maxmustermann/minio
1
0
Fork 0
Code Issues Pull requests Projects Releases Wiki Activity

Update reedsolomon/highwayhash to start using ppc64le support (#7003)

Browse source 
Thanks to @fwessels for the upstream work on reedsolomon and
highwayhash which has resulted in 10x performance improvement
on ppc64 architecture.
This commit is contained in:
Harshavardhana 2018-12-20 09:47:05 -08:00 committed by Nitish Tiwari
parent bc67410548
commit def04f01cf
17 changed files with 460 additions and 14 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

18
vendor/github.com/klauspost/reedsolomon/README.md generated vendored
View file

@ -24,6 +24,10 @@ go get -u github.com/klauspost/reedsolomon
# Changes
## December 18, 2018
Assembly code for ppc64le has been contributed, this boosts performance by about 10x on this platform.
## November 18, 2017
Added [WithAutoGoroutines](https://godoc.org/github.com/klauspost/reedsolomon#WithAutoGoroutines) which will attempt to calculate the optimal number of goroutines to use based on your expected shard size and detected CPU.
@ -259,6 +263,18 @@ By exploiting NEON instructions the performance for ARM has been accelerated. Be
| 10 | 2 | 20% | 188 | 1738 | 925% |
| 10 | 4 | 40% | 96 | 839 | 877% |
# Performance on ppc64le
The performance for ppc64le has been accelerated. This gives roughly a 10x performance improvement on this architecture as can been seen below:
```
benchmark old MB/s new MB/s speedup
BenchmarkGalois128K-160 948.87 8878.85 9.36x
BenchmarkGalois1M-160 968.85 9041.92 9.33x
BenchmarkGaloisXor128K-160 862.02 7905.00 9.17x
BenchmarkGaloisXor1M-160 784.60 6296.65 8.03x
```
# asm2plan9s
[asm2plan9s](https://github.com/fwessels/asm2plan9s) is used for assembling the AVX2 instructions into their BYTE/WORD/LONG equivalents.
@ -266,8 +282,10 @@ By exploiting NEON instructions the performance for ARM has been accelerated. Be
# Links
* [Backblaze Open Sources Reed-Solomon Erasure Coding Source Code](https://www.backblaze.com/blog/reed-solomon/).
* [JavaReedSolomon](https://github.com/Backblaze/JavaReedSolomon). Compatible java library by Backblaze.
* [ocaml-reed-solomon-erasure](https://gitlab.com/darrenldl/ocaml-reed-solomon-erasure). Compatible OCaml implementation.
* [reedsolomon-c](https://github.com/jannson/reedsolomon-c). C version, compatible with output from this package.
* [Reed-Solomon Erasure Coding in Haskell](https://github.com/NicolasT/reedsolomon). Haskell port of the package with similar performance.
* [reed-solomon-erasure](https://github.com/darrenldl/reed-solomon-erasure). Compatible Rust implementation.
* [go-erasure](https://github.com/somethingnew2-0/go-erasure). A similar library using cgo, slower in my tests.
* [rsraid](https://github.com/goayame/rsraid). A similar library written in Go. Slower, but supports more shards.
* [Screaming Fast Galois Field Arithmetic](http://www.snia.org/sites/default/files2/SDC2013/presentations/NewThinking/EthanMiller_Screaming_Fast_Galois_Field%20Arithmetic_SIMD%20Instructions.pdf). Basis for SSE3 optimizations.

1
vendor/github.com/klauspost/reedsolomon/galois_amd64.go generated vendored
View file

@ -1,5 +1,6 @@
//+build !noasm
//+build !appengine
//+build !gccgo
// Copyright 2015, Klaus Post, see LICENSE for details.

2
vendor/github.com/klauspost/reedsolomon/galois_amd64.s generated vendored
View file

@ -1,4 +1,4 @@
//+build !noasm !appengine
//+build !noasm !appengine !gccgo
// Copyright 2015, Klaus Post, see LICENSE for details.

1
vendor/github.com/klauspost/reedsolomon/galois_arm64.go generated vendored
View file

@ -1,5 +1,6 @@
//+build !noasm
//+build !appengine
//+build !gccgo
// Copyright 2015, Klaus Post, see LICENSE for details.
// Copyright 2017, Minio, Inc.

2
vendor/github.com/klauspost/reedsolomon/galois_arm64.s generated vendored
View file

@ -1,4 +1,4 @@
//+build !noasm !appengine
//+build !noasm !appengine !gccgo
// Copyright 2015, Klaus Post, see LICENSE for details.
// Copyright 2017, Minio, Inc.

5
vendor/github.com/klauspost/reedsolomon/galois_noasm.go generated vendored
View file

@ -1,5 +1,6 @@
//+build !amd64 noasm appengine
//+build !arm64 noasm appengine
//+build !amd64 noasm appengine gccgo
//+build !arm64 noasm appengine gccgo
//+build !ppc64le noasm appengine gccgo
// Copyright 2015, Klaus Post, see LICENSE for details.

67
vendor/github.com/klauspost/reedsolomon/galois_ppc64le.go generated vendored Normal file
View file

@ -0,0 +1,67 @@
//+build !noasm
//+build !appengine
//+build !gccgo
// Copyright 2015, Klaus Post, see LICENSE for details.
// Copyright 2018, Minio, Inc.
package reedsolomon
//go:noescape
func galMulPpc(low, high, in, out []byte)
//go:noescape
func galMulPpcXor(low, high, in, out []byte)
// This is what the assembler routines do in blocks of 16 bytes:
/*
func galMulPpc(low, high, in, out []byte) {
for n, input := range in {
l := input & 0xf
h := input >> 4
out[n] = low[l] ^ high[h]
}
}
func galMulPpcXor(low, high, in, out []byte) {
for n, input := range in {
l := input & 0xf
h := input >> 4
out[n] ^= low[l] ^ high[h]
}
}
*/
func galMulSlice(c byte, in, out []byte, ssse3, avx2 bool) {
done := (len(in) >> 4) << 4
if done > 0 {
galMulPpc(mulTableLow[c][:], mulTableHigh[c][:], in[:done], out)
}
remain := len(in) - done
if remain > 0 {
mt := mulTable[c]
for i := done; i < len(in); i++ {
out[i] = mt[in[i]]
}
}
}
func galMulSliceXor(c byte, in, out []byte, ssse3, avx2 bool) {
done := (len(in) >> 4) << 4
if done > 0 {
galMulPpcXor(mulTableLow[c][:], mulTableHigh[c][:], in[:done], out)
}
remain := len(in) - done
if remain > 0 {
mt := mulTable[c]
for i := done; i < len(in); i++ {
out[i] ^= mt[in[i]]
}
}
}
// slice galois add
func sliceXor(in, out []byte, sse2 bool) {
for n, input := range in {
out[n] ^= input
}
}

126
vendor/github.com/klauspost/reedsolomon/galois_ppc64le.s generated vendored Normal file
View file

@ -0,0 +1,126 @@
//+build !noasm !appengine !gccgo
// Copyright 2015, Klaus Post, see LICENSE for details.
// Copyright 2018, Minio, Inc.
#include "textflag.h"
#define LOW R3
#define HIGH R4
#define IN R5
#define LEN R6
#define OUT R7
#define CONSTANTS R8
#define OFFSET R9
#define OFFSET1 R10
#define OFFSET2 R11
#define X6 VS34
#define X6_ V2
#define X7 VS35
#define X7_ V3
#define MSG VS36
#define MSG_ V4
#define MSG_HI VS37
#define MSG_HI_ V5
#define RESULT VS38
#define RESULT_ V6
#define ROTATE VS39
#define ROTATE_ V7
#define MASK VS40
#define MASK_ V8
#define FLIP VS41
#define FLIP_ V9
// func galMulPpc(low, high, in, out []byte)
TEXT ·galMulPpc(SB), NOFRAME|NOSPLIT, $0-96
MOVD low+0(FP), LOW
MOVD high+24(FP), HIGH
MOVD in+48(FP), IN
MOVD in_len+56(FP), LEN
MOVD out+72(FP), OUT
MOVD $16, OFFSET1
MOVD $32, OFFSET2
MOVD $·constants(SB), CONSTANTS
LXVD2X (CONSTANTS)(R0), ROTATE
LXVD2X (CONSTANTS)(OFFSET1), MASK
LXVD2X (CONSTANTS)(OFFSET2), FLIP
LXVD2X (LOW)(R0), X6
LXVD2X (HIGH)(R0), X7
VPERM X6_, V31, FLIP_, X6_
VPERM X7_, V31, FLIP_, X7_
MOVD $0, OFFSET
loop:
LXVD2X (IN)(OFFSET), MSG
VSRB MSG_, ROTATE_, MSG_HI_
VAND MSG_, MASK_, MSG_
VPERM X6_, V31, MSG_, MSG_
VPERM X7_, V31, MSG_HI_, MSG_HI_
VXOR MSG_, MSG_HI_, MSG_
STXVD2X MSG, (OUT)(OFFSET)
ADD $16, OFFSET, OFFSET
CMP LEN, OFFSET
BGT loop
RET
// func galMulPpcXorlow, high, in, out []byte)
TEXT ·galMulPpcXor(SB), NOFRAME|NOSPLIT, $0-96
MOVD low+0(FP), LOW
MOVD high+24(FP), HIGH
MOVD in+48(FP), IN
MOVD in_len+56(FP), LEN
MOVD out+72(FP), OUT
MOVD $16, OFFSET1
MOVD $32, OFFSET2
MOVD $·constants(SB), CONSTANTS
LXVD2X (CONSTANTS)(R0), ROTATE
LXVD2X (CONSTANTS)(OFFSET1), MASK
LXVD2X (CONSTANTS)(OFFSET2), FLIP
LXVD2X (LOW)(R0), X6
LXVD2X (HIGH)(R0), X7
VPERM X6_, V31, FLIP_, X6_
VPERM X7_, V31, FLIP_, X7_
MOVD $0, OFFSET
loopXor:
LXVD2X (IN)(OFFSET), MSG
LXVD2X (OUT)(OFFSET), RESULT
VSRB MSG_, ROTATE_, MSG_HI_
VAND MSG_, MASK_, MSG_
VPERM X6_, V31, MSG_, MSG_
VPERM X7_, V31, MSG_HI_, MSG_HI_
VXOR MSG_, MSG_HI_, MSG_
VXOR MSG_, RESULT_, RESULT_
STXVD2X RESULT, (OUT)(OFFSET)
ADD $16, OFFSET, OFFSET
CMP LEN, OFFSET
BGT loopXor
RET
DATA ·constants+0x0(SB)/8, $0x0404040404040404
DATA ·constants+0x8(SB)/8, $0x0404040404040404
DATA ·constants+0x10(SB)/8, $0x0f0f0f0f0f0f0f0f
DATA ·constants+0x18(SB)/8, $0x0f0f0f0f0f0f0f0f
DATA ·constants+0x20(SB)/8, $0x0706050403020100
DATA ·constants+0x28(SB)/8, $0x0f0e0d0c0b0a0908
GLOBL ·constants(SB), 8, $48

6
vendor/github.com/klauspost/reedsolomon/reedsolomon.go generated vendored
View file

@ -471,12 +471,12 @@ func (r reedSolomon) codeSomeShardsP(matrixRows, inputs, outputs [][]byte, outpu
wg.Add(1)
go func(start, stop int) {
for c := 0; c < r.DataShards; c++ {
in := inputs[c]
in := inputs[c][start:stop]
for iRow := 0; iRow < outputCount; iRow++ {
if c == 0 {
galMulSlice(matrixRows[iRow][c], in[start:stop], outputs[iRow][start:stop], r.o.useSSSE3, r.o.useAVX2)
galMulSlice(matrixRows[iRow][c], in, outputs[iRow][start:stop], r.o.useSSSE3, r.o.useAVX2)
} else {
galMulSliceXor(matrixRows[iRow][c], in[start:stop], outputs[iRow][start:stop], r.o.useSSSE3, r.o.useAVX2)
galMulSliceXor(matrixRows[iRow][c], in, outputs[iRow][start:stop], r.o.useSSSE3, r.o.useAVX2)
}
}
}

13
vendor/github.com/minio/highwayhash/README.md generated vendored
View file

@ -7,7 +7,7 @@
It can be used to prevent hash-flooding attacks or authenticate short-lived messages. Additionally it can be used as a fingerprinting function. HighwayHash is not a general purpose cryptographic hash function (such as Blake2b, SHA-3 or SHA-2) and should not be used if strong collision resistance is required.
This repository contains a native Go version and optimized assembly implementations on both Intel and ARM platforms.
This repository contains a native Go version and optimized assembly implementations for Intel, ARM and ppc64le architectures.
### High performance
@ -50,6 +50,17 @@ ARM64 NEON | 384 MB/s | 955 MB/s | 1053 MB/s
*Note: For now just the (main) update loop is implemented in assembly, so for small messages there is still considerable overhead due to initialization and finalization.*
### ppc64le Performance
The ppc64le accelerated version is roughly 10x faster compared to the non-optimized version:
```
benchmark old MB/s new MB/s speedup
BenchmarkWrite_8K 531.19 5566.41 10.48x
BenchmarkSum64_8K 518.86 4971.88 9.58x
BenchmarkSum256_8K 502.45 4474.20 8.90x
```
### Performance compared to other hashing techniques
On a Skylake CPU (3.0 GHz Xeon Platinum 8124M) the table below shows how HighwayHash compares to other hashing techniques for 5 MB messages (single core performance, all Golang implementations, see [benchmark](https://github.com/fwessels/HashCompare/blob/master/benchmarks_test.go)).

1
vendor/github.com/minio/highwayhash/highwayhashAVX2_amd64.go generated vendored
View file

@ -13,6 +13,7 @@ var (
useSSE4 = cpu.X86.HasSSE41
useAVX2 = cpu.X86.HasAVX2
useNEON = false
useVMX = false
)
//go:noescape

1
vendor/github.com/minio/highwayhash/highwayhash_amd64.go generated vendored
View file

@ -13,6 +13,7 @@ var (
useSSE4 = cpu.X86.HasSSE41
useAVX2 = false
useNEON = false
useVMX = false
)
//go:noescape

1
vendor/github.com/minio/highwayhash/highwayhash_arm64.go generated vendored
View file

@ -10,6 +10,7 @@ var (
useSSE4 = false
useAVX2 = false
useNEON = true
useVMX = false
)
//go:noescape

33
vendor/github.com/minio/highwayhash/highwayhash_ppc64le.go generated vendored Normal file
View file

@ -0,0 +1,33 @@
//+build !noasm
// Copyright (c) 2017 Minio Inc. All rights reserved.
// Use of this source code is governed by a license that can be
// found in the LICENSE file.
package highwayhash
var (
useSSE4 = false
useAVX2 = false
useNEON = false
useVMX = true
)
//go:noescape
func updatePpc64Le(state *[16]uint64, msg []byte)
func initialize(state *[16]uint64, key []byte) {
initializeGeneric(state, key)
}
func update(state *[16]uint64, msg []byte) {
if useVMX {
updatePpc64Le(state, msg)
} else {
updateGeneric(state, msg)
}
}
func finalize(out []byte, state *[16]uint64) {
finalizeGeneric(out, state)
}

183
vendor/github.com/minio/highwayhash/highwayhash_ppc64le.s generated vendored Normal file
View file

@ -0,0 +1,183 @@
//+build !noasm !appengine
//
// Minio Cloud Storage, (C) 2018 Minio, Inc.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
//
#include "textflag.h"
// Definition of registers
#define V0_LO VS32
#define V0_LO_ V0
#define V0_HI VS33
#define V0_HI_ V1
#define V1_LO VS34
#define V1_LO_ V2
#define V1_HI VS35
#define V1_HI_ V3
#define MUL0_LO VS36
#define MUL0_LO_ V4
#define MUL0_HI VS37
#define MUL0_HI_ V5
#define MUL1_LO VS38
#define MUL1_LO_ V6
#define MUL1_HI VS39
#define MUL1_HI_ V7
// Message
#define MSG_LO VS40
#define MSG_LO_ V8
#define MSG_HI VS41
// Constants
#define ROTATE VS42
#define ROTATE_ V10
#define MASK VS43
#define MASK_ V11
// Temps
#define TEMP1 VS44
#define TEMP1_ V12
#define TEMP2 VS45
#define TEMP2_ V13
#define TEMP3 VS46
#define TEMP3_ V14
#define TEMP4_ V15
#define TEMP5_ V16
#define TEMP6_ V17
#define TEMP7_ V18
// Regular registers
#define STATE R3
#define MSG_BASE R4
#define MSG_LEN R5
#define CONSTANTS R6
#define P1 R7
#define P2 R8
#define P3 R9
#define P4 R10
#define P5 R11
#define P6 R12
#define P7 R14 // avoid using R13
TEXT ·updatePpc64Le(SB), NOFRAME|NOSPLIT, $0-32
MOVD state+0(FP), STATE
MOVD msg_base+8(FP), MSG_BASE
MOVD msg_len+16(FP), MSG_LEN // length of message
// Sanity check for length
CMPU MSG_LEN, $31
BLE complete
// Setup offsets
MOVD $16, P1
MOVD $32, P2
MOVD $48, P3
MOVD $64, P4
MOVD $80, P5
MOVD $96, P6
MOVD $112, P7
// Load state
LXVD2X (STATE)(R0), V0_LO
LXVD2X (STATE)(P1), V0_HI
LXVD2X (STATE)(P2), V1_LO
LXVD2X (STATE)(P3), V1_HI
LXVD2X (STATE)(P4), MUL0_LO
LXVD2X (STATE)(P5), MUL0_HI
LXVD2X (STATE)(P6), MUL1_LO
LXVD2X (STATE)(P7), MUL1_HI
XXPERMDI V0_LO, V0_LO, $2, V0_LO
XXPERMDI V0_HI, V0_HI, $2, V0_HI
XXPERMDI V1_LO, V1_LO, $2, V1_LO
XXPERMDI V1_HI, V1_HI, $2, V1_HI
XXPERMDI MUL0_LO, MUL0_LO, $2, MUL0_LO
XXPERMDI MUL0_HI, MUL0_HI, $2, MUL0_HI
XXPERMDI MUL1_LO, MUL1_LO, $2, MUL1_LO
XXPERMDI MUL1_HI, MUL1_HI, $2, MUL1_HI
// Load constants table pointer
MOVD $·constants(SB), CONSTANTS
LXVD2X (CONSTANTS)(R0), ROTATE
LXVD2X (CONSTANTS)(P1), MASK
XXLNAND MASK, MASK, MASK
loop:
// Main highwayhash update loop
LXVD2X (MSG_BASE)(R0), MSG_LO
VADDUDM V0_LO_, MUL1_LO_, TEMP1_
VRLD V0_LO_, ROTATE_, TEMP2_
VADDUDM MUL1_HI_, V0_HI_, TEMP3_
LXVD2X (MSG_BASE)(P1), MSG_HI
ADD $32, MSG_BASE, MSG_BASE
XXPERMDI MSG_LO, MSG_LO, $2, MSG_LO
XXPERMDI MSG_HI, MSG_HI, $2, V0_LO
VADDUDM MSG_LO_, MUL0_LO_, MSG_LO_
VADDUDM V0_LO_, MUL0_HI_, V0_LO_
VADDUDM MSG_LO_, V1_LO_, V1_LO_
VSRD V0_HI_, ROTATE_, MSG_LO_
VADDUDM V0_LO_, V1_HI_, V1_HI_
VPERM V1_LO_, V1_LO_, MASK_, V0_LO_
VMULOUW V1_LO_, TEMP2_, TEMP2_
VPERM V1_HI_, V1_HI_, MASK_, TEMP7_
VADDUDM V0_LO_, TEMP1_, V0_LO_
VMULOUW V1_HI_, MSG_LO_, MSG_LO_
VADDUDM TEMP7_, TEMP3_, V0_HI_
VPERM V0_LO_, V0_LO_, MASK_, TEMP6_
VRLD V1_LO_, ROTATE_, TEMP4_
VSRD V1_HI_, ROTATE_, TEMP5_
VPERM V0_HI_, V0_HI_, MASK_, TEMP7_
XXLXOR MUL0_LO, TEMP2, MUL0_LO
VMULOUW TEMP1_, TEMP4_, TEMP1_
VMULOUW TEMP3_, TEMP5_, TEMP3_
XXLXOR MUL0_HI, MSG_LO, MUL0_HI
XXLXOR MUL1_LO, TEMP1, MUL1_LO
XXLXOR MUL1_HI, TEMP3, MUL1_HI
VADDUDM TEMP6_, V1_LO_, V1_LO_
VADDUDM TEMP7_, V1_HI_, V1_HI_
SUB $32, MSG_LEN, MSG_LEN
CMPU MSG_LEN, $32
BGE loop
// Save state
XXPERMDI V0_LO, V0_LO, $2, V0_LO
XXPERMDI V0_HI, V0_HI, $2, V0_HI
XXPERMDI V1_LO, V1_LO, $2, V1_LO
XXPERMDI V1_HI, V1_HI, $2, V1_HI
XXPERMDI MUL0_LO, MUL0_LO, $2, MUL0_LO
XXPERMDI MUL0_HI, MUL0_HI, $2, MUL0_HI
XXPERMDI MUL1_LO, MUL1_LO, $2, MUL1_LO
XXPERMDI MUL1_HI, MUL1_HI, $2, MUL1_HI
STXVD2X V0_LO, (STATE)(R0)
STXVD2X V0_HI, (STATE)(P1)
STXVD2X V1_LO, (STATE)(P2)
STXVD2X V1_HI, (STATE)(P3)
STXVD2X MUL0_LO, (STATE)(P4)
STXVD2X MUL0_HI, (STATE)(P5)
STXVD2X MUL1_LO, (STATE)(P6)
STXVD2X MUL1_HI, (STATE)(P7)
complete:
RET
// Constants table
DATA ·constants+0x0(SB)/8, $0x0000000000000020
DATA ·constants+0x8(SB)/8, $0x0000000000000020
DATA ·constants+0x10(SB)/8, $0x070806090d0a040b // zipper merge constant
DATA ·constants+0x18(SB)/8, $0x000f010e05020c03 // zipper merge constant
GLOBL ·constants(SB), 8, $32

2
vendor/github.com/minio/highwayhash/highwayhash_ref.go generated vendored
View file

@ -4,6 +4,7 @@
// +build !amd64
// +build !arm64
// +build !ppc64le
package highwayhash
@ -11,6 +12,7 @@ var (
useSSE4 = false
useAVX2 = false
useNEON = false
useVMX = false
)
func initialize(state *[16]uint64, k []byte) {

12
vendor/vendor.json vendored
View file

@ -563,10 +563,10 @@
"revisionTime": "2017-10-07T12:43:06Z"
},
{
"checksumSHA1": "ehsrWipiGIWqa4To8TmelIx06vI=",
"checksumSHA1": "KiQa3vguztElzJkoqeIGHlfLFJA=",
"path": "github.com/klauspost/reedsolomon",
"revision": "0b30fa71cc8e4e9010c9aba6d0320e2e5b163b29",
"revisionTime": "2017-12-19T13:34:37Z"
"revision": "8885f3a1c73882e6f11b766242c69a1eb8f44b28",
"revisionTime": "2018-12-18T19:39:59Z"
},
{
"checksumSHA1": "xxLSo5tKtXc7jGrR70yoEfza8Cw=",
@ -634,10 +634,10 @@
"revisionTime": "2018-01-23T12:12:34Z"
},
{
"checksumSHA1": "2Fu1GmLwDo6FFdahjnlWnPkwJTE=",
"checksumSHA1": "CD2MtlgA8h0z6hYJHURS5eOmZ1k=",
"path": "github.com/minio/highwayhash",
"revision": "85fc8a2dacad36a6beb2865793cd81363a496696",
"revisionTime": "2018-05-01T08:09:13Z"
"revision": "93ed73d641695483ab4438817457b6586ee5765c",
"revisionTime": "2018-12-20T01:13:08Z"
},
{
"checksumSHA1": "7/Hdd23/j4/yt4BXa+h0kqz1yjw=",