This patch replaces the old register lifetime estiamtion and
rename mapping evaluation with the new one.
Performance to compare between the current and the new implementation
were measured by running the shader-db in one thread.
-----------------------------------------------------------
old new(std::sort)
---------------- time ./run -j shaders --------------------
real 5.80s 5.75s
user 5.75s 5.70s
sys 0.05s 0.05s
---- valgrind --tool=callgrind --dump-instr=yes------------
merge 0.08% 0.18%
estimate lifetime 0.02% 0.11%
evaluate mapping (incl=0.3%) 0.04%
apply mapping 0.03% 0.02%
--- perf (approximate because of statistic sampling) ----
merge (total) 0.09% 0.16%
estimate lifetime 0.03% 0.10%
evaluate mapping (incl=0.02%) 0.04%
apply mapping 0.04% 0.04%
---
src/mesa/state_tracker/st_glsl_to_tgsi.cpp | 58 ++++++------------------------
1 file changed, 11 insertions(+), 47 deletions(-)
diff --git a/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
b/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
index a9c2584f26..3f311d6668 100644
--- a/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
+++ b/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
@@ -55,7 +55,7 @@
#include "st_glsl_types.h"
#include "st_nir.h"
#include "st_shader_cache.h"
-#include "st_glsl_to_tgsi_private.h"
+#include "st_glsl_to_tgsi_temprename.h"
#include "util/hash_table.h"
#include <algorithm>
@@ -5145,54 +5145,18 @@ glsl_to_tgsi_visitor::merge_two_dsts(void)
void
glsl_to_tgsi_visitor::merge_registers(void)
{
- int *last_reads = ralloc_array(mem_ctx, int, this->next_temp);
- int *first_writes = ralloc_array(mem_ctx, int, this->next_temp);
- struct rename_reg_pair *renames = rzalloc_array(mem_ctx, struct
rename_reg_pair, this->next_temp);
- int i, j;
-
- /* Read the indices of the last read and first write to each temp register
- * into an array so that we don't have to traverse the instruction list as
- * much. */
- for (i = 0; i < this->next_temp; i++) {
- last_reads[i] = -1;
- first_writes[i] = -1;
- }
- get_last_temp_read_first_temp_write(last_reads, first_writes);
-
- /* Start looking for registers with non-overlapping usages that can be
- * merged together. */
- for (i = 0; i < this->next_temp; i++) {
- /* Don't touch unused registers. */
- if (last_reads[i] < 0 || first_writes[i] < 0) continue;
-
- for (j = 0; j < this->next_temp; j++) {
- /* Don't touch unused registers. */
- if (last_reads[j] < 0 || first_writes[j] < 0) continue;
-
- /* We can merge the two registers if the first write to j is after or
- * in the same instruction as the last read from i. Note that the
- * register at index i will always be used earlier or at the same time
- * as the register at index j. */
- if (first_writes[i] <= first_writes[j] &&
- last_reads[i] <= first_writes[j]) {
- renames[j].new_reg = i;
- renames[j].valid = true;
-
- /* Update the first_writes and last_reads arrays with the new
- * values for the merged register index, and mark the newly unused
- * register index as such. */
- assert(last_reads[j] >= last_reads[i]);
- last_reads[i] = last_reads[j];
- first_writes[j] = -1;
- last_reads[j] = -1;
- }
- }
- }
-
+ struct rename_reg_pair *renames =
+ rzalloc_array(mem_ctx, struct rename_reg_pair, this->next_temp);
+ struct lifetime *lifetimes =
+ rzalloc_array(mem_ctx, struct lifetime, this->next_temp);
+
+ get_temp_registers_required_lifetimes(mem_ctx, &this->instructions,
+ this->next_temp, lifetimes);
+ get_temp_registers_remapping(mem_ctx, this->next_temp, lifetimes, renames);
rename_temp_registers(renames);
+
+ ralloc_free(lifetimes);
ralloc_free(renames);
- ralloc_free(last_reads);
- ralloc_free(first_writes);
}
/* Reassign indices to temporary registers by reusing unused indices created
--
2.13.0
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev
