 */
// ================================================================================================
void VirtualGPU::submitKernel(amd::NDRangeKernelCommand& vcmd) {
  if (vcmd.cooperativeGroups()) {
    // Wait for the execution on the current queue, since the coop groups will use the device queue
    releaseGpuMemoryFence(kSkipCpuWait);

    // Get device queue for exclusive GPU access
    VirtualGPU* queue = dev().xferQueue();
    if (!queue) {
      LogError("Runtime failed to acquire a cooperative queue!");
      vcmd.setStatus(CL_INVALID_OPERATION);
      return;
    }

    // Lock the queue, using the blit manager lock
    std::scoped_lock k(*(queue->blitMgr().lockXfer()));

    queue->profilingBegin(vcmd);

    // Add a dependency into the device queue on the current queue
    queue->Barriers().AddExternalSignal(Barriers().GetLastSignal());

    if (dev().settings().gwsInitSupported_ == true) {
      uint32_t workgroups = vcmd.numWorkgroups();
      static_cast<KernelBlitManager&>(queue->blitMgr()).RunGwsInit(workgroups - 1);
    }

    // Sync AQL packets
    queue->setAqlHeader(dispatchPacketHeader_);

    // Submit kernel to HW
    if (!queue->submitKernelInternal(vcmd.sizes(), vcmd.kernel(), vcmd.parameters(),
                                     static_cast<void*>(as_cl(&vcmd.event())),
                                     vcmd.sharedMemBytes(), &vcmd)) {
      LogError("AQL dispatch failed!");
      vcmd.setStatus(CL_INVALID_OPERATION);
    }
    // Wait for the execution on the device queue. Keep the current queue in-order
    queue->releaseGpuMemoryFence(kSkipCpuWait);

    // Add a dependency into the current queue on the coop queue
    Barriers().AddExternalSignal(queue->Barriers().GetLastSignal());
    hasPendingDispatch_ = true;
    retainExternalSignals_ = true;

    queue->profilingEnd();
  } else {
    // Make sure VirtualGPU has an exclusive access to the resources
    std::scoped_lock lock(execution());

    profilingBegin(vcmd);

    if (vcmd.dynDataPrefetchConfig().isEnabled()) {
      metadata_preloader_.SetDynDataPrefetchRegions(vcmd.dynDataPrefetchConfig());
    }

    // Submit kernel to HW
    if (!submitKernelInternal(vcmd.sizes(), vcmd.kernel(), vcmd.parameters(),
                              static_cast<void*>(as_cl(&vcmd.event())), vcmd.sharedMemBytes(),
                              &vcmd)) {
      LogError("AQL dispatch failed!");
      vcmd.setStatus(CL_INVALID_OPERATION);
    }

    metadata_preloader_.ClearDynDataPrefetchConfig();

    profilingEnd();
  }
}

// ================================================================================================
void VirtualGPU::submitNativeFn(amd::NativeFnCommand& cmd) {}

// ================================================================================================
void VirtualGPU::submitMarker(amd::Marker& vcmd) {
  // Make sure VirtualGPU has an exclusive access to the resources
  std::scoped_lock lock(execution());
  if (vcmd.CpuWaitRequested()) {
    force_irq_ = IS_WINDOWS;
    // It should be safe to call flush directly if there are not pending dispatches without
    // HSA signal callback
    if (!dedicated_queue_ && gpu_queue_ == nullptr) {
      SetGpuQueue(roc_device_.AcquireActiveQueue(priority_));
    }
    flush(vcmd.GetBatchHead());
    SetCoalesceWindow(0, nullptr);
  } else {
    profilingBegin(vcmd);
    const Settings& settings = dev().settings();
    hsa_signal_t ipc_s{0};
    if (vcmd.ipcCompletionSignal() != nullptr) {
      ipc_s.handle = static_cast<uint64_t>(
          reinterpret_cast<uintptr_t>(vcmd.ipcCompletionSignal()->getHandle()));
    }

    if (vcmd.ipcDepSignal() != nullptr) {
      hsa_signal_t s;
      s.handle = static_cast<uint64_t>(reinterpret_cast<uintptr_t>(
          vcmd.ipcDepSignal()->getHandle()));
      WaitCompleteSignal(s);
    } else if (timestamp_ != nullptr || vcmd.profilingInfo().marker_ts_ || ipc_s.handle != 0) {
      // Skip a redundant barrier when this marker records the same client event as
      // the previous barrier with no intervening work. IPC records always dispatch.
      if (ipc_s.handle == 0 && ShouldCoalesceMarker(vcmd)) {
        // Reuse the prior barrier's completion signal so hipEventQuery/Synchronize
        // observe correct readiness even though no new barrier was dispatched.
        AttachHwEvent(command_, last_barrier_hw_event_);
        profilingEnd();
        force_irq_ = false;
        return;
      }
      // IPC event record: if ipc_s is non-zero, first dispatch a NOP barrier with
      // the IPC signal as completion_signal; it fires when prior work completes.
      if (ipc_s.handle != 0) {
        if (settings.barrier_value_packet_) {
          dispatchBarrierValuePacket(kBarrierVendorPacketNopScopeHeader, false, hsa_signal_t{0}, 0,
                                     0, HSA_SIGNAL_CONDITION_EQ, true, ipc_s);
        } else {
          dispatchBarrierPacket(kNopPacketHeader, true, ipc_s);
        }
      }

      int32_t releaseFlags = vcmd.getCommandEntryScope();
      if (releaseFlags == Device::CacheState::kCacheStateIgnore) {
        if (settings.barrier_value_packet_ && vcmd.profilingInfo().marker_ts_) {
          dispatchBarrierValuePacket(kBarrierVendorPacketNopScopeHeader, true);
        } else {
          dispatchBarrierPacket(kNopPacketHeader, false);
        }
      } else {
        // Submit a barrier with a cache flushes.
        force_irq_ = IS_WINDOWS;
        if (settings.barrier_value_packet_ && vcmd.profilingInfo().marker_ts_) {
          dispatchBarrierValuePacket(kBarrierVendorPacketHeader, true);
        } else {
          dispatchBarrierPacket(kBarrierPacketHeader, false);
        }
        hasPendingDispatch_ = false;
      }
    }
    // A record sets the window to its own event and barrier signal; a wait/other
    // marker has a zero coalesceEvent() and so clears it, which a wait relies on.
    uint64_t ev = vcmd.coalesceEvent();
    SetCoalesceWindow(ev, ev != 0 ? command_->HwEvent() : nullptr);
    profilingEnd();
  }
  force_irq_ = false;
}

// ================================================================================================
void VirtualGPU::submitAccumulate(amd::AccumulateCommand& vcmd) {
  // Make sure VirtualGPU has an exclusive access to the resources
  std::scoped_lock lock(execution());
  profilingBegin(vcmd);

  // Register pre-patched HW event signals with the Timestamp for profiling.
  // These signals were configured by ApplyHwEventPatches (isPacketDispatch_,
  // done_ flags set there) but bypass ActiveSignal, so they must be added
  // here so checkGpuTime → ExtractSignalTiming → addTimestamps picks them up.
  if (timestamp_ != nullptr) {
    for (const auto& [_, events] : vcmd.getHwEvents()) {
      for (void* hw_event : events) {
        auto* ps = reinterpret_cast<ProfilingSignal*>(hw_event);
        if (ps != nullptr) {
          timestamp_->AddProfilingSignal(ps);
        }
      }
    }
  }

  const Settings& settings = dev().settings();
  if (settings.barrier_value_packet_) {
    dispatchBarrierValuePacket(kBarrierVendorPacketNopScopeHeader, true);
  } else {
    dispatchBarrierPacket(kNopPacketHeader, false);
  }

  profilingEnd();
}

// ================================================================================================
void VirtualGPU::submitExternalSemaphoreCmd(amd::ExternalSemaphoreCmd& cmd) {
  // Serialize queue access like the other submit* paths.
  std::scoped_lock lock(execution());

  // Unwrap the CLR-visible handle; null guard for holder/queue.
  const auto* holder =
      static_cast<const hsa_amd_external_semaphore_t*>(cmd.sem_ptr());
  if (holder == nullptr || gpu_queue_ == nullptr) {
    cmd.setStatus(CL_INVALID_OPERATION);
    return;
  }

  // A failed HSA call is logged but does not fail the command: some KMDs reject
  // GPU-side waits even when signal works, and higher-level sync gates the data
  // dependency. The barriers reproduce the needed cross-engine ordering.
  if (cmd.semaphoreCmd() == amd::ExternalSemaphoreCmd::COMMAND_SIGNAL_EXTSEMAPHORE) {
    // Leading barrier: drain prior work (incl. async SDMA via dep_signal[])
    // before the signal so external observers see it last.
    dispatchBarrierPacket(kBarrierPacketHeader);

    hsa_status_t s =
        hsa_amd_queue_signal_external_semaphore(gpu_queue_, *holder, cmd.fence());
    if (s != HSA_STATUS_SUCCESS) {
      LogError("Failed to signal external semaphore");
    }
  } else {
    hsa_status_t s =
        hsa_amd_queue_wait_external_semaphore(gpu_queue_, *holder, cmd.fence());
    if (s != HSA_STATUS_SUCCESS) {
      LogError("Failed to wait on external semaphore");
    } else {
      // Trailing acquire barrier (successful wait only): orders later
      // cross-engine deps after the wait and makes producer writes visible.
      dispatchBarrierPacket(kBarrierPacketAcquireHeader);
    }
  }
}

// ================================================================================================
void VirtualGPU::submitAcquireExtObjects(amd::AcquireExtObjectsCommand& vcmd) {
  // Make sure VirtualGPU has an exclusive access to the resources
  std::scoped_lock lock(execution());

  profilingBegin(vcmd);
  addSystemScope();

  for (auto& mem : vcmd.getMemList()) {
    amd::InteropObject* interop = mem->getInteropObj();
    if (!interop) continue;
    // Copy data from the original D3D resource to the shared intermediate resource
    // (no-op for GL and for resources that are already directly shared)
    interop->copyOrigToShared();
  }

  profilingEnd();
}

// ================================================================================================
void VirtualGPU::submitReleaseExtObjects(amd::ReleaseExtObjectsCommand& vcmd) {
  // Make sure VirtualGPU has an exclusive access to the resources
  std::scoped_lock lock(execution());
  profilingBegin(vcmd);

  for (auto& mem : vcmd.getMemList()) {
    amd::InteropObject* interop = mem->getInteropObj();
    if (!interop) continue;
    // Copy data from the shared intermediate resource back to the original D3D resource
    interop->copySharedToOrig();
  }

  profilingEnd();
}

// ================================================================================================
void VirtualGPU::flush(amd::Command* list, bool wait) {
  // If barrier is requested, then wait for everything, otherwise
  // a per disaptch wait will occur later in updateCommandsState()
  releaseGpuMemoryFence();
  updateCommandsState(list);
}

// ================================================================================================
void VirtualGPU::addPinnedMem(amd::Memory* mem) {
  // Optimize pinning path for Linux only(KFD has special tracking for pinned memory) or OpenCL,
  // since OpenCL always waits for completion on CPU
  if ((command_ != nullptr) && (IS_LINUX || !amd::IS_HIP)) {
    command_->AddPinnedMemory(mem);
  } else {
    //! @note: ROCr backend doesn't have per resource busy tracking, hence runtime has to wait
    //!        unconditionally, before it can release pinned memory
    releaseGpuMemoryFence();
    mem->release();
  }
}

// ================================================================================================
void VirtualGPU::enableSyncBlit() const { blitMgr_->enableSynchronization(); }

// ================================================================================================
void VirtualGPU::submitPerfCounter(amd::PerfCounterCommand& vcmd) {
  // Make sure VirtualGPU has an exclusive access to the resources
  std::scoped_lock lock(execution());

  const amd::PerfCounterCommand::PerfCounterList counters = vcmd.getCounters();

  if (vcmd.getState() == amd::PerfCounterCommand::Begin) {
    // Create a profile for the profiling AQL packet
    PerfCounterProfile* profileRef = new PerfCounterProfile(roc_device_);
    if (profileRef == nullptr || !profileRef->Create()) {
      LogError("Failed to create performance counter profile");
      vcmd.setStatus(CL_INVALID_OPERATION);
      return;
    }

    // Make sure all performance counter objects to use the same profile
    PerfCounter* counter = nullptr;
    for (uint i = 0; i < vcmd.getNumCounters(); ++i) {
      amd::PerfCounter* amdCounter = static_cast<amd::PerfCounter*>(counters[i]);
      counter = static_cast<PerfCounter*>(amdCounter->getDeviceCounter());

      if (nullptr == counter) {
        amd::PerfCounter::Properties prop = amdCounter->properties();
        PerfCounter* rocCounter = new PerfCounter(roc_device_, prop[CL_PERFCOUNTER_GPU_BLOCK_INDEX],
                                                  prop[CL_PERFCOUNTER_GPU_COUNTER_INDEX],
                                                  prop[CL_PERFCOUNTER_GPU_EVENT_INDEX]);

        if (nullptr == rocCounter || rocCounter->gfxVersion() == PerfCounter::ROC_UNSUPPORTED) {
          LogError("Failed to create the performance counter");
          vcmd.setStatus(CL_INVALID_OPERATION);
          delete rocCounter;
          return;
        }

        amdCounter->setDeviceCounter(rocCounter);
        counter = rocCounter;
      }

      counter->setProfile(profileRef);
    }

    if (!profileRef->initialize()) {
      LogError("Failed to initialize performance counter");
      vcmd.setStatus(CL_INVALID_OPERATION);
    } else if (profileRef->createStartPacket() == nullptr) {
      LogError("Failed to create AQL packet for start profiling");
      vcmd.setStatus(CL_INVALID_OPERATION);
    } else {
      dispatchCounterAqlPacket(profileRef->prePacket(), counter->gfxVersion(), false,
                               profileRef->api());
    }

    profileRef->release();
  } else if (vcmd.getState() == amd::PerfCounterCommand::End) {
    // Since all performance counters should use the same profile, use the 1st
    // one to get the profile object
    amd::PerfCounter* amdCounter = static_cast<amd::PerfCounter*>(counters[0]);
    PerfCounter* counter = static_cast<PerfCounter*>(amdCounter->getDeviceCounter());
    if (counter == nullptr) {
      LogError("Invalid Performance Counter");
      vcmd.setStatus(CL_INVALID_OPERATION);
      return;
    }
    PerfCounterProfile* profileRef = counter->profileRef();

    // create the AQL packet for stop profiling
    if (profileRef->createStopPacket() == nullptr) {
      LogError("Failed to create AQL packet for stop profiling");
      vcmd.setStatus(CL_INVALID_OPERATION);
    }
    dispatchCounterAqlPacket(profileRef->postPacket(), counter->gfxVersion(), true,
                             profileRef->api());
  } else {
    LogError("Unsupported performance counter state");
    vcmd.setStatus(CL_INVALID_OPERATION);
  }
}

// ================================================================================================
void *VirtualGPU::getOrCreateHostcallBuffer() {
  if (hostcallBuffer_ != nullptr) {
    return hostcallBuffer_;
  }

  // The number of packets required in each buffer is at least equal to the
  // maximum number of waves supported by the device.
  auto wavesPerCu =
      dev().info().maxThreadsPerCU_ / dev().info().wavefrontWidth_;
  auto numPackets = dev().info().maxComputeUnits_ * wavesPerCu;

  auto size = amd::getHostcallBufferSize(numPackets);
  auto align = amd::getHostcallBufferAlignment();

  hostcallBuffer_ = dev().hostAlloc(
      size, align, Device::MemorySegment::kAtomics, nullptr, false);
  if (!hostcallBuffer_) {
    ClPrint(amd::LOG_ERROR, amd::LOG_QUEUE, "Failed to create hostcall buffer");
    return nullptr;
  }
  hostcallBufferSize_ = size;

#ifdef USE_NEW_HOSTCALL_IMPL
  uint32_t occupiedWords = (numPackets + 31) / 32;
  size_t occupiedSize = occupiedWords * sizeof(uint32_t);
  amd::Buffer* occupiedBuf =
      new (dev().context()) amd::Buffer(dev().context(), CL_MEM_READ_WRITE, occupiedSize, nullptr);
  if (occupiedBuf == nullptr || !occupiedBuf->create()) {
    ClPrint(amd::LOG_ERROR, amd::LOG_QUEUE,
            "Failed to allocate occupied bitfield for hostcall buffer");
    if (occupiedBuf != nullptr) {
      occupiedBuf->release();
    }
    dev().hostFree(hostcallBuffer_, hostcallBufferSize_);
    hostcallBuffer_ = nullptr;
    hostcallBufferSize_ = 0;
    return nullptr;
  }

  device::Memory* occupiedMem = occupiedBuf->getDeviceMemory(dev());
  if (occupiedMem == nullptr || occupiedMem->virtualAddress() == 0) {
    ClPrint(amd::LOG_ERROR, amd::LOG_QUEUE,
            "Failed to get device memory for hostcall occupied bitfield");
    occupiedBuf->release();
    dev().hostFree(hostcallBuffer_, hostcallBufferSize_);
    hostcallBuffer_ = nullptr;
    hostcallBufferSize_ = 0;
    return nullptr;
  }

  uint32_t zero = 0;
  // xferMgr() is synchronous, so the fill completes before any kernel reads it.
  if (!dev().xferMgr().fillBuffer(*occupiedMem, &zero, sizeof(zero), amd::Coord3D(occupiedSize, 1, 1),
                                  amd::Coord3D(0, 0, 0), amd::Coord3D(occupiedSize, 1, 1), true)) {
    ClPrint(amd::LOG_ERROR, amd::LOG_QUEUE,
            "Failed to zero-initialize hostcall occupied bitfield");
    occupiedBuf->release();
    dev().hostFree(hostcallBuffer_, hostcallBufferSize_);
    hostcallBuffer_ = nullptr;
    hostcallBufferSize_ = 0;
    return nullptr;
  }

#endif  // USE_NEW_HOSTCALL_IMPL
  ClPrint(amd::LOG_INFO, amd::LOG_QUEUE,
          "Created hostcall buffer %p (numPackets == %d, size == %d, align == "
          "%d) for virtual "
          "queue %p\n",
          hostcallBuffer_, numPackets, size, align, this);

#ifdef USE_NEW_HOSTCALL_IMPL
  if (!amd::enableHostcalls(dev(), hostcallBuffer_, numPackets, occupiedBuf)) {
    ClPrint(amd::LOG_ERROR, amd::LOG_QUEUE,
            "Failed to register hostcall buffer %p with listener",
            hostcallBuffer_);
    occupiedBuf->release();
    dev().hostFree(hostcallBuffer_, hostcallBufferSize_);
    hostcallBuffer_ = nullptr;
    hostcallBufferSize_ = 0;
    return nullptr;
  }
#else  // !USE_NEW_HOSTCALL_IMPL
  if (!amd::enableHostcalls(dev(), hostcallBuffer_, numPackets)) {
    ClPrint(amd::LOG_ERROR, amd::LOG_QUEUE,
            "Failed to register hostcall buffer %p with listener",
            hostcallBuffer_);
    dev().hostFree(hostcallBuffer_, hostcallBufferSize_);
    hostcallBuffer_ = nullptr;
    hostcallBufferSize_ = 0;
    return nullptr;
  }
#endif  // USE_NEW_HOSTCALL_IMPL
  return hostcallBuffer_;
}

// ================================================================================================
static void convertDynDataPrefetchToHsa(const amd::DynDataPrefetchRegion* regions,
                                        uint8_t hints,
                                        amd_data_prefetch_t* hw,
                                        uint32_t numRegions) {
  for (uint32_t i = 0; i < numRegions && i < amd::kDynDataPrefetchMaxRegions; ++i) {
    const auto& r = regions[i];

    uintptr_t addr = reinterpret_cast<uintptr_t>(r.baseAddress);
    // addr_lo is VA[31:8] (256B aligned), addr_hi is VA[56:32]
    hw[i].addr_lo = static_cast<uint32_t>((addr >> 8) & 0xFFFFFFu);
    hw[i].addr_hi = static_cast<uint32_t>((addr >> 32) & 0x1FFFFFFu);
    // 256B is the hardware prefetch engine request unit
    hw[i].burst_size = static_cast<uint32_t>((r.burstSize / 256u) - 1u);
    hw[i].num_burst = r.numBursts - 1u;
    hw[i].stride = static_cast<uint32_t>(r.stride / 256u);
    hw[i].cooperative = 1;
    hw[i].temporal = hints & 0x3u;
    hw[i].scope = 2; // DEVICE
    hw[i].mode = 1; // ABSOLUTE_VA
  }
}

// ================================================================================================
// Fill the body (all fields except the 4 header dwords) of a kernel-dispatch
// metadata packet from |aql| and the staged descriptor/preload state, returning
// the header dword the caller must publish into header0..header3.  Pure field
// assembly with no ring-buffer transfer or fences — callers (SetPacket for the
// queue write, CaptureMetadata for graph capture) handle the destination.
uint32_t VirtualGPU::MetaDataPreloader::FillKernelDispatchMetadata(
    hsa_kernel_dispatch_packet_t* aql, uint16_t header,
    hsa_amd_metadata_kernel_dispatch_packet_t* m, bool target_is_zeroed) {
  assert(pending_descriptor_ != nullptr);
  {
    // Headers must remain HSA_PACKET_TYPE_INVALID until we publish valid metadata headers
    // at the end. Only clear fields that could otherwise contain stale required-zero data.
    // When staging, the stack-local is already zero-initialized so these are unnecessary.
    if (!target_is_zeroed) {
      std::memset(m->reserved0, 0, sizeof(m->reserved0));
      std::memset(&m->launch_descriptor, 0, sizeof(m->launch_descriptor));
    }

    if (dyn_data_prefetch_enabled_ && dyn_data_prefetch_num_regions_ > 0) {
      m->launch_descriptor.version = launch_descriptor_version_;
      convertDynDataPrefetchToHsa(
          dyn_data_prefetch_regions_,
          dyn_data_prefetch_hints_,
          m->launch_descriptor.prefetch,
          dyn_data_prefetch_num_regions_);
    }

    // Write event_id from amd_signal_t directly (non-interrupt signals have event_id == 0).
    if (aql->completion_signal.handle) {
      auto* signal = reinterpret_cast<amd_signal_t*>(aql->completion_signal.handle);
      m->event_id = signal->event_id;
    } else {
      m->event_id = 0;
    }

    // Fill hsa_amd_metadata_kernel_dispatch_packet->kernel_descriptor fields.
    // The metadata packet kernel descriptor fields are a subset of
    // kernel_descriptor_t (Code Object V3 Kernel Descriptor) from the AQL packet, from bytes
    // llvm::amdhsa::KERNEL_CODE_ENTRY_BYTE_OFFSET_OFFSET(16) to sizeof(kernel_descriptor_t).
    // See include/llvm/Support/AMDHSAKernelDescriptor.h.
    std::memcpy(&m->kernel_descriptor, pending_descriptor_,
                sizeof(m->kernel_descriptor));

    // Fill hsa_amd_metadata_kernel_dispatch_packet->kernarg_preload_* fields
    constexpr uint16_t kPreload_limit =
        (sizeof(m->kernarg_preload_0_14) +
         sizeof(m->kernarg_preload_15_29) +
         sizeof(m->kernarg_preload_30_31)) / sizeof(uint32_t);

    uint16_t preload_length = pending_preload_length_;
    if (preload_length > kPreload_limit) {
      m->kernel_descriptor.kernarg_preload.length = kPreload_limit;
      preload_length = kPreload_limit;
    }
    ClPrint(amd::LOG_DEBUG, amd::LOG_AQL,
            "metadata prefetch: preload_length=%u, preload_offset=%u, preload_limit=%u",
            preload_length, pending_preload_offset_, kPreload_limit);

    if (preload_length > 0) {
      const uint8_t* kernargs = reinterpret_cast<const uint8_t*>(aql->kernarg_address);
      assert(kernargs);
      // Kernarg preload offset is in DWORDs
      const uint8_t* src = kernargs + pending_preload_offset_ * sizeof(uint32_t);
      uint16_t remain = preload_length;

      constexpr uint16_t kSec0 = sizeof(m->kernarg_preload_0_14) / sizeof(uint32_t);
      constexpr uint16_t kSec1 = sizeof(m->kernarg_preload_15_29) / sizeof(uint32_t);
      constexpr uint16_t kSec2 = sizeof(m->kernarg_preload_30_31) / sizeof(uint32_t);

      // Copy the prefix first, then zero only the tail — avoids double-writing
      // the prefix that a memset(whole)+memcpy(prefix) pattern caused. When
      // target_is_zeroed (staging buffer is already zero-initialized), skip the
      // tail zeroing entirely.
      uint16_t n = std::min<uint16_t>(remain, kSec0);
      std::memcpy(m->kernarg_preload_0_14, src, n * sizeof(uint32_t));
      if (!target_is_zeroed && n < kSec0) {
        std::memset(&m->kernarg_preload_0_14[n], 0, (kSec0 - n) * sizeof(uint32_t));
      }
      remain -= n;

      if (remain > 0) {
        src += n * sizeof(uint32_t);
        n = std::min<uint16_t>(remain, kSec1);
        std::memcpy(m->kernarg_preload_15_29, src, n * sizeof(uint32_t));
        if (!target_is_zeroed && n < kSec1) {
          std::memset(&m->kernarg_preload_15_29[n], 0, (kSec1 - n) * sizeof(uint32_t));
        }
        remain -= n;

        if (remain > 0) {
          src += n * sizeof(uint32_t);
          n = std::min<uint16_t>(remain, kSec2);
          std::memcpy(m->kernarg_preload_30_31, src, n * sizeof(uint32_t));
          if (!target_is_zeroed && n < kSec2) {
            std::memset(&m->kernarg_preload_30_31[n], 0, (kSec2 - n) * sizeof(uint32_t));
          }
        }
      }
    }

    return GetType(header) | metadata_version_header_;
  }
}

// ================================================================================================
// Unified metadata write for kernel dispatch packets.  Body assembly is shared
// via FillKernelDispatchMetadata; only the final transfer to the ring buffer
// differs:
//
// (A) Device memory (WC over PCIe BAR): stage locally, then copy 4 x 64B
//     segments via MOVDIR64B (if available) or NT stores.  Each segment's
//     body + header are in the same atomic/NT copy — no sfence between them.
//     Avoids dozens of scattered WC stores over the bus.
//
// (B) System memory (WB): write fields directly into the ring buffer (all
//     stores hit L1 cache).  SFENCE between body and headers because the
//     CP's prefetcher may observe WB memory via a non-cached path.
//
// The ROCR allocator guarantees the metadata ring buffer is 4KiB-aligned with
// 256B packets, so every slot is naturally 64B-aligned.
void VirtualGPU::MetaDataPreloader::SetPacket(
    hsa_kernel_dispatch_packet_t* aql,  uint16_t header,
    hsa_amd_metadata_kernel_dispatch_packet_t* metadata,
    bool use_movdir64b, bool device_mem_ring_buf) {
  assert(pending_descriptor_ != nullptr);

  // Staging is only needed for device-memory (WC) ring buffers to avoid
  // scattered WC stores over PCIe.  For system-memory (WB) queues, direct
  // writes hit L1 cache and are fast — no staging overhead needed.
  const bool use_staging = device_mem_ring_buf;

  if (use_staging) {
    alignas(64) hsa_amd_metadata_kernel_dispatch_packet_t stg = {};
    uint32_t metadata_header = FillKernelDispatchMetadata(aql, header, &stg, true);

    // Write headers into the staging buffer before copying out.
    stg.header0 = metadata_header;
    stg.header1 = metadata_header;
    stg.header2 = metadata_header;
    stg.header3 = metadata_header;

    assert(reinterpret_cast<uintptr_t>(metadata) % 64 == 0 &&
           "metadata ring buffer slot must be 64-byte aligned");
    static_assert(sizeof(stg) % 64 == 0);
    auto* s = reinterpret_cast<const char*>(&stg);
    auto* d = reinterpret_cast<char*>(metadata);

    if (use_movdir64b) {
      for (size_t off = 0; off < sizeof(stg); off += 64) {
        amd::movdir64b_copy64(d + off, s + off);
      }
    } else {
      for (size_t off = 0; off < sizeof(stg); off += 64) {
        amd::nontemporalMemcpy(d + off, s + off, 64);
      }
    }
  } else {
    uint32_t metadata_header = FillKernelDispatchMetadata(aql, header, metadata, false);

    // Write headers last — arms the metadata packet for the CP.
    // SFENCE between body and headers: the metadata ring buffer may be WC-mapped
    // (device memory) and the CP's prefetcher reads it independently. Without a
    // fence, WC reordering could make valid headers visible before the body is
    // flushed, causing the prefetcher to read garbage. This fence ensures all
    // body writes are ordered before the header writes (ordering point #7).
    amd::nontemporalStoreFence();
    metadata->header3 = metadata_header;
    metadata->header2 = metadata_header;
    metadata->header1 = metadata_header;
    metadata->header0 = metadata_header;
  }
}
}  // End of roc namespace