xenia-canary/src/xenia/gpu/texture_cache.cc

/**
 ******************************************************************************
 * Xenia : Xbox 360 Emulator Research Project                                 *
 ******************************************************************************
 * Copyright 2022 Ben Vanik. All rights reserved.                             *
 * Released under the BSD license - see LICENSE in the root for more details. *
 ******************************************************************************
 */

#include "xenia/gpu/texture_cache.h"

#include <algorithm>
#include <cstdint>
#include <utility>

#include "xenia/base/assert.h"
#include "xenia/base/clock.h"
#include "xenia/base/cvar.h"
#include "xenia/base/logging.h"
#include "xenia/base/math.h"
#include "xenia/base/profiling.h"
#include "xenia/gpu/gpu_flags.h"
#include "xenia/gpu/register_file.h"
#include "xenia/gpu/texture_info.h"
#include "xenia/gpu/texture_util.h"
#include "xenia/gpu/xenos.h"

DEFINE_int32(
    draw_resolution_scale_x, 1,
    "Integer pixel width scale used for scaling the rendering resolution "
    "opaquely to the game.\n"
    "1, 2 and 3 may be supported, but support of anything above 1 depends on "
    "the device properties, such as whether it supports sparse binding / tiled "
    "resources, the number of virtual address bits per resource, and other "
    "factors.\n"
    "Various effects and parts of game rendering pipelines may work "
    "incorrectly as pixels become ambiguous from the game's perspective and "
    "because half-pixel offset (which normally doesn't affect coverage when "
    "MSAA isn't used) becomes full-pixel.",
    "GPU");
DEFINE_int32(
    draw_resolution_scale_y, 1,
    "Integer pixel width scale used for scaling the rendering resolution "
    "opaquely to the game.\n"
    "See draw_resolution_scale_x for more information.",
    "GPU");
DEFINE_uint32(
    texture_cache_memory_limit_soft, 384,
    "Maximum host texture memory usage (in megabytes) above which old textures "
    "will be destroyed.",
    "GPU");
DEFINE_uint32(
    texture_cache_memory_limit_soft_lifetime, 30,
    "Seconds a texture should be unused to be considered old enough to be "
    "deleted if texture memory usage exceeds texture_cache_memory_limit_soft.",
    "GPU");
DEFINE_uint32(
    texture_cache_memory_limit_hard, 768,
    "Maximum host texture memory usage (in megabytes) above which textures "
    "will be destroyed as soon as possible.",
    "GPU");
DEFINE_uint32(
    texture_cache_memory_limit_render_to_texture, 24,
    "Part of the host texture memory budget (in megabytes) that will be scaled "
    "by the current drawing resolution scale.\n"
    "If texture_cache_memory_limit_soft, for instance, is 384, and this is 24, "
    "it will be assumed that the game will be using roughly 24 MB of "
    "render-to-texture (resolve) targets and 384 - 24 = 360 MB of regular "
    "textures - so with 2x2 resolution scaling, the soft limit will be 360 + "
    "96 MB, and with 3x3, it will be 360 + 216 MB.",
    "GPU");

namespace xe {
namespace gpu {

const TextureCache::LoadShaderInfo
    TextureCache::load_shader_info_[kLoadShaderCount] = {
        // k8bpb
        {3, 4, 1, 4},
        // k16bpb
        {4, 4, 2, 4},
        // k32bpb
        {4, 4, 4, 3},
        // k64bpb
        {4, 4, 8, 2},
        // k128bpb
        {4, 4, 16, 1},
        // kR5G5B5A1ToB5G5R5A1
        {4, 4, 2, 4},
        // kR5G6B5ToB5G6R5
        {4, 4, 2, 4},
        // kR5G5B6ToB5G6R5WithRBGASwizzle
        {4, 4, 2, 4},
        // kRGBA4ToBGRA4
        {4, 4, 2, 4},
        // kRGBA4ToARGB4
        {4, 4, 2, 4},
        // kGBGR8ToGRGB8
        {4, 4, 4, 3},
        // kGBGR8ToRGB8
        {4, 4, 8, 3},
        // kBGRG8ToRGBG8
        {4, 4, 4, 3},
        // kBGRG8ToRGB8
        {4, 4, 8, 3},
        // kR10G11B11ToRGBA16
        {4, 4, 8, 3},
        // kR10G11B11ToRGBA16SNorm
        {4, 4, 8, 3},
        // kR11G11B10ToRGBA16
        {4, 4, 8, 3},
        // kR11G11B10ToRGBA16SNorm
        {4, 4, 8, 3},
        // kR16UNormToFloat
        {4, 4, 2, 4},
        // kR16SNormToFloat
        {4, 4, 2, 4},
        // kRG16UNormToFloat
        {4, 4, 4, 3},
        // kRG16SNormToFloat
        {4, 4, 4, 3},
        // kRGBA16UNormToFloat
        {4, 4, 8, 2},
        // kRGBA16SNormToFloat
        {4, 4, 8, 2},
        // kDXT1ToRGBA8
        {4, 4, 4, 2},
        // kDXT3ToRGBA8
        {4, 4, 4, 1},
        // kDXT5ToRGBA8
        {4, 4, 4, 1},
        // kDXNToRG8
        {4, 4, 2, 1},
        // kDXT3A
        {4, 4, 1, 2},
        // kDXT3AAs1111ToBGRA4
        {4, 4, 2, 2},
        // kDXT3AAs1111ToARGB4
        {4, 4, 2, 2},
        // kDXT5AToR8
        {4, 4, 1, 2},
        // kCTX1
        {4, 4, 2, 2},
        // kDepthUnorm
        {4, 4, 4, 3},
        // kDepthFloat
        {4, 4, 4, 3},
};

TextureCache::TextureCache(const RegisterFile& register_file,
                           SharedMemory& shared_memory,
                           uint32_t draw_resolution_scale_x,
                           uint32_t draw_resolution_scale_y)
    : register_file_(register_file),
      shared_memory_(shared_memory),
      draw_resolution_scale_x_(draw_resolution_scale_x),
      draw_resolution_scale_y_(draw_resolution_scale_y),
      draw_resolution_scale_x_divisor_(draw_resolution_scale_x),
      draw_resolution_scale_y_divisor_(draw_resolution_scale_y) {
  assert_true(draw_resolution_scale_x >= 1);
  assert_true(draw_resolution_scale_x <= kMaxDrawResolutionScaleAlongAxis);
  assert_true(draw_resolution_scale_y >= 1);
  assert_true(draw_resolution_scale_y <= kMaxDrawResolutionScaleAlongAxis);

  if (draw_resolution_scale_x > 1 || draw_resolution_scale_y > 1) {
    constexpr uint32_t kScaledResolvePageDwordCount =
        SharedMemory::kBufferSize / 4096 / 32;
    scaled_resolve_pages_ =
        std::unique_ptr<uint32_t[]>(new uint32_t[kScaledResolvePageDwordCount]);
    std::memset(scaled_resolve_pages_.get(), 0,
                kScaledResolvePageDwordCount * sizeof(uint32_t));
    std::memset(scaled_resolve_pages_l2_, 0, sizeof(scaled_resolve_pages_l2_));
    scaled_resolve_global_watch_handle_ = shared_memory.RegisterGlobalWatch(
        ScaledResolveGlobalWatchCallbackThunk, this);
  }
}

TextureCache::~TextureCache() {
  DestroyAllTextures(true);

  if (scaled_resolve_global_watch_handle_) {
    shared_memory().UnregisterGlobalWatch(scaled_resolve_global_watch_handle_);
  }
}

bool TextureCache::GetConfigDrawResolutionScale(uint32_t& x_out,
                                                uint32_t& y_out) {
  uint32_t config_x =
      uint32_t(std::max(INT32_C(1), cvars::draw_resolution_scale_x));
  uint32_t config_y =
      uint32_t(std::max(INT32_C(1), cvars::draw_resolution_scale_y));

  uint32_t clamped_x = std::min(kMaxDrawResolutionScaleAlongAxis, config_x);
  uint32_t clamped_y = std::min(kMaxDrawResolutionScaleAlongAxis, config_y);
  x_out = clamped_x;
  y_out = clamped_y;
  return clamped_x == config_x && clamped_y == config_y;
}

void TextureCache::ClearCache() { DestroyAllTextures(); }

void TextureCache::CompletedSubmissionUpdated(
    uint64_t completed_submission_index) {
  // If memory usage is too high, destroy unused textures.
  uint64_t current_time = xe::Clock::QueryHostUptimeMillis();
  // texture_cache_memory_limit_render_to_texture is assumed to be included in
  // texture_cache_memory_limit_soft and texture_cache_memory_limit_hard, at 1x,
  // so subtracting 1 from the scale.
  uint32_t limit_scaled_resolve_add_mb =
      cvars::texture_cache_memory_limit_render_to_texture *
      (draw_resolution_scale_x() * draw_resolution_scale_y() - 1);
  uint32_t limit_soft_mb =
      cvars::texture_cache_memory_limit_soft + limit_scaled_resolve_add_mb;
  uint32_t limit_hard_mb =
      cvars::texture_cache_memory_limit_hard + limit_scaled_resolve_add_mb;
  uint32_t limit_soft_lifetime =
      cvars::texture_cache_memory_limit_soft_lifetime * 1000;
  bool destroyed_any = false;
  while (texture_used_first_ != nullptr) {
    uint64_t total_host_memory_usage_mb =
        (textures_total_host_memory_usage_ + ((UINT32_C(1) << 20) - 1)) >> 20;
    bool limit_hard_exceeded = total_host_memory_usage_mb > limit_hard_mb;
    if (total_host_memory_usage_mb <= limit_soft_mb && !limit_hard_exceeded) {
      break;
    }
    Texture* texture = texture_used_first_;
    if (texture->last_usage_submission_index() > completed_submission_index) {
      break;
    }
    if (!limit_hard_exceeded &&
        (texture->last_usage_time() + limit_soft_lifetime) > current_time) {
      break;
    }
    if (!destroyed_any) {
      destroyed_any = true;
      // The texture being destroyed might have been bound in the previous
      // submissions, and nothing has overwritten the binding yet, so completion
      // of the submission where the texture was last actually used on the GPU
      // doesn't imply that it's not bound currently. Reset bindings if
      // any texture has been destroyed.
      ResetTextureBindings();
    }
    // Remove the texture from the map and destroy it via its unique_ptr.
    auto found_texture_it = textures_.find(texture->key());
    assert_true(found_texture_it != textures_.end());
    if (found_texture_it != textures_.end()) {
      assert_true(found_texture_it->second.get() == texture);
      textures_.erase(found_texture_it);
      // `texture` is invalid now.
    }
  }
  if (destroyed_any) {
    COUNT_profile_set("gpu/texture_cache/textures", textures_.size());
  }
}

void TextureCache::BeginSubmission(uint64_t new_submission_index) {
  assert_true(new_submission_index > current_submission_index_);
  current_submission_index_ = new_submission_index;
  current_submission_time_ = xe::Clock::QueryHostUptimeMillis();
}

void TextureCache::BeginFrame() {
  // In case there was a failure to create something in the previous frame, make
  // sure bindings are reset so a new attempt will surely be made if the texture
  // is requested again.
  ResetTextureBindings();
}

void TextureCache::MarkRangeAsResolved(uint32_t start_unscaled,
                                       uint32_t length_unscaled) {
  if (length_unscaled == 0) {
    return;
  }
  start_unscaled &= 0x1FFFFFFF;
  length_unscaled = std::min(length_unscaled, 0x20000000 - start_unscaled);

  if (IsDrawResolutionScaled()) {
    uint32_t page_first = start_unscaled >> 12;
    uint32_t page_last = (start_unscaled + length_unscaled - 1) >> 12;
    uint32_t block_first = page_first >> 5;
    uint32_t block_last = page_last >> 5;
    auto global_lock = global_critical_region_.Acquire();
    for (uint32_t i = block_first; i <= block_last; ++i) {
      uint32_t add_bits = UINT32_MAX;
      if (i == block_first) {
        add_bits &= ~((UINT32_C(1) << (page_first & 31)) - 1);
      }
      if (i == block_last && (page_last & 31) != 31) {
        add_bits &= (UINT32_C(1) << ((page_last & 31) + 1)) - 1;
      }
      scaled_resolve_pages_[i] |= add_bits;
      scaled_resolve_pages_l2_[i >> 6] |= UINT64_C(1) << (i & 63);
    }
  }

  // Invalidate textures. Toggling individual textures between scaled and
  // unscaled also relies on invalidation through shared memory.
  shared_memory().RangeWrittenByGpu(start_unscaled, length_unscaled, true);
}

uint32_t TextureCache::GuestToHostSwizzle(uint32_t guest_swizzle,
                                          uint32_t host_format_swizzle) {
  uint32_t host_swizzle = 0;
  for (uint32_t i = 0; i < 4; ++i) {
    uint32_t guest_swizzle_component = (guest_swizzle >> (3 * i)) & 0b111;
    uint32_t host_swizzle_component;
    if (guest_swizzle_component >= xenos::XE_GPU_TEXTURE_SWIZZLE_0) {
      // Get rid of 6 and 7 values (to prevent host GPU errors if the game has
      // something broken) the simple way - by changing them to 4 (0) and 5 (1).
      host_swizzle_component = guest_swizzle_component & 0b101;
    } else {
      host_swizzle_component =
          (host_format_swizzle >> (3 * guest_swizzle_component)) & 0b111;
    }
    host_swizzle |= host_swizzle_component << (3 * i);
  }
  return host_swizzle;
}

void TextureCache::RequestTextures(uint32_t used_texture_mask) {
  const auto& regs = register_file();

  if (texture_became_outdated_.exchange(false, std::memory_order_acquire)) {
    // A texture has become outdated - make sure whether textures are outdated
    // is rechecked in this draw and in subsequent ones to reload the new data
    // if needed.
    ResetTextureBindings();
  }

  // Update the texture keys and the textures.
  uint32_t bindings_changed = 0;
  uint32_t textures_remaining = used_texture_mask & ~texture_bindings_in_sync_;
  uint32_t index = 0;
  while (xe::bit_scan_forward(textures_remaining, &index)) {
    uint32_t index_bit = UINT32_C(1) << index;
    textures_remaining &= ~index_bit;
    TextureBinding& binding = texture_bindings_[index];
    const auto& fetch = regs.Get<xenos::xe_gpu_texture_fetch_t>(
        XE_GPU_REG_SHADER_CONSTANT_FETCH_00_0 + index * 6);
    TextureKey old_key = binding.key;
    uint8_t old_swizzled_signs = binding.swizzled_signs;
    BindingInfoFromFetchConstant(fetch, binding.key, &binding.swizzled_signs);
    texture_bindings_in_sync_ |= index_bit;
    if (!binding.key.is_valid) {
      if (old_key.is_valid) {
        bindings_changed |= index_bit;
      }
      binding.Reset();
      continue;
    }
    uint32_t old_host_swizzle = binding.host_swizzle;
    binding.host_swizzle =
        GuestToHostSwizzle(fetch.swizzle, GetHostFormatSwizzle(binding.key));

    // Check if need to load the unsigned and the signed versions of the texture
    // (if the format is emulated with different host bit representations for
    // signed and unsigned - otherwise only the unsigned one is loaded).
    bool key_changed = binding.key != old_key;
    bool any_sign_was_not_signed =
        texture_util::IsAnySignNotSigned(old_swizzled_signs);
    bool any_sign_was_signed =
        texture_util::IsAnySignSigned(old_swizzled_signs);
    bool any_sign_is_not_signed =
        texture_util::IsAnySignNotSigned(binding.swizzled_signs);
    bool any_sign_is_signed =
        texture_util::IsAnySignSigned(binding.swizzled_signs);
    if (key_changed || binding.host_swizzle != old_host_swizzle ||
        any_sign_is_not_signed != any_sign_was_not_signed ||
        any_sign_is_signed != any_sign_was_signed) {
      bindings_changed |= index_bit;
    }
    bool load_unsigned_data = false, load_signed_data = false;
    if (IsSignedVersionSeparateForFormat(binding.key)) {
      // Can reuse previously loaded unsigned/signed versions if the key is the
      // same and the texture was previously bound as unsigned/signed
      // respectively (checking the previous values of signedness rather than
      // binding.texture != nullptr and binding.texture_signed != nullptr also
      // prevents repeated attempts to load the texture if it has failed to
      // load).
      if (any_sign_is_not_signed) {
        if (key_changed || !any_sign_was_not_signed) {
          binding.texture = FindOrCreateTexture(binding.key);
          load_unsigned_data = true;
        }
      } else {
        binding.texture = nullptr;
      }
      if (any_sign_is_signed) {
        if (key_changed || !any_sign_was_signed) {
          TextureKey signed_key = binding.key;
          signed_key.signed_separate = 1;
          binding.texture_signed = FindOrCreateTexture(signed_key);
          load_signed_data = true;
        }
      } else {
        binding.texture_signed = nullptr;
      }
    } else {
      // Same resource for both unsigned and signed, but descriptor formats may
      // be different.
      if (key_changed) {
        binding.texture = FindOrCreateTexture(binding.key);
        load_unsigned_data = true;
      }
      binding.texture_signed = nullptr;
    }
    if (load_unsigned_data && binding.texture != nullptr) {
      LoadTextureData(*binding.texture);
    }
    if (load_signed_data && binding.texture_signed != nullptr) {
      LoadTextureData(*binding.texture_signed);
    }
  }
  if (bindings_changed) {
    UpdateTextureBindingsImpl(bindings_changed);
  }
}

const char* TextureCache::TextureKey::GetLogDimensionName(
    xenos::DataDimension dimension) {
  switch (dimension) {
    case xenos::DataDimension::k1D:
      return "1D";
    case xenos::DataDimension::k2DOrStacked:
      return "2D";
    case xenos::DataDimension::k3D:
      return "3D";
    case xenos::DataDimension::kCube:
      return "cube";
    default:
      assert_unhandled_case(dimension);
      return "unknown";
  }
}

void TextureCache::TextureKey::LogAction(const char* action) const {
  XELOGGPU(
      "{} {} {}{}x{}x{} {} {} texture with {} {}packed mip level{}, "
      "base at 0x{:08X} (pitch {}), mips at 0x{:08X}",
      action, tiled ? "tiled" : "linear", scaled_resolve ? "scaled " : "",
      GetWidth(), GetHeight(), GetDepthOrArraySize(), GetLogDimensionName(),
      FormatInfo::Get(format)->name, mip_max_level + 1, packed_mips ? "" : "un",
      mip_max_level != 0 ? "s" : "", base_page << 12, pitch << 5,
      mip_page << 12);
}

void TextureCache::Texture::LogAction(const char* action) const {
  XELOGGPU(
      "{} {} {}{}x{}x{} {} {} texture with {} {}packed mip level{}, "
      "base at 0x{:08X} (pitch {}, size 0x{:08X}), mips at 0x{:08X} (size "
      "0x{:08X})",
      action, key_.tiled ? "tiled" : "linear",
      key_.scaled_resolve ? "scaled " : "", key_.GetWidth(), key_.GetHeight(),
      key_.GetDepthOrArraySize(), key_.GetLogDimensionName(),
      FormatInfo::Get(key_.format)->name, key_.mip_max_level + 1,
      key_.packed_mips ? "" : "un", key_.mip_max_level != 0 ? "s" : "",
      key_.base_page << 12, key_.pitch << 5, GetGuestBaseSize(),
      key_.mip_page << 12, GetGuestMipsSize());
}

// The texture must be in the recent usage list. Place it in front now because
// after creation, the texture will likely be used immediately, and it should
// not be destroyed immediately after creation if dropping of old textures is
// performed somehow. The list is maintained by the Texture, not the
// TextureCache itself (unlike the `textures_` container).
TextureCache::Texture::Texture(TextureCache& texture_cache,
                               const TextureKey& key)
    : texture_cache_(texture_cache),
      key_(key),
      guest_layout_(key.GetGuestLayout()),
      base_resolved_(key.scaled_resolve),
      mips_resolved_(key.scaled_resolve),
      last_usage_submission_index_(texture_cache.current_submission_index_),
      last_usage_time_(texture_cache.current_submission_time_),
      used_previous_(texture_cache.texture_used_last_),
      used_next_(nullptr) {
  if (texture_cache.texture_used_last_) {
    texture_cache.texture_used_last_->used_next_ = this;
  } else {
    texture_cache.texture_used_first_ = this;
  }
  texture_cache.texture_used_last_ = this;

  // Never try to upload data that doesn't exist.
  base_outdated_ = guest_layout().base.level_data_extent_bytes != 0;
  mips_outdated_ = guest_layout().mips_total_extent_bytes != 0;
}

TextureCache::Texture::~Texture() {
  if (mips_watch_handle_) {
    texture_cache().shared_memory().UnwatchMemoryRange(mips_watch_handle_);
  }
  if (base_watch_handle_) {
    texture_cache().shared_memory().UnwatchMemoryRange(base_watch_handle_);
  }

  if (used_previous_) {
    used_previous_->used_next_ = used_next_;
  } else {
    texture_cache_.texture_used_first_ = used_next_;
  }
  if (used_next_) {
    used_next_->used_previous_ = used_previous_;
  } else {
    texture_cache_.texture_used_last_ = used_previous_;
  }

  texture_cache_.UpdateTexturesTotalHostMemoryUsage(0, host_memory_usage_);
}

void TextureCache::Texture::MakeUpToDateAndWatch(
    const global_unique_lock_type& global_lock) {
  SharedMemory& shared_memory = texture_cache().shared_memory();
  if (base_outdated_) {
    assert_not_zero(GetGuestBaseSize());
    base_outdated_ = false;
    base_watch_handle_ = shared_memory.WatchMemoryRange(
        key().base_page << 12, GetGuestBaseSize(), TextureCache::WatchCallback,
        this, nullptr, 0);
  }
  if (mips_outdated_) {
    assert_not_zero(GetGuestMipsSize());
    mips_outdated_ = false;
    mips_watch_handle_ = shared_memory.WatchMemoryRange(
        key().mip_page << 12, GetGuestMipsSize(), TextureCache::WatchCallback,
        this, nullptr, 1);
  }
}

void TextureCache::Texture::MarkAsUsed() {
  assert_true(last_usage_submission_index_ <=
              texture_cache_.current_submission_index_);
  // This is called very frequently, don't relink unless needed for caching.
  if (last_usage_submission_index_ >=
      texture_cache_.current_submission_index_) {
    return;
  }
  last_usage_submission_index_ = texture_cache_.current_submission_index_;
  last_usage_time_ = texture_cache_.current_submission_time_;
  if (used_next_ == nullptr) {
    // Already the most recently used.
    return;
  }
  if (used_previous_ != nullptr) {
    used_previous_->used_next_ = used_next_;
  } else {
    texture_cache_.texture_used_first_ = used_next_;
  }
  used_next_->used_previous_ = used_previous_;
  used_previous_ = texture_cache_.texture_used_last_;
  used_next_ = nullptr;
  texture_cache_.texture_used_last_->used_next_ = this;
  texture_cache_.texture_used_last_ = this;
}

void TextureCache::Texture::WatchCallback(
    [[maybe_unused]] const global_unique_lock_type& global_lock, bool is_mip) {
  if (is_mip) {
    assert_not_zero(GetGuestMipsSize());
    mips_outdated_ = true;
    mips_watch_handle_ = nullptr;
  } else {
    assert_not_zero(GetGuestBaseSize());
    base_outdated_ = true;
    base_watch_handle_ = nullptr;
  }
}

void TextureCache::WatchCallback(const global_unique_lock_type& global_lock,
                                 void* context, void* data, uint64_t argument,
                                 bool invalidated_by_gpu) {
  Texture& texture = *static_cast<Texture*>(context);
  texture.WatchCallback(global_lock, argument != 0);
  texture.texture_cache().texture_became_outdated_.store(
      true, std::memory_order_release);
}

void TextureCache::DestroyAllTextures(bool from_destructor) {
  ResetTextureBindings(from_destructor);
  textures_.clear();
  COUNT_profile_set("gpu/texture_cache/textures", 0);
}

TextureCache::Texture* TextureCache::FindOrCreateTexture(TextureKey key) {
  // Check if the texture is a scaled resolve texture.
  if (IsDrawResolutionScaled() && key.tiled &&
      IsScaledResolveSupportedForFormat(key)) {
    texture_util::TextureGuestLayout scaled_resolve_guest_layout =
        key.GetGuestLayout();
    if ((scaled_resolve_guest_layout.base.level_data_extent_bytes &&
         IsRangeScaledResolved(
             key.base_page << 12,
             scaled_resolve_guest_layout.base.level_data_extent_bytes)) ||
        (scaled_resolve_guest_layout.mips_total_extent_bytes &&
         IsRangeScaledResolved(
             key.mip_page << 12,
             scaled_resolve_guest_layout.mips_total_extent_bytes))) {
      key.scaled_resolve = 1;
    }
  }

  uint32_t host_width = key.GetWidth();
  uint32_t host_height = key.GetHeight();
  if (key.scaled_resolve) {
    host_width *= draw_resolution_scale_x();
    host_height *= draw_resolution_scale_y();
  }
  // With 3x resolution scaling, a 2D texture may become bigger than the
  // Direct3D 11 limit, and with 2x, a 3D one as well.
  // TODO(Triang3l): Skip mips on Vulkan in this case - the minimum requirement
  // there is 4096, which is below the Xenos maximum texture size of 8192.
  uint32_t max_host_width_height = GetMaxHostTextureWidthHeight(key.dimension);
  uint32_t max_host_depth_or_array_size =
      GetMaxHostTextureDepthOrArraySize(key.dimension);
  if (host_width > max_host_width_height ||
      host_height > max_host_width_height ||
      key.GetDepthOrArraySize() > max_host_depth_or_array_size) {
    return nullptr;
  }

  // Try to find an existing texture.
  // TODO(Triang3l): Reuse a texture with mip_page unchanged, but base_page
  // previously 0, now not 0, to save memory - common case in streaming.
  auto found_texture_it = textures_.find(key);
  if (found_texture_it != textures_.end()) {
    return found_texture_it->second.get();
  }

  // Create the texture and add it to the map.
  Texture* texture;
  {
    std::unique_ptr<Texture> new_texture = CreateTexture(key);
    if (!new_texture) {
      key.LogAction("Failed to create");
      return nullptr;
    }
    assert_true(new_texture->key() == key);
    texture =
        textures_.emplace(key, std::move(new_texture)).first->second.get();
  }
  COUNT_profile_set("gpu/texture_cache/textures", textures_.size());
  texture->LogAction("Created");
  return texture;
}

bool TextureCache::LoadTextureData(Texture& texture) {
  // Check what needs to be uploaded.
  bool base_outdated, mips_outdated;
  {
    auto global_lock = global_critical_region_.Acquire();
    base_outdated = texture.base_outdated(global_lock);
    mips_outdated = texture.mips_outdated(global_lock);
  }
  if (!base_outdated && !mips_outdated) {
    return true;
  }

  TextureKey texture_key = texture.key();

  // Implementation may load multiple blocks at once via accesses of up to 128
  // bits (R32G32B32A32_UINT), so aligning the size to this value to make sure
  // if the texture is small (especially if it's linear), the last blocks won't
  // be cut off (hosts may return 0, 0, 0, 0 for the whole R32G32B32A32_UINT
  // access for the non-16-aligned tail even if 1...15 bytes are actually
  // provided for it).

  // Request uploading of the texture data to the shared memory.
  // This is also necessary when resolution scaling is used - the texture cache
  // relies on shared memory for invalidation of both unscaled and scaled
  // textures. Plus a texture may be unscaled partially, when only a portion of
  // its pages is invalidated, in this case we'll need the texture from the
  // shared memory to load the unscaled parts.
  // TODO(Triang3l): Load unscaled parts.
  bool base_resolved = texture.GetBaseResolved();
  if (base_outdated) {
    if (!shared_memory().RequestRange(
            texture_key.base_page << 12,
            xe::align(texture.GetGuestBaseSize(), UINT32_C(16)),
            texture_key.scaled_resolve ? nullptr : &base_resolved)) {
      return false;
    }
  }
  bool mips_resolved = texture.GetMipsResolved();
  if (mips_outdated) {
    if (!shared_memory().RequestRange(
            texture_key.mip_page << 12,
            xe::align(texture.GetGuestMipsSize(), UINT32_C(16)),
            texture_key.scaled_resolve ? nullptr : &mips_resolved)) {
      return false;
    }
  }
  if (texture_key.scaled_resolve) {
    // Make sure all the scaled resolve memory is resident and accessible from
    // the shader, including any possible padding that hasn't yet been touched
    // by an actual resolve, but is still included in the texture size, so the
    // GPU won't be trying to access unmapped memory.
    if (!EnsureScaledResolveMemoryCommitted(texture_key.base_page << 12,
                                            texture.GetGuestBaseSize(), 4)) {
      return false;
    }
    if (!EnsureScaledResolveMemoryCommitted(texture_key.mip_page << 12,
                                            texture.GetGuestMipsSize(), 4)) {
      return false;
    }
  }

  // Actually load the texture data.
  if (!LoadTextureDataFromResidentMemoryImpl(texture, base_outdated,
                                             mips_outdated)) {
    return false;
  }

  // Update the source of the texture (resolve vs. CPU or memexport) for
  // purposes of handling piecewise gamma emulation via sRGB and for resolution
  // scale in sampling offsets.
  if (!texture_key.scaled_resolve) {
    texture.SetBaseResolved(base_resolved);
    texture.SetMipsResolved(mips_resolved);
  }

  // Mark the ranges as uploaded and watch them. This is needed for scaled
  // resolves as well to detect when the CPU wants to reuse the memory for a
  // regular texture or a vertex buffer, and thus the scaled resolve version is
  // not up to date anymore.
  texture.MakeUpToDateAndWatch(global_critical_region_.Acquire());

  texture.LogAction("Loaded");

  return true;
}

void TextureCache::BindingInfoFromFetchConstant(
    const xenos::xe_gpu_texture_fetch_t& fetch, TextureKey& key_out,
    uint8_t* swizzled_signs_out) {
  // Reset the key and the signedness.
  key_out.MakeInvalid();
  if (swizzled_signs_out != nullptr) {
    *swizzled_signs_out =
        uint8_t(xenos::TextureSign::kUnsigned) * uint8_t(0b01010101);
  }

  switch (fetch.type) {
    case xenos::FetchConstantType::kTexture:
      break;
    case xenos::FetchConstantType::kInvalidTexture:
      if (cvars::gpu_allow_invalid_fetch_constants) {
        break;
      }
      XELOGW(
          "Texture fetch constant ({:08X} {:08X} {:08X} {:08X} {:08X} {:08X}) "
          "has \"invalid\" type! This is incorrect behavior, but you can try "
          "bypassing this by launching Xenia with "
          "--gpu_allow_invalid_fetch_constants=true.",
          fetch.dword_0, fetch.dword_1, fetch.dword_2, fetch.dword_3,
          fetch.dword_4, fetch.dword_5);
      return;
    default:
      XELOGW(
          "Texture fetch constant ({:08X} {:08X} {:08X} {:08X} {:08X} {:08X}) "
          "is completely invalid!",
          fetch.dword_0, fetch.dword_1, fetch.dword_2, fetch.dword_3,
          fetch.dword_4, fetch.dword_5);
      return;
  }

  uint32_t width_minus_1, height_minus_1, depth_or_array_size_minus_1;
  uint32_t base_page, mip_page, mip_max_level;
  texture_util::GetSubresourcesFromFetchConstant(
      fetch, &width_minus_1, &height_minus_1, &depth_or_array_size_minus_1,
      &base_page, &mip_page, nullptr, &mip_max_level);
  if (base_page == 0 && mip_page == 0) {
    // No texture data at all.
    return;
  }
  if (fetch.dimension == xenos::DataDimension::k1D) {
    bool is_invalid_1d = false;
    // TODO(Triang3l): Support long 1D textures.
    if (width_minus_1 >= xenos::kTexture2DCubeMaxWidthHeight) {
      XELOGE(
          "1D texture is too wide ({}) - ignoring! Report the game to Xenia "
          "developers",
          width_minus_1 + 1);
      is_invalid_1d = true;
    }
    assert_false(fetch.tiled);
    if (fetch.tiled) {
      XELOGE(
          "1D texture has tiling enabled in the fetch constant, but this "
          "appears to be completely wrong - ignoring! Report the game to Xenia "
          "developers");
      is_invalid_1d = true;
    }
    assert_false(fetch.packed_mips);
    if (fetch.packed_mips) {
      XELOGE(
          "1D texture has packed mips enabled in the fetch constant, but this "
          "appears to be completely wrong - ignoring! Report the game to Xenia "
          "developers");
      is_invalid_1d = true;
    }
    if (is_invalid_1d) {
      return;
    }
  }

  xenos::TextureFormat format = GetBaseFormat(fetch.format);

  key_out.base_page = base_page;
  key_out.mip_page = mip_page;
  key_out.dimension = fetch.dimension;
  key_out.width_minus_1 = width_minus_1;
  key_out.height_minus_1 = height_minus_1;
  key_out.depth_or_array_size_minus_1 = depth_or_array_size_minus_1;
  key_out.pitch = fetch.pitch;
  key_out.mip_max_level = mip_max_level;
  key_out.tiled = fetch.tiled;
  key_out.packed_mips = fetch.packed_mips;
  key_out.format = format;
  key_out.endianness = fetch.endianness;

  key_out.is_valid = 1;

  if (swizzled_signs_out != nullptr) {
    *swizzled_signs_out = texture_util::SwizzleSigns(fetch);
  }
}

void TextureCache::ResetTextureBindings(bool from_destructor) {
  uint32_t bindings_reset = 0;
  for (size_t i = 0; i < texture_bindings_.size(); ++i) {
    TextureBinding& binding = texture_bindings_[i];
    if (!binding.key.is_valid) {
      continue;
    }
    binding.Reset();
    bindings_reset |= UINT32_C(1) << i;
  }
  texture_bindings_in_sync_ &= ~bindings_reset;
  if (!from_destructor && bindings_reset) {
    UpdateTextureBindingsImpl(bindings_reset);
  }
}

void TextureCache::UpdateTexturesTotalHostMemoryUsage(uint64_t add,
                                                      uint64_t subtract) {
  textures_total_host_memory_usage_ =
      textures_total_host_memory_usage_ - subtract + add;
  COUNT_profile_set("gpu/texture_cache/total_host_memory_usage_mb",
                    uint32_t((textures_total_host_memory_usage_ +
                              ((UINT32_C(1) << 20) - 1)) >>
                             20));
}

bool TextureCache::IsRangeScaledResolved(uint32_t start_unscaled,
                                         uint32_t length_unscaled) {
  if (!IsDrawResolutionScaled()) {
    return false;
  }

  start_unscaled = std::min(start_unscaled, SharedMemory::kBufferSize);
  length_unscaled =
      std::min(length_unscaled, SharedMemory::kBufferSize - start_unscaled);
  if (!length_unscaled) {
    return false;
  }

  // Two-level check for faster rejection since resolve targets are usually
  // placed in relatively small and localized memory portions (confirmed by
  // testing - pretty much all times the deeper level was entered, the texture
  // was a resolve target).
  uint32_t page_first = start_unscaled >> 12;
  uint32_t page_last = (start_unscaled + length_unscaled - 1) >> 12;
  uint32_t block_first = page_first >> 5;
  uint32_t block_last = page_last >> 5;
  uint32_t l2_block_first = block_first >> 6;
  uint32_t l2_block_last = block_last >> 6;
  auto global_lock = global_critical_region_.Acquire();
  for (uint32_t i = l2_block_first; i <= l2_block_last; ++i) {
    uint64_t l2_block = scaled_resolve_pages_l2_[i];
    if (i == l2_block_first) {
      l2_block &= ~((UINT64_C(1) << (block_first & 63)) - 1);
    }
    if (i == l2_block_last && (block_last & 63) != 63) {
      l2_block &= (UINT64_C(1) << ((block_last & 63) + 1)) - 1;
    }
    uint32_t block_relative_index;
    while (xe::bit_scan_forward(l2_block, &block_relative_index)) {
      l2_block &= ~(UINT64_C(1) << block_relative_index);
      uint32_t block_index = (i << 6) + block_relative_index;
      uint32_t check_bits = UINT32_MAX;
      if (block_index == block_first) {
        check_bits &= ~((UINT32_C(1) << (page_first & 31)) - 1);
      }
      if (block_index == block_last && (page_last & 31) != 31) {
        check_bits &= (UINT32_C(1) << ((page_last & 31) + 1)) - 1;
      }
      if (scaled_resolve_pages_[block_index] & check_bits) {
        return true;
      }
    }
  }
  return false;
}

void TextureCache::ScaledResolveGlobalWatchCallbackThunk(
    const global_unique_lock_type& global_lock, void* context,
    uint32_t address_first, uint32_t address_last, bool invalidated_by_gpu) {
  TextureCache* texture_cache = reinterpret_cast<TextureCache*>(context);
  texture_cache->ScaledResolveGlobalWatchCallback(
      global_lock, address_first, address_last, invalidated_by_gpu);
}

void TextureCache::ScaledResolveGlobalWatchCallback(
    const global_unique_lock_type& global_lock, uint32_t address_first,
    uint32_t address_last, bool invalidated_by_gpu) {
  assert_true(IsDrawResolutionScaled());
  if (invalidated_by_gpu) {
    // Resolves themselves do exactly the opposite of what this should do.
    return;
  }
  // Mark scaled resolve ranges as non-scaled. Textures themselves will be
  // invalidated by their shared memory watches.
  uint32_t resolve_page_first = address_first >> 12;
  uint32_t resolve_page_last = address_last >> 12;
  uint32_t resolve_block_first = resolve_page_first >> 5;
  uint32_t resolve_block_last = resolve_page_last >> 5;
  uint32_t resolve_l2_block_first = resolve_block_first >> 6;
  uint32_t resolve_l2_block_last = resolve_block_last >> 6;
  for (uint32_t i = resolve_l2_block_first; i <= resolve_l2_block_last; ++i) {
    uint64_t resolve_l2_block = scaled_resolve_pages_l2_[i];
    uint32_t resolve_block_relative_index;
    while (
        xe::bit_scan_forward(resolve_l2_block, &resolve_block_relative_index)) {
      resolve_l2_block &= ~(UINT64_C(1) << resolve_block_relative_index);
      uint32_t resolve_block_index = (i << 6) + resolve_block_relative_index;
      uint32_t resolve_keep_bits = 0;
      if (resolve_block_index == resolve_block_first) {
        resolve_keep_bits |= (UINT32_C(1) << (resolve_page_first & 31)) - 1;
      }
      if (resolve_block_index == resolve_block_last &&
          (resolve_page_last & 31) != 31) {
        resolve_keep_bits |=
            ~((UINT32_C(1) << ((resolve_page_last & 31) + 1)) - 1);
      }
      scaled_resolve_pages_[resolve_block_index] &= resolve_keep_bits;
      if (scaled_resolve_pages_[resolve_block_index] == 0) {
        scaled_resolve_pages_l2_[i] &=
            ~(UINT64_C(1) << resolve_block_relative_index);
      }
    }
  }
}

}  // namespace gpu
}  // namespace xe