7237N/A# ==================================================================== 7237N/A# Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL 7237N/A# project. The module is, however, dual licensed under OpenSSL and 7237N/A# CRYPTOGAMS licenses depending on where you obtain it. For further 7237N/A# Hardware SPARC T4 support by David S. Miller <davem@davemloft.net>. 7237N/A# ==================================================================== 7237N/A# SHA256 performance improvement over compiler generated code varies 7237N/A# from 40% for Sun C [32-bit build] to 70% for gcc [3.3, 64-bit 7237N/A# build]. Just like in SHA1 module I aim to ensure scalability on 7237N/A# UltraSPARC T1 by packing X[16] to 8 64-bit registers. 7237N/A# SHA512 on pre-T1 UltraSPARC. 7237N/A# Performance is >75% better than 64-bit code generated by Sun C and 7237N/A# over 2x than 32-bit code. X[16] resides on stack, but access to it 7237N/A# is scheduled for L2 latency and staged through 32 least significant 7237N/A# bits of %l0-%l7. The latter is done to achieve 32-/64-bit ABI 7237N/A# duality. Nevetheless it's ~40% faster than SHA256, which is pretty 7237N/A# good [optimal coefficient is 50%]. 7237N/A# It's not any faster than 64-bit code generated by Sun C 5.8. This is 7237N/A# because 64-bit code generator has the advantage of using 64-bit 7237N/A# loads(*) to access X[16], which I consciously traded for 32-/64-bit 7237N/A# ABI duality [as per above]. But it surpasses 32-bit Sun C generated 7237N/A# code by 60%, not to mention that it doesn't suffer from severe decay 7237N/A# when running 4 times physical cores threads and that it leaves gcc 7237N/A# [3.4] behind by over 4x factor! If compared to SHA256, single thread 7237N/A# performance is only 10% better, but overall throughput for maximum 7237N/A# amount of threads for given CPU exceeds corresponding one of SHA256 7237N/A# by 30% [again, optimal coefficient is 50%]. 7237N/A# (*) Unlike pre-T1 UltraSPARC loads on T1 are executed strictly 7237N/A# in-order, i.e. load instruction has to complete prior next 7237N/A# instruction in given thread is executed, even if the latter is 7237N/A# not dependent on load result! This means that on T1 two 32-bit 7237N/A# loads are always slower than one 64-bit load. Once again this 7237N/A# is unlike pre-T1 UltraSPARC, where, if scheduled appropriately, 7237N/A# 2x32-bit loads can be as fast as 1x64-bit ones. 7237N/A# SPARC T4 SHA256/512 hardware achieves 3.17/2.01 cycles per byte, 7237N/A# which is 9.3x/11.1x faster than software. Multi-process benchmark 7237N/A# saturates at 11.5x single-process result on 8-core processor, or 7237N/A# ~11/16GBps per 2.85GHz socket. 7237N/A @V=($A,$B,$C,$D,$E,$F,$G,$H);
7237N/A @X=(
"%o0",
"%o1",
"%o2",
"%o3",
"%o4",
"%o5",
"%g1",
"%o7");
7237N/A @V=($A,$B,$C,$D,$E,$F,$G,$H);
7237N/Amy ($i,$a,$b,$c,$d,$e,$f,$g,$h)=@_;
7237N/A $
code.=
"\tsrlx @X[$i/2],32,$T1\n\tadd $h,$T1,$T1\n";
7237N/Amy ($i,$a,$b,$c,$d,$e,$f,$g,$h)=@_;
7237N/Amy @
pair=(
"%l".
eval(($i*
2)%
8),
"%l".
eval(($i*
2)%
8+
1),
"%l".
eval((($i+
1)*
2)%
8));
7237N/A `
"ld [$inp+".
eval(
32+
0+$i*
8).
"],@pair[0]" if ($i<
12)`
7237N/A `
"ld [$inp+".
eval(
32+
4+$i*
8).
"],@pair[1]" if ($i<
12)`
7237N/Amy ($i,$a,$b,$c,$d,$e,$f,$g,$h)=@_;
7237N/Amy @
pair=(
"%l".
eval(($i*
2)%
8),
"%l".
eval(($i*
2)%
8+
1));
7237N/A .
long 0x428a2f98,
0x71374491,
0xb5c0fbcf,
0xe9b5dba5 7237N/A .
long 0x3956c25b,
0x59f111f1,
0x923f82a4,
0xab1c5ed5 7237N/A .
long 0xd807aa98,
0x12835b01,
0x243185be,
0x550c7dc3 7237N/A .
long 0x72be5d74,
0x80deb1fe,
0x9bdc06a7,
0xc19bf174 7237N/A .
long 0xe49b69c1,
0xefbe4786,
0x0fc19dc6,
0x240ca1cc 7237N/A .
long 0x2de92c6f,
0x4a7484aa,
0x5cb0a9dc,
0x76f988da 7237N/A .
long 0x983e5152,
0xa831c66d,
0xb00327c8,
0xbf597fc7 7237N/A .
long 0xc6e00bf3,
0xd5a79147,
0x06ca6351,
0x14292967 7237N/A .
long 0x27b70a85,
0x2e1b2138,
0x4d2c6dfc,
0x53380d13 7237N/A .
long 0x650a7354,
0x766a0abb,
0x81c2c92e,
0x92722c85 7237N/A .
long 0xa2bfe8a1,
0xa81a664b,
0xc24b8b70,
0xc76c51a3 7237N/A .
long 0xd192e819,
0xd6990624,
0xf40e3585,
0x106aa070 7237N/A .
long 0x19a4c116,
0x1e376c08,
0x2748774c,
0x34b0bcb5 7237N/A .
long 0x391c0cb3,
0x4ed8aa4a,
0x5b9cca4f,
0x682e6ff3 7237N/A .
long 0x748f82ee,
0x78a5636f,
0x84c87814,
0x8cc70208 7237N/A .
long 0x90befffa,
0xa4506ceb,
0xbef9a3f7,
0xc67178f2 7237N/A .
long 0x428a2f98,
0xd728ae22,
0x71374491,
0x23ef65cd 7237N/A .
long 0xb5c0fbcf,
0xec4d3b2f,
0xe9b5dba5,
0x8189dbbc 7237N/A .
long 0x3956c25b,
0xf348b538,
0x59f111f1,
0xb605d019 7237N/A .
long 0x923f82a4,
0xaf194f9b,
0xab1c5ed5,
0xda6d8118 7237N/A .
long 0xd807aa98,
0xa3030242,
0x12835b01,
0x45706fbe 7237N/A .
long 0x243185be,
0x4ee4b28c,
0x550c7dc3,
0xd5ffb4e2 7237N/A .
long 0x72be5d74,
0xf27b896f,
0x80deb1fe,
0x3b1696b1 7237N/A .
long 0x9bdc06a7,
0x25c71235,
0xc19bf174,
0xcf692694 7237N/A .
long 0xe49b69c1,
0x9ef14ad2,
0xefbe4786,
0x384f25e3 7237N/A .
long 0x0fc19dc6,
0x8b8cd5b5,
0x240ca1cc,
0x77ac9c65 7237N/A .
long 0x2de92c6f,
0x592b0275,
0x4a7484aa,
0x6ea6e483 7237N/A .
long 0x5cb0a9dc,
0xbd41fbd4,
0x76f988da,
0x831153b5 7237N/A .
long 0x983e5152,
0xee66dfab,
0xa831c66d,
0x2db43210 7237N/A .
long 0xb00327c8,
0x98fb213f,
0xbf597fc7,
0xbeef0ee4 7237N/A .
long 0xc6e00bf3,
0x3da88fc2,
0xd5a79147,
0x930aa725 7237N/A .
long 0x06ca6351,
0xe003826f,
0x14292967,
0x0a0e6e70 7237N/A .
long 0x27b70a85,
0x46d22ffc,
0x2e1b2138,
0x5c26c926 7237N/A .
long 0x4d2c6dfc,
0x5ac42aed,
0x53380d13,
0x9d95b3df 7237N/A .
long 0x650a7354,
0x8baf63de,
0x766a0abb,
0x3c77b2a8 7237N/A .
long 0x81c2c92e,
0x47edaee6,
0x92722c85,
0x1482353b 7237N/A .
long 0xa2bfe8a1,
0x4cf10364,
0xa81a664b,
0xbc423001 7237N/A .
long 0xc24b8b70,
0xd0f89791,
0xc76c51a3,
0x0654be30 7237N/A .
long 0xd192e819,
0xd6ef5218,
0xd6990624,
0x5565a910 7237N/A .
long 0xf40e3585,
0x5771202a,
0x106aa070,
0x32bbd1b8 7237N/A .
long 0x19a4c116,
0xb8d2d0c8,
0x1e376c08,
0x5141ab53 7237N/A .
long 0x2748774c,
0xdf8eeb99,
0x34b0bcb5,
0xe19b48a8 7237N/A .
long 0x391c0cb3,
0xc5c95a63,
0x4ed8aa4a,
0xe3418acb 7237N/A .
long 0x5b9cca4f,
0x7763e373,
0x682e6ff3,
0xd6b2b8a3 7237N/A .
long 0x748f82ee,
0x5defb2fc,
0x78a5636f,
0x43172f60 7237N/A .
long 0x84c87814,
0xa1f0ab72,
0x8cc70208,
0x1a6439ec 7237N/A .
long 0x90befffa,
0x23631e28,
0xa4506ceb,
0xde82bde9 7237N/A .
long 0xbef9a3f7,
0xb2c67915,
0xc67178f2,
0xe372532b 7237N/A .
long 0xca273ece,
0xea26619c,
0xd186b8c7,
0x21c0c207 7237N/A .
long 0xeada7dd6,
0xcde0eb1e,
0xf57d4f7f,
0xee6ed178 7237N/A .
long 0x06f067aa,
0x72176fba,
0x0a637dc5,
0xa2c898a6 7237N/A .
long 0x113f9804,
0xbef90dae,
0x1b710b35,
0x131c471b 7237N/A .
long 0x28db77f5,
0x23047d84,
0x32caab7b,
0x40c72493 7237N/A .
long 0x3c9ebe0a,
0x15c9bebc,
0x431d67c4,
0x9c100d4c 7237N/A .
long 0x4cc5d4be,
0xcb3e42b6,
0x597f299c,
0xfc657e2a 7237N/A .
long 0x5fcb6fab,
0x3ad6faec,
0x6c44198c,
0x4a475817 7237N/A.
asciz "SHA${label} block transform for SPARCv9, CRYPTOGAMS by <appro\@openssl.org>" 7237N/A# Purpose of these subroutines is to explicitly encode VIS instructions, 7237N/A# so that one can compile the module without having to specify VIS 7237N/A# extentions on compiler command line, e.g. -xarch=v9 vs. -xarch=v9a. 7237N/A# Idea is to reserve for option to produce "universal" binary and let 7237N/A# programmer detect if current CPU is VIS capable at run-time. 7237N/A # re-encode for upper double register addressing 7237N/A return sprintf ".word\t0x%08x !%s",
7237N/Amy %
bias = (
"g" =>
0,
"o" =>
8,
"l" =>
16,
"i" =>
24 );
7237N/A return sprintf ".word\t0x%08x !%s",
7237N/A s/\b(f[^\s]*)\s+(%f[
0-
9]{
1,
2}),\s*(%f[
0-
9]{
1,
2}),\s*(%f[
0-
9]{
1,
2})/