1
0
Fork 0
forked from suyu/suyu

gl_buffer_cache: Drop interop based parameter buffer workarounds

Sacrify runtime performance to avoid generating kernel exceptions on
Windows due to our abusive aliasing of interop buffer objects.
This commit is contained in:
ReinUsesLisp 2021-01-18 19:00:00 -03:00
parent 2b95c137ff
commit 3da87d3f12
3 changed files with 45 additions and 65 deletions

View file

@ -91,7 +91,7 @@ class BufferCache {
}; };
public: public:
static constexpr size_t SKIP_CACHE_SIZE = 4096; static constexpr u32 SKIP_CACHE_SIZE = 4096;
explicit BufferCache(VideoCore::RasterizerInterface& rasterizer_, explicit BufferCache(VideoCore::RasterizerInterface& rasterizer_,
Tegra::Engines::Maxwell3D& maxwell3d_, Tegra::Engines::Maxwell3D& maxwell3d_,
@ -671,7 +671,7 @@ void BufferCache<P>::BindHostGraphicsUniformBuffer(size_t stage, u32 index, u32
const VAddr cpu_addr = binding.cpu_addr; const VAddr cpu_addr = binding.cpu_addr;
const u32 size = binding.size; const u32 size = binding.size;
Buffer& buffer = slot_buffers[binding.buffer_id]; Buffer& buffer = slot_buffers[binding.buffer_id];
if (size <= runtime.SkipCacheSize() && !buffer.IsRegionGpuModified(cpu_addr, size)) { if (size <= SKIP_CACHE_SIZE && !buffer.IsRegionGpuModified(cpu_addr, size)) {
if constexpr (IS_OPENGL) { if constexpr (IS_OPENGL) {
if (runtime.HasFastBufferSubData()) { if (runtime.HasFastBufferSubData()) {
// Fast path for Nvidia // Fast path for Nvidia

View file

@ -36,13 +36,8 @@ Buffer::Buffer(BufferCacheRuntime& runtime, VideoCore::RasterizerInterface& rast
buffer.Create(); buffer.Create();
const std::string name = fmt::format("Buffer 0x{:x}", CpuAddr()); const std::string name = fmt::format("Buffer 0x{:x}", CpuAddr());
glObjectLabel(GL_BUFFER, buffer.handle, static_cast<GLsizei>(name.size()), name.data()); glObjectLabel(GL_BUFFER, buffer.handle, static_cast<GLsizei>(name.size()), name.data());
if (runtime.device.UseAssemblyShaders()) { glNamedBufferData(buffer.handle, SizeBytes(), nullptr, GL_DYNAMIC_DRAW);
CreateMemoryObjects(runtime);
glNamedBufferStorageMemEXT(buffer.handle, SizeBytes(), memory_commit.ExportOpenGLHandle(),
memory_commit.Offset());
} else {
glNamedBufferData(buffer.handle, SizeBytes(), nullptr, GL_DYNAMIC_DRAW);
}
if (runtime.has_unified_vertex_buffers) { if (runtime.has_unified_vertex_buffers) {
glGetNamedBufferParameterui64vNV(buffer.handle, GL_BUFFER_GPU_ADDRESS_NV, &address); glGetNamedBufferParameterui64vNV(buffer.handle, GL_BUFFER_GPU_ADDRESS_NV, &address);
} }
@ -71,61 +66,33 @@ void Buffer::MakeResident(GLenum access) noexcept {
glMakeNamedBufferResidentNV(buffer.handle, access); glMakeNamedBufferResidentNV(buffer.handle, access);
} }
GLuint Buffer::SubBuffer(u32 offset) {
if (offset == 0) {
return buffer.handle;
}
for (const auto& [sub_buffer, sub_offset] : subs) {
if (sub_offset == offset) {
return sub_buffer.handle;
}
}
OGLBuffer sub_buffer;
sub_buffer.Create();
glNamedBufferStorageMemEXT(sub_buffer.handle, SizeBytes() - offset,
memory_commit.ExportOpenGLHandle(), memory_commit.Offset() + offset);
return subs.emplace_back(std::move(sub_buffer), offset).first.handle;
}
void Buffer::CreateMemoryObjects(BufferCacheRuntime& runtime) {
auto& allocator = runtime.vulkan_memory_allocator;
auto& device = runtime.vulkan_device->GetLogical();
auto vulkan_buffer = device.CreateBuffer(VkBufferCreateInfo{
.sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
.pNext = nullptr,
.flags = 0,
.size = SizeBytes(),
.usage = VK_BUFFER_USAGE_TRANSFER_SRC_BIT | VK_BUFFER_USAGE_TRANSFER_DST_BIT |
VK_BUFFER_USAGE_UNIFORM_TEXEL_BUFFER_BIT |
VK_BUFFER_USAGE_STORAGE_TEXEL_BUFFER_BIT | VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT |
VK_BUFFER_USAGE_STORAGE_BUFFER_BIT | VK_BUFFER_USAGE_INDEX_BUFFER_BIT |
VK_BUFFER_USAGE_VERTEX_BUFFER_BIT,
.sharingMode = VK_SHARING_MODE_EXCLUSIVE,
.queueFamilyIndexCount = 0,
.pQueueFamilyIndices = nullptr,
});
const VkMemoryRequirements requirements = device.GetBufferMemoryRequirements(*vulkan_buffer);
memory_commit = allocator->Commit(requirements, Vulkan::MemoryUsage::DeviceLocal);
}
BufferCacheRuntime::BufferCacheRuntime(const Device& device_, const Vulkan::Device* vulkan_device_, BufferCacheRuntime::BufferCacheRuntime(const Device& device_, const Vulkan::Device* vulkan_device_,
Vulkan::MemoryAllocator* vulkan_memory_allocator_) Vulkan::MemoryAllocator* vulkan_memory_allocator_)
: device{device_}, vulkan_device{vulkan_device_}, : device{device_}, vulkan_device{vulkan_device_},
vulkan_memory_allocator{vulkan_memory_allocator_}, vulkan_memory_allocator{vulkan_memory_allocator_},
stream_buffer{device.HasFastBufferSubData() ? std::nullopt has_fast_buffer_sub_data{device.HasFastBufferSubData()},
: std::make_optional<StreamBuffer>()} { use_assembly_shaders{device.UseAssemblyShaders()},
has_unified_vertex_buffers{device.HasVertexBufferUnifiedMemory()},
stream_buffer{has_fast_buffer_sub_data ? std::nullopt : std::make_optional<StreamBuffer>()} {
GLint gl_max_attributes; GLint gl_max_attributes;
glGetIntegerv(GL_MAX_VERTEX_ATTRIBS, &gl_max_attributes); glGetIntegerv(GL_MAX_VERTEX_ATTRIBS, &gl_max_attributes);
max_attributes = static_cast<u32>(gl_max_attributes); max_attributes = static_cast<u32>(gl_max_attributes);
use_assembly_shaders = device.UseAssemblyShaders();
has_unified_vertex_buffers = device.HasVertexBufferUnifiedMemory();
for (auto& stage_uniforms : fast_uniforms) { for (auto& stage_uniforms : fast_uniforms) {
for (OGLBuffer& buffer : stage_uniforms) { for (OGLBuffer& buffer : stage_uniforms) {
buffer.Create(); buffer.Create();
glNamedBufferData(buffer.handle, BufferCache::SKIP_CACHE_SIZE, nullptr, GL_STREAM_DRAW); glNamedBufferData(buffer.handle, BufferCache::SKIP_CACHE_SIZE, nullptr, GL_STREAM_DRAW);
} }
} }
for (auto& stage_uniforms : copy_uniforms) {
for (OGLBuffer& buffer : stage_uniforms) {
buffer.Create();
glNamedBufferData(buffer.handle, 0x10'000, nullptr, GL_STREAM_COPY);
}
}
for (OGLBuffer& buffer : copy_compute_uniforms) {
buffer.Create();
glNamedBufferData(buffer.handle, 0x10'000, nullptr, GL_STREAM_COPY);
}
} }
void BufferCacheRuntime::CopyBuffer(Buffer& dst_buffer, Buffer& src_buffer, void BufferCacheRuntime::CopyBuffer(Buffer& dst_buffer, Buffer& src_buffer,
@ -167,8 +134,14 @@ void BufferCacheRuntime::BindVertexBuffer(u32 index, Buffer& buffer, u32 offset,
void BufferCacheRuntime::BindUniformBuffer(size_t stage, u32 binding_index, Buffer& buffer, void BufferCacheRuntime::BindUniformBuffer(size_t stage, u32 binding_index, Buffer& buffer,
u32 offset, u32 size) { u32 offset, u32 size) {
if (use_assembly_shaders) { if (use_assembly_shaders) {
const GLuint sub_buffer = buffer.SubBuffer(offset); GLuint handle;
glBindBufferRangeNV(PABO_LUT[stage], binding_index, sub_buffer, 0, if (offset != 0) {
handle = copy_uniforms[stage][binding_index].handle;
glCopyNamedBufferSubData(buffer.Handle(), handle, offset, 0, size);
} else {
handle = buffer.Handle();
}
glBindBufferRangeNV(PABO_LUT[stage], binding_index, handle, 0,
static_cast<GLsizeiptr>(size)); static_cast<GLsizeiptr>(size));
} else { } else {
const GLuint base_binding = device.GetBaseBindings(stage).uniform_buffer; const GLuint base_binding = device.GetBaseBindings(stage).uniform_buffer;
@ -181,8 +154,15 @@ void BufferCacheRuntime::BindUniformBuffer(size_t stage, u32 binding_index, Buff
void BufferCacheRuntime::BindComputeUniformBuffer(u32 binding_index, Buffer& buffer, u32 offset, void BufferCacheRuntime::BindComputeUniformBuffer(u32 binding_index, Buffer& buffer, u32 offset,
u32 size) { u32 size) {
if (use_assembly_shaders) { if (use_assembly_shaders) {
glBindBufferRangeNV(GL_COMPUTE_PROGRAM_PARAMETER_BUFFER_NV, binding_index, GLuint handle;
buffer.SubBuffer(offset), 0, static_cast<GLsizeiptr>(size)); if (offset != 0) {
handle = copy_compute_uniforms[binding_index].handle;
glCopyNamedBufferSubData(buffer.Handle(), handle, offset, 0, size);
} else {
handle = buffer.Handle();
}
glBindBufferRangeNV(GL_COMPUTE_PROGRAM_PARAMETER_BUFFER_NV, binding_index, handle, 0,
static_cast<GLsizeiptr>(size));
} else { } else {
glBindBufferRange(GL_UNIFORM_BUFFER, binding_index, buffer.Handle(), glBindBufferRange(GL_UNIFORM_BUFFER, binding_index, buffer.Handle(),
static_cast<GLintptr>(offset), static_cast<GLsizeiptr>(size)); static_cast<GLintptr>(offset), static_cast<GLsizeiptr>(size));

View file

@ -39,8 +39,6 @@ public:
void MakeResident(GLenum access) noexcept; void MakeResident(GLenum access) noexcept;
[[nodiscard]] GLuint SubBuffer(u32 offset);
[[nodiscard]] GLuint64EXT HostGpuAddr() const noexcept { [[nodiscard]] GLuint64EXT HostGpuAddr() const noexcept {
return address; return address;
} }
@ -50,13 +48,9 @@ public:
} }
private: private:
void CreateMemoryObjects(BufferCacheRuntime& runtime);
GLuint64EXT address = 0; GLuint64EXT address = 0;
Vulkan::MemoryCommit memory_commit;
OGLBuffer buffer; OGLBuffer buffer;
GLenum current_residency_access = GL_NONE; GLenum current_residency_access = GL_NONE;
std::vector<std::pair<OGLBuffer, u32>> subs;
}; };
class BufferCacheRuntime { class BufferCacheRuntime {
@ -127,7 +121,7 @@ public:
} }
[[nodiscard]] bool HasFastBufferSubData() const noexcept { [[nodiscard]] bool HasFastBufferSubData() const noexcept {
return device.HasFastBufferSubData(); return has_fast_buffer_sub_data;
} }
private: private:
@ -140,16 +134,22 @@ private:
const Device& device; const Device& device;
const Vulkan::Device* vulkan_device; const Vulkan::Device* vulkan_device;
Vulkan::MemoryAllocator* vulkan_memory_allocator; Vulkan::MemoryAllocator* vulkan_memory_allocator;
std::optional<StreamBuffer> stream_buffer;
bool has_fast_buffer_sub_data = false;
bool use_assembly_shaders = false;
bool has_unified_vertex_buffers = false;
u32 max_attributes = 0; u32 max_attributes = 0;
bool use_assembly_shaders = false; std::optional<StreamBuffer> stream_buffer;
bool has_unified_vertex_buffers = false;
std::array<std::array<OGLBuffer, VideoCommon::NUM_GRAPHICS_UNIFORM_BUFFERS>, std::array<std::array<OGLBuffer, VideoCommon::NUM_GRAPHICS_UNIFORM_BUFFERS>,
VideoCommon::NUM_STAGES> VideoCommon::NUM_STAGES>
fast_uniforms; fast_uniforms;
std::array<std::array<OGLBuffer, VideoCommon::NUM_GRAPHICS_UNIFORM_BUFFERS>,
VideoCommon::NUM_STAGES>
copy_uniforms;
std::array<OGLBuffer, VideoCommon::NUM_COMPUTE_UNIFORM_BUFFERS> copy_compute_uniforms;
u32 index_buffer_offset = 0; u32 index_buffer_offset = 0;
}; };