forked from ShuriZma/suyu
1
0
Fork 0

Merge pull request #2121 from FernandoS27/texception2

Improve the Accuracy of the Rasterizer Cache through a Texception Pass
This commit is contained in:
bunnei 2019-02-27 21:17:55 -05:00 committed by GitHub
commit 1f5d6a8fed
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
4 changed files with 213 additions and 16 deletions

View File

@ -129,6 +129,15 @@ protected:
return ++modified_ticks;
}
/// Flushes the specified object, updating appropriate cache state as needed
void FlushObject(const T& object) {
if (!object->IsDirty()) {
return;
}
object->Flush();
object->MarkAsModified(false, *this);
}
private:
/// Returns a list of cached objects from the specified memory region, ordered by access time
std::vector<T> GetSortedObjectsFromRegion(VAddr addr, u64 size) {
@ -154,15 +163,6 @@ private:
return objects;
}
/// Flushes the specified object, updating appropriate cache state as needed
void FlushObject(const T& object) {
if (!object->IsDirty()) {
return;
}
object->Flush();
object->MarkAsModified(false, *this);
}
using ObjectSet = std::set<T>;
using ObjectCache = std::unordered_map<VAddr, T>;
using IntervalCache = boost::icl::interval_map<VAddr, ObjectSet>;

View File

@ -738,9 +738,13 @@ void RasterizerOpenGL::DrawArrays() {
shader_program_manager->ApplyTo(state);
state.Apply();
res_cache.SignalPreDrawCall();
// Execute draw call
params.DispatchDraw();
res_cache.SignalPostDrawCall();
// Disable scissor test
state.viewports[0].scissor.enabled = false;

View File

@ -3,6 +3,7 @@
// Refer to the license.txt file included.
#include <algorithm>
#include <optional>
#include <glad/glad.h>
#include "common/alignment.h"
@ -549,6 +550,8 @@ CachedSurface::CachedSurface(const SurfaceParams& params)
// alternatives. This signals a bug on those functions.
const auto width = static_cast<GLsizei>(params.MipWidth(0));
const auto height = static_cast<GLsizei>(params.MipHeight(0));
memory_size = params.MemorySize();
reinterpreted = false;
const auto& format_tuple = GetFormatTuple(params.pixel_format, params.component_type);
gl_internal_format = format_tuple.internal_format;
@ -970,22 +973,23 @@ Surface RasterizerCacheOpenGL::GetColorBufferSurface(std::size_t index, bool pre
ASSERT(index < Tegra::Engines::Maxwell3D::Regs::NumRenderTargets);
if (index >= regs.rt_control.count) {
return last_color_buffers[index] = {};
return current_color_buffers[index] = {};
}
if (regs.rt[index].Address() == 0 || regs.rt[index].format == Tegra::RenderTargetFormat::NONE) {
return last_color_buffers[index] = {};
return current_color_buffers[index] = {};
}
const SurfaceParams color_params{SurfaceParams::CreateForFramebuffer(index)};
return last_color_buffers[index] = GetSurface(color_params, preserve_contents);
return current_color_buffers[index] = GetSurface(color_params, preserve_contents);
}
void RasterizerCacheOpenGL::LoadSurface(const Surface& surface) {
surface->LoadGLBuffer();
surface->UploadGLTexture(read_framebuffer.handle, draw_framebuffer.handle);
surface->MarkAsModified(false, *this);
surface->MarkForReload(false);
}
Surface RasterizerCacheOpenGL::GetSurface(const SurfaceParams& params, bool preserve_contents) {
@ -997,18 +1001,23 @@ Surface RasterizerCacheOpenGL::GetSurface(const SurfaceParams& params, bool pres
Surface surface{TryGet(params.addr)};
if (surface) {
if (surface->GetSurfaceParams().IsCompatibleSurface(params)) {
// Use the cached surface as-is
// Use the cached surface as-is unless it's not synced with memory
if (surface->MustReload())
LoadSurface(surface);
return surface;
} else if (preserve_contents) {
// If surface parameters changed and we care about keeping the previous data, recreate
// the surface from the old one
Surface new_surface{RecreateSurface(surface, params)};
Unregister(surface);
UnregisterSurface(surface);
Register(new_surface);
if (new_surface->IsUploaded()) {
RegisterReinterpretSurface(new_surface);
}
return new_surface;
} else {
// Delete the old surface before creating a new one to prevent collisions.
Unregister(surface);
UnregisterSurface(surface);
}
}
@ -1290,4 +1299,107 @@ Surface RasterizerCacheOpenGL::TryGetReservedSurface(const SurfaceParams& params
return {};
}
static std::optional<u32> TryFindBestMipMap(std::size_t memory, const SurfaceParams params,
u32 height) {
for (u32 i = 0; i < params.max_mip_level; i++) {
if (memory == params.GetMipmapSingleSize(i) && params.MipHeight(i) == height) {
return {i};
}
}
return {};
}
static std::optional<u32> TryFindBestLayer(VAddr addr, const SurfaceParams params, u32 mipmap) {
const std::size_t size = params.LayerMemorySize();
VAddr start = params.addr + params.GetMipmapLevelOffset(mipmap);
for (u32 i = 0; i < params.depth; i++) {
if (start == addr) {
return {i};
}
start += size;
}
return {};
}
static bool LayerFitReinterpretSurface(RasterizerCacheOpenGL& cache, const Surface render_surface,
const Surface blitted_surface) {
const auto& dst_params = blitted_surface->GetSurfaceParams();
const auto& src_params = render_surface->GetSurfaceParams();
const std::size_t src_memory_size = src_params.size_in_bytes;
const std::optional<u32> level =
TryFindBestMipMap(src_memory_size, dst_params, src_params.height);
if (level.has_value()) {
if (src_params.width == dst_params.MipWidthGobAligned(*level) &&
src_params.height == dst_params.MipHeight(*level) &&
src_params.block_height >= dst_params.MipBlockHeight(*level)) {
const std::optional<u32> slot =
TryFindBestLayer(render_surface->GetAddr(), dst_params, *level);
if (slot.has_value()) {
glCopyImageSubData(render_surface->Texture().handle,
SurfaceTargetToGL(src_params.target), 0, 0, 0, 0,
blitted_surface->Texture().handle,
SurfaceTargetToGL(dst_params.target), *level, 0, 0, *slot,
dst_params.MipWidth(*level), dst_params.MipHeight(*level), 1);
blitted_surface->MarkAsModified(true, cache);
return true;
}
}
}
return false;
}
static bool IsReinterpretInvalid(const Surface render_surface, const Surface blitted_surface) {
const VAddr bound1 = blitted_surface->GetAddr() + blitted_surface->GetMemorySize();
const VAddr bound2 = render_surface->GetAddr() + render_surface->GetMemorySize();
if (bound2 > bound1)
return true;
const auto& dst_params = blitted_surface->GetSurfaceParams();
const auto& src_params = render_surface->GetSurfaceParams();
return (dst_params.component_type != src_params.component_type);
}
static bool IsReinterpretInvalidSecond(const Surface render_surface,
const Surface blitted_surface) {
const auto& dst_params = blitted_surface->GetSurfaceParams();
const auto& src_params = render_surface->GetSurfaceParams();
return (dst_params.height > src_params.height && dst_params.width > src_params.width);
}
bool RasterizerCacheOpenGL::PartialReinterpretSurface(Surface triggering_surface,
Surface intersect) {
if (IsReinterpretInvalid(triggering_surface, intersect)) {
UnregisterSurface(intersect);
return false;
}
if (!LayerFitReinterpretSurface(*this, triggering_surface, intersect)) {
if (IsReinterpretInvalidSecond(triggering_surface, intersect)) {
UnregisterSurface(intersect);
return false;
}
FlushObject(intersect);
FlushObject(triggering_surface);
intersect->MarkForReload(true);
}
return true;
}
void RasterizerCacheOpenGL::SignalPreDrawCall() {
if (texception && GLAD_GL_ARB_texture_barrier) {
glTextureBarrier();
}
texception = false;
}
void RasterizerCacheOpenGL::SignalPostDrawCall() {
for (u32 i = 0; i < Maxwell::NumRenderTargets; i++) {
if (current_color_buffers[i] != nullptr) {
Surface intersect = CollideOnReinterpretedSurface(current_color_buffers[i]->GetAddr());
if (intersect != nullptr) {
PartialReinterpretSurface(current_color_buffers[i], intersect);
texception = true;
}
}
}
}
} // namespace OpenGL

View File

@ -34,6 +34,7 @@ using SurfaceTarget = VideoCore::Surface::SurfaceTarget;
using SurfaceType = VideoCore::Surface::SurfaceType;
using PixelFormat = VideoCore::Surface::PixelFormat;
using ComponentType = VideoCore::Surface::ComponentType;
using Maxwell = Tegra::Engines::Maxwell3D::Regs;
struct SurfaceParams {
enum class SurfaceClass {
@ -140,10 +141,18 @@ struct SurfaceParams {
return offset;
}
std::size_t GetMipmapSingleSize(u32 mip_level) const {
return InnerMipmapMemorySize(mip_level, false, is_layered);
}
u32 MipWidth(u32 mip_level) const {
return std::max(1U, width >> mip_level);
}
u32 MipWidthGobAligned(u32 mip_level) const {
return Common::AlignUp(std::max(1U, width >> mip_level), 64U * 8U / GetFormatBpp());
}
u32 MipHeight(u32 mip_level) const {
return std::max(1U, height >> mip_level);
}
@ -346,6 +355,10 @@ public:
return cached_size_in_bytes;
}
std::size_t GetMemorySize() const {
return memory_size;
}
void Flush() override {
FlushGLBuffer();
}
@ -395,6 +408,26 @@ public:
Tegra::Texture::SwizzleSource swizzle_z,
Tegra::Texture::SwizzleSource swizzle_w);
void MarkReinterpreted() {
reinterpreted = true;
}
bool IsReinterpreted() const {
return reinterpreted;
}
void MarkForReload(bool reload) {
must_reload = reload;
}
bool MustReload() const {
return must_reload;
}
bool IsUploaded() const {
return params.identity == SurfaceParams::SurfaceClass::Uploaded;
}
private:
void UploadGLMipmapTexture(u32 mip_map, GLuint read_fb_handle, GLuint draw_fb_handle);
@ -408,6 +441,9 @@ private:
GLenum gl_internal_format{};
std::size_t cached_size_in_bytes{};
std::array<GLenum, 4> swizzle{GL_RED, GL_GREEN, GL_BLUE, GL_ALPHA};
std::size_t memory_size;
bool reinterpreted = false;
bool must_reload = false;
};
class RasterizerCacheOpenGL final : public RasterizerCache<Surface> {
@ -433,6 +469,9 @@ public:
const Common::Rectangle<u32>& src_rect,
const Common::Rectangle<u32>& dst_rect);
void SignalPreDrawCall();
void SignalPostDrawCall();
private:
void LoadSurface(const Surface& surface);
Surface GetSurface(const SurfaceParams& params, bool preserve_contents = true);
@ -449,6 +488,10 @@ private:
/// Tries to get a reserved surface for the specified parameters
Surface TryGetReservedSurface(const SurfaceParams& params);
// Partialy reinterpret a surface based on a triggering_surface that collides with it.
// returns true if the reinterpret was successful, false in case it was not.
bool PartialReinterpretSurface(Surface triggering_surface, Surface intersect);
/// Performs a slow but accurate surface copy, flushing to RAM and reinterpreting the data
void AccurateCopySurface(const Surface& src_surface, const Surface& dst_surface);
void FastLayeredCopySurface(const Surface& src_surface, const Surface& dst_surface);
@ -465,12 +508,50 @@ private:
OGLFramebuffer read_framebuffer;
OGLFramebuffer draw_framebuffer;
bool texception = false;
/// Use a Pixel Buffer Object to download the previous texture and then upload it to the new one
/// using the new format.
OGLBuffer copy_pbo;
std::array<Surface, Tegra::Engines::Maxwell3D::Regs::NumRenderTargets> last_color_buffers;
std::array<Surface, Maxwell::NumRenderTargets> last_color_buffers;
std::array<Surface, Maxwell::NumRenderTargets> current_color_buffers;
Surface last_depth_buffer;
using SurfaceIntervalCache = boost::icl::interval_map<VAddr, Surface>;
using SurfaceInterval = typename SurfaceIntervalCache::interval_type;
static auto GetReinterpretInterval(const Surface& object) {
return SurfaceInterval::right_open(object->GetAddr() + 1,
object->GetAddr() + object->GetMemorySize() - 1);
}
// Reinterpreted surfaces are very fragil as the game may keep rendering into them.
SurfaceIntervalCache reinterpreted_surfaces;
void RegisterReinterpretSurface(Surface reinterpret_surface) {
auto interval = GetReinterpretInterval(reinterpret_surface);
reinterpreted_surfaces.insert({interval, reinterpret_surface});
reinterpret_surface->MarkReinterpreted();
}
Surface CollideOnReinterpretedSurface(VAddr addr) const {
const SurfaceInterval interval{addr};
for (auto& pair :
boost::make_iterator_range(reinterpreted_surfaces.equal_range(interval))) {
return pair.second;
}
return nullptr;
}
/// Unregisters an object from the cache
void UnregisterSurface(const Surface& object) {
if (object->IsReinterpreted()) {
auto interval = GetReinterpretInterval(object);
reinterpreted_surfaces.erase(interval);
}
Unregister(object);
}
};
} // namespace OpenGL