2D Extension of 1D Cooley-Tukey FFT. Refactoring original not necesss…#1
2D Extension of 1D Cooley-Tukey FFT. Refactoring original not necesss…#1bertiethorpe wants to merge 1 commit intomarkp-gc:mainfrom
Conversation
…ary as extension makes use of spatial variable separability. Included also is bash script for performing sweeps for data analysis
markp-gc
left a comment
markp-gc
left a comment
There was a problem hiding this comment.
Just need to avoid losing some of my recent chagnes: if you like I can merge just the new files (but then it won't credit you as the author).
| // Can only do the second expression in-place: | ||
| auto tmpReal = popops::map(graph, complexMulExprRe, {real, v.real, imag, v.imag}, | ||
| prog, debugPrefix + "/complex_mul_re"); | ||
| popops::mapInPlace(graph, complexMulExprRe, {real, v.real, imag, v.imag}, |
There was a problem hiding this comment.
Need to keep my change here. Only the second operation can be done in place because it needs the input from the first.
| poplar::Tensor partial = | ||
| poplin::matMul(graph, matrix.real, realBatch, prog, | ||
| elemType, debugStr + "/real_matmul", matmulOptions); | ||
| elemType, debugStr + "/real_matmul"); |
There was a problem hiding this comment.
Need to keep my change here (matmul options).
|
|
||
| poplin::matMulAcc(graph, partial, 1.f, matrix.imag, imagBatch, prog, | ||
| debugStr + "/imag_matmul", matmulOptions); | ||
| debugStr + "/imag_matmul"); |
| result_odd_remapped.mapLinearly(graph); | ||
| prog.add(copy(result_odd, result_odd_remapped)); | ||
| result_odd = result_odd_remapped; | ||
| } |
There was a problem hiding this comment.
Keep my change here: it uses a better layout for small FFTs to improve their performance and memory use.
| auto result_odd_remapped = ComplexTensor(graph, result_even.elementType(), result_even.shape(), "dft_even_remapped"); | ||
| result_odd_remapped.mapLinearly(graph); | ||
| prog.add(copy(result_odd, result_odd_remapped)); | ||
| result_odd = result_odd_remapped; |
There was a problem hiding this comment.
Keep this as it improves memory use and perf.
|
|
||
| ComplexTensor FFTBuilder::inverseFourierMatrices( | ||
| std::size_t length, poplar::Type elemType) { | ||
| const double twoPi_over_length = (2.0L / length) * 3.141592653589793238462643383279502884L; |
There was a problem hiding this comment.
Keep my change as it improves prexision significantly (same below).
| availableMemoryProportion(-1.f), flopEstimate(0) {} | ||
|
|
||
| /// Set the proportion of memory available for the inner DFT matrix-multiplies. | ||
| void setAvailableMemoryProportion(float proportion) { availableMemoryProportion = proportion; } |
| @@ -0,0 +1,202 @@ | |||
| // Copyright (c) 2022 Graphcore Ltd. All rights reserved. | |||
There was a problem hiding this comment.
If you like I can just merge the new files and leave everythign else unchaged?
| @@ -0,0 +1,202 @@ | |||
| // Copyright (c) 2022 Graphcore Ltd. All rights reserved. | |||
|
|
|||
There was a problem hiding this comment.
Needs Graphcore copyright notice (copy format from one of the other files).
2 participants
…ary as extension makes use of spatial variable separability.
Included also is bash script for performing sweeps for data analysis