Add prefetch to sha1_block_data_order_shaext
Similar idea to https://boringssl-review.googlesource.com/c/boringssl/+/55466
Results are pretty close to the current state,
e.g. tool speed goes from
Did 74000 SHA-1 (16384 bytes) operations in 1004094us (73698.3 ops/sec): 1207.5 MB/s
to
Did 75000 SHA-1 (16384 bytes) operations in 1004022us (74699.6 ops/sec): 1223.9 MB/s
But on AMD with prefetchers disabled and large enough data size,
to force cache misses this gives ~3x improvement:
name old time/op new time/op delta
BM_SHA1Hash/2 141ns ± 1% 143ns ± 2% ~ (p=0.421 n=5+5)
BM_SHA1Hash/4 143ns ± 2% 143ns ± 3% ~ (p=0.841 n=5+5)
BM_SHA1Hash/8 141ns ± 1% 141ns ± 2% ~ (p=1.000 n=5+5)
BM_SHA1Hash/16 141ns ± 1% 141ns ± 1% ~ (p=0.841 n=5+5)
BM_SHA1Hash/32 143ns ± 2% 143ns ± 1% ~ (p=0.690 n=5+5)
BM_SHA1Hash/64 178ns ± 1% 179ns ± 1% ~ (p=0.151 n=5+5)
BM_SHA1Hash/512 454ns ± 1% 454ns ± 1% ~ (p=0.841 n=5+5)
BM_SHA1Hash/4k 2.66µs ± 1% 2.65µs ± 1% ~ (p=1.000 n=5+5)
BM_SHA1Hash/32k 20.3µs ± 1% 20.3µs ± 2% ~ (p=1.000 n=5+5)
BM_SHA1Hash/256k 162µs ± 1% 161µs ± 1% ~ (p=0.548 n=5+5)
BM_SHA1Hash/1M 644µs ± 1% 645µs ± 1% ~ (p=0.841 n=5+5)
BM_SHA1Hash/2M 1.29ms ± 1% 1.29ms ± 2% ~ (p=0.690 n=5+5)
BM_SHA1Hash/4M 2.58ms ± 1% 2.58ms ± 1% ~ (p=0.841 n=5+5)
BM_SHA1Hash/8M 5.14ms ± 0% 5.15ms ± 1% ~ (p=0.286 n=4+5)
BM_SHA1Hash/16M 11.4ms ± 3% 10.3ms ± 1% -9.04% (p=0.016 n=4+5)
BM_SHA1Hash/128M 249ms ± 0% 83ms ± 1% -66.73% (p=0.008 n=5+5)
Change-Id: I7cae746b6d8a705d6bf2d5c5df6a2dca6d44791a
Reviewed-on: https://boringssl-review.googlesource.com/c/boringssl/+/57826
Commit-Queue: Adam Langley <agl@google.com>
Reviewed-by: Adam Langley <agl@google.com>
diff --git a/crypto/fipsmodule/sha/asm/sha1-x86_64.pl b/crypto/fipsmodule/sha/asm/sha1-x86_64.pl
index 6ee7887..d9afacb 100755
--- a/crypto/fipsmodule/sha/asm/sha1-x86_64.pl
+++ b/crypto/fipsmodule/sha/asm/sha1-x86_64.pl
@@ -389,6 +389,7 @@
lea 0x40($inp),%r8 # next input block
paddd @MSG[0],$E
cmovne %r8,$inp
+ prefetcht0 512($inp)
movdqa $ABCD,$ABCD_SAVE # offload $ABCD
___
for($i=0;$i<20-4;$i+=2) {
