Merge pull request #2318 from ReinUsesLisp/sampler-cache

gl_sampler_cache: Port sampler cache to OpenGL
merge-requests/60/head
bunnei 2019-04-17 21:45:56 +07:00 committed by GitHub
commit 4294062516
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
9 changed files with 183 additions and 181 deletions

@ -46,6 +46,8 @@ add_library(video_core STATIC
renderer_opengl/gl_rasterizer_cache.h renderer_opengl/gl_rasterizer_cache.h
renderer_opengl/gl_resource_manager.cpp renderer_opengl/gl_resource_manager.cpp
renderer_opengl/gl_resource_manager.h renderer_opengl/gl_resource_manager.h
renderer_opengl/gl_sampler_cache.cpp
renderer_opengl/gl_sampler_cache.h
renderer_opengl/gl_shader_cache.cpp renderer_opengl/gl_shader_cache.cpp
renderer_opengl/gl_shader_cache.h renderer_opengl/gl_shader_cache.h
renderer_opengl/gl_shader_decompiler.cpp renderer_opengl/gl_shader_decompiler.cpp
@ -67,6 +69,8 @@ add_library(video_core STATIC
renderer_opengl/renderer_opengl.h renderer_opengl/renderer_opengl.h
renderer_opengl/utils.cpp renderer_opengl/utils.cpp
renderer_opengl/utils.h renderer_opengl/utils.h
sampler_cache.cpp
sampler_cache.h
shader/decode/arithmetic.cpp shader/decode/arithmetic.cpp
shader/decode/arithmetic_immediate.cpp shader/decode/arithmetic_immediate.cpp
shader/decode/bfe.cpp shader/decode/bfe.cpp

@ -101,12 +101,6 @@ struct FramebufferCacheKey {
RasterizerOpenGL::RasterizerOpenGL(Core::System& system, ScreenInfo& info) RasterizerOpenGL::RasterizerOpenGL(Core::System& system, ScreenInfo& info)
: res_cache{*this}, shader_cache{*this, system}, global_cache{*this}, system{system}, : res_cache{*this}, shader_cache{*this, system}, global_cache{*this}, system{system},
screen_info{info}, buffer_cache(*this, STREAM_BUFFER_SIZE) { screen_info{info}, buffer_cache(*this, STREAM_BUFFER_SIZE) {
// Create sampler objects
for (std::size_t i = 0; i < texture_samplers.size(); ++i) {
texture_samplers[i].Create();
state.texture_units[i].sampler = texture_samplers[i].sampler.handle;
}
OpenGLState::ApplyDefaultState(); OpenGLState::ApplyDefaultState();
shader_program_manager = std::make_unique<GLShader::ProgramManager>(); shader_program_manager = std::make_unique<GLShader::ProgramManager>();
@ -813,92 +807,6 @@ bool RasterizerOpenGL::AccelerateDisplay(const Tegra::FramebufferConfig& config,
return true; return true;
} }
void RasterizerOpenGL::SamplerInfo::Create() {
sampler.Create();
mag_filter = Tegra::Texture::TextureFilter::Linear;
min_filter = Tegra::Texture::TextureFilter::Linear;
wrap_u = Tegra::Texture::WrapMode::Wrap;
wrap_v = Tegra::Texture::WrapMode::Wrap;
wrap_p = Tegra::Texture::WrapMode::Wrap;
use_depth_compare = false;
depth_compare_func = Tegra::Texture::DepthCompareFunc::Never;
// OpenGL's default is GL_LINEAR_MIPMAP_LINEAR
glSamplerParameteri(sampler.handle, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
glSamplerParameteri(sampler.handle, GL_TEXTURE_COMPARE_FUNC, GL_NEVER);
// Other attributes have correct defaults
}
void RasterizerOpenGL::SamplerInfo::SyncWithConfig(const Tegra::Texture::TSCEntry& config) {
const GLuint sampler_id = sampler.handle;
if (mag_filter != config.mag_filter) {
mag_filter = config.mag_filter;
glSamplerParameteri(
sampler_id, GL_TEXTURE_MAG_FILTER,
MaxwellToGL::TextureFilterMode(mag_filter, Tegra::Texture::TextureMipmapFilter::None));
}
if (min_filter != config.min_filter || mipmap_filter != config.mipmap_filter) {
min_filter = config.min_filter;
mipmap_filter = config.mipmap_filter;
glSamplerParameteri(sampler_id, GL_TEXTURE_MIN_FILTER,
MaxwellToGL::TextureFilterMode(min_filter, mipmap_filter));
}
if (wrap_u != config.wrap_u) {
wrap_u = config.wrap_u;
glSamplerParameteri(sampler_id, GL_TEXTURE_WRAP_S, MaxwellToGL::WrapMode(wrap_u));
}
if (wrap_v != config.wrap_v) {
wrap_v = config.wrap_v;
glSamplerParameteri(sampler_id, GL_TEXTURE_WRAP_T, MaxwellToGL::WrapMode(wrap_v));
}
if (wrap_p != config.wrap_p) {
wrap_p = config.wrap_p;
glSamplerParameteri(sampler_id, GL_TEXTURE_WRAP_R, MaxwellToGL::WrapMode(wrap_p));
}
if (const bool enabled = config.depth_compare_enabled == 1; use_depth_compare != enabled) {
use_depth_compare = enabled;
glSamplerParameteri(sampler_id, GL_TEXTURE_COMPARE_MODE,
use_depth_compare ? GL_COMPARE_REF_TO_TEXTURE : GL_NONE);
}
if (depth_compare_func != config.depth_compare_func) {
depth_compare_func = config.depth_compare_func;
glSamplerParameteri(sampler_id, GL_TEXTURE_COMPARE_FUNC,
MaxwellToGL::DepthCompareFunc(depth_compare_func));
}
if (const auto new_border_color = config.GetBorderColor(); border_color != new_border_color) {
border_color = new_border_color;
glSamplerParameterfv(sampler_id, GL_TEXTURE_BORDER_COLOR, border_color.data());
}
if (const float anisotropic = config.GetMaxAnisotropy(); max_anisotropic != anisotropic) {
max_anisotropic = anisotropic;
if (GLAD_GL_ARB_texture_filter_anisotropic) {
glSamplerParameterf(sampler_id, GL_TEXTURE_MAX_ANISOTROPY, max_anisotropic);
} else if (GLAD_GL_EXT_texture_filter_anisotropic) {
glSamplerParameterf(sampler_id, GL_TEXTURE_MAX_ANISOTROPY_EXT, max_anisotropic);
}
}
if (const float min = config.GetMinLod(); min_lod != min) {
min_lod = min;
glSamplerParameterf(sampler_id, GL_TEXTURE_MIN_LOD, min_lod);
}
if (const float max = config.GetMaxLod(); max_lod != max) {
max_lod = max;
glSamplerParameterf(sampler_id, GL_TEXTURE_MAX_LOD, max_lod);
}
if (const float bias = config.GetLodBias(); lod_bias != bias) {
lod_bias = bias;
glSamplerParameterf(sampler_id, GL_TEXTURE_LOD_BIAS, lod_bias);
}
}
void RasterizerOpenGL::SetupConstBuffers(Tegra::Engines::Maxwell3D::Regs::ShaderStage stage, void RasterizerOpenGL::SetupConstBuffers(Tegra::Engines::Maxwell3D::Regs::ShaderStage stage,
const Shader& shader, GLuint program_handle, const Shader& shader, GLuint program_handle,
BaseBindings base_bindings) { BaseBindings base_bindings) {
@ -985,7 +893,7 @@ void RasterizerOpenGL::SetupTextures(Maxwell::ShaderStage stage, const Shader& s
} }
const u32 current_bindpoint = base_bindings.sampler + bindpoint; const u32 current_bindpoint = base_bindings.sampler + bindpoint;
texture_samplers[current_bindpoint].SyncWithConfig(texture.tsc); state.texture_units[current_bindpoint].sampler = sampler_cache.GetSampler(texture.tsc);
if (Surface surface = res_cache.GetTextureSurface(texture, entry); surface) { if (Surface surface = res_cache.GetTextureSurface(texture, entry); surface) {
state.texture_units[current_bindpoint].texture = state.texture_units[current_bindpoint].texture =

@ -25,6 +25,7 @@
#include "video_core/renderer_opengl/gl_primitive_assembler.h" #include "video_core/renderer_opengl/gl_primitive_assembler.h"
#include "video_core/renderer_opengl/gl_rasterizer_cache.h" #include "video_core/renderer_opengl/gl_rasterizer_cache.h"
#include "video_core/renderer_opengl/gl_resource_manager.h" #include "video_core/renderer_opengl/gl_resource_manager.h"
#include "video_core/renderer_opengl/gl_sampler_cache.h"
#include "video_core/renderer_opengl/gl_shader_cache.h" #include "video_core/renderer_opengl/gl_shader_cache.h"
#include "video_core/renderer_opengl/gl_shader_manager.h" #include "video_core/renderer_opengl/gl_shader_manager.h"
#include "video_core/renderer_opengl/gl_state.h" #include "video_core/renderer_opengl/gl_state.h"
@ -72,34 +73,6 @@ public:
"The maximum size of a constbuffer must be a multiple of the size of GLvec4"); "The maximum size of a constbuffer must be a multiple of the size of GLvec4");
private: private:
class SamplerInfo {
public:
OGLSampler sampler;
/// Creates the sampler object, initializing its state so that it's in sync with the
/// SamplerInfo struct.
void Create();
/// Syncs the sampler object with the config, updating any necessary state.
void SyncWithConfig(const Tegra::Texture::TSCEntry& info);
private:
Tegra::Texture::TextureFilter mag_filter = Tegra::Texture::TextureFilter::Nearest;
Tegra::Texture::TextureFilter min_filter = Tegra::Texture::TextureFilter::Nearest;
Tegra::Texture::TextureMipmapFilter mipmap_filter =
Tegra::Texture::TextureMipmapFilter::None;
Tegra::Texture::WrapMode wrap_u = Tegra::Texture::WrapMode::ClampToEdge;
Tegra::Texture::WrapMode wrap_v = Tegra::Texture::WrapMode::ClampToEdge;
Tegra::Texture::WrapMode wrap_p = Tegra::Texture::WrapMode::ClampToEdge;
bool use_depth_compare = false;
Tegra::Texture::DepthCompareFunc depth_compare_func =
Tegra::Texture::DepthCompareFunc::Always;
GLvec4 border_color = {};
float min_lod = 0.0f;
float max_lod = 16.0f;
float lod_bias = 0.0f;
float max_anisotropic = 1.0f;
};
struct FramebufferConfigState { struct FramebufferConfigState {
bool using_color_fb{}; bool using_color_fb{};
bool using_depth_fb{}; bool using_depth_fb{};
@ -204,6 +177,7 @@ private:
RasterizerCacheOpenGL res_cache; RasterizerCacheOpenGL res_cache;
ShaderCacheOpenGL shader_cache; ShaderCacheOpenGL shader_cache;
GlobalRegionCacheOpenGL global_cache; GlobalRegionCacheOpenGL global_cache;
SamplerCacheOpenGL sampler_cache;
Core::System& system; Core::System& system;
@ -219,8 +193,6 @@ private:
FramebufferConfigState current_framebuffer_config_state; FramebufferConfigState current_framebuffer_config_state;
std::pair<bool, bool> current_depth_stencil_usage{}; std::pair<bool, bool> current_depth_stencil_usage{};
std::array<SamplerInfo, Tegra::Engines::Maxwell3D::Regs::NumTextureSamplers> texture_samplers;
static constexpr std::size_t STREAM_BUFFER_SIZE = 128 * 1024 * 1024; static constexpr std::size_t STREAM_BUFFER_SIZE = 128 * 1024 * 1024;
OGLBufferCache buffer_cache; OGLBufferCache buffer_cache;
PrimitiveAssembler primitive_assembler{buffer_cache}; PrimitiveAssembler primitive_assembler{buffer_cache};

@ -0,0 +1,52 @@
// Copyright 2019 yuzu Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
#include "common/logging/log.h"
#include "video_core/renderer_opengl/gl_resource_manager.h"
#include "video_core/renderer_opengl/gl_sampler_cache.h"
#include "video_core/renderer_opengl/maxwell_to_gl.h"
namespace OpenGL {
SamplerCacheOpenGL::SamplerCacheOpenGL() = default;
SamplerCacheOpenGL::~SamplerCacheOpenGL() = default;
OGLSampler SamplerCacheOpenGL::CreateSampler(const Tegra::Texture::TSCEntry& tsc) const {
OGLSampler sampler;
sampler.Create();
const GLuint sampler_id{sampler.handle};
glSamplerParameteri(
sampler_id, GL_TEXTURE_MAG_FILTER,
MaxwellToGL::TextureFilterMode(tsc.mag_filter, Tegra::Texture::TextureMipmapFilter::None));
glSamplerParameteri(sampler_id, GL_TEXTURE_MIN_FILTER,
MaxwellToGL::TextureFilterMode(tsc.min_filter, tsc.mipmap_filter));
glSamplerParameteri(sampler_id, GL_TEXTURE_WRAP_S, MaxwellToGL::WrapMode(tsc.wrap_u));
glSamplerParameteri(sampler_id, GL_TEXTURE_WRAP_T, MaxwellToGL::WrapMode(tsc.wrap_v));
glSamplerParameteri(sampler_id, GL_TEXTURE_WRAP_R, MaxwellToGL::WrapMode(tsc.wrap_p));
glSamplerParameteri(sampler_id, GL_TEXTURE_COMPARE_MODE,
tsc.depth_compare_enabled == 1 ? GL_COMPARE_REF_TO_TEXTURE : GL_NONE);
glSamplerParameteri(sampler_id, GL_TEXTURE_COMPARE_FUNC,
MaxwellToGL::DepthCompareFunc(tsc.depth_compare_func));
glSamplerParameterfv(sampler_id, GL_TEXTURE_BORDER_COLOR, tsc.GetBorderColor().data());
glSamplerParameterf(sampler_id, GL_TEXTURE_MIN_LOD, tsc.GetMinLod());
glSamplerParameterf(sampler_id, GL_TEXTURE_MAX_LOD, tsc.GetMaxLod());
glSamplerParameterf(sampler_id, GL_TEXTURE_LOD_BIAS, tsc.GetLodBias());
if (GLAD_GL_ARB_texture_filter_anisotropic) {
glSamplerParameterf(sampler_id, GL_TEXTURE_MAX_ANISOTROPY, tsc.GetMaxAnisotropy());
} else if (GLAD_GL_EXT_texture_filter_anisotropic) {
glSamplerParameterf(sampler_id, GL_TEXTURE_MAX_ANISOTROPY_EXT, tsc.GetMaxAnisotropy());
} else if (tsc.GetMaxAnisotropy() != 1) {
LOG_WARNING(Render_OpenGL, "Anisotropy not supported by host GPU driver");
}
return sampler;
}
GLuint SamplerCacheOpenGL::ToSamplerType(const OGLSampler& sampler) const {
return sampler.handle;
}
} // namespace OpenGL

@ -0,0 +1,25 @@
// Copyright 2019 yuzu Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
#pragma once
#include <glad/glad.h>
#include "video_core/renderer_opengl/gl_resource_manager.h"
#include "video_core/sampler_cache.h"
namespace OpenGL {
class SamplerCacheOpenGL final : public VideoCommon::SamplerCache<GLuint, OGLSampler> {
public:
explicit SamplerCacheOpenGL();
~SamplerCacheOpenGL();
protected:
OGLSampler CreateSampler(const Tegra::Texture::TSCEntry& tsc) const;
GLuint ToSamplerType(const OGLSampler& sampler) const;
};
} // namespace OpenGL

@ -7,7 +7,6 @@
#include <unordered_map> #include <unordered_map>
#include "common/assert.h" #include "common/assert.h"
#include "common/cityhash.h"
#include "video_core/renderer_vulkan/declarations.h" #include "video_core/renderer_vulkan/declarations.h"
#include "video_core/renderer_vulkan/maxwell_to_vk.h" #include "video_core/renderer_vulkan/maxwell_to_vk.h"
#include "video_core/renderer_vulkan/vk_sampler_cache.h" #include "video_core/renderer_vulkan/vk_sampler_cache.h"
@ -28,39 +27,20 @@ static std::optional<vk::BorderColor> TryConvertBorderColor(std::array<float, 4>
} }
} }
std::size_t SamplerCacheKey::Hash() const {
static_assert(sizeof(raw) % sizeof(u64) == 0);
return static_cast<std::size_t>(
Common::CityHash64(reinterpret_cast<const char*>(raw.data()), sizeof(raw) / sizeof(u64)));
}
bool SamplerCacheKey::operator==(const SamplerCacheKey& rhs) const {
return raw == rhs.raw;
}
VKSamplerCache::VKSamplerCache(const VKDevice& device) : device{device} {} VKSamplerCache::VKSamplerCache(const VKDevice& device) : device{device} {}
VKSamplerCache::~VKSamplerCache() = default; VKSamplerCache::~VKSamplerCache() = default;
vk::Sampler VKSamplerCache::GetSampler(const Tegra::Texture::TSCEntry& tsc) { UniqueSampler VKSamplerCache::CreateSampler(const Tegra::Texture::TSCEntry& tsc) const {
const auto [entry, is_cache_miss] = cache.try_emplace(SamplerCacheKey{tsc}); const float max_anisotropy{tsc.GetMaxAnisotropy()};
auto& sampler = entry->second; const bool has_anisotropy{max_anisotropy > 1.0f};
if (is_cache_miss) {
sampler = CreateSampler(tsc);
}
return *sampler;
}
UniqueSampler VKSamplerCache::CreateSampler(const Tegra::Texture::TSCEntry& tsc) { const auto border_color{tsc.GetBorderColor()};
const float max_anisotropy = tsc.GetMaxAnisotropy(); const auto vk_border_color{TryConvertBorderColor(border_color)};
const bool has_anisotropy = max_anisotropy > 1.0f;
const auto border_color = tsc.GetBorderColor();
const auto vk_border_color = TryConvertBorderColor(border_color);
UNIMPLEMENTED_IF_MSG(!vk_border_color, "Unimplemented border color {} {} {} {}", UNIMPLEMENTED_IF_MSG(!vk_border_color, "Unimplemented border color {} {} {} {}",
border_color[0], border_color[1], border_color[2], border_color[3]); border_color[0], border_color[1], border_color[2], border_color[3]);
constexpr bool unnormalized_coords = false; constexpr bool unnormalized_coords{false};
const vk::SamplerCreateInfo sampler_ci( const vk::SamplerCreateInfo sampler_ci(
{}, MaxwellToVK::Sampler::Filter(tsc.mag_filter), {}, MaxwellToVK::Sampler::Filter(tsc.mag_filter),
@ -73,9 +53,13 @@ UniqueSampler VKSamplerCache::CreateSampler(const Tegra::Texture::TSCEntry& tsc)
tsc.GetMaxLod(), vk_border_color.value_or(vk::BorderColor::eFloatTransparentBlack), tsc.GetMaxLod(), vk_border_color.value_or(vk::BorderColor::eFloatTransparentBlack),
unnormalized_coords); unnormalized_coords);
const auto& dld = device.GetDispatchLoader(); const auto& dld{device.GetDispatchLoader()};
const auto dev = device.GetLogical(); const auto dev{device.GetLogical()};
return dev.createSamplerUnique(sampler_ci, nullptr, dld); return dev.createSamplerUnique(sampler_ci, nullptr, dld);
} }
vk::Sampler VKSamplerCache::ToSamplerType(const UniqueSampler& sampler) const {
return *sampler;
}
} // namespace Vulkan } // namespace Vulkan

@ -8,49 +8,25 @@
#include "common/common_types.h" #include "common/common_types.h"
#include "video_core/renderer_vulkan/declarations.h" #include "video_core/renderer_vulkan/declarations.h"
#include "video_core/sampler_cache.h"
#include "video_core/textures/texture.h" #include "video_core/textures/texture.h"
namespace Vulkan { namespace Vulkan {
class VKDevice; class VKDevice;
struct SamplerCacheKey final : public Tegra::Texture::TSCEntry { class VKSamplerCache final : public VideoCommon::SamplerCache<vk::Sampler, UniqueSampler> {
std::size_t Hash() const;
bool operator==(const SamplerCacheKey& rhs) const;
bool operator!=(const SamplerCacheKey& rhs) const {
return !operator==(rhs);
}
};
} // namespace Vulkan
namespace std {
template <>
struct hash<Vulkan::SamplerCacheKey> {
std::size_t operator()(const Vulkan::SamplerCacheKey& k) const noexcept {
return k.Hash();
}
};
} // namespace std
namespace Vulkan {
class VKSamplerCache {
public: public:
explicit VKSamplerCache(const VKDevice& device); explicit VKSamplerCache(const VKDevice& device);
~VKSamplerCache(); ~VKSamplerCache();
vk::Sampler GetSampler(const Tegra::Texture::TSCEntry& tsc); protected:
UniqueSampler CreateSampler(const Tegra::Texture::TSCEntry& tsc) const;
vk::Sampler ToSamplerType(const UniqueSampler& sampler) const;
private: private:
UniqueSampler CreateSampler(const Tegra::Texture::TSCEntry& tsc);
const VKDevice& device; const VKDevice& device;
std::unordered_map<SamplerCacheKey, UniqueSampler> cache;
}; };
} // namespace Vulkan } // namespace Vulkan

@ -0,0 +1,21 @@
// Copyright 2019 yuzu Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
#include "common/cityhash.h"
#include "common/common_types.h"
#include "video_core/sampler_cache.h"
namespace VideoCommon {
std::size_t SamplerCacheKey::Hash() const {
static_assert(sizeof(raw) % sizeof(u64) == 0);
return static_cast<std::size_t>(
Common::CityHash64(reinterpret_cast<const char*>(raw.data()), sizeof(raw) / sizeof(u64)));
}
bool SamplerCacheKey::operator==(const SamplerCacheKey& rhs) const {
return raw == rhs.raw;
}
} // namespace VideoCommon

@ -0,0 +1,60 @@
// Copyright 2019 yuzu Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
#pragma once
#include <cstddef>
#include <unordered_map>
#include "video_core/textures/texture.h"
namespace VideoCommon {
struct SamplerCacheKey final : public Tegra::Texture::TSCEntry {
std::size_t Hash() const;
bool operator==(const SamplerCacheKey& rhs) const;
bool operator!=(const SamplerCacheKey& rhs) const {
return !operator==(rhs);
}
};
} // namespace VideoCommon
namespace std {
template <>
struct hash<VideoCommon::SamplerCacheKey> {
std::size_t operator()(const VideoCommon::SamplerCacheKey& k) const noexcept {
return k.Hash();
}
};
} // namespace std
namespace VideoCommon {
template <typename SamplerType, typename SamplerStorageType>
class SamplerCache {
public:
SamplerType GetSampler(const Tegra::Texture::TSCEntry& tsc) {
const auto [entry, is_cache_miss] = cache.try_emplace(SamplerCacheKey{tsc});
auto& sampler = entry->second;
if (is_cache_miss) {
sampler = CreateSampler(tsc);
}
return ToSamplerType(sampler);
}
protected:
virtual SamplerStorageType CreateSampler(const Tegra::Texture::TSCEntry& tsc) const = 0;
virtual SamplerType ToSamplerType(const SamplerStorageType& sampler) const = 0;
private:
std::unordered_map<SamplerCacheKey, SamplerStorageType> cache;
};
} // namespace VideoCommon