hipTensor code for ROCm 7.0.1 did not change. The library was rebuilt for the updated ROCm 7.0.1 stack.

Added

• Added element-wise binary operation support.
• Added element-wise trinary operation support.
• Added support for new GPU target gfx950.
• Added dynamic unary and binary operator support for element-wise operations and permutation.
• Added a CMake check for f8 datatype availability.
• Added hiptensorDestroyOperationDescriptor to free all resources related to the provided descriptor.
• Added hiptensorOperationDescriptorSetAttribute to set attribute of a hiptensorOperationDescriptor_t object.
• Added hiptensorOperationDescriptorGetAttribute to retrieve an attribute of the provided hiptensorOperationDescriptor_t object.
• Added hiptensorCreatePlanPreference to allocate the hiptensorPlanPreference_t and enabled users to limit the applicable kernels for a given plan or operation.
• Added hiptensorDestroyPlanPreference to free all resources related to the provided preference.
• Added hiptensorPlanPreferenceSetAttribute to set attribute of a hiptensorPlanPreference_t object.
• Added hiptensorPlanGetAttribute to retrieve information about an already-created plan.
• Added hiptensorEstimateWorkspaceSize to determine the required workspaceSize for the given operation.
• Added hiptensorCreatePlan to allocate a hiptensorPlan_t object, select an appropriate kernel for a given operation and prepare a plan that encodes the execution.
• Added hiptensorDestroyPlan to free all resources related to the provided plan.

Changed

• Removed architecture support for gfx940 and gfx941.
• Generalized opaque buffer now for any descriptor.
• Replaced hipDataType with hiptensorDataType_t for all supported types, for example, HIP_R_32F to HIPTENSOR_R_32F.
• Replaced hiptensorComputeType_t with hiptensorComputeDescriptor_t for all supported types.
• Replaced hiptensorInitTensorDescriptor with hiptensorCreateTensorDescriptor.
• Changed handle type and API usage from *handle to handle.
• Replaced hiptensorContractionDescriptor_t with hipTensorOperationDescriptor_t.
• Replaced hiptensorInitContractionDescriptor with hiptensorCreateContraction.
• Replaced hiptensorContractionFind_t with hiptensorPlanPreference_t.
• Replaced hiptensorInitContractionFind with hiptensorCreatePlanPreference.
• Replaced hiptensorContractionGetWorkspaceSize with hiptensorEstimateWorkspaceSize.
• Replaced HIPTENSOR_WORKSPACE_RECOMMENDED with HIPTENSOR_WORKSPACE_DEFAULT.
• Replaced hiptensorContractionPlan_t with hiptensorPlan_t.
• Replaced hiptensorInitContractionPlan with hiptensorCreatePlan.
• Replaced hiptensorContraction with hiptensorContract.
• Replaced hiptensorPermutation with hiptensorPermute.
• Replaced hiptensorReduction with hiptensorReduce.
• Replaced hiptensorElementwiseBinary with hiptensorElementwiseBinaryExecute.
• Replaced hiptensorElementwiseTrinary with hiptensorElementwiseTrinaryExecute.
• Removed function hiptensorReductionGetWorkspaceSize.

hipTensor code for ROCm 6.4.4 did not change. The library was rebuilt for the updated ROCm 6.4.4 stack.

hipTensor code for ROCm 6.4.3 did not change. The library was rebuilt for the updated ROCm 6.4.3 stack.

hipTensor code for ROCm 6.4.2 did not change. The library was rebuilt for the updated ROCm 6.4.2 stack.

hipTensor code for ROCm 6.4.1 did not change. The library was rebuilt for the updated ROCm 6.4.1 stack.

ROCm release v6.1.5

Added

• Added benchmarking suites for contraction, permutation, and reduction. YAML files are categorized into bench and validation folders for organization
• Added emulation test suites for contraction, permutation, and reduction
• Support has been added for changing the default data layout using the HIPTENSOR_DEFAULT_STRIDES_COL_MAJOR environment variable

Changed

• Used GPU_TARGETS instead of AMDGPU_TARGETS in cmakelists.txt

Optimized

• Optimized the hyper-parameter selection algorithm for permutation

Resolved issues

• For CMake bug workaround, set CMAKE_NO_BUILTIN_CHRPATH when BUILD_OFFLOAD_COMPRESS is unset

hipTensor code for ROCm 6.3.3 did not change. The library was rebuilt for the updated ROCm 6.3.3 stack.

hipTensor code for ROCm 6.3.2 did not change. The library was rebuilt for the updated ROCm 6.3.2 stack.