nil/copy.rs: Add host copy support for Fermi-Volta

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35756>

src/nouveau/nil/copy.rs

@@ -121,6 +121,50 @@ trait CopyBytes {
     }
 }
 
+/// Implements copies for [`GOBType::FermiColor`]
+struct CopyGOBFermi<C: CopyBytes> {
+    phantom: std::marker::PhantomData<C>,
+}
+
+impl<C: CopyBytes> CopyGOBLines for CopyGOBFermi<C> {
+    const GOB_EXTENT_B: Extent4D<units::Bytes> = Extent4D::new(64, 8, 1, 1);
+    const LINE_WIDTH_B: u32 = 16;
+    const X_DIVISOR: u32 = C::X_DIVISOR;
+
+    unsafe fn copy(tiled: *mut u8, linear: *mut u8, bytes: usize) {
+        C::copy(tiled, linear, bytes);
+    }
+
+    unsafe fn copy_whole_line(tiled: *mut u8, linear: *mut u8) {
+        C::copy_16b(tiled as *mut _, linear as *mut _);
+    }
+
+    #[inline(always)]
+    fn for_each_gob_line(mut f: impl FnMut(u32, u32, u32, u32)) {
+        for i in 0..2 {
+            f(i * 0x100 + 0x00, i * 32 + 0, 2, 0);
+            f(i * 0x100 + 0x10, i * 32 + 0, 3, 0);
+            f(i * 0x100 + 0x20, i * 32 + 0, 0, 0);
+            f(i * 0x100 + 0x30, i * 32 + 0, 1, 0);
+
+            f(i * 0x100 + 0x40, i * 32 + 16, 0, 0);
+            f(i * 0x100 + 0x50, i * 32 + 16, 1, 0);
+            f(i * 0x100 + 0x60, i * 32 + 16, 2, 0);
+            f(i * 0x100 + 0x70, i * 32 + 16, 3, 0);
+
+            f(i * 0x100 + 0xc0, i * 32 + 0, 6, 0);
+            f(i * 0x100 + 0xd0, i * 32 + 0, 7, 0);
+            f(i * 0x100 + 0xe0, i * 32 + 0, 4, 0);
+            f(i * 0x100 + 0xf0, i * 32 + 0, 5, 0);
+
+            f(i * 0x100 + 0x80, i * 32 + 16, 4, 0);
+            f(i * 0x100 + 0x90, i * 32 + 16, 5, 0);
+            f(i * 0x100 + 0xa0, i * 32 + 16, 6, 0);
+            f(i * 0x100 + 0xb0, i * 32 + 16, 7, 0);
+        }
+    }
+}
+
 /// Implements copies for [`GOBType::TuringColor2D`]
 struct CopyGOBTuring2D<C: CopyBytes> {
     phantom: std::marker::PhantomData<C>,
@@ -597,6 +641,16 @@ pub unsafe extern "C" fn nil_copy_linear_to_tiled(
                 end_B,
             );
         }
+        GOBType::FermiColor => {
+            copy_tiled::<CopyGOBFermi<RawCopyToTiled>>(
+                *tiling,
+                level_extent_B,
+                tiled_dst,
+                linear_pointer,
+                offset_B,
+                end_B,
+            );
+        }
         _ => panic!("Unsupported GOB type"),
     }
 }
@@ -654,6 +708,16 @@ pub unsafe extern "C" fn nil_copy_tiled_to_linear(
                 end_B,
             );
         }
+        GOBType::FermiColor => {
+            copy_tiled::<CopyGOBFermi<RawCopyToLinear>>(
+                *tiling,
+                level_extent_B,
+                tiled_src,
+                linear_pointer,
+                offset_B,
+                end_B,
+            );
+        }
         _ => panic!("Unsupported GOB type"),
     }
 }

src/nouveau/nil/tiling.rs

@@ -22,6 +22,21 @@ pub enum GOBType {
     FermiZS,
 
     /// The Fermi GOB format for color images
+    ///
+    /// A `FermiColor` GOB is 512 bytes, arranged in a 64x8 layout and split
+    /// into Sectors. Each Sector is 32 Bytes, and the Sectors in a GOB are
+    /// arranged in a 16x2 layout (i.e., two 16B lines on top of each other).
+    /// It's then arranged into two columns that are 2 sectors by 4, leading to
+    /// a 4x4 grid of sectors:
+    ///
+    /// |           |           |           |           |
+    /// |-----------|-----------|-----------|-----------|
+    /// | Sector  1 | Sector  2 | Sector  9 | Sector 10 |
+    /// | Sector  0 | Sector  3 | Sector  8 | Sector 11 |
+    /// | Sector  5 | Sector  6 | Sector 13 | Sector 14 |
+    /// | Sector  4 | Sector  7 | Sector 12 | Sector 15 |
+    ///
+    /// `CopyGOBFermi` implements CPU copies for Fermi color GOBs.
     FermiColor,
 
     /// The Turing 2D GOB format for color images