Merge pull request #9849 from ameerj/async-astc

texture_cache: Add asynchronous ASTC texture decoding
This commit is contained in:
liamwhite 2023-02-26 09:20:12 -05:00 committed by GitHub
commit 26c1edf2f0
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
15 changed files with 138 additions and 8 deletions

View File

@ -23,6 +23,7 @@ public:
buffer{Common::make_unique_for_overwrite<T[]>(initial_capacity)} {}
~ScratchBuffer() = default;
ScratchBuffer(ScratchBuffer&&) = default;
/// This will only grow the buffer's capacity if size is greater than the current capacity.
/// The previously held data will remain intact.

View File

@ -59,6 +59,7 @@ void LogSettings() {
values.use_asynchronous_gpu_emulation.GetValue());
log_setting("Renderer_NvdecEmulation", values.nvdec_emulation.GetValue());
log_setting("Renderer_AccelerateASTC", values.accelerate_astc.GetValue());
log_setting("Renderer_AsyncASTC", values.async_astc.GetValue());
log_setting("Renderer_UseVsync", values.use_vsync.GetValue());
log_setting("Renderer_ShaderBackend", values.shader_backend.GetValue());
log_setting("Renderer_UseAsynchronousShaders", values.use_asynchronous_shaders.GetValue());
@ -219,6 +220,7 @@ void RestoreGlobalState(bool is_powered_on) {
values.use_asynchronous_gpu_emulation.SetGlobal(true);
values.nvdec_emulation.SetGlobal(true);
values.accelerate_astc.SetGlobal(true);
values.async_astc.SetGlobal(true);
values.use_vsync.SetGlobal(true);
values.shader_backend.SetGlobal(true);
values.use_asynchronous_shaders.SetGlobal(true);

View File

@ -453,6 +453,7 @@ struct Values {
SwitchableSetting<bool> use_asynchronous_gpu_emulation{true, "use_asynchronous_gpu_emulation"};
SwitchableSetting<NvdecEmulation> nvdec_emulation{NvdecEmulation::GPU, "nvdec_emulation"};
SwitchableSetting<bool> accelerate_astc{true, "accelerate_astc"};
SwitchableSetting<bool> async_astc{false, "async_astc"};
SwitchableSetting<bool> use_vsync{true, "use_vsync"};
SwitchableSetting<ShaderBackend, true> shader_backend{ShaderBackend::GLSL, ShaderBackend::GLSL,
ShaderBackend::SPIRV, "shader_backend"};

View File

@ -228,8 +228,9 @@ void ApplySwizzle(GLuint handle, PixelFormat format, std::array<SwizzleSource, 4
[[nodiscard]] bool CanBeAccelerated(const TextureCacheRuntime& runtime,
const VideoCommon::ImageInfo& info) {
if (IsPixelFormatASTC(info.format)) {
return !runtime.HasNativeASTC() && Settings::values.accelerate_astc.GetValue();
if (IsPixelFormatASTC(info.format) && !runtime.HasNativeASTC()) {
return Settings::values.accelerate_astc.GetValue() &&
!Settings::values.async_astc.GetValue();
}
// Disable other accelerated uploads for now as they don't implement swizzled uploads
return false;
@ -258,6 +259,14 @@ void ApplySwizzle(GLuint handle, PixelFormat format, std::array<SwizzleSource, 4
return format_info.compatibility_class == store_class;
}
[[nodiscard]] bool CanBeDecodedAsync(const TextureCacheRuntime& runtime,
const VideoCommon::ImageInfo& info) {
if (IsPixelFormatASTC(info.format) && !runtime.HasNativeASTC()) {
return Settings::values.async_astc.GetValue();
}
return false;
}
[[nodiscard]] CopyOrigin MakeCopyOrigin(VideoCommon::Offset3D offset,
VideoCommon::SubresourceLayers subresource, GLenum target) {
switch (target) {
@ -721,7 +730,9 @@ std::optional<size_t> TextureCacheRuntime::StagingBuffers::FindBuffer(size_t req
Image::Image(TextureCacheRuntime& runtime_, const VideoCommon::ImageInfo& info_, GPUVAddr gpu_addr_,
VAddr cpu_addr_)
: VideoCommon::ImageBase(info_, gpu_addr_, cpu_addr_), runtime{&runtime_} {
if (CanBeAccelerated(*runtime, info)) {
if (CanBeDecodedAsync(*runtime, info)) {
flags |= ImageFlagBits::AsynchronousDecode;
} else if (CanBeAccelerated(*runtime, info)) {
flags |= ImageFlagBits::AcceleratedUpload;
}
if (IsConverted(runtime->device, info.format, info.type)) {

View File

@ -1256,11 +1256,12 @@ Image::Image(TextureCacheRuntime& runtime_, const ImageInfo& info_, GPUVAddr gpu
commit(runtime_.memory_allocator.Commit(original_image, MemoryUsage::DeviceLocal)),
aspect_mask(ImageAspectMask(info.format)) {
if (IsPixelFormatASTC(info.format) && !runtime->device.IsOptimalAstcSupported()) {
if (Settings::values.accelerate_astc.GetValue()) {
if (Settings::values.async_astc.GetValue()) {
flags |= VideoCommon::ImageFlagBits::AsynchronousDecode;
} else if (Settings::values.accelerate_astc.GetValue()) {
flags |= VideoCommon::ImageFlagBits::AcceleratedUpload;
} else {
flags |= VideoCommon::ImageFlagBits::Converted;
}
flags |= VideoCommon::ImageFlagBits::Converted;
flags |= VideoCommon::ImageFlagBits::CostlyLoad;
}
if (runtime->device.HasDebuggingToolAttached()) {

View File

@ -38,6 +38,9 @@ enum class ImageFlagBits : u32 {
Rescaled = 1 << 13,
CheckingRescalable = 1 << 14,
IsRescalable = 1 << 15,
AsynchronousDecode = 1 << 16,
IsDecoding = 1 << 17, ///< Is currently being decoded asynchornously.
};
DECLARE_ENUM_FLAG_OPERATORS(ImageFlagBits)

View File

@ -85,6 +85,11 @@ void TextureCache<P>::RunGarbageCollector() {
}
--num_iterations;
auto& image = slot_images[image_id];
if (True(image.flags & ImageFlagBits::IsDecoding)) {
// This image is still being decoded, deleting it will invalidate the slot
// used by the async decoder thread.
return false;
}
const bool must_download =
image.IsSafeDownload() && False(image.flags & ImageFlagBits::BadOverlap);
if (!high_priority_mode &&
@ -133,6 +138,8 @@ void TextureCache<P>::TickFrame() {
sentenced_images.Tick();
sentenced_framebuffers.Tick();
sentenced_image_view.Tick();
TickAsyncDecode();
runtime.TickFrame();
critical_gc = 0;
++frame_tick;
@ -777,6 +784,10 @@ void TextureCache<P>::RefreshContents(Image& image, ImageId image_id) {
LOG_WARNING(HW_GPU, "MSAA image uploads are not implemented");
return;
}
if (True(image.flags & ImageFlagBits::AsynchronousDecode)) {
QueueAsyncDecode(image, image_id);
return;
}
auto staging = runtime.UploadStagingBuffer(MapSizeBytes(image));
UploadImageContents(image, staging);
runtime.InsertUploadMemoryBarrier();
@ -989,6 +1000,65 @@ u64 TextureCache<P>::GetScaledImageSizeBytes(const ImageBase& image) {
return fitted_size;
}
template <class P>
void TextureCache<P>::QueueAsyncDecode(Image& image, ImageId image_id) {
UNIMPLEMENTED_IF(False(image.flags & ImageFlagBits::Converted));
LOG_INFO(HW_GPU, "Queuing async texture decode");
image.flags |= ImageFlagBits::IsDecoding;
auto decode = std::make_unique<AsyncDecodeContext>();
auto* decode_ptr = decode.get();
decode->image_id = image_id;
async_decodes.push_back(std::move(decode));
Common::ScratchBuffer<u8> local_unswizzle_data_buffer(image.unswizzled_size_bytes);
const size_t guest_size_bytes = image.guest_size_bytes;
swizzle_data_buffer.resize_destructive(guest_size_bytes);
gpu_memory->ReadBlockUnsafe(image.gpu_addr, swizzle_data_buffer.data(), guest_size_bytes);
auto copies = UnswizzleImage(*gpu_memory, image.gpu_addr, image.info, swizzle_data_buffer,
local_unswizzle_data_buffer);
const size_t out_size = MapSizeBytes(image);
auto func = [out_size, copies, info = image.info,
input = std::move(local_unswizzle_data_buffer),
async_decode = decode_ptr]() mutable {
async_decode->decoded_data.resize_destructive(out_size);
std::span copies_span{copies.data(), copies.size()};
ConvertImage(input, info, async_decode->decoded_data, copies_span);
// TODO: Do we need this lock?
std::unique_lock lock{async_decode->mutex};
async_decode->copies = std::move(copies);
async_decode->complete = true;
};
texture_decode_worker.QueueWork(std::move(func));
}
template <class P>
void TextureCache<P>::TickAsyncDecode() {
bool has_uploads{};
auto i = async_decodes.begin();
while (i != async_decodes.end()) {
auto* async_decode = i->get();
std::unique_lock lock{async_decode->mutex};
if (!async_decode->complete) {
++i;
continue;
}
Image& image = slot_images[async_decode->image_id];
auto staging = runtime.UploadStagingBuffer(MapSizeBytes(image));
std::memcpy(staging.mapped_span.data(), async_decode->decoded_data.data(),
async_decode->decoded_data.size());
image.UploadMemory(staging, async_decode->copies);
image.flags &= ~ImageFlagBits::IsDecoding;
has_uploads = true;
i = async_decodes.erase(i);
}
if (has_uploads) {
runtime.InsertUploadMemoryBarrier();
}
}
template <class P>
bool TextureCache<P>::ScaleUp(Image& image) {
const bool has_copy = image.HasScaled();

View File

@ -3,6 +3,7 @@
#pragma once
#include <atomic>
#include <deque>
#include <limits>
#include <mutex>
@ -18,6 +19,7 @@
#include "common/lru_cache.h"
#include "common/polyfill_ranges.h"
#include "common/scratch_buffer.h"
#include "common/thread_worker.h"
#include "video_core/compatible_formats.h"
#include "video_core/control/channel_state_cache.h"
#include "video_core/delayed_destruction_ring.h"
@ -54,6 +56,14 @@ struct ImageViewInOut {
ImageViewId id{};
};
struct AsyncDecodeContext {
ImageId image_id;
Common::ScratchBuffer<u8> decoded_data;
std::vector<BufferImageCopy> copies;
std::mutex mutex;
std::atomic_bool complete;
};
using TextureCacheGPUMap = std::unordered_map<u64, std::vector<ImageId>, Common::IdentityHash<u64>>;
class TextureCacheChannelInfo : public ChannelInfo {
@ -377,6 +387,9 @@ private:
bool ScaleDown(Image& image);
u64 GetScaledImageSizeBytes(const ImageBase& image);
void QueueAsyncDecode(Image& image, ImageId image_id);
void TickAsyncDecode();
Runtime& runtime;
VideoCore::RasterizerInterface& rasterizer;
@ -430,6 +443,9 @@ private:
u64 modification_tick = 0;
u64 frame_tick = 0;
Common::ThreadWorker texture_decode_worker{1, "TextureDecoder"};
std::vector<std::unique_ptr<AsyncDecodeContext>> async_decodes;
};
} // namespace VideoCommon

View File

@ -1656,7 +1656,7 @@ void Decompress(std::span<const uint8_t> data, uint32_t width, uint32_t height,
const u32 rows = Common::DivideUp(height, block_height);
const u32 cols = Common::DivideUp(width, block_width);
Common::ThreadWorker workers{std::max(std::thread::hardware_concurrency(), 2U) / 2,
static Common::ThreadWorker workers{std::max(std::thread::hardware_concurrency(), 2U) / 2,
"ASTCDecompress"};
for (u32 z = 0; z < depth; ++z) {

View File

@ -702,6 +702,7 @@ void Config::ReadRendererValues() {
ReadGlobalSetting(Settings::values.use_asynchronous_gpu_emulation);
ReadGlobalSetting(Settings::values.nvdec_emulation);
ReadGlobalSetting(Settings::values.accelerate_astc);
ReadGlobalSetting(Settings::values.async_astc);
ReadGlobalSetting(Settings::values.use_vsync);
ReadGlobalSetting(Settings::values.shader_backend);
ReadGlobalSetting(Settings::values.use_asynchronous_shaders);
@ -1343,6 +1344,7 @@ void Config::SaveRendererValues() {
static_cast<u32>(Settings::values.nvdec_emulation.GetDefault()),
Settings::values.nvdec_emulation.UsingGlobal());
WriteGlobalSetting(Settings::values.accelerate_astc);
WriteGlobalSetting(Settings::values.async_astc);
WriteGlobalSetting(Settings::values.use_vsync);
WriteSetting(QString::fromStdString(Settings::values.shader_backend.GetLabel()),
static_cast<u32>(Settings::values.shader_backend.GetValue(global)),

View File

@ -23,11 +23,13 @@ void ConfigureGraphicsAdvanced::SetConfiguration() {
const bool runtime_lock = !system.IsPoweredOn();
ui->use_vsync->setEnabled(runtime_lock);
ui->renderer_force_max_clock->setEnabled(runtime_lock);
ui->async_astc->setEnabled(runtime_lock);
ui->use_asynchronous_shaders->setEnabled(runtime_lock);
ui->anisotropic_filtering_combobox->setEnabled(runtime_lock);
ui->renderer_force_max_clock->setChecked(Settings::values.renderer_force_max_clock.GetValue());
ui->use_vsync->setChecked(Settings::values.use_vsync.GetValue());
ui->async_astc->setChecked(Settings::values.async_astc.GetValue());
ui->use_asynchronous_shaders->setChecked(Settings::values.use_asynchronous_shaders.GetValue());
ui->use_fast_gpu_time->setChecked(Settings::values.use_fast_gpu_time.GetValue());
ui->use_pessimistic_flushes->setChecked(Settings::values.use_pessimistic_flushes.GetValue());
@ -58,6 +60,8 @@ void ConfigureGraphicsAdvanced::ApplyConfiguration() {
ConfigurationShared::ApplyPerGameSetting(&Settings::values.max_anisotropy,
ui->anisotropic_filtering_combobox);
ConfigurationShared::ApplyPerGameSetting(&Settings::values.use_vsync, ui->use_vsync, use_vsync);
ConfigurationShared::ApplyPerGameSetting(&Settings::values.async_astc, ui->async_astc,
async_astc);
ConfigurationShared::ApplyPerGameSetting(&Settings::values.use_asynchronous_shaders,
ui->use_asynchronous_shaders,
use_asynchronous_shaders);
@ -89,6 +93,7 @@ void ConfigureGraphicsAdvanced::SetupPerGameUI() {
ui->renderer_force_max_clock->setEnabled(
Settings::values.renderer_force_max_clock.UsingGlobal());
ui->use_vsync->setEnabled(Settings::values.use_vsync.UsingGlobal());
ui->async_astc->setEnabled(Settings::values.async_astc.UsingGlobal());
ui->use_asynchronous_shaders->setEnabled(
Settings::values.use_asynchronous_shaders.UsingGlobal());
ui->use_fast_gpu_time->setEnabled(Settings::values.use_fast_gpu_time.UsingGlobal());
@ -106,6 +111,8 @@ void ConfigureGraphicsAdvanced::SetupPerGameUI() {
Settings::values.renderer_force_max_clock,
renderer_force_max_clock);
ConfigurationShared::SetColoredTristate(ui->use_vsync, Settings::values.use_vsync, use_vsync);
ConfigurationShared::SetColoredTristate(ui->async_astc, Settings::values.async_astc,
async_astc);
ConfigurationShared::SetColoredTristate(ui->use_asynchronous_shaders,
Settings::values.use_asynchronous_shaders,
use_asynchronous_shaders);

View File

@ -38,6 +38,7 @@ private:
ConfigurationShared::CheckState renderer_force_max_clock;
ConfigurationShared::CheckState use_vsync;
ConfigurationShared::CheckState async_astc;
ConfigurationShared::CheckState use_asynchronous_shaders;
ConfigurationShared::CheckState use_fast_gpu_time;
ConfigurationShared::CheckState use_pessimistic_flushes;

View File

@ -89,6 +89,16 @@
</property>
</widget>
</item>
<item>
<widget class="QCheckBox" name="async_astc">
<property name="toolTip">
<string>Enables asynchronous ASTC texture decoding, which may reduce load time stutter. This feature is experimental.</string>
</property>
<property name="text">
<string>Decode ASTC textures asynchronously (Hack)</string>
</property>
</widget>
</item>
<item>
<widget class="QCheckBox" name="use_asynchronous_shaders">
<property name="toolTip">

View File

@ -324,6 +324,7 @@ void Config::ReadValues() {
ReadSetting("Renderer", Settings::values.use_asynchronous_shaders);
ReadSetting("Renderer", Settings::values.nvdec_emulation);
ReadSetting("Renderer", Settings::values.accelerate_astc);
ReadSetting("Renderer", Settings::values.async_astc);
ReadSetting("Renderer", Settings::values.use_fast_gpu_time);
ReadSetting("Renderer", Settings::values.use_pessimistic_flushes);
ReadSetting("Renderer", Settings::values.use_vulkan_driver_pipeline_cache);

View File

@ -342,6 +342,10 @@ nvdec_emulation =
# 0: Off, 1 (default): On
accelerate_astc =
# Decode ASTC textures asynchronously.
# 0 (default): Off, 1: On
async_astc =
# Turns on the speed limiter, which will limit the emulation speed to the desired speed limit value
# 0: Off, 1: On (default)
use_speed_limit =