# February 2011.
#
# Profiler-assisted and platform-specific optimization resulted in 16%
-# improvement on Cortex A8 core and ~17 cycles per processed byte.
+# improvement on Cortex A8 core and ~16.4 cycles per processed byte.
while (($output=shift) && ($output!~/^\w[\w\-]*\.\w+$/)) {}
open STDOUT,">$output";
$ctx="r0"; $t0="r0";
-$inp="r1"; $t3="r1";
+$inp="r1"; $t4="r1";
$len="r2"; $t1="r2";
-$T1="r3";
+$T1="r3"; $t3="r3";
$A="r4";
$B="r5";
$C="r6";
$code.=<<___ if ($i<16);
#if __ARM_ARCH__>=7
- ldr $T1,[$inp],#4
+ @ ldr $t1,[$inp],#4 @ $i
+# if $i==15
+ str $inp,[sp,#17*4] @ make room for $t4
+# endif
+ mov $t0,$e,ror#$Sigma1[0]
+ add $a,$a,$t2 @ h+=Maj(a,b,c) from the past
+ rev $t1,$t1
+ eor $t0,$t0,$e,ror#$Sigma1[1]
#else
- ldrb $T1,[$inp,#3] @ $i
+ @ ldrb $t1,[$inp,#3] @ $i
+ add $a,$a,$t2 @ h+=Maj(a,b,c) from the past
ldrb $t2,[$inp,#2]
- ldrb $t1,[$inp,#1]
- ldrb $t0,[$inp],#4
- orr $T1,$T1,$t2,lsl#8
- orr $T1,$T1,$t1,lsl#16
- orr $T1,$T1,$t0,lsl#24
+ ldrb $t0,[$inp,#1]
+ orr $t1,$t1,$t2,lsl#8
+ ldrb $t2,[$inp],#4
+ orr $t1,$t1,$t0,lsl#16
+# if $i==15
+ str $inp,[sp,#17*4] @ make room for $t4
+# endif
+ mov $t0,$e,ror#$Sigma1[0]
+ orr $t1,$t1,$t2,lsl#24
+ eor $t0,$t0,$e,ror#$Sigma1[1]
#endif
___
$code.=<<___;
- mov $t0,$e,ror#$Sigma1[0]
ldr $t2,[$Ktbl],#4 @ *K256++
- eor $t0,$t0,$e,ror#$Sigma1[1]
+ add $h,$h,$t1 @ h+=X[i]
+ str $t1,[sp,#`$i%16`*4]
eor $t1,$f,$g
-#if $i>=16
- add $T1,$T1,$t3 @ from BODY_16_xx
-#elif __ARM_ARCH__>=7 && defined(__ARMEL__)
- rev $T1,$T1
-#endif
-#if $i==15
- str $inp,[sp,#17*4] @ leave room for $t3
-#endif
eor $t0,$t0,$e,ror#$Sigma1[2] @ Sigma1(e)
and $t1,$t1,$e
- str $T1,[sp,#`$i%16`*4]
- add $T1,$T1,$t0
+ add $h,$h,$t0 @ h+=Sigma1(e)
eor $t1,$t1,$g @ Ch(e,f,g)
- add $T1,$T1,$h
- mov $h,$a,ror#$Sigma0[0]
- add $T1,$T1,$t1
- eor $h,$h,$a,ror#$Sigma0[1]
- add $T1,$T1,$t2
- eor $h,$h,$a,ror#$Sigma0[2] @ Sigma0(a)
-#if $i>=15
- ldr $t3,[sp,#`($i+2)%16`*4] @ from BODY_16_xx
+ add $h,$h,$t2 @ h+=K256[i]
+ mov $t0,$a,ror#$Sigma0[0]
+ add $h,$h,$t1 @ h+=Ch(e,f,g)
+#if $i==31
+ and $t2,$t2,#0xff
+ cmp $t2,#0xf2 @ done?
#endif
- orr $t0,$a,$b
- and $t1,$a,$b
- and $t0,$t0,$c
- add $h,$h,$T1
- orr $t0,$t0,$t1 @ Maj(a,b,c)
- add $d,$d,$T1
- add $h,$h,$t0
+#if $i<15
+# if __ARM_ARCH__>=7
+ ldr $t1,[$inp],#4 @ prefetch
+# else
+ ldrb $t1,[$inp,#3]
+# endif
+ eor $t2,$a,$b @ a^b, b^c in next round
+#else
+ ldr $t1,[sp,#`($i+2)%16`*4] @ from future BODY_16_xx
+ eor $t2,$a,$b @ a^b, b^c in next round
+ ldr $t4,[sp,#`($i+15)%16`*4] @ from future BODY_16_xx
+#endif
+ eor $t0,$a,ror#$Sigma0[1]
+ and $t3,$t3,$t2 @ (b^c)&=(a^b)
+ add $d,$d,$h @ d+=h
+ eor $t0,$a,ror#$Sigma0[2] @ Sigma0(a)
+ eor $t3,$t3,$b @ Maj(a,b,c)
+ add $h,$h,$t0 @ h+=Sigma0(a)
+ @ add $h,$h,$t3 @ h+=Maj(a,b,c)
___
+ ($t2,$t3)=($t3,$t2);
}
sub BODY_16_XX {
my ($i,$a,$b,$c,$d,$e,$f,$g,$h) = @_;
$code.=<<___;
- @ ldr $t3,[sp,#`($i+1)%16`*4] @ $i
- ldr $t2,[sp,#`($i+14)%16`*4]
- mov $t0,$t3,ror#$sigma0[0]
- ldr $T1,[sp,#`($i+0)%16`*4]
- eor $t0,$t0,$t3,ror#$sigma0[1]
- ldr $t1,[sp,#`($i+9)%16`*4]
- eor $t0,$t0,$t3,lsr#$sigma0[2] @ sigma0(X[i+1])
- mov $t3,$t2,ror#$sigma1[0]
- add $T1,$T1,$t0
- eor $t3,$t3,$t2,ror#$sigma1[1]
- add $T1,$T1,$t1
- eor $t3,$t3,$t2,lsr#$sigma1[2] @ sigma1(X[i+14])
- @ add $T1,$T1,$t3
+ @ ldr $t1,[sp,#`($i+1)%16`*4] @ $i
+ @ ldr $t4,[sp,#`($i+14)%16`*4]
+ mov $t0,$t1,ror#$sigma0[0]
+ add $a,$a,$t2 @ h+=Maj(a,b,c) from the past
+ mov $t2,$t4,ror#$sigma1[0]
+ eor $t0,$t0,$t1,ror#$sigma0[1]
+ eor $t2,$t2,$t4,ror#$sigma1[1]
+ eor $t0,$t0,$t1,lsr#$sigma0[2] @ sigma0(X[i+1])
+ ldr $t1,[sp,#`($i+0)%16`*4]
+ eor $t2,$t2,$t4,lsr#$sigma1[2] @ sigma1(X[i+14])
+ ldr $t4,[sp,#`($i+9)%16`*4]
+
+ add $t2,$t2,$t0
+ mov $t0,$e,ror#$Sigma1[0] @ from BODY_00_15
+ add $t1,$t1,$t2
+ eor $t0,$t0,$e,ror#$Sigma1[1] @ from BODY_00_15
+ add $t1,$t1,$t4 @ X[i]
___
&BODY_00_15(@_);
}
sub $Ktbl,r3,#256 @ K256
sub sp,sp,#16*4 @ alloca(X[16])
.Loop:
+# if __ARM_ARCH__>=7
+ ldr $t1,[$inp],#4
+# else
+ ldrb $t1,[$inp,#3]
+# endif
+ eor $t3,$B,$C @ magic
+ eor $t2,$t2,$t2
___
for($i=0;$i<16;$i++) { &BODY_00_15($i,@V); unshift(@V,pop(@V)); }
$code.=".Lrounds_16_xx:\n";
for (;$i<32;$i++) { &BODY_16_XX($i,@V); unshift(@V,pop(@V)); }
$code.=<<___;
- and $t2,$t2,#0xff
- cmp $t2,#0xf2
+ ldreq $t3,[sp,#16*4] @ pull ctx
bne .Lrounds_16_xx
- ldr $T1,[sp,#16*4] @ pull ctx
- ldr $t0,[$T1,#0]
- ldr $t1,[$T1,#4]
- ldr $t2,[$T1,#8]
+ add $A,$A,$t2 @ h+=Maj(a,b,c) from the past
+ ldr $t0,[$t3,#0]
+ ldr $t1,[$t3,#4]
+ ldr $t2,[$t3,#8]
add $A,$A,$t0
- ldr $t0,[$T1,#12]
+ ldr $t0,[$t3,#12]
add $B,$B,$t1
- ldr $t1,[$T1,#16]
+ ldr $t1,[$t3,#16]
add $C,$C,$t2
- ldr $t2,[$T1,#20]
+ ldr $t2,[$t3,#20]
add $D,$D,$t0
- ldr $t0,[$T1,#24]
+ ldr $t0,[$t3,#24]
add $E,$E,$t1
- ldr $t1,[$T1,#28]
+ ldr $t1,[$t3,#28]
add $F,$F,$t2
ldr $inp,[sp,#17*4] @ pull inp
ldr $t2,[sp,#18*4] @ pull inp+len
add $G,$G,$t0
add $H,$H,$t1
- stmia $T1,{$A,$B,$C,$D,$E,$F,$G,$H}
+ stmia $t3,{$A,$B,$C,$D,$E,$F,$G,$H}
cmp $inp,$t2
sub $Ktbl,$Ktbl,#256 @ rewind Ktbl
bne .Loop
#!/usr/bin/env perl
# ====================================================================
-# Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL
+# Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
# project. The module is, however, dual licensed under OpenSSL and
# CRYPTOGAMS licenses depending on where you obtain it. For further
# details see http://www.openssl.org/~appro/cryptogams/.
# March 2011.
#
# Add NEON implementation. On Cortex A8 it was measured to process
-# one byte in 25.5 cycles or 47% faster than integer-only code.
+# one byte in 23.3 cycles or ~60% faster than integer-only code.
+
+# August 2012.
+#
+# Improve NEON performance by 12% on Snapdragon S4. In absolute
+# terms it's 22.6 cycles per byte, which is disappointing result.
+# Technical writers asserted that 3-way S4 pipeline can sustain
+# multiple NEON instructions per cycle, but dual NEON issue could
+# not be observed, and for NEON-only sequences IPC(*) was found to
+# be limited by 1:-( 0.33 and 0.66 were measured for sequences with
+# ILPs(*) of 1 and 2 respectively. This in turn means that you can
+# even find yourself striving, as I did here, for achieving IPC
+# adequate to one delivered by Cortex A8 [for reference, it's
+# 0.5 for ILP of 1, and 1 for higher ILPs].
+#
+# (*) ILP, instruction-level parallelism, how many instructions
+# *can* execute at the same time. IPC, instructions per cycle,
+# indicates how many instructions actually execute.
# Byte order [in]dependence. =========================================
#
vld1.64 {@X[$i%16]},[$inp]! @ handles unaligned
#endif
vshr.u64 $t1,$e,#@Sigma1[1]
+#if $i>0
+ vadd.i64 $a,$Maj @ h+=Maj from the past
+#endif
vshr.u64 $t2,$e,#@Sigma1[2]
___
$code.=<<___;
vld1.64 {$K},[$Ktbl,:64]! @ K[i++]
vsli.64 $t0,$e,#`64-@Sigma1[0]`
vsli.64 $t1,$e,#`64-@Sigma1[1]`
+ vmov $Ch,$e
vsli.64 $t2,$e,#`64-@Sigma1[2]`
#if $i<16 && defined(__ARMEL__)
vrev64.8 @X[$i],@X[$i]
#endif
- vadd.i64 $T1,$K,$h
- veor $Ch,$f,$g
- veor $t0,$t1
- vand $Ch,$e
- veor $t0,$t2 @ Sigma1(e)
- veor $Ch,$g @ Ch(e,f,g)
- vadd.i64 $T1,$t0
+ veor $t1,$t0
+ vbsl $Ch,$f,$g @ Ch(e,f,g)
vshr.u64 $t0,$a,#@Sigma0[0]
- vadd.i64 $T1,$Ch
+ veor $t2,$t1 @ Sigma1(e)
+ vadd.i64 $T1,$Ch,$h
vshr.u64 $t1,$a,#@Sigma0[1]
- vshr.u64 $t2,$a,#@Sigma0[2]
vsli.64 $t0,$a,#`64-@Sigma0[0]`
+ vadd.i64 $T1,$t2
+ vshr.u64 $t2,$a,#@Sigma0[2]
+ vadd.i64 $K,@X[$i%16]
vsli.64 $t1,$a,#`64-@Sigma0[1]`
+ veor $Maj,$a,$b
vsli.64 $t2,$a,#`64-@Sigma0[2]`
- vadd.i64 $T1,@X[$i%16]
- vorr $Maj,$a,$c
- vand $Ch,$a,$c
veor $h,$t0,$t1
- vand $Maj,$b
+ vadd.i64 $T1,$K
+ vbsl $Maj,$c,$b @ Maj(a,b,c)
veor $h,$t2 @ Sigma0(a)
- vorr $Maj,$Ch @ Maj(a,b,c)
- vadd.i64 $h,$T1
vadd.i64 $d,$T1
- vadd.i64 $h,$Maj
+ vadd.i64 $Maj,$T1
+ @ vadd.i64 $h,$Maj
___
}
$code.=<<___;
vshr.u64 $t0,@X[($i+7)%8],#@sigma1[0]
vshr.u64 $t1,@X[($i+7)%8],#@sigma1[1]
+ vadd.i64 @_[0],d30 @ h+=Maj from the past
vshr.u64 $s1,@X[($i+7)%8],#@sigma1[2]
vsli.64 $t0,@X[($i+7)%8],#`64-@sigma1[0]`
vext.8 $s0,@X[$i%8],@X[($i+1)%8],#8 @ X[i+1]
$code.=<<___;
bne .L16_79_neon
+ vadd.i64 $A,d30 @ h+=Maj from the past
vldmia $ctx,{d24-d31} @ load context to temp
vadd.i64 q8,q12 @ vectorized accumulate
vadd.i64 q9,q13
projects / openssl / commitdiff