Merge pull request #1121 from Subv/tex_reinterpret
Rasterizer: Use PBOs to reinterpret texture formats when games re-use the same memory.
This commit is contained in:
commit
ca58929eb0
|
@ -304,7 +304,8 @@ bool RasterizerOpenGL::AccelerateDrawBatch(bool is_indexed) {
|
||||||
}
|
}
|
||||||
|
|
||||||
std::pair<Surface, Surface> RasterizerOpenGL::ConfigureFramebuffers(bool using_color_fb,
|
std::pair<Surface, Surface> RasterizerOpenGL::ConfigureFramebuffers(bool using_color_fb,
|
||||||
bool using_depth_fb) {
|
bool using_depth_fb,
|
||||||
|
bool preserve_contents) {
|
||||||
const auto& regs = Core::System::GetInstance().GPU().Maxwell3D().regs;
|
const auto& regs = Core::System::GetInstance().GPU().Maxwell3D().regs;
|
||||||
|
|
||||||
if (regs.rt[0].format == Tegra::RenderTargetFormat::NONE) {
|
if (regs.rt[0].format == Tegra::RenderTargetFormat::NONE) {
|
||||||
|
@ -327,7 +328,7 @@ std::pair<Surface, Surface> RasterizerOpenGL::ConfigureFramebuffers(bool using_c
|
||||||
Surface depth_surface;
|
Surface depth_surface;
|
||||||
MathUtil::Rectangle<u32> surfaces_rect;
|
MathUtil::Rectangle<u32> surfaces_rect;
|
||||||
std::tie(color_surface, depth_surface, surfaces_rect) =
|
std::tie(color_surface, depth_surface, surfaces_rect) =
|
||||||
res_cache.GetFramebufferSurfaces(using_color_fb, using_depth_fb);
|
res_cache.GetFramebufferSurfaces(using_color_fb, using_depth_fb, preserve_contents);
|
||||||
|
|
||||||
const MathUtil::Rectangle<s32> viewport_rect{regs.viewport_transform[0].GetRect()};
|
const MathUtil::Rectangle<s32> viewport_rect{regs.viewport_transform[0].GetRect()};
|
||||||
const MathUtil::Rectangle<u32> draw_rect{
|
const MathUtil::Rectangle<u32> draw_rect{
|
||||||
|
@ -390,7 +391,7 @@ void RasterizerOpenGL::Clear() {
|
||||||
ScopeAcquireGLContext acquire_context{emu_window};
|
ScopeAcquireGLContext acquire_context{emu_window};
|
||||||
|
|
||||||
auto [dirty_color_surface, dirty_depth_surface] =
|
auto [dirty_color_surface, dirty_depth_surface] =
|
||||||
ConfigureFramebuffers(use_color_fb, use_depth_fb);
|
ConfigureFramebuffers(use_color_fb, use_depth_fb, false);
|
||||||
|
|
||||||
// TODO(Subv): Support clearing only partial colors.
|
// TODO(Subv): Support clearing only partial colors.
|
||||||
glClearColor(regs.clear_color[0], regs.clear_color[1], regs.clear_color[2],
|
glClearColor(regs.clear_color[0], regs.clear_color[1], regs.clear_color[2],
|
||||||
|
@ -445,7 +446,7 @@ void RasterizerOpenGL::DrawArrays() {
|
||||||
ScopeAcquireGLContext acquire_context{emu_window};
|
ScopeAcquireGLContext acquire_context{emu_window};
|
||||||
|
|
||||||
auto [dirty_color_surface, dirty_depth_surface] =
|
auto [dirty_color_surface, dirty_depth_surface] =
|
||||||
ConfigureFramebuffers(true, regs.zeta.Address() != 0 && regs.zeta_enable != 0);
|
ConfigureFramebuffers(true, regs.zeta.Address() != 0 && regs.zeta_enable != 0, true);
|
||||||
|
|
||||||
SyncDepthTestState();
|
SyncDepthTestState();
|
||||||
SyncBlendState();
|
SyncBlendState();
|
||||||
|
|
|
@ -87,7 +87,8 @@ private:
|
||||||
|
|
||||||
/// Configures the color and depth framebuffer states and returns the dirty <Color, Depth>
|
/// Configures the color and depth framebuffer states and returns the dirty <Color, Depth>
|
||||||
/// surfaces if writing was enabled.
|
/// surfaces if writing was enabled.
|
||||||
std::pair<Surface, Surface> ConfigureFramebuffers(bool using_color_fb, bool using_depth_fb);
|
std::pair<Surface, Surface> ConfigureFramebuffers(bool using_color_fb, bool using_depth_fb,
|
||||||
|
bool preserve_contents);
|
||||||
|
|
||||||
/// Binds the framebuffer color and depth surface
|
/// Binds the framebuffer color and depth surface
|
||||||
void BindFramebufferSurfaces(const Surface& color_surface, const Surface& depth_surface,
|
void BindFramebufferSurfaces(const Surface& color_surface, const Surface& depth_surface,
|
||||||
|
|
|
@ -686,7 +686,8 @@ Surface RasterizerCacheOpenGL::GetTextureSurface(const Tegra::Texture::FullTextu
|
||||||
}
|
}
|
||||||
|
|
||||||
SurfaceSurfaceRect_Tuple RasterizerCacheOpenGL::GetFramebufferSurfaces(bool using_color_fb,
|
SurfaceSurfaceRect_Tuple RasterizerCacheOpenGL::GetFramebufferSurfaces(bool using_color_fb,
|
||||||
bool using_depth_fb) {
|
bool using_depth_fb,
|
||||||
|
bool preserve_contents) {
|
||||||
const auto& regs = Core::System::GetInstance().GPU().Maxwell3D().regs;
|
const auto& regs = Core::System::GetInstance().GPU().Maxwell3D().regs;
|
||||||
|
|
||||||
// TODO(bunnei): This is hard corded to use just the first render buffer
|
// TODO(bunnei): This is hard corded to use just the first render buffer
|
||||||
|
@ -708,7 +709,7 @@ SurfaceSurfaceRect_Tuple RasterizerCacheOpenGL::GetFramebufferSurfaces(bool usin
|
||||||
MathUtil::Rectangle<u32> color_rect{};
|
MathUtil::Rectangle<u32> color_rect{};
|
||||||
Surface color_surface;
|
Surface color_surface;
|
||||||
if (using_color_fb) {
|
if (using_color_fb) {
|
||||||
color_surface = GetSurface(color_params);
|
color_surface = GetSurface(color_params, preserve_contents);
|
||||||
if (color_surface) {
|
if (color_surface) {
|
||||||
color_rect = color_surface->GetSurfaceParams().GetRect();
|
color_rect = color_surface->GetSurfaceParams().GetRect();
|
||||||
}
|
}
|
||||||
|
@ -717,7 +718,7 @@ SurfaceSurfaceRect_Tuple RasterizerCacheOpenGL::GetFramebufferSurfaces(bool usin
|
||||||
MathUtil::Rectangle<u32> depth_rect{};
|
MathUtil::Rectangle<u32> depth_rect{};
|
||||||
Surface depth_surface;
|
Surface depth_surface;
|
||||||
if (using_depth_fb) {
|
if (using_depth_fb) {
|
||||||
depth_surface = GetSurface(depth_params);
|
depth_surface = GetSurface(depth_params, preserve_contents);
|
||||||
if (depth_surface) {
|
if (depth_surface) {
|
||||||
depth_rect = depth_surface->GetSurfaceParams().GetRect();
|
depth_rect = depth_surface->GetSurfaceParams().GetRect();
|
||||||
}
|
}
|
||||||
|
@ -752,7 +753,7 @@ void RasterizerCacheOpenGL::FlushSurface(const Surface& surface) {
|
||||||
surface->FlushGLBuffer();
|
surface->FlushGLBuffer();
|
||||||
}
|
}
|
||||||
|
|
||||||
Surface RasterizerCacheOpenGL::GetSurface(const SurfaceParams& params) {
|
Surface RasterizerCacheOpenGL::GetSurface(const SurfaceParams& params, bool preserve_contents) {
|
||||||
if (params.addr == 0 || params.height * params.width == 0) {
|
if (params.addr == 0 || params.height * params.width == 0) {
|
||||||
return {};
|
return {};
|
||||||
}
|
}
|
||||||
|
@ -774,9 +775,13 @@ Surface RasterizerCacheOpenGL::GetSurface(const SurfaceParams& params) {
|
||||||
} else if (surface->GetSurfaceParams().IsCompatibleSurface(params)) {
|
} else if (surface->GetSurfaceParams().IsCompatibleSurface(params)) {
|
||||||
// Use the cached surface as-is
|
// Use the cached surface as-is
|
||||||
return surface;
|
return surface;
|
||||||
} else {
|
} else if (preserve_contents) {
|
||||||
// If surface parameters changed, recreate the surface from the old one
|
// If surface parameters changed and we care about keeping the previous data, recreate
|
||||||
|
// the surface from the old one
|
||||||
return RecreateSurface(surface, params);
|
return RecreateSurface(surface, params);
|
||||||
|
} else {
|
||||||
|
// Delete the old surface before creating a new one to prevent collisions.
|
||||||
|
UnregisterSurface(surface);
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
|
|
||||||
|
@ -793,12 +798,58 @@ Surface RasterizerCacheOpenGL::RecreateSurface(const Surface& surface,
|
||||||
// Verify surface is compatible for blitting
|
// Verify surface is compatible for blitting
|
||||||
const auto& params{surface->GetSurfaceParams()};
|
const auto& params{surface->GetSurfaceParams()};
|
||||||
ASSERT(params.type == new_params.type);
|
ASSERT(params.type == new_params.type);
|
||||||
|
ASSERT_MSG(params.GetCompressionFactor(params.pixel_format) == 1,
|
||||||
|
"Compressed texture reinterpretation is not supported");
|
||||||
|
|
||||||
// Create a new surface with the new parameters, and blit the previous surface to it
|
// Create a new surface with the new parameters, and blit the previous surface to it
|
||||||
Surface new_surface{std::make_shared<CachedSurface>(new_params)};
|
Surface new_surface{std::make_shared<CachedSurface>(new_params)};
|
||||||
BlitTextures(surface->Texture().handle, params.GetRect(), new_surface->Texture().handle,
|
|
||||||
new_surface->GetSurfaceParams().GetRect(), params.type, read_framebuffer.handle,
|
auto source_format = GetFormatTuple(params.pixel_format, params.component_type);
|
||||||
draw_framebuffer.handle);
|
auto dest_format = GetFormatTuple(new_params.pixel_format, new_params.component_type);
|
||||||
|
|
||||||
|
size_t buffer_size = std::max(params.SizeInBytes(), new_params.SizeInBytes());
|
||||||
|
|
||||||
|
// Use a Pixel Buffer Object to download the previous texture and then upload it to the new one
|
||||||
|
// using the new format.
|
||||||
|
OGLBuffer pbo;
|
||||||
|
pbo.Create();
|
||||||
|
|
||||||
|
glBindBuffer(GL_PIXEL_PACK_BUFFER, pbo.handle);
|
||||||
|
glBufferData(GL_PIXEL_PACK_BUFFER, buffer_size, nullptr, GL_STREAM_DRAW_ARB);
|
||||||
|
glGetTextureImage(surface->Texture().handle, 0, source_format.format, source_format.type,
|
||||||
|
params.SizeInBytes(), nullptr);
|
||||||
|
|
||||||
|
// If the new texture is bigger than the previous one, we need to fill in the rest with data
|
||||||
|
// from the CPU.
|
||||||
|
if (params.SizeInBytes() < new_params.SizeInBytes()) {
|
||||||
|
// Upload the rest of the memory.
|
||||||
|
if (new_params.is_tiled) {
|
||||||
|
// TODO(Subv): We might have to de-tile the subtexture and re-tile it with the rest of
|
||||||
|
// the data in this case. Games like Super Mario Odyssey seem to hit this case when
|
||||||
|
// drawing, it re-uses the memory of a previous texture as a bigger framebuffer but it
|
||||||
|
// doesn't clear it beforehand, the texture is already full of zeros.
|
||||||
|
LOG_CRITICAL(HW_GPU, "Trying to upload extra texture data from the CPU during "
|
||||||
|
"reinterpretation but the texture is tiled.");
|
||||||
|
}
|
||||||
|
size_t remaining_size = new_params.SizeInBytes() - params.SizeInBytes();
|
||||||
|
auto address = Core::System::GetInstance().GPU().memory_manager->GpuToCpuAddress(
|
||||||
|
new_params.addr + params.SizeInBytes());
|
||||||
|
std::vector<u8> data(remaining_size);
|
||||||
|
Memory::ReadBlock(*address, data.data(), data.size());
|
||||||
|
glBufferSubData(GL_PIXEL_PACK_BUFFER, params.SizeInBytes(), remaining_size, data.data());
|
||||||
|
}
|
||||||
|
|
||||||
|
glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
|
||||||
|
|
||||||
|
const auto& dest_rect{new_params.GetRect()};
|
||||||
|
|
||||||
|
glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pbo.handle);
|
||||||
|
glTextureSubImage2D(
|
||||||
|
new_surface->Texture().handle, 0, 0, 0, static_cast<GLsizei>(dest_rect.GetWidth()),
|
||||||
|
static_cast<GLsizei>(dest_rect.GetHeight()), dest_format.format, dest_format.type, nullptr);
|
||||||
|
glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
|
||||||
|
|
||||||
|
pbo.Release();
|
||||||
|
|
||||||
// Update cache accordingly
|
// Update cache accordingly
|
||||||
UnregisterSurface(surface);
|
UnregisterSurface(surface);
|
||||||
|
|
|
@ -722,7 +722,8 @@ public:
|
||||||
Surface GetTextureSurface(const Tegra::Texture::FullTextureInfo& config);
|
Surface GetTextureSurface(const Tegra::Texture::FullTextureInfo& config);
|
||||||
|
|
||||||
/// Get the color and depth surfaces based on the framebuffer configuration
|
/// Get the color and depth surfaces based on the framebuffer configuration
|
||||||
SurfaceSurfaceRect_Tuple GetFramebufferSurfaces(bool using_color_fb, bool using_depth_fb);
|
SurfaceSurfaceRect_Tuple GetFramebufferSurfaces(bool using_color_fb, bool using_depth_fb,
|
||||||
|
bool preserve_contents);
|
||||||
|
|
||||||
/// Flushes the surface to Switch memory
|
/// Flushes the surface to Switch memory
|
||||||
void FlushSurface(const Surface& surface);
|
void FlushSurface(const Surface& surface);
|
||||||
|
@ -738,7 +739,7 @@ public:
|
||||||
|
|
||||||
private:
|
private:
|
||||||
void LoadSurface(const Surface& surface);
|
void LoadSurface(const Surface& surface);
|
||||||
Surface GetSurface(const SurfaceParams& params);
|
Surface GetSurface(const SurfaceParams& params, bool preserve_contents = true);
|
||||||
|
|
||||||
/// Recreates a surface with new parameters
|
/// Recreates a surface with new parameters
|
||||||
Surface RecreateSurface(const Surface& surface, const SurfaceParams& new_params);
|
Surface RecreateSurface(const Surface& surface, const SurfaceParams& new_params);
|
||||||
|
|
Reference in New Issue