From 763e18886c06be0e55d2c2618cf16e6fe544cb13 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Gr=C3=A9goire=20KUBLER?= <gregoire.kubler@proton.me>
Date: Thu, 17 Apr 2025 10:48:41 +0100
Subject: [PATCH] chore : alternative implementation to reduce boilerplate code
sadly this also lengthen exec time as functions cannot be properly inlined when called through std::function.
See https://gitlab.eclipse.org/eclipse/aidge/aidge_backend_cpu/-/merge_requests/160#note_3653885
for more information.
---
.../backend/cpu/operator/PadImpl_kernels.hpp | 321 +++++++-----------
1 file changed, 114 insertions(+), 207 deletions(-)
diff --git a/include/aidge/backend/cpu/operator/PadImpl_kernels.hpp b/include/aidge/backend/cpu/operator/PadImpl_kernels.hpp
index 762d41fa..dfbd8edb 100644
--- a/include/aidge/backend/cpu/operator/PadImpl_kernels.hpp
+++ b/include/aidge/backend/cpu/operator/PadImpl_kernels.hpp
@@ -20,8 +20,8 @@
#include <cstddef> // std::size_t
#include <cstdint> // std::int32_t
#include <fmt/base.h>
+#include <functional>
#include <stdexcept>
-#include <type_traits>
#include "aidge/backend/cpu/operator/PadImpl.hpp"
#include "aidge/utils/Registrar.hpp"
@@ -216,39 +216,51 @@ pad3DForwardConstant(const std::array<DimSize_t, 6> &beginEndBorders,
std::array<DimSize_t, 4> &oOffsets,
O *output) {
- for (DimSize_t oX = 0; oX < oDims[0]; ++oX) {
- oOffsets[2] = oX * oStrides[2] + oOffsets[1];
-
- const SignedDimSize_t iX = oX - beginEndBorders[0];
- if (iX >= 0 && iX < static_cast<SignedDimSize_t>(iDims[2])) {
- iOffsets[2] = iX * iStrides[2] + iOffsets[1];
- } else {
- std::fill(output + oOffsets[2],
- output + oOffsets[2] + oStrides[2],
- borderValue);
- continue;
- }
+ for (std::size_t batch = 0; batch < iDims[0]; ++batch) {
+ oOffsets[0] = batch * oStrides[0];
+ iOffsets[0] = batch * iStrides[0];
- for (DimSize_t oY = 0; oY < oDims[1]; ++oY) {
- oOffsets[3] = oY * oStrides[3] + oOffsets[2];
- const SignedDimSize_t iY = oY - beginEndBorders[1];
- if (iY >= 0 && iY < static_cast<SignedDimSize_t>(iDims[3])) {
- iOffsets[3] = iY * iStrides[3] + iOffsets[2];
- } else {
- std::fill(output + oOffsets[3],
- output + oOffsets[3] + oStrides[3],
- borderValue);
- continue;
- }
+ for (std::size_t ch = 0; ch < iDims[1]; ++ch) {
+ iOffsets[1] = ch * iStrides[1] + iOffsets[0];
+ oOffsets[1] = ch * oStrides[1] + oOffsets[0];
+
+ for (DimSize_t oX = 0; oX < oDims[0]; ++oX) {
+ oOffsets[2] = oX * oStrides[2] + oOffsets[1];
+
+ const SignedDimSize_t iX = oX - beginEndBorders[0];
+ if (iX >= 0 && iX < static_cast<SignedDimSize_t>(iDims[2])) {
+ iOffsets[2] = iX * iStrides[2] + iOffsets[1];
+ } else {
+ std::fill(output + oOffsets[2],
+ output + oOffsets[2] + oStrides[2],
+ borderValue);
+ continue;
+ }
- for (DimSize_t oZ = 0; oZ < oDims[2]; ++oZ) {
- const SignedDimSize_t iZ = oZ - beginEndBorders[2];
- // if in bounds, takes corresponding value, otherwise takes
- // default value
- output[oOffsets[3] + oZ] =
- (iZ >= 0 && iZ < static_cast<SignedDimSize_t>(iDims[4]))
- ? input[iOffsets[3] + iZ]
- : borderValue;
+ for (DimSize_t oY = 0; oY < oDims[1]; ++oY) {
+ oOffsets[3] = oY * oStrides[3] + oOffsets[2];
+ const SignedDimSize_t iY = oY - beginEndBorders[1];
+ if (iY >= 0 &&
+ iY < static_cast<SignedDimSize_t>(iDims[3])) {
+ iOffsets[3] = iY * iStrides[3] + iOffsets[2];
+ } else {
+ std::fill(output + oOffsets[3],
+ output + oOffsets[3] + oStrides[3],
+ borderValue);
+ continue;
+ }
+
+ for (DimSize_t oZ = 0; oZ < oDims[2]; ++oZ) {
+ const SignedDimSize_t iZ = oZ - beginEndBorders[2];
+ // if in bounds, takes corresponding value, otherwise
+ // takes default value
+ output[oOffsets[3] + oZ] =
+ (iZ >= 0 &&
+ iZ < static_cast<SignedDimSize_t>(iDims[4]))
+ ? input[iOffsets[3] + iZ]
+ : borderValue;
+ }
+ }
}
}
}
@@ -270,39 +282,6 @@ static inline DimSize_t padEdgeComputeInputCoord(const DimSize_t beginBorder,
static_cast<SignedDimSize_t>(oCoord - beginBorder))));
}
-template <typename I, typename O>
-static inline void
-pad3DForwardEdge(const std::array<DimSize_t, 6> &beginEndBorders,
- const std::array<DimSize_t, 5> &iDims,
- const std::array<DimSize_t, 4> &iStrides,
- std::array<DimSize_t, 4> &iOffsets,
- const I *input,
- const std::array<DimSize_t, 3> &oDims,
- const std::array<DimSize_t, 4> &oStrides,
- std::array<DimSize_t, 4> &oOffsets,
- O *output) {
- for (DimSize_t oX = 0; oX < oDims[0]; ++oX) {
- oOffsets[2] = oX * oStrides[2] + oOffsets[1];
- const DimSize_t iX =
- padEdgeComputeInputCoord(beginEndBorders[0], iDims[2], oX);
- iOffsets[2] = iX * iStrides[2] + iOffsets[1];
-
- for (DimSize_t oY = 0; oY < oDims[1]; ++oY) {
- oOffsets[3] = oY * oStrides[3] + oOffsets[2];
- const DimSize_t iY =
- padEdgeComputeInputCoord(beginEndBorders[1], iDims[3], oY);
- iOffsets[3] = iY * iStrides[3] + iOffsets[2];
-
- for (DimSize_t oZ = 0; oZ < oDims[2]; ++oZ) {
- const DimSize_t iZ =
- padEdgeComputeInputCoord(beginEndBorders[2], iDims[4], oZ);
-
- output[oOffsets[3] + oZ] = input[iOffsets[3] + iZ];
- }
- }
- }
-}
-
/**
* @brief small inline fctn to generate the corresponding input coordinates of
* an output coord in reflect padding along a given dimension.
@@ -316,10 +295,8 @@ padReflectComputeInputCoord(const DimSize_t beginBorder,
const DimSize_t oCoord) {
SignedDimSize_t iCoord =
std::abs(static_cast<SignedDimSize_t>(oCoord - beginBorder));
-
- // Handle case where iCoord > iDim
- // If so iCoord must be changed to (iDim - 1) - delta
- // With delta = |iDim - 1 - icoord|
+ // Handle case where iCoord > iDim If so iCoord must be changed to (iDim -
+ // 1) - delta With delta = |iDim - 1 - icoord|
//
// Since iCoord > iDim - 1, |(iDim - 1) - iCoord| <=> iCoord - (iDim - 1)
// <=> iCoord + 1 - iDim
@@ -332,40 +309,6 @@ padReflectComputeInputCoord(const DimSize_t beginBorder,
return iCoord;
}
-template <typename I, typename O>
-static inline void
-pad3DForwardReflect(const std::array<DimSize_t, 6> &beginEndBorders,
- const std::array<DimSize_t, 5> &iDims,
- const std::array<DimSize_t, 4> &iStrides,
- std::array<DimSize_t, 4> &iOffsets,
- const I *input,
- const std::array<DimSize_t, 3> &oDims,
- const std::array<DimSize_t, 4> &oStrides,
- std::array<DimSize_t, 4> &oOffsets,
- O *output) {
-
- for (DimSize_t oX = 0; oX < oDims[0]; ++oX) {
- oOffsets[2] = oX * oStrides[2] + oOffsets[1];
- DimSize_t iX =
- padReflectComputeInputCoord(beginEndBorders[0], iDims[2], oX);
- iOffsets[2] = iX * iStrides[2] + iOffsets[1];
-
- for (DimSize_t oY = 0; oY < oDims[1]; ++oY) {
- oOffsets[3] = oY * oStrides[3] + oOffsets[2];
- DimSize_t iY =
- padReflectComputeInputCoord(beginEndBorders[1], iDims[3], oY);
- iOffsets[3] = iY * iStrides[3] + iOffsets[2];
-
- for (DimSize_t oZ = 0; oZ < oDims[2]; ++oZ) {
- DimSize_t iZ = padReflectComputeInputCoord(beginEndBorders[2],
- iDims[4],
- oZ);
- output[oOffsets[3] + oZ] = input[iOffsets[3] + iZ];
- }
- }
- }
-}
-
/**
* @brief small inline fctn to generate the corresponding input coordinates of
* an output coord in wrap padding along a given dimension.
@@ -379,41 +322,6 @@ static inline DimSize_t padWrapComputeInputCoord(const DimSize_t beginBorder,
return (iDim + oCoord - beginBorder) % iDim;
}
-template <typename I, typename O>
-static inline void
-pad3DForwardWrap(const std::array<DimSize_t, 6> &beginEndBorders,
- const std::array<DimSize_t, 5> &iDims,
- const std::array<DimSize_t, 4> &iStrides,
- std::array<DimSize_t, 4> &iOffsets,
- const I *input,
- const std::array<DimSize_t, 3> &oDims,
- const std::array<DimSize_t, 4> &oStrides,
- std::array<DimSize_t, 4> &oOffsets,
- O *output) {
- for (DimSize_t oX = 0; oX < oDims[0]; ++oX) {
- oOffsets[2] = oX * oStrides[2] + oOffsets[1];
-
- const DimSize_t iX =
- padWrapComputeInputCoord(beginEndBorders[0], iDims[2], oX);
- iOffsets[2] = iX * iStrides[2] + iOffsets[1];
-
- for (DimSize_t oY = 0; oY < oDims[1]; ++oY) {
- oOffsets[3] = oY * oStrides[3] + oOffsets[2];
-
- const DimSize_t iY =
- padWrapComputeInputCoord(beginEndBorders[1], iDims[3], oY);
-
- iOffsets[3] = iY * iStrides[3] + iOffsets[2];
-
- for (DimSize_t oZ = 0; oZ < oDims[2]; ++oZ) {
- const DimSize_t iZ =
- padWrapComputeInputCoord(beginEndBorders[2], iDims[4], oZ);
- output[oOffsets[3] + oZ] = input[iOffsets[3] + iZ];
- }
- }
- }
-}
-
/**
* @brief Forward kernel for 2D Padding on CPU backend.
* @tparam I Input data type.
@@ -456,6 +364,55 @@ void PadImpl3D_cpu_forward_kernel(
std::array<DimSize_t, 4> oOffsets = {0, 0, 0, 0};
std::array<DimSize_t, 4> iOffsets = {0, 0, 0, 0};
+ std::function<DimSize_t(const DimSize_t, const DimSize_t, const DimSize_t)>
+ getInputCoord;
+
+ switch (borderType) {
+ case PadBorderType::Constant: {
+ pad3DForwardConstant(beginEndBorders,
+ static_cast<O>(borderValue),
+ iDims,
+ iStrides,
+ iOffsets,
+ input,
+ oDims,
+ oStrides,
+ oOffsets,
+ output);
+ return;
+ }
+ case PadBorderType::Zero: {
+ pad3DForwardConstant(beginEndBorders,
+ static_cast<O>(0),
+ iDims,
+ iStrides,
+ iOffsets,
+ input,
+ oDims,
+ oStrides,
+ oOffsets,
+ output);
+ return;
+ }
+ case PadBorderType::Edge: {
+ getInputCoord = padEdgeComputeInputCoord;
+ break;
+ }
+ case PadBorderType::Reflect: {
+ getInputCoord = padReflectComputeInputCoord;
+ break;
+ }
+ case PadBorderType::Wrap: {
+ getInputCoord = padWrapComputeInputCoord;
+ break;
+ }
+ default: {
+ AIDGE_THROW_OR_ABORT(std::runtime_error,
+ "Pad3D : unsupported padding method : {}.",
+ borderType);
+ }
+ }
+
for (std::size_t batch = 0; batch < iDims[0]; ++batch) {
oOffsets[0] = batch * oStrides[0];
iOffsets[0] = batch * iStrides[0];
@@ -464,75 +421,25 @@ void PadImpl3D_cpu_forward_kernel(
iOffsets[1] = ch * iStrides[1] + iOffsets[0];
oOffsets[1] = ch * oStrides[1] + oOffsets[0];
- switch (borderType) {
- case PadBorderType::Constant: {
- pad3DForwardConstant(beginEndBorders,
- static_cast<O>(borderValue),
- iDims,
- iStrides,
- iOffsets,
- input,
- oDims,
- oStrides,
- oOffsets,
- output);
- break;
- }
- case PadBorderType::Zero: {
- pad3DForwardConstant(beginEndBorders,
- static_cast<O>(0),
- iDims,
- iStrides,
- iOffsets,
- input,
- oDims,
- oStrides,
- oOffsets,
- output);
- break;
- }
- case PadBorderType::Edge: {
- pad3DForwardEdge(beginEndBorders,
- iDims,
- iStrides,
- iOffsets,
- input,
- oDims,
- oStrides,
- oOffsets,
- output);
- break;
- }
- case PadBorderType::Reflect: {
- pad3DForwardReflect(beginEndBorders,
- iDims,
- iStrides,
- iOffsets,
- input,
- oDims,
- oStrides,
- oOffsets,
- output);
- break;
- }
- case PadBorderType::Wrap: {
- pad3DForwardWrap(beginEndBorders,
- iDims,
- iStrides,
- iOffsets,
- input,
- oDims,
- oStrides,
- oOffsets,
- output);
- break;
- }
- default: {
- AIDGE_THROW_OR_ABORT(
- std::runtime_error,
- "Pad3D : unsupported padding method : {}.",
- borderType);
- }
+ for (DimSize_t oX = 0; oX < oDims[0]; ++oX) {
+ oOffsets[2] = oX * oStrides[2] + oOffsets[1];
+ const DimSize_t iX =
+ getInputCoord(beginEndBorders[0], iDims[2], oX);
+ iOffsets[2] = iX * iStrides[2] + iOffsets[1];
+
+ for (DimSize_t oY = 0; oY < oDims[1]; ++oY) {
+ oOffsets[3] = oY * oStrides[3] + oOffsets[2];
+ const DimSize_t iY =
+ getInputCoord(beginEndBorders[1], iDims[3], oY);
+ iOffsets[3] = iY * iStrides[3] + iOffsets[2];
+
+ for (DimSize_t oZ = 0; oZ < oDims[2]; ++oZ) {
+ const DimSize_t iZ =
+ getInputCoord(beginEndBorders[2], iDims[4], oZ);
+
+ output[oOffsets[3] + oZ] = input[iOffsets[3] + iZ];
+ }
+ }
}
}
}
--
GitLab