diff options
author | Matthias Hopf <mhopf@suse.de> | 2009-01-29 18:36:29 +0100 |
---|---|---|
committer | Matthias Hopf <mhopf@suse.de> | 2009-01-29 18:36:29 +0100 |
commit | 55e13dea3f69597bad60b2b4f2ac6dfe8a07d64f (patch) | |
tree | 2bc41c83f80a8240b86144a1431bdcb879e205ca | |
parent | 5d79809fa5c471c283363e4746b965cf755c6606 (diff) |
Start performance test with 3 ALU instructions per clause.
1 ALU instruction (5 entries) is never saturating the ALU units, and on RV770
you need a clause pretty full (approx. 50 of 128 entries) to saturate them,
while on RV670 10 entries was enough.
When starting with 15 entries, we can fill up 120 entries (24 ALU
instructions), while when starting with 5 entries, we end up with a maximum
of 80.
-rw-r--r-- | r600_perf.c | 2 |
1 files changed, 1 insertions, 1 deletions
diff --git a/r600_perf.c b/r600_perf.c index c386d92..c0d6303 100644 --- a/r600_perf.c +++ b/r600_perf.c @@ -755,7 +755,7 @@ void test_alu_quad_perf(adapter_t *adapt) /* Render n times, duplicate alu instructions per clause each time */ render_num = 64; - for (alu_num = 1; alu_num <= MAX_NUM_ALUS_PER_CLAUSE; alu_num *= 2) { + for (alu_num = 3; alu_num <= MAX_NUM_ALUS_PER_CLAUSE; alu_num *= 2) { /* Set number of alu insts per clause */ for (i = 0; i < NUM_ALU_CLAUSES; i++) { |