gpu: nvgpu: optimize barrier in batch pramin writes

Move wmb() before the loop in pramin-accessed batch writes and use
writel_relaxed() directly, instead of calling gk20a_writel() that would
do wmb() on each iteration separately.

Jira DNVGPU-24

Change-Id: I4c1375a819266727f97e2f109d3132b5b0974ac6
Signed-off-by: Konsta Holtta <kholtta@nvidia.com>
Reviewed-on: http://git-master/r/1213600
(cherry picked from commit 79e3e38e0c5384ababfd55b8e6cd9723eb8f7b66)
Reviewed-on: http://git-master/r/1184343
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>

1 files changed, 14 insertions, 2 deletions

diff --git a/drivers/gpu/nvgpu/gk20a/mm_gk20a.c b/drivers/gpu/nvgpu/gk20a/mm_gk20a.c
index 1551dd16..a84d8ff0 100644
--- a/drivers/gpu/nvgpu/gk20a/mm_gk20a.c
+++ b/drivers/gpu/nvgpu/gk20a/mm_gk20a.c
@@ -229,8 +229,14 @@ static inline void pramin_access_batch_wr_n(struct gk20a *g, u32 start,
 {
         u32 r = start, *src_u32 = *arg;
 
+        /*
+         * Barrier moved here from gk20a_writel in the loop. The writes don't
+         * have to be ordered.
+         */
+        wmb();
+
         while (words--) {
-                gk20a_writel(g, r, *src_u32++);
+                writel_relaxed(*src_u32++, g->regs + r);
                 r += sizeof(u32);
         }
 
@@ -242,8 +248,14 @@ static inline void pramin_access_batch_set(struct gk20a *g, u32 start,
 {
         u32 r = start, repeat = **arg;
 
+        /*
+         * Barrier moved here from gk20a_writel in the loop. The writes don't
+         * have to be ordered.
+         */
+        wmb();
+
         while (words--) {
-                gk20a_writel(g, r, repeat);
+                writel_relaxed(repeat, g->regs + r);
                 r += sizeof(u32);
         }
 }