Memory corruption

[ojwoodford]

I am seeing some memory corruption in a function that I believe should be working. The code is:

using LoopVectorization

function At_mul_A!(C, A)
   @turbo for n in indices((C, A), 2), m in indices((C, A), (1, 2))
       Cmn = zero(eltype(C))
       for k in indices(A, 1)
           Cmn += A[k,m] * A[k,n]
       end
       C[m,n] = Cmn
   end
end

function test()
A = randn(16, 64)
C = randn(64, 64)
for a = 1:1000
    At_mul_A!(C, A)
end
end
test()

If I run this once, or sometimes a few times, I get the error:

julia(57934,0x1f15fe080) malloc: Incorrect checksum for freed object 0x12b391000: probably modified after being freed.
Corrupt value: 0x4037203caf449798
julia(57934,0x1f15fe080) malloc: *** set a breakpoint in malloc_error_break to debug

and julia crashes. I'm running Julia 1.10.0-beta2 and LoopVectorization v0.12.165.

I cannot see any reason that this shouldn't work - apologies if there is, but I really have looked it over several times, and run it without @turbo to check it's valid.

[chriselrod]

I got

julia> function test()
       A = randn(16, 64)
       C = randn(64, 64)
       for a = 1:1000
           At_mul_A!(C, A)
       end
       end
test (generic function with 1 method)

julia> @time test()
  0.004525 seconds (3 allocations: 40.172 KiB)

julia> @time test()
  0.004688 seconds (3 allocations: 40.172 KiB)

julia> @time test()
  0.004536 seconds (3 allocations: 40.172 KiB)

julia> @time test()
  0.004673 seconds (3 allocations: 40.172 KiB)

julia> @time foreach(_->test(),1:1000)
  4.794809 seconds (5.07 k allocations: 39.368 MiB, 0.20% gc time, 0.09% compilation time)

julia> function test2()
       A = randn(64, 16)' # 16x64
       C = randn(64, 64)
       for a = 1:1000
           At_mul_A!(C, A)
       end
       end
test2 (generic function with 1 method)

julia> @time test2()
  0.001879 seconds (3 allocations: 40.172 KiB)

julia> malloc(): invalid size (unsorted)

[7231] signal (6.-6): Aborted
in expression starting at none:0

It'd be faster to use a different memory layout if possible, which you can see: test2 is over twice as fast as test. That is, A_mul_At is faster than At_mul_A due to the better memory layout. This surprises most people who only think about a matrix multiply as a bunch of dot products; i.e. some people naively suspect that setting the layout to make a dot product of a view as fast as possible would be best, but column-major A' B is actually the worst of the four combinations, even `A' B'is better. A' * Arequires traversing memory much more quickly, increasing bandwdith requirements, and requires reductions at the end of thek` loop, requiring extra FLOPs. None of the other 3 orientations need this.

Anyway, I got a segfault, too.

I tried replaces indices with axes, and didn't get a segfault. It might be a bug in the optimizations it does for indices.

julia> using LoopVectorization

julia> function At_mul_A!(C, A)
          @turbo for n in axes(C, 2), m in axes(C, 1)
              Cmn = zero(eltype(C))
              for k in axes(A, 1)
                  Cmn += A[k,m] * A[k,n]
              end
              C[m,n] = Cmn
          end
       end
At_mul_A! (generic function with 1 method)

julia> function test()
       A = randn(16, 64)
       C = randn(64, 64)
       for a = 1:1000
           At_mul_A!(C, A)
       end
       end
test (generic function with 1 method)

julia> @time test()
  0.004558 seconds (3 allocations: 40.172 KiB)

julia> @time test()
  0.004533 seconds (3 allocations: 40.172 KiB)

julia> @time test()
  0.004505 seconds (3 allocations: 40.172 KiB)

julia> @time test()
  0.004553 seconds (3 allocations: 40.172 KiB)

julia> @time foreach(_->test(),1:1000)
  4.710555 seconds (12.46 k allocations: 39.802 MiB, 0.24% gc time, 0.08% compilation time)

julia> function test2()
       A = randn(64, 16)' # 16x64
       C = randn(64, 64)
       for a = 1:1000
           At_mul_A!(C, A)
       end
       end
test2 (generic function with 1 method)

julia> @time test2()
  0.002044 seconds (3 allocations: 40.172 KiB)

julia> @time test2()
  0.001877 seconds (3 allocations: 40.172 KiB)

julia> @time test2()
  0.001880 seconds (3 allocations: 40.172 KiB)

julia> @time test2()
  0.001868 seconds (3 allocations: 40.172 KiB)

julia> @time foreach(_->test2(),1:1000)
  2.011197 seconds (5.07 k allocations: 39.368 MiB, 0.45% gc time, 0.20% compilation time)

[ojwoodford]

Thanks, @chriselrod . Especially for the tips on faster layouts. However, I also get the memory corruption using axes instead of indices.