3 files changed, 66 insertions, 17 deletions

diff --git a/src/video_core/engines/maxwell_dma.cpp b/src/video_core/engines/maxwell_dma.cpp
index bf2a21bb6..103cd110e 100644
--- a/src/video_core/engines/maxwell_dma.cpp
+++ b/src/video_core/engines/maxwell_dma.cpp
@@ -4,12 +4,14 @@
 
 #include "core/memory.h"
 #include "video_core/engines/maxwell_dma.h"
+#include "video_core/rasterizer_interface.h"
 #include "video_core/textures/decoders.h"
 
 namespace Tegra {
 namespace Engines {
 
-MaxwellDMA::MaxwellDMA(MemoryManager& memory_manager) : memory_manager(memory_manager) {}
+MaxwellDMA::MaxwellDMA(VideoCore::RasterizerInterface& rasterizer, MemoryManager& memory_manager)
+    : memory_manager(memory_manager), rasterizer{rasterizer} {}
 
 void MaxwellDMA::WriteReg(u32 method, u32 value) {
     ASSERT_MSG(method < Regs::NUM_REGS,
@@ -44,38 +46,79 @@ void MaxwellDMA::HandleCopy() {
     ASSERT(regs.exec.query_mode == Regs::QueryMode::None);
     ASSERT(regs.exec.query_intr == Regs::QueryIntr::None);
     ASSERT(regs.exec.copy_mode == Regs::CopyMode::Unk2);
-    ASSERT(regs.src_params.pos_x == 0);
-    ASSERT(regs.src_params.pos_y == 0);
     ASSERT(regs.dst_params.pos_x == 0);
     ASSERT(regs.dst_params.pos_y == 0);
 
-    if (regs.exec.is_dst_linear == regs.exec.is_src_linear) {
-        std::size_t copy_size = regs.x_count;
+    if (!regs.exec.is_dst_linear && !regs.exec.is_src_linear) {
+        // If both the source and the destination are in block layout, assert.
+        UNREACHABLE_MSG("Tiled->Tiled DMA transfers are not yet implemented");
+        return;
+    }
 
+    if (regs.exec.is_dst_linear && regs.exec.is_src_linear) {
         // When the enable_2d bit is disabled, the copy is performed as if we were copying a 1D
-        // buffer of length `x_count`, otherwise we copy a 2D buffer of size (x_count, y_count).
-        if (regs.exec.enable_2d) {
-            copy_size = copy_size * regs.y_count;
+        // buffer of length `x_count`, otherwise we copy a 2D image of dimensions (x_count,
+        // y_count).
+        if (!regs.exec.enable_2d) {
+            Memory::CopyBlock(dest_cpu, source_cpu, regs.x_count);
+            return;
         }
 
-        Memory::CopyBlock(dest_cpu, source_cpu, copy_size);
+        // If both the source and the destination are in linear layout, perform a line-by-line
+        // copy. We're going to take a subrect of size (x_count, y_count) from the source
+        // rectangle. There is no need to manually flush/invalidate the regions because
+        // CopyBlock does that for us.
+        for (u32 line = 0; line < regs.y_count; ++line) {
+            const VAddr source_line = source_cpu + line * regs.src_pitch;
+            const VAddr dest_line = dest_cpu + line * regs.dst_pitch;
+            Memory::CopyBlock(dest_line, source_line, regs.x_count);
+        }
         return;
     }
 
     ASSERT(regs.exec.enable_2d == 1);
+
+    std::size_t copy_size = regs.x_count * regs.y_count;
+
+    const auto FlushAndInvalidate = [&](u32 src_size, u32 dst_size) {
+        // TODO(Subv): For now, manually flush the regions until we implement GPU-accelerated
+        // copying.
+        rasterizer.FlushRegion(source_cpu, src_size);
+
+        // We have to invalidate the destination region to evict any outdated surfaces from the
+        // cache. We do this before actually writing the new data because the destination address
+        // might contain a dirty surface that will have to be written back to memory.
+        rasterizer.InvalidateRegion(dest_cpu, dst_size);
+    };
+
     u8* src_buffer = Memory::GetPointer(source_cpu);
     u8* dst_buffer = Memory::GetPointer(dest_cpu);
 
     if (regs.exec.is_dst_linear && !regs.exec.is_src_linear) {
+        ASSERT(regs.src_params.size_z == 1);
         // If the input is tiled and the output is linear, deswizzle the input and copy it over.
-        Texture::CopySwizzledData(regs.src_params.size_x, regs.src_params.size_y,
-                                  regs.src_params.size_z, 1, 1, src_buffer, dst_buffer, true,
-                                  regs.src_params.BlockHeight(), regs.src_params.BlockDepth());
+
+        u32 src_bytes_per_pixel = regs.src_pitch / regs.src_params.size_x;
+
+        FlushAndInvalidate(regs.src_pitch * regs.src_params.size_y,
+                           copy_size * src_bytes_per_pixel);
+
+        Texture::UnswizzleSubrect(regs.x_count, regs.y_count, regs.dst_pitch,
+                                  regs.src_params.size_x, src_bytes_per_pixel, source_cpu, dest_cpu,
+                                  regs.src_params.BlockHeight(), regs.src_params.pos_x,
+                                  regs.src_params.pos_y);
     } else {
+        ASSERT(regs.dst_params.size_z == 1);
+        ASSERT(regs.src_pitch == regs.x_count);
+
+        u32 src_bpp = regs.src_pitch / regs.x_count;
+
+        FlushAndInvalidate(regs.src_pitch * regs.y_count,
+                           regs.dst_params.size_x * regs.dst_params.size_y * src_bpp);
+
         // If the input is linear and the output is tiled, swizzle the input and copy it over.
-        Texture::CopySwizzledData(regs.dst_params.size_x, regs.dst_params.size_y,
-                                  regs.dst_params.size_z, 1, 1, dst_buffer, src_buffer, false,
-                                  regs.dst_params.BlockHeight(), regs.dst_params.BlockDepth());
+        Texture::SwizzleSubrect(regs.x_count, regs.y_count, regs.src_pitch, regs.dst_params.size_x,
+                                src_bpp, dest_cpu, source_cpu, regs.dst_params.BlockHeight());
     }
 }
 
diff --git a/src/video_core/engines/maxwell_dma.h b/src/video_core/engines/maxwell_dma.h
index df19e02e2..5f3704f05 100644
--- a/src/video_core/engines/maxwell_dma.h
+++ b/src/video_core/engines/maxwell_dma.h
@@ -12,11 +12,15 @@
 #include "video_core/gpu.h"
 #include "video_core/memory_manager.h"
 
+namespace VideoCore {
+class RasterizerInterface;
+}
+
 namespace Tegra::Engines {
 
 class MaxwellDMA final {
 public:
-    explicit MaxwellDMA(MemoryManager& memory_manager);
+    explicit MaxwellDMA(VideoCore::RasterizerInterface& rasterizer, MemoryManager& memory_manager);
     ~MaxwellDMA() = default;
 
     /// Write the value to the register identified by method.
@@ -133,6 +137,8 @@ public:
     MemoryManager& memory_manager;
 
 private:
+    VideoCore::RasterizerInterface& rasterizer;
+
     /// Performs the copy from the source buffer to the destination buffer as configured in the
     /// registers.
     void HandleCopy();
diff --git a/src/video_core/gpu.cpp b/src/video_core/gpu.cpp
index 2d70a833b..83c7e5b0b 100644
--- a/src/video_core/gpu.cpp
+++ b/src/video_core/gpu.cpp
@@ -27,7 +27,7 @@ GPU::GPU(VideoCore::RasterizerInterface& rasterizer) {
     maxwell_3d = std::make_unique<Engines::Maxwell3D>(rasterizer, *memory_manager);
     fermi_2d = std::make_unique<Engines::Fermi2D>(rasterizer, *memory_manager);
     maxwell_compute = std::make_unique<Engines::MaxwellCompute>();
-    maxwell_dma = std::make_unique<Engines::MaxwellDMA>(*memory_manager);
+    maxwell_dma = std::make_unique<Engines::MaxwellDMA>(rasterizer, *memory_manager);
     kepler_memory = std::make_unique<Engines::KeplerMemory>(rasterizer, *memory_manager);
 }