aboutsummaryrefslogtreecommitdiff
path: root/src/video_core/renderer_opengl
diff options
context:
space:
mode:
Diffstat (limited to 'src/video_core/renderer_opengl')
-rw-r--r--src/video_core/renderer_opengl/gl_buffer_cache.cpp52
-rw-r--r--src/video_core/renderer_opengl/gl_buffer_cache.h39
-rw-r--r--src/video_core/renderer_opengl/gl_device.cpp3
-rw-r--r--src/video_core/renderer_opengl/gl_device.h5
-rw-r--r--src/video_core/renderer_opengl/gl_rasterizer.cpp8
-rw-r--r--src/video_core/renderer_opengl/gl_rasterizer.h1
-rw-r--r--src/video_core/renderer_opengl/gl_shader_cache.cpp14
-rw-r--r--src/video_core/renderer_opengl/gl_shader_cache.h1
-rw-r--r--src/video_core/renderer_opengl/gl_shader_decompiler.cpp67
-rw-r--r--src/video_core/renderer_opengl/gl_texture_cache.cpp4
-rw-r--r--src/video_core/renderer_opengl/gl_texture_cache.h2
11 files changed, 157 insertions, 39 deletions
diff --git a/src/video_core/renderer_opengl/gl_buffer_cache.cpp b/src/video_core/renderer_opengl/gl_buffer_cache.cpp
index 2a9b523f5..f8a807c84 100644
--- a/src/video_core/renderer_opengl/gl_buffer_cache.cpp
+++ b/src/video_core/renderer_opengl/gl_buffer_cache.cpp
@@ -7,28 +7,41 @@
#include <glad/glad.h>
#include "common/assert.h"
+#include "common/microprofile.h"
+#include "video_core/rasterizer_interface.h"
#include "video_core/renderer_opengl/gl_buffer_cache.h"
#include "video_core/renderer_opengl/gl_rasterizer.h"
#include "video_core/renderer_opengl/gl_resource_manager.h"
namespace OpenGL {
+MICROPROFILE_DEFINE(OpenGL_Buffer_Download, "OpenGL", "Buffer Download", MP_RGB(192, 192, 128));
+
+CachedBufferBlock::CachedBufferBlock(CacheAddr cache_addr, const std::size_t size)
+ : VideoCommon::BufferBlock{cache_addr, size} {
+ gl_buffer.Create();
+ glNamedBufferData(gl_buffer.handle, static_cast<GLsizeiptr>(size), nullptr, GL_DYNAMIC_DRAW);
+}
+
+CachedBufferBlock::~CachedBufferBlock() = default;
+
OGLBufferCache::OGLBufferCache(RasterizerOpenGL& rasterizer, Core::System& system,
std::size_t stream_size)
- : VideoCommon::BufferCache<OGLBuffer, GLuint, OGLStreamBuffer>{
+ : VideoCommon::BufferCache<Buffer, GLuint, OGLStreamBuffer>{
rasterizer, system, std::make_unique<OGLStreamBuffer>(stream_size, true)} {}
OGLBufferCache::~OGLBufferCache() = default;
-OGLBuffer OGLBufferCache::CreateBuffer(std::size_t size) {
- OGLBuffer buffer;
- buffer.Create();
- glNamedBufferData(buffer.handle, static_cast<GLsizeiptr>(size), nullptr, GL_DYNAMIC_DRAW);
- return buffer;
+Buffer OGLBufferCache::CreateBlock(CacheAddr cache_addr, std::size_t size) {
+ return std::make_shared<CachedBufferBlock>(cache_addr, size);
+}
+
+void OGLBufferCache::WriteBarrier() {
+ glMemoryBarrier(GL_ALL_BARRIER_BITS);
}
-const GLuint* OGLBufferCache::ToHandle(const OGLBuffer& buffer) {
- return &buffer.handle;
+const GLuint* OGLBufferCache::ToHandle(const Buffer& buffer) {
+ return buffer->GetHandle();
}
const GLuint* OGLBufferCache::GetEmptyBuffer(std::size_t) {
@@ -36,23 +49,24 @@ const GLuint* OGLBufferCache::GetEmptyBuffer(std::size_t) {
return &null_buffer;
}
-void OGLBufferCache::UploadBufferData(const OGLBuffer& buffer, std::size_t offset, std::size_t size,
- const u8* data) {
- glNamedBufferSubData(buffer.handle, static_cast<GLintptr>(offset),
+void OGLBufferCache::UploadBlockData(const Buffer& buffer, std::size_t offset, std::size_t size,
+ const u8* data) {
+ glNamedBufferSubData(*buffer->GetHandle(), static_cast<GLintptr>(offset),
static_cast<GLsizeiptr>(size), data);
}
-void OGLBufferCache::DownloadBufferData(const OGLBuffer& buffer, std::size_t offset,
- std::size_t size, u8* data) {
- glGetNamedBufferSubData(buffer.handle, static_cast<GLintptr>(offset),
+void OGLBufferCache::DownloadBlockData(const Buffer& buffer, std::size_t offset, std::size_t size,
+ u8* data) {
+ MICROPROFILE_SCOPE(OpenGL_Buffer_Download);
+ glGetNamedBufferSubData(*buffer->GetHandle(), static_cast<GLintptr>(offset),
static_cast<GLsizeiptr>(size), data);
}
-void OGLBufferCache::CopyBufferData(const OGLBuffer& src, const OGLBuffer& dst,
- std::size_t src_offset, std::size_t dst_offset,
- std::size_t size) {
- glCopyNamedBufferSubData(src.handle, dst.handle, static_cast<GLintptr>(src_offset),
- static_cast<GLintptr>(dst_offset), static_cast<GLsizeiptr>(size));
+void OGLBufferCache::CopyBlock(const Buffer& src, const Buffer& dst, std::size_t src_offset,
+ std::size_t dst_offset, std::size_t size) {
+ glCopyNamedBufferSubData(*src->GetHandle(), *dst->GetHandle(),
+ static_cast<GLintptr>(src_offset), static_cast<GLintptr>(dst_offset),
+ static_cast<GLsizeiptr>(size));
}
} // namespace OpenGL
diff --git a/src/video_core/renderer_opengl/gl_buffer_cache.h b/src/video_core/renderer_opengl/gl_buffer_cache.h
index 8c8ac4038..022e7bfa9 100644
--- a/src/video_core/renderer_opengl/gl_buffer_cache.h
+++ b/src/video_core/renderer_opengl/gl_buffer_cache.h
@@ -7,7 +7,7 @@
#include <memory>
#include "common/common_types.h"
-#include "video_core/buffer_cache.h"
+#include "video_core/buffer_cache/buffer_cache.h"
#include "video_core/rasterizer_cache.h"
#include "video_core/renderer_opengl/gl_resource_manager.h"
#include "video_core/renderer_opengl/gl_stream_buffer.h"
@@ -21,7 +21,24 @@ namespace OpenGL {
class OGLStreamBuffer;
class RasterizerOpenGL;
-class OGLBufferCache final : public VideoCommon::BufferCache<OGLBuffer, GLuint, OGLStreamBuffer> {
+class CachedBufferBlock;
+
+using Buffer = std::shared_ptr<CachedBufferBlock>;
+
+class CachedBufferBlock : public VideoCommon::BufferBlock {
+public:
+ explicit CachedBufferBlock(CacheAddr cache_addr, const std::size_t size);
+ ~CachedBufferBlock();
+
+ const GLuint* GetHandle() const {
+ return &gl_buffer.handle;
+ }
+
+private:
+ OGLBuffer gl_buffer{};
+};
+
+class OGLBufferCache final : public VideoCommon::BufferCache<Buffer, GLuint, OGLStreamBuffer> {
public:
explicit OGLBufferCache(RasterizerOpenGL& rasterizer, Core::System& system,
std::size_t stream_size);
@@ -30,18 +47,20 @@ public:
const GLuint* GetEmptyBuffer(std::size_t) override;
protected:
- OGLBuffer CreateBuffer(std::size_t size) override;
+ Buffer CreateBlock(CacheAddr cache_addr, std::size_t size) override;
+
+ void WriteBarrier() override;
- const GLuint* ToHandle(const OGLBuffer& buffer) override;
+ const GLuint* ToHandle(const Buffer& buffer) override;
- void UploadBufferData(const OGLBuffer& buffer, std::size_t offset, std::size_t size,
- const u8* data) override;
+ void UploadBlockData(const Buffer& buffer, std::size_t offset, std::size_t size,
+ const u8* data) override;
- void DownloadBufferData(const OGLBuffer& buffer, std::size_t offset, std::size_t size,
- u8* data) override;
+ void DownloadBlockData(const Buffer& buffer, std::size_t offset, std::size_t size,
+ u8* data) override;
- void CopyBufferData(const OGLBuffer& src, const OGLBuffer& dst, std::size_t src_offset,
- std::size_t dst_offset, std::size_t size) override;
+ void CopyBlock(const Buffer& src, const Buffer& dst, std::size_t src_offset,
+ std::size_t dst_offset, std::size_t size) override;
};
} // namespace OpenGL
diff --git a/src/video_core/renderer_opengl/gl_device.cpp b/src/video_core/renderer_opengl/gl_device.cpp
index 85424a4c9..03d434b28 100644
--- a/src/video_core/renderer_opengl/gl_device.cpp
+++ b/src/video_core/renderer_opengl/gl_device.cpp
@@ -27,6 +27,8 @@ Device::Device() {
shader_storage_alignment = GetInteger<std::size_t>(GL_SHADER_STORAGE_BUFFER_OFFSET_ALIGNMENT);
max_vertex_attributes = GetInteger<u32>(GL_MAX_VERTEX_ATTRIBS);
max_varyings = GetInteger<u32>(GL_MAX_VARYING_VECTORS);
+ has_warp_intrinsics = GLAD_GL_NV_gpu_shader5 && GLAD_GL_NV_shader_thread_group &&
+ GLAD_GL_NV_shader_thread_shuffle;
has_vertex_viewport_layer = GLAD_GL_ARB_shader_viewport_layer_array;
has_variable_aoffi = TestVariableAoffi();
has_component_indexing_bug = TestComponentIndexingBug();
@@ -36,6 +38,7 @@ Device::Device(std::nullptr_t) {
uniform_buffer_alignment = 0;
max_vertex_attributes = 16;
max_varyings = 15;
+ has_warp_intrinsics = true;
has_vertex_viewport_layer = true;
has_variable_aoffi = true;
has_component_indexing_bug = false;
diff --git a/src/video_core/renderer_opengl/gl_device.h b/src/video_core/renderer_opengl/gl_device.h
index dc883722d..3ef7c6dd8 100644
--- a/src/video_core/renderer_opengl/gl_device.h
+++ b/src/video_core/renderer_opengl/gl_device.h
@@ -30,6 +30,10 @@ public:
return max_varyings;
}
+ bool HasWarpIntrinsics() const {
+ return has_warp_intrinsics;
+ }
+
bool HasVertexViewportLayer() const {
return has_vertex_viewport_layer;
}
@@ -50,6 +54,7 @@ private:
std::size_t shader_storage_alignment{};
u32 max_vertex_attributes{};
u32 max_varyings{};
+ bool has_warp_intrinsics{};
bool has_vertex_viewport_layer{};
bool has_variable_aoffi{};
bool has_component_indexing_bug{};
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index c28ae795c..bb09ecd52 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -708,8 +708,6 @@ void RasterizerOpenGL::DrawArrays() {
return;
}
- const auto& regs = gpu.regs;
-
SyncColorMask();
SyncFragmentColorClampState();
SyncMultiSampleState();
@@ -863,6 +861,10 @@ void RasterizerOpenGL::FlushAndInvalidateRegion(CacheAddr addr, u64 size) {
InvalidateRegion(addr, size);
}
+void RasterizerOpenGL::FlushCommands() {
+ glFlush();
+}
+
void RasterizerOpenGL::TickFrame() {
buffer_cache.TickFrame();
}
@@ -976,7 +978,7 @@ void RasterizerOpenGL::SetupGlobalMemory(const GLShader::GlobalMemoryEntry& entr
GPUVAddr gpu_addr, std::size_t size) {
const auto alignment{device.GetShaderStorageBufferAlignment()};
const auto [ssbo, buffer_offset] =
- buffer_cache.UploadMemory(gpu_addr, size, alignment, true, entry.IsWritten());
+ buffer_cache.UploadMemory(gpu_addr, size, alignment, entry.IsWritten());
bind_ssbo_pushbuffer.Push(ssbo, buffer_offset, static_cast<GLsizeiptr>(size));
}
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.h b/src/video_core/renderer_opengl/gl_rasterizer.h
index 8b123c48d..9d20a4fbf 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.h
+++ b/src/video_core/renderer_opengl/gl_rasterizer.h
@@ -63,6 +63,7 @@ public:
void FlushRegion(CacheAddr addr, u64 size) override;
void InvalidateRegion(CacheAddr addr, u64 size) override;
void FlushAndInvalidateRegion(CacheAddr addr, u64 size) override;
+ void FlushCommands() override;
void TickFrame() override;
bool AccelerateSurfaceCopy(const Tegra::Engines::Fermi2D::Regs::Surface& src,
const Tegra::Engines::Fermi2D::Regs::Surface& dst,
diff --git a/src/video_core/renderer_opengl/gl_shader_cache.cpp b/src/video_core/renderer_opengl/gl_shader_cache.cpp
index 1c90facc3..cf6a5cddf 100644
--- a/src/video_core/renderer_opengl/gl_shader_cache.cpp
+++ b/src/video_core/renderer_opengl/gl_shader_cache.cpp
@@ -212,7 +212,9 @@ CachedProgram SpecializeShader(const std::string& code, const GLShader::ShaderEn
const auto texture_buffer_usage{variant.texture_buffer_usage};
std::string source = "#version 430 core\n"
- "#extension GL_ARB_separate_shader_objects : enable\n";
+ "#extension GL_ARB_separate_shader_objects : enable\n"
+ "#extension GL_NV_gpu_shader5 : enable\n"
+ "#extension GL_NV_shader_thread_group : enable\n";
if (entries.shader_viewport_layer_array) {
source += "#extension GL_ARB_shader_viewport_layer_array : enable\n";
}
@@ -247,20 +249,24 @@ CachedProgram SpecializeShader(const std::string& code, const GLShader::ShaderEn
if (!texture_buffer_usage.test(i)) {
continue;
}
- source += fmt::format("#define SAMPLER_{}_IS_BUFFER", i);
+ source += fmt::format("#define SAMPLER_{}_IS_BUFFER\n", i);
+ }
+ if (texture_buffer_usage.any()) {
+ source += '\n';
}
if (program_type == ProgramType::Geometry) {
const auto [glsl_topology, debug_name, max_vertices] =
GetPrimitiveDescription(primitive_mode);
- source += "layout (" + std::string(glsl_topology) + ") in;\n";
+ source += "layout (" + std::string(glsl_topology) + ") in;\n\n";
source += "#define MAX_VERTEX_INPUT " + std::to_string(max_vertices) + '\n';
}
if (program_type == ProgramType::Compute) {
source += "layout (local_size_variable) in;\n";
}
+ source += '\n';
source += code;
OGLShader shader;
@@ -289,7 +295,7 @@ std::set<GLenum> GetSupportedFormats() {
CachedShader::CachedShader(const ShaderParameters& params, ProgramType program_type,
GLShader::ProgramResult result)
- : RasterizerCacheObject{params.host_ptr}, host_ptr{params.host_ptr}, cpu_addr{params.cpu_addr},
+ : RasterizerCacheObject{params.host_ptr}, cpu_addr{params.cpu_addr},
unique_identifier{params.unique_identifier}, program_type{program_type},
disk_cache{params.disk_cache}, precompiled_programs{params.precompiled_programs},
entries{result.second}, code{std::move(result.first)}, shader_length{entries.shader_length} {}
diff --git a/src/video_core/renderer_opengl/gl_shader_cache.h b/src/video_core/renderer_opengl/gl_shader_cache.h
index a3106a0ff..2c8faf855 100644
--- a/src/video_core/renderer_opengl/gl_shader_cache.h
+++ b/src/video_core/renderer_opengl/gl_shader_cache.h
@@ -106,7 +106,6 @@ private:
ShaderDiskCacheUsage GetUsage(const ProgramVariant& variant) const;
- u8* host_ptr{};
VAddr cpu_addr{};
u64 unique_identifier{};
ProgramType program_type{};
diff --git a/src/video_core/renderer_opengl/gl_shader_decompiler.cpp b/src/video_core/renderer_opengl/gl_shader_decompiler.cpp
index ffe26b241..359d58cbe 100644
--- a/src/video_core/renderer_opengl/gl_shader_decompiler.cpp
+++ b/src/video_core/renderer_opengl/gl_shader_decompiler.cpp
@@ -565,7 +565,7 @@ private:
case Tegra::Shader::ImageType::Texture1D:
return "image1D";
case Tegra::Shader::ImageType::TextureBuffer:
- return "bufferImage";
+ return "imageBuffer";
case Tegra::Shader::ImageType::Texture1DArray:
return "image1DArray";
case Tegra::Shader::ImageType::Texture2D:
@@ -1136,6 +1136,16 @@ private:
Type::Float);
}
+ std::string FCastHalf0(Operation operation) {
+ const std::string op_a = VisitOperand(operation, 0, Type::HalfFloat);
+ return fmt::format("({})[0]", op_a);
+ }
+
+ std::string FCastHalf1(Operation operation) {
+ const std::string op_a = VisitOperand(operation, 0, Type::HalfFloat);
+ return fmt::format("({})[1]", op_a);
+ }
+
template <Type type>
std::string Min(Operation operation) {
return GenerateBinaryCall(operation, "min", type, type, type);
@@ -1292,6 +1302,11 @@ private:
return ApplyPrecise(operation, BitwiseCastResult(clamped, Type::HalfFloat));
}
+ std::string HCastFloat(Operation operation) {
+ const std::string op_a = VisitOperand(operation, 0, Type::Float);
+ return fmt::format("fromHalf2(vec2({}, 0.0f))", op_a);
+ }
+
std::string HUnpack(Operation operation) {
const std::string operand{VisitOperand(operation, 0, Type::HalfFloat)};
const auto value = [&]() -> std::string {
@@ -1720,6 +1735,48 @@ private:
return "utof(gl_WorkGroupID"s + GetSwizzle(element) + ')';
}
+ std::string BallotThread(Operation operation) {
+ const std::string value = VisitOperand(operation, 0, Type::Bool);
+ if (!device.HasWarpIntrinsics()) {
+ LOG_ERROR(Render_OpenGL,
+ "Nvidia warp intrinsics are not available and its required by a shader");
+ // Stub on non-Nvidia devices by simulating all threads voting the same as the active
+ // one.
+ return fmt::format("utof({} ? 0xFFFFFFFFU : 0U)", value);
+ }
+ return fmt::format("utof(ballotThreadNV({}))", value);
+ }
+
+ std::string Vote(Operation operation, const char* func) {
+ const std::string value = VisitOperand(operation, 0, Type::Bool);
+ if (!device.HasWarpIntrinsics()) {
+ LOG_ERROR(Render_OpenGL,
+ "Nvidia vote intrinsics are not available and its required by a shader");
+ // Stub with a warp size of one.
+ return value;
+ }
+ return fmt::format("{}({})", func, value);
+ }
+
+ std::string VoteAll(Operation operation) {
+ return Vote(operation, "allThreadsNV");
+ }
+
+ std::string VoteAny(Operation operation) {
+ return Vote(operation, "anyThreadNV");
+ }
+
+ std::string VoteEqual(Operation operation) {
+ if (!device.HasWarpIntrinsics()) {
+ LOG_ERROR(Render_OpenGL,
+ "Nvidia vote intrinsics are not available and its required by a shader");
+ // We must return true here since a stub for a theoretical warp size of 1 will always
+ // return an equal result for all its votes.
+ return "true";
+ }
+ return Vote(operation, "allThreadsEqualNV");
+ }
+
static constexpr std::array operation_decompilers = {
&GLSLDecompiler::Assign,
@@ -1732,6 +1789,8 @@ private:
&GLSLDecompiler::Negate<Type::Float>,
&GLSLDecompiler::Absolute<Type::Float>,
&GLSLDecompiler::FClamp,
+ &GLSLDecompiler::FCastHalf0,
+ &GLSLDecompiler::FCastHalf1,
&GLSLDecompiler::Min<Type::Float>,
&GLSLDecompiler::Max<Type::Float>,
&GLSLDecompiler::FCos,
@@ -1792,6 +1851,7 @@ private:
&GLSLDecompiler::Absolute<Type::HalfFloat>,
&GLSLDecompiler::HNegate,
&GLSLDecompiler::HClamp,
+ &GLSLDecompiler::HCastFloat,
&GLSLDecompiler::HUnpack,
&GLSLDecompiler::HMergeF32,
&GLSLDecompiler::HMergeH0,
@@ -1867,6 +1927,11 @@ private:
&GLSLDecompiler::WorkGroupId<0>,
&GLSLDecompiler::WorkGroupId<1>,
&GLSLDecompiler::WorkGroupId<2>,
+
+ &GLSLDecompiler::BallotThread,
+ &GLSLDecompiler::VoteAll,
+ &GLSLDecompiler::VoteAny,
+ &GLSLDecompiler::VoteEqual,
};
static_assert(operation_decompilers.size() == static_cast<std::size_t>(OperationCode::Amount));
diff --git a/src/video_core/renderer_opengl/gl_texture_cache.cpp b/src/video_core/renderer_opengl/gl_texture_cache.cpp
index 408332f90..4f135fe03 100644
--- a/src/video_core/renderer_opengl/gl_texture_cache.cpp
+++ b/src/video_core/renderer_opengl/gl_texture_cache.cpp
@@ -184,6 +184,9 @@ GLint GetSwizzleSource(SwizzleSource source) {
}
void ApplyTextureDefaults(const SurfaceParams& params, GLuint texture) {
+ if (params.IsBuffer()) {
+ return;
+ }
glTextureParameteri(texture, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
glTextureParameteri(texture, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
glTextureParameteri(texture, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
@@ -208,6 +211,7 @@ OGLTexture CreateTexture(const SurfaceParams& params, GLenum target, GLenum inte
glNamedBufferStorage(texture_buffer.handle, params.width * params.GetBytesPerPixel(),
nullptr, GL_DYNAMIC_STORAGE_BIT);
glTextureBuffer(texture.handle, internal_format, texture_buffer.handle);
+ break;
case SurfaceTarget::Texture2D:
case SurfaceTarget::TextureCubemap:
glTextureStorage2D(texture.handle, params.emulated_levels, internal_format, params.width,
diff --git a/src/video_core/renderer_opengl/gl_texture_cache.h b/src/video_core/renderer_opengl/gl_texture_cache.h
index ff6ab6988..21324488a 100644
--- a/src/video_core/renderer_opengl/gl_texture_cache.h
+++ b/src/video_core/renderer_opengl/gl_texture_cache.h
@@ -51,7 +51,7 @@ public:
}
protected:
- void DecorateSurfaceName();
+ void DecorateSurfaceName() override;
View CreateView(const ViewParams& view_key) override;
View CreateViewInner(const ViewParams& view_key, bool is_proxy);
generated by cgit v1.2.3 (git 2.25.1) at 2025-11-09 16:57:29 +0000