iris: Skip emitting 3DSTATE_INDEX_BUFFER if possible

We were emitting 3DSTATE_INDEX_BUFFER on every indexed draw, even if
back-to-back draws referred to the same index buffer.  This improves
drawoverhead scores in the DrawElements cases by about 10%, by giving
us even more minimal batches.
This commit is contained in:
Kenneth Graunke 2019-02-16 00:57:54 -08:00
parent 8af1990ad7
commit b61f17d362
3 changed files with 27 additions and 11 deletions

View File

@ -100,6 +100,7 @@ iris_lost_context_state(struct iris_batch *batch)
ice->state.dirty = ~0ull;
memset(ice->state.last_grid, 0, sizeof(ice->state.last_grid));
batch->last_surface_base_address = ~0ull;
ice->vtbl.lost_genx_state(ice, batch);
}
static enum pipe_reset_status

View File

@ -481,6 +481,7 @@ struct iris_vtable {
void (*populate_cs_key)(const struct iris_context *ice,
struct brw_cs_prog_key *key);
uint32_t (*mocs)(const struct iris_bo *bo);
void (*lost_genx_state)(struct iris_context *ice, struct iris_batch *batch);
};
/**

View File

@ -784,6 +784,7 @@ struct iris_depth_buffer_state {
*/
struct iris_genx_state {
struct iris_vertex_buffer_state vertex_buffers[33];
uint32_t last_index_buffer[GENX(3DSTATE_INDEX_BUFFER_length)];
struct iris_depth_buffer_state depth_buffer;
@ -4249,13 +4250,7 @@ iris_restore_render_saved_bos(struct iris_context *ice,
pin_depth_and_stencil_buffers(batch, cso_fb->zsbuf, ice->state.cso_zsa);
}
if (draw->index_size == 0 && ice->state.last_res.index_buffer) {
/* This draw didn't emit a new index buffer, so we are inheriting the
* older index buffer. This draw didn't need it, but future ones may.
*/
struct iris_bo *bo = iris_resource_bo(ice->state.last_res.index_buffer);
iris_use_pinned_bo(batch, bo, false);
}
iris_use_optional_res(batch, ice->state.last_res.index_buffer, false);
if (clean & IRIS_DIRTY_VERTEX_BUFFERS) {
uint64_t bound = ice->state.bound_vertex_buffers;
@ -5185,13 +5180,21 @@ iris_upload_render_state(struct iris_context *ice,
offset = 0;
}
struct iris_genx_state *genx = ice->state.genx;
struct iris_bo *bo = iris_resource_bo(ice->state.last_res.index_buffer);
iris_emit_cmd(batch, GENX(3DSTATE_INDEX_BUFFER), ib) {
uint32_t ib_packet[GENX(3DSTATE_INDEX_BUFFER_length)];
iris_pack_command(GENX(3DSTATE_INDEX_BUFFER), ib_packet, ib) {
ib.IndexFormat = draw->index_size >> 1;
ib.MOCS = mocs(bo);
ib.BufferSize = bo->size - offset;
ib.BufferStartingAddress = ro_bo(bo, offset);
ib.BufferStartingAddress = ro_bo(NULL, bo->gtt_offset + offset);
}
if (memcmp(genx->last_index_buffer, ib_packet, sizeof(ib_packet)) != 0) {
memcpy(genx->last_index_buffer, ib_packet, sizeof(ib_packet));
iris_batch_emit(batch, ib_packet, sizeof(ib_packet));
iris_use_pinned_bo(batch, bo, false);
}
/* The VF cache key only uses 32-bits, see vertex buffer comment above */
@ -5633,8 +5636,10 @@ iris_rebind_buffer(struct iris_context *ice,
}
}
/* No need to handle these:
* - PIPE_BIND_INDEX_BUFFER (emitted for every indexed draw)
/* We don't need to handle PIPE_BIND_INDEX_BUFFER here: we re-emit
* the 3DSTATE_INDEX_BUFFER packet whenever the address changes.
*
* There is also no need to handle these:
* - PIPE_BIND_COMMAND_ARGS_BUFFER (emitted for every indirect draw)
* - PIPE_BIND_QUERY_BUFFER (no persistent state references)
*/
@ -6385,6 +6390,14 @@ gen9_toggle_preemption(struct iris_context *ice,
}
#endif
static void
iris_lost_genx_state(struct iris_context *ice, struct iris_batch *batch)
{
struct iris_genx_state *genx = ice->state.genx;
memset(genx->last_index_buffer, 0, sizeof(genx->last_index_buffer));
}
void
genX(init_state)(struct iris_context *ice)
{
@ -6459,6 +6472,7 @@ genX(init_state)(struct iris_context *ice)
ice->vtbl.populate_fs_key = iris_populate_fs_key;
ice->vtbl.populate_cs_key = iris_populate_cs_key;
ice->vtbl.mocs = mocs;
ice->vtbl.lost_genx_state = iris_lost_genx_state;
ice->state.dirty = ~0ull;