panfrost: Allocate polygon lists on-demand

Rather than alloacting a huge (64MB) polygon list on context creation
and sharing it across framebuffers, we instead allocate polygon lists as
BOs (which consistently hit the cache) sized appropriately; for about a
month, we've known how to calculate the polygon list size so this has
only recently become possible.

The good news is we can render to truly massive framebuffers without
crashing and, more importantly, we eliminate the 64MB upfront overhead.
If a list that size isn't actually needed, it's not allocated.

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

6 files changed, 36 insertions, 10 deletions

diff --git a/src/gallium/drivers/panfrost/pan_context.c b/src/gallium/drivers/panfrost/pan_context.c
index d261a284212..e9d18605dd0 100644
--- a/src/gallium/drivers/panfrost/pan_context.c
+++ b/src/gallium/drivers/panfrost/pan_context.c
@@ -62,6 +62,7 @@ panfrost_emit_midg_tiler(
         unsigned vertex_count)
 {
         struct midgard_tiler_descriptor t = {};
+        struct panfrost_job *batch = panfrost_get_job_for_fbo(ctx);
 
         t.hierarchy_mask =
                 panfrost_choose_hierarchy_mask(width, height, vertex_count);
@@ -77,10 +78,7 @@ panfrost_emit_midg_tiler(
         /* Sanity check */
 
         if (t.hierarchy_mask) {
-                assert(ctx->tiler_polygon_list.bo->size >= (header_size + body_size));
-
-                /* Specify allocated tiler structures */
-                t.polygon_list = ctx->tiler_polygon_list.bo->gpu;
+                t.polygon_list = panfrost_job_get_polygon_list(batch, header_size + body_size);
 
                 /* Allow the entire tiler heap */
                 t.heap_start = ctx->tiler_heap.bo->gpu;
@@ -2532,7 +2530,6 @@ panfrost_destroy(struct pipe_context *pipe)
         panfrost_drm_free_slab(screen, &panfrost->scratchpad);
         panfrost_drm_free_slab(screen, &panfrost->shaders);
         panfrost_drm_free_slab(screen, &panfrost->tiler_heap);
-        panfrost_drm_free_slab(screen, &panfrost->tiler_polygon_list);
         panfrost_drm_free_slab(screen, &panfrost->tiler_dummy);
 
         ralloc_free(pipe);
@@ -2678,7 +2675,6 @@ panfrost_setup_hardware(struct panfrost_context *ctx)
         panfrost_drm_allocate_slab(screen, &ctx->scratchpad, 64*4, false, 0, 0, 0);
         panfrost_drm_allocate_slab(screen, &ctx->shaders, 4096, true, PAN_ALLOCATE_EXECUTE, 0, 0);
         panfrost_drm_allocate_slab(screen, &ctx->tiler_heap, 4096, false, PAN_ALLOCATE_INVISIBLE | PAN_ALLOCATE_GROWABLE, 1, 128);
-        panfrost_drm_allocate_slab(screen, &ctx->tiler_polygon_list, 128*128, false, PAN_ALLOCATE_INVISIBLE | PAN_ALLOCATE_GROWABLE, 1, 128);
         panfrost_drm_allocate_slab(screen, &ctx->tiler_dummy, 1, false, PAN_ALLOCATE_INVISIBLE, 0, 0);
 }
 
diff --git a/src/gallium/drivers/panfrost/pan_context.h b/src/gallium/drivers/panfrost/pan_context.h
index d8dbd4d66f6..e8d2417e0cb 100644
--- a/src/gallium/drivers/panfrost/pan_context.h
+++ b/src/gallium/drivers/panfrost/pan_context.h
@@ -111,7 +111,6 @@ struct panfrost_context {
         struct panfrost_memory shaders;
         struct panfrost_memory scratchpad;
         struct panfrost_memory tiler_heap;
-        struct panfrost_memory tiler_polygon_list;
         struct panfrost_memory tiler_dummy;
         struct panfrost_memory depth_stencil_buffer;
 
diff --git a/src/gallium/drivers/panfrost/pan_drm.c b/src/gallium/drivers/panfrost/pan_drm.c
index 89c7019dd9c..42cf1750334 100644
--- a/src/gallium/drivers/panfrost/pan_drm.c
+++ b/src/gallium/drivers/panfrost/pan_drm.c
@@ -288,7 +288,7 @@ panfrost_drm_submit_vs_fs_job(struct panfrost_context *ctx, bool has_draws, bool
         panfrost_job_add_bo(job, ctx->shaders.bo);
         panfrost_job_add_bo(job, ctx->scratchpad.bo);
         panfrost_job_add_bo(job, ctx->tiler_heap.bo);
-        panfrost_job_add_bo(job, ctx->tiler_polygon_list.bo);
+        panfrost_job_add_bo(job, job->polygon_list);
 
         if (job->first_job.gpu) {
                 ret = panfrost_drm_submit_job(ctx, job->first_job.gpu, 0);
diff --git a/src/gallium/drivers/panfrost/pan_job.c b/src/gallium/drivers/panfrost/pan_job.c
index d75af0c5330..661f8ae154e 100644
--- a/src/gallium/drivers/panfrost/pan_job.c
+++ b/src/gallium/drivers/panfrost/pan_job.c
@@ -72,6 +72,9 @@ panfrost_free_job(struct panfrost_context *ctx, struct panfrost_job *job)
                 BITSET_SET(screen->free_transient, *index);
         }
 
+        /* Unreference the polygon list */
+        panfrost_bo_unreference(ctx->base.screen, job->polygon_list);
+
         _mesa_hash_table_remove_key(ctx->jobs, &job->key);
 
         if (ctx->job == job)
@@ -160,6 +163,27 @@ panfrost_job_add_bo(struct panfrost_job *job, struct panfrost_bo *bo)
         _mesa_set_add(job->bos, bo);
 }
 
+/* Returns the polygon list's GPU address if available, or otherwise allocates
+ * the polygon list.  It's perfectly fast to use allocate/free BO directly,
+ * since we'll hit the BO cache and this is one-per-batch anyway. */
+
+mali_ptr
+panfrost_job_get_polygon_list(struct panfrost_job *batch, unsigned size)
+{
+        if (batch->polygon_list) {
+                assert(batch->polygon_list->size >= size);
+        } else {
+                struct panfrost_screen *screen = pan_screen(batch->ctx->base.screen);
+
+                /* Create the BO as invisible, as there's no reason to map */
+
+                batch->polygon_list = panfrost_drm_create_bo(screen,
+                                size, PAN_ALLOCATE_INVISIBLE);
+        }
+
+        return batch->polygon_list->gpu;
+}
+
 void
 panfrost_flush_jobs_writing_resource(struct panfrost_context *panfrost,
                                      struct pipe_resource *prsc)
diff --git a/src/gallium/drivers/panfrost/pan_job.h b/src/gallium/drivers/panfrost/pan_job.h
index c6ae2a4eb9f..a5ea5bf41cc 100644
--- a/src/gallium/drivers/panfrost/pan_job.h
+++ b/src/gallium/drivers/panfrost/pan_job.h
@@ -112,6 +112,9 @@ struct panfrost_job {
 
         /* Within the topmost transient BO, how much has been used? */
         unsigned transient_offset;
+
+        /* Polygon list bound to the batch, or NULL if none bound yet */
+        struct panfrost_bo *polygon_list;
 };
 
 /* Functions for managing the above */
@@ -150,6 +153,9 @@ void
 panfrost_job_set_requirements(struct panfrost_context *ctx,
                               struct panfrost_job *job);
 
+mali_ptr
+panfrost_job_get_polygon_list(struct panfrost_job *batch, unsigned size);
+
 void
 panfrost_job_clear(struct panfrost_context *ctx,
                    struct panfrost_job *job,
diff --git a/src/gallium/drivers/panfrost/pan_scoreboard.c b/src/gallium/drivers/panfrost/pan_scoreboard.c
index 1d15ca80338..bae9dfb3c25 100644
--- a/src/gallium/drivers/panfrost/pan_scoreboard.c
+++ b/src/gallium/drivers/panfrost/pan_scoreboard.c
@@ -303,10 +303,11 @@ panfrost_scoreboard_set_value(struct panfrost_job *batch)
         if (!batch->last_tiler.gpu)
                 return;
 
-        /* Okay, we do. Let's generate it */
+        /* Okay, we do. Let's generate it. We'll need the job's polygon list
+         * regardless of size. */
 
         struct panfrost_context *ctx = batch->ctx;
-        mali_ptr polygon_list = ctx->tiler_polygon_list.bo->gpu;
+        mali_ptr polygon_list = panfrost_job_get_polygon_list(batch, 0);
 
         struct panfrost_transfer job =
                 panfrost_set_value_job(ctx, polygon_list);