radv: do not sync CP DMA when copying buffers

We already track if the DMA engine is busy/idle with a flag,
and we emit a packet that waits for all CP DMA operations
to be complete. This is done at end of command buffer because
the kernel doesn't wait for them, and also when emitting
barriers, so it should be safe.

This improves small copies for both aligned and unaligned sizes.

Aligned sizes:
BEFORE:
1 KB: 59.840000 ms
2 KB: 71.200000 ms
AFTER:
1 KB: 31.200000 ms
2 KB: 31.040000 ms

Unaligned sizes:
BEFORE:
2 KB: 68.3200 ms
3 KB: 79.3600 ms
5 KB: 76.6400 ms
9 KB: 90.8800 ms
17 KB: 116.0000 ms
AFTER:
2 KB: 31.0400 ms
3 KB: 32.0000 ms
5 KB: 30.8800 ms
9 KB: 30.5600 ms
17 KB: 29.6000 ms

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

diff --git a/src/amd/vulkan/si_cmd_buffer.c b/src/amd/vulkan/si_cmd_buffer.c
index e5bc151f64bb5e71e87f9552a1240d09bb803ada..e0d474756a32059aa8364eea0ee1e44af150520e 100644 (file)
--- a/src/amd/vulkan/si_cmd_buffer.c
+++ b/src/amd/vulkan/si_cmd_buffer.c
@@ -1215,6 +1215,8 @@ void si_cp_dma_buffer_copy(struct radv_cmd_buffer *cmd_buffer,
                                  size + skipped_size + realign_size,
                                  &dma_flags);
 
+               dma_flags &= ~CP_DMA_SYNC;
+
                si_emit_cp_dma(cmd_buffer, main_dest_va, main_src_va,
                               byte_count, dma_flags);