4 files changed, 35 insertions, 3 deletions

diff --git a/src/core/hle/service/gsp_gpu.cpp b/src/core/hle/service/gsp_gpu.cpp
index c11c5faba..c56475ae4 100644
--- a/src/core/hle/service/gsp_gpu.cpp
+++ b/src/core/hle/service/gsp_gpu.cpp
@@ -15,6 +15,7 @@
 #include "core/hw/lcd.h"
 
 #include "video_core/gpu_debugger.h"
+#include "video_core/video_core.h"
 
 // Main graphics debugger object - TODO: Here is probably not the best place for this
 GraphicsDebugger g_debugger;
@@ -264,6 +265,8 @@ static void FlushDataCache(Service::Interface* self) {
     u32 size    = cmd_buff[2];
     u32 process = cmd_buff[4];
 
+    VideoCore::g_renderer->hw_rasterizer->NotifyFlush(Memory::VirtualToPhysicalAddress(address), size);
+
     // TODO(purpasmart96): Verify return header on HW
 
     cmd_buff[1] = RESULT_SUCCESS.raw; // No error
@@ -352,10 +355,16 @@ static void ExecuteCommand(const Command& command, u32 thread_id) {
 
     // GX request DMA - typically used for copying memory from GSP heap to VRAM
     case CommandId::REQUEST_DMA:
+        VideoCore::g_renderer->hw_rasterizer->NotifyPreRead(Memory::VirtualToPhysicalAddress(command.dma_request.source_address),
+                                                            command.dma_request.size);
+
         memcpy(Memory::GetPointer(command.dma_request.dest_address),
                Memory::GetPointer(command.dma_request.source_address),
                command.dma_request.size);
         SignalInterrupt(InterruptId::DMA);
+
+        VideoCore::g_renderer->hw_rasterizer->NotifyFlush(Memory::VirtualToPhysicalAddress(command.dma_request.dest_address),
+                                                          command.dma_request.size);
         break;
 
     // ctrulib homebrew sends all relevant command list data with this command,
diff --git a/src/core/hle/service/y2r_u.cpp b/src/core/hle/service/y2r_u.cpp
index ce822e990..15987e028 100644
--- a/src/core/hle/service/y2r_u.cpp
+++ b/src/core/hle/service/y2r_u.cpp
@@ -9,7 +9,11 @@
 #include "core/hle/hle.h"
 #include "core/hle/kernel/event.h"
 #include "core/hle/service/y2r_u.h"
+#include "core/mem_map.h"
+#include "core/memory.h"
+
 #include "video_core/utils.h"
+#include "video_core/video_core.h"
 
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // Namespace Y2R_U
@@ -260,6 +264,13 @@ static void StartConversion(Service::Interface* self) {
         break;
     }
     }
+
+    // dst_image_size would seem to be perfect for this, but it doesn't include the stride :(
+    u32 total_output_size = conversion_params.input_lines *
+        (conversion_params.dst_transfer_unit + conversion_params.dst_stride);
+    VideoCore::g_renderer->hw_rasterizer->NotifyFlush(
+        Memory::VirtualToPhysicalAddress(conversion_params.dst_address), total_output_size);
+
     LOG_DEBUG(Service_Y2R, "called");
     completion_event->Signal();
 
diff --git a/src/core/hw/gpu.cpp b/src/core/hw/gpu.cpp
index 8ef1f70df..ddc5d647e 100644
--- a/src/core/hw/gpu.cpp
+++ b/src/core/hw/gpu.cpp
@@ -106,6 +106,8 @@ inline void Write(u32 addr, const T data) {
             } else {
                 GSP_GPU::SignalInterrupt(GSP_GPU::InterruptId::PSC1);
             }
+
+            VideoCore::g_renderer->hw_rasterizer->NotifyFlush(config.GetStartAddress(), config.GetEndAddress() - config.GetStartAddress());
         }
         break;
     }
@@ -129,19 +131,25 @@ inline void Write(u32 addr, const T data) {
             u32 output_width = config.output_width / horizontal_scale;
             u32 output_height = config.output_height / vertical_scale;
 
+            u32 input_size = config.input_width * config.input_height * GPU::Regs::BytesPerPixel(config.input_format);
+            u32 output_size = output_width * output_height * GPU::Regs::BytesPerPixel(config.output_format);
+
+            VideoCore::g_renderer->hw_rasterizer->NotifyPreRead(config.GetPhysicalInputAddress(), input_size);
+
             if (config.raw_copy) {
                 // Raw copies do not perform color conversion nor tiled->linear / linear->tiled conversions
                 // TODO(Subv): Verify if raw copies perform scaling
-                memcpy(dst_pointer, src_pointer, config.output_width * config.output_height * 
-                        GPU::Regs::BytesPerPixel(config.output_format));
+                memcpy(dst_pointer, src_pointer, output_size);
                 
                 LOG_TRACE(HW_GPU, "DisplayTriggerTransfer: 0x%08x bytes from 0x%08x(%ux%u)-> 0x%08x(%ux%u), output format: %x, flags 0x%08X, Raw copy",
-                    config.output_height * output_width * GPU::Regs::BytesPerPixel(config.output_format),
+                    output_size,
                     config.GetPhysicalInputAddress(), config.input_width.Value(), config.input_height.Value(),
                     config.GetPhysicalOutputAddress(), config.output_width.Value(), config.output_height.Value(),
                     config.output_format.Value(), config.flags);
 
                 GSP_GPU::SignalInterrupt(GSP_GPU::InterruptId::PPF);
+
+                VideoCore::g_renderer->hw_rasterizer->NotifyFlush(config.GetPhysicalOutputAddress(), output_size);
                 break;
             }
 
@@ -247,6 +255,8 @@ inline void Write(u32 addr, const T data) {
                       config.output_format.Value(), config.flags);
 
             GSP_GPU::SignalInterrupt(GSP_GPU::InterruptId::PPF);
+
+            VideoCore::g_renderer->hw_rasterizer->NotifyFlush(config.GetPhysicalOutputAddress(), output_size);
         }
         break;
     }
diff --git a/src/core/settings.h b/src/core/settings.h
index 0f4700241..54c1023b8 100644
--- a/src/core/settings.h
+++ b/src/core/settings.h
@@ -45,6 +45,8 @@ struct Values {
     int region_value;
 
     // Renderer
+    bool use_hw_renderer;
+
     float bg_red;
     float bg_green;
     float bg_blue;