Faster MTTKRPs Algorithm #40

alexmul1114 · 2024-02-26T20:20:54Z

Addresses #17.

codecov · 2024-02-26T20:24:43Z

Codecov Report

Attention: Patch coverage is 96.36364% with 2 lines in your changes are missing coverage. Please review.

Project coverage is 99.34%. Comparing base (bb701f2) to head (49d34eb).

Files	Patch %	Lines
src/tensor-kernels/mttkrps.jl	95.65%	2 Missing ⚠️

Additional details and impacted files

@@             Coverage Diff             @@
##            master      #40      +/-   ##
===========================================
- Coverage   100.00%   99.34%   -0.66%     
===========================================
  Files           12       12              
  Lines          258      306      +48     
===========================================
+ Hits           258      304      +46     
- Misses           0        2       +2

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

alexmul1114 · 2024-03-04T21:16:19Z

For some reason, although running one iteration of MTTKRPs with the new algorithm versus the old implementation is much faster, but the gcp benchmarks show no improvement:

julia> using GCPDecompositions
Precompiling GCPDecompositions
  1 dependency successfully precompiled in 3 seconds. 20 already precompiled.

julia> using Random

julia> using BenchmarkTools

julia> T=Float64;

julia> r=2;

julia> using LinearAlgebra;

julia> function test_new(GU, M, X)
           N = ndims(X)
           R = size(M.U[1])[2]

           # Determine order of modes of MTTKRP to compute
           Jns = [prod(size(X)[1:n]) for n in 1:N]
           Kns = [prod(size(X)[n+1:end]) for n in 1:N]
           Kn_minus_ones = [prod(size(X)[n:end]) for n in 1:N]
           comp = Jns .<= Kn_minus_ones
               n_star = maximum(map(x -> comp[x] ? x : 0, 1:N))
           order = vcat([i for i in n_star:-1:1], [i for i in n_star+1:N])

           # Compute MTTKRPs recursively
           saved = similar(M.U[1], Jns[n_star], R)
           for n in order
               if n == n_star
                   saved = reshape(X, (Jns[n], Kns[n])) * khatrirao(M.U[reverse(n+1:N)]...)
                   GCPDecompositions.mttkrps_helper!(GU, saved, M, n, "right", N, Jns, Kns)
               elseif n == n_star + 1
                   if n == N
                       mul!(GU[n], reshape(X, (Jns[n-1], Kns[n-1]))', khatrirao(M.U[reverse(1:n-1)]...))
                   else
                       saved = (khatrirao(M.U[reverse(1:n-1)]...)' * reshape(X, (Jns[n-1], Kns[n-1])))'
                       GCPDecompositions.mttkrps_helper!(GU, saved, M, n, "left", N, Jns, Kns)
                   end
               elseif n < n_star
                   if n == 1
                       for r in 1:R
                           mul!(view(GU[n], :, r), reshape(view(saved, :, r), (Jns[n], size(X)[n+1])), view(M.U[n+1], :, r))
                           #GU[n][:, r] = reshape(view(saved, :, r), (Jns[n], size(X)[n+1])) * view(M.U[n+1], :, r)
                       end
                   else
                       saved = stack(reshape(view(saved, :, r), (Jns[n], size(X)[n+1])) * view(M.U[n+1], :, r) for r in 1:R)
                       GCPDecompositions.mttkrps_helper!(GU, saved, M, n, "right", N, Jns, Kns)
                   end
               else
                   if n == N
                       for r in 1:R
                           mul!(view(GU[n], :, r), reshape(view(saved, :, r), (size(X)[n-1], Kns[n-1]))', view(M.U[n-1], :, r))
                       end
                       #GU[n] = stack(reshape(view(saved, :, r), (size(X)[n-1], Kns[n-1]))' * view(M.U[n-1], :, r) for r in 1:R)
                   else
                       saved = stack(reshape(view(saved, :, r), (size(X)[n-1], Kns[n-1]))' * view(M.U[n-1], :, r) for r in 1:R)
                       GCPDecompositions.mttkrps_helper!(GU, saved, M, n, "left", N, Jns, Kns)
                   end
               end
           end
       end
test_new (generic function with 1 method)

julia> function test_old(GU, M, X)
           N=ndims(X)
               for k in 1:ndims(X)
               Yk = reshape(PermutedDimsArray(X, [k; setdiff(1:N, k)]), size(X, k), :)
               Zk = similar(Yk, prod(size(X)[setdiff(1:N, k)]), ncomponents(M))
               for j in Base.OneTo(ncomponents(M))
                           Zk[:, j] = reduce(kron, [view(M.U[i], :, j) for i in reverse(setdiff(1:N, k))])
               end
               mul!(GU[k], Yk, Zk)
           end
       end
test_old (generic function with 1 method)

julia> function khatrirao(A::Vararg{T,N}) where {T<:AbstractMatrix,N}
           r = size(A[1],2)
           # @boundscheck all(==(r),size.(A,2)) || throw(DimensionMismatch())
               R = ntuple(Val(N)) do k
               dims = (ntuple(i->1,Val(N-k))..., :, ntuple(i->1,Val(k-1))..., r)
               return reshape(A[k],dims)
                   end
           return reshape(broadcast(*, R...),:,r)
       end
khatrirao (generic function with 1 method)

julia> X = rand(15,20,25);

julia> Random.seed!(0);

julia> M_old = CPD(ones(T, r), rand.(T, size(X), r));

julia> Random.seed!(0);

julia> M_new = CPD(ones(T, r), rand.(T, size(X), r));

julia> GU_old = [similar(M_old.U[k]) for k in 1:ndims(X)];

julia> GU_new = [similar(M_new.U[k]) for k in 1:ndims(X)];

julia> @btime test_old(GU_old, M_old, X);
  143.400 μs (135 allocations: 139.67 KiB)

julia> @btime test_new(GU_new, M_new, X);
  8.800 μs (75 allocations: 18.31 KiB)

julia> maximum(GU_new[1]-GU_old[1])
1.4210854715202004e-14

julia> maximum(GU_new[2]-GU_old[2])
1.4210854715202004e-14

julia> maximum(GU_new[3]-GU_old[3])
7.105427357601002e-15

Benchmark Report for `GCPDecompositions`

Job Properties

Time of benchmarks:
- Target: 4 Mar 2024 - 13:23
- Baseline: 4 Mar 2024 - 13:24
Package commits:
- Target: 4003e9
- Baseline: 62ce59
Julia commits:
- Target: 312098
- Baseline: 312098
Julia command flags:
- Target: None
- Baseline: None
Environment variables:
- Target: GCP_BENCHMARK_SUITES => gcp
- Baseline: GCP_BENCHMARK_SUITES => gcp

Results

A ratio greater than 1.0 denotes a possible regression (marked with ❌), while a ratio less
than 1.0 denotes a possible improvement (marked with ✅). Only significant results - results
that indicate possible regressions or improvements - are shown below (thus, an empty table means that all
benchmark results remained invariant between builds).

ID	time ratio	memory ratio
`["gcp", "bernoulliOdds-size(X)=(15, 20, 25), rank(X)=1"]`	0.85 (5%) ✅	0.94 (1%) ✅
`["gcp", "bernoulliOdds-size(X)=(15, 20, 25), rank(X)=2"]`	0.86 (5%) ✅	1.07 (1%) ❌
`["gcp", "bernoulliOdds-size(X)=(30, 40, 50), rank(X)=1"]`	0.80 (5%) ✅	1.04 (1%) ❌
`["gcp", "bernoulliOdds-size(X)=(30, 40, 50), rank(X)=2"]`	0.86 (5%) ✅	0.97 (1%) ✅
`["gcp", "gamma-size(X)=(15, 20, 25), rank(X)=1"]`	1.21 (5%) ❌	0.97 (1%) ✅
`["gcp", "gamma-size(X)=(15, 20, 25), rank(X)=2"]`	0.91 (5%) ✅	0.90 (1%) ✅
`["gcp", "gamma-size(X)=(30, 40, 50), rank(X)=1"]`	1.21 (5%) ❌	1.22 (1%) ❌
`["gcp", "gamma-size(X)=(30, 40, 50), rank(X)=2"]`	0.82 (5%) ✅	1.02 (1%) ❌
`["gcp", "least-squares-size(X)=(15, 20, 25), rank(X)=2"]`	0.89 (5%) ✅	1.00 (1%)
`["gcp", "least-squares-size(X)=(30, 40, 50), rank(X)=1"]`	1.22 (5%) ❌	1.00 (1%)
`["gcp", "poisson-size(X)=(15, 20, 25), rank(X)=1"]`	1.08 (5%) ❌	1.00 (1%)
`["gcp", "poisson-size(X)=(15, 20, 25), rank(X)=2"]`	0.99 (5%)	1.25 (1%) ❌
`["gcp", "poisson-size(X)=(30, 40, 50), rank(X)=1"]`	0.82 (5%) ✅	0.90 (1%) ✅
`["gcp", "poisson-size(X)=(30, 40, 50), rank(X)=2"]`	1.31 (5%) ❌	1.29 (1%) ❌

Benchmark Group List

Here's a list of all the benchmark groups executed by this job:

["gcp"]

Julia versioninfo

Target

Julia Version 1.10.0
Commit 3120989f39 (2023-12-25 18:01 UTC)
Build Info:
  Official https://julialang.org/ release
Platform Info:
  OS: Windows (x86_64-w64-mingw32)
      Microsoft Windows [Version 10.0.22621.3155]
  CPU: 11th Gen Intel(R) Core(TM) i7-11800H @ 2.30GHz: 
                 speed         user         nice          sys         idle          irq
       #1-16  2304 MHz   44819604            0     44240962    7935611915      1134929  ticks
  Memory: 31.726390838623047 GB (13396.0078125 MB free)
  Uptime: 1.04030714e6 sec
  Load Avg:  0.0  0.0  0.0
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-15.0.7 (ORCJIT, tigerlake)
  Threads: 1 on 16 virtual cores

Baseline

Julia Version 1.10.0
Commit 3120989f39 (2023-12-25 18:01 UTC)
Build Info:
  Official https://julialang.org/ release
Platform Info:
  OS: Windows (x86_64-w64-mingw32)
      Microsoft Windows [Version 10.0.22621.3155]
  CPU: 11th Gen Intel(R) Core(TM) i7-11800H @ 2.30GHz: 
                 speed         user         nice          sys         idle          irq
       #1-16  2304 MHz   44924368            0     44263742    7937208507      1135085  ticks
  Memory: 31.726390838623047 GB (13700.35546875 MB free)
  Uptime: 1.04041489e6 sec
  Load Avg:  0.0  0.0  0.0
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-15.0.7 (ORCJIT, tigerlake)
  Threads: 1 on 16 virtual cores

alexmul1114 · 2024-03-05T14:14:25Z

For the gcp benchmark with sz=(15,20,25), r=2, Poisson loss, it looks like most of the time is spent in evaluating the function and filling out Y (tensor of elementwise derivatives) rather than doing MTTKRPs. Below is the profiling result for the new MTTKRPs:

And below is the profiling result using the old implementation:

The MTTKRPs don't show up in the profiling for the new version but for the old version they take up maybe 5% of the total time (the ntuple part is the loop over the N modes for MTTKRP).
Using r=100 with the same size and Poisson loss, the MTTKRPs take up a somewhat bigger chunk of the time but still only about 5% for new MTTKRPs (yellow gcp_grad_U!):

and 15% for old (yellow gcp_grad_U!):

alexmul1114 · 2024-03-05T17:18:18Z

Changing the gcp_func method from using sum(value(loss, X[I], M[I]) for I in CartesianIndices(X) if !ismissing(X[I])) to mapreduce(I -> !ismissing(X[I]) ? value(loss, X[I], M[I]) : 0, +, CartesianIndices(X)) provides a decent memory benefit and some speed-up (both target and baseline here using old MTTKRPs):

Benchmark Report for `GCPDecompositions`

Job Properties

Time of benchmarks:
- Target: 5 Mar 2024 - 12:11
- Baseline: 5 Mar 2024 - 12:13
Package commits:
- Target: 2d8cd8
- Baseline: 62ce59
Julia commits:
- Target: 312098
- Baseline: 312098
Julia command flags:
- Target: None
- Baseline: None
Environment variables:
- Target: GCP_BENCHMARK_SUITES => gcp
- Baseline: GCP_BENCHMARK_SUITES => gcp

Results

A ratio greater than 1.0 denotes a possible regression (marked with ❌), while a ratio less
than 1.0 denotes a possible improvement (marked with ✅). Only significant results - results
that indicate possible regressions or improvements - are shown below (thus, an empty table means that all
benchmark results remained invariant between builds).

ID	time ratio	memory ratio
`["gcp", "bernoulliOdds-size(X)=(15, 20, 25), rank(X)=1"]`	0.90 (5%) ✅	0.35 (1%) ✅
`["gcp", "bernoulliOdds-size(X)=(15, 20, 25), rank(X)=2"]`	0.82 (5%) ✅	0.33 (1%) ✅
`["gcp", "bernoulliOdds-size(X)=(30, 40, 50), rank(X)=1"]`	0.88 (5%) ✅	0.34 (1%) ✅
`["gcp", "bernoulliOdds-size(X)=(30, 40, 50), rank(X)=2"]`	0.87 (5%) ✅	0.34 (1%) ✅
`["gcp", "gamma-size(X)=(15, 20, 25), rank(X)=1"]`	1.01 (5%)	0.40 (1%) ✅
`["gcp", "gamma-size(X)=(15, 20, 25), rank(X)=2"]`	0.90 (5%) ✅	0.41 (1%) ✅
`["gcp", "gamma-size(X)=(30, 40, 50), rank(X)=1"]`	1.01 (5%)	0.46 (1%) ✅
`["gcp", "gamma-size(X)=(30, 40, 50), rank(X)=2"]`	0.78 (5%) ✅	0.34 (1%) ✅
`["gcp", "least-squares-size(X)=(15, 20, 25), rank(X)=2"]`	1.10 (5%) ❌	1.00 (1%)
`["gcp", "least-squares-size(X)=(30, 40, 50), rank(X)=1"]`	1.26 (5%) ❌	1.00 (1%)
`["gcp", "least-squares-size(X)=(30, 40, 50), rank(X)=2"]`	1.35 (5%) ❌	1.00 (1%)
`["gcp", "poisson-size(X)=(15, 20, 25), rank(X)=1"]`	0.84 (5%) ✅	0.56 (1%) ✅
`["gcp", "poisson-size(X)=(15, 20, 25), rank(X)=2"]`	1.05 (5%) ❌	0.83 (1%) ✅
`["gcp", "poisson-size(X)=(30, 40, 50), rank(X)=1"]`	0.97 (5%)	0.56 (1%) ✅
`["gcp", "poisson-size(X)=(30, 40, 50), rank(X)=2"]`	0.84 (5%) ✅	0.55 (1%) ✅

Benchmark Group List

Here's a list of all the benchmark groups executed by this job:

["gcp"]

Julia versioninfo

Target

Julia Version 1.10.0
Commit 3120989f39 (2023-12-25 18:01 UTC)
Build Info:
  Official https://julialang.org/ release
Platform Info:
  OS: Windows (x86_64-w64-mingw32)
      Microsoft Windows [Version 10.0.22621.3155]
  CPU: 11th Gen Intel(R) Core(TM) i7-11800H @ 2.30GHz: 
                 speed         user         nice          sys         idle          irq
       #1-16  2304 MHz   49475134            0     49141931    8601542102      1332603  ticks
  Memory: 31.726390838623047 GB (13258.26171875 MB free)
  Uptime: 1.122432015e6 sec
  Load Avg:  0.0  0.0  0.0
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-15.0.7 (ORCJIT, tigerlake)
  Threads: 1 on 16 virtual cores

Baseline

Julia Version 1.10.0
Commit 3120989f39 (2023-12-25 18:01 UTC)
Build Info:
  Official https://julialang.org/ release
Platform Info:
  OS: Windows (x86_64-w64-mingw32)
      Microsoft Windows [Version 10.0.22621.3155]
  CPU: 11th Gen Intel(R) Core(TM) i7-11800H @ 2.30GHz: 
                 speed         user         nice          sys         idle          irq
       #1-16  2304 MHz   49564492            0     49151838    8603080366      1332882  ticks
  Memory: 31.726390838623047 GB (13464.0390625 MB free)
  Uptime: 1.122534359e6 sec
  Load Avg:  0.0  0.0  0.0
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-15.0.7 (ORCJIT, tigerlake)
  Threads: 1 on 16 virtual cores

alexmul1114 · 2024-03-08T00:37:39Z

Benchmarking comparison of old getindex for CPD versus new getindex that doesn't use sum:

Benchmark Report for `GCPDecompositions`

Job Properties

Time of benchmarks:
- Target: 7 Mar 2024 - 19:33
- Baseline: 7 Mar 2024 - 19:35
Package commits:
- Target: 6aa305
- Baseline: aef707
Julia commits:
- Target: 312098
- Baseline: 312098
Julia command flags:
- Target: None
- Baseline: None
Environment variables:
- Target: GCP_BENCHMARK_SUITES => gcp
- Baseline: GCP_BENCHMARK_SUITES => gcp

Results

A ratio greater than 1.0 denotes a possible regression (marked with ❌), while a ratio less
than 1.0 denotes a possible improvement (marked with ✅). Only significant results - results
that indicate possible regressions or improvements - are shown below (thus, an empty table means that all
benchmark results remained invariant between builds).

ID	time ratio	memory ratio
`["gcp", "bernoulliOdds-size(X)=(15, 20, 25), rank(X)=1"]`	0.33 (5%) ✅	0.08 (1%) ✅
`["gcp", "bernoulliOdds-size(X)=(15, 20, 25), rank(X)=2"]`	0.50 (5%) ✅	0.09 (1%) ✅
`["gcp", "bernoulliOdds-size(X)=(30, 40, 50), rank(X)=1"]`	0.29 (5%) ✅	0.08 (1%) ✅
`["gcp", "bernoulliOdds-size(X)=(30, 40, 50), rank(X)=2"]`	0.48 (5%) ✅	0.07 (1%) ✅
`["gcp", "gamma-size(X)=(15, 20, 25), rank(X)=1"]`	0.47 (5%) ✅	0.06 (1%) ✅
`["gcp", "gamma-size(X)=(15, 20, 25), rank(X)=2"]`	0.48 (5%) ✅	0.07 (1%) ✅
`["gcp", "gamma-size(X)=(30, 40, 50), rank(X)=1"]`	0.44 (5%) ✅	0.06 (1%) ✅
`["gcp", "gamma-size(X)=(30, 40, 50), rank(X)=2"]`	0.45 (5%) ✅	0.05 (1%) ✅
`["gcp", "poisson-size(X)=(15, 20, 25), rank(X)=1"]`	0.42 (5%) ✅	0.02 (1%) ✅
`["gcp", "poisson-size(X)=(15, 20, 25), rank(X)=2"]`	0.97 (5%)	0.83 (1%) ✅
`["gcp", "poisson-size(X)=(30, 40, 50), rank(X)=1"]`	0.39 (5%) ✅	0.03 (1%) ✅
`["gcp", "poisson-size(X)=(30, 40, 50), rank(X)=2"]`	0.46 (5%) ✅	0.03 (1%) ✅

Benchmark Group List

Here's a list of all the benchmark groups executed by this job:

["gcp"]

Julia versioninfo

Target

Julia Version 1.10.0
Commit 3120989f39 (2023-12-25 18:01 UTC)
Build Info:
  Official https://julialang.org/ release
Platform Info:
  OS: Windows (x86_64-w64-mingw32)
      Microsoft Windows [Version 10.0.22621.3155]
  CPU: 11th Gen Intel(R) Core(TM) i7-11800H @ 2.30GHz: 
                 speed         user         nice          sys         idle          irq
       #1-16  2304 MHz   70951977            0     64394166    10622768227      1747463  ticks
  Memory: 31.726390838623047 GB (12502.3671875 MB free)
  Uptime: 1.321744875e6 sec
  Load Avg:  0.0  0.0  0.0
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-15.0.7 (ORCJIT, tigerlake)
  Threads: 1 on 16 virtual cores

Baseline

Julia Version 1.10.0
Commit 3120989f39 (2023-12-25 18:01 UTC)
Build Info:
  Official https://julialang.org/ release
Platform Info:
  OS: Windows (x86_64-w64-mingw32)
      Microsoft Windows [Version 10.0.22621.3155]
  CPU: 11th Gen Intel(R) Core(TM) i7-11800H @ 2.30GHz: 
                 speed         user         nice          sys         idle          irq
       #1-16  2304 MHz   71135820            0     64411196    10624224353      1747806  ticks
  Memory: 31.726390838623047 GB (12590.13671875 MB free)
  Uptime: 1.321848437e6 sec
  Load Avg:  0.0  0.0  0.0
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-15.0.7 (ORCJIT, tigerlake)
  Threads: 1 on 16 virtual cores

…er mttkrps in grad_U

alexmul1114 · 2024-03-13T15:49:39Z

Benchmark Report using new get index for CPD, comparing MTTKRPs implementations.

Benchmark Report for `GCPDecompositions`

Job Properties

Time of benchmarks:
- Target: 11 Mar 2024 - 19:22
- Baseline: 11 Mar 2024 - 19:24
Package commits:
- Target: 49d34e
- Baseline: 6aa305
Julia commits:
- Target: 312098
- Baseline: 312098
Julia command flags:
- Target: None
- Baseline: None
Environment variables:
- Target: GCP_BENCHMARK_SUITES => gcp
- Baseline: GCP_BENCHMARK_SUITES => gcp

Results

A ratio greater than 1.0 denotes a possible regression (marked with ❌), while a ratio less
than 1.0 denotes a possible improvement (marked with ✅). Only significant results - results
that indicate possible regressions or improvements - are shown below (thus, an empty table means that all
benchmark results remained invariant between builds).

ID	time ratio	memory ratio
`["gcp", "bernoulliOdds-size(X)=(15, 20, 25), rank(X)=1"]`	1.11 (5%) ❌	1.07 (1%) ❌
`["gcp", "bernoulliOdds-size(X)=(15, 20, 25), rank(X)=2"]`	0.87 (5%) ✅	0.88 (1%) ✅
`["gcp", "bernoulliOdds-size(X)=(30, 40, 50), rank(X)=1"]`	0.97 (5%)	0.96 (1%) ✅
`["gcp", "gamma-size(X)=(15, 20, 25), rank(X)=1"]`	1.02 (5%)	1.02 (1%) ❌
`["gcp", "gamma-size(X)=(15, 20, 25), rank(X)=2"]`	1.02 (5%)	1.11 (1%) ❌
`["gcp", "gamma-size(X)=(30, 40, 50), rank(X)=1"]`	0.94 (5%) ✅	0.92 (1%) ✅
`["gcp", "gamma-size(X)=(30, 40, 50), rank(X)=2"]`	0.94 (5%) ✅	0.95 (1%) ✅
`["gcp", "poisson-size(X)=(15, 20, 25), rank(X)=1"]`	1.04 (5%)	1.08 (1%) ❌
`["gcp", "poisson-size(X)=(15, 20, 25), rank(X)=2"]`	0.63 (5%) ✅	0.46 (1%) ✅
`["gcp", "poisson-size(X)=(30, 40, 50), rank(X)=1"]`	0.96 (5%)	0.97 (1%) ✅
`["gcp", "poisson-size(X)=(30, 40, 50), rank(X)=2"]`	0.87 (5%) ✅	0.96 (1%) ✅

Benchmark Group List

Here's a list of all the benchmark groups executed by this job:

["gcp"]

Julia versioninfo

Target

Julia Version 1.10.0
Commit 3120989f39 (2023-12-25 18:01 UTC)
Build Info:
  Official https://julialang.org/ release
Platform Info:
  OS: Windows (x86_64-w64-mingw32)
      Microsoft Windows [Version 10.0.22621.3155]
  CPU: 11th Gen Intel(R) Core(TM) i7-11800H @ 2.30GHz: 
                 speed         user         nice          sys         idle          irq
       #1-16  2304 MHz  107470103            0     79055961    13638613649      2131432  ticks
  Memory: 31.726390838623047 GB (12977.23828125 MB free)
  Uptime: 1.663074234e6 sec
  Load Avg:  0.0  0.0  0.0
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-15.0.7 (ORCJIT, tigerlake)
  Threads: 1 on 16 virtual cores

Baseline

Julia Version 1.10.0
Commit 3120989f39 (2023-12-25 18:01 UTC)
Build Info:
  Official https://julialang.org/ release
Platform Info:
  OS: Windows (x86_64-w64-mingw32)
      Microsoft Windows [Version 10.0.22621.3155]
  CPU: 11th Gen Intel(R) Core(TM) i7-11800H @ 2.30GHz: 
                 speed         user         nice          sys         idle          irq
       #1-16  2304 MHz  107567853            0     79065914    13640122431      2131572  ticks
  Memory: 31.726390838623047 GB (12880.9140625 MB free)
  Uptime: 1.663175265e6 sec
  Load Avg:  0.0  0.0  0.0
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-15.0.7 (ORCJIT, tigerlake)
  Threads: 1 on 16 virtual cores

Add faster mttkrps algorithm

80adc72

alexmul1114 added 11 commits February 29, 2024 12:35

Replace mul(GU[n][:,r], ...) with GU[n][:,r]=...

0807001

Remove unnecessary copies

043575a

Fix typo

afbd963

Use mul!(view(GU[n]),:,r)...)

d6cc7e3

Add larger gcp benchmarks (baseline)

bb738e3

Add larger gcp benchmarks (new)

4bb9716

Rename benchmark file (baseline)

9b25f93

Rename benchmark file (new)

a9f9d2a

Replace stack with loop for n==N case

4e78b0b

Reduce number of large benchmarks (baseline)

62ce59c

Reduce number of large benchmarks (new)

4003e95

alexmul1114 added 2 commits March 5, 2024 11:49

Redo gcp_func to use mapreduce instead of sum (baseline)

b3af06c

Change ismissing to !ismissing

2d8cd8e

alexmul1114 and others added 6 commits March 5, 2024 12:30

Replace sum with mapreduce in type-cpdt to index into M (baseline)

678b542

Merge branch 'master' into pr/alexmul1114/40

e4c03d0

Change getindex for CPD

750fcbd

Update benchmark manifest

2ba4e48

Baseline getindex

aef707d

New getindex

6aa3056

Add in extra special case to mttkrps algorithm for n=n_star, use fast…

49d34eb

…er mttkrps in grad_U

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Faster MTTKRPs Algorithm #40

Faster MTTKRPs Algorithm #40

alexmul1114 commented Feb 26, 2024

codecov bot commented Feb 26, 2024 •

edited

Loading

alexmul1114 commented Mar 4, 2024

Target

Baseline

alexmul1114 commented Mar 5, 2024 •

edited

Loading

alexmul1114 commented Mar 5, 2024

Target

Baseline

alexmul1114 commented Mar 8, 2024

Target

Baseline

alexmul1114 commented Mar 13, 2024

Target

Baseline

Faster MTTKRPs Algorithm #40

Are you sure you want to change the base?

Faster MTTKRPs Algorithm #40

Conversation

alexmul1114 commented Feb 26, 2024

codecov bot commented Feb 26, 2024 • edited Loading

Codecov Report

alexmul1114 commented Mar 4, 2024

Benchmark Report for GCPDecompositions

Job Properties

Results

Benchmark Group List

Julia versioninfo

Target

Baseline

alexmul1114 commented Mar 5, 2024 • edited Loading

alexmul1114 commented Mar 5, 2024

Benchmark Report for GCPDecompositions

Job Properties

Results

Benchmark Group List

Julia versioninfo

Target

Baseline

alexmul1114 commented Mar 8, 2024

Benchmark Report for GCPDecompositions

Job Properties

Results

Benchmark Group List

Julia versioninfo

Target

Baseline

alexmul1114 commented Mar 13, 2024

Benchmark Report for GCPDecompositions

Job Properties

Results

Benchmark Group List

Julia versioninfo

Target

Baseline

codecov bot commented Feb 26, 2024 •

edited

Loading

Benchmark Report for `GCPDecompositions`

alexmul1114 commented Mar 5, 2024 •

edited

Loading

Benchmark Report for `GCPDecompositions`

Benchmark Report for `GCPDecompositions`

Benchmark Report for `GCPDecompositions`