// Copyright 2019 The Chromium Authors
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.

#include "gpu/command_buffer/service/dawn_platform.h"

#include <mutex>
#include <string_view>

#include "base/compiler_specific.h"
#include "base/containers/fixed_flat_map.h"
#include "base/metrics/histogram.h"
#include "base/metrics/histogram_functions.h"
#include "base/rand_util.h"
#include "base/strings/strcat.h"
#include "base/strings/string_util.h"
#include "base/synchronization/waitable_event.h"
#include "base/task/post_job.h"
#include "base/task/thread_pool.h"
#include "base/time/time.h"
#include "base/trace_event/trace_event.h"
#include "gpu/command_buffer/service/dawn_caching_interface.h"
#include "gpu/config/gpu_finch_features.h"
#include "gpu/vulkan/buildflags.h"
#include "ui/gl/progress_reporter.h"

#if BUILDFLAG(ENABLE_VULKAN)
#include "gpu/vulkan/vulkan_util.h"
#endif

namespace gpu::webgpu {

namespace {

class AsyncWaitableEventImpl
    : public base::RefCountedThreadSafe<AsyncWaitableEventImpl> {
 public:
  AsyncWaitableEventImpl()
      : waitable_event_(base::WaitableEvent::ResetPolicy::MANUAL,
                        base::WaitableEvent::InitialState::NOT_SIGNALED) {}

  void Wait() { waitable_event_.Wait(); }

  bool IsComplete() { return waitable_event_.IsSignaled(); }

  void MarkAsComplete() { waitable_event_.Signal(); }

 private:
  friend class base::RefCountedThreadSafe<AsyncWaitableEventImpl>;
  ~AsyncWaitableEventImpl() = default;

  base::WaitableEvent waitable_event_;
};

class AsyncWaitableEvent : public dawn::platform::WaitableEvent {
 public:
  AsyncWaitableEvent()
      : waitable_event_impl_(base::MakeRefCounted<AsyncWaitableEventImpl>()) {}
  ~AsyncWaitableEvent() override = default;

  void Wait() override { waitable_event_impl_->Wait(); }

  bool IsComplete() override { return waitable_event_impl_->IsComplete(); }

  scoped_refptr<AsyncWaitableEventImpl> GetWaitableEventImpl() const {
    return waitable_event_impl_;
  }

 private:
  scoped_refptr<AsyncWaitableEventImpl> waitable_event_impl_;
};

class AsyncJobHandle : public dawn::platform::JobHandle {
 public:
  explicit AsyncJobHandle(base::JobHandle&& job_handle)
      : job_handle_(std::move(job_handle)) {}
  ~AsyncJobHandle() override = default;

  // Chromium's JobHandle can either be cancelled or joined, but not both (it
  // crashes if you try to |Join| after |Cancel|). The Dawn platform's JobHandle
  // allows both to be called without error, so we wrap it here and avoid
  // calling them more than once.
  void Cancel() override {
    std::call_once(once_flag_, [&]() { job_handle_.Cancel(); });
  }
  void Join() override {
    std::call_once(once_flag_, [&]() { job_handle_.Join(); });
  }

 private:
  std::once_flag once_flag_;
  base::JobHandle job_handle_;
};

class AsyncWorkerTaskPool : public dawn::platform::WorkerTaskPool {
 public:
  explicit AsyncWorkerTaskPool(gl::ProgressReporter* progress_reporter)
      : progress_reporter_(progress_reporter) {}
  ~AsyncWorkerTaskPool() override = default;

  std::unique_ptr<dawn::platform::WaitableEvent> PostWorkerTask(
      dawn::platform::PostWorkerTaskCallback callback,
      void* user_data) override {
    std::unique_ptr<AsyncWaitableEvent> waitable_event =
        std::make_unique<AsyncWaitableEvent>();
    base::ThreadPool::PostTask(
        FROM_HERE, {base::MayBlock(), base::TaskPriority::USER_BLOCKING},
        base::BindOnce(&RunWorkerTask, callback, user_data, progress_reporter_,
                       waitable_event->GetWaitableEventImpl()));
    return waitable_event;
  }

  std::unique_ptr<dawn::platform::JobHandle> PostWorkerJob(
      dawn::platform::PostWorkerJobCallback callback,
      void* user_data) override {
    std::unique_ptr<AsyncJobHandle> job_handle =
        std::make_unique<AsyncJobHandle>(base::PostJob(
            FROM_HERE, {base::TaskPriority::USER_VISIBLE},
            base::BindRepeating(&RunWorkerJob, callback, user_data),
            base::BindRepeating([](size_t) -> size_t { return 1; })));
    return job_handle;
  }

 private:
  static void RunWorkerTask(
      dawn::platform::PostWorkerTaskCallback callback,
      void* user_data,
      gl::ProgressReporter* progress_reporter,
      scoped_refptr<AsyncWaitableEventImpl> waitable_event_impl) {
    TRACE_EVENT0("toplevel", "DawnPlatformImpl::RunWorkerTask");
    callback(user_data);
    if (progress_reporter) {
      progress_reporter->ReportProgress();
    }
    waitable_event_impl->MarkAsComplete();
  }

  static void RunWorkerJob(dawn::platform::PostWorkerJobCallback callback,
                           void* user_data,
                           base::JobDelegate* delegate) {
    dawn::platform::JobStatus status = dawn::platform::JobStatus::Continue;
    while (!delegate->ShouldYield()) {
      if (status != dawn::platform::JobStatus::Continue) {
        return;
      }
      status = callback(user_data);
    }
  }

  const raw_ptr<gl::ProgressReporter> progress_reporter_;
};

void RecordDelayedUMA(scoped_refptr<DawnPlatform::CacheCountsMap> cache_map,
                      std::string uma_prefix) {
  base::AutoLock autolock(cache_map->lock);
  for (auto const& [base_name, cache_counts] : cache_map->counts) {
    // Report Hit counts
    base::UmaHistogramCounts10000(
        uma_prefix + base_name + "CacheHit.Counts.90SecondsPostStartup",
        cache_counts.cache_hit_count);

    // Report Miss counts
    base::UmaHistogramCounts10000(
        uma_prefix + base_name + "CacheMiss.Counts.90SecondsPostStartup",
        cache_counts.cache_miss_count);

    // Report Percentage
    int total_counts =
        cache_counts.cache_hit_count + cache_counts.cache_miss_count;
    if (total_counts > 0) {
      int hit_percentage = (cache_counts.cache_hit_count * 100) / total_counts;

      base::UmaHistogramPercentage(
          uma_prefix + base_name + "CacheHit.Percentage.90SecondsPostStartup",
          hit_percentage);
    }
  }
}

// Some metrics are subsampled as they are on critical path.
bool ShouldRecordMetric(std::string_view name) {
  static constexpr double kSubsamplingProb = 0.01;
  return name == "Vulkan.VkQueueSubmitUS"
             ? base::ShouldRecordSubsampledMetric(kSubsamplingProb)
             : true;
}

#if BUILDFLAG(ENABLE_VULKAN) && BUILDFLAG(IS_ANDROID)
// A collection for which a unified metric is emitted for Ganesh/Graphite Vulkan
// backends.
constexpr auto kUnifiedSkiaMetrics =
    base::MakeFixedFlatMap<std::string_view, void (*)(base::TimeDelta)>(
        // Also emitted from CreateGraphicsPipelinesHook with same name.
        {{"Vulkan.CreateGraphicsPipelines.CacheHit",
          &gpu::EmitVkCreateGraphicsPipelinesUMA},
         {"Vulkan.CreateGraphicsPipelines.CacheMiss",
          &gpu::EmitVkCreateGraphicsPipelinesUMA},
         {"Vulkan.VkQueueSubmitUS", &gpu::EmitVkQueueSubmitUMA}});

bool ShouldEmitUnifiedHistogram(const std::string& uma_prefix,
                                const char* name) {
  if (uma_prefix == "GPU.GraphiteDawn." && kUnifiedSkiaMetrics.contains(name)) {
    return true;
  }
  return false;
}
#endif  // BUILDFLAG(ENABLE_VULKAN) && BUILDFLAG(IS_ANDROID)

void EmitUnifiedHistogram(const std::string& uma_prefix,
                          const char* name,
                          int sample) {
#if BUILDFLAG(ENABLE_VULKAN) && BUILDFLAG(IS_ANDROID)
  if (!ShouldEmitUnifiedHistogram(uma_prefix, name)) {
    return;
  }
  kUnifiedSkiaMetrics.at(name)(base::Microseconds(sample));
#endif  // BUILDFLAG(ENABLE_VULKAN) && BUILDFLAG(IS_ANDROID)
}

}  // anonymous namespace

DawnPlatform::CacheCountsMap::CacheCountsMap() = default;
DawnPlatform::CacheCountsMap::~CacheCountsMap() = default;

DawnPlatform::DawnPlatform(
    std::unique_ptr<DawnCachingInterface> dawn_caching_interface,
    gl::ProgressReporter* progress_reporter,
    const char* uma_prefix,
    bool record_cache_count_uma)
    : dawn_caching_interface_(std::move(dawn_caching_interface)),
      progress_reporter_(progress_reporter),
      uma_prefix_(uma_prefix),
      cache_map_(base::MakeRefCounted<CacheCountsMap>()),
      startup_time_(base::TimeTicks::Now()) {
  if (record_cache_count_uma) {
    base::ThreadPool::PostDelayedTask(
        FROM_HERE, base::BindOnce(&RecordDelayedUMA, cache_map_, uma_prefix_),
        base::Seconds(90));
  }
}

DawnPlatform::~DawnPlatform() = default;

double DawnPlatform::MonotonicallyIncreasingTime() {
  return (base::TimeTicks::Now() - base::TimeTicks()).InSecondsF();
}

void DawnPlatform::HistogramCacheCountHelper(std::string_view name,
                                             int sample,
                                             int min,
                                             int max,
                                             int bucketCount) {
  if (name.find("Cache") != std::string_view::npos) {
    base::AutoLock autolock(cache_map_->lock);
    std::string_view base_name = name;
    bool is_hit = false;
    size_t pos = base_name.find("CacheHit");
    if (pos != std::string_view::npos) {
      base_name = base_name.substr(0, pos);
      is_hit = true;
    } else {
      pos = base_name.find("CacheMiss");
      if (pos != std::string_view::npos) {
        base_name = base_name.substr(0, pos);
      }
    }

    auto& cache_counts = cache_map_->counts[base_name];
    if (is_hit) {
      ++cache_counts.cache_hit_count;
    } else {
      ++cache_counts.cache_miss_count;
    }

    if (base::TimeTicks::Now() - startup_time_ <= base::Seconds(90)) {
      base::UmaHistogramCustomCounts(
          base::StrCat({uma_prefix_, name, ".90SecondsPostStartup"}), sample,
          min, max, bucketCount);
    }
  }
}

void DawnPlatform::HistogramCustomCounts(const char* name,
                                         int sample,
                                         int min,
                                         int max,
                                         int bucketCount) {
  base::UmaHistogramCustomCounts(uma_prefix_ + name, sample, min, max,
                                 bucketCount);
  HistogramCacheCountHelper(name, sample, min, max, bucketCount);
  EmitUnifiedHistogram(uma_prefix_, name, sample);
}

void DawnPlatform::HistogramCustomCountsHPC(const char* name,
                                            int sample,
                                            int min,
                                            int max,
                                            int bucketCount) {
  if (base::TimeTicks::IsHighResolution()) {
    if (ShouldRecordMetric(name)) {
      base::UmaHistogramCustomCounts(uma_prefix_ + name, sample, min, max,
                                     bucketCount);
      HistogramCacheCountHelper(name, sample, min, max, bucketCount);
    }
    EmitUnifiedHistogram(uma_prefix_, name, sample);
  }
}

void DawnPlatform::HistogramEnumeration(const char* name,
                                        int sample,
                                        int boundaryValue) {
  base::UmaHistogramExactLinear(uma_prefix_ + name, sample, boundaryValue);
}

void DawnPlatform::HistogramSparse(const char* name, int sample) {
  base::UmaHistogramSparse(uma_prefix_ + name, sample);
}

void DawnPlatform::HistogramBoolean(const char* name, bool sample) {
  base::UmaHistogramBoolean(uma_prefix_ + name, sample);
}

dawn::platform::CachingInterface* DawnPlatform::GetCachingInterface() {
  return dawn_caching_interface_.get();
}

std::unique_ptr<dawn::platform::WorkerTaskPool>
DawnPlatform::CreateWorkerTaskPool() {
  return std::make_unique<AsyncWorkerTaskPool>(progress_reporter_);
}

bool DawnPlatform::IsFeatureEnabled(dawn::platform::Features feature) {
  switch (feature) {
    case dawn::platform::Features::kWebGPUUseDXC:
      // (crbug.com/362247692): Feature is enabled by default.
      return true;
    case dawn::platform::Features::kWebGPUEnableRangeAnalysisForRobustness:
      return base::FeatureList::IsEnabled(
          features::kWebGPUEnableRangeAnalysisForRobustness);
    case dawn::platform::Features::kWebGPUUseSpirv14:
      return base::FeatureList::IsEnabled(features::kWebGPUUseSpirv14);
    case dawn::platform::Features::kWebGPUDecomposeUniformBuffers:
      return base::FeatureList::IsEnabled(
          features::kWebGPUDecomposeUniformBuffers);
    case dawn::platform::Features::kWebGPUUseHLSL2021:
      return base::FeatureList::IsEnabled(features::kWebGPUUseHLSL2021);
    case dawn::platform::Features::kWebGPUUseSpirvReconvergenceMode:
      return base::FeatureList::IsEnabled(
          features::kWebGPUUseSpirvReconvergenceMode);
    default:
      return false;
  }
}

void DawnPlatform::ReportProgress() {
  if (progress_reporter_) {
    progress_reporter_->ReportProgress();
  }
}

void DawnPlatform::OnFramePresented() {
  // Report cache hits' stats but only for first presentation.
  base::AutoLock autolock(cache_map_->lock);
  if (did_report_1st_present_cache_stats_) {
    return;
  }
  did_report_1st_present_cache_stats_ = true;

  for (auto const& [base_name, cache_counts] : cache_map_->counts) {
    // Report Hit counts
    base::UmaHistogramCounts10000(
        uma_prefix_ + base_name + "CacheHit.Counts.1stPresent",
        cache_counts.cache_hit_count);

    // Report Percentage
    int total_counts =
        cache_counts.cache_hit_count + cache_counts.cache_miss_count;
    if (total_counts > 0) {
      int hit_percentage = (cache_counts.cache_hit_count * 100) / total_counts;

      base::UmaHistogramPercentage(
          uma_prefix_ + base_name + "CacheHit.Percentage.1stPresent",
          hit_percentage);
    }
  }
}

}  // namespace gpu::webgpu