TextureCache: Remove distinction between resolve textures and regular textures.

This commit is contained in:
Dr. Chat 2017-03-14 23:18:53 -05:00
parent 4a815e62cb
commit b34350e9e7
3 changed files with 32 additions and 146 deletions

View File

@ -430,19 +430,26 @@ bool TextureCache::FreeTexture(Texture* texture) {
} }
TextureCache::Texture* TextureCache::DemandResolveTexture( TextureCache::Texture* TextureCache::DemandResolveTexture(
const TextureInfo& texture_info, TextureFormat format, const TextureInfo& texture_info, TextureFormat format) {
VkOffset2D* out_offset) { auto texture_hash = texture_info.hash();
// Check to see if we've already used a texture at this location. for (auto it = textures_.find(texture_hash); it != textures_.end(); ++it) {
auto texture = LookupAddress( if (it->second->texture_info == texture_info) {
texture_info.guest_address, texture_info.size_2d.block_width, if (it->second->pending_invalidation) {
texture_info.size_2d.block_height, format, out_offset); // This texture has been invalidated!
if (texture) { RemoveInvalidatedTextures();
return texture; break;
}
// Tell the trace writer to "cache" this memory (but not read it)
trace_writer_->WriteMemoryReadCachedNop(texture_info.guest_address,
texture_info.input_length);
return it->second;
}
} }
// No texture at this location. Make a new one. // No texture at this location. Make a new one.
texture = AllocateTexture(texture_info); auto texture = AllocateTexture(texture_info);
texture->is_full_texture = false;
// Setup a debug name for the texture. // Setup a debug name for the texture.
device_->DbgSetObjectName( device_->DbgSetObjectName(
@ -461,16 +468,18 @@ TextureCache::Texture* TextureCache::DemandResolveTexture(
auto touched_texture = reinterpret_cast<Texture*>(data_ptr); auto touched_texture = reinterpret_cast<Texture*>(data_ptr);
// Clear watch handle first so we don't redundantly // Clear watch handle first so we don't redundantly
// remove. // remove.
assert_not_zero(touched_texture->access_watch_handle);
touched_texture->access_watch_handle = 0; touched_texture->access_watch_handle = 0;
touched_texture->pending_invalidation = true; touched_texture->pending_invalidation = true;
// Add to pending list so Scavenge will clean it up. // Add to pending list so Scavenge will clean it up.
self->invalidated_resolve_textures_mutex_.lock(); self->invalidated_textures_mutex_.lock();
self->invalidated_resolve_textures_.push_back(touched_texture); self->invalidated_textures_->push_back(touched_texture);
self->invalidated_resolve_textures_mutex_.unlock(); self->invalidated_textures_mutex_.unlock();
}, },
this, texture); this, texture);
resolve_textures_.push_back(texture); textures_[texture_hash] = texture;
return texture; return texture;
} }
@ -494,58 +503,6 @@ TextureCache::Texture* TextureCache::Demand(const TextureInfo& texture_info,
} }
} }
// Check resolve textures.
for (auto it = resolve_textures_.begin(); it != resolve_textures_.end();
++it) {
auto texture = (*it);
if (texture_info.guest_address == texture->texture_info.guest_address &&
texture_info.size_2d.logical_width ==
texture->texture_info.size_2d.logical_width &&
texture_info.size_2d.logical_height ==
texture->texture_info.size_2d.logical_height) {
if (texture->pending_invalidation) {
// Texture invalidated! Remove.
RemoveInvalidatedTextures();
break;
}
// Exact match.
// TODO: Lazy match (at an offset)
// Upgrade this texture to a full texture.
texture->is_full_texture = true;
texture->texture_info = texture_info;
if (texture->access_watch_handle) {
memory_->CancelAccessWatch(texture->access_watch_handle);
}
// Tell the trace writer to cache this memory but don't read it
trace_writer_->WriteMemoryReadCachedNop(texture_info.guest_address,
texture_info.input_length);
texture->access_watch_handle = memory_->AddPhysicalAccessWatch(
texture_info.guest_address, texture_info.input_length,
cpu::MMIOHandler::kWatchWrite,
[](void* context_ptr, void* data_ptr, uint32_t address) {
auto self = reinterpret_cast<TextureCache*>(context_ptr);
auto touched_texture = reinterpret_cast<Texture*>(data_ptr);
// Clear watch handle first so we don't redundantly
// remove.
touched_texture->access_watch_handle = 0;
touched_texture->pending_invalidation = true;
// Add to pending list so Scavenge will clean it up.
self->invalidated_textures_mutex_.lock();
self->invalidated_textures_->push_back(touched_texture);
self->invalidated_textures_mutex_.unlock();
},
this, texture);
textures_[texture_hash] = *it;
it = resolve_textures_.erase(it);
return textures_[texture_hash];
}
}
if (!command_buffer) { if (!command_buffer) {
// Texture not found and no command buffer was passed, preventing us from // Texture not found and no command buffer was passed, preventing us from
// uploading a new one. // uploading a new one.
@ -598,22 +555,6 @@ TextureCache::Texture* TextureCache::Demand(const TextureInfo& texture_info,
"0x%.8X - 0x%.8X", texture_info.guest_address, "0x%.8X - 0x%.8X", texture_info.guest_address,
texture_info.guest_address + texture_info.output_length)); texture_info.guest_address + texture_info.output_length));
// Copy in overlapping resolve textures.
// FIXME: RDR appears to take textures from small chunks of a resolve texture?
if (texture_info.dimension == Dimension::k2D) {
for (auto it = resolve_textures_.begin(); it != resolve_textures_.end();
++it) {
auto texture = (*it);
if (texture_info.guest_address >= texture->texture_info.guest_address &&
texture_info.guest_address < texture->texture_info.guest_address +
texture->texture_info.input_length) {
// Lazy matched a resolve texture. Copy it in and destroy it.
// Future resolves will just copy directly into this texture.
// assert_always();
}
}
}
// Though we didn't find an exact match, that doesn't mean we're out of the // Though we didn't find an exact match, that doesn't mean we're out of the
// woods yet. This texture could either be a portion of another texture or // woods yet. This texture could either be a portion of another texture or
// vice versa. Copy any overlapping textures into this texture. // vice versa. Copy any overlapping textures into this texture.
@ -634,6 +575,7 @@ TextureCache::Texture* TextureCache::Demand(const TextureInfo& texture_info,
auto touched_texture = reinterpret_cast<Texture*>(data_ptr); auto touched_texture = reinterpret_cast<Texture*>(data_ptr);
// Clear watch handle first so we don't redundantly // Clear watch handle first so we don't redundantly
// remove. // remove.
assert_not_zero(touched_texture->access_watch_handle);
touched_texture->access_watch_handle = 0; touched_texture->access_watch_handle = 0;
touched_texture->pending_invalidation = true; touched_texture->pending_invalidation = true;
// Add to pending list so Scavenge will clean it up. // Add to pending list so Scavenge will clean it up.
@ -693,8 +635,9 @@ TextureCache::TextureView* TextureCache::DemandView(Texture* texture,
swiz_component_map[(swizzle >> 9) & 0x7], swiz_component_map[(swizzle >> 9) & 0x7],
}; };
view_info.subresourceRange = {VK_IMAGE_ASPECT_COLOR_BIT, 0, 1, 0, 1}; view_info.subresourceRange = {VK_IMAGE_ASPECT_COLOR_BIT, 0, 1, 0, 1};
if (texture->format == VK_FORMAT_D24_UNORM_S8_UINT) { if (texture->format == VK_FORMAT_D24_UNORM_S8_UINT ||
// This applies to any depth/stencil format, but we only use D24S8. texture->format == VK_FORMAT_D32_SFLOAT_S8_UINT) {
// This applies to any depth/stencil format, but we only use D24S8 / D32FS8.
view_info.subresourceRange.aspectMask = VK_IMAGE_ASPECT_DEPTH_BIT; view_info.subresourceRange.aspectMask = VK_IMAGE_ASPECT_DEPTH_BIT;
} }
@ -804,7 +747,7 @@ TextureCache::Sampler* TextureCache::Demand(const SamplerInfo& sampler_info) {
sampler_create_info.addressModeW = sampler_create_info.addressModeW =
address_mode_map[static_cast<int>(sampler_info.clamp_w)]; address_mode_map[static_cast<int>(sampler_info.clamp_w)];
sampler_create_info.mipLodBias = 0.0f; sampler_create_info.mipLodBias = sampler_info.lod_bias;
float aniso = 0.f; float aniso = 0.f;
switch (sampler_info.aniso_filter) { switch (sampler_info.aniso_filter) {
@ -895,23 +838,6 @@ TextureCache::Texture* TextureCache::LookupAddress(uint32_t guest_address,
} }
} }
// Check resolve textures
for (auto it = resolve_textures_.begin(); it != resolve_textures_.end();
++it) {
const auto& texture_info = (*it)->texture_info;
if (texture_info.guest_address == guest_address &&
texture_info.dimension == Dimension::k2D &&
texture_info.size_2d.input_width == width &&
texture_info.size_2d.input_height == height) {
if (out_offset) {
out_offset->x = 0;
out_offset->y = 0;
}
return (*it);
}
}
return nullptr; return nullptr;
} }
@ -1239,8 +1165,7 @@ bool TextureCache::UploadTexture2D(VkCommandBuffer command_buffer,
barrier.sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER; barrier.sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER;
barrier.pNext = nullptr; barrier.pNext = nullptr;
barrier.srcAccessMask = 0; barrier.srcAccessMask = 0;
barrier.dstAccessMask = barrier.dstAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT;
VK_ACCESS_TRANSFER_WRITE_BIT | VK_ACCESS_HOST_WRITE_BIT;
barrier.oldLayout = dest->image_layout; barrier.oldLayout = dest->image_layout;
barrier.newLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL; barrier.newLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
barrier.srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED; barrier.srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
@ -1419,7 +1344,6 @@ VkDescriptorSet TextureCache::PrepareTextureSet(
descriptor_pool_->BeginBatch(completion_fence); descriptor_pool_->BeginBatch(completion_fence);
} }
// TODO(benvanik): reuse.
auto descriptor_set = auto descriptor_set =
descriptor_pool_->AcquireEntry(texture_descriptor_set_layout_); descriptor_pool_->AcquireEntry(texture_descriptor_set_layout_);
if (!descriptor_set) { if (!descriptor_set) {
@ -1557,24 +1481,6 @@ void TextureCache::RemoveInvalidatedTextures() {
invalidated_textures.clear(); invalidated_textures.clear();
} }
// Invalidated resolve textures.
invalidated_resolve_textures_mutex_.lock();
if (!invalidated_resolve_textures_.empty()) {
for (auto it = invalidated_resolve_textures_.begin();
it != invalidated_resolve_textures_.end(); ++it) {
pending_delete_textures_.push_back(*it);
auto tex =
std::find(resolve_textures_.begin(), resolve_textures_.end(), *it);
if (tex != resolve_textures_.end()) {
resolve_textures_.erase(tex);
}
}
invalidated_resolve_textures_.clear();
}
invalidated_resolve_textures_mutex_.unlock();
} }
void TextureCache::ClearCache() { void TextureCache::ClearCache() {

View File

@ -38,10 +38,6 @@ class TextureCache {
TextureInfo texture_info; TextureInfo texture_info;
std::vector<std::unique_ptr<TextureView>> views; std::vector<std::unique_ptr<TextureView>> views;
// True if we know all info about this texture, false otherwise.
// (e.g. we resolve to system memory and may not know the full details about
// this texture)
bool is_full_texture;
VkFormat format; VkFormat format;
VkImage image; VkImage image;
VkImageLayout image_layout; VkImageLayout image_layout;
@ -93,8 +89,6 @@ class TextureCache {
const std::vector<Shader::TextureBinding>& vertex_bindings, const std::vector<Shader::TextureBinding>& vertex_bindings,
const std::vector<Shader::TextureBinding>& pixel_bindings); const std::vector<Shader::TextureBinding>& pixel_bindings);
// TODO(benvanik): UploadTexture.
// TODO(benvanik): Resolve.
// TODO(benvanik): ReadTexture. // TODO(benvanik): ReadTexture.
// Looks for a texture either containing or matching these parameters. // Looks for a texture either containing or matching these parameters.
@ -107,17 +101,9 @@ class TextureCache {
VkOffset2D* out_offset = nullptr); VkOffset2D* out_offset = nullptr);
// Demands a texture for the purpose of resolving from EDRAM. This either // Demands a texture for the purpose of resolving from EDRAM. This either
// creates a new texture or returns a previously created texture. texture_info // creates a new texture or returns a previously created texture.
// is not required to be completely filled out, just guest_address and all
// sizes.
//
// It's possible that this may return an image that is larger than the
// requested size (e.g. resolving into a bigger texture) or an image that
// must have an offset applied. If so, the caller must handle this.
// At the very least, it's guaranteed that the image will be large enough to
// hold the requested size.
Texture* DemandResolveTexture(const TextureInfo& texture_info, Texture* DemandResolveTexture(const TextureInfo& texture_info,
TextureFormat format, VkOffset2D* out_offset); TextureFormat format);
// Clears all cached content. // Clears all cached content.
void ClearCache(); void ClearCache();
@ -203,16 +189,12 @@ class TextureCache {
ui::vulkan::CircularBuffer staging_buffer_; ui::vulkan::CircularBuffer staging_buffer_;
std::unordered_map<uint64_t, Texture*> textures_; std::unordered_map<uint64_t, Texture*> textures_;
std::unordered_map<uint64_t, Sampler*> samplers_; std::unordered_map<uint64_t, Sampler*> samplers_;
std::vector<Texture*> resolve_textures_;
std::list<Texture*> pending_delete_textures_; std::list<Texture*> pending_delete_textures_;
std::mutex invalidated_textures_mutex_; std::mutex invalidated_textures_mutex_;
std::vector<Texture*>* invalidated_textures_; std::vector<Texture*>* invalidated_textures_;
std::vector<Texture*> invalidated_textures_sets_[2]; std::vector<Texture*> invalidated_textures_sets_[2];
std::mutex invalidated_resolve_textures_mutex_;
std::vector<Texture*> invalidated_resolve_textures_;
struct UpdateSetInfo { struct UpdateSetInfo {
// Bitmap of all 32 fetch constants and whether they have been setup yet. // Bitmap of all 32 fetch constants and whether they have been setup yet.
// This prevents duplication across the vertex and pixel shader. // This prevents duplication across the vertex and pixel shader.

View File

@ -850,8 +850,6 @@ bool VulkanCommandProcessor::IssueCopy() {
// but I can't seem to find something similar. // but I can't seem to find something similar.
uint32_t dest_logical_width = copy_dest_pitch; uint32_t dest_logical_width = copy_dest_pitch;
uint32_t dest_logical_height = copy_dest_height; uint32_t dest_logical_height = copy_dest_height;
uint32_t dest_block_width = xe::round_up(dest_logical_width, 32);
uint32_t dest_block_height = /*xe::round_up(*/ dest_logical_height /*, 32)*/;
uint32_t window_offset = regs[XE_GPU_REG_PA_SC_WINDOW_OFFSET].u32; uint32_t window_offset = regs[XE_GPU_REG_PA_SC_WINDOW_OFFSET].u32;
int16_t window_offset_x = window_offset & 0x7FFF; int16_t window_offset_x = window_offset & 0x7FFF;
@ -951,7 +949,7 @@ bool VulkanCommandProcessor::IssueCopy() {
&tex_info); &tex_info);
auto texture = auto texture =
texture_cache_->DemandResolveTexture(tex_info, copy_dest_format, nullptr); texture_cache_->DemandResolveTexture(tex_info, copy_dest_format);
assert_not_null(texture); assert_not_null(texture);
texture->in_flight_fence = current_batch_fence_; texture->in_flight_fence = current_batch_fence_;