From d5531357487a144cf962ce08a912417fd5e61570 Mon Sep 17 00:00:00 2001 From: Yuri Kunde Schlesner Date: Wed, 21 Jun 2017 22:48:00 -0700 Subject: [PATCH] Memory: Add function to flush a virtual range from the rasterizer cache This is slightly more ergonomic to use, correctly handles virtual regions which are disjoint in physical addressing space, and checks only regions which can be cached by the rasterizer. --- src/core/hle/service/gsp_gpu.cpp | 11 ++-- src/core/hle/service/y2r_u.cpp | 4 +- src/core/memory.cpp | 91 ++++++++++++++++++-------------- src/core/memory.h | 13 +++++ 4 files changed, 72 insertions(+), 47 deletions(-) diff --git a/src/core/hle/service/gsp_gpu.cpp b/src/core/hle/service/gsp_gpu.cpp index bc964ec601..88684b82db 100644 --- a/src/core/hle/service/gsp_gpu.cpp +++ b/src/core/hle/service/gsp_gpu.cpp @@ -475,12 +475,11 @@ static void ExecuteCommand(const Command& command, u32 thread_id) { // TODO: Consider attempting rasterizer-accelerated surface blit if that usage is ever // possible/likely - Memory::RasterizerFlushRegion( - Memory::VirtualToPhysicalAddress(command.dma_request.source_address), - command.dma_request.size); - Memory::RasterizerFlushAndInvalidateRegion( - Memory::VirtualToPhysicalAddress(command.dma_request.dest_address), - command.dma_request.size); + Memory::RasterizerFlushVirtualRegion(command.dma_request.source_address, + command.dma_request.size, Memory::FlushMode::Flush); + Memory::RasterizerFlushVirtualRegion(command.dma_request.dest_address, + command.dma_request.size, + Memory::FlushMode::FlushAndInvalidate); // TODO(Subv): These memory accesses should not go through the application's memory mapping. // They should go through the GSP module's memory mapping. diff --git a/src/core/hle/service/y2r_u.cpp b/src/core/hle/service/y2r_u.cpp index e73971d5fb..57172ddd6d 100644 --- a/src/core/hle/service/y2r_u.cpp +++ b/src/core/hle/service/y2r_u.cpp @@ -587,8 +587,8 @@ static void StartConversion(Interface* self) { // dst_image_size would seem to be perfect for this, but it doesn't include the gap :( u32 total_output_size = conversion.input_lines * (conversion.dst.transfer_unit + conversion.dst.gap); - Memory::RasterizerFlushAndInvalidateRegion( - Memory::VirtualToPhysicalAddress(conversion.dst.address), total_output_size); + Memory::RasterizerFlushVirtualRegion(conversion.dst.address, total_output_size, + Memory::FlushMode::FlushAndInvalidate); HW::Y2R::PerformConversion(conversion); diff --git a/src/core/memory.cpp b/src/core/memory.cpp index 42ca69e006..38fe9e2317 100644 --- a/src/core/memory.cpp +++ b/src/core/memory.cpp @@ -83,19 +83,13 @@ static void MapPages(u32 base, u32 size, u8* memory, PageType type) { LOG_DEBUG(HW_Memory, "Mapping %p onto %08X-%08X", memory, base * PAGE_SIZE, (base + size) * PAGE_SIZE); - u32 end = base + size; + RasterizerFlushVirtualRegion(base << PAGE_BITS, size * PAGE_SIZE, + FlushMode::FlushAndInvalidate); + u32 end = base + size; while (base != end) { ASSERT_MSG(base < PAGE_TABLE_NUM_ENTRIES, "out of range mapping at %08X", base); - // Since pages are unmapped on shutdown after video core is shutdown, the renderer may be - // null here - if (current_page_table->attributes[base] == PageType::RasterizerCachedMemory || - current_page_table->attributes[base] == PageType::RasterizerCachedSpecial) { - RasterizerFlushAndInvalidateRegion(VirtualToPhysicalAddress(base << PAGE_BITS), - PAGE_SIZE); - } - current_page_table->attributes[base] = type; current_page_table->pointers[base] = memory; current_page_table->cached_res_count[base] = 0; @@ -189,7 +183,7 @@ T Read(const VAddr vaddr) { ASSERT_MSG(false, "Mapped memory page without a pointer @ %08X", vaddr); break; case PageType::RasterizerCachedMemory: { - RasterizerFlushRegion(VirtualToPhysicalAddress(vaddr), sizeof(T)); + RasterizerFlushVirtualRegion(vaddr, sizeof(T), FlushMode::Flush); T value; std::memcpy(&value, GetPointerFromVMA(vaddr), sizeof(T)); @@ -198,8 +192,7 @@ T Read(const VAddr vaddr) { case PageType::Special: return ReadMMIO(GetMMIOHandler(vaddr), vaddr); case PageType::RasterizerCachedSpecial: { - RasterizerFlushRegion(VirtualToPhysicalAddress(vaddr), sizeof(T)); - + RasterizerFlushVirtualRegion(vaddr, sizeof(T), FlushMode::Flush); return ReadMMIO(GetMMIOHandler(vaddr), vaddr); } default: @@ -229,8 +222,7 @@ void Write(const VAddr vaddr, const T data) { ASSERT_MSG(false, "Mapped memory page without a pointer @ %08X", vaddr); break; case PageType::RasterizerCachedMemory: { - RasterizerFlushAndInvalidateRegion(VirtualToPhysicalAddress(vaddr), sizeof(T)); - + RasterizerFlushVirtualRegion(vaddr, sizeof(T), FlushMode::FlushAndInvalidate); std::memcpy(GetPointerFromVMA(vaddr), &data, sizeof(T)); break; } @@ -238,8 +230,7 @@ void Write(const VAddr vaddr, const T data) { WriteMMIO(GetMMIOHandler(vaddr), vaddr, data); break; case PageType::RasterizerCachedSpecial: { - RasterizerFlushAndInvalidateRegion(VirtualToPhysicalAddress(vaddr), sizeof(T)); - + RasterizerFlushVirtualRegion(vaddr, sizeof(T), FlushMode::FlushAndInvalidate); WriteMMIO(GetMMIOHandler(vaddr), vaddr, data); break; } @@ -369,11 +360,48 @@ void RasterizerFlushRegion(PAddr start, u32 size) { } void RasterizerFlushAndInvalidateRegion(PAddr start, u32 size) { + // Since pages are unmapped on shutdown after video core is shutdown, the renderer may be + // null here if (VideoCore::g_renderer != nullptr) { VideoCore::g_renderer->Rasterizer()->FlushAndInvalidateRegion(start, size); } } +void RasterizerFlushVirtualRegion(VAddr start, u32 size, FlushMode mode) { + // Since pages are unmapped on shutdown after video core is shutdown, the renderer may be + // null here + if (VideoCore::g_renderer != nullptr) { + VAddr end = start + size; + + auto CheckRegion = [&](VAddr region_start, VAddr region_end) { + if (start >= region_end || end <= region_start) { + // No overlap with region + return; + } + + VAddr overlap_start = std::max(start, region_start); + VAddr overlap_end = std::min(end, region_end); + + PAddr physical_start = TryVirtualToPhysicalAddress(overlap_start).value(); + u32 overlap_size = overlap_end - overlap_start; + + auto* rasterizer = VideoCore::g_renderer->Rasterizer(); + switch (mode) { + case FlushMode::Flush: + rasterizer->FlushRegion(physical_start, overlap_size); + break; + case FlushMode::FlushAndInvalidate: + rasterizer->FlushAndInvalidateRegion(physical_start, overlap_size); + break; + } + }; + + CheckRegion(LINEAR_HEAP_VADDR, LINEAR_HEAP_VADDR_END); + CheckRegion(NEW_LINEAR_HEAP_VADDR, NEW_LINEAR_HEAP_VADDR_END); + CheckRegion(VRAM_VADDR, VRAM_VADDR_END); + } +} + u8 Read8(const VAddr addr) { return Read(addr); } @@ -420,16 +448,13 @@ void ReadBlock(const VAddr src_addr, void* dest_buffer, const size_t size) { break; } case PageType::RasterizerCachedMemory: { - RasterizerFlushRegion(VirtualToPhysicalAddress(current_vaddr), copy_amount); - + RasterizerFlushVirtualRegion(current_vaddr, copy_amount, FlushMode::Flush); std::memcpy(dest_buffer, GetPointerFromVMA(current_vaddr), copy_amount); break; } case PageType::RasterizerCachedSpecial: { DEBUG_ASSERT(GetMMIOHandler(current_vaddr)); - - RasterizerFlushRegion(VirtualToPhysicalAddress(current_vaddr), copy_amount); - + RasterizerFlushVirtualRegion(current_vaddr, copy_amount, FlushMode::Flush); GetMMIOHandler(current_vaddr)->ReadBlock(current_vaddr, dest_buffer, copy_amount); break; } @@ -490,18 +515,13 @@ void WriteBlock(const VAddr dest_addr, const void* src_buffer, const size_t size break; } case PageType::RasterizerCachedMemory: { - RasterizerFlushAndInvalidateRegion(VirtualToPhysicalAddress(current_vaddr), - copy_amount); - + RasterizerFlushVirtualRegion(current_vaddr, copy_amount, FlushMode::FlushAndInvalidate); std::memcpy(GetPointerFromVMA(current_vaddr), src_buffer, copy_amount); break; } case PageType::RasterizerCachedSpecial: { DEBUG_ASSERT(GetMMIOHandler(current_vaddr)); - - RasterizerFlushAndInvalidateRegion(VirtualToPhysicalAddress(current_vaddr), - copy_amount); - + RasterizerFlushVirtualRegion(current_vaddr, copy_amount, FlushMode::FlushAndInvalidate); GetMMIOHandler(current_vaddr)->WriteBlock(current_vaddr, src_buffer, copy_amount); break; } @@ -547,18 +567,13 @@ void ZeroBlock(const VAddr dest_addr, const size_t size) { break; } case PageType::RasterizerCachedMemory: { - RasterizerFlushAndInvalidateRegion(VirtualToPhysicalAddress(current_vaddr), - copy_amount); - + RasterizerFlushVirtualRegion(current_vaddr, copy_amount, FlushMode::FlushAndInvalidate); std::memset(GetPointerFromVMA(current_vaddr), 0, copy_amount); break; } case PageType::RasterizerCachedSpecial: { DEBUG_ASSERT(GetMMIOHandler(current_vaddr)); - - RasterizerFlushAndInvalidateRegion(VirtualToPhysicalAddress(current_vaddr), - copy_amount); - + RasterizerFlushVirtualRegion(current_vaddr, copy_amount, FlushMode::FlushAndInvalidate); GetMMIOHandler(current_vaddr)->WriteBlock(current_vaddr, zeros.data(), copy_amount); break; } @@ -603,15 +618,13 @@ void CopyBlock(VAddr dest_addr, VAddr src_addr, const size_t size) { break; } case PageType::RasterizerCachedMemory: { - RasterizerFlushRegion(VirtualToPhysicalAddress(current_vaddr), copy_amount); - + RasterizerFlushVirtualRegion(current_vaddr, copy_amount, FlushMode::Flush); WriteBlock(dest_addr, GetPointerFromVMA(current_vaddr), copy_amount); break; } case PageType::RasterizerCachedSpecial: { DEBUG_ASSERT(GetMMIOHandler(current_vaddr)); - - RasterizerFlushRegion(VirtualToPhysicalAddress(current_vaddr), copy_amount); + RasterizerFlushVirtualRegion(current_vaddr, copy_amount, FlushMode::Flush); std::vector buffer(copy_amount); GetMMIOHandler(current_vaddr)->ReadBlock(current_vaddr, buffer.data(), buffer.size()); diff --git a/src/core/memory.h b/src/core/memory.h index 96ce9e52e0..c8c56babd9 100644 --- a/src/core/memory.h +++ b/src/core/memory.h @@ -190,6 +190,19 @@ void RasterizerFlushRegion(PAddr start, u32 size); */ void RasterizerFlushAndInvalidateRegion(PAddr start, u32 size); +enum class FlushMode { + /// Write back modified surfaces to RAM + Flush, + /// Write back modified surfaces to RAM, and also remove them from the cache + FlushAndInvalidate, +}; + +/** + * Flushes and invalidates any externally cached rasterizer resources touching the given virtual + * address region. + */ +void RasterizerFlushVirtualRegion(VAddr start, u32 size, FlushMode mode); + /** * Dynarmic has an optimization to memory accesses when the pointer to the page exists that * can be used by setting up the current page table as a callback. This function is used to