derek/TheRock
1
0 Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity
Files
1a182e2dac58879f215c0fad0f5f78c76dcafe56
TheRock/rocm-libraries/projects/hipdnn/backend/tests/descriptors/TestGraphDescriptorOps.cpp
T
derek.holloway 7502bf64a4
CodeQL / Analyze (actions) (push) Canceled after 0s
Details
CodeQL / Analyze (python) (push) Canceled after 0s
Details
pre-commit / pre-commit (push) Canceled after 0s
Details
Multi-Arch CI / setup (push) Canceled after 0s
Details
Publish build_manylinux_x86_64 images / publish_build_manylinux_x86_64 (push) Canceled after 0s
Details
Setup (Multi-Arch) / setup (push) Canceled after 0s
Details
Multi-Arch CI / Linux::${{ fromJSON(needs.setup.outputs.linux_build_config || '{}').build_variant_label || 'skip' }} (push) Canceled after 0s
Details
Publish TheRock Docker image / build-and-push-image (push) Canceled after 0s
Details
Publish no_rocm_image_ubi10 image / publish_no_rocm_image_ubi10 (push) Canceled after 0s
Details
Multi-Arch CI / Windows::${{ fromJSON(needs.setup.outputs.windows_build_config || '{}').build_variant_label || 'skip' }} (push) Canceled after 0s
Details
Publish TheRock Docker image / build-and-push-image (push) Canceled after 0s
Details
Multi-Arch CI / Manifest Diff (push) Canceled after 0s
Details
TheRock Manifest Diff Report / Generate Manifest Diff Report (push) Canceled after 0s
Details
Multi-Arch CI / CI Summary (push) Canceled after 0s
Details
Publish no_rocm_image_ubuntu24_04 images / publish_no_rocm_image_ubuntu24_04 (push) Canceled after 0s
Details
Publish no_rocm_image_ubuntu24_04_ocl_rt images / publish_no_rocm_image_ubuntu24_04_ocl_rt (push) Canceled after 0s
Details
Publish TheRock Docker image / build-and-push-image (push) Canceled after 0s
Details
Publish TheRock Docker image / build-and-push-image (push) Canceled after 0s
Details
Publish no_rocm_image_ubuntu24_04 images / publish_no_rocm_image_ubuntu24_04_rocgdb (push) Canceled after 0s
Details
Publish no_rocm_image_ubuntu24_04_rocgdb images / publish_no_rocm_image_ubuntu24_04_rocgdb (push) Canceled after 0s
Details
Publish no_rocm_image_ubuntu24_04 images / publish_no_rocm_image_ubuntu24_04_ocl_rt (push) Canceled after 0s
Details
Publish TheRock Docker image / build-and-push-image (push) Canceled after 0s
Details
Unit Tests / Unit Tests :: ubuntu-24.04 (push) Canceled after 0s
Details
Unit Tests / Unit Tests :: windows-2022 (push) Canceled after 0s
Details
Publish no_rocm_image_ubuntu24_04 images / publish_no_rocm_image_ubuntu24_04_media (push) Canceled after 0s
Details
Unit Tests / Unit Tests Summary (push) Canceled after 0s
Details
Publish no_rocm_image_ubuntu24_04 images / publish_no_rocm_image_ubuntu24_04_openmpi (push) Canceled after 0s
Details
Flatten subprojects into a single base monorepo
2026-09-02 16:34:47 -07:00

2060 lines
84 KiB
C++
Raw Blame History

// Copyright © Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
#include "DescriptorTestUtils.hpp"
#include "GraphTestUtils.hpp"
#include "HipdnnException.hpp"
#include "HipdnnOperationType.h"
#include "TensorDescriptorTestUtils.hpp"
#include "TestMacros.hpp"
#include "descriptors/ConvolutionFwdOperationDescriptor.hpp"
#include "descriptors/DataTypeConversion.hpp"
#include "descriptors/GraphDescriptor.hpp"
#include "descriptors/TensorDescriptor.hpp"
#include "hipdnn_backend.h"
#include "mocks/MockHandle.hpp"
#include <flatbuffers/flatbuffers.h>
#include <gtest/gtest.h>
#include <hipdnn_flatbuffers_sdk/data_objects/convolution_fwd_attributes_generated.h>
#include <hipdnn_flatbuffers_sdk/data_objects/graph_generated.h>
#include <hipdnn_flatbuffers_sdk/data_objects/tensor_attributes_generated.h>
#include <hipdnn_test_sdk/constants/ConvFpropConstants.hpp>
#include <hipdnn_test_sdk/utilities/ToVec.hpp>
#include <array>
#include <map>
#include <memory>
#include <set>
#include <vector>
using namespace hipdnn_backend;
using namespace hipdnn_backend::test_utilities;
using namespace hipdnn_flatbuffers_sdk::data_objects;
using namespace hipdnn_tests::constants;
using hipdnn_tests::toVec;
namespace
{
// Alternate UIDs for testing attribute-order independence
constexpr int64_t K_ALT_TENSOR_X_UID = 11;
constexpr int64_t K_ALT_TENSOR_W_UID = 12;
constexpr int64_t K_ALT_TENSOR_Y_UID = 13;
// UIDs for the second conv op in SharedTensorDifferentPositions
constexpr int64_t K_SHARED_TENSOR_W_UID = 14;
constexpr int64_t K_SHARED_TENSOR_Y_UID = 15;
} // namespace
class TestGraphDescriptorOps : public ::testing::Test
{
public:
// Validate a tensor's fields against expected values
static void verifyTensor(const TensorAttributesT* tensor,
int64_t expectedUid,
const std::vector<int64_t>& expectedDims,
const std::vector<int64_t>& expectedStrides,
DataType expectedDataType,
bool expectedVirtual = false)
{
ASSERT_NE(tensor, nullptr) << "Tensor with UID " << expectedUid << " not found";
EXPECT_EQ(tensor->uid, expectedUid);
EXPECT_EQ(tensor->dims, expectedDims);
EXPECT_EQ(tensor->strides, expectedStrides);
EXPECT_EQ(tensor->data_type, expectedDataType);
EXPECT_EQ(tensor->virtual_, expectedVirtual);
}
// Validate a node's convolution forward attributes
static void verifyConvFwdNode(const NodeT& node,
DataType expectedComputeType,
int64_t expectedXUid,
int64_t expectedWUid,
int64_t expectedYUid,
const std::vector<int64_t>& expectedPrePadding,
const std::vector<int64_t>& expectedPostPadding,
const std::vector<int64_t>& expectedStride,
const std::vector<int64_t>& expectedDilation)
{
EXPECT_EQ(node.compute_data_type, expectedComputeType);
ASSERT_EQ(node.attributes.type, NodeAttributes::ConvolutionFwdAttributes);
auto* convAttrs = node.attributes.AsConvolutionFwdAttributes();
ASSERT_NE(convAttrs, nullptr);
EXPECT_EQ(convAttrs->x_tensor_uid, expectedXUid);
EXPECT_EQ(convAttrs->w_tensor_uid, expectedWUid);
EXPECT_EQ(convAttrs->y_tensor_uid, expectedYUid);
EXPECT_EQ(convAttrs->pre_padding, expectedPrePadding);
EXPECT_EQ(convAttrs->post_padding, expectedPostPadding);
EXPECT_EQ(convAttrs->stride, expectedStride);
EXPECT_EQ(convAttrs->dilation, expectedDilation);
}
std::shared_ptr<GraphDescriptor> getDescriptor() const
{
return _wrapper->asDescriptor<GraphDescriptor>();
}
void setHandle() const
{
auto desc = getDescriptor();
hipdnnHandle_t handle = &_mockHandle;
desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_HANDLE,
HIPDNN_TYPE_HANDLE,
1,
static_cast<const void*>(&handle));
}
protected:
std::unique_ptr<HipdnnBackendDescriptor> _wrapper = nullptr;
mutable MockHandle _mockHandle;
void SetUp() override
{
_wrapper = createDescriptor<GraphDescriptor>();
}
void TearDown() override
{
_wrapper.reset();
}
};
// =============================================================================
// Build From Operations Tests
// =============================================================================
TEST_F(TestGraphDescriptorOps, BuildFromSingleOperation)
{
auto conv = createDefaultConvOp();
auto desc = getDescriptor();
setHandle();
std::array<HipdnnBackendDescriptor*, 1> ops = {conv.convOp.get()};
ASSERT_NO_THROW(desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(ops.data())));
ASSERT_NO_THROW(desc->finalize());
// Verify the built graph
auto serialized = desc->getSerializedGraph();
ASSERT_NE(serialized.ptr, nullptr);
ASSERT_GT(serialized.size, 0UL);
flatbuffers::Verifier verifier(static_cast<const uint8_t*>(serialized.ptr), serialized.size);
ASSERT_TRUE(verifier.VerifyBuffer<Graph>());
auto graphT = UnPackGraph(serialized.ptr);
ASSERT_EQ(graphT->nodes.size(), 1);
ASSERT_EQ(graphT->tensors.size(), 3);
// Verify each tensor has correct fields
verifyTensor(findTensorByUid(*graphT, K_FPROP_TENSOR_X_UID),
K_FPROP_TENSOR_X_UID,
toVec(K_FPROP_TENSOR_X_DIMS),
toVec(K_FPROP_TENSOR_X_STRIDES),
DataType::FLOAT);
verifyTensor(findTensorByUid(*graphT, K_FPROP_TENSOR_W_UID),
K_FPROP_TENSOR_W_UID,
toVec(K_FPROP_TENSOR_W_DIMS),
toVec(K_FPROP_TENSOR_W_STRIDES),
DataType::FLOAT);
verifyTensor(findTensorByUid(*graphT, K_FPROP_TENSOR_Y_UID),
K_FPROP_TENSOR_Y_UID,
toVec(K_FPROP_TENSOR_Y_DIMS),
toVec(K_FPROP_TENSOR_Y_STRIDES),
DataType::FLOAT);
// Verify the node's convolution attributes and tensor UID references
verifyConvFwdNode(*graphT->nodes[0],
DataType::FLOAT,
K_FPROP_TENSOR_X_UID,
K_FPROP_TENSOR_W_UID,
K_FPROP_TENSOR_Y_UID,
toVec(K_FPROP_CONV_PADDING),
toVec(K_FPROP_CONV_PADDING),
toVec(K_FPROP_CONV_STRIDE),
toVec(K_FPROP_CONV_DILATION));
}
TEST_F(TestGraphDescriptorOps, BuildFromMultipleOperations)
{
// First conv: primary tensor set
auto conv1 = createDefaultConvOp();
// Second conv: second tensor set
auto xDesc2 = createFinalizedTensor(
K_FPROP_TENSOR_X2_UID, toVec(K_FPROP_TENSOR_X2_DIMS), toVec(K_FPROP_TENSOR_X2_STRIDES));
auto wDesc2 = createFinalizedTensor(
K_FPROP_TENSOR_W2_UID, toVec(K_FPROP_TENSOR_W2_DIMS), toVec(K_FPROP_TENSOR_W2_STRIDES));
auto yDesc2 = createFinalizedTensor(
K_FPROP_TENSOR_Y2_UID, toVec(K_FPROP_TENSOR_Y2_DIMS), toVec(K_FPROP_TENSOR_Y2_STRIDES));
auto convOp2 = createFinalizedConvOp(xDesc2.get(), wDesc2.get(), yDesc2.get());
auto desc = getDescriptor();
setHandle();
std::array<HipdnnBackendDescriptor*, 2> ops = {conv1.convOp.get(), convOp2.get()};
ASSERT_NO_THROW(desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
2,
static_cast<const void*>(ops.data())));
ASSERT_NO_THROW(desc->finalize());
auto serialized = desc->getSerializedGraph();
auto graphT = UnPackGraph(serialized.ptr);
ASSERT_EQ(graphT->nodes.size(), 2);
ASSERT_EQ(graphT->tensors.size(), 6);
// Verify all tensors from first conv op
verifyTensor(findTensorByUid(*graphT, K_FPROP_TENSOR_X_UID),
K_FPROP_TENSOR_X_UID,
toVec(K_FPROP_TENSOR_X_DIMS),
toVec(K_FPROP_TENSOR_X_STRIDES),
DataType::FLOAT);
verifyTensor(findTensorByUid(*graphT, K_FPROP_TENSOR_W_UID),
K_FPROP_TENSOR_W_UID,
toVec(K_FPROP_TENSOR_W_DIMS),
toVec(K_FPROP_TENSOR_W_STRIDES),
DataType::FLOAT);
verifyTensor(findTensorByUid(*graphT, K_FPROP_TENSOR_Y_UID),
K_FPROP_TENSOR_Y_UID,
toVec(K_FPROP_TENSOR_Y_DIMS),
toVec(K_FPROP_TENSOR_Y_STRIDES),
DataType::FLOAT);
// Verify all tensors from second conv op
verifyTensor(findTensorByUid(*graphT, K_FPROP_TENSOR_X2_UID),
K_FPROP_TENSOR_X2_UID,
toVec(K_FPROP_TENSOR_X2_DIMS),
toVec(K_FPROP_TENSOR_X2_STRIDES),
DataType::FLOAT);
verifyTensor(findTensorByUid(*graphT, K_FPROP_TENSOR_W2_UID),
K_FPROP_TENSOR_W2_UID,
toVec(K_FPROP_TENSOR_W2_DIMS),
toVec(K_FPROP_TENSOR_W2_STRIDES),
DataType::FLOAT);
verifyTensor(findTensorByUid(*graphT, K_FPROP_TENSOR_Y2_UID),
K_FPROP_TENSOR_Y2_UID,
toVec(K_FPROP_TENSOR_Y2_DIMS),
toVec(K_FPROP_TENSOR_Y2_STRIDES),
DataType::FLOAT);
// Verify first node references primary tensors
verifyConvFwdNode(*graphT->nodes[0],
DataType::FLOAT,
K_FPROP_TENSOR_X_UID,
K_FPROP_TENSOR_W_UID,
K_FPROP_TENSOR_Y_UID,
toVec(K_FPROP_CONV_PADDING),
toVec(K_FPROP_CONV_PADDING),
toVec(K_FPROP_CONV_STRIDE),
toVec(K_FPROP_CONV_DILATION));
// Verify second node references second tensor set
verifyConvFwdNode(*graphT->nodes[1],
DataType::FLOAT,
K_FPROP_TENSOR_X2_UID,
K_FPROP_TENSOR_W2_UID,
K_FPROP_TENSOR_Y2_UID,
toVec(K_FPROP_CONV_PADDING),
toVec(K_FPROP_CONV_PADDING),
toVec(K_FPROP_CONV_STRIDE),
toVec(K_FPROP_CONV_DILATION));
}
TEST_F(TestGraphDescriptorOps, TensorDeduplication)
{
// Two operations sharing the same output tensor (Y from first op)
auto conv1 = createDefaultConvOp();
auto xDesc2 = createFinalizedTensor(
K_FPROP_TENSOR_X2_UID, toVec(K_FPROP_TENSOR_X2_DIMS), toVec(K_FPROP_TENSOR_X2_STRIDES));
auto wDesc2 = createFinalizedTensor(
K_FPROP_TENSOR_W2_UID, toVec(K_FPROP_TENSOR_W2_DIMS), toVec(K_FPROP_TENSOR_W2_STRIDES));
// Reuse conv1.yDesc (uid 3) as Y for the second op
auto convOp2 = createFinalizedConvOp(xDesc2.get(), wDesc2.get(), conv1.yDesc.get());
auto desc = getDescriptor();
setHandle();
std::array<HipdnnBackendDescriptor*, 2> ops = {conv1.convOp.get(), convOp2.get()};
ASSERT_NO_THROW(desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
2,
static_cast<const void*>(ops.data())));
ASSERT_NO_THROW(desc->finalize());
auto serialized = desc->getSerializedGraph();
auto graphT = UnPackGraph(serialized.ptr);
ASSERT_EQ(graphT->nodes.size(), 2);
// Should have 5 unique tensors, not 6 (tensor Y deduplicated)
ASSERT_EQ(graphT->tensors.size(), 5);
// Verify tensor UIDs are unique
std::set<int64_t> tensorUids;
for(const auto& tensor : graphT->tensors)
{
tensorUids.insert(tensor->uid);
}
EXPECT_EQ(tensorUids.size(), 5);
// Verify each tensor's fields
verifyTensor(findTensorByUid(*graphT, K_FPROP_TENSOR_X_UID),
K_FPROP_TENSOR_X_UID,
toVec(K_FPROP_TENSOR_X_DIMS),
toVec(K_FPROP_TENSOR_X_STRIDES),
DataType::FLOAT);
verifyTensor(findTensorByUid(*graphT, K_FPROP_TENSOR_W_UID),
K_FPROP_TENSOR_W_UID,
toVec(K_FPROP_TENSOR_W_DIMS),
toVec(K_FPROP_TENSOR_W_STRIDES),
DataType::FLOAT);
verifyTensor(findTensorByUid(*graphT, K_FPROP_TENSOR_Y_UID),
K_FPROP_TENSOR_Y_UID,
toVec(K_FPROP_TENSOR_Y_DIMS),
toVec(K_FPROP_TENSOR_Y_STRIDES),
DataType::FLOAT);
verifyTensor(findTensorByUid(*graphT, K_FPROP_TENSOR_X2_UID),
K_FPROP_TENSOR_X2_UID,
toVec(K_FPROP_TENSOR_X2_DIMS),
toVec(K_FPROP_TENSOR_X2_STRIDES),
DataType::FLOAT);
verifyTensor(findTensorByUid(*graphT, K_FPROP_TENSOR_W2_UID),
K_FPROP_TENSOR_W2_UID,
toVec(K_FPROP_TENSOR_W2_DIMS),
toVec(K_FPROP_TENSOR_W2_STRIDES),
DataType::FLOAT);
// Verify first node: primary tensors
verifyConvFwdNode(*graphT->nodes[0],
DataType::FLOAT,
K_FPROP_TENSOR_X_UID,
K_FPROP_TENSOR_W_UID,
K_FPROP_TENSOR_Y_UID,
toVec(K_FPROP_CONV_PADDING),
toVec(K_FPROP_CONV_PADDING),
toVec(K_FPROP_CONV_STRIDE),
toVec(K_FPROP_CONV_DILATION));
// Verify second node also references Y (the shared tensor)
verifyConvFwdNode(*graphT->nodes[1],
DataType::FLOAT,
K_FPROP_TENSOR_X2_UID,
K_FPROP_TENSOR_W2_UID,
K_FPROP_TENSOR_Y_UID,
toVec(K_FPROP_CONV_PADDING),
toVec(K_FPROP_CONV_PADDING),
toVec(K_FPROP_CONV_STRIDE),
toVec(K_FPROP_CONV_DILATION));
}
TEST_F(TestGraphDescriptorOps, ComputeDataTypePreserved)
{
auto conv = createDefaultConvOp(HIPDNN_DATA_HALF);
auto desc = getDescriptor();
setHandle();
std::array<HipdnnBackendDescriptor*, 1> ops = {conv.convOp.get()};
desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(ops.data()));
desc->finalize();
auto serialized = desc->getSerializedGraph();
auto graphT = UnPackGraph(serialized.ptr);
ASSERT_EQ(graphT->nodes.size(), 1);
ASSERT_EQ(graphT->tensors.size(), 3);
// Verify tensors retain FLOAT data type (tensor data type is independent of compute type)
verifyTensor(findTensorByUid(*graphT, K_FPROP_TENSOR_X_UID),
K_FPROP_TENSOR_X_UID,
toVec(K_FPROP_TENSOR_X_DIMS),
toVec(K_FPROP_TENSOR_X_STRIDES),
DataType::FLOAT);
verifyTensor(findTensorByUid(*graphT, K_FPROP_TENSOR_W_UID),
K_FPROP_TENSOR_W_UID,
toVec(K_FPROP_TENSOR_W_DIMS),
toVec(K_FPROP_TENSOR_W_STRIDES),
DataType::FLOAT);
verifyTensor(findTensorByUid(*graphT, K_FPROP_TENSOR_Y_UID),
K_FPROP_TENSOR_Y_UID,
toVec(K_FPROP_TENSOR_Y_DIMS),
toVec(K_FPROP_TENSOR_Y_STRIDES),
DataType::FLOAT);
// Verify node compute data type is HALF and all conv attributes are correct
verifyConvFwdNode(*graphT->nodes[0],
DataType::HALF,
K_FPROP_TENSOR_X_UID,
K_FPROP_TENSOR_W_UID,
K_FPROP_TENSOR_Y_UID,
toVec(K_FPROP_CONV_PADDING),
toVec(K_FPROP_CONV_PADDING),
toVec(K_FPROP_CONV_STRIDE),
toVec(K_FPROP_CONV_DILATION));
}
TEST_F(TestGraphDescriptorOps, ConvolutionAttributesPreserved)
{
auto conv = createDefaultConvOp();
// Create conv op with non-default parameters to test graph roundtrip
auto wrapper = createDescriptor<ConvolutionFwdOperationDescriptor>();
auto convDesc = wrapper->asDescriptor<ConvolutionFwdOperationDescriptor>();
HipdnnBackendDescriptor* x = conv.xDesc.get();
HipdnnBackendDescriptor* w = conv.wDesc.get();
HipdnnBackendDescriptor* y = conv.yDesc.get();
convDesc->setAttribute(HIPDNN_ATTR_OPERATION_CONVOLUTION_FORWARD_X,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(&x));
convDesc->setAttribute(HIPDNN_ATTR_OPERATION_CONVOLUTION_FORWARD_W,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(&w));
convDesc->setAttribute(HIPDNN_ATTR_OPERATION_CONVOLUTION_FORWARD_Y,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(&y));
const std::vector<int64_t> kCustomPrePadding = {2, 3};
const std::vector<int64_t> kCustomPostPadding = {4, 5};
const std::vector<int64_t> kCustomStride = {2, 2};
auto dilation = toVec(K_FPROP_CONV_DILATION);
convDesc->setAttribute(
HIPDNN_ATTR_CONVOLUTION_PRE_PADDINGS, HIPDNN_TYPE_INT64, 2, kCustomPrePadding.data());
convDesc->setAttribute(
HIPDNN_ATTR_CONVOLUTION_POST_PADDINGS, HIPDNN_TYPE_INT64, 2, kCustomPostPadding.data());
convDesc->setAttribute(
HIPDNN_ATTR_CONVOLUTION_FILTER_STRIDES, HIPDNN_TYPE_INT64, 2, kCustomStride.data());
convDesc->setAttribute(
HIPDNN_ATTR_CONVOLUTION_DILATIONS, HIPDNN_TYPE_INT64, 2, dilation.data());
auto computeType = HIPDNN_DATA_FLOAT;
convDesc->setAttribute(
HIPDNN_ATTR_CONVOLUTION_COMP_TYPE, HIPDNN_TYPE_DATA_TYPE, 1, &computeType);
hipdnnConvolutionMode_t convMode = HIPDNN_CROSS_CORRELATION;
convDesc->setAttribute(
HIPDNN_ATTR_CONVOLUTION_CONV_MODE, HIPDNN_TYPE_CONVOLUTION_MODE, 1, &convMode);
convDesc->finalize();
auto desc = getDescriptor();
setHandle();
std::array<HipdnnBackendDescriptor*, 1> ops = {wrapper.get()};
desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(ops.data()));
desc->finalize();
auto serialized = desc->getSerializedGraph();
auto graphT = UnPackGraph(serialized.ptr);
ASSERT_EQ(graphT->nodes.size(), 1);
ASSERT_EQ(graphT->tensors.size(), 3);
// Verify tensors
verifyTensor(findTensorByUid(*graphT, K_FPROP_TENSOR_X_UID),
K_FPROP_TENSOR_X_UID,
toVec(K_FPROP_TENSOR_X_DIMS),
toVec(K_FPROP_TENSOR_X_STRIDES),
DataType::FLOAT);
verifyTensor(findTensorByUid(*graphT, K_FPROP_TENSOR_W_UID),
K_FPROP_TENSOR_W_UID,
toVec(K_FPROP_TENSOR_W_DIMS),
toVec(K_FPROP_TENSOR_W_STRIDES),
DataType::FLOAT);
verifyTensor(findTensorByUid(*graphT, K_FPROP_TENSOR_Y_UID),
K_FPROP_TENSOR_Y_UID,
toVec(K_FPROP_TENSOR_Y_DIMS),
toVec(K_FPROP_TENSOR_Y_STRIDES),
DataType::FLOAT);
// Verify node with asymmetric padding and non-unit stride
verifyConvFwdNode(*graphT->nodes[0],
DataType::FLOAT,
K_FPROP_TENSOR_X_UID,
K_FPROP_TENSOR_W_UID,
K_FPROP_TENSOR_Y_UID,
kCustomPrePadding,
kCustomPostPadding,
kCustomStride,
toVec(K_FPROP_CONV_DILATION));
}
// =============================================================================
// SetAttribute Order Tests
// =============================================================================
TEST_F(TestGraphDescriptorOps, SetOperationsAndHandleAnyOrder)
{
// Test: operations first, then handle
{
auto graphWrapper = createDescriptor<GraphDescriptor>();
auto desc = graphWrapper->asDescriptor<GraphDescriptor>();
auto conv = createDefaultConvOp();
std::array<HipdnnBackendDescriptor*, 1> ops = {conv.convOp.get()};
ASSERT_NO_THROW(desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(ops.data())));
hipdnnHandle_t handle = &_mockHandle;
ASSERT_NO_THROW(desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_HANDLE,
HIPDNN_TYPE_HANDLE,
1,
static_cast<const void*>(&handle)));
ASSERT_NO_THROW(desc->finalize());
auto serialized = desc->getSerializedGraph();
auto graphT = UnPackGraph(serialized.ptr);
ASSERT_EQ(graphT->tensors.size(), 3);
ASSERT_EQ(graphT->nodes.size(), 1);
verifyTensor(findTensorByUid(*graphT, K_FPROP_TENSOR_X_UID),
K_FPROP_TENSOR_X_UID,
toVec(K_FPROP_TENSOR_X_DIMS),
toVec(K_FPROP_TENSOR_X_STRIDES),
DataType::FLOAT);
verifyTensor(findTensorByUid(*graphT, K_FPROP_TENSOR_W_UID),
K_FPROP_TENSOR_W_UID,
toVec(K_FPROP_TENSOR_W_DIMS),
toVec(K_FPROP_TENSOR_W_STRIDES),
DataType::FLOAT);
verifyTensor(findTensorByUid(*graphT, K_FPROP_TENSOR_Y_UID),
K_FPROP_TENSOR_Y_UID,
toVec(K_FPROP_TENSOR_Y_DIMS),
toVec(K_FPROP_TENSOR_Y_STRIDES),
DataType::FLOAT);
verifyConvFwdNode(*graphT->nodes[0],
DataType::FLOAT,
K_FPROP_TENSOR_X_UID,
K_FPROP_TENSOR_W_UID,
K_FPROP_TENSOR_Y_UID,
toVec(K_FPROP_CONV_PADDING),
toVec(K_FPROP_CONV_PADDING),
toVec(K_FPROP_CONV_STRIDE),
toVec(K_FPROP_CONV_DILATION));
}
// Test: handle first, then operations (with alternate UIDs)
{
auto graphWrapper = createDescriptor<GraphDescriptor>();
auto desc = graphWrapper->asDescriptor<GraphDescriptor>();
auto xDesc = createFinalizedTensor(K_ALT_TENSOR_X_UID);
auto wDesc = createFinalizedTensor(
K_ALT_TENSOR_W_UID, toVec(K_FPROP_TENSOR_W_DIMS), toVec(K_FPROP_TENSOR_W_STRIDES));
auto yDesc = createFinalizedTensor(
K_ALT_TENSOR_Y_UID, toVec(K_FPROP_TENSOR_Y_DIMS), toVec(K_FPROP_TENSOR_Y_STRIDES));
auto convOp = createFinalizedConvOp(xDesc.get(), wDesc.get(), yDesc.get());
hipdnnHandle_t handle = &_mockHandle;
ASSERT_NO_THROW(desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_HANDLE,
HIPDNN_TYPE_HANDLE,
1,
static_cast<const void*>(&handle)));
std::array<HipdnnBackendDescriptor*, 1> ops = {convOp.get()};
ASSERT_NO_THROW(desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(ops.data())));
ASSERT_NO_THROW(desc->finalize());
auto serialized = desc->getSerializedGraph();
auto graphT = UnPackGraph(serialized.ptr);
ASSERT_EQ(graphT->tensors.size(), 3);
ASSERT_EQ(graphT->nodes.size(), 1);
verifyTensor(findTensorByUid(*graphT, K_ALT_TENSOR_X_UID),
K_ALT_TENSOR_X_UID,
toVec(K_FPROP_TENSOR_X_DIMS),
toVec(K_FPROP_TENSOR_X_STRIDES),
DataType::FLOAT);
verifyTensor(findTensorByUid(*graphT, K_ALT_TENSOR_W_UID),
K_ALT_TENSOR_W_UID,
toVec(K_FPROP_TENSOR_W_DIMS),
toVec(K_FPROP_TENSOR_W_STRIDES),
DataType::FLOAT);
verifyTensor(findTensorByUid(*graphT, K_ALT_TENSOR_Y_UID),
K_ALT_TENSOR_Y_UID,
toVec(K_FPROP_TENSOR_Y_DIMS),
toVec(K_FPROP_TENSOR_Y_STRIDES),
DataType::FLOAT);
verifyConvFwdNode(*graphT->nodes[0],
DataType::FLOAT,
K_ALT_TENSOR_X_UID,
K_ALT_TENSOR_W_UID,
K_ALT_TENSOR_Y_UID,
toVec(K_FPROP_CONV_PADDING),
toVec(K_FPROP_CONV_PADDING),
toVec(K_FPROP_CONV_STRIDE),
toVec(K_FPROP_CONV_DILATION));
}
}
TEST_F(TestGraphDescriptorOps, SetOperationsMultipleBatches)
{
auto conv1 = createDefaultConvOp();
auto xDesc2 = createFinalizedTensor(
K_FPROP_TENSOR_X2_UID, toVec(K_FPROP_TENSOR_X2_DIMS), toVec(K_FPROP_TENSOR_X2_STRIDES));
auto wDesc2 = createFinalizedTensor(
K_FPROP_TENSOR_W2_UID, toVec(K_FPROP_TENSOR_W2_DIMS), toVec(K_FPROP_TENSOR_W2_STRIDES));
auto yDesc2 = createFinalizedTensor(
K_FPROP_TENSOR_Y2_UID, toVec(K_FPROP_TENSOR_Y2_DIMS), toVec(K_FPROP_TENSOR_Y2_STRIDES));
auto convOp2 = createFinalizedConvOp(xDesc2.get(), wDesc2.get(), yDesc2.get());
auto desc = getDescriptor();
setHandle();
// Set multiple operations in a single setAttribute call
std::array<HipdnnBackendDescriptor*, 2> ops = {conv1.convOp.get(), convOp2.get()};
ASSERT_NO_THROW(desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
2,
static_cast<const void*>(ops.data())));
ASSERT_NO_THROW(desc->finalize());
auto serialized = desc->getSerializedGraph();
auto graphT = UnPackGraph(serialized.ptr);
// Both operations should be present in the graph
ASSERT_EQ(graphT->nodes.size(), 2);
ASSERT_EQ(graphT->tensors.size(), 6);
// Verify tensors from first operation
verifyTensor(findTensorByUid(*graphT, K_FPROP_TENSOR_X_UID),
K_FPROP_TENSOR_X_UID,
toVec(K_FPROP_TENSOR_X_DIMS),
toVec(K_FPROP_TENSOR_X_STRIDES),
DataType::FLOAT);
verifyTensor(findTensorByUid(*graphT, K_FPROP_TENSOR_W_UID),
K_FPROP_TENSOR_W_UID,
toVec(K_FPROP_TENSOR_W_DIMS),
toVec(K_FPROP_TENSOR_W_STRIDES),
DataType::FLOAT);
verifyTensor(findTensorByUid(*graphT, K_FPROP_TENSOR_Y_UID),
K_FPROP_TENSOR_Y_UID,
toVec(K_FPROP_TENSOR_Y_DIMS),
toVec(K_FPROP_TENSOR_Y_STRIDES),
DataType::FLOAT);
// Verify tensors from second operation
verifyTensor(findTensorByUid(*graphT, K_FPROP_TENSOR_X2_UID),
K_FPROP_TENSOR_X2_UID,
toVec(K_FPROP_TENSOR_X2_DIMS),
toVec(K_FPROP_TENSOR_X2_STRIDES),
DataType::FLOAT);
verifyTensor(findTensorByUid(*graphT, K_FPROP_TENSOR_W2_UID),
K_FPROP_TENSOR_W2_UID,
toVec(K_FPROP_TENSOR_W2_DIMS),
toVec(K_FPROP_TENSOR_W2_STRIDES),
DataType::FLOAT);
verifyTensor(findTensorByUid(*graphT, K_FPROP_TENSOR_Y2_UID),
K_FPROP_TENSOR_Y2_UID,
toVec(K_FPROP_TENSOR_Y2_DIMS),
toVec(K_FPROP_TENSOR_Y2_STRIDES),
DataType::FLOAT);
// Verify both nodes reference correct tensor UIDs
verifyConvFwdNode(*graphT->nodes[0],
DataType::FLOAT,
K_FPROP_TENSOR_X_UID,
K_FPROP_TENSOR_W_UID,
K_FPROP_TENSOR_Y_UID,
toVec(K_FPROP_CONV_PADDING),
toVec(K_FPROP_CONV_PADDING),
toVec(K_FPROP_CONV_STRIDE),
toVec(K_FPROP_CONV_DILATION));
verifyConvFwdNode(*graphT->nodes[1],
DataType::FLOAT,
K_FPROP_TENSOR_X2_UID,
K_FPROP_TENSOR_W2_UID,
K_FPROP_TENSOR_Y2_UID,
toVec(K_FPROP_CONV_PADDING),
toVec(K_FPROP_CONV_PADDING),
toVec(K_FPROP_CONV_STRIDE),
toVec(K_FPROP_CONV_DILATION));
}
// =============================================================================
// Error Cases
// =============================================================================
TEST_F(TestGraphDescriptorOps, SetOperationsFailsUnfinalized)
{
auto desc = getDescriptor();
// Create unfinalized operation
auto unfinalizedOp = createDescriptor<ConvolutionFwdOperationDescriptor>();
std::array<HipdnnBackendDescriptor*, 1> ops = {unfinalizedOp.get()};
ASSERT_THROW_HIPDNN_STATUS(desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(ops.data())),
HIPDNN_STATUS_BAD_PARAM_NOT_FINALIZED);
}
TEST_F(TestGraphDescriptorOps, SetOperationsFailsNullDescriptor)
{
auto desc = getDescriptor();
std::array<HipdnnBackendDescriptor*, 1> ops = {nullptr};
ASSERT_THROW_HIPDNN_STATUS(desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(ops.data())),
HIPDNN_STATUS_BAD_PARAM_NULL_POINTER);
}
TEST_F(TestGraphDescriptorOps, SetOperationsFailsWrongType)
{
auto desc = getDescriptor();
// Use a TensorDescriptor instead of an operation descriptor
auto tensorDesc = createFinalizedTensor(K_FPROP_TENSOR_X_UID);
std::array<HipdnnBackendDescriptor*, 1> ops = {tensorDesc.get()};
ASSERT_THROW_HIPDNN_STATUS(desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(ops.data())),
HIPDNN_STATUS_NOT_SUPPORTED);
}
TEST_F(TestGraphDescriptorOps, SetOperationsFailsAfterFinalize)
{
auto conv = createDefaultConvOp();
auto desc = getDescriptor();
setHandle();
std::array<HipdnnBackendDescriptor*, 1> ops = {conv.convOp.get()};
desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(ops.data()));
desc->finalize();
// Try to set operations after finalize
ASSERT_THROW_HIPDNN_STATUS(desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(ops.data())),
HIPDNN_STATUS_NOT_INITIALIZED);
}
TEST_F(TestGraphDescriptorOps, FinalizeFailsWithoutHandle)
{
auto conv = createDefaultConvOp();
auto desc = getDescriptor();
// Don't set handle
std::array<HipdnnBackendDescriptor*, 1> ops = {conv.convOp.get()};
desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(ops.data()));
ASSERT_THROW_HIPDNN_STATUS(desc->finalize(), HIPDNN_STATUS_BAD_PARAM);
}
TEST_F(TestGraphDescriptorOps, FinalizeFailsWithoutOperationsOrGraph)
{
auto desc = getDescriptor();
setHandle();
// Don't set operations or deserialize graph
ASSERT_THROW_HIPDNN_STATUS(desc->finalize(), HIPDNN_STATUS_BAD_PARAM);
}
// =============================================================================
// Serialization Tests
// =============================================================================
TEST_F(TestGraphDescriptorOps, SerializedGraphVerifiable)
{
auto conv = createDefaultConvOp();
auto desc = getDescriptor();
setHandle();
std::array<HipdnnBackendDescriptor*, 1> ops = {conv.convOp.get()};
desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(ops.data()));
desc->finalize();
auto serialized = desc->getSerializedGraph();
// Verify FlatBuffer is valid
flatbuffers::Verifier verifier(static_cast<const uint8_t*>(serialized.ptr), serialized.size);
ASSERT_TRUE(verifier.VerifyBuffer<Graph>());
// Verify the verified buffer contains the expected graph structure
auto graphT = UnPackGraph(serialized.ptr);
ASSERT_EQ(graphT->tensors.size(), 3);
ASSERT_EQ(graphT->nodes.size(), 1);
verifyConvFwdNode(*graphT->nodes[0],
DataType::FLOAT,
K_FPROP_TENSOR_X_UID,
K_FPROP_TENSOR_W_UID,
K_FPROP_TENSOR_Y_UID,
toVec(K_FPROP_CONV_PADDING),
toVec(K_FPROP_CONV_PADDING),
toVec(K_FPROP_CONV_STRIDE),
toVec(K_FPROP_CONV_DILATION));
}
TEST_F(TestGraphDescriptorOps, SerializedGraphUnpackable)
{
auto conv = createDefaultConvOp();
auto desc = getDescriptor();
setHandle();
std::array<HipdnnBackendDescriptor*, 1> ops = {conv.convOp.get()};
desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(ops.data()));
desc->finalize();
auto serialized = desc->getSerializedGraph();
ASSERT_NE(GetGraph(serialized.ptr), nullptr);
// Unpack should work and produce correct values
auto graphT = UnPackGraph(serialized.ptr);
ASSERT_NE(graphT, nullptr);
ASSERT_EQ(graphT->tensors.size(), 3);
ASSERT_EQ(graphT->nodes.size(), 1);
// Verify unpacked tensor values match input
verifyTensor(findTensorByUid(*graphT, K_FPROP_TENSOR_X_UID),
K_FPROP_TENSOR_X_UID,
toVec(K_FPROP_TENSOR_X_DIMS),
toVec(K_FPROP_TENSOR_X_STRIDES),
DataType::FLOAT);
verifyTensor(findTensorByUid(*graphT, K_FPROP_TENSOR_W_UID),
K_FPROP_TENSOR_W_UID,
toVec(K_FPROP_TENSOR_W_DIMS),
toVec(K_FPROP_TENSOR_W_STRIDES),
DataType::FLOAT);
verifyTensor(findTensorByUid(*graphT, K_FPROP_TENSOR_Y_UID),
K_FPROP_TENSOR_Y_UID,
toVec(K_FPROP_TENSOR_Y_DIMS),
toVec(K_FPROP_TENSOR_Y_STRIDES),
DataType::FLOAT);
// Verify unpacked node values match input
verifyConvFwdNode(*graphT->nodes[0],
DataType::FLOAT,
K_FPROP_TENSOR_X_UID,
K_FPROP_TENSOR_W_UID,
K_FPROP_TENSOR_Y_UID,
toVec(K_FPROP_CONV_PADDING),
toVec(K_FPROP_CONV_PADDING),
toVec(K_FPROP_CONV_STRIDE),
toVec(K_FPROP_CONV_DILATION));
}
TEST_F(TestGraphDescriptorOps, GetSerializedGraphMultipleCalls)
{
auto conv = createDefaultConvOp();
auto desc = getDescriptor();
setHandle();
std::array<HipdnnBackendDescriptor*, 1> ops = {conv.convOp.get()};
desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(ops.data()));
desc->finalize();
// Call getSerializedGraph multiple times
auto serialized1 = desc->getSerializedGraph();
auto serialized2 = desc->getSerializedGraph();
// Should return same data
EXPECT_EQ(serialized1.ptr, serialized2.ptr);
EXPECT_EQ(serialized1.size, serialized2.size);
// Both should unpack to identical graph values
auto graphT1 = UnPackGraph(serialized1.ptr);
auto graphT2 = UnPackGraph(serialized2.ptr);
ASSERT_EQ(graphT1->tensors.size(), 3);
ASSERT_EQ(graphT2->tensors.size(), 3);
ASSERT_EQ(graphT1->nodes.size(), 1);
ASSERT_EQ(graphT2->nodes.size(), 1);
// Verify both unpacked graphs contain identical tensor data
for(size_t i = 0; i < graphT1->tensors.size(); ++i)
{
EXPECT_EQ(graphT1->tensors[i]->uid, graphT2->tensors[i]->uid);
EXPECT_EQ(graphT1->tensors[i]->dims, graphT2->tensors[i]->dims);
EXPECT_EQ(graphT1->tensors[i]->strides, graphT2->tensors[i]->strides);
EXPECT_EQ(graphT1->tensors[i]->data_type, graphT2->tensors[i]->data_type);
}
// Verify both unpacked graphs contain identical node data
EXPECT_EQ(graphT1->nodes[0]->compute_data_type, graphT2->nodes[0]->compute_data_type);
EXPECT_EQ(graphT1->nodes[0]->attributes.type, graphT2->nodes[0]->attributes.type);
}
// =============================================================================
// Graph Structure Verification Tests
// =============================================================================
TEST_F(TestGraphDescriptorOps, GraphHasCorrectNodeCount)
{
auto conv = createDefaultConvOp();
auto desc = getDescriptor();
setHandle();
std::array<HipdnnBackendDescriptor*, 1> ops = {conv.convOp.get()};
desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(ops.data()));
desc->finalize();
auto serialized = desc->getSerializedGraph();
auto graphT = UnPackGraph(serialized.ptr);
ASSERT_EQ(graphT->nodes.size(), 1);
// Verify node has ConvolutionFwdAttributes and correct tensor UID references
verifyConvFwdNode(*graphT->nodes[0],
DataType::FLOAT,
K_FPROP_TENSOR_X_UID,
K_FPROP_TENSOR_W_UID,
K_FPROP_TENSOR_Y_UID,
toVec(K_FPROP_CONV_PADDING),
toVec(K_FPROP_CONV_PADDING),
toVec(K_FPROP_CONV_STRIDE),
toVec(K_FPROP_CONV_DILATION));
}
TEST_F(TestGraphDescriptorOps, GraphHasCorrectTensorCount)
{
auto conv = createDefaultConvOp();
auto desc = getDescriptor();
setHandle();
std::array<HipdnnBackendDescriptor*, 1> ops = {conv.convOp.get()};
desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(ops.data()));
desc->finalize();
auto serialized = desc->getSerializedGraph();
auto graphT = UnPackGraph(serialized.ptr);
ASSERT_EQ(graphT->tensors.size(), 3);
// Verify each tensor's full field values (not just UIDs)
verifyTensor(findTensorByUid(*graphT, K_FPROP_TENSOR_X_UID),
K_FPROP_TENSOR_X_UID,
toVec(K_FPROP_TENSOR_X_DIMS),
toVec(K_FPROP_TENSOR_X_STRIDES),
DataType::FLOAT);
verifyTensor(findTensorByUid(*graphT, K_FPROP_TENSOR_W_UID),
K_FPROP_TENSOR_W_UID,
toVec(K_FPROP_TENSOR_W_DIMS),
toVec(K_FPROP_TENSOR_W_STRIDES),
DataType::FLOAT);
verifyTensor(findTensorByUid(*graphT, K_FPROP_TENSOR_Y_UID),
K_FPROP_TENSOR_Y_UID,
toVec(K_FPROP_TENSOR_Y_DIMS),
toVec(K_FPROP_TENSOR_Y_STRIDES),
DataType::FLOAT);
}
// =============================================================================
// Equivalence Tests - Compare FlatBuffer path vs Descriptor path
// =============================================================================
struct ConvEquivalenceParams
{
std::string name;
int64_t xUid;
int64_t wUid;
int64_t yUid;
std::vector<int64_t> xDims;
std::vector<int64_t> xStrides;
std::vector<int64_t> wDims;
std::vector<int64_t> wStrides;
std::vector<int64_t> yDims;
std::vector<int64_t> yStrides;
std::vector<int64_t> prePadding;
std::vector<int64_t> postPadding;
std::vector<int64_t> stride;
std::vector<int64_t> dilation;
hipdnnDataType_t tensorDataType;
hipdnnDataType_t computeDataType;
};
// Outside anonymous namespace so ADL finds it for gtest printing
// NOLINTNEXTLINE(misc-use-internal-linkage)
std::ostream& operator<<(std::ostream& os, const ConvEquivalenceParams& p)
{
return os << p.name;
}
class TestGraphDescriptorEquivalence : public TestGraphDescriptorOps,
public ::testing::WithParamInterface<ConvEquivalenceParams>
{
public:
// Build a graph via FlatBuffer serialization (the "old" path)
static flatbuffers::DetachedBuffer
buildGraphViaFlatBuffer(const TensorAttributesT& xTensor,
const TensorAttributesT& wTensor,
const TensorAttributesT& yTensor,
const ConvolutionFwdAttributesT& convAttrs,
DataType computeDataType)
{
flatbuffers::FlatBufferBuilder builder;
// Build tensors
std::vector<flatbuffers::Offset<TensorAttributes>> tensorOffsets;
tensorOffsets.push_back(TensorAttributes::Pack(builder, &xTensor));
tensorOffsets.push_back(TensorAttributes::Pack(builder, &wTensor));
tensorOffsets.push_back(TensorAttributes::Pack(builder, &yTensor));
// Build node with conv attributes
NodeT nodeT;
nodeT.compute_data_type = computeDataType;
nodeT.attributes.Set(ConvolutionFwdAttributesT(convAttrs));
std::vector<flatbuffers::Offset<Node>> nodeOffsets;
nodeOffsets.push_back(Node::Pack(builder, &nodeT));
// Build graph
auto graphOffset = CreateGraphDirect(builder,
nullptr, // name
DataType::UNSET,
DataType::UNSET,
DataType::UNSET,
&tensorOffsets,
&nodeOffsets);
builder.Finish(graphOffset);
return builder.Release();
}
// Build a graph via descriptors (the "new" path)
std::unique_ptr<GraphT> buildGraphViaDescriptors(int64_t xUid,
int64_t wUid,
int64_t yUid,
const std::vector<int64_t>& xDims,
const std::vector<int64_t>& xStrides,
const std::vector<int64_t>& wDims,
const std::vector<int64_t>& wStrides,
const std::vector<int64_t>& yDims,
const std::vector<int64_t>& yStrides,
const std::vector<int64_t>& prePadding,
const std::vector<int64_t>& postPadding,
const std::vector<int64_t>& stride,
const std::vector<int64_t>& dilation,
hipdnnDataType_t tensorDataType,
hipdnnDataType_t computeDataType)
{
// Create tensor descriptors
auto xDesc = createFinalizedTensor(xUid, xDims, xStrides, tensorDataType);
auto wDesc = createFinalizedTensor(wUid, wDims, wStrides, tensorDataType);
auto yDesc = createFinalizedTensor(yUid, yDims, yStrides, tensorDataType);
// Create conv op descriptor
auto convWrapper = createDescriptor<ConvolutionFwdOperationDescriptor>();
auto convDesc = convWrapper->asDescriptor<ConvolutionFwdOperationDescriptor>();
HipdnnBackendDescriptor* x = xDesc.get();
HipdnnBackendDescriptor* w = wDesc.get();
HipdnnBackendDescriptor* y = yDesc.get();
convDesc->setAttribute(HIPDNN_ATTR_OPERATION_CONVOLUTION_FORWARD_X,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(&x));
convDesc->setAttribute(HIPDNN_ATTR_OPERATION_CONVOLUTION_FORWARD_W,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(&w));
convDesc->setAttribute(HIPDNN_ATTR_OPERATION_CONVOLUTION_FORWARD_Y,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(&y));
convDesc->setAttribute(HIPDNN_ATTR_CONVOLUTION_PRE_PADDINGS,
HIPDNN_TYPE_INT64,
static_cast<int64_t>(prePadding.size()),
prePadding.data());
convDesc->setAttribute(HIPDNN_ATTR_CONVOLUTION_POST_PADDINGS,
HIPDNN_TYPE_INT64,
static_cast<int64_t>(postPadding.size()),
postPadding.data());
convDesc->setAttribute(HIPDNN_ATTR_CONVOLUTION_FILTER_STRIDES,
HIPDNN_TYPE_INT64,
static_cast<int64_t>(stride.size()),
stride.data());
convDesc->setAttribute(HIPDNN_ATTR_CONVOLUTION_DILATIONS,
HIPDNN_TYPE_INT64,
static_cast<int64_t>(dilation.size()),
dilation.data());
convDesc->setAttribute(
HIPDNN_ATTR_CONVOLUTION_COMP_TYPE, HIPDNN_TYPE_DATA_TYPE, 1, &computeDataType);
hipdnnConvolutionMode_t convMode = HIPDNN_CROSS_CORRELATION;
convDesc->setAttribute(
HIPDNN_ATTR_CONVOLUTION_CONV_MODE, HIPDNN_TYPE_CONVOLUTION_MODE, 1, &convMode);
convDesc->finalize();
// Build graph via GraphDescriptor
auto graphWrapper = createDescriptor<GraphDescriptor>();
auto graphDesc = graphWrapper->asDescriptor<GraphDescriptor>();
hipdnnHandle_t handle = &_mockHandle;
graphDesc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_HANDLE,
HIPDNN_TYPE_HANDLE,
1,
static_cast<const void*>(&handle));
std::array<HipdnnBackendDescriptor*, 1> ops = {convWrapper.get()};
graphDesc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(ops.data()));
graphDesc->finalize();
auto serialized = graphDesc->getSerializedGraph();
return UnPackGraph(serialized.ptr);
}
void verifyEquivalence(const ConvEquivalenceParams& p)
{
// Build via FlatBuffer path
auto sdkTensorDt = hipdnn_backend::toSdkDataType(p.tensorDataType);
auto sdkComputeDt = hipdnn_backend::toSdkDataType(p.computeDataType);
TensorAttributesT xTensor;
xTensor.uid = p.xUid;
xTensor.dims = p.xDims;
xTensor.strides = p.xStrides;
xTensor.data_type = sdkTensorDt;
TensorAttributesT wTensor;
wTensor.uid = p.wUid;
wTensor.dims = p.wDims;
wTensor.strides = p.wStrides;
wTensor.data_type = sdkTensorDt;
TensorAttributesT yTensor;
yTensor.uid = p.yUid;
yTensor.dims = p.yDims;
yTensor.strides = p.yStrides;
yTensor.data_type = sdkTensorDt;
ConvolutionFwdAttributesT convAttrs;
convAttrs.x_tensor_uid = p.xUid;
convAttrs.w_tensor_uid = p.wUid;
convAttrs.y_tensor_uid = p.yUid;
convAttrs.pre_padding = p.prePadding;
convAttrs.post_padding = p.postPadding;
convAttrs.stride = p.stride;
convAttrs.dilation = p.dilation;
auto flatbufferBuffer
= buildGraphViaFlatBuffer(xTensor, wTensor, yTensor, convAttrs, sdkComputeDt);
auto serializedGraphT = UnPackGraph(flatbufferBuffer.data());
// Build via descriptor path
auto descriptorGraphT = buildGraphViaDescriptors(p.xUid,
p.wUid,
p.yUid,
p.xDims,
p.xStrides,
p.wDims,
p.wStrides,
p.yDims,
p.yStrides,
p.prePadding,
p.postPadding,
p.stride,
p.dilation,
p.tensorDataType,
p.computeDataType);
// Verify structural equivalence
ASSERT_EQ(serializedGraphT->tensors.size(), descriptorGraphT->tensors.size());
ASSERT_EQ(serializedGraphT->nodes.size(), descriptorGraphT->nodes.size());
// Compare tensors (order may differ, so compare by UID)
std::map<int64_t, const TensorAttributesT*> fbTensors;
for(const auto& t : serializedGraphT->tensors)
{
fbTensors[t->uid] = t.get();
}
for(const auto& descTensor : descriptorGraphT->tensors)
{
auto it = fbTensors.find(descTensor->uid);
ASSERT_NE(it, fbTensors.end())
<< "Tensor UID " << descTensor->uid << " not found in FlatBuffer graph";
const auto* fbTensor = it->second;
EXPECT_EQ(fbTensor->uid, descTensor->uid);
EXPECT_EQ(fbTensor->dims, descTensor->dims);
EXPECT_EQ(fbTensor->strides, descTensor->strides);
EXPECT_EQ(fbTensor->data_type, descTensor->data_type);
}
// Compare nodes
ASSERT_EQ(serializedGraphT->nodes.size(), 1);
ASSERT_EQ(descriptorGraphT->nodes.size(), 1);
const auto& fbNode = serializedGraphT->nodes[0];
const auto& descNode = descriptorGraphT->nodes[0];
EXPECT_EQ(fbNode->compute_data_type, descNode->compute_data_type);
EXPECT_EQ(fbNode->attributes.type, descNode->attributes.type);
EXPECT_EQ(fbNode->attributes.type, NodeAttributes::ConvolutionFwdAttributes);
const auto* fbConv = fbNode->attributes.AsConvolutionFwdAttributes();
const auto* descConv = descNode->attributes.AsConvolutionFwdAttributes();
ASSERT_NE(fbConv, nullptr);
ASSERT_NE(descConv, nullptr);
EXPECT_EQ(fbConv->x_tensor_uid, descConv->x_tensor_uid);
EXPECT_EQ(fbConv->w_tensor_uid, descConv->w_tensor_uid);
EXPECT_EQ(fbConv->y_tensor_uid, descConv->y_tensor_uid);
EXPECT_EQ(fbConv->pre_padding, descConv->pre_padding);
EXPECT_EQ(fbConv->post_padding, descConv->post_padding);
EXPECT_EQ(fbConv->stride, descConv->stride);
EXPECT_EQ(fbConv->dilation, descConv->dilation);
}
};
TEST_P(TestGraphDescriptorEquivalence, ConvOpEquivalence)
{
verifyEquivalence(GetParam());
}
namespace
{
std::string convEquivalenceParamName(const ::testing::TestParamInfo<ConvEquivalenceParams>& info)
{
return info.param.name;
}
} // namespace
INSTANTIATE_TEST_SUITE_P(ConvOps,
TestGraphDescriptorEquivalence,
::testing::Values(ConvEquivalenceParams{"SingleConvOp",
K_FPROP_TENSOR_X_UID,
K_FPROP_TENSOR_W_UID,
K_FPROP_TENSOR_Y_UID,
toVec(K_FPROP_TENSOR_X_DIMS),
toVec(K_FPROP_TENSOR_X_STRIDES),
toVec(K_FPROP_TENSOR_W_DIMS),
toVec(K_FPROP_TENSOR_W_STRIDES),
toVec(K_FPROP_TENSOR_Y_DIMS),
toVec(K_FPROP_TENSOR_Y_STRIDES),
toVec(K_FPROP_CONV_PADDING),
toVec(K_FPROP_CONV_PADDING),
toVec(K_FPROP_CONV_STRIDE),
toVec(K_FPROP_CONV_DILATION),
HIPDNN_DATA_FLOAT,
HIPDNN_DATA_FLOAT},
ConvEquivalenceParams{"HalfPrecision",
10,
20,
30,
{2, 64, 56, 56},
{200704, 3136, 56, 1},
{128, 64, 3, 3},
{576, 9, 3, 1},
{2, 128, 56, 56},
{401408, 3136, 56, 1},
{1, 1},
{1, 1},
{1, 1},
{1, 1},
HIPDNN_DATA_HALF,
HIPDNN_DATA_HALF},
ConvEquivalenceParams{"NonUnitStrideAndDilation",
100,
200,
300,
{1, 256, 28, 28},
{200704, 784, 28, 1},
{512, 256, 3, 3},
{2304, 9, 3, 1},
{1, 512, 14, 14},
{100352, 196, 14, 1},
{2, 2},
{2, 2},
{2, 2},
{2, 2},
HIPDNN_DATA_FLOAT,
HIPDNN_DATA_FLOAT}),
convEquivalenceParamName);
// =============================================================================
// Graph-Level Attribute Tests
// =============================================================================
TEST_F(TestGraphDescriptorOps, GraphLevelDataTypesPreserved)
{
auto conv = createDefaultConvOp();
auto desc = getDescriptor();
setHandle();
std::array<HipdnnBackendDescriptor*, 1> ops = {conv.convOp.get()};
desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(ops.data()));
// Set graph-level data types before finalize
auto computeDt = HIPDNN_DATA_HALF;
desc->setAttribute(
HIPDNN_ATTR_OPERATIONGRAPH_COMPUTE_DATA_TYPE_EXT, HIPDNN_TYPE_DATA_TYPE, 1, &computeDt);
auto intermediateDt = HIPDNN_DATA_BFLOAT16;
desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_INTERMEDIATE_DATA_TYPE_EXT,
HIPDNN_TYPE_DATA_TYPE,
1,
&intermediateDt);
auto ioDt = HIPDNN_DATA_FLOAT;
desc->setAttribute(
HIPDNN_ATTR_OPERATIONGRAPH_IO_DATA_TYPE_EXT, HIPDNN_TYPE_DATA_TYPE, 1, &ioDt);
desc->finalize();
auto serialized = desc->getSerializedGraph();
auto graphT = UnPackGraph(serialized.ptr);
EXPECT_EQ(graphT->compute_data_type, DataType::HALF);
EXPECT_EQ(graphT->intermediate_data_type, DataType::BFLOAT16);
EXPECT_EQ(graphT->io_data_type, DataType::FLOAT);
}
TEST_F(TestGraphDescriptorOps, PreferredEngineIdPreserved)
{
auto conv = createDefaultConvOp();
auto desc = getDescriptor();
setHandle();
std::array<HipdnnBackendDescriptor*, 1> ops = {conv.convOp.get()};
desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(ops.data()));
int64_t engineId = 42;
desc->setAttribute(
HIPDNN_ATTR_OPERATIONGRAPH_PREFERRED_ENGINE_ID_EXT, HIPDNN_TYPE_INT64, 1, &engineId);
desc->finalize();
auto serialized = desc->getSerializedGraph();
auto graphT = UnPackGraph(serialized.ptr);
EXPECT_EQ(graphT->preferred_engine_id, 42);
}
TEST_F(TestGraphDescriptorOps, GraphLevelDataTypesDefaultToUnset)
{
auto conv = createDefaultConvOp();
auto desc = getDescriptor();
setHandle();
std::array<HipdnnBackendDescriptor*, 1> ops = {conv.convOp.get()};
desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(ops.data()));
// Finalize without setting any graph-level data types
desc->finalize();
auto serialized = desc->getSerializedGraph();
auto graphT = UnPackGraph(serialized.ptr);
EXPECT_EQ(graphT->compute_data_type, DataType::UNSET);
EXPECT_EQ(graphT->intermediate_data_type, DataType::UNSET);
EXPECT_EQ(graphT->io_data_type, DataType::UNSET);
}
TEST_F(TestGraphDescriptorOps, SharedTensorDifferentPositions)
{
// Same tensor used as Y in first op and X in second op
auto graphWrapper = createDescriptor<GraphDescriptor>();
auto graphDesc = graphWrapper->asDescriptor<GraphDescriptor>();
auto handle = reinterpret_cast<hipdnnHandle_t>(&_mockHandle);
graphDesc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_HANDLE,
HIPDNN_TYPE_HANDLE,
1,
static_cast<const void*>(&handle));
// Create tensors (Y tensor shared between ops)
auto xDesc1 = createFinalizedTensor(K_FPROP_TENSOR_X_UID);
auto wDesc1 = createFinalizedTensor(
K_FPROP_TENSOR_W_UID, toVec(K_FPROP_TENSOR_W_DIMS), toVec(K_FPROP_TENSOR_W_STRIDES));
auto sharedTensor = createFinalizedTensor(
K_FPROP_TENSOR_Y_UID, toVec(K_FPROP_TENSOR_Y_DIMS), toVec(K_FPROP_TENSOR_Y_STRIDES));
// 64->64 channel weights unique to this test (not in the shared constant set)
auto wDesc2 = createFinalizedTensor(K_SHARED_TENSOR_W_UID, {64, 64, 3, 3}, {576, 9, 3, 1});
auto yDesc2 = createFinalizedTensor(
K_SHARED_TENSOR_Y_UID, toVec(K_FPROP_TENSOR_Y_DIMS), toVec(K_FPROP_TENSOR_Y_STRIDES));
// Op1: x=1, w=2, y=3
auto op1 = createFinalizedConvOp(xDesc1.get(), wDesc1.get(), sharedTensor.get());
// Op2: x=3, w=4, y=5 (shares tensor 3 with op1 in a different position)
auto op2 = createFinalizedConvOp(sharedTensor.get(), wDesc2.get(), yDesc2.get());
std::array<HipdnnBackendDescriptor*, 2> opDescs = {op1.get(), op2.get()};
graphDesc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
2,
static_cast<const void*>(opDescs.data()));
graphDesc->finalize();
auto serialized = graphDesc->getSerializedGraph();
ASSERT_NE(GetGraph(serialized.ptr), nullptr);
auto graphT = UnPackGraph(serialized.ptr);
// 5 unique tensors, shared tensor is deduplicated
ASSERT_EQ(graphT->tensors.size(), 5);
ASSERT_EQ(graphT->nodes.size(), 2);
// Verify each tensor's fields
verifyTensor(findTensorByUid(*graphT, K_FPROP_TENSOR_X_UID),
K_FPROP_TENSOR_X_UID,
toVec(K_FPROP_TENSOR_X_DIMS),
toVec(K_FPROP_TENSOR_X_STRIDES),
DataType::FLOAT);
verifyTensor(findTensorByUid(*graphT, K_FPROP_TENSOR_W_UID),
K_FPROP_TENSOR_W_UID,
toVec(K_FPROP_TENSOR_W_DIMS),
toVec(K_FPROP_TENSOR_W_STRIDES),
DataType::FLOAT);
verifyTensor(findTensorByUid(*graphT, K_FPROP_TENSOR_Y_UID),
K_FPROP_TENSOR_Y_UID,
toVec(K_FPROP_TENSOR_Y_DIMS),
toVec(K_FPROP_TENSOR_Y_STRIDES),
DataType::FLOAT);
verifyTensor(findTensorByUid(*graphT, K_SHARED_TENSOR_W_UID),
K_SHARED_TENSOR_W_UID,
{64, 64, 3, 3},
{576, 9, 3, 1},
DataType::FLOAT);
verifyTensor(findTensorByUid(*graphT, K_SHARED_TENSOR_Y_UID),
K_SHARED_TENSOR_Y_UID,
toVec(K_FPROP_TENSOR_Y_DIMS),
toVec(K_FPROP_TENSOR_Y_STRIDES),
DataType::FLOAT);
// Verify first node: x=1, w=2, y=3 (shared tensor used as Y here)
verifyConvFwdNode(*graphT->nodes[0],
DataType::FLOAT,
K_FPROP_TENSOR_X_UID,
K_FPROP_TENSOR_W_UID,
K_FPROP_TENSOR_Y_UID,
toVec(K_FPROP_CONV_PADDING),
toVec(K_FPROP_CONV_PADDING),
toVec(K_FPROP_CONV_STRIDE),
toVec(K_FPROP_CONV_DILATION));
// Verify second node: x=3, w=4, y=5 (shared tensor reused as X here)
verifyConvFwdNode(*graphT->nodes[1],
DataType::FLOAT,
K_FPROP_TENSOR_Y_UID,
K_SHARED_TENSOR_W_UID,
K_SHARED_TENSOR_Y_UID,
toVec(K_FPROP_CONV_PADDING),
toVec(K_FPROP_CONV_PADDING),
toVec(K_FPROP_CONV_STRIDE),
toVec(K_FPROP_CONV_DILATION));
}
TEST_F(TestGraphDescriptorOps, FinalizeFailsDuplicateTensorUidDifferentDescriptors)
{
// Create two distinct tensor descriptors with the same UID
auto xDesc1 = createFinalizedTensor(K_FPROP_TENSOR_X_UID);
auto xDesc2 = createFinalizedTensor(K_FPROP_TENSOR_X_UID);
auto wDesc = createFinalizedTensor(
K_FPROP_TENSOR_W_UID, toVec(K_FPROP_TENSOR_W_DIMS), toVec(K_FPROP_TENSOR_W_STRIDES));
auto yDesc = createFinalizedTensor(
K_FPROP_TENSOR_Y_UID, toVec(K_FPROP_TENSOR_Y_DIMS), toVec(K_FPROP_TENSOR_Y_STRIDES));
// Op1 uses xDesc1, Op2 uses xDesc2 (different object, same UID)
auto op1 = createFinalizedConvOp(xDesc1.get(), wDesc.get(), yDesc.get());
auto xDesc2a = createFinalizedTensor(
K_FPROP_TENSOR_X2_UID, toVec(K_FPROP_TENSOR_X2_DIMS), toVec(K_FPROP_TENSOR_X2_STRIDES));
auto wDesc2 = createFinalizedTensor(
K_FPROP_TENSOR_W2_UID, toVec(K_FPROP_TENSOR_W2_DIMS), toVec(K_FPROP_TENSOR_W2_STRIDES));
// Use xDesc2 (same UID as xDesc1, different descriptor object) as Y
auto op2 = createFinalizedConvOp(xDesc2a.get(), wDesc2.get(), xDesc2.get());
auto desc = getDescriptor();
setHandle();
std::array<HipdnnBackendDescriptor*, 2> ops = {op1.get(), op2.get()};
desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
2,
static_cast<const void*>(ops.data()));
ASSERT_THROW_HIPDNN_STATUS(desc->finalize(), HIPDNN_STATUS_BAD_PARAM);
}
TEST_F(TestGraphDescriptorOps, SetOperationsRejectsNonOperationDescriptor)
{
auto graphWrapper = createDescriptor<GraphDescriptor>();
auto graphDesc = graphWrapper->asDescriptor<GraphDescriptor>();
// Set handle first
auto handle = reinterpret_cast<hipdnnHandle_t>(&_mockHandle);
graphDesc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_HANDLE,
HIPDNN_TYPE_HANDLE,
1,
static_cast<const void*>(&handle));
// Create a tensor descriptor (does NOT implement IGraphOperation)
auto tensorWrapper = createFinalizedTensor(99);
HipdnnBackendDescriptor* tensorDescPtr = tensorWrapper.get();
// Attempting to set a non-operation descriptor as an operation should throw
ASSERT_THROW_HIPDNN_STATUS(graphDesc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(&tensorDescPtr)),
HIPDNN_STATUS_NOT_SUPPORTED);
}
// =============================================================================
// GetAttribute Tests
// =============================================================================
TEST_F(TestGraphDescriptorOps, GetAttributeReturnsOperationCount)
{
auto conv = createDefaultConvOp();
auto desc = getDescriptor();
setHandle();
std::array<HipdnnBackendDescriptor*, 1> ops = {conv.convOp.get()};
desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(ops.data()));
desc->finalize();
// Query operation count by passing nullptr arrayOfElements
int64_t elementCount = 0;
ASSERT_NO_THROW(desc->getAttribute(
HIPDNN_ATTR_OPERATIONGRAPH_OPS, HIPDNN_TYPE_BACKEND_DESCRIPTOR, 0, &elementCount, nullptr));
EXPECT_EQ(elementCount, 1);
}
TEST_F(TestGraphDescriptorOps, GetAttributeReturnsOperationCountMultiple)
{
auto conv1 = createDefaultConvOp();
auto xDesc2 = createFinalizedTensor(
K_FPROP_TENSOR_X2_UID, toVec(K_FPROP_TENSOR_X2_DIMS), toVec(K_FPROP_TENSOR_X2_STRIDES));
auto wDesc2 = createFinalizedTensor(
K_FPROP_TENSOR_W2_UID, toVec(K_FPROP_TENSOR_W2_DIMS), toVec(K_FPROP_TENSOR_W2_STRIDES));
auto yDesc2 = createFinalizedTensor(
K_FPROP_TENSOR_Y2_UID, toVec(K_FPROP_TENSOR_Y2_DIMS), toVec(K_FPROP_TENSOR_Y2_STRIDES));
auto convOp2 = createFinalizedConvOp(xDesc2.get(), wDesc2.get(), yDesc2.get());
auto desc = getDescriptor();
setHandle();
std::array<HipdnnBackendDescriptor*, 2> ops = {conv1.convOp.get(), convOp2.get()};
desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
2,
static_cast<const void*>(ops.data()));
desc->finalize();
int64_t elementCount = 0;
ASSERT_NO_THROW(desc->getAttribute(
HIPDNN_ATTR_OPERATIONGRAPH_OPS, HIPDNN_TYPE_BACKEND_DESCRIPTOR, 0, &elementCount, nullptr));
EXPECT_EQ(elementCount, 2);
}
TEST_F(TestGraphDescriptorOps, GetAttributeReturnsOperations)
{
auto conv = createDefaultConvOp();
auto desc = getDescriptor();
setHandle();
std::array<HipdnnBackendDescriptor*, 1> ops = {conv.convOp.get()};
desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(ops.data()));
desc->finalize();
// Query the actual operations
int64_t elementCount = 0;
std::array<HipdnnBackendDescriptor*, 1> returnedOps = {nullptr};
ASSERT_NO_THROW(desc->getAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
&elementCount,
static_cast<void*>(returnedOps.data())));
EXPECT_EQ(elementCount, 1);
ASSERT_NE(returnedOps[0], nullptr);
// Verify the returned operation is a conv forward operation
int64_t opTypeCount = 0;
hipdnnOperationType_ext_t opType = HIPDNN_OPERATION_TYPE_NOT_SET_EXT;
ASSERT_NO_THROW(returnedOps[0]->getAttribute(
HIPDNN_ATTR_OPERATION_TYPE_EXT, HIPDNN_TYPE_OPERATION_TYPE_EXT, 1, &opTypeCount, &opType));
EXPECT_EQ(opType, HIPDNN_OPERATION_TYPE_CONVOLUTION_FORWARD_EXT);
auto returnedOp0 = std::unique_ptr<HipdnnBackendDescriptor>(returnedOps[0]);
EXPECT_TRUE(returnedOp0->isFinalized());
}
TEST_F(TestGraphDescriptorOps, GetAttributeReturnsComputeDataType)
{
auto conv = createDefaultConvOp();
auto desc = getDescriptor();
setHandle();
std::array<HipdnnBackendDescriptor*, 1> ops = {conv.convOp.get()};
desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(ops.data()));
auto computeDt = HIPDNN_DATA_HALF;
desc->setAttribute(
HIPDNN_ATTR_OPERATIONGRAPH_COMPUTE_DATA_TYPE_EXT, HIPDNN_TYPE_DATA_TYPE, 1, &computeDt);
desc->finalize();
// Query compute data type
int64_t elementCount = 0;
hipdnnDataType_t queriedType = HIPDNN_DATA_FLOAT;
ASSERT_NO_THROW(desc->getAttribute(HIPDNN_ATTR_OPERATIONGRAPH_COMPUTE_DATA_TYPE_EXT,
HIPDNN_TYPE_DATA_TYPE,
1,
&elementCount,
&queriedType));
EXPECT_EQ(elementCount, 1);
EXPECT_EQ(queriedType, HIPDNN_DATA_HALF);
}
TEST_F(TestGraphDescriptorOps, GetAttributeReturnsIntermediateDataType)
{
auto conv = createDefaultConvOp();
auto desc = getDescriptor();
setHandle();
std::array<HipdnnBackendDescriptor*, 1> ops = {conv.convOp.get()};
desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(ops.data()));
auto intermediateDt = HIPDNN_DATA_BFLOAT16;
desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_INTERMEDIATE_DATA_TYPE_EXT,
HIPDNN_TYPE_DATA_TYPE,
1,
&intermediateDt);
desc->finalize();
// Query intermediate data type
int64_t elementCount = 0;
hipdnnDataType_t queriedType = HIPDNN_DATA_FLOAT;
ASSERT_NO_THROW(desc->getAttribute(HIPDNN_ATTR_OPERATIONGRAPH_INTERMEDIATE_DATA_TYPE_EXT,
HIPDNN_TYPE_DATA_TYPE,
1,
&elementCount,
&queriedType));
EXPECT_EQ(elementCount, 1);
EXPECT_EQ(queriedType, HIPDNN_DATA_BFLOAT16);
}
TEST_F(TestGraphDescriptorOps, GetAttributeReturnsIoDataType)
{
auto conv = createDefaultConvOp();
auto desc = getDescriptor();
setHandle();
std::array<HipdnnBackendDescriptor*, 1> ops = {conv.convOp.get()};
desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(ops.data()));
auto ioDt = HIPDNN_DATA_DOUBLE;
desc->setAttribute(
HIPDNN_ATTR_OPERATIONGRAPH_IO_DATA_TYPE_EXT, HIPDNN_TYPE_DATA_TYPE, 1, &ioDt);
desc->finalize();
// Query IO data type
int64_t elementCount = 0;
hipdnnDataType_t queriedType = HIPDNN_DATA_FLOAT;
ASSERT_NO_THROW(desc->getAttribute(HIPDNN_ATTR_OPERATIONGRAPH_IO_DATA_TYPE_EXT,
HIPDNN_TYPE_DATA_TYPE,
1,
&elementCount,
&queriedType));
EXPECT_EQ(elementCount, 1);
EXPECT_EQ(queriedType, HIPDNN_DATA_DOUBLE);
}
TEST_F(TestGraphDescriptorOps, GetAttributeReturnsPreferredEngineId)
{
auto conv = createDefaultConvOp();
auto desc = getDescriptor();
setHandle();
std::array<HipdnnBackendDescriptor*, 1> ops = {conv.convOp.get()};
desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(ops.data()));
int64_t engineId = 42;
desc->setAttribute(
HIPDNN_ATTR_OPERATIONGRAPH_PREFERRED_ENGINE_ID_EXT, HIPDNN_TYPE_INT64, 1, &engineId);
desc->finalize();
// Query preferred engine ID
int64_t elementCount = 0;
int64_t queriedEngineId = 0;
ASSERT_NO_THROW(desc->getAttribute(HIPDNN_ATTR_OPERATIONGRAPH_PREFERRED_ENGINE_ID_EXT,
HIPDNN_TYPE_INT64,
1,
&elementCount,
&queriedEngineId));
EXPECT_EQ(elementCount, 1);
EXPECT_EQ(queriedEngineId, 42);
}
TEST_F(TestGraphDescriptorOps, GetAttributePreferredEngineIdCountWhenUnset)
{
auto conv = createDefaultConvOp();
auto desc = getDescriptor();
setHandle();
std::array<HipdnnBackendDescriptor*, 1> ops = {conv.convOp.get()};
desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(ops.data()));
// Do not set preferred engine ID
desc->finalize();
// Query count (should be 0 when not set)
int64_t elementCount = 99;
ASSERT_NO_THROW(desc->getAttribute(HIPDNN_ATTR_OPERATIONGRAPH_PREFERRED_ENGINE_ID_EXT,
HIPDNN_TYPE_INT64,
0,
&elementCount,
nullptr));
EXPECT_EQ(elementCount, 0);
}
TEST_F(TestGraphDescriptorOps, OperationsPreservedAfterFinalize)
{
auto conv = createDefaultConvOp();
auto desc = getDescriptor();
setHandle();
std::array<HipdnnBackendDescriptor*, 1> ops = {conv.convOp.get()};
desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(ops.data()));
desc->finalize();
// Verify operations are still accessible after finalize + buildGraphFromOperations
int64_t elementCount = 0;
std::array<HipdnnBackendDescriptor*, 1> returnedOps = {nullptr};
ASSERT_NO_THROW(desc->getAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
&elementCount,
static_cast<void*>(returnedOps.data())));
EXPECT_EQ(elementCount, 1);
ASSERT_NE(returnedOps[0], nullptr);
// Verify the returned operation is a conv forward operation
int64_t opTypeCount = 0;
hipdnnOperationType_ext_t opType = HIPDNN_OPERATION_TYPE_NOT_SET_EXT;
ASSERT_NO_THROW(returnedOps[0]->getAttribute(
HIPDNN_ATTR_OPERATION_TYPE_EXT, HIPDNN_TYPE_OPERATION_TYPE_EXT, 1, &opTypeCount, &opType));
EXPECT_EQ(opType, HIPDNN_OPERATION_TYPE_CONVOLUTION_FORWARD_EXT);
auto returnedOp0 = std::unique_ptr<HipdnnBackendDescriptor>(returnedOps[0]);
EXPECT_TRUE(returnedOp0->isFinalized());
}
TEST_F(TestGraphDescriptorOps, GetAttributeWrongTypeForDataType)
{
// Build a graph with conv op, set compute data type to HALF, finalize
auto conv = createDefaultConvOp(HIPDNN_DATA_HALF);
auto desc = getDescriptor();
setHandle();
std::array<HipdnnBackendDescriptor*, 1> ops = {conv.convOp.get()};
desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(ops.data()));
auto computeDt = HIPDNN_DATA_HALF;
desc->setAttribute(
HIPDNN_ATTR_OPERATIONGRAPH_COMPUTE_DATA_TYPE_EXT, HIPDNN_TYPE_DATA_TYPE, 1, &computeDt);
desc->finalize();
// Call getAttribute with HIPDNN_ATTR_OPERATIONGRAPH_COMPUTE_DATA_TYPE_EXT
// but pass HIPDNN_TYPE_INT64 instead of HIPDNN_TYPE_DATA_TYPE
int64_t elementCount = 0;
int64_t wrongTypeBuffer = 0;
ASSERT_THROW_HIPDNN_STATUS(desc->getAttribute(HIPDNN_ATTR_OPERATIONGRAPH_COMPUTE_DATA_TYPE_EXT,
HIPDNN_TYPE_INT64,
1,
&elementCount,
&wrongTypeBuffer),
HIPDNN_STATUS_BAD_PARAM);
}
TEST_F(TestGraphDescriptorOps, GetAttributeWrongTypeForPreferredEngineId)
{
// Build a graph with conv op, set preferred engine id, finalize
auto conv = createDefaultConvOp();
auto desc = getDescriptor();
setHandle();
std::array<HipdnnBackendDescriptor*, 1> ops = {conv.convOp.get()};
desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(ops.data()));
int64_t engineId = 42;
desc->setAttribute(
HIPDNN_ATTR_OPERATIONGRAPH_PREFERRED_ENGINE_ID_EXT, HIPDNN_TYPE_INT64, 1, &engineId);
desc->finalize();
// Call getAttribute with HIPDNN_ATTR_OPERATIONGRAPH_PREFERRED_ENGINE_ID_EXT
// but pass HIPDNN_TYPE_DATA_TYPE instead of HIPDNN_TYPE_INT64
int64_t elementCount = 0;
hipdnnDataType_t wrongTypeBuffer = HIPDNN_DATA_FLOAT;
ASSERT_THROW_HIPDNN_STATUS(
desc->getAttribute(HIPDNN_ATTR_OPERATIONGRAPH_PREFERRED_ENGINE_ID_EXT,
HIPDNN_TYPE_DATA_TYPE,
1,
&elementCount,
&wrongTypeBuffer),
HIPDNN_STATUS_BAD_PARAM);
}
TEST_F(TestGraphDescriptorOps, GetAttributeReturnsName)
{
auto conv = createDefaultConvOp();
auto desc = getDescriptor();
setHandle();
std::array<HipdnnBackendDescriptor*, 1> ops = {conv.convOp.get()};
desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(ops.data()));
const std::string graphName = "MyTestGraph";
desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_NAME_EXT,
HIPDNN_TYPE_CHAR,
static_cast<int64_t>(graphName.size()),
graphName.c_str());
desc->finalize();
// Query count (getString returns size+1 for null terminator)
int64_t elementCount = 0;
ASSERT_NO_THROW(desc->getAttribute(
HIPDNN_ATTR_OPERATIONGRAPH_NAME_EXT, HIPDNN_TYPE_CHAR, 0, &elementCount, nullptr));
EXPECT_EQ(elementCount, static_cast<int64_t>(graphName.size() + 1));
// Query value
std::vector<char> nameBuffer(static_cast<size_t>(elementCount));
int64_t actualCount = 0;
ASSERT_NO_THROW(desc->getAttribute(HIPDNN_ATTR_OPERATIONGRAPH_NAME_EXT,
HIPDNN_TYPE_CHAR,
elementCount,
&actualCount,
nameBuffer.data()));
EXPECT_STREQ(nameBuffer.data(), graphName.c_str());
}
TEST_F(TestGraphDescriptorOps, GetAttributeNameCountWhenUnset)
{
auto conv = createDefaultConvOp();
auto desc = getDescriptor();
setHandle();
std::array<HipdnnBackendDescriptor*, 1> ops = {conv.convOp.get()};
desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(ops.data()));
// Do not set name
desc->finalize();
// Query count (empty string = just null terminator)
int64_t elementCount = 99;
ASSERT_NO_THROW(desc->getAttribute(
HIPDNN_ATTR_OPERATIONGRAPH_NAME_EXT, HIPDNN_TYPE_CHAR, 0, &elementCount, nullptr));
EXPECT_EQ(elementCount, 1);
}
TEST_F(TestGraphDescriptorOps, AppendOpsAfterSerialization)
{
// Set a single op, build the serialized buffer, then append more ops
auto conv1 = createDefaultConvOp();
auto desc = getDescriptor();
setHandle();
std::array<HipdnnBackendDescriptor*, 1> ops1 = {conv1.convOp.get()};
ASSERT_NO_THROW(desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(ops1.data())));
// Build the serialized buffer explicitly
desc->buildSerializedGraph();
auto serialized1 = desc->getSerializedGraph();
ASSERT_NE(serialized1.ptr, nullptr);
ASSERT_GT(serialized1.size, 0UL);
// Append a second operation (should succeed and invalidate the cache)
auto xDesc2 = createFinalizedTensor(
K_ALT_TENSOR_X_UID, toVec(K_FPROP_TENSOR_X_DIMS), toVec(K_FPROP_TENSOR_X_STRIDES));
auto wDesc2 = createFinalizedTensor(
K_ALT_TENSOR_W_UID, toVec(K_FPROP_TENSOR_W_DIMS), toVec(K_FPROP_TENSOR_W_STRIDES));
auto yDesc2 = createFinalizedTensor(
K_ALT_TENSOR_Y_UID, toVec(K_FPROP_TENSOR_Y_DIMS), toVec(K_FPROP_TENSOR_Y_STRIDES));
auto convOp2 = createFinalizedConvOp(xDesc2.get(), wDesc2.get(), yDesc2.get());
std::array<HipdnnBackendDescriptor*, 1> ops2 = {convOp2.get()};
ASSERT_NO_THROW(desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(ops2.data())));
// Verify cache was invalidated by the append
ASSERT_THROW_HIPDNN_STATUS(desc->getSerializedGraph(), HIPDNN_STATUS_BAD_PARAM);
// Verify operations count is 2
int64_t elementCount = 0;
ASSERT_NO_THROW(desc->getAttribute(
HIPDNN_ATTR_OPERATIONGRAPH_OPS, HIPDNN_TYPE_BACKEND_DESCRIPTOR, 0, &elementCount, nullptr));
EXPECT_EQ(elementCount, 2);
// Verify re-serialization reflects both operations
desc->buildSerializedGraph();
auto serialized2 = desc->getSerializedGraph();
auto graphT = UnPackGraph(serialized2.ptr);
ASSERT_EQ(graphT->nodes.size(), 2);
ASSERT_EQ(graphT->tensors.size(), 6);
}
TEST_F(TestGraphDescriptorOps, GetAttributeWrongTypeForName)
{
auto conv = createDefaultConvOp();
auto desc = getDescriptor();
setHandle();
std::array<HipdnnBackendDescriptor*, 1> ops = {conv.convOp.get()};
desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_OPS,
HIPDNN_TYPE_BACKEND_DESCRIPTOR,
1,
static_cast<const void*>(ops.data()));
const std::string graphName = "MyTestGraph";
desc->setAttribute(HIPDNN_ATTR_OPERATIONGRAPH_NAME_EXT,
HIPDNN_TYPE_CHAR,
static_cast<int64_t>(graphName.size()),
graphName.c_str());
desc->finalize();
// Call getAttribute with HIPDNN_ATTR_OPERATIONGRAPH_NAME_EXT
// but pass HIPDNN_TYPE_INT64 instead of HIPDNN_TYPE_CHAR
int64_t elementCount = 0;
int64_t wrongTypeBuffer = 0;
ASSERT_THROW_HIPDNN_STATUS(desc->getAttribute(HIPDNN_ATTR_OPERATIONGRAPH_NAME_EXT,
HIPDNN_TYPE_INT64,
1,
&elementCount,
&wrongTypeBuffer),
HIPDNN_STATUS_BAD_PARAM);
}
Reference in new issue View Git Blame Copy Permalink