Merge pull request #2256 from bunnei/gpu-vmm

gpu: Rewrite MemoryManager based on the VMManager implementation.
This commit is contained in:
bunnei 2019-03-22 18:41:12 -04:00 committed by GitHub
commit e5893db3e6
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
28 changed files with 561 additions and 268 deletions

View file

@ -40,10 +40,9 @@ using s64 = std::int64_t; ///< 64-bit signed int
using f32 = float; ///< 32-bit floating point using f32 = float; ///< 32-bit floating point
using f64 = double; ///< 64-bit floating point using f64 = double; ///< 64-bit floating point
// TODO: It would be nice to eventually replace these with strong types that prevent accidental
// conversion between each other.
using VAddr = u64; ///< Represents a pointer in the userspace virtual address space. using VAddr = u64; ///< Represents a pointer in the userspace virtual address space.
using PAddr = u64; ///< Represents a pointer in the ARM11 physical address space. using PAddr = u64; ///< Represents a pointer in the ARM11 physical address space.
using GPUVAddr = u64; ///< Represents a pointer in the GPU virtual address space.
using u128 = std::array<std::uint64_t, 2>; using u128 = std::array<std::uint64_t, 2>;
static_assert(sizeof(u128) == 16, "u128 must be 128 bits wide"); static_assert(sizeof(u128) == 16, "u128 must be 128 bits wide");

View file

@ -16,6 +16,7 @@ void PageTable::Resize(std::size_t address_space_width_in_bits) {
pointers.resize(num_page_table_entries); pointers.resize(num_page_table_entries);
attributes.resize(num_page_table_entries); attributes.resize(num_page_table_entries);
backing_addr.resize(num_page_table_entries);
// The default is a 39-bit address space, which causes an initial 1GB allocation size. If the // The default is a 39-bit address space, which causes an initial 1GB allocation size. If the
// vector size is subsequently decreased (via resize), the vector might not automatically // vector size is subsequently decreased (via resize), the vector might not automatically
@ -24,6 +25,7 @@ void PageTable::Resize(std::size_t address_space_width_in_bits) {
pointers.shrink_to_fit(); pointers.shrink_to_fit();
attributes.shrink_to_fit(); attributes.shrink_to_fit();
backing_addr.shrink_to_fit();
} }
} // namespace Common } // namespace Common

View file

@ -21,6 +21,8 @@ enum class PageType : u8 {
RasterizerCachedMemory, RasterizerCachedMemory,
/// Page is mapped to a I/O region. Writing and reading to this page is handled by functions. /// Page is mapped to a I/O region. Writing and reading to this page is handled by functions.
Special, Special,
/// Page is allocated for use.
Allocated,
}; };
struct SpecialRegion { struct SpecialRegion {
@ -66,7 +68,7 @@ struct PageTable {
* Contains MMIO handlers that back memory regions whose entries in the `attribute` vector is * Contains MMIO handlers that back memory regions whose entries in the `attribute` vector is
* of type `Special`. * of type `Special`.
*/ */
boost::icl::interval_map<VAddr, std::set<SpecialRegion>> special_regions; boost::icl::interval_map<u64, std::set<SpecialRegion>> special_regions;
/** /**
* Vector of fine grained page attributes. If it is set to any value other than `Memory`, then * Vector of fine grained page attributes. If it is set to any value other than `Memory`, then
@ -74,6 +76,8 @@ struct PageTable {
*/ */
std::vector<PageType> attributes; std::vector<PageType> attributes;
std::vector<u64> backing_addr;
const std::size_t page_size_in_bits{}; const std::size_t page_size_in_bits{};
}; };

View file

@ -89,7 +89,7 @@ u32 nvhost_as_gpu::Remap(const std::vector<u8>& input, std::vector<u8>& output)
for (const auto& entry : entries) { for (const auto& entry : entries) {
LOG_WARNING(Service_NVDRV, "remap entry, offset=0x{:X} handle=0x{:X} pages=0x{:X}", LOG_WARNING(Service_NVDRV, "remap entry, offset=0x{:X} handle=0x{:X} pages=0x{:X}",
entry.offset, entry.nvmap_handle, entry.pages); entry.offset, entry.nvmap_handle, entry.pages);
Tegra::GPUVAddr offset = static_cast<Tegra::GPUVAddr>(entry.offset) << 0x10; GPUVAddr offset = static_cast<GPUVAddr>(entry.offset) << 0x10;
auto object = nvmap_dev->GetObject(entry.nvmap_handle); auto object = nvmap_dev->GetObject(entry.nvmap_handle);
if (!object) { if (!object) {
LOG_CRITICAL(Service_NVDRV, "nvmap {} is an invalid handle!", entry.nvmap_handle); LOG_CRITICAL(Service_NVDRV, "nvmap {} is an invalid handle!", entry.nvmap_handle);
@ -102,7 +102,7 @@ u32 nvhost_as_gpu::Remap(const std::vector<u8>& input, std::vector<u8>& output)
u64 size = static_cast<u64>(entry.pages) << 0x10; u64 size = static_cast<u64>(entry.pages) << 0x10;
ASSERT(size <= object->size); ASSERT(size <= object->size);
Tegra::GPUVAddr returned = gpu.MemoryManager().MapBufferEx(object->addr, offset, size); GPUVAddr returned = gpu.MemoryManager().MapBufferEx(object->addr, offset, size);
ASSERT(returned == offset); ASSERT(returned == offset);
} }
std::memcpy(output.data(), entries.data(), output.size()); std::memcpy(output.data(), entries.data(), output.size());
@ -173,16 +173,8 @@ u32 nvhost_as_gpu::UnmapBuffer(const std::vector<u8>& input, std::vector<u8>& ou
return 0; return 0;
} }
auto& system_instance = Core::System::GetInstance(); params.offset = Core::System::GetInstance().GPU().MemoryManager().UnmapBuffer(params.offset,
itr->second.size);
// Remove this memory region from the rasterizer cache.
auto& gpu = system_instance.GPU();
auto cpu_addr = gpu.MemoryManager().GpuToCpuAddress(params.offset);
ASSERT(cpu_addr);
gpu.FlushAndInvalidateRegion(ToCacheAddr(Memory::GetPointer(*cpu_addr)), itr->second.size);
params.offset = gpu.MemoryManager().UnmapBuffer(params.offset, itr->second.size);
buffer_mappings.erase(itr->second.offset); buffer_mappings.erase(itr->second.offset);
std::memcpy(output.data(), &params, output.size()); std::memcpy(output.data(), &params, output.size());

View file

@ -48,7 +48,7 @@ static void MapPages(Common::PageTable& page_table, VAddr base, u64 size, u8* me
(base + size) * PAGE_SIZE); (base + size) * PAGE_SIZE);
// During boot, current_page_table might not be set yet, in which case we need not flush // During boot, current_page_table might not be set yet, in which case we need not flush
if (current_page_table) { if (Core::System::GetInstance().IsPoweredOn()) {
Core::System::GetInstance().GPU().FlushAndInvalidateRegion(base << PAGE_BITS, Core::System::GetInstance().GPU().FlushAndInvalidateRegion(base << PAGE_BITS,
size * PAGE_SIZE); size * PAGE_SIZE);
} }

View file

@ -9,7 +9,6 @@
#include "common/bit_field.h" #include "common/bit_field.h"
#include "common/common_types.h" #include "common/common_types.h"
#include "video_core/memory_manager.h"
namespace Tegra { namespace Tegra {

View file

@ -46,7 +46,7 @@ void KeplerMemory::ProcessData(u32 data) {
// contain a dirty surface that will have to be written back to memory. // contain a dirty surface that will have to be written back to memory.
const GPUVAddr address{regs.dest.Address() + state.write_offset * sizeof(u32)}; const GPUVAddr address{regs.dest.Address() + state.write_offset * sizeof(u32)};
rasterizer.InvalidateRegion(ToCacheAddr(memory_manager.GetPointer(address)), sizeof(u32)); rasterizer.InvalidateRegion(ToCacheAddr(memory_manager.GetPointer(address)), sizeof(u32));
memory_manager.Write32(address, data); memory_manager.Write<u32>(address, data);
system.GPU().Maxwell3D().dirty_flags.OnMemoryWrite(); system.GPU().Maxwell3D().dirty_flags.OnMemoryWrite();

View file

@ -307,7 +307,7 @@ void Maxwell3D::ProcessQueryGet() {
// Write the current query sequence to the sequence address. // Write the current query sequence to the sequence address.
// TODO(Subv): Find out what happens if you use a long query type but mark it as a short // TODO(Subv): Find out what happens if you use a long query type but mark it as a short
// query. // query.
memory_manager.Write32(sequence_address, sequence); memory_manager.Write<u32>(sequence_address, sequence);
} else { } else {
// Write the 128-bit result structure in long mode. Note: We emulate an infinitely fast // Write the 128-bit result structure in long mode. Note: We emulate an infinitely fast
// GPU, this command may actually take a while to complete in real hardware due to GPU // GPU, this command may actually take a while to complete in real hardware due to GPU
@ -395,7 +395,7 @@ void Maxwell3D::ProcessCBData(u32 value) {
u8* ptr{memory_manager.GetPointer(address)}; u8* ptr{memory_manager.GetPointer(address)};
rasterizer.InvalidateRegion(ToCacheAddr(ptr), sizeof(u32)); rasterizer.InvalidateRegion(ToCacheAddr(ptr), sizeof(u32));
memory_manager.Write32(address, value); memory_manager.Write<u32>(address, value);
dirty_flags.OnMemoryWrite(); dirty_flags.OnMemoryWrite();
@ -447,7 +447,7 @@ std::vector<Texture::FullTextureInfo> Maxwell3D::GetStageTextures(Regs::ShaderSt
for (GPUVAddr current_texture = tex_info_buffer.address + TextureInfoOffset; for (GPUVAddr current_texture = tex_info_buffer.address + TextureInfoOffset;
current_texture < tex_info_buffer_end; current_texture += sizeof(Texture::TextureHandle)) { current_texture < tex_info_buffer_end; current_texture += sizeof(Texture::TextureHandle)) {
const Texture::TextureHandle tex_handle{memory_manager.Read32(current_texture)}; const Texture::TextureHandle tex_handle{memory_manager.Read<u32>(current_texture)};
Texture::FullTextureInfo tex_info{}; Texture::FullTextureInfo tex_info{};
// TODO(Subv): Use the shader to determine which textures are actually accessed. // TODO(Subv): Use the shader to determine which textures are actually accessed.
@ -482,7 +482,7 @@ Texture::FullTextureInfo Maxwell3D::GetStageTexture(Regs::ShaderStage stage,
ASSERT(tex_info_address < tex_info_buffer.address + tex_info_buffer.size); ASSERT(tex_info_address < tex_info_buffer.address + tex_info_buffer.size);
const Texture::TextureHandle tex_handle{memory_manager.Read32(tex_info_address)}; const Texture::TextureHandle tex_handle{memory_manager.Read<u32>(tex_info_address)};
Texture::FullTextureInfo tex_info{}; Texture::FullTextureInfo tex_info{};
tex_info.index = static_cast<u32>(offset); tex_info.index = static_cast<u32>(offset);

View file

@ -88,6 +88,16 @@ void MaxwellDMA::HandleCopy() {
auto source_ptr{memory_manager.GetPointer(source)}; auto source_ptr{memory_manager.GetPointer(source)};
auto dst_ptr{memory_manager.GetPointer(dest)}; auto dst_ptr{memory_manager.GetPointer(dest)};
if (!source_ptr) {
LOG_ERROR(HW_GPU, "source_ptr is invalid");
return;
}
if (!dst_ptr) {
LOG_ERROR(HW_GPU, "dst_ptr is invalid");
return;
}
const auto FlushAndInvalidate = [&](u32 src_size, u64 dst_size) { const auto FlushAndInvalidate = [&](u32 src_size, u64 dst_size) {
// TODO(Subv): For now, manually flush the regions until we implement GPU-accelerated // TODO(Subv): For now, manually flush the regions until we implement GPU-accelerated
// copying. // copying.

View file

@ -12,6 +12,7 @@
#include "video_core/engines/maxwell_3d.h" #include "video_core/engines/maxwell_3d.h"
#include "video_core/engines/maxwell_dma.h" #include "video_core/engines/maxwell_dma.h"
#include "video_core/gpu.h" #include "video_core/gpu.h"
#include "video_core/memory_manager.h"
#include "video_core/renderer_base.h" #include "video_core/renderer_base.h"
namespace Tegra { namespace Tegra {
@ -287,7 +288,7 @@ void GPU::ProcessSemaphoreTriggerMethod() {
block.timestamp = Core::System::GetInstance().CoreTiming().GetTicks(); block.timestamp = Core::System::GetInstance().CoreTiming().GetTicks();
memory_manager->WriteBlock(regs.smaphore_address.SmaphoreAddress(), &block, sizeof(block)); memory_manager->WriteBlock(regs.smaphore_address.SmaphoreAddress(), &block, sizeof(block));
} else { } else {
const u32 word{memory_manager->Read32(regs.smaphore_address.SmaphoreAddress())}; const u32 word{memory_manager->Read<u32>(regs.smaphore_address.SmaphoreAddress())};
if ((op == GpuSemaphoreOperation::AcquireEqual && word == regs.semaphore_sequence) || if ((op == GpuSemaphoreOperation::AcquireEqual && word == regs.semaphore_sequence) ||
(op == GpuSemaphoreOperation::AcquireGequal && (op == GpuSemaphoreOperation::AcquireGequal &&
static_cast<s32>(word - regs.semaphore_sequence) > 0) || static_cast<s32>(word - regs.semaphore_sequence) > 0) ||
@ -314,11 +315,11 @@ void GPU::ProcessSemaphoreTriggerMethod() {
} }
void GPU::ProcessSemaphoreRelease() { void GPU::ProcessSemaphoreRelease() {
memory_manager->Write32(regs.smaphore_address.SmaphoreAddress(), regs.semaphore_release); memory_manager->Write<u32>(regs.smaphore_address.SmaphoreAddress(), regs.semaphore_release);
} }
void GPU::ProcessSemaphoreAcquire() { void GPU::ProcessSemaphoreAcquire() {
const u32 word = memory_manager->Read32(regs.smaphore_address.SmaphoreAddress()); const u32 word = memory_manager->Read<u32>(regs.smaphore_address.SmaphoreAddress());
const auto value = regs.semaphore_acquire; const auto value = regs.semaphore_acquire;
if (word != value) { if (word != value) {
regs.acquire_active = true; regs.acquire_active = true;

View file

@ -9,7 +9,6 @@
#include "common/common_types.h" #include "common/common_types.h"
#include "core/hle/service/nvflinger/buffer_queue.h" #include "core/hle/service/nvflinger/buffer_queue.h"
#include "video_core/dma_pusher.h" #include "video_core/dma_pusher.h"
#include "video_core/memory_manager.h"
using CacheAddr = std::uintptr_t; using CacheAddr = std::uintptr_t;
inline CacheAddr ToCacheAddr(const void* host_ptr) { inline CacheAddr ToCacheAddr(const void* host_ptr) {
@ -124,6 +123,8 @@ enum class EngineID {
MAXWELL_DMA_COPY_A = 0xB0B5, MAXWELL_DMA_COPY_A = 0xB0B5,
}; };
class MemoryManager;
class GPU { class GPU {
public: public:
explicit GPU(Core::System& system, VideoCore::RendererBase& renderer); explicit GPU(Core::System& system, VideoCore::RendererBase& renderer);
@ -244,9 +245,8 @@ protected:
private: private:
std::unique_ptr<Tegra::MemoryManager> memory_manager; std::unique_ptr<Tegra::MemoryManager> memory_manager;
/// Mapping of command subchannels to their bound engine ids. /// Mapping of command subchannels to their bound engine ids
std::array<EngineID, 8> bound_engines = {}; std::array<EngineID, 8> bound_engines = {};
/// 3D engine /// 3D engine
std::unique_ptr<Engines::Maxwell3D> maxwell_3d; std::unique_ptr<Engines::Maxwell3D> maxwell_3d;
/// 2D engine /// 2D engine

View file

@ -5,198 +5,187 @@
#include "common/alignment.h" #include "common/alignment.h"
#include "common/assert.h" #include "common/assert.h"
#include "common/logging/log.h" #include "common/logging/log.h"
#include "core/core.h"
#include "core/memory.h" #include "core/memory.h"
#include "video_core/gpu.h"
#include "video_core/memory_manager.h" #include "video_core/memory_manager.h"
#include "video_core/rasterizer_interface.h"
#include "video_core/renderer_base.h"
namespace Tegra { namespace Tegra {
MemoryManager::MemoryManager() { MemoryManager::MemoryManager() {
// Mark the first page as reserved, so that 0 is not a valid GPUVAddr. Otherwise, games might std::fill(page_table.pointers.begin(), page_table.pointers.end(), nullptr);
// try to use 0 as a valid address, which is also used to mean nullptr. This fixes a bug with std::fill(page_table.attributes.begin(), page_table.attributes.end(),
// Undertale using 0 for a render target. Common::PageType::Unmapped);
PageSlot(0) = static_cast<u64>(PageStatus::Reserved); page_table.Resize(address_space_width);
// Initialize the map with a single free region covering the entire managed space.
VirtualMemoryArea initial_vma;
initial_vma.size = address_space_end;
vma_map.emplace(initial_vma.base, initial_vma);
UpdatePageTableForVMA(initial_vma);
} }
GPUVAddr MemoryManager::AllocateSpace(u64 size, u64 align) { GPUVAddr MemoryManager::AllocateSpace(u64 size, u64 align) {
const std::optional<GPUVAddr> gpu_addr{FindFreeBlock(0, size, align, PageStatus::Unmapped)}; const u64 aligned_size{Common::AlignUp(size, page_size)};
const GPUVAddr gpu_addr{FindFreeRegion(address_space_base, aligned_size)};
ASSERT_MSG(gpu_addr, "unable to find available GPU memory"); AllocateMemory(gpu_addr, 0, aligned_size);
for (u64 offset{}; offset < size; offset += PAGE_SIZE) { return gpu_addr;
VAddr& slot{PageSlot(*gpu_addr + offset)};
ASSERT(slot == static_cast<u64>(PageStatus::Unmapped));
slot = static_cast<u64>(PageStatus::Allocated);
}
return *gpu_addr;
} }
GPUVAddr MemoryManager::AllocateSpace(GPUVAddr gpu_addr, u64 size, u64 align) { GPUVAddr MemoryManager::AllocateSpace(GPUVAddr gpu_addr, u64 size, u64 align) {
for (u64 offset{}; offset < size; offset += PAGE_SIZE) { const u64 aligned_size{Common::AlignUp(size, page_size)};
VAddr& slot{PageSlot(gpu_addr + offset)};
ASSERT(slot == static_cast<u64>(PageStatus::Unmapped)); AllocateMemory(gpu_addr, 0, aligned_size);
slot = static_cast<u64>(PageStatus::Allocated);
}
return gpu_addr; return gpu_addr;
} }
GPUVAddr MemoryManager::MapBufferEx(VAddr cpu_addr, u64 size) { GPUVAddr MemoryManager::MapBufferEx(VAddr cpu_addr, u64 size) {
const std::optional<GPUVAddr> gpu_addr{FindFreeBlock(0, size, PAGE_SIZE, PageStatus::Unmapped)}; const u64 aligned_size{Common::AlignUp(size, page_size)};
const GPUVAddr gpu_addr{FindFreeRegion(address_space_base, aligned_size)};
ASSERT_MSG(gpu_addr, "unable to find available GPU memory"); MapBackingMemory(gpu_addr, Memory::GetPointer(cpu_addr), aligned_size, cpu_addr);
for (u64 offset{}; offset < size; offset += PAGE_SIZE) { return gpu_addr;
VAddr& slot{PageSlot(*gpu_addr + offset)};
ASSERT(slot == static_cast<u64>(PageStatus::Unmapped));
slot = cpu_addr + offset;
}
const MappedRegion region{cpu_addr, *gpu_addr, size};
mapped_regions.push_back(region);
return *gpu_addr;
} }
GPUVAddr MemoryManager::MapBufferEx(VAddr cpu_addr, GPUVAddr gpu_addr, u64 size) { GPUVAddr MemoryManager::MapBufferEx(VAddr cpu_addr, GPUVAddr gpu_addr, u64 size) {
ASSERT((gpu_addr & PAGE_MASK) == 0); ASSERT((gpu_addr & page_mask) == 0);
if (PageSlot(gpu_addr) != static_cast<u64>(PageStatus::Allocated)) { const u64 aligned_size{Common::AlignUp(size, page_size)};
// Page has been already mapped. In this case, we must find a new area of memory to use that
// is different than the specified one. Super Mario Odyssey hits this scenario when changing
// areas, but we do not want to overwrite the old pages.
// TODO(bunnei): We need to write a hardware test to confirm this behavior.
LOG_ERROR(HW_GPU, "attempting to map addr 0x{:016X}, which is not available!", gpu_addr); MapBackingMemory(gpu_addr, Memory::GetPointer(cpu_addr), aligned_size, cpu_addr);
const std::optional<GPUVAddr> new_gpu_addr{
FindFreeBlock(gpu_addr, size, PAGE_SIZE, PageStatus::Allocated)};
ASSERT_MSG(new_gpu_addr, "unable to find available GPU memory");
gpu_addr = *new_gpu_addr;
}
for (u64 offset{}; offset < size; offset += PAGE_SIZE) {
VAddr& slot{PageSlot(gpu_addr + offset)};
ASSERT(slot == static_cast<u64>(PageStatus::Allocated));
slot = cpu_addr + offset;
}
const MappedRegion region{cpu_addr, gpu_addr, size};
mapped_regions.push_back(region);
return gpu_addr; return gpu_addr;
} }
GPUVAddr MemoryManager::UnmapBuffer(GPUVAddr gpu_addr, u64 size) { GPUVAddr MemoryManager::UnmapBuffer(GPUVAddr gpu_addr, u64 size) {
ASSERT((gpu_addr & PAGE_MASK) == 0); ASSERT((gpu_addr & page_mask) == 0);
for (u64 offset{}; offset < size; offset += PAGE_SIZE) { const u64 aligned_size{Common::AlignUp(size, page_size)};
VAddr& slot{PageSlot(gpu_addr + offset)}; const CacheAddr cache_addr{ToCacheAddr(GetPointer(gpu_addr))};
ASSERT(slot != static_cast<u64>(PageStatus::Allocated) && Core::System::GetInstance().Renderer().Rasterizer().FlushAndInvalidateRegion(cache_addr,
slot != static_cast<u64>(PageStatus::Unmapped)); aligned_size);
UnmapRange(gpu_addr, aligned_size);
slot = static_cast<u64>(PageStatus::Unmapped);
}
// Delete the region mappings that are contained within the unmapped region
mapped_regions.erase(std::remove_if(mapped_regions.begin(), mapped_regions.end(),
[&](const MappedRegion& region) {
return region.gpu_addr <= gpu_addr &&
region.gpu_addr + region.size < gpu_addr + size;
}),
mapped_regions.end());
return gpu_addr; return gpu_addr;
} }
GPUVAddr MemoryManager::GetRegionEnd(GPUVAddr region_start) const { GPUVAddr MemoryManager::FindFreeRegion(GPUVAddr region_start, u64 size) {
for (const auto& region : mapped_regions) { // Find the first Free VMA.
const GPUVAddr region_end{region.gpu_addr + region.size}; const VMAHandle vma_handle{std::find_if(vma_map.begin(), vma_map.end(), [&](const auto& vma) {
if (region_start >= region.gpu_addr && region_start < region_end) { if (vma.second.type != VirtualMemoryArea::Type::Unmapped) {
return region_end; return false;
}
} }
const VAddr vma_end{vma.second.base + vma.second.size};
return vma_end > region_start && vma_end >= region_start + size;
})};
if (vma_handle == vma_map.end()) {
return {}; return {};
}
return std::max(region_start, vma_handle->second.base);
} }
std::optional<GPUVAddr> MemoryManager::FindFreeBlock(GPUVAddr region_start, u64 size, u64 align, bool MemoryManager::IsAddressValid(GPUVAddr addr) const {
PageStatus status) { return (addr >> page_bits) < page_table.pointers.size();
GPUVAddr gpu_addr{region_start}; }
u64 free_space{};
align = (align + PAGE_MASK) & ~PAGE_MASK;
while (gpu_addr + free_space < MAX_ADDRESS) { std::optional<VAddr> MemoryManager::GpuToCpuAddress(GPUVAddr addr) {
if (PageSlot(gpu_addr + free_space) == static_cast<u64>(status)) { if (!IsAddressValid(addr)) {
free_space += PAGE_SIZE; return {};
if (free_space >= size) {
return gpu_addr;
}
} else {
gpu_addr += free_space + PAGE_SIZE;
free_space = 0;
gpu_addr = Common::AlignUp(gpu_addr, align);
} }
VAddr cpu_addr{page_table.backing_addr[addr >> page_bits]};
if (cpu_addr) {
return cpu_addr + (addr & page_mask);
} }
return {}; return {};
} }
std::optional<VAddr> MemoryManager::GpuToCpuAddress(GPUVAddr gpu_addr) { template <typename T>
const VAddr base_addr{PageSlot(gpu_addr)}; T MemoryManager::Read(GPUVAddr addr) {
if (!IsAddressValid(addr)) {
if (base_addr == static_cast<u64>(PageStatus::Allocated) ||
base_addr == static_cast<u64>(PageStatus::Unmapped) ||
base_addr == static_cast<u64>(PageStatus::Reserved)) {
return {}; return {};
} }
return base_addr + (gpu_addr & PAGE_MASK); const u8* page_pointer{page_table.pointers[addr >> page_bits]};
if (page_pointer) {
// NOTE: Avoid adding any extra logic to this fast-path block
T value;
std::memcpy(&value, &page_pointer[addr & page_mask], sizeof(T));
return value;
}
switch (page_table.attributes[addr >> page_bits]) {
case Common::PageType::Unmapped:
LOG_ERROR(HW_GPU, "Unmapped Read{} @ 0x{:08X}", sizeof(T) * 8, addr);
return 0;
case Common::PageType::Memory:
ASSERT_MSG(false, "Mapped memory page without a pointer @ {:016X}", addr);
break;
default:
UNREACHABLE();
}
return {};
} }
u8 MemoryManager::Read8(GPUVAddr addr) { template <typename T>
return Memory::Read8(*GpuToCpuAddress(addr)); void MemoryManager::Write(GPUVAddr addr, T data) {
if (!IsAddressValid(addr)) {
return;
}
u8* page_pointer{page_table.pointers[addr >> page_bits]};
if (page_pointer) {
// NOTE: Avoid adding any extra logic to this fast-path block
std::memcpy(&page_pointer[addr & page_mask], &data, sizeof(T));
return;
}
switch (page_table.attributes[addr >> page_bits]) {
case Common::PageType::Unmapped:
LOG_ERROR(HW_GPU, "Unmapped Write{} 0x{:08X} @ 0x{:016X}", sizeof(data) * 8,
static_cast<u32>(data), addr);
return;
case Common::PageType::Memory:
ASSERT_MSG(false, "Mapped memory page without a pointer @ {:016X}", addr);
break;
default:
UNREACHABLE();
}
} }
u16 MemoryManager::Read16(GPUVAddr addr) { template u8 MemoryManager::Read<u8>(GPUVAddr addr);
return Memory::Read16(*GpuToCpuAddress(addr)); template u16 MemoryManager::Read<u16>(GPUVAddr addr);
} template u32 MemoryManager::Read<u32>(GPUVAddr addr);
template u64 MemoryManager::Read<u64>(GPUVAddr addr);
u32 MemoryManager::Read32(GPUVAddr addr) { template void MemoryManager::Write<u8>(GPUVAddr addr, u8 data);
return Memory::Read32(*GpuToCpuAddress(addr)); template void MemoryManager::Write<u16>(GPUVAddr addr, u16 data);
} template void MemoryManager::Write<u32>(GPUVAddr addr, u32 data);
template void MemoryManager::Write<u64>(GPUVAddr addr, u64 data);
u64 MemoryManager::Read64(GPUVAddr addr) {
return Memory::Read64(*GpuToCpuAddress(addr));
}
void MemoryManager::Write8(GPUVAddr addr, u8 data) {
Memory::Write8(*GpuToCpuAddress(addr), data);
}
void MemoryManager::Write16(GPUVAddr addr, u16 data) {
Memory::Write16(*GpuToCpuAddress(addr), data);
}
void MemoryManager::Write32(GPUVAddr addr, u32 data) {
Memory::Write32(*GpuToCpuAddress(addr), data);
}
void MemoryManager::Write64(GPUVAddr addr, u64 data) {
Memory::Write64(*GpuToCpuAddress(addr), data);
}
u8* MemoryManager::GetPointer(GPUVAddr addr) { u8* MemoryManager::GetPointer(GPUVAddr addr) {
return Memory::GetPointer(*GpuToCpuAddress(addr)); if (!IsAddressValid(addr)) {
return {};
}
u8* page_pointer{page_table.pointers[addr >> page_bits]};
if (page_pointer) {
return page_pointer + (addr & page_mask);
}
LOG_ERROR(HW_GPU, "Unknown GetPointer @ 0x{:016X}", addr);
return {};
} }
void MemoryManager::ReadBlock(GPUVAddr src_addr, void* dest_buffer, std::size_t size) { void MemoryManager::ReadBlock(GPUVAddr src_addr, void* dest_buffer, std::size_t size) {
@ -210,13 +199,252 @@ void MemoryManager::CopyBlock(GPUVAddr dest_addr, GPUVAddr src_addr, std::size_t
std::memcpy(GetPointer(dest_addr), GetPointer(src_addr), size); std::memcpy(GetPointer(dest_addr), GetPointer(src_addr), size);
} }
VAddr& MemoryManager::PageSlot(GPUVAddr gpu_addr) { void MemoryManager::MapPages(GPUVAddr base, u64 size, u8* memory, Common::PageType type,
auto& block{page_table[(gpu_addr >> (PAGE_BITS + PAGE_TABLE_BITS)) & PAGE_TABLE_MASK]}; VAddr backing_addr) {
if (!block) { LOG_DEBUG(HW_GPU, "Mapping {} onto {:016X}-{:016X}", fmt::ptr(memory), base * page_size,
block = std::make_unique<PageBlock>(); (base + size) * page_size);
block->fill(static_cast<VAddr>(PageStatus::Unmapped));
const VAddr end{base + size};
ASSERT_MSG(end <= page_table.pointers.size(), "out of range mapping at {:016X}",
base + page_table.pointers.size());
std::fill(page_table.attributes.begin() + base, page_table.attributes.begin() + end, type);
if (memory == nullptr) {
std::fill(page_table.pointers.begin() + base, page_table.pointers.begin() + end, memory);
std::fill(page_table.backing_addr.begin() + base, page_table.backing_addr.begin() + end,
backing_addr);
} else {
while (base != end) {
page_table.pointers[base] = memory;
page_table.backing_addr[base] = backing_addr;
base += 1;
memory += page_size;
backing_addr += page_size;
}
}
}
void MemoryManager::MapMemoryRegion(GPUVAddr base, u64 size, u8* target, VAddr backing_addr) {
ASSERT_MSG((size & page_mask) == 0, "non-page aligned size: {:016X}", size);
ASSERT_MSG((base & page_mask) == 0, "non-page aligned base: {:016X}", base);
MapPages(base / page_size, size / page_size, target, Common::PageType::Memory, backing_addr);
}
void MemoryManager::UnmapRegion(GPUVAddr base, u64 size) {
ASSERT_MSG((size & page_mask) == 0, "non-page aligned size: {:016X}", size);
ASSERT_MSG((base & page_mask) == 0, "non-page aligned base: {:016X}", base);
MapPages(base / page_size, size / page_size, nullptr, Common::PageType::Unmapped);
}
bool VirtualMemoryArea::CanBeMergedWith(const VirtualMemoryArea& next) const {
ASSERT(base + size == next.base);
if (type != next.type) {
return {};
}
if (type == VirtualMemoryArea::Type::Allocated && (offset + size != next.offset)) {
return {};
}
if (type == VirtualMemoryArea::Type::Mapped && backing_memory + size != next.backing_memory) {
return {};
}
return true;
}
MemoryManager::VMAHandle MemoryManager::FindVMA(GPUVAddr target) const {
if (target >= address_space_end) {
return vma_map.end();
} else {
return std::prev(vma_map.upper_bound(target));
}
}
MemoryManager::VMAIter MemoryManager::Allocate(VMAIter vma_handle) {
VirtualMemoryArea& vma{vma_handle->second};
vma.type = VirtualMemoryArea::Type::Allocated;
vma.backing_addr = 0;
vma.backing_memory = {};
UpdatePageTableForVMA(vma);
return MergeAdjacent(vma_handle);
}
MemoryManager::VMAHandle MemoryManager::AllocateMemory(GPUVAddr target, std::size_t offset,
u64 size) {
// This is the appropriately sized VMA that will turn into our allocation.
VMAIter vma_handle{CarveVMA(target, size)};
VirtualMemoryArea& vma{vma_handle->second};
ASSERT(vma.size == size);
vma.offset = offset;
return Allocate(vma_handle);
}
MemoryManager::VMAHandle MemoryManager::MapBackingMemory(GPUVAddr target, u8* memory, u64 size,
VAddr backing_addr) {
// This is the appropriately sized VMA that will turn into our allocation.
VMAIter vma_handle{CarveVMA(target, size)};
VirtualMemoryArea& vma{vma_handle->second};
ASSERT(vma.size == size);
vma.type = VirtualMemoryArea::Type::Mapped;
vma.backing_memory = memory;
vma.backing_addr = backing_addr;
UpdatePageTableForVMA(vma);
return MergeAdjacent(vma_handle);
}
void MemoryManager::UnmapRange(GPUVAddr target, u64 size) {
VMAIter vma{CarveVMARange(target, size)};
const VAddr target_end{target + size};
const VMAIter end{vma_map.end()};
// The comparison against the end of the range must be done using addresses since VMAs can be
// merged during this process, causing invalidation of the iterators.
while (vma != end && vma->second.base < target_end) {
// Unmapped ranges return to allocated state and can be reused
// This behavior is used by Super Mario Odyssey, Sonic Forces, and likely other games
vma = std::next(Allocate(vma));
}
ASSERT(FindVMA(target)->second.size >= size);
}
MemoryManager::VMAIter MemoryManager::StripIterConstness(const VMAHandle& iter) {
// This uses a neat C++ trick to convert a const_iterator to a regular iterator, given
// non-const access to its container.
return vma_map.erase(iter, iter); // Erases an empty range of elements
}
MemoryManager::VMAIter MemoryManager::CarveVMA(GPUVAddr base, u64 size) {
ASSERT_MSG((size & page_mask) == 0, "non-page aligned size: 0x{:016X}", size);
ASSERT_MSG((base & page_mask) == 0, "non-page aligned base: 0x{:016X}", base);
VMAIter vma_handle{StripIterConstness(FindVMA(base))};
if (vma_handle == vma_map.end()) {
// Target address is outside the managed range
return {};
}
const VirtualMemoryArea& vma{vma_handle->second};
if (vma.type == VirtualMemoryArea::Type::Mapped) {
// Region is already allocated
return {};
}
const VAddr start_in_vma{base - vma.base};
const VAddr end_in_vma{start_in_vma + size};
ASSERT_MSG(end_in_vma <= vma.size, "region size 0x{:016X} is less than required size 0x{:016X}",
vma.size, end_in_vma);
if (end_in_vma < vma.size) {
// Split VMA at the end of the allocated region
SplitVMA(vma_handle, end_in_vma);
}
if (start_in_vma != 0) {
// Split VMA at the start of the allocated region
vma_handle = SplitVMA(vma_handle, start_in_vma);
}
return vma_handle;
}
MemoryManager::VMAIter MemoryManager::CarveVMARange(GPUVAddr target, u64 size) {
ASSERT_MSG((size & page_mask) == 0, "non-page aligned size: 0x{:016X}", size);
ASSERT_MSG((target & page_mask) == 0, "non-page aligned base: 0x{:016X}", target);
const VAddr target_end{target + size};
ASSERT(target_end >= target);
ASSERT(size > 0);
VMAIter begin_vma{StripIterConstness(FindVMA(target))};
const VMAIter i_end{vma_map.lower_bound(target_end)};
if (std::any_of(begin_vma, i_end, [](const auto& entry) {
return entry.second.type == VirtualMemoryArea::Type::Unmapped;
})) {
return {};
}
if (target != begin_vma->second.base) {
begin_vma = SplitVMA(begin_vma, target - begin_vma->second.base);
}
VMAIter end_vma{StripIterConstness(FindVMA(target_end))};
if (end_vma != vma_map.end() && target_end != end_vma->second.base) {
end_vma = SplitVMA(end_vma, target_end - end_vma->second.base);
}
return begin_vma;
}
MemoryManager::VMAIter MemoryManager::SplitVMA(VMAIter vma_handle, u64 offset_in_vma) {
VirtualMemoryArea& old_vma{vma_handle->second};
VirtualMemoryArea new_vma{old_vma}; // Make a copy of the VMA
// For now, don't allow no-op VMA splits (trying to split at a boundary) because it's probably
// a bug. This restriction might be removed later.
ASSERT(offset_in_vma < old_vma.size);
ASSERT(offset_in_vma > 0);
old_vma.size = offset_in_vma;
new_vma.base += offset_in_vma;
new_vma.size -= offset_in_vma;
switch (new_vma.type) {
case VirtualMemoryArea::Type::Unmapped:
break;
case VirtualMemoryArea::Type::Allocated:
new_vma.offset += offset_in_vma;
break;
case VirtualMemoryArea::Type::Mapped:
new_vma.backing_memory += offset_in_vma;
break;
}
ASSERT(old_vma.CanBeMergedWith(new_vma));
return vma_map.emplace_hint(std::next(vma_handle), new_vma.base, new_vma);
}
MemoryManager::VMAIter MemoryManager::MergeAdjacent(VMAIter iter) {
const VMAIter next_vma{std::next(iter)};
if (next_vma != vma_map.end() && iter->second.CanBeMergedWith(next_vma->second)) {
iter->second.size += next_vma->second.size;
vma_map.erase(next_vma);
}
if (iter != vma_map.begin()) {
VMAIter prev_vma{std::prev(iter)};
if (prev_vma->second.CanBeMergedWith(iter->second)) {
prev_vma->second.size += iter->second.size;
vma_map.erase(iter);
iter = prev_vma;
}
}
return iter;
}
void MemoryManager::UpdatePageTableForVMA(const VirtualMemoryArea& vma) {
switch (vma.type) {
case VirtualMemoryArea::Type::Unmapped:
UnmapRegion(vma.base, vma.size);
break;
case VirtualMemoryArea::Type::Allocated:
MapMemoryRegion(vma.base, vma.size, nullptr, vma.backing_addr);
break;
case VirtualMemoryArea::Type::Mapped:
MapMemoryRegion(vma.base, vma.size, vma.backing_memory, vma.backing_addr);
break;
} }
return (*block)[(gpu_addr >> PAGE_BITS) & PAGE_BLOCK_MASK];
} }
} // namespace Tegra } // namespace Tegra

View file

@ -1,82 +1,148 @@
// Copyright 2018 yuzu emulator team // Copyright 2018 yuzu emulator team
// Licensed under GPLv2 or any later version // Licensed under GPLv2 or any later version
// Refer to the license.txt file included. // Refer to the license.txt file included.
#pragma once #pragma once
#include <array> #include <map>
#include <memory>
#include <optional> #include <optional>
#include <vector>
#include "common/common_types.h" #include "common/common_types.h"
#include "common/page_table.h"
namespace Tegra { namespace Tegra {
/// Virtual addresses in the GPU's memory map are 64 bit. /**
using GPUVAddr = u64; * Represents a VMA in an address space. A VMA is a contiguous region of virtual addressing space
* with homogeneous attributes across its extents. In this particular implementation each VMA is
* also backed by a single host memory allocation.
*/
struct VirtualMemoryArea {
enum class Type : u8 {
Unmapped,
Allocated,
Mapped,
};
/// Virtual base address of the region.
GPUVAddr base{};
/// Size of the region.
u64 size{};
/// Memory area mapping type.
Type type{Type::Unmapped};
/// CPU memory mapped address corresponding to this memory area.
VAddr backing_addr{};
/// Offset into the backing_memory the mapping starts from.
std::size_t offset{};
/// Pointer backing this VMA.
u8* backing_memory{};
/// Tests if this area can be merged to the right with `next`.
bool CanBeMergedWith(const VirtualMemoryArea& next) const;
};
class MemoryManager final { class MemoryManager final {
public: public:
MemoryManager(); MemoryManager();
GPUVAddr AllocateSpace(u64 size, u64 align); GPUVAddr AllocateSpace(u64 size, u64 align);
GPUVAddr AllocateSpace(GPUVAddr gpu_addr, u64 size, u64 align); GPUVAddr AllocateSpace(GPUVAddr addr, u64 size, u64 align);
GPUVAddr MapBufferEx(VAddr cpu_addr, u64 size); GPUVAddr MapBufferEx(VAddr cpu_addr, u64 size);
GPUVAddr MapBufferEx(VAddr cpu_addr, GPUVAddr gpu_addr, u64 size); GPUVAddr MapBufferEx(VAddr cpu_addr, GPUVAddr addr, u64 size);
GPUVAddr UnmapBuffer(GPUVAddr gpu_addr, u64 size); GPUVAddr UnmapBuffer(GPUVAddr addr, u64 size);
GPUVAddr GetRegionEnd(GPUVAddr region_start) const; std::optional<VAddr> GpuToCpuAddress(GPUVAddr addr);
std::optional<VAddr> GpuToCpuAddress(GPUVAddr gpu_addr);
static constexpr u64 PAGE_BITS = 16; template <typename T>
static constexpr u64 PAGE_SIZE = 1 << PAGE_BITS; T Read(GPUVAddr addr);
static constexpr u64 PAGE_MASK = PAGE_SIZE - 1;
u8 Read8(GPUVAddr addr); template <typename T>
u16 Read16(GPUVAddr addr); void Write(GPUVAddr addr, T data);
u32 Read32(GPUVAddr addr);
u64 Read64(GPUVAddr addr);
void Write8(GPUVAddr addr, u8 data); u8* GetPointer(GPUVAddr addr);
void Write16(GPUVAddr addr, u16 data);
void Write32(GPUVAddr addr, u32 data);
void Write64(GPUVAddr addr, u64 data);
u8* GetPointer(GPUVAddr vaddr);
void ReadBlock(GPUVAddr src_addr, void* dest_buffer, std::size_t size); void ReadBlock(GPUVAddr src_addr, void* dest_buffer, std::size_t size);
void WriteBlock(GPUVAddr dest_addr, const void* src_buffer, std::size_t size); void WriteBlock(GPUVAddr dest_addr, const void* src_buffer, std::size_t size);
void CopyBlock(VAddr dest_addr, VAddr src_addr, std::size_t size); void CopyBlock(GPUVAddr dest_addr, GPUVAddr src_addr, std::size_t size);
private: private:
enum class PageStatus : u64 { using VMAMap = std::map<GPUVAddr, VirtualMemoryArea>;
Unmapped = 0xFFFFFFFFFFFFFFFFULL, using VMAHandle = VMAMap::const_iterator;
Allocated = 0xFFFFFFFFFFFFFFFEULL, using VMAIter = VMAMap::iterator;
Reserved = 0xFFFFFFFFFFFFFFFDULL,
};
std::optional<GPUVAddr> FindFreeBlock(GPUVAddr region_start, u64 size, u64 align, bool IsAddressValid(GPUVAddr addr) const;
PageStatus status); void MapPages(GPUVAddr base, u64 size, u8* memory, Common::PageType type,
VAddr& PageSlot(GPUVAddr gpu_addr); VAddr backing_addr = 0);
void MapMemoryRegion(GPUVAddr base, u64 size, u8* target, VAddr backing_addr);
void UnmapRegion(GPUVAddr base, u64 size);
static constexpr u64 MAX_ADDRESS{0x10000000000ULL}; /// Finds the VMA in which the given address is included in, or `vma_map.end()`.
static constexpr u64 PAGE_TABLE_BITS{10}; VMAHandle FindVMA(GPUVAddr target) const;
static constexpr u64 PAGE_TABLE_SIZE{1 << PAGE_TABLE_BITS};
static constexpr u64 PAGE_TABLE_MASK{PAGE_TABLE_SIZE - 1};
static constexpr u64 PAGE_BLOCK_BITS{14};
static constexpr u64 PAGE_BLOCK_SIZE{1 << PAGE_BLOCK_BITS};
static constexpr u64 PAGE_BLOCK_MASK{PAGE_BLOCK_SIZE - 1};
using PageBlock = std::array<VAddr, PAGE_BLOCK_SIZE>; VMAHandle AllocateMemory(GPUVAddr target, std::size_t offset, u64 size);
std::array<std::unique_ptr<PageBlock>, PAGE_TABLE_SIZE> page_table{};
struct MappedRegion { /**
VAddr cpu_addr; * Maps an unmanaged host memory pointer at a given address.
GPUVAddr gpu_addr; *
u64 size; * @param target The guest address to start the mapping at.
}; * @param memory The memory to be mapped.
* @param size Size of the mapping.
* @param state MemoryState tag to attach to the VMA.
*/
VMAHandle MapBackingMemory(GPUVAddr target, u8* memory, u64 size, VAddr backing_addr);
std::vector<MappedRegion> mapped_regions; /// Unmaps a range of addresses, splitting VMAs as necessary.
void UnmapRange(GPUVAddr target, u64 size);
/// Converts a VMAHandle to a mutable VMAIter.
VMAIter StripIterConstness(const VMAHandle& iter);
/// Marks as the specfied VMA as allocated.
VMAIter Allocate(VMAIter vma);
/**
* Carves a VMA of a specific size at the specified address by splitting Free VMAs while doing
* the appropriate error checking.
*/
VMAIter CarveVMA(GPUVAddr base, u64 size);
/**
* Splits the edges of the given range of non-Free VMAs so that there is a VMA split at each
* end of the range.
*/
VMAIter CarveVMARange(GPUVAddr base, u64 size);
/**
* Splits a VMA in two, at the specified offset.
* @returns the right side of the split, with the original iterator becoming the left side.
*/
VMAIter SplitVMA(VMAIter vma, u64 offset_in_vma);
/**
* Checks for and merges the specified VMA with adjacent ones if possible.
* @returns the merged VMA or the original if no merging was possible.
*/
VMAIter MergeAdjacent(VMAIter vma);
/// Updates the pages corresponding to this VMA so they match the VMA's attributes.
void UpdatePageTableForVMA(const VirtualMemoryArea& vma);
/// Finds a free (unmapped region) of the specified size starting at the specified address.
GPUVAddr FindFreeRegion(GPUVAddr region_start, u64 size);
private:
static constexpr u64 page_bits{16};
static constexpr u64 page_size{1 << page_bits};
static constexpr u64 page_mask{page_size - 1};
/// Address space in bits, this is fairly arbitrary but sufficiently large.
static constexpr u32 address_space_width{39};
/// Start address for mapping, this is fairly arbitrary but must be non-zero.
static constexpr GPUVAddr address_space_base{0x100000};
/// End of address space, based on address space in bits.
static constexpr GPUVAddr address_space_end{1ULL << address_space_width};
Common::PageTable page_table{page_bits};
VMAMap vma_map;
}; };
} // namespace Tegra } // namespace Tegra

View file

@ -9,7 +9,6 @@
#include "common/common_types.h" #include "common/common_types.h"
#include "video_core/engines/fermi_2d.h" #include "video_core/engines/fermi_2d.h"
#include "video_core/gpu.h" #include "video_core/gpu.h"
#include "video_core/memory_manager.h"
namespace VideoCore { namespace VideoCore {

View file

@ -21,8 +21,8 @@ CachedBufferEntry::CachedBufferEntry(VAddr cpu_addr, std::size_t size, GLintptr
OGLBufferCache::OGLBufferCache(RasterizerOpenGL& rasterizer, std::size_t size) OGLBufferCache::OGLBufferCache(RasterizerOpenGL& rasterizer, std::size_t size)
: RasterizerCache{rasterizer}, stream_buffer(size, true) {} : RasterizerCache{rasterizer}, stream_buffer(size, true) {}
GLintptr OGLBufferCache::UploadMemory(Tegra::GPUVAddr gpu_addr, std::size_t size, GLintptr OGLBufferCache::UploadMemory(GPUVAddr gpu_addr, std::size_t size, std::size_t alignment,
std::size_t alignment, bool cache) { bool cache) {
auto& memory_manager = Core::System::GetInstance().GPU().MemoryManager(); auto& memory_manager = Core::System::GetInstance().GPU().MemoryManager();
// Cache management is a big overhead, so only cache entries with a given size. // Cache management is a big overhead, so only cache entries with a given size.

View file

@ -58,7 +58,7 @@ public:
/// Uploads data from a guest GPU address. Returns host's buffer offset where it's been /// Uploads data from a guest GPU address. Returns host's buffer offset where it's been
/// allocated. /// allocated.
GLintptr UploadMemory(Tegra::GPUVAddr gpu_addr, std::size_t size, std::size_t alignment = 4, GLintptr UploadMemory(GPUVAddr gpu_addr, std::size_t size, std::size_t alignment = 4,
bool cache = true); bool cache = true);
/// Uploads from a host memory. Returns host's buffer offset where it's been allocated. /// Uploads from a host memory. Returns host's buffer offset where it's been allocated.

View file

@ -46,7 +46,7 @@ GlobalRegion GlobalRegionCacheOpenGL::TryGetReservedGlobalRegion(CacheAddr addr,
return search->second; return search->second;
} }
GlobalRegion GlobalRegionCacheOpenGL::GetUncachedGlobalRegion(Tegra::GPUVAddr addr, u32 size, GlobalRegion GlobalRegionCacheOpenGL::GetUncachedGlobalRegion(GPUVAddr addr, u32 size,
u8* host_ptr) { u8* host_ptr) {
GlobalRegion region{TryGetReservedGlobalRegion(ToCacheAddr(host_ptr), size)}; GlobalRegion region{TryGetReservedGlobalRegion(ToCacheAddr(host_ptr), size)};
if (!region) { if (!region) {
@ -76,8 +76,8 @@ GlobalRegion GlobalRegionCacheOpenGL::GetGlobalRegion(
const auto cbufs{gpu.Maxwell3D().state.shader_stages[static_cast<u64>(stage)]}; const auto cbufs{gpu.Maxwell3D().state.shader_stages[static_cast<u64>(stage)]};
const auto addr{cbufs.const_buffers[global_region.GetCbufIndex()].address + const auto addr{cbufs.const_buffers[global_region.GetCbufIndex()].address +
global_region.GetCbufOffset()}; global_region.GetCbufOffset()};
const auto actual_addr{memory_manager.Read64(addr)}; const auto actual_addr{memory_manager.Read<u64>(addr)};
const auto size{memory_manager.Read32(addr + 8)}; const auto size{memory_manager.Read<u32>(addr + 8)};
// Look up global region in the cache based on address // Look up global region in the cache based on address
const auto& host_ptr{memory_manager.GetPointer(actual_addr)}; const auto& host_ptr{memory_manager.GetPointer(actual_addr)};

View file

@ -66,7 +66,7 @@ public:
private: private:
GlobalRegion TryGetReservedGlobalRegion(CacheAddr addr, u32 size) const; GlobalRegion TryGetReservedGlobalRegion(CacheAddr addr, u32 size) const;
GlobalRegion GetUncachedGlobalRegion(Tegra::GPUVAddr addr, u32 size, u8* host_ptr); GlobalRegion GetUncachedGlobalRegion(GPUVAddr addr, u32 size, u8* host_ptr);
void ReserveGlobalRegion(GlobalRegion region); void ReserveGlobalRegion(GlobalRegion region);
std::unordered_map<CacheAddr, GlobalRegion> reserve; std::unordered_map<CacheAddr, GlobalRegion> reserve;

View file

@ -40,8 +40,7 @@ GLintptr PrimitiveAssembler::MakeQuadArray(u32 first, u32 count) {
return index_offset; return index_offset;
} }
GLintptr PrimitiveAssembler::MakeQuadIndexed(Tegra::GPUVAddr gpu_addr, std::size_t index_size, GLintptr PrimitiveAssembler::MakeQuadIndexed(GPUVAddr gpu_addr, std::size_t index_size, u32 count) {
u32 count) {
const std::size_t map_size{CalculateQuadSize(count)}; const std::size_t map_size{CalculateQuadSize(count)};
auto [dst_pointer, index_offset] = buffer_cache.ReserveMemory(map_size); auto [dst_pointer, index_offset] = buffer_cache.ReserveMemory(map_size);

View file

@ -24,7 +24,7 @@ public:
GLintptr MakeQuadArray(u32 first, u32 count); GLintptr MakeQuadArray(u32 first, u32 count);
GLintptr MakeQuadIndexed(Tegra::GPUVAddr gpu_addr, std::size_t index_size, u32 count); GLintptr MakeQuadIndexed(GPUVAddr gpu_addr, std::size_t index_size, u32 count);
private: private:
OGLBufferCache& buffer_cache; OGLBufferCache& buffer_cache;

View file

@ -225,8 +225,8 @@ void RasterizerOpenGL::SetupVertexBuffer(GLuint vao) {
if (!vertex_array.IsEnabled()) if (!vertex_array.IsEnabled())
continue; continue;
const Tegra::GPUVAddr start = vertex_array.StartAddress(); const GPUVAddr start = vertex_array.StartAddress();
const Tegra::GPUVAddr end = regs.vertex_array_limit[index].LimitAddress(); const GPUVAddr end = regs.vertex_array_limit[index].LimitAddress();
ASSERT(end > start); ASSERT(end > start);
const u64 size = end - start + 1; const u64 size = end - start + 1;
@ -421,8 +421,8 @@ std::size_t RasterizerOpenGL::CalculateVertexArraysSize() const {
if (!regs.vertex_array[index].IsEnabled()) if (!regs.vertex_array[index].IsEnabled())
continue; continue;
const Tegra::GPUVAddr start = regs.vertex_array[index].StartAddress(); const GPUVAddr start = regs.vertex_array[index].StartAddress();
const Tegra::GPUVAddr end = regs.vertex_array_limit[index].LimitAddress(); const GPUVAddr end = regs.vertex_array_limit[index].LimitAddress();
ASSERT(end > start); ASSERT(end > start);
size += end - start + 1; size += end - start + 1;

View file

@ -55,7 +55,7 @@ static void ApplyTextureDefaults(GLuint texture, u32 max_mip_level) {
} }
} }
void SurfaceParams::InitCacheParameters(Tegra::GPUVAddr gpu_addr_) { void SurfaceParams::InitCacheParameters(GPUVAddr gpu_addr_) {
auto& memory_manager{Core::System::GetInstance().GPU().MemoryManager()}; auto& memory_manager{Core::System::GetInstance().GPU().MemoryManager()};
gpu_addr = gpu_addr_; gpu_addr = gpu_addr_;
@ -222,7 +222,7 @@ std::size_t SurfaceParams::InnerMemorySize(bool force_gl, bool layer_only,
} }
/*static*/ SurfaceParams SurfaceParams::CreateForDepthBuffer( /*static*/ SurfaceParams SurfaceParams::CreateForDepthBuffer(
u32 zeta_width, u32 zeta_height, Tegra::GPUVAddr zeta_address, Tegra::DepthFormat format, u32 zeta_width, u32 zeta_height, GPUVAddr zeta_address, Tegra::DepthFormat format,
u32 block_width, u32 block_height, u32 block_depth, u32 block_width, u32 block_height, u32 block_depth,
Tegra::Engines::Maxwell3D::Regs::InvMemoryLayout type) { Tegra::Engines::Maxwell3D::Regs::InvMemoryLayout type) {
SurfaceParams params{}; SurfaceParams params{};
@ -564,6 +564,12 @@ void RasterizerCacheOpenGL::CopySurface(const Surface& src_surface, const Surfac
CachedSurface::CachedSurface(const SurfaceParams& params) CachedSurface::CachedSurface(const SurfaceParams& params)
: params{params}, gl_target{SurfaceTargetToGL(params.target)}, : params{params}, gl_target{SurfaceTargetToGL(params.target)},
cached_size_in_bytes{params.size_in_bytes}, RasterizerCacheObject{params.host_ptr} { cached_size_in_bytes{params.size_in_bytes}, RasterizerCacheObject{params.host_ptr} {
const auto optional_cpu_addr{
Core::System::GetInstance().GPU().MemoryManager().GpuToCpuAddress(params.gpu_addr)};
ASSERT_MSG(optional_cpu_addr, "optional_cpu_addr is invalid");
cpu_addr = *optional_cpu_addr;
texture.Create(gl_target); texture.Create(gl_target);
// TODO(Rodrigo): Using params.GetRect() returns a different size than using its Mip*(0) // TODO(Rodrigo): Using params.GetRect() returns a different size than using its Mip*(0)
@ -603,20 +609,6 @@ CachedSurface::CachedSurface(const SurfaceParams& params)
ApplyTextureDefaults(texture.handle, params.max_mip_level); ApplyTextureDefaults(texture.handle, params.max_mip_level);
OpenGL::LabelGLObject(GL_TEXTURE, texture.handle, params.gpu_addr, params.IdentityString()); OpenGL::LabelGLObject(GL_TEXTURE, texture.handle, params.gpu_addr, params.IdentityString());
// Clamp size to mapped GPU memory region
// TODO(bunnei): Super Mario Odyssey maps a 0x40000 byte region and then uses it for a 0x80000
// R32F render buffer. We do not yet know if this is a game bug or something else, but this
// check is necessary to prevent flushing from overwriting unmapped memory.
auto& memory_manager{Core::System::GetInstance().GPU().MemoryManager()};
const u64 max_size{memory_manager.GetRegionEnd(params.gpu_addr) - params.gpu_addr};
if (cached_size_in_bytes > max_size) {
LOG_ERROR(HW_GPU, "Surface size {} exceeds region size {}", params.size_in_bytes, max_size);
cached_size_in_bytes = max_size;
}
cpu_addr = *memory_manager.GpuToCpuAddress(params.gpu_addr);
} }
MICROPROFILE_DEFINE(OpenGL_SurfaceLoad, "OpenGL", "Surface Load", MP_RGB(128, 192, 64)); MICROPROFILE_DEFINE(OpenGL_SurfaceLoad, "OpenGL", "Surface Load", MP_RGB(128, 192, 64));
@ -925,7 +917,7 @@ void RasterizerCacheOpenGL::LoadSurface(const Surface& surface) {
} }
Surface RasterizerCacheOpenGL::GetSurface(const SurfaceParams& params, bool preserve_contents) { Surface RasterizerCacheOpenGL::GetSurface(const SurfaceParams& params, bool preserve_contents) {
if (params.gpu_addr == 0 || params.height * params.width == 0) { if (!params.IsValid()) {
return {}; return {};
} }
@ -980,11 +972,11 @@ void RasterizerCacheOpenGL::FastLayeredCopySurface(const Surface& src_surface,
const auto& init_params{src_surface->GetSurfaceParams()}; const auto& init_params{src_surface->GetSurfaceParams()};
const auto& dst_params{dst_surface->GetSurfaceParams()}; const auto& dst_params{dst_surface->GetSurfaceParams()};
auto& memory_manager{Core::System::GetInstance().GPU().MemoryManager()}; auto& memory_manager{Core::System::GetInstance().GPU().MemoryManager()};
Tegra::GPUVAddr address{init_params.gpu_addr}; GPUVAddr address{init_params.gpu_addr};
const std::size_t layer_size{dst_params.LayerMemorySize()}; const std::size_t layer_size{dst_params.LayerMemorySize()};
for (u32 layer = 0; layer < dst_params.depth; layer++) { for (u32 layer = 0; layer < dst_params.depth; layer++) {
for (u32 mipmap = 0; mipmap < dst_params.max_mip_level; mipmap++) { for (u32 mipmap = 0; mipmap < dst_params.max_mip_level; mipmap++) {
const Tegra::GPUVAddr sub_address{address + dst_params.GetMipmapLevelOffset(mipmap)}; const GPUVAddr sub_address{address + dst_params.GetMipmapLevelOffset(mipmap)};
const Surface& copy{TryGet(memory_manager.GetPointer(sub_address))}; const Surface& copy{TryGet(memory_manager.GetPointer(sub_address))};
if (!copy) { if (!copy) {
continue; continue;
@ -1244,10 +1236,9 @@ static std::optional<u32> TryFindBestMipMap(std::size_t memory, const SurfacePar
return {}; return {};
} }
static std::optional<u32> TryFindBestLayer(Tegra::GPUVAddr addr, const SurfaceParams params, static std::optional<u32> TryFindBestLayer(GPUVAddr addr, const SurfaceParams params, u32 mipmap) {
u32 mipmap) {
const std::size_t size{params.LayerMemorySize()}; const std::size_t size{params.LayerMemorySize()};
Tegra::GPUVAddr start{params.gpu_addr + params.GetMipmapLevelOffset(mipmap)}; GPUVAddr start{params.gpu_addr + params.GetMipmapLevelOffset(mipmap)};
for (u32 i = 0; i < params.depth; i++) { for (u32 i = 0; i < params.depth; i++) {
if (start == addr) { if (start == addr) {
return {i}; return {i};

View file

@ -109,6 +109,11 @@ struct SurfaceParams {
return size; return size;
} }
/// Returns true if the parameters constitute a valid rasterizer surface.
bool IsValid() const {
return gpu_addr && host_ptr && height && width;
}
/// Returns the exact size of the memory occupied by a layer in a texture in VRAM, including /// Returns the exact size of the memory occupied by a layer in a texture in VRAM, including
/// mipmaps. /// mipmaps.
std::size_t LayerMemorySize() const { std::size_t LayerMemorySize() const {
@ -210,7 +215,7 @@ struct SurfaceParams {
/// Creates SurfaceParams for a depth buffer configuration /// Creates SurfaceParams for a depth buffer configuration
static SurfaceParams CreateForDepthBuffer( static SurfaceParams CreateForDepthBuffer(
u32 zeta_width, u32 zeta_height, Tegra::GPUVAddr zeta_address, Tegra::DepthFormat format, u32 zeta_width, u32 zeta_height, GPUVAddr zeta_address, Tegra::DepthFormat format,
u32 block_width, u32 block_height, u32 block_depth, u32 block_width, u32 block_height, u32 block_depth,
Tegra::Engines::Maxwell3D::Regs::InvMemoryLayout type); Tegra::Engines::Maxwell3D::Regs::InvMemoryLayout type);
@ -232,7 +237,7 @@ struct SurfaceParams {
} }
/// Initializes parameters for caching, should be called after everything has been initialized /// Initializes parameters for caching, should be called after everything has been initialized
void InitCacheParameters(Tegra::GPUVAddr gpu_addr); void InitCacheParameters(GPUVAddr gpu_addr);
std::string TargetName() const { std::string TargetName() const {
switch (target) { switch (target) {
@ -297,7 +302,7 @@ struct SurfaceParams {
bool srgb_conversion; bool srgb_conversion;
// Parameters used for caching // Parameters used for caching
u8* host_ptr; u8* host_ptr;
Tegra::GPUVAddr gpu_addr; GPUVAddr gpu_addr;
std::size_t size_in_bytes; std::size_t size_in_bytes;
std::size_t size_in_bytes_gl; std::size_t size_in_bytes_gl;

View file

@ -32,7 +32,7 @@ struct UnspecializedShader {
namespace { namespace {
/// Gets the address for the specified shader stage program /// Gets the address for the specified shader stage program
Tegra::GPUVAddr GetShaderAddress(Maxwell::ShaderProgram program) { GPUVAddr GetShaderAddress(Maxwell::ShaderProgram program) {
const auto& gpu{Core::System::GetInstance().GPU().Maxwell3D()}; const auto& gpu{Core::System::GetInstance().GPU().Maxwell3D()};
const auto& shader_config{gpu.regs.shader_config[static_cast<std::size_t>(program)]}; const auto& shader_config{gpu.regs.shader_config[static_cast<std::size_t>(program)]};
return gpu.regs.code_address.CodeAddress() + shader_config.offset; return gpu.regs.code_address.CodeAddress() + shader_config.offset;
@ -486,7 +486,7 @@ Shader ShaderCacheOpenGL::GetStageProgram(Maxwell::ShaderProgram program) {
} }
auto& memory_manager{Core::System::GetInstance().GPU().MemoryManager()}; auto& memory_manager{Core::System::GetInstance().GPU().MemoryManager()};
const Tegra::GPUVAddr program_addr{GetShaderAddress(program)}; const GPUVAddr program_addr{GetShaderAddress(program)};
// Look up shader in the cache based on address // Look up shader in the cache based on address
const auto& host_ptr{memory_manager.GetPointer(program_addr)}; const auto& host_ptr{memory_manager.GetPointer(program_addr)};

View file

@ -39,8 +39,7 @@ VKBufferCache::VKBufferCache(Tegra::MemoryManager& tegra_memory_manager,
VKBufferCache::~VKBufferCache() = default; VKBufferCache::~VKBufferCache() = default;
u64 VKBufferCache::UploadMemory(Tegra::GPUVAddr gpu_addr, std::size_t size, u64 alignment, u64 VKBufferCache::UploadMemory(GPUVAddr gpu_addr, std::size_t size, u64 alignment, bool cache) {
bool cache) {
const auto cpu_addr{tegra_memory_manager.GpuToCpuAddress(gpu_addr)}; const auto cpu_addr{tegra_memory_manager.GpuToCpuAddress(gpu_addr)};
ASSERT_MSG(cpu_addr, "Invalid GPU address"); ASSERT_MSG(cpu_addr, "Invalid GPU address");

View file

@ -68,8 +68,7 @@ public:
/// Uploads data from a guest GPU address. Returns host's buffer offset where it's been /// Uploads data from a guest GPU address. Returns host's buffer offset where it's been
/// allocated. /// allocated.
u64 UploadMemory(Tegra::GPUVAddr gpu_addr, std::size_t size, u64 alignment = 4, u64 UploadMemory(GPUVAddr gpu_addr, std::size_t size, u64 alignment = 4, bool cache = true);
bool cache = true);
/// Uploads from a host memory. Returns host's buffer offset where it's been allocated. /// Uploads from a host memory. Returns host's buffer offset where it's been allocated.
u64 UploadHostMemory(const u8* raw_pointer, std::size_t size, u64 alignment = 4); u64 UploadHostMemory(const u8* raw_pointer, std::size_t size, u64 alignment = 4);

View file

@ -261,7 +261,7 @@ void GraphicsSurfaceWidget::OnSurfaceSourceChanged(int new_value) {
void GraphicsSurfaceWidget::OnSurfaceAddressChanged(qint64 new_value) { void GraphicsSurfaceWidget::OnSurfaceAddressChanged(qint64 new_value) {
if (surface_address != new_value) { if (surface_address != new_value) {
surface_address = static_cast<Tegra::GPUVAddr>(new_value); surface_address = static_cast<GPUVAddr>(new_value);
surface_source_list->setCurrentIndex(static_cast<int>(Source::Custom)); surface_source_list->setCurrentIndex(static_cast<int>(Source::Custom));
emit Update(); emit Update();

View file

@ -87,7 +87,7 @@ private:
QPushButton* save_surface; QPushButton* save_surface;
Source surface_source; Source surface_source;
Tegra::GPUVAddr surface_address; GPUVAddr surface_address;
unsigned surface_width; unsigned surface_width;
unsigned surface_height; unsigned surface_height;
Tegra::Texture::TextureFormat surface_format; Tegra::Texture::TextureFormat surface_format;