Skip to content

Fix dilation>1 in Conv operator (GPU & CPU) #34

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
harz05 wants to merge 3 commits into
base: gpu/alpaka
Choose a base branch
from
+66 −8
Open

Fix dilation>1 in Conv operator (GPU & CPU) #34

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
9979d37
fix dilation>1 in conv for cpu and gpu + test
harz05 Jun 9, 2026
4b4c110
add cpu dilation test
harz05 Jun 9, 2026
ec97e1e
remove cpu dilation test
harz05 Jun 9, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
31 changes: 23 additions & 8 deletions core/inc/SOFIE/ROperator_Conv.hxx
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -430,6 +430,9 @@ public:
out << SP << SP << "}\n";
out << SP << "}\n";

// dilation already folded into the expanded kernel and dilated _f layout above
fAttrDilations = std::vector<size_t>(3, 1);

//out << SP << "char " << OpName << "_transA = 'T';\n";
out << SP << "char " << OpName << "_transA = 'N';\n";
out << SP << "char " << OpName << "_transB = 'N';\n";
Expand Down Expand Up @@ -634,6 +637,10 @@ public:
size_t ocstride = fShapeW[1] * icstride;
size_t wTotalElements = ConvertShapeToLength(fShapeW);

// effective (dilation-expanded) kernel extents, used for the dense im2col decode
size_t kHeightEff = (fDim > 1) ? fAttrKernelShape[ih] : 1;
size_t kWidthEff = fAttrKernelShape[iw];

std::string op;

// Kernel 1: Weight vectorisation — reorder W into _f with dilation layout
Expand Down Expand Up @@ -709,13 +716,13 @@ public:
op += SP + SP + SP + SP + "std::size_t const ic = col_row / " + std::to_string(kernelSize) + "u;\n";
op += SP + SP + SP + SP + "std::size_t const k_rem = col_row % " + std::to_string(kernelSize) + "u;\n";
if (fDim > 2) {
op += SP + SP + SP + SP + "std::size_t const kd = k_rem / " + std::to_string(kHeight * kWidth) + "u;\n";
op += SP + SP + SP + SP + "std::size_t const kh = (k_rem / " + std::to_string(kWidth) + "u) % " + std::to_string(kHeight) + "u;\n";
op += SP + SP + SP + SP + "std::size_t const kw = k_rem % " + std::to_string(kWidth) + "u;\n\n";
op += SP + SP + SP + SP + "std::size_t const kd = k_rem / " + std::to_string(kHeightEff * kWidthEff) + "u;\n";
op += SP + SP + SP + SP + "std::size_t const kh = (k_rem / " + std::to_string(kWidthEff) + "u) % " + std::to_string(kHeightEff) + "u;\n";
op += SP + SP + SP + SP + "std::size_t const kw = k_rem % " + std::to_string(kWidthEff) + "u;\n\n";
} else if (fDim > 1) {
op += SP + SP + SP + SP + "std::size_t const kd = 0u;\n";
op += SP + SP + SP + SP + "std::size_t const kh = k_rem / " + std::to_string(kWidth) + "u;\n";
op += SP + SP + SP + SP + "std::size_t const kw = k_rem % " + std::to_string(kWidth) + "u;\n\n";
op += SP + SP + SP + SP + "std::size_t const kh = k_rem / " + std::to_string(kWidthEff) + "u;\n";
op += SP + SP + SP + SP + "std::size_t const kw = k_rem % " + std::to_string(kWidthEff) + "u;\n\n";
} else {
op += SP + SP + SP + SP + "std::size_t const kd = 0u;\n";
op += SP + SP + SP + SP + "std::size_t const kh = 0u;\n";
Expand All @@ -740,7 +747,7 @@ public:
// applying it here would make id_in negative and zero the whole output.
if (fDim >= 3) {
op += SP + SP + SP + SP + "int64_t const id_in = static_cast<int64_t>(od * " + std::to_string(fAttrStrides[0])
+ "u + kd * " + std::to_string(fAttrDilations[0]) + "u) - " + std::to_string(fAttrPads[0]) + ";\n";
+ "u + kd) - " + std::to_string(fAttrPads[0]) + ";\n";
} else {
op += SP + SP + SP + SP + "int64_t const id_in = 0;\n";
}
Expand All @@ -750,7 +757,7 @@ public:
size_t const hIdx = (fDim > 2) ? 1 : 0;
if (fDim >= 2) {
op += SP + SP + SP + SP + "int64_t const ih_in = static_cast<int64_t>(oh * " + std::to_string(fAttrStrides[hIdx])
+ "u + kh * " + std::to_string(fAttrDilations[hIdx]) + "u) - " + std::to_string(fAttrPads[hIdx]) + ";\n";
+ "u + kh) - " + std::to_string(fAttrPads[hIdx]) + ";\n";
} else {
op += SP + SP + SP + SP + "int64_t const ih_in = 0;\n";
}
Expand All @@ -759,7 +766,7 @@ public:
{
size_t const wIdx = fDim - 1;
op += SP + SP + SP + SP + "int64_t const iw_in = static_cast<int64_t>(ow * " + std::to_string(fAttrStrides[wIdx])
+ "u + kw * " + std::to_string(fAttrDilations[wIdx]) + "u) - " + std::to_string(fAttrPads[wIdx]) + ";\n\n";
+ "u + kw) - " + std::to_string(fAttrPads[wIdx]) + ";\n\n";
}

op += SP + SP + SP + SP + "bool const in_bounds =\n";
Expand Down Expand Up @@ -846,6 +853,14 @@ public:
std::stringstream out;
out << "\n//------ CONV_GPU_ALPAKA\n";

// dilation>1 leaves gaps in the dilated _f layout; zero it so those slots stay 0
bool hasDilation = false;
for (size_t d = 0; d < fDim; ++d) if (fAttrDilations[d] > 1) hasDilation = true;
if (hasDilation) {
out << SP << "alpaka::memset(queue, deviceBuf_" << convK << ", 0);\n";
out << SP << "alpaka::wait(queue);\n";
}

// -----------------------------------------------------------------------
// Step 1: Weight vectorisation kernel — runs once, fully on GPU
// -----------------------------------------------------------------------
Expand Down
40 changes: 40 additions & 0 deletions core/test/TestCustomModelsFromONNXForAlpakaCuda.cxx
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -164,6 +164,9 @@
#include "ConvWithAsymmetricPadding_FromONNX_GPU_ALPAKA.hxx"
#include "input_models/references/ConvWithAsymmetricPadding.ref.hxx"

#include "ConvWithDilation_FromONNX_GPU_ALPAKA.hxx"
#include "input_models/references/ConvWithDilation.ref.hxx"

#include "BatchNorm_FromONNX_GPU_ALPAKA.hxx"
#include "BatchNormRelu_FromONNX_GPU_ALPAKA.hxx"

Expand Down Expand Up @@ -2442,6 +2445,43 @@ TEST_F(SofieAlpakaTest, ConvWithAsymmetricPadding)
}
}

// Exercises the Conv GPU dilation>1 path (issue #32). x[1,1,7,7], W[2,1,3,3] iota, 3x3
// kernel, dilation 2, no padding. Before the fix the dilation was double-counted.
TEST_F(SofieAlpakaTest, ConvWithDilation)
{
constexpr float TOLERANCE = DEFAULT_TOLERANCE;

std::vector<float> input(49);
std::iota(input.begin(), input.end(), 0.0f);

auto input_h = alpaka::allocBuf<float, Idx>(host, Ext1D::all(Idx{input.size()}));
float* input_ptr = reinterpret_cast<float*>(alpaka::getPtrNative(input_h));
for (Idx i = 0; i < input.size(); ++i) input_ptr[i] = input[i];

auto input_d = alpaka::allocBuf<float, Idx>(device, Ext1D::all(Idx{input.size()}));
alpaka::memcpy(queue, input_d, input_h);
alpaka::wait(queue);

auto result_h = alpaka::allocBuf<float, Idx>(host, Ext1D::all(Idx{sizeof(ConvWithDilation_ExpectedOutput::all_ones) / sizeof(float)}));

{
SOFIE_ConvWithDilation::Session<alpaka::TagGpuCudaRt> session("ConvWithDilation_FromONNX_GPU_ALPAKA.dat");
auto result = session.infer(input_d);
alpaka::wait(queue);
cudaDeviceSynchronize();
alpaka::memcpy(queue, result_h, result);
alpaka::wait(queue);
}

float* res_ptr = reinterpret_cast<float*>(alpaka::getPtrNative(result_h));
float *correct = ConvWithDilation_ExpectedOutput::all_ones;
constexpr size_t nOut_dilation = sizeof(ConvWithDilation_ExpectedOutput::all_ones) / sizeof(float);

for (size_t i = 0; i < nOut_dilation; ++i) {
EXPECT_LE(std::abs(res_ptr[i] - correct[i]), TOLERANCE) << "i=" << i;
}
}

TEST_F(SofieAlpakaTest, BatchNormalization)
{
constexpr float TOLERANCE = DEFAULT_TOLERANCE;
Expand Down
Binary file added core/test/input_models/ConvWithDilation.onnx
View file
Open in desktop
Binary file not shown.
3 changes: 3 additions & 0 deletions core/test/input_models/references/ConvWithDilation.ref.hxx
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,3 @@
namespace ConvWithDilation_ExpectedOutput {
float all_ones[] = {840.0, 876.0, 912.0, 1092.0, 1128.0, 1164.0, 1344.0, 1380.0, 1416.0, 2136.0, 2253.0, 2370.0, 2955.0, 3072.0, 3189.0, 3774.0, 3891.0, 4008.0};
} // namespace ConvWithDilation_ExpectedOutput