Skip to content
Snippets Groups Projects
Commit b8ab127c authored by Grégoire Kubler's avatar Grégoire Kubler
Browse files

feat : refactored conv1D forward to use only DimSize_t

parent 839fa0aa
No related branches found
No related tags found
No related merge requests found
This commit is part of merge request !142. Comments created here will be created in the context of that merge request.
Hide whitespace changes
Inline Side-by-side
include/aidge/backend/cpu/operator/ConvImpl.hpp
+ 10
− 10
View file @ b8ab127c
...@@ -26,16 +26,16 @@ ...@@ -26,16 +26,16 @@
namespace Aidge { namespace Aidge {
// Operator implementation entry point for the backend // Operator implementation entry point for the backend
using Conv1D_Op = Conv_Op<1>; using Conv1D_Op = Conv_Op<1>;
using ConvImpl1D_cpu = OperatorImpl_cpu<Conv_Op<1>, using ConvImpl1D_cpu = OperatorImpl_cpu<Conv1D_Op,
void(const std::array<DimSize_t, 1>&, void(const DimSize_t &,
const std::array<DimSize_t, 1>&, const DimSize_t &,
const std::array<DimSize_t, 1>&, const DimSize_t &,
const std::array<DimSize_t, 3> &, const std::array<DimSize_t, 3> &,
DimSize_t, DimSize_t,
const void *, const void *,
const void *, const void *,
const void *, const void *,
void *)>; void *),
using Conv2D_Op = Conv_Op<2>; using Conv2D_Op = Conv_Op<2>;
using ConvImpl2D_cpu = OperatorImpl_cpu<Conv_Op<2>, using ConvImpl2D_cpu = OperatorImpl_cpu<Conv_Op<2>,
......
include/aidge/backend/cpu/operator/ConvImpl_kernels.hpp
+ 43
− 27
View file @ b8ab127c
...@@ -39,16 +39,15 @@ namespace Aidge { ...@@ -39,16 +39,15 @@ namespace Aidge {
* @param output_ Output Tensor. * @param output_ Output Tensor.
*/ */
template <class I, class W, class B, class O> template <class I, class W, class B, class O>
void ConvImpl1D_cpu_forward_kernel(const std::array<DimSize_t, 1>& strideDims, void ConvImpl1D_cpu_forward_kernel(const DimSize_t &strideDim,
const std::array<DimSize_t, 1>& dilationDims, const DimSize_t &dilationDim,
const std::array<DimSize_t, 1>& kernelDims, const DimSize_t &kernelDim,
const std::array<DimSize_t, 3>& inputDims, const std::array<DimSize_t, 3> &inputDims,
DimSize_t outChannels, DimSize_t outChannels,
const void *input_, const void *input_,
const void *weights_, const void *weights_,
const void *biases_, const void *biases_,
void *output_) void *output_) {
{
// FIXME: missing convolution attributes as arguments // FIXME: missing convolution attributes as arguments
const I *input = static_cast<const I *>(input_); const I *input = static_cast<const I *>(input_);
const W *weights = static_cast<const W *>(weights_); const W *weights = static_cast<const W *>(weights_);
...@@ -56,34 +55,51 @@ void ConvImpl1D_cpu_forward_kernel(const std::array<DimSize_t, 1>& strideDims, ...@@ -56,34 +55,51 @@ void ConvImpl1D_cpu_forward_kernel(const std::array<DimSize_t, 1>& strideDims,
O *output = static_cast<O *>(output_); O *output = static_cast<O *>(output_);
// output H size // output H size
const std::size_t oxSize = const std::size_t oxSize = static_cast<std::size_t>(std::floor(
static_cast<std::size_t>(std::floor(static_cast<float>(inputDims[2] - dilationDims[0]*(kernelDims[0] - 1) - 1 + strideDims[0]) / static_cast<float>(inputDims[2] - dilationDim * (kernelDim - 1) - 1 +
static_cast<float>(strideDims[0]))); strideDim) /
const DimSize_t dilated_kernel_x = dilationDims[0]*(kernelDims[0] - 1) + 1; static_cast<float>(strideDim)));
const DimSize_t dilated_kernel_x = dilationDim * (kernelDim - 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; using signedsize = std::make_signed<std::size_t>::type;
for (std::size_t batch = 0; batch < inputDims[0]; ++batch) { for (std::size_t batch = 0; batch < inputDims[0]; ++batch) {
for (std::size_t outCh = 0; outCh < outChannels; ++outCh) { 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) // If bias = nullptr, set B(0)
B biasVal = (biases != nullptr) ? biases[outCh] : 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) { 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 iIndex =
const std::size_t wIndex = (inCh + outCh*inputDims[1]) * kernelDims[0]; (inCh + batch * inputDims[1]) * inputDims[2];
const std::size_t wIndex =
(inCh + outCh * inputDims[1]) * kernelDim;
for (std::size_t ox = 0; ox < oxSize; ++ox) { for (std::size_t ox = 0; ox < oxSize; ++ox) {
// const signedsize difx = static_cast<signedsize>(- ox * strideDims[0]); // const signedsize difx = static_cast<signedsize>(- ox *
// const std::size_t sxMin = static_cast<std::size_t>(std::max(difx, signedsize(0))); // strideDims); const std::size_t sxMin =
// const std::size_t sxMax = (static_cast<signedsize>(inputDims[2]) + difx) < 0 ? 0 : ((inputDims[2] + difx) > kernelDims[0] ? kernelDims[0] : inputDims[2] + difx); // 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 :
// inputDims[2] + difx);
const std::size_t sxMin = 0; const std::size_t sxMin = 0;
const std::size_t sxMax = dilated_kernel_x; const std::size_t sxMax = dilated_kernel_x;
const std::size_t oIndexFull = oIndex + ox; const std::size_t oIndexFull = oIndex + ox;
const signedsize ix = static_cast<signedsize>(ox * strideDims[0]); const signedsize ix =
static_cast<signedsize>(ox * strideDim);
for (std::size_t sx = sxMin; sx * dilationDim < sxMax;
++sx) {
output[oIndexFull] +=
weights[wIndex + sx] *
input[iIndex + static_cast<std::size_t>(
ix + static_cast<signedsize>(
sx * dilationDim))];
}
}
}
}
}
}
for (std::size_t sx = sxMin; sx*dilationDims[0] < sxMax; ++sx) { for (std::size_t sx = sxMin; sx*dilationDims[0] < sxMax; ++sx) {
output[oIndexFull] += weights[wIndex + sx] * output[oIndexFull] += weights[wIndex + sx] *
......
src/operator/ConvImpl.cpp
+ 11
− 10
View file @ b8ab127c
...@@ -43,16 +43,17 @@ void Aidge::ConvImpl1D_cpu::forward() { ...@@ -43,16 +43,17 @@ void Aidge::ConvImpl1D_cpu::forward() {
const auto& input2 = (op_.getInput(2)) ? op_.getInput(2)->refCastFrom(input2Fallback, *op_.getOutput(0)) : Tensor(); const auto& input2 = (op_.getInput(2)) ? op_.getInput(2)->refCastFrom(input2Fallback, *op_.getOutput(0)) : Tensor();
// Call kernel // Call kernel
impl.forward(op_.strideDims(), impl.forward(
op_.dilationDims(), op_.strideDims()[0],
op_.kernelDims(), op_.dilationDims()[0],
op_.getInput(0)->template dims<3>(), // input dimensions op_.kernelDims()[0],
dynamic_cast<const Conv_Op<1>&>(mOp).outChannels(), // outChannels op_.getInput(0)->template dims<3>(), // input dimensions
input0.getImpl()->rawPtr(), // input dynamic_cast<const Conv_Op<1> &>(mOp).outChannels(), // outChannels
input1.getImpl()->rawPtr(), // weight input0.getImpl()->rawPtr(), // input
op_.getInput(2) ? input2.getImpl()->rawPtr() : nullptr, // bias input1.getImpl()->rawPtr(), // weight
getCPUPtr(mOp.getRawOutput(0)) // output op_.getInput(2) ? input2.getImpl()->rawPtr() : nullptr, // bias
); getCPUPtr(mOp.getRawOutput(0)) // output
);
} }
template <> template <>
......
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment

Copyright © Eclipse Foundation, Inc. All Rights Reserved.     Privacy Policy | Terms of Use | Copyright Agent