repository

Contents

Verbose Message Catalogue

The following catalogue lists verbose messages, explanations, and additional information for: primitive creation and dispatch checks for primitive implementations; and implementation failures that occur during engine/memory object creation.

Primitive Creation/Dispatching

VERBOSE MESSAGE	SUBSTRING	PRIMITIVE	EXPLANATION
Bad/Invalid Arguments
bad algorithm		all	Bad or invalid algorithm dnnl::algorithm selected for the current primitive implementation. The choice and availability of the algorithm depends on the specific implementation selected for the primitive. For example, oneDNN supports Winograd convolution only on GPU and AArch64 CPU systems.
bad propagation kind		all	Incorrect propagation kind dnnl::prop_kind ( forward_training , backward_weights , backward_data , etc.) selected for the current primitive implementation.
bad param <p>	p - initialization parameter for dnnl::primitive_desc	all	Invalid parameter passed for the initialization of the primitive descriptor. Example : For the group_normalization primitive, this message is displayed when the value passed for the groups parameter does not evenly divide the number of channels for the source tensor.
one of the mandatory arguments is nullptr		all	A NULL pointer argument exception for the primitive methods.
bad flags		all	Bad or unsupported flags specified for the initialization of the primitive descriptor dnnl::primitive_desc .
Unsupported Arguments/Parameters
unsupported isa		all	Primitive implementation does not support the current ISA. This typically results in the dispatching to the next best supporting implementation for the ISA.
unsupported datatype		all	Tensor data type is not supported for the current implementation of the primitive. Depending on the primitive, this may correspond to the source, weights, bias or destination tensors.
unsupported datatype combination		all	Datatype combination for source, weights, bias or destination tensors is not supported for the current primitive implementation.
unsupported attr		all	Bad or unsupported attributes dnnl::primitive_attr passed to the primitive descriptor for the current implementation. Since attributes are separately created from the corresponding primitive descriptor, the selected primitive implementation may not support the attribute configuration.
unsupported post-ops		all	Unsupported post-ops configuration dnnl::post_ops passed on to the primitive descriptor. Similar to dnnl::primitive_attr , the selected implementation may not support the postop configuration during primitve creation.
unsupported scales configuration		all	Unsupported scales configuration specified for the primitive attributes dnnl::primitive_attr .
unsupported zero-point configuration		all	Unsupported zero-point configuration specified for the primitive attributes dnnl::primitive_attr .
unsupported bias configuration		compute-based primitives	Unsupported bias data configuration specified for the descriptors.
unsupported sparse md configuration		matmul	Current primitive implementation does not support sparse data operations.
unsupported format tag		all	Unsupported format tag dnnl::memory::format_tag encountered during primitive operation.
unsupported format tag for <t>	t - tensor	all	Unsupported format tag dnnl::memory::format_tag for specified tensor during primitive operation.
unsupported format kind		all	Unsupported format kind dnnl::memory::format_kind encountered during primitive operation.
runtime dimension is not supported		all	Current implementation does not support processing runtime-specified shapes and strides using DNNL_RUNTIME_DIM_VAL .
Tensor Operations
tensor <t> has no elements	t - tensor	all	Empty tensor passed as data to the primitive. Depending on the primitive, this may correspond to the source, weights or destination tensors.
<t> has a bad number of dimensions <ndims>	t - tensor, ndims - number of tensor dimensions	all	Tensor data has bad or invalid number of dimensions for the current primitive operation. Example : The convolution primitive expects only 1D-, 2D- or 3D-spatial tensors for operations and prints this message for any other data with higher dimensions.
bad dimensions <t>:<axis>	t - tensor, axis - axis	all	Tensor <t> has an invalid dimension along the specified axis. Example : The concat primitive prints this message when the destination tensor dimension along the concatenated axis does not match the sum of the dimensions of the concatenated tensors.
dimension <t0>:<a0> is inconsistent with <t1>:<a1>	t0, t1 - tensors, a0, a1 - tensor axes	all	Tensors t0, t1 have inconsistent dimensions along axes a0 and a1 respectively. Example : This is encountered for the matmul primitive when the input matrices have mismatching dimensions.
tensors <t0> and <t1> have inconsistent number of dimensions	t0, t1 - tensors	all	Tensors t0, t1 have inconsistent dimensions for primitive operation.
tensors <t0> and <t1> have inconsistent datatypes	t0, t1 - tensors	all	Tensors t0, t1 have inconsistent data types for primitive operation.
Unsupported Combinations
sparse encoding is not supported on this isa		all	(self-explanatory)
datatype configuration not supported on this isa		all	(self-explanatory)
datatype and propagation kind mismatch		all	(self-explanatory)
inconsistent <t0> and <t1> mds	t0, t1 - tensors	all	Tensors t0, t1 have inconsistent memory descriptors dnnl::memory::desc for the primitive operation.
Implementation Heuristics/Features
unsupported feature for implementation: <msg>	msg - feature description	all	Current implementation is skipped because it does not support the specified feature for the primitive operation.
<msg> feature unavailable for device	msg - feature description	all	Current implementation is skipped because the selected device does not support the specified feature for primitive operation.
unsupported feature for padding: <msg>	msg - feature description	all	Current implementation is skipped because of a padding inconsistency or unsupported feature related to padding.
heuristic fail: <h>	h - implementation heuristic	all	Implementation skipped due to specified heuristic.
blocking heuristic fail: <h>	h - implementation heuristic	all	Current implementation is skipped because of the specified inconsistency or implementation heuristic related to blocking.
Primitive-Specific Messages
heuristic fail for 1x1 convolution: <h>	h - implementation heuristic	convolution	Implementation skipped due to specified heuristic related to 1x1 convolution.
<o> offsets do not fit into <dt> datatype	o - { input , output }, dt - data type	convolution	I/O dimension offsets do not fit into the specified datatype range for the kernel implementation.
failed shape restrictions		convolution , gnorm	Implementation skipped because the current data layout/shapes exceeds the range supported by the current implementation.
alpha and beta parameters are not properly set		eltwise	Alpha and beta parameters are not properly set for the elementwise algorithm.
large shapes fall back		gemm	Heuristic to skip current implementation for large tensor shapes for better performance.
only trivial strides are supported		gemm , rnn	Current implementation for the primitive does not process non-trivial stride values.
unsupported fpmath mode		matmul	Floating-point math mode is not supported by the current primitive implementation.
small shapes fall back		matmul	Heuristic to skip current implementation for small tensor shapes for better performance.
incompatible gemm format		matmul , ip	Specified GeMM format is incompatible with the current primitive implementation.
unsupported <t> tensor layout	t - tensor	reorder	The data layout for the source/destination tensor is not supported by the current implementation.
bad axis		softmax , shuffle	Bad or invalid axis specified for softmax/shuffle operation.
unsupported <d> architecture	d - dnnl::engine::kind	gemm	Unsupported architecture for specified device-type. Typically encountered when current GPU device does not support the primitive.
Miscellaneous
failed to create nested primitive <pm>	pm - dnnl::primitive	all	Descriptor initialization for the nested primitive implementation was unsuccessful.
failed to create <pm> descriptor	pm - dnnl::primitive , dnnl::memory	all	Descriptor initialization for the primitive or memory object was unsuccessful.
bad accumulation mode		all	Bad or invalid accumulation mode specified for primitive attribute dnnl::primitive_attr .
unsupported <t> md flag	t - tensor	all	Bad or unsupported flags specified for the memory descriptor dnnl::memory::desc .
problem is not mathematically consistent		all	(self-explanatory)
workspace mismatch between forward and backward primitive descriptors		all	(self-explanatory)
workspace initialization failed		all	Workspace descriptor initialization was unsuccessful during primitive creation.
invalid datatype for <t>	t - tensor	all	The data type for the tensor/data processed by the primitive is invalid. Example : This is encountered when an undefined data type data_type::undef is specified for the accumulator.
failed to run kernel deterministically		all	failed to run application in the deterministic mode .

Engine Creation

VERBOSE MESSAGE	SUBSTRING	ENGINE	EXPLANATION
bad engine kind		all	Invalid value for dnnl::engine::kind encountered during engine creation.
invalid <d> device in environment: index <i>	d - dnnl::engine::kind , i - device index	all	Device of type dnnl::engine::kind and index i is invalid for the current environment.
no <d> device is available	d - dnnl::engine::kind	all	No device of type dnnl::engine::kind was found during engine creation.
<n> <d> devices are available but <i> was queried	d - dnnl::engine::kind , n - number of d devices, i - queried device index	all	Queried index is out-of-range for device of type dnnl::engine::kind .
device not found in the given context		all	(self-explanatory)
unsupported <d> platform (expected <d0> got <d1>)	d - dnnl::engine::kind , d0 - queried platform, d1 - available platform	sycl , opencl	Unsupported device platform encountered during engine creation.
failed to create <d> engine with index <i>	d - dnnl::engine::kind , i - device index	all	Engine creation was unsuccessful for specified device index and kind.
unsupported <d> backend	d - dnnl::engine::kind	sycl	(self-explanatory)
profiling capabilities are not supported		all	Experimental profiling ( ONEDNN_EXPERIMENTAL_PROFILING ) is not enabled for the application.

Memory Creation and Related Operations

VERBOSE MESSAGE	EXPLANATION
bad arguments for memory descriptor	Bad or unsupported values passed to the memory descriptor dnnl::memory::desc during memory object creation.
invalid memory index	An out-of-range value encountered for memory handle during data mapping.
unsupported memory stride	Memory descriptor initialization failed due to unsupported value for memory strides.
scratchpad memory limit exceeded	Scratchpad space is exhausted during GEMM kernel initialization.
scratchpad initialization unsuccessful	(self-explanatory)

CPU ISA Hints Advanced Topics