wiicompiled/aurora-main/lib/gfx/efb_ram_copy.cpp

#include "efb_ram_copy.hpp"

#include "efb_ram_encoder.hpp"
#include "tex_copy_conv.hpp"
#include "../internal.hpp"
#include "../webgpu/gpu.hpp"

#include <algorithm>
#include <array>
#include <cstring>
#include <mutex>
#include <unordered_map>
#include <vector>

#include <magic_enum.hpp>

namespace aurora::gfx::efb_ram {
namespace {

Module Log("aurora::gfx::efb_ram");

using webgpu::g_device;
using webgpu::g_instance;

// A CPU-consumed EFB copy this small is the lens-flare occlusion probe (4x4 Z24X8, 64 bytes).
// Resolving it synchronously stalls the producer and kills interpolation, so read it back async.
constexpr size_t kAsyncReadbackMaxBytes = 256;
// Each destination keeps its readback buffer forever. Only a handful are expected, and the cap
// stops an unexpected pattern of one-shot destinations from leaking GPU buffers.
constexpr size_t kMaxAsyncSlots = 32;

struct PendingCopy {
  void* dest = nullptr;
  uint32_t width = 0;
  uint32_t height = 0;
  GXTexFmt format = GX_TF_RGBA8;
  TextureHandle texture;
  TextureHandle nativeTexture;
  Range nativeBlitUniform;
};

struct Download {
  PendingCopy copy;
  wgpu::Buffer buffer;
  uint32_t bytesPerRow = 0;
  uint64_t bufferSize = 0;
};

enum class AsyncState : uint8_t {
  Idle,
  CopySubmitted,
  MapPending,
};

// One readback per destination address, reused for that destination's lifetime: the probe hits
// the same buffer every frame, so pooling avoids a per-frame allocation.
struct AsyncSlot {
  wgpu::Buffer buffer;
  TextureHandle nativeTexture;
  uint64_t bufferSize = 0;
  uint32_t bytesPerRow = 0;
  // Latched at encode time and read by the map callback.
  void* dest = nullptr;
  GXTexFmt format = GX_TF_RGBA8;
  uint32_t width = 0;
  uint32_t height = 0;
  uint32_t hostWidth = 0;
  uint32_t hostHeight = 0;
  HostPixelOrder order = HostPixelOrder::RGBA;
  AsyncState state = AsyncState::Idle;
};

std::vector<PendingCopy> g_pending;
std::vector<Download> g_downloads;

// Recorded by the producer, moved to the worker-private sealed list while the
// producer is excluded.
std::vector<PendingCopy> g_asyncPending;
std::vector<PendingCopy> g_asyncSealed;

std::mutex g_asyncMutex;
std::unordered_map<void*, AsyncSlot> g_asyncSlots;
uint32_t g_asyncMapsInFlight = 0;

uint32_t align_to(uint32_t value, uint32_t alignment) noexcept { return (value + alignment - 1) & ~(alignment - 1); }

// An internally scaled copy cannot be read back directly, so blit it down to native-resolution
// RGBA8 first. Pushes a uniform, so it must run while the staging buffers are still mapped.
void ensure_native_texture(PendingCopy& pending, TextureHandle* cache = nullptr) noexcept {
  if (pending.texture->size.width == pending.width && pending.texture->size.height == pending.height) {
    return;
  }
  if (pending.nativeTexture) {
    return;
  }
  if (cache != nullptr && *cache && (*cache)->size.width == pending.width &&
      (*cache)->size.height == pending.height) {
    pending.nativeTexture = *cache;
  } else {
    pending.nativeTexture = new_conv_texture(pending.width, pending.height, GX_TF_RGBA8, "GX EFB RAM native readback");
    if (cache != nullptr) {
      *cache = pending.nativeTexture;
    }
  }
  const std::array nativeBlitUniform{
      0.0f, 0.0f, 1.0f, 1.0f, 0.0f, 64.0f, 0.0f, 0.0f, 0.0f, 1.0f, 0.0f, 0.0f,
  };
  pending.nativeBlitUniform = push_uniform(nativeBlitUniform);
}

void encode_native_blit(const wgpu::CommandEncoder& encoder, const PendingCopy& pending) noexcept {
  const tex_copy_conv::ConvRequest blitRequest{
      .fmt = GX_TF_RGBA8,
      .srcView = pending.texture->sampleTextureView,
      .uniformRange = pending.nativeBlitUniform,
      .dst = pending.nativeTexture,
      .sampleFilter = pending.format == GX_TF_Z16 || pending.format == GX_TF_Z24X8
                          ? tex_copy_conv::SampleFilter::Nearest
                          : tex_copy_conv::SampleFilter::Linear,
  };
  tex_copy_conv::blit(encoder, blitRequest);
}

HostPixelOrder texture_pixel_order(const TextureHandle& texture) noexcept {
  return texture->format == wgpu::TextureFormat::BGRA8Unorm ? HostPixelOrder::BGRA : HostPixelOrder::RGBA;
}

void complete_async_slot(void* dest, wgpu::MapAsyncStatus status, wgpu::StringView message) noexcept {
  std::lock_guard lock{g_asyncMutex};
  if (g_asyncMapsInFlight > 0) {
    --g_asyncMapsInFlight;
  }
  const auto it = g_asyncSlots.find(dest);
  if (it == g_asyncSlots.end()) {
    return;
  }
  auto& slot = it->second;
  if (status == wgpu::MapAsyncStatus::Success) {
    const auto* pixels = static_cast<const uint8_t*>(slot.buffer.GetConstMappedRange(0, slot.bufferSize));
    if (pixels != nullptr) {
      // Writes guest RAM from the event-queue thread while the guest may be reading it. The only
      // consumer min/maxes depth for a fade factor, so a torn tile just mixes two frames' depths.
      const size_t outputSize = encoded_size(slot.format, slot.width, slot.height);
      if (!encode(slot.dest, outputSize, slot.format, slot.width, slot.height, pixels, slot.hostWidth, slot.hostHeight,
                  slot.bytesPerRow, slot.order)) {
        Log.error("Failed to encode async EFB RAM copy format=0x{:x} size={}x{}", static_cast<unsigned>(slot.format),
                  slot.width, slot.height);
      }
      // Guest RAM written from outside the embedder, so nothing bumps its write generation and a
      // texture cached over this range would keep its digest.
      notify_guest_write(slot.dest, outputSize);
    }
    slot.buffer.Unmap();
  } else if (status != wgpu::MapAsyncStatus::CallbackCancelled && status != wgpu::MapAsyncStatus::Aborted) {
    Log.warn("Async EFB RAM readback mapping failed {}: {}", magic_enum::enum_name(status), message);
  }
  slot.state = AsyncState::Idle;
}

// Runs the completion callbacks for maps whose GPU work has finished. Must not
// be called while g_asyncMutex is held: complete_async_slot takes it.
void drain_async_events() noexcept {
  {
    std::lock_guard lock{g_asyncMutex};
    if (g_asyncMapsInFlight == 0) {
      return;
    }
  }
  if (g_instance) {
    g_instance.ProcessEvents();
  }
}

} // namespace

void schedule(void* dest, uint32_t width, uint32_t height, GXTexFmt format, TextureHandle texture) noexcept {
  if (dest == nullptr || width == 0 || height == 0 || !texture) {
    return;
  }
  if (!supports_format(format)) {
    Log.fatal("Unsupported CPU-visible EFB copy format 0x{:x}", static_cast<unsigned>(format));
  }

  PendingCopy request{
      .dest = dest,
      .width = width,
      .height = height,
      .format = format,
      .texture = std::move(texture),
  };

  const size_t encodedSize = encoded_size(format, width, height);
  // Offscreen copies resolve on a separate pass list that the sealed frame's
  // native render does not replay, so they stay on the lazy path.
  bool async = !is_offscreen() && encodedSize != 0 && encodedSize <= kAsyncReadbackMaxBytes;
  if (async) {
    std::lock_guard lock{g_asyncMutex};
    const auto it = g_asyncSlots.find(dest);
    if (it == g_asyncSlots.end()) {
      if (g_asyncSlots.size() >= kMaxAsyncSlots) {
        async = false;
      } else {
        g_asyncSlots.try_emplace(dest);
        // No readback has landed here yet. 0xff decodes to far Z, which the probe reads as unobstructed
        // so a new flare fades in; zero-filled RAM would decode as fully occluded.
        std::memset(dest, 0xff, encodedSize);
        notify_guest_write(dest, encodedSize);
      }
    }
  }

  auto& list = async ? g_asyncPending : g_pending;
  if (auto it = std::find_if(list.begin(), list.end(),
                             [dest](const PendingCopy& pending) { return pending.dest == dest; });
      it != list.end()) {
    *it = std::move(request);
  } else {
    list.push_back(std::move(request));
  }
}

bool has_pending(void* dest) noexcept {
  if (dest == nullptr) return !g_pending.empty() || !g_downloads.empty();
  return std::any_of(g_pending.begin(), g_pending.end(),
                     [dest](const PendingCopy& pending) { return pending.dest == dest; }) ||
         std::any_of(g_downloads.begin(), g_downloads.end(),
                     [dest](const Download& download) { return download.copy.dest == dest; });
}

bool prepare_downloads(void* dest) noexcept {
  bool found = false;
  for (auto& pending : g_pending) {
    if (dest != nullptr && pending.dest != dest) continue;
    found = true;
    ensure_native_texture(pending);
  }
  return found;
}

void encode_downloads(const wgpu::CommandEncoder& encoder, void* dest) noexcept {
  if (!g_downloads.empty()) {
    Log.fatal("Attempted to encode new EFB RAM downloads before completing the previous batch");
  }

  g_downloads.reserve(g_pending.size());
  for (auto it = g_pending.begin(); it != g_pending.end();) {
    if (dest != nullptr && it->dest != dest) {
      ++it;
      continue;
    }
    PendingCopy pending = std::move(*it);
    it = g_pending.erase(it);
    if (pending.nativeTexture) {
      encode_native_blit(encoder, pending);
    }
    const auto& texture = pending.nativeTexture ? pending.nativeTexture : pending.texture;
    const uint32_t bytesPerRow = align_to(texture->size.width * 4, 256);
    const uint64_t bufferSize = static_cast<uint64_t>(bytesPerRow) * texture->size.height;
    const wgpu::BufferDescriptor descriptor{
        .label = "GX EFB RAM copy readback",
        .usage = wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::MapRead,
        .size = bufferSize,
    };
    auto buffer = g_device.CreateBuffer(&descriptor);
    const wgpu::TexelCopyTextureInfo source{
        .texture = texture->texture,
        .mipLevel = 0,
        .origin = {0, 0, 0},
        .aspect = wgpu::TextureAspect::All,
    };
    const wgpu::TexelCopyBufferInfo destination{
        .layout =
            wgpu::TexelCopyBufferLayout{
                .offset = 0,
                .bytesPerRow = bytesPerRow,
                .rowsPerImage = texture->size.height,
            },
        .buffer = buffer,
    };
    encoder.CopyTextureToBuffer(&source, &destination, &texture->size);
    g_downloads.push_back({
        .copy = std::move(pending),
        .buffer = std::move(buffer),
        .bytesPerRow = bytesPerRow,
        .bufferSize = bufferSize,
    });
  }
}

bool complete_downloads() noexcept {
  bool success = true;
  for (auto& download : g_downloads) {
    wgpu::MapAsyncStatus mapStatus = wgpu::MapAsyncStatus::CallbackCancelled;
    wgpu::StringView mapMessage{};
    const auto future =
        download.buffer.MapAsync(wgpu::MapMode::Read, 0, download.bufferSize, wgpu::CallbackMode::WaitAnyOnly,
                                 [&mapStatus, &mapMessage](wgpu::MapAsyncStatus status, wgpu::StringView message) {
                                   mapStatus = status;
                                   mapMessage = message;
                                 });
    const auto waitStatus = g_instance.WaitAny(future, 5000000000);
    if (waitStatus != wgpu::WaitStatus::Success || mapStatus != wgpu::MapAsyncStatus::Success) {
      Log.error("EFB RAM readback failed wait={} map={} message={}", magic_enum::enum_name(waitStatus),
                magic_enum::enum_name(mapStatus), mapMessage);
      success = false;
      continue;
    }

    const auto* pixels = static_cast<const uint8_t*>(download.buffer.GetConstMappedRange(0, download.bufferSize));
    const auto& readbackTexture = download.copy.nativeTexture ? download.copy.nativeTexture : download.copy.texture;
    const auto order = texture_pixel_order(readbackTexture);
    const size_t outputSize = encoded_size(download.copy.format, download.copy.width, download.copy.height);
    if (!encode(download.copy.dest, outputSize, download.copy.format, download.copy.width, download.copy.height, pixels,
                readbackTexture->size.width, readbackTexture->size.height, download.bytesPerRow, order)) {
      Log.error("Failed to encode EFB RAM copy format=0x{:x} size={}x{}", static_cast<unsigned>(download.copy.format),
                download.copy.width, download.copy.height);
      success = false;
    }
    notify_guest_write(download.copy.dest, outputSize);
    download.buffer.Unmap();
  }
  g_downloads.clear();
  return success;
}

void cancel() noexcept {
  g_pending.clear();
  g_downloads.clear();
  g_asyncPending.clear();
}

void seal_async_downloads() noexcept {
  {
    std::lock_guard lock{g_asyncMutex};
    for (auto& pending : g_asyncPending) {
      const auto it = g_asyncSlots.find(pending.dest);
      ensure_native_texture(pending, it != g_asyncSlots.end() ? &it->second.nativeTexture : nullptr);
    }
  }
  g_asyncSealed = std::move(g_asyncPending);
  g_asyncPending.clear();
}

void encode_async_downloads(const wgpu::CommandEncoder& encoder) noexcept {
  if (g_asyncSealed.empty()) {
    return;
  }
  // Retire the previous frame's map first so its slot is free here, otherwise a single-slot
  // destination could only be sampled every other frame.
  drain_async_events();
  std::lock_guard lock{g_asyncMutex};
  for (auto& pending : g_asyncSealed) {
    const auto it = g_asyncSlots.find(pending.dest);
    if (it == g_asyncSlots.end()) {
      continue;
    }
    auto& slot = it->second;
    // Only one readback per destination may be outstanding and the previous map still owns the pooled
    // buffer. Dropping this request just leaves the guest on a slightly older probe.
    if (slot.state != AsyncState::Idle) {
      continue;
    }
    if (pending.nativeTexture) {
      encode_native_blit(encoder, pending);
    }
    const auto& texture = pending.nativeTexture ? pending.nativeTexture : pending.texture;
    const uint32_t bytesPerRow = align_to(texture->size.width * 4, 256);
    const uint64_t bufferSize = static_cast<uint64_t>(bytesPerRow) * texture->size.height;
    if (!slot.buffer || slot.bufferSize != bufferSize) {
      const wgpu::BufferDescriptor descriptor{
          .label = "GX EFB RAM async readback",
          .usage = wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::MapRead,
          .size = bufferSize,
      };
      slot.buffer = g_device.CreateBuffer(&descriptor);
      slot.bufferSize = bufferSize;
    }
    if (!slot.buffer) {
      continue;
    }
    const wgpu::TexelCopyTextureInfo source{
        .texture = texture->texture,
        .mipLevel = 0,
        .origin = {0, 0, 0},
        .aspect = wgpu::TextureAspect::All,
    };
    const wgpu::TexelCopyBufferInfo destination{
        .layout =
            wgpu::TexelCopyBufferLayout{
                .offset = 0,
                .bytesPerRow = bytesPerRow,
                .rowsPerImage = texture->size.height,
            },
        .buffer = slot.buffer,
    };
    encoder.CopyTextureToBuffer(&source, &destination, &texture->size);
    slot.bytesPerRow = bytesPerRow;
    slot.dest = pending.dest;
    slot.format = pending.format;
    slot.width = pending.width;
    slot.height = pending.height;
    slot.hostWidth = texture->size.width;
    slot.hostHeight = texture->size.height;
    slot.order = texture_pixel_order(texture);
    slot.state = AsyncState::CopySubmitted;
  }
  g_asyncSealed.clear();
}

void after_submit() noexcept {
  struct PendingMap {
    void* dest;
    wgpu::Buffer buffer;
    uint64_t bufferSize;
  };
  std::vector<PendingMap> pendingMaps;
  {
    std::lock_guard lock{g_asyncMutex};
    for (auto& [dest, slot] : g_asyncSlots) {
      if (slot.state != AsyncState::CopySubmitted) {
        continue;
      }
      slot.state = AsyncState::MapPending;
      ++g_asyncMapsInFlight;
      pendingMaps.push_back({dest, slot.buffer, slot.bufferSize});
    }
  }

  for (const auto& pending : pendingMaps) {
    pending.buffer.MapAsync(wgpu::MapMode::Read, 0, pending.bufferSize, wgpu::CallbackMode::AllowSpontaneous,
                            [dest = pending.dest](wgpu::MapAsyncStatus status, wgpu::StringView message) {
                              complete_async_slot(dest, status, message);
                            });
  }

  // Nothing else pumps the WebGPU event queue on this thread, so a completed map would sit
  // unharvested with its slot busy. ProcessEvents does not block.
  drain_async_events();
}

void abort_async() noexcept { g_asyncSealed.clear(); }

void shutdown() noexcept {
  cancel();
  g_asyncSealed.clear();
  std::lock_guard lock{g_asyncMutex};
  g_asyncSlots.clear();
  g_asyncMapsInFlight = 0;
}

} // namespace aurora::gfx::efb_ram