panfrost: Share tiler_heap across batches/contexts

There's only one tiler, so this is safe. (The blob does the same
optimization.) This avoids allocating multiple heaps for multiple
batches, which wastes memory and CPU time.

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Reviewed-by: Tomeu Vizoso <tomeu.vizoso@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6373>

src/gallium/drivers/panfrost/pan_context.c

@@ -76,20 +76,14 @@ panfrost_emit_midg_tiler(struct panfrost_batch *batch, unsigned vertex_count)
         t.polygon_list_size = panfrost_tiler_full_size(
                                      width, height, t.hierarchy_mask, hierarchy);
 
-        /* Sanity check */
-
         if (vertex_count) {
-                struct panfrost_bo *tiler_heap;
-
-                tiler_heap = panfrost_batch_get_tiler_heap(batch);
                 t.polygon_list = panfrost_batch_get_polygon_list(batch,
                                                                  header_size +
                                                                  t.polygon_list_size);
 
 
-                /* Allow the entire tiler heap */
+                t.heap_start = device->tiler_heap->gpu;
-                t.heap_start = tiler_heap->gpu;
+                t.heap_end = device->tiler_heap->gpu + device->tiler_heap->size;
-                t.heap_end = tiler_heap->gpu + tiler_heap->size;
         } else {
                 struct panfrost_bo *tiler_dummy;
 

src/gallium/drivers/panfrost/pan_job.c

@@ -679,23 +679,6 @@ panfrost_batch_get_shared_memory(struct panfrost_batch *batch,
         return batch->shared_memory;
 }
 
-struct panfrost_bo *
-panfrost_batch_get_tiler_heap(struct panfrost_batch *batch)
-{
-        if (batch->tiler_heap)
-                return batch->tiler_heap;
-
-        batch->tiler_heap = panfrost_batch_create_bo(batch, 4096 * 4096,
-                                                     PAN_BO_INVISIBLE |
-                                                     PAN_BO_GROWABLE,
-                                                     PAN_BO_ACCESS_PRIVATE |
-                                                     PAN_BO_ACCESS_RW |
-                                                     PAN_BO_ACCESS_VERTEX_TILER |
-                                                     PAN_BO_ACCESS_FRAGMENT);
-        assert(batch->tiler_heap);
-        return batch->tiler_heap;
-}
-
 mali_ptr
 panfrost_batch_get_tiler_meta(struct panfrost_batch *batch, unsigned vertex_count)
 {
@@ -705,14 +688,13 @@ panfrost_batch_get_tiler_meta(struct panfrost_batch *batch, unsigned vertex_coun
         if (batch->tiler_meta)
                 return batch->tiler_meta;
 
-        struct panfrost_bo *tiler_heap;
+        struct panfrost_device *dev = pan_device(batch->ctx->base.screen);
-        tiler_heap = panfrost_batch_get_tiler_heap(batch);
 
         struct bifrost_tiler_heap_meta tiler_heap_meta = {
-            .heap_size = tiler_heap->size,
+            .heap_size = dev->tiler_heap->size,
-            .tiler_heap_start = tiler_heap->gpu,
+            .tiler_heap_start = dev->tiler_heap->gpu,
-            .tiler_heap_free = tiler_heap->gpu,
+            .tiler_heap_free = dev->tiler_heap->gpu,
-            .tiler_heap_end = tiler_heap->gpu + tiler_heap->size,
+            .tiler_heap_end = dev->tiler_heap->gpu + dev->tiler_heap->size,
             .unk1 = 0x1,
             .unk7e007e = 0x7e007e,
         };
@@ -997,7 +979,7 @@ panfrost_batch_submit_ioctl(struct panfrost_batch *batch,
         submit.jc = first_job_desc;
         submit.requirements = reqs;
 
-        bo_handles = calloc(batch->pool.bos->entries + batch->invisible_pool.bos->entries + batch->bos->entries, sizeof(*bo_handles));
+        bo_handles = calloc(batch->pool.bos->entries + batch->invisible_pool.bos->entries + batch->bos->entries + 1, sizeof(*bo_handles));
         assert(bo_handles);
 
         hash_table_foreach(batch->bos, entry)
@@ -1009,6 +991,10 @@ panfrost_batch_submit_ioctl(struct panfrost_batch *batch,
         hash_table_foreach(batch->invisible_pool.bos, entry)
                 panfrost_batch_record_bo(entry, bo_handles, submit.bo_handle_count++);
 
+        /* Used by all tiler jobs (XXX: skip for compute-only) */
+        if (!(reqs & PANFROST_JD_REQ_FS))
+                bo_handles[submit.bo_handle_count++] = dev->tiler_heap->gem_handle;
+
         submit.bo_handles = (u64) (uintptr_t) bo_handles;
         ret = drmIoctl(dev->fd, DRM_IOCTL_PANFROST_SUBMIT, &submit);
         free(bo_handles);

src/gallium/drivers/panfrost/pan_job.h

@@ -182,9 +182,6 @@ panfrost_batch_get_shared_memory(struct panfrost_batch *batch, unsigned size, un
 mali_ptr
 panfrost_batch_get_polygon_list(struct panfrost_batch *batch, unsigned size);
 
-struct panfrost_bo *
-panfrost_batch_get_tiler_heap(struct panfrost_batch *batch);
-
 struct panfrost_bo *
 panfrost_batch_get_tiler_dummy(struct panfrost_batch *batch);
 

src/panfrost/lib/pan_device.h

@@ -127,6 +127,14 @@ struct panfrost_device {
         } bo_cache;
 
         struct pan_blit_shaders blit_shaders;
+
+        /* Tiler heap shared across all tiler jobs, allocated against the
+         * device since there's only a single tiler. Since this is invisible to
+         * the CPU, it's okay for multiple contexts to reference it
+         * simultaneously; by keeping on the device struct, we eliminate a
+         * costly per-context allocation. */
+
+        struct panfrost_bo *tiler_heap;
 };
 
 void

src/panfrost/lib/pan_props.c

@@ -172,12 +172,20 @@ panfrost_open_device(void *memctx, int fd, struct panfrost_device *dev)
 
         for (unsigned i = 0; i < ARRAY_SIZE(dev->bo_cache.buckets); ++i)
                 list_inithead(&dev->bo_cache.buckets[i]);
+
+        /* Tiler heap is internally required by the tiler, which can only be
+         * active for a single job chain at once, so a single heap can be
+         * shared across batches/contextes */
+
+        dev->tiler_heap = panfrost_bo_create(dev, 4096 * 4096,
+                        PAN_BO_INVISIBLE | PAN_BO_GROWABLE);
 }
 
 void
 panfrost_close_device(struct panfrost_device *dev)
 {
         panfrost_bo_unreference(dev->blit_shaders.bo);
+        panfrost_bo_unreference(dev->tiler_heap);
         panfrost_bo_cache_evict_all(dev);
         pthread_mutex_destroy(&dev->bo_cache.lock);
         drmFreeVersion(dev->kernel_version);