radv: Fix dword alignment in SDMA buffer copy.

Also add a comment that explains the dword aligned mode.

Note that the SDMA shader uploads are always dword aligned
so this commit doesn't fix any issues but just prepares this
function for more general use.

Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Tatsuyuki Ishi <ishitatsuyuki@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22551>

diff --git a/src/amd/vulkan/radv_sdma_copy_image.c b/src/amd/vulkan/radv_sdma_copy_image.c
index 20cc2c0..ae92ede 100644 (file)
--- a/src/amd/vulkan/radv_sdma_copy_image.c
+++ b/src/amd/vulkan/radv_sdma_copy_image.c
@@ -159,8 +159,14 @@ radv_sdma_copy_buffer(struct radv_device *device, struct radeon_cmdbuf *cs, uint
 
    assert(gfx_level >= GFX7);
 
-   /* Align copy size to dw if src/dst address are dw aligned */
-   if ((src_va & 0x3) == 0 && (src_va & 0x3) == 0 && size > 4 && (size & 3) != 0) {
+   /* SDMA FW automatically enables a faster dword copy mode when
+    * source, destination and size are all dword-aligned.
+    *
+    * When source and destination are dword-aligned, round down the size to
+    * take advantage of faster dword copy, and copy the remaining few bytes
+    * with the last copy packet.
+    */
+   if ((src_va & 0x3) == 0 && (dst_va & 0x3) == 0 && size > 4 && (size & 0x3) != 0) {
       align = ~0x3u;
       ncopy++;
    }