Split blocks into size divisible by 16
Older systems (typically without AVX2) are more sensitive to misaligned load+stores. Add parameter to automatically set the number of goroutines. name old time/op new time/op delta Encode10x2x10000-8 18.4µs ± 1% 16.1µs ± 1% -12.43% (p=0.000 n=9+9) Encode100x20x10000-8 692µs ± 1% 608µs ± 1% -12.10% (p=0.000 n=10+10) Encode17x3x1M-8 1.78ms ± 5% 1.49ms ± 1% -16.63% (p=0.000 n=10+10) Encode10x4x16M-8 21.5ms ± 5% 19.6ms ± 4% -8.74% (p=0.000 n=10+9) Encode5x2x1M-8 343µs ± 2% 267µs ± 2% -22.22% (p=0.000 n=9+10) Encode10x2x1M-8 858µs ± 5% 701µs ± 5% -18.34% (p=0.000 n=10+10) Encode10x4x1M-8 1.34ms ± 1% 1.16ms ± 1% -13.19% (p=0.000 n=9+9) Encode50x20x1M-8 30.3ms ± 4% 25.0ms ± 2% -17.51% (p=0.000 n=10+8) Encode17x3x16M-8 26.9ms ± 1% 24.5ms ± 4% -9.13% (p=0.000 n=8+10) name old speed new speed delta Encode10x2x10000-8 5.45GB/s ± 1% 6.22GB/s ± 1% +14.20% (p=0.000 n=9+9) Encode100x20x10000-8 1.44GB/s ± 1% 1.64GB/s ± 1% +13.77% (p=0.000 n=10+10) Encode17x3x1M-8 10.0GB/s ± 5% 12.0GB/s ± 1% +19.88% (p=0.000 n=10+10) Encode10x4x16M-8 7.81GB/s ± 5% 8.56GB/s ± 5% +9.58% (p=0.000 n=10+9) Encode5x2x1M-8 15.3GB/s ± 2% 19.6GB/s ± 2% +28.57% (p=0.000 n=9+10) Encode10x2x1M-8 12.2GB/s ± 5% 15.0GB/s ± 5% +22.45% (p=0.000 n=10+10) Encode10x4x1M-8 7.84GB/s ± 1% 9.03GB/s ± 1% +15.19% (p=0.000 n=9+9) Encode50x20x1M-8 1.73GB/s ± 4% 2.09GB/s ± 4% +20.59% (p=0.000 n=10+9) Encode17x3x16M-8 10.6GB/s ± 1% 11.7GB/s ± 4% +10.12% (p=0.000 n=8+10)master
Klaus Post
parent
e52c150f96
commit
f5e73dcfe2
|
|
@ -24,6 +24,10 @@ go get -u github.com/klauspost/reedsolomon
|
|||
|
||||
# Changes
|
||||
|
||||
## November 18, 2017
|
||||
|
||||
Added [WithAutoGoroutines](https://godoc.org/github.com/klauspost/reedsolomon#WithAutoGoroutines) which will attempt to calculate the optimal number og goroutines to use based on your expected shard size and detected CPU.
|
||||
|
||||
## October 1, 2017
|
||||
|
||||
* [Cauchy Matrix](https://godoc.org/github.com/klauspost/reedsolomon#WithCauchyMatrix) is now an option. Thanks to [templexxx](https://github.com/templexxx) for the basis of this.
|
||||
|
|
|
|||
|
|
@ -19,8 +19,35 @@ TEXT ·galMulSSSE3Xor(SB), 7, $0
|
|||
MOVQ out+72(FP), DX // DX: &out
|
||||
PSHUFB X5, X8 // X8: lomask (unpacked)
|
||||
SHRQ $4, R9 // len(in) / 16
|
||||
MOVQ SI, AX
|
||||
MOVQ DX, BX
|
||||
ANDQ $15, AX
|
||||
ANDQ $15, BX
|
||||
CMPQ R9, $0
|
||||
JEQ done_xor
|
||||
ORQ AX, BX
|
||||
CMPQ BX, $0
|
||||
JNZ loopback_xor
|
||||
|
||||
loopback_xor_aligned:
|
||||
MOVOA (SI), X0 // in[x]
|
||||
MOVOA (DX), X4 // out[x]
|
||||
MOVOA X0, X1 // in[x]
|
||||
MOVOA X6, X2 // low copy
|
||||
MOVOA X7, X3 // high copy
|
||||
PSRLQ $4, X1 // X1: high input
|
||||
PAND X8, X0 // X0: low input
|
||||
PAND X8, X1 // X0: high input
|
||||
PSHUFB X0, X2 // X2: mul low part
|
||||
PSHUFB X1, X3 // X3: mul high part
|
||||
PXOR X2, X3 // X3: Result
|
||||
PXOR X4, X3 // X3: Result xor existing out
|
||||
MOVOA X3, (DX) // Store
|
||||
ADDQ $16, SI // in+=16
|
||||
ADDQ $16, DX // out+=16
|
||||
SUBQ $1, R9
|
||||
JNZ loopback_xor_aligned
|
||||
JMP done_xor
|
||||
|
||||
loopback_xor:
|
||||
MOVOU (SI), X0 // in[x]
|
||||
|
|
@ -57,15 +84,40 @@ TEXT ·galMulSSSE3(SB), 7, $0
|
|||
MOVQ in_len+56(FP), R9 // R9: len(in)
|
||||
MOVQ out+72(FP), DX // DX: &out
|
||||
PSHUFB X5, X8 // X8: lomask (unpacked)
|
||||
MOVQ SI, AX
|
||||
MOVQ DX, BX
|
||||
SHRQ $4, R9 // len(in) / 16
|
||||
ANDQ $15, AX
|
||||
ANDQ $15, BX
|
||||
CMPQ R9, $0
|
||||
JEQ done
|
||||
ORQ AX, BX
|
||||
CMPQ BX, $0
|
||||
JNZ loopback
|
||||
|
||||
loopback_aligned:
|
||||
MOVOA (SI), X0 // in[x]
|
||||
MOVOA X0, X1 // in[x]
|
||||
MOVOA X6, X2 // low copy
|
||||
MOVOA X7, X3 // high copy
|
||||
PSRLQ $4, X1 // X1: high input
|
||||
PAND X8, X0 // X0: low input
|
||||
PAND X8, X1 // X0: high input
|
||||
PSHUFB X0, X2 // X2: mul low part
|
||||
PSHUFB X1, X3 // X3: mul high part
|
||||
PXOR X2, X3 // X3: Result
|
||||
MOVOA X3, (DX) // Store
|
||||
ADDQ $16, SI // in+=16
|
||||
ADDQ $16, DX // out+=16
|
||||
SUBQ $1, R9
|
||||
JNZ loopback_aligned
|
||||
JMP done
|
||||
|
||||
loopback:
|
||||
MOVOU (SI), X0 // in[x]
|
||||
MOVOU X0, X1 // in[x]
|
||||
MOVOU X6, X2 // low copy
|
||||
MOVOU X7, X3 // high copy
|
||||
MOVOA X6, X2 // low copy
|
||||
MOVOA X7, X3 // high copy
|
||||
PSRLQ $4, X1 // X1: high input
|
||||
PAND X8, X0 // X0: low input
|
||||
PAND X8, X1 // X0: high input
|
||||
|
|
|
|||
15
options.go
15
options.go
|
|
@ -15,11 +15,12 @@ type options struct {
|
|||
useAVX2, useSSSE3, useSSE2 bool
|
||||
usePAR1Matrix bool
|
||||
useCauchy bool
|
||||
shardSize int
|
||||
}
|
||||
|
||||
var defaultOptions = options{
|
||||
maxGoroutines: 384,
|
||||
minSplitSize: 512,
|
||||
minSplitSize: 1024,
|
||||
}
|
||||
|
||||
func init() {
|
||||
|
|
@ -46,6 +47,18 @@ func WithMaxGoroutines(n int) Option {
|
|||
}
|
||||
}
|
||||
|
||||
// WithAutoGoroutines will adjust the number of goroutines for optimal speed with a
|
||||
// specific shard size.
|
||||
// Send in the shard size you expect to send. Other shard sizes will work, but may not
|
||||
// run at the optimal speed.
|
||||
// Overwrites WithMaxGoroutines.
|
||||
// If shardSize <= 0, it is ignored.
|
||||
func WithAutoGoroutines(shardSize int) Option {
|
||||
return func(o *options) {
|
||||
o.shardSize = shardSize
|
||||
}
|
||||
}
|
||||
|
||||
// WithMinSplitSize is the minimum encoding size in bytes per goroutine.
|
||||
// See WithMaxGoroutines on how jobs are split.
|
||||
// If n <= 0, it is ignored.
|
||||
|
|
|
|||
|
|
@ -15,7 +15,10 @@ import (
|
|||
"bytes"
|
||||
"errors"
|
||||
"io"
|
||||
"runtime"
|
||||
"sync"
|
||||
|
||||
"github.com/klauspost/cpuid"
|
||||
)
|
||||
|
||||
// Encoder is an interface to encode Reed-Salomon parity sets for your data.
|
||||
|
|
@ -239,6 +242,33 @@ func New(dataShards, parityShards int, opts ...Option) (Encoder, error) {
|
|||
if err != nil {
|
||||
return nil, err
|
||||
}
|
||||
if r.o.shardSize > 0 {
|
||||
cacheSize := cpuid.CPU.Cache.L2
|
||||
if cacheSize <= 0 {
|
||||
// Set to 128K if undetectable.
|
||||
cacheSize = 128 << 10
|
||||
}
|
||||
p := runtime.NumCPU()
|
||||
|
||||
// 1 input + parity must fit in cache, and we add one more to be safer.
|
||||
shards := 1 + parityShards
|
||||
g := (r.o.shardSize * shards) / (cacheSize - (cacheSize >> 4))
|
||||
|
||||
if cpuid.CPU.ThreadsPerCore > 1 {
|
||||
// If multiple threads per core, make sure they don't contend for cache.
|
||||
g *= cpuid.CPU.ThreadsPerCore
|
||||
}
|
||||
g *= 2
|
||||
if g < p {
|
||||
g = p
|
||||
}
|
||||
|
||||
// Have g be multiple of p
|
||||
g += p - 1
|
||||
g -= g % p
|
||||
|
||||
r.o.maxGoroutines = g
|
||||
}
|
||||
|
||||
// Inverted matrices are cached in a tree keyed by the indices
|
||||
// of the invalid rows of the data to reconstruct.
|
||||
|
|
@ -431,6 +461,8 @@ func (r reedSolomon) codeSomeShardsP(matrixRows, inputs, outputs [][]byte, outpu
|
|||
if do < r.o.minSplitSize {
|
||||
do = r.o.minSplitSize
|
||||
}
|
||||
// Make sizes divisible by 16
|
||||
do = (do + 15) & (^15)
|
||||
start := 0
|
||||
for start < byteCount {
|
||||
if start+do > byteCount {
|
||||
|
|
@ -490,6 +522,8 @@ func (r reedSolomon) checkSomeShardsP(matrixRows, inputs, toCheck [][]byte, outp
|
|||
if do < r.o.minSplitSize {
|
||||
do = r.o.minSplitSize
|
||||
}
|
||||
// Make sizes divisible by 16
|
||||
do = (do + 15) & (^15)
|
||||
start := 0
|
||||
for start < byteCount {
|
||||
if start+do > byteCount {
|
||||
|
|
|
|||
|
|
@ -113,6 +113,7 @@ func testOpts() [][]Option {
|
|||
{WithMaxGoroutines(5000), WithMinSplitSize(50), withSSE3(false), withAVX2(false)},
|
||||
{WithMaxGoroutines(5000), WithMinSplitSize(500000), withSSE3(false), withAVX2(false)},
|
||||
{WithMaxGoroutines(1), WithMinSplitSize(500000), withSSE3(false), withAVX2(false)},
|
||||
{WithAutoGoroutines(50000), WithMinSplitSize(500)},
|
||||
}
|
||||
for _, o := range opts[:] {
|
||||
if defaultOptions.useSSSE3 {
|
||||
|
|
@ -655,7 +656,7 @@ func fillRandom(p []byte) {
|
|||
}
|
||||
|
||||
func benchmarkEncode(b *testing.B, dataShards, parityShards, shardSize int) {
|
||||
r, err := New(dataShards, parityShards)
|
||||
r, err := New(dataShards, parityShards, WithAutoGoroutines(shardSize))
|
||||
if err != nil {
|
||||
b.Fatal(err)
|
||||
}
|
||||
|
|
@ -722,7 +723,7 @@ func BenchmarkEncode17x3x16M(b *testing.B) {
|
|||
}
|
||||
|
||||
func benchmarkVerify(b *testing.B, dataShards, parityShards, shardSize int) {
|
||||
r, err := New(dataShards, parityShards)
|
||||
r, err := New(dataShards, parityShards, WithAutoGoroutines(shardSize))
|
||||
if err != nil {
|
||||
b.Fatal(err)
|
||||
}
|
||||
|
|
@ -793,7 +794,7 @@ func corruptRandom(shards [][]byte, dataShards, parityShards int) {
|
|||
}
|
||||
|
||||
func benchmarkReconstruct(b *testing.B, dataShards, parityShards, shardSize int) {
|
||||
r, err := New(dataShards, parityShards)
|
||||
r, err := New(dataShards, parityShards, WithAutoGoroutines(shardSize))
|
||||
if err != nil {
|
||||
b.Fatal(err)
|
||||
}
|
||||
|
|
@ -873,7 +874,7 @@ func corruptRandomData(shards [][]byte, dataShards, parityShards int) {
|
|||
}
|
||||
|
||||
func benchmarkReconstructData(b *testing.B, dataShards, parityShards, shardSize int) {
|
||||
r, err := New(dataShards, parityShards)
|
||||
r, err := New(dataShards, parityShards, WithAutoGoroutines(shardSize))
|
||||
if err != nil {
|
||||
b.Fatal(err)
|
||||
}
|
||||
|
|
@ -939,7 +940,7 @@ func BenchmarkReconstructData10x4x16M(b *testing.B) {
|
|||
}
|
||||
|
||||
func benchmarkReconstructP(b *testing.B, dataShards, parityShards, shardSize int) {
|
||||
r, err := New(dataShards, parityShards)
|
||||
r, err := New(dataShards, parityShards, WithAutoGoroutines(shardSize))
|
||||
if err != nil {
|
||||
b.Fatal(err)
|
||||
}
|
||||
|
|
|
|||
Loading…
Reference in New Issue