freedreno/ir3: Don't leave holes the UBO upload plan.
Shaders may not use a particular region of a UBO in a given shader (think
UBOs shared between stages, or between shaders), and by just always
extending the existing range for a given UBO, we'd waste bandwidth
uploading it, and also waste our precious const space in storing the
unused data.
Instead, only upload exactly the ranges we can use, and merge ranges when
they're neighbors. We may end up with more upload packets, but the
bandwidth savings is surely going to be worth it (and if find we want a
distance threshold for merging with nearby uploads, that would be easy to
add).
total instructions in shared programs: 9266114 -> 9255092 (-0.12%)
total full in shared programs: 343162 -> 341709 (-0.42%)
total constlen in shared programs: 1454368 -> 1275236 (-12.32%)
total cat6 in shared programs: 93073 -> 82589 (-11.26%)
total (ss) in shared programs: 212402 -> 206404 (-2.82%)
total (sy) in shared programs: 122905 -> 114007 (-7.24%)
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7036>
diff --git a/src/freedreno/ir3/ir3_nir_analyze_ubo_ranges.c b/src/freedreno/ir3/ir3_nir_analyze_ubo_ranges.c
index 13ee7b5..de62271 100644
--- a/src/freedreno/ir3/ir3_nir_analyze_ubo_ranges.c
+++ b/src/freedreno/ir3/ir3_nir_analyze_ubo_ranges.c
@@ -72,12 +72,12 @@
}
/**
- * Get an existing range, but don't create a new range associated with
- * the ubo, but don't create a new one if one does not already exist.
+ * Finds the given instruction's UBO load in the UBO upload plan, if any.
*/
static const struct ir3_ubo_range *
get_existing_range(nir_intrinsic_instr *instr,
- const struct ir3_ubo_analysis_state *state)
+ const struct ir3_ubo_analysis_state *state,
+ struct ir3_ubo_range *r)
{
struct ir3_ubo_info ubo = {};
@@ -86,14 +86,46 @@
for (int i = 0; i < state->num_enabled; i++) {
const struct ir3_ubo_range *range = &state->range[i];
- if (!memcmp(&range->ubo, &ubo, sizeof(ubo)))
+ if (!memcmp(&range->ubo, &ubo, sizeof(ubo)) &&
+ r->start >= range->start &&
+ r->end <= range->end) {
return range;
+ }
}
return NULL;
}
/**
+ * Merges together neighboring/overlapping ranges in the range plan with a
+ * newly updated range.
+ */
+static void
+merge_neighbors(struct ir3_ubo_analysis_state *state, int index)
+{
+ struct ir3_ubo_range *a = &state->range[index];
+
+ /* index is always the first slot that would have neighbored/overlapped with
+ * the new range.
+ */
+ for (int i = index + 1; i < state->num_enabled; i++) {
+ struct ir3_ubo_range *b = &state->range[i];
+ if (memcmp(&a->ubo, &b->ubo, sizeof(a->ubo)))
+ continue;
+
+ if (a->start > b->end || a->end < b->start)
+ continue;
+
+ /* Merge B into A. */
+ a->start = MIN2(a->start, b->start);
+ a->end = MAX2(a->end, b->end);
+
+ /* Swap the last enabled range into B's now unused slot */
+ *b = state->range[--state->num_enabled];
+ }
+}
+
+/**
* During the first pass over the shader, makes the plan of which UBO upload
* should include the range covering this UBO load.
*
@@ -122,6 +154,12 @@
if (memcmp(&plan_r->ubo, &ubo, sizeof(ubo)))
continue;
+ /* Don't extend existing uploads unless they're
+ * neighboring/overlapping.
+ */
+ if (r.start > plan_r->end || r.end < plan_r->start)
+ continue;
+
r.start = MIN2(r.start, plan_r->start);
r.end = MAX2(r.end, plan_r->end);
@@ -132,6 +170,8 @@
plan_r->start = r.start;
plan_r->end = r.end;
*upload_remaining -= added;
+
+ merge_neighbors(state, i);
return;
}
@@ -235,26 +275,18 @@
{
b->cursor = nir_before_instr(&instr->instr);
- /* We don't lower dynamic block index UBO loads to load_uniform, but we
- * could probably with some effort determine a block stride in number of
- * registers.
- */
- const struct ir3_ubo_range *range = get_existing_range(instr, state);
- if (!range) {
- track_ubo_use(instr, b, num_ubos);
- return false;
- }
-
struct ir3_ubo_range r;
if (!get_ubo_load_range(b->shader, instr, alignment, &r)) {
track_ubo_use(instr, b, num_ubos);
return false;
}
- /* After gathering the UBO access ranges, we limit the total
- * upload. Don't lower if this load is outside the range.
+ /* We don't lower dynamic block index UBO loads to load_uniform, but we
+ * could probably with some effort determine a block stride in number of
+ * registers.
*/
- if (!(range->start <= r.start && r.end <= range->end)) {
+ const struct ir3_ubo_range *range = get_existing_range(instr, state, &r);
+ if (!range) {
track_ubo_use(instr, b, num_ubos);
return false;
}
