summaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
Diffstat
-rw-r--r--.gitmodules8
m---------externals/opus0
-rw-r--r--externals/opus/CMakeLists.txt250
m---------externals/opus/opus0
-rw-r--r--src/core/file_sys/savedata_factory.cpp12
-rw-r--r--src/video_core/CMakeLists.txt2
-rw-r--r--src/video_core/engines/kepler_compute.cpp12
-rw-r--r--src/video_core/engines/kepler_compute.h7
-rw-r--r--src/video_core/engines/maxwell_3d.cpp61
-rw-r--r--src/video_core/engines/maxwell_3d.h8
-rw-r--r--src/video_core/morton.cpp2
-rw-r--r--src/video_core/rasterizer_accelerated.cpp63
-rw-r--r--src/video_core/rasterizer_accelerated.h31
-rw-r--r--src/video_core/renderer_opengl/gl_rasterizer.cpp50
-rw-r--r--src/video_core/renderer_opengl/gl_rasterizer.h11
-rw-r--r--src/video_core/renderer_opengl/gl_state.cpp223
-rw-r--r--src/video_core/renderer_opengl/gl_state.h222
-rw-r--r--src/video_core/renderer_opengl/gl_texture_cache.cpp1
-rw-r--r--src/video_core/surface.cpp8
-rw-r--r--src/video_core/surface.h16
-rw-r--r--src/video_core/textures/texture.h1
21 files changed, 562 insertions, 426 deletions
diff --git a/.gitmodules b/.gitmodules
index 35e0d1240..ee0dc6c19 100644
--- a/.gitmodules
+++ b/.gitmodules
@@ -26,11 +26,11 @@
path = externals/mbedtls
url = https://github.com/DarkLordZach/mbedtls
[submodule "opus"]
- path = externals/opus
- url = https://github.com/ogniK5377/opus.git
+ path = externals/opus/opus
+ url = https://github.com/xiph/opus.git
[submodule "soundtouch"]
- path = externals/soundtouch
- url = https://github.com/citra-emu/ext-soundtouch.git
+ path = externals/soundtouch
+ url = https://github.com/citra-emu/ext-soundtouch.git
[submodule "libressl"]
path = externals/libressl
url = https://github.com/citra-emu/ext-libressl-portable.git
diff --git a/externals/opus b/externals/opus
deleted file mode 160000
-Subproject 562f8ba555c4181e1b57e82e496e4a959b9c019
diff --git a/externals/opus/CMakeLists.txt b/externals/opus/CMakeLists.txt
new file mode 100644
index 000000000..cbb393272
--- /dev/null
+++ b/externals/opus/CMakeLists.txt
@@ -0,0 +1,250 @@
+cmake_minimum_required(VERSION 3.8)
+
+project(opus)
+
+option(OPUS_STACK_PROTECTOR "Use stack protection" OFF)
+option(OPUS_USE_ALLOCA "Use alloca for stack arrays (on non-C99 compilers)" OFF)
+option(OPUS_CUSTOM_MODES "Enable non-Opus modes, e.g. 44.1 kHz & 2^n frames" OFF)
+option(OPUS_FIXED_POINT "Compile as fixed-point (for machines without a fast enough FPU)" OFF)
+option(OPUS_ENABLE_FLOAT_API "Compile with the floating point API (for machines with float library" ON)
+
+include(opus/opus_functions.cmake)
+
+if(OPUS_STACK_PROTECTOR)
+ if(NOT MSVC) # GC on by default on MSVC
+ check_and_set_flag(STACK_PROTECTION_STRONG -fstack-protector-strong)
+ endif()
+else()
+ if(MSVC)
+ check_and_set_flag(BUFFER_SECURITY_CHECK /GS-)
+ endif()
+endif()
+
+add_library(opus STATIC
+ # CELT sources
+ opus/celt/bands.c
+ opus/celt/celt.c
+ opus/celt/celt_decoder.c
+ opus/celt/celt_encoder.c
+ opus/celt/celt_lpc.c
+ opus/celt/cwrs.c
+ opus/celt/entcode.c
+ opus/celt/entdec.c
+ opus/celt/entenc.c
+ opus/celt/kiss_fft.c
+ opus/celt/laplace.c
+ opus/celt/mathops.c
+ opus/celt/mdct.c
+ opus/celt/modes.c
+ opus/celt/pitch.c
+ opus/celt/quant_bands.c
+ opus/celt/rate.c
+ opus/celt/vq.c
+
+ # SILK sources
+ opus/silk/A2NLSF.c
+ opus/silk/CNG.c
+ opus/silk/HP_variable_cutoff.c
+ opus/silk/LPC_analysis_filter.c
+ opus/silk/LPC_fit.c
+ opus/silk/LPC_inv_pred_gain.c
+ opus/silk/LP_variable_cutoff.c
+ opus/silk/NLSF2A.c
+ opus/silk/NLSF_VQ.c
+ opus/silk/NLSF_VQ_weights_laroia.c
+ opus/silk/NLSF_decode.c
+ opus/silk/NLSF_del_dec_quant.c
+ opus/silk/NLSF_encode.c
+ opus/silk/NLSF_stabilize.c
+ opus/silk/NLSF_unpack.c
+ opus/silk/NSQ.c
+ opus/silk/NSQ_del_dec.c
+ opus/silk/PLC.c
+ opus/silk/VAD.c
+ opus/silk/VQ_WMat_EC.c
+ opus/silk/ana_filt_bank_1.c
+ opus/silk/biquad_alt.c
+ opus/silk/bwexpander.c
+ opus/silk/bwexpander_32.c
+ opus/silk/check_control_input.c
+ opus/silk/code_signs.c
+ opus/silk/control_SNR.c
+ opus/silk/control_audio_bandwidth.c
+ opus/silk/control_codec.c
+ opus/silk/dec_API.c
+ opus/silk/decode_core.c
+ opus/silk/decode_frame.c
+ opus/silk/decode_indices.c
+ opus/silk/decode_parameters.c
+ opus/silk/decode_pitch.c
+ opus/silk/decode_pulses.c
+ opus/silk/decoder_set_fs.c
+ opus/silk/enc_API.c
+ opus/silk/encode_indices.c
+ opus/silk/encode_pulses.c
+ opus/silk/gain_quant.c
+ opus/silk/init_decoder.c
+ opus/silk/init_encoder.c
+ opus/silk/inner_prod_aligned.c
+ opus/silk/interpolate.c
+ opus/silk/lin2log.c
+ opus/silk/log2lin.c
+ opus/silk/pitch_est_tables.c
+ opus/silk/process_NLSFs.c
+ opus/silk/quant_LTP_gains.c
+ opus/silk/resampler.c
+ opus/silk/resampler_down2.c
+ opus/silk/resampler_down2_3.c
+ opus/silk/resampler_private_AR2.c
+ opus/silk/resampler_private_IIR_FIR.c
+ opus/silk/resampler_private_down_FIR.c
+ opus/silk/resampler_private_up2_HQ.c
+ opus/silk/resampler_rom.c
+ opus/silk/shell_coder.c
+ opus/silk/sigm_Q15.c
+ opus/silk/sort.c
+ opus/silk/stereo_LR_to_MS.c
+ opus/silk/stereo_MS_to_LR.c
+ opus/silk/stereo_decode_pred.c
+ opus/silk/stereo_encode_pred.c
+ opus/silk/stereo_find_predictor.c
+ opus/silk/stereo_quant_pred.c
+ opus/silk/sum_sqr_shift.c
+ opus/silk/table_LSF_cos.c
+ opus/silk/tables_LTP.c
+ opus/silk/tables_NLSF_CB_NB_MB.c
+ opus/silk/tables_NLSF_CB_WB.c
+ opus/silk/tables_gain.c
+ opus/silk/tables_other.c
+ opus/silk/tables_pitch_lag.c
+ opus/silk/tables_pulses_per_block.c
+
+ # Opus sources
+ opus/src/analysis.c
+ opus/src/mapping_matrix.c
+ opus/src/mlp.c
+ opus/src/mlp_data.c
+ opus/src/opus.c
+ opus/src/opus_decoder.c
+ opus/src/opus_encoder.c
+ opus/src/opus_multistream.c
+ opus/src/opus_multistream_decoder.c
+ opus/src/opus_multistream_encoder.c
+ opus/src/opus_projection_decoder.c
+ opus/src/opus_projection_encoder.c
+ opus/src/repacketizer.c
+)
+
+if (DEBUG)
+ target_sources(opus PRIVATE opus/silk/debug.c)
+endif()
+
+if (OPUS_FIXED_POINT)
+ target_sources(opus PRIVATE
+ opus/silk/fixed/LTP_analysis_filter_FIX.c
+ opus/silk/fixed/LTP_scale_ctrl_FIX.c
+ opus/silk/fixed/apply_sine_window_FIX.c
+ opus/silk/fixed/autocorr_FIX.c
+ opus/silk/fixed/burg_modified_FIX.c
+ opus/silk/fixed/corrMatrix_FIX.c
+ opus/silk/fixed/encode_frame_FIX.c
+ opus/silk/fixed/find_LPC_FIX.c
+ opus/silk/fixed/find_LTP_FIX.c
+ opus/silk/fixed/find_pitch_lags_FIX.c
+ opus/silk/fixed/find_pred_coefs_FIX.c
+ opus/silk/fixed/k2a_FIX.c
+ opus/silk/fixed/k2a_Q16_FIX.c
+ opus/silk/fixed/noise_shape_analysis_FIX.c
+ opus/silk/fixed/pitch_analysis_core_FIX.c
+ opus/silk/fixed/prefilter_FIX.c
+ opus/silk/fixed/process_gains_FIX.c
+ opus/silk/fixed/regularize_correlations_FIX.c
+ opus/silk/fixed/residual_energy16_FIX.c
+ opus/silk/fixed/residual_energy_FIX.c
+ opus/silk/fixed/schur64_FIX.c
+ opus/silk/fixed/schur_FIX.c
+ opus/silk/fixed/solve_LS_FIX.c
+ opus/silk/fixed/vector_ops_FIX.c
+ opus/silk/fixed/warped_autocorrelation_FIX.c
+ )
+else()
+ target_sources(opus PRIVATE
+ opus/silk/float/LPC_analysis_filter_FLP.c
+ opus/silk/float/LPC_inv_pred_gain_FLP.c
+ opus/silk/float/LTP_analysis_filter_FLP.c
+ opus/silk/float/LTP_scale_ctrl_FLP.c
+ opus/silk/float/apply_sine_window_FLP.c
+ opus/silk/float/autocorrelation_FLP.c
+ opus/silk/float/burg_modified_FLP.c
+ opus/silk/float/bwexpander_FLP.c
+ opus/silk/float/corrMatrix_FLP.c
+ opus/silk/float/encode_frame_FLP.c
+ opus/silk/float/energy_FLP.c
+ opus/silk/float/find_LPC_FLP.c
+ opus/silk/float/find_LTP_FLP.c
+ opus/silk/float/find_pitch_lags_FLP.c
+ opus/silk/float/find_pred_coefs_FLP.c
+ opus/silk/float/inner_product_FLP.c
+ opus/silk/float/k2a_FLP.c
+ opus/silk/float/noise_shape_analysis_FLP.c
+ opus/silk/float/pitch_analysis_core_FLP.c
+ opus/silk/float/process_gains_FLP.c
+ opus/silk/float/regularize_correlations_FLP.c
+ opus/silk/float/residual_energy_FLP.c
+ opus/silk/float/scale_copy_vector_FLP.c
+ opus/silk/float/scale_vector_FLP.c
+ opus/silk/float/schur_FLP.c
+ opus/silk/float/sort_FLP.c
+ opus/silk/float/warped_autocorrelation_FLP.c
+ opus/silk/float/wrappers_FLP.c
+ )
+endif()
+
+target_compile_definitions(opus PRIVATE OPUS_BUILD ENABLE_HARDENING)
+
+if(NOT MSVC)
+ target_compile_definitions(opus PRIVATE _FORTIFY_SOURCE=2)
+endif()
+
+# It is strongly recommended to uncomment one of these VAR_ARRAYS: Use C99
+# variable-length arrays for stack allocation USE_ALLOCA: Use alloca() for stack
+# allocation If none is defined, then the fallback is a non-threadsafe global
+# array
+if(OPUS_USE_ALLOCA OR MSVC)
+ target_compile_definitions(opus PRIVATE USE_ALLOCA)
+else()
+ target_compile_definitions(opus PRIVATE VAR_ARRAYS)
+endif()
+
+if(OPUS_CUSTOM_MODES)
+ target_compile_definitions(opus PRIVATE CUSTOM_MODES)
+endif()
+
+if(NOT OPUS_ENABLE_FLOAT_API)
+ target_compile_definitions(opus PRIVATE DISABLE_FLOAT_API)
+endif()
+
+target_compile_definitions(opus
+PUBLIC
+ -DOPUS_VERSION="\\"1.3.1\\""
+
+PRIVATE
+ # Use C99 intrinsics to speed up float-to-int conversion
+ HAVE_LRINTF
+)
+
+if (FIXED_POINT)
+ target_compile_definitions(opus PRIVATE -DFIXED_POINT=1 -DDISABLE_FLOAT_API)
+endif()
+
+target_include_directories(opus
+PUBLIC
+ opus/include
+
+PRIVATE
+ opus/celt
+ opus/silk
+ opus/silk/fixed
+ opus/silk/float
+ opus/src
+)
diff --git a/externals/opus/opus b/externals/opus/opus
new file mode 160000
+Subproject ad8fe90db79b7d2a135e3dfd2ed6631b0c5662a
diff --git a/src/core/file_sys/savedata_factory.cpp b/src/core/file_sys/savedata_factory.cpp
index fc8755c78..e2a7eaf7b 100644
--- a/src/core/file_sys/savedata_factory.cpp
+++ b/src/core/file_sys/savedata_factory.cpp
@@ -16,6 +16,7 @@ namespace FileSys {
constexpr char SAVE_DATA_SIZE_FILENAME[] = ".yuzu_save_size";
namespace {
+
void PrintSaveDataDescriptorWarnings(SaveDataDescriptor meta) {
if (meta.type == SaveDataType::SystemSaveData || meta.type == SaveDataType::SaveData) {
if (meta.zero_1 != 0) {
@@ -52,6 +53,13 @@ void PrintSaveDataDescriptorWarnings(SaveDataDescriptor meta) {
meta.user_id[1], meta.user_id[0]);
}
}
+
+bool ShouldSaveDataBeAutomaticallyCreated(SaveDataSpaceId space, const SaveDataDescriptor& desc) {
+ return desc.type == SaveDataType::CacheStorage || desc.type == SaveDataType::TemporaryStorage ||
+ (space == SaveDataSpaceId::NandUser && ///< Normal Save Data -- Current Title & User
+ desc.type == SaveDataType::SaveData && desc.title_id == 0 && desc.save_id == 0);
+}
+
} // Anonymous namespace
std::string SaveDataDescriptor::DebugInfo() const {
@@ -96,6 +104,10 @@ ResultVal<VirtualDir> SaveDataFactory::Open(SaveDataSpaceId space,
auto out = dir->GetDirectoryRelative(save_directory);
+ if (out == nullptr && ShouldSaveDataBeAutomaticallyCreated(space, meta)) {
+ return Create(space, meta);
+ }
+
// Return an error if the save data doesn't actually exist.
if (out == nullptr) {
// TODO(Subv): Find out correct error code.
diff --git a/src/video_core/CMakeLists.txt b/src/video_core/CMakeLists.txt
index cb6eda1b8..c911c6ec4 100644
--- a/src/video_core/CMakeLists.txt
+++ b/src/video_core/CMakeLists.txt
@@ -36,6 +36,8 @@ add_library(video_core STATIC
memory_manager.h
morton.cpp
morton.h
+ rasterizer_accelerated.cpp
+ rasterizer_accelerated.h
rasterizer_cache.cpp
rasterizer_cache.h
rasterizer_interface.h
diff --git a/src/video_core/engines/kepler_compute.cpp b/src/video_core/engines/kepler_compute.cpp
index 91adef360..3a39aeabe 100644
--- a/src/video_core/engines/kepler_compute.cpp
+++ b/src/video_core/engines/kepler_compute.cpp
@@ -50,7 +50,7 @@ void KeplerCompute::CallMethod(const GPU::MethodCall& method_call) {
}
}
-Tegra::Texture::FullTextureInfo KeplerCompute::GetTexture(std::size_t offset) const {
+Texture::FullTextureInfo KeplerCompute::GetTexture(std::size_t offset) const {
const std::bitset<8> cbuf_mask = launch_description.const_buffer_enable_mask.Value();
ASSERT(cbuf_mask[regs.tex_cb_index]);
@@ -61,13 +61,11 @@ Tegra::Texture::FullTextureInfo KeplerCompute::GetTexture(std::size_t offset) co
ASSERT(address < texinfo.Address() + texinfo.size);
const Texture::TextureHandle tex_handle{memory_manager.Read<u32>(address)};
- return GetTextureInfo(tex_handle, offset);
+ return GetTextureInfo(tex_handle);
}
-Texture::FullTextureInfo KeplerCompute::GetTextureInfo(const Texture::TextureHandle tex_handle,
- std::size_t offset) const {
- return Texture::FullTextureInfo{static_cast<u32>(offset), GetTICEntry(tex_handle.tic_id),
- GetTSCEntry(tex_handle.tsc_id)};
+Texture::FullTextureInfo KeplerCompute::GetTextureInfo(Texture::TextureHandle tex_handle) const {
+ return Texture::FullTextureInfo{GetTICEntry(tex_handle.tic_id), GetTSCEntry(tex_handle.tsc_id)};
}
u32 KeplerCompute::AccessConstBuffer32(ShaderType stage, u64 const_buffer, u64 offset) const {
@@ -89,7 +87,7 @@ SamplerDescriptor KeplerCompute::AccessBindlessSampler(ShaderType stage, u64 con
const GPUVAddr tex_info_address = tex_info_buffer.Address() + offset;
const Texture::TextureHandle tex_handle{memory_manager.Read<u32>(tex_info_address)};
- const Texture::FullTextureInfo tex_info = GetTextureInfo(tex_handle, offset);
+ const Texture::FullTextureInfo tex_info = GetTextureInfo(tex_handle);
SamplerDescriptor result = SamplerDescriptor::FromTicTexture(tex_info.tic.texture_type.Value());
result.is_shadow.Assign(tex_info.tsc.depth_compare_enabled.Value());
return result;
diff --git a/src/video_core/engines/kepler_compute.h b/src/video_core/engines/kepler_compute.h
index 8e7182727..b185c98c7 100644
--- a/src/video_core/engines/kepler_compute.h
+++ b/src/video_core/engines/kepler_compute.h
@@ -196,11 +196,10 @@ public:
/// Write the value to the register identified by method.
void CallMethod(const GPU::MethodCall& method_call);
- Tegra::Texture::FullTextureInfo GetTexture(std::size_t offset) const;
+ Texture::FullTextureInfo GetTexture(std::size_t offset) const;
- /// Given a Texture Handle, returns the TSC and TIC entries.
- Texture::FullTextureInfo GetTextureInfo(const Texture::TextureHandle tex_handle,
- std::size_t offset) const;
+ /// Given a texture handle, returns the TSC and TIC entries.
+ Texture::FullTextureInfo GetTextureInfo(Texture::TextureHandle tex_handle) const;
u32 AccessConstBuffer32(ShaderType stage, u64 const_buffer, u64 offset) const override;
diff --git a/src/video_core/engines/maxwell_3d.cpp b/src/video_core/engines/maxwell_3d.cpp
index 514ed93fa..2bed6cb38 100644
--- a/src/video_core/engines/maxwell_3d.cpp
+++ b/src/video_core/engines/maxwell_3d.cpp
@@ -760,61 +760,8 @@ Texture::TSCEntry Maxwell3D::GetTSCEntry(u32 tsc_index) const {
return tsc_entry;
}
-std::vector<Texture::FullTextureInfo> Maxwell3D::GetStageTextures(Regs::ShaderStage stage) const {
- std::vector<Texture::FullTextureInfo> textures;
-
- auto& fragment_shader = state.shader_stages[static_cast<std::size_t>(stage)];
- auto& tex_info_buffer = fragment_shader.const_buffers[regs.tex_cb_index];
- ASSERT(tex_info_buffer.enabled && tex_info_buffer.address != 0);
-
- GPUVAddr tex_info_buffer_end = tex_info_buffer.address + tex_info_buffer.size;
-
- // Offset into the texture constbuffer where the texture info begins.
- static constexpr std::size_t TextureInfoOffset = 0x20;
-
- for (GPUVAddr current_texture = tex_info_buffer.address + TextureInfoOffset;
- current_texture < tex_info_buffer_end; current_texture += sizeof(Texture::TextureHandle)) {
-
- const Texture::TextureHandle tex_handle{memory_manager.Read<u32>(current_texture)};
-
- Texture::FullTextureInfo tex_info{};
- // TODO(Subv): Use the shader to determine which textures are actually accessed.
- tex_info.index =
- static_cast<u32>(current_texture - tex_info_buffer.address - TextureInfoOffset) /
- sizeof(Texture::TextureHandle);
-
- // Load the TIC data.
- auto tic_entry = GetTICEntry(tex_handle.tic_id);
- // TODO(Subv): Workaround for BitField's move constructor being deleted.
- std::memcpy(&tex_info.tic, &tic_entry, sizeof(tic_entry));
-
- // Load the TSC data
- auto tsc_entry = GetTSCEntry(tex_handle.tsc_id);
- // TODO(Subv): Workaround for BitField's move constructor being deleted.
- std::memcpy(&tex_info.tsc, &tsc_entry, sizeof(tsc_entry));
-
- textures.push_back(tex_info);
- }
-
- return textures;
-}
-
-Texture::FullTextureInfo Maxwell3D::GetTextureInfo(const Texture::TextureHandle tex_handle,
- std::size_t offset) const {
- Texture::FullTextureInfo tex_info{};
- tex_info.index = static_cast<u32>(offset);
-
- // Load the TIC data.
- auto tic_entry = GetTICEntry(tex_handle.tic_id);
- // TODO(Subv): Workaround for BitField's move constructor being deleted.
- std::memcpy(&tex_info.tic, &tic_entry, sizeof(tic_entry));
-
- // Load the TSC data
- auto tsc_entry = GetTSCEntry(tex_handle.tsc_id);
- // TODO(Subv): Workaround for BitField's move constructor being deleted.
- std::memcpy(&tex_info.tsc, &tsc_entry, sizeof(tsc_entry));
-
- return tex_info;
+Texture::FullTextureInfo Maxwell3D::GetTextureInfo(Texture::TextureHandle tex_handle) const {
+ return Texture::FullTextureInfo{GetTICEntry(tex_handle.tic_id), GetTSCEntry(tex_handle.tsc_id)};
}
Texture::FullTextureInfo Maxwell3D::GetStageTexture(Regs::ShaderStage stage,
@@ -830,7 +777,7 @@ Texture::FullTextureInfo Maxwell3D::GetStageTexture(Regs::ShaderStage stage,
const Texture::TextureHandle tex_handle{memory_manager.Read<u32>(tex_info_address)};
- return GetTextureInfo(tex_handle, offset);
+ return GetTextureInfo(tex_handle);
}
u32 Maxwell3D::GetRegisterValue(u32 method) const {
@@ -867,7 +814,7 @@ SamplerDescriptor Maxwell3D::AccessBindlessSampler(ShaderType stage, u64 const_b
const GPUVAddr tex_info_address = tex_info_buffer.address + offset;
const Texture::TextureHandle tex_handle{memory_manager.Read<u32>(tex_info_address)};
- const Texture::FullTextureInfo tex_info = GetTextureInfo(tex_handle, offset);
+ const Texture::FullTextureInfo tex_info = GetTextureInfo(tex_handle);
SamplerDescriptor result = SamplerDescriptor::FromTicTexture(tex_info.tic.texture_type.Value());
result.is_shadow.Assign(tex_info.tsc.depth_compare_enabled.Value());
return result;
diff --git a/src/video_core/engines/maxwell_3d.h b/src/video_core/engines/maxwell_3d.h
index 987ad77b2..8cc842684 100644
--- a/src/video_core/engines/maxwell_3d.h
+++ b/src/video_core/engines/maxwell_3d.h
@@ -1250,12 +1250,8 @@ public:
void FlushMMEInlineDraw();
- /// Given a Texture Handle, returns the TSC and TIC entries.
- Texture::FullTextureInfo GetTextureInfo(const Texture::TextureHandle tex_handle,
- std::size_t offset) const;
-
- /// Returns a list of enabled textures for the specified shader stage.
- std::vector<Texture::FullTextureInfo> GetStageTextures(Regs::ShaderStage stage) const;
+ /// Given a texture handle, returns the TSC and TIC entries.
+ Texture::FullTextureInfo GetTextureInfo(Texture::TextureHandle tex_handle) const;
/// Returns the texture information for a specific texture in a specific shader stage.
Texture::FullTextureInfo GetStageTexture(Regs::ShaderStage stage, std::size_t offset) const;
diff --git a/src/video_core/morton.cpp b/src/video_core/morton.cpp
index fe5f08ace..2f2fe6859 100644
--- a/src/video_core/morton.cpp
+++ b/src/video_core/morton.cpp
@@ -112,6 +112,7 @@ static constexpr ConversionArray morton_to_linear_fns = {
MortonCopy<true, PixelFormat::ASTC_2D_8X6_SRGB>,
MortonCopy<true, PixelFormat::ASTC_2D_6X5>,
MortonCopy<true, PixelFormat::ASTC_2D_6X5_SRGB>,
+ MortonCopy<true, PixelFormat::E5B9G9R9F>,
MortonCopy<true, PixelFormat::Z32F>,
MortonCopy<true, PixelFormat::Z16>,
MortonCopy<true, PixelFormat::Z24S8>,
@@ -192,6 +193,7 @@ static constexpr ConversionArray linear_to_morton_fns = {
nullptr,
nullptr,
nullptr,
+ MortonCopy<false, PixelFormat::E5B9G9R9F>,
MortonCopy<false, PixelFormat::Z32F>,
MortonCopy<false, PixelFormat::Z16>,
MortonCopy<false, PixelFormat::Z24S8>,
diff --git a/src/video_core/rasterizer_accelerated.cpp b/src/video_core/rasterizer_accelerated.cpp
new file mode 100644
index 000000000..b230dcc18
--- /dev/null
+++ b/src/video_core/rasterizer_accelerated.cpp
@@ -0,0 +1,63 @@
+// Copyright 2019 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include <mutex>
+
+#include <boost/icl/interval_map.hpp>
+
+#include "common/assert.h"
+#include "common/common_types.h"
+#include "core/memory.h"
+#include "video_core/rasterizer_accelerated.h"
+
+namespace VideoCore {
+
+namespace {
+
+template <typename Map, typename Interval>
+constexpr auto RangeFromInterval(Map& map, const Interval& interval) {
+ return boost::make_iterator_range(map.equal_range(interval));
+}
+
+} // Anonymous namespace
+
+RasterizerAccelerated::RasterizerAccelerated() = default;
+
+RasterizerAccelerated::~RasterizerAccelerated() = default;
+
+void RasterizerAccelerated::UpdatePagesCachedCount(VAddr addr, u64 size, int delta) {
+ std::lock_guard lock{pages_mutex};
+ const u64 page_start{addr >> Memory::PAGE_BITS};
+ const u64 page_end{(addr + size + Memory::PAGE_SIZE - 1) >> Memory::PAGE_BITS};
+
+ // Interval maps will erase segments if count reaches 0, so if delta is negative we have to
+ // subtract after iterating
+ const auto pages_interval = CachedPageMap::interval_type::right_open(page_start, page_end);
+ if (delta > 0) {
+ cached_pages.add({pages_interval, delta});
+ }
+
+ for (const auto& pair : RangeFromInterval(cached_pages, pages_interval)) {
+ const auto interval = pair.first & pages_interval;
+ const int count = pair.second;
+
+ const VAddr interval_start_addr = boost::icl::first(interval) << Memory::PAGE_BITS;
+ const VAddr interval_end_addr = boost::icl::last_next(interval) << Memory::PAGE_BITS;
+ const u64 interval_size = interval_end_addr - interval_start_addr;
+
+ if (delta > 0 && count == delta) {
+ Memory::RasterizerMarkRegionCached(interval_start_addr, interval_size, true);
+ } else if (delta < 0 && count == -delta) {
+ Memory::RasterizerMarkRegionCached(interval_start_addr, interval_size, false);
+ } else {
+ ASSERT(count >= 0);
+ }
+ }
+
+ if (delta < 0) {
+ cached_pages.add({pages_interval, delta});
+ }
+}
+
+} // namespace VideoCore
diff --git a/src/video_core/rasterizer_accelerated.h b/src/video_core/rasterizer_accelerated.h
new file mode 100644
index 000000000..8f7e3547e
--- /dev/null
+++ b/src/video_core/rasterizer_accelerated.h
@@ -0,0 +1,31 @@
+// Copyright 2019 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include <mutex>
+
+#include <boost/icl/interval_map.hpp>
+
+#include "common/common_types.h"
+#include "video_core/rasterizer_interface.h"
+
+namespace VideoCore {
+
+/// Implements the shared part in GPU accelerated rasterizers in RasterizerInterface.
+class RasterizerAccelerated : public RasterizerInterface {
+public:
+ explicit RasterizerAccelerated();
+ ~RasterizerAccelerated() override;
+
+ void UpdatePagesCachedCount(VAddr addr, u64 size, int delta) override;
+
+private:
+ using CachedPageMap = boost::icl::interval_map<u64, int>;
+ CachedPageMap cached_pages;
+
+ std::mutex pages_mutex;
+};
+
+} // namespace VideoCore
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index 9431d64ac..6a4d2c83a 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -68,8 +68,6 @@ RasterizerOpenGL::RasterizerOpenGL(Core::System& system, Core::Frontend::EmuWind
ScreenInfo& info)
: texture_cache{system, *this, device}, shader_cache{*this, system, emu_window, device},
system{system}, screen_info{info}, buffer_cache{*this, system, STREAM_BUFFER_SIZE} {
- OpenGLState::ApplyDefaultState();
-
shader_program_manager = std::make_unique<GLShader::ProgramManager>();
state.draw.shader_program = 0;
state.Apply();
@@ -342,42 +340,6 @@ std::size_t RasterizerOpenGL::CalculateIndexBufferSize() const {
static_cast<std::size_t>(regs.index_array.FormatSizeInBytes());
}
-template <typename Map, typename Interval>
-static constexpr auto RangeFromInterval(Map& map, const Interval& interval) {
- return boost::make_iterator_range(map.equal_range(interval));
-}
-
-void RasterizerOpenGL::UpdatePagesCachedCount(VAddr addr, u64 size, int delta) {
- std::lock_guard lock{pages_mutex};
- const u64 page_start{addr >> Memory::PAGE_BITS};
- const u64 page_end{(addr + size + Memory::PAGE_SIZE - 1) >> Memory::PAGE_BITS};
-
- // Interval maps will erase segments if count reaches 0, so if delta is negative we have to
- // subtract after iterating
- const auto pages_interval = CachedPageMap::interval_type::right_open(page_start, page_end);
- if (delta > 0)
- cached_pages.add({pages_interval, delta});
-
- for (const auto& pair : RangeFromInterval(cached_pages, pages_interval)) {
- const auto interval = pair.first & pages_interval;
- const int count = pair.second;
-
- const VAddr interval_start_addr = boost::icl::first(interval) << Memory::PAGE_BITS;
- const VAddr interval_end_addr = boost::icl::last_next(interval) << Memory::PAGE_BITS;
- const u64 interval_size = interval_end_addr - interval_start_addr;
-
- if (delta > 0 && count == delta)
- Memory::RasterizerMarkRegionCached(interval_start_addr, interval_size, true);
- else if (delta < 0 && count == -delta)
- Memory::RasterizerMarkRegionCached(interval_start_addr, interval_size, false);
- else
- ASSERT(count >= 0);
- }
-
- if (delta < 0)
- cached_pages.add({pages_interval, delta});
-}
-
void RasterizerOpenGL::LoadDiskResources(const std::atomic_bool& stop_loading,
const VideoCore::DiskResourceLoadCallback& callback) {
shader_cache.LoadDiskCache(stop_loading, callback);
@@ -969,7 +931,7 @@ TextureBufferUsage RasterizerOpenGL::SetupDrawTextures(Maxwell::ShaderStage stag
for (u32 bindpoint = 0; bindpoint < entries.size(); ++bindpoint) {
const auto& entry = entries[bindpoint];
- const auto texture = [&]() {
+ const auto texture = [&] {
if (!entry.IsBindless()) {
return maxwell3d.GetStageTexture(stage, entry.GetOffset());
}
@@ -977,7 +939,7 @@ TextureBufferUsage RasterizerOpenGL::SetupDrawTextures(Maxwell::ShaderStage stag
Tegra::Texture::TextureHandle tex_handle;
Tegra::Engines::ShaderType shader_type = static_cast<Tegra::Engines::ShaderType>(stage);
tex_handle.raw = maxwell3d.AccessConstBuffer32(shader_type, cbuf.first, cbuf.second);
- return maxwell3d.GetTextureInfo(tex_handle, entry.GetOffset());
+ return maxwell3d.GetTextureInfo(tex_handle);
}();
if (SetupTexture(base_bindings.sampler + bindpoint, texture, entry)) {
@@ -1000,7 +962,7 @@ TextureBufferUsage RasterizerOpenGL::SetupComputeTextures(const Shader& kernel)
for (u32 bindpoint = 0; bindpoint < entries.size(); ++bindpoint) {
const auto& entry = entries[bindpoint];
- const auto texture = [&]() {
+ const auto texture = [&] {
if (!entry.IsBindless()) {
return compute.GetTexture(entry.GetOffset());
}
@@ -1008,7 +970,7 @@ TextureBufferUsage RasterizerOpenGL::SetupComputeTextures(const Shader& kernel)
Tegra::Texture::TextureHandle tex_handle;
tex_handle.raw = compute.AccessConstBuffer32(Tegra::Engines::ShaderType::Compute,
cbuf.first, cbuf.second);
- return compute.GetTextureInfo(tex_handle, entry.GetOffset());
+ return compute.GetTextureInfo(tex_handle);
}();
if (SetupTexture(bindpoint, texture, entry)) {
@@ -1046,7 +1008,7 @@ void RasterizerOpenGL::SetupComputeImages(const Shader& shader) {
const auto& entries = shader->GetShaderEntries().images;
for (u32 bindpoint = 0; bindpoint < entries.size(); ++bindpoint) {
const auto& entry = entries[bindpoint];
- const auto tic = [&]() {
+ const auto tic = [&] {
if (!entry.IsBindless()) {
return compute.GetTexture(entry.GetOffset()).tic;
}
@@ -1054,7 +1016,7 @@ void RasterizerOpenGL::SetupComputeImages(const Shader& shader) {
Tegra::Texture::TextureHandle tex_handle;
tex_handle.raw = compute.AccessConstBuffer32(Tegra::Engines::ShaderType::Compute,
cbuf.first, cbuf.second);
- return compute.GetTextureInfo(tex_handle, entry.GetOffset()).tic;
+ return compute.GetTextureInfo(tex_handle).tic;
}();
SetupImage(bindpoint, tic, entry);
}
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.h b/src/video_core/renderer_opengl/gl_rasterizer.h
index c24a02d71..bd6fe5c3a 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.h
+++ b/src/video_core/renderer_opengl/gl_rasterizer.h
@@ -9,17 +9,16 @@
#include <cstddef>
#include <map>
#include <memory>
-#include <mutex>
#include <optional>
#include <tuple>
#include <utility>
-#include <boost/icl/interval_map.hpp>
#include <glad/glad.h>
#include "common/common_types.h"
#include "video_core/engines/const_buffer_info.h"
#include "video_core/engines/maxwell_3d.h"
+#include "video_core/rasterizer_accelerated.h"
#include "video_core/rasterizer_cache.h"
#include "video_core/rasterizer_interface.h"
#include "video_core/renderer_opengl/gl_buffer_cache.h"
@@ -52,7 +51,7 @@ namespace OpenGL {
struct ScreenInfo;
struct DrawParameters;
-class RasterizerOpenGL : public VideoCore::RasterizerInterface {
+class RasterizerOpenGL : public VideoCore::RasterizerAccelerated {
public:
explicit RasterizerOpenGL(Core::System& system, Core::Frontend::EmuWindow& emu_window,
ScreenInfo& info);
@@ -73,7 +72,6 @@ public:
const Tegra::Engines::Fermi2D::Config& copy_config) override;
bool AccelerateDisplay(const Tegra::FramebufferConfig& config, VAddr framebuffer_addr,
u32 pixel_stride) override;
- void UpdatePagesCachedCount(VAddr addr, u64 size, int delta) override;
void LoadDiskResources(const std::atomic_bool& stop_loading,
const VideoCore::DiskResourceLoadCallback& callback) override;
@@ -228,11 +226,6 @@ private:
AccelDraw accelerate_draw = AccelDraw::Disabled;
OGLFramebuffer clear_framebuffer;
-
- using CachedPageMap = boost::icl::interval_map<u64, int>;
- CachedPageMap cached_pages;
-
- std::mutex pages_mutex;
};
} // namespace OpenGL
diff --git a/src/video_core/renderer_opengl/gl_state.cpp b/src/video_core/renderer_opengl/gl_state.cpp
index bf86b5a0b..f25148362 100644
--- a/src/video_core/renderer_opengl/gl_state.cpp
+++ b/src/video_core/renderer_opengl/gl_state.cpp
@@ -2,6 +2,7 @@
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
+#include <algorithm>
#include <iterator>
#include <glad/glad.h>
#include "common/assert.h"
@@ -69,147 +70,29 @@ void Enable(GLenum cap, GLuint index, bool enable) {
}
void Enable(GLenum cap, bool& current_value, bool new_value) {
- if (UpdateValue(current_value, new_value))
+ if (UpdateValue(current_value, new_value)) {
Enable(cap, new_value);
+ }
}
void Enable(GLenum cap, GLuint index, bool& current_value, bool new_value) {
- if (UpdateValue(current_value, new_value))
+ if (UpdateValue(current_value, new_value)) {
Enable(cap, index, new_value);
-}
-
-} // namespace
-
-OpenGLState::OpenGLState() {
- // These all match default OpenGL values
- framebuffer_srgb.enabled = false;
-
- multisample_control.alpha_to_coverage = false;
- multisample_control.alpha_to_one = false;
-
- cull.enabled = false;
- cull.mode = GL_BACK;
- cull.front_face = GL_CCW;
-
- depth.test_enabled = false;
- depth.test_func = GL_LESS;
- depth.write_mask = GL_TRUE;
-
- primitive_restart.enabled = false;
- primitive_restart.index = 0;
-
- for (auto& item : color_mask) {
- item.red_enabled = GL_TRUE;
- item.green_enabled = GL_TRUE;
- item.blue_enabled = GL_TRUE;
- item.alpha_enabled = GL_TRUE;
}
+}
- const auto ResetStencil = [](auto& config) {
- config.test_func = GL_ALWAYS;
- config.test_ref = 0;
- config.test_mask = 0xFFFFFFFF;
- config.write_mask = 0xFFFFFFFF;
- config.action_depth_fail = GL_KEEP;
- config.action_depth_pass = GL_KEEP;
- config.action_stencil_fail = GL_KEEP;
- };
- stencil.test_enabled = false;
- ResetStencil(stencil.front);
- ResetStencil(stencil.back);
-
- for (auto& item : viewports) {
- item.x = 0;
- item.y = 0;
- item.width = 0;
- item.height = 0;
- item.depth_range_near = 0.0f;
- item.depth_range_far = 1.0f;
- item.scissor.enabled = false;
- item.scissor.x = 0;
- item.scissor.y = 0;
- item.scissor.width = 0;
- item.scissor.height = 0;
- }
-
- for (auto& item : blend) {
- item.enabled = true;
- item.rgb_equation = GL_FUNC_ADD;
- item.a_equation = GL_FUNC_ADD;
- item.src_rgb_func = GL_ONE;
- item.dst_rgb_func = GL_ZERO;
- item.src_a_func = GL_ONE;
- item.dst_a_func = GL_ZERO;
- }
-
- independant_blend.enabled = false;
-
- blend_color.red = 0.0f;
- blend_color.green = 0.0f;
- blend_color.blue = 0.0f;
- blend_color.alpha = 0.0f;
-
- logic_op.enabled = false;
- logic_op.operation = GL_COPY;
-
- draw.read_framebuffer = 0;
- draw.draw_framebuffer = 0;
- draw.vertex_array = 0;
- draw.shader_program = 0;
- draw.program_pipeline = 0;
-
- clip_distance = {};
-
- point.size = 1;
-
- fragment_color_clamp.enabled = false;
-
- depth_clamp.far_plane = false;
- depth_clamp.near_plane = false;
-
- polygon_offset.fill_enable = false;
- polygon_offset.line_enable = false;
- polygon_offset.point_enable = false;
- polygon_offset.factor = 0.0f;
- polygon_offset.units = 0.0f;
- polygon_offset.clamp = 0.0f;
+} // Anonymous namespace
- alpha_test.enabled = false;
- alpha_test.func = GL_ALWAYS;
- alpha_test.ref = 0.0f;
-}
+OpenGLState::OpenGLState() = default;
void OpenGLState::SetDefaultViewports() {
- for (auto& item : viewports) {
- item.x = 0;
- item.y = 0;
- item.width = 0;
- item.height = 0;
- item.depth_range_near = 0.0f;
- item.depth_range_far = 1.0f;
- item.scissor.enabled = false;
- item.scissor.x = 0;
- item.scissor.y = 0;
- item.scissor.width = 0;
- item.scissor.height = 0;
- }
+ viewports.fill(Viewport{});
depth_clamp.far_plane = false;
depth_clamp.near_plane = false;
}
-void OpenGLState::ApplyDefaultState() {
- glEnable(GL_BLEND);
- glDisable(GL_FRAMEBUFFER_SRGB);
- glDisable(GL_CULL_FACE);
- glDisable(GL_DEPTH_TEST);
- glDisable(GL_PRIMITIVE_RESTART);
- glDisable(GL_STENCIL_TEST);
- glDisable(GL_COLOR_LOGIC_OP);
- glDisable(GL_SCISSOR_TEST);
-}
-
-void OpenGLState::ApplyFramebufferState() const {
+void OpenGLState::ApplyFramebufferState() {
if (UpdateValue(cur_state.draw.read_framebuffer, draw.read_framebuffer)) {
glBindFramebuffer(GL_READ_FRAMEBUFFER, draw.read_framebuffer);
}
@@ -218,52 +101,52 @@ void OpenGLState::ApplyFramebufferState() const {
}
}
-void OpenGLState::ApplyVertexArrayState() const {
+void OpenGLState::ApplyVertexArrayState() {
if (UpdateValue(cur_state.draw.vertex_array, draw.vertex_array)) {
glBindVertexArray(draw.vertex_array);
}
}
-void OpenGLState::ApplyShaderProgram() const {
+void OpenGLState::ApplyShaderProgram() {
if (UpdateValue(cur_state.draw.shader_program, draw.shader_program)) {
glUseProgram(draw.shader_program);
}
}
-void OpenGLState::ApplyProgramPipeline() const {
+void OpenGLState::ApplyProgramPipeline() {
if (UpdateValue(cur_state.draw.program_pipeline, draw.program_pipeline)) {
glBindProgramPipeline(draw.program_pipeline);
}
}
-void OpenGLState::ApplyClipDistances() const {
+void OpenGLState::ApplyClipDistances() {
for (std::size_t i = 0; i < clip_distance.size(); ++i) {
Enable(GL_CLIP_DISTANCE0 + static_cast<GLenum>(i), cur_state.clip_distance[i],
clip_distance[i]);
}
}
-void OpenGLState::ApplyPointSize() const {
+void OpenGLState::ApplyPointSize() {
if (UpdateValue(cur_state.point.size, point.size)) {
glPointSize(point.size);
}
}
-void OpenGLState::ApplyFragmentColorClamp() const {
+void OpenGLState::ApplyFragmentColorClamp() {
if (UpdateValue(cur_state.fragment_color_clamp.enabled, fragment_color_clamp.enabled)) {
glClampColor(GL_CLAMP_FRAGMENT_COLOR_ARB,
fragment_color_clamp.enabled ? GL_TRUE : GL_FALSE);
}
}
-void OpenGLState::ApplyMultisample() const {
+void OpenGLState::ApplyMultisample() {
Enable(GL_SAMPLE_ALPHA_TO_COVERAGE, cur_state.multisample_control.alpha_to_coverage,
multisample_control.alpha_to_coverage);
Enable(GL_SAMPLE_ALPHA_TO_ONE, cur_state.multisample_control.alpha_to_one,
multisample_control.alpha_to_one);
}
-void OpenGLState::ApplyDepthClamp() const {
+void OpenGLState::ApplyDepthClamp() {
if (depth_clamp.far_plane == cur_state.depth_clamp.far_plane &&
depth_clamp.near_plane == cur_state.depth_clamp.near_plane) {
return;
@@ -276,7 +159,7 @@ void OpenGLState::ApplyDepthClamp() const {
Enable(GL_DEPTH_CLAMP, depth_clamp.far_plane || depth_clamp.near_plane);
}
-void OpenGLState::ApplySRgb() const {
+void OpenGLState::ApplySRgb() {
if (cur_state.framebuffer_srgb.enabled == framebuffer_srgb.enabled)
return;
cur_state.framebuffer_srgb.enabled = framebuffer_srgb.enabled;
@@ -287,7 +170,7 @@ void OpenGLState::ApplySRgb() const {
}
}
-void OpenGLState::ApplyCulling() const {
+void OpenGLState::ApplyCulling() {
Enable(GL_CULL_FACE, cur_state.cull.enabled, cull.enabled);
if (UpdateValue(cur_state.cull.mode, cull.mode)) {
@@ -299,7 +182,12 @@ void OpenGLState::ApplyCulling() const {
}
}
-void OpenGLState::ApplyColorMask() const {
+void OpenGLState::ApplyColorMask() {
+ if (!dirty.color_mask) {
+ return;
+ }
+ dirty.color_mask = false;
+
for (std::size_t i = 0; i < Maxwell::NumRenderTargets; ++i) {
const auto& updated = color_mask[i];
auto& current = cur_state.color_mask[i];
@@ -314,7 +202,7 @@ void OpenGLState::ApplyColorMask() const {
}
}
-void OpenGLState::ApplyDepth() const {
+void OpenGLState::ApplyDepth() {
Enable(GL_DEPTH_TEST, cur_state.depth.test_enabled, depth.test_enabled);
if (cur_state.depth.test_func != depth.test_func) {
@@ -328,7 +216,7 @@ void OpenGLState::ApplyDepth() const {
}
}
-void OpenGLState::ApplyPrimitiveRestart() const {
+void OpenGLState::ApplyPrimitiveRestart() {
Enable(GL_PRIMITIVE_RESTART, cur_state.primitive_restart.enabled, primitive_restart.enabled);
if (cur_state.primitive_restart.index != primitive_restart.index) {
@@ -337,7 +225,12 @@ void OpenGLState::ApplyPrimitiveRestart() const {
}
}
-void OpenGLState::ApplyStencilTest() const {
+void OpenGLState::ApplyStencilTest() {
+ if (!dirty.stencil_state) {
+ return;
+ }
+ dirty.stencil_state = false;
+
Enable(GL_STENCIL_TEST, cur_state.stencil.test_enabled, stencil.test_enabled);
const auto ConfigStencil = [](GLenum face, const auto& config, auto& current) {
@@ -366,7 +259,7 @@ void OpenGLState::ApplyStencilTest() const {
ConfigStencil(GL_BACK, stencil.back, cur_state.stencil.back);
}
-void OpenGLState::ApplyViewport() const {
+void OpenGLState::ApplyViewport() {
for (GLuint i = 0; i < static_cast<GLuint>(Maxwell::NumViewports); ++i) {
const auto& updated = viewports[i];
auto& current = cur_state.viewports[i];
@@ -403,7 +296,7 @@ void OpenGLState::ApplyViewport() const {
}
}
-void OpenGLState::ApplyGlobalBlending() const {
+void OpenGLState::ApplyGlobalBlending() {
const Blend& updated = blend[0];
Blend& current = cur_state.blend[0];
@@ -427,7 +320,7 @@ void OpenGLState::ApplyGlobalBlending() const {
}
}
-void OpenGLState::ApplyTargetBlending(std::size_t target, bool force) const {
+void OpenGLState::ApplyTargetBlending(std::size_t target, bool force) {
const Blend& updated = blend[target];
Blend& current = cur_state.blend[target];
@@ -451,7 +344,12 @@ void OpenGLState::ApplyTargetBlending(std::size_t target, bool force) const {
}
}
-void OpenGLState::ApplyBlending() const {
+void OpenGLState::ApplyBlending() {
+ if (!dirty.blend_state) {
+ return;
+ }
+ dirty.blend_state = false;
+
if (independant_blend.enabled) {
const bool force = independant_blend.enabled != cur_state.independant_blend.enabled;
for (std::size_t target = 0; target < Maxwell::NumRenderTargets; ++target) {
@@ -470,7 +368,7 @@ void OpenGLState::ApplyBlending() const {
}
}
-void OpenGLState::ApplyLogicOp() const {
+void OpenGLState::ApplyLogicOp() {
Enable(GL_COLOR_LOGIC_OP, cur_state.logic_op.enabled, logic_op.enabled);
if (UpdateValue(cur_state.logic_op.operation, logic_op.operation)) {
@@ -478,7 +376,12 @@ void OpenGLState::ApplyLogicOp() const {
}
}
-void OpenGLState::ApplyPolygonOffset() const {
+void OpenGLState::ApplyPolygonOffset() {
+ if (!dirty.polygon_offset) {
+ return;
+ }
+ dirty.polygon_offset = false;
+
Enable(GL_POLYGON_OFFSET_FILL, cur_state.polygon_offset.fill_enable,
polygon_offset.fill_enable);
Enable(GL_POLYGON_OFFSET_LINE, cur_state.polygon_offset.line_enable,
@@ -499,7 +402,7 @@ void OpenGLState::ApplyPolygonOffset() const {
}
}
-void OpenGLState::ApplyAlphaTest() const {
+void OpenGLState::ApplyAlphaTest() {
Enable(GL_ALPHA_TEST, cur_state.alpha_test.enabled, alpha_test.enabled);
if (UpdateTie(std::tie(cur_state.alpha_test.func, cur_state.alpha_test.ref),
std::tie(alpha_test.func, alpha_test.ref))) {
@@ -507,19 +410,19 @@ void OpenGLState::ApplyAlphaTest() const {
}
}
-void OpenGLState::ApplyTextures() const {
+void OpenGLState::ApplyTextures() {
if (const auto update = UpdateArray(cur_state.textures, textures)) {
glBindTextures(update->first, update->second, textures.data() + update->first);
}
}
-void OpenGLState::ApplySamplers() const {
+void OpenGLState::ApplySamplers() {
if (const auto update = UpdateArray(cur_state.samplers, samplers)) {
glBindSamplers(update->first, update->second, samplers.data() + update->first);
}
}
-void OpenGLState::ApplyImages() const {
+void OpenGLState::ApplyImages() {
if (const auto update = UpdateArray(cur_state.images, images)) {
glBindImageTextures(update->first, update->second, images.data() + update->first);
}
@@ -535,32 +438,20 @@ void OpenGLState::Apply() {
ApplyPointSize();
ApplyFragmentColorClamp();
ApplyMultisample();
- if (dirty.color_mask) {
- ApplyColorMask();
- dirty.color_mask = false;
- }
+ ApplyColorMask();
ApplyDepthClamp();
ApplyViewport();
- if (dirty.stencil_state) {
- ApplyStencilTest();
- dirty.stencil_state = false;
- }
+ ApplyStencilTest();
ApplySRgb();
ApplyCulling();
ApplyDepth();
ApplyPrimitiveRestart();
- if (dirty.blend_state) {
- ApplyBlending();
- dirty.blend_state = false;
- }
+ ApplyBlending();
ApplyLogicOp();
ApplyTextures();
ApplySamplers();
ApplyImages();
- if (dirty.polygon_offset) {
- ApplyPolygonOffset();
- dirty.polygon_offset = false;
- }
+ ApplyPolygonOffset();
ApplyAlphaTest();
}
diff --git a/src/video_core/renderer_opengl/gl_state.h b/src/video_core/renderer_opengl/gl_state.h
index c358d3b38..cca25206b 100644
--- a/src/video_core/renderer_opengl/gl_state.h
+++ b/src/video_core/renderer_opengl/gl_state.h
@@ -5,168 +5,146 @@
#pragma once
#include <array>
+#include <type_traits>
#include <glad/glad.h>
#include "video_core/engines/maxwell_3d.h"
namespace OpenGL {
-namespace TextureUnits {
-
-struct TextureUnit {
- GLint id;
- constexpr GLenum Enum() const {
- return static_cast<GLenum>(GL_TEXTURE0 + id);
- }
-};
-
-constexpr TextureUnit MaxwellTexture(int unit) {
- return TextureUnit{unit};
-}
-
-constexpr TextureUnit LightingLUT{3};
-constexpr TextureUnit FogLUT{4};
-constexpr TextureUnit ProcTexNoiseLUT{5};
-constexpr TextureUnit ProcTexColorMap{6};
-constexpr TextureUnit ProcTexAlphaMap{7};
-constexpr TextureUnit ProcTexLUT{8};
-constexpr TextureUnit ProcTexDiffLUT{9};
-
-} // namespace TextureUnits
-
class OpenGLState {
public:
struct {
- bool enabled; // GL_FRAMEBUFFER_SRGB
+ bool enabled = false; // GL_FRAMEBUFFER_SRGB
} framebuffer_srgb;
struct {
- bool alpha_to_coverage; // GL_ALPHA_TO_COVERAGE
- bool alpha_to_one; // GL_ALPHA_TO_ONE
+ bool alpha_to_coverage = false; // GL_ALPHA_TO_COVERAGE
+ bool alpha_to_one = false; // GL_ALPHA_TO_ONE
} multisample_control;
struct {
- bool enabled; // GL_CLAMP_FRAGMENT_COLOR_ARB
+ bool enabled = false; // GL_CLAMP_FRAGMENT_COLOR_ARB
} fragment_color_clamp;
struct {
- bool far_plane;
- bool near_plane;
+ bool far_plane = false;
+ bool near_plane = false;
} depth_clamp; // GL_DEPTH_CLAMP
struct {
- bool enabled; // GL_CULL_FACE
- GLenum mode; // GL_CULL_FACE_MODE
- GLenum front_face; // GL_FRONT_FACE
+ bool enabled = false; // GL_CULL_FACE
+ GLenum mode = GL_BACK; // GL_CULL_FACE_MODE
+ GLenum front_face = GL_CCW; // GL_FRONT_FACE
} cull;
struct {
- bool test_enabled; // GL_DEPTH_TEST
- GLenum test_func; // GL_DEPTH_FUNC
- GLboolean write_mask; // GL_DEPTH_WRITEMASK
+ bool test_enabled = false; // GL_DEPTH_TEST
+ GLboolean write_mask = GL_TRUE; // GL_DEPTH_WRITEMASK
+ GLenum test_func = GL_LESS; // GL_DEPTH_FUNC
} depth;
struct {
- bool enabled;
- GLuint index;
+ bool enabled = false;
+ GLuint index = 0;
} primitive_restart; // GL_PRIMITIVE_RESTART
struct ColorMask {
- GLboolean red_enabled;
- GLboolean green_enabled;
- GLboolean blue_enabled;
- GLboolean alpha_enabled;
+ GLboolean red_enabled = GL_TRUE;
+ GLboolean green_enabled = GL_TRUE;
+ GLboolean blue_enabled = GL_TRUE;
+ GLboolean alpha_enabled = GL_TRUE;
};
std::array<ColorMask, Tegra::Engines::Maxwell3D::Regs::NumRenderTargets>
color_mask; // GL_COLOR_WRITEMASK
struct {
- bool test_enabled; // GL_STENCIL_TEST
+ bool test_enabled = false; // GL_STENCIL_TEST
struct {
- GLenum test_func; // GL_STENCIL_FUNC
- GLint test_ref; // GL_STENCIL_REF
- GLuint test_mask; // GL_STENCIL_VALUE_MASK
- GLuint write_mask; // GL_STENCIL_WRITEMASK
- GLenum action_stencil_fail; // GL_STENCIL_FAIL
- GLenum action_depth_fail; // GL_STENCIL_PASS_DEPTH_FAIL
- GLenum action_depth_pass; // GL_STENCIL_PASS_DEPTH_PASS
+ GLenum test_func = GL_ALWAYS; // GL_STENCIL_FUNC
+ GLint test_ref = 0; // GL_STENCIL_REF
+ GLuint test_mask = 0xFFFFFFFF; // GL_STENCIL_VALUE_MASK
+ GLuint write_mask = 0xFFFFFFFF; // GL_STENCIL_WRITEMASK
+ GLenum action_stencil_fail = GL_KEEP; // GL_STENCIL_FAIL
+ GLenum action_depth_fail = GL_KEEP; // GL_STENCIL_PASS_DEPTH_FAIL
+ GLenum action_depth_pass = GL_KEEP; // GL_STENCIL_PASS_DEPTH_PASS
} front, back;
} stencil;
struct Blend {
- bool enabled; // GL_BLEND
- GLenum rgb_equation; // GL_BLEND_EQUATION_RGB
- GLenum a_equation; // GL_BLEND_EQUATION_ALPHA
- GLenum src_rgb_func; // GL_BLEND_SRC_RGB
- GLenum dst_rgb_func; // GL_BLEND_DST_RGB
- GLenum src_a_func; // GL_BLEND_SRC_ALPHA
- GLenum dst_a_func; // GL_BLEND_DST_ALPHA
+ bool enabled = false; // GL_BLEND
+ GLenum rgb_equation = GL_FUNC_ADD; // GL_BLEND_EQUATION_RGB
+ GLenum a_equation = GL_FUNC_ADD; // GL_BLEND_EQUATION_ALPHA
+ GLenum src_rgb_func = GL_ONE; // GL_BLEND_SRC_RGB
+ GLenum dst_rgb_func = GL_ZERO; // GL_BLEND_DST_RGB
+ GLenum src_a_func = GL_ONE; // GL_BLEND_SRC_ALPHA
+ GLenum dst_a_func = GL_ZERO; // GL_BLEND_DST_ALPHA
};
std::array<Blend, Tegra::Engines::Maxwell3D::Regs::NumRenderTargets> blend;
struct {
- bool enabled;
+ bool enabled = false;
} independant_blend;
struct {
- GLclampf red;
- GLclampf green;
- GLclampf blue;
- GLclampf alpha;
+ GLclampf red = 0.0f;
+ GLclampf green = 0.0f;
+ GLclampf blue = 0.0f;
+ GLclampf alpha = 0.0f;
} blend_color; // GL_BLEND_COLOR
struct {
- bool enabled; // GL_LOGIC_OP_MODE
- GLenum operation;
+ bool enabled = false; // GL_LOGIC_OP_MODE
+ GLenum operation = GL_COPY;
} logic_op;
- std::array<GLuint, Tegra::Engines::Maxwell3D::Regs::NumTextureSamplers> textures{};
- std::array<GLuint, Tegra::Engines::Maxwell3D::Regs::NumTextureSamplers> samplers{};
- std::array<GLuint, Tegra::Engines::Maxwell3D::Regs::NumImages> images{};
+ std::array<GLuint, Tegra::Engines::Maxwell3D::Regs::NumTextureSamplers> textures = {};
+ std::array<GLuint, Tegra::Engines::Maxwell3D::Regs::NumTextureSamplers> samplers = {};
+ std::array<GLuint, Tegra::Engines::Maxwell3D::Regs::NumImages> images = {};
struct {
- GLuint read_framebuffer; // GL_READ_FRAMEBUFFER_BINDING
- GLuint draw_framebuffer; // GL_DRAW_FRAMEBUFFER_BINDING
- GLuint vertex_array; // GL_VERTEX_ARRAY_BINDING
- GLuint shader_program; // GL_CURRENT_PROGRAM
- GLuint program_pipeline; // GL_PROGRAM_PIPELINE_BINDING
+ GLuint read_framebuffer = 0; // GL_READ_FRAMEBUFFER_BINDING
+ GLuint draw_framebuffer = 0; // GL_DRAW_FRAMEBUFFER_BINDING
+ GLuint vertex_array = 0; // GL_VERTEX_ARRAY_BINDING
+ GLuint shader_program = 0; // GL_CURRENT_PROGRAM
+ GLuint program_pipeline = 0; // GL_PROGRAM_PIPELINE_BINDING
} draw;
- struct viewport {
- GLint x;
- GLint y;
- GLint width;
- GLint height;
- GLfloat depth_range_near; // GL_DEPTH_RANGE
- GLfloat depth_range_far; // GL_DEPTH_RANGE
+ struct Viewport {
+ GLint x = 0;
+ GLint y = 0;
+ GLint width = 0;
+ GLint height = 0;
+ GLfloat depth_range_near = 0.0f; // GL_DEPTH_RANGE
+ GLfloat depth_range_far = 1.0f; // GL_DEPTH_RANGE
struct {
- bool enabled; // GL_SCISSOR_TEST
- GLint x;
- GLint y;
- GLsizei width;
- GLsizei height;
+ bool enabled = false; // GL_SCISSOR_TEST
+ GLint x = 0;
+ GLint y = 0;
+ GLsizei width = 0;
+ GLsizei height = 0;
} scissor;
};
- std::array<viewport, Tegra::Engines::Maxwell3D::Regs::NumViewports> viewports;
+ std::array<Viewport, Tegra::Engines::Maxwell3D::Regs::NumViewports> viewports;
struct {
- float size; // GL_POINT_SIZE
+ float size = 1.0f; // GL_POINT_SIZE
} point;
struct {
- bool point_enable;
- bool line_enable;
- bool fill_enable;
- GLfloat units;
- GLfloat factor;
- GLfloat clamp;
+ bool point_enable = false;
+ bool line_enable = false;
+ bool fill_enable = false;
+ GLfloat units = 0.0f;
+ GLfloat factor = 0.0f;
+ GLfloat clamp = 0.0f;
} polygon_offset;
struct {
- bool enabled; // GL_ALPHA_TEST
- GLenum func; // GL_ALPHA_TEST_FUNC
- GLfloat ref; // GL_ALPHA_TEST_REF
+ bool enabled = false; // GL_ALPHA_TEST
+ GLenum func = GL_ALWAYS; // GL_ALPHA_TEST_FUNC
+ GLfloat ref = 0.0f; // GL_ALPHA_TEST_REF
} alpha_test;
- std::array<bool, 8> clip_distance; // GL_CLIP_DISTANCE
+ std::array<bool, 8> clip_distance = {}; // GL_CLIP_DISTANCE
OpenGLState();
@@ -179,34 +157,31 @@ public:
/// Apply this state as the current OpenGL state
void Apply();
- void ApplyFramebufferState() const;
- void ApplyVertexArrayState() const;
- void ApplyShaderProgram() const;
- void ApplyProgramPipeline() const;
- void ApplyClipDistances() const;
- void ApplyPointSize() const;
- void ApplyFragmentColorClamp() const;
- void ApplyMultisample() const;
- void ApplySRgb() const;
- void ApplyCulling() const;
- void ApplyColorMask() const;
- void ApplyDepth() const;
- void ApplyPrimitiveRestart() const;
- void ApplyStencilTest() const;
- void ApplyViewport() const;
- void ApplyTargetBlending(std::size_t target, bool force) const;
- void ApplyGlobalBlending() const;
- void ApplyBlending() const;
- void ApplyLogicOp() const;
- void ApplyTextures() const;
- void ApplySamplers() const;
- void ApplyImages() const;
- void ApplyDepthClamp() const;
- void ApplyPolygonOffset() const;
- void ApplyAlphaTest() const;
-
- /// Set the initial OpenGL state
- static void ApplyDefaultState();
+ void ApplyFramebufferState();
+ void ApplyVertexArrayState();
+ void ApplyShaderProgram();
+ void ApplyProgramPipeline();
+ void ApplyClipDistances();
+ void ApplyPointSize();
+ void ApplyFragmentColorClamp();
+ void ApplyMultisample();
+ void ApplySRgb();
+ void ApplyCulling();
+ void ApplyColorMask();
+ void ApplyDepth();
+ void ApplyPrimitiveRestart();
+ void ApplyStencilTest();
+ void ApplyViewport();
+ void ApplyTargetBlending(std::size_t target, bool force);
+ void ApplyGlobalBlending();
+ void ApplyBlending();
+ void ApplyLogicOp();
+ void ApplyTextures();
+ void ApplySamplers();
+ void ApplyImages();
+ void ApplyDepthClamp();
+ void ApplyPolygonOffset();
+ void ApplyAlphaTest();
/// Resets any references to the given resource
OpenGLState& UnbindTexture(GLuint handle);
@@ -253,5 +228,6 @@ private:
bool color_mask;
} dirty{};
};
+static_assert(std::is_trivially_copyable_v<OpenGLState>);
} // namespace OpenGL
diff --git a/src/video_core/renderer_opengl/gl_texture_cache.cpp b/src/video_core/renderer_opengl/gl_texture_cache.cpp
index 2f9bfd7e4..55b3e58b2 100644
--- a/src/video_core/renderer_opengl/gl_texture_cache.cpp
+++ b/src/video_core/renderer_opengl/gl_texture_cache.cpp
@@ -131,6 +131,7 @@ constexpr std::array<FormatTuple, VideoCore::Surface::MaxPixelFormat> tex_format
{GL_SRGB8_ALPHA8, GL_RGBA, GL_UNSIGNED_BYTE, ComponentType::UNorm, false}, // ASTC_2D_8X6_SRGB
{GL_RGBA8, GL_RGBA, GL_UNSIGNED_BYTE, ComponentType::UNorm, false}, // ASTC_2D_6X5
{GL_SRGB8_ALPHA8, GL_RGBA, GL_UNSIGNED_BYTE, ComponentType::UNorm, false}, // ASTC_2D_6X5_SRGB
+ {GL_RGB9_E5, GL_RGB, GL_UNSIGNED_INT_5_9_9_9_REV, ComponentType::Float, false}, // E5B9G9R9F
// Depth formats
{GL_DEPTH_COMPONENT32F, GL_DEPTH_COMPONENT, GL_FLOAT, ComponentType::Float, false}, // Z32F
diff --git a/src/video_core/surface.cpp b/src/video_core/surface.cpp
index 9a3c05288..621136b6e 100644
--- a/src/video_core/surface.cpp
+++ b/src/video_core/surface.cpp
@@ -315,6 +315,14 @@ PixelFormat PixelFormatFromTextureFormat(Tegra::Texture::TextureFormat format,
break;
}
break;
+ case Tegra::Texture::TextureFormat::E5B9G9R9_SHAREDEXP:
+ switch (component_type) {
+ case Tegra::Texture::ComponentType::FLOAT:
+ return PixelFormat::E5B9G9R9F;
+ default:
+ break;
+ }
+ break;
case Tegra::Texture::TextureFormat::ZF32:
return PixelFormat::Z32F;
case Tegra::Texture::TextureFormat::Z16:
diff --git a/src/video_core/surface.h b/src/video_core/surface.h
index 97668f802..d3bcd38c5 100644
--- a/src/video_core/surface.h
+++ b/src/video_core/surface.h
@@ -86,19 +86,20 @@ enum class PixelFormat {
ASTC_2D_8X6_SRGB = 68,
ASTC_2D_6X5 = 69,
ASTC_2D_6X5_SRGB = 70,
+ E5B9G9R9F = 71,
MaxColorFormat,
// Depth formats
- Z32F = 71,
- Z16 = 72,
+ Z32F = 72,
+ Z16 = 73,
MaxDepthFormat,
// DepthStencil formats
- Z24S8 = 73,
- S8Z24 = 74,
- Z32FS8 = 75,
+ Z24S8 = 74,
+ S8Z24 = 75,
+ Z32FS8 = 76,
MaxDepthStencilFormat,
@@ -207,6 +208,7 @@ constexpr std::array<u32, MaxPixelFormat> compression_factor_shift_table = {{
2, // ASTC_2D_8X6_SRGB
2, // ASTC_2D_6X5
2, // ASTC_2D_6X5_SRGB
+ 0, // E5B9G9R9F
0, // Z32F
0, // Z16
0, // Z24S8
@@ -302,6 +304,7 @@ constexpr std::array<u32, MaxPixelFormat> block_width_table = {{
8, // ASTC_2D_8X6_SRGB
6, // ASTC_2D_6X5
6, // ASTC_2D_6X5_SRGB
+ 1, // E5B9G9R9F
1, // Z32F
1, // Z16
1, // Z24S8
@@ -389,6 +392,7 @@ constexpr std::array<u32, MaxPixelFormat> block_height_table = {{
6, // ASTC_2D_8X6_SRGB
5, // ASTC_2D_6X5
5, // ASTC_2D_6X5_SRGB
+ 1, // E5B9G9R9F
1, // Z32F
1, // Z16
1, // Z24S8
@@ -476,6 +480,7 @@ constexpr std::array<u32, MaxPixelFormat> bpp_table = {{
128, // ASTC_2D_8X6_SRGB
128, // ASTC_2D_6X5
128, // ASTC_2D_6X5_SRGB
+ 32, // E5B9G9R9F
32, // Z32F
16, // Z16
32, // Z24S8
@@ -578,6 +583,7 @@ constexpr std::array<SurfaceCompression, MaxPixelFormat> compression_type_table
SurfaceCompression::Converted, // ASTC_2D_8X6_SRGB
SurfaceCompression::Converted, // ASTC_2D_6X5
SurfaceCompression::Converted, // ASTC_2D_6X5_SRGB
+ SurfaceCompression::None, // E5B9G9R9F
SurfaceCompression::None, // Z32F
SurfaceCompression::None, // Z16
SurfaceCompression::None, // Z24S8
diff --git a/src/video_core/textures/texture.h b/src/video_core/textures/texture.h
index e36bc2c04..0429af9c1 100644
--- a/src/video_core/textures/texture.h
+++ b/src/video_core/textures/texture.h
@@ -354,7 +354,6 @@ struct TSCEntry {
static_assert(sizeof(TSCEntry) == 0x20, "TSCEntry has wrong size");
struct FullTextureInfo {
- u32 index;
TICEntry tic;
TSCEntry tsc;
};
generated by cgit v1.2.3 (git 2.25.1) at 2024-11-13 09:10:36 +0100