summaryrefslogtreecommitdiff
path: root/mpi/alpha/README
diff options
context:
space:
mode:
Diffstat (limited to 'mpi/alpha/README')
-rw-r--r--mpi/alpha/README53
1 files changed, 53 insertions, 0 deletions
diff --git a/mpi/alpha/README b/mpi/alpha/README
new file mode 100644
index 0000000..55c0a29
--- /dev/null
+++ b/mpi/alpha/README
@@ -0,0 +1,53 @@
+This directory contains mpn functions optimized for DEC Alpha processors.
+
+RELEVANT OPTIMIZATION ISSUES
+
+EV4
+
+1. This chip has very limited store bandwidth. The on-chip L1 cache is
+write-through, and a cache line is transfered from the store buffer to the
+off-chip L2 in as much 15 cycles on most systems. This delay hurts
+mpn_add_n, mpn_sub_n, mpn_lshift, and mpn_rshift.
+
+2. Pairing is possible between memory instructions and integer arithmetic
+instructions.
+
+3. mulq and umulh is documented to have a latency of 23 cycles, but 2 of
+these cycles are pipelined. Thus, multiply instructions can be issued at a
+rate of one each 21nd cycle.
+
+EV5
+
+1. The memory bandwidth of this chip seems excellent, both for loads and
+stores. Even when the working set is larger than the on-chip L1 and L2
+caches, the perfromance remain almost unaffected.
+
+2. mulq has a measured latency of 13 cycles and an issue rate of 1 each 8th
+cycle. umulh has a measured latency of 15 cycles and an issue rate of 1
+each 10th cycle. But the exact timing is somewhat confusing.
+
+3. mpn_add_n. With 4-fold unrolling, we need 37 instructions, whereof 12
+ are memory operations. This will take at least
+ ceil(37/2) [dual issue] + 1 [taken branch] = 20 cycles
+ We have 12 memory cycles, plus 4 after-store conflict cycles, or 16 data
+ cache cycles, which should be completely hidden in the 20 issue cycles.
+ The computation is inherently serial, with these dependencies:
+ addq
+ / \
+ addq cmpult
+ | |
+ cmpult |
+ \ /
+ or
+ I.e., there is a 4 cycle path for each limb, making 16 cycles the absolute
+ minimum. We could replace the `or' with a cmoveq/cmovne, which would save
+ a cycle on EV5, but that might waste a cycle on EV4. Also, cmov takes 2
+ cycles.
+ addq
+ / \
+ addq cmpult
+ | \
+ cmpult -> cmovne
+
+STATUS
+