radeonsi: don't use the CP DMA workaround on Fiji and newer

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>

diff --git a/src/gallium/drivers/radeonsi/si_cp_dma.c b/src/gallium/drivers/radeonsi/si_cp_dma.c
index 0bf85a04db7af4c80bea11cd5c473505f2af0fbd..a5e3d79aa4d9a2fc4590192324050928b9db747e 100644 (file)
--- a/src/gallium/drivers/radeonsi/si_cp_dma.c
+++ b/src/gallium/drivers/radeonsi/si_cp_dma.c
@@ -273,22 +273,26 @@ void si_copy_buffer(struct si_context *sctx,
        dst_offset += r600_resource(dst)->gpu_address;
        src_offset += r600_resource(src)->gpu_address;
 
-       /* If the size is not aligned, we must add a dummy copy at the end
-        * just to align the internal counter. Otherwise, the DMA engine
-        * would slow down by an order of magnitude for following copies.
-        */
-       if (size % CP_DMA_ALIGNMENT)
-               realign_size = CP_DMA_ALIGNMENT - (size % CP_DMA_ALIGNMENT);
-
-       /* If the copy begins unaligned, we must start copying from the next
-        * aligned block and the skipped part should be copied after everything
-        * else has been copied. Only the src alignment matters, not dst.
-        */
-       if (src_offset % CP_DMA_ALIGNMENT) {
-               skipped_size = CP_DMA_ALIGNMENT - (src_offset % CP_DMA_ALIGNMENT);
-               /* The main part will be skipped if the size is too small. */
-               skipped_size = MIN2(skipped_size, size);
-               size -= skipped_size;
+       /* The workarounds aren't needed on Fiji and beyond. */
+       if (sctx->b.family <= CHIP_CARRIZO ||
+           sctx->b.family == CHIP_STONEY) {
+               /* If the size is not aligned, we must add a dummy copy at the end
+                * just to align the internal counter. Otherwise, the DMA engine
+                * would slow down by an order of magnitude for following copies.
+                */
+               if (size % CP_DMA_ALIGNMENT)
+                       realign_size = CP_DMA_ALIGNMENT - (size % CP_DMA_ALIGNMENT);
+
+               /* If the copy begins unaligned, we must start copying from the next
+                * aligned block and the skipped part should be copied after everything
+                * else has been copied. Only the src alignment matters, not dst.
+                */
+               if (src_offset % CP_DMA_ALIGNMENT) {
+                       skipped_size = CP_DMA_ALIGNMENT - (src_offset % CP_DMA_ALIGNMENT);
+                       /* The main part will be skipped if the size is too small. */
+                       skipped_size = MIN2(skipped_size, size);
+                       size -= skipped_size;
+               }
        }
 
        /* Flush the caches. */