GPU evaluation of `Recurrence()` broken on Metal #473

mchitre · 2023-12-28T14:45:30Z

This works:

m = Dense(10 => 10)
ps, st = Lux.setup(rng, m)
data = rand(10,10,10)
Lux.apply(m, data, ps, st)  # works
Lux.apply(m, gpu_device()(data), gpu_device()(ps), gpu_device()(st))  # works

but this doesn't:

m = Recurrence(RNNCell(10 => 10))
ps, st = Lux.setup(rng, m)
data = rand(10,10,10)
Lux.apply(m, data, ps, st)  # works
Lux.apply(m, gpu_device()(data), gpu_device()(ps), gpu_device()(st))  # fails

Stack trace:

ERROR: ArgumentError: cannot take the CPU address of a MtlMatrix{Float32, Metal.MTL.MTLResourceStorageModePrivate}
Stacktrace:
  [1] unsafe_convert(::Type{Ptr{Float32}}, x::MtlMatrix{Float32, Metal.MTL.MTLResourceStorageModePrivate})
    @ Metal ~/.julia/packages/Metal/lnkVP/src/array.jl:148
  [2] gemm!(transA::Char, transB::Char, alpha::Float32, A::MtlMatrix{…}, B::SubArray{…}, beta::Float32, C::MtlMatrix{…})
    @ LinearAlgebra.BLAS ~/.julia/juliaup/julia-1.10.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.10/LinearAlgebra/src/blas.jl:1524
  [3] gemm_wrapper!(C::MtlMatrix{…}, tA::Char, tB::Char, A::MtlMatrix{…}, B::SubArray{…}, _add::LinearAlgebra.MulAddMul{…})
    @ LinearAlgebra ~/.julia/juliaup/julia-1.10.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.10/LinearAlgebra/src/matmul.jl:605
  [4] generic_matmatmul!(C::MtlMatrix{…}, tA::Char, tB::Char, A::MtlMatrix{…}, B::SubArray{…}, _add::LinearAlgebra.MulAddMul{…})
    @ LinearAlgebra ~/.julia/juliaup/julia-1.10.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.10/LinearAlgebra/src/matmul.jl:352
  [5] mul!
    @ ~/.julia/juliaup/julia-1.10.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.10/LinearAlgebra/src/matmul.jl:263 [inlined]
  [6] mul!
    @ ~/.julia/juliaup/julia-1.10.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.10/LinearAlgebra/src/matmul.jl:237 [inlined]
  [7] *(A::MtlMatrix{Float32, Metal.MTL.MTLResourceStorageModePrivate}, B::SubArray{Float32, 2, MtlArray{…}, Tuple{…}, false})
    @ LinearAlgebra ~/.julia/juliaup/julia-1.10.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.10/LinearAlgebra/src/matmul.jl:113
  [8] (::RNNCell{…})(::Tuple{…}, ps::@NamedTuple{…}, st::@NamedTuple{…})
    @ Lux ~/.julia/packages/Lux/VFyfk/src/layers/recurrent.jl:259
  [9] (::RNNCell{true})(::Tuple{AbstractMatrix, Tuple{AbstractMatrix}}, ps::Any, st::NamedTuple)
    @ Lux ~/.julia/packages/Lux/VFyfk/src/layers/recurrent.jl:245 [inlined]
 [10] apply
    @ ~/.julia/packages/LuxCore/aumFq/src/LuxCore.jl:115 [inlined]
 [11] (::Recurrence{…})(x::Vector{…}, ps::@NamedTuple{…}, st::@NamedTuple{…})
    @ Lux ~/.julia/packages/Lux/VFyfk/src/layers/recurrent.jl:78
 [12] apply
    @ LuxCore ~/.julia/packages/LuxCore/aumFq/src/LuxCore.jl:115 [inlined]
 [13] Recurrence
    @ Lux ~/.julia/packages/Lux/VFyfk/src/layers/recurrent.jl:74 [inlined]
 [14] apply(model::Recurrence{…}, x::MtlArray{…}, ps::@NamedTuple{…}, st::@NamedTuple{…})
    @ LuxCore ~/.julia/packages/LuxCore/aumFq/src/LuxCore.jl:115
 [15] top-level scope
    @ REPL[93]:1

The text was updated successfully, but these errors were encountered:

avik-pal · 2023-12-28T16:37:08Z

Can you also post the stacktrace via show(err)?

mchitre · 2023-12-28T16:51:50Z

julia> show(err)
1-element ExceptionStack:
ArgumentError: cannot take the CPU address of a MtlMatrix{Float32, Metal.MTL.MTLResourceStorageModePrivate}
Stacktrace:
  [1] unsafe_convert(::Type{Ptr{Float32}}, x::MtlMatrix{Float32, Metal.MTL.MTLResourceStorageModePrivate})
    @ Metal ~/.julia/packages/Metal/lnkVP/src/array.jl:148
  [2] gemm!(transA::Char, transB::Char, alpha::Float32, A::MtlMatrix{Float32, Metal.MTL.MTLResourceStorageModePrivate}, B::SubArray{Float32, 2, MtlArray{Float32, 3, Metal.MTL.MTLResourceStorageModePrivate}, Tuple{Base.Slice{Base.OneTo{Int64}}, Int64, Base.Slice{Base.OneTo{Int64}}}, false}, beta::Float32, C::MtlMatrix{Float32, Metal.MTL.MTLResourceStorageModePrivate})
    @ LinearAlgebra.BLAS ~/.julia/juliaup/julia-1.10.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.10/LinearAlgebra/src/blas.jl:1524
  [3] gemm_wrapper!(C::MtlMatrix{Float32, Metal.MTL.MTLResourceStorageModePrivate}, tA::Char, tB::Char, A::MtlMatrix{Float32, Metal.MTL.MTLResourceStorageModePrivate}, B::SubArray{Float32, 2, MtlArray{Float32, 3, Metal.MTL.MTLResourceStorageModePrivate}, Tuple{Base.Slice{Base.OneTo{Int64}}, Int64, Base.Slice{Base.OneTo{Int64}}}, false}, _add::LinearAlgebra.MulAddMul{true, true, Bool, Bool})
    @ LinearAlgebra ~/.julia/juliaup/julia-1.10.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.10/LinearAlgebra/src/matmul.jl:605
  [4] generic_matmatmul!(C::MtlMatrix{Float32, Metal.MTL.MTLResourceStorageModePrivate}, tA::Char, tB::Char, A::MtlMatrix{Float32, Metal.MTL.MTLResourceStorageModePrivate}, B::SubArray{Float32, 2, MtlArray{Float32, 3, Metal.MTL.MTLResourceStorageModePrivate}, Tuple{Base.Slice{Base.OneTo{Int64}}, Int64, Base.Slice{Base.OneTo{Int64}}}, false}, _add::LinearAlgebra.MulAddMul{true, true, Bool, Bool})
    @ LinearAlgebra ~/.julia/juliaup/julia-1.10.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.10/LinearAlgebra/src/matmul.jl:352
  [5] mul!
    @ ~/.julia/juliaup/julia-1.10.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.10/LinearAlgebra/src/matmul.jl:263 [inlined]
  [6] mul!
    @ ~/.julia/juliaup/julia-1.10.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.10/LinearAlgebra/src/matmul.jl:237 [inlined]
  [7] *(A::MtlMatrix{Float32, Metal.MTL.MTLResourceStorageModePrivate}, B::SubArray{Float32, 2, MtlArray{Float32, 3, Metal.MTL.MTLResourceStorageModePrivate}, Tuple{Base.Slice{Base.OneTo{Int64}}, Int64, Base.Slice{Base.OneTo{Int64}}}, false})
    @ LinearAlgebra ~/.julia/juliaup/julia-1.10.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.10/LinearAlgebra/src/matmul.jl:113
  [8] (::RNNCell{true, false, typeof(tanh), typeof(zeros32), typeof(glorot_uniform), typeof(ones32)})(::Tuple{SubArray{Float32, 2, MtlArray{Float32, 3, Metal.MTL.MTLResourceStorageModePrivate}, Tuple{Base.Slice{Base.OneTo{Int64}}, Int64, Base.Slice{Base.OneTo{Int64}}}, false}, Tuple{Matrix{Float32}}}, ps::@NamedTuple{weight_ih::MtlMatrix{Float32, Metal.MTL.MTLResourceStorageModePrivate}, weight_hh::MtlMatrix{Float32, Metal.MTL.MTLResourceStorageModePrivate}, bias::MtlVector{Float32, Metal.MTL.MTLResourceStorageModePrivate}}, st::@NamedTuple{rng::Xoshiro})
    @ Lux ~/.julia/packages/Lux/VFyfk/src/layers/recurrent.jl:259
  [9] (::RNNCell{true})(::Tuple{AbstractMatrix, Tuple{AbstractMatrix}}, ps::Any, st::NamedTuple)
    @ Lux ~/.julia/packages/Lux/VFyfk/src/layers/recurrent.jl:245 [inlined]
 [10] apply
    @ ~/.julia/packages/LuxCore/aumFq/src/LuxCore.jl:115 [inlined]
 [11] (::Recurrence{false, RNNCell{true, false, typeof(tanh), typeof(zeros32), typeof(glorot_uniform), typeof(ones32)}, BatchLastIndex})(x::Vector{SubArray{Float32, 2, MtlArray{Float32, 3, Metal.MTL.MTLResourceStorageModePrivate}, Tuple{Base.Slice{Base.OneTo{Int64}}, Int64, Base.Slice{Base.OneTo{Int64}}}, false}}, ps::@NamedTuple{weight_ih::MtlMatrix{Float32, Metal.MTL.MTLResourceStorageModePrivate}, weight_hh::MtlMatrix{Float32, Metal.MTL.MTLResourceStorageModePrivate}, bias::MtlVector{Float32, Metal.MTL.MTLResourceStorageModePrivate}}, st::@NamedTuple{rng::Xoshiro})
    @ Lux ~/.julia/packages/Lux/VFyfk/src/layers/recurrent.jl:78
 [12] apply
    @ LuxCore ~/.julia/packages/LuxCore/aumFq/src/LuxCore.jl:115 [inlined]
 [13] Recurrence
    @ Lux ~/.julia/packages/Lux/VFyfk/src/layers/recurrent.jl:74 [inlined]
 [14] apply(model::Recurrence{false, RNNCell{true, false, typeof(tanh), typeof(zeros32), typeof(glorot_uniform), typeof(ones32)}, BatchLastIndex}, x::MtlArray{Float32, 3, Metal.MTL.MTLResourceStorageModePrivate}, ps::@NamedTuple{weight_ih::MtlMatrix{Float32, Metal.MTL.MTLResourceStorageModePrivate}, weight_hh::MtlMatrix{Float32, Metal.MTL.MTLResourceStorageModePrivate}, bias::MtlVector{Float32, Metal.MTL.MTLResourceStorageModePrivate}}, st::@NamedTuple{rng::Xoshiro})
    @ LuxCore ~/.julia/packages/LuxCore/aumFq/src/LuxCore.jl:115
 [15] top-level scope
    @ REPL[86]:1
 [16] eval
    @ ./boot.jl:385 [inlined]
 [17] eval
    @ ./Base.jl:88 [inlined]
 [18] repleval(m::Module, code::Expr, ::String)
    @ VSCodeServer ~/.vscode/extensions/julialang.language-julia-1.65.2/scripts/packages/VSCodeServer/src/repl.jl:229
 [19] (::VSCodeServer.var"#110#112"{Module, Expr, REPL.LineEditREPL, REPL.LineEdit.Prompt})()
    @ VSCodeServer ~/.vscode/extensions/julialang.language-julia-1.65.2/scripts/packages/VSCodeServer/src/repl.jl:192
 [20] with_logstate(f::Function, logstate::Any)
    @ Base.CoreLogging ./logging.jl:515
 [21] with_logger
    @ ./logging.jl:627 [inlined]
 [22] (::VSCodeServer.var"#109#111"{Module, Expr, REPL.LineEditREPL, REPL.LineEdit.Prompt})()
    @ VSCodeServer ~/.vscode/extensions/julialang.language-julia-1.65.2/scripts/packages/VSCodeServer/src/repl.jl:193
 [23] #invokelatest#2
    @ Base ./essentials.jl:887 [inlined]
 [24] invokelatest(::Any)
    @ Base ./essentials.jl:884
 [25] (::VSCodeServer.var"#62#63")()
    @ VSCodeServer ~/.vscode/extensions/julialang.language-julia-1.65.2/scripts/packages/VSCodeServer/src/eval.jl:34

avik-pal · 2023-12-28T16:58:54Z

Ah I see, it is Metal not handling wrappers properly. CUDA is quite good it this as it doesn't return a wrapper type if the storage is contiguous. For AMD I had to patch it at one point, I will see if that can be handled similarly for Metal.

mchitre · 2023-12-28T17:07:32Z

Yes, I confirm it works with CUDA. Changing issue title to reflect this.

avik-pal · 2023-12-28T17:19:01Z

As a temporary workaround, you could pass in a vector of matrices instead of the 3D Array.

To solve this problem at its core, either:

Metal needs to handle wrappers correctly and not dispatch to julia generic matmuls (upstream problem, not much I can do here)
Inside the core lux layers wherever a view is created, for Metal we just force a copy (maybe enabling it via a preferences)

mchitre · 2023-12-28T17:29:37Z

You mean like this?

data2 = [data[:,i,:] for i ∈ 1:10]
Lux.apply(m, gpu_device()(data2), gpu_device()(ps), gpu_device()(st))

On CPU they are equivalent. On Metal, it still doesn't work:

ERROR: ArgumentError: cannot take the CPU address of a MtlMatrix{Float32, Metal.MTL.MTLResourceStorageModePrivate}
Stacktrace:
  [1] unsafe_convert(::Type{Ptr{Float32}}, x::MtlMatrix{Float32, Metal.MTL.MTLResourceStorageModePrivate})
    @ Metal ~/.julia/packages/Metal/lnkVP/src/array.jl:148
  [2] gemm!(transA::Char, transB::Char, alpha::Float32, A::MtlMatrix{…}, B::Matrix{…}, beta::Float32, C::Matrix{…})
    @ LinearAlgebra.BLAS ~/.julia/juliaup/julia-1.10.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.10/LinearAlgebra/src/blas.jl:1524
  [3] gemm_wrapper!(C::Matrix{…}, tA::Char, tB::Char, A::MtlMatrix{…}, B::Matrix{…}, _add::LinearAlgebra.MulAddMul{…})
    @ LinearAlgebra ~/.julia/juliaup/julia-1.10.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.10/LinearAlgebra/src/matmul.jl:605
  [4] generic_matmatmul!
    @ ~/.julia/juliaup/julia-1.10.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.10/LinearAlgebra/src/matmul.jl:352 [inlined]
  [5] mul!
    @ ~/.julia/juliaup/julia-1.10.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.10/LinearAlgebra/src/matmul.jl:263 [inlined]
  [6] mul!
    @ ~/.julia/juliaup/julia-1.10.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.10/LinearAlgebra/src/matmul.jl:237 [inlined]
  [7] *(A::MtlMatrix{Float32, Metal.MTL.MTLResourceStorageModePrivate}, B::Matrix{Float32})
    @ LinearAlgebra ~/.julia/juliaup/julia-1.10.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.10/LinearAlgebra/src/matmul.jl:113
  [8] (::RNNCell{…})(::Tuple{…}, ps::@NamedTuple{…}, st::@NamedTuple{…})
    @ Lux ~/.julia/packages/Lux/VFyfk/src/layers/recurrent.jl:259
  [9] (::RNNCell{true})(::Tuple{AbstractMatrix, Tuple{AbstractMatrix}}, ps::Any, st::NamedTuple)
    @ Lux ~/.julia/packages/Lux/VFyfk/src/layers/recurrent.jl:245 [inlined]
 [10] apply
    @ ~/.julia/packages/LuxCore/aumFq/src/LuxCore.jl:115 [inlined]
 [11] (::Recurrence{…})(x::Vector{…}, ps::@NamedTuple{…}, st::@NamedTuple{…})
    @ Lux ~/.julia/packages/Lux/VFyfk/src/layers/recurrent.jl:78
 [12] apply(model::Recurrence{…}, x::Vector{…}, ps::@NamedTuple{…}, st::@NamedTuple{…})
    @ LuxCore ~/.julia/packages/LuxCore/aumFq/src/LuxCore.jl:115
 [13] top-level scope
    @ REPL[11]:1

mchitre · 2023-12-28T17:31:40Z

For upstream issue, worth opening an issue on Metal.jl? I don't understand details enough to do it, so might need you to open one if you think it worthwhile.

avik-pal · 2023-12-28T17:44:25Z

try data2 .|> gpu_device()

mchitre · 2023-12-31T04:07:54Z

data2 = [data[:,i,:] for i ∈ 1:10]
data2 = data2 .|> gpu_device()
Lux.apply(m, data2, gpu_device()(ps), gpu_device()(st))

still gives:

ERROR: ArgumentError: cannot take the CPU address of a MtlMatrix{Float32, Metal.MTL.MTLResourceStorageModePrivate}
Stacktrace:
  [1] unsafe_convert(::Type{Ptr{Float32}}, x::MtlMatrix{Float32, Metal.MTL.MTLResourceStorageModePrivate})
    @ Metal ~/.julia/packages/Metal/lnkVP/src/array.jl:148
  [2] gemm!(transA::Char, transB::Char, alpha::Float32, A::MtlMatrix{…}, B::Matrix{…}, beta::Float32, C::Matrix{…})
    @ LinearAlgebra.BLAS ~/.julia/juliaup/julia-1.10.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.10/LinearAlgebra/src/blas.jl:1524
  [3] gemm_wrapper!(C::Matrix{…}, tA::Char, tB::Char, A::MtlMatrix{…}, B::Matrix{…}, _add::LinearAlgebra.MulAddMul{…})
    @ LinearAlgebra ~/.julia/juliaup/julia-1.10.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.10/LinearAlgebra/src/matmul.jl:605
  [4] generic_matmatmul!
    @ ~/.julia/juliaup/julia-1.10.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.10/LinearAlgebra/src/matmul.jl:352 [inlined]
  [5] mul!
    @ ~/.julia/juliaup/julia-1.10.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.10/LinearAlgebra/src/matmul.jl:263 [inlined]
  [6] mul!
    @ ~/.julia/juliaup/julia-1.10.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.10/LinearAlgebra/src/matmul.jl:237 [inlined]
  [7] *(A::MtlMatrix{Float32, Metal.MTL.MTLResourceStorageModePrivate}, B::Matrix{Float32})
    @ LinearAlgebra ~/.julia/juliaup/julia-1.10.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.10/LinearAlgebra/src/matmul.jl:113
  [8] (::RNNCell{…})(::Tuple{…}, ps::@NamedTuple{…}, st::@NamedTuple{…})
    @ Lux ~/.julia/packages/Lux/VFyfk/src/layers/recurrent.jl:259
  [9] (::RNNCell{true})(::Tuple{AbstractMatrix, Tuple{AbstractMatrix}}, ps::Any, st::NamedTuple)
    @ Lux ~/.julia/packages/Lux/VFyfk/src/layers/recurrent.jl:245 [inlined]
 [10] apply
    @ ~/.julia/packages/LuxCore/aumFq/src/LuxCore.jl:115 [inlined]
 [11] (::Recurrence{…})(x::Vector{…}, ps::@NamedTuple{…}, st::@NamedTuple{…})
    @ Lux ~/.julia/packages/Lux/VFyfk/src/layers/recurrent.jl:78
 [12] apply(model::Recurrence{…}, x::Vector{…}, ps::@NamedTuple{…}, st::@NamedTuple{…})
    @ LuxCore ~/.julia/packages/LuxCore/aumFq/src/LuxCore.jl:115
 [13] top-level scope
    @ REPL[26]:1
 [14] top-level scope
    @ ~/.julia/packages/Metal/lnkVP/src/initialization.jl:57

prbzrg · 2024-01-01T17:01:32Z

What if we use Metal APIs for RNG and data?

m = Recurrence(RNNCell(10 => 10))
ps, st = Lux.setup(Metal.gpuarrays_rng(), m)
data = Metal.rand(10,10,10)
Lux.apply(m, data, ps, st)

mchitre · 2024-01-02T03:06:34Z

julia> ps, st = Lux.setup(Metal.gpuarrays_rng(), m)
ERROR: MethodError: no method matching Random.Sampler(::Type{GPUArrays.RNG}, ::Random.SamplerType{UInt32}, ::Val{1})

Closest candidates are:
  Random.Sampler(::Type{<:AbstractRNG}, ::Random.Sampler, ::Union{Val{1}, Val{Inf}})
   @ Random ~/.julia/juliaup/julia-1.10.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.10/Random/src/Random.jl:147
  Random.Sampler(::Type{<:AbstractRNG}, ::Any, ::Union{Val{1}, Val{Inf}})
   @ Random ~/.julia/juliaup/julia-1.10.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.10/Random/src/Random.jl:183
  Random.Sampler(::Type{<:AbstractRNG}, ::BitSet, ::Union{Val{1}, Val{Inf}})
   @ Random ~/.julia/juliaup/julia-1.10.0+0.aarch64.apple.darwin14/share/julia/stdlib/v1.10/Random/src/generation.jl:450
  ...

avik-pal · 2024-01-02T09:27:00Z

We need a dispatch like

Lux.jl/ext/LuxLuxAMDGPUExt.jl

Lines 9 to 11 in f11c407

    
           @inline function Lux._init_hidden_state(rng::AbstractRNG, rnn, x::AMDGPU.AnyROCArray) 
        
               return ROCArray(rnn.init_state(rng, rnn.out_dims, size(x, 2))) 
        
           end

for Metal

avik-pal · 2024-01-02T10:24:22Z

@mchitre can you try out #474? I don't have Apple hardware handy currently to test out the patch. This should fix both the problems.

Regarding the patch @prbzrg posted, if that doesn't work for you it might be worthwhile opening an issue on Metal.jl

mchitre · 2024-01-03T17:45:30Z

@avik-pal yes, I confirm that #474 fixes the problem. The original code in the description runs with it.

avik-pal · 2024-01-06T06:45:44Z

The patch will be available from the next release

mchitre changed the title ~~GPU evaluation of Recurrence() broken?~~ GPU evaluation of Recurrence() broken on Metal Dec 28, 2023

avik-pal added the bug Something isn't working label Dec 28, 2023

avik-pal mentioned this issue Jan 2, 2024

Patch Metal Recurrent Neural Networks #474

Merged

avik-pal closed this as completed in #474 Jan 6, 2024

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

GPU evaluation of `Recurrence()` broken on Metal #473

GPU evaluation of `Recurrence()` broken on Metal #473

mchitre commented Dec 28, 2023

avik-pal commented Dec 28, 2023

mchitre commented Dec 28, 2023

avik-pal commented Dec 28, 2023

mchitre commented Dec 28, 2023

avik-pal commented Dec 28, 2023

mchitre commented Dec 28, 2023

mchitre commented Dec 28, 2023

avik-pal commented Dec 28, 2023

mchitre commented Dec 31, 2023

prbzrg commented Jan 1, 2024

mchitre commented Jan 2, 2024

avik-pal commented Jan 2, 2024

avik-pal commented Jan 2, 2024

mchitre commented Jan 3, 2024

avik-pal commented Jan 6, 2024

GPU evaluation of Recurrence() broken on Metal #473

GPU evaluation of Recurrence() broken on Metal #473

Comments

mchitre commented Dec 28, 2023

avik-pal commented Dec 28, 2023

mchitre commented Dec 28, 2023

avik-pal commented Dec 28, 2023

mchitre commented Dec 28, 2023

avik-pal commented Dec 28, 2023

mchitre commented Dec 28, 2023

mchitre commented Dec 28, 2023

avik-pal commented Dec 28, 2023

mchitre commented Dec 31, 2023

prbzrg commented Jan 1, 2024

mchitre commented Jan 2, 2024

avik-pal commented Jan 2, 2024

avik-pal commented Jan 2, 2024

mchitre commented Jan 3, 2024

avik-pal commented Jan 6, 2024

GPU evaluation of `Recurrence()` broken on Metal #473

GPU evaluation of `Recurrence()` broken on Metal #473