panfrost: Do fine-grained flushing when preparing BO for CPU accesses

We don't have to flush all batches when we're only interested in
reading/writing a specific BO. Thanks to the
panfrost_flush_batches_accessing_bo() and panfrost_bo_wait() helpers
we can now flush only the batches touching the BO we want to access
from the CPU.

This fixes the dEQP-GLES2.functional.fbo.render.texsubimage.* tests.

Signed-off-by: Boris Brezillon <[email protected]>
Reviewed-by: Alyssa Rosenzweig <[email protected]>

2 files changed, 12 insertions, 19 deletions

diff --git a/src/gallium/drivers/panfrost/ci/expected-failures.txt b/src/gallium/drivers/panfrost/ci/expected-failures.txt
index 0cadaa20cdb..7e7dbd62307 100644
--- a/src/gallium/drivers/panfrost/ci/expected-failures.txt
+++ b/src/gallium/drivers/panfrost/ci/expected-failures.txt
@@ -51,10 +51,6 @@ dEQP-GLES2.functional.fbo.render.shared_colorbuffer.tex2d_rgb_depth_component16
 dEQP-GLES2.functional.fbo.render.shared_depthbuffer.rbo_rgb565_depth_component16 Fail
 dEQP-GLES2.functional.fbo.render.shared_depthbuffer.tex2d_rgba_depth_component16 Fail
 dEQP-GLES2.functional.fbo.render.shared_depthbuffer.tex2d_rgb_depth_component16 Fail
-dEQP-GLES2.functional.fbo.render.texsubimage.after_render_tex2d_rgba Fail
-dEQP-GLES2.functional.fbo.render.texsubimage.after_render_tex2d_rgb Fail
-dEQP-GLES2.functional.fbo.render.texsubimage.between_render_tex2d_rgba Fail
-dEQP-GLES2.functional.fbo.render.texsubimage.between_render_tex2d_rgb Fail
 dEQP-GLES2.functional.fragment_ops.depth_stencil.random.0 Fail
 dEQP-GLES2.functional.fragment_ops.depth_stencil.random.10 Fail
 dEQP-GLES2.functional.fragment_ops.depth_stencil.random.11 Fail
diff --git a/src/gallium/drivers/panfrost/pan_resource.c b/src/gallium/drivers/panfrost/pan_resource.c
index 1f7605adcd5..49088618ae4 100644
--- a/src/gallium/drivers/panfrost/pan_resource.c
+++ b/src/gallium/drivers/panfrost/pan_resource.c
@@ -578,10 +578,8 @@ panfrost_transfer_map(struct pipe_context *pctx,
                         is_bound |= fb->cbufs[c]->texture == resource;
         }
 
-        if (is_bound && (usage & PIPE_TRANSFER_READ)) {
-                assert(level == 0);
-                panfrost_flush_all_batches(ctx, true);
-        }
+        if (is_bound && (usage & PIPE_TRANSFER_READ))
+                 assert(level == 0);
 
         /* TODO: Respect usage flags */
 
@@ -594,11 +592,11 @@ panfrost_transfer_map(struct pipe_context *pctx,
                 /* No flush for writes to uninitialized */
         } else if (!(usage & PIPE_TRANSFER_UNSYNCHRONIZED)) {
                 if (usage & PIPE_TRANSFER_WRITE) {
-                        /* STUB: flush reading */
-                        //printf("debug: missed reading flush %d\n", resource->target);
+                        panfrost_flush_batches_accessing_bo(ctx, bo, PAN_BO_ACCESS_RW);
+                        panfrost_bo_wait(bo, INT64_MAX, PAN_BO_ACCESS_RW);
                 } else if (usage & PIPE_TRANSFER_READ) {
-                        /* STUB: flush writing */
-                        //printf("debug: missed writing flush %d (%d-%d)\n", resource->target, box->x, box->x + box->width);
+                        panfrost_flush_batches_accessing_bo(ctx, bo, PAN_BO_ACCESS_WRITE);
+                        panfrost_bo_wait(bo, INT64_MAX, PAN_BO_ACCESS_WRITE);
                 } else {
                         /* Why are you even mapping?! */
                 }
@@ -748,11 +746,8 @@ panfrost_generate_mipmap(
          * reorder-type optimizations in place. But for now prioritize
          * correctness. */
 
-        struct panfrost_batch *batch = panfrost_get_batch_for_fbo(ctx);
-        bool has_draws = batch->last_job.gpu;
-
-        if (has_draws)
-                panfrost_flush_all_batches(ctx, true);
+        panfrost_flush_batches_accessing_bo(ctx, rsrc->bo, PAN_BO_ACCESS_RW);
+        panfrost_bo_wait(rsrc->bo, INT64_MAX, PAN_BO_ACCESS_RW);
 
         /* We've flushed the original buffer if needed, now trigger a blit */
 
@@ -765,8 +760,10 @@ panfrost_generate_mipmap(
         /* If the blit was successful, flush once more. If it wasn't, well, let
          * the state tracker deal with it. */
 
-        if (blit_res)
-                panfrost_flush_all_batches(ctx, true);
+        if (blit_res) {
+                panfrost_flush_batches_accessing_bo(ctx, rsrc->bo, PAN_BO_ACCESS_WRITE);
+                panfrost_bo_wait(rsrc->bo, INT64_MAX, PAN_BO_ACCESS_WRITE);
+        }
 
         return blit_res;
 }