#include "util/set.h"
#include "broadcom/clif/clif_dump.h"
-static void
+void
v3d_job_free(struct v3d_context *v3d, struct v3d_job *job)
{
set_foreach(job->bos, entry) {
ralloc_free(job);
}
-static struct v3d_job *
+struct v3d_job *
v3d_job_create(struct v3d_context *v3d)
{
struct v3d_job *job = rzalloc(v3d, struct v3d_job);
{
assert(v3d->prim_counts);
- perf_debug("stalling on TF counts readback");
+ perf_debug("stalling on TF counts readback\n");
struct v3d_resource *rsc = v3d_resource(v3d->prim_counts);
if (v3d_bo_wait(rsc->bo, PIPE_TIMEOUT_INFINITE, "prim-counts")) {
- uint32_t *map = v3d_bo_map(rsc->bo);
+ uint32_t *map = v3d_bo_map(rsc->bo) + v3d->prim_counts_offset;
v3d->tf_prims_generated += map[V3D_PRIM_COUNTS_TF_WRITTEN];
+ /* When we only have a vertex shader we determine the primitive
+ * count in the CPU so don't update it here again.
+ */
+ if (v3d->prog.gs)
+ v3d->prims_generated += map[V3D_PRIM_COUNTS_WRITTEN];
}
}
job->submit.bcl_end = job->bcl.bo->offset + cl_offset(&job->bcl);
job->submit.rcl_end = job->rcl.bo->offset + cl_offset(&job->rcl);
+ job->submit.flags = 0;
+ if (job->tmu_dirty_rcl && screen->has_cache_flush)
+ job->submit.flags |= DRM_V3D_SUBMIT_CL_FLUSH_CACHE;
+
/* On V3D 4.1, the tile alloc/state setup moved to register writes
* instead of binner packets.
*/
* feedback we need to read the primitive counts and accumulate
* them, otherwise they will be reset at the start of the next
* draw when we emit the Tile Binning Mode Configuration packet.
+ *
+ * If the job doesn't have any TF draw calls, then we know
+ * the primitive count must be zero and we can skip stalling
+ * for this. This also fixes a problem because it seems that
+ * in this scenario the counters are not reset with the Tile
+ * Binning Mode Configuration packet, which would translate
+ * to us reading an obsolete (possibly non-zero) value from
+ * the GPU counters.
*/
- if (v3d->streamout.num_targets)
+ if (v3d->streamout.num_targets && job->tf_draw_calls_queued > 0)
v3d_read_and_accumulate_primitive_counters(v3d);
}