MaxwellDMA: Optimize micro copies.

src/video_core/engines/maxwell_dma.cpp

@@ -90,7 +90,47 @@ void MaxwellDMA::HandleCopy() {
     ASSERT(regs.exec.enable_2d == 1);
 
     if (regs.exec.is_dst_linear && !regs.exec.is_src_linear) {
+
         ASSERT(regs.src_params.BlockDepth() == 0);
+        // Optimized path for micro copies.
+        if (regs.dst_pitch * regs.y_count < Texture::GetGOBSize() && regs.dst_pitch <= 64) {
+            const u32 bytes_per_pixel = regs.dst_pitch / regs.x_count;
+            const std::size_t src_size = Texture::GetGOBSize();
+            const std::size_t dst_size = regs.dst_pitch * regs.y_count;
+            u32 pos_x = regs.src_params.pos_x;
+            u32 pos_y = regs.src_params.pos_y;
+            const u64 offset =
+                Texture::GetGOBOffset(regs.src_params.size_x, regs.src_params.size_y, pos_x, pos_y,
+                                      regs.src_params.BlockDepth(), bytes_per_pixel);
+            const u32 x_in_gob = 64 / bytes_per_pixel;
+            pos_x = pos_x % x_in_gob;
+            pos_y = pos_y % 8;
+
+            if (read_buffer.size() < src_size) {
+                read_buffer.resize(src_size);
+            }
+
+            if (write_buffer.size() < dst_size) {
+                write_buffer.resize(dst_size);
+            }
+
+            if (Settings::IsGPULevelExtreme()) {
+                memory_manager.ReadBlock(source + offset, read_buffer.data(), src_size);
+                memory_manager.ReadBlock(dest, write_buffer.data(), dst_size);
+            } else {
+                memory_manager.ReadBlockUnsafe(source + offset, read_buffer.data(), src_size);
+                memory_manager.ReadBlockUnsafe(dest, write_buffer.data(), dst_size);
+            }
+
+            Texture::UnswizzleSubrect(regs.x_count, regs.y_count, regs.dst_pitch,
+                                      regs.src_params.size_x, bytes_per_pixel, read_buffer.data(),
+                                      write_buffer.data(), regs.src_params.BlockHeight(), pos_x,
+                                      pos_y);
+
+            memory_manager.WriteBlock(dest, write_buffer.data(), dst_size);
+
+            return;
+        }
         // If the input is tiled and the output is linear, deswizzle the input and copy it over.
         const u32 bytes_per_pixel = regs.dst_pitch / regs.x_count;
         const std::size_t src_size = Texture::CalculateSize(

src/video_core/textures/decoders.cpp

@@ -382,4 +382,18 @@ std::size_t CalculateSize(bool tiled, u32 bytes_per_pixel, u32 width, u32 height
     }
 }
 
+u64 GetGOBOffset(u32 width, u32 height, u32 dst_x, u32 dst_y, u32 block_height,
+                 u32 bytes_per_pixel) {
+    auto div_ceil = [](const u32 x, const u32 y) { return ((x + y - 1) / y); };
+    const u32 gobs_in_block = 1 << block_height;
+    const u32 y_blocks = gob_size_y << block_height;
+    const u32 x_per_gob = gob_size_x / bytes_per_pixel;
+    const u32 x_blocks = div_ceil(width, x_per_gob);
+    const u32 block_size = gob_size * gobs_in_block;
+    const u32 stride = block_size * x_blocks;
+    const u32 base = (dst_y / y_blocks) * stride + (dst_x / x_per_gob) * block_size;
+    const u32 relative_y = dst_y % y_blocks;
+    return base + (relative_y / gob_size_y) * gob_size;
+}
+
 } // namespace Tegra::Texture

src/video_core/textures/decoders.h

@@ -59,4 +59,7 @@ void UnswizzleSubrect(u32 subrect_width, u32 subrect_height, u32 dest_pitch, u32
 void SwizzleKepler(u32 width, u32 height, u32 dst_x, u32 dst_y, u32 block_height,
                    std::size_t copy_size, const u8* source_data, u8* swizzle_data);
 
+u64 GetGOBOffset(u32 width, u32 height, u32 dst_x, u32 dst_y, u32 block_height,
+                 u32 bytes_per_pixel);
+
 } // namespace Tegra::Texture