Aidge backend CUDA#

When imported automatically update availables implementations for aidge core objects.

66% (97/148)
Supported operators: ✔️ Yes (∇ = grad), ✅ Generic impl., ☑️ MetaOp impl., ❌ No#

Operator (⧖ = custom prod. conso.)

Support
Impl specs
(— = any ; ✱ = additional requirements)
Abs
(aidge_core.Abs)

✔️
1. float64 — → float64 —
2. float32 — → float32 —
3. float16 — → float16 —
Add
(aidge_core.Add)

✔️ ∇
1. float64 — → float64 —
2. float32 — → float32 —
3. float16 — → float16 —
And
(aidge_core.And)

❌
ArgMax
(aidge_core.ArgMax)

✔️
1. float64 — → int64 —
2. float32 — → int64 —
3. float16 — → int64 —
ArgMin
(aidge_core.ArgMin)

❌
Asin
(aidge_core.Asin)

❌
Atan
(aidge_core.Atan)

❌
AvgPooling1D
(aidge_core.AvgPooling1D)

❌
AvgPooling2D
(aidge_core.AvgPooling2D)

✔️ ∇
1. float64 — → float64 —
2. float64 — → float64 — | int64 —
3. float32 — → float32 —
4. float32 — → float32 — | int64 —
5. float16 — → float16 —
6. float16 — → float16 — | int64 —
AvgPooling3D
(aidge_core.AvgPooling3D)

✅
BatchNorm2D
(aidge_core.BatchNorm2D)

✔️ ∇
1. float64 — → float64 —
2. float32 — → float32 —
3. float16 — → float16 —
⚙ BatchNorm2D impl. attrs
BitErrorRate
(aidge_core.BitErrorRate)

❌
BitShift
(aidge_core.BitShift)

✔️
1. int32 — → int32 —
2. int64 — → int64 —
CastLike
(aidge_core.CastLike)

✅
Cast
(aidge_core.Cast)

✅
Ceil
(aidge_core.Ceil)

❌
Clip
(aidge_core.Clip)

✔️ ∇
1. float64 — → float64 —
2. float32 — → float32 —
3. float16 — → float16 —
4. int32 — → int32 —
5. int64 — → int64 —
ComplexToInnerPair
(aidge_core.ComplexToInnerPair)

✅
Concat
(aidge_core.Concat)

✅
ConnectedComponentLabeling
(aidge_core.ConnectedComponentLabeling)

❌
ConstantOfShape
(aidge_core.ConstantOfShape)

✅
Conv1D
(aidge_core.Conv1D)

✔️ ∇
1. float64 nchw → float64 nchw
2. float32 nchw → float32 nchw
3. float16 nchw → float16 nchw
4. int32 nchw → int32 nchw
Conv2D
(aidge_core.Conv2D)

✔️ ∇
1. float64 nchw → float64 nchw
2. float32 nchw → float32 nchw
3. float16 nchw → float16 nchw
4. int32 nchw → int32 nchw
⚙ Conv2D impl. attrs
Conv3D
(aidge_core.Conv3D)

❌
ConvDepthWise1D
(aidge_core.ConvDepthWise1D)

✔️ ∇
1. float64 nchw → float64 nchw
2. float32 nchw → float32 nchw
3. float16 nchw → float16 nchw
4. int32 nchw → int32 nchw
ConvDepthWise2D
(aidge_core.ConvDepthWise2D)

✔️ ∇
1. float64 nchw → float64 nchw
2. float32 nchw → float32 nchw
3. float16 nchw → float16 nchw
4. int32 nchw → int32 nchw
ConvTranspose1D
(aidge_core.ConvTranspose1D)

❌
ConvTranspose2D
(aidge_core.ConvTranspose2D)

❌
ConvTranspose3D
(aidge_core.ConvTranspose3D)

❌
Cosh
(aidge_core.Cosh)

❌
Cos
(aidge_core.Cos)

❌
CryptoHash
(aidge_core.CryptoHash)

✅
Cusum
(aidge_core.Cusum)

☑️
DepthToSpace
(aidge_core.DepthToSpace)

✅
DFT
(aidge_core.DFT)

❌
Div
(aidge_core.Div)

✔️
1. float64 — → float64 —
2. float32 — → float32 —
3. float16 — → float16 —
DropBlock
(aidge_core.DropBlock)

✔️
1. float64 — → float64 — | boolean —
2. float32 — → float32 — | boolean —
3. float16 — → float16 — | boolean —
Dropout
(aidge_core.Dropout)

✔️
1. float64 — | — — | boolean — → float64 — | boolean —
2. float32 — | — — | boolean — → float32 — | boolean —
3. float16 — | — — | boolean — → float16 — | boolean —
Equal
(aidge_core.Equal)

✔️
1. float64 — → boolean —
2. float32 — → boolean —
3. float16 — → boolean —
4. boolean — → boolean —
5. int32 — → boolean —
6. int64 — → boolean —
Erf
(aidge_core.Erf)

✔️
1. float64 — → float64 —
2. float32 — → float32 —
3. int8 — → int8 —
4. int32 — → int32 —
Expand
(aidge_core.Expand)

❌
Exp
(aidge_core.Exp)

✔️ ∇
1. float64 — → float64 —
2. float32 — → float32 —
3. float16 — → float16 —
FC
(aidge_core.FC)

✔️ ∇
1. float64 — → float64 —
2. float32 — → float32 —
3. float16 — → float16 —
FixedNBitFlip
(aidge_core.FixedNBitFlip)

❌
Flatten
(aidge_core.Flatten)

✅
Floor
(aidge_core.Floor)

❌
Fold2D
(aidge_core.Fold2D)

❌
GatherElements
(aidge_core.GatherElements)

✅✔️ ∇
1. float64 — | int64 — → float64 —
2. float32 — | int64 — → float32 —
3. float16 — | int64 — → float16 —
4. int32 — | int64 — → int32 —
5. int64 — | int64 — → int64 —
6. uint32 — | int64 — → uint32 —
7. uint64 — | int64 — → uint64 —
GatherND
(aidge_core.GatherND)

✅
Gather
(aidge_core.Gather)

✅
Gemm
(aidge_core.Gemm)

☑️
GlobalAveragePooling
(aidge_core.GlobalAveragePooling)

✔️ ∇
1. float64 — → float64 —
2. float32 — → float32 —
3. float16 — → float16 —
Greater
(aidge_core.Greater)

❌
GreaterOrEqual
(aidge_core.GreaterOrEqual)

❌
GridSample
(aidge_core.GridSample)

❌
GRU
(aidge_core.GRU)

☑️
HannWindow
(aidge_core.HannWindow)

☑️
Hardmax
(aidge_core.Hardmax)

❌
HardSigmoid
(aidge_core.HardSigmoid)

❌
HardSwish
(aidge_core.HardSwish)

☑️
Heaviside
(aidge_core.Heaviside)

✔️
1. float64 — → float64 —
2. float32 — → float32 —
3. float16 — → float16 —
Identity ⧖
(aidge_core.Identity)

✅
InnerPairToComplex
(aidge_core.InnerPairToComplex)

✅
InstanceNorm
(aidge_core.InstanceNorm)

❌
LayerNorm
(aidge_core.LayerNorm)

☑️
LeakyReLU
(aidge_core.LeakyReLU)

✔️
1. float64 — → float64 —
2. float32 — → float32 —
Less
(aidge_core.Less)

❌
LessOrEqual
(aidge_core.LessOrEqual)

❌
Ln
(aidge_core.Ln)

✔️
1. float64 — → float64 —
2. float32 — → float32 —
3. float16 — → float16 —
LogSoftmax
(aidge_core.LogSoftmax)

✔️ ∇
1. float64 — → float64 —
2. float32 — → float32 —
3. float16 — → float16 —
LRN
(aidge_core.LRN)

✔️ ∇
1. float64 — → float64 —
2. float32 — → float32 —
LSTM
(aidge_core.LSTM)

☑️
MatMul
(aidge_core.MatMul)

✔️
1. float64 — → float64 —
2. float32 — → float32 —
3. float16 — → float16 —
Max
(aidge_core.Max)

✔️
1. float64 — → float64 —
2. float32 — → float32 —
3. float16 — → float16 —
MaxPooling1D
(aidge_core.MaxPooling1D)

❌
MaxPooling2D
(aidge_core.MaxPooling2D)

✔️ ∇
1. float64 — → float64 —
2. float64 — → float64 — | int64 —
3. float32 — → float32 —
4. float32 — → float32 — | int64 —
5. float16 — → float16 —
6. float16 — → float16 — | int64 —
MaxPooling3D
(aidge_core.MaxPooling3D)

✅
Memorize ⧖
(aidge_core.Memorize)

✅
Min
(aidge_core.Min)

✔️
1. float64 — → float64 —
2. float32 — → float32 —
3. float16 — → float16 —
Mod
(aidge_core.Mod)

❌
Mul
(aidge_core.Mul)

✔️ ∇
1. float64 — → float64 —
2. float32 — → float32 —
3. float16 — → float16 —
NBitFlip
(aidge_core.NBitFlip)

❌
Neg
(aidge_core.Neg)

❌
NonMaxSuppression
(aidge_core.NonMaxSuppression)

❌
NonZero
(aidge_core.NonZero)

❌
Not
(aidge_core.Not)

❌
OneHot
(aidge_core.OneHot)

✔️
1. float64 — → float64 —
2. float32 — → float32 —
3. float16 — → float16 —
4. int32 — → int32 —
5. int64 — → int64 —
Pack
(aidge_core.Pack)

❌
PaddedAvgPooling2D
(aidge_core.PaddedAvgPooling2D)

☑️
PaddedConv1D
(aidge_core.PaddedConv1D)

☑️
PaddedConv2D
(aidge_core.PaddedConv2D)

☑️
PaddedConv3D
(aidge_core.PaddedConv3D)

☑️
PaddedConvDepthWise1D
(aidge_core.PaddedConvDepthWise1D)

☑️
PaddedConvDepthWise2D
(aidge_core.PaddedConvDepthWise2D)

☑️
PaddedConvTranspose1D
(aidge_core.PaddedConvTranspose1D)

☑️
PaddedConvTranspose2D
(aidge_core.PaddedConvTranspose2D)

☑️
PaddedConvTranspose3D
(aidge_core.PaddedConvTranspose3D)

☑️
PaddedMaxPooling2D
(aidge_core.PaddedMaxPooling2D)

☑️
Pad
(aidge_core.Pad)

✔️ ∇
1. float64 — | int64 — | — — | — — → float64 —
2. float32 — | int64 — | — — | — — → float32 —
3. float16 — | int64 — | — — | — — → float16 —
4. int8 — | int64 — | — — | — — → int8 —
5. int32 — | int64 — | — — | — — → int32 —
Pop ⧖
(aidge_core.Pop)

✅
Pow
(aidge_core.Pow)

✔️
1. float64 — → float64 —
2. float32 — → float32 —
3. float16 — → float16 —
Producer
(aidge_core.Producer)

✅
RandomNormalLike
(aidge_core.RandomNormalLike)

✔️
1. float64 — → float64 —
2. float32 — → float32 —
3. float16 — → float16 —
Range
(aidge_core.Range)

✅
Reciprocal
(aidge_core.Reciprocal)

✔️ ∇
1. float64 — → float64 —
2. float32 — → float32 —
3. float16 — → float16 —
ReduceL1
(aidge_core.ReduceL1)

☑️
ReduceL2
(aidge_core.ReduceL2)

☑️
ReduceMax
(aidge_core.ReduceMax)

✔️ ∇
1. float64 — → float64 —
2. float32 — → float32 —
3. float16 — → float16 —
ReduceMean
(aidge_core.ReduceMean)

✔️ ∇
1. float64 — | int64 — → float64 —
2. float32 — | int64 — → float32 —
3. float16 — | int64 — → float16 —
ReduceMin
(aidge_core.ReduceMin)

✔️ ∇
1. float64 — → float64 —
2. float32 — → float32 —
3. float16 — → float16 —
ReduceSum
(aidge_core.ReduceSum)

✔️ ∇
1. float64 — → float64 —
2. float32 — → float32 —
3. float16 — → float16 —
ReLU
(aidge_core.ReLU)

✔️ ∇
1. float64 — → float64 —
2. float32 — → float32 —
3. float16 — → float16 —
Reshape
(aidge_core.Reshape)

✅
Resize
(aidge_core.Resize)

✔️ ∇
1. float64 — | — — | — — | — — → float64 —
2. float32 — | — — | — — | — — → float32 —
3. float16 — | — — | — — | — — → float16 —
Round
(aidge_core.Round)

✔️
1. float64 — → float64 —
2. float32 — → float32 —
Scatter
(aidge_core.Scatter)

❌
SDLayerNorm
(aidge_core.SDLayerNorm)

❌
SDShiftGELU
(aidge_core.SDShiftGELU)

❌
SDShiftmax
(aidge_core.SDShiftmax)

❌
SDShiftSiLU
(aidge_core.SDShiftSiLU)

❌
Select
(aidge_core.Select)

✅
Shape
(aidge_core.Shape)

✅
Sigmoid
(aidge_core.Sigmoid)

✔️ ∇
1. float64 — → float64 —
2. float32 — → float32 —
3. float16 — → float16 —
Sign
(aidge_core.Sign)

✔️
1. float64 — → float64 —
2. float32 — → float32 —
3. float16 — → float16 —
Sinh
(aidge_core.Sinh)

❌
Sin
(aidge_core.Sin)

❌
Size
(aidge_core.Size)

✅
Slice
(aidge_core.Slice)

❌
SLSTM
(aidge_core.SLSTM)

☑️
Softmax
(aidge_core.Softmax)

✔️ ∇
1. float64 — → float64 —
2. float32 — → float32 —
3. float16 — → float16 —
Split
(aidge_core.Split)

✅✔️
1. float64 — | — — → float64 —
2. float32 — | — — → float32 —
3. float16 — | — — → float16 —
4. int32 — | — — → int32 —
5. int64 — | — — → int64 —
6. uint32 — | — — → uint32 —
7. uint64 — | — — → uint64 —
Sqrt
(aidge_core.Sqrt)

✔️ ∇
1. float64 — → float64 —
2. float32 — → float32 —
Squeeze
(aidge_core.Squeeze)

✅
Stack ⧖
(aidge_core.Stack)

✅
STFT
(aidge_core.STFT)

❌
Sub
(aidge_core.Sub)

✔️ ∇
1. float64 — → float64 —
2. float32 — → float32 —
3. float16 — → float16 —
Sum
(aidge_core.Sum)

❌
SVMRegressor
(aidge_core.SVMRegressor)

❌
Tanh
(aidge_core.Tanh)

✔️ ∇
1. float64 — → float64 —
2. float32 — → float32 —
3. float16 — → float16 —
Tan
(aidge_core.Tan)

❌
Tile
(aidge_core.Tile)

✅
TopK
(aidge_core.TopK)

❌
Transpose
(aidge_core.Transpose)

✔️ ∇
1. float64 — → float64 —
2. float32 — → float32 —
3. float16 — → float16 —
4. int32 — → int32 —
5. int64 — → int64 —
6. uint32 — → uint32 —
7. uint64 — → uint64 —
Trilu
(aidge_core.Trilu)

❌
Unfold2D
(aidge_core.Unfold2D)

✅
Unpack
(aidge_core.Unpack)

❌
Unsqueeze
(aidge_core.Unsqueeze)

✅
Where
(aidge_core.Where)

❌

Additional implementation attributes for Conv2D

  • impl.cuda.no_auto_fwd_algo: Disable automatic selection of the best forward algorithm (CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_GEMM is used in this case)

Additional implementation attributes for BatchNorm2D

  • impl.cuda.no_cache: Disable cache of intermediate results during forward (cache is used to speed-up backward)