From 7bd9c550694f560301abe4bd04aaeb1c81fd8cea Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Gr=C3=A9goire=20KUBLER?= <gregoire.kubler@proton.me>
Date: Tue, 25 Feb 2025 13:31:12 +0000
Subject: [PATCH] chore : conv forward 1/2D formatting
---
.../aidge/backend/cpu/operator/ConvImpl.hpp | 1 +
.../backend/cpu/operator/ConvImpl_kernels.hpp | 249 ++++++++++++------
src/operator/ConvImpl.cpp | 21 +-
3 files changed, 177 insertions(+), 94 deletions(-)
diff --git a/include/aidge/backend/cpu/operator/ConvImpl.hpp b/include/aidge/backend/cpu/operator/ConvImpl.hpp
index 8bf11ac0..e480697b 100644
--- a/include/aidge/backend/cpu/operator/ConvImpl.hpp
+++ b/include/aidge/backend/cpu/operator/ConvImpl.hpp
@@ -20,6 +20,7 @@
#include "aidge/utils/Types.h"
namespace Aidge {
+
// Operator implementation entry point for the backend
using Conv1D_Op = Conv_Op<1>;
using ConvImpl1D_cpu = OperatorImpl_cpu<Conv_Op<1>,
diff --git a/include/aidge/backend/cpu/operator/ConvImpl_kernels.hpp b/include/aidge/backend/cpu/operator/ConvImpl_kernels.hpp
index 4e1861b4..274f5f4f 100644
--- a/include/aidge/backend/cpu/operator/ConvImpl_kernels.hpp
+++ b/include/aidge/backend/cpu/operator/ConvImpl_kernels.hpp
@@ -37,16 +37,15 @@ using std::array;
* @param output_ Output Tensor.
*/
template <class I, class W, class B, class O>
-void ConvImpl1D_cpu_forward_kernel(const std::array<DimSize_t, 1>& strideDims,
- const std::array<DimSize_t, 1>& dilationDims,
- const std::array<DimSize_t, 1>& kernelDims,
- const std::array<DimSize_t, 3>& inputDims,
- DimSize_t outChannels,
- const void *input_,
- const void *weights_,
- const void *biases_,
- void *output_)
-{
+void ConvImpl1D_cpu_forward_kernel(const array<DimSize_t, 1> &strideDim,
+ const array<DimSize_t, 1> &dilationDim,
+ const array<DimSize_t, 1> &kernelDim,
+ const std::array<DimSize_t, 3> &inputDims,
+ DimSize_t outChannels,
+ const void *input_,
+ const void *weights_,
+ const void *biases_,
+ void *output_) {
// FIXME: missing convolution attributes as arguments
const I *input = static_cast<const I *>(input_);
const W *weights = static_cast<const W *>(weights_);
@@ -54,34 +53,51 @@ void ConvImpl1D_cpu_forward_kernel(const std::array<DimSize_t, 1>& strideDims,
O *output = static_cast<O *>(output_);
// output H size
- const std::size_t oxSize =
- static_cast<std::size_t>(std::floor(static_cast<float>(inputDims[2] - dilationDims[0]*(kernelDims[0] - 1) - 1 + strideDims[0]) /
- static_cast<float>(strideDims[0])));
- const DimSize_t dilated_kernel_x = dilationDims[0]*(kernelDims[0] - 1) + 1;
+ const std::size_t oxSize = static_cast<std::size_t>(std::floor(
+ static_cast<float>(inputDims[2] - dilationDim[0] * (kernelDim[0] - 1) -
+ 1 + strideDim[0]) /
+ static_cast<float>(strideDim[0])));
+ const DimSize_t dilated_kernel_x = dilationDim[0] * (kernelDim[0] - 1) + 1;
- // TODO: kernel computation
- // output (batch, outCh, Xout, Yout)
- // input (batch, inCh, Xin, Yin)
- // weight (outCh, inCh, kernelX, kernelY)
- // does not take Dilation attribute into account
using signedsize = std::make_signed<std::size_t>::type;
for (std::size_t batch = 0; batch < inputDims[0]; ++batch) {
for (std::size_t outCh = 0; outCh < outChannels; ++outCh) {
- const std::size_t oIndex = (outCh + batch*outChannels) * oxSize;
+ const std::size_t oIndex = (outCh + batch * outChannels) * oxSize;
// If bias = nullptr, set B(0)
B biasVal = (biases != nullptr) ? biases[outCh] : B(0);
- std::fill(output + oIndex, output+(oIndex+oxSize), biasVal);
+ std::fill(output + oIndex, output + (oIndex + oxSize), biasVal);
for (std::size_t inCh = 0; inCh < inputDims[1]; ++inCh) {
- const std::size_t iIndex = (inCh + batch*inputDims[1]) * inputDims[2];
- const std::size_t wIndex = (inCh + outCh*inputDims[1]) * kernelDims[0];
+ const std::size_t iIndex =
+ (inCh + batch * inputDims[1]) * inputDims[2];
+ const std::size_t wIndex =
+ (inCh + outCh * inputDims[1]) * kernelDim[0];
for (std::size_t ox = 0; ox < oxSize; ++ox) {
- // const signedsize difx = static_cast<signedsize>(- ox * strideDims[0]);
- // const std::size_t sxMin = static_cast<std::size_t>(std::max(difx, signedsize(0)));
- // const std::size_t sxMax = (static_cast<signedsize>(inputDims[2]) + difx) < 0 ? 0 : ((inputDims[2] + difx) > kernelDims[0] ? kernelDims[0] : inputDims[2] + difx);
+ // const signedsize difx = static_cast<signedsize>(- ox *
+ // strideDim[0s); const std::size_t sxMin =
+ // static_cast<std::size_t>(std::max(difx, signedsize(0)));
+ // const std::size_t sxMax =
+ // (static_cast<signedsize>(inputDims[2]) + difx) < 0 ? 0 :
+ // ((inputDims[2] + difx) > kernelDim[0s[0] ? kernelDim[0s
+ // : inputDims[2] + difx);
const std::size_t sxMin = 0;
const std::size_t sxMax = dilated_kernel_x;
const std::size_t oIndexFull = oIndex + ox;
- const signedsize ix = static_cast<signedsize>(ox * strideDims[0]);
+ const signedsize ix =
+ static_cast<signedsize>(ox * strideDim[0]);
+
+ for (std::size_t sx = sxMin; sx * dilationDim[0] < sxMax;
+ ++sx) {
+ output[oIndexFull] +=
+ weights[wIndex + sx] *
+ input[iIndex + static_cast<std::size_t>(
+ ix + static_cast<signedsize>(
+ sx * dilationDim[0]))];
+ }
+ }
+ }
+ }
+ }
+}
/**
* @brief perform 1D backpropagation for the data input
@@ -119,9 +135,9 @@ void ConvImpl1D_cpu_forward_kernel(const std::array<DimSize_t, 1>& strideDims,
* @param[inout] iGrad gradients of the input to update
*/
template <class I, class W, class O>
-void conv1DBackwardInput(const DimSize_t &stride,
- const DimSize_t &dilation,
- const DimSize_t &kDim,
+void conv1DBackwardInput(const array<DimSize_t, 1> &stride,
+ const array<DimSize_t, 1> &dilation,
+ const array<DimSize_t, 1> &kDim,
const array<DimSize_t, 2> &kStrides,
const W *weights,
const array<DimSize_t, 3> &oDims,
@@ -434,16 +450,15 @@ REGISTRAR(ConvImpl1D_cpu,
* @param output_ Output Tensor.
*/
template <class I, class W, class B, class O>
-void ConvImpl2D_cpu_forward_kernel(const std::array<DimSize_t, 2>& strideDims,
- const std::array<DimSize_t, 2>& dilationDims,
- const std::array<DimSize_t, 2>& kernelDims,
- const std::array<DimSize_t, 4> &inputDims,
- DimSize_t outChannels,
- const void *input_,
- const void *weights_,
- const void *biases_,
- void *output_)
-{
+void ConvImpl2D_cpu_forward_kernel(const array<DimSize_t, 2> &strideDims,
+ const array<DimSize_t, 2> &dilationDims,
+ const array<DimSize_t, 2> &kernelDims,
+ const array<DimSize_t, 4> &inputDims,
+ DimSize_t outChannels,
+ const void *input_,
+ const void *weights_,
+ const void *biases_,
+ void *output_) {
// FIXME: missing convolution attributes as arguments
const I *input = static_cast<const I *>(input_);
const W *weights = static_cast<const W *>(weights_);
@@ -451,59 +466,102 @@ void ConvImpl2D_cpu_forward_kernel(const std::array<DimSize_t, 2>& strideDims,
O *output = static_cast<O *>(output_);
// output H size
- const DimSize_t dilated_kernel_x = dilationDims[0]*(kernelDims[0] - 1) + 1;
- const std::size_t oxSize =
- static_cast<std::size_t>(std::floor(static_cast<float>(inputDims[2] - dilated_kernel_x + strideDims[0]) /
- static_cast<float>(strideDims[0])));
+ const DimSize_t dilated_kernel_x =
+ dilationDims[0] * (kernelDims[0] - 1) + 1;
+ const std::size_t oxSize = static_cast<std::size_t>(std::floor(
+ static_cast<float>(inputDims[2] - dilated_kernel_x + strideDims[0]) /
+ static_cast<float>(strideDims[0])));
// output W size
- const DimSize_t dilated_kernel_y = dilationDims[1]*(kernelDims[1] - 1) + 1;
- const std::size_t oySize =
- static_cast<std::size_t>(std::floor(static_cast<float>(inputDims[3] - dilated_kernel_y + strideDims[1]) /
- static_cast<float>(strideDims[1])));
-
+ const DimSize_t dilated_kernel_y =
+ dilationDims[1] * (kernelDims[1] - 1) + 1;
+ const std::size_t oySize = static_cast<std::size_t>(std::floor(
+ static_cast<float>(inputDims[3] - dilated_kernel_y + strideDims[1]) /
+ static_cast<float>(strideDims[1])));
// TODO: kernel computation
// output (batch, outCh, Xout, Yout)
// input (batch, inCh, Xin, Yin)
// weight (outCh, inCh, kernelX, kernelY)
// does not take Dilation attribute into account
- const std::size_t outChannels_s = oxSize * oySize;
+ const std::size_t outChannels_s = oxSize * oySize;
if (dilated_kernel_x == 3 && dilated_kernel_y == 3) {
for (std::size_t batch = 0; batch < inputDims[0]; ++batch) {
for (std::size_t outCh = 0; outCh < outChannels; ++outCh) {
// If bias = nullptr, set B(0)
B biasVal = (biases != nullptr) ? biases[outCh] : B(0);
- std::fill(output, output+outChannels_s, biasVal);
+ std::fill(output, output + outChannels_s, biasVal);
for (std::size_t inCh = 0; inCh < inputDims[1]; ++inCh) {
- std::size_t iIndex = (inCh + batch*inputDims[1]) * inputDims[2] * inputDims[3];
- const std::size_t wIndex = (inCh + outCh*inputDims[1]) * 9;
- if (strideDims[0] == 1 && strideDims[1]==1) {
- for (std::size_t ox = 0, oIndex = 0; ox < oxSize; ++ox, oIndex+=oySize, iIndex-=inputDims[3]) {
+ std::size_t iIndex = (inCh + batch * inputDims[1]) *
+ inputDims[2] * inputDims[3];
+ const std::size_t wIndex =
+ (inCh + outCh * inputDims[1]) * 9;
+ if (strideDims[0] == 1 && strideDims[1] == 1) {
+ for (std::size_t ox = 0, oIndex = 0; ox < oxSize;
+ ++ox, oIndex += oySize, iIndex -= inputDims[3]) {
for (std::size_t oy = 0; oy < oySize; ++oy) {
- output[oIndex + oy] += weights[wIndex+0]*input[iIndex+oy]+weights[wIndex+1]*input[iIndex+oy+1]+weights[wIndex+2]*input[iIndex+oy+2];
+ output[oIndex + oy] +=
+ weights[wIndex + 0] * input[iIndex + oy] +
+ weights[wIndex + 1] *
+ input[iIndex + oy + 1] +
+ weights[wIndex + 2] *
+ input[iIndex + oy + 2];
}
- iIndex+=inputDims[3];
+ iIndex += inputDims[3];
for (std::size_t oy = 0; oy < oySize; ++oy) {
- output[oIndex + oy] += weights[wIndex+3]*input[iIndex+oy]+weights[wIndex+4]*input[iIndex+oy+1]+weights[wIndex+5]*input[iIndex+oy+2];
+ output[oIndex + oy] +=
+ weights[wIndex + 3] * input[iIndex + oy] +
+ weights[wIndex + 4] *
+ input[iIndex + oy + 1] +
+ weights[wIndex + 5] *
+ input[iIndex + oy + 2];
}
- iIndex+=inputDims[3];
+ iIndex += inputDims[3];
for (std::size_t oy = 0; oy < oySize; ++oy) {
- output[oIndex + oy] += weights[wIndex+6]*input[iIndex+oy]+weights[wIndex+7]*input[iIndex+oy+1]+weights[wIndex+8]*input[iIndex+oy+2];
+ output[oIndex + oy] +=
+ weights[wIndex + 6] * input[iIndex + oy] +
+ weights[wIndex + 7] *
+ input[iIndex + oy + 1] +
+ weights[wIndex + 8] *
+ input[iIndex + oy + 2];
}
}
} else {
- for (std::size_t ox = 0, oIndex = 0; ox < oxSize; ++ox, oIndex+=oySize, iIndex+=(strideDims[0]-2)*inputDims[3]) {
+ for (std::size_t ox = 0, oIndex = 0; ox < oxSize; ++ox,
+ oIndex += oySize,
+ iIndex += (strideDims[0] -
+ 2) * inputDims[3]) {
for (std::size_t oy = 0; oy < oySize; ++oy) {
- output[oIndex + oy] += weights[wIndex+0]*input[iIndex+oy*strideDims[1]]+weights[wIndex+1]*input[iIndex+oy*strideDims[1]+1]+weights[wIndex+2]*input[iIndex+oy*strideDims[1]+2];
+ output[oIndex + oy] +=
+ weights[wIndex + 0] *
+ input[iIndex + oy * strideDims[1]] +
+ weights[wIndex + 1] *
+ input[iIndex + oy * strideDims[1] +
+ 1] +
+ weights[wIndex + 2] *
+ input[iIndex + oy * strideDims[1] + 2];
}
- iIndex+=inputDims[3];
+ iIndex += inputDims[3];
for (std::size_t oy = 0; oy < oySize; ++oy) {
- output[oIndex + oy] += weights[wIndex+3]*input[iIndex+oy*strideDims[1]]+weights[wIndex+4]*input[iIndex+oy*strideDims[1]+1]+weights[wIndex+5]*input[iIndex+oy*strideDims[1]+2];
+ output[oIndex + oy] +=
+ weights[wIndex + 3] *
+ input[iIndex + oy * strideDims[1]] +
+ weights[wIndex + 4] *
+ input[iIndex + oy * strideDims[1] +
+ 1] +
+ weights[wIndex + 5] *
+ input[iIndex + oy * strideDims[1] + 2];
}
- iIndex+=inputDims[3];
+ iIndex += inputDims[3];
for (std::size_t oy = 0; oy < oySize; ++oy) {
- output[oIndex + oy] += weights[wIndex+6]*input[iIndex+oy*strideDims[1]]+weights[wIndex+7]*input[iIndex+oy*strideDims[1]+1]+weights[wIndex+8]*input[iIndex+oy*strideDims[1]+2];
+ output[oIndex + oy] +=
+ weights[wIndex + 6] *
+ input[iIndex + oy * strideDims[1]] +
+ weights[wIndex + 7] *
+ input[iIndex + oy * strideDims[1] +
+ 1] +
+ weights[wIndex + 8] *
+ input[iIndex + oy * strideDims[1] + 2];
}
}
}
@@ -516,18 +574,26 @@ void ConvImpl2D_cpu_forward_kernel(const std::array<DimSize_t, 2>& strideDims,
for (std::size_t outCh = 0; outCh < outChannels; ++outCh) {
// If bias = nullptr, set B(0)
B biasVal = (biases != nullptr) ? biases[outCh] : B(0);
- std::fill(output, output+outChannels_s, biasVal);
+ std::fill(output, output + outChannels_s, biasVal);
for (std::size_t inCh = 0; inCh < inputDims[1]; ++inCh) {
- std::size_t iIndex = (inCh + batch*inputDims[1]) * inputDims[2] * inputDims[3];
- const std::size_t wIndex = (inCh + outCh*inputDims[1]);
+ std::size_t iIndex = (inCh + batch * inputDims[1]) *
+ inputDims[2] * inputDims[3];
+ const std::size_t wIndex = (inCh + outCh * inputDims[1]);
if (strideDims[0] == 1 && strideDims[1] == 1) {
- for (std::size_t oIndex = 0; oIndex < oxSize*oySize; ++oIndex, ++iIndex) {
+ for (std::size_t oIndex = 0; oIndex < oxSize * oySize;
+ ++oIndex, ++iIndex) {
output[oIndex] += weights[wIndex] * input[iIndex];
}
- } else {
- for (std::size_t ox = 0, oIndex = 0; ox < oxSize; ++ox, oIndex+=oySize, iIndex+=inputDims[3]*strideDims[0]) {
- for (std::size_t oy = 0, iy = 0; oy < oySize; ++oy, iy+=strideDims[1]) {
- output[oIndex + oy] += weights[wIndex+0]*input[iIndex+iy];
+ } else {
+ for (std::size_t ox = 0, oIndex = 0; ox < oxSize;
+ ++ox,
+ oIndex += oySize,
+ iIndex +=
+ inputDims[3] * strideDims[0]) {
+ for (std::size_t oy = 0, iy = 0; oy < oySize;
+ ++oy, iy += strideDims[1]) {
+ output[oIndex + oy] +=
+ weights[wIndex + 0] * input[iIndex + iy];
}
}
}
@@ -540,21 +606,36 @@ void ConvImpl2D_cpu_forward_kernel(const std::array<DimSize_t, 2>& strideDims,
for (std::size_t outCh = 0; outCh < outChannels; ++outCh) {
// If bias = nullptr, set B(0)
B biasVal = (biases != nullptr) ? biases[outCh] : B(0);
- std::fill(output, output+outChannels_s, biasVal);
+ std::fill(output, output + outChannels_s, biasVal);
for (std::size_t inCh = 0; inCh < inputDims[1]; ++inCh) {
- std::size_t iIndex_channel = (inCh + batch*inputDims[1]) * inputDims[2] * inputDims[3];
- const std::size_t wIndex = (inCh + outCh*inputDims[1]) * kernelDims[0] * kernelDims[1];
+ std::size_t iIndex_channel =
+ (inCh + batch * inputDims[1]) * inputDims[2] *
+ inputDims[3];
+ const std::size_t wIndex = (inCh + outCh * inputDims[1]) *
+ kernelDims[0] * kernelDims[1];
// loop over each ouput line
- for (std::size_t ox = 0, oIndex = 0; ox < oxSize; ++ox, oIndex+=oySize, iIndex_channel+=inputDims[3]*strideDims[0]) {
+ for (std::size_t ox = 0, oIndex = 0; ox < oxSize;
+ ++ox,
+ oIndex += oySize,
+ iIndex_channel +=
+ inputDims[3] * strideDims[0]) {
// loop over associated input line
- for (std::size_t ky = 0, ix = 0; ky < kernelDims[0]; ++ky, ix += inputDims[3]*dilationDims[0]) {
+ for (std::size_t ky = 0, ix = 0; ky < kernelDims[0];
+ ++ky, ix += inputDims[3] * dilationDims[0]) {
// loop over the entire line
- for (std::size_t oy = 0, iy = 0; oy < oySize; ++oy, iy+=strideDims[1]) {
- const std::size_t iIndex = iIndex_channel + ix + iy;
- // loop over elements assosicated with one output
- for (std::size_t kx = 0; kx < kernelDims[0]; ++kx) {
- output[oIndex + oy] += weights[wIndex+kernelDims[0]*ky+kx]*input[iIndex+kx*dilationDims[1]];
+ for (std::size_t oy = 0, iy = 0; oy < oySize;
+ ++oy, iy += strideDims[1]) {
+ const std::size_t iIndex =
+ iIndex_channel + ix + iy;
+ // loop over elements assosicated with one
+ // output
+ for (std::size_t kx = 0; kx < kernelDims[0];
+ ++kx) {
+ output[oIndex + oy] +=
+ weights[wIndex + kernelDims[0] * ky +
+ kx] *
+ input[iIndex + kx * dilationDims[1]];
}
}
}
diff --git a/src/operator/ConvImpl.cpp b/src/operator/ConvImpl.cpp
index ffbb75c1..d23a9968 100644
--- a/src/operator/ConvImpl.cpp
+++ b/src/operator/ConvImpl.cpp
@@ -40,16 +40,17 @@ void Aidge::ConvImpl1D_cpu::forward() {
const auto& input2 = (op_.getInput(2)) ? op_.getInput(2)->refCastFrom(input2Fallback, *op_.getOutput(0)) : Tensor();
// Call kernel
- impl.forward(op_.strideDims(),
- op_.dilationDims(),
- op_.kernelDims(),
- op_.getInput(0)->template dims<3>(), // input dimensions
- dynamic_cast<const Conv_Op<1>&>(mOp).outChannels(), // outChannels
- input0.getImpl()->rawPtr(), // input
- input1.getImpl()->rawPtr(), // weight
- op_.getInput(2) ? input2.getImpl()->rawPtr() : nullptr, // bias
- getCPUPtr(mOp.getRawOutput(0)) // output
- );
+ impl.forward(
+ op_.strideDims(),
+ op_.dilationDims(),
+ op_.kernelDims(),
+ op_.getInput(0)->template dims<3>(), // input dimensions
+ dynamic_cast<const Conv_Op<1> &>(mOp).outChannels(), // outChannels
+ input0.getImpl()->rawPtr(), // input
+ input1.getImpl()->rawPtr(), // weight
+ op_.getInput(2) ? input2.getImpl()->rawPtr() : nullptr, // bias
+ getCPUPtr(mOp.getRawOutput(0)) // output
+ );
}
template <> void ConvImpl1D_cpu::backward() {
--
GitLab