accelerateDMA: Fixes and feedback.

This commit is contained in:
Fernando Sahmkow 2021-07-12 04:10:42 +02:00
parent be1a3f7a0f
commit bc19d28963
3 changed files with 62 additions and 88 deletions

View File

@ -202,6 +202,36 @@ private:
} }
} }
template <typename Func>
void ForEachWrittenRange(VAddr cpu_addr, u64 size, Func&& func) {
const VAddr start_address = cpu_addr;
const VAddr end_address = start_address + size;
const VAddr search_base =
static_cast<VAddr>(std::min<s64>(0LL, static_cast<s64>(start_address - size)));
const IntervalType search_interval{search_base, search_base + 1};
auto it = common_ranges.lower_bound(search_interval);
if (it == common_ranges.end()) {
it = common_ranges.begin();
}
for (; it != common_ranges.end(); it++) {
VAddr inter_addr_end = it->upper();
VAddr inter_addr = it->lower();
if (inter_addr >= end_address) {
break;
}
if (inter_addr_end <= start_address) {
continue;
}
if (inter_addr_end > end_address) {
inter_addr_end = end_address;
}
if (inter_addr < start_address) {
inter_addr = start_address;
}
func(inter_addr, inter_addr_end);
}
}
static bool IsRangeGranular(VAddr cpu_addr, size_t size) { static bool IsRangeGranular(VAddr cpu_addr, size_t size) {
return (cpu_addr & ~Core::Memory::PAGE_MASK) == return (cpu_addr & ~Core::Memory::PAGE_MASK) ==
((cpu_addr + size) & ~Core::Memory::PAGE_MASK); ((cpu_addr + size) & ~Core::Memory::PAGE_MASK);
@ -441,12 +471,15 @@ bool BufferCache<P>::DMACopy(GPUVAddr src_address, GPUVAddr dest_address, u64 am
} }
const bool source_dirty = IsRegionGpuModified(*cpu_src_address, amount); const bool source_dirty = IsRegionGpuModified(*cpu_src_address, amount);
const bool dest_dirty = IsRegionGpuModified(*cpu_dest_address, amount); const bool dest_dirty = IsRegionGpuModified(*cpu_dest_address, amount);
if (!(source_dirty || dest_dirty)) { if (!source_dirty && !dest_dirty) {
return false; return false;
} }
const IntervalType subtract_interval{*cpu_dest_address, *cpu_dest_address + amount}; const IntervalType subtract_interval{*cpu_dest_address, *cpu_dest_address + amount};
common_ranges.subtract(subtract_interval); uncommitted_ranges.subtract(subtract_interval);
for (auto& interval_set : committed_ranges) {
interval_set.subtract(subtract_interval);
}
BufferId buffer_a; BufferId buffer_a;
BufferId buffer_b; BufferId buffer_b;
@ -457,46 +490,28 @@ bool BufferCache<P>::DMACopy(GPUVAddr src_address, GPUVAddr dest_address, u64 am
} while (has_deleted_buffers); } while (has_deleted_buffers);
auto& src_buffer = slot_buffers[buffer_a]; auto& src_buffer = slot_buffers[buffer_a];
auto& dest_buffer = slot_buffers[buffer_b]; auto& dest_buffer = slot_buffers[buffer_b];
SynchronizeBuffer(src_buffer, *cpu_src_address, amount); SynchronizeBuffer(src_buffer, *cpu_src_address, static_cast<u32>(amount));
SynchronizeBuffer(dest_buffer, *cpu_dest_address, amount); SynchronizeBuffer(dest_buffer, *cpu_dest_address, static_cast<u32>(amount));
std::array copies{BufferCopy{ std::array copies{BufferCopy{
.src_offset = src_buffer.Offset(*cpu_src_address), .src_offset = src_buffer.Offset(*cpu_src_address),
.dst_offset = dest_buffer.Offset(*cpu_dest_address), .dst_offset = dest_buffer.Offset(*cpu_dest_address),
.size = amount, .size = amount,
}}; }};
auto mirror = [&](VAddr base_address, u64 size) { boost::container::small_vector<IntervalType, 4> tmp_intervals;
VAddr diff = base_address - *cpu_src_address; auto mirror = [&](VAddr base_address, VAddr base_address_end) {
VAddr new_base_address = *cpu_dest_address + diff; const u64 size = base_address_end - base_address;
const VAddr diff = base_address - *cpu_src_address;
const VAddr new_base_address = *cpu_dest_address + diff;
const IntervalType add_interval{new_base_address, new_base_address + size}; const IntervalType add_interval{new_base_address, new_base_address + size};
common_ranges.add(add_interval); uncommitted_ranges.add(add_interval);
tmp_intervals.push_back(add_interval);
}; };
ForEachWrittenRange(*cpu_src_address, amount, mirror);
const VAddr start_address = *cpu_src_address; // This subtraction in this order is important for overlapping copies.
const VAddr end_address = start_address + amount; common_ranges.subtract(subtract_interval);
const IntervalType search_interval{start_address - amount, 1}; for (const IntervalType add_interval : tmp_intervals) {
auto it = common_ranges.lower_bound(search_interval); common_ranges.add(add_interval);
if (it == common_ranges.end()) {
it = common_ranges.begin();
}
while (it != common_ranges.end()) {
VAddr inter_addr_end = it->upper();
VAddr inter_addr = it->lower();
if (inter_addr >= end_address) {
break;
}
if (inter_addr_end <= start_address) {
it++;
continue;
}
if (inter_addr_end > end_address) {
inter_addr_end = end_address;
}
if (inter_addr < start_address) {
inter_addr = start_address;
}
mirror(inter_addr, inter_addr_end - inter_addr);
it++;
} }
runtime.CopyBuffer(dest_buffer, src_buffer, copies); runtime.CopyBuffer(dest_buffer, src_buffer, copies);
@ -695,30 +710,7 @@ void BufferCache<P>::CommitAsyncFlushesHigh() {
const VAddr start_address = buffer_addr + range_offset; const VAddr start_address = buffer_addr + range_offset;
const VAddr end_address = start_address + range_size; const VAddr end_address = start_address + range_size;
const IntervalType search_interval{cpu_addr, 1}; ForEachWrittenRange(start_address, range_size, add_download);
auto it = common_ranges.lower_bound(search_interval);
if (it == common_ranges.end()) {
it = common_ranges.begin();
}
while (it != common_ranges.end()) {
VAddr inter_addr_end = it->upper();
VAddr inter_addr = it->lower();
if (inter_addr >= end_address) {
break;
}
if (inter_addr_end <= start_address) {
it++;
continue;
}
if (inter_addr_end > end_address) {
inter_addr_end = end_address;
}
if (inter_addr < start_address) {
inter_addr = start_address;
}
add_download(inter_addr, inter_addr_end);
it++;
}
const IntervalType subtract_interval{start_address, end_address}; const IntervalType subtract_interval{start_address, end_address};
common_ranges.subtract(subtract_interval); common_ranges.subtract(subtract_interval);
}); });
@ -816,7 +808,9 @@ void BufferCache<P>::BindHostIndexBuffer() {
const u32 size = index_buffer.size; const u32 size = index_buffer.size;
SynchronizeBuffer(buffer, index_buffer.cpu_addr, size); SynchronizeBuffer(buffer, index_buffer.cpu_addr, size);
if constexpr (HAS_FULL_INDEX_AND_PRIMITIVE_SUPPORT) { if constexpr (HAS_FULL_INDEX_AND_PRIMITIVE_SUPPORT) {
runtime.BindIndexBuffer(buffer, offset, size); const u32 new_offset = offset + maxwell3d.regs.index_array.first *
maxwell3d.regs.index_array.FormatSizeInBytes();
runtime.BindIndexBuffer(buffer, new_offset, size);
} else { } else {
runtime.BindIndexBuffer(maxwell3d.regs.draw.topology, maxwell3d.regs.index_array.format, runtime.BindIndexBuffer(maxwell3d.regs.draw.topology, maxwell3d.regs.index_array.format,
maxwell3d.regs.index_array.first, maxwell3d.regs.index_array.count, maxwell3d.regs.index_array.first, maxwell3d.regs.index_array.count,
@ -1429,30 +1423,7 @@ void BufferCache<P>::DownloadBufferMemory(Buffer& buffer, VAddr cpu_addr, u64 si
const VAddr start_address = buffer_addr + range_offset; const VAddr start_address = buffer_addr + range_offset;
const VAddr end_address = start_address + range_size; const VAddr end_address = start_address + range_size;
const IntervalType search_interval{start_address - range_size, 1}; ForEachWrittenRange(start_address, range_size, add_download);
auto it = common_ranges.lower_bound(search_interval);
if (it == common_ranges.end()) {
it = common_ranges.begin();
}
while (it != common_ranges.end()) {
VAddr inter_addr_end = it->upper();
VAddr inter_addr = it->lower();
if (inter_addr >= end_address) {
break;
}
if (inter_addr_end <= start_address) {
it++;
continue;
}
if (inter_addr_end > end_address) {
inter_addr_end = end_address;
}
if (inter_addr < start_address) {
inter_addr = start_address;
}
add_download(inter_addr, inter_addr_end);
it++;
}
const IntervalType subtract_interval{start_address, end_address}; const IntervalType subtract_interval{start_address, end_address};
common_ranges.subtract(subtract_interval); common_ranges.subtract(subtract_interval);
}); });

View File

@ -10,13 +10,15 @@
#include <stop_token> #include <stop_token>
#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/engines/maxwell_dma.h"
#include "video_core/gpu.h" #include "video_core/gpu.h"
#include "video_core/guest_driver.h" #include "video_core/guest_driver.h"
namespace Tegra { namespace Tegra {
class MemoryManager; class MemoryManager;
namespace Engines {
class AccelerateDMAInterface;
} }
} // namespace Tegra
namespace VideoCore { namespace VideoCore {

View File

@ -13,6 +13,7 @@
#include <boost/container/static_vector.hpp> #include <boost/container/static_vector.hpp>
#include "common/common_types.h" #include "common/common_types.h"
#include "video_core/engines/maxwell_dma.h"
#include "video_core/rasterizer_accelerated.h" #include "video_core/rasterizer_accelerated.h"
#include "video_core/rasterizer_interface.h" #include "video_core/rasterizer_interface.h"
#include "video_core/renderer_vulkan/blit_image.h" #include "video_core/renderer_vulkan/blit_image.h"
@ -51,12 +52,12 @@ class StateTracker;
class AccelerateDMA : public Tegra::Engines::AccelerateDMAInterface { class AccelerateDMA : public Tegra::Engines::AccelerateDMAInterface {
public: public:
explicit AccelerateDMA(BufferCache& buffer_cache); explicit AccelerateDMA(BufferCache& buffer_cache);
bool BufferCopy(GPUVAddr start_address, GPUVAddr end_address, u64 amount) override; bool BufferCopy(GPUVAddr start_address, GPUVAddr end_address, u64 amount) override;
private: private:
BufferCache& buffer_cache; BufferCache& buffer_cache;
}; };
class RasterizerVulkan final : public VideoCore::RasterizerAccelerated { class RasterizerVulkan final : public VideoCore::RasterizerAccelerated {