Improving gonum perfs
115 views
Skip to first unread message
Sebastien Binet
Oct 5, 2017, 2:25:14 AM10/5/17
to gonu...@googlegroups.com
hi there,
I've stumbled on this piece of code from reddit:
I've tried to see what it would look like with Gonum:
it's not always super pretty (especially the part that updates the weights row/row).
and, it's even slower than the original (initially b/c of the `mat.Dense.RowView` that I've modified to reuse a `mat.VecDense`+`RowViewOf`, and now it's slow b/c of interface-checks in `mat.Dot` https://github.com/sbinet-staging/goAdaline/blob/9b8316625f4170edc195b4e3cd5bb6ca6bb13f8c/main.go#L50)
see:
```
$> time ./goAdaline-ref -cycles=1000
Test error:
0.017159516933323248
Weights:
[0.629646421847802 0.5111203968481454 0.21903157034908186 -0.24872268718943477 0.113694617824979 0.03994836902175236 0.09902375643857968 0.5027786680714972 -0.015058660602257478]
real 0m0.080s
user 0m0.077s
sys 0m0.003s
$> time ./goAdaline-gonum -cycles=1000
Test error:
0.01715951693332323
Weights:
[0.6296464218478025 0.5111203968481459 0.21903157034908222 -0.2487226871894338 0.11369461782497912 0.03994836902175287 0.09902375643858036 0.5027786680714972 -0.015058660602259038]
real 0m0.278s
user 0m0.276s
sys 0m0.003s
```
here is a `pprof` output:
```
$> go tool pprof ./cpu.prof
File: goAdaline-gonum
Type: cpu
Time: Oct 3, 2017 at 9:29am (CEST)
Duration: 400.60ms, Total samples = 260ms (64.90%)
Entering interactive mode (type "help" for commands, "o" for options)
(pprof) top50
Showing nodes accounting for 260ms, 100% of 260ms total
flat flat% sum% cum cum%
100ms 38.46% 38.46% 260ms 100% main.main /home/binet/tmp/go/src/github.com/Plorenzo/goAdaline/main.go
30ms 11.54% 50.00% 30ms 11.54% gonum.org/v1/gonum/mat.(*VecDense).At /home/binet/tmp/go/src/gonum.org/v1/gonum/mat/index_no_bound_checks.go
30ms 11.54% 61.54% 40ms 15.38% gonum.org/v1/gonum/mat.(*VecDense).RowViewOf /home/binet/tmp/go/src/gonum.org/v1/gonum/mat/vector.go
20ms 7.69% 69.23% 20ms 7.69% gonum.org/v1/gonum/blas/gonum.Implementation.Ddot /home/binet/tmp/go/src/gonum.org/v1/gonum/blas/gonum/level1double_ddot.go
10ms 3.85% 73.08% 30ms 11.54% gonum.org/v1/gonum/blas/gonum.(*Implementation).Ddot <autogenerated>
10ms 3.85% 76.92% 10ms 3.85% gonum.org/v1/gonum/internal/asm/f64.DotUnitary /home/binet/tmp/go/src/gonum.org/v1/gonum/internal/asm/f64/dot_amd64.s
10ms 3.85% 80.77% 10ms 3.85% gonum.org/v1/gonum/mat.(*Dense).At /home/binet/tmp/go/src/gonum.org/v1/gonum/mat/index_no_bound_checks.go
10ms 3.85% 84.62% 10ms 3.85% gonum.org/v1/gonum/mat.(*Dense).RawMatrix /home/binet/tmp/go/src/gonum.org/v1/gonum/mat/dense.go
10ms 3.85% 88.46% 10ms 3.85% gonum.org/v1/gonum/mat.(*VecDense).RawVector /home/binet/tmp/go/src/gonum.org/v1/gonum/mat/vector.go
10ms 3.85% 92.31% 20ms 7.69% gonum.org/v1/gonum/mat.Sum /home/binet/tmp/go/src/gonum.org/v1/gonum/mat/matrix.go
10ms 3.85% 96.15% 20ms 7.69% runtime.assertI2I2 /usr/lib/go/src/runtime/iface.go
10ms 3.85% 100% 10ms 3.85% runtime.getitab /usr/lib/go/src/runtime/iface.go
0 0% 100% 30ms 11.54% gonum.org/v1/gonum/blas/blas64.Dot /home/binet/tmp/go/src/gonum.org/v1/gonum/blas/blas64/blas64.go
0 0% 100% 10ms 3.85% gonum.org/v1/gonum/blas/blas64.Gemv /home/binet/tmp/go/src/gonum.org/v1/gonum/blas/blas64/blas64.go
0 0% 100% 10ms 3.85% gonum.org/v1/gonum/blas/gonum.(*Implementation).Dgemv <autogenerated>
0 0% 100% 10ms 3.85% gonum.org/v1/gonum/blas/gonum.Implementation.Dgemv /home/binet/tmp/go/src/gonum.org/v1/gonum/blas/gonum/level2double.go
0 0% 100% 10ms 3.85% gonum.org/v1/gonum/mat.(*VecDense).MulVec /home/binet/tmp/go/src/gonum.org/v1/gonum/mat/vector.go
0 0% 100% 60ms 23.08% gonum.org/v1/gonum/mat.Dot /home/binet/tmp/go/src/gonum.org/v1/gonum/mat/matrix.go
0 0% 100% 30ms 11.54% main.computeError /home/binet/tmp/go/src/github.com/Plorenzo/goAdaline/main.go
0 0% 100% 260ms 100% runtime.main /usr/lib/go/src/runtime/proc.go
(pprof)
(pprof) list main
Total: 260ms
ROUTINE ======================== main.computeError in /home/binet/tmp/go/src/github.com/Plorenzo/goAdaline/main.go
0 30ms (flat, cum) 11.54% of Total
. . 198:}
. . 199:
. . 200:func computeError(data *mat.Dense, expected, weights *mat.VecDense) float64 {. . 201:
. . 202: var errs mat.VecDense
. 10ms 203: errs.MulVec(data, weights)
. . 204: errs.SubVec(expected, &errs)
. . 205: errs.MulElemVec(&errs, &errs)
. . 206:
. 20ms 207: return mat.Sum(&errs) / float64(errs.Len())
. . 208:}
ROUTINE ======================== main.main in /home/binet/tmp/go/src/github.com/Plorenzo/goAdaline/main.go
100ms 260ms (flat, cum) 100% of Total
. . 56: var errorsTrain []float64
. . 57: var errorsValidate []float64
. . 58: var errorsTest float64
. . 59:
. . 60: // Learning
10ms 10ms 61: for cycle := 0; cycle < *cycles; cycle++ {. . 62: var row mat.VecDense
. . 63: for i := 0; i < nrows; i++ {. 40ms 64: row.RowViewOf(data, i)
. . 65: // Calculate estimate
. 60ms 66: estimate := mat.Dot(&row, weights)
. . 67: // Update weights (range passes values as a copy)
. . 68: raw := weights.RawVector().Data
20ms 20ms 69: for x := range raw {70ms 100ms 70: raw[x] += *learningRate * (expectedY.At(i, 0) - estimate) * data.At(i, x)
. . 71: }
. . 72: }
. . 73:
. . 74: // Compute cycle train error
. 30ms 75: errorsTrain = append(errorsTrain, computeError(data, expectedY, weights))
. . 76: errorsValidate = append(errorsValidate, computeError(validateData, valExpectedY, weights))
. . 77: }
. . 78:
. . 79: errorsTest = computeError(testData, testExpectedY, weights)
. . 80:
```
not sure whether there is any Gonum-based more performant way to do this...
sent from my droid
Dan Kortschak
Oct 5, 2017, 2:44:05 AM10/5/17
to Sebastien Binet, gonu...@googlegroups.com
If mat is causing the problems, don't use it; we provide blas64 (and
blas for that matter) as a public package for a reason.
```untested
rate := *learningRate
row := mat.VecDense{
N: ncols,
Inc: 1,
}
raw := data.RawMatrix()
expRaw : expectedY.RawMatrix()
// Calculate estimate
estimate := blas64.Dot(ncols, row, weights) # Initialise weights as a blas64.Vector.
raw.Data[j] += rate * (expRaw.Data[i*expRaw.Stride] - estimate) * raw.Data[i*raw.Stride+j]
It's not ideal, but if you want very high performance in tight loops,
sometimes it's unnecessary.
blas for that matter) as a public package for a reason.
```untested
rate := *learningRate
row := mat.VecDense{
N: ncols,
Inc: 1,
}
raw := data.RawMatrix()
expRaw : expectedY.RawMatrix()
for cycle := 0; cycle < *cycles; cycle++ {
for i := 0; i < nrows; i++ {
row.Data = raw.Data[i*raw.Stride:]
// Calculate estimate
estimate := blas64.Dot(ncols, row, weights) # Initialise weights as a blas64.Vector.
// Update weights (range passes values as a copy)
for j := range row.Data {
raw.Data[j] += rate * (expRaw.Data[i*expRaw.Stride] - estimate) * raw.Data[i*raw.Stride+j]
}
}
// Compute cycle train error
}
// Compute cycle train error
errorsTrain = append(errorsTrain, computeError(data, expectedY, weights))
errorsValidate = append(errorsValidate, computeError(validateData, valExpectedY, weights))
}
```
}
It's not ideal, but if you want very high performance in tight loops,
sometimes it's unnecessary.
Sebastien Binet
Oct 5, 2017, 3:42:06 AM10/5/17
to Dan Kortschak, gonu...@googlegroups.com
Thanks Dan.
I guess this is good for thought for multi-dim slices for Go 2 :)
I'll check how this translates performance wise.
I guess this is good for thought for multi-dim slices for Go 2 :)
Cheers,
-s
sent from my droid
--
You received this message because you are subscribed to the Google Groups "gonum-dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to gonum-dev+unsubscribe@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.
Sebastien Binet
Oct 7, 2017, 6:02:54 AM10/7/17
to Dan Kortschak, gonu...@googlegroups.com
On Thu, Oct 5, 2017 at 9:42 AM, Sebastien Binet <seb....@gmail.com> wrote:
Thanks Dan.I'll check how this translates performance wise.
ok. so it's better but still no cigar :)
$> time ./goAdaline-ref -cycles=100000
Test error:
0.017159516933323238
Weights:
[0.6296464218478021 0.5111203968481455 0.2190315703490819 -0.2487226871894348 0.11369461782497882 0.03994836902175239 0.09902375643857965 0.5027786680714972 -0.015058660602257527]
real 0m3.312s
user 0m3.500s
sys 0m0.011s
$> time ./goAdaline-gonum -cycles=100000
Test error:
0.017159516933323238
Weights:
[0.6296464218478026 0.511120396848146 0.21903157034908224 -0.24872268718943405 0.11369461782497912 0.0399483690217529 0.09902375643858032 0.5027786680714972 -0.015058660602258935]
real 0m12.858s
user 0m12.902s
sys 0m0.044s
$> time ./goAdaline-blas64 -cycles=100000
Test error:
0.017159516933323238
Weights:
[0.6296464218478026 0.511120396848146 0.21903157034908224 -0.24872268718943405 0.11369461782497912 0.0399483690217529 0.09902375643858032 0.5027786680714972 -0.015058660602258935]
real 0m4.734s
user 0m4.794s
sys 0m0.024s
goAdaline-blas64 is using gonum/blas instead of gonum/mat.
everything is in the "use-gonum" branch.
There is probably some more gains to be have by migrating everything to gonum/blas (especially in the "computeErrors" func) but there isn't much to be shaved off:
(pprof) top20
Showing nodes accounting for 4.55s, 95.99% of 4.74s total
Dropped 37 nodes (cum <= 0.02s)
Showing top 20 nodes out of 39
flat flat% sum% cum cum%
1.16s 24.47% 24.47% 4.67s 98.52% main.main /home/binet/work/gonum/src/github.com/Plorenzo/goAdaline/main.go
1.13s 23.84% 48.31% 1.13s 23.84% gonum.org/v1/gonum/internal/asm/f64.DotUnitary /home/binet/work/gonum/src/gonum.org/v1/gonum/internal/asm/f64/dot_amd64.s
0.49s 10.34% 58.65% 0.78s 16.46% gonum.org/v1/gonum/blas/gonum.Implementation.Dgemv /home/binet/work/gonum/src/gonum.org/v1/gonum/blas/gonum/level2double.go
0.36s 7.59% 66.24% 1.21s 25.53% gonum.org/v1/gonum/blas/gonum.Implementation.Ddot /home/binet/work/gonum/src/gonum.org/v1/gonum/blas/gonum/level1double_ddot.go
0.34s 7.17% 73.42% 0.52s 10.97% gonum.org/v1/gonum/mat.Sum /home/binet/work/gonum/src/gonum.org/v1/gonum/mat/matrix.go
0.31s 6.54% 79.96% 1.78s 37.55% gonum.org/v1/gonum/blas/blas64.Dot /home/binet/work/gonum/src/gonum.org/v1/gonum/blas/blas64/blas64.go
0.26s 5.49% 85.44% 1.47s 31.01% gonum.org/v1/gonum/blas/gonum.(*Implementation).Ddot <autogenerated>
0.16s 3.38% 88.82% 0.16s 3.38% gonum.org/v1/gonum/mat.(*VecDense).At /home/binet/work/gonum/src/gonum.org/v1/gonum/mat/index_no_bound_checks.go
0.10s 2.11% 90.93% 0.10s 2.11% gonum.org/v1/gonum/mat.(*VecDense).MulElemVec /home/binet/work/gonum/src/gonum.org/v1/gonum/mat/vector.go
0.10s 2.11% 93.04% 0.10s 2.11% runtime.memclrNoHeapPointers /usr/lib/go/src/runtime/memclr_amd64.s
0.03s 0.63% 93.67% 0.03s 0.63% runtime.getitab /usr/lib/go/src/runtime/iface.go
0.03s 0.63% 94.30% 0.03s 0.63% runtime.greyobject /usr/lib/go/src/runtime/mgcmark.go
0.02s 0.42% 94.73% 0.04s 0.84% gonum.org/v1/gonum/mat.(*VecDense).SubVec /home/binet/work/gonum/src/gonum.org/v1/gonum/mat/vector.go
0.02s 0.42% 95.15% 1.63s 34.39% main.computeError /home/binet/work/gonum/src/github.com/Plorenzo/goAdaline/main.go
0.01s 0.21% 95.36% 0.04s 0.84% runtime.assertI2I2 /usr/lib/go/src/runtime/iface.go
0.01s 0.21% 95.57% 0.03s 0.63% runtime.gcDrain /usr/lib/go/src/runtime/mgcmark.go
0.01s 0.21% 95.78% 0.03s 0.63% runtime.scanobject /usr/lib/go/src/runtime/mgcmark.go
0.01s 0.21% 95.99% 0.03s 0.63% strconv.formatDigits /usr/lib/go/src/strconv/ftoa.go
0 0% 95.99% 0.78s 16.46% gonum.org/v1/gonum/blas/blas64.Gemv /home/binet/work/gonum/src/gonum.org/v1/gonum/blas/blas64/blas64.go
0 0% 95.99% 0.78s 16.46% gonum.org/v1/gonum/blas/gonum.(*Implementation).Dgemv <autogenerated>
-s
Kunde21
Oct 7, 2017, 8:19:41 PM10/7/17
to gonum-dev
Can you pull the `f64/gemv` branch and run it with that?
To unsubscribe from this group and stop receiving emails from it, send an email to gonum-dev+...@googlegroups.com.
Sebastien Binet
Oct 8, 2017, 3:36:29 AM10/8/17
to Kunde21, gonum-dev
On Sun, Oct 8, 2017 at 2:19 AM, Kunde21 <kun...@gmail.com> wrote:
Can you pull the `f64/gemv` branch and run it with that?
it improves a bit:
$> time ./goAdaline-no-gemv-branch -cycles=100000
Test error:
0.017159516933323238
Weights:
[0.6296464218478026 0.511120396848146 0.21903157034908224 -0.24872268718943405 0.11369461782497912 0.0399483690217529 0.09902375643858032 0.5027786680714972 -0.015058660602258935]
real 0m4.755s
user 0m4.852s
sys 0m0.017s
$> time ./goAdaline-gemv-branch -cycles=100000
Test error:
0.01715951693332324
Weights:
[0.6296464218478026 0.511120396848146 0.21903157034908224 -0.24872268718943405 0.11369461782497912 0.0399483690217529 0.09902375643858032 0.5027786680714972 -0.015058660602258935]
real 0m4.281s
user 0m4.374s
sys 0m0.020s
but we are still not there:
$> time ./goAdaline-ref -cycles=100000
Test error:
0.017159516933323238
Weights:
[0.6296464218478021 0.5111203968481455 0.2190315703490819 -0.2487226871894348 0.11369461782497882 0.03994836902175239 0.09902375643857965 0.5027786680714972 -0.015058660602257527]
real 0m3.318s
user 0m4.253s
sys 0m0.057s
To unsubscribe from this group and stop receiving emails from it, send an email to gonum-dev+unsubscribe@googlegroups.com.
Kunde21
Oct 8, 2017, 5:44:32 PM10/8/17
to gonum-dev
I think the biggest problem is the narrow dataset. Any function calls inside the training loop won't make up the cost of the call, because there's just not enough data in each row.
If you were to change the algorithm from streaming to batch learning, you could use the bigger matrix calculations to process enough data to see improvements.
Dan Kortschak
Oct 8, 2017, 8:54:35 PM10/8/17
to Sebastien Binet, Kunde21, gonum-dev
Just for kicks, what happens with `-tag noasm`?
Brendan Tracey
Oct 10, 2017, 5:50:57 PM10/10/17
to Dan Kortschak, Sebastien Binet, Kunde21, gonum-dev
You mean, `-tags noasm`
Dan Kortschak
Oct 10, 2017, 6:02:46 PM10/10/17
to Brendan Tracey, Sebastien Binet, Kunde21, gonum-dev
Yes.
On Tue, 2017-10-10 at 15:50 -0600, Brendan Tracey wrote:
> You mean, `-tags noasm`
On Tue, 2017-10-10 at 15:50 -0600, Brendan Tracey wrote:
> You mean, `-tags noasm`
Reply all
Reply to author
Forward
0 new messages