poly1305/asm/poly1305-x86_64.pl: add Knights Landing AVX512 result.
authorAndy Polyakov <appro@openssl.org>
Wed, 6 Dec 2017 14:51:32 +0000 (15:51 +0100)
committerAndy Polyakov <appro@openssl.org>
Sat, 23 Dec 2017 15:06:25 +0000 (16:06 +0100)
Hardware used for benchmarking courtesy of Atos, experiments run by
Romain Dolbeau <romain.dolbeau@atos.net>. Kudos!

Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/4855)

crypto/poly1305/asm/poly1305-x86_64.pl

index 4519084..0d1c0de 100755 (executable)
@@ -45,7 +45,7 @@
 # Haswell      1.14/+175%      1.11            0.65
 # Skylake[-X]  1.13/+120%      0.96            0.51    [0.35]
 # Silvermont   2.83/+95%       -
-# Knights L    3.60/?          1.65            1.10    ?
+# Knights L    3.60/?          1.65            1.10    0.41(***)
 # Goldmont     1.70/+180%      -
 # VIA Nano     1.82/+150%      -
 # Sledgehammer 1.38/+160%      -
@@ -60,6 +60,8 @@
 #      Core processors, 50-30%, less newer processor is, but slower on
 #      contemporary ones, for example almost 2x slower on Atom, and as
 #      former are naturally disappearing, SSE2 is deemed unnecessary;
+# (***)        strangely enough performance seems to vary from core to core,
+#      listed result is best case;
 
 $flavour = shift;
 $output  = shift;
@@ -2160,33 +2162,33 @@ $code.=<<___;
        vmovdqa         96(%rcx),%y#$T2         # .Lpermd_avx2
 
        # expand pre-calculated table
-       vmovdqu32       `16*0-64`($ctx),${R0}{%k2}{z}
+       vmovdqu         `16*0-64`($ctx),%x#$D0  # will become expanded ${R0}
        and             \$-512,%rsp
-       vmovdqu32       `16*1-64`($ctx),${R1}{%k2}{z}
+       vmovdqu         `16*1-64`($ctx),%x#$D1  # will become ... ${R1}
        mov             \$0x20,%rax
-       vmovdqu32       `16*2-64`($ctx),${S1}{%k2}{z}
-       vmovdqu32       `16*3-64`($ctx),${R2}{%k2}{z}
-       vmovdqu32       `16*4-64`($ctx),${S2}{%k2}{z}
-       vmovdqu32       `16*5-64`($ctx),${R3}{%k2}{z}
-       vmovdqu32       `16*6-64`($ctx),${S3}{%k2}{z}
-       vmovdqu32       `16*7-64`($ctx),${R4}{%k2}{z}
-       vmovdqu32       `16*8-64`($ctx),${S4}{%k2}{z}
-       vpermd          $R0,$T2,$R0             # 00003412 -> 14243444
+       vmovdqu         `16*2-64`($ctx),%x#$T0  # ... ${S1}
+       vmovdqu         `16*3-64`($ctx),%x#$D2  # ... ${R2}
+       vmovdqu         `16*4-64`($ctx),%x#$T1  # ... ${S2}
+       vmovdqu         `16*5-64`($ctx),%x#$D3  # ... ${R3}
+       vmovdqu         `16*6-64`($ctx),%x#$T3  # ... ${S3}
+       vmovdqu         `16*7-64`($ctx),%x#$D4  # ... ${R4}
+       vmovdqu         `16*8-64`($ctx),%x#$T4  # ... ${S4}
+       vpermd          $D0,$T2,$R0             # 00003412 -> 14243444
        vpbroadcastq    64(%rcx),$MASK          # .Lmask26
-       vpermd          $R1,$T2,$R1
-       vpermd          $S1,$T2,$S1
-       vpermd          $R2,$T2,$R2
+       vpermd          $D1,$T2,$R1
+       vpermd          $T0,$T2,$S1
+       vpermd          $D2,$T2,$R2
        vmovdqa64       $R0,0x00(%rsp){%k2}     # save in case $len%128 != 0
         vpsrlq         \$32,$R0,$T0            # 14243444 -> 01020304
-       vpermd          $S2,$T2,$S2
+       vpermd          $T1,$T2,$S2
        vmovdqu64       $R1,0x00(%rsp,%rax){%k2}
         vpsrlq         \$32,$R1,$T1
-       vpermd          $R3,$T2,$R3
+       vpermd          $D3,$T2,$R3
        vmovdqa64       $S1,0x40(%rsp){%k2}
-       vpermd          $S3,$T2,$S3
-       vpermd          $R4,$T2,$R4
+       vpermd          $T3,$T2,$S3
+       vpermd          $D4,$T2,$R4
        vmovdqu64       $R2,0x40(%rsp,%rax){%k2}
-       vpermd          $S4,$T2,$S4
+       vpermd          $T4,$T2,$S4
        vmovdqa64       $S2,0x80(%rsp){%k2}
        vmovdqu64       $R3,0x80(%rsp,%rax){%k2}
        vmovdqa64       $S3,0xc0(%rsp){%k2}