Skip to content

Float vs Complex Performance

Compiled speedups split by dtype. Complex-native operators use complex64; the rest use their natural real dtype.

Performance by dtype

torch.complex64

Operator Level Method Device Speedup Class Details
Interpolate wrapper fwd cpu 3.96x works details
Interpolate inner fwd cpu 3.83x works details
Interpolate inner adj cpu 3.66x works details
Interpolate wrapper adj cpu 3.57x works details
ArrayToBlocks wrapper fwd cpu 3.01x works details
ArrayToBlocks wrapper adj cpu 2.74x works details
BlocksToArray wrapper adj cpu 2.72x works details
functional.ungrid inner fwd cpu 2.57x works details
BlocksToArray wrapper fwd cpu 2.53x works details
functional.grid inner fwd cpu 2.37x works details
ArrayToBlocks wrapper fwd cuda 2.22x works details
BlocksToArray wrapper adj cuda 2.10x works details
ArrayToBlocks wrapper adj cuda 2.03x works details
BlocksToArray wrapper fwd cuda 1.91x works details
FastSamplingNUFFT wrapper fwd cuda 1.82x works details
FastSamplingNUFFT inner fwd cuda 1.71x works details
Interpolate wrapper adj cuda 1.68x works details
Interpolate wrapper fwd cuda 1.60x works details
Interpolate inner adj cuda 1.57x works details
Interpolate inner fwd cuda 1.47x works details
NUFFT wrapper adj cpu 1.36x works details
NUFFT wrapper fwd cpu 1.36x works details
FastSamplingNUFFT wrapper fwd cpu 1.34x works details
NUFFT inner fwd cpu 1.29x works details
functional.fold inner fwd cpu 1.29x works details
FastSamplingNUFFT inner fwd cpu 1.26x works details
NUFFT inner adj cpu 1.26x works details
functional.unfold inner fwd cuda 1.25x works details
functional.unfold inner fwd cpu 1.18x works details
functional.fold inner fwd cuda 1.13x works details
SamplingNUFFT wrapper fwd cpu 0.97x works details
SamplingNUFFT inner fwd cpu 0.94x works details
NUFFT wrapper fwd cuda 0.92x silent_degradation details
NUFFT wrapper adj cuda 0.88x compiles_slower details
NUFFT inner fwd cuda 0.87x compiles_slower details
SamplingNUFFT wrapper fwd cuda 0.85x compiles_slower details
functional.ungrid inner fwd cuda 0.84x compiles_slower details
ToeplitzPSF wrapper fwd cpu 0.83x compiles_slower details
SamplingNUFFT wrapper adj cpu 0.83x works_with_breaks details
functional.grid inner fwd cuda 0.83x compiles_slower details
FastSamplingNUFFT wrapper adj cpu 0.78x works_with_breaks details
NUFFT inner adj cuda 0.77x compiles_slower details
SamplingNUFFT inner adj cpu 0.77x works_with_breaks details
FastSamplingNUFFT inner adj cpu 0.75x works_with_breaks details
ToeplitzPSF wrapper adj cpu 0.75x compiles_slower details
SamplingNUFFT inner fwd cuda 0.73x compiles_slower details
ToeplitzPSF wrapper fwd cuda 0.72x compiles_slower details
FastSamplingNUFFT wrapper adj cuda 0.67x works_with_breaks details
ToeplitzPSF wrapper adj cuda 0.67x compiles_slower details
SamplingNUFFT wrapper adj cuda 0.66x works_with_breaks details
FastSamplingNUFFT inner adj cuda 0.64x works_with_breaks details
SamplingNUFFT inner adj cuda 0.60x works_with_breaks details
ToeplitzPSF inner adj cpu 0.50x compiles_slower details
ToeplitzPSF inner fwd cuda 0.50x compiles_slower details
ToeplitzPSF inner fwd cpu 0.49x compiles_slower details
ToeplitzPSF inner adj cuda 0.49x compiles_slower details
FFT wrapper adj cpu 0.44x compiles_slower details
FFT wrapper fwd cpu 0.44x compiles_slower details
FFT wrapper fwd cuda 0.43x compiles_slower details
FFT wrapper adj cuda 0.43x compiles_slower details
FFT inner fwd cpu 0.40x compiles_slower details
FFT inner adj cpu 0.40x compiles_slower details
FFT inner fwd cuda 0.39x compiles_slower details
FFT inner adj cuda 0.39x compiles_slower details

torch.float32

Operator Level Method Device Speedup Class Details
Add wrapper fwd cpu 5.65x works details
Add wrapper adj cpu 5.30x works details
Chain wrapper fwd cpu 4.80x works details
Chain wrapper adj cpu 4.67x works details
Sampling wrapper fwd cuda 3.00x works details
Sampling wrapper fwd cpu 2.69x works details
Add wrapper fwd cuda 2.62x works details
Add wrapper adj cuda 2.43x works details
Chain wrapper fwd cuda 2.37x works details
Chain wrapper adj cuda 2.33x works details
DenseSimple wrapper fwd cpu 1.72x works details
Dense wrapper fwd cpu 1.68x works details
Dense wrapper adj cpu 1.62x works details
DenseSimple wrapper adj cpu 1.54x works details
DenseSimple wrapper fwd cuda 1.43x works details
Dense wrapper adj cuda 1.38x works details
DenseSimple wrapper adj cuda 1.33x works details
Dense wrapper fwd cuda 1.30x works details
Sampling wrapper adj cpu 1.28x works details
SumReduce wrapper fwd cpu 1.20x works details
Repeat wrapper adj cpu 1.17x works details
Sampling wrapper adj cuda 1.15x works details
Pad wrapper fwd cpu 1.14x works details
SimpleLinop wrapper adj cpu 1.12x works details
ToDevice wrapper fwd cpu 1.12x works details
ToDevice wrapper adj cpu 1.10x works details
Repeat wrapper fwd cpu 1.01x works details
SumReduce wrapper adj cpu 1.01x works details
Convolution wrapper fwd cpu 0.99x works details
Concat wrapper fwd cpu 0.98x works details
Truncate wrapper adj cpu 0.98x works details
Convolution wrapper adj cpu 0.97x works details
Concat wrapper adj cuda 0.97x works_with_breaks details
PadDim wrapper fwd cuda 0.97x works details
SumReduce wrapper fwd cuda 0.96x works details
PadDim wrapper fwd cpu 0.96x works details
ToDevice wrapper fwd cuda 0.95x works details
Stack wrapper adj cuda 0.95x works_with_breaks details
Pad wrapper fwd cuda 0.95x works details
Truncate wrapper adj cuda 0.94x works details
Concat wrapper fwd cuda 0.94x works_with_breaks details
ToDevice wrapper adj cuda 0.94x works details
Stack wrapper fwd cpu 0.93x works details
Concat wrapper adj cpu 0.92x works details
SimpleLinop wrapper fwd cpu 0.92x works details
Repeat wrapper adj cuda 0.92x works details
SumReduce wrapper adj cuda 0.91x works details
Repeat wrapper fwd cuda 0.91x works details
Convolution wrapper adj cuda 0.88x compiles_slower details
Stack wrapper fwd cuda 0.88x works_with_breaks details
SimpleLinop wrapper fwd cuda 0.85x compiles_slower details
Stack wrapper adj cpu 0.85x compiles_slower details
SimpleLinop wrapper adj cuda 0.85x compiles_slower details
Convolution inner adj cpu 0.84x compiles_slower details
Rearrange wrapper adj cpu 0.84x compiles_slower details
Convolution inner fwd cpu 0.83x compiles_slower details
Convolution wrapper fwd cuda 0.83x compiles_slower details
Scalar wrapper fwd cpu 0.82x compiles_slower details
Zero wrapper adj cpu 0.82x compiles_slower details
Diagonal wrapper fwd cpu 0.82x compiles_slower details
Scalar wrapper adj cpu 0.81x compiles_slower details
Diagonal wrapper adj cpu 0.81x compiles_slower details
Zero wrapper fwd cpu 0.80x compiles_slower details
Rearrange wrapper fwd cpu 0.77x compiles_slower details
Diagonal wrapper fwd cuda 0.76x compiles_slower details
Pad wrapper adj cpu 0.73x compiles_slower details
Truncate wrapper fwd cpu 0.72x compiles_slower details
ShapeSpec wrapper fwd cpu 0.72x compiles_slower details
Scalar wrapper fwd cuda 0.72x compiles_slower details
Diagonal wrapper adj cuda 0.72x compiles_slower details
Zero wrapper fwd cuda 0.72x compiles_slower details
Rearrange wrapper adj cuda 0.71x compiles_slower details
Identity wrapper adj cpu 0.70x compiles_slower details
Convolution inner adj cuda 0.70x compiles_slower details
ShapeSpec wrapper adj cuda 0.70x compiles_slower details
Rearrange wrapper fwd cuda 0.69x compiles_slower details
Identity wrapper fwd cpu 0.69x compiles_slower details
PadDim wrapper adj cpu 0.69x compiles_slower details
Identity wrapper adj cuda 0.69x compiles_slower details
ShapeSpec wrapper fwd cuda 0.67x compiles_slower details
ShapeSpec wrapper adj cpu 0.67x compiles_slower details
Identity wrapper fwd cuda 0.66x compiles_slower details
Scalar wrapper adj cuda 0.65x compiles_slower details
Zero wrapper adj cuda 0.65x compiles_slower details
Pad wrapper adj cuda 0.64x compiles_slower details
Truncate wrapper fwd cuda 0.64x compiles_slower details
PadDim wrapper adj cuda 0.63x compiles_slower details
Convolution inner fwd cuda 0.59x compiles_slower details

Charts

CPU

Compiled speedup — CPU

GPU

Compiled speedup — GPU


See Methodology for timing, graph-break, and level definitions.