Optimized cached write-through

- Allows grabbing an unsynchronized memory block if overwriting contents
anyway
- Allows flushing only specified range of memory

rpcs3/Emu/RSX/GL/GLTextureCache.h

@@ -478,7 +478,7 @@ namespace gl
 			flushed = true;
 
 			const auto valid_range = get_confirmed_range();
-			void *dst = get_raw_ptr(valid_range.first);
+			void *dst = get_raw_ptr(valid_range.first, true);
 
 			glBindBuffer(GL_PIXEL_PACK_BUFFER, pbo_id);
 			void *src = glMapBufferRange(GL_PIXEL_PACK_BUFFER, valid_range.first, valid_range.second, GL_MAP_READ_BIT);
@@ -559,7 +559,7 @@ namespace gl
 				}
 			}
 
-			flush_io();
+			flush_io(valid_range.first, valid_range.second);
 			glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
 			glBindBuffer(GL_PIXEL_PACK_BUFFER, GL_NONE);
 

rpcs3/Emu/RSX/VK/VKTextureCache.h

@@ -245,7 +245,7 @@ namespace vk
 
 			const auto valid_range = get_confirmed_range();
 			void* pixels_src = dma_buffer->map(valid_range.first, valid_range.second);
-			void* pixels_dst = get_raw_ptr(valid_range.first);
+			void* pixels_dst = get_raw_ptr(valid_range.first, true);
 
 			const auto texel_layout = vk::get_format_element_size(vram_texture->info.format);
 			const auto elem_size = texel_layout.first;
@@ -323,7 +323,7 @@ namespace vk
 				}
 			}
 
-			flush_io();
+			flush_io(valid_range.first, valid_range.second);
 			dma_buffer->unmap();
 			reset_write_statistics();
 

rpcs3/Emu/RSX/rsx_cache.h

@@ -38,13 +38,14 @@ namespace rsx
 			if (locked_memory_ptr)
 			{
 				const u32 valid_limit = (confirmed_range.second) ? confirmed_range.first + confirmed_range.second : cpu_address_range;
-				u32* first = locked_memory_ptr.get<u32>(confirmed_range.first);
-				u32* last = locked_memory_ptr.get<u32>(valid_limit - 4);
+				u32* first = locked_memory_ptr.get<u32>(confirmed_range.first, true);
+				u32* last = locked_memory_ptr.get<u32>(valid_limit - 4, true);
 
 				*first = cpu_address_base + confirmed_range.first;
 				*last = cpu_address_base + valid_limit - 4;
 
-				locked_memory_ptr.flush();
+				locked_memory_ptr.flush(confirmed_range.first, 4);
+				locked_memory_ptr.flush(valid_limit - 4, 4);
 			}
 		}
 
@@ -321,10 +322,10 @@ namespace rsx
 		}
 
 		template <typename T = void>
-		T* get_raw_ptr(u32 offset = 0)
+		T* get_raw_ptr(u32 offset = 0, bool no_sync = false)
 		{
 			verify(HERE), locked_memory_ptr;
-			return locked_memory_ptr.get<T>(offset);
+			return locked_memory_ptr.get<T>(offset, no_sync);
 		}
 
 		bool test_memory_head()
@@ -350,9 +351,9 @@ namespace rsx
 			return (*last == (cpu_address_base + valid_limit - 4));
 		}
 
-		void flush_io() const
+		void flush_io(u32 offset = 0, u32 len = 0) const
 		{
-			locked_memory_ptr.flush();
+			locked_memory_ptr.flush(offset, len);
 		}
 
 		std::pair<u32, u32> get_confirmed_range() const

rpcs3/Emu/RSX/rsx_utils.h

@@ -97,7 +97,7 @@ namespace rsx
 		}
 
 		template <typename T = void>
-		T* get(u32 offset = 0)
+		T* get(u32 offset = 0, bool no_sync = false)
 		{
 			if (contiguous)
 			{
@@ -105,7 +105,7 @@ namespace rsx
 			}
 			else
 			{
-				if (!synchronized)
+				if (!synchronized && !no_sync)
 					sync();
 
 				return (T*)(io_cache.data() + offset);
@@ -127,16 +127,70 @@ namespace rsx
 			synchronized = true;
 		}
 
-		void flush() const
+		void flush(u32 offset = 0, u32 len = 0) const
 		{
 			if (contiguous)
 				return;
 
 			u8* src = (u8*)io_cache.data();
-			for (const auto &block : _blocks)
+
+			if (!offset && (!len || len == io_cache.size()))
 			{
-				memcpy(block.first.get(), src, block.second);
-				src += block.second;
+				for (const auto &block : _blocks)
+				{
+					memcpy(block.first.get(), src, block.second);
+					src += block.second;
+				}
+			}
+			else
+			{
+				auto remaining_bytes = len? len : io_cache.size() - offset;
+				const auto write_end = remaining_bytes + offset;
+
+				u32 write_offset;
+				u32 write_length;
+				u32 base_offset = 0;
+
+				for (const auto &block : _blocks)
+				{
+					const u32 block_end = base_offset + block.second;
+
+					if (offset >= base_offset && offset < block_end)
+					{
+						// Head
+						write_offset = (offset - base_offset);
+						write_length = std::min<u32>(block.second - write_offset, remaining_bytes);
+					}
+					else if (base_offset > offset && block_end <= write_end)
+					{
+						// Completely spanned
+						write_offset = 0;
+						write_length = block.second;
+					}
+					else if (base_offset > offset && write_end < block_end)
+					{
+						// Tail
+						write_offset = 0;
+						write_length = remaining_bytes;
+					}
+					else
+					{
+						// No overlap; skip
+						write_length = 0;
+					}
+
+					if (write_length)
+					{
+						memcpy(block.first.get() + write_offset, src + (base_offset + write_offset), write_length);
+
+						verify(HERE), write_length <= remaining_bytes;
+						remaining_bytes -= write_length;
+						if (!remaining_bytes)
+							break;
+					}
+
+					base_offset += block.second;
+				}
 			}
 		}